ref: 50adfdf5baf4243eecacb94282141568ef4b666a
parent: db8e731b8d5a9e15e369e3921d4326e32b9efa10
author: Jingning Han <[email protected]>
date: Wed Jul 15 15:14:54 EDT 2015
Migrate loop filter functions from vp9/ to vpx_dsp/ The various tap loop filter operations are common functions across codec. This commit moves them along with SIMD optimizations to vpx_dsp folder. Change-Id: Ia5fa0b2e5289cdb98467502a549c380b9c60e92c
--- a/test/lpf_8_test.cc
+++ b/test/lpf_8_test.cc
@@ -19,7 +19,7 @@
#include "test/util.h"
#include "./vpx_config.h"
-#include "./vp9_rtcd.h"
+#include "./vpx_dsp_rtcd.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_loopfilter.h"
#include "vpx/vpx_integer.h"
@@ -665,11 +665,11 @@
&wrapper_vertical_16_c, 8, 1),
make_tuple(&wrapper_vertical_16_dual_neon,
&wrapper_vertical_16_dual_c, 8, 1),
+#endif // HAVE_NEON_ASM
make_tuple(&vp9_lpf_horizontal_8_neon,
&vp9_lpf_horizontal_8_c, 8, 1),
make_tuple(&vp9_lpf_vertical_8_neon,
&vp9_lpf_vertical_8_c, 8, 1),
-#endif // HAVE_NEON_ASM
make_tuple(&vp9_lpf_horizontal_4_neon,
&vp9_lpf_horizontal_4_c, 8, 1),
make_tuple(&vp9_lpf_vertical_4_neon,
--- a/vp9/common/arm/neon/vp9_loopfilter_16_neon.c
+++ /dev/null
@@ -1,179 +1,0 @@
-/*
- * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <arm_neon.h>
-
-#include "./vp9_rtcd.h"
-#include "./vpx_config.h"
-#include "vpx/vpx_integer.h"
-
-static INLINE void vp9_loop_filter_neon_16(
- uint8x16_t qblimit, // blimit
- uint8x16_t qlimit, // limit
- uint8x16_t qthresh, // thresh
- uint8x16_t q3, // p3
- uint8x16_t q4, // p2
- uint8x16_t q5, // p1
- uint8x16_t q6, // p0
- uint8x16_t q7, // q0
- uint8x16_t q8, // q1
- uint8x16_t q9, // q2
- uint8x16_t q10, // q3
- uint8x16_t *q5r, // p1
- uint8x16_t *q6r, // p0
- uint8x16_t *q7r, // q0
- uint8x16_t *q8r) { // q1- uint8x16_t q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
- int16x8_t q2s16, q11s16;
- uint16x8_t q4u16;
- int8x16_t q0s8, q1s8, q2s8, q11s8, q12s8, q13s8;
- int8x8_t d2s8, d3s8;
-
- q11u8 = vabdq_u8(q3, q4);
- q12u8 = vabdq_u8(q4, q5);
- q13u8 = vabdq_u8(q5, q6);
- q14u8 = vabdq_u8(q8, q7);
- q3 = vabdq_u8(q9, q8);
- q4 = vabdq_u8(q10, q9);
-
- q11u8 = vmaxq_u8(q11u8, q12u8);
- q12u8 = vmaxq_u8(q13u8, q14u8);
- q3 = vmaxq_u8(q3, q4);
- q15u8 = vmaxq_u8(q11u8, q12u8);
-
- q9 = vabdq_u8(q6, q7);
-
- // vp8_hevmask
- q13u8 = vcgtq_u8(q13u8, qthresh);
- q14u8 = vcgtq_u8(q14u8, qthresh);
- q15u8 = vmaxq_u8(q15u8, q3);
-
- q2u8 = vabdq_u8(q5, q8);
- q9 = vqaddq_u8(q9, q9);
-
- q15u8 = vcgeq_u8(qlimit, q15u8);
-
- // vp8_filter() function
- // convert to signed
- q10 = vdupq_n_u8(0x80);
- q8 = veorq_u8(q8, q10);
- q7 = veorq_u8(q7, q10);
- q6 = veorq_u8(q6, q10);
- q5 = veorq_u8(q5, q10);
-
- q2u8 = vshrq_n_u8(q2u8, 1);
- q9 = vqaddq_u8(q9, q2u8);
-
- q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
- vget_low_s8(vreinterpretq_s8_u8(q6)));
- q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
- vget_high_s8(vreinterpretq_s8_u8(q6)));
-
- q9 = vcgeq_u8(qblimit, q9);
-
- q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
- vreinterpretq_s8_u8(q8));
-
- q14u8 = vorrq_u8(q13u8, q14u8);
-
- q4u16 = vdupq_n_u16(3);
- q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
- q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
-
- q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
- q15u8 = vandq_u8(q15u8, q9);
-
- q1s8 = vreinterpretq_s8_u8(q1u8);
- q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
- q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
-
- q4 = vdupq_n_u8(3);
- q9 = vdupq_n_u8(4);
- // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
- d2s8 = vqmovn_s16(q2s16);
- d3s8 = vqmovn_s16(q11s16);
- q1s8 = vcombine_s8(d2s8, d3s8);
- q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
- q1s8 = vreinterpretq_s8_u8(q1u8);
-
- q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q4));
- q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
- q2s8 = vshrq_n_s8(q2s8, 3);
- q1s8 = vshrq_n_s8(q1s8, 3);
-
- q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
- q0s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
-
- q1s8 = vrshrq_n_s8(q1s8, 1);
- q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
-
- q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
- q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
-
- *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q10);
- *q7r = veorq_u8(vreinterpretq_u8_s8(q0s8), q10);
- *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q10);
- *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q10);
- return;
-}
-
-void vp9_lpf_horizontal_4_dual_neon(uint8_t *s, int p /* pitch */,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1) {- uint8x8_t dblimit0, dlimit0, dthresh0, dblimit1, dlimit1, dthresh1;
- uint8x16_t qblimit, qlimit, qthresh;
- uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8;
-
- dblimit0 = vld1_u8(blimit0);
- dlimit0 = vld1_u8(limit0);
- dthresh0 = vld1_u8(thresh0);
- dblimit1 = vld1_u8(blimit1);
- dlimit1 = vld1_u8(limit1);
- dthresh1 = vld1_u8(thresh1);
- qblimit = vcombine_u8(dblimit0, dblimit1);
- qlimit = vcombine_u8(dlimit0, dlimit1);
- qthresh = vcombine_u8(dthresh0, dthresh1);
-
- s -= (p << 2);
-
- q3u8 = vld1q_u8(s);
- s += p;
- q4u8 = vld1q_u8(s);
- s += p;
- q5u8 = vld1q_u8(s);
- s += p;
- q6u8 = vld1q_u8(s);
- s += p;
- q7u8 = vld1q_u8(s);
- s += p;
- q8u8 = vld1q_u8(s);
- s += p;
- q9u8 = vld1q_u8(s);
- s += p;
- q10u8 = vld1q_u8(s);
-
- vp9_loop_filter_neon_16(qblimit, qlimit, qthresh,
- q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8,
- &q5u8, &q6u8, &q7u8, &q8u8);
-
- s -= (p * 5);
- vst1q_u8(s, q5u8);
- s += p;
- vst1q_u8(s, q6u8);
- s += p;
- vst1q_u8(s, q7u8);
- s += p;
- vst1q_u8(s, q8u8);
- return;
-}
--- a/vp9/common/arm/neon/vp9_loopfilter_16_neon_asm.asm
+++ /dev/null
@@ -1,199 +1,0 @@
-;
-; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license
-; that can be found in the LICENSE file in the root of the source
-; tree. An additional intellectual property rights grant can be found
-; in the file PATENTS. All contributing project authors may
-; be found in the AUTHORS file in the root of the source tree.
-;
-
- EXPORT |vp9_lpf_horizontal_4_dual_neon|
- ARM
-
- AREA ||.text||, CODE, READONLY, ALIGN=2
-
-;void vp9_lpf_horizontal_4_dual_neon(uint8_t *s, int p,
-; const uint8_t *blimit0,
-; const uint8_t *limit0,
-; const uint8_t *thresh0,
-; const uint8_t *blimit1,
-; const uint8_t *limit1,
-; const uint8_t *thresh1)
-; r0 uint8_t *s,
-; r1 int p,
-; r2 const uint8_t *blimit0,
-; r3 const uint8_t *limit0,
-; sp const uint8_t *thresh0,
-; sp+4 const uint8_t *blimit1,
-; sp+8 const uint8_t *limit1,
-; sp+12 const uint8_t *thresh1,
-
-|vp9_lpf_horizontal_4_dual_neon| PROC
- push {lr}-
- ldr r12, [sp, #4] ; load thresh0
- vld1.8 {d0}, [r2] ; load blimit0 to first half q- vld1.8 {d2}, [r3] ; load limit0 to first half q-
- add r1, r1, r1 ; double pitch
- ldr r2, [sp, #8] ; load blimit1
-
- vld1.8 {d4}, [r12] ; load thresh0 to first half q-
- ldr r3, [sp, #12] ; load limit1
- ldr r12, [sp, #16] ; load thresh1
- vld1.8 {d1}, [r2] ; load blimit1 to 2nd half q-
- sub r2, r0, r1, lsl #1 ; s[-4 * p]
-
- vld1.8 {d3}, [r3] ; load limit1 to 2nd half q- vld1.8 {d5}, [r12] ; load thresh1 to 2nd half q-
- vpush {d8-d15} ; save neon registers-
- add r3, r2, r1, lsr #1 ; s[-3 * p]
-
- vld1.u8 {q3}, [r2@64], r1 ; p3- vld1.u8 {q4}, [r3@64], r1 ; p2- vld1.u8 {q5}, [r2@64], r1 ; p1- vld1.u8 {q6}, [r3@64], r1 ; p0- vld1.u8 {q7}, [r2@64], r1 ; q0- vld1.u8 {q8}, [r3@64], r1 ; q1- vld1.u8 {q9}, [r2@64] ; q2- vld1.u8 {q10}, [r3@64] ; q3-
- sub r2, r2, r1, lsl #1
- sub r3, r3, r1, lsl #1
-
- bl vp9_loop_filter_neon_16
-
- vst1.u8 {q5}, [r2@64], r1 ; store op1- vst1.u8 {q6}, [r3@64], r1 ; store op0- vst1.u8 {q7}, [r2@64], r1 ; store oq0- vst1.u8 {q8}, [r3@64], r1 ; store oq1-
- vpop {d8-d15} ; restore neon registers-
- pop {pc}- ENDP ; |vp9_lpf_horizontal_4_dual_neon|
-
-; void vp9_loop_filter_neon_16();
-; This is a helper function for the loopfilters. The invidual functions do the
-; necessary load, transpose (if necessary) and store. This function uses
-; registers d8-d15, so the calling function must save those registers.
-;
-; r0-r3, r12 PRESERVE
-; q0 blimit
-; q1 limit
-; q2 thresh
-; q3 p3
-; q4 p2
-; q5 p1
-; q6 p0
-; q7 q0
-; q8 q1
-; q9 q2
-; q10 q3
-;
-; Outputs:
-; q5 op1
-; q6 op0
-; q7 oq0
-; q8 oq1
-|vp9_loop_filter_neon_16| PROC
-
- ; filter_mask
- vabd.u8 q11, q3, q4 ; m1 = abs(p3 - p2)
- vabd.u8 q12, q4, q5 ; m2 = abs(p2 - p1)
- vabd.u8 q13, q5, q6 ; m3 = abs(p1 - p0)
- vabd.u8 q14, q8, q7 ; m4 = abs(q1 - q0)
- vabd.u8 q3, q9, q8 ; m5 = abs(q2 - q1)
- vabd.u8 q4, q10, q9 ; m6 = abs(q3 - q2)
-
- ; only compare the largest value to limit
- vmax.u8 q11, q11, q12 ; m7 = max(m1, m2)
- vmax.u8 q12, q13, q14 ; m8 = max(m3, m4)
-
- vabd.u8 q9, q6, q7 ; abs(p0 - q0)
-
- vmax.u8 q3, q3, q4 ; m9 = max(m5, m6)
-
- vmov.u8 q10, #0x80
-
- vmax.u8 q15, q11, q12 ; m10 = max(m7, m8)
-
- vcgt.u8 q13, q13, q2 ; (abs(p1 - p0) > thresh)*-1
- vcgt.u8 q14, q14, q2 ; (abs(q1 - q0) > thresh)*-1
- vmax.u8 q15, q15, q3 ; m11 = max(m10, m9)
-
- vabd.u8 q2, q5, q8 ; a = abs(p1 - q1)
- vqadd.u8 q9, q9, q9 ; b = abs(p0 - q0) * 2
-
- veor q7, q7, q10 ; qs0
-
- vcge.u8 q15, q1, q15 ; abs(m11) > limit
-
- vshr.u8 q2, q2, #1 ; a = a / 2
- veor q6, q6, q10 ; ps0
-
- veor q5, q5, q10 ; ps1
- vqadd.u8 q9, q9, q2 ; a = b + a
-
- veor q8, q8, q10 ; qs1
-
- vmov.u16 q4, #3
-
- vsubl.s8 q2, d14, d12 ; ( qs0 - ps0)
- vsubl.s8 q11, d15, d13
-
- vcge.u8 q9, q0, q9 ; a > blimit
-
- vqsub.s8 q1, q5, q8 ; filter = clamp(ps1-qs1)
- vorr q14, q13, q14 ; hev
-
- vmul.i16 q2, q2, q4 ; 3 * ( qs0 - ps0)
- vmul.i16 q11, q11, q4
-
- vand q1, q1, q14 ; filter &= hev
- vand q15, q15, q9 ; mask
-
- vmov.u8 q4, #3
-
- vaddw.s8 q2, q2, d2 ; filter + 3 * (qs0 - ps0)
- vaddw.s8 q11, q11, d3
-
- vmov.u8 q9, #4
-
- ; filter = clamp(filter + 3 * ( qs0 - ps0))
- vqmovn.s16 d2, q2
- vqmovn.s16 d3, q11
- vand q1, q1, q15 ; filter &= mask
-
- vqadd.s8 q2, q1, q4 ; filter2 = clamp(filter+3)
- vqadd.s8 q1, q1, q9 ; filter1 = clamp(filter+4)
- vshr.s8 q2, q2, #3 ; filter2 >>= 3
- vshr.s8 q1, q1, #3 ; filter1 >>= 3
-
-
- vqadd.s8 q11, q6, q2 ; u = clamp(ps0 + filter2)
- vqsub.s8 q0, q7, q1 ; u = clamp(qs0 - filter1)
-
- ; outer tap adjustments
- vrshr.s8 q1, q1, #1 ; filter = ++filter1 >> 1
-
- veor q7, q0, q10 ; *oq0 = u^0x80
-
- vbic q1, q1, q14 ; filter &= ~hev
-
- vqadd.s8 q13, q5, q1 ; u = clamp(ps1 + filter)
- vqsub.s8 q12, q8, q1 ; u = clamp(qs1 - filter)
-
- veor q6, q11, q10 ; *op0 = u^0x80
- veor q5, q13, q10 ; *op1 = u^0x80
- veor q8, q12, q10 ; *oq1 = u^0x80
-
- bx lr
- ENDP ; |vp9_loop_filter_neon_16|
-
- END
--- a/vp9/common/arm/neon/vp9_loopfilter_4_neon.c
+++ /dev/null
@@ -1,274 +1,0 @@
-/*
- * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <arm_neon.h>
-
-#include "./vpx_config.h"
-
-static INLINE void vp9_loop_filter_neon(
- uint8x8_t dblimit, // flimit
- uint8x8_t dlimit, // limit
- uint8x8_t dthresh, // thresh
- uint8x8_t d3u8, // p3
- uint8x8_t d4u8, // p2
- uint8x8_t d5u8, // p1
- uint8x8_t d6u8, // p0
- uint8x8_t d7u8, // q0
- uint8x8_t d16u8, // q1
- uint8x8_t d17u8, // q2
- uint8x8_t d18u8, // q3
- uint8x8_t *d4ru8, // p1
- uint8x8_t *d5ru8, // p0
- uint8x8_t *d6ru8, // q0
- uint8x8_t *d7ru8) { // q1- uint8x8_t d19u8, d20u8, d21u8, d22u8, d23u8, d27u8, d28u8;
- int16x8_t q12s16;
- int8x8_t d19s8, d20s8, d21s8, d26s8, d27s8, d28s8;
-
- d19u8 = vabd_u8(d3u8, d4u8);
- d20u8 = vabd_u8(d4u8, d5u8);
- d21u8 = vabd_u8(d5u8, d6u8);
- d22u8 = vabd_u8(d16u8, d7u8);
- d3u8 = vabd_u8(d17u8, d16u8);
- d4u8 = vabd_u8(d18u8, d17u8);
-
- d19u8 = vmax_u8(d19u8, d20u8);
- d20u8 = vmax_u8(d21u8, d22u8);
- d3u8 = vmax_u8(d3u8, d4u8);
- d23u8 = vmax_u8(d19u8, d20u8);
-
- d17u8 = vabd_u8(d6u8, d7u8);
-
- d21u8 = vcgt_u8(d21u8, dthresh);
- d22u8 = vcgt_u8(d22u8, dthresh);
- d23u8 = vmax_u8(d23u8, d3u8);
-
- d28u8 = vabd_u8(d5u8, d16u8);
- d17u8 = vqadd_u8(d17u8, d17u8);
-
- d23u8 = vcge_u8(dlimit, d23u8);
-
- d18u8 = vdup_n_u8(0x80);
- d5u8 = veor_u8(d5u8, d18u8);
- d6u8 = veor_u8(d6u8, d18u8);
- d7u8 = veor_u8(d7u8, d18u8);
- d16u8 = veor_u8(d16u8, d18u8);
-
- d28u8 = vshr_n_u8(d28u8, 1);
- d17u8 = vqadd_u8(d17u8, d28u8);
-
- d19u8 = vdup_n_u8(3);
-
- d28s8 = vsub_s8(vreinterpret_s8_u8(d7u8),
- vreinterpret_s8_u8(d6u8));
-
- d17u8 = vcge_u8(dblimit, d17u8);
-
- d27s8 = vqsub_s8(vreinterpret_s8_u8(d5u8),
- vreinterpret_s8_u8(d16u8));
-
- d22u8 = vorr_u8(d21u8, d22u8);
-
- q12s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d19u8));
-
- d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d22u8);
- d23u8 = vand_u8(d23u8, d17u8);
-
- q12s16 = vaddw_s8(q12s16, vreinterpret_s8_u8(d27u8));
-
- d17u8 = vdup_n_u8(4);
-
- d27s8 = vqmovn_s16(q12s16);
- d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d23u8);
- d27s8 = vreinterpret_s8_u8(d27u8);
-
- d28s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d19u8));
- d27s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d17u8));
- d28s8 = vshr_n_s8(d28s8, 3);
- d27s8 = vshr_n_s8(d27s8, 3);
-
- d19s8 = vqadd_s8(vreinterpret_s8_u8(d6u8), d28s8);
- d26s8 = vqsub_s8(vreinterpret_s8_u8(d7u8), d27s8);
-
- d27s8 = vrshr_n_s8(d27s8, 1);
- d27s8 = vbic_s8(d27s8, vreinterpret_s8_u8(d22u8));
-
- d21s8 = vqadd_s8(vreinterpret_s8_u8(d5u8), d27s8);
- d20s8 = vqsub_s8(vreinterpret_s8_u8(d16u8), d27s8);
-
- *d4ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d18u8);
- *d5ru8 = veor_u8(vreinterpret_u8_s8(d19s8), d18u8);
- *d6ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d18u8);
- *d7ru8 = veor_u8(vreinterpret_u8_s8(d20s8), d18u8);
- return;
-}
-
-void vp9_lpf_horizontal_4_neon(
- unsigned char *src,
- int pitch,
- unsigned char *blimit,
- unsigned char *limit,
- unsigned char *thresh,
- int count) {- int i;
- uint8_t *s, *psrc;
- uint8x8_t dblimit, dlimit, dthresh;
- uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
-
- if (count == 0) // end_vp9_lf_h_edge
- return;
-
- dblimit = vld1_u8(blimit);
- dlimit = vld1_u8(limit);
- dthresh = vld1_u8(thresh);
-
- psrc = src - (pitch << 2);
- for (i = 0; i < count; i++) {- s = psrc + i * 8;
-
- d3u8 = vld1_u8(s);
- s += pitch;
- d4u8 = vld1_u8(s);
- s += pitch;
- d5u8 = vld1_u8(s);
- s += pitch;
- d6u8 = vld1_u8(s);
- s += pitch;
- d7u8 = vld1_u8(s);
- s += pitch;
- d16u8 = vld1_u8(s);
- s += pitch;
- d17u8 = vld1_u8(s);
- s += pitch;
- d18u8 = vld1_u8(s);
-
- vp9_loop_filter_neon(dblimit, dlimit, dthresh,
- d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
- &d4u8, &d5u8, &d6u8, &d7u8);
-
- s -= (pitch * 5);
- vst1_u8(s, d4u8);
- s += pitch;
- vst1_u8(s, d5u8);
- s += pitch;
- vst1_u8(s, d6u8);
- s += pitch;
- vst1_u8(s, d7u8);
- }
- return;
-}
-
-void vp9_lpf_vertical_4_neon(
- unsigned char *src,
- int pitch,
- unsigned char *blimit,
- unsigned char *limit,
- unsigned char *thresh,
- int count) {- int i, pitch8;
- uint8_t *s;
- uint8x8_t dblimit, dlimit, dthresh;
- uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
- uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
- uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
- uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
- uint8x8x4_t d4Result;
-
- if (count == 0) // end_vp9_lf_h_edge
- return;
-
- dblimit = vld1_u8(blimit);
- dlimit = vld1_u8(limit);
- dthresh = vld1_u8(thresh);
-
- pitch8 = pitch * 8;
- for (i = 0; i < count; i++, src += pitch8) {- s = src - (i + 1) * 4;
-
- d3u8 = vld1_u8(s);
- s += pitch;
- d4u8 = vld1_u8(s);
- s += pitch;
- d5u8 = vld1_u8(s);
- s += pitch;
- d6u8 = vld1_u8(s);
- s += pitch;
- d7u8 = vld1_u8(s);
- s += pitch;
- d16u8 = vld1_u8(s);
- s += pitch;
- d17u8 = vld1_u8(s);
- s += pitch;
- d18u8 = vld1_u8(s);
-
- d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
- vreinterpret_u32_u8(d7u8));
- d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
- vreinterpret_u32_u8(d16u8));
- d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
- vreinterpret_u32_u8(d17u8));
- d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
- vreinterpret_u32_u8(d18u8));
-
- d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
- vreinterpret_u16_u32(d2tmp2.val[0]));
- d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
- vreinterpret_u16_u32(d2tmp3.val[0]));
- d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
- vreinterpret_u16_u32(d2tmp2.val[1]));
- d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
- vreinterpret_u16_u32(d2tmp3.val[1]));
-
- d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
- vreinterpret_u8_u16(d2tmp5.val[0]));
- d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
- vreinterpret_u8_u16(d2tmp5.val[1]));
- d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
- vreinterpret_u8_u16(d2tmp7.val[0]));
- d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
- vreinterpret_u8_u16(d2tmp7.val[1]));
-
- d3u8 = d2tmp8.val[0];
- d4u8 = d2tmp8.val[1];
- d5u8 = d2tmp9.val[0];
- d6u8 = d2tmp9.val[1];
- d7u8 = d2tmp10.val[0];
- d16u8 = d2tmp10.val[1];
- d17u8 = d2tmp11.val[0];
- d18u8 = d2tmp11.val[1];
-
- vp9_loop_filter_neon(dblimit, dlimit, dthresh,
- d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
- &d4u8, &d5u8, &d6u8, &d7u8);
-
- d4Result.val[0] = d4u8;
- d4Result.val[1] = d5u8;
- d4Result.val[2] = d6u8;
- d4Result.val[3] = d7u8;
-
- src -= 2;
- vst4_lane_u8(src, d4Result, 0);
- src += pitch;
- vst4_lane_u8(src, d4Result, 1);
- src += pitch;
- vst4_lane_u8(src, d4Result, 2);
- src += pitch;
- vst4_lane_u8(src, d4Result, 3);
- src += pitch;
- vst4_lane_u8(src, d4Result, 4);
- src += pitch;
- vst4_lane_u8(src, d4Result, 5);
- src += pitch;
- vst4_lane_u8(src, d4Result, 6);
- src += pitch;
- vst4_lane_u8(src, d4Result, 7);
- }
- return;
-}
--- a/vp9/common/arm/neon/vp9_loopfilter_4_neon_asm.asm
+++ /dev/null
@@ -1,277 +1,0 @@
-;
-; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license
-; that can be found in the LICENSE file in the root of the source
-; tree. An additional intellectual property rights grant can be found
-; in the file PATENTS. All contributing project authors may
-; be found in the AUTHORS file in the root of the source tree.
-;
-
- EXPORT |vp9_lpf_horizontal_4_neon|
- EXPORT |vp9_lpf_vertical_4_neon|
- ARM
-
- AREA ||.text||, CODE, READONLY, ALIGN=2
-
-; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
-; works on 16 iterations at a time.
-; TODO(fgalligan): See about removing the count code as this function is only
-; called with a count of 1.
-;
-; void vp9_lpf_horizontal_4_neon(uint8_t *s,
-; int p /* pitch */,
-; const uint8_t *blimit,
-; const uint8_t *limit,
-; const uint8_t *thresh,
-; int count)
-;
-; r0 uint8_t *s,
-; r1 int p, /* pitch */
-; r2 const uint8_t *blimit,
-; r3 const uint8_t *limit,
-; sp const uint8_t *thresh,
-; sp+4 int count
-|vp9_lpf_horizontal_4_neon| PROC
- push {lr}-
- vld1.8 {d0[]}, [r2] ; duplicate *blimit- ldr r12, [sp, #8] ; load count
- ldr r2, [sp, #4] ; load thresh
- add r1, r1, r1 ; double pitch
-
- cmp r12, #0
- beq end_vp9_lf_h_edge
-
- vld1.8 {d1[]}, [r3] ; duplicate *limit- vld1.8 {d2[]}, [r2] ; duplicate *thresh-
-count_lf_h_loop
- sub r2, r0, r1, lsl #1 ; move src pointer down by 4 lines
- add r3, r2, r1, lsr #1 ; set to 3 lines down
-
- vld1.u8 {d3}, [r2@64], r1 ; p3- vld1.u8 {d4}, [r3@64], r1 ; p2- vld1.u8 {d5}, [r2@64], r1 ; p1- vld1.u8 {d6}, [r3@64], r1 ; p0- vld1.u8 {d7}, [r2@64], r1 ; q0- vld1.u8 {d16}, [r3@64], r1 ; q1- vld1.u8 {d17}, [r2@64] ; q2- vld1.u8 {d18}, [r3@64] ; q3-
- sub r2, r2, r1, lsl #1
- sub r3, r3, r1, lsl #1
-
- bl vp9_loop_filter_neon
-
- vst1.u8 {d4}, [r2@64], r1 ; store op1- vst1.u8 {d5}, [r3@64], r1 ; store op0- vst1.u8 {d6}, [r2@64], r1 ; store oq0- vst1.u8 {d7}, [r3@64], r1 ; store oq1-
- add r0, r0, #8
- subs r12, r12, #1
- bne count_lf_h_loop
-
-end_vp9_lf_h_edge
- pop {pc}- ENDP ; |vp9_lpf_horizontal_4_neon|
-
-; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
-; works on 16 iterations at a time.
-; TODO(fgalligan): See about removing the count code as this function is only
-; called with a count of 1.
-;
-; void vp9_lpf_vertical_4_neon(uint8_t *s,
-; int p /* pitch */,
-; const uint8_t *blimit,
-; const uint8_t *limit,
-; const uint8_t *thresh,
-; int count)
-;
-; r0 uint8_t *s,
-; r1 int p, /* pitch */
-; r2 const uint8_t *blimit,
-; r3 const uint8_t *limit,
-; sp const uint8_t *thresh,
-; sp+4 int count
-|vp9_lpf_vertical_4_neon| PROC
- push {lr}-
- vld1.8 {d0[]}, [r2] ; duplicate *blimit- ldr r12, [sp, #8] ; load count
- vld1.8 {d1[]}, [r3] ; duplicate *limit-
- ldr r3, [sp, #4] ; load thresh
- sub r2, r0, #4 ; move s pointer down by 4 columns
- cmp r12, #0
- beq end_vp9_lf_v_edge
-
- vld1.8 {d2[]}, [r3] ; duplicate *thresh-
-count_lf_v_loop
- vld1.u8 {d3}, [r2], r1 ; load s data- vld1.u8 {d4}, [r2], r1- vld1.u8 {d5}, [r2], r1- vld1.u8 {d6}, [r2], r1- vld1.u8 {d7}, [r2], r1- vld1.u8 {d16}, [r2], r1- vld1.u8 {d17}, [r2], r1- vld1.u8 {d18}, [r2]-
- ;transpose to 8x16 matrix
- vtrn.32 d3, d7
- vtrn.32 d4, d16
- vtrn.32 d5, d17
- vtrn.32 d6, d18
-
- vtrn.16 d3, d5
- vtrn.16 d4, d6
- vtrn.16 d7, d17
- vtrn.16 d16, d18
-
- vtrn.8 d3, d4
- vtrn.8 d5, d6
- vtrn.8 d7, d16
- vtrn.8 d17, d18
-
- bl vp9_loop_filter_neon
-
- sub r0, r0, #2
-
- ;store op1, op0, oq0, oq1
- vst4.8 {d4[0], d5[0], d6[0], d7[0]}, [r0], r1- vst4.8 {d4[1], d5[1], d6[1], d7[1]}, [r0], r1- vst4.8 {d4[2], d5[2], d6[2], d7[2]}, [r0], r1- vst4.8 {d4[3], d5[3], d6[3], d7[3]}, [r0], r1- vst4.8 {d4[4], d5[4], d6[4], d7[4]}, [r0], r1- vst4.8 {d4[5], d5[5], d6[5], d7[5]}, [r0], r1- vst4.8 {d4[6], d5[6], d6[6], d7[6]}, [r0], r1- vst4.8 {d4[7], d5[7], d6[7], d7[7]}, [r0]-
- add r0, r0, r1, lsl #3 ; s += pitch * 8
- subs r12, r12, #1
- subne r2, r0, #4 ; move s pointer down by 4 columns
- bne count_lf_v_loop
-
-end_vp9_lf_v_edge
- pop {pc}- ENDP ; |vp9_lpf_vertical_4_neon|
-
-; void vp9_loop_filter_neon();
-; This is a helper function for the loopfilters. The invidual functions do the
-; necessary load, transpose (if necessary) and store. The function does not use
-; registers d8-d15.
-;
-; Inputs:
-; r0-r3, r12 PRESERVE
-; d0 blimit
-; d1 limit
-; d2 thresh
-; d3 p3
-; d4 p2
-; d5 p1
-; d6 p0
-; d7 q0
-; d16 q1
-; d17 q2
-; d18 q3
-;
-; Outputs:
-; d4 op1
-; d5 op0
-; d6 oq0
-; d7 oq1
-|vp9_loop_filter_neon| PROC
- ; filter_mask
- vabd.u8 d19, d3, d4 ; m1 = abs(p3 - p2)
- vabd.u8 d20, d4, d5 ; m2 = abs(p2 - p1)
- vabd.u8 d21, d5, d6 ; m3 = abs(p1 - p0)
- vabd.u8 d22, d16, d7 ; m4 = abs(q1 - q0)
- vabd.u8 d3, d17, d16 ; m5 = abs(q2 - q1)
- vabd.u8 d4, d18, d17 ; m6 = abs(q3 - q2)
-
- ; only compare the largest value to limit
- vmax.u8 d19, d19, d20 ; m1 = max(m1, m2)
- vmax.u8 d20, d21, d22 ; m2 = max(m3, m4)
-
- vabd.u8 d17, d6, d7 ; abs(p0 - q0)
-
- vmax.u8 d3, d3, d4 ; m3 = max(m5, m6)
-
- vmov.u8 d18, #0x80
-
- vmax.u8 d23, d19, d20 ; m1 = max(m1, m2)
-
- ; hevmask
- vcgt.u8 d21, d21, d2 ; (abs(p1 - p0) > thresh)*-1
- vcgt.u8 d22, d22, d2 ; (abs(q1 - q0) > thresh)*-1
- vmax.u8 d23, d23, d3 ; m1 = max(m1, m3)
-
- vabd.u8 d28, d5, d16 ; a = abs(p1 - q1)
- vqadd.u8 d17, d17, d17 ; b = abs(p0 - q0) * 2
-
- veor d7, d7, d18 ; qs0
-
- vcge.u8 d23, d1, d23 ; abs(m1) > limit
-
- ; filter() function
- ; convert to signed
-
- vshr.u8 d28, d28, #1 ; a = a / 2
- veor d6, d6, d18 ; ps0
-
- veor d5, d5, d18 ; ps1
- vqadd.u8 d17, d17, d28 ; a = b + a
-
- veor d16, d16, d18 ; qs1
-
- vmov.u8 d19, #3
-
- vsub.s8 d28, d7, d6 ; ( qs0 - ps0)
-
- vcge.u8 d17, d0, d17 ; a > blimit
-
- vqsub.s8 d27, d5, d16 ; filter = clamp(ps1-qs1)
- vorr d22, d21, d22 ; hevmask
-
- vmull.s8 q12, d28, d19 ; 3 * ( qs0 - ps0)
-
- vand d27, d27, d22 ; filter &= hev
- vand d23, d23, d17 ; filter_mask
-
- vaddw.s8 q12, q12, d27 ; filter + 3 * (qs0 - ps0)
-
- vmov.u8 d17, #4
-
- ; filter = clamp(filter + 3 * ( qs0 - ps0))
- vqmovn.s16 d27, q12
-
- vand d27, d27, d23 ; filter &= mask
-
- vqadd.s8 d28, d27, d19 ; filter2 = clamp(filter+3)
- vqadd.s8 d27, d27, d17 ; filter1 = clamp(filter+4)
- vshr.s8 d28, d28, #3 ; filter2 >>= 3
- vshr.s8 d27, d27, #3 ; filter1 >>= 3
-
- vqadd.s8 d19, d6, d28 ; u = clamp(ps0 + filter2)
- vqsub.s8 d26, d7, d27 ; u = clamp(qs0 - filter1)
-
- ; outer tap adjustments
- vrshr.s8 d27, d27, #1 ; filter = ++filter1 >> 1
-
- veor d6, d26, d18 ; *oq0 = u^0x80
-
- vbic d27, d27, d22 ; filter &= ~hev
-
- vqadd.s8 d21, d5, d27 ; u = clamp(ps1 + filter)
- vqsub.s8 d20, d16, d27 ; u = clamp(qs1 - filter)
-
- veor d5, d19, d18 ; *op0 = u^0x80
- veor d4, d21, d18 ; *op1 = u^0x80
- veor d7, d20, d18 ; *oq1 = u^0x80
-
- bx lr
- ENDP ; |vp9_loop_filter_neon|
-
- END
--- a/vp9/common/arm/neon/vp9_loopfilter_8_neon.c
+++ /dev/null
@@ -1,453 +1,0 @@
-/*
- * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <arm_neon.h>
-
-#include "./vpx_config.h"
-
-static INLINE void vp9_mbloop_filter_neon(
- uint8x8_t dblimit, // mblimit
- uint8x8_t dlimit, // limit
- uint8x8_t dthresh, // thresh
- uint8x8_t d3u8, // p2
- uint8x8_t d4u8, // p2
- uint8x8_t d5u8, // p1
- uint8x8_t d6u8, // p0
- uint8x8_t d7u8, // q0
- uint8x8_t d16u8, // q1
- uint8x8_t d17u8, // q2
- uint8x8_t d18u8, // q3
- uint8x8_t *d0ru8, // p1
- uint8x8_t *d1ru8, // p1
- uint8x8_t *d2ru8, // p0
- uint8x8_t *d3ru8, // q0
- uint8x8_t *d4ru8, // q1
- uint8x8_t *d5ru8) { // q1- uint32_t flat;
- uint8x8_t d0u8, d1u8, d2u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8;
- uint8x8_t d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
- int16x8_t q15s16;
- uint16x8_t q10u16, q14u16;
- int8x8_t d21s8, d24s8, d25s8, d26s8, d28s8, d29s8, d30s8;
-
- d19u8 = vabd_u8(d3u8, d4u8);
- d20u8 = vabd_u8(d4u8, d5u8);
- d21u8 = vabd_u8(d5u8, d6u8);
- d22u8 = vabd_u8(d16u8, d7u8);
- d23u8 = vabd_u8(d17u8, d16u8);
- d24u8 = vabd_u8(d18u8, d17u8);
-
- d19u8 = vmax_u8(d19u8, d20u8);
- d20u8 = vmax_u8(d21u8, d22u8);
-
- d25u8 = vabd_u8(d6u8, d4u8);
-
- d23u8 = vmax_u8(d23u8, d24u8);
-
- d26u8 = vabd_u8(d7u8, d17u8);
-
- d19u8 = vmax_u8(d19u8, d20u8);
-
- d24u8 = vabd_u8(d6u8, d7u8);
- d27u8 = vabd_u8(d3u8, d6u8);
- d28u8 = vabd_u8(d18u8, d7u8);
-
- d19u8 = vmax_u8(d19u8, d23u8);
-
- d23u8 = vabd_u8(d5u8, d16u8);
- d24u8 = vqadd_u8(d24u8, d24u8);
-
-
- d19u8 = vcge_u8(dlimit, d19u8);
-
-
- d25u8 = vmax_u8(d25u8, d26u8);
- d26u8 = vmax_u8(d27u8, d28u8);
-
- d23u8 = vshr_n_u8(d23u8, 1);
-
- d25u8 = vmax_u8(d25u8, d26u8);
-
- d24u8 = vqadd_u8(d24u8, d23u8);
-
- d20u8 = vmax_u8(d20u8, d25u8);
-
- d23u8 = vdup_n_u8(1);
- d24u8 = vcge_u8(dblimit, d24u8);
-
- d21u8 = vcgt_u8(d21u8, dthresh);
-
- d20u8 = vcge_u8(d23u8, d20u8);
-
- d19u8 = vand_u8(d19u8, d24u8);
-
- d23u8 = vcgt_u8(d22u8, dthresh);
-
- d20u8 = vand_u8(d20u8, d19u8);
-
- d22u8 = vdup_n_u8(0x80);
-
- d23u8 = vorr_u8(d21u8, d23u8);
-
- q10u16 = vcombine_u16(vreinterpret_u16_u8(d20u8),
- vreinterpret_u16_u8(d21u8));
-
- d30u8 = vshrn_n_u16(q10u16, 4);
- flat = vget_lane_u32(vreinterpret_u32_u8(d30u8), 0);
-
- if (flat == 0xffffffff) { // Check for all 1's, power_branch_only- d27u8 = vdup_n_u8(3);
- d21u8 = vdup_n_u8(2);
- q14u16 = vaddl_u8(d6u8, d7u8);
- q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
- q14u16 = vmlal_u8(q14u16, d4u8, d21u8);
- q14u16 = vaddw_u8(q14u16, d5u8);
- *d0ru8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d3u8);
- q14u16 = vsubw_u8(q14u16, d4u8);
- q14u16 = vaddw_u8(q14u16, d5u8);
- q14u16 = vaddw_u8(q14u16, d16u8);
- *d1ru8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d3u8);
- q14u16 = vsubw_u8(q14u16, d5u8);
- q14u16 = vaddw_u8(q14u16, d6u8);
- q14u16 = vaddw_u8(q14u16, d17u8);
- *d2ru8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d3u8);
- q14u16 = vsubw_u8(q14u16, d6u8);
- q14u16 = vaddw_u8(q14u16, d7u8);
- q14u16 = vaddw_u8(q14u16, d18u8);
- *d3ru8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d4u8);
- q14u16 = vsubw_u8(q14u16, d7u8);
- q14u16 = vaddw_u8(q14u16, d16u8);
- q14u16 = vaddw_u8(q14u16, d18u8);
- *d4ru8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d5u8);
- q14u16 = vsubw_u8(q14u16, d16u8);
- q14u16 = vaddw_u8(q14u16, d17u8);
- q14u16 = vaddw_u8(q14u16, d18u8);
- *d5ru8 = vqrshrn_n_u16(q14u16, 3);
- } else {- d21u8 = veor_u8(d7u8, d22u8);
- d24u8 = veor_u8(d6u8, d22u8);
- d25u8 = veor_u8(d5u8, d22u8);
- d26u8 = veor_u8(d16u8, d22u8);
-
- d27u8 = vdup_n_u8(3);
-
- d28s8 = vsub_s8(vreinterpret_s8_u8(d21u8), vreinterpret_s8_u8(d24u8));
- d29s8 = vqsub_s8(vreinterpret_s8_u8(d25u8), vreinterpret_s8_u8(d26u8));
-
- q15s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d27u8));
-
- d29s8 = vand_s8(d29s8, vreinterpret_s8_u8(d23u8));
-
- q15s16 = vaddw_s8(q15s16, d29s8);
-
- d29u8 = vdup_n_u8(4);
-
- d28s8 = vqmovn_s16(q15s16);
-
- d28s8 = vand_s8(d28s8, vreinterpret_s8_u8(d19u8));
-
- d30s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d27u8));
- d29s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d29u8));
- d30s8 = vshr_n_s8(d30s8, 3);
- d29s8 = vshr_n_s8(d29s8, 3);
-
- d24s8 = vqadd_s8(vreinterpret_s8_u8(d24u8), d30s8);
- d21s8 = vqsub_s8(vreinterpret_s8_u8(d21u8), d29s8);
-
- d29s8 = vrshr_n_s8(d29s8, 1);
- d29s8 = vbic_s8(d29s8, vreinterpret_s8_u8(d23u8));
-
- d25s8 = vqadd_s8(vreinterpret_s8_u8(d25u8), d29s8);
- d26s8 = vqsub_s8(vreinterpret_s8_u8(d26u8), d29s8);
-
- if (flat == 0) { // filter_branch_only- *d0ru8 = d4u8;
- *d1ru8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
- *d2ru8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
- *d3ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
- *d4ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
- *d5ru8 = d17u8;
- return;
- }
-
- d21u8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
- d24u8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
- d25u8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
- d26u8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
-
- d23u8 = vdup_n_u8(2);
- q14u16 = vaddl_u8(d6u8, d7u8);
- q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
- q14u16 = vmlal_u8(q14u16, d4u8, d23u8);
-
- d0u8 = vbsl_u8(d20u8, dblimit, d4u8);
-
- q14u16 = vaddw_u8(q14u16, d5u8);
-
- d1u8 = vbsl_u8(d20u8, dlimit, d25u8);
-
- d30u8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d3u8);
- q14u16 = vsubw_u8(q14u16, d4u8);
- q14u16 = vaddw_u8(q14u16, d5u8);
- q14u16 = vaddw_u8(q14u16, d16u8);
-
- d2u8 = vbsl_u8(d20u8, dthresh, d24u8);
-
- d31u8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d3u8);
- q14u16 = vsubw_u8(q14u16, d5u8);
- q14u16 = vaddw_u8(q14u16, d6u8);
- q14u16 = vaddw_u8(q14u16, d17u8);
-
- *d0ru8 = vbsl_u8(d20u8, d30u8, d0u8);
-
- d23u8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d3u8);
- q14u16 = vsubw_u8(q14u16, d6u8);
- q14u16 = vaddw_u8(q14u16, d7u8);
-
- *d1ru8 = vbsl_u8(d20u8, d31u8, d1u8);
-
- q14u16 = vaddw_u8(q14u16, d18u8);
-
- *d2ru8 = vbsl_u8(d20u8, d23u8, d2u8);
-
- d22u8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d4u8);
- q14u16 = vsubw_u8(q14u16, d7u8);
- q14u16 = vaddw_u8(q14u16, d16u8);
-
- d3u8 = vbsl_u8(d20u8, d3u8, d21u8);
-
- q14u16 = vaddw_u8(q14u16, d18u8);
-
- d4u8 = vbsl_u8(d20u8, d4u8, d26u8);
-
- d6u8 = vqrshrn_n_u16(q14u16, 3);
-
- q14u16 = vsubw_u8(q14u16, d5u8);
- q14u16 = vsubw_u8(q14u16, d16u8);
- q14u16 = vaddw_u8(q14u16, d17u8);
- q14u16 = vaddw_u8(q14u16, d18u8);
-
- d5u8 = vbsl_u8(d20u8, d5u8, d17u8);
-
- d7u8 = vqrshrn_n_u16(q14u16, 3);
-
- *d3ru8 = vbsl_u8(d20u8, d22u8, d3u8);
- *d4ru8 = vbsl_u8(d20u8, d6u8, d4u8);
- *d5ru8 = vbsl_u8(d20u8, d7u8, d5u8);
- }
- return;
-}
-
-void vp9_lpf_horizontal_8_neon(
- unsigned char *src,
- int pitch,
- unsigned char *blimit,
- unsigned char *limit,
- unsigned char *thresh,
- int count) {- int i;
- uint8_t *s, *psrc;
- uint8x8_t dblimit, dlimit, dthresh;
- uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
- uint8x8_t d16u8, d17u8, d18u8;
-
- if (count == 0) // end_vp9_mblf_h_edge
- return;
-
- dblimit = vld1_u8(blimit);
- dlimit = vld1_u8(limit);
- dthresh = vld1_u8(thresh);
-
- psrc = src - (pitch << 2);
- for (i = 0; i < count; i++) {- s = psrc + i * 8;
-
- d3u8 = vld1_u8(s);
- s += pitch;
- d4u8 = vld1_u8(s);
- s += pitch;
- d5u8 = vld1_u8(s);
- s += pitch;
- d6u8 = vld1_u8(s);
- s += pitch;
- d7u8 = vld1_u8(s);
- s += pitch;
- d16u8 = vld1_u8(s);
- s += pitch;
- d17u8 = vld1_u8(s);
- s += pitch;
- d18u8 = vld1_u8(s);
-
- vp9_mbloop_filter_neon(dblimit, dlimit, dthresh,
- d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
- &d0u8, &d1u8, &d2u8, &d3u8, &d4u8, &d5u8);
-
- s -= (pitch * 6);
- vst1_u8(s, d0u8);
- s += pitch;
- vst1_u8(s, d1u8);
- s += pitch;
- vst1_u8(s, d2u8);
- s += pitch;
- vst1_u8(s, d3u8);
- s += pitch;
- vst1_u8(s, d4u8);
- s += pitch;
- vst1_u8(s, d5u8);
- }
- return;
-}
-
-void vp9_lpf_vertical_8_neon(
- unsigned char *src,
- int pitch,
- unsigned char *blimit,
- unsigned char *limit,
- unsigned char *thresh,
- int count) {- int i;
- uint8_t *s;
- uint8x8_t dblimit, dlimit, dthresh;
- uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
- uint8x8_t d16u8, d17u8, d18u8;
- uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
- uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
- uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
- uint8x8x4_t d4Result;
- uint8x8x2_t d2Result;
-
- if (count == 0)
- return;
-
- dblimit = vld1_u8(blimit);
- dlimit = vld1_u8(limit);
- dthresh = vld1_u8(thresh);
-
- for (i = 0; i < count; i++) {- s = src + (i * (pitch << 3)) - 4;
-
- d3u8 = vld1_u8(s);
- s += pitch;
- d4u8 = vld1_u8(s);
- s += pitch;
- d5u8 = vld1_u8(s);
- s += pitch;
- d6u8 = vld1_u8(s);
- s += pitch;
- d7u8 = vld1_u8(s);
- s += pitch;
- d16u8 = vld1_u8(s);
- s += pitch;
- d17u8 = vld1_u8(s);
- s += pitch;
- d18u8 = vld1_u8(s);
-
- d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
- vreinterpret_u32_u8(d7u8));
- d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
- vreinterpret_u32_u8(d16u8));
- d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
- vreinterpret_u32_u8(d17u8));
- d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
- vreinterpret_u32_u8(d18u8));
-
- d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
- vreinterpret_u16_u32(d2tmp2.val[0]));
- d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
- vreinterpret_u16_u32(d2tmp3.val[0]));
- d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
- vreinterpret_u16_u32(d2tmp2.val[1]));
- d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
- vreinterpret_u16_u32(d2tmp3.val[1]));
-
- d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
- vreinterpret_u8_u16(d2tmp5.val[0]));
- d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
- vreinterpret_u8_u16(d2tmp5.val[1]));
- d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
- vreinterpret_u8_u16(d2tmp7.val[0]));
- d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
- vreinterpret_u8_u16(d2tmp7.val[1]));
-
- d3u8 = d2tmp8.val[0];
- d4u8 = d2tmp8.val[1];
- d5u8 = d2tmp9.val[0];
- d6u8 = d2tmp9.val[1];
- d7u8 = d2tmp10.val[0];
- d16u8 = d2tmp10.val[1];
- d17u8 = d2tmp11.val[0];
- d18u8 = d2tmp11.val[1];
-
- vp9_mbloop_filter_neon(dblimit, dlimit, dthresh,
- d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
- &d0u8, &d1u8, &d2u8, &d3u8, &d4u8, &d5u8);
-
- d4Result.val[0] = d0u8;
- d4Result.val[1] = d1u8;
- d4Result.val[2] = d2u8;
- d4Result.val[3] = d3u8;
-
- d2Result.val[0] = d4u8;
- d2Result.val[1] = d5u8;
-
- s = src - 3;
- vst4_lane_u8(s, d4Result, 0);
- s += pitch;
- vst4_lane_u8(s, d4Result, 1);
- s += pitch;
- vst4_lane_u8(s, d4Result, 2);
- s += pitch;
- vst4_lane_u8(s, d4Result, 3);
- s += pitch;
- vst4_lane_u8(s, d4Result, 4);
- s += pitch;
- vst4_lane_u8(s, d4Result, 5);
- s += pitch;
- vst4_lane_u8(s, d4Result, 6);
- s += pitch;
- vst4_lane_u8(s, d4Result, 7);
-
- s = src + 1;
- vst2_lane_u8(s, d2Result, 0);
- s += pitch;
- vst2_lane_u8(s, d2Result, 1);
- s += pitch;
- vst2_lane_u8(s, d2Result, 2);
- s += pitch;
- vst2_lane_u8(s, d2Result, 3);
- s += pitch;
- vst2_lane_u8(s, d2Result, 4);
- s += pitch;
- vst2_lane_u8(s, d2Result, 5);
- s += pitch;
- vst2_lane_u8(s, d2Result, 6);
- s += pitch;
- vst2_lane_u8(s, d2Result, 7);
- }
- return;
-}
--- a/vp9/common/arm/neon/vp9_loopfilter_8_neon_asm.asm
+++ /dev/null
@@ -1,451 +1,0 @@
-;
-; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license
-; that can be found in the LICENSE file in the root of the source
-; tree. An additional intellectual property rights grant can be found
-; in the file PATENTS. All contributing project authors may
-; be found in the AUTHORS file in the root of the source tree.
-;
-
- EXPORT |vp9_lpf_horizontal_8_neon|
- EXPORT |vp9_lpf_vertical_8_neon|
- ARM
-
- AREA ||.text||, CODE, READONLY, ALIGN=2
-
-; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
-; works on 16 iterations at a time.
-; TODO(fgalligan): See about removing the count code as this function is only
-; called with a count of 1.
-;
-; void vp9_lpf_horizontal_8_neon(uint8_t *s, int p,
-; const uint8_t *blimit,
-; const uint8_t *limit,
-; const uint8_t *thresh,
-; int count)
-; r0 uint8_t *s,
-; r1 int p, /* pitch */
-; r2 const uint8_t *blimit,
-; r3 const uint8_t *limit,
-; sp const uint8_t *thresh,
-; sp+4 int count
-|vp9_lpf_horizontal_8_neon| PROC
- push {r4-r5, lr}-
- vld1.8 {d0[]}, [r2] ; duplicate *blimit- ldr r12, [sp, #16] ; load count
- ldr r2, [sp, #12] ; load thresh
- add r1, r1, r1 ; double pitch
-
- cmp r12, #0
- beq end_vp9_mblf_h_edge
-
- vld1.8 {d1[]}, [r3] ; duplicate *limit- vld1.8 {d2[]}, [r2] ; duplicate *thresh-
-count_mblf_h_loop
- sub r3, r0, r1, lsl #1 ; move src pointer down by 4 lines
- add r2, r3, r1, lsr #1 ; set to 3 lines down
-
- vld1.u8 {d3}, [r3@64], r1 ; p3- vld1.u8 {d4}, [r2@64], r1 ; p2- vld1.u8 {d5}, [r3@64], r1 ; p1- vld1.u8 {d6}, [r2@64], r1 ; p0- vld1.u8 {d7}, [r3@64], r1 ; q0- vld1.u8 {d16}, [r2@64], r1 ; q1- vld1.u8 {d17}, [r3@64] ; q2- vld1.u8 {d18}, [r2@64], r1 ; q3-
- sub r3, r3, r1, lsl #1
- sub r2, r2, r1, lsl #2
-
- bl vp9_mbloop_filter_neon
-
- vst1.u8 {d0}, [r2@64], r1 ; store op2- vst1.u8 {d1}, [r3@64], r1 ; store op1- vst1.u8 {d2}, [r2@64], r1 ; store op0- vst1.u8 {d3}, [r3@64], r1 ; store oq0- vst1.u8 {d4}, [r2@64], r1 ; store oq1- vst1.u8 {d5}, [r3@64], r1 ; store oq2-
- add r0, r0, #8
- subs r12, r12, #1
- bne count_mblf_h_loop
-
-end_vp9_mblf_h_edge
- pop {r4-r5, pc}-
- ENDP ; |vp9_lpf_horizontal_8_neon|
-
-; void vp9_lpf_vertical_8_neon(uint8_t *s,
-; int pitch,
-; const uint8_t *blimit,
-; const uint8_t *limit,
-; const uint8_t *thresh,
-; int count)
-;
-; r0 uint8_t *s,
-; r1 int pitch,
-; r2 const uint8_t *blimit,
-; r3 const uint8_t *limit,
-; sp const uint8_t *thresh,
-; sp+4 int count
-|vp9_lpf_vertical_8_neon| PROC
- push {r4-r5, lr}-
- vld1.8 {d0[]}, [r2] ; duplicate *blimit- ldr r12, [sp, #16] ; load count
- vld1.8 {d1[]}, [r3] ; duplicate *limit-
- ldr r3, [sp, #12] ; load thresh
- sub r2, r0, #4 ; move s pointer down by 4 columns
- cmp r12, #0
- beq end_vp9_mblf_v_edge
-
- vld1.8 {d2[]}, [r3] ; duplicate *thresh-
-count_mblf_v_loop
- vld1.u8 {d3}, [r2], r1 ; load s data- vld1.u8 {d4}, [r2], r1- vld1.u8 {d5}, [r2], r1- vld1.u8 {d6}, [r2], r1- vld1.u8 {d7}, [r2], r1- vld1.u8 {d16}, [r2], r1- vld1.u8 {d17}, [r2], r1- vld1.u8 {d18}, [r2]-
- ;transpose to 8x16 matrix
- vtrn.32 d3, d7
- vtrn.32 d4, d16
- vtrn.32 d5, d17
- vtrn.32 d6, d18
-
- vtrn.16 d3, d5
- vtrn.16 d4, d6
- vtrn.16 d7, d17
- vtrn.16 d16, d18
-
- vtrn.8 d3, d4
- vtrn.8 d5, d6
- vtrn.8 d7, d16
- vtrn.8 d17, d18
-
- sub r2, r0, #3
- add r3, r0, #1
-
- bl vp9_mbloop_filter_neon
-
- ;store op2, op1, op0, oq0
- vst4.8 {d0[0], d1[0], d2[0], d3[0]}, [r2], r1- vst4.8 {d0[1], d1[1], d2[1], d3[1]}, [r2], r1- vst4.8 {d0[2], d1[2], d2[2], d3[2]}, [r2], r1- vst4.8 {d0[3], d1[3], d2[3], d3[3]}, [r2], r1- vst4.8 {d0[4], d1[4], d2[4], d3[4]}, [r2], r1- vst4.8 {d0[5], d1[5], d2[5], d3[5]}, [r2], r1- vst4.8 {d0[6], d1[6], d2[6], d3[6]}, [r2], r1- vst4.8 {d0[7], d1[7], d2[7], d3[7]}, [r2]-
- ;store oq1, oq2
- vst2.8 {d4[0], d5[0]}, [r3], r1- vst2.8 {d4[1], d5[1]}, [r3], r1- vst2.8 {d4[2], d5[2]}, [r3], r1- vst2.8 {d4[3], d5[3]}, [r3], r1- vst2.8 {d4[4], d5[4]}, [r3], r1- vst2.8 {d4[5], d5[5]}, [r3], r1- vst2.8 {d4[6], d5[6]}, [r3], r1- vst2.8 {d4[7], d5[7]}, [r3]-
- add r0, r0, r1, lsl #3 ; s += pitch * 8
- subs r12, r12, #1
- subne r2, r0, #4 ; move s pointer down by 4 columns
- bne count_mblf_v_loop
-
-end_vp9_mblf_v_edge
- pop {r4-r5, pc}- ENDP ; |vp9_lpf_vertical_8_neon|
-
-; void vp9_mbloop_filter_neon();
-; This is a helper function for the loopfilters. The invidual functions do the
-; necessary load, transpose (if necessary) and store. The function does not use
-; registers d8-d15.
-;
-; Inputs:
-; r0-r3, r12 PRESERVE
-; d0 blimit
-; d1 limit
-; d2 thresh
-; d3 p3
-; d4 p2
-; d5 p1
-; d6 p0
-; d7 q0
-; d16 q1
-; d17 q2
-; d18 q3
-;
-; Outputs:
-; d0 op2
-; d1 op1
-; d2 op0
-; d3 oq0
-; d4 oq1
-; d5 oq2
-|vp9_mbloop_filter_neon| PROC
- ; filter_mask
- vabd.u8 d19, d3, d4 ; m1 = abs(p3 - p2)
- vabd.u8 d20, d4, d5 ; m2 = abs(p2 - p1)
- vabd.u8 d21, d5, d6 ; m3 = abs(p1 - p0)
- vabd.u8 d22, d16, d7 ; m4 = abs(q1 - q0)
- vabd.u8 d23, d17, d16 ; m5 = abs(q2 - q1)
- vabd.u8 d24, d18, d17 ; m6 = abs(q3 - q2)
-
- ; only compare the largest value to limit
- vmax.u8 d19, d19, d20 ; m1 = max(m1, m2)
- vmax.u8 d20, d21, d22 ; m2 = max(m3, m4)
-
- vabd.u8 d25, d6, d4 ; m7 = abs(p0 - p2)
-
- vmax.u8 d23, d23, d24 ; m3 = max(m5, m6)
-
- vabd.u8 d26, d7, d17 ; m8 = abs(q0 - q2)
-
- vmax.u8 d19, d19, d20
-
- vabd.u8 d24, d6, d7 ; m9 = abs(p0 - q0)
- vabd.u8 d27, d3, d6 ; m10 = abs(p3 - p0)
- vabd.u8 d28, d18, d7 ; m11 = abs(q3 - q0)
-
- vmax.u8 d19, d19, d23
-
- vabd.u8 d23, d5, d16 ; a = abs(p1 - q1)
- vqadd.u8 d24, d24, d24 ; b = abs(p0 - q0) * 2
-
- ; abs () > limit
- vcge.u8 d19, d1, d19
-
- ; only compare the largest value to thresh
- vmax.u8 d25, d25, d26 ; m4 = max(m7, m8)
- vmax.u8 d26, d27, d28 ; m5 = max(m10, m11)
-
- vshr.u8 d23, d23, #1 ; a = a / 2
-
- vmax.u8 d25, d25, d26 ; m4 = max(m4, m5)
-
- vqadd.u8 d24, d24, d23 ; a = b + a
-
- vmax.u8 d20, d20, d25 ; m2 = max(m2, m4)
-
- vmov.u8 d23, #1
- vcge.u8 d24, d0, d24 ; a > blimit
-
- vcgt.u8 d21, d21, d2 ; (abs(p1 - p0) > thresh)*-1
-
- vcge.u8 d20, d23, d20 ; flat
-
- vand d19, d19, d24 ; mask
-
- vcgt.u8 d23, d22, d2 ; (abs(q1 - q0) > thresh)*-1
-
- vand d20, d20, d19 ; flat & mask
-
- vmov.u8 d22, #0x80
-
- vorr d23, d21, d23 ; hev
-
- ; This instruction will truncate the "flat & mask" masks down to 4 bits
- ; each to fit into one 32 bit arm register. The values are stored in
- ; q10.64[0].
- vshrn.u16 d30, q10, #4
- vmov.u32 r4, d30[0] ; flat & mask 4bits
-
- adds r5, r4, #1 ; Check for all 1's
-
- ; If mask and flat are 1's for all vectors, then we only need to execute
- ; the power branch for all vectors.
- beq power_branch_only
-
- cmp r4, #0 ; Check for 0, set flag for later
-
- ; mbfilter() function
- ; filter() function
- ; convert to signed
- veor d21, d7, d22 ; qs0
- veor d24, d6, d22 ; ps0
- veor d25, d5, d22 ; ps1
- veor d26, d16, d22 ; qs1
-
- vmov.u8 d27, #3
-
- vsub.s8 d28, d21, d24 ; ( qs0 - ps0)
-
- vqsub.s8 d29, d25, d26 ; filter = clamp(ps1-qs1)
-
- vmull.s8 q15, d28, d27 ; 3 * ( qs0 - ps0)
-
- vand d29, d29, d23 ; filter &= hev
-
- vaddw.s8 q15, q15, d29 ; filter + 3 * (qs0 - ps0)
-
- vmov.u8 d29, #4
-
- ; filter = clamp(filter + 3 * ( qs0 - ps0))
- vqmovn.s16 d28, q15
-
- vand d28, d28, d19 ; filter &= mask
-
- vqadd.s8 d30, d28, d27 ; filter2 = clamp(filter+3)
- vqadd.s8 d29, d28, d29 ; filter1 = clamp(filter+4)
- vshr.s8 d30, d30, #3 ; filter2 >>= 3
- vshr.s8 d29, d29, #3 ; filter1 >>= 3
-
- vqadd.s8 d24, d24, d30 ; op0 = clamp(ps0 + filter2)
- vqsub.s8 d21, d21, d29 ; oq0 = clamp(qs0 - filter1)
-
- ; outer tap adjustments: ++filter1 >> 1
- vrshr.s8 d29, d29, #1
- vbic d29, d29, d23 ; filter &= ~hev
-
- vqadd.s8 d25, d25, d29 ; op1 = clamp(ps1 + filter)
- vqsub.s8 d26, d26, d29 ; oq1 = clamp(qs1 - filter)
-
- ; If mask and flat are 0's for all vectors, then we only need to execute
- ; the filter branch for all vectors.
- beq filter_branch_only
-
- ; If mask and flat are mixed then we must perform both branches and
- ; combine the data.
- veor d24, d24, d22 ; *f_op0 = u^0x80
- veor d21, d21, d22 ; *f_oq0 = u^0x80
- veor d25, d25, d22 ; *f_op1 = u^0x80
- veor d26, d26, d22 ; *f_oq1 = u^0x80
-
- ; At this point we have already executed the filter branch. The filter
- ; branch does not set op2 or oq2, so use p2 and q2. Execute the power
- ; branch and combine the data.
- vmov.u8 d23, #2
- vaddl.u8 q14, d6, d7 ; r_op2 = p0 + q0
- vmlal.u8 q14, d3, d27 ; r_op2 += p3 * 3
- vmlal.u8 q14, d4, d23 ; r_op2 += p2 * 2
-
- vbif d0, d4, d20 ; op2 |= p2 & ~(flat & mask)
-
- vaddw.u8 q14, d5 ; r_op2 += p1
-
- vbif d1, d25, d20 ; op1 |= f_op1 & ~(flat & mask)
-
- vqrshrn.u16 d30, q14, #3 ; r_op2
-
- vsubw.u8 q14, d3 ; r_op1 = r_op2 - p3
- vsubw.u8 q14, d4 ; r_op1 -= p2
- vaddw.u8 q14, d5 ; r_op1 += p1
- vaddw.u8 q14, d16 ; r_op1 += q1
-
- vbif d2, d24, d20 ; op0 |= f_op0 & ~(flat & mask)
-
- vqrshrn.u16 d31, q14, #3 ; r_op1
-
- vsubw.u8 q14, d3 ; r_op0 = r_op1 - p3
- vsubw.u8 q14, d5 ; r_op0 -= p1
- vaddw.u8 q14, d6 ; r_op0 += p0
- vaddw.u8 q14, d17 ; r_op0 += q2
-
- vbit d0, d30, d20 ; op2 |= r_op2 & (flat & mask)
-
- vqrshrn.u16 d23, q14, #3 ; r_op0
-
- vsubw.u8 q14, d3 ; r_oq0 = r_op0 - p3
- vsubw.u8 q14, d6 ; r_oq0 -= p0
- vaddw.u8 q14, d7 ; r_oq0 += q0
-
- vbit d1, d31, d20 ; op1 |= r_op1 & (flat & mask)
-
- vaddw.u8 q14, d18 ; oq0 += q3
-
- vbit d2, d23, d20 ; op0 |= r_op0 & (flat & mask)
-
- vqrshrn.u16 d22, q14, #3 ; r_oq0
-
- vsubw.u8 q14, d4 ; r_oq1 = r_oq0 - p2
- vsubw.u8 q14, d7 ; r_oq1 -= q0
- vaddw.u8 q14, d16 ; r_oq1 += q1
-
- vbif d3, d21, d20 ; oq0 |= f_oq0 & ~(flat & mask)
-
- vaddw.u8 q14, d18 ; r_oq1 += q3
-
- vbif d4, d26, d20 ; oq1 |= f_oq1 & ~(flat & mask)
-
- vqrshrn.u16 d6, q14, #3 ; r_oq1
-
- vsubw.u8 q14, d5 ; r_oq2 = r_oq1 - p1
- vsubw.u8 q14, d16 ; r_oq2 -= q1
- vaddw.u8 q14, d17 ; r_oq2 += q2
- vaddw.u8 q14, d18 ; r_oq2 += q3
-
- vbif d5, d17, d20 ; oq2 |= q2 & ~(flat & mask)
-
- vqrshrn.u16 d7, q14, #3 ; r_oq2
-
- vbit d3, d22, d20 ; oq0 |= r_oq0 & (flat & mask)
- vbit d4, d6, d20 ; oq1 |= r_oq1 & (flat & mask)
- vbit d5, d7, d20 ; oq2 |= r_oq2 & (flat & mask)
-
- bx lr
-
-power_branch_only
- vmov.u8 d27, #3
- vmov.u8 d21, #2
- vaddl.u8 q14, d6, d7 ; op2 = p0 + q0
- vmlal.u8 q14, d3, d27 ; op2 += p3 * 3
- vmlal.u8 q14, d4, d21 ; op2 += p2 * 2
- vaddw.u8 q14, d5 ; op2 += p1
- vqrshrn.u16 d0, q14, #3 ; op2
-
- vsubw.u8 q14, d3 ; op1 = op2 - p3
- vsubw.u8 q14, d4 ; op1 -= p2
- vaddw.u8 q14, d5 ; op1 += p1
- vaddw.u8 q14, d16 ; op1 += q1
- vqrshrn.u16 d1, q14, #3 ; op1
-
- vsubw.u8 q14, d3 ; op0 = op1 - p3
- vsubw.u8 q14, d5 ; op0 -= p1
- vaddw.u8 q14, d6 ; op0 += p0
- vaddw.u8 q14, d17 ; op0 += q2
- vqrshrn.u16 d2, q14, #3 ; op0
-
- vsubw.u8 q14, d3 ; oq0 = op0 - p3
- vsubw.u8 q14, d6 ; oq0 -= p0
- vaddw.u8 q14, d7 ; oq0 += q0
- vaddw.u8 q14, d18 ; oq0 += q3
- vqrshrn.u16 d3, q14, #3 ; oq0
-
- vsubw.u8 q14, d4 ; oq1 = oq0 - p2
- vsubw.u8 q14, d7 ; oq1 -= q0
- vaddw.u8 q14, d16 ; oq1 += q1
- vaddw.u8 q14, d18 ; oq1 += q3
- vqrshrn.u16 d4, q14, #3 ; oq1
-
- vsubw.u8 q14, d5 ; oq2 = oq1 - p1
- vsubw.u8 q14, d16 ; oq2 -= q1
- vaddw.u8 q14, d17 ; oq2 += q2
- vaddw.u8 q14, d18 ; oq2 += q3
- vqrshrn.u16 d5, q14, #3 ; oq2
-
- bx lr
-
-filter_branch_only
- ; TODO(fgalligan): See if we can rearange registers so we do not need to
- ; do the 2 vswp.
- vswp d0, d4 ; op2
- vswp d5, d17 ; oq2
- veor d2, d24, d22 ; *op0 = u^0x80
- veor d3, d21, d22 ; *oq0 = u^0x80
- veor d1, d25, d22 ; *op1 = u^0x80
- veor d4, d26, d22 ; *oq1 = u^0x80
-
- bx lr
-
- ENDP ; |vp9_mbloop_filter_neon|
-
- END
--- a/vp9/common/arm/neon/vp9_loopfilter_neon.c
+++ /dev/null
@@ -1,58 +1,0 @@
-/*
- * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <arm_neon.h>
-
-#include "./vp9_rtcd.h"
-#include "./vpx_config.h"
-#include "vpx/vpx_integer.h"
-
-void vp9_lpf_vertical_4_dual_neon(uint8_t *s, int p,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1) {- vp9_lpf_vertical_4_neon(s, p, blimit0, limit0, thresh0, 1);
- vp9_lpf_vertical_4_neon(s + 8 * p, p, blimit1, limit1, thresh1, 1);
-}
-
-#if HAVE_NEON_ASM
-void vp9_lpf_horizontal_8_dual_neon(uint8_t *s, int p /* pitch */,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1) {- vp9_lpf_horizontal_8_neon(s, p, blimit0, limit0, thresh0, 1);
- vp9_lpf_horizontal_8_neon(s + 8, p, blimit1, limit1, thresh1, 1);
-}
-
-void vp9_lpf_vertical_8_dual_neon(uint8_t *s, int p,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1) {- vp9_lpf_vertical_8_neon(s, p, blimit0, limit0, thresh0, 1);
- vp9_lpf_vertical_8_neon(s + 8 * p, p, blimit1, limit1, thresh1, 1);
-}
-
-void vp9_lpf_vertical_16_dual_neon(uint8_t *s, int p,
- const uint8_t *blimit,
- const uint8_t *limit,
- const uint8_t *thresh) {- vp9_lpf_vertical_16_neon(s, p, blimit, limit, thresh);
- vp9_lpf_vertical_16_neon(s + 8 * p, p, blimit, limit, thresh);
-}
-#endif // HAVE_NEON_ASM
--- a/vp9/common/arm/neon/vp9_mb_lpf_neon.asm
+++ /dev/null
@@ -1,606 +1,0 @@
-;
-; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license
-; that can be found in the LICENSE file in the root of the source
-; tree. An additional intellectual property rights grant can be found
-; in the file PATENTS. All contributing project authors may
-; be found in the AUTHORS file in the root of the source tree.
-;
-
- EXPORT |vp9_lpf_horizontal_16_neon|
- EXPORT |vp9_lpf_vertical_16_neon|
- ARM
-
- AREA ||.text||, CODE, READONLY, ALIGN=2
-
-; void vp9_lpf_horizontal_16_neon(uint8_t *s, int p,
-; const uint8_t *blimit,
-; const uint8_t *limit,
-; const uint8_t *thresh
-; int count)
-; r0 uint8_t *s,
-; r1 int p, /* pitch */
-; r2 const uint8_t *blimit,
-; r3 const uint8_t *limit,
-; sp const uint8_t *thresh,
-|vp9_lpf_horizontal_16_neon| PROC
- push {r4-r8, lr}- vpush {d8-d15}- ldr r4, [sp, #88] ; load thresh
- ldr r12, [sp, #92] ; load count
-
-h_count
- vld1.8 {d16[]}, [r2] ; load *blimit- vld1.8 {d17[]}, [r3] ; load *limit- vld1.8 {d18[]}, [r4] ; load *thresh-
- sub r8, r0, r1, lsl #3 ; move src pointer down by 8 lines
-
- vld1.u8 {d0}, [r8@64], r1 ; p7- vld1.u8 {d1}, [r8@64], r1 ; p6- vld1.u8 {d2}, [r8@64], r1 ; p5- vld1.u8 {d3}, [r8@64], r1 ; p4- vld1.u8 {d4}, [r8@64], r1 ; p3- vld1.u8 {d5}, [r8@64], r1 ; p2- vld1.u8 {d6}, [r8@64], r1 ; p1- vld1.u8 {d7}, [r8@64], r1 ; p0- vld1.u8 {d8}, [r8@64], r1 ; q0- vld1.u8 {d9}, [r8@64], r1 ; q1- vld1.u8 {d10}, [r8@64], r1 ; q2- vld1.u8 {d11}, [r8@64], r1 ; q3- vld1.u8 {d12}, [r8@64], r1 ; q4- vld1.u8 {d13}, [r8@64], r1 ; q5- vld1.u8 {d14}, [r8@64], r1 ; q6- vld1.u8 {d15}, [r8@64], r1 ; q7-
- bl vp9_wide_mbfilter_neon
-
- tst r7, #1
- beq h_mbfilter
-
- ; flat && mask were not set for any of the channels. Just store the values
- ; from filter.
- sub r8, r0, r1, lsl #1
-
- vst1.u8 {d25}, [r8@64], r1 ; store op1- vst1.u8 {d24}, [r8@64], r1 ; store op0- vst1.u8 {d23}, [r8@64], r1 ; store oq0- vst1.u8 {d26}, [r8@64], r1 ; store oq1-
- b h_next
-
-h_mbfilter
- tst r7, #2
- beq h_wide_mbfilter
-
- ; flat2 was not set for any of the channels. Just store the values from
- ; mbfilter.
- sub r8, r0, r1, lsl #1
- sub r8, r8, r1
-
- vst1.u8 {d18}, [r8@64], r1 ; store op2- vst1.u8 {d19}, [r8@64], r1 ; store op1- vst1.u8 {d20}, [r8@64], r1 ; store op0- vst1.u8 {d21}, [r8@64], r1 ; store oq0- vst1.u8 {d22}, [r8@64], r1 ; store oq1- vst1.u8 {d23}, [r8@64], r1 ; store oq2-
- b h_next
-
-h_wide_mbfilter
- sub r8, r0, r1, lsl #3
- add r8, r8, r1
-
- vst1.u8 {d16}, [r8@64], r1 ; store op6- vst1.u8 {d24}, [r8@64], r1 ; store op5- vst1.u8 {d25}, [r8@64], r1 ; store op4- vst1.u8 {d26}, [r8@64], r1 ; store op3- vst1.u8 {d27}, [r8@64], r1 ; store op2- vst1.u8 {d18}, [r8@64], r1 ; store op1- vst1.u8 {d19}, [r8@64], r1 ; store op0- vst1.u8 {d20}, [r8@64], r1 ; store oq0- vst1.u8 {d21}, [r8@64], r1 ; store oq1- vst1.u8 {d22}, [r8@64], r1 ; store oq2- vst1.u8 {d23}, [r8@64], r1 ; store oq3- vst1.u8 {d1}, [r8@64], r1 ; store oq4- vst1.u8 {d2}, [r8@64], r1 ; store oq5- vst1.u8 {d3}, [r8@64], r1 ; store oq6-
-h_next
- add r0, r0, #8
- subs r12, r12, #1
- bne h_count
-
- vpop {d8-d15}- pop {r4-r8, pc}-
- ENDP ; |vp9_lpf_horizontal_16_neon|
-
-; void vp9_lpf_vertical_16_neon(uint8_t *s, int p,
-; const uint8_t *blimit,
-; const uint8_t *limit,
-; const uint8_t *thresh)
-; r0 uint8_t *s,
-; r1 int p, /* pitch */
-; r2 const uint8_t *blimit,
-; r3 const uint8_t *limit,
-; sp const uint8_t *thresh,
-|vp9_lpf_vertical_16_neon| PROC
- push {r4-r8, lr}- vpush {d8-d15}- ldr r4, [sp, #88] ; load thresh
-
- vld1.8 {d16[]}, [r2] ; load *blimit- vld1.8 {d17[]}, [r3] ; load *limit- vld1.8 {d18[]}, [r4] ; load *thresh-
- sub r8, r0, #8
-
- vld1.8 {d0}, [r8@64], r1- vld1.8 {d8}, [r0@64], r1- vld1.8 {d1}, [r8@64], r1- vld1.8 {d9}, [r0@64], r1- vld1.8 {d2}, [r8@64], r1- vld1.8 {d10}, [r0@64], r1- vld1.8 {d3}, [r8@64], r1- vld1.8 {d11}, [r0@64], r1- vld1.8 {d4}, [r8@64], r1- vld1.8 {d12}, [r0@64], r1- vld1.8 {d5}, [r8@64], r1- vld1.8 {d13}, [r0@64], r1- vld1.8 {d6}, [r8@64], r1- vld1.8 {d14}, [r0@64], r1- vld1.8 {d7}, [r8@64], r1- vld1.8 {d15}, [r0@64], r1-
- sub r0, r0, r1, lsl #3
-
- vtrn.32 q0, q2
- vtrn.32 q1, q3
- vtrn.32 q4, q6
- vtrn.32 q5, q7
-
- vtrn.16 q0, q1
- vtrn.16 q2, q3
- vtrn.16 q4, q5
- vtrn.16 q6, q7
-
- vtrn.8 d0, d1
- vtrn.8 d2, d3
- vtrn.8 d4, d5
- vtrn.8 d6, d7
-
- vtrn.8 d8, d9
- vtrn.8 d10, d11
- vtrn.8 d12, d13
- vtrn.8 d14, d15
-
- bl vp9_wide_mbfilter_neon
-
- tst r7, #1
- beq v_mbfilter
-
- ; flat && mask were not set for any of the channels. Just store the values
- ; from filter.
- sub r8, r0, #2
-
- vswp d23, d25
-
- vst4.8 {d23[0], d24[0], d25[0], d26[0]}, [r8], r1- vst4.8 {d23[1], d24[1], d25[1], d26[1]}, [r8], r1- vst4.8 {d23[2], d24[2], d25[2], d26[2]}, [r8], r1- vst4.8 {d23[3], d24[3], d25[3], d26[3]}, [r8], r1- vst4.8 {d23[4], d24[4], d25[4], d26[4]}, [r8], r1- vst4.8 {d23[5], d24[5], d25[5], d26[5]}, [r8], r1- vst4.8 {d23[6], d24[6], d25[6], d26[6]}, [r8], r1- vst4.8 {d23[7], d24[7], d25[7], d26[7]}, [r8], r1-
- b v_end
-
-v_mbfilter
- tst r7, #2
- beq v_wide_mbfilter
-
- ; flat2 was not set for any of the channels. Just store the values from
- ; mbfilter.
- sub r8, r0, #3
-
- vst3.8 {d18[0], d19[0], d20[0]}, [r8], r1- vst3.8 {d21[0], d22[0], d23[0]}, [r0], r1- vst3.8 {d18[1], d19[1], d20[1]}, [r8], r1- vst3.8 {d21[1], d22[1], d23[1]}, [r0], r1- vst3.8 {d18[2], d19[2], d20[2]}, [r8], r1- vst3.8 {d21[2], d22[2], d23[2]}, [r0], r1- vst3.8 {d18[3], d19[3], d20[3]}, [r8], r1- vst3.8 {d21[3], d22[3], d23[3]}, [r0], r1- vst3.8 {d18[4], d19[4], d20[4]}, [r8], r1- vst3.8 {d21[4], d22[4], d23[4]}, [r0], r1- vst3.8 {d18[5], d19[5], d20[5]}, [r8], r1- vst3.8 {d21[5], d22[5], d23[5]}, [r0], r1- vst3.8 {d18[6], d19[6], d20[6]}, [r8], r1- vst3.8 {d21[6], d22[6], d23[6]}, [r0], r1- vst3.8 {d18[7], d19[7], d20[7]}, [r8], r1- vst3.8 {d21[7], d22[7], d23[7]}, [r0], r1-
- b v_end
-
-v_wide_mbfilter
- sub r8, r0, #8
-
- vtrn.32 d0, d26
- vtrn.32 d16, d27
- vtrn.32 d24, d18
- vtrn.32 d25, d19
-
- vtrn.16 d0, d24
- vtrn.16 d16, d25
- vtrn.16 d26, d18
- vtrn.16 d27, d19
-
- vtrn.8 d0, d16
- vtrn.8 d24, d25
- vtrn.8 d26, d27
- vtrn.8 d18, d19
-
- vtrn.32 d20, d1
- vtrn.32 d21, d2
- vtrn.32 d22, d3
- vtrn.32 d23, d15
-
- vtrn.16 d20, d22
- vtrn.16 d21, d23
- vtrn.16 d1, d3
- vtrn.16 d2, d15
-
- vtrn.8 d20, d21
- vtrn.8 d22, d23
- vtrn.8 d1, d2
- vtrn.8 d3, d15
-
- vst1.8 {d0}, [r8@64], r1- vst1.8 {d20}, [r0@64], r1- vst1.8 {d16}, [r8@64], r1- vst1.8 {d21}, [r0@64], r1- vst1.8 {d24}, [r8@64], r1- vst1.8 {d22}, [r0@64], r1- vst1.8 {d25}, [r8@64], r1- vst1.8 {d23}, [r0@64], r1- vst1.8 {d26}, [r8@64], r1- vst1.8 {d1}, [r0@64], r1- vst1.8 {d27}, [r8@64], r1- vst1.8 {d2}, [r0@64], r1- vst1.8 {d18}, [r8@64], r1- vst1.8 {d3}, [r0@64], r1- vst1.8 {d19}, [r8@64], r1- vst1.8 {d15}, [r0@64], r1-
-v_end
- vpop {d8-d15}- pop {r4-r8, pc}-
- ENDP ; |vp9_lpf_vertical_16_neon|
-
-; void vp9_wide_mbfilter_neon();
-; This is a helper function for the loopfilters. The invidual functions do the
-; necessary load, transpose (if necessary) and store.
-;
-; r0-r3 PRESERVE
-; d16 blimit
-; d17 limit
-; d18 thresh
-; d0 p7
-; d1 p6
-; d2 p5
-; d3 p4
-; d4 p3
-; d5 p2
-; d6 p1
-; d7 p0
-; d8 q0
-; d9 q1
-; d10 q2
-; d11 q3
-; d12 q4
-; d13 q5
-; d14 q6
-; d15 q7
-|vp9_wide_mbfilter_neon| PROC
- mov r7, #0
-
- ; filter_mask
- vabd.u8 d19, d4, d5 ; abs(p3 - p2)
- vabd.u8 d20, d5, d6 ; abs(p2 - p1)
- vabd.u8 d21, d6, d7 ; abs(p1 - p0)
- vabd.u8 d22, d9, d8 ; abs(q1 - q0)
- vabd.u8 d23, d10, d9 ; abs(q2 - q1)
- vabd.u8 d24, d11, d10 ; abs(q3 - q2)
-
- ; only compare the largest value to limit
- vmax.u8 d19, d19, d20 ; max(abs(p3 - p2), abs(p2 - p1))
- vmax.u8 d20, d21, d22 ; max(abs(p1 - p0), abs(q1 - q0))
- vmax.u8 d23, d23, d24 ; max(abs(q2 - q1), abs(q3 - q2))
- vmax.u8 d19, d19, d20
-
- vabd.u8 d24, d7, d8 ; abs(p0 - q0)
-
- vmax.u8 d19, d19, d23
-
- vabd.u8 d23, d6, d9 ; a = abs(p1 - q1)
- vqadd.u8 d24, d24, d24 ; b = abs(p0 - q0) * 2
-
- ; abs () > limit
- vcge.u8 d19, d17, d19
-
- ; flatmask4
- vabd.u8 d25, d7, d5 ; abs(p0 - p2)
- vabd.u8 d26, d8, d10 ; abs(q0 - q2)
- vabd.u8 d27, d4, d7 ; abs(p3 - p0)
- vabd.u8 d28, d11, d8 ; abs(q3 - q0)
-
- ; only compare the largest value to thresh
- vmax.u8 d25, d25, d26 ; max(abs(p0 - p2), abs(q0 - q2))
- vmax.u8 d26, d27, d28 ; max(abs(p3 - p0), abs(q3 - q0))
- vmax.u8 d25, d25, d26
- vmax.u8 d20, d20, d25
-
- vshr.u8 d23, d23, #1 ; a = a / 2
- vqadd.u8 d24, d24, d23 ; a = b + a
-
- vmov.u8 d30, #1
- vcge.u8 d24, d16, d24 ; (a > blimit * 2 + limit) * -1
-
- vcge.u8 d20, d30, d20 ; flat
-
- vand d19, d19, d24 ; mask
-
- ; hevmask
- vcgt.u8 d21, d21, d18 ; (abs(p1 - p0) > thresh)*-1
- vcgt.u8 d22, d22, d18 ; (abs(q1 - q0) > thresh)*-1
- vorr d21, d21, d22 ; hev
-
- vand d16, d20, d19 ; flat && mask
- vmov r5, r6, d16
-
- ; flatmask5(1, p7, p6, p5, p4, p0, q0, q4, q5, q6, q7)
- vabd.u8 d22, d3, d7 ; abs(p4 - p0)
- vabd.u8 d23, d12, d8 ; abs(q4 - q0)
- vabd.u8 d24, d7, d2 ; abs(p0 - p5)
- vabd.u8 d25, d8, d13 ; abs(q0 - q5)
- vabd.u8 d26, d1, d7 ; abs(p6 - p0)
- vabd.u8 d27, d14, d8 ; abs(q6 - q0)
- vabd.u8 d28, d0, d7 ; abs(p7 - p0)
- vabd.u8 d29, d15, d8 ; abs(q7 - q0)
-
- ; only compare the largest value to thresh
- vmax.u8 d22, d22, d23 ; max(abs(p4 - p0), abs(q4 - q0))
- vmax.u8 d23, d24, d25 ; max(abs(p0 - p5), abs(q0 - q5))
- vmax.u8 d24, d26, d27 ; max(abs(p6 - p0), abs(q6 - q0))
- vmax.u8 d25, d28, d29 ; max(abs(p7 - p0), abs(q7 - q0))
-
- vmax.u8 d26, d22, d23
- vmax.u8 d27, d24, d25
- vmax.u8 d23, d26, d27
-
- vcge.u8 d18, d30, d23 ; flat2
-
- vmov.u8 d22, #0x80
-
- orrs r5, r5, r6 ; Check for 0
- orreq r7, r7, #1 ; Only do filter branch
-
- vand d17, d18, d16 ; flat2 && flat && mask
- vmov r5, r6, d17
-
- ; mbfilter() function
-
- ; filter() function
- ; convert to signed
- veor d23, d8, d22 ; qs0
- veor d24, d7, d22 ; ps0
- veor d25, d6, d22 ; ps1
- veor d26, d9, d22 ; qs1
-
- vmov.u8 d27, #3
-
- vsub.s8 d28, d23, d24 ; ( qs0 - ps0)
- vqsub.s8 d29, d25, d26 ; filter = clamp(ps1-qs1)
- vmull.s8 q15, d28, d27 ; 3 * ( qs0 - ps0)
- vand d29, d29, d21 ; filter &= hev
- vaddw.s8 q15, q15, d29 ; filter + 3 * (qs0 - ps0)
- vmov.u8 d29, #4
-
- ; filter = clamp(filter + 3 * ( qs0 - ps0))
- vqmovn.s16 d28, q15
-
- vand d28, d28, d19 ; filter &= mask
-
- vqadd.s8 d30, d28, d27 ; filter2 = clamp(filter+3)
- vqadd.s8 d29, d28, d29 ; filter1 = clamp(filter+4)
- vshr.s8 d30, d30, #3 ; filter2 >>= 3
- vshr.s8 d29, d29, #3 ; filter1 >>= 3
-
-
- vqadd.s8 d24, d24, d30 ; op0 = clamp(ps0 + filter2)
- vqsub.s8 d23, d23, d29 ; oq0 = clamp(qs0 - filter1)
-
- ; outer tap adjustments: ++filter1 >> 1
- vrshr.s8 d29, d29, #1
- vbic d29, d29, d21 ; filter &= ~hev
-
- vqadd.s8 d25, d25, d29 ; op1 = clamp(ps1 + filter)
- vqsub.s8 d26, d26, d29 ; oq1 = clamp(qs1 - filter)
-
- veor d24, d24, d22 ; *f_op0 = u^0x80
- veor d23, d23, d22 ; *f_oq0 = u^0x80
- veor d25, d25, d22 ; *f_op1 = u^0x80
- veor d26, d26, d22 ; *f_oq1 = u^0x80
-
- tst r7, #1
- bxne lr
-
- orrs r5, r5, r6 ; Check for 0
- orreq r7, r7, #2 ; Only do mbfilter branch
-
- ; mbfilter flat && mask branch
- ; TODO(fgalligan): Can I decrease the cycles shifting to consective d's
- ; and using vibt on the q's?
- vmov.u8 d29, #2
- vaddl.u8 q15, d7, d8 ; op2 = p0 + q0
- vmlal.u8 q15, d4, d27 ; op2 = p0 + q0 + p3 * 3
- vmlal.u8 q15, d5, d29 ; op2 = p0 + q0 + p3 * 3 + p2 * 2
- vaddl.u8 q10, d4, d5
- vaddw.u8 q15, d6 ; op2=p1 + p0 + q0 + p3 * 3 + p2 *2
- vaddl.u8 q14, d6, d9
- vqrshrn.u16 d18, q15, #3 ; r_op2
-
- vsub.i16 q15, q10
- vaddl.u8 q10, d4, d6
- vadd.i16 q15, q14
- vaddl.u8 q14, d7, d10
- vqrshrn.u16 d19, q15, #3 ; r_op1
-
- vsub.i16 q15, q10
- vadd.i16 q15, q14
- vaddl.u8 q14, d8, d11
- vqrshrn.u16 d20, q15, #3 ; r_op0
-
- vsubw.u8 q15, d4 ; oq0 = op0 - p3
- vsubw.u8 q15, d7 ; oq0 -= p0
- vadd.i16 q15, q14
- vaddl.u8 q14, d9, d11
- vqrshrn.u16 d21, q15, #3 ; r_oq0
-
- vsubw.u8 q15, d5 ; oq1 = oq0 - p2
- vsubw.u8 q15, d8 ; oq1 -= q0
- vadd.i16 q15, q14
- vaddl.u8 q14, d10, d11
- vqrshrn.u16 d22, q15, #3 ; r_oq1
-
- vsubw.u8 q15, d6 ; oq2 = oq0 - p1
- vsubw.u8 q15, d9 ; oq2 -= q1
- vadd.i16 q15, q14
- vqrshrn.u16 d27, q15, #3 ; r_oq2
-
- ; Filter does not set op2 or oq2, so use p2 and q2.
- vbif d18, d5, d16 ; t_op2 |= p2 & ~(flat & mask)
- vbif d19, d25, d16 ; t_op1 |= f_op1 & ~(flat & mask)
- vbif d20, d24, d16 ; t_op0 |= f_op0 & ~(flat & mask)
- vbif d21, d23, d16 ; t_oq0 |= f_oq0 & ~(flat & mask)
- vbif d22, d26, d16 ; t_oq1 |= f_oq1 & ~(flat & mask)
-
- vbit d23, d27, d16 ; t_oq2 |= r_oq2 & (flat & mask)
- vbif d23, d10, d16 ; t_oq2 |= q2 & ~(flat & mask)
-
- tst r7, #2
- bxne lr
-
- ; wide_mbfilter flat2 && flat && mask branch
- vmov.u8 d16, #7
- vaddl.u8 q15, d7, d8 ; op6 = p0 + q0
- vaddl.u8 q12, d2, d3
- vaddl.u8 q13, d4, d5
- vaddl.u8 q14, d1, d6
- vmlal.u8 q15, d0, d16 ; op6 += p7 * 3
- vadd.i16 q12, q13
- vadd.i16 q15, q14
- vaddl.u8 q14, d2, d9
- vadd.i16 q15, q12
- vaddl.u8 q12, d0, d1
- vaddw.u8 q15, d1
- vaddl.u8 q13, d0, d2
- vadd.i16 q14, q15, q14
- vqrshrn.u16 d16, q15, #4 ; w_op6
-
- vsub.i16 q15, q14, q12
- vaddl.u8 q14, d3, d10
- vqrshrn.u16 d24, q15, #4 ; w_op5
-
- vsub.i16 q15, q13
- vaddl.u8 q13, d0, d3
- vadd.i16 q15, q14
- vaddl.u8 q14, d4, d11
- vqrshrn.u16 d25, q15, #4 ; w_op4
-
- vadd.i16 q15, q14
- vaddl.u8 q14, d0, d4
- vsub.i16 q15, q13
- vsub.i16 q14, q15, q14
- vqrshrn.u16 d26, q15, #4 ; w_op3
-
- vaddw.u8 q15, q14, d5 ; op2 += p2
- vaddl.u8 q14, d0, d5
- vaddw.u8 q15, d12 ; op2 += q4
- vbif d26, d4, d17 ; op3 |= p3 & ~(f2 & f & m)
- vqrshrn.u16 d27, q15, #4 ; w_op2
-
- vsub.i16 q15, q14
- vaddl.u8 q14, d0, d6
- vaddw.u8 q15, d6 ; op1 += p1
- vaddw.u8 q15, d13 ; op1 += q5
- vbif d27, d18, d17 ; op2 |= t_op2 & ~(f2 & f & m)
- vqrshrn.u16 d18, q15, #4 ; w_op1
-
- vsub.i16 q15, q14
- vaddl.u8 q14, d0, d7
- vaddw.u8 q15, d7 ; op0 += p0
- vaddw.u8 q15, d14 ; op0 += q6
- vbif d18, d19, d17 ; op1 |= t_op1 & ~(f2 & f & m)
- vqrshrn.u16 d19, q15, #4 ; w_op0
-
- vsub.i16 q15, q14
- vaddl.u8 q14, d1, d8
- vaddw.u8 q15, d8 ; oq0 += q0
- vaddw.u8 q15, d15 ; oq0 += q7
- vbif d19, d20, d17 ; op0 |= t_op0 & ~(f2 & f & m)
- vqrshrn.u16 d20, q15, #4 ; w_oq0
-
- vsub.i16 q15, q14
- vaddl.u8 q14, d2, d9
- vaddw.u8 q15, d9 ; oq1 += q1
- vaddl.u8 q4, d10, d15
- vaddw.u8 q15, d15 ; oq1 += q7
- vbif d20, d21, d17 ; oq0 |= t_oq0 & ~(f2 & f & m)
- vqrshrn.u16 d21, q15, #4 ; w_oq1
-
- vsub.i16 q15, q14
- vaddl.u8 q14, d3, d10
- vadd.i16 q15, q4
- vaddl.u8 q4, d11, d15
- vbif d21, d22, d17 ; oq1 |= t_oq1 & ~(f2 & f & m)
- vqrshrn.u16 d22, q15, #4 ; w_oq2
-
- vsub.i16 q15, q14
- vaddl.u8 q14, d4, d11
- vadd.i16 q15, q4
- vaddl.u8 q4, d12, d15
- vbif d22, d23, d17 ; oq2 |= t_oq2 & ~(f2 & f & m)
- vqrshrn.u16 d23, q15, #4 ; w_oq3
-
- vsub.i16 q15, q14
- vaddl.u8 q14, d5, d12
- vadd.i16 q15, q4
- vaddl.u8 q4, d13, d15
- vbif d16, d1, d17 ; op6 |= p6 & ~(f2 & f & m)
- vqrshrn.u16 d1, q15, #4 ; w_oq4
-
- vsub.i16 q15, q14
- vaddl.u8 q14, d6, d13
- vadd.i16 q15, q4
- vaddl.u8 q4, d14, d15
- vbif d24, d2, d17 ; op5 |= p5 & ~(f2 & f & m)
- vqrshrn.u16 d2, q15, #4 ; w_oq5
-
- vsub.i16 q15, q14
- vbif d25, d3, d17 ; op4 |= p4 & ~(f2 & f & m)
- vadd.i16 q15, q4
- vbif d23, d11, d17 ; oq3 |= q3 & ~(f2 & f & m)
- vqrshrn.u16 d3, q15, #4 ; w_oq6
- vbif d1, d12, d17 ; oq4 |= q4 & ~(f2 & f & m)
- vbif d2, d13, d17 ; oq5 |= q5 & ~(f2 & f & m)
- vbif d3, d14, d17 ; oq6 |= q6 & ~(f2 & f & m)
-
- bx lr
- ENDP ; |vp9_wide_mbfilter_neon|
-
- END
--- a/vp9/common/mips/msa/vp9_loopfilter_16_msa.c
+++ /dev/null
@@ -1,1480 +1,0 @@
-/*
- * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "vpx_ports/mem.h"
-#include "vp9/common/mips/msa/vp9_loopfilter_msa.h"
-
-int32_t vp9_hz_lpf_t4_and_t8_16w(uint8_t *src, int32_t pitch,
- uint8_t *filter48,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr) {- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
- v16u8 flat, mask, hev, thresh, b_limit, limit;
- v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
- v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
- v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
- v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
- v16u8 zero = { 0 };-
- /* load vector elements */
- LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh_ptr);
- b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
- limit = (v16u8)__msa_fill_b(*limit_ptr);
-
- /* mask and hev */
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
- VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
-
- if (__msa_test_bz_v(flat)) {- ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
-
- return 1;
- } else {- ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
- zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
- q2_r, q3_r);
- VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
- p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
-
- ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
- ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
- VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
- p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
-
- /* convert 16 bit output data into 8 bit */
- PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
- p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
- p0_filt8_r, q0_filt8_r);
- PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
- q2_filt8_r);
-
- /* store pixel values */
- p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
- p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
- p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
- q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
- q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
- q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
-
- ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
- filter48 += (4 * 16);
- ST_UB2(q1_out, q2_out, filter48, 16);
- filter48 += (2 * 16);
- ST_UB(flat, filter48);
-
- return 0;
- }
-}
-
-void vp9_hz_lpf_t16_16w(uint8_t *src, int32_t pitch, uint8_t *filter48) {- v16u8 flat, flat2, filter8;
- v16i8 zero = { 0 };- v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
- v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
- v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
- v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in;
- v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in;
- v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l;
- v8i16 l_out, r_out;
-
- flat = LD_UB(filter48 + 96);
-
- LD_UB8((src - 8 * pitch), pitch, p7, p6, p5, p4, p3, p2, p1, p0);
- LD_UB8(src, pitch, q0, q1, q2, q3, q4, q5, q6, q7);
- VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
-
- if (__msa_test_bz_v(flat2)) {- LD_UB4(filter48, 16, p2, p1, p0, q0);
- LD_UB2(filter48 + 4 * 16, 16, q1, q2);
-
- src -= 3 * pitch;
- ST_UB4(p2, p1, p0, q0, src, pitch);
- src += (4 * pitch);
- ST_UB2(q1, q2, src, pitch);
- } else {- src -= 7 * pitch;
-
- ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
- zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in,
- p2_r_in, p1_r_in, p0_r_in);
-
- q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
-
- tmp0_r = p7_r_in << 3;
- tmp0_r -= p7_r_in;
- tmp0_r += p6_r_in;
- tmp0_r += q0_r_in;
- tmp1_r = p6_r_in + p5_r_in;
- tmp1_r += p4_r_in;
- tmp1_r += p3_r_in;
- tmp1_r += p2_r_in;
- tmp1_r += p1_r_in;
- tmp1_r += p0_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in,
- p5_l_in, p4_l_in);
- ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in,
- p1_l_in, p0_l_in);
- q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0);
-
- tmp0_l = p7_l_in << 3;
- tmp0_l -= p7_l_in;
- tmp0_l += p6_l_in;
- tmp0_l += q0_l_in;
- tmp1_l = p6_l_in + p5_l_in;
- tmp1_l += p4_l_in;
- tmp1_l += p3_l_in;
- tmp1_l += p2_l_in;
- tmp1_l += p1_l_in;
- tmp1_l += p0_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
- ST_UB(p6, src);
- src += pitch;
-
- /* p5 */
- q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
- tmp0_r = p5_r_in - p6_r_in;
- tmp0_r += q1_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1);
- tmp0_l = p5_l_in - p6_l_in;
- tmp0_l += q1_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
- ST_UB(p5, src);
- src += pitch;
-
- /* p4 */
- q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
- tmp0_r = p4_r_in - p5_r_in;
- tmp0_r += q2_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = (v8i16)__msa_srari_h((v8i16)tmp1_r, 4);
-
- q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2);
- tmp0_l = p4_l_in - p5_l_in;
- tmp0_l += q2_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
- ST_UB(p4, src);
- src += pitch;
-
- /* p3 */
- q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
- tmp0_r = p3_r_in - p4_r_in;
- tmp0_r += q3_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3);
- tmp0_l = p3_l_in - p4_l_in;
- tmp0_l += q3_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
- ST_UB(p3, src);
- src += pitch;
-
- /* p2 */
- q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
- filter8 = LD_UB(filter48);
- tmp0_r = p2_r_in - p3_r_in;
- tmp0_r += q4_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4);
- tmp0_l = p2_l_in - p3_l_in;
- tmp0_l += q4_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += pitch;
-
- /* p1 */
- q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
- filter8 = LD_UB(filter48 + 16);
- tmp0_r = p1_r_in - p2_r_in;
- tmp0_r += q5_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5);
- tmp0_l = p1_l_in - p2_l_in;
- tmp0_l += q5_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += pitch;
-
- /* p0 */
- q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
- filter8 = LD_UB(filter48 + 32);
- tmp0_r = p0_r_in - p1_r_in;
- tmp0_r += q6_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6);
- tmp0_l = p0_l_in - p1_l_in;
- tmp0_l += q6_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += pitch;
-
- /* q0 */
- q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
- filter8 = LD_UB(filter48 + 48);
- tmp0_r = q7_r_in - p0_r_in;
- tmp0_r += q0_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7);
- tmp0_l = q7_l_in - p0_l_in;
- tmp0_l += q0_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += pitch;
-
- /* q1 */
- filter8 = LD_UB(filter48 + 64);
- tmp0_r = q7_r_in - q0_r_in;
- tmp0_r += q1_r_in;
- tmp0_r -= p6_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- tmp0_l = q7_l_in - q0_l_in;
- tmp0_l += q1_l_in;
- tmp0_l -= p6_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += pitch;
-
- /* q2 */
- filter8 = LD_UB(filter48 + 80);
- tmp0_r = q7_r_in - q1_r_in;
- tmp0_r += q2_r_in;
- tmp0_r -= p5_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- tmp0_l = q7_l_in - q1_l_in;
- tmp0_l += q2_l_in;
- tmp0_l -= p5_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += pitch;
-
- /* q3 */
- tmp0_r = q7_r_in - q2_r_in;
- tmp0_r += q3_r_in;
- tmp0_r -= p4_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- tmp0_l = q7_l_in - q2_l_in;
- tmp0_l += q3_l_in;
- tmp0_l -= p4_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
- ST_UB(q3, src);
- src += pitch;
-
- /* q4 */
- tmp0_r = q7_r_in - q3_r_in;
- tmp0_r += q4_r_in;
- tmp0_r -= p3_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- tmp0_l = q7_l_in - q3_l_in;
- tmp0_l += q4_l_in;
- tmp0_l -= p3_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
- ST_UB(q4, src);
- src += pitch;
-
- /* q5 */
- tmp0_r = q7_r_in - q4_r_in;
- tmp0_r += q5_r_in;
- tmp0_r -= p2_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- tmp0_l = q7_l_in - q4_l_in;
- tmp0_l += q5_l_in;
- tmp0_l -= p2_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
- ST_UB(q5, src);
- src += pitch;
-
- /* q6 */
- tmp0_r = q7_r_in - q5_r_in;
- tmp0_r += q6_r_in;
- tmp0_r -= p1_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- tmp0_l = q7_l_in - q5_l_in;
- tmp0_l += q6_l_in;
- tmp0_l -= p1_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
- ST_UB(q6, src);
- }
-}
-
-void vp9_lpf_horizontal_16_dual_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr,
- int32_t count) {- DECLARE_ALIGNED(32, uint8_t, filter48[16 * 8]);
- uint8_t early_exit = 0;
-
- (void)count;
-
- early_exit = vp9_hz_lpf_t4_and_t8_16w(src, pitch, &filter48[0], b_limit_ptr,
- limit_ptr, thresh_ptr);
-
- if (0 == early_exit) {- vp9_hz_lpf_t16_16w(src, pitch, filter48);
- }
-}
-
-void vp9_lpf_horizontal_16_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr,
- int32_t count) {- if (1 == count) {- uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
- uint64_t dword0, dword1;
- v16u8 flat2, mask, hev, flat, thresh, b_limit, limit;
- v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p7, p6, p5, p4, q4, q5, q6, q7;
- v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
- v16u8 p0_filter16, p1_filter16;
- v8i16 p2_filter8, p1_filter8, p0_filter8;
- v8i16 q0_filter8, q1_filter8, q2_filter8;
- v8u16 p7_r, p6_r, p5_r, p4_r, q7_r, q6_r, q5_r, q4_r;
- v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
- v16i8 zero = { 0 };- v8u16 tmp0, tmp1, tmp2;
-
- /* load vector elements */
- LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh_ptr);
- b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
- limit = (v16u8)__msa_fill_b(*limit_ptr);
-
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
- VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out,
- q1_out);
-
- flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
-
- if (__msa_test_bz_v(flat)) {- p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
- p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
- q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
- q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
- SD4(p1_d, p0_d, q0_d, q1_d, src - 2 * pitch, pitch);
- } else {- /* convert 8 bit input data into 16 bit */
- ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
- zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
- q3_r);
- VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
- p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
-
- /* convert 16 bit output data into 8 bit */
- PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8,
- zero, q0_filter8, p2_filter8, p1_filter8, p0_filter8,
- q0_filter8);
- PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
-
- /* store pixel values */
- p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
- p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
- p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
- q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
- q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
- q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
-
- /* load 16 vector elements */
- LD_UB4((src - 8 * pitch), pitch, p7, p6, p5, p4);
- LD_UB4(src + (4 * pitch), pitch, q4, q5, q6, q7);
-
- VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
-
- if (__msa_test_bz_v(flat2)) {- p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
- p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
- p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
- q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
- q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
- q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
-
- SD4(p2_d, p1_d, p0_d, q0_d, src - 3 * pitch, pitch);
- SD(q1_d, src + pitch);
- SD(q2_d, src + 2 * pitch);
- } else {- /* LSB(right) 8 pixel operation */
- ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, q4, zero, q5,
- zero, q6, zero, q7, p7_r, p6_r, p5_r, p4_r, q4_r, q5_r, q6_r,
- q7_r);
-
- tmp0 = p7_r << 3;
- tmp0 -= p7_r;
- tmp0 += p6_r;
- tmp0 += q0_r;
-
- src -= 7 * pitch;
-
- /* calculation of p6 and p5 */
- tmp1 = p6_r + p5_r + p4_r + p3_r;
- tmp1 += (p2_r + p1_r + p0_r);
- tmp1 += tmp0;
- p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- tmp0 = p5_r - p6_r + q1_r - p7_r;
- tmp1 += tmp0;
- p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
- p1_filter16);
- p0_filter16 = __msa_bmnz_v(p6, p0_filter16, flat2);
- p1_filter16 = __msa_bmnz_v(p5, p1_filter16, flat2);
- dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
- dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
- SD(dword0, src);
- src += pitch;
- SD(dword1, src);
- src += pitch;
-
- /* calculation of p4 and p3 */
- tmp0 = p4_r - p5_r + q2_r - p7_r;
- tmp2 = p3_r - p4_r + q3_r - p7_r;
- tmp1 += tmp0;
- p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- tmp1 += tmp2;
- p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
- p1_filter16);
- p0_filter16 = __msa_bmnz_v(p4, p0_filter16, flat2);
- p1_filter16 = __msa_bmnz_v(p3, p1_filter16, flat2);
- dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
- dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
- SD(dword0, src);
- src += pitch;
- SD(dword1, src);
- src += pitch;
-
- /* calculation of p2 and p1 */
- tmp0 = p2_r - p3_r + q4_r - p7_r;
- tmp2 = p1_r - p2_r + q5_r - p7_r;
- tmp1 += tmp0;
- p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- tmp1 += tmp2;
- p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
- p1_filter16);
- p0_filter16 = __msa_bmnz_v(p2_out, p0_filter16, flat2);
- p1_filter16 = __msa_bmnz_v(p1_out, p1_filter16, flat2);
- dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
- dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
- SD(dword0, src);
- src += pitch;
- SD(dword1, src);
- src += pitch;
-
- /* calculation of p0 and q0 */
- tmp0 = (p0_r - p1_r) + (q6_r - p7_r);
- tmp2 = (q7_r - p0_r) + (q0_r - p7_r);
- tmp1 += tmp0;
- p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- tmp1 += tmp2;
- p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
- p1_filter16);
- p0_filter16 = __msa_bmnz_v(p0_out, p0_filter16, flat2);
- p1_filter16 = __msa_bmnz_v(q0_out, p1_filter16, flat2);
- dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
- dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
- SD(dword0, src);
- src += pitch;
- SD(dword1, src);
- src += pitch;
-
- /* calculation of q1 and q2 */
- tmp0 = q7_r - q0_r + q1_r - p6_r;
- tmp2 = q7_r - q1_r + q2_r - p5_r;
- tmp1 += tmp0;
- p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- tmp1 += tmp2;
- p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
- p1_filter16);
- p0_filter16 = __msa_bmnz_v(q1_out, p0_filter16, flat2);
- p1_filter16 = __msa_bmnz_v(q2_out, p1_filter16, flat2);
- dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
- dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
- SD(dword0, src);
- src += pitch;
- SD(dword1, src);
- src += pitch;
-
- /* calculation of q3 and q4 */
- tmp0 = (q7_r - q2_r) + (q3_r - p4_r);
- tmp2 = (q7_r - q3_r) + (q4_r - p3_r);
- tmp1 += tmp0;
- p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- tmp1 += tmp2;
- p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
- p1_filter16);
- p0_filter16 = __msa_bmnz_v(q3, p0_filter16, flat2);
- p1_filter16 = __msa_bmnz_v(q4, p1_filter16, flat2);
- dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
- dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
- SD(dword0, src);
- src += pitch;
- SD(dword1, src);
- src += pitch;
-
- /* calculation of q5 and q6 */
- tmp0 = (q7_r - q4_r) + (q5_r - p2_r);
- tmp2 = (q7_r - q5_r) + (q6_r - p1_r);
- tmp1 += tmp0;
- p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- tmp1 += tmp2;
- p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
- PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
- p1_filter16);
- p0_filter16 = __msa_bmnz_v(q5, p0_filter16, flat2);
- p1_filter16 = __msa_bmnz_v(q6, p1_filter16, flat2);
- dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
- dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
- SD(dword0, src);
- src += pitch;
- SD(dword1, src);
- }
- }
- } else {- vp9_lpf_horizontal_16_dual_msa(src, pitch, b_limit_ptr, limit_ptr,
- thresh_ptr, count);
- }
-}
-
-static void vp9_transpose_16x8_to_8x16(uint8_t *input, int32_t in_pitch,
- uint8_t *output, int32_t out_pitch) {- v16u8 p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, p0_org;
- v16i8 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
- v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
-
- LD_UB8(input, in_pitch,
- p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, p0_org);
- /* 8x8 transpose */
- TRANSPOSE8x8_UB_UB(p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org,
- p0_org, p7, p6, p5, p4, p3, p2, p1, p0);
- /* 8x8 transpose */
- ILVL_B4_SB(p5_org, p7_org, p4_org, p6_org, p1_org, p3_org, p0_org, p2_org,
- tmp0, tmp1, tmp2, tmp3);
- ILVR_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp4, tmp6);
- ILVL_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp5, tmp7);
- ILVR_W2_UB(tmp6, tmp4, tmp7, tmp5, q0, q4);
- ILVL_W2_UB(tmp6, tmp4, tmp7, tmp5, q2, q6);
- SLDI_B4_0_UB(q0, q2, q4, q6, q1, q3, q5, q7, 8);
-
- ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch);
- output += (8 * out_pitch);
- ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch);
-}
-
-static void vp9_transpose_8x16_to_16x8(uint8_t *input, int32_t in_pitch,
- uint8_t *output, int32_t out_pitch) {- v16u8 p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o;
- v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
-
- LD_UB8(input, in_pitch, p7, p6, p5, p4, p3, p2, p1, p0);
- LD_UB8(input + (8 * in_pitch), in_pitch, q0, q1, q2, q3, q4, q5, q6, q7);
- TRANSPOSE16x8_UB_UB(p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5,
- q6, q7, p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o);
- ST_UB8(p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o, output, out_pitch);
-}
-
-static void vp9_transpose_16x16(uint8_t *input, int32_t in_pitch,
- uint8_t *output, int32_t out_pitch) {- v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
- v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
- v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
- v8i16 tmp0, tmp1, tmp4, tmp5, tmp6, tmp7;
- v4i32 tmp2, tmp3;
-
- LD_UB8(input, in_pitch, row0, row1, row2, row3, row4, row5, row6, row7);
- input += (8 * in_pitch);
- LD_UB8(input, in_pitch,
- row8, row9, row10, row11, row12, row13, row14, row15);
-
- TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
- row8, row9, row10, row11, row12, row13, row14, row15,
- p7, p6, p5, p4, p3, p2, p1, p0);
-
- /* transpose 16x8 matrix into 8x16 */
- /* total 8 intermediate register and 32 instructions */
- q7 = (v16u8)__msa_ilvod_d((v2i64)row8, (v2i64)row0);
- q6 = (v16u8)__msa_ilvod_d((v2i64)row9, (v2i64)row1);
- q5 = (v16u8)__msa_ilvod_d((v2i64)row10, (v2i64)row2);
- q4 = (v16u8)__msa_ilvod_d((v2i64)row11, (v2i64)row3);
- q3 = (v16u8)__msa_ilvod_d((v2i64)row12, (v2i64)row4);
- q2 = (v16u8)__msa_ilvod_d((v2i64)row13, (v2i64)row5);
- q1 = (v16u8)__msa_ilvod_d((v2i64)row14, (v2i64)row6);
- q0 = (v16u8)__msa_ilvod_d((v2i64)row15, (v2i64)row7);
-
- ILVEV_B2_SH(q7, q6, q5, q4, tmp0, tmp1);
- tmp4 = (v8i16)__msa_ilvod_b((v16i8)q6, (v16i8)q7);
- tmp5 = (v8i16)__msa_ilvod_b((v16i8)q4, (v16i8)q5);
-
- ILVEV_B2_UB(q3, q2, q1, q0, q5, q7);
- tmp6 = (v8i16)__msa_ilvod_b((v16i8)q2, (v16i8)q3);
- tmp7 = (v8i16)__msa_ilvod_b((v16i8)q0, (v16i8)q1);
-
- ILVEV_H2_SW(tmp0, tmp1, q5, q7, tmp2, tmp3);
- q0 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
- q4 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
-
- tmp2 = (v4i32)__msa_ilvod_h(tmp1, tmp0);
- tmp3 = (v4i32)__msa_ilvod_h((v8i16)q7, (v8i16)q5);
- q2 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
- q6 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
-
- ILVEV_H2_SW(tmp4, tmp5, tmp6, tmp7, tmp2, tmp3);
- q1 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
- q5 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
-
- tmp2 = (v4i32)__msa_ilvod_h(tmp5, tmp4);
- tmp3 = (v4i32)__msa_ilvod_h(tmp7, tmp6);
- q3 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
- q7 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
-
- ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch);
- output += (8 * out_pitch);
- ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch);
-}
-
-int32_t vp9_vt_lpf_t4_and_t8_8w(uint8_t *src, uint8_t *filter48,
- uint8_t *src_org, int32_t pitch_org,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr) {- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
- v16u8 flat, mask, hev, thresh, b_limit, limit;
- v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
- v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
- v16i8 zero = { 0 };- v8i16 vec0, vec1, vec2, vec3;
-
- /* load vector elements */
- LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh_ptr);
- b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
- limit = (v16u8)__msa_fill_b(*limit_ptr);
-
- /* mask and hev */
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- /* flat4 */
- VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
- /* filter4 */
- VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
-
- flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
-
- if (__msa_test_bz_v(flat)) {- ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec2, vec3);
- ST4x8_UB(vec2, vec3, (src_org - 2), pitch_org);
- return 1;
- } else {- ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
- zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
- q3_r);
- VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
- p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
-
- /* convert 16 bit output data into 8 bit */
- p2_r = (v8u16)__msa_pckev_b((v16i8)p2_filt8_r, (v16i8)p2_filt8_r);
- p1_r = (v8u16)__msa_pckev_b((v16i8)p1_filt8_r, (v16i8)p1_filt8_r);
- p0_r = (v8u16)__msa_pckev_b((v16i8)p0_filt8_r, (v16i8)p0_filt8_r);
- q0_r = (v8u16)__msa_pckev_b((v16i8)q0_filt8_r, (v16i8)q0_filt8_r);
- q1_r = (v8u16)__msa_pckev_b((v16i8)q1_filt8_r, (v16i8)q1_filt8_r);
- q2_r = (v8u16)__msa_pckev_b((v16i8)q2_filt8_r, (v16i8)q2_filt8_r);
-
- /* store pixel values */
- p2_out = __msa_bmnz_v(p2, (v16u8)p2_r, flat);
- p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_r, flat);
- p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_r, flat);
- q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_r, flat);
- q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_r, flat);
- q2_out = __msa_bmnz_v(q2, (v16u8)q2_r, flat);
-
- ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
- filter48 += (4 * 16);
- ST_UB2(q1_out, q2_out, filter48, 16);
- filter48 += (2 * 16);
- ST_UB(flat, filter48);
-
- return 0;
- }
-}
-
-int32_t vp9_vt_lpf_t16_8w(uint8_t *src, uint8_t *src_org, int32_t pitch,
- uint8_t *filter48) {- v16i8 zero = { 0 };- v16u8 filter8, flat, flat2;
- v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
- v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
- v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
- v8u16 tmp0_r, tmp1_r;
- v8i16 r_out;
-
- flat = LD_UB(filter48 + 6 * 16);
-
- LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0);
- LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7);
-
- VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
-
- if (__msa_test_bz_v(flat2)) {- v8i16 vec0, vec1, vec2, vec3, vec4;
-
- LD_UB4(filter48, 16, p2, p1, p0, q0);
- LD_UB2(filter48 + 4 * 16, 16, q1, q2);
-
- ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec3, vec4);
- vec2 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
-
- src_org -= 3;
- ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch);
- ST2x4_UB(vec2, 0, (src_org + 4), pitch);
- src_org += (4 * pitch);
- ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch);
- ST2x4_UB(vec2, 4, (src_org + 4), pitch);
-
- return 1;
- } else {- src -= 7 * 16;
-
- ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
- zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in,
- p3_r_in, p2_r_in, p1_r_in, p0_r_in);
- q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
-
- tmp0_r = p7_r_in << 3;
- tmp0_r -= p7_r_in;
- tmp0_r += p6_r_in;
- tmp0_r += q0_r_in;
- tmp1_r = p6_r_in + p5_r_in;
- tmp1_r += p4_r_in;
- tmp1_r += p3_r_in;
- tmp1_r += p2_r_in;
- tmp1_r += p1_r_in;
- tmp1_r += p0_r_in;
- tmp1_r += tmp0_r;
-
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
- ST8x1_UB(p6, src);
- src += 16;
-
- /* p5 */
- q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
- tmp0_r = p5_r_in - p6_r_in;
- tmp0_r += q1_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
- ST8x1_UB(p5, src);
- src += 16;
-
- /* p4 */
- q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
- tmp0_r = p4_r_in - p5_r_in;
- tmp0_r += q2_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
- ST8x1_UB(p4, src);
- src += 16;
-
- /* p3 */
- q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
- tmp0_r = p3_r_in - p4_r_in;
- tmp0_r += q3_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
- ST8x1_UB(p3, src);
- src += 16;
-
- /* p2 */
- q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
- filter8 = LD_UB(filter48);
- tmp0_r = p2_r_in - p3_r_in;
- tmp0_r += q4_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST8x1_UB(filter8, src);
- src += 16;
-
- /* p1 */
- q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
- filter8 = LD_UB(filter48 + 16);
- tmp0_r = p1_r_in - p2_r_in;
- tmp0_r += q5_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST8x1_UB(filter8, src);
- src += 16;
-
- /* p0 */
- q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
- filter8 = LD_UB(filter48 + 32);
- tmp0_r = p0_r_in - p1_r_in;
- tmp0_r += q6_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST8x1_UB(filter8, src);
- src += 16;
-
- /* q0 */
- q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
- filter8 = LD_UB(filter48 + 48);
- tmp0_r = q7_r_in - p0_r_in;
- tmp0_r += q0_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST8x1_UB(filter8, src);
- src += 16;
-
- /* q1 */
- filter8 = LD_UB(filter48 + 64);
- tmp0_r = q7_r_in - q0_r_in;
- tmp0_r += q1_r_in;
- tmp0_r -= p6_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST8x1_UB(filter8, src);
- src += 16;
-
- /* q2 */
- filter8 = LD_UB(filter48 + 80);
- tmp0_r = q7_r_in - q1_r_in;
- tmp0_r += q2_r_in;
- tmp0_r -= p5_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST8x1_UB(filter8, src);
- src += 16;
-
- /* q3 */
- tmp0_r = q7_r_in - q2_r_in;
- tmp0_r += q3_r_in;
- tmp0_r -= p4_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
- ST8x1_UB(q3, src);
- src += 16;
-
- /* q4 */
- tmp0_r = q7_r_in - q3_r_in;
- tmp0_r += q4_r_in;
- tmp0_r -= p3_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
- ST8x1_UB(q4, src);
- src += 16;
-
- /* q5 */
- tmp0_r = q7_r_in - q4_r_in;
- tmp0_r += q5_r_in;
- tmp0_r -= p2_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
- ST8x1_UB(q5, src);
- src += 16;
-
- /* q6 */
- tmp0_r = q7_r_in - q5_r_in;
- tmp0_r += q6_r_in;
- tmp0_r -= p1_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
- q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
- ST8x1_UB(q6, src);
-
- return 0;
- }
-}
-
-void vp9_lpf_vertical_16_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr) {- uint8_t early_exit = 0;
- DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]);
- uint8_t *filter48 = &transposed_input[16 * 16];
-
- vp9_transpose_16x8_to_8x16(src - 8, pitch, transposed_input, 16);
-
- early_exit = vp9_vt_lpf_t4_and_t8_8w((transposed_input + 16 * 8),
- &filter48[0], src, pitch, b_limit_ptr,
- limit_ptr, thresh_ptr);
-
- if (0 == early_exit) {- early_exit = vp9_vt_lpf_t16_8w((transposed_input + 16 * 8), src, pitch,
- &filter48[0]);
-
- if (0 == early_exit) {- vp9_transpose_8x16_to_16x8(transposed_input, 16, src - 8, pitch);
- }
- }
-}
-
-int32_t vp9_vt_lpf_t4_and_t8_16w(uint8_t *src, uint8_t *filter48,
- uint8_t *src_org, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr) {- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
- v16u8 flat, mask, hev, thresh, b_limit, limit;
- v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
- v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
- v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
- v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
- v16i8 zero = { 0 };- v8i16 vec0, vec1, vec2, vec3, vec4, vec5;
-
- /* load vector elements */
- LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh_ptr);
- b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
- limit = (v16u8)__msa_fill_b(*limit_ptr);
-
- /* mask and hev */
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- /* flat4 */
- VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
- /* filter4 */
- VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
-
- if (__msa_test_bz_v(flat)) {- ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec2, vec3);
- ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec4, vec5);
-
- src_org -= 2;
- ST4x8_UB(vec2, vec3, src_org, pitch);
- src_org += 8 * pitch;
- ST4x8_UB(vec4, vec5, src_org, pitch);
-
- return 1;
- } else {- ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
- zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
- q3_r);
- VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
- p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
- ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
- ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
- VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
- p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
-
- /* convert 16 bit output data into 8 bit */
- PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
- p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
- p0_filt8_r, q0_filt8_r);
- PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
- q2_filt8_r);
-
- /* store pixel values */
- p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
- p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
- p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
- q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
- q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
- q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
-
- ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
- filter48 += (4 * 16);
- ST_UB2(q1_out, q2_out, filter48, 16);
- filter48 += (2 * 16);
- ST_UB(flat, filter48);
-
- return 0;
- }
-}
-
-int32_t vp9_vt_lpf_t16_16w(uint8_t *src, uint8_t *src_org, int32_t pitch,
- uint8_t *filter48) {- v16u8 flat, flat2, filter8;
- v16i8 zero = { 0 };- v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
- v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
- v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
- v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in;
- v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in;
- v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l;
- v8i16 l_out, r_out;
-
- flat = LD_UB(filter48 + 6 * 16);
-
- LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0);
- LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7);
-
- VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
-
- if (__msa_test_bz_v(flat2)) {- v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
-
- LD_UB4(filter48, 16, p2, p1, p0, q0);
- LD_UB2(filter48 + 4 * 16, 16, q1, q2);
-
- ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec3, vec4);
- ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec6, vec7);
- ILVRL_B2_SH(q2, q1, vec2, vec5);
-
- src_org -= 3;
- ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch);
- ST2x4_UB(vec2, 0, (src_org + 4), pitch);
- src_org += (4 * pitch);
- ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch);
- ST2x4_UB(vec2, 4, (src_org + 4), pitch);
- src_org += (4 * pitch);
- ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src_org, pitch);
- ST2x4_UB(vec5, 0, (src_org + 4), pitch);
- src_org += (4 * pitch);
- ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src_org, pitch);
- ST2x4_UB(vec5, 4, (src_org + 4), pitch);
-
- return 1;
- } else {- src -= 7 * 16;
-
- ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
- zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in,
- p3_r_in, p2_r_in, p1_r_in, p0_r_in);
- q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
-
- tmp0_r = p7_r_in << 3;
- tmp0_r -= p7_r_in;
- tmp0_r += p6_r_in;
- tmp0_r += q0_r_in;
- tmp1_r = p6_r_in + p5_r_in;
- tmp1_r += p4_r_in;
- tmp1_r += p3_r_in;
- tmp1_r += p2_r_in;
- tmp1_r += p1_r_in;
- tmp1_r += p0_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
-
- ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in,
- p5_l_in, p4_l_in);
- ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in,
- p1_l_in, p0_l_in);
- q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0);
-
- tmp0_l = p7_l_in << 3;
- tmp0_l -= p7_l_in;
- tmp0_l += p6_l_in;
- tmp0_l += q0_l_in;
- tmp1_l = p6_l_in + p5_l_in;
- tmp1_l += p4_l_in;
- tmp1_l += p3_l_in;
- tmp1_l += p2_l_in;
- tmp1_l += p1_l_in;
- tmp1_l += p0_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
-
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
- ST_UB(p6, src);
- src += 16;
-
- /* p5 */
- q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
- tmp0_r = p5_r_in - p6_r_in;
- tmp0_r += q1_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1);
- tmp0_l = p5_l_in - p6_l_in;
- tmp0_l += q1_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
- ST_UB(p5, src);
- src += 16;
-
- /* p4 */
- q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
- tmp0_r = p4_r_in - p5_r_in;
- tmp0_r += q2_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2);
- tmp0_l = p4_l_in - p5_l_in;
- tmp0_l += q2_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
- ST_UB(p4, src);
- src += 16;
-
- /* p3 */
- q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
- tmp0_r = p3_r_in - p4_r_in;
- tmp0_r += q3_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3);
- tmp0_l = p3_l_in - p4_l_in;
- tmp0_l += q3_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
- ST_UB(p3, src);
- src += 16;
-
- /* p2 */
- q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
- filter8 = LD_UB(filter48);
- tmp0_r = p2_r_in - p3_r_in;
- tmp0_r += q4_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4);
- tmp0_l = p2_l_in - p3_l_in;
- tmp0_l += q4_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += 16;
-
- /* p1 */
- q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
- filter8 = LD_UB(filter48 + 16);
- tmp0_r = p1_r_in - p2_r_in;
- tmp0_r += q5_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5);
- tmp0_l = p1_l_in - p2_l_in;
- tmp0_l += q5_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)(tmp1_l), 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += 16;
-
- /* p0 */
- q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
- filter8 = LD_UB(filter48 + 32);
- tmp0_r = p0_r_in - p1_r_in;
- tmp0_r += q6_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6);
- tmp0_l = p0_l_in - p1_l_in;
- tmp0_l += q6_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += 16;
-
- /* q0 */
- q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
- filter8 = LD_UB(filter48 + 48);
- tmp0_r = q7_r_in - p0_r_in;
- tmp0_r += q0_r_in;
- tmp0_r -= p7_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7);
- tmp0_l = q7_l_in - p0_l_in;
- tmp0_l += q0_l_in;
- tmp0_l -= p7_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += 16;
-
- /* q1 */
- filter8 = LD_UB(filter48 + 64);
- tmp0_r = q7_r_in - q0_r_in;
- tmp0_r += q1_r_in;
- tmp0_r -= p6_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- tmp0_l = q7_l_in - q0_l_in;
- tmp0_l += q1_l_in;
- tmp0_l -= p6_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += 16;
-
- /* q2 */
- filter8 = LD_UB(filter48 + 80);
- tmp0_r = q7_r_in - q1_r_in;
- tmp0_r += q2_r_in;
- tmp0_r -= p5_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- tmp0_l = q7_l_in - q1_l_in;
- tmp0_l += q2_l_in;
- tmp0_l -= p5_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
- ST_UB(filter8, src);
- src += 16;
-
- /* q3 */
- tmp0_r = q7_r_in - q2_r_in;
- tmp0_r += q3_r_in;
- tmp0_r -= p4_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- tmp0_l = q7_l_in - q2_l_in;
- tmp0_l += q3_l_in;
- tmp0_l -= p4_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
- ST_UB(q3, src);
- src += 16;
-
- /* q4 */
- tmp0_r = q7_r_in - q3_r_in;
- tmp0_r += q4_r_in;
- tmp0_r -= p3_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- tmp0_l = q7_l_in - q3_l_in;
- tmp0_l += q4_l_in;
- tmp0_l -= p3_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
- ST_UB(q4, src);
- src += 16;
-
- /* q5 */
- tmp0_r = q7_r_in - q4_r_in;
- tmp0_r += q5_r_in;
- tmp0_r -= p2_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- tmp0_l = q7_l_in - q4_l_in;
- tmp0_l += q5_l_in;
- tmp0_l -= p2_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
- ST_UB(q5, src);
- src += 16;
-
- /* q6 */
- tmp0_r = q7_r_in - q5_r_in;
- tmp0_r += q6_r_in;
- tmp0_r -= p1_r_in;
- tmp1_r += tmp0_r;
- r_out = __msa_srari_h((v8i16)tmp1_r, 4);
- tmp0_l = q7_l_in - q5_l_in;
- tmp0_l += q6_l_in;
- tmp0_l -= p1_l_in;
- tmp1_l += tmp0_l;
- l_out = __msa_srari_h((v8i16)tmp1_l, 4);
- r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
- q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
- ST_UB(q6, src);
-
- return 0;
- }
-}
-
-void vp9_lpf_vertical_16_dual_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr) {- uint8_t early_exit = 0;
- DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]);
- uint8_t *filter48 = &transposed_input[16 * 16];
-
- vp9_transpose_16x16((src - 8), pitch, &transposed_input[0], 16);
-
- early_exit = vp9_vt_lpf_t4_and_t8_16w((transposed_input + 16 * 8),
- &filter48[0], src, pitch, b_limit_ptr,
- limit_ptr, thresh_ptr);
-
- if (0 == early_exit) {- early_exit = vp9_vt_lpf_t16_16w((transposed_input + 16 * 8), src, pitch,
- &filter48[0]);
-
- if (0 == early_exit) {- vp9_transpose_16x16(transposed_input, 16, (src - 8), pitch);
- }
- }
-}
--- a/vp9/common/mips/msa/vp9_loopfilter_4_msa.c
+++ /dev/null
@@ -1,152 +1,0 @@
-/*
- * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "vp9/common/mips/msa/vp9_loopfilter_msa.h"
-
-void vp9_lpf_horizontal_4_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr,
- int32_t count) {- uint64_t p1_d, p0_d, q0_d, q1_d;
- v16u8 mask, hev, flat, thresh, b_limit, limit;
- v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p1_out, p0_out, q0_out, q1_out;
-
- (void)count;
-
- /* load vector elements */
- LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh_ptr);
- b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
- limit = (v16u8)__msa_fill_b(*limit_ptr);
-
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
-
- p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
- p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
- q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
- q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
- SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
-}
-
-void vp9_lpf_horizontal_4_dual_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit0_ptr,
- const uint8_t *limit0_ptr,
- const uint8_t *thresh0_ptr,
- const uint8_t *b_limit1_ptr,
- const uint8_t *limit1_ptr,
- const uint8_t *thresh1_ptr) {- v16u8 mask, hev, flat, thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
-
- /* load vector elements */
- LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
- thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
- thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
-
- b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
- b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
- b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
-
- limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
- limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
- limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
-
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
- hev, mask, flat);
- VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
-
- ST_UB4(p1, p0, q0, q1, (src - 2 * pitch), pitch);
-}
-
-void vp9_lpf_vertical_4_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr,
- int32_t count) {- v16u8 mask, hev, flat, limit, thresh, b_limit;
- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v8i16 vec0, vec1, vec2, vec3;
-
- (void)count;
-
- LD_UB8((src - 4), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh_ptr);
- b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
- limit = (v16u8)__msa_fill_b(*limit_ptr);
-
- TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3,
- p3, p2, p1, p0, q0, q1, q2, q3);
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
- ILVR_B2_SH(p0, p1, q1, q0, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec2, vec3);
-
- src -= 2;
- ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
- src += 4 * pitch;
- ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
-}
-
-void vp9_lpf_vertical_4_dual_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit0_ptr,
- const uint8_t *limit0_ptr,
- const uint8_t *thresh0_ptr,
- const uint8_t *b_limit1_ptr,
- const uint8_t *limit1_ptr,
- const uint8_t *thresh1_ptr) {- v16u8 mask, hev, flat;
- v16u8 thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
- v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
- v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
-
- LD_UB8(src - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
- LD_UB8(src - 4 + (8 * pitch), pitch,
- row8, row9, row10, row11, row12, row13, row14, row15);
-
- TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
- row8, row9, row10, row11, row12, row13, row14, row15,
- p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
- thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
- thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
-
- b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
- b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
- b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
-
- limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
- limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
- limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
-
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
- hev, mask, flat);
- VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
- ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
- ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3);
- ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
- ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5);
-
- src -= 2;
-
- ST4x8_UB(tmp2, tmp3, src, pitch);
- src += (8 * pitch);
- ST4x8_UB(tmp4, tmp5, src, pitch);
-}
--- a/vp9/common/mips/msa/vp9_loopfilter_8_msa.c
+++ /dev/null
@@ -1,348 +1,0 @@
-/*
- * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "vp9/common/mips/msa/vp9_loopfilter_msa.h"
-
-void vp9_lpf_horizontal_8_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr,
- int32_t count) {- uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
- v16u8 mask, hev, flat, thresh, b_limit, limit;
- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
- v8i16 p2_filter8, p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8;
- v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
- v16i8 zero = { 0 };-
- (void)count;
-
- /* load vector elements */
- LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh_ptr);
- b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
- limit = (v16u8)__msa_fill_b(*limit_ptr);
-
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
- VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
-
- flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
-
- if (__msa_test_bz_v(flat)) {- p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
- p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
- q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
- q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
- SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
- } else {- ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
- zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
- q2_r, q3_r);
- VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
- p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
-
- /* convert 16 bit output data into 8 bit */
- PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8,
- zero, q0_filter8, p2_filter8, p1_filter8, p0_filter8,
- q0_filter8);
- PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
-
- /* store pixel values */
- p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
- p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
- p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
- q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
- q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
- q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
-
- p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
- p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
- p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
- q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
- q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
- q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
-
- src -= 3 * pitch;
-
- SD4(p2_d, p1_d, p0_d, q0_d, src, pitch);
- src += (4 * pitch);
- SD(q1_d, src);
- src += pitch;
- SD(q2_d, src);
- }
-}
-
-void vp9_lpf_horizontal_8_dual_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *b_limit1,
- const uint8_t *limit1,
- const uint8_t *thresh1) {- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
- v16u8 flat, mask, hev, tmp, thresh, b_limit, limit;
- v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
- v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
- v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
- v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
- v16u8 zero = { 0 };-
- /* load vector elements */
- LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh0);
- tmp = (v16u8)__msa_fill_b(*thresh1);
- thresh = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)thresh);
-
- b_limit = (v16u8)__msa_fill_b(*b_limit0);
- tmp = (v16u8)__msa_fill_b(*b_limit1);
- b_limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)b_limit);
-
- limit = (v16u8)__msa_fill_b(*limit0);
- tmp = (v16u8)__msa_fill_b(*limit1);
- limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)limit);
-
- /* mask and hev */
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
- VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
-
- if (__msa_test_bz_v(flat)) {- ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
- } else {- ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
- zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
- q2_r, q3_r);
- VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
- p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
-
- ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
- ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
- VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
- p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
-
- /* convert 16 bit output data into 8 bit */
- PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
- p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
- p0_filt8_r, q0_filt8_r);
- PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
- q2_filt8_r);
-
- /* store pixel values */
- p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
- p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
- p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
- q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
- q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
- q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
-
- src -= 3 * pitch;
-
- ST_UB4(p2_out, p1_out, p0_out, q0_out, src, pitch);
- src += (4 * pitch);
- ST_UB2(q1_out, q2_out, src, pitch);
- src += (2 * pitch);
- }
-}
-
-void vp9_lpf_vertical_8_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit_ptr,
- const uint8_t *limit_ptr,
- const uint8_t *thresh_ptr,
- int32_t count) {- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v16u8 p1_out, p0_out, q0_out, q1_out;
- v16u8 flat, mask, hev, thresh, b_limit, limit;
- v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
- v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
- v16u8 zero = { 0 };- v8i16 vec0, vec1, vec2, vec3, vec4;
-
- (void)count;
-
- /* load vector elements */
- LD_UB8(src - 4, pitch, p3, p2, p1, p0, q0, q1, q2, q3);
-
- TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3,
- p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh_ptr);
- b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
- limit = (v16u8)__msa_fill_b(*limit_ptr);
-
- /* mask and hev */
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- /* flat4 */
- VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
- /* filter4 */
- VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
-
- flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
-
- if (__msa_test_bz_v(flat)) {- /* Store 4 pixels p1-_q1 */
- ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec2, vec3);
-
- src -= 2;
- ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
- src += 4 * pitch;
- ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
- } else {- ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
- zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
- q3_r);
- VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
- p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
- /* convert 16 bit output data into 8 bit */
- PCKEV_B4_SH(p2_filt8_r, p2_filt8_r, p1_filt8_r, p1_filt8_r, p0_filt8_r,
- p0_filt8_r, q0_filt8_r, q0_filt8_r, p2_filt8_r, p1_filt8_r,
- p0_filt8_r, q0_filt8_r);
- PCKEV_B2_SH(q1_filt8_r, q1_filt8_r, q2_filt8_r, q2_filt8_r, q1_filt8_r,
- q2_filt8_r);
-
- /* store pixel values */
- p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
- p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
- p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
- q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
- q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
- q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
-
- /* Store 6 pixels p2-_q2 */
- ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec2, vec3);
- vec4 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
-
- src -= 3;
- ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
- ST2x4_UB(vec4, 0, src + 4, pitch);
- src += (4 * pitch);
- ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
- ST2x4_UB(vec4, 4, src + 4, pitch);
- }
-}
-
-void vp9_lpf_vertical_8_dual_msa(uint8_t *src, int32_t pitch,
- const uint8_t *b_limit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *b_limit1,
- const uint8_t *limit1,
- const uint8_t *thresh1) {- uint8_t *temp_src;
- v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
- v16u8 p1_out, p0_out, q0_out, q1_out;
- v16u8 flat, mask, hev, thresh, b_limit, limit;
- v16u8 row4, row5, row6, row7, row12, row13, row14, row15;
- v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
- v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
- v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
- v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
- v16u8 zero = { 0 };- v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
-
- temp_src = src - 4;
-
- LD_UB8(temp_src, pitch, p0, p1, p2, p3, row4, row5, row6, row7);
- temp_src += (8 * pitch);
- LD_UB8(temp_src, pitch, q3, q2, q1, q0, row12, row13, row14, row15);
-
- /* transpose 16x8 matrix into 8x16 */
- TRANSPOSE16x8_UB_UB(p0, p1, p2, p3, row4, row5, row6, row7,
- q3, q2, q1, q0, row12, row13, row14, row15,
- p3, p2, p1, p0, q0, q1, q2, q3);
-
- thresh = (v16u8)__msa_fill_b(*thresh0);
- vec0 = (v8i16)__msa_fill_b(*thresh1);
- thresh = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)thresh);
-
- b_limit = (v16u8)__msa_fill_b(*b_limit0);
- vec0 = (v8i16)__msa_fill_b(*b_limit1);
- b_limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)b_limit);
-
- limit = (v16u8)__msa_fill_b(*limit0);
- vec0 = (v8i16)__msa_fill_b(*limit1);
- limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)limit);
-
- /* mask and hev */
- LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
- hev, mask, flat);
- /* flat4 */
- VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
- /* filter4 */
- VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
-
- if (__msa_test_bz_v(flat)) {- ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec2, vec3);
- ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec4, vec5);
-
- src -= 2;
- ST4x8_UB(vec2, vec3, src, pitch);
- src += 8 * pitch;
- ST4x8_UB(vec4, vec5, src, pitch);
- } else {- ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
- zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
- q3_r);
- VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
- p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
-
- ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
- ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
-
- /* filter8 */
- VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
- p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
-
- /* convert 16 bit output data into 8 bit */
- PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
- p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
- p0_filt8_r, q0_filt8_r);
- PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
- q2_filt8_r);
-
- /* store pixel values */
- p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
- p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
- p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
- q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
- q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
- q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
-
- ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec3, vec4);
- ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
- ILVRL_H2_SH(vec1, vec0, vec6, vec7);
- ILVRL_B2_SH(q2, q1, vec2, vec5);
-
- src -= 3;
- ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
- ST2x4_UB(vec2, 0, src + 4, pitch);
- src += (4 * pitch);
- ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src, pitch);
- ST2x4_UB(vec2, 4, src + 4, pitch);
- src += (4 * pitch);
- ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src, pitch);
- ST2x4_UB(vec5, 0, src + 4, pitch);
- src += (4 * pitch);
- ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src, pitch);
- ST2x4_UB(vec5, 4, src + 4, pitch);
- }
-}
--- a/vp9/common/mips/msa/vp9_loopfilter_msa.h
+++ /dev/null
@@ -1,246 +1,0 @@
-/*
- * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#ifndef VP9_COMMON_MIPS_MSA_VP9_LOOPFILTER_MSA_H_
-#define VP9_COMMON_MIPS_MSA_VP9_LOOPFILTER_MSA_H_
-
-#include "vpx_dsp/mips/macros_msa.h"
-
-#define VP9_LPF_FILTER4_8W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \
- p1_out, p0_out, q0_out, q1_out) { \- v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign; \
- v16i8 filt, filt1, filt2, cnst4b, cnst3b; \
- v8i16 q0_sub_p0_r, filt_r, cnst3h; \
- \
- p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \
- p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \
- q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \
- q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \
- \
- filt = __msa_subs_s_b(p1_m, q1_m); \
- filt = filt & (v16i8)hev_in; \
- q0_sub_p0 = q0_m - p0_m; \
- filt_sign = __msa_clti_s_b(filt, 0); \
- \
- cnst3h = __msa_ldi_h(3); \
- q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0); \
- q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h); \
- filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt); \
- filt_r += q0_sub_p0_r; \
- filt_r = __msa_sat_s_h(filt_r, 7); \
- \
- /* combine left and right part */ \
- filt = __msa_pckev_b((v16i8)filt_r, (v16i8)filt_r); \
- \
- filt = filt & (v16i8)mask_in; \
- cnst4b = __msa_ldi_b(4); \
- filt1 = __msa_adds_s_b(filt, cnst4b); \
- filt1 >>= 3; \
- \
- cnst3b = __msa_ldi_b(3); \
- filt2 = __msa_adds_s_b(filt, cnst3b); \
- filt2 >>= 3; \
- \
- q0_m = __msa_subs_s_b(q0_m, filt1); \
- q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \
- p0_m = __msa_adds_s_b(p0_m, filt2); \
- p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \
- \
- filt = __msa_srari_b(filt1, 1); \
- hev_in = __msa_xori_b((v16u8)hev_in, 0xff); \
- filt = filt & (v16i8)hev_in; \
- \
- q1_m = __msa_subs_s_b(q1_m, filt); \
- q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \
- p1_m = __msa_adds_s_b(p1_m, filt); \
- p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \
-}
-
-#define VP9_LPF_FILTER4_4W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \
- p1_out, p0_out, q0_out, q1_out) { \- v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign; \
- v16i8 filt, filt1, filt2, cnst4b, cnst3b; \
- v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_l, filt_r, cnst3h; \
- \
- p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \
- p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \
- q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \
- q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \
- \
- filt = __msa_subs_s_b(p1_m, q1_m); \
- \
- filt = filt & (v16i8)hev_in; \
- \
- q0_sub_p0 = q0_m - p0_m; \
- filt_sign = __msa_clti_s_b(filt, 0); \
- \
- cnst3h = __msa_ldi_h(3); \
- q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0); \
- q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h); \
- filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt); \
- filt_r += q0_sub_p0_r; \
- filt_r = __msa_sat_s_h(filt_r, 7); \
- \
- q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0, q0_sub_p0); \
- q0_sub_p0_l = __msa_dotp_s_h((v16i8)q0_sub_p0_l, (v16i8)cnst3h); \
- filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt); \
- filt_l += q0_sub_p0_l; \
- filt_l = __msa_sat_s_h(filt_l, 7); \
- \
- filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r); \
- filt = filt & (v16i8)mask_in; \
- \
- cnst4b = __msa_ldi_b(4); \
- filt1 = __msa_adds_s_b(filt, cnst4b); \
- filt1 >>= 3; \
- \
- cnst3b = __msa_ldi_b(3); \
- filt2 = __msa_adds_s_b(filt, cnst3b); \
- filt2 >>= 3; \
- \
- q0_m = __msa_subs_s_b(q0_m, filt1); \
- q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \
- p0_m = __msa_adds_s_b(p0_m, filt2); \
- p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \
- \
- filt = __msa_srari_b(filt1, 1); \
- hev_in = __msa_xori_b((v16u8)hev_in, 0xff); \
- filt = filt & (v16i8)hev_in; \
- \
- q1_m = __msa_subs_s_b(q1_m, filt); \
- q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \
- p1_m = __msa_adds_s_b(p1_m, filt); \
- p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \
-}
-
-#define VP9_FLAT4(p3_in, p2_in, p0_in, q0_in, q2_in, q3_in, flat_out) { \- v16u8 tmp, p2_a_sub_p0, q2_a_sub_q0, p3_a_sub_p0, q3_a_sub_q0; \
- v16u8 zero_in = { 0 }; \- \
- tmp = __msa_ori_b(zero_in, 1); \
- p2_a_sub_p0 = __msa_asub_u_b(p2_in, p0_in); \
- q2_a_sub_q0 = __msa_asub_u_b(q2_in, q0_in); \
- p3_a_sub_p0 = __msa_asub_u_b(p3_in, p0_in); \
- q3_a_sub_q0 = __msa_asub_u_b(q3_in, q0_in); \
- \
- p2_a_sub_p0 = __msa_max_u_b(p2_a_sub_p0, q2_a_sub_q0); \
- flat_out = __msa_max_u_b(p2_a_sub_p0, flat_out); \
- p3_a_sub_p0 = __msa_max_u_b(p3_a_sub_p0, q3_a_sub_q0); \
- flat_out = __msa_max_u_b(p3_a_sub_p0, flat_out); \
- \
- flat_out = (tmp < (v16u8)flat_out); \
- flat_out = __msa_xori_b(flat_out, 0xff); \
- flat_out = flat_out & (mask); \
-}
-
-#define VP9_FLAT5(p7_in, p6_in, p5_in, p4_in, p0_in, q0_in, q4_in, \
- q5_in, q6_in, q7_in, flat_in, flat2_out) { \- v16u8 tmp, zero_in = { 0 }; \- v16u8 p4_a_sub_p0, q4_a_sub_q0, p5_a_sub_p0, q5_a_sub_q0; \
- v16u8 p6_a_sub_p0, q6_a_sub_q0, p7_a_sub_p0, q7_a_sub_q0; \
- \
- tmp = __msa_ori_b(zero_in, 1); \
- p4_a_sub_p0 = __msa_asub_u_b(p4_in, p0_in); \
- q4_a_sub_q0 = __msa_asub_u_b(q4_in, q0_in); \
- p5_a_sub_p0 = __msa_asub_u_b(p5_in, p0_in); \
- q5_a_sub_q0 = __msa_asub_u_b(q5_in, q0_in); \
- p6_a_sub_p0 = __msa_asub_u_b(p6_in, p0_in); \
- q6_a_sub_q0 = __msa_asub_u_b(q6_in, q0_in); \
- p7_a_sub_p0 = __msa_asub_u_b(p7_in, p0_in); \
- q7_a_sub_q0 = __msa_asub_u_b(q7_in, q0_in); \
- \
- p4_a_sub_p0 = __msa_max_u_b(p4_a_sub_p0, q4_a_sub_q0); \
- flat2_out = __msa_max_u_b(p5_a_sub_p0, q5_a_sub_q0); \
- flat2_out = __msa_max_u_b(p4_a_sub_p0, flat2_out); \
- p6_a_sub_p0 = __msa_max_u_b(p6_a_sub_p0, q6_a_sub_q0); \
- flat2_out = __msa_max_u_b(p6_a_sub_p0, flat2_out); \
- p7_a_sub_p0 = __msa_max_u_b(p7_a_sub_p0, q7_a_sub_q0); \
- flat2_out = __msa_max_u_b(p7_a_sub_p0, flat2_out); \
- \
- flat2_out = (tmp < (v16u8)flat2_out); \
- flat2_out = __msa_xori_b(flat2_out, 0xff); \
- flat2_out = flat2_out & flat_in; \
-}
-
-#define VP9_FILTER8(p3_in, p2_in, p1_in, p0_in, \
- q0_in, q1_in, q2_in, q3_in, \
- p2_filt8_out, p1_filt8_out, p0_filt8_out, \
- q0_filt8_out, q1_filt8_out, q2_filt8_out) { \- v8u16 tmp0, tmp1, tmp2; \
- \
- tmp2 = p2_in + p1_in + p0_in; \
- tmp0 = p3_in << 1; \
- \
- tmp0 = tmp0 + tmp2 + q0_in; \
- tmp1 = tmp0 + p3_in + p2_in; \
- p2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
- \
- tmp1 = tmp0 + p1_in + q1_in; \
- p1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
- \
- tmp1 = q2_in + q1_in + q0_in; \
- tmp2 = tmp2 + tmp1; \
- tmp0 = tmp2 + (p0_in); \
- tmp0 = tmp0 + (p3_in); \
- p0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp0, 3); \
- \
- tmp0 = q2_in + q3_in; \
- tmp0 = p0_in + tmp1 + tmp0; \
- tmp1 = q3_in + q3_in; \
- tmp1 = tmp1 + tmp0; \
- q2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
- \
- tmp0 = tmp2 + q3_in; \
- tmp1 = tmp0 + q0_in; \
- q0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
- \
- tmp1 = tmp0 - p2_in; \
- tmp0 = q1_in + q3_in; \
- tmp1 = tmp0 + tmp1; \
- q1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
-}
-
-#define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in, \
- q0_in, q1_in, q2_in, q3_in, \
- limit_in, b_limit_in, thresh_in, \
- hev_out, mask_out, flat_out) { \- v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m; \
- v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m; \
- \
- /* absolute subtraction of pixel values */ \
- p3_asub_p2_m = __msa_asub_u_b(p3_in, p2_in); \
- p2_asub_p1_m = __msa_asub_u_b(p2_in, p1_in); \
- p1_asub_p0_m = __msa_asub_u_b(p1_in, p0_in); \
- q1_asub_q0_m = __msa_asub_u_b(q1_in, q0_in); \
- q2_asub_q1_m = __msa_asub_u_b(q2_in, q1_in); \
- q3_asub_q2_m = __msa_asub_u_b(q3_in, q2_in); \
- p0_asub_q0_m = __msa_asub_u_b(p0_in, q0_in); \
- p1_asub_q1_m = __msa_asub_u_b(p1_in, q1_in); \
- \
- /* calculation of hev */ \
- flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m); \
- hev_out = thresh_in < (v16u8)flat_out; \
- \
- /* calculation of mask */ \
- p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m); \
- p1_asub_q1_m >>= 1; \
- p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m); \
- \
- mask_out = b_limit_in < p0_asub_q0_m; \
- mask_out = __msa_max_u_b(flat_out, mask_out); \
- p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m); \
- mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out); \
- q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m); \
- mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out); \
- \
- mask_out = limit_in < (v16u8)mask_out; \
- mask_out = __msa_xori_b(mask_out, 0xff); \
-}
-#endif /* VP9_COMMON_MIPS_MSA_VP9_LOOPFILTER_MSA_H_ */
--- a/vp9/common/vp9_loopfilter.c
+++ b/vp9/common/vp9_loopfilter.c
@@ -9,6 +9,7 @@
*/
#include "./vpx_config.h"
+#include "./vpx_dsp_rtcd.h"
#include "vp9/common/vp9_loopfilter.h"
#include "vp9/common/vp9_onyxc_int.h"
#include "vp9/common/vp9_reconinter.h"
--- a/vp9/common/vp9_loopfilter_filters.c
+++ /dev/null
@@ -1,743 +1,0 @@
-/*
- * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include "./vpx_config.h"
-#include "vpx_dsp/vpx_dsp_common.h"
-#include "vpx_ports/mem.h"
-
-static INLINE int8_t signed_char_clamp(int t) {- return (int8_t)clamp(t, -128, 127);
-}
-
-#if CONFIG_VP9_HIGHBITDEPTH
-static INLINE int16_t signed_char_clamp_high(int t, int bd) {- switch (bd) {- case 10:
- return (int16_t)clamp(t, -128*4, 128*4-1);
- case 12:
- return (int16_t)clamp(t, -128*16, 128*16-1);
- case 8:
- default:
- return (int16_t)clamp(t, -128, 128-1);
- }
-}
-#endif
-
-// should we apply any filter at all: 11111111 yes, 00000000 no
-static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit,
- uint8_t p3, uint8_t p2,
- uint8_t p1, uint8_t p0,
- uint8_t q0, uint8_t q1,
- uint8_t q2, uint8_t q3) {- int8_t mask = 0;
- mask |= (abs(p3 - p2) > limit) * -1;
- mask |= (abs(p2 - p1) > limit) * -1;
- mask |= (abs(p1 - p0) > limit) * -1;
- mask |= (abs(q1 - q0) > limit) * -1;
- mask |= (abs(q2 - q1) > limit) * -1;
- mask |= (abs(q3 - q2) > limit) * -1;
- mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- return ~mask;
-}
-
-static INLINE int8_t flat_mask4(uint8_t thresh,
- uint8_t p3, uint8_t p2,
- uint8_t p1, uint8_t p0,
- uint8_t q0, uint8_t q1,
- uint8_t q2, uint8_t q3) {- int8_t mask = 0;
- mask |= (abs(p1 - p0) > thresh) * -1;
- mask |= (abs(q1 - q0) > thresh) * -1;
- mask |= (abs(p2 - p0) > thresh) * -1;
- mask |= (abs(q2 - q0) > thresh) * -1;
- mask |= (abs(p3 - p0) > thresh) * -1;
- mask |= (abs(q3 - q0) > thresh) * -1;
- return ~mask;
-}
-
-static INLINE int8_t flat_mask5(uint8_t thresh,
- uint8_t p4, uint8_t p3,
- uint8_t p2, uint8_t p1,
- uint8_t p0, uint8_t q0,
- uint8_t q1, uint8_t q2,
- uint8_t q3, uint8_t q4) {- int8_t mask = ~flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3);
- mask |= (abs(p4 - p0) > thresh) * -1;
- mask |= (abs(q4 - q0) > thresh) * -1;
- return ~mask;
-}
-
-// is there high edge variance internal edge: 11111111 yes, 00000000 no
-static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0,
- uint8_t q0, uint8_t q1) {- int8_t hev = 0;
- hev |= (abs(p1 - p0) > thresh) * -1;
- hev |= (abs(q1 - q0) > thresh) * -1;
- return hev;
-}
-
-static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1,
- uint8_t *op0, uint8_t *oq0, uint8_t *oq1) {- int8_t filter1, filter2;
-
- const int8_t ps1 = (int8_t) *op1 ^ 0x80;
- const int8_t ps0 = (int8_t) *op0 ^ 0x80;
- const int8_t qs0 = (int8_t) *oq0 ^ 0x80;
- const int8_t qs1 = (int8_t) *oq1 ^ 0x80;
- const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1);
-
- // add outer taps if we have high edge variance
- int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
-
- // inner taps
- filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
-
- // save bottom 3 bits so that we round one side +4 and the other +3
- // if it equals 4 we'll set to adjust by -1 to account for the fact
- // we'd round 3 the other way
- filter1 = signed_char_clamp(filter + 4) >> 3;
- filter2 = signed_char_clamp(filter + 3) >> 3;
-
- *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
- *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
-
- // outer tap adjustments
- filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
-
- *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
- *op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
-}
-
-void vp9_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */,
- const uint8_t *blimit, const uint8_t *limit,
- const uint8_t *thresh, int count) {- int i;
-
- // loop filter designed to work using chars so that we can make maximum use
- // of 8 bit simd instructions.
- for (i = 0; i < 8 * count; ++i) {- const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
- const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
- const int8_t mask = filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3);
- filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p);
- ++s;
- }
-}
-
-void vp9_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
- const uint8_t *limit0, const uint8_t *thresh0,
- const uint8_t *blimit1, const uint8_t *limit1,
- const uint8_t *thresh1) {- vp9_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1);
- vp9_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1);
-}
-
-void vp9_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int count) {- int i;
-
- // loop filter designed to work using chars so that we can make maximum use
- // of 8 bit simd instructions.
- for (i = 0; i < 8 * count; ++i) {- const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
- const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
- const int8_t mask = filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3);
- filter4(mask, *thresh, s - 2, s - 1, s, s + 1);
- s += pitch;
- }
-}
-
-void vp9_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
- const uint8_t *limit0, const uint8_t *thresh0,
- const uint8_t *blimit1, const uint8_t *limit1,
- const uint8_t *thresh1) {- vp9_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1);
- vp9_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
- thresh1, 1);
-}
-
-static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat,
- uint8_t *op3, uint8_t *op2,
- uint8_t *op1, uint8_t *op0,
- uint8_t *oq0, uint8_t *oq1,
- uint8_t *oq2, uint8_t *oq3) {- if (flat && mask) {- const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
- const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
-
- // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
- *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
- *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
- *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
- *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
- *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
- *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
- } else {- filter4(mask, thresh, op1, op0, oq0, oq1);
- }
-}
-
-void vp9_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int count) {- int i;
-
- // loop filter designed to work using chars so that we can make maximum use
- // of 8 bit simd instructions.
- for (i = 0; i < 8 * count; ++i) {- const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
- const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
-
- const int8_t mask = filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3);
- const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
- filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
- s, s + 1 * p, s + 2 * p, s + 3 * p);
- ++s;
- }
-}
-
-void vp9_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
- const uint8_t *limit0, const uint8_t *thresh0,
- const uint8_t *blimit1, const uint8_t *limit1,
- const uint8_t *thresh1) {- vp9_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1);
- vp9_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1);
-}
-
-void vp9_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int count) {- int i;
-
- for (i = 0; i < 8 * count; ++i) {- const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
- const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
- const int8_t mask = filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3);
- const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
- filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1,
- s, s + 1, s + 2, s + 3);
- s += pitch;
- }
-}
-
-void vp9_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
- const uint8_t *limit0, const uint8_t *thresh0,
- const uint8_t *blimit1, const uint8_t *limit1,
- const uint8_t *thresh1) {- vp9_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1);
- vp9_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
- thresh1, 1);
-}
-
-static INLINE void filter16(int8_t mask, uint8_t thresh,
- uint8_t flat, uint8_t flat2,
- uint8_t *op7, uint8_t *op6,
- uint8_t *op5, uint8_t *op4,
- uint8_t *op3, uint8_t *op2,
- uint8_t *op1, uint8_t *op0,
- uint8_t *oq0, uint8_t *oq1,
- uint8_t *oq2, uint8_t *oq3,
- uint8_t *oq4, uint8_t *oq5,
- uint8_t *oq6, uint8_t *oq7) {- if (flat2 && flat && mask) {- const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4,
- p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
-
- const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3,
- q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
-
- // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
- *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
- q0, 4);
- *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
- q0 + q1, 4);
- *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
- q0 + q1 + q2, 4);
- *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
- q0 + q1 + q2 + q3, 4);
- *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
- q0 + q1 + q2 + q3 + q4, 4);
- *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
- q0 + q1 + q2 + q3 + q4 + q5, 4);
- *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
- q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
- *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
- q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
- *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
- q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
- *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
- q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
- *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
- q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
- *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
- q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
- *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
- q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
- *oq6 = ROUND_POWER_OF_TWO(p0 +
- q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
- } else {- filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
- }
-}
-
-void vp9_lpf_horizontal_16_c(uint8_t *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int count) {- int i;
-
- // loop filter designed to work using chars so that we can make maximum use
- // of 8 bit simd instructions.
- for (i = 0; i < 8 * count; ++i) {- const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
- const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
- const int8_t mask = filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3);
- const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
- const int8_t flat2 = flat_mask5(1,
- s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
- q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p]);
-
- filter16(mask, *thresh, flat, flat2,
- s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
- s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
- s, s + 1 * p, s + 2 * p, s + 3 * p,
- s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p);
- ++s;
- }
-}
-
-static void mb_lpf_vertical_edge_w(uint8_t *s, int p,
- const uint8_t *blimit,
- const uint8_t *limit,
- const uint8_t *thresh,
- int count) {- int i;
-
- for (i = 0; i < count; ++i) {- const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
- const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
- const int8_t mask = filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3);
- const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
- const int8_t flat2 = flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
- q0, s[4], s[5], s[6], s[7]);
-
- filter16(mask, *thresh, flat, flat2,
- s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
- s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7);
- s += p;
- }
-}
-
-void vp9_lpf_vertical_16_c(uint8_t *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh) {- mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8);
-}
-
-void vp9_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh) {- mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16);
-}
-
-#if CONFIG_VP9_HIGHBITDEPTH
-// Should we apply any filter at all: 11111111 yes, 00000000 no ?
-static INLINE int8_t highbd_filter_mask(uint8_t limit, uint8_t blimit,
- uint16_t p3, uint16_t p2,
- uint16_t p1, uint16_t p0,
- uint16_t q0, uint16_t q1,
- uint16_t q2, uint16_t q3, int bd) {- int8_t mask = 0;
- int16_t limit16 = (uint16_t)limit << (bd - 8);
- int16_t blimit16 = (uint16_t)blimit << (bd - 8);
- mask |= (abs(p3 - p2) > limit16) * -1;
- mask |= (abs(p2 - p1) > limit16) * -1;
- mask |= (abs(p1 - p0) > limit16) * -1;
- mask |= (abs(q1 - q0) > limit16) * -1;
- mask |= (abs(q2 - q1) > limit16) * -1;
- mask |= (abs(q3 - q2) > limit16) * -1;
- mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1;
- return ~mask;
-}
-
-static INLINE int8_t highbd_flat_mask4(uint8_t thresh,
- uint16_t p3, uint16_t p2,
- uint16_t p1, uint16_t p0,
- uint16_t q0, uint16_t q1,
- uint16_t q2, uint16_t q3, int bd) {- int8_t mask = 0;
- int16_t thresh16 = (uint16_t)thresh << (bd - 8);
- mask |= (abs(p1 - p0) > thresh16) * -1;
- mask |= (abs(q1 - q0) > thresh16) * -1;
- mask |= (abs(p2 - p0) > thresh16) * -1;
- mask |= (abs(q2 - q0) > thresh16) * -1;
- mask |= (abs(p3 - p0) > thresh16) * -1;
- mask |= (abs(q3 - q0) > thresh16) * -1;
- return ~mask;
-}
-
-static INLINE int8_t highbd_flat_mask5(uint8_t thresh,
- uint16_t p4, uint16_t p3,
- uint16_t p2, uint16_t p1,
- uint16_t p0, uint16_t q0,
- uint16_t q1, uint16_t q2,
- uint16_t q3, uint16_t q4, int bd) {- int8_t mask = ~highbd_flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3, bd);
- int16_t thresh16 = (uint16_t)thresh << (bd - 8);
- mask |= (abs(p4 - p0) > thresh16) * -1;
- mask |= (abs(q4 - q0) > thresh16) * -1;
- return ~mask;
-}
-
-// Is there high edge variance internal edge:
-// 11111111_11111111 yes, 00000000_00000000 no ?
-static INLINE int16_t highbd_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0,
- uint16_t q0, uint16_t q1, int bd) {- int16_t hev = 0;
- int16_t thresh16 = (uint16_t)thresh << (bd - 8);
- hev |= (abs(p1 - p0) > thresh16) * -1;
- hev |= (abs(q1 - q0) > thresh16) * -1;
- return hev;
-}
-
-static INLINE void highbd_filter4(int8_t mask, uint8_t thresh, uint16_t *op1,
- uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
- int bd) {- int16_t filter1, filter2;
- // ^0x80 equivalent to subtracting 0x80 from the values to turn them
- // into -128 to +127 instead of 0 to 255.
- int shift = bd - 8;
- const int16_t ps1 = (int16_t)*op1 - (0x80 << shift);
- const int16_t ps0 = (int16_t)*op0 - (0x80 << shift);
- const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift);
- const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift);
- const uint16_t hev = highbd_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd);
-
- // Add outer taps if we have high edge variance.
- int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev;
-
- // Inner taps.
- filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask;
-
- // Save bottom 3 bits so that we round one side +4 and the other +3
- // if it equals 4 we'll set to adjust by -1 to account for the fact
- // we'd round 3 the other way.
- filter1 = signed_char_clamp_high(filter + 4, bd) >> 3;
- filter2 = signed_char_clamp_high(filter + 3, bd) >> 3;
-
- *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift);
- *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift);
-
- // Outer tap adjustments.
- filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
-
- *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift);
- *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift);
-}
-
-void vp9_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */,
- const uint8_t *blimit, const uint8_t *limit,
- const uint8_t *thresh, int count, int bd) {- int i;
-
- // loop filter designed to work using chars so that we can make maximum use
- // of 8 bit simd instructions.
- for (i = 0; i < 8 * count; ++i) {- const uint16_t p3 = s[-4 * p];
- const uint16_t p2 = s[-3 * p];
- const uint16_t p1 = s[-2 * p];
- const uint16_t p0 = s[-p];
- const uint16_t q0 = s[0 * p];
- const uint16_t q1 = s[1 * p];
- const uint16_t q2 = s[2 * p];
- const uint16_t q3 = s[3 * p];
- const int8_t mask = highbd_filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3, bd);
- highbd_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd);
- ++s;
- }
-}
-
-void vp9_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int p,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1,
- int bd) {- vp9_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1, bd);
- vp9_highbd_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1, bd);
-}
-
-void vp9_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int count, int bd) {- int i;
-
- // loop filter designed to work using chars so that we can make maximum use
- // of 8 bit simd instructions.
- for (i = 0; i < 8 * count; ++i) {- const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
- const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
- const int8_t mask = highbd_filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3, bd);
- highbd_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd);
- s += pitch;
- }
-}
-
-void vp9_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1,
- int bd) {- vp9_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1, bd);
- vp9_highbd_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
- thresh1, 1, bd);
-}
-
-static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, uint8_t flat,
- uint16_t *op3, uint16_t *op2,
- uint16_t *op1, uint16_t *op0,
- uint16_t *oq0, uint16_t *oq1,
- uint16_t *oq2, uint16_t *oq3, int bd) {- if (flat && mask) {- const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
- const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
-
- // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
- *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
- *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
- *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
- *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
- *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
- *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
- } else {- highbd_filter4(mask, thresh, op1, op0, oq0, oq1, bd);
- }
-}
-
-void vp9_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int count, int bd) {- int i;
-
- // loop filter designed to work using chars so that we can make maximum use
- // of 8 bit simd instructions.
- for (i = 0; i < 8 * count; ++i) {- const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
- const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
-
- const int8_t mask = highbd_filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3, bd);
- const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
- bd);
- highbd_filter8(mask, *thresh, flat,
- s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
- s, s + 1 * p, s + 2 * p, s + 3 * p, bd);
- ++s;
- }
-}
-
-void vp9_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int p,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1,
- int bd) {- vp9_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1, bd);
- vp9_highbd_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1, bd);
-}
-
-void vp9_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int count, int bd) {- int i;
-
- for (i = 0; i < 8 * count; ++i) {- const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
- const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
- const int8_t mask = highbd_filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3, bd);
- const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
- bd);
- highbd_filter8(mask, *thresh, flat,
- s - 4, s - 3, s - 2, s - 1,
- s, s + 1, s + 2, s + 3,
- bd);
- s += pitch;
- }
-}
-
-void vp9_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1,
- int bd) {- vp9_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1, bd);
- vp9_highbd_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
- thresh1, 1, bd);
-}
-
-static INLINE void highbd_filter16(int8_t mask, uint8_t thresh,
- uint8_t flat, uint8_t flat2,
- uint16_t *op7, uint16_t *op6,
- uint16_t *op5, uint16_t *op4,
- uint16_t *op3, uint16_t *op2,
- uint16_t *op1, uint16_t *op0,
- uint16_t *oq0, uint16_t *oq1,
- uint16_t *oq2, uint16_t *oq3,
- uint16_t *oq4, uint16_t *oq5,
- uint16_t *oq6, uint16_t *oq7, int bd) {- if (flat2 && flat && mask) {- const uint16_t p7 = *op7;
- const uint16_t p6 = *op6;
- const uint16_t p5 = *op5;
- const uint16_t p4 = *op4;
- const uint16_t p3 = *op3;
- const uint16_t p2 = *op2;
- const uint16_t p1 = *op1;
- const uint16_t p0 = *op0;
- const uint16_t q0 = *oq0;
- const uint16_t q1 = *oq1;
- const uint16_t q2 = *oq2;
- const uint16_t q3 = *oq3;
- const uint16_t q4 = *oq4;
- const uint16_t q5 = *oq5;
- const uint16_t q6 = *oq6;
- const uint16_t q7 = *oq7;
-
- // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
- *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
- q0, 4);
- *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
- q0 + q1, 4);
- *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
- q0 + q1 + q2, 4);
- *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
- q0 + q1 + q2 + q3, 4);
- *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
- q0 + q1 + q2 + q3 + q4, 4);
- *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
- q0 + q1 + q2 + q3 + q4 + q5, 4);
- *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
- q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
- *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
- q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
- *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
- q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
- *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
- q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
- *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
- q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
- *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
- q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
- *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
- q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
- *oq6 = ROUND_POWER_OF_TWO(p0 +
- q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
- } else {- highbd_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3,
- bd);
- }
-}
-
-void vp9_highbd_lpf_horizontal_16_c(uint16_t *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int count, int bd) {- int i;
-
- // loop filter designed to work using chars so that we can make maximum use
- // of 8 bit simd instructions.
- for (i = 0; i < 8 * count; ++i) {- const uint16_t p3 = s[-4 * p];
- const uint16_t p2 = s[-3 * p];
- const uint16_t p1 = s[-2 * p];
- const uint16_t p0 = s[-p];
- const uint16_t q0 = s[0 * p];
- const uint16_t q1 = s[1 * p];
- const uint16_t q2 = s[2 * p];
- const uint16_t q3 = s[3 * p];
- const int8_t mask = highbd_filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3, bd);
- const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
- bd);
- const int8_t flat2 = highbd_flat_mask5(
- 1, s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
- q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p], bd);
-
- highbd_filter16(mask, *thresh, flat, flat2,
- s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
- s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
- s, s + 1 * p, s + 2 * p, s + 3 * p,
- s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p,
- bd);
- ++s;
- }
-}
-
-static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p,
- const uint8_t *blimit,
- const uint8_t *limit,
- const uint8_t *thresh,
- int count, int bd) {- int i;
-
- for (i = 0; i < count; ++i) {- const uint16_t p3 = s[-4];
- const uint16_t p2 = s[-3];
- const uint16_t p1 = s[-2];
- const uint16_t p0 = s[-1];
- const uint16_t q0 = s[0];
- const uint16_t q1 = s[1];
- const uint16_t q2 = s[2];
- const uint16_t q3 = s[3];
- const int8_t mask = highbd_filter_mask(*limit, *blimit,
- p3, p2, p1, p0, q0, q1, q2, q3, bd);
- const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
- bd);
- const int8_t flat2 = highbd_flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
- q0, s[4], s[5], s[6], s[7], bd);
-
- highbd_filter16(mask, *thresh, flat, flat2,
- s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
- s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7,
- bd);
- s += p;
- }
-}
-
-void vp9_highbd_lpf_vertical_16_c(uint16_t *s, int p, const uint8_t *blimit,
- const uint8_t *limit, const uint8_t *thresh,
- int bd) {- highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd);
-}
-
-void vp9_highbd_lpf_vertical_16_dual_c(uint16_t *s, int p,
- const uint8_t *blimit,
- const uint8_t *limit,
- const uint8_t *thresh,
- int bd) {- highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16, bd);
-}
-#endif // CONFIG_VP9_HIGHBITDEPTH
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -220,49 +220,6 @@
specialize qw/vp9_dc_128_predictor_32x32 msa neon/, "$sse2_x86inc";
#
-# Loopfilter
-#
-add_proto qw/void vp9_lpf_vertical_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/vp9_lpf_vertical_16 sse2 neon_asm dspr2 msa/;
-$vp9_lpf_vertical_16_neon_asm=vp9_lpf_vertical_16_neon;
-
-add_proto qw/void vp9_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/vp9_lpf_vertical_16_dual sse2 neon_asm dspr2 msa/;
-$vp9_lpf_vertical_16_dual_neon_asm=vp9_lpf_vertical_16_dual_neon;
-
-add_proto qw/void vp9_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_vertical_8 sse2 neon_asm dspr2 msa/;
-$vp9_lpf_vertical_8_neon_asm=vp9_lpf_vertical_8_neon;
-
-add_proto qw/void vp9_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
-specialize qw/vp9_lpf_vertical_8_dual sse2 neon_asm dspr2 msa/;
-$vp9_lpf_vertical_8_dual_neon_asm=vp9_lpf_vertical_8_dual_neon;
-
-add_proto qw/void vp9_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_vertical_4 mmx neon dspr2 msa/;
-
-add_proto qw/void vp9_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
-specialize qw/vp9_lpf_vertical_4_dual sse2 neon dspr2 msa/;
-
-add_proto qw/void vp9_lpf_horizontal_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_horizontal_16 sse2 avx2 neon_asm dspr2 msa/;
-$vp9_lpf_horizontal_16_neon_asm=vp9_lpf_horizontal_16_neon;
-
-add_proto qw/void vp9_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_horizontal_8 sse2 neon_asm dspr2 msa/;
-$vp9_lpf_horizontal_8_neon_asm=vp9_lpf_horizontal_8_neon;
-
-add_proto qw/void vp9_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
-specialize qw/vp9_lpf_horizontal_8_dual sse2 neon_asm dspr2 msa/;
-$vp9_lpf_horizontal_8_dual_neon_asm=vp9_lpf_horizontal_8_dual_neon;
-
-add_proto qw/void vp9_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
-specialize qw/vp9_lpf_horizontal_4 mmx neon dspr2 msa/;
-
-add_proto qw/void vp9_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
-specialize qw/vp9_lpf_horizontal_4_dual sse2 neon dspr2 msa/;
-
-#
# post proc
#
if (vpx_config("CONFIG_VP9_POSTPROC") eq "yes") {@@ -666,42 +623,6 @@
add_proto qw/void vp9_highbd_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
specialize qw/vp9_highbd_convolve8_avg_vert/, "$sse2_x86_64";
-
- #
- # Loopfilter
- #
- add_proto qw/void vp9_highbd_lpf_vertical_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
- specialize qw/vp9_highbd_lpf_vertical_16 sse2/;
-
- add_proto qw/void vp9_highbd_lpf_vertical_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
- specialize qw/vp9_highbd_lpf_vertical_16_dual sse2/;
-
- add_proto qw/void vp9_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
- specialize qw/vp9_highbd_lpf_vertical_8 sse2/;
-
- add_proto qw/void vp9_highbd_lpf_vertical_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
- specialize qw/vp9_highbd_lpf_vertical_8_dual sse2/;
-
- add_proto qw/void vp9_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
- specialize qw/vp9_highbd_lpf_vertical_4 sse2/;
-
- add_proto qw/void vp9_highbd_lpf_vertical_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
- specialize qw/vp9_highbd_lpf_vertical_4_dual sse2/;
-
- add_proto qw/void vp9_highbd_lpf_horizontal_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
- specialize qw/vp9_highbd_lpf_horizontal_16 sse2/;
-
- add_proto qw/void vp9_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
- specialize qw/vp9_highbd_lpf_horizontal_8 sse2/;
-
- add_proto qw/void vp9_highbd_lpf_horizontal_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
- specialize qw/vp9_highbd_lpf_horizontal_8_dual sse2/;
-
- add_proto qw/void vp9_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
- specialize qw/vp9_highbd_lpf_horizontal_4 sse2/;
-
- add_proto qw/void vp9_highbd_lpf_horizontal_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
- specialize qw/vp9_highbd_lpf_horizontal_4_dual sse2/;
#
# post proc
--- a/vp9/common/x86/vp9_high_loopfilter_intrin_sse2.c
+++ /dev/null
@@ -1,1215 +1,0 @@
-/*
- * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <emmintrin.h> // SSE2
-
-#include "./vp9_rtcd.h"
-#include "vpx_ports/mem.h"
-#include "vp9/common/vp9_loopfilter.h"
-#include "vpx_ports/emmintrin_compat.h"
-
-static INLINE __m128i signed_char_clamp_bd_sse2(__m128i value, int bd) {- __m128i ubounded;
- __m128i lbounded;
- __m128i retval;
-
- const __m128i zero = _mm_set1_epi16(0);
- const __m128i one = _mm_set1_epi16(1);
- __m128i t80, max, min;
-
- if (bd == 8) {- t80 = _mm_set1_epi16(0x80);
- max = _mm_subs_epi16(
- _mm_subs_epi16(_mm_slli_epi16(one, 8), one), t80);
- } else if (bd == 10) {- t80 = _mm_set1_epi16(0x200);
- max = _mm_subs_epi16(
- _mm_subs_epi16(_mm_slli_epi16(one, 10), one), t80);
- } else { // bd == 12- t80 = _mm_set1_epi16(0x800);
- max = _mm_subs_epi16(
- _mm_subs_epi16(_mm_slli_epi16(one, 12), one), t80);
- }
-
- min = _mm_subs_epi16(zero, t80);
-
- ubounded = _mm_cmpgt_epi16(value, max);
- lbounded = _mm_cmplt_epi16(value, min);
- retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value);
- ubounded = _mm_and_si128(ubounded, max);
- lbounded = _mm_and_si128(lbounded, min);
- retval = _mm_or_si128(retval, ubounded);
- retval = _mm_or_si128(retval, lbounded);
- return retval;
-}
-
-// TODO(debargha, peter): Break up large functions into smaller ones
-// in this file.
-static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
- int p,
- const uint8_t *_blimit,
- const uint8_t *_limit,
- const uint8_t *_thresh,
- int bd) {- const __m128i zero = _mm_set1_epi16(0);
- const __m128i one = _mm_set1_epi16(1);
- __m128i blimit, limit, thresh;
- __m128i q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0;
- __m128i mask, hev, flat, flat2, abs_p1p0, abs_q1q0;
- __m128i ps1, qs1, ps0, qs0;
- __m128i abs_p0q0, abs_p1q1, ffff, work;
- __m128i filt, work_a, filter1, filter2;
- __m128i flat2_q6, flat2_p6, flat2_q5, flat2_p5, flat2_q4, flat2_p4;
- __m128i flat2_q3, flat2_p3, flat2_q2, flat2_p2, flat2_q1, flat2_p1;
- __m128i flat2_q0, flat2_p0;
- __m128i flat_q2, flat_p2, flat_q1, flat_p1, flat_q0, flat_p0;
- __m128i pixelFilter_p, pixelFilter_q;
- __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
- __m128i sum_p7, sum_q7, sum_p3, sum_q3;
- __m128i t4, t3, t80, t1;
- __m128i eight, four;
-
- if (bd == 8) {- blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
- limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
- thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
- } else if (bd == 10) {- blimit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
- limit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
- thresh = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
- } else { // bd == 12- blimit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
- limit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
- thresh = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
- }
-
- q4 = _mm_load_si128((__m128i *)(s + 4 * p));
- p4 = _mm_load_si128((__m128i *)(s - 5 * p));
- q3 = _mm_load_si128((__m128i *)(s + 3 * p));
- p3 = _mm_load_si128((__m128i *)(s - 4 * p));
- q2 = _mm_load_si128((__m128i *)(s + 2 * p));
- p2 = _mm_load_si128((__m128i *)(s - 3 * p));
- q1 = _mm_load_si128((__m128i *)(s + 1 * p));
- p1 = _mm_load_si128((__m128i *)(s - 2 * p));
- q0 = _mm_load_si128((__m128i *)(s + 0 * p));
- p0 = _mm_load_si128((__m128i *)(s - 1 * p));
-
- // highbd_filter_mask
- abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1));
- abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1));
-
- ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
-
- abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0));
- abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1));
-
- // highbd_hev_mask (in C code this is actually called from highbd_filter4)
- flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu16(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
-
- abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0); // abs(p0 - q0) * 2
- abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); // abs(p1 - q1) / 2
- mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
- mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p1, p0),
- _mm_subs_epu16(p0, p1)),
- _mm_or_si128(_mm_subs_epu16(q1, q0),
- _mm_subs_epu16(q0, q1)));
- mask = _mm_max_epi16(work, mask);
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
- _mm_subs_epu16(p1, p2)),
- _mm_or_si128(_mm_subs_epu16(q2, q1),
- _mm_subs_epu16(q1, q2)));
- mask = _mm_max_epi16(work, mask);
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p2),
- _mm_subs_epu16(p2, p3)),
- _mm_or_si128(_mm_subs_epu16(q3, q2),
- _mm_subs_epu16(q2, q3)));
- mask = _mm_max_epi16(work, mask);
-
- mask = _mm_subs_epu16(mask, limit);
- mask = _mm_cmpeq_epi16(mask, zero); // return ~mask
-
- // lp filter
- // highbd_filter4
- t4 = _mm_set1_epi16(4);
- t3 = _mm_set1_epi16(3);
- if (bd == 8)
- t80 = _mm_set1_epi16(0x80);
- else if (bd == 10)
- t80 = _mm_set1_epi16(0x200);
- else // bd == 12
- t80 = _mm_set1_epi16(0x800);
-
- t1 = _mm_set1_epi16(0x1);
-
- ps1 = _mm_subs_epi16(p1, t80);
- qs1 = _mm_subs_epi16(q1, t80);
- ps0 = _mm_subs_epi16(p0, t80);
- qs0 = _mm_subs_epi16(q0, t80);
-
- filt = _mm_and_si128(
- signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd), hev);
- work_a = _mm_subs_epi16(qs0, ps0);
- filt = _mm_adds_epi16(filt, work_a);
- filt = _mm_adds_epi16(filt, work_a);
- filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
- filt = _mm_and_si128(filt, mask);
- filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
- filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
-
- // Filter1 >> 3
- filter1 = _mm_srai_epi16(filter1, 0x3);
- filter2 = _mm_srai_epi16(filter2, 0x3);
-
- qs0 = _mm_adds_epi16(
- signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd),
- t80);
- ps0 = _mm_adds_epi16(
- signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd),
- t80);
- filt = _mm_adds_epi16(filter1, t1);
- filt = _mm_srai_epi16(filt, 1);
- filt = _mm_andnot_si128(hev, filt);
- qs1 = _mm_adds_epi16(
- signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd),
- t80);
- ps1 = _mm_adds_epi16(
- signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd),
- t80);
-
- // end highbd_filter4
- // loopfilter done
-
- // highbd_flat_mask4
- flat = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p0),
- _mm_subs_epu16(p0, p2)),
- _mm_or_si128(_mm_subs_epu16(p3, p0),
- _mm_subs_epu16(p0, p3)));
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(q2, q0),
- _mm_subs_epu16(q0, q2)),
- _mm_or_si128(_mm_subs_epu16(q3, q0),
- _mm_subs_epu16(q0, q3)));
- flat = _mm_max_epi16(work, flat);
- work = _mm_max_epi16(abs_p1p0, abs_q1q0);
- flat = _mm_max_epi16(work, flat);
-
- if (bd == 8)
- flat = _mm_subs_epu16(flat, one);
- else if (bd == 10)
- flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
- else // bd == 12
- flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
-
- flat = _mm_cmpeq_epi16(flat, zero);
- // end flat_mask4
-
- // flat & mask = flat && mask (as used in filter8)
- // (because, in both vars, each block of 16 either all 1s or all 0s)
- flat = _mm_and_si128(flat, mask);
-
- p5 = _mm_load_si128((__m128i *)(s - 6 * p));
- q5 = _mm_load_si128((__m128i *)(s + 5 * p));
- p6 = _mm_load_si128((__m128i *)(s - 7 * p));
- q6 = _mm_load_si128((__m128i *)(s + 6 * p));
- p7 = _mm_load_si128((__m128i *)(s - 8 * p));
- q7 = _mm_load_si128((__m128i *)(s + 7 * p));
-
- // highbd_flat_mask5 (arguments passed in are p0, q0, p4-p7, q4-q7
- // but referred to as p0-p4 & q0-q4 in fn)
- flat2 = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p4, p0),
- _mm_subs_epu16(p0, p4)),
- _mm_or_si128(_mm_subs_epu16(q4, q0),
- _mm_subs_epu16(q0, q4)));
-
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p5, p0),
- _mm_subs_epu16(p0, p5)),
- _mm_or_si128(_mm_subs_epu16(q5, q0),
- _mm_subs_epu16(q0, q5)));
- flat2 = _mm_max_epi16(work, flat2);
-
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p6, p0),
- _mm_subs_epu16(p0, p6)),
- _mm_or_si128(_mm_subs_epu16(q6, q0),
- _mm_subs_epu16(q0, q6)));
- flat2 = _mm_max_epi16(work, flat2);
-
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p7, p0),
- _mm_subs_epu16(p0, p7)),
- _mm_or_si128(_mm_subs_epu16(q7, q0),
- _mm_subs_epu16(q0, q7)));
- flat2 = _mm_max_epi16(work, flat2);
-
- if (bd == 8)
- flat2 = _mm_subs_epu16(flat2, one);
- else if (bd == 10)
- flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 2));
- else // bd == 12
- flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 4));
-
- flat2 = _mm_cmpeq_epi16(flat2, zero);
- flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
- // end highbd_flat_mask5
-
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- // flat and wide flat calculations
- eight = _mm_set1_epi16(8);
- four = _mm_set1_epi16(4);
-
- pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6, p5),
- _mm_add_epi16(p4, p3));
- pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5),
- _mm_add_epi16(q4, q3));
-
- pixetFilter_p2p1p0 = _mm_add_epi16(p0, _mm_add_epi16(p2, p1));
- pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
-
- pixetFilter_q2q1q0 = _mm_add_epi16(q0, _mm_add_epi16(q2, q1));
- pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
- pixelFilter_p = _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
- pixelFilter_q));
- pixetFilter_p2p1p0 = _mm_add_epi16(four,
- _mm_add_epi16(pixetFilter_p2p1p0,
- pixetFilter_q2q1q0));
- flat2_p0 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(p7, p0)), 4);
- flat2_q0 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(q7, q0)), 4);
- flat_p0 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(p3, p0)), 3);
- flat_q0 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(q3, q0)), 3);
-
- sum_p7 = _mm_add_epi16(p7, p7);
- sum_q7 = _mm_add_epi16(q7, q7);
- sum_p3 = _mm_add_epi16(p3, p3);
- sum_q3 = _mm_add_epi16(q3, q3);
-
- pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6);
- flat2_p1 = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1)), 4);
- flat2_q1 = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1)), 4);
-
- pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2);
- pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2);
- flat_p1 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(sum_p3, p1)), 3);
- flat_q1 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
- _mm_add_epi16(sum_q3, q1)), 3);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7);
- sum_q7 = _mm_add_epi16(sum_q7, q7);
- sum_p3 = _mm_add_epi16(sum_p3, p3);
- sum_q3 = _mm_add_epi16(sum_q3, q3);
-
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5);
- flat2_p2 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p2)), 4);
- flat2_q2 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q2)), 4);
-
- pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1);
- pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1);
- flat_p2 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(sum_p3, p2)), 3);
- flat_q2 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
- _mm_add_epi16(sum_q3, q2)), 3);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7);
- sum_q7 = _mm_add_epi16(sum_q7, q7);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4);
- flat2_p3 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p3)), 4);
- flat2_q3 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q3)), 4);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7);
- sum_q7 = _mm_add_epi16(sum_q7, q7);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3);
- flat2_p4 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p4)), 4);
- flat2_q4 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q4)), 4);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7);
- sum_q7 = _mm_add_epi16(sum_q7, q7);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2);
- flat2_p5 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p5)), 4);
- flat2_q5 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q5)), 4);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7);
- sum_q7 = _mm_add_epi16(sum_q7, q7);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1);
- flat2_p6 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p6)), 4);
- flat2_q6 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q6)), 4);
-
- // wide flat
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
- // highbd_filter8
- p2 = _mm_andnot_si128(flat, p2);
- // p2 remains unchanged if !(flat && mask)
- flat_p2 = _mm_and_si128(flat, flat_p2);
- // when (flat && mask)
- p2 = _mm_or_si128(p2, flat_p2); // full list of p2 values
- q2 = _mm_andnot_si128(flat, q2);
- flat_q2 = _mm_and_si128(flat, flat_q2);
- q2 = _mm_or_si128(q2, flat_q2); // full list of q2 values
-
- ps1 = _mm_andnot_si128(flat, ps1);
- // p1 takes the value assigned to in in filter4 if !(flat && mask)
- flat_p1 = _mm_and_si128(flat, flat_p1);
- // when (flat && mask)
- p1 = _mm_or_si128(ps1, flat_p1); // full list of p1 values
- qs1 = _mm_andnot_si128(flat, qs1);
- flat_q1 = _mm_and_si128(flat, flat_q1);
- q1 = _mm_or_si128(qs1, flat_q1); // full list of q1 values
-
- ps0 = _mm_andnot_si128(flat, ps0);
- // p0 takes the value assigned to in in filter4 if !(flat && mask)
- flat_p0 = _mm_and_si128(flat, flat_p0);
- // when (flat && mask)
- p0 = _mm_or_si128(ps0, flat_p0); // full list of p0 values
- qs0 = _mm_andnot_si128(flat, qs0);
- flat_q0 = _mm_and_si128(flat, flat_q0);
- q0 = _mm_or_si128(qs0, flat_q0); // full list of q0 values
- // end highbd_filter8
-
- // highbd_filter16
- p6 = _mm_andnot_si128(flat2, p6);
- // p6 remains unchanged if !(flat2 && flat && mask)
- flat2_p6 = _mm_and_si128(flat2, flat2_p6);
- // get values for when (flat2 && flat && mask)
- p6 = _mm_or_si128(p6, flat2_p6); // full list of p6 values
- q6 = _mm_andnot_si128(flat2, q6);
- // q6 remains unchanged if !(flat2 && flat && mask)
- flat2_q6 = _mm_and_si128(flat2, flat2_q6);
- // get values for when (flat2 && flat && mask)
- q6 = _mm_or_si128(q6, flat2_q6); // full list of q6 values
- _mm_store_si128((__m128i *)(s - 7 * p), p6);
- _mm_store_si128((__m128i *)(s + 6 * p), q6);
-
- p5 = _mm_andnot_si128(flat2, p5);
- // p5 remains unchanged if !(flat2 && flat && mask)
- flat2_p5 = _mm_and_si128(flat2, flat2_p5);
- // get values for when (flat2 && flat && mask)
- p5 = _mm_or_si128(p5, flat2_p5);
- // full list of p5 values
- q5 = _mm_andnot_si128(flat2, q5);
- // q5 remains unchanged if !(flat2 && flat && mask)
- flat2_q5 = _mm_and_si128(flat2, flat2_q5);
- // get values for when (flat2 && flat && mask)
- q5 = _mm_or_si128(q5, flat2_q5);
- // full list of q5 values
- _mm_store_si128((__m128i *)(s - 6 * p), p5);
- _mm_store_si128((__m128i *)(s + 5 * p), q5);
-
- p4 = _mm_andnot_si128(flat2, p4);
- // p4 remains unchanged if !(flat2 && flat && mask)
- flat2_p4 = _mm_and_si128(flat2, flat2_p4);
- // get values for when (flat2 && flat && mask)
- p4 = _mm_or_si128(p4, flat2_p4); // full list of p4 values
- q4 = _mm_andnot_si128(flat2, q4);
- // q4 remains unchanged if !(flat2 && flat && mask)
- flat2_q4 = _mm_and_si128(flat2, flat2_q4);
- // get values for when (flat2 && flat && mask)
- q4 = _mm_or_si128(q4, flat2_q4); // full list of q4 values
- _mm_store_si128((__m128i *)(s - 5 * p), p4);
- _mm_store_si128((__m128i *)(s + 4 * p), q4);
-
- p3 = _mm_andnot_si128(flat2, p3);
- // p3 takes value from highbd_filter8 if !(flat2 && flat && mask)
- flat2_p3 = _mm_and_si128(flat2, flat2_p3);
- // get values for when (flat2 && flat && mask)
- p3 = _mm_or_si128(p3, flat2_p3); // full list of p3 values
- q3 = _mm_andnot_si128(flat2, q3);
- // q3 takes value from highbd_filter8 if !(flat2 && flat && mask)
- flat2_q3 = _mm_and_si128(flat2, flat2_q3);
- // get values for when (flat2 && flat && mask)
- q3 = _mm_or_si128(q3, flat2_q3); // full list of q3 values
- _mm_store_si128((__m128i *)(s - 4 * p), p3);
- _mm_store_si128((__m128i *)(s + 3 * p), q3);
-
- p2 = _mm_andnot_si128(flat2, p2);
- // p2 takes value from highbd_filter8 if !(flat2 && flat && mask)
- flat2_p2 = _mm_and_si128(flat2, flat2_p2);
- // get values for when (flat2 && flat && mask)
- p2 = _mm_or_si128(p2, flat2_p2);
- // full list of p2 values
- q2 = _mm_andnot_si128(flat2, q2);
- // q2 takes value from highbd_filter8 if !(flat2 && flat && mask)
- flat2_q2 = _mm_and_si128(flat2, flat2_q2);
- // get values for when (flat2 && flat && mask)
- q2 = _mm_or_si128(q2, flat2_q2); // full list of q2 values
- _mm_store_si128((__m128i *)(s - 3 * p), p2);
- _mm_store_si128((__m128i *)(s + 2 * p), q2);
-
- p1 = _mm_andnot_si128(flat2, p1);
- // p1 takes value from highbd_filter8 if !(flat2 && flat && mask)
- flat2_p1 = _mm_and_si128(flat2, flat2_p1);
- // get values for when (flat2 && flat && mask)
- p1 = _mm_or_si128(p1, flat2_p1); // full list of p1 values
- q1 = _mm_andnot_si128(flat2, q1);
- // q1 takes value from highbd_filter8 if !(flat2 && flat && mask)
- flat2_q1 = _mm_and_si128(flat2, flat2_q1);
- // get values for when (flat2 && flat && mask)
- q1 = _mm_or_si128(q1, flat2_q1); // full list of q1 values
- _mm_store_si128((__m128i *)(s - 2 * p), p1);
- _mm_store_si128((__m128i *)(s + 1 * p), q1);
-
- p0 = _mm_andnot_si128(flat2, p0);
- // p0 takes value from highbd_filter8 if !(flat2 && flat && mask)
- flat2_p0 = _mm_and_si128(flat2, flat2_p0);
- // get values for when (flat2 && flat && mask)
- p0 = _mm_or_si128(p0, flat2_p0); // full list of p0 values
- q0 = _mm_andnot_si128(flat2, q0);
- // q0 takes value from highbd_filter8 if !(flat2 && flat && mask)
- flat2_q0 = _mm_and_si128(flat2, flat2_q0);
- // get values for when (flat2 && flat && mask)
- q0 = _mm_or_si128(q0, flat2_q0); // full list of q0 values
- _mm_store_si128((__m128i *)(s - 1 * p), p0);
- _mm_store_si128((__m128i *)(s - 0 * p), q0);
-}
-
-static void highbd_mb_lpf_horizontal_edge_w_sse2_16(uint16_t *s,
- int p,
- const uint8_t *_blimit,
- const uint8_t *_limit,
- const uint8_t *_thresh,
- int bd) {- highbd_mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh, bd);
- highbd_mb_lpf_horizontal_edge_w_sse2_8(s + 8, p, _blimit, _limit, _thresh,
- bd);
-}
-
-// TODO(yunqingwang): remove count and call these 2 functions(8 or 16) directly.
-void vp9_highbd_lpf_horizontal_16_sse2(uint16_t *s, int p,
- const uint8_t *_blimit,
- const uint8_t *_limit,
- const uint8_t *_thresh,
- int count, int bd) {- if (count == 1)
- highbd_mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh, bd);
- else
- highbd_mb_lpf_horizontal_edge_w_sse2_16(s, p, _blimit, _limit, _thresh, bd);
-}
-
-void vp9_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p,
- const uint8_t *_blimit,
- const uint8_t *_limit,
- const uint8_t *_thresh,
- int count, int bd) {- DECLARE_ALIGNED(16, uint16_t, flat_op2[16]);
- DECLARE_ALIGNED(16, uint16_t, flat_op1[16]);
- DECLARE_ALIGNED(16, uint16_t, flat_op0[16]);
- DECLARE_ALIGNED(16, uint16_t, flat_oq2[16]);
- DECLARE_ALIGNED(16, uint16_t, flat_oq1[16]);
- DECLARE_ALIGNED(16, uint16_t, flat_oq0[16]);
- const __m128i zero = _mm_set1_epi16(0);
- __m128i blimit, limit, thresh;
- __m128i mask, hev, flat;
- __m128i p3 = _mm_load_si128((__m128i *)(s - 4 * p));
- __m128i q3 = _mm_load_si128((__m128i *)(s + 3 * p));
- __m128i p2 = _mm_load_si128((__m128i *)(s - 3 * p));
- __m128i q2 = _mm_load_si128((__m128i *)(s + 2 * p));
- __m128i p1 = _mm_load_si128((__m128i *)(s - 2 * p));
- __m128i q1 = _mm_load_si128((__m128i *)(s + 1 * p));
- __m128i p0 = _mm_load_si128((__m128i *)(s - 1 * p));
- __m128i q0 = _mm_load_si128((__m128i *)(s + 0 * p));
- const __m128i one = _mm_set1_epi16(1);
- const __m128i ffff = _mm_cmpeq_epi16(one, one);
- __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
- const __m128i four = _mm_set1_epi16(4);
- __m128i workp_a, workp_b, workp_shft;
-
- const __m128i t4 = _mm_set1_epi16(4);
- const __m128i t3 = _mm_set1_epi16(3);
- __m128i t80;
- const __m128i t1 = _mm_set1_epi16(0x1);
- __m128i ps1, ps0, qs0, qs1;
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
-
- (void)count;
-
- if (bd == 8) {- blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
- limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
- thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
- t80 = _mm_set1_epi16(0x80);
- } else if (bd == 10) {- blimit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
- limit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
- thresh = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
- t80 = _mm_set1_epi16(0x200);
- } else { // bd == 12- blimit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
- limit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
- thresh = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
- t80 = _mm_set1_epi16(0x800);
- }
-
- ps1 = _mm_subs_epi16(p1, t80);
- ps0 = _mm_subs_epi16(p0, t80);
- qs0 = _mm_subs_epi16(q0, t80);
- qs1 = _mm_subs_epi16(q1, t80);
-
- // filter_mask and hev_mask
- abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0),
- _mm_subs_epu16(p0, p1));
- abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0),
- _mm_subs_epu16(q0, q1));
-
- abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0),
- _mm_subs_epu16(q0, p0));
- abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1),
- _mm_subs_epu16(q1, p1));
- flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu16(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
-
- abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
- mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- // So taking maximums continues to work:
- mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
- mask = _mm_max_epi16(abs_p1p0, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- mask = _mm_max_epi16(abs_q1q0, mask);
- // mask |= (abs(q1 - q0) > limit) * -1;
-
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
- _mm_subs_epu16(p1, p2)),
- _mm_or_si128(_mm_subs_epu16(q2, q1),
- _mm_subs_epu16(q1, q2)));
- mask = _mm_max_epi16(work, mask);
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p2),
- _mm_subs_epu16(p2, p3)),
- _mm_or_si128(_mm_subs_epu16(q3, q2),
- _mm_subs_epu16(q2, q3)));
- mask = _mm_max_epi16(work, mask);
- mask = _mm_subs_epu16(mask, limit);
- mask = _mm_cmpeq_epi16(mask, zero);
-
- // flat_mask4
- flat = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p0),
- _mm_subs_epu16(p0, p2)),
- _mm_or_si128(_mm_subs_epu16(q2, q0),
- _mm_subs_epu16(q0, q2)));
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p0),
- _mm_subs_epu16(p0, p3)),
- _mm_or_si128(_mm_subs_epu16(q3, q0),
- _mm_subs_epu16(q0, q3)));
- flat = _mm_max_epi16(work, flat);
- flat = _mm_max_epi16(abs_p1p0, flat);
- flat = _mm_max_epi16(abs_q1q0, flat);
-
- if (bd == 8)
- flat = _mm_subs_epu16(flat, one);
- else if (bd == 10)
- flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
- else // bd == 12
- flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
-
- flat = _mm_cmpeq_epi16(flat, zero);
- flat = _mm_and_si128(flat, mask); // flat & mask
-
- // Added before shift for rounding part of ROUND_POWER_OF_TWO
-
- workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
- workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
- workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_store_si128((__m128i *)&flat_op2[0], workp_shft);
-
- workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_store_si128((__m128i *)&flat_op1[0], workp_shft);
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_store_si128((__m128i *)&flat_op0[0], workp_shft);
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_store_si128((__m128i *)&flat_oq0[0], workp_shft);
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_store_si128((__m128i *)&flat_oq1[0], workp_shft);
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_store_si128((__m128i *)&flat_oq2[0], workp_shft);
-
- // lp filter
- filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
- filt = _mm_and_si128(filt, hev);
- work_a = _mm_subs_epi16(qs0, ps0);
- filt = _mm_adds_epi16(filt, work_a);
- filt = _mm_adds_epi16(filt, work_a);
- filt = _mm_adds_epi16(filt, work_a);
- // (vp9_filter + 3 * (qs0 - ps0)) & mask
- filt = signed_char_clamp_bd_sse2(filt, bd);
- filt = _mm_and_si128(filt, mask);
-
- filter1 = _mm_adds_epi16(filt, t4);
- filter2 = _mm_adds_epi16(filt, t3);
-
- // Filter1 >> 3
- filter1 = signed_char_clamp_bd_sse2(filter1, bd);
- filter1 = _mm_srai_epi16(filter1, 3);
-
- // Filter2 >> 3
- filter2 = signed_char_clamp_bd_sse2(filter2, bd);
- filter2 = _mm_srai_epi16(filter2, 3);
-
- // filt >> 1
- filt = _mm_adds_epi16(filter1, t1);
- filt = _mm_srai_epi16(filt, 1);
- // filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
- filt = _mm_andnot_si128(hev, filt);
-
- work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd);
- work_a = _mm_adds_epi16(work_a, t80);
- q0 = _mm_load_si128((__m128i *)flat_oq0);
- work_a = _mm_andnot_si128(flat, work_a);
- q0 = _mm_and_si128(flat, q0);
- q0 = _mm_or_si128(work_a, q0);
-
- work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd);
- work_a = _mm_adds_epi16(work_a, t80);
- q1 = _mm_load_si128((__m128i *)flat_oq1);
- work_a = _mm_andnot_si128(flat, work_a);
- q1 = _mm_and_si128(flat, q1);
- q1 = _mm_or_si128(work_a, q1);
-
- work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
- q2 = _mm_load_si128((__m128i *)flat_oq2);
- work_a = _mm_andnot_si128(flat, work_a);
- q2 = _mm_and_si128(flat, q2);
- q2 = _mm_or_si128(work_a, q2);
-
- work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd);
- work_a = _mm_adds_epi16(work_a, t80);
- p0 = _mm_load_si128((__m128i *)flat_op0);
- work_a = _mm_andnot_si128(flat, work_a);
- p0 = _mm_and_si128(flat, p0);
- p0 = _mm_or_si128(work_a, p0);
-
- work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd);
- work_a = _mm_adds_epi16(work_a, t80);
- p1 = _mm_load_si128((__m128i *)flat_op1);
- work_a = _mm_andnot_si128(flat, work_a);
- p1 = _mm_and_si128(flat, p1);
- p1 = _mm_or_si128(work_a, p1);
-
- work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
- p2 = _mm_load_si128((__m128i *)flat_op2);
- work_a = _mm_andnot_si128(flat, work_a);
- p2 = _mm_and_si128(flat, p2);
- p2 = _mm_or_si128(work_a, p2);
-
- _mm_store_si128((__m128i *)(s - 3 * p), p2);
- _mm_store_si128((__m128i *)(s - 2 * p), p1);
- _mm_store_si128((__m128i *)(s - 1 * p), p0);
- _mm_store_si128((__m128i *)(s + 0 * p), q0);
- _mm_store_si128((__m128i *)(s + 1 * p), q1);
- _mm_store_si128((__m128i *)(s + 2 * p), q2);
-}
-
-void vp9_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int p,
- const uint8_t *_blimit0,
- const uint8_t *_limit0,
- const uint8_t *_thresh0,
- const uint8_t *_blimit1,
- const uint8_t *_limit1,
- const uint8_t *_thresh1,
- int bd) {- vp9_highbd_lpf_horizontal_8_sse2(s, p, _blimit0, _limit0, _thresh0, 1, bd);
- vp9_highbd_lpf_horizontal_8_sse2(s + 8, p, _blimit1, _limit1, _thresh1,
- 1, bd);
-}
-
-void vp9_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
- const uint8_t *_blimit,
- const uint8_t *_limit,
- const uint8_t *_thresh,
- int count, int bd) {- const __m128i zero = _mm_set1_epi16(0);
- __m128i blimit, limit, thresh;
- __m128i mask, hev, flat;
- __m128i p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
- __m128i p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
- __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
- __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
- __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
- __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
- __m128i q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
- __m128i q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
- const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0),
- _mm_subs_epu16(p0, p1));
- const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0),
- _mm_subs_epu16(q0, q1));
- const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
- const __m128i one = _mm_set1_epi16(1);
- __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0),
- _mm_subs_epu16(q0, p0));
- __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1),
- _mm_subs_epu16(q1, p1));
- __m128i work;
- const __m128i t4 = _mm_set1_epi16(4);
- const __m128i t3 = _mm_set1_epi16(3);
- __m128i t80;
- __m128i tff80;
- __m128i tffe0;
- __m128i t1f;
- // equivalent to shifting 0x1f left by bitdepth - 8
- // and setting new bits to 1
- const __m128i t1 = _mm_set1_epi16(0x1);
- __m128i t7f;
- // equivalent to shifting 0x7f left by bitdepth - 8
- // and setting new bits to 1
- __m128i ps1, ps0, qs0, qs1;
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
-
- (void)count;
-
- if (bd == 8) {- blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
- limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
- thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
- t80 = _mm_set1_epi16(0x80);
- tff80 = _mm_set1_epi16(0xff80);
- tffe0 = _mm_set1_epi16(0xffe0);
- t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 8);
- t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 8);
- } else if (bd == 10) {- blimit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
- limit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
- thresh = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
- t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 2);
- tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 2);
- tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 2);
- t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 6);
- t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 6);
- } else { // bd == 12- blimit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
- limit = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
- thresh = _mm_slli_epi16(
- _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
- t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 4);
- tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 4);
- tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 4);
- t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 4);
- t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 4);
- }
-
- ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80);
- ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80);
- qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80);
- qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80);
-
- // filter_mask and hev_mask
- flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu16(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
-
- abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
- mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- // So taking maximums continues to work:
- mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
- mask = _mm_max_epi16(flat, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- // mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
- _mm_subs_epu16(p1, p2)),
- _mm_or_si128(_mm_subs_epu16(p3, p2),
- _mm_subs_epu16(p2, p3)));
- mask = _mm_max_epi16(work, mask);
- work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(q2, q1),
- _mm_subs_epu16(q1, q2)),
- _mm_or_si128(_mm_subs_epu16(q3, q2),
- _mm_subs_epu16(q2, q3)));
- mask = _mm_max_epi16(work, mask);
- mask = _mm_subs_epu16(mask, limit);
- mask = _mm_cmpeq_epi16(mask, zero);
-
- // filter4
- filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
- filt = _mm_and_si128(filt, hev);
- work_a = _mm_subs_epi16(qs0, ps0);
- filt = _mm_adds_epi16(filt, work_a);
- filt = _mm_adds_epi16(filt, work_a);
- filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
-
- // (vp9_filter + 3 * (qs0 - ps0)) & mask
- filt = _mm_and_si128(filt, mask);
-
- filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
- filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
-
- // Filter1 >> 3
- work_a = _mm_cmpgt_epi16(zero, filter1); // get the values that are <0
- filter1 = _mm_srli_epi16(filter1, 3);
- work_a = _mm_and_si128(work_a, tffe0); // sign bits for the values < 0
- filter1 = _mm_and_si128(filter1, t1f); // clamp the range
- filter1 = _mm_or_si128(filter1, work_a); // reinsert the sign bits
-
- // Filter2 >> 3
- work_a = _mm_cmpgt_epi16(zero, filter2);
- filter2 = _mm_srli_epi16(filter2, 3);
- work_a = _mm_and_si128(work_a, tffe0);
- filter2 = _mm_and_si128(filter2, t1f);
- filter2 = _mm_or_si128(filter2, work_a);
-
- // filt >> 1
- filt = _mm_adds_epi16(filter1, t1);
- work_a = _mm_cmpgt_epi16(zero, filt);
- filt = _mm_srli_epi16(filt, 1);
- work_a = _mm_and_si128(work_a, tff80);
- filt = _mm_and_si128(filt, t7f);
- filt = _mm_or_si128(filt, work_a);
-
- filt = _mm_andnot_si128(hev, filt);
-
- q0 = _mm_adds_epi16(
- signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80);
- q1 = _mm_adds_epi16(
- signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), t80);
- p0 = _mm_adds_epi16(
- signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80);
- p1 = _mm_adds_epi16(
- signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd), t80);
-
- _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
- _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
- _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
- _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
-}
-
-void vp9_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int p,
- const uint8_t *_blimit0,
- const uint8_t *_limit0,
- const uint8_t *_thresh0,
- const uint8_t *_blimit1,
- const uint8_t *_limit1,
- const uint8_t *_thresh1,
- int bd) {- vp9_highbd_lpf_horizontal_4_sse2(s, p, _blimit0, _limit0, _thresh0, 1, bd);
- vp9_highbd_lpf_horizontal_4_sse2(s + 8, p, _blimit1, _limit1, _thresh1, 1,
- bd);
-}
-
-static INLINE void highbd_transpose(uint16_t *src[], int in_p,
- uint16_t *dst[], int out_p,
- int num_8x8_to_transpose) {- int idx8x8 = 0;
- __m128i p0, p1, p2, p3, p4, p5, p6, p7, x0, x1, x2, x3, x4, x5, x6, x7;
- do {- uint16_t *in = src[idx8x8];
- uint16_t *out = dst[idx8x8];
-
- p0 = _mm_loadu_si128((__m128i *)(in + 0*in_p)); // 00 01 02 03 04 05 06 07
- p1 = _mm_loadu_si128((__m128i *)(in + 1*in_p)); // 10 11 12 13 14 15 16 17
- p2 = _mm_loadu_si128((__m128i *)(in + 2*in_p)); // 20 21 22 23 24 25 26 27
- p3 = _mm_loadu_si128((__m128i *)(in + 3*in_p)); // 30 31 32 33 34 35 36 37
- p4 = _mm_loadu_si128((__m128i *)(in + 4*in_p)); // 40 41 42 43 44 45 46 47
- p5 = _mm_loadu_si128((__m128i *)(in + 5*in_p)); // 50 51 52 53 54 55 56 57
- p6 = _mm_loadu_si128((__m128i *)(in + 6*in_p)); // 60 61 62 63 64 65 66 67
- p7 = _mm_loadu_si128((__m128i *)(in + 7*in_p)); // 70 71 72 73 74 75 76 77
- // 00 10 01 11 02 12 03 13
- x0 = _mm_unpacklo_epi16(p0, p1);
- // 20 30 21 31 22 32 23 33
- x1 = _mm_unpacklo_epi16(p2, p3);
- // 40 50 41 51 42 52 43 53
- x2 = _mm_unpacklo_epi16(p4, p5);
- // 60 70 61 71 62 72 63 73
- x3 = _mm_unpacklo_epi16(p6, p7);
- // 00 10 20 30 01 11 21 31
- x4 = _mm_unpacklo_epi32(x0, x1);
- // 40 50 60 70 41 51 61 71
- x5 = _mm_unpacklo_epi32(x2, x3);
- // 00 10 20 30 40 50 60 70
- x6 = _mm_unpacklo_epi64(x4, x5);
- // 01 11 21 31 41 51 61 71
- x7 = _mm_unpackhi_epi64(x4, x5);
-
- _mm_storeu_si128((__m128i *)(out + 0*out_p), x6);
- // 00 10 20 30 40 50 60 70
- _mm_storeu_si128((__m128i *)(out + 1*out_p), x7);
- // 01 11 21 31 41 51 61 71
-
- // 02 12 22 32 03 13 23 33
- x4 = _mm_unpackhi_epi32(x0, x1);
- // 42 52 62 72 43 53 63 73
- x5 = _mm_unpackhi_epi32(x2, x3);
- // 02 12 22 32 42 52 62 72
- x6 = _mm_unpacklo_epi64(x4, x5);
- // 03 13 23 33 43 53 63 73
- x7 = _mm_unpackhi_epi64(x4, x5);
-
- _mm_storeu_si128((__m128i *)(out + 2*out_p), x6);
- // 02 12 22 32 42 52 62 72
- _mm_storeu_si128((__m128i *)(out + 3*out_p), x7);
- // 03 13 23 33 43 53 63 73
-
- // 04 14 05 15 06 16 07 17
- x0 = _mm_unpackhi_epi16(p0, p1);
- // 24 34 25 35 26 36 27 37
- x1 = _mm_unpackhi_epi16(p2, p3);
- // 44 54 45 55 46 56 47 57
- x2 = _mm_unpackhi_epi16(p4, p5);
- // 64 74 65 75 66 76 67 77
- x3 = _mm_unpackhi_epi16(p6, p7);
- // 04 14 24 34 05 15 25 35
- x4 = _mm_unpacklo_epi32(x0, x1);
- // 44 54 64 74 45 55 65 75
- x5 = _mm_unpacklo_epi32(x2, x3);
- // 04 14 24 34 44 54 64 74
- x6 = _mm_unpacklo_epi64(x4, x5);
- // 05 15 25 35 45 55 65 75
- x7 = _mm_unpackhi_epi64(x4, x5);
-
- _mm_storeu_si128((__m128i *)(out + 4*out_p), x6);
- // 04 14 24 34 44 54 64 74
- _mm_storeu_si128((__m128i *)(out + 5*out_p), x7);
- // 05 15 25 35 45 55 65 75
-
- // 06 16 26 36 07 17 27 37
- x4 = _mm_unpackhi_epi32(x0, x1);
- // 46 56 66 76 47 57 67 77
- x5 = _mm_unpackhi_epi32(x2, x3);
- // 06 16 26 36 46 56 66 76
- x6 = _mm_unpacklo_epi64(x4, x5);
- // 07 17 27 37 47 57 67 77
- x7 = _mm_unpackhi_epi64(x4, x5);
-
- _mm_storeu_si128((__m128i *)(out + 6*out_p), x6);
- // 06 16 26 36 46 56 66 76
- _mm_storeu_si128((__m128i *)(out + 7*out_p), x7);
- // 07 17 27 37 47 57 67 77
- } while (++idx8x8 < num_8x8_to_transpose);
-}
-
-static INLINE void highbd_transpose8x16(uint16_t *in0, uint16_t *in1,
- int in_p, uint16_t *out, int out_p) {- uint16_t *src0[1];
- uint16_t *src1[1];
- uint16_t *dest0[1];
- uint16_t *dest1[1];
- src0[0] = in0;
- src1[0] = in1;
- dest0[0] = out;
- dest1[0] = out + 8;
- highbd_transpose(src0, in_p, dest0, out_p, 1);
- highbd_transpose(src1, in_p, dest1, out_p, 1);
-}
-
-void vp9_highbd_lpf_vertical_4_sse2(uint16_t *s, int p,
- const uint8_t *blimit,
- const uint8_t *limit,
- const uint8_t *thresh,
- int count, int bd) {- DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
- uint16_t *src[1];
- uint16_t *dst[1];
- (void)count;
-
- // Transpose 8x8
- src[0] = s - 4;
- dst[0] = t_dst;
-
- highbd_transpose(src, p, dst, 8, 1);
-
- // Loop filtering
- vp9_highbd_lpf_horizontal_4_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1,
- bd);
-
- src[0] = t_dst;
- dst[0] = s - 4;
-
- // Transpose back
- highbd_transpose(src, 8, dst, p, 1);
-}
-
-void vp9_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int p,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1,
- int bd) {- DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
- uint16_t *src[2];
- uint16_t *dst[2];
-
- // Transpose 8x16
- highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
-
- // Loop filtering
- vp9_highbd_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
- thresh0, blimit1, limit1, thresh1, bd);
- src[0] = t_dst;
- src[1] = t_dst + 8;
- dst[0] = s - 4;
- dst[1] = s - 4 + p * 8;
-
- // Transpose back
- highbd_transpose(src, 16, dst, p, 2);
-}
-
-void vp9_highbd_lpf_vertical_8_sse2(uint16_t *s, int p,
- const uint8_t *blimit,
- const uint8_t *limit,
- const uint8_t *thresh,
- int count, int bd) {- DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
- uint16_t *src[1];
- uint16_t *dst[1];
- (void)count;
-
- // Transpose 8x8
- src[0] = s - 4;
- dst[0] = t_dst;
-
- highbd_transpose(src, p, dst, 8, 1);
-
- // Loop filtering
- vp9_highbd_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1,
- bd);
-
- src[0] = t_dst;
- dst[0] = s - 4;
-
- // Transpose back
- highbd_transpose(src, 8, dst, p, 1);
-}
-
-void vp9_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int p,
- const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1,
- int bd) {- DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
- uint16_t *src[2];
- uint16_t *dst[2];
-
- // Transpose 8x16
- highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
-
- // Loop filtering
- vp9_highbd_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
- thresh0, blimit1, limit1, thresh1, bd);
- src[0] = t_dst;
- src[1] = t_dst + 8;
-
- dst[0] = s - 4;
- dst[1] = s - 4 + p * 8;
-
- // Transpose back
- highbd_transpose(src, 16, dst, p, 2);
-}
-
-void vp9_highbd_lpf_vertical_16_sse2(uint16_t *s, int p,
- const uint8_t *blimit,
- const uint8_t *limit,
- const uint8_t *thresh,
- int bd) {- DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 16]);
- uint16_t *src[2];
- uint16_t *dst[2];
-
- src[0] = s - 8;
- src[1] = s;
- dst[0] = t_dst;
- dst[1] = t_dst + 8 * 8;
-
- // Transpose 16x8
- highbd_transpose(src, p, dst, 8, 2);
-
- // Loop filtering
- highbd_mb_lpf_horizontal_edge_w_sse2_8(t_dst + 8 * 8, 8, blimit, limit,
- thresh, bd);
- src[0] = t_dst;
- src[1] = t_dst + 8 * 8;
- dst[0] = s - 8;
- dst[1] = s;
-
- // Transpose back
- highbd_transpose(src, 8, dst, p, 2);
-}
-
-void vp9_highbd_lpf_vertical_16_dual_sse2(uint16_t *s,
- int p,
- const uint8_t *blimit,
- const uint8_t *limit,
- const uint8_t *thresh,
- int bd) {- DECLARE_ALIGNED(16, uint16_t, t_dst[256]);
-
- // Transpose 16x16
- highbd_transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
- highbd_transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
-
- // Loop filtering
- highbd_mb_lpf_horizontal_edge_w_sse2_16(t_dst + 8 * 16, 16, blimit, limit,
- thresh, bd);
-
- // Transpose back
- highbd_transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
- highbd_transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
-}
--- a/vp9/common/x86/vp9_loopfilter_intrin_avx2.c
+++ /dev/null
@@ -1,986 +1,0 @@
-/*
- * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <immintrin.h> /* AVX2 */
-
-#include "./vp9_rtcd.h"
-#include "vpx_ports/mem.h"
-
-static void mb_lpf_horizontal_edge_w_avx2_8(unsigned char *s, int p,
- const unsigned char *_blimit, const unsigned char *_limit,
- const unsigned char *_thresh) {- __m128i mask, hev, flat, flat2;
- const __m128i zero = _mm_set1_epi16(0);
- const __m128i one = _mm_set1_epi8(1);
- __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
- __m128i abs_p1p0;
-
- const __m128i thresh = _mm_broadcastb_epi8(
- _mm_cvtsi32_si128((int) _thresh[0]));
- const __m128i limit = _mm_broadcastb_epi8(
- _mm_cvtsi32_si128((int) _limit[0]));
- const __m128i blimit = _mm_broadcastb_epi8(
- _mm_cvtsi32_si128((int) _blimit[0]));
-
- q4p4 = _mm_loadl_epi64((__m128i *) (s - 5 * p));
- q4p4 = _mm_castps_si128(
- _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *) (s + 4 * p)));
- q3p3 = _mm_loadl_epi64((__m128i *) (s - 4 * p));
- q3p3 = _mm_castps_si128(
- _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *) (s + 3 * p)));
- q2p2 = _mm_loadl_epi64((__m128i *) (s - 3 * p));
- q2p2 = _mm_castps_si128(
- _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *) (s + 2 * p)));
- q1p1 = _mm_loadl_epi64((__m128i *) (s - 2 * p));
- q1p1 = _mm_castps_si128(
- _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *) (s + 1 * p)));
- p1q1 = _mm_shuffle_epi32(q1p1, 78);
- q0p0 = _mm_loadl_epi64((__m128i *) (s - 1 * p));
- q0p0 = _mm_castps_si128(
- _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *) (s - 0 * p)));
- p0q0 = _mm_shuffle_epi32(q0p0, 78);
-
- {- __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
- abs_p1p0 = _mm_or_si128(_mm_subs_epu8(q1p1, q0p0),
- _mm_subs_epu8(q0p0, q1p1));
- abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
- fe = _mm_set1_epi8(0xfe);
- ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
- abs_p0q0 = _mm_or_si128(_mm_subs_epu8(q0p0, p0q0),
- _mm_subs_epu8(p0q0, q0p0));
- abs_p1q1 = _mm_or_si128(_mm_subs_epu8(q1p1, p1q1),
- _mm_subs_epu8(p1q1, q1p1));
- flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu8(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
-
- abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
- mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(abs_p1p0, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- // mask |= (abs(q1 - q0) > limit) * -1;
-
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(q2p2, q1p1),
- _mm_subs_epu8(q1p1, q2p2)),
- _mm_or_si128(_mm_subs_epu8(q3p3, q2p2),
- _mm_subs_epu8(q2p2, q3p3)));
- mask = _mm_max_epu8(work, mask);
- mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
- mask = _mm_subs_epu8(mask, limit);
- mask = _mm_cmpeq_epi8(mask, zero);
- }
-
- // lp filter
- {- const __m128i t4 = _mm_set1_epi8(4);
- const __m128i t3 = _mm_set1_epi8(3);
- const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i t1 = _mm_set1_epi16(0x1);
- __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
- __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
- __m128i qs0 = _mm_xor_si128(p0q0, t80);
- __m128i qs1 = _mm_xor_si128(p1q1, t80);
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
- __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
- __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
-
- filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, qs0ps0);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- /* (vp9_filter + 3 * (qs0 - ps0)) & mask */
- filt = _mm_and_si128(filt, mask);
-
- filter1 = _mm_adds_epi8(filt, t4);
- filter2 = _mm_adds_epi8(filt, t3);
-
- filter1 = _mm_unpacklo_epi8(zero, filter1);
- filter1 = _mm_srai_epi16(filter1, 0xB);
- filter2 = _mm_unpacklo_epi8(zero, filter2);
- filter2 = _mm_srai_epi16(filter2, 0xB);
-
- /* Filter1 >> 3 */
- filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
- qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
-
- /* filt >> 1 */
- filt = _mm_adds_epi16(filter1, t1);
- filt = _mm_srai_epi16(filt, 1);
- filt = _mm_andnot_si128(
- _mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8), filt);
- filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
- qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
- // loopfilter done
-
- {- __m128i work;
- flat = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(q2p2, q0p0),
- _mm_subs_epu8(q0p0, q2p2)),
- _mm_or_si128(_mm_subs_epu8(q3p3, q0p0),
- _mm_subs_epu8(q0p0, q3p3)));
- flat = _mm_max_epu8(abs_p1p0, flat);
- flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
- flat = _mm_subs_epu8(flat, one);
- flat = _mm_cmpeq_epi8(flat, zero);
- flat = _mm_and_si128(flat, mask);
-
- q5p5 = _mm_loadl_epi64((__m128i *) (s - 6 * p));
- q5p5 = _mm_castps_si128(
- _mm_loadh_pi(_mm_castsi128_ps(q5p5),
- (__m64 *) (s + 5 * p)));
-
- q6p6 = _mm_loadl_epi64((__m128i *) (s - 7 * p));
- q6p6 = _mm_castps_si128(
- _mm_loadh_pi(_mm_castsi128_ps(q6p6),
- (__m64 *) (s + 6 * p)));
-
- flat2 = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(q4p4, q0p0),
- _mm_subs_epu8(q0p0, q4p4)),
- _mm_or_si128(_mm_subs_epu8(q5p5, q0p0),
- _mm_subs_epu8(q0p0, q5p5)));
-
- q7p7 = _mm_loadl_epi64((__m128i *) (s - 8 * p));
- q7p7 = _mm_castps_si128(
- _mm_loadh_pi(_mm_castsi128_ps(q7p7),
- (__m64 *) (s + 7 * p)));
-
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(q6p6, q0p0),
- _mm_subs_epu8(q0p0, q6p6)),
- _mm_or_si128(_mm_subs_epu8(q7p7, q0p0),
- _mm_subs_epu8(q0p0, q7p7)));
-
- flat2 = _mm_max_epu8(work, flat2);
- flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
- flat2 = _mm_subs_epu8(flat2, one);
- flat2 = _mm_cmpeq_epi8(flat2, zero);
- flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
- }
-
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- // flat and wide flat calculations
- {- const __m128i eight = _mm_set1_epi16(8);
- const __m128i four = _mm_set1_epi16(4);
- __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
- __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
- __m128i pixelFilter_p, pixelFilter_q;
- __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
- __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
-
- p7_16 = _mm_unpacklo_epi8(q7p7, zero);
- p6_16 = _mm_unpacklo_epi8(q6p6, zero);
- p5_16 = _mm_unpacklo_epi8(q5p5, zero);
- p4_16 = _mm_unpacklo_epi8(q4p4, zero);
- p3_16 = _mm_unpacklo_epi8(q3p3, zero);
- p2_16 = _mm_unpacklo_epi8(q2p2, zero);
- p1_16 = _mm_unpacklo_epi8(q1p1, zero);
- p0_16 = _mm_unpacklo_epi8(q0p0, zero);
- q0_16 = _mm_unpackhi_epi8(q0p0, zero);
- q1_16 = _mm_unpackhi_epi8(q1p1, zero);
- q2_16 = _mm_unpackhi_epi8(q2p2, zero);
- q3_16 = _mm_unpackhi_epi8(q3p3, zero);
- q4_16 = _mm_unpackhi_epi8(q4p4, zero);
- q5_16 = _mm_unpackhi_epi8(q5p5, zero);
- q6_16 = _mm_unpackhi_epi8(q6p6, zero);
- q7_16 = _mm_unpackhi_epi8(q7p7, zero);
-
- pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
- _mm_add_epi16(p4_16, p3_16));
- pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
- _mm_add_epi16(q4_16, q3_16));
-
- pixetFilter_p2p1p0 = _mm_add_epi16(p0_16,
- _mm_add_epi16(p2_16, p1_16));
- pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
-
- pixetFilter_q2q1q0 = _mm_add_epi16(q0_16,
- _mm_add_epi16(q2_16, q1_16));
- pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
- pixelFilter_p = _mm_add_epi16(eight,
- _mm_add_epi16(pixelFilter_p, pixelFilter_q));
- pixetFilter_p2p1p0 = _mm_add_epi16(four,
- _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)),
- 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)),
- 4);
- flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(p3_16, p0_16)), 3);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(q3_16, q0_16)), 3);
-
- flat_q0p0 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(p7_16, p7_16);
- sum_q7 = _mm_add_epi16(q7_16, q7_16);
- sum_p3 = _mm_add_epi16(p3_16, p3_16);
- sum_q3 = _mm_add_epi16(q3_16, q3_16);
-
- pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)),
- 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)),
- 4);
- flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
-
- pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
- pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(sum_p3, p1_16)), 3);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixetFilter_q2q1q0,
- _mm_add_epi16(sum_q3, q1_16)), 3);
- flat_q1p1 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- sum_p3 = _mm_add_epi16(sum_p3, p3_16);
- sum_q3 = _mm_add_epi16(sum_q3, q3_16);
-
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)),
- 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)),
- 4);
- flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
-
- pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
- pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
-
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(sum_p3, p2_16)), 3);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixetFilter_q2q1q0,
- _mm_add_epi16(sum_q3, q2_16)), 3);
- flat_q2p2 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)),
- 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)),
- 4);
- flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)),
- 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)),
- 4);
- flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)),
- 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)),
- 4);
- flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
- res_p = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)),
- 4);
- res_q = _mm_srli_epi16(
- _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)),
- 4);
- flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
- }
- // wide flat
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
- flat = _mm_shuffle_epi32(flat, 68);
- flat2 = _mm_shuffle_epi32(flat2, 68);
-
- q2p2 = _mm_andnot_si128(flat, q2p2);
- flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
- q2p2 = _mm_or_si128(q2p2, flat_q2p2);
-
- qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
- flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
- q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
-
- qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
- flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
- q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
-
- q6p6 = _mm_andnot_si128(flat2, q6p6);
- flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
- q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
- _mm_storel_epi64((__m128i *) (s - 7 * p), q6p6);
- _mm_storeh_pi((__m64 *) (s + 6 * p), _mm_castsi128_ps(q6p6));
-
- q5p5 = _mm_andnot_si128(flat2, q5p5);
- flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
- q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
- _mm_storel_epi64((__m128i *) (s - 6 * p), q5p5);
- _mm_storeh_pi((__m64 *) (s + 5 * p), _mm_castsi128_ps(q5p5));
-
- q4p4 = _mm_andnot_si128(flat2, q4p4);
- flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
- q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
- _mm_storel_epi64((__m128i *) (s - 5 * p), q4p4);
- _mm_storeh_pi((__m64 *) (s + 4 * p), _mm_castsi128_ps(q4p4));
-
- q3p3 = _mm_andnot_si128(flat2, q3p3);
- flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
- q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
- _mm_storel_epi64((__m128i *) (s - 4 * p), q3p3);
- _mm_storeh_pi((__m64 *) (s + 3 * p), _mm_castsi128_ps(q3p3));
-
- q2p2 = _mm_andnot_si128(flat2, q2p2);
- flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
- q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
- _mm_storel_epi64((__m128i *) (s - 3 * p), q2p2);
- _mm_storeh_pi((__m64 *) (s + 2 * p), _mm_castsi128_ps(q2p2));
-
- q1p1 = _mm_andnot_si128(flat2, q1p1);
- flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
- q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
- _mm_storel_epi64((__m128i *) (s - 2 * p), q1p1);
- _mm_storeh_pi((__m64 *) (s + 1 * p), _mm_castsi128_ps(q1p1));
-
- q0p0 = _mm_andnot_si128(flat2, q0p0);
- flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
- q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
- _mm_storel_epi64((__m128i *) (s - 1 * p), q0p0);
- _mm_storeh_pi((__m64 *) (s - 0 * p), _mm_castsi128_ps(q0p0));
- }
-}
-
-DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = {- 0, 128, 1, 128, 2, 128, 3, 128, 4, 128, 5, 128, 6, 128, 7, 128,
- 8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128
-};
-
-static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
- const unsigned char *_blimit, const unsigned char *_limit,
- const unsigned char *_thresh) {- __m128i mask, hev, flat, flat2;
- const __m128i zero = _mm_set1_epi16(0);
- const __m128i one = _mm_set1_epi8(1);
- __m128i p7, p6, p5;
- __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
- __m128i q5, q6, q7;
- __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4,
- q256_4, p256_3, q256_3, p256_2, q256_2, p256_1, q256_1,
- p256_0, q256_0;
-
- const __m128i thresh = _mm_broadcastb_epi8(
- _mm_cvtsi32_si128((int) _thresh[0]));
- const __m128i limit = _mm_broadcastb_epi8(
- _mm_cvtsi32_si128((int) _limit[0]));
- const __m128i blimit = _mm_broadcastb_epi8(
- _mm_cvtsi32_si128((int) _blimit[0]));
-
- p256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 5 * p)));
- p256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 4 * p)));
- p256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 3 * p)));
- p256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 2 * p)));
- p256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 1 * p)));
- q256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 0 * p)));
- q256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s + 1 * p)));
- q256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s + 2 * p)));
- q256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s + 3 * p)));
- q256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s + 4 * p)));
-
- p4 = _mm256_castsi256_si128(p256_4);
- p3 = _mm256_castsi256_si128(p256_3);
- p2 = _mm256_castsi256_si128(p256_2);
- p1 = _mm256_castsi256_si128(p256_1);
- p0 = _mm256_castsi256_si128(p256_0);
- q0 = _mm256_castsi256_si128(q256_0);
- q1 = _mm256_castsi256_si128(q256_1);
- q2 = _mm256_castsi256_si128(q256_2);
- q3 = _mm256_castsi256_si128(q256_3);
- q4 = _mm256_castsi256_si128(q256_4);
-
- {- const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
- _mm_subs_epu8(p0, p1));
- const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
- _mm_subs_epu8(q0, q1));
- const __m128i fe = _mm_set1_epi8(0xfe);
- const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
- __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
- _mm_subs_epu8(q0, p0));
- __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
- _mm_subs_epu8(q1, p1));
- __m128i work;
- flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu8(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
-
- abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
- mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(flat, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- // mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
- _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
- mask = _mm_max_epu8(work, mask);
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
- _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
- mask = _mm_max_epu8(work, mask);
- mask = _mm_subs_epu8(mask, limit);
- mask = _mm_cmpeq_epi8(mask, zero);
- }
-
- // lp filter
- {- const __m128i t4 = _mm_set1_epi8(4);
- const __m128i t3 = _mm_set1_epi8(3);
- const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i te0 = _mm_set1_epi8(0xe0);
- const __m128i t1f = _mm_set1_epi8(0x1f);
- const __m128i t1 = _mm_set1_epi8(0x1);
- const __m128i t7f = _mm_set1_epi8(0x7f);
-
- __m128i ps1 = _mm_xor_si128(p1, t80);
- __m128i ps0 = _mm_xor_si128(p0, t80);
- __m128i qs0 = _mm_xor_si128(q0, t80);
- __m128i qs1 = _mm_xor_si128(q1, t80);
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
- __m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1,
- flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4,
- flat2_q5, flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1,
- flat_q2;
-
- filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, ps0);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- /* (vp9_filter + 3 * (qs0 - ps0)) & mask */
- filt = _mm_and_si128(filt, mask);
-
- filter1 = _mm_adds_epi8(filt, t4);
- filter2 = _mm_adds_epi8(filt, t3);
-
- /* Filter1 >> 3 */
- work_a = _mm_cmpgt_epi8(zero, filter1);
- filter1 = _mm_srli_epi16(filter1, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter1 = _mm_and_si128(filter1, t1f);
- filter1 = _mm_or_si128(filter1, work_a);
- qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
-
- /* Filter2 >> 3 */
- work_a = _mm_cmpgt_epi8(zero, filter2);
- filter2 = _mm_srli_epi16(filter2, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter2 = _mm_and_si128(filter2, t1f);
- filter2 = _mm_or_si128(filter2, work_a);
- ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
-
- /* filt >> 1 */
- filt = _mm_adds_epi8(filter1, t1);
- work_a = _mm_cmpgt_epi8(zero, filt);
- filt = _mm_srli_epi16(filt, 1);
- work_a = _mm_and_si128(work_a, t80);
- filt = _mm_and_si128(filt, t7f);
- filt = _mm_or_si128(filt, work_a);
- filt = _mm_andnot_si128(hev, filt);
- ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
- qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
- // loopfilter done
-
- {- __m128i work;
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
- _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
- flat = _mm_max_epu8(work, flat);
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
- _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
- flat = _mm_max_epu8(work, flat);
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)),
- _mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4)));
- flat = _mm_subs_epu8(flat, one);
- flat = _mm_cmpeq_epi8(flat, zero);
- flat = _mm_and_si128(flat, mask);
-
- p256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 6 * p)));
- q256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s + 5 * p)));
- p5 = _mm256_castsi256_si128(p256_5);
- q5 = _mm256_castsi256_si128(q256_5);
- flat2 = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
- _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
-
- flat2 = _mm_max_epu8(work, flat2);
- p256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 7 * p)));
- q256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s + 6 * p)));
- p6 = _mm256_castsi256_si128(p256_6);
- q6 = _mm256_castsi256_si128(q256_6);
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
- _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
-
- flat2 = _mm_max_epu8(work, flat2);
-
- p256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s - 8 * p)));
- q256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
- (__m128d const *)(s + 7 * p)));
- p7 = _mm256_castsi256_si128(p256_7);
- q7 = _mm256_castsi256_si128(q256_7);
- work = _mm_max_epu8(
- _mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
- _mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
-
- flat2 = _mm_max_epu8(work, flat2);
- flat2 = _mm_subs_epu8(flat2, one);
- flat2 = _mm_cmpeq_epi8(flat2, zero);
- flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
- }
-
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- // flat and wide flat calculations
- {- const __m256i eight = _mm256_set1_epi16(8);
- const __m256i four = _mm256_set1_epi16(4);
- __m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
- pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p,
- res_q;
-
- const __m256i filter = _mm256_load_si256(
- (__m256i const *)filt_loopfilter_avx2);
- p256_7 = _mm256_shuffle_epi8(p256_7, filter);
- p256_6 = _mm256_shuffle_epi8(p256_6, filter);
- p256_5 = _mm256_shuffle_epi8(p256_5, filter);
- p256_4 = _mm256_shuffle_epi8(p256_4, filter);
- p256_3 = _mm256_shuffle_epi8(p256_3, filter);
- p256_2 = _mm256_shuffle_epi8(p256_2, filter);
- p256_1 = _mm256_shuffle_epi8(p256_1, filter);
- p256_0 = _mm256_shuffle_epi8(p256_0, filter);
- q256_0 = _mm256_shuffle_epi8(q256_0, filter);
- q256_1 = _mm256_shuffle_epi8(q256_1, filter);
- q256_2 = _mm256_shuffle_epi8(q256_2, filter);
- q256_3 = _mm256_shuffle_epi8(q256_3, filter);
- q256_4 = _mm256_shuffle_epi8(q256_4, filter);
- q256_5 = _mm256_shuffle_epi8(q256_5, filter);
- q256_6 = _mm256_shuffle_epi8(q256_6, filter);
- q256_7 = _mm256_shuffle_epi8(q256_7, filter);
-
- pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
- _mm256_add_epi16(p256_4, p256_3));
- pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5),
- _mm256_add_epi16(q256_4, q256_3));
-
- pixetFilter_p2p1p0 = _mm256_add_epi16(p256_0,
- _mm256_add_epi16(p256_2, p256_1));
- pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
-
- pixetFilter_q2q1q0 = _mm256_add_epi16(q256_0,
- _mm256_add_epi16(q256_2, q256_1));
- pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
-
- pixelFilter_p = _mm256_add_epi16(eight,
- _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
-
- pixetFilter_p2p1p0 = _mm256_add_epi16(four,
- _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_p,
- _mm256_add_epi16(p256_7, p256_0)), 4);
-
- flat2_p0 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_p,
- _mm256_add_epi16(q256_7, q256_0)), 4);
-
- flat2_q0 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixetFilter_p2p1p0,
- _mm256_add_epi16(p256_3, p256_0)), 3);
-
- flat_p0 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixetFilter_p2p1p0,
- _mm256_add_epi16(q256_3, q256_0)), 3);
-
- flat_q0 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- sum_p7 = _mm256_add_epi16(p256_7, p256_7);
-
- sum_q7 = _mm256_add_epi16(q256_7, q256_7);
-
- sum_p3 = _mm256_add_epi16(p256_3, p256_3);
-
- sum_q3 = _mm256_add_epi16(q256_3, q256_3);
-
- pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6);
-
- pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6);
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_p,
- _mm256_add_epi16(sum_p7, p256_1)), 4);
-
- flat2_p1 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_q,
- _mm256_add_epi16(sum_q7, q256_1)), 4);
-
- flat2_q1 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2);
-
- pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2);
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixetFilter_p2p1p0,
- _mm256_add_epi16(sum_p3, p256_1)), 3);
-
- flat_p1 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixetFilter_q2q1q0,
- _mm256_add_epi16(sum_q3, q256_1)), 3);
-
- flat_q1 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
-
- sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
-
- sum_p3 = _mm256_add_epi16(sum_p3, p256_3);
-
- sum_q3 = _mm256_add_epi16(sum_q3, q256_3);
-
- pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5);
-
- pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5);
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_p,
- _mm256_add_epi16(sum_p7, p256_2)), 4);
-
- flat2_p2 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_q,
- _mm256_add_epi16(sum_q7, q256_2)), 4);
-
- flat2_q2 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1);
-
- pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1);
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixetFilter_p2p1p0,
- _mm256_add_epi16(sum_p3, p256_2)), 3);
-
- flat_p2 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixetFilter_q2q1q0,
- _mm256_add_epi16(sum_q3, q256_2)), 3);
-
- flat_q2 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
-
- sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
-
- pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4);
-
- pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4);
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_p,
- _mm256_add_epi16(sum_p7, p256_3)), 4);
-
- flat2_p3 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_q,
- _mm256_add_epi16(sum_q7, q256_3)), 4);
-
- flat2_q3 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
-
- sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
-
- pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3);
-
- pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3);
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_p,
- _mm256_add_epi16(sum_p7, p256_4)), 4);
-
- flat2_p4 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_q,
- _mm256_add_epi16(sum_q7, q256_4)), 4);
-
- flat2_q4 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
-
- sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
-
- pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2);
-
- pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2);
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_p,
- _mm256_add_epi16(sum_p7, p256_5)), 4);
-
- flat2_p5 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_q,
- _mm256_add_epi16(sum_q7, q256_5)), 4);
-
- flat2_q5 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
-
- sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
-
- sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
-
- pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1);
-
- pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1);
-
- res_p = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_p,
- _mm256_add_epi16(sum_p7, p256_6)), 4);
-
- flat2_p6 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
- 168));
-
- res_q = _mm256_srli_epi16(
- _mm256_add_epi16(pixelFilter_q,
- _mm256_add_epi16(sum_q7, q256_6)), 4);
-
- flat2_q6 = _mm256_castsi256_si128(
- _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
- 168));
- }
-
- // wide flat
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
- p2 = _mm_andnot_si128(flat, p2);
- flat_p2 = _mm_and_si128(flat, flat_p2);
- p2 = _mm_or_si128(flat_p2, p2);
-
- p1 = _mm_andnot_si128(flat, ps1);
- flat_p1 = _mm_and_si128(flat, flat_p1);
- p1 = _mm_or_si128(flat_p1, p1);
-
- p0 = _mm_andnot_si128(flat, ps0);
- flat_p0 = _mm_and_si128(flat, flat_p0);
- p0 = _mm_or_si128(flat_p0, p0);
-
- q0 = _mm_andnot_si128(flat, qs0);
- flat_q0 = _mm_and_si128(flat, flat_q0);
- q0 = _mm_or_si128(flat_q0, q0);
-
- q1 = _mm_andnot_si128(flat, qs1);
- flat_q1 = _mm_and_si128(flat, flat_q1);
- q1 = _mm_or_si128(flat_q1, q1);
-
- q2 = _mm_andnot_si128(flat, q2);
- flat_q2 = _mm_and_si128(flat, flat_q2);
- q2 = _mm_or_si128(flat_q2, q2);
-
- p6 = _mm_andnot_si128(flat2, p6);
- flat2_p6 = _mm_and_si128(flat2, flat2_p6);
- p6 = _mm_or_si128(flat2_p6, p6);
- _mm_storeu_si128((__m128i *) (s - 7 * p), p6);
-
- p5 = _mm_andnot_si128(flat2, p5);
- flat2_p5 = _mm_and_si128(flat2, flat2_p5);
- p5 = _mm_or_si128(flat2_p5, p5);
- _mm_storeu_si128((__m128i *) (s - 6 * p), p5);
-
- p4 = _mm_andnot_si128(flat2, p4);
- flat2_p4 = _mm_and_si128(flat2, flat2_p4);
- p4 = _mm_or_si128(flat2_p4, p4);
- _mm_storeu_si128((__m128i *) (s - 5 * p), p4);
-
- p3 = _mm_andnot_si128(flat2, p3);
- flat2_p3 = _mm_and_si128(flat2, flat2_p3);
- p3 = _mm_or_si128(flat2_p3, p3);
- _mm_storeu_si128((__m128i *) (s - 4 * p), p3);
-
- p2 = _mm_andnot_si128(flat2, p2);
- flat2_p2 = _mm_and_si128(flat2, flat2_p2);
- p2 = _mm_or_si128(flat2_p2, p2);
- _mm_storeu_si128((__m128i *) (s - 3 * p), p2);
-
- p1 = _mm_andnot_si128(flat2, p1);
- flat2_p1 = _mm_and_si128(flat2, flat2_p1);
- p1 = _mm_or_si128(flat2_p1, p1);
- _mm_storeu_si128((__m128i *) (s - 2 * p), p1);
-
- p0 = _mm_andnot_si128(flat2, p0);
- flat2_p0 = _mm_and_si128(flat2, flat2_p0);
- p0 = _mm_or_si128(flat2_p0, p0);
- _mm_storeu_si128((__m128i *) (s - 1 * p), p0);
-
- q0 = _mm_andnot_si128(flat2, q0);
- flat2_q0 = _mm_and_si128(flat2, flat2_q0);
- q0 = _mm_or_si128(flat2_q0, q0);
- _mm_storeu_si128((__m128i *) (s - 0 * p), q0);
-
- q1 = _mm_andnot_si128(flat2, q1);
- flat2_q1 = _mm_and_si128(flat2, flat2_q1);
- q1 = _mm_or_si128(flat2_q1, q1);
- _mm_storeu_si128((__m128i *) (s + 1 * p), q1);
-
- q2 = _mm_andnot_si128(flat2, q2);
- flat2_q2 = _mm_and_si128(flat2, flat2_q2);
- q2 = _mm_or_si128(flat2_q2, q2);
- _mm_storeu_si128((__m128i *) (s + 2 * p), q2);
-
- q3 = _mm_andnot_si128(flat2, q3);
- flat2_q3 = _mm_and_si128(flat2, flat2_q3);
- q3 = _mm_or_si128(flat2_q3, q3);
- _mm_storeu_si128((__m128i *) (s + 3 * p), q3);
-
- q4 = _mm_andnot_si128(flat2, q4);
- flat2_q4 = _mm_and_si128(flat2, flat2_q4);
- q4 = _mm_or_si128(flat2_q4, q4);
- _mm_storeu_si128((__m128i *) (s + 4 * p), q4);
-
- q5 = _mm_andnot_si128(flat2, q5);
- flat2_q5 = _mm_and_si128(flat2, flat2_q5);
- q5 = _mm_or_si128(flat2_q5, q5);
- _mm_storeu_si128((__m128i *) (s + 5 * p), q5);
-
- q6 = _mm_andnot_si128(flat2, q6);
- flat2_q6 = _mm_and_si128(flat2, flat2_q6);
- q6 = _mm_or_si128(flat2_q6, q6);
- _mm_storeu_si128((__m128i *) (s + 6 * p), q6);
- }
-}
-
-void vp9_lpf_horizontal_16_avx2(unsigned char *s, int p,
- const unsigned char *_blimit, const unsigned char *_limit,
- const unsigned char *_thresh, int count) {- if (count == 1)
- mb_lpf_horizontal_edge_w_avx2_8(s, p, _blimit, _limit, _thresh);
- else
- mb_lpf_horizontal_edge_w_avx2_16(s, p, _blimit, _limit, _thresh);
-}
--- a/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
+++ /dev/null
@@ -1,1587 +1,0 @@
-/*
- * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
- *
- * Use of this source code is governed by a BSD-style license
- * that can be found in the LICENSE file in the root of the source
- * tree. An additional intellectual property rights grant can be found
- * in the file PATENTS. All contributing project authors may
- * be found in the AUTHORS file in the root of the source tree.
- */
-
-#include <emmintrin.h> // SSE2
-
-#include "./vp9_rtcd.h"
-#include "vp9/common/vp9_loopfilter.h"
-#include "vpx_ports/emmintrin_compat.h"
-
-static INLINE __m128i abs_diff(__m128i a, __m128i b) {- return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a));
-}
-
-static void mb_lpf_horizontal_edge_w_sse2_8(unsigned char *s,
- int p,
- const unsigned char *_blimit,
- const unsigned char *_limit,
- const unsigned char *_thresh) {- const __m128i zero = _mm_set1_epi16(0);
- const __m128i one = _mm_set1_epi8(1);
- const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
- const __m128i limit = _mm_load_si128((const __m128i *)_limit);
- const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
- __m128i mask, hev, flat, flat2;
- __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
- __m128i abs_p1p0;
-
- q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
- q4p4 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q4p4),
- (__m64 *)(s + 4 * p)));
- q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
- q3p3 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q3p3),
- (__m64 *)(s + 3 * p)));
- q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
- q2p2 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q2p2),
- (__m64 *)(s + 2 * p)));
- q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
- q1p1 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q1p1),
- (__m64 *)(s + 1 * p)));
- p1q1 = _mm_shuffle_epi32(q1p1, 78);
- q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
- q0p0 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q0p0),
- (__m64 *)(s - 0 * p)));
- p0q0 = _mm_shuffle_epi32(q0p0, 78);
-
- {- __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
- abs_p1p0 = abs_diff(q1p1, q0p0);
- abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
- fe = _mm_set1_epi8(0xfe);
- ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
- abs_p0q0 = abs_diff(q0p0, p0q0);
- abs_p1q1 = abs_diff(q1p1, p1q1);
- flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu8(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
-
- abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
- mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(abs_p1p0, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- // mask |= (abs(q1 - q0) > limit) * -1;
-
- work = _mm_max_epu8(abs_diff(q2p2, q1p1),
- abs_diff(q3p3, q2p2));
- mask = _mm_max_epu8(work, mask);
- mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
- mask = _mm_subs_epu8(mask, limit);
- mask = _mm_cmpeq_epi8(mask, zero);
- }
-
- // lp filter
- {- const __m128i t4 = _mm_set1_epi8(4);
- const __m128i t3 = _mm_set1_epi8(3);
- const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i t1 = _mm_set1_epi16(0x1);
- __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
- __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
- __m128i qs0 = _mm_xor_si128(p0q0, t80);
- __m128i qs1 = _mm_xor_si128(p1q1, t80);
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
- __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
- __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
-
- filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, qs0ps0);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- // (vp9_filter + 3 * (qs0 - ps0)) & mask
- filt = _mm_and_si128(filt, mask);
-
- filter1 = _mm_adds_epi8(filt, t4);
- filter2 = _mm_adds_epi8(filt, t3);
-
- filter1 = _mm_unpacklo_epi8(zero, filter1);
- filter1 = _mm_srai_epi16(filter1, 0xB);
- filter2 = _mm_unpacklo_epi8(zero, filter2);
- filter2 = _mm_srai_epi16(filter2, 0xB);
-
- // Filter1 >> 3
- filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
- qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
-
- // filt >> 1
- filt = _mm_adds_epi16(filter1, t1);
- filt = _mm_srai_epi16(filt, 1);
- filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
- filt);
- filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
- qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
- // loopfilter done
-
- {- __m128i work;
- flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
- flat = _mm_max_epu8(abs_p1p0, flat);
- flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
- flat = _mm_subs_epu8(flat, one);
- flat = _mm_cmpeq_epi8(flat, zero);
- flat = _mm_and_si128(flat, mask);
-
- q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
- q5p5 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q5p5),
- (__m64 *)(s + 5 * p)));
-
- q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
- q6p6 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q6p6),
- (__m64 *)(s + 6 * p)));
- flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0));
-
- q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
- q7p7 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q7p7),
- (__m64 *)(s + 7 * p)));
- work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0));
- flat2 = _mm_max_epu8(work, flat2);
- flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
- flat2 = _mm_subs_epu8(flat2, one);
- flat2 = _mm_cmpeq_epi8(flat2, zero);
- flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
- }
-
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- // flat and wide flat calculations
- {- const __m128i eight = _mm_set1_epi16(8);
- const __m128i four = _mm_set1_epi16(4);
- __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
- __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
- __m128i pixelFilter_p, pixelFilter_q;
- __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
- __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
-
- p7_16 = _mm_unpacklo_epi8(q7p7, zero);;
- p6_16 = _mm_unpacklo_epi8(q6p6, zero);
- p5_16 = _mm_unpacklo_epi8(q5p5, zero);
- p4_16 = _mm_unpacklo_epi8(q4p4, zero);
- p3_16 = _mm_unpacklo_epi8(q3p3, zero);
- p2_16 = _mm_unpacklo_epi8(q2p2, zero);
- p1_16 = _mm_unpacklo_epi8(q1p1, zero);
- p0_16 = _mm_unpacklo_epi8(q0p0, zero);
- q0_16 = _mm_unpackhi_epi8(q0p0, zero);
- q1_16 = _mm_unpackhi_epi8(q1p1, zero);
- q2_16 = _mm_unpackhi_epi8(q2p2, zero);
- q3_16 = _mm_unpackhi_epi8(q3p3, zero);
- q4_16 = _mm_unpackhi_epi8(q4p4, zero);
- q5_16 = _mm_unpackhi_epi8(q5p5, zero);
- q6_16 = _mm_unpackhi_epi8(q6p6, zero);
- q7_16 = _mm_unpackhi_epi8(q7p7, zero);
-
- pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
- _mm_add_epi16(p4_16, p3_16));
- pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
- _mm_add_epi16(q4_16, q3_16));
-
- pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
- pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
-
- pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
- pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
- pixelFilter_p = _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
- pixelFilter_q));
- pixetFilter_p2p1p0 = _mm_add_epi16(four,
- _mm_add_epi16(pixetFilter_p2p1p0,
- pixetFilter_q2q1q0));
- res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(p7_16, p0_16)), 4);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(q7_16, q0_16)), 4);
- flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
- res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(p3_16, p0_16)), 3);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(q3_16, q0_16)), 3);
-
- flat_q0p0 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(p7_16, p7_16);
- sum_q7 = _mm_add_epi16(q7_16, q7_16);
- sum_p3 = _mm_add_epi16(p3_16, p3_16);
- sum_q3 = _mm_add_epi16(q3_16, q3_16);
-
- pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
- res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p1_16)), 4);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q1_16)), 4);
- flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
-
- pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
- pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
- res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(sum_p3, p1_16)), 3);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
- _mm_add_epi16(sum_q3, q1_16)), 3);
- flat_q1p1 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- sum_p3 = _mm_add_epi16(sum_p3, p3_16);
- sum_q3 = _mm_add_epi16(sum_q3, q3_16);
-
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
- res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p2_16)), 4);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q2_16)), 4);
- flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
-
- pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
- pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
-
- res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
- _mm_add_epi16(sum_p3, p2_16)), 3);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
- _mm_add_epi16(sum_q3, q2_16)), 3);
- flat_q2p2 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
- res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p3_16)), 4);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q3_16)), 4);
- flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
- res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p4_16)), 4);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q4_16)), 4);
- flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
- res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p5_16)), 4);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q5_16)), 4);
- flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
-
- sum_p7 = _mm_add_epi16(sum_p7, p7_16);
- sum_q7 = _mm_add_epi16(sum_q7, q7_16);
- pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
- pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
- res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
- _mm_add_epi16(sum_p7, p6_16)), 4);
- res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
- _mm_add_epi16(sum_q7, q6_16)), 4);
- flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
- }
- // wide flat
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
- flat = _mm_shuffle_epi32(flat, 68);
- flat2 = _mm_shuffle_epi32(flat2, 68);
-
- q2p2 = _mm_andnot_si128(flat, q2p2);
- flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
- q2p2 = _mm_or_si128(q2p2, flat_q2p2);
-
- qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
- flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
- q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
-
- qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
- flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
- q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
-
- q6p6 = _mm_andnot_si128(flat2, q6p6);
- flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
- q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
- _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
- _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
-
- q5p5 = _mm_andnot_si128(flat2, q5p5);
- flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
- q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
- _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
- _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
-
- q4p4 = _mm_andnot_si128(flat2, q4p4);
- flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
- q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
- _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
- _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
-
- q3p3 = _mm_andnot_si128(flat2, q3p3);
- flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
- q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
- _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
- _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
-
- q2p2 = _mm_andnot_si128(flat2, q2p2);
- flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
- q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
- _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
- _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
-
- q1p1 = _mm_andnot_si128(flat2, q1p1);
- flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
- q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
- _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
- _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
-
- q0p0 = _mm_andnot_si128(flat2, q0p0);
- flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
- q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
- _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
- _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0));
- }
-}
-
-static INLINE __m128i filter_add2_sub2(const __m128i *const total,
- const __m128i *const a1,
- const __m128i *const a2,
- const __m128i *const s1,
- const __m128i *const s2) {- __m128i x = _mm_add_epi16(*a1, *total);
- x = _mm_add_epi16(_mm_sub_epi16(x, _mm_add_epi16(*s1, *s2)), *a2);
- return x;
-}
-
-static INLINE __m128i filter8_mask(const __m128i *const flat,
- const __m128i *const other_filt,
- const __m128i *const f8_lo,
- const __m128i *const f8_hi) {- const __m128i f8 = _mm_packus_epi16(_mm_srli_epi16(*f8_lo, 3),
- _mm_srli_epi16(*f8_hi, 3));
- const __m128i result = _mm_and_si128(*flat, f8);
- return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
-}
-
-static INLINE __m128i filter16_mask(const __m128i *const flat,
- const __m128i *const other_filt,
- const __m128i *const f_lo,
- const __m128i *const f_hi) {- const __m128i f = _mm_packus_epi16(_mm_srli_epi16(*f_lo, 4),
- _mm_srli_epi16(*f_hi, 4));
- const __m128i result = _mm_and_si128(*flat, f);
- return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
-}
-
-static void mb_lpf_horizontal_edge_w_sse2_16(unsigned char *s,
- int p,
- const unsigned char *_blimit,
- const unsigned char *_limit,
- const unsigned char *_thresh) {- const __m128i zero = _mm_set1_epi16(0);
- const __m128i one = _mm_set1_epi8(1);
- const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
- const __m128i limit = _mm_load_si128((const __m128i *)_limit);
- const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
- __m128i mask, hev, flat, flat2;
- __m128i p7, p6, p5;
- __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
- __m128i q5, q6, q7;
-
- __m128i op2, op1, op0, oq0, oq1, oq2;
-
- __m128i max_abs_p1p0q1q0;
-
- p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
- p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
- p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
- p4 = _mm_loadu_si128((__m128i *)(s - 5 * p));
- p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
- p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
- p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
- p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
- q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
- q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
- q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
- q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
- q4 = _mm_loadu_si128((__m128i *)(s + 4 * p));
- q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
- q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
- q7 = _mm_loadu_si128((__m128i *)(s + 7 * p));
-
- {- const __m128i abs_p1p0 = abs_diff(p1, p0);
- const __m128i abs_q1q0 = abs_diff(q1, q0);
- const __m128i fe = _mm_set1_epi8(0xfe);
- const __m128i ff = _mm_cmpeq_epi8(zero, zero);
- __m128i abs_p0q0 = abs_diff(p0, q0);
- __m128i abs_p1q1 = abs_diff(p1, q1);
- __m128i work;
- max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
-
- abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
- mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- // mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2));
- mask = _mm_max_epu8(work, mask);
- work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2));
- mask = _mm_max_epu8(work, mask);
- mask = _mm_subs_epu8(mask, limit);
- mask = _mm_cmpeq_epi8(mask, zero);
- }
-
- {- __m128i work;
- work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
- flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
- work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0));
- flat = _mm_max_epu8(work, flat);
- work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0));
- flat = _mm_subs_epu8(flat, one);
- flat = _mm_cmpeq_epi8(flat, zero);
- flat = _mm_and_si128(flat, mask);
- flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0));
- flat2 = _mm_max_epu8(work, flat2);
- work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0));
- flat2 = _mm_max_epu8(work, flat2);
- work = _mm_max_epu8(abs_diff(p7, p0), abs_diff(q7, q0));
- flat2 = _mm_max_epu8(work, flat2);
- flat2 = _mm_subs_epu8(flat2, one);
- flat2 = _mm_cmpeq_epi8(flat2, zero);
- flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
- }
-
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- // filter4
- {- const __m128i t4 = _mm_set1_epi8(4);
- const __m128i t3 = _mm_set1_epi8(3);
- const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i te0 = _mm_set1_epi8(0xe0);
- const __m128i t1f = _mm_set1_epi8(0x1f);
- const __m128i t1 = _mm_set1_epi8(0x1);
- const __m128i t7f = _mm_set1_epi8(0x7f);
- const __m128i ff = _mm_cmpeq_epi8(t4, t4);
-
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
-
- op1 = _mm_xor_si128(p1, t80);
- op0 = _mm_xor_si128(p0, t80);
- oq0 = _mm_xor_si128(q0, t80);
- oq1 = _mm_xor_si128(q1, t80);
-
- hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
- filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
-
- work_a = _mm_subs_epi8(oq0, op0);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- // (vp9_filter + 3 * (qs0 - ps0)) & mask
- filt = _mm_and_si128(filt, mask);
- filter1 = _mm_adds_epi8(filt, t4);
- filter2 = _mm_adds_epi8(filt, t3);
-
- // Filter1 >> 3
- work_a = _mm_cmpgt_epi8(zero, filter1);
- filter1 = _mm_srli_epi16(filter1, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter1 = _mm_and_si128(filter1, t1f);
- filter1 = _mm_or_si128(filter1, work_a);
- oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
-
- // Filter2 >> 3
- work_a = _mm_cmpgt_epi8(zero, filter2);
- filter2 = _mm_srli_epi16(filter2, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter2 = _mm_and_si128(filter2, t1f);
- filter2 = _mm_or_si128(filter2, work_a);
- op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
-
- // filt >> 1
- filt = _mm_adds_epi8(filter1, t1);
- work_a = _mm_cmpgt_epi8(zero, filt);
- filt = _mm_srli_epi16(filt, 1);
- work_a = _mm_and_si128(work_a, t80);
- filt = _mm_and_si128(filt, t7f);
- filt = _mm_or_si128(filt, work_a);
- filt = _mm_andnot_si128(hev, filt);
- op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
- oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
- // loopfilter done
-
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- // filter8
- {- const __m128i four = _mm_set1_epi16(4);
- const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
- const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
- const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
- const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
- const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
- const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
- const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
- const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
-
- const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
- const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
- const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
- const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
- const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
- const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
- const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
- const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
- __m128i f8_lo, f8_hi;
-
- f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four),
- _mm_add_epi16(p3_lo, p2_lo));
- f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo),
- _mm_add_epi16(p2_lo, p1_lo));
- f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
-
- f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four),
- _mm_add_epi16(p3_hi, p2_hi));
- f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi),
- _mm_add_epi16(p2_hi, p1_hi));
- f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
-
- op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi);
-
- f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo);
- f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi);
- op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi);
-
- f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo);
- f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi);
- op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi);
-
- f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo);
- f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi);
- oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi);
-
- f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo);
- f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi);
- oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi);
-
- f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo);
- f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi);
- oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi);
- }
-
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- // wide flat calculations
- {- const __m128i eight = _mm_set1_epi16(8);
- const __m128i p7_lo = _mm_unpacklo_epi8(p7, zero);
- const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero);
- const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero);
- const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero);
- const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
- const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
- const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
- const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
- const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
- const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
- const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
- const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
- const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero);
- const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero);
- const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero);
- const __m128i q7_lo = _mm_unpacklo_epi8(q7, zero);
-
- const __m128i p7_hi = _mm_unpackhi_epi8(p7, zero);
- const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero);
- const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero);
- const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero);
- const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
- const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
- const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
- const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
- const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
- const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
- const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
- const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
- const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero);
- const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero);
- const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero);
- const __m128i q7_hi = _mm_unpackhi_epi8(q7, zero);
-
- __m128i f_lo;
- __m128i f_hi;
-
- f_lo = _mm_sub_epi16(_mm_slli_epi16(p7_lo, 3), p7_lo); // p7 * 7
- f_lo = _mm_add_epi16(_mm_slli_epi16(p6_lo, 1),
- _mm_add_epi16(p4_lo, f_lo));
- f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo),
- _mm_add_epi16(p2_lo, p1_lo));
- f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo);
- f_lo = _mm_add_epi16(_mm_add_epi16(p5_lo, eight), f_lo);
-
- f_hi = _mm_sub_epi16(_mm_slli_epi16(p7_hi, 3), p7_hi); // p7 * 7
- f_hi = _mm_add_epi16(_mm_slli_epi16(p6_hi, 1),
- _mm_add_epi16(p4_hi, f_hi));
- f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi),
- _mm_add_epi16(p2_hi, p1_hi));
- f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi);
- f_hi = _mm_add_epi16(_mm_add_epi16(p5_hi, eight), f_hi);
-
- p6 = filter16_mask(&flat2, &p6, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s - 7 * p), p6);
-
- f_lo = filter_add2_sub2(&f_lo, &q1_lo, &p5_lo, &p6_lo, &p7_lo);
- f_hi = filter_add2_sub2(&f_hi, &q1_hi, &p5_hi, &p6_hi, &p7_hi);
- p5 = filter16_mask(&flat2, &p5, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
-
- f_lo = filter_add2_sub2(&f_lo, &q2_lo, &p4_lo, &p5_lo, &p7_lo);
- f_hi = filter_add2_sub2(&f_hi, &q2_hi, &p4_hi, &p5_hi, &p7_hi);
- p4 = filter16_mask(&flat2, &p4, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
-
- f_lo = filter_add2_sub2(&f_lo, &q3_lo, &p3_lo, &p4_lo, &p7_lo);
- f_hi = filter_add2_sub2(&f_hi, &q3_hi, &p3_hi, &p4_hi, &p7_hi);
- p3 = filter16_mask(&flat2, &p3, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
-
- f_lo = filter_add2_sub2(&f_lo, &q4_lo, &p2_lo, &p3_lo, &p7_lo);
- f_hi = filter_add2_sub2(&f_hi, &q4_hi, &p2_hi, &p3_hi, &p7_hi);
- op2 = filter16_mask(&flat2, &op2, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
-
- f_lo = filter_add2_sub2(&f_lo, &q5_lo, &p1_lo, &p2_lo, &p7_lo);
- f_hi = filter_add2_sub2(&f_hi, &q5_hi, &p1_hi, &p2_hi, &p7_hi);
- op1 = filter16_mask(&flat2, &op1, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
-
- f_lo = filter_add2_sub2(&f_lo, &q6_lo, &p0_lo, &p1_lo, &p7_lo);
- f_hi = filter_add2_sub2(&f_hi, &q6_hi, &p0_hi, &p1_hi, &p7_hi);
- op0 = filter16_mask(&flat2, &op0, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
-
- f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q0_lo, &p0_lo, &p7_lo);
- f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q0_hi, &p0_hi, &p7_hi);
- oq0 = filter16_mask(&flat2, &oq0, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
-
- f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q1_lo, &p6_lo, &q0_lo);
- f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q1_hi, &p6_hi, &q0_hi);
- oq1 = filter16_mask(&flat2, &oq1, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
-
- f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q2_lo, &p5_lo, &q1_lo);
- f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q2_hi, &p5_hi, &q1_hi);
- oq2 = filter16_mask(&flat2, &oq2, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
-
- f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q3_lo, &p4_lo, &q2_lo);
- f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q3_hi, &p4_hi, &q2_hi);
- q3 = filter16_mask(&flat2, &q3, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
-
- f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q4_lo, &p3_lo, &q3_lo);
- f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q4_hi, &p3_hi, &q3_hi);
- q4 = filter16_mask(&flat2, &q4, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
-
- f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q5_lo, &p2_lo, &q4_lo);
- f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q5_hi, &p2_hi, &q4_hi);
- q5 = filter16_mask(&flat2, &q5, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
-
- f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q6_lo, &p1_lo, &q5_lo);
- f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q6_hi, &p1_hi, &q5_hi);
- q6 = filter16_mask(&flat2, &q6, &f_lo, &f_hi);
- _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
- }
- // wide flat
- // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- }
-}
-
-// TODO(yunqingwang): remove count and call these 2 functions(8 or 16) directly.
-void vp9_lpf_horizontal_16_sse2(unsigned char *s, int p,
- const unsigned char *_blimit,
- const unsigned char *_limit,
- const unsigned char *_thresh, int count) {- if (count == 1)
- mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh);
- else
- mb_lpf_horizontal_edge_w_sse2_16(s, p, _blimit, _limit, _thresh);
-}
-
-void vp9_lpf_horizontal_8_sse2(unsigned char *s, int p,
- const unsigned char *_blimit,
- const unsigned char *_limit,
- const unsigned char *_thresh, int count) {- DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
- const __m128i zero = _mm_set1_epi16(0);
- const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
- const __m128i limit = _mm_load_si128((const __m128i *)_limit);
- const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
- __m128i mask, hev, flat;
- __m128i p3, p2, p1, p0, q0, q1, q2, q3;
- __m128i q3p3, q2p2, q1p1, q0p0, p1q1, p0q0;
-
- (void)count;
-
- q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)),
- _mm_loadl_epi64((__m128i *)(s + 3 * p)));
- q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
- _mm_loadl_epi64((__m128i *)(s + 2 * p)));
- q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
- _mm_loadl_epi64((__m128i *)(s + 1 * p)));
- q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
- _mm_loadl_epi64((__m128i *)(s - 0 * p)));
- p1q1 = _mm_shuffle_epi32(q1p1, 78);
- p0q0 = _mm_shuffle_epi32(q0p0, 78);
-
- {- // filter_mask and hev_mask
- const __m128i one = _mm_set1_epi8(1);
- const __m128i fe = _mm_set1_epi8(0xfe);
- const __m128i ff = _mm_cmpeq_epi8(fe, fe);
- __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
- abs_p1p0 = abs_diff(q1p1, q0p0);
- abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
-
- abs_p0q0 = abs_diff(q0p0, p0q0);
- abs_p1q1 = abs_diff(q1p1, p1q1);
- flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu8(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
-
- abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
- mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(abs_p1p0, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- // mask |= (abs(q1 - q0) > limit) * -1;
-
- work = _mm_max_epu8(abs_diff(q2p2, q1p1),
- abs_diff(q3p3, q2p2));
- mask = _mm_max_epu8(work, mask);
- mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
- mask = _mm_subs_epu8(mask, limit);
- mask = _mm_cmpeq_epi8(mask, zero);
-
- // flat_mask4
-
- flat = _mm_max_epu8(abs_diff(q2p2, q0p0),
- abs_diff(q3p3, q0p0));
- flat = _mm_max_epu8(abs_p1p0, flat);
- flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
- flat = _mm_subs_epu8(flat, one);
- flat = _mm_cmpeq_epi8(flat, zero);
- flat = _mm_and_si128(flat, mask);
- }
-
- {- const __m128i four = _mm_set1_epi16(4);
- unsigned char *src = s;
- {- __m128i workp_a, workp_b, workp_shft;
- p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
- p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
- p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
- p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
- q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
- q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
- q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
- q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
-
- workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
- workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
- workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op2[0],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op1[0],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op0[0],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq0[0],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq1[0],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq2[0],
- _mm_packus_epi16(workp_shft, workp_shft));
- }
- }
- // lp filter
- {- const __m128i t4 = _mm_set1_epi8(4);
- const __m128i t3 = _mm_set1_epi8(3);
- const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i t1 = _mm_set1_epi8(0x1);
- const __m128i ps1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
- t80);
- const __m128i ps0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
- t80);
- const __m128i qs0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 0 * p)),
- t80);
- const __m128i qs1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 1 * p)),
- t80);
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
-
- filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, ps0);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- // (vp9_filter + 3 * (qs0 - ps0)) & mask
- filt = _mm_and_si128(filt, mask);
-
- filter1 = _mm_adds_epi8(filt, t4);
- filter2 = _mm_adds_epi8(filt, t3);
-
- // Filter1 >> 3
- filter1 = _mm_unpacklo_epi8(zero, filter1);
- filter1 = _mm_srai_epi16(filter1, 11);
- filter1 = _mm_packs_epi16(filter1, filter1);
-
- // Filter2 >> 3
- filter2 = _mm_unpacklo_epi8(zero, filter2);
- filter2 = _mm_srai_epi16(filter2, 11);
- filter2 = _mm_packs_epi16(filter2, zero);
-
- // filt >> 1
- filt = _mm_adds_epi8(filter1, t1);
- filt = _mm_unpacklo_epi8(zero, filt);
- filt = _mm_srai_epi16(filt, 9);
- filt = _mm_packs_epi16(filt, zero);
-
- filt = _mm_andnot_si128(hev, filt);
-
- work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
- q0 = _mm_loadl_epi64((__m128i *)flat_oq0);
- work_a = _mm_andnot_si128(flat, work_a);
- q0 = _mm_and_si128(flat, q0);
- q0 = _mm_or_si128(work_a, q0);
-
- work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
- q1 = _mm_loadl_epi64((__m128i *)flat_oq1);
- work_a = _mm_andnot_si128(flat, work_a);
- q1 = _mm_and_si128(flat, q1);
- q1 = _mm_or_si128(work_a, q1);
-
- work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
- q2 = _mm_loadl_epi64((__m128i *)flat_oq2);
- work_a = _mm_andnot_si128(flat, work_a);
- q2 = _mm_and_si128(flat, q2);
- q2 = _mm_or_si128(work_a, q2);
-
- work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
- p0 = _mm_loadl_epi64((__m128i *)flat_op0);
- work_a = _mm_andnot_si128(flat, work_a);
- p0 = _mm_and_si128(flat, p0);
- p0 = _mm_or_si128(work_a, p0);
-
- work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
- p1 = _mm_loadl_epi64((__m128i *)flat_op1);
- work_a = _mm_andnot_si128(flat, work_a);
- p1 = _mm_and_si128(flat, p1);
- p1 = _mm_or_si128(work_a, p1);
-
- work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
- p2 = _mm_loadl_epi64((__m128i *)flat_op2);
- work_a = _mm_andnot_si128(flat, work_a);
- p2 = _mm_and_si128(flat, p2);
- p2 = _mm_or_si128(work_a, p2);
-
- _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
- _mm_storel_epi64((__m128i *)(s - 2 * p), p1);
- _mm_storel_epi64((__m128i *)(s - 1 * p), p0);
- _mm_storel_epi64((__m128i *)(s + 0 * p), q0);
- _mm_storel_epi64((__m128i *)(s + 1 * p), q1);
- _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
- }
-}
-
-void vp9_lpf_horizontal_8_dual_sse2(uint8_t *s, int p,
- const uint8_t *_blimit0,
- const uint8_t *_limit0,
- const uint8_t *_thresh0,
- const uint8_t *_blimit1,
- const uint8_t *_limit1,
- const uint8_t *_thresh1) {- DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
- DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
- const __m128i zero = _mm_set1_epi16(0);
- const __m128i blimit =
- _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
- _mm_load_si128((const __m128i *)_blimit1));
- const __m128i limit =
- _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
- _mm_load_si128((const __m128i *)_limit1));
- const __m128i thresh =
- _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
- _mm_load_si128((const __m128i *)_thresh1));
-
- __m128i mask, hev, flat;
- __m128i p3, p2, p1, p0, q0, q1, q2, q3;
-
- p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
- p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
- p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
- p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
- q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
- q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
- q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
- q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
- {- const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
- _mm_subs_epu8(p0, p1));
- const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
- _mm_subs_epu8(q0, q1));
- const __m128i one = _mm_set1_epi8(1);
- const __m128i fe = _mm_set1_epi8(0xfe);
- const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
- __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
- _mm_subs_epu8(q0, p0));
- __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
- _mm_subs_epu8(q1, p1));
- __m128i work;
-
- // filter_mask and hev_mask
- flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu8(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
-
- abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
- mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(flat, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- // mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
- _mm_subs_epu8(p1, p2)),
- _mm_or_si128(_mm_subs_epu8(p3, p2),
- _mm_subs_epu8(p2, p3)));
- mask = _mm_max_epu8(work, mask);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
- _mm_subs_epu8(q1, q2)),
- _mm_or_si128(_mm_subs_epu8(q3, q2),
- _mm_subs_epu8(q2, q3)));
- mask = _mm_max_epu8(work, mask);
- mask = _mm_subs_epu8(mask, limit);
- mask = _mm_cmpeq_epi8(mask, zero);
-
- // flat_mask4
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
- _mm_subs_epu8(p0, p2)),
- _mm_or_si128(_mm_subs_epu8(q2, q0),
- _mm_subs_epu8(q0, q2)));
- flat = _mm_max_epu8(work, flat);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
- _mm_subs_epu8(p0, p3)),
- _mm_or_si128(_mm_subs_epu8(q3, q0),
- _mm_subs_epu8(q0, q3)));
- flat = _mm_max_epu8(work, flat);
- flat = _mm_subs_epu8(flat, one);
- flat = _mm_cmpeq_epi8(flat, zero);
- flat = _mm_and_si128(flat, mask);
- }
- {- const __m128i four = _mm_set1_epi16(4);
- unsigned char *src = s;
- int i = 0;
-
- do {- __m128i workp_a, workp_b, workp_shft;
- p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
- p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
- p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
- p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
- q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
- q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
- q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
- q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
-
- workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
- workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
- workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op2[i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op1[i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_op0[i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq0[i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq1[i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
- workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
- _mm_storel_epi64((__m128i *)&flat_oq2[i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- src += 8;
- } while (++i < 2);
- }
- // lp filter
- {- const __m128i t4 = _mm_set1_epi8(4);
- const __m128i t3 = _mm_set1_epi8(3);
- const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i te0 = _mm_set1_epi8(0xe0);
- const __m128i t1f = _mm_set1_epi8(0x1f);
- const __m128i t1 = _mm_set1_epi8(0x1);
- const __m128i t7f = _mm_set1_epi8(0x7f);
-
- const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
- t80);
- const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
- t80);
- const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
- t80);
- const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
- t80);
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
-
- filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, ps0);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- // (vp9_filter + 3 * (qs0 - ps0)) & mask
- filt = _mm_and_si128(filt, mask);
-
- filter1 = _mm_adds_epi8(filt, t4);
- filter2 = _mm_adds_epi8(filt, t3);
-
- // Filter1 >> 3
- work_a = _mm_cmpgt_epi8(zero, filter1);
- filter1 = _mm_srli_epi16(filter1, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter1 = _mm_and_si128(filter1, t1f);
- filter1 = _mm_or_si128(filter1, work_a);
-
- // Filter2 >> 3
- work_a = _mm_cmpgt_epi8(zero, filter2);
- filter2 = _mm_srli_epi16(filter2, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter2 = _mm_and_si128(filter2, t1f);
- filter2 = _mm_or_si128(filter2, work_a);
-
- // filt >> 1
- filt = _mm_adds_epi8(filter1, t1);
- work_a = _mm_cmpgt_epi8(zero, filt);
- filt = _mm_srli_epi16(filt, 1);
- work_a = _mm_and_si128(work_a, t80);
- filt = _mm_and_si128(filt, t7f);
- filt = _mm_or_si128(filt, work_a);
-
- filt = _mm_andnot_si128(hev, filt);
-
- work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
- q0 = _mm_load_si128((__m128i *)flat_oq0);
- work_a = _mm_andnot_si128(flat, work_a);
- q0 = _mm_and_si128(flat, q0);
- q0 = _mm_or_si128(work_a, q0);
-
- work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
- q1 = _mm_load_si128((__m128i *)flat_oq1);
- work_a = _mm_andnot_si128(flat, work_a);
- q1 = _mm_and_si128(flat, q1);
- q1 = _mm_or_si128(work_a, q1);
-
- work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
- q2 = _mm_load_si128((__m128i *)flat_oq2);
- work_a = _mm_andnot_si128(flat, work_a);
- q2 = _mm_and_si128(flat, q2);
- q2 = _mm_or_si128(work_a, q2);
-
- work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
- p0 = _mm_load_si128((__m128i *)flat_op0);
- work_a = _mm_andnot_si128(flat, work_a);
- p0 = _mm_and_si128(flat, p0);
- p0 = _mm_or_si128(work_a, p0);
-
- work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
- p1 = _mm_load_si128((__m128i *)flat_op1);
- work_a = _mm_andnot_si128(flat, work_a);
- p1 = _mm_and_si128(flat, p1);
- p1 = _mm_or_si128(work_a, p1);
-
- work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
- p2 = _mm_load_si128((__m128i *)flat_op2);
- work_a = _mm_andnot_si128(flat, work_a);
- p2 = _mm_and_si128(flat, p2);
- p2 = _mm_or_si128(work_a, p2);
-
- _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
- _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
- _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
- _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
- _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
- _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
- }
-}
-
-void vp9_lpf_horizontal_4_dual_sse2(unsigned char *s, int p,
- const unsigned char *_blimit0,
- const unsigned char *_limit0,
- const unsigned char *_thresh0,
- const unsigned char *_blimit1,
- const unsigned char *_limit1,
- const unsigned char *_thresh1) {- const __m128i blimit =
- _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
- _mm_load_si128((const __m128i *)_blimit1));
- const __m128i limit =
- _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
- _mm_load_si128((const __m128i *)_limit1));
- const __m128i thresh =
- _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
- _mm_load_si128((const __m128i *)_thresh1));
- const __m128i zero = _mm_set1_epi16(0);
- __m128i p3, p2, p1, p0, q0, q1, q2, q3;
- __m128i mask, hev, flat;
-
- p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
- p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
- p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
- p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
- q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
- q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
- q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
- q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
-
- // filter_mask and hev_mask
- {- const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
- _mm_subs_epu8(p0, p1));
- const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
- _mm_subs_epu8(q0, q1));
- const __m128i fe = _mm_set1_epi8(0xfe);
- const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
- __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
- _mm_subs_epu8(q0, p0));
- __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
- _mm_subs_epu8(q1, p1));
- __m128i work;
-
- flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu8(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
-
- abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
- abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
- mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
- mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
- // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(flat, mask);
- // mask |= (abs(p1 - p0) > limit) * -1;
- // mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
- _mm_subs_epu8(p1, p2)),
- _mm_or_si128(_mm_subs_epu8(p3, p2),
- _mm_subs_epu8(p2, p3)));
- mask = _mm_max_epu8(work, mask);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
- _mm_subs_epu8(q1, q2)),
- _mm_or_si128(_mm_subs_epu8(q3, q2),
- _mm_subs_epu8(q2, q3)));
- mask = _mm_max_epu8(work, mask);
- mask = _mm_subs_epu8(mask, limit);
- mask = _mm_cmpeq_epi8(mask, zero);
- }
-
- // filter4
- {- const __m128i t4 = _mm_set1_epi8(4);
- const __m128i t3 = _mm_set1_epi8(3);
- const __m128i t80 = _mm_set1_epi8(0x80);
- const __m128i te0 = _mm_set1_epi8(0xe0);
- const __m128i t1f = _mm_set1_epi8(0x1f);
- const __m128i t1 = _mm_set1_epi8(0x1);
- const __m128i t7f = _mm_set1_epi8(0x7f);
-
- const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
- t80);
- const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
- t80);
- const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
- t80);
- const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
- t80);
- __m128i filt;
- __m128i work_a;
- __m128i filter1, filter2;
-
- filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, ps0);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- filt = _mm_adds_epi8(filt, work_a);
- // (vp9_filter + 3 * (qs0 - ps0)) & mask
- filt = _mm_and_si128(filt, mask);
-
- filter1 = _mm_adds_epi8(filt, t4);
- filter2 = _mm_adds_epi8(filt, t3);
-
- // Filter1 >> 3
- work_a = _mm_cmpgt_epi8(zero, filter1);
- filter1 = _mm_srli_epi16(filter1, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter1 = _mm_and_si128(filter1, t1f);
- filter1 = _mm_or_si128(filter1, work_a);
-
- // Filter2 >> 3
- work_a = _mm_cmpgt_epi8(zero, filter2);
- filter2 = _mm_srli_epi16(filter2, 3);
- work_a = _mm_and_si128(work_a, te0);
- filter2 = _mm_and_si128(filter2, t1f);
- filter2 = _mm_or_si128(filter2, work_a);
-
- // filt >> 1
- filt = _mm_adds_epi8(filter1, t1);
- work_a = _mm_cmpgt_epi8(zero, filt);
- filt = _mm_srli_epi16(filt, 1);
- work_a = _mm_and_si128(work_a, t80);
- filt = _mm_and_si128(filt, t7f);
- filt = _mm_or_si128(filt, work_a);
-
- filt = _mm_andnot_si128(hev, filt);
-
- q0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
- q1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
- p0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
- p1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
-
- _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
- _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
- _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
- _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
- }
-}
-
-static INLINE void transpose8x16(unsigned char *in0, unsigned char *in1,
- int in_p, unsigned char *out, int out_p) {- __m128i x0, x1, x2, x3, x4, x5, x6, x7;
- __m128i x8, x9, x10, x11, x12, x13, x14, x15;
-
- // 2-way interleave w/hoisting of unpacks
- x0 = _mm_loadl_epi64((__m128i *)in0); // 1
- x1 = _mm_loadl_epi64((__m128i *)(in0 + in_p)); // 3
- x0 = _mm_unpacklo_epi8(x0, x1); // 1
-
- x2 = _mm_loadl_epi64((__m128i *)(in0 + 2 * in_p)); // 5
- x3 = _mm_loadl_epi64((__m128i *)(in0 + 3*in_p)); // 7
- x1 = _mm_unpacklo_epi8(x2, x3); // 2
-
- x4 = _mm_loadl_epi64((__m128i *)(in0 + 4*in_p)); // 9
- x5 = _mm_loadl_epi64((__m128i *)(in0 + 5*in_p)); // 11
- x2 = _mm_unpacklo_epi8(x4, x5); // 3
-
- x6 = _mm_loadl_epi64((__m128i *)(in0 + 6*in_p)); // 13
- x7 = _mm_loadl_epi64((__m128i *)(in0 + 7*in_p)); // 15
- x3 = _mm_unpacklo_epi8(x6, x7); // 4
- x4 = _mm_unpacklo_epi16(x0, x1); // 9
-
- x8 = _mm_loadl_epi64((__m128i *)in1); // 2
- x9 = _mm_loadl_epi64((__m128i *)(in1 + in_p)); // 4
- x8 = _mm_unpacklo_epi8(x8, x9); // 5
- x5 = _mm_unpacklo_epi16(x2, x3); // 10
-
- x10 = _mm_loadl_epi64((__m128i *)(in1 + 2 * in_p)); // 6
- x11 = _mm_loadl_epi64((__m128i *)(in1 + 3*in_p)); // 8
- x9 = _mm_unpacklo_epi8(x10, x11); // 6
-
- x12 = _mm_loadl_epi64((__m128i *)(in1 + 4*in_p)); // 10
- x13 = _mm_loadl_epi64((__m128i *)(in1 + 5*in_p)); // 12
- x10 = _mm_unpacklo_epi8(x12, x13); // 7
- x12 = _mm_unpacklo_epi16(x8, x9); // 11
-
- x14 = _mm_loadl_epi64((__m128i *)(in1 + 6*in_p)); // 14
- x15 = _mm_loadl_epi64((__m128i *)(in1 + 7*in_p)); // 16
- x11 = _mm_unpacklo_epi8(x14, x15); // 8
- x13 = _mm_unpacklo_epi16(x10, x11); // 12
-
- x6 = _mm_unpacklo_epi32(x4, x5); // 13
- x7 = _mm_unpackhi_epi32(x4, x5); // 14
- x14 = _mm_unpacklo_epi32(x12, x13); // 15
- x15 = _mm_unpackhi_epi32(x12, x13); // 16
-
- // Store first 4-line result
- _mm_storeu_si128((__m128i *)out, _mm_unpacklo_epi64(x6, x14));
- _mm_storeu_si128((__m128i *)(out + out_p), _mm_unpackhi_epi64(x6, x14));
- _mm_storeu_si128((__m128i *)(out + 2 * out_p), _mm_unpacklo_epi64(x7, x15));
- _mm_storeu_si128((__m128i *)(out + 3 * out_p), _mm_unpackhi_epi64(x7, x15));
-
- x4 = _mm_unpackhi_epi16(x0, x1);
- x5 = _mm_unpackhi_epi16(x2, x3);
- x12 = _mm_unpackhi_epi16(x8, x9);
- x13 = _mm_unpackhi_epi16(x10, x11);
-
- x6 = _mm_unpacklo_epi32(x4, x5);
- x7 = _mm_unpackhi_epi32(x4, x5);
- x14 = _mm_unpacklo_epi32(x12, x13);
- x15 = _mm_unpackhi_epi32(x12, x13);
-
- // Store second 4-line result
- _mm_storeu_si128((__m128i *)(out + 4 * out_p), _mm_unpacklo_epi64(x6, x14));
- _mm_storeu_si128((__m128i *)(out + 5 * out_p), _mm_unpackhi_epi64(x6, x14));
- _mm_storeu_si128((__m128i *)(out + 6 * out_p), _mm_unpacklo_epi64(x7, x15));
- _mm_storeu_si128((__m128i *)(out + 7 * out_p), _mm_unpackhi_epi64(x7, x15));
-}
-
-static INLINE void transpose(unsigned char *src[], int in_p,
- unsigned char *dst[], int out_p,
- int num_8x8_to_transpose) {- int idx8x8 = 0;
- __m128i x0, x1, x2, x3, x4, x5, x6, x7;
- do {- unsigned char *in = src[idx8x8];
- unsigned char *out = dst[idx8x8];
-
- x0 = _mm_loadl_epi64((__m128i *)(in + 0*in_p)); // 00 01 02 03 04 05 06 07
- x1 = _mm_loadl_epi64((__m128i *)(in + 1*in_p)); // 10 11 12 13 14 15 16 17
- // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
- x0 = _mm_unpacklo_epi8(x0, x1);
-
- x2 = _mm_loadl_epi64((__m128i *)(in + 2*in_p)); // 20 21 22 23 24 25 26 27
- x3 = _mm_loadl_epi64((__m128i *)(in + 3*in_p)); // 30 31 32 33 34 35 36 37
- // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
- x1 = _mm_unpacklo_epi8(x2, x3);
-
- x4 = _mm_loadl_epi64((__m128i *)(in + 4*in_p)); // 40 41 42 43 44 45 46 47
- x5 = _mm_loadl_epi64((__m128i *)(in + 5*in_p)); // 50 51 52 53 54 55 56 57
- // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
- x2 = _mm_unpacklo_epi8(x4, x5);
-
- x6 = _mm_loadl_epi64((__m128i *)(in + 6*in_p)); // 60 61 62 63 64 65 66 67
- x7 = _mm_loadl_epi64((__m128i *)(in + 7*in_p)); // 70 71 72 73 74 75 76 77
- // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
- x3 = _mm_unpacklo_epi8(x6, x7);
-
- // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
- x4 = _mm_unpacklo_epi16(x0, x1);
- // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
- x5 = _mm_unpacklo_epi16(x2, x3);
- // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
- x6 = _mm_unpacklo_epi32(x4, x5);
- _mm_storel_pd((double *)(out + 0*out_p),
- _mm_castsi128_pd(x6)); // 00 10 20 30 40 50 60 70
- _mm_storeh_pd((double *)(out + 1*out_p),
- _mm_castsi128_pd(x6)); // 01 11 21 31 41 51 61 71
- // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
- x7 = _mm_unpackhi_epi32(x4, x5);
- _mm_storel_pd((double *)(out + 2*out_p),
- _mm_castsi128_pd(x7)); // 02 12 22 32 42 52 62 72
- _mm_storeh_pd((double *)(out + 3*out_p),
- _mm_castsi128_pd(x7)); // 03 13 23 33 43 53 63 73
-
- // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
- x4 = _mm_unpackhi_epi16(x0, x1);
- // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77
- x5 = _mm_unpackhi_epi16(x2, x3);
- // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75
- x6 = _mm_unpacklo_epi32(x4, x5);
- _mm_storel_pd((double *)(out + 4*out_p),
- _mm_castsi128_pd(x6)); // 04 14 24 34 44 54 64 74
- _mm_storeh_pd((double *)(out + 5*out_p),
- _mm_castsi128_pd(x6)); // 05 15 25 35 45 55 65 75
- // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77
- x7 = _mm_unpackhi_epi32(x4, x5);
-
- _mm_storel_pd((double *)(out + 6*out_p),
- _mm_castsi128_pd(x7)); // 06 16 26 36 46 56 66 76
- _mm_storeh_pd((double *)(out + 7*out_p),
- _mm_castsi128_pd(x7)); // 07 17 27 37 47 57 67 77
- } while (++idx8x8 < num_8x8_to_transpose);
-}
-
-void vp9_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1) {- DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
- unsigned char *src[2];
- unsigned char *dst[2];
-
- // Transpose 8x16
- transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
-
- // Loop filtering
- vp9_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
- blimit1, limit1, thresh1);
- src[0] = t_dst;
- src[1] = t_dst + 8;
- dst[0] = s - 4;
- dst[1] = s - 4 + p * 8;
-
- // Transpose back
- transpose(src, 16, dst, p, 2);
-}
-
-void vp9_lpf_vertical_8_sse2(unsigned char *s, int p,
- const unsigned char *blimit,
- const unsigned char *limit,
- const unsigned char *thresh, int count) {- DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 8]);
- unsigned char *src[1];
- unsigned char *dst[1];
- (void)count;
-
- // Transpose 8x8
- src[0] = s - 4;
- dst[0] = t_dst;
-
- transpose(src, p, dst, 8, 1);
-
- // Loop filtering
- vp9_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1);
-
- src[0] = t_dst;
- dst[0] = s - 4;
-
- // Transpose back
- transpose(src, 8, dst, p, 1);
-}
-
-void vp9_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
- const uint8_t *limit0,
- const uint8_t *thresh0,
- const uint8_t *blimit1,
- const uint8_t *limit1,
- const uint8_t *thresh1) {- DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
- unsigned char *src[2];
- unsigned char *dst[2];
-
- // Transpose 8x16
- transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
-
- // Loop filtering
- vp9_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
- blimit1, limit1, thresh1);
- src[0] = t_dst;
- src[1] = t_dst + 8;
-
- dst[0] = s - 4;
- dst[1] = s - 4 + p * 8;
-
- // Transpose back
- transpose(src, 16, dst, p, 2);
-}
-
-void vp9_lpf_vertical_16_sse2(unsigned char *s, int p,
- const unsigned char *blimit,
- const unsigned char *limit,
- const unsigned char *thresh) {- DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 16]);
- unsigned char *src[2];
- unsigned char *dst[2];
-
- src[0] = s - 8;
- src[1] = s;
- dst[0] = t_dst;
- dst[1] = t_dst + 8 * 8;
-
- // Transpose 16x8
- transpose(src, p, dst, 8, 2);
-
- // Loop filtering
- mb_lpf_horizontal_edge_w_sse2_8(t_dst + 8 * 8, 8, blimit, limit, thresh);
-
- src[0] = t_dst;
- src[1] = t_dst + 8 * 8;
- dst[0] = s - 8;
- dst[1] = s;
-
- // Transpose back
- transpose(src, 8, dst, p, 2);
-}
-
-void vp9_lpf_vertical_16_dual_sse2(unsigned char *s, int p,
- const uint8_t *blimit, const uint8_t *limit,
- const uint8_t *thresh) {- DECLARE_ALIGNED(16, unsigned char, t_dst[256]);
-
- // Transpose 16x16
- transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
- transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
-
- // Loop filtering
- mb_lpf_horizontal_edge_w_sse2_16(t_dst + 8 * 16, 16, blimit, limit,
- thresh);
-
- // Transpose back
- transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
- transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
-}
--- a/vp9/common/x86/vp9_loopfilter_mmx.asm
+++ /dev/null
@@ -1,611 +1,0 @@
-;
-; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
-;
-; Use of this source code is governed by a BSD-style license
-; that can be found in the LICENSE file in the root of the source
-; tree. An additional intellectual property rights grant can be found
-; in the file PATENTS. All contributing project authors may
-; be found in the AUTHORS file in the root of the source tree.
-;
-
-
-%include "vpx_ports/x86_abi_support.asm"
-
-
-;void vp9_lpf_horizontal_4_mmx
-;(
-; unsigned char *src_ptr,
-; int src_pixel_step,
-; const char *blimit,
-; const char *limit,
-; const char *thresh,
-; int count
-;)
-global sym(vp9_lpf_horizontal_4_mmx) PRIVATE
-sym(vp9_lpf_horizontal_4_mmx):
- push rbp
- mov rbp, rsp
- SHADOW_ARGS_TO_STACK 6
- GET_GOT rbx
- push rsi
- push rdi
- ; end prolog
-
- ALIGN_STACK 16, rax
- sub rsp, 32 ; reserve 32 bytes
- %define t0 [rsp + 0] ;__declspec(align(16)) char t0[8];
- %define t1 [rsp + 16] ;__declspec(align(16)) char t1[8];
-
- mov rsi, arg(0) ;src_ptr
- movsxd rax, dword ptr arg(1) ;src_pixel_step ; destination pitch?
-
- movsxd rcx, dword ptr arg(5) ;count
-.next8_h:
- mov rdx, arg(3) ;limit
- movq mm7, [rdx]
- mov rdi, rsi ; rdi points to row +1 for indirect addressing
- add rdi, rax
-
- ; calculate breakout conditions
- movq mm2, [rdi+2*rax] ; q3
- movq mm1, [rsi+2*rax] ; q2
- movq mm6, mm1 ; q2
- psubusb mm1, mm2 ; q2-=q3
- psubusb mm2, mm6 ; q3-=q2
- por mm1, mm2 ; abs(q3-q2)
- psubusb mm1, mm7 ;
-
-
- movq mm4, [rsi+rax] ; q1
- movq mm3, mm4 ; q1
- psubusb mm4, mm6 ; q1-=q2
- psubusb mm6, mm3 ; q2-=q1
- por mm4, mm6 ; abs(q2-q1)
-
- psubusb mm4, mm7
- por mm1, mm4
-
- movq mm4, [rsi] ; q0
- movq mm0, mm4 ; q0
- psubusb mm4, mm3 ; q0-=q1
- psubusb mm3, mm0 ; q1-=q0
- por mm4, mm3 ; abs(q0-q1)
- movq t0, mm4 ; save to t0
- psubusb mm4, mm7
- por mm1, mm4
-
-
- neg rax ; negate pitch to deal with above border
-
- movq mm2, [rsi+4*rax] ; p3
- movq mm4, [rdi+4*rax] ; p2
- movq mm5, mm4 ; p2
- psubusb mm4, mm2 ; p2-=p3
- psubusb mm2, mm5 ; p3-=p2
- por mm4, mm2 ; abs(p3 - p2)
- psubusb mm4, mm7
- por mm1, mm4
-
-
- movq mm4, [rsi+2*rax] ; p1
- movq mm3, mm4 ; p1
- psubusb mm4, mm5 ; p1-=p2
- psubusb mm5, mm3 ; p2-=p1
- por mm4, mm5 ; abs(p2 - p1)
- psubusb mm4, mm7
- por mm1, mm4
-
- movq mm2, mm3 ; p1
-
- movq mm4, [rsi+rax] ; p0
- movq mm5, mm4 ; p0
- psubusb mm4, mm3 ; p0-=p1
- psubusb mm3, mm5 ; p1-=p0
- por mm4, mm3 ; abs(p1 - p0)
- movq t1, mm4 ; save to t1
- psubusb mm4, mm7
- por mm1, mm4
-
- movq mm3, [rdi] ; q1
- movq mm4, mm3 ; q1
- psubusb mm3, mm2 ; q1-=p1
- psubusb mm2, mm4 ; p1-=q1
- por mm2, mm3 ; abs(p1-q1)
- pand mm2, [GLOBAL(tfe)] ; set lsb of each byte to zero
- psrlw mm2, 1 ; abs(p1-q1)/2
-
- movq mm6, mm5 ; p0
- movq mm3, [rsi] ; q0
- psubusb mm5, mm3 ; p0-=q0
- psubusb mm3, mm6 ; q0-=p0
- por mm5, mm3 ; abs(p0 - q0)
- paddusb mm5, mm5 ; abs(p0-q0)*2
- paddusb mm5, mm2 ; abs (p0 - q0) *2 + abs(p1-q1)/2
-
- mov rdx, arg(2) ;blimit ; get blimit
- movq mm7, [rdx] ; blimit
-
- psubusb mm5, mm7 ; abs (p0 - q0) *2 + abs(p1-q1)/2 > blimit
- por mm1, mm5
- pxor mm5, mm5
- pcmpeqb mm1, mm5 ; mask mm1
-
- ; calculate high edge variance
- mov rdx, arg(4) ;thresh ; get thresh
- movq mm7, [rdx] ;
- movq mm4, t0 ; get abs (q1 - q0)
- psubusb mm4, mm7
- movq mm3, t1 ; get abs (p1 - p0)
- psubusb mm3, mm7
- paddb mm4, mm3 ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
-
- pcmpeqb mm4, mm5
-
- pcmpeqb mm5, mm5
- pxor mm4, mm5
-
-
- ; start work on filters
- movq mm2, [rsi+2*rax] ; p1
- movq mm7, [rdi] ; q1
- pxor mm2, [GLOBAL(t80)] ; p1 offset to convert to signed values
- pxor mm7, [GLOBAL(t80)] ; q1 offset to convert to signed values
- psubsb mm2, mm7 ; p1 - q1
- pand mm2, mm4 ; high var mask (hvm)(p1 - q1)
- pxor mm6, [GLOBAL(t80)] ; offset to convert to signed values
- pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
- movq mm3, mm0 ; q0
- psubsb mm0, mm6 ; q0 - p0
- paddsb mm2, mm0 ; 1 * (q0 - p0) + hvm(p1 - q1)
- paddsb mm2, mm0 ; 2 * (q0 - p0) + hvm(p1 - q1)
- paddsb mm2, mm0 ; 3 * (q0 - p0) + hvm(p1 - q1)
- pand mm1, mm2 ; mask filter values we don't care about
- movq mm2, mm1
- paddsb mm1, [GLOBAL(t4)] ; 3* (q0 - p0) + hvm(p1 - q1) + 4
- paddsb mm2, [GLOBAL(t3)] ; 3* (q0 - p0) + hvm(p1 - q1) + 3
-
- pxor mm0, mm0 ;
- pxor mm5, mm5
- punpcklbw mm0, mm2 ;
- punpckhbw mm5, mm2 ;
- psraw mm0, 11 ;
- psraw mm5, 11
- packsswb mm0, mm5
- movq mm2, mm0 ; (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
-
- pxor mm0, mm0 ; 0
- movq mm5, mm1 ; abcdefgh
- punpcklbw mm0, mm1 ; e0f0g0h0
- psraw mm0, 11 ; sign extended shift right by 3
- pxor mm1, mm1 ; 0
- punpckhbw mm1, mm5 ; a0b0c0d0
- psraw mm1, 11 ; sign extended shift right by 3
- movq mm5, mm0 ; save results
-
- packsswb mm0, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
- paddsw mm5, [GLOBAL(ones)]
- paddsw mm1, [GLOBAL(ones)]
- psraw mm5, 1 ; partial shifted one more time for 2nd tap
- psraw mm1, 1 ; partial shifted one more time for 2nd tap
- packsswb mm5, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
- pandn mm4, mm5 ; high edge variance additive
-
- paddsb mm6, mm2 ; p0+= p0 add
- pxor mm6, [GLOBAL(t80)] ; unoffset
- movq [rsi+rax], mm6 ; write back
-
- movq mm6, [rsi+2*rax] ; p1
- pxor mm6, [GLOBAL(t80)] ; reoffset
- paddsb mm6, mm4 ; p1+= p1 add
- pxor mm6, [GLOBAL(t80)] ; unoffset
- movq [rsi+2*rax], mm6 ; write back
-
- psubsb mm3, mm0 ; q0-= q0 add
- pxor mm3, [GLOBAL(t80)] ; unoffset
- movq [rsi], mm3 ; write back
-
- psubsb mm7, mm4 ; q1-= q1 add
- pxor mm7, [GLOBAL(t80)] ; unoffset
- movq [rdi], mm7 ; write back
-
- add rsi,8
- neg rax
- dec rcx
- jnz .next8_h
-
- add rsp, 32
- pop rsp
- ; begin epilog
- pop rdi
- pop rsi
- RESTORE_GOT
- UNSHADOW_ARGS
- pop rbp
- ret
-
-
-;void vp9_lpf_vertical_4_mmx
-;(
-; unsigned char *src_ptr,
-; int src_pixel_step,
-; const char *blimit,
-; const char *limit,
-; const char *thresh,
-; int count
-;)
-global sym(vp9_lpf_vertical_4_mmx) PRIVATE
-sym(vp9_lpf_vertical_4_mmx):
- push rbp
- mov rbp, rsp
- SHADOW_ARGS_TO_STACK 6
- GET_GOT rbx
- push rsi
- push rdi
- ; end prolog
-
- ALIGN_STACK 16, rax
- sub rsp, 64 ; reserve 64 bytes
- %define t0 [rsp + 0] ;__declspec(align(16)) char t0[8];
- %define t1 [rsp + 16] ;__declspec(align(16)) char t1[8];
- %define srct [rsp + 32] ;__declspec(align(16)) char srct[32];
-
- mov rsi, arg(0) ;src_ptr
- movsxd rax, dword ptr arg(1) ;src_pixel_step ; destination pitch?
-
- lea rsi, [rsi + rax*4 - 4]
-
- movsxd rcx, dword ptr arg(5) ;count
-.next8_v:
- mov rdi, rsi ; rdi points to row +1 for indirect addressing
- add rdi, rax
-
-
- ;transpose
- movq mm6, [rsi+2*rax] ; 67 66 65 64 63 62 61 60
- movq mm7, mm6 ; 77 76 75 74 73 72 71 70
-
- punpckhbw mm7, [rdi+2*rax] ; 77 67 76 66 75 65 74 64
- punpcklbw mm6, [rdi+2*rax] ; 73 63 72 62 71 61 70 60
-
- movq mm4, [rsi] ; 47 46 45 44 43 42 41 40
- movq mm5, mm4 ; 47 46 45 44 43 42 41 40
-
- punpckhbw mm5, [rsi+rax] ; 57 47 56 46 55 45 54 44
- punpcklbw mm4, [rsi+rax] ; 53 43 52 42 51 41 50 40
-
- movq mm3, mm5 ; 57 47 56 46 55 45 54 44
- punpckhwd mm5, mm7 ; 77 67 57 47 76 66 56 46
-
- punpcklwd mm3, mm7 ; 75 65 55 45 74 64 54 44
- movq mm2, mm4 ; 53 43 52 42 51 41 50 40
-
- punpckhwd mm4, mm6 ; 73 63 53 43 72 62 52 42
- punpcklwd mm2, mm6 ; 71 61 51 41 70 60 50 40
-
- neg rax
- movq mm6, [rsi+rax*2] ; 27 26 25 24 23 22 21 20
-
- movq mm1, mm6 ; 27 26 25 24 23 22 21 20
- punpckhbw mm6, [rsi+rax] ; 37 27 36 36 35 25 34 24
-
- punpcklbw mm1, [rsi+rax] ; 33 23 32 22 31 21 30 20
- movq mm7, [rsi+rax*4]; ; 07 06 05 04 03 02 01 00
-
- punpckhbw mm7, [rdi+rax*4] ; 17 07 16 06 15 05 14 04
- movq mm0, mm7 ; 17 07 16 06 15 05 14 04
-
- punpckhwd mm7, mm6 ; 37 27 17 07 36 26 16 06
- punpcklwd mm0, mm6 ; 35 25 15 05 34 24 14 04
-
- movq mm6, mm7 ; 37 27 17 07 36 26 16 06
- punpckhdq mm7, mm5 ; 77 67 57 47 37 27 17 07 = q3
-
- punpckldq mm6, mm5 ; 76 66 56 46 36 26 16 06 = q2
-
- movq mm5, mm6 ; 76 66 56 46 36 26 16 06
- psubusb mm5, mm7 ; q2-q3
-
- psubusb mm7, mm6 ; q3-q2
- por mm7, mm5; ; mm7=abs (q3-q2)
-
- movq mm5, mm0 ; 35 25 15 05 34 24 14 04
- punpckhdq mm5, mm3 ; 75 65 55 45 35 25 15 05 = q1
-
- punpckldq mm0, mm3 ; 74 64 54 44 34 24 15 04 = q0
- movq mm3, mm5 ; 75 65 55 45 35 25 15 05 = q1
-
- psubusb mm3, mm6 ; q1-q2
- psubusb mm6, mm5 ; q2-q1
-
- por mm6, mm3 ; mm6=abs(q2-q1)
- lea rdx, srct
-
- movq [rdx+24], mm5 ; save q1
- movq [rdx+16], mm0 ; save q0
-
- movq mm3, [rsi+rax*4] ; 07 06 05 04 03 02 01 00
- punpcklbw mm3, [rdi+rax*4] ; 13 03 12 02 11 01 10 00
-
- movq mm0, mm3 ; 13 03 12 02 11 01 10 00
- punpcklwd mm0, mm1 ; 31 21 11 01 30 20 10 00
-
- punpckhwd mm3, mm1 ; 33 23 13 03 32 22 12 02
- movq mm1, mm0 ; 31 21 11 01 30 20 10 00
-
- punpckldq mm0, mm2 ; 70 60 50 40 30 20 10 00 =p3
- punpckhdq mm1, mm2 ; 71 61 51 41 31 21 11 01 =p2
-
- movq mm2, mm1 ; 71 61 51 41 31 21 11 01 =p2
- psubusb mm2, mm0 ; p2-p3
-
- psubusb mm0, mm1 ; p3-p2
- por mm0, mm2 ; mm0=abs(p3-p2)
-
- movq mm2, mm3 ; 33 23 13 03 32 22 12 02
- punpckldq mm2, mm4 ; 72 62 52 42 32 22 12 02 = p1
-
- punpckhdq mm3, mm4 ; 73 63 53 43 33 23 13 03 = p0
- movq [rdx+8], mm3 ; save p0
-
- movq [rdx], mm2 ; save p1
- movq mm5, mm2 ; mm5 = p1
-
- psubusb mm2, mm1 ; p1-p2
- psubusb mm1, mm5 ; p2-p1
-
- por mm1, mm2 ; mm1=abs(p2-p1)
- mov rdx, arg(3) ;limit
-
- movq mm4, [rdx] ; mm4 = limit
- psubusb mm7, mm4
-
- psubusb mm0, mm4
- psubusb mm1, mm4
-
- psubusb mm6, mm4
- por mm7, mm6
-
- por mm0, mm1
- por mm0, mm7 ; abs(q3-q2) > limit || abs(p3-p2) > limit ||abs(p2-p1) > limit || abs(q2-q1) > limit
-
- movq mm1, mm5 ; p1
-
- movq mm7, mm3 ; mm3=mm7=p0
- psubusb mm7, mm5 ; p0 - p1
-
- psubusb mm5, mm3 ; p1 - p0
- por mm5, mm7 ; abs(p1-p0)
-
- movq t0, mm5 ; save abs(p1-p0)
- lea rdx, srct
-
- psubusb mm5, mm4
- por mm0, mm5 ; mm0=mask
-
- movq mm5, [rdx+16] ; mm5=q0
- movq mm7, [rdx+24] ; mm7=q1
-
- movq mm6, mm5 ; mm6=q0
- movq mm2, mm7 ; q1
- psubusb mm5, mm7 ; q0-q1
-
- psubusb mm7, mm6 ; q1-q0
- por mm7, mm5 ; abs(q1-q0)
-
- movq t1, mm7 ; save abs(q1-q0)
- psubusb mm7, mm4
-
- por mm0, mm7 ; mask
-
- movq mm5, mm2 ; q1
- psubusb mm5, mm1 ; q1-=p1
- psubusb mm1, mm2 ; p1-=q1
- por mm5, mm1 ; abs(p1-q1)
- pand mm5, [GLOBAL(tfe)] ; set lsb of each byte to zero
- psrlw mm5, 1 ; abs(p1-q1)/2
-
- mov rdx, arg(2) ;blimit ;
-
- movq mm4, [rdx] ;blimit
- movq mm1, mm3 ; mm1=mm3=p0
-
- movq mm7, mm6 ; mm7=mm6=q0
- psubusb mm1, mm7 ; p0-q0
-
- psubusb mm7, mm3 ; q0-p0
- por mm1, mm7 ; abs(q0-p0)
- paddusb mm1, mm1 ; abs(q0-p0)*2
- paddusb mm1, mm5 ; abs (p0 - q0) *2 + abs(p1-q1)/2
-
- psubusb mm1, mm4 ; abs (p0 - q0) *2 + abs(p1-q1)/2 > blimit
- por mm1, mm0; ; mask
-
- pxor mm0, mm0
- pcmpeqb mm1, mm0
-
- ; calculate high edge variance
- mov rdx, arg(4) ;thresh ; get thresh
- movq mm7, [rdx]
- ;
- movq mm4, t0 ; get abs (q1 - q0)
- psubusb mm4, mm7
-
- movq mm3, t1 ; get abs (p1 - p0)
- psubusb mm3, mm7
-
- por mm4, mm3 ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
- pcmpeqb mm4, mm0
-
- pcmpeqb mm0, mm0
- pxor mm4, mm0
-
-
-
- ; start work on filters
- lea rdx, srct
-
- movq mm2, [rdx] ; p1
- movq mm7, [rdx+24] ; q1
-
- movq mm6, [rdx+8] ; p0
- movq mm0, [rdx+16] ; q0
-
- pxor mm2, [GLOBAL(t80)] ; p1 offset to convert to signed values
- pxor mm7, [GLOBAL(t80)] ; q1 offset to convert to signed values
-
- psubsb mm2, mm7 ; p1 - q1
- pand mm2, mm4 ; high var mask (hvm)(p1 - q1)
-
- pxor mm6, [GLOBAL(t80)] ; offset to convert to signed values
- pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
-
- movq mm3, mm0 ; q0
- psubsb mm0, mm6 ; q0 - p0
-
- paddsb mm2, mm0 ; 1 * (q0 - p0) + hvm(p1 - q1)
- paddsb mm2, mm0 ; 2 * (q0 - p0) + hvm(p1 - q1)
-
- paddsb mm2, mm0 ; 3 * (q0 - p0) + hvm(p1 - q1)
- pand mm1, mm2 ; mask filter values we don't care about
-
- movq mm2, mm1
- paddsb mm1, [GLOBAL(t4)] ; 3* (q0 - p0) + hvm(p1 - q1) + 4
-
- paddsb mm2, [GLOBAL(t3)] ; 3* (q0 - p0) + hvm(p1 - q1) + 3
- pxor mm0, mm0 ;
-
- pxor mm5, mm5
- punpcklbw mm0, mm2 ;
-
- punpckhbw mm5, mm2 ;
- psraw mm0, 11 ;
-
- psraw mm5, 11
- packsswb mm0, mm5
-
- movq mm2, mm0 ; (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
-
- pxor mm0, mm0 ; 0
- movq mm5, mm1 ; abcdefgh
-
- punpcklbw mm0, mm1 ; e0f0g0h0
- psraw mm0, 11 ; sign extended shift right by 3
-
- pxor mm1, mm1 ; 0
- punpckhbw mm1, mm5 ; a0b0c0d0
-
- psraw mm1, 11 ; sign extended shift right by 3
- movq mm5, mm0 ; save results
-
- packsswb mm0, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
- paddsw mm5, [GLOBAL(ones)]
-
- paddsw mm1, [GLOBAL(ones)]
- psraw mm5, 1 ; partial shifted one more time for 2nd tap
-
- psraw mm1, 1 ; partial shifted one more time for 2nd tap
- packsswb mm5, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
-
- pandn mm4, mm5 ; high edge variance additive
-
- paddsb mm6, mm2 ; p0+= p0 add
- pxor mm6, [GLOBAL(t80)] ; unoffset
-
- ; mm6=p0 ;
- movq mm1, [rdx] ; p1
- pxor mm1, [GLOBAL(t80)] ; reoffset
-
- paddsb mm1, mm4 ; p1+= p1 add
- pxor mm1, [GLOBAL(t80)] ; unoffset
- ; mm6 = p0 mm1 = p1
-
- psubsb mm3, mm0 ; q0-= q0 add
- pxor mm3, [GLOBAL(t80)] ; unoffset
-
- ; mm3 = q0
- psubsb mm7, mm4 ; q1-= q1 add
- pxor mm7, [GLOBAL(t80)] ; unoffset
- ; mm7 = q1
-
- ; transpose and write back
- ; mm1 = 72 62 52 42 32 22 12 02
- ; mm6 = 73 63 53 43 33 23 13 03
- ; mm3 = 74 64 54 44 34 24 14 04
- ; mm7 = 75 65 55 45 35 25 15 05
-
- movq mm2, mm1 ; 72 62 52 42 32 22 12 02
- punpcklbw mm2, mm6 ; 33 32 23 22 13 12 03 02
-
- movq mm4, mm3 ; 74 64 54 44 34 24 14 04
- punpckhbw mm1, mm6 ; 73 72 63 62 53 52 43 42
-
- punpcklbw mm4, mm7 ; 35 34 25 24 15 14 05 04
- punpckhbw mm3, mm7 ; 75 74 65 64 55 54 45 44
-
- movq mm6, mm2 ; 33 32 23 22 13 12 03 02
- punpcklwd mm2, mm4 ; 15 14 13 12 05 04 03 02
-
- punpckhwd mm6, mm4 ; 35 34 33 32 25 24 23 22
- movq mm5, mm1 ; 73 72 63 62 53 52 43 42
-
- punpcklwd mm1, mm3 ; 55 54 53 52 45 44 43 42
- punpckhwd mm5, mm3 ; 75 74 73 72 65 64 63 62
-
-
- ; mm2 = 15 14 13 12 05 04 03 02
- ; mm6 = 35 34 33 32 25 24 23 22
- ; mm5 = 55 54 53 52 45 44 43 42
- ; mm1 = 75 74 73 72 65 64 63 62
-
-
-
- movd [rsi+rax*4+2], mm2
- psrlq mm2, 32
-
- movd [rdi+rax*4+2], mm2
- movd [rsi+rax*2+2], mm6
-
- psrlq mm6, 32
- movd [rsi+rax+2],mm6
-
- movd [rsi+2], mm1
- psrlq mm1, 32
-
- movd [rdi+2], mm1
- neg rax
-
- movd [rdi+rax+2],mm5
- psrlq mm5, 32
-
- movd [rdi+rax*2+2], mm5
-
- lea rsi, [rsi+rax*8]
- dec rcx
- jnz .next8_v
-
- add rsp, 64
- pop rsp
- ; begin epilog
- pop rdi
- pop rsi
- RESTORE_GOT
- UNSHADOW_ARGS
- pop rbp
- ret
-
-SECTION_RODATA
-align 16
-tfe:
- times 8 db 0xfe
-align 16
-t80:
- times 8 db 0x80
-align 16
-t3:
- times 8 db 0x03
-align 16
-t4:
- times 8 db 0x04
-align 16
-ones:
- times 4 dw 0x0001
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -54,7 +54,6 @@
VP9_COMMON_SRCS-yes += common/vp9_tile_common.h
VP9_COMMON_SRCS-yes += common/vp9_tile_common.c
VP9_COMMON_SRCS-yes += common/vp9_loopfilter.c
-VP9_COMMON_SRCS-yes += common/vp9_loopfilter_filters.c
VP9_COMMON_SRCS-yes += common/vp9_thread_common.c
VP9_COMMON_SRCS-yes += common/vp9_mvref_common.c
VP9_COMMON_SRCS-yes += common/vp9_mvref_common.h
@@ -69,14 +68,11 @@
VP9_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/convolve.h
VP9_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/vp9_asm_stubs.c
-VP9_COMMON_SRCS-$(ARCH_X86)$(ARCH_X86_64) += common/x86/vp9_loopfilter_intrin_sse2.c
-VP9_COMMON_SRCS-$(HAVE_AVX2) += common/x86/vp9_loopfilter_intrin_avx2.c
VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_postproc.h
VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_postproc.c
VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_mfqe.h
VP9_COMMON_SRCS-$(CONFIG_VP9_POSTPROC) += common/vp9_mfqe.c
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_sse2.asm
-VP9_COMMON_SRCS-$(HAVE_MMX) += common/x86/vp9_loopfilter_mmx.asm
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_subpixel_8t_sse2.asm
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_subpixel_bilinear_sse2.asm
VP9_COMMON_SRCS-$(HAVE_SSSE3) += common/x86/vp9_subpixel_8t_ssse3.asm
@@ -95,7 +91,6 @@
endif
ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
-VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_high_loopfilter_intrin_sse2.c
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_high_subpixel_8t_sse2.asm
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_high_subpixel_bilinear_sse2.asm
ifeq ($(CONFIG_USE_X86INC),yes)
@@ -147,10 +142,6 @@
VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_idct32x32_msa.c
VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_idct_msa.h
VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_intra_predict_msa.c
-VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_loopfilter_4_msa.c
-VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_loopfilter_8_msa.c
-VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_loopfilter_16_msa.c
-VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_loopfilter_msa.h
ifeq ($(CONFIG_VP9_POSTPROC),yes)
VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_mfqe_msa.c
@@ -165,9 +156,6 @@
endif
endif
-VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_loopfilter_16_neon_asm$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_loopfilter_8_neon_asm$(ASM)
-VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_mb_lpf_neon$(ASM)
VP9_COMMON_SRCS-$(HAVE_NEON_ASM) += common/arm/neon/vp9_save_reg_neon$(ASM)
ifneq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
@@ -174,7 +162,6 @@
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht4x4_add_neon.c
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht8x8_add_neon.c
endif
-VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_loopfilter_neon.c
# neon with assembly and intrinsics implementations. If both are available
# prefer assembly.
@@ -193,7 +180,6 @@
VP9_COMMON_SRCS-yes += common/arm/neon/vp9_idct4x4_add_neon_asm$(ASM)
VP9_COMMON_SRCS-yes += common/arm/neon/vp9_idct8x8_1_add_neon_asm$(ASM)
VP9_COMMON_SRCS-yes += common/arm/neon/vp9_idct8x8_add_neon_asm$(ASM)
-VP9_COMMON_SRCS-yes += common/arm/neon/vp9_loopfilter_4_neon_asm$(ASM)
VP9_COMMON_SRCS-yes += common/arm/neon/vp9_reconintra_neon_asm$(ASM)
else
ifeq ($(HAVE_NEON), yes)
@@ -211,11 +197,6 @@
VP9_COMMON_SRCS-yes += common/arm/neon/vp9_idct4x4_add_neon.c
VP9_COMMON_SRCS-yes += common/arm/neon/vp9_idct8x8_1_add_neon.c
VP9_COMMON_SRCS-yes += common/arm/neon/vp9_idct8x8_add_neon.c
-VP9_COMMON_SRCS-yes += common/arm/neon/vp9_loopfilter_16_neon.c
-VP9_COMMON_SRCS-yes += common/arm/neon/vp9_loopfilter_4_neon.c
-# TODO(johannkoenig): re-enable when chromium build is fixed
-# # https://code.google.com/p/chromium/issues/detail?id=443839
-#VP9_COMMON_SRCS-yes += common/arm/neon/vp9_loopfilter_8_neon.c
endif # HAVE_NEON
endif # HAVE_NEON_ASM
--- /dev/null
+++ b/vpx_dsp/arm/loopfilter_16_neon.asm
@@ -1,0 +1,199 @@
+;
+; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+ EXPORT |vp9_lpf_horizontal_4_dual_neon|
+ ARM
+
+ AREA ||.text||, CODE, READONLY, ALIGN=2
+
+;void vp9_lpf_horizontal_4_dual_neon(uint8_t *s, int p,
+; const uint8_t *blimit0,
+; const uint8_t *limit0,
+; const uint8_t *thresh0,
+; const uint8_t *blimit1,
+; const uint8_t *limit1,
+; const uint8_t *thresh1)
+; r0 uint8_t *s,
+; r1 int p,
+; r2 const uint8_t *blimit0,
+; r3 const uint8_t *limit0,
+; sp const uint8_t *thresh0,
+; sp+4 const uint8_t *blimit1,
+; sp+8 const uint8_t *limit1,
+; sp+12 const uint8_t *thresh1,
+
+|vp9_lpf_horizontal_4_dual_neon| PROC
+ push {lr}+
+ ldr r12, [sp, #4] ; load thresh0
+ vld1.8 {d0}, [r2] ; load blimit0 to first half q+ vld1.8 {d2}, [r3] ; load limit0 to first half q+
+ add r1, r1, r1 ; double pitch
+ ldr r2, [sp, #8] ; load blimit1
+
+ vld1.8 {d4}, [r12] ; load thresh0 to first half q+
+ ldr r3, [sp, #12] ; load limit1
+ ldr r12, [sp, #16] ; load thresh1
+ vld1.8 {d1}, [r2] ; load blimit1 to 2nd half q+
+ sub r2, r0, r1, lsl #1 ; s[-4 * p]
+
+ vld1.8 {d3}, [r3] ; load limit1 to 2nd half q+ vld1.8 {d5}, [r12] ; load thresh1 to 2nd half q+
+ vpush {d8-d15} ; save neon registers+
+ add r3, r2, r1, lsr #1 ; s[-3 * p]
+
+ vld1.u8 {q3}, [r2@64], r1 ; p3+ vld1.u8 {q4}, [r3@64], r1 ; p2+ vld1.u8 {q5}, [r2@64], r1 ; p1+ vld1.u8 {q6}, [r3@64], r1 ; p0+ vld1.u8 {q7}, [r2@64], r1 ; q0+ vld1.u8 {q8}, [r3@64], r1 ; q1+ vld1.u8 {q9}, [r2@64] ; q2+ vld1.u8 {q10}, [r3@64] ; q3+
+ sub r2, r2, r1, lsl #1
+ sub r3, r3, r1, lsl #1
+
+ bl vp9_loop_filter_neon_16
+
+ vst1.u8 {q5}, [r2@64], r1 ; store op1+ vst1.u8 {q6}, [r3@64], r1 ; store op0+ vst1.u8 {q7}, [r2@64], r1 ; store oq0+ vst1.u8 {q8}, [r3@64], r1 ; store oq1+
+ vpop {d8-d15} ; restore neon registers+
+ pop {pc}+ ENDP ; |vp9_lpf_horizontal_4_dual_neon|
+
+; void vp9_loop_filter_neon_16();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. This function uses
+; registers d8-d15, so the calling function must save those registers.
+;
+; r0-r3, r12 PRESERVE
+; q0 blimit
+; q1 limit
+; q2 thresh
+; q3 p3
+; q4 p2
+; q5 p1
+; q6 p0
+; q7 q0
+; q8 q1
+; q9 q2
+; q10 q3
+;
+; Outputs:
+; q5 op1
+; q6 op0
+; q7 oq0
+; q8 oq1
+|vp9_loop_filter_neon_16| PROC
+
+ ; filter_mask
+ vabd.u8 q11, q3, q4 ; m1 = abs(p3 - p2)
+ vabd.u8 q12, q4, q5 ; m2 = abs(p2 - p1)
+ vabd.u8 q13, q5, q6 ; m3 = abs(p1 - p0)
+ vabd.u8 q14, q8, q7 ; m4 = abs(q1 - q0)
+ vabd.u8 q3, q9, q8 ; m5 = abs(q2 - q1)
+ vabd.u8 q4, q10, q9 ; m6 = abs(q3 - q2)
+
+ ; only compare the largest value to limit
+ vmax.u8 q11, q11, q12 ; m7 = max(m1, m2)
+ vmax.u8 q12, q13, q14 ; m8 = max(m3, m4)
+
+ vabd.u8 q9, q6, q7 ; abs(p0 - q0)
+
+ vmax.u8 q3, q3, q4 ; m9 = max(m5, m6)
+
+ vmov.u8 q10, #0x80
+
+ vmax.u8 q15, q11, q12 ; m10 = max(m7, m8)
+
+ vcgt.u8 q13, q13, q2 ; (abs(p1 - p0) > thresh)*-1
+ vcgt.u8 q14, q14, q2 ; (abs(q1 - q0) > thresh)*-1
+ vmax.u8 q15, q15, q3 ; m11 = max(m10, m9)
+
+ vabd.u8 q2, q5, q8 ; a = abs(p1 - q1)
+ vqadd.u8 q9, q9, q9 ; b = abs(p0 - q0) * 2
+
+ veor q7, q7, q10 ; qs0
+
+ vcge.u8 q15, q1, q15 ; abs(m11) > limit
+
+ vshr.u8 q2, q2, #1 ; a = a / 2
+ veor q6, q6, q10 ; ps0
+
+ veor q5, q5, q10 ; ps1
+ vqadd.u8 q9, q9, q2 ; a = b + a
+
+ veor q8, q8, q10 ; qs1
+
+ vmov.u16 q4, #3
+
+ vsubl.s8 q2, d14, d12 ; ( qs0 - ps0)
+ vsubl.s8 q11, d15, d13
+
+ vcge.u8 q9, q0, q9 ; a > blimit
+
+ vqsub.s8 q1, q5, q8 ; filter = clamp(ps1-qs1)
+ vorr q14, q13, q14 ; hev
+
+ vmul.i16 q2, q2, q4 ; 3 * ( qs0 - ps0)
+ vmul.i16 q11, q11, q4
+
+ vand q1, q1, q14 ; filter &= hev
+ vand q15, q15, q9 ; mask
+
+ vmov.u8 q4, #3
+
+ vaddw.s8 q2, q2, d2 ; filter + 3 * (qs0 - ps0)
+ vaddw.s8 q11, q11, d3
+
+ vmov.u8 q9, #4
+
+ ; filter = clamp(filter + 3 * ( qs0 - ps0))
+ vqmovn.s16 d2, q2
+ vqmovn.s16 d3, q11
+ vand q1, q1, q15 ; filter &= mask
+
+ vqadd.s8 q2, q1, q4 ; filter2 = clamp(filter+3)
+ vqadd.s8 q1, q1, q9 ; filter1 = clamp(filter+4)
+ vshr.s8 q2, q2, #3 ; filter2 >>= 3
+ vshr.s8 q1, q1, #3 ; filter1 >>= 3
+
+
+ vqadd.s8 q11, q6, q2 ; u = clamp(ps0 + filter2)
+ vqsub.s8 q0, q7, q1 ; u = clamp(qs0 - filter1)
+
+ ; outer tap adjustments
+ vrshr.s8 q1, q1, #1 ; filter = ++filter1 >> 1
+
+ veor q7, q0, q10 ; *oq0 = u^0x80
+
+ vbic q1, q1, q14 ; filter &= ~hev
+
+ vqadd.s8 q13, q5, q1 ; u = clamp(ps1 + filter)
+ vqsub.s8 q12, q8, q1 ; u = clamp(qs1 - filter)
+
+ veor q6, q11, q10 ; *op0 = u^0x80
+ veor q5, q13, q10 ; *op1 = u^0x80
+ veor q8, q12, q10 ; *oq1 = u^0x80
+
+ bx lr
+ ENDP ; |vp9_loop_filter_neon_16|
+
+ END
--- /dev/null
+++ b/vpx_dsp/arm/loopfilter_16_neon.c
@@ -1,0 +1,179 @@
+/*
+ * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+static INLINE void vp9_loop_filter_neon_16(
+ uint8x16_t qblimit, // blimit
+ uint8x16_t qlimit, // limit
+ uint8x16_t qthresh, // thresh
+ uint8x16_t q3, // p3
+ uint8x16_t q4, // p2
+ uint8x16_t q5, // p1
+ uint8x16_t q6, // p0
+ uint8x16_t q7, // q0
+ uint8x16_t q8, // q1
+ uint8x16_t q9, // q2
+ uint8x16_t q10, // q3
+ uint8x16_t *q5r, // p1
+ uint8x16_t *q6r, // p0
+ uint8x16_t *q7r, // q0
+ uint8x16_t *q8r) { // q1+ uint8x16_t q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
+ int16x8_t q2s16, q11s16;
+ uint16x8_t q4u16;
+ int8x16_t q0s8, q1s8, q2s8, q11s8, q12s8, q13s8;
+ int8x8_t d2s8, d3s8;
+
+ q11u8 = vabdq_u8(q3, q4);
+ q12u8 = vabdq_u8(q4, q5);
+ q13u8 = vabdq_u8(q5, q6);
+ q14u8 = vabdq_u8(q8, q7);
+ q3 = vabdq_u8(q9, q8);
+ q4 = vabdq_u8(q10, q9);
+
+ q11u8 = vmaxq_u8(q11u8, q12u8);
+ q12u8 = vmaxq_u8(q13u8, q14u8);
+ q3 = vmaxq_u8(q3, q4);
+ q15u8 = vmaxq_u8(q11u8, q12u8);
+
+ q9 = vabdq_u8(q6, q7);
+
+ // vp8_hevmask
+ q13u8 = vcgtq_u8(q13u8, qthresh);
+ q14u8 = vcgtq_u8(q14u8, qthresh);
+ q15u8 = vmaxq_u8(q15u8, q3);
+
+ q2u8 = vabdq_u8(q5, q8);
+ q9 = vqaddq_u8(q9, q9);
+
+ q15u8 = vcgeq_u8(qlimit, q15u8);
+
+ // vp8_filter() function
+ // convert to signed
+ q10 = vdupq_n_u8(0x80);
+ q8 = veorq_u8(q8, q10);
+ q7 = veorq_u8(q7, q10);
+ q6 = veorq_u8(q6, q10);
+ q5 = veorq_u8(q5, q10);
+
+ q2u8 = vshrq_n_u8(q2u8, 1);
+ q9 = vqaddq_u8(q9, q2u8);
+
+ q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
+ vget_low_s8(vreinterpretq_s8_u8(q6)));
+ q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
+ vget_high_s8(vreinterpretq_s8_u8(q6)));
+
+ q9 = vcgeq_u8(qblimit, q9);
+
+ q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5),
+ vreinterpretq_s8_u8(q8));
+
+ q14u8 = vorrq_u8(q13u8, q14u8);
+
+ q4u16 = vdupq_n_u16(3);
+ q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
+ q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
+
+ q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
+ q15u8 = vandq_u8(q15u8, q9);
+
+ q1s8 = vreinterpretq_s8_u8(q1u8);
+ q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
+ q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
+
+ q4 = vdupq_n_u8(3);
+ q9 = vdupq_n_u8(4);
+ // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
+ d2s8 = vqmovn_s16(q2s16);
+ d3s8 = vqmovn_s16(q11s16);
+ q1s8 = vcombine_s8(d2s8, d3s8);
+ q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
+ q1s8 = vreinterpretq_s8_u8(q1u8);
+
+ q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q4));
+ q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
+ q2s8 = vshrq_n_s8(q2s8, 3);
+ q1s8 = vshrq_n_s8(q1s8, 3);
+
+ q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
+ q0s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
+
+ q1s8 = vrshrq_n_s8(q1s8, 1);
+ q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
+
+ q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
+ q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
+
+ *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q10);
+ *q7r = veorq_u8(vreinterpretq_u8_s8(q0s8), q10);
+ *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q10);
+ *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q10);
+ return;
+}
+
+void vp9_lpf_horizontal_4_dual_neon(uint8_t *s, int p /* pitch */,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1) {+ uint8x8_t dblimit0, dlimit0, dthresh0, dblimit1, dlimit1, dthresh1;
+ uint8x16_t qblimit, qlimit, qthresh;
+ uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8;
+
+ dblimit0 = vld1_u8(blimit0);
+ dlimit0 = vld1_u8(limit0);
+ dthresh0 = vld1_u8(thresh0);
+ dblimit1 = vld1_u8(blimit1);
+ dlimit1 = vld1_u8(limit1);
+ dthresh1 = vld1_u8(thresh1);
+ qblimit = vcombine_u8(dblimit0, dblimit1);
+ qlimit = vcombine_u8(dlimit0, dlimit1);
+ qthresh = vcombine_u8(dthresh0, dthresh1);
+
+ s -= (p << 2);
+
+ q3u8 = vld1q_u8(s);
+ s += p;
+ q4u8 = vld1q_u8(s);
+ s += p;
+ q5u8 = vld1q_u8(s);
+ s += p;
+ q6u8 = vld1q_u8(s);
+ s += p;
+ q7u8 = vld1q_u8(s);
+ s += p;
+ q8u8 = vld1q_u8(s);
+ s += p;
+ q9u8 = vld1q_u8(s);
+ s += p;
+ q10u8 = vld1q_u8(s);
+
+ vp9_loop_filter_neon_16(qblimit, qlimit, qthresh,
+ q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8,
+ &q5u8, &q6u8, &q7u8, &q8u8);
+
+ s -= (p * 5);
+ vst1q_u8(s, q5u8);
+ s += p;
+ vst1q_u8(s, q6u8);
+ s += p;
+ vst1q_u8(s, q7u8);
+ s += p;
+ vst1q_u8(s, q8u8);
+ return;
+}
--- /dev/null
+++ b/vpx_dsp/arm/loopfilter_4_neon.asm
@@ -1,0 +1,277 @@
+;
+; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+ EXPORT |vp9_lpf_horizontal_4_neon|
+ EXPORT |vp9_lpf_vertical_4_neon|
+ ARM
+
+ AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+; TODO(fgalligan): See about removing the count code as this function is only
+; called with a count of 1.
+;
+; void vp9_lpf_horizontal_4_neon(uint8_t *s,
+; int p /* pitch */,
+; const uint8_t *blimit,
+; const uint8_t *limit,
+; const uint8_t *thresh,
+; int count)
+;
+; r0 uint8_t *s,
+; r1 int p, /* pitch */
+; r2 const uint8_t *blimit,
+; r3 const uint8_t *limit,
+; sp const uint8_t *thresh,
+; sp+4 int count
+|vp9_lpf_horizontal_4_neon| PROC
+ push {lr}+
+ vld1.8 {d0[]}, [r2] ; duplicate *blimit+ ldr r12, [sp, #8] ; load count
+ ldr r2, [sp, #4] ; load thresh
+ add r1, r1, r1 ; double pitch
+
+ cmp r12, #0
+ beq end_vp9_lf_h_edge
+
+ vld1.8 {d1[]}, [r3] ; duplicate *limit+ vld1.8 {d2[]}, [r2] ; duplicate *thresh+
+count_lf_h_loop
+ sub r2, r0, r1, lsl #1 ; move src pointer down by 4 lines
+ add r3, r2, r1, lsr #1 ; set to 3 lines down
+
+ vld1.u8 {d3}, [r2@64], r1 ; p3+ vld1.u8 {d4}, [r3@64], r1 ; p2+ vld1.u8 {d5}, [r2@64], r1 ; p1+ vld1.u8 {d6}, [r3@64], r1 ; p0+ vld1.u8 {d7}, [r2@64], r1 ; q0+ vld1.u8 {d16}, [r3@64], r1 ; q1+ vld1.u8 {d17}, [r2@64] ; q2+ vld1.u8 {d18}, [r3@64] ; q3+
+ sub r2, r2, r1, lsl #1
+ sub r3, r3, r1, lsl #1
+
+ bl vp9_loop_filter_neon
+
+ vst1.u8 {d4}, [r2@64], r1 ; store op1+ vst1.u8 {d5}, [r3@64], r1 ; store op0+ vst1.u8 {d6}, [r2@64], r1 ; store oq0+ vst1.u8 {d7}, [r3@64], r1 ; store oq1+
+ add r0, r0, #8
+ subs r12, r12, #1
+ bne count_lf_h_loop
+
+end_vp9_lf_h_edge
+ pop {pc}+ ENDP ; |vp9_lpf_horizontal_4_neon|
+
+; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+; TODO(fgalligan): See about removing the count code as this function is only
+; called with a count of 1.
+;
+; void vp9_lpf_vertical_4_neon(uint8_t *s,
+; int p /* pitch */,
+; const uint8_t *blimit,
+; const uint8_t *limit,
+; const uint8_t *thresh,
+; int count)
+;
+; r0 uint8_t *s,
+; r1 int p, /* pitch */
+; r2 const uint8_t *blimit,
+; r3 const uint8_t *limit,
+; sp const uint8_t *thresh,
+; sp+4 int count
+|vp9_lpf_vertical_4_neon| PROC
+ push {lr}+
+ vld1.8 {d0[]}, [r2] ; duplicate *blimit+ ldr r12, [sp, #8] ; load count
+ vld1.8 {d1[]}, [r3] ; duplicate *limit+
+ ldr r3, [sp, #4] ; load thresh
+ sub r2, r0, #4 ; move s pointer down by 4 columns
+ cmp r12, #0
+ beq end_vp9_lf_v_edge
+
+ vld1.8 {d2[]}, [r3] ; duplicate *thresh+
+count_lf_v_loop
+ vld1.u8 {d3}, [r2], r1 ; load s data+ vld1.u8 {d4}, [r2], r1+ vld1.u8 {d5}, [r2], r1+ vld1.u8 {d6}, [r2], r1+ vld1.u8 {d7}, [r2], r1+ vld1.u8 {d16}, [r2], r1+ vld1.u8 {d17}, [r2], r1+ vld1.u8 {d18}, [r2]+
+ ;transpose to 8x16 matrix
+ vtrn.32 d3, d7
+ vtrn.32 d4, d16
+ vtrn.32 d5, d17
+ vtrn.32 d6, d18
+
+ vtrn.16 d3, d5
+ vtrn.16 d4, d6
+ vtrn.16 d7, d17
+ vtrn.16 d16, d18
+
+ vtrn.8 d3, d4
+ vtrn.8 d5, d6
+ vtrn.8 d7, d16
+ vtrn.8 d17, d18
+
+ bl vp9_loop_filter_neon
+
+ sub r0, r0, #2
+
+ ;store op1, op0, oq0, oq1
+ vst4.8 {d4[0], d5[0], d6[0], d7[0]}, [r0], r1+ vst4.8 {d4[1], d5[1], d6[1], d7[1]}, [r0], r1+ vst4.8 {d4[2], d5[2], d6[2], d7[2]}, [r0], r1+ vst4.8 {d4[3], d5[3], d6[3], d7[3]}, [r0], r1+ vst4.8 {d4[4], d5[4], d6[4], d7[4]}, [r0], r1+ vst4.8 {d4[5], d5[5], d6[5], d7[5]}, [r0], r1+ vst4.8 {d4[6], d5[6], d6[6], d7[6]}, [r0], r1+ vst4.8 {d4[7], d5[7], d6[7], d7[7]}, [r0]+
+ add r0, r0, r1, lsl #3 ; s += pitch * 8
+ subs r12, r12, #1
+ subne r2, r0, #4 ; move s pointer down by 4 columns
+ bne count_lf_v_loop
+
+end_vp9_lf_v_edge
+ pop {pc}+ ENDP ; |vp9_lpf_vertical_4_neon|
+
+; void vp9_loop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0 blimit
+; d1 limit
+; d2 thresh
+; d3 p3
+; d4 p2
+; d5 p1
+; d6 p0
+; d7 q0
+; d16 q1
+; d17 q2
+; d18 q3
+;
+; Outputs:
+; d4 op1
+; d5 op0
+; d6 oq0
+; d7 oq1
+|vp9_loop_filter_neon| PROC
+ ; filter_mask
+ vabd.u8 d19, d3, d4 ; m1 = abs(p3 - p2)
+ vabd.u8 d20, d4, d5 ; m2 = abs(p2 - p1)
+ vabd.u8 d21, d5, d6 ; m3 = abs(p1 - p0)
+ vabd.u8 d22, d16, d7 ; m4 = abs(q1 - q0)
+ vabd.u8 d3, d17, d16 ; m5 = abs(q2 - q1)
+ vabd.u8 d4, d18, d17 ; m6 = abs(q3 - q2)
+
+ ; only compare the largest value to limit
+ vmax.u8 d19, d19, d20 ; m1 = max(m1, m2)
+ vmax.u8 d20, d21, d22 ; m2 = max(m3, m4)
+
+ vabd.u8 d17, d6, d7 ; abs(p0 - q0)
+
+ vmax.u8 d3, d3, d4 ; m3 = max(m5, m6)
+
+ vmov.u8 d18, #0x80
+
+ vmax.u8 d23, d19, d20 ; m1 = max(m1, m2)
+
+ ; hevmask
+ vcgt.u8 d21, d21, d2 ; (abs(p1 - p0) > thresh)*-1
+ vcgt.u8 d22, d22, d2 ; (abs(q1 - q0) > thresh)*-1
+ vmax.u8 d23, d23, d3 ; m1 = max(m1, m3)
+
+ vabd.u8 d28, d5, d16 ; a = abs(p1 - q1)
+ vqadd.u8 d17, d17, d17 ; b = abs(p0 - q0) * 2
+
+ veor d7, d7, d18 ; qs0
+
+ vcge.u8 d23, d1, d23 ; abs(m1) > limit
+
+ ; filter() function
+ ; convert to signed
+
+ vshr.u8 d28, d28, #1 ; a = a / 2
+ veor d6, d6, d18 ; ps0
+
+ veor d5, d5, d18 ; ps1
+ vqadd.u8 d17, d17, d28 ; a = b + a
+
+ veor d16, d16, d18 ; qs1
+
+ vmov.u8 d19, #3
+
+ vsub.s8 d28, d7, d6 ; ( qs0 - ps0)
+
+ vcge.u8 d17, d0, d17 ; a > blimit
+
+ vqsub.s8 d27, d5, d16 ; filter = clamp(ps1-qs1)
+ vorr d22, d21, d22 ; hevmask
+
+ vmull.s8 q12, d28, d19 ; 3 * ( qs0 - ps0)
+
+ vand d27, d27, d22 ; filter &= hev
+ vand d23, d23, d17 ; filter_mask
+
+ vaddw.s8 q12, q12, d27 ; filter + 3 * (qs0 - ps0)
+
+ vmov.u8 d17, #4
+
+ ; filter = clamp(filter + 3 * ( qs0 - ps0))
+ vqmovn.s16 d27, q12
+
+ vand d27, d27, d23 ; filter &= mask
+
+ vqadd.s8 d28, d27, d19 ; filter2 = clamp(filter+3)
+ vqadd.s8 d27, d27, d17 ; filter1 = clamp(filter+4)
+ vshr.s8 d28, d28, #3 ; filter2 >>= 3
+ vshr.s8 d27, d27, #3 ; filter1 >>= 3
+
+ vqadd.s8 d19, d6, d28 ; u = clamp(ps0 + filter2)
+ vqsub.s8 d26, d7, d27 ; u = clamp(qs0 - filter1)
+
+ ; outer tap adjustments
+ vrshr.s8 d27, d27, #1 ; filter = ++filter1 >> 1
+
+ veor d6, d26, d18 ; *oq0 = u^0x80
+
+ vbic d27, d27, d22 ; filter &= ~hev
+
+ vqadd.s8 d21, d5, d27 ; u = clamp(ps1 + filter)
+ vqsub.s8 d20, d16, d27 ; u = clamp(qs1 - filter)
+
+ veor d5, d19, d18 ; *op0 = u^0x80
+ veor d4, d21, d18 ; *op1 = u^0x80
+ veor d7, d20, d18 ; *oq1 = u^0x80
+
+ bx lr
+ ENDP ; |vp9_loop_filter_neon|
+
+ END
--- /dev/null
+++ b/vpx_dsp/arm/loopfilter_4_neon.c
@@ -1,0 +1,274 @@
+/*
+ * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+
+static INLINE void vp9_loop_filter_neon(
+ uint8x8_t dblimit, // flimit
+ uint8x8_t dlimit, // limit
+ uint8x8_t dthresh, // thresh
+ uint8x8_t d3u8, // p3
+ uint8x8_t d4u8, // p2
+ uint8x8_t d5u8, // p1
+ uint8x8_t d6u8, // p0
+ uint8x8_t d7u8, // q0
+ uint8x8_t d16u8, // q1
+ uint8x8_t d17u8, // q2
+ uint8x8_t d18u8, // q3
+ uint8x8_t *d4ru8, // p1
+ uint8x8_t *d5ru8, // p0
+ uint8x8_t *d6ru8, // q0
+ uint8x8_t *d7ru8) { // q1+ uint8x8_t d19u8, d20u8, d21u8, d22u8, d23u8, d27u8, d28u8;
+ int16x8_t q12s16;
+ int8x8_t d19s8, d20s8, d21s8, d26s8, d27s8, d28s8;
+
+ d19u8 = vabd_u8(d3u8, d4u8);
+ d20u8 = vabd_u8(d4u8, d5u8);
+ d21u8 = vabd_u8(d5u8, d6u8);
+ d22u8 = vabd_u8(d16u8, d7u8);
+ d3u8 = vabd_u8(d17u8, d16u8);
+ d4u8 = vabd_u8(d18u8, d17u8);
+
+ d19u8 = vmax_u8(d19u8, d20u8);
+ d20u8 = vmax_u8(d21u8, d22u8);
+ d3u8 = vmax_u8(d3u8, d4u8);
+ d23u8 = vmax_u8(d19u8, d20u8);
+
+ d17u8 = vabd_u8(d6u8, d7u8);
+
+ d21u8 = vcgt_u8(d21u8, dthresh);
+ d22u8 = vcgt_u8(d22u8, dthresh);
+ d23u8 = vmax_u8(d23u8, d3u8);
+
+ d28u8 = vabd_u8(d5u8, d16u8);
+ d17u8 = vqadd_u8(d17u8, d17u8);
+
+ d23u8 = vcge_u8(dlimit, d23u8);
+
+ d18u8 = vdup_n_u8(0x80);
+ d5u8 = veor_u8(d5u8, d18u8);
+ d6u8 = veor_u8(d6u8, d18u8);
+ d7u8 = veor_u8(d7u8, d18u8);
+ d16u8 = veor_u8(d16u8, d18u8);
+
+ d28u8 = vshr_n_u8(d28u8, 1);
+ d17u8 = vqadd_u8(d17u8, d28u8);
+
+ d19u8 = vdup_n_u8(3);
+
+ d28s8 = vsub_s8(vreinterpret_s8_u8(d7u8),
+ vreinterpret_s8_u8(d6u8));
+
+ d17u8 = vcge_u8(dblimit, d17u8);
+
+ d27s8 = vqsub_s8(vreinterpret_s8_u8(d5u8),
+ vreinterpret_s8_u8(d16u8));
+
+ d22u8 = vorr_u8(d21u8, d22u8);
+
+ q12s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+ d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d22u8);
+ d23u8 = vand_u8(d23u8, d17u8);
+
+ q12s16 = vaddw_s8(q12s16, vreinterpret_s8_u8(d27u8));
+
+ d17u8 = vdup_n_u8(4);
+
+ d27s8 = vqmovn_s16(q12s16);
+ d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d23u8);
+ d27s8 = vreinterpret_s8_u8(d27u8);
+
+ d28s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d19u8));
+ d27s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d17u8));
+ d28s8 = vshr_n_s8(d28s8, 3);
+ d27s8 = vshr_n_s8(d27s8, 3);
+
+ d19s8 = vqadd_s8(vreinterpret_s8_u8(d6u8), d28s8);
+ d26s8 = vqsub_s8(vreinterpret_s8_u8(d7u8), d27s8);
+
+ d27s8 = vrshr_n_s8(d27s8, 1);
+ d27s8 = vbic_s8(d27s8, vreinterpret_s8_u8(d22u8));
+
+ d21s8 = vqadd_s8(vreinterpret_s8_u8(d5u8), d27s8);
+ d20s8 = vqsub_s8(vreinterpret_s8_u8(d16u8), d27s8);
+
+ *d4ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d18u8);
+ *d5ru8 = veor_u8(vreinterpret_u8_s8(d19s8), d18u8);
+ *d6ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d18u8);
+ *d7ru8 = veor_u8(vreinterpret_u8_s8(d20s8), d18u8);
+ return;
+}
+
+void vp9_lpf_horizontal_4_neon(
+ uint8_t *src,
+ int pitch,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int count) {+ int i;
+ uint8_t *s, *psrc;
+ uint8x8_t dblimit, dlimit, dthresh;
+ uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+
+ if (count == 0) // end_vp9_lf_h_edge
+ return;
+
+ dblimit = vld1_u8(blimit);
+ dlimit = vld1_u8(limit);
+ dthresh = vld1_u8(thresh);
+
+ psrc = src - (pitch << 2);
+ for (i = 0; i < count; i++) {+ s = psrc + i * 8;
+
+ d3u8 = vld1_u8(s);
+ s += pitch;
+ d4u8 = vld1_u8(s);
+ s += pitch;
+ d5u8 = vld1_u8(s);
+ s += pitch;
+ d6u8 = vld1_u8(s);
+ s += pitch;
+ d7u8 = vld1_u8(s);
+ s += pitch;
+ d16u8 = vld1_u8(s);
+ s += pitch;
+ d17u8 = vld1_u8(s);
+ s += pitch;
+ d18u8 = vld1_u8(s);
+
+ vp9_loop_filter_neon(dblimit, dlimit, dthresh,
+ d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+ &d4u8, &d5u8, &d6u8, &d7u8);
+
+ s -= (pitch * 5);
+ vst1_u8(s, d4u8);
+ s += pitch;
+ vst1_u8(s, d5u8);
+ s += pitch;
+ vst1_u8(s, d6u8);
+ s += pitch;
+ vst1_u8(s, d7u8);
+ }
+ return;
+}
+
+void vp9_lpf_vertical_4_neon(
+ uint8_t *src,
+ int pitch,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int count) {+ int i, pitch8;
+ uint8_t *s;
+ uint8x8_t dblimit, dlimit, dthresh;
+ uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
+ uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+ uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+ uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+ uint8x8x4_t d4Result;
+
+ if (count == 0) // end_vp9_lf_h_edge
+ return;
+
+ dblimit = vld1_u8(blimit);
+ dlimit = vld1_u8(limit);
+ dthresh = vld1_u8(thresh);
+
+ pitch8 = pitch * 8;
+ for (i = 0; i < count; i++, src += pitch8) {+ s = src - (i + 1) * 4;
+
+ d3u8 = vld1_u8(s);
+ s += pitch;
+ d4u8 = vld1_u8(s);
+ s += pitch;
+ d5u8 = vld1_u8(s);
+ s += pitch;
+ d6u8 = vld1_u8(s);
+ s += pitch;
+ d7u8 = vld1_u8(s);
+ s += pitch;
+ d16u8 = vld1_u8(s);
+ s += pitch;
+ d17u8 = vld1_u8(s);
+ s += pitch;
+ d18u8 = vld1_u8(s);
+
+ d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
+ vreinterpret_u32_u8(d7u8));
+ d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
+ vreinterpret_u32_u8(d16u8));
+ d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
+ vreinterpret_u32_u8(d17u8));
+ d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
+ vreinterpret_u32_u8(d18u8));
+
+ d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+ vreinterpret_u16_u32(d2tmp2.val[0]));
+ d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+ vreinterpret_u16_u32(d2tmp3.val[0]));
+ d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+ vreinterpret_u16_u32(d2tmp2.val[1]));
+ d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+ vreinterpret_u16_u32(d2tmp3.val[1]));
+
+ d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+ vreinterpret_u8_u16(d2tmp5.val[0]));
+ d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+ vreinterpret_u8_u16(d2tmp5.val[1]));
+ d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+ vreinterpret_u8_u16(d2tmp7.val[0]));
+ d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+ vreinterpret_u8_u16(d2tmp7.val[1]));
+
+ d3u8 = d2tmp8.val[0];
+ d4u8 = d2tmp8.val[1];
+ d5u8 = d2tmp9.val[0];
+ d6u8 = d2tmp9.val[1];
+ d7u8 = d2tmp10.val[0];
+ d16u8 = d2tmp10.val[1];
+ d17u8 = d2tmp11.val[0];
+ d18u8 = d2tmp11.val[1];
+
+ vp9_loop_filter_neon(dblimit, dlimit, dthresh,
+ d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+ &d4u8, &d5u8, &d6u8, &d7u8);
+
+ d4Result.val[0] = d4u8;
+ d4Result.val[1] = d5u8;
+ d4Result.val[2] = d6u8;
+ d4Result.val[3] = d7u8;
+
+ src -= 2;
+ vst4_lane_u8(src, d4Result, 0);
+ src += pitch;
+ vst4_lane_u8(src, d4Result, 1);
+ src += pitch;
+ vst4_lane_u8(src, d4Result, 2);
+ src += pitch;
+ vst4_lane_u8(src, d4Result, 3);
+ src += pitch;
+ vst4_lane_u8(src, d4Result, 4);
+ src += pitch;
+ vst4_lane_u8(src, d4Result, 5);
+ src += pitch;
+ vst4_lane_u8(src, d4Result, 6);
+ src += pitch;
+ vst4_lane_u8(src, d4Result, 7);
+ }
+ return;
+}
--- /dev/null
+++ b/vpx_dsp/arm/loopfilter_8_neon.asm
@@ -1,0 +1,451 @@
+;
+; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+ EXPORT |vp9_lpf_horizontal_8_neon|
+ EXPORT |vp9_lpf_vertical_8_neon|
+ ARM
+
+ AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; Currently vp9 only works on iterations 8 at a time. The vp8 loop filter
+; works on 16 iterations at a time.
+; TODO(fgalligan): See about removing the count code as this function is only
+; called with a count of 1.
+;
+; void vp9_lpf_horizontal_8_neon(uint8_t *s, int p,
+; const uint8_t *blimit,
+; const uint8_t *limit,
+; const uint8_t *thresh,
+; int count)
+; r0 uint8_t *s,
+; r1 int p, /* pitch */
+; r2 const uint8_t *blimit,
+; r3 const uint8_t *limit,
+; sp const uint8_t *thresh,
+; sp+4 int count
+|vp9_lpf_horizontal_8_neon| PROC
+ push {r4-r5, lr}+
+ vld1.8 {d0[]}, [r2] ; duplicate *blimit+ ldr r12, [sp, #16] ; load count
+ ldr r2, [sp, #12] ; load thresh
+ add r1, r1, r1 ; double pitch
+
+ cmp r12, #0
+ beq end_vp9_mblf_h_edge
+
+ vld1.8 {d1[]}, [r3] ; duplicate *limit+ vld1.8 {d2[]}, [r2] ; duplicate *thresh+
+count_mblf_h_loop
+ sub r3, r0, r1, lsl #1 ; move src pointer down by 4 lines
+ add r2, r3, r1, lsr #1 ; set to 3 lines down
+
+ vld1.u8 {d3}, [r3@64], r1 ; p3+ vld1.u8 {d4}, [r2@64], r1 ; p2+ vld1.u8 {d5}, [r3@64], r1 ; p1+ vld1.u8 {d6}, [r2@64], r1 ; p0+ vld1.u8 {d7}, [r3@64], r1 ; q0+ vld1.u8 {d16}, [r2@64], r1 ; q1+ vld1.u8 {d17}, [r3@64] ; q2+ vld1.u8 {d18}, [r2@64], r1 ; q3+
+ sub r3, r3, r1, lsl #1
+ sub r2, r2, r1, lsl #2
+
+ bl vp9_mbloop_filter_neon
+
+ vst1.u8 {d0}, [r2@64], r1 ; store op2+ vst1.u8 {d1}, [r3@64], r1 ; store op1+ vst1.u8 {d2}, [r2@64], r1 ; store op0+ vst1.u8 {d3}, [r3@64], r1 ; store oq0+ vst1.u8 {d4}, [r2@64], r1 ; store oq1+ vst1.u8 {d5}, [r3@64], r1 ; store oq2+
+ add r0, r0, #8
+ subs r12, r12, #1
+ bne count_mblf_h_loop
+
+end_vp9_mblf_h_edge
+ pop {r4-r5, pc}+
+ ENDP ; |vp9_lpf_horizontal_8_neon|
+
+; void vp9_lpf_vertical_8_neon(uint8_t *s,
+; int pitch,
+; const uint8_t *blimit,
+; const uint8_t *limit,
+; const uint8_t *thresh,
+; int count)
+;
+; r0 uint8_t *s,
+; r1 int pitch,
+; r2 const uint8_t *blimit,
+; r3 const uint8_t *limit,
+; sp const uint8_t *thresh,
+; sp+4 int count
+|vp9_lpf_vertical_8_neon| PROC
+ push {r4-r5, lr}+
+ vld1.8 {d0[]}, [r2] ; duplicate *blimit+ ldr r12, [sp, #16] ; load count
+ vld1.8 {d1[]}, [r3] ; duplicate *limit+
+ ldr r3, [sp, #12] ; load thresh
+ sub r2, r0, #4 ; move s pointer down by 4 columns
+ cmp r12, #0
+ beq end_vp9_mblf_v_edge
+
+ vld1.8 {d2[]}, [r3] ; duplicate *thresh+
+count_mblf_v_loop
+ vld1.u8 {d3}, [r2], r1 ; load s data+ vld1.u8 {d4}, [r2], r1+ vld1.u8 {d5}, [r2], r1+ vld1.u8 {d6}, [r2], r1+ vld1.u8 {d7}, [r2], r1+ vld1.u8 {d16}, [r2], r1+ vld1.u8 {d17}, [r2], r1+ vld1.u8 {d18}, [r2]+
+ ;transpose to 8x16 matrix
+ vtrn.32 d3, d7
+ vtrn.32 d4, d16
+ vtrn.32 d5, d17
+ vtrn.32 d6, d18
+
+ vtrn.16 d3, d5
+ vtrn.16 d4, d6
+ vtrn.16 d7, d17
+ vtrn.16 d16, d18
+
+ vtrn.8 d3, d4
+ vtrn.8 d5, d6
+ vtrn.8 d7, d16
+ vtrn.8 d17, d18
+
+ sub r2, r0, #3
+ add r3, r0, #1
+
+ bl vp9_mbloop_filter_neon
+
+ ;store op2, op1, op0, oq0
+ vst4.8 {d0[0], d1[0], d2[0], d3[0]}, [r2], r1+ vst4.8 {d0[1], d1[1], d2[1], d3[1]}, [r2], r1+ vst4.8 {d0[2], d1[2], d2[2], d3[2]}, [r2], r1+ vst4.8 {d0[3], d1[3], d2[3], d3[3]}, [r2], r1+ vst4.8 {d0[4], d1[4], d2[4], d3[4]}, [r2], r1+ vst4.8 {d0[5], d1[5], d2[5], d3[5]}, [r2], r1+ vst4.8 {d0[6], d1[6], d2[6], d3[6]}, [r2], r1+ vst4.8 {d0[7], d1[7], d2[7], d3[7]}, [r2]+
+ ;store oq1, oq2
+ vst2.8 {d4[0], d5[0]}, [r3], r1+ vst2.8 {d4[1], d5[1]}, [r3], r1+ vst2.8 {d4[2], d5[2]}, [r3], r1+ vst2.8 {d4[3], d5[3]}, [r3], r1+ vst2.8 {d4[4], d5[4]}, [r3], r1+ vst2.8 {d4[5], d5[5]}, [r3], r1+ vst2.8 {d4[6], d5[6]}, [r3], r1+ vst2.8 {d4[7], d5[7]}, [r3]+
+ add r0, r0, r1, lsl #3 ; s += pitch * 8
+ subs r12, r12, #1
+ subne r2, r0, #4 ; move s pointer down by 4 columns
+ bne count_mblf_v_loop
+
+end_vp9_mblf_v_edge
+ pop {r4-r5, pc}+ ENDP ; |vp9_lpf_vertical_8_neon|
+
+; void vp9_mbloop_filter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store. The function does not use
+; registers d8-d15.
+;
+; Inputs:
+; r0-r3, r12 PRESERVE
+; d0 blimit
+; d1 limit
+; d2 thresh
+; d3 p3
+; d4 p2
+; d5 p1
+; d6 p0
+; d7 q0
+; d16 q1
+; d17 q2
+; d18 q3
+;
+; Outputs:
+; d0 op2
+; d1 op1
+; d2 op0
+; d3 oq0
+; d4 oq1
+; d5 oq2
+|vp9_mbloop_filter_neon| PROC
+ ; filter_mask
+ vabd.u8 d19, d3, d4 ; m1 = abs(p3 - p2)
+ vabd.u8 d20, d4, d5 ; m2 = abs(p2 - p1)
+ vabd.u8 d21, d5, d6 ; m3 = abs(p1 - p0)
+ vabd.u8 d22, d16, d7 ; m4 = abs(q1 - q0)
+ vabd.u8 d23, d17, d16 ; m5 = abs(q2 - q1)
+ vabd.u8 d24, d18, d17 ; m6 = abs(q3 - q2)
+
+ ; only compare the largest value to limit
+ vmax.u8 d19, d19, d20 ; m1 = max(m1, m2)
+ vmax.u8 d20, d21, d22 ; m2 = max(m3, m4)
+
+ vabd.u8 d25, d6, d4 ; m7 = abs(p0 - p2)
+
+ vmax.u8 d23, d23, d24 ; m3 = max(m5, m6)
+
+ vabd.u8 d26, d7, d17 ; m8 = abs(q0 - q2)
+
+ vmax.u8 d19, d19, d20
+
+ vabd.u8 d24, d6, d7 ; m9 = abs(p0 - q0)
+ vabd.u8 d27, d3, d6 ; m10 = abs(p3 - p0)
+ vabd.u8 d28, d18, d7 ; m11 = abs(q3 - q0)
+
+ vmax.u8 d19, d19, d23
+
+ vabd.u8 d23, d5, d16 ; a = abs(p1 - q1)
+ vqadd.u8 d24, d24, d24 ; b = abs(p0 - q0) * 2
+
+ ; abs () > limit
+ vcge.u8 d19, d1, d19
+
+ ; only compare the largest value to thresh
+ vmax.u8 d25, d25, d26 ; m4 = max(m7, m8)
+ vmax.u8 d26, d27, d28 ; m5 = max(m10, m11)
+
+ vshr.u8 d23, d23, #1 ; a = a / 2
+
+ vmax.u8 d25, d25, d26 ; m4 = max(m4, m5)
+
+ vqadd.u8 d24, d24, d23 ; a = b + a
+
+ vmax.u8 d20, d20, d25 ; m2 = max(m2, m4)
+
+ vmov.u8 d23, #1
+ vcge.u8 d24, d0, d24 ; a > blimit
+
+ vcgt.u8 d21, d21, d2 ; (abs(p1 - p0) > thresh)*-1
+
+ vcge.u8 d20, d23, d20 ; flat
+
+ vand d19, d19, d24 ; mask
+
+ vcgt.u8 d23, d22, d2 ; (abs(q1 - q0) > thresh)*-1
+
+ vand d20, d20, d19 ; flat & mask
+
+ vmov.u8 d22, #0x80
+
+ vorr d23, d21, d23 ; hev
+
+ ; This instruction will truncate the "flat & mask" masks down to 4 bits
+ ; each to fit into one 32 bit arm register. The values are stored in
+ ; q10.64[0].
+ vshrn.u16 d30, q10, #4
+ vmov.u32 r4, d30[0] ; flat & mask 4bits
+
+ adds r5, r4, #1 ; Check for all 1's
+
+ ; If mask and flat are 1's for all vectors, then we only need to execute
+ ; the power branch for all vectors.
+ beq power_branch_only
+
+ cmp r4, #0 ; Check for 0, set flag for later
+
+ ; mbfilter() function
+ ; filter() function
+ ; convert to signed
+ veor d21, d7, d22 ; qs0
+ veor d24, d6, d22 ; ps0
+ veor d25, d5, d22 ; ps1
+ veor d26, d16, d22 ; qs1
+
+ vmov.u8 d27, #3
+
+ vsub.s8 d28, d21, d24 ; ( qs0 - ps0)
+
+ vqsub.s8 d29, d25, d26 ; filter = clamp(ps1-qs1)
+
+ vmull.s8 q15, d28, d27 ; 3 * ( qs0 - ps0)
+
+ vand d29, d29, d23 ; filter &= hev
+
+ vaddw.s8 q15, q15, d29 ; filter + 3 * (qs0 - ps0)
+
+ vmov.u8 d29, #4
+
+ ; filter = clamp(filter + 3 * ( qs0 - ps0))
+ vqmovn.s16 d28, q15
+
+ vand d28, d28, d19 ; filter &= mask
+
+ vqadd.s8 d30, d28, d27 ; filter2 = clamp(filter+3)
+ vqadd.s8 d29, d28, d29 ; filter1 = clamp(filter+4)
+ vshr.s8 d30, d30, #3 ; filter2 >>= 3
+ vshr.s8 d29, d29, #3 ; filter1 >>= 3
+
+ vqadd.s8 d24, d24, d30 ; op0 = clamp(ps0 + filter2)
+ vqsub.s8 d21, d21, d29 ; oq0 = clamp(qs0 - filter1)
+
+ ; outer tap adjustments: ++filter1 >> 1
+ vrshr.s8 d29, d29, #1
+ vbic d29, d29, d23 ; filter &= ~hev
+
+ vqadd.s8 d25, d25, d29 ; op1 = clamp(ps1 + filter)
+ vqsub.s8 d26, d26, d29 ; oq1 = clamp(qs1 - filter)
+
+ ; If mask and flat are 0's for all vectors, then we only need to execute
+ ; the filter branch for all vectors.
+ beq filter_branch_only
+
+ ; If mask and flat are mixed then we must perform both branches and
+ ; combine the data.
+ veor d24, d24, d22 ; *f_op0 = u^0x80
+ veor d21, d21, d22 ; *f_oq0 = u^0x80
+ veor d25, d25, d22 ; *f_op1 = u^0x80
+ veor d26, d26, d22 ; *f_oq1 = u^0x80
+
+ ; At this point we have already executed the filter branch. The filter
+ ; branch does not set op2 or oq2, so use p2 and q2. Execute the power
+ ; branch and combine the data.
+ vmov.u8 d23, #2
+ vaddl.u8 q14, d6, d7 ; r_op2 = p0 + q0
+ vmlal.u8 q14, d3, d27 ; r_op2 += p3 * 3
+ vmlal.u8 q14, d4, d23 ; r_op2 += p2 * 2
+
+ vbif d0, d4, d20 ; op2 |= p2 & ~(flat & mask)
+
+ vaddw.u8 q14, d5 ; r_op2 += p1
+
+ vbif d1, d25, d20 ; op1 |= f_op1 & ~(flat & mask)
+
+ vqrshrn.u16 d30, q14, #3 ; r_op2
+
+ vsubw.u8 q14, d3 ; r_op1 = r_op2 - p3
+ vsubw.u8 q14, d4 ; r_op1 -= p2
+ vaddw.u8 q14, d5 ; r_op1 += p1
+ vaddw.u8 q14, d16 ; r_op1 += q1
+
+ vbif d2, d24, d20 ; op0 |= f_op0 & ~(flat & mask)
+
+ vqrshrn.u16 d31, q14, #3 ; r_op1
+
+ vsubw.u8 q14, d3 ; r_op0 = r_op1 - p3
+ vsubw.u8 q14, d5 ; r_op0 -= p1
+ vaddw.u8 q14, d6 ; r_op0 += p0
+ vaddw.u8 q14, d17 ; r_op0 += q2
+
+ vbit d0, d30, d20 ; op2 |= r_op2 & (flat & mask)
+
+ vqrshrn.u16 d23, q14, #3 ; r_op0
+
+ vsubw.u8 q14, d3 ; r_oq0 = r_op0 - p3
+ vsubw.u8 q14, d6 ; r_oq0 -= p0
+ vaddw.u8 q14, d7 ; r_oq0 += q0
+
+ vbit d1, d31, d20 ; op1 |= r_op1 & (flat & mask)
+
+ vaddw.u8 q14, d18 ; oq0 += q3
+
+ vbit d2, d23, d20 ; op0 |= r_op0 & (flat & mask)
+
+ vqrshrn.u16 d22, q14, #3 ; r_oq0
+
+ vsubw.u8 q14, d4 ; r_oq1 = r_oq0 - p2
+ vsubw.u8 q14, d7 ; r_oq1 -= q0
+ vaddw.u8 q14, d16 ; r_oq1 += q1
+
+ vbif d3, d21, d20 ; oq0 |= f_oq0 & ~(flat & mask)
+
+ vaddw.u8 q14, d18 ; r_oq1 += q3
+
+ vbif d4, d26, d20 ; oq1 |= f_oq1 & ~(flat & mask)
+
+ vqrshrn.u16 d6, q14, #3 ; r_oq1
+
+ vsubw.u8 q14, d5 ; r_oq2 = r_oq1 - p1
+ vsubw.u8 q14, d16 ; r_oq2 -= q1
+ vaddw.u8 q14, d17 ; r_oq2 += q2
+ vaddw.u8 q14, d18 ; r_oq2 += q3
+
+ vbif d5, d17, d20 ; oq2 |= q2 & ~(flat & mask)
+
+ vqrshrn.u16 d7, q14, #3 ; r_oq2
+
+ vbit d3, d22, d20 ; oq0 |= r_oq0 & (flat & mask)
+ vbit d4, d6, d20 ; oq1 |= r_oq1 & (flat & mask)
+ vbit d5, d7, d20 ; oq2 |= r_oq2 & (flat & mask)
+
+ bx lr
+
+power_branch_only
+ vmov.u8 d27, #3
+ vmov.u8 d21, #2
+ vaddl.u8 q14, d6, d7 ; op2 = p0 + q0
+ vmlal.u8 q14, d3, d27 ; op2 += p3 * 3
+ vmlal.u8 q14, d4, d21 ; op2 += p2 * 2
+ vaddw.u8 q14, d5 ; op2 += p1
+ vqrshrn.u16 d0, q14, #3 ; op2
+
+ vsubw.u8 q14, d3 ; op1 = op2 - p3
+ vsubw.u8 q14, d4 ; op1 -= p2
+ vaddw.u8 q14, d5 ; op1 += p1
+ vaddw.u8 q14, d16 ; op1 += q1
+ vqrshrn.u16 d1, q14, #3 ; op1
+
+ vsubw.u8 q14, d3 ; op0 = op1 - p3
+ vsubw.u8 q14, d5 ; op0 -= p1
+ vaddw.u8 q14, d6 ; op0 += p0
+ vaddw.u8 q14, d17 ; op0 += q2
+ vqrshrn.u16 d2, q14, #3 ; op0
+
+ vsubw.u8 q14, d3 ; oq0 = op0 - p3
+ vsubw.u8 q14, d6 ; oq0 -= p0
+ vaddw.u8 q14, d7 ; oq0 += q0
+ vaddw.u8 q14, d18 ; oq0 += q3
+ vqrshrn.u16 d3, q14, #3 ; oq0
+
+ vsubw.u8 q14, d4 ; oq1 = oq0 - p2
+ vsubw.u8 q14, d7 ; oq1 -= q0
+ vaddw.u8 q14, d16 ; oq1 += q1
+ vaddw.u8 q14, d18 ; oq1 += q3
+ vqrshrn.u16 d4, q14, #3 ; oq1
+
+ vsubw.u8 q14, d5 ; oq2 = oq1 - p1
+ vsubw.u8 q14, d16 ; oq2 -= q1
+ vaddw.u8 q14, d17 ; oq2 += q2
+ vaddw.u8 q14, d18 ; oq2 += q3
+ vqrshrn.u16 d5, q14, #3 ; oq2
+
+ bx lr
+
+filter_branch_only
+ ; TODO(fgalligan): See if we can rearange registers so we do not need to
+ ; do the 2 vswp.
+ vswp d0, d4 ; op2
+ vswp d5, d17 ; oq2
+ veor d2, d24, d22 ; *op0 = u^0x80
+ veor d3, d21, d22 ; *oq0 = u^0x80
+ veor d1, d25, d22 ; *op1 = u^0x80
+ veor d4, d26, d22 ; *oq1 = u^0x80
+
+ bx lr
+
+ ENDP ; |vp9_mbloop_filter_neon|
+
+ END
--- /dev/null
+++ b/vpx_dsp/arm/loopfilter_8_neon.c
@@ -1,0 +1,453 @@
+/*
+ * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+
+static INLINE void vp9_mbloop_filter_neon(
+ uint8x8_t dblimit, // mblimit
+ uint8x8_t dlimit, // limit
+ uint8x8_t dthresh, // thresh
+ uint8x8_t d3u8, // p2
+ uint8x8_t d4u8, // p2
+ uint8x8_t d5u8, // p1
+ uint8x8_t d6u8, // p0
+ uint8x8_t d7u8, // q0
+ uint8x8_t d16u8, // q1
+ uint8x8_t d17u8, // q2
+ uint8x8_t d18u8, // q3
+ uint8x8_t *d0ru8, // p1
+ uint8x8_t *d1ru8, // p1
+ uint8x8_t *d2ru8, // p0
+ uint8x8_t *d3ru8, // q0
+ uint8x8_t *d4ru8, // q1
+ uint8x8_t *d5ru8) { // q1+ uint32_t flat;
+ uint8x8_t d0u8, d1u8, d2u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8;
+ uint8x8_t d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
+ int16x8_t q15s16;
+ uint16x8_t q10u16, q14u16;
+ int8x8_t d21s8, d24s8, d25s8, d26s8, d28s8, d29s8, d30s8;
+
+ d19u8 = vabd_u8(d3u8, d4u8);
+ d20u8 = vabd_u8(d4u8, d5u8);
+ d21u8 = vabd_u8(d5u8, d6u8);
+ d22u8 = vabd_u8(d16u8, d7u8);
+ d23u8 = vabd_u8(d17u8, d16u8);
+ d24u8 = vabd_u8(d18u8, d17u8);
+
+ d19u8 = vmax_u8(d19u8, d20u8);
+ d20u8 = vmax_u8(d21u8, d22u8);
+
+ d25u8 = vabd_u8(d6u8, d4u8);
+
+ d23u8 = vmax_u8(d23u8, d24u8);
+
+ d26u8 = vabd_u8(d7u8, d17u8);
+
+ d19u8 = vmax_u8(d19u8, d20u8);
+
+ d24u8 = vabd_u8(d6u8, d7u8);
+ d27u8 = vabd_u8(d3u8, d6u8);
+ d28u8 = vabd_u8(d18u8, d7u8);
+
+ d19u8 = vmax_u8(d19u8, d23u8);
+
+ d23u8 = vabd_u8(d5u8, d16u8);
+ d24u8 = vqadd_u8(d24u8, d24u8);
+
+
+ d19u8 = vcge_u8(dlimit, d19u8);
+
+
+ d25u8 = vmax_u8(d25u8, d26u8);
+ d26u8 = vmax_u8(d27u8, d28u8);
+
+ d23u8 = vshr_n_u8(d23u8, 1);
+
+ d25u8 = vmax_u8(d25u8, d26u8);
+
+ d24u8 = vqadd_u8(d24u8, d23u8);
+
+ d20u8 = vmax_u8(d20u8, d25u8);
+
+ d23u8 = vdup_n_u8(1);
+ d24u8 = vcge_u8(dblimit, d24u8);
+
+ d21u8 = vcgt_u8(d21u8, dthresh);
+
+ d20u8 = vcge_u8(d23u8, d20u8);
+
+ d19u8 = vand_u8(d19u8, d24u8);
+
+ d23u8 = vcgt_u8(d22u8, dthresh);
+
+ d20u8 = vand_u8(d20u8, d19u8);
+
+ d22u8 = vdup_n_u8(0x80);
+
+ d23u8 = vorr_u8(d21u8, d23u8);
+
+ q10u16 = vcombine_u16(vreinterpret_u16_u8(d20u8),
+ vreinterpret_u16_u8(d21u8));
+
+ d30u8 = vshrn_n_u16(q10u16, 4);
+ flat = vget_lane_u32(vreinterpret_u32_u8(d30u8), 0);
+
+ if (flat == 0xffffffff) { // Check for all 1's, power_branch_only+ d27u8 = vdup_n_u8(3);
+ d21u8 = vdup_n_u8(2);
+ q14u16 = vaddl_u8(d6u8, d7u8);
+ q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+ q14u16 = vmlal_u8(q14u16, d4u8, d21u8);
+ q14u16 = vaddw_u8(q14u16, d5u8);
+ *d0ru8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d3u8);
+ q14u16 = vsubw_u8(q14u16, d4u8);
+ q14u16 = vaddw_u8(q14u16, d5u8);
+ q14u16 = vaddw_u8(q14u16, d16u8);
+ *d1ru8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d3u8);
+ q14u16 = vsubw_u8(q14u16, d5u8);
+ q14u16 = vaddw_u8(q14u16, d6u8);
+ q14u16 = vaddw_u8(q14u16, d17u8);
+ *d2ru8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d3u8);
+ q14u16 = vsubw_u8(q14u16, d6u8);
+ q14u16 = vaddw_u8(q14u16, d7u8);
+ q14u16 = vaddw_u8(q14u16, d18u8);
+ *d3ru8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d4u8);
+ q14u16 = vsubw_u8(q14u16, d7u8);
+ q14u16 = vaddw_u8(q14u16, d16u8);
+ q14u16 = vaddw_u8(q14u16, d18u8);
+ *d4ru8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d5u8);
+ q14u16 = vsubw_u8(q14u16, d16u8);
+ q14u16 = vaddw_u8(q14u16, d17u8);
+ q14u16 = vaddw_u8(q14u16, d18u8);
+ *d5ru8 = vqrshrn_n_u16(q14u16, 3);
+ } else {+ d21u8 = veor_u8(d7u8, d22u8);
+ d24u8 = veor_u8(d6u8, d22u8);
+ d25u8 = veor_u8(d5u8, d22u8);
+ d26u8 = veor_u8(d16u8, d22u8);
+
+ d27u8 = vdup_n_u8(3);
+
+ d28s8 = vsub_s8(vreinterpret_s8_u8(d21u8), vreinterpret_s8_u8(d24u8));
+ d29s8 = vqsub_s8(vreinterpret_s8_u8(d25u8), vreinterpret_s8_u8(d26u8));
+
+ q15s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d27u8));
+
+ d29s8 = vand_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+ q15s16 = vaddw_s8(q15s16, d29s8);
+
+ d29u8 = vdup_n_u8(4);
+
+ d28s8 = vqmovn_s16(q15s16);
+
+ d28s8 = vand_s8(d28s8, vreinterpret_s8_u8(d19u8));
+
+ d30s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d27u8));
+ d29s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d29u8));
+ d30s8 = vshr_n_s8(d30s8, 3);
+ d29s8 = vshr_n_s8(d29s8, 3);
+
+ d24s8 = vqadd_s8(vreinterpret_s8_u8(d24u8), d30s8);
+ d21s8 = vqsub_s8(vreinterpret_s8_u8(d21u8), d29s8);
+
+ d29s8 = vrshr_n_s8(d29s8, 1);
+ d29s8 = vbic_s8(d29s8, vreinterpret_s8_u8(d23u8));
+
+ d25s8 = vqadd_s8(vreinterpret_s8_u8(d25u8), d29s8);
+ d26s8 = vqsub_s8(vreinterpret_s8_u8(d26u8), d29s8);
+
+ if (flat == 0) { // filter_branch_only+ *d0ru8 = d4u8;
+ *d1ru8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+ *d2ru8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+ *d3ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+ *d4ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+ *d5ru8 = d17u8;
+ return;
+ }
+
+ d21u8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
+ d24u8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
+ d25u8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
+ d26u8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
+
+ d23u8 = vdup_n_u8(2);
+ q14u16 = vaddl_u8(d6u8, d7u8);
+ q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
+ q14u16 = vmlal_u8(q14u16, d4u8, d23u8);
+
+ d0u8 = vbsl_u8(d20u8, dblimit, d4u8);
+
+ q14u16 = vaddw_u8(q14u16, d5u8);
+
+ d1u8 = vbsl_u8(d20u8, dlimit, d25u8);
+
+ d30u8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d3u8);
+ q14u16 = vsubw_u8(q14u16, d4u8);
+ q14u16 = vaddw_u8(q14u16, d5u8);
+ q14u16 = vaddw_u8(q14u16, d16u8);
+
+ d2u8 = vbsl_u8(d20u8, dthresh, d24u8);
+
+ d31u8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d3u8);
+ q14u16 = vsubw_u8(q14u16, d5u8);
+ q14u16 = vaddw_u8(q14u16, d6u8);
+ q14u16 = vaddw_u8(q14u16, d17u8);
+
+ *d0ru8 = vbsl_u8(d20u8, d30u8, d0u8);
+
+ d23u8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d3u8);
+ q14u16 = vsubw_u8(q14u16, d6u8);
+ q14u16 = vaddw_u8(q14u16, d7u8);
+
+ *d1ru8 = vbsl_u8(d20u8, d31u8, d1u8);
+
+ q14u16 = vaddw_u8(q14u16, d18u8);
+
+ *d2ru8 = vbsl_u8(d20u8, d23u8, d2u8);
+
+ d22u8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d4u8);
+ q14u16 = vsubw_u8(q14u16, d7u8);
+ q14u16 = vaddw_u8(q14u16, d16u8);
+
+ d3u8 = vbsl_u8(d20u8, d3u8, d21u8);
+
+ q14u16 = vaddw_u8(q14u16, d18u8);
+
+ d4u8 = vbsl_u8(d20u8, d4u8, d26u8);
+
+ d6u8 = vqrshrn_n_u16(q14u16, 3);
+
+ q14u16 = vsubw_u8(q14u16, d5u8);
+ q14u16 = vsubw_u8(q14u16, d16u8);
+ q14u16 = vaddw_u8(q14u16, d17u8);
+ q14u16 = vaddw_u8(q14u16, d18u8);
+
+ d5u8 = vbsl_u8(d20u8, d5u8, d17u8);
+
+ d7u8 = vqrshrn_n_u16(q14u16, 3);
+
+ *d3ru8 = vbsl_u8(d20u8, d22u8, d3u8);
+ *d4ru8 = vbsl_u8(d20u8, d6u8, d4u8);
+ *d5ru8 = vbsl_u8(d20u8, d7u8, d5u8);
+ }
+ return;
+}
+
+void vp9_lpf_horizontal_8_neon(
+ uint8_t *src,
+ int pitch,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int count) {+ int i;
+ uint8_t *s, *psrc;
+ uint8x8_t dblimit, dlimit, dthresh;
+ uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+ uint8x8_t d16u8, d17u8, d18u8;
+
+ if (count == 0) // end_vp9_mblf_h_edge
+ return;
+
+ dblimit = vld1_u8(blimit);
+ dlimit = vld1_u8(limit);
+ dthresh = vld1_u8(thresh);
+
+ psrc = src - (pitch << 2);
+ for (i = 0; i < count; i++) {+ s = psrc + i * 8;
+
+ d3u8 = vld1_u8(s);
+ s += pitch;
+ d4u8 = vld1_u8(s);
+ s += pitch;
+ d5u8 = vld1_u8(s);
+ s += pitch;
+ d6u8 = vld1_u8(s);
+ s += pitch;
+ d7u8 = vld1_u8(s);
+ s += pitch;
+ d16u8 = vld1_u8(s);
+ s += pitch;
+ d17u8 = vld1_u8(s);
+ s += pitch;
+ d18u8 = vld1_u8(s);
+
+ vp9_mbloop_filter_neon(dblimit, dlimit, dthresh,
+ d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+ &d0u8, &d1u8, &d2u8, &d3u8, &d4u8, &d5u8);
+
+ s -= (pitch * 6);
+ vst1_u8(s, d0u8);
+ s += pitch;
+ vst1_u8(s, d1u8);
+ s += pitch;
+ vst1_u8(s, d2u8);
+ s += pitch;
+ vst1_u8(s, d3u8);
+ s += pitch;
+ vst1_u8(s, d4u8);
+ s += pitch;
+ vst1_u8(s, d5u8);
+ }
+ return;
+}
+
+void vp9_lpf_vertical_8_neon(
+ uint8_t *src,
+ int pitch,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int count) {+ int i;
+ uint8_t *s;
+ uint8x8_t dblimit, dlimit, dthresh;
+ uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
+ uint8x8_t d16u8, d17u8, d18u8;
+ uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
+ uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
+ uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
+ uint8x8x4_t d4Result;
+ uint8x8x2_t d2Result;
+
+ if (count == 0)
+ return;
+
+ dblimit = vld1_u8(blimit);
+ dlimit = vld1_u8(limit);
+ dthresh = vld1_u8(thresh);
+
+ for (i = 0; i < count; i++) {+ s = src + (i * (pitch << 3)) - 4;
+
+ d3u8 = vld1_u8(s);
+ s += pitch;
+ d4u8 = vld1_u8(s);
+ s += pitch;
+ d5u8 = vld1_u8(s);
+ s += pitch;
+ d6u8 = vld1_u8(s);
+ s += pitch;
+ d7u8 = vld1_u8(s);
+ s += pitch;
+ d16u8 = vld1_u8(s);
+ s += pitch;
+ d17u8 = vld1_u8(s);
+ s += pitch;
+ d18u8 = vld1_u8(s);
+
+ d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8),
+ vreinterpret_u32_u8(d7u8));
+ d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8),
+ vreinterpret_u32_u8(d16u8));
+ d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8),
+ vreinterpret_u32_u8(d17u8));
+ d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8),
+ vreinterpret_u32_u8(d18u8));
+
+ d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
+ vreinterpret_u16_u32(d2tmp2.val[0]));
+ d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
+ vreinterpret_u16_u32(d2tmp3.val[0]));
+ d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
+ vreinterpret_u16_u32(d2tmp2.val[1]));
+ d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
+ vreinterpret_u16_u32(d2tmp3.val[1]));
+
+ d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
+ vreinterpret_u8_u16(d2tmp5.val[0]));
+ d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
+ vreinterpret_u8_u16(d2tmp5.val[1]));
+ d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
+ vreinterpret_u8_u16(d2tmp7.val[0]));
+ d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
+ vreinterpret_u8_u16(d2tmp7.val[1]));
+
+ d3u8 = d2tmp8.val[0];
+ d4u8 = d2tmp8.val[1];
+ d5u8 = d2tmp9.val[0];
+ d6u8 = d2tmp9.val[1];
+ d7u8 = d2tmp10.val[0];
+ d16u8 = d2tmp10.val[1];
+ d17u8 = d2tmp11.val[0];
+ d18u8 = d2tmp11.val[1];
+
+ vp9_mbloop_filter_neon(dblimit, dlimit, dthresh,
+ d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8,
+ &d0u8, &d1u8, &d2u8, &d3u8, &d4u8, &d5u8);
+
+ d4Result.val[0] = d0u8;
+ d4Result.val[1] = d1u8;
+ d4Result.val[2] = d2u8;
+ d4Result.val[3] = d3u8;
+
+ d2Result.val[0] = d4u8;
+ d2Result.val[1] = d5u8;
+
+ s = src - 3;
+ vst4_lane_u8(s, d4Result, 0);
+ s += pitch;
+ vst4_lane_u8(s, d4Result, 1);
+ s += pitch;
+ vst4_lane_u8(s, d4Result, 2);
+ s += pitch;
+ vst4_lane_u8(s, d4Result, 3);
+ s += pitch;
+ vst4_lane_u8(s, d4Result, 4);
+ s += pitch;
+ vst4_lane_u8(s, d4Result, 5);
+ s += pitch;
+ vst4_lane_u8(s, d4Result, 6);
+ s += pitch;
+ vst4_lane_u8(s, d4Result, 7);
+
+ s = src + 1;
+ vst2_lane_u8(s, d2Result, 0);
+ s += pitch;
+ vst2_lane_u8(s, d2Result, 1);
+ s += pitch;
+ vst2_lane_u8(s, d2Result, 2);
+ s += pitch;
+ vst2_lane_u8(s, d2Result, 3);
+ s += pitch;
+ vst2_lane_u8(s, d2Result, 4);
+ s += pitch;
+ vst2_lane_u8(s, d2Result, 5);
+ s += pitch;
+ vst2_lane_u8(s, d2Result, 6);
+ s += pitch;
+ vst2_lane_u8(s, d2Result, 7);
+ }
+ return;
+}
--- /dev/null
+++ b/vpx_dsp/arm/loopfilter_mb_neon.asm
@@ -1,0 +1,606 @@
+;
+; Copyright (c) 2013 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+ EXPORT |vp9_lpf_horizontal_16_neon|
+ EXPORT |vp9_lpf_vertical_16_neon|
+ ARM
+
+ AREA ||.text||, CODE, READONLY, ALIGN=2
+
+; void vp9_lpf_horizontal_16_neon(uint8_t *s, int p,
+; const uint8_t *blimit,
+; const uint8_t *limit,
+; const uint8_t *thresh
+; int count)
+; r0 uint8_t *s,
+; r1 int p, /* pitch */
+; r2 const uint8_t *blimit,
+; r3 const uint8_t *limit,
+; sp const uint8_t *thresh,
+|vp9_lpf_horizontal_16_neon| PROC
+ push {r4-r8, lr}+ vpush {d8-d15}+ ldr r4, [sp, #88] ; load thresh
+ ldr r12, [sp, #92] ; load count
+
+h_count
+ vld1.8 {d16[]}, [r2] ; load *blimit+ vld1.8 {d17[]}, [r3] ; load *limit+ vld1.8 {d18[]}, [r4] ; load *thresh+
+ sub r8, r0, r1, lsl #3 ; move src pointer down by 8 lines
+
+ vld1.u8 {d0}, [r8@64], r1 ; p7+ vld1.u8 {d1}, [r8@64], r1 ; p6+ vld1.u8 {d2}, [r8@64], r1 ; p5+ vld1.u8 {d3}, [r8@64], r1 ; p4+ vld1.u8 {d4}, [r8@64], r1 ; p3+ vld1.u8 {d5}, [r8@64], r1 ; p2+ vld1.u8 {d6}, [r8@64], r1 ; p1+ vld1.u8 {d7}, [r8@64], r1 ; p0+ vld1.u8 {d8}, [r8@64], r1 ; q0+ vld1.u8 {d9}, [r8@64], r1 ; q1+ vld1.u8 {d10}, [r8@64], r1 ; q2+ vld1.u8 {d11}, [r8@64], r1 ; q3+ vld1.u8 {d12}, [r8@64], r1 ; q4+ vld1.u8 {d13}, [r8@64], r1 ; q5+ vld1.u8 {d14}, [r8@64], r1 ; q6+ vld1.u8 {d15}, [r8@64], r1 ; q7+
+ bl vp9_wide_mbfilter_neon
+
+ tst r7, #1
+ beq h_mbfilter
+
+ ; flat && mask were not set for any of the channels. Just store the values
+ ; from filter.
+ sub r8, r0, r1, lsl #1
+
+ vst1.u8 {d25}, [r8@64], r1 ; store op1+ vst1.u8 {d24}, [r8@64], r1 ; store op0+ vst1.u8 {d23}, [r8@64], r1 ; store oq0+ vst1.u8 {d26}, [r8@64], r1 ; store oq1+
+ b h_next
+
+h_mbfilter
+ tst r7, #2
+ beq h_wide_mbfilter
+
+ ; flat2 was not set for any of the channels. Just store the values from
+ ; mbfilter.
+ sub r8, r0, r1, lsl #1
+ sub r8, r8, r1
+
+ vst1.u8 {d18}, [r8@64], r1 ; store op2+ vst1.u8 {d19}, [r8@64], r1 ; store op1+ vst1.u8 {d20}, [r8@64], r1 ; store op0+ vst1.u8 {d21}, [r8@64], r1 ; store oq0+ vst1.u8 {d22}, [r8@64], r1 ; store oq1+ vst1.u8 {d23}, [r8@64], r1 ; store oq2+
+ b h_next
+
+h_wide_mbfilter
+ sub r8, r0, r1, lsl #3
+ add r8, r8, r1
+
+ vst1.u8 {d16}, [r8@64], r1 ; store op6+ vst1.u8 {d24}, [r8@64], r1 ; store op5+ vst1.u8 {d25}, [r8@64], r1 ; store op4+ vst1.u8 {d26}, [r8@64], r1 ; store op3+ vst1.u8 {d27}, [r8@64], r1 ; store op2+ vst1.u8 {d18}, [r8@64], r1 ; store op1+ vst1.u8 {d19}, [r8@64], r1 ; store op0+ vst1.u8 {d20}, [r8@64], r1 ; store oq0+ vst1.u8 {d21}, [r8@64], r1 ; store oq1+ vst1.u8 {d22}, [r8@64], r1 ; store oq2+ vst1.u8 {d23}, [r8@64], r1 ; store oq3+ vst1.u8 {d1}, [r8@64], r1 ; store oq4+ vst1.u8 {d2}, [r8@64], r1 ; store oq5+ vst1.u8 {d3}, [r8@64], r1 ; store oq6+
+h_next
+ add r0, r0, #8
+ subs r12, r12, #1
+ bne h_count
+
+ vpop {d8-d15}+ pop {r4-r8, pc}+
+ ENDP ; |vp9_lpf_horizontal_16_neon|
+
+; void vp9_lpf_vertical_16_neon(uint8_t *s, int p,
+; const uint8_t *blimit,
+; const uint8_t *limit,
+; const uint8_t *thresh)
+; r0 uint8_t *s,
+; r1 int p, /* pitch */
+; r2 const uint8_t *blimit,
+; r3 const uint8_t *limit,
+; sp const uint8_t *thresh,
+|vp9_lpf_vertical_16_neon| PROC
+ push {r4-r8, lr}+ vpush {d8-d15}+ ldr r4, [sp, #88] ; load thresh
+
+ vld1.8 {d16[]}, [r2] ; load *blimit+ vld1.8 {d17[]}, [r3] ; load *limit+ vld1.8 {d18[]}, [r4] ; load *thresh+
+ sub r8, r0, #8
+
+ vld1.8 {d0}, [r8@64], r1+ vld1.8 {d8}, [r0@64], r1+ vld1.8 {d1}, [r8@64], r1+ vld1.8 {d9}, [r0@64], r1+ vld1.8 {d2}, [r8@64], r1+ vld1.8 {d10}, [r0@64], r1+ vld1.8 {d3}, [r8@64], r1+ vld1.8 {d11}, [r0@64], r1+ vld1.8 {d4}, [r8@64], r1+ vld1.8 {d12}, [r0@64], r1+ vld1.8 {d5}, [r8@64], r1+ vld1.8 {d13}, [r0@64], r1+ vld1.8 {d6}, [r8@64], r1+ vld1.8 {d14}, [r0@64], r1+ vld1.8 {d7}, [r8@64], r1+ vld1.8 {d15}, [r0@64], r1+
+ sub r0, r0, r1, lsl #3
+
+ vtrn.32 q0, q2
+ vtrn.32 q1, q3
+ vtrn.32 q4, q6
+ vtrn.32 q5, q7
+
+ vtrn.16 q0, q1
+ vtrn.16 q2, q3
+ vtrn.16 q4, q5
+ vtrn.16 q6, q7
+
+ vtrn.8 d0, d1
+ vtrn.8 d2, d3
+ vtrn.8 d4, d5
+ vtrn.8 d6, d7
+
+ vtrn.8 d8, d9
+ vtrn.8 d10, d11
+ vtrn.8 d12, d13
+ vtrn.8 d14, d15
+
+ bl vp9_wide_mbfilter_neon
+
+ tst r7, #1
+ beq v_mbfilter
+
+ ; flat && mask were not set for any of the channels. Just store the values
+ ; from filter.
+ sub r8, r0, #2
+
+ vswp d23, d25
+
+ vst4.8 {d23[0], d24[0], d25[0], d26[0]}, [r8], r1+ vst4.8 {d23[1], d24[1], d25[1], d26[1]}, [r8], r1+ vst4.8 {d23[2], d24[2], d25[2], d26[2]}, [r8], r1+ vst4.8 {d23[3], d24[3], d25[3], d26[3]}, [r8], r1+ vst4.8 {d23[4], d24[4], d25[4], d26[4]}, [r8], r1+ vst4.8 {d23[5], d24[5], d25[5], d26[5]}, [r8], r1+ vst4.8 {d23[6], d24[6], d25[6], d26[6]}, [r8], r1+ vst4.8 {d23[7], d24[7], d25[7], d26[7]}, [r8], r1+
+ b v_end
+
+v_mbfilter
+ tst r7, #2
+ beq v_wide_mbfilter
+
+ ; flat2 was not set for any of the channels. Just store the values from
+ ; mbfilter.
+ sub r8, r0, #3
+
+ vst3.8 {d18[0], d19[0], d20[0]}, [r8], r1+ vst3.8 {d21[0], d22[0], d23[0]}, [r0], r1+ vst3.8 {d18[1], d19[1], d20[1]}, [r8], r1+ vst3.8 {d21[1], d22[1], d23[1]}, [r0], r1+ vst3.8 {d18[2], d19[2], d20[2]}, [r8], r1+ vst3.8 {d21[2], d22[2], d23[2]}, [r0], r1+ vst3.8 {d18[3], d19[3], d20[3]}, [r8], r1+ vst3.8 {d21[3], d22[3], d23[3]}, [r0], r1+ vst3.8 {d18[4], d19[4], d20[4]}, [r8], r1+ vst3.8 {d21[4], d22[4], d23[4]}, [r0], r1+ vst3.8 {d18[5], d19[5], d20[5]}, [r8], r1+ vst3.8 {d21[5], d22[5], d23[5]}, [r0], r1+ vst3.8 {d18[6], d19[6], d20[6]}, [r8], r1+ vst3.8 {d21[6], d22[6], d23[6]}, [r0], r1+ vst3.8 {d18[7], d19[7], d20[7]}, [r8], r1+ vst3.8 {d21[7], d22[7], d23[7]}, [r0], r1+
+ b v_end
+
+v_wide_mbfilter
+ sub r8, r0, #8
+
+ vtrn.32 d0, d26
+ vtrn.32 d16, d27
+ vtrn.32 d24, d18
+ vtrn.32 d25, d19
+
+ vtrn.16 d0, d24
+ vtrn.16 d16, d25
+ vtrn.16 d26, d18
+ vtrn.16 d27, d19
+
+ vtrn.8 d0, d16
+ vtrn.8 d24, d25
+ vtrn.8 d26, d27
+ vtrn.8 d18, d19
+
+ vtrn.32 d20, d1
+ vtrn.32 d21, d2
+ vtrn.32 d22, d3
+ vtrn.32 d23, d15
+
+ vtrn.16 d20, d22
+ vtrn.16 d21, d23
+ vtrn.16 d1, d3
+ vtrn.16 d2, d15
+
+ vtrn.8 d20, d21
+ vtrn.8 d22, d23
+ vtrn.8 d1, d2
+ vtrn.8 d3, d15
+
+ vst1.8 {d0}, [r8@64], r1+ vst1.8 {d20}, [r0@64], r1+ vst1.8 {d16}, [r8@64], r1+ vst1.8 {d21}, [r0@64], r1+ vst1.8 {d24}, [r8@64], r1+ vst1.8 {d22}, [r0@64], r1+ vst1.8 {d25}, [r8@64], r1+ vst1.8 {d23}, [r0@64], r1+ vst1.8 {d26}, [r8@64], r1+ vst1.8 {d1}, [r0@64], r1+ vst1.8 {d27}, [r8@64], r1+ vst1.8 {d2}, [r0@64], r1+ vst1.8 {d18}, [r8@64], r1+ vst1.8 {d3}, [r0@64], r1+ vst1.8 {d19}, [r8@64], r1+ vst1.8 {d15}, [r0@64], r1+
+v_end
+ vpop {d8-d15}+ pop {r4-r8, pc}+
+ ENDP ; |vp9_lpf_vertical_16_neon|
+
+; void vp9_wide_mbfilter_neon();
+; This is a helper function for the loopfilters. The invidual functions do the
+; necessary load, transpose (if necessary) and store.
+;
+; r0-r3 PRESERVE
+; d16 blimit
+; d17 limit
+; d18 thresh
+; d0 p7
+; d1 p6
+; d2 p5
+; d3 p4
+; d4 p3
+; d5 p2
+; d6 p1
+; d7 p0
+; d8 q0
+; d9 q1
+; d10 q2
+; d11 q3
+; d12 q4
+; d13 q5
+; d14 q6
+; d15 q7
+|vp9_wide_mbfilter_neon| PROC
+ mov r7, #0
+
+ ; filter_mask
+ vabd.u8 d19, d4, d5 ; abs(p3 - p2)
+ vabd.u8 d20, d5, d6 ; abs(p2 - p1)
+ vabd.u8 d21, d6, d7 ; abs(p1 - p0)
+ vabd.u8 d22, d9, d8 ; abs(q1 - q0)
+ vabd.u8 d23, d10, d9 ; abs(q2 - q1)
+ vabd.u8 d24, d11, d10 ; abs(q3 - q2)
+
+ ; only compare the largest value to limit
+ vmax.u8 d19, d19, d20 ; max(abs(p3 - p2), abs(p2 - p1))
+ vmax.u8 d20, d21, d22 ; max(abs(p1 - p0), abs(q1 - q0))
+ vmax.u8 d23, d23, d24 ; max(abs(q2 - q1), abs(q3 - q2))
+ vmax.u8 d19, d19, d20
+
+ vabd.u8 d24, d7, d8 ; abs(p0 - q0)
+
+ vmax.u8 d19, d19, d23
+
+ vabd.u8 d23, d6, d9 ; a = abs(p1 - q1)
+ vqadd.u8 d24, d24, d24 ; b = abs(p0 - q0) * 2
+
+ ; abs () > limit
+ vcge.u8 d19, d17, d19
+
+ ; flatmask4
+ vabd.u8 d25, d7, d5 ; abs(p0 - p2)
+ vabd.u8 d26, d8, d10 ; abs(q0 - q2)
+ vabd.u8 d27, d4, d7 ; abs(p3 - p0)
+ vabd.u8 d28, d11, d8 ; abs(q3 - q0)
+
+ ; only compare the largest value to thresh
+ vmax.u8 d25, d25, d26 ; max(abs(p0 - p2), abs(q0 - q2))
+ vmax.u8 d26, d27, d28 ; max(abs(p3 - p0), abs(q3 - q0))
+ vmax.u8 d25, d25, d26
+ vmax.u8 d20, d20, d25
+
+ vshr.u8 d23, d23, #1 ; a = a / 2
+ vqadd.u8 d24, d24, d23 ; a = b + a
+
+ vmov.u8 d30, #1
+ vcge.u8 d24, d16, d24 ; (a > blimit * 2 + limit) * -1
+
+ vcge.u8 d20, d30, d20 ; flat
+
+ vand d19, d19, d24 ; mask
+
+ ; hevmask
+ vcgt.u8 d21, d21, d18 ; (abs(p1 - p0) > thresh)*-1
+ vcgt.u8 d22, d22, d18 ; (abs(q1 - q0) > thresh)*-1
+ vorr d21, d21, d22 ; hev
+
+ vand d16, d20, d19 ; flat && mask
+ vmov r5, r6, d16
+
+ ; flatmask5(1, p7, p6, p5, p4, p0, q0, q4, q5, q6, q7)
+ vabd.u8 d22, d3, d7 ; abs(p4 - p0)
+ vabd.u8 d23, d12, d8 ; abs(q4 - q0)
+ vabd.u8 d24, d7, d2 ; abs(p0 - p5)
+ vabd.u8 d25, d8, d13 ; abs(q0 - q5)
+ vabd.u8 d26, d1, d7 ; abs(p6 - p0)
+ vabd.u8 d27, d14, d8 ; abs(q6 - q0)
+ vabd.u8 d28, d0, d7 ; abs(p7 - p0)
+ vabd.u8 d29, d15, d8 ; abs(q7 - q0)
+
+ ; only compare the largest value to thresh
+ vmax.u8 d22, d22, d23 ; max(abs(p4 - p0), abs(q4 - q0))
+ vmax.u8 d23, d24, d25 ; max(abs(p0 - p5), abs(q0 - q5))
+ vmax.u8 d24, d26, d27 ; max(abs(p6 - p0), abs(q6 - q0))
+ vmax.u8 d25, d28, d29 ; max(abs(p7 - p0), abs(q7 - q0))
+
+ vmax.u8 d26, d22, d23
+ vmax.u8 d27, d24, d25
+ vmax.u8 d23, d26, d27
+
+ vcge.u8 d18, d30, d23 ; flat2
+
+ vmov.u8 d22, #0x80
+
+ orrs r5, r5, r6 ; Check for 0
+ orreq r7, r7, #1 ; Only do filter branch
+
+ vand d17, d18, d16 ; flat2 && flat && mask
+ vmov r5, r6, d17
+
+ ; mbfilter() function
+
+ ; filter() function
+ ; convert to signed
+ veor d23, d8, d22 ; qs0
+ veor d24, d7, d22 ; ps0
+ veor d25, d6, d22 ; ps1
+ veor d26, d9, d22 ; qs1
+
+ vmov.u8 d27, #3
+
+ vsub.s8 d28, d23, d24 ; ( qs0 - ps0)
+ vqsub.s8 d29, d25, d26 ; filter = clamp(ps1-qs1)
+ vmull.s8 q15, d28, d27 ; 3 * ( qs0 - ps0)
+ vand d29, d29, d21 ; filter &= hev
+ vaddw.s8 q15, q15, d29 ; filter + 3 * (qs0 - ps0)
+ vmov.u8 d29, #4
+
+ ; filter = clamp(filter + 3 * ( qs0 - ps0))
+ vqmovn.s16 d28, q15
+
+ vand d28, d28, d19 ; filter &= mask
+
+ vqadd.s8 d30, d28, d27 ; filter2 = clamp(filter+3)
+ vqadd.s8 d29, d28, d29 ; filter1 = clamp(filter+4)
+ vshr.s8 d30, d30, #3 ; filter2 >>= 3
+ vshr.s8 d29, d29, #3 ; filter1 >>= 3
+
+
+ vqadd.s8 d24, d24, d30 ; op0 = clamp(ps0 + filter2)
+ vqsub.s8 d23, d23, d29 ; oq0 = clamp(qs0 - filter1)
+
+ ; outer tap adjustments: ++filter1 >> 1
+ vrshr.s8 d29, d29, #1
+ vbic d29, d29, d21 ; filter &= ~hev
+
+ vqadd.s8 d25, d25, d29 ; op1 = clamp(ps1 + filter)
+ vqsub.s8 d26, d26, d29 ; oq1 = clamp(qs1 - filter)
+
+ veor d24, d24, d22 ; *f_op0 = u^0x80
+ veor d23, d23, d22 ; *f_oq0 = u^0x80
+ veor d25, d25, d22 ; *f_op1 = u^0x80
+ veor d26, d26, d22 ; *f_oq1 = u^0x80
+
+ tst r7, #1
+ bxne lr
+
+ orrs r5, r5, r6 ; Check for 0
+ orreq r7, r7, #2 ; Only do mbfilter branch
+
+ ; mbfilter flat && mask branch
+ ; TODO(fgalligan): Can I decrease the cycles shifting to consective d's
+ ; and using vibt on the q's?
+ vmov.u8 d29, #2
+ vaddl.u8 q15, d7, d8 ; op2 = p0 + q0
+ vmlal.u8 q15, d4, d27 ; op2 = p0 + q0 + p3 * 3
+ vmlal.u8 q15, d5, d29 ; op2 = p0 + q0 + p3 * 3 + p2 * 2
+ vaddl.u8 q10, d4, d5
+ vaddw.u8 q15, d6 ; op2=p1 + p0 + q0 + p3 * 3 + p2 *2
+ vaddl.u8 q14, d6, d9
+ vqrshrn.u16 d18, q15, #3 ; r_op2
+
+ vsub.i16 q15, q10
+ vaddl.u8 q10, d4, d6
+ vadd.i16 q15, q14
+ vaddl.u8 q14, d7, d10
+ vqrshrn.u16 d19, q15, #3 ; r_op1
+
+ vsub.i16 q15, q10
+ vadd.i16 q15, q14
+ vaddl.u8 q14, d8, d11
+ vqrshrn.u16 d20, q15, #3 ; r_op0
+
+ vsubw.u8 q15, d4 ; oq0 = op0 - p3
+ vsubw.u8 q15, d7 ; oq0 -= p0
+ vadd.i16 q15, q14
+ vaddl.u8 q14, d9, d11
+ vqrshrn.u16 d21, q15, #3 ; r_oq0
+
+ vsubw.u8 q15, d5 ; oq1 = oq0 - p2
+ vsubw.u8 q15, d8 ; oq1 -= q0
+ vadd.i16 q15, q14
+ vaddl.u8 q14, d10, d11
+ vqrshrn.u16 d22, q15, #3 ; r_oq1
+
+ vsubw.u8 q15, d6 ; oq2 = oq0 - p1
+ vsubw.u8 q15, d9 ; oq2 -= q1
+ vadd.i16 q15, q14
+ vqrshrn.u16 d27, q15, #3 ; r_oq2
+
+ ; Filter does not set op2 or oq2, so use p2 and q2.
+ vbif d18, d5, d16 ; t_op2 |= p2 & ~(flat & mask)
+ vbif d19, d25, d16 ; t_op1 |= f_op1 & ~(flat & mask)
+ vbif d20, d24, d16 ; t_op0 |= f_op0 & ~(flat & mask)
+ vbif d21, d23, d16 ; t_oq0 |= f_oq0 & ~(flat & mask)
+ vbif d22, d26, d16 ; t_oq1 |= f_oq1 & ~(flat & mask)
+
+ vbit d23, d27, d16 ; t_oq2 |= r_oq2 & (flat & mask)
+ vbif d23, d10, d16 ; t_oq2 |= q2 & ~(flat & mask)
+
+ tst r7, #2
+ bxne lr
+
+ ; wide_mbfilter flat2 && flat && mask branch
+ vmov.u8 d16, #7
+ vaddl.u8 q15, d7, d8 ; op6 = p0 + q0
+ vaddl.u8 q12, d2, d3
+ vaddl.u8 q13, d4, d5
+ vaddl.u8 q14, d1, d6
+ vmlal.u8 q15, d0, d16 ; op6 += p7 * 3
+ vadd.i16 q12, q13
+ vadd.i16 q15, q14
+ vaddl.u8 q14, d2, d9
+ vadd.i16 q15, q12
+ vaddl.u8 q12, d0, d1
+ vaddw.u8 q15, d1
+ vaddl.u8 q13, d0, d2
+ vadd.i16 q14, q15, q14
+ vqrshrn.u16 d16, q15, #4 ; w_op6
+
+ vsub.i16 q15, q14, q12
+ vaddl.u8 q14, d3, d10
+ vqrshrn.u16 d24, q15, #4 ; w_op5
+
+ vsub.i16 q15, q13
+ vaddl.u8 q13, d0, d3
+ vadd.i16 q15, q14
+ vaddl.u8 q14, d4, d11
+ vqrshrn.u16 d25, q15, #4 ; w_op4
+
+ vadd.i16 q15, q14
+ vaddl.u8 q14, d0, d4
+ vsub.i16 q15, q13
+ vsub.i16 q14, q15, q14
+ vqrshrn.u16 d26, q15, #4 ; w_op3
+
+ vaddw.u8 q15, q14, d5 ; op2 += p2
+ vaddl.u8 q14, d0, d5
+ vaddw.u8 q15, d12 ; op2 += q4
+ vbif d26, d4, d17 ; op3 |= p3 & ~(f2 & f & m)
+ vqrshrn.u16 d27, q15, #4 ; w_op2
+
+ vsub.i16 q15, q14
+ vaddl.u8 q14, d0, d6
+ vaddw.u8 q15, d6 ; op1 += p1
+ vaddw.u8 q15, d13 ; op1 += q5
+ vbif d27, d18, d17 ; op2 |= t_op2 & ~(f2 & f & m)
+ vqrshrn.u16 d18, q15, #4 ; w_op1
+
+ vsub.i16 q15, q14
+ vaddl.u8 q14, d0, d7
+ vaddw.u8 q15, d7 ; op0 += p0
+ vaddw.u8 q15, d14 ; op0 += q6
+ vbif d18, d19, d17 ; op1 |= t_op1 & ~(f2 & f & m)
+ vqrshrn.u16 d19, q15, #4 ; w_op0
+
+ vsub.i16 q15, q14
+ vaddl.u8 q14, d1, d8
+ vaddw.u8 q15, d8 ; oq0 += q0
+ vaddw.u8 q15, d15 ; oq0 += q7
+ vbif d19, d20, d17 ; op0 |= t_op0 & ~(f2 & f & m)
+ vqrshrn.u16 d20, q15, #4 ; w_oq0
+
+ vsub.i16 q15, q14
+ vaddl.u8 q14, d2, d9
+ vaddw.u8 q15, d9 ; oq1 += q1
+ vaddl.u8 q4, d10, d15
+ vaddw.u8 q15, d15 ; oq1 += q7
+ vbif d20, d21, d17 ; oq0 |= t_oq0 & ~(f2 & f & m)
+ vqrshrn.u16 d21, q15, #4 ; w_oq1
+
+ vsub.i16 q15, q14
+ vaddl.u8 q14, d3, d10
+ vadd.i16 q15, q4
+ vaddl.u8 q4, d11, d15
+ vbif d21, d22, d17 ; oq1 |= t_oq1 & ~(f2 & f & m)
+ vqrshrn.u16 d22, q15, #4 ; w_oq2
+
+ vsub.i16 q15, q14
+ vaddl.u8 q14, d4, d11
+ vadd.i16 q15, q4
+ vaddl.u8 q4, d12, d15
+ vbif d22, d23, d17 ; oq2 |= t_oq2 & ~(f2 & f & m)
+ vqrshrn.u16 d23, q15, #4 ; w_oq3
+
+ vsub.i16 q15, q14
+ vaddl.u8 q14, d5, d12
+ vadd.i16 q15, q4
+ vaddl.u8 q4, d13, d15
+ vbif d16, d1, d17 ; op6 |= p6 & ~(f2 & f & m)
+ vqrshrn.u16 d1, q15, #4 ; w_oq4
+
+ vsub.i16 q15, q14
+ vaddl.u8 q14, d6, d13
+ vadd.i16 q15, q4
+ vaddl.u8 q4, d14, d15
+ vbif d24, d2, d17 ; op5 |= p5 & ~(f2 & f & m)
+ vqrshrn.u16 d2, q15, #4 ; w_oq5
+
+ vsub.i16 q15, q14
+ vbif d25, d3, d17 ; op4 |= p4 & ~(f2 & f & m)
+ vadd.i16 q15, q4
+ vbif d23, d11, d17 ; oq3 |= q3 & ~(f2 & f & m)
+ vqrshrn.u16 d3, q15, #4 ; w_oq6
+ vbif d1, d12, d17 ; oq4 |= q4 & ~(f2 & f & m)
+ vbif d2, d13, d17 ; oq5 |= q5 & ~(f2 & f & m)
+ vbif d3, d14, d17 ; oq6 |= q6 & ~(f2 & f & m)
+
+ bx lr
+ ENDP ; |vp9_wide_mbfilter_neon|
+
+ END
--- /dev/null
+++ b/vpx_dsp/arm/loopfilter_neon.c
@@ -1,0 +1,58 @@
+/*
+ * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "./vpx_config.h"
+#include "vpx/vpx_integer.h"
+
+void vp9_lpf_vertical_4_dual_neon(uint8_t *s, int p,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1) {+ vp9_lpf_vertical_4_neon(s, p, blimit0, limit0, thresh0, 1);
+ vp9_lpf_vertical_4_neon(s + 8 * p, p, blimit1, limit1, thresh1, 1);
+}
+
+#if HAVE_NEON_ASM
+void vp9_lpf_horizontal_8_dual_neon(uint8_t *s, int p /* pitch */,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1) {+ vp9_lpf_horizontal_8_neon(s, p, blimit0, limit0, thresh0, 1);
+ vp9_lpf_horizontal_8_neon(s + 8, p, blimit1, limit1, thresh1, 1);
+}
+
+void vp9_lpf_vertical_8_dual_neon(uint8_t *s, int p,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1) {+ vp9_lpf_vertical_8_neon(s, p, blimit0, limit0, thresh0, 1);
+ vp9_lpf_vertical_8_neon(s + 8 * p, p, blimit1, limit1, thresh1, 1);
+}
+
+void vp9_lpf_vertical_16_dual_neon(uint8_t *s, int p,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh) {+ vp9_lpf_vertical_16_neon(s, p, blimit, limit, thresh);
+ vp9_lpf_vertical_16_neon(s + 8 * p, p, blimit, limit, thresh);
+}
+#endif // HAVE_NEON_ASM
--- /dev/null
+++ b/vpx_dsp/loopfilter.c
@@ -1,0 +1,743 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+#include "vpx_dsp/vpx_dsp_common.h"
+#include "vpx_ports/mem.h"
+
+static INLINE int8_t signed_char_clamp(int t) {+ return (int8_t)clamp(t, -128, 127);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static INLINE int16_t signed_char_clamp_high(int t, int bd) {+ switch (bd) {+ case 10:
+ return (int16_t)clamp(t, -128*4, 128*4-1);
+ case 12:
+ return (int16_t)clamp(t, -128*16, 128*16-1);
+ case 8:
+ default:
+ return (int16_t)clamp(t, -128, 128-1);
+ }
+}
+#endif
+
+// should we apply any filter at all: 11111111 yes, 00000000 no
+static INLINE int8_t filter_mask(uint8_t limit, uint8_t blimit,
+ uint8_t p3, uint8_t p2,
+ uint8_t p1, uint8_t p0,
+ uint8_t q0, uint8_t q1,
+ uint8_t q2, uint8_t q3) {+ int8_t mask = 0;
+ mask |= (abs(p3 - p2) > limit) * -1;
+ mask |= (abs(p2 - p1) > limit) * -1;
+ mask |= (abs(p1 - p0) > limit) * -1;
+ mask |= (abs(q1 - q0) > limit) * -1;
+ mask |= (abs(q2 - q1) > limit) * -1;
+ mask |= (abs(q3 - q2) > limit) * -1;
+ mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ return ~mask;
+}
+
+static INLINE int8_t flat_mask4(uint8_t thresh,
+ uint8_t p3, uint8_t p2,
+ uint8_t p1, uint8_t p0,
+ uint8_t q0, uint8_t q1,
+ uint8_t q2, uint8_t q3) {+ int8_t mask = 0;
+ mask |= (abs(p1 - p0) > thresh) * -1;
+ mask |= (abs(q1 - q0) > thresh) * -1;
+ mask |= (abs(p2 - p0) > thresh) * -1;
+ mask |= (abs(q2 - q0) > thresh) * -1;
+ mask |= (abs(p3 - p0) > thresh) * -1;
+ mask |= (abs(q3 - q0) > thresh) * -1;
+ return ~mask;
+}
+
+static INLINE int8_t flat_mask5(uint8_t thresh,
+ uint8_t p4, uint8_t p3,
+ uint8_t p2, uint8_t p1,
+ uint8_t p0, uint8_t q0,
+ uint8_t q1, uint8_t q2,
+ uint8_t q3, uint8_t q4) {+ int8_t mask = ~flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3);
+ mask |= (abs(p4 - p0) > thresh) * -1;
+ mask |= (abs(q4 - q0) > thresh) * -1;
+ return ~mask;
+}
+
+// is there high edge variance internal edge: 11111111 yes, 00000000 no
+static INLINE int8_t hev_mask(uint8_t thresh, uint8_t p1, uint8_t p0,
+ uint8_t q0, uint8_t q1) {+ int8_t hev = 0;
+ hev |= (abs(p1 - p0) > thresh) * -1;
+ hev |= (abs(q1 - q0) > thresh) * -1;
+ return hev;
+}
+
+static INLINE void filter4(int8_t mask, uint8_t thresh, uint8_t *op1,
+ uint8_t *op0, uint8_t *oq0, uint8_t *oq1) {+ int8_t filter1, filter2;
+
+ const int8_t ps1 = (int8_t) *op1 ^ 0x80;
+ const int8_t ps0 = (int8_t) *op0 ^ 0x80;
+ const int8_t qs0 = (int8_t) *oq0 ^ 0x80;
+ const int8_t qs1 = (int8_t) *oq1 ^ 0x80;
+ const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1);
+
+ // add outer taps if we have high edge variance
+ int8_t filter = signed_char_clamp(ps1 - qs1) & hev;
+
+ // inner taps
+ filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask;
+
+ // save bottom 3 bits so that we round one side +4 and the other +3
+ // if it equals 4 we'll set to adjust by -1 to account for the fact
+ // we'd round 3 the other way
+ filter1 = signed_char_clamp(filter + 4) >> 3;
+ filter2 = signed_char_clamp(filter + 3) >> 3;
+
+ *oq0 = signed_char_clamp(qs0 - filter1) ^ 0x80;
+ *op0 = signed_char_clamp(ps0 + filter2) ^ 0x80;
+
+ // outer tap adjustments
+ filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+ *oq1 = signed_char_clamp(qs1 - filter) ^ 0x80;
+ *op1 = signed_char_clamp(ps1 + filter) ^ 0x80;
+}
+
+void vp9_lpf_horizontal_4_c(uint8_t *s, int p /* pitch */,
+ const uint8_t *blimit, const uint8_t *limit,
+ const uint8_t *thresh, int count) {+ int i;
+
+ // loop filter designed to work using chars so that we can make maximum use
+ // of 8 bit simd instructions.
+ for (i = 0; i < 8 * count; ++i) {+ const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+ const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p);
+ ++s;
+ }
+}
+
+void vp9_lpf_horizontal_4_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
+ const uint8_t *limit0, const uint8_t *thresh0,
+ const uint8_t *blimit1, const uint8_t *limit1,
+ const uint8_t *thresh1) {+ vp9_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1);
+ vp9_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1);
+}
+
+void vp9_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int count) {+ int i;
+
+ // loop filter designed to work using chars so that we can make maximum use
+ // of 8 bit simd instructions.
+ for (i = 0; i < 8 * count; ++i) {+ const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+ const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ filter4(mask, *thresh, s - 2, s - 1, s, s + 1);
+ s += pitch;
+ }
+}
+
+void vp9_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
+ const uint8_t *limit0, const uint8_t *thresh0,
+ const uint8_t *blimit1, const uint8_t *limit1,
+ const uint8_t *thresh1) {+ vp9_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1);
+ vp9_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
+ thresh1, 1);
+}
+
+static INLINE void filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+ uint8_t *op3, uint8_t *op2,
+ uint8_t *op1, uint8_t *op0,
+ uint8_t *oq0, uint8_t *oq1,
+ uint8_t *oq2, uint8_t *oq3) {+ if (flat && mask) {+ const uint8_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+ const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+ // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+ *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+ *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+ *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+ *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+ *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+ *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+ } else {+ filter4(mask, thresh, op1, op0, oq0, oq1);
+ }
+}
+
+void vp9_lpf_horizontal_8_c(uint8_t *s, int p, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int count) {+ int i;
+
+ // loop filter designed to work using chars so that we can make maximum use
+ // of 8 bit simd instructions.
+ for (i = 0; i < 8 * count; ++i) {+ const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+ const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+ filter8(mask, *thresh, flat, s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+ s, s + 1 * p, s + 2 * p, s + 3 * p);
+ ++s;
+ }
+}
+
+void vp9_lpf_horizontal_8_dual_c(uint8_t *s, int p, const uint8_t *blimit0,
+ const uint8_t *limit0, const uint8_t *thresh0,
+ const uint8_t *blimit1, const uint8_t *limit1,
+ const uint8_t *thresh1) {+ vp9_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1);
+ vp9_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1);
+}
+
+void vp9_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int count) {+ int i;
+
+ for (i = 0; i < 8 * count; ++i) {+ const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+ const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+ filter8(mask, *thresh, flat, s - 4, s - 3, s - 2, s - 1,
+ s, s + 1, s + 2, s + 3);
+ s += pitch;
+ }
+}
+
+void vp9_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0,
+ const uint8_t *limit0, const uint8_t *thresh0,
+ const uint8_t *blimit1, const uint8_t *limit1,
+ const uint8_t *thresh1) {+ vp9_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1);
+ vp9_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
+ thresh1, 1);
+}
+
+static INLINE void filter16(int8_t mask, uint8_t thresh,
+ uint8_t flat, uint8_t flat2,
+ uint8_t *op7, uint8_t *op6,
+ uint8_t *op5, uint8_t *op4,
+ uint8_t *op3, uint8_t *op2,
+ uint8_t *op1, uint8_t *op0,
+ uint8_t *oq0, uint8_t *oq1,
+ uint8_t *oq2, uint8_t *oq3,
+ uint8_t *oq4, uint8_t *oq5,
+ uint8_t *oq6, uint8_t *oq7) {+ if (flat2 && flat && mask) {+ const uint8_t p7 = *op7, p6 = *op6, p5 = *op5, p4 = *op4,
+ p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+
+ const uint8_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3,
+ q4 = *oq4, q5 = *oq5, q6 = *oq6, q7 = *oq7;
+
+ // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+ *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
+ q0, 4);
+ *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
+ q0 + q1, 4);
+ *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
+ q0 + q1 + q2, 4);
+ *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
+ q0 + q1 + q2 + q3, 4);
+ *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
+ q0 + q1 + q2 + q3 + q4, 4);
+ *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+ q0 + q1 + q2 + q3 + q4 + q5, 4);
+ *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+ q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
+ *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
+ q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
+ *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
+ q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
+ *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
+ q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
+ *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
+ q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+ *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
+ q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+ *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
+ q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+ *oq6 = ROUND_POWER_OF_TWO(p0 +
+ q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+ } else {+ filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3);
+ }
+}
+
+void vp9_lpf_horizontal_16_c(uint8_t *s, int p, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int count) {+ int i;
+
+ // loop filter designed to work using chars so that we can make maximum use
+ // of 8 bit simd instructions.
+ for (i = 0; i < 8 * count; ++i) {+ const uint8_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+ const uint8_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t flat2 = flat_mask5(1,
+ s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
+ q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p]);
+
+ filter16(mask, *thresh, flat, flat2,
+ s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
+ s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+ s, s + 1 * p, s + 2 * p, s + 3 * p,
+ s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p);
+ ++s;
+ }
+}
+
+static void mb_lpf_vertical_edge_w(uint8_t *s, int p,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int count) {+ int i;
+
+ for (i = 0; i < count; ++i) {+ const uint8_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+ const uint8_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+ const int8_t mask = filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t flat = flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3);
+ const int8_t flat2 = flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+ q0, s[4], s[5], s[6], s[7]);
+
+ filter16(mask, *thresh, flat, flat2,
+ s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
+ s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7);
+ s += p;
+ }
+}
+
+void vp9_lpf_vertical_16_c(uint8_t *s, int p, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh) {+ mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8);
+}
+
+void vp9_lpf_vertical_16_dual_c(uint8_t *s, int p, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh) {+ mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+// Should we apply any filter at all: 11111111 yes, 00000000 no ?
+static INLINE int8_t highbd_filter_mask(uint8_t limit, uint8_t blimit,
+ uint16_t p3, uint16_t p2,
+ uint16_t p1, uint16_t p0,
+ uint16_t q0, uint16_t q1,
+ uint16_t q2, uint16_t q3, int bd) {+ int8_t mask = 0;
+ int16_t limit16 = (uint16_t)limit << (bd - 8);
+ int16_t blimit16 = (uint16_t)blimit << (bd - 8);
+ mask |= (abs(p3 - p2) > limit16) * -1;
+ mask |= (abs(p2 - p1) > limit16) * -1;
+ mask |= (abs(p1 - p0) > limit16) * -1;
+ mask |= (abs(q1 - q0) > limit16) * -1;
+ mask |= (abs(q2 - q1) > limit16) * -1;
+ mask |= (abs(q3 - q2) > limit16) * -1;
+ mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit16) * -1;
+ return ~mask;
+}
+
+static INLINE int8_t highbd_flat_mask4(uint8_t thresh,
+ uint16_t p3, uint16_t p2,
+ uint16_t p1, uint16_t p0,
+ uint16_t q0, uint16_t q1,
+ uint16_t q2, uint16_t q3, int bd) {+ int8_t mask = 0;
+ int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+ mask |= (abs(p1 - p0) > thresh16) * -1;
+ mask |= (abs(q1 - q0) > thresh16) * -1;
+ mask |= (abs(p2 - p0) > thresh16) * -1;
+ mask |= (abs(q2 - q0) > thresh16) * -1;
+ mask |= (abs(p3 - p0) > thresh16) * -1;
+ mask |= (abs(q3 - q0) > thresh16) * -1;
+ return ~mask;
+}
+
+static INLINE int8_t highbd_flat_mask5(uint8_t thresh,
+ uint16_t p4, uint16_t p3,
+ uint16_t p2, uint16_t p1,
+ uint16_t p0, uint16_t q0,
+ uint16_t q1, uint16_t q2,
+ uint16_t q3, uint16_t q4, int bd) {+ int8_t mask = ~highbd_flat_mask4(thresh, p3, p2, p1, p0, q0, q1, q2, q3, bd);
+ int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+ mask |= (abs(p4 - p0) > thresh16) * -1;
+ mask |= (abs(q4 - q0) > thresh16) * -1;
+ return ~mask;
+}
+
+// Is there high edge variance internal edge:
+// 11111111_11111111 yes, 00000000_00000000 no ?
+static INLINE int16_t highbd_hev_mask(uint8_t thresh, uint16_t p1, uint16_t p0,
+ uint16_t q0, uint16_t q1, int bd) {+ int16_t hev = 0;
+ int16_t thresh16 = (uint16_t)thresh << (bd - 8);
+ hev |= (abs(p1 - p0) > thresh16) * -1;
+ hev |= (abs(q1 - q0) > thresh16) * -1;
+ return hev;
+}
+
+static INLINE void highbd_filter4(int8_t mask, uint8_t thresh, uint16_t *op1,
+ uint16_t *op0, uint16_t *oq0, uint16_t *oq1,
+ int bd) {+ int16_t filter1, filter2;
+ // ^0x80 equivalent to subtracting 0x80 from the values to turn them
+ // into -128 to +127 instead of 0 to 255.
+ int shift = bd - 8;
+ const int16_t ps1 = (int16_t)*op1 - (0x80 << shift);
+ const int16_t ps0 = (int16_t)*op0 - (0x80 << shift);
+ const int16_t qs0 = (int16_t)*oq0 - (0x80 << shift);
+ const int16_t qs1 = (int16_t)*oq1 - (0x80 << shift);
+ const uint16_t hev = highbd_hev_mask(thresh, *op1, *op0, *oq0, *oq1, bd);
+
+ // Add outer taps if we have high edge variance.
+ int16_t filter = signed_char_clamp_high(ps1 - qs1, bd) & hev;
+
+ // Inner taps.
+ filter = signed_char_clamp_high(filter + 3 * (qs0 - ps0), bd) & mask;
+
+ // Save bottom 3 bits so that we round one side +4 and the other +3
+ // if it equals 4 we'll set to adjust by -1 to account for the fact
+ // we'd round 3 the other way.
+ filter1 = signed_char_clamp_high(filter + 4, bd) >> 3;
+ filter2 = signed_char_clamp_high(filter + 3, bd) >> 3;
+
+ *oq0 = signed_char_clamp_high(qs0 - filter1, bd) + (0x80 << shift);
+ *op0 = signed_char_clamp_high(ps0 + filter2, bd) + (0x80 << shift);
+
+ // Outer tap adjustments.
+ filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+
+ *oq1 = signed_char_clamp_high(qs1 - filter, bd) + (0x80 << shift);
+ *op1 = signed_char_clamp_high(ps1 + filter, bd) + (0x80 << shift);
+}
+
+void vp9_highbd_lpf_horizontal_4_c(uint16_t *s, int p /* pitch */,
+ const uint8_t *blimit, const uint8_t *limit,
+ const uint8_t *thresh, int count, int bd) {+ int i;
+
+ // loop filter designed to work using chars so that we can make maximum use
+ // of 8 bit simd instructions.
+ for (i = 0; i < 8 * count; ++i) {+ const uint16_t p3 = s[-4 * p];
+ const uint16_t p2 = s[-3 * p];
+ const uint16_t p1 = s[-2 * p];
+ const uint16_t p0 = s[-p];
+ const uint16_t q0 = s[0 * p];
+ const uint16_t q1 = s[1 * p];
+ const uint16_t q2 = s[2 * p];
+ const uint16_t q3 = s[3 * p];
+ const int8_t mask = highbd_filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3, bd);
+ highbd_filter4(mask, *thresh, s - 2 * p, s - 1 * p, s, s + 1 * p, bd);
+ ++s;
+ }
+}
+
+void vp9_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int p,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1,
+ int bd) {+ vp9_highbd_lpf_horizontal_4_c(s, p, blimit0, limit0, thresh0, 1, bd);
+ vp9_highbd_lpf_horizontal_4_c(s + 8, p, blimit1, limit1, thresh1, 1, bd);
+}
+
+void vp9_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int count, int bd) {+ int i;
+
+ // loop filter designed to work using chars so that we can make maximum use
+ // of 8 bit simd instructions.
+ for (i = 0; i < 8 * count; ++i) {+ const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+ const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+ const int8_t mask = highbd_filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3, bd);
+ highbd_filter4(mask, *thresh, s - 2, s - 1, s, s + 1, bd);
+ s += pitch;
+ }
+}
+
+void vp9_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1,
+ int bd) {+ vp9_highbd_lpf_vertical_4_c(s, pitch, blimit0, limit0, thresh0, 1, bd);
+ vp9_highbd_lpf_vertical_4_c(s + 8 * pitch, pitch, blimit1, limit1,
+ thresh1, 1, bd);
+}
+
+static INLINE void highbd_filter8(int8_t mask, uint8_t thresh, uint8_t flat,
+ uint16_t *op3, uint16_t *op2,
+ uint16_t *op1, uint16_t *op0,
+ uint16_t *oq0, uint16_t *oq1,
+ uint16_t *oq2, uint16_t *oq3, int bd) {+ if (flat && mask) {+ const uint16_t p3 = *op3, p2 = *op2, p1 = *op1, p0 = *op0;
+ const uint16_t q0 = *oq0, q1 = *oq1, q2 = *oq2, q3 = *oq3;
+
+ // 7-tap filter [1, 1, 1, 2, 1, 1, 1]
+ *op2 = ROUND_POWER_OF_TWO(p3 + p3 + p3 + 2 * p2 + p1 + p0 + q0, 3);
+ *op1 = ROUND_POWER_OF_TWO(p3 + p3 + p2 + 2 * p1 + p0 + q0 + q1, 3);
+ *op0 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + 2 * p0 + q0 + q1 + q2, 3);
+ *oq0 = ROUND_POWER_OF_TWO(p2 + p1 + p0 + 2 * q0 + q1 + q2 + q3, 3);
+ *oq1 = ROUND_POWER_OF_TWO(p1 + p0 + q0 + 2 * q1 + q2 + q3 + q3, 3);
+ *oq2 = ROUND_POWER_OF_TWO(p0 + q0 + q1 + 2 * q2 + q3 + q3 + q3, 3);
+ } else {+ highbd_filter4(mask, thresh, op1, op0, oq0, oq1, bd);
+ }
+}
+
+void vp9_highbd_lpf_horizontal_8_c(uint16_t *s, int p, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int count, int bd) {+ int i;
+
+ // loop filter designed to work using chars so that we can make maximum use
+ // of 8 bit simd instructions.
+ for (i = 0; i < 8 * count; ++i) {+ const uint16_t p3 = s[-4 * p], p2 = s[-3 * p], p1 = s[-2 * p], p0 = s[-p];
+ const uint16_t q0 = s[0 * p], q1 = s[1 * p], q2 = s[2 * p], q3 = s[3 * p];
+
+ const int8_t mask = highbd_filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3, bd);
+ const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+ bd);
+ highbd_filter8(mask, *thresh, flat,
+ s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+ s, s + 1 * p, s + 2 * p, s + 3 * p, bd);
+ ++s;
+ }
+}
+
+void vp9_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int p,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1,
+ int bd) {+ vp9_highbd_lpf_horizontal_8_c(s, p, blimit0, limit0, thresh0, 1, bd);
+ vp9_highbd_lpf_horizontal_8_c(s + 8, p, blimit1, limit1, thresh1, 1, bd);
+}
+
+void vp9_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int count, int bd) {+ int i;
+
+ for (i = 0; i < 8 * count; ++i) {+ const uint16_t p3 = s[-4], p2 = s[-3], p1 = s[-2], p0 = s[-1];
+ const uint16_t q0 = s[0], q1 = s[1], q2 = s[2], q3 = s[3];
+ const int8_t mask = highbd_filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3, bd);
+ const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+ bd);
+ highbd_filter8(mask, *thresh, flat,
+ s - 4, s - 3, s - 2, s - 1,
+ s, s + 1, s + 2, s + 3,
+ bd);
+ s += pitch;
+ }
+}
+
+void vp9_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1,
+ int bd) {+ vp9_highbd_lpf_vertical_8_c(s, pitch, blimit0, limit0, thresh0, 1, bd);
+ vp9_highbd_lpf_vertical_8_c(s + 8 * pitch, pitch, blimit1, limit1,
+ thresh1, 1, bd);
+}
+
+static INLINE void highbd_filter16(int8_t mask, uint8_t thresh,
+ uint8_t flat, uint8_t flat2,
+ uint16_t *op7, uint16_t *op6,
+ uint16_t *op5, uint16_t *op4,
+ uint16_t *op3, uint16_t *op2,
+ uint16_t *op1, uint16_t *op0,
+ uint16_t *oq0, uint16_t *oq1,
+ uint16_t *oq2, uint16_t *oq3,
+ uint16_t *oq4, uint16_t *oq5,
+ uint16_t *oq6, uint16_t *oq7, int bd) {+ if (flat2 && flat && mask) {+ const uint16_t p7 = *op7;
+ const uint16_t p6 = *op6;
+ const uint16_t p5 = *op5;
+ const uint16_t p4 = *op4;
+ const uint16_t p3 = *op3;
+ const uint16_t p2 = *op2;
+ const uint16_t p1 = *op1;
+ const uint16_t p0 = *op0;
+ const uint16_t q0 = *oq0;
+ const uint16_t q1 = *oq1;
+ const uint16_t q2 = *oq2;
+ const uint16_t q3 = *oq3;
+ const uint16_t q4 = *oq4;
+ const uint16_t q5 = *oq5;
+ const uint16_t q6 = *oq6;
+ const uint16_t q7 = *oq7;
+
+ // 15-tap filter [1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1]
+ *op6 = ROUND_POWER_OF_TWO(p7 * 7 + p6 * 2 + p5 + p4 + p3 + p2 + p1 + p0 +
+ q0, 4);
+ *op5 = ROUND_POWER_OF_TWO(p7 * 6 + p6 + p5 * 2 + p4 + p3 + p2 + p1 + p0 +
+ q0 + q1, 4);
+ *op4 = ROUND_POWER_OF_TWO(p7 * 5 + p6 + p5 + p4 * 2 + p3 + p2 + p1 + p0 +
+ q0 + q1 + q2, 4);
+ *op3 = ROUND_POWER_OF_TWO(p7 * 4 + p6 + p5 + p4 + p3 * 2 + p2 + p1 + p0 +
+ q0 + q1 + q2 + q3, 4);
+ *op2 = ROUND_POWER_OF_TWO(p7 * 3 + p6 + p5 + p4 + p3 + p2 * 2 + p1 + p0 +
+ q0 + q1 + q2 + q3 + q4, 4);
+ *op1 = ROUND_POWER_OF_TWO(p7 * 2 + p6 + p5 + p4 + p3 + p2 + p1 * 2 + p0 +
+ q0 + q1 + q2 + q3 + q4 + q5, 4);
+ *op0 = ROUND_POWER_OF_TWO(p7 + p6 + p5 + p4 + p3 + p2 + p1 + p0 * 2 +
+ q0 + q1 + q2 + q3 + q4 + q5 + q6, 4);
+ *oq0 = ROUND_POWER_OF_TWO(p6 + p5 + p4 + p3 + p2 + p1 + p0 +
+ q0 * 2 + q1 + q2 + q3 + q4 + q5 + q6 + q7, 4);
+ *oq1 = ROUND_POWER_OF_TWO(p5 + p4 + p3 + p2 + p1 + p0 +
+ q0 + q1 * 2 + q2 + q3 + q4 + q5 + q6 + q7 * 2, 4);
+ *oq2 = ROUND_POWER_OF_TWO(p4 + p3 + p2 + p1 + p0 +
+ q0 + q1 + q2 * 2 + q3 + q4 + q5 + q6 + q7 * 3, 4);
+ *oq3 = ROUND_POWER_OF_TWO(p3 + p2 + p1 + p0 +
+ q0 + q1 + q2 + q3 * 2 + q4 + q5 + q6 + q7 * 4, 4);
+ *oq4 = ROUND_POWER_OF_TWO(p2 + p1 + p0 +
+ q0 + q1 + q2 + q3 + q4 * 2 + q5 + q6 + q7 * 5, 4);
+ *oq5 = ROUND_POWER_OF_TWO(p1 + p0 +
+ q0 + q1 + q2 + q3 + q4 + q5 * 2 + q6 + q7 * 6, 4);
+ *oq6 = ROUND_POWER_OF_TWO(p0 +
+ q0 + q1 + q2 + q3 + q4 + q5 + q6 * 2 + q7 * 7, 4);
+ } else {+ highbd_filter8(mask, thresh, flat, op3, op2, op1, op0, oq0, oq1, oq2, oq3,
+ bd);
+ }
+}
+
+void vp9_highbd_lpf_horizontal_16_c(uint16_t *s, int p, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int count, int bd) {+ int i;
+
+ // loop filter designed to work using chars so that we can make maximum use
+ // of 8 bit simd instructions.
+ for (i = 0; i < 8 * count; ++i) {+ const uint16_t p3 = s[-4 * p];
+ const uint16_t p2 = s[-3 * p];
+ const uint16_t p1 = s[-2 * p];
+ const uint16_t p0 = s[-p];
+ const uint16_t q0 = s[0 * p];
+ const uint16_t q1 = s[1 * p];
+ const uint16_t q2 = s[2 * p];
+ const uint16_t q3 = s[3 * p];
+ const int8_t mask = highbd_filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3, bd);
+ const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+ bd);
+ const int8_t flat2 = highbd_flat_mask5(
+ 1, s[-8 * p], s[-7 * p], s[-6 * p], s[-5 * p], p0,
+ q0, s[4 * p], s[5 * p], s[6 * p], s[7 * p], bd);
+
+ highbd_filter16(mask, *thresh, flat, flat2,
+ s - 8 * p, s - 7 * p, s - 6 * p, s - 5 * p,
+ s - 4 * p, s - 3 * p, s - 2 * p, s - 1 * p,
+ s, s + 1 * p, s + 2 * p, s + 3 * p,
+ s + 4 * p, s + 5 * p, s + 6 * p, s + 7 * p,
+ bd);
+ ++s;
+ }
+}
+
+static void highbd_mb_lpf_vertical_edge_w(uint16_t *s, int p,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int count, int bd) {+ int i;
+
+ for (i = 0; i < count; ++i) {+ const uint16_t p3 = s[-4];
+ const uint16_t p2 = s[-3];
+ const uint16_t p1 = s[-2];
+ const uint16_t p0 = s[-1];
+ const uint16_t q0 = s[0];
+ const uint16_t q1 = s[1];
+ const uint16_t q2 = s[2];
+ const uint16_t q3 = s[3];
+ const int8_t mask = highbd_filter_mask(*limit, *blimit,
+ p3, p2, p1, p0, q0, q1, q2, q3, bd);
+ const int8_t flat = highbd_flat_mask4(1, p3, p2, p1, p0, q0, q1, q2, q3,
+ bd);
+ const int8_t flat2 = highbd_flat_mask5(1, s[-8], s[-7], s[-6], s[-5], p0,
+ q0, s[4], s[5], s[6], s[7], bd);
+
+ highbd_filter16(mask, *thresh, flat, flat2,
+ s - 8, s - 7, s - 6, s - 5, s - 4, s - 3, s - 2, s - 1,
+ s, s + 1, s + 2, s + 3, s + 4, s + 5, s + 6, s + 7,
+ bd);
+ s += p;
+ }
+}
+
+void vp9_highbd_lpf_vertical_16_c(uint16_t *s, int p, const uint8_t *blimit,
+ const uint8_t *limit, const uint8_t *thresh,
+ int bd) {+ highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 8, bd);
+}
+
+void vp9_highbd_lpf_vertical_16_dual_c(uint16_t *s, int p,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int bd) {+ highbd_mb_lpf_vertical_edge_w(s, p, blimit, limit, thresh, 16, bd);
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
--- /dev/null
+++ b/vpx_dsp/mips/loopfilter_16_msa.c
@@ -1,0 +1,1480 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_ports/mem.h"
+#include "vpx_dsp/mips/loopfilter_msa.h"
+
+int32_t vp9_hz_lpf_t4_and_t8_16w(uint8_t *src, int32_t pitch,
+ uint8_t *filter48,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr) {+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+ v16u8 flat, mask, hev, thresh, b_limit, limit;
+ v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+ v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+ v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+ v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+ v16u8 zero = { 0 };+
+ /* load vector elements */
+ LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+ b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+ limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+ /* mask and hev */
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+ VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+ if (__msa_test_bz_v(flat)) {+ ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
+
+ return 1;
+ } else {+ ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+ zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
+ q2_r, q3_r);
+ VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+ p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+ ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+ ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+ VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+ p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+ /* convert 16 bit output data into 8 bit */
+ PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+ p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+ p0_filt8_r, q0_filt8_r);
+ PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+ q2_filt8_r);
+
+ /* store pixel values */
+ p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+ p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+ p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+ q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+ q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+ q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+ ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+ filter48 += (4 * 16);
+ ST_UB2(q1_out, q2_out, filter48, 16);
+ filter48 += (2 * 16);
+ ST_UB(flat, filter48);
+
+ return 0;
+ }
+}
+
+void vp9_hz_lpf_t16_16w(uint8_t *src, int32_t pitch, uint8_t *filter48) {+ v16u8 flat, flat2, filter8;
+ v16i8 zero = { 0 };+ v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+ v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+ v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+ v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in;
+ v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in;
+ v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l;
+ v8i16 l_out, r_out;
+
+ flat = LD_UB(filter48 + 96);
+
+ LD_UB8((src - 8 * pitch), pitch, p7, p6, p5, p4, p3, p2, p1, p0);
+ LD_UB8(src, pitch, q0, q1, q2, q3, q4, q5, q6, q7);
+ VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+ if (__msa_test_bz_v(flat2)) {+ LD_UB4(filter48, 16, p2, p1, p0, q0);
+ LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+ src -= 3 * pitch;
+ ST_UB4(p2, p1, p0, q0, src, pitch);
+ src += (4 * pitch);
+ ST_UB2(q1, q2, src, pitch);
+ } else {+ src -= 7 * pitch;
+
+ ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
+ zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in,
+ p2_r_in, p1_r_in, p0_r_in);
+
+ q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+ tmp0_r = p7_r_in << 3;
+ tmp0_r -= p7_r_in;
+ tmp0_r += p6_r_in;
+ tmp0_r += q0_r_in;
+ tmp1_r = p6_r_in + p5_r_in;
+ tmp1_r += p4_r_in;
+ tmp1_r += p3_r_in;
+ tmp1_r += p2_r_in;
+ tmp1_r += p1_r_in;
+ tmp1_r += p0_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in,
+ p5_l_in, p4_l_in);
+ ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in,
+ p1_l_in, p0_l_in);
+ q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0);
+
+ tmp0_l = p7_l_in << 3;
+ tmp0_l -= p7_l_in;
+ tmp0_l += p6_l_in;
+ tmp0_l += q0_l_in;
+ tmp1_l = p6_l_in + p5_l_in;
+ tmp1_l += p4_l_in;
+ tmp1_l += p3_l_in;
+ tmp1_l += p2_l_in;
+ tmp1_l += p1_l_in;
+ tmp1_l += p0_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+ ST_UB(p6, src);
+ src += pitch;
+
+ /* p5 */
+ q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+ tmp0_r = p5_r_in - p6_r_in;
+ tmp0_r += q1_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1);
+ tmp0_l = p5_l_in - p6_l_in;
+ tmp0_l += q1_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+ ST_UB(p5, src);
+ src += pitch;
+
+ /* p4 */
+ q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+ tmp0_r = p4_r_in - p5_r_in;
+ tmp0_r += q2_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = (v8i16)__msa_srari_h((v8i16)tmp1_r, 4);
+
+ q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2);
+ tmp0_l = p4_l_in - p5_l_in;
+ tmp0_l += q2_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+ ST_UB(p4, src);
+ src += pitch;
+
+ /* p3 */
+ q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+ tmp0_r = p3_r_in - p4_r_in;
+ tmp0_r += q3_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3);
+ tmp0_l = p3_l_in - p4_l_in;
+ tmp0_l += q3_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+ ST_UB(p3, src);
+ src += pitch;
+
+ /* p2 */
+ q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+ filter8 = LD_UB(filter48);
+ tmp0_r = p2_r_in - p3_r_in;
+ tmp0_r += q4_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4);
+ tmp0_l = p2_l_in - p3_l_in;
+ tmp0_l += q4_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += pitch;
+
+ /* p1 */
+ q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+ filter8 = LD_UB(filter48 + 16);
+ tmp0_r = p1_r_in - p2_r_in;
+ tmp0_r += q5_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5);
+ tmp0_l = p1_l_in - p2_l_in;
+ tmp0_l += q5_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += pitch;
+
+ /* p0 */
+ q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+ filter8 = LD_UB(filter48 + 32);
+ tmp0_r = p0_r_in - p1_r_in;
+ tmp0_r += q6_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6);
+ tmp0_l = p0_l_in - p1_l_in;
+ tmp0_l += q6_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += pitch;
+
+ /* q0 */
+ q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+ filter8 = LD_UB(filter48 + 48);
+ tmp0_r = q7_r_in - p0_r_in;
+ tmp0_r += q0_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7);
+ tmp0_l = q7_l_in - p0_l_in;
+ tmp0_l += q0_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += pitch;
+
+ /* q1 */
+ filter8 = LD_UB(filter48 + 64);
+ tmp0_r = q7_r_in - q0_r_in;
+ tmp0_r += q1_r_in;
+ tmp0_r -= p6_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ tmp0_l = q7_l_in - q0_l_in;
+ tmp0_l += q1_l_in;
+ tmp0_l -= p6_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += pitch;
+
+ /* q2 */
+ filter8 = LD_UB(filter48 + 80);
+ tmp0_r = q7_r_in - q1_r_in;
+ tmp0_r += q2_r_in;
+ tmp0_r -= p5_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ tmp0_l = q7_l_in - q1_l_in;
+ tmp0_l += q2_l_in;
+ tmp0_l -= p5_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += pitch;
+
+ /* q3 */
+ tmp0_r = q7_r_in - q2_r_in;
+ tmp0_r += q3_r_in;
+ tmp0_r -= p4_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ tmp0_l = q7_l_in - q2_l_in;
+ tmp0_l += q3_l_in;
+ tmp0_l -= p4_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+ ST_UB(q3, src);
+ src += pitch;
+
+ /* q4 */
+ tmp0_r = q7_r_in - q3_r_in;
+ tmp0_r += q4_r_in;
+ tmp0_r -= p3_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ tmp0_l = q7_l_in - q3_l_in;
+ tmp0_l += q4_l_in;
+ tmp0_l -= p3_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+ ST_UB(q4, src);
+ src += pitch;
+
+ /* q5 */
+ tmp0_r = q7_r_in - q4_r_in;
+ tmp0_r += q5_r_in;
+ tmp0_r -= p2_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ tmp0_l = q7_l_in - q4_l_in;
+ tmp0_l += q5_l_in;
+ tmp0_l -= p2_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+ ST_UB(q5, src);
+ src += pitch;
+
+ /* q6 */
+ tmp0_r = q7_r_in - q5_r_in;
+ tmp0_r += q6_r_in;
+ tmp0_r -= p1_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ tmp0_l = q7_l_in - q5_l_in;
+ tmp0_l += q6_l_in;
+ tmp0_l -= p1_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+ ST_UB(q6, src);
+ }
+}
+
+void vp9_lpf_horizontal_16_dual_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr,
+ int32_t count) {+ DECLARE_ALIGNED(32, uint8_t, filter48[16 * 8]);
+ uint8_t early_exit = 0;
+
+ (void)count;
+
+ early_exit = vp9_hz_lpf_t4_and_t8_16w(src, pitch, &filter48[0], b_limit_ptr,
+ limit_ptr, thresh_ptr);
+
+ if (0 == early_exit) {+ vp9_hz_lpf_t16_16w(src, pitch, filter48);
+ }
+}
+
+void vp9_lpf_horizontal_16_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr,
+ int32_t count) {+ if (1 == count) {+ uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
+ uint64_t dword0, dword1;
+ v16u8 flat2, mask, hev, flat, thresh, b_limit, limit;
+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p7, p6, p5, p4, q4, q5, q6, q7;
+ v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+ v16u8 p0_filter16, p1_filter16;
+ v8i16 p2_filter8, p1_filter8, p0_filter8;
+ v8i16 q0_filter8, q1_filter8, q2_filter8;
+ v8u16 p7_r, p6_r, p5_r, p4_r, q7_r, q6_r, q5_r, q4_r;
+ v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
+ v16i8 zero = { 0 };+ v8u16 tmp0, tmp1, tmp2;
+
+ /* load vector elements */
+ LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+ b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+ limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+ VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out,
+ q1_out);
+
+ flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+ if (__msa_test_bz_v(flat)) {+ p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+ p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+ q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+ q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+ SD4(p1_d, p0_d, q0_d, q1_d, src - 2 * pitch, pitch);
+ } else {+ /* convert 8 bit input data into 16 bit */
+ ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+ zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+ q3_r);
+ VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
+ p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
+
+ /* convert 16 bit output data into 8 bit */
+ PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8,
+ zero, q0_filter8, p2_filter8, p1_filter8, p0_filter8,
+ q0_filter8);
+ PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
+
+ /* store pixel values */
+ p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
+ p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
+ p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
+ q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
+ q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
+ q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
+
+ /* load 16 vector elements */
+ LD_UB4((src - 8 * pitch), pitch, p7, p6, p5, p4);
+ LD_UB4(src + (4 * pitch), pitch, q4, q5, q6, q7);
+
+ VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+ if (__msa_test_bz_v(flat2)) {+ p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
+ p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+ p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+ q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+ q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+ q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
+
+ SD4(p2_d, p1_d, p0_d, q0_d, src - 3 * pitch, pitch);
+ SD(q1_d, src + pitch);
+ SD(q2_d, src + 2 * pitch);
+ } else {+ /* LSB(right) 8 pixel operation */
+ ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, q4, zero, q5,
+ zero, q6, zero, q7, p7_r, p6_r, p5_r, p4_r, q4_r, q5_r, q6_r,
+ q7_r);
+
+ tmp0 = p7_r << 3;
+ tmp0 -= p7_r;
+ tmp0 += p6_r;
+ tmp0 += q0_r;
+
+ src -= 7 * pitch;
+
+ /* calculation of p6 and p5 */
+ tmp1 = p6_r + p5_r + p4_r + p3_r;
+ tmp1 += (p2_r + p1_r + p0_r);
+ tmp1 += tmp0;
+ p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ tmp0 = p5_r - p6_r + q1_r - p7_r;
+ tmp1 += tmp0;
+ p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+ p1_filter16);
+ p0_filter16 = __msa_bmnz_v(p6, p0_filter16, flat2);
+ p1_filter16 = __msa_bmnz_v(p5, p1_filter16, flat2);
+ dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+ dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+ SD(dword0, src);
+ src += pitch;
+ SD(dword1, src);
+ src += pitch;
+
+ /* calculation of p4 and p3 */
+ tmp0 = p4_r - p5_r + q2_r - p7_r;
+ tmp2 = p3_r - p4_r + q3_r - p7_r;
+ tmp1 += tmp0;
+ p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ tmp1 += tmp2;
+ p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+ p1_filter16);
+ p0_filter16 = __msa_bmnz_v(p4, p0_filter16, flat2);
+ p1_filter16 = __msa_bmnz_v(p3, p1_filter16, flat2);
+ dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+ dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+ SD(dword0, src);
+ src += pitch;
+ SD(dword1, src);
+ src += pitch;
+
+ /* calculation of p2 and p1 */
+ tmp0 = p2_r - p3_r + q4_r - p7_r;
+ tmp2 = p1_r - p2_r + q5_r - p7_r;
+ tmp1 += tmp0;
+ p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ tmp1 += tmp2;
+ p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+ p1_filter16);
+ p0_filter16 = __msa_bmnz_v(p2_out, p0_filter16, flat2);
+ p1_filter16 = __msa_bmnz_v(p1_out, p1_filter16, flat2);
+ dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+ dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+ SD(dword0, src);
+ src += pitch;
+ SD(dword1, src);
+ src += pitch;
+
+ /* calculation of p0 and q0 */
+ tmp0 = (p0_r - p1_r) + (q6_r - p7_r);
+ tmp2 = (q7_r - p0_r) + (q0_r - p7_r);
+ tmp1 += tmp0;
+ p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ tmp1 += tmp2;
+ p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+ p1_filter16);
+ p0_filter16 = __msa_bmnz_v(p0_out, p0_filter16, flat2);
+ p1_filter16 = __msa_bmnz_v(q0_out, p1_filter16, flat2);
+ dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+ dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+ SD(dword0, src);
+ src += pitch;
+ SD(dword1, src);
+ src += pitch;
+
+ /* calculation of q1 and q2 */
+ tmp0 = q7_r - q0_r + q1_r - p6_r;
+ tmp2 = q7_r - q1_r + q2_r - p5_r;
+ tmp1 += tmp0;
+ p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ tmp1 += tmp2;
+ p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+ p1_filter16);
+ p0_filter16 = __msa_bmnz_v(q1_out, p0_filter16, flat2);
+ p1_filter16 = __msa_bmnz_v(q2_out, p1_filter16, flat2);
+ dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+ dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+ SD(dword0, src);
+ src += pitch;
+ SD(dword1, src);
+ src += pitch;
+
+ /* calculation of q3 and q4 */
+ tmp0 = (q7_r - q2_r) + (q3_r - p4_r);
+ tmp2 = (q7_r - q3_r) + (q4_r - p3_r);
+ tmp1 += tmp0;
+ p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ tmp1 += tmp2;
+ p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+ p1_filter16);
+ p0_filter16 = __msa_bmnz_v(q3, p0_filter16, flat2);
+ p1_filter16 = __msa_bmnz_v(q4, p1_filter16, flat2);
+ dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+ dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+ SD(dword0, src);
+ src += pitch;
+ SD(dword1, src);
+ src += pitch;
+
+ /* calculation of q5 and q6 */
+ tmp0 = (q7_r - q4_r) + (q5_r - p2_r);
+ tmp2 = (q7_r - q5_r) + (q6_r - p1_r);
+ tmp1 += tmp0;
+ p0_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ tmp1 += tmp2;
+ p1_filter16 = (v16u8)__msa_srari_h((v8i16)tmp1, 4);
+ PCKEV_B2_UB(zero, p0_filter16, zero, p1_filter16, p0_filter16,
+ p1_filter16);
+ p0_filter16 = __msa_bmnz_v(q5, p0_filter16, flat2);
+ p1_filter16 = __msa_bmnz_v(q6, p1_filter16, flat2);
+ dword0 = __msa_copy_u_d((v2i64)p0_filter16, 0);
+ dword1 = __msa_copy_u_d((v2i64)p1_filter16, 0);
+ SD(dword0, src);
+ src += pitch;
+ SD(dword1, src);
+ }
+ }
+ } else {+ vp9_lpf_horizontal_16_dual_msa(src, pitch, b_limit_ptr, limit_ptr,
+ thresh_ptr, count);
+ }
+}
+
+static void vp9_transpose_16x8_to_8x16(uint8_t *input, int32_t in_pitch,
+ uint8_t *output, int32_t out_pitch) {+ v16u8 p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, p0_org;
+ v16i8 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
+ v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+
+ LD_UB8(input, in_pitch,
+ p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org, p0_org);
+ /* 8x8 transpose */
+ TRANSPOSE8x8_UB_UB(p7_org, p6_org, p5_org, p4_org, p3_org, p2_org, p1_org,
+ p0_org, p7, p6, p5, p4, p3, p2, p1, p0);
+ /* 8x8 transpose */
+ ILVL_B4_SB(p5_org, p7_org, p4_org, p6_org, p1_org, p3_org, p0_org, p2_org,
+ tmp0, tmp1, tmp2, tmp3);
+ ILVR_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp4, tmp6);
+ ILVL_B2_SB(tmp1, tmp0, tmp3, tmp2, tmp5, tmp7);
+ ILVR_W2_UB(tmp6, tmp4, tmp7, tmp5, q0, q4);
+ ILVL_W2_UB(tmp6, tmp4, tmp7, tmp5, q2, q6);
+ SLDI_B4_0_UB(q0, q2, q4, q6, q1, q3, q5, q7, 8);
+
+ ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch);
+ output += (8 * out_pitch);
+ ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch);
+}
+
+static void vp9_transpose_8x16_to_16x8(uint8_t *input, int32_t in_pitch,
+ uint8_t *output, int32_t out_pitch) {+ v16u8 p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o;
+ v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+
+ LD_UB8(input, in_pitch, p7, p6, p5, p4, p3, p2, p1, p0);
+ LD_UB8(input + (8 * in_pitch), in_pitch, q0, q1, q2, q3, q4, q5, q6, q7);
+ TRANSPOSE16x8_UB_UB(p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5,
+ q6, q7, p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o);
+ ST_UB8(p7_o, p6_o, p5_o, p4_o, p3_o, p2_o, p1_o, p0_o, output, out_pitch);
+}
+
+static void vp9_transpose_16x16(uint8_t *input, int32_t in_pitch,
+ uint8_t *output, int32_t out_pitch) {+ v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
+ v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
+ v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+ v8i16 tmp0, tmp1, tmp4, tmp5, tmp6, tmp7;
+ v4i32 tmp2, tmp3;
+
+ LD_UB8(input, in_pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+ input += (8 * in_pitch);
+ LD_UB8(input, in_pitch,
+ row8, row9, row10, row11, row12, row13, row14, row15);
+
+ TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+ row8, row9, row10, row11, row12, row13, row14, row15,
+ p7, p6, p5, p4, p3, p2, p1, p0);
+
+ /* transpose 16x8 matrix into 8x16 */
+ /* total 8 intermediate register and 32 instructions */
+ q7 = (v16u8)__msa_ilvod_d((v2i64)row8, (v2i64)row0);
+ q6 = (v16u8)__msa_ilvod_d((v2i64)row9, (v2i64)row1);
+ q5 = (v16u8)__msa_ilvod_d((v2i64)row10, (v2i64)row2);
+ q4 = (v16u8)__msa_ilvod_d((v2i64)row11, (v2i64)row3);
+ q3 = (v16u8)__msa_ilvod_d((v2i64)row12, (v2i64)row4);
+ q2 = (v16u8)__msa_ilvod_d((v2i64)row13, (v2i64)row5);
+ q1 = (v16u8)__msa_ilvod_d((v2i64)row14, (v2i64)row6);
+ q0 = (v16u8)__msa_ilvod_d((v2i64)row15, (v2i64)row7);
+
+ ILVEV_B2_SH(q7, q6, q5, q4, tmp0, tmp1);
+ tmp4 = (v8i16)__msa_ilvod_b((v16i8)q6, (v16i8)q7);
+ tmp5 = (v8i16)__msa_ilvod_b((v16i8)q4, (v16i8)q5);
+
+ ILVEV_B2_UB(q3, q2, q1, q0, q5, q7);
+ tmp6 = (v8i16)__msa_ilvod_b((v16i8)q2, (v16i8)q3);
+ tmp7 = (v8i16)__msa_ilvod_b((v16i8)q0, (v16i8)q1);
+
+ ILVEV_H2_SW(tmp0, tmp1, q5, q7, tmp2, tmp3);
+ q0 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+ q4 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+ tmp2 = (v4i32)__msa_ilvod_h(tmp1, tmp0);
+ tmp3 = (v4i32)__msa_ilvod_h((v8i16)q7, (v8i16)q5);
+ q2 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+ q6 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+ ILVEV_H2_SW(tmp4, tmp5, tmp6, tmp7, tmp2, tmp3);
+ q1 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+ q5 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+ tmp2 = (v4i32)__msa_ilvod_h(tmp5, tmp4);
+ tmp3 = (v4i32)__msa_ilvod_h(tmp7, tmp6);
+ q3 = (v16u8)__msa_ilvev_w(tmp3, tmp2);
+ q7 = (v16u8)__msa_ilvod_w(tmp3, tmp2);
+
+ ST_UB8(p7, p6, p5, p4, p3, p2, p1, p0, output, out_pitch);
+ output += (8 * out_pitch);
+ ST_UB8(q0, q1, q2, q3, q4, q5, q6, q7, output, out_pitch);
+}
+
+int32_t vp9_vt_lpf_t4_and_t8_8w(uint8_t *src, uint8_t *filter48,
+ uint8_t *src_org, int32_t pitch_org,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr) {+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+ v16u8 flat, mask, hev, thresh, b_limit, limit;
+ v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+ v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+ v16i8 zero = { 0 };+ v8i16 vec0, vec1, vec2, vec3;
+
+ /* load vector elements */
+ LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+ b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+ limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+ /* mask and hev */
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ /* flat4 */
+ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+ /* filter4 */
+ VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+ flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+ if (__msa_test_bz_v(flat)) {+ ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+ ST4x8_UB(vec2, vec3, (src_org - 2), pitch_org);
+ return 1;
+ } else {+ ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+ zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+ q3_r);
+ VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+ p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+ /* convert 16 bit output data into 8 bit */
+ p2_r = (v8u16)__msa_pckev_b((v16i8)p2_filt8_r, (v16i8)p2_filt8_r);
+ p1_r = (v8u16)__msa_pckev_b((v16i8)p1_filt8_r, (v16i8)p1_filt8_r);
+ p0_r = (v8u16)__msa_pckev_b((v16i8)p0_filt8_r, (v16i8)p0_filt8_r);
+ q0_r = (v8u16)__msa_pckev_b((v16i8)q0_filt8_r, (v16i8)q0_filt8_r);
+ q1_r = (v8u16)__msa_pckev_b((v16i8)q1_filt8_r, (v16i8)q1_filt8_r);
+ q2_r = (v8u16)__msa_pckev_b((v16i8)q2_filt8_r, (v16i8)q2_filt8_r);
+
+ /* store pixel values */
+ p2_out = __msa_bmnz_v(p2, (v16u8)p2_r, flat);
+ p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_r, flat);
+ p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_r, flat);
+ q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_r, flat);
+ q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_r, flat);
+ q2_out = __msa_bmnz_v(q2, (v16u8)q2_r, flat);
+
+ ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+ filter48 += (4 * 16);
+ ST_UB2(q1_out, q2_out, filter48, 16);
+ filter48 += (2 * 16);
+ ST_UB(flat, filter48);
+
+ return 0;
+ }
+}
+
+int32_t vp9_vt_lpf_t16_8w(uint8_t *src, uint8_t *src_org, int32_t pitch,
+ uint8_t *filter48) {+ v16i8 zero = { 0 };+ v16u8 filter8, flat, flat2;
+ v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+ v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+ v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+ v8u16 tmp0_r, tmp1_r;
+ v8i16 r_out;
+
+ flat = LD_UB(filter48 + 6 * 16);
+
+ LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0);
+ LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7);
+
+ VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+ if (__msa_test_bz_v(flat2)) {+ v8i16 vec0, vec1, vec2, vec3, vec4;
+
+ LD_UB4(filter48, 16, p2, p1, p0, q0);
+ LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+ ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+ vec2 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
+
+ src_org -= 3;
+ ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch);
+ ST2x4_UB(vec2, 0, (src_org + 4), pitch);
+ src_org += (4 * pitch);
+ ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch);
+ ST2x4_UB(vec2, 4, (src_org + 4), pitch);
+
+ return 1;
+ } else {+ src -= 7 * 16;
+
+ ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
+ zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in,
+ p3_r_in, p2_r_in, p1_r_in, p0_r_in);
+ q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+ tmp0_r = p7_r_in << 3;
+ tmp0_r -= p7_r_in;
+ tmp0_r += p6_r_in;
+ tmp0_r += q0_r_in;
+ tmp1_r = p6_r_in + p5_r_in;
+ tmp1_r += p4_r_in;
+ tmp1_r += p3_r_in;
+ tmp1_r += p2_r_in;
+ tmp1_r += p1_r_in;
+ tmp1_r += p0_r_in;
+ tmp1_r += tmp0_r;
+
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+ ST8x1_UB(p6, src);
+ src += 16;
+
+ /* p5 */
+ q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+ tmp0_r = p5_r_in - p6_r_in;
+ tmp0_r += q1_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+ ST8x1_UB(p5, src);
+ src += 16;
+
+ /* p4 */
+ q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+ tmp0_r = p4_r_in - p5_r_in;
+ tmp0_r += q2_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+ ST8x1_UB(p4, src);
+ src += 16;
+
+ /* p3 */
+ q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+ tmp0_r = p3_r_in - p4_r_in;
+ tmp0_r += q3_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+ ST8x1_UB(p3, src);
+ src += 16;
+
+ /* p2 */
+ q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+ filter8 = LD_UB(filter48);
+ tmp0_r = p2_r_in - p3_r_in;
+ tmp0_r += q4_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST8x1_UB(filter8, src);
+ src += 16;
+
+ /* p1 */
+ q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+ filter8 = LD_UB(filter48 + 16);
+ tmp0_r = p1_r_in - p2_r_in;
+ tmp0_r += q5_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST8x1_UB(filter8, src);
+ src += 16;
+
+ /* p0 */
+ q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+ filter8 = LD_UB(filter48 + 32);
+ tmp0_r = p0_r_in - p1_r_in;
+ tmp0_r += q6_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST8x1_UB(filter8, src);
+ src += 16;
+
+ /* q0 */
+ q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+ filter8 = LD_UB(filter48 + 48);
+ tmp0_r = q7_r_in - p0_r_in;
+ tmp0_r += q0_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST8x1_UB(filter8, src);
+ src += 16;
+
+ /* q1 */
+ filter8 = LD_UB(filter48 + 64);
+ tmp0_r = q7_r_in - q0_r_in;
+ tmp0_r += q1_r_in;
+ tmp0_r -= p6_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST8x1_UB(filter8, src);
+ src += 16;
+
+ /* q2 */
+ filter8 = LD_UB(filter48 + 80);
+ tmp0_r = q7_r_in - q1_r_in;
+ tmp0_r += q2_r_in;
+ tmp0_r -= p5_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST8x1_UB(filter8, src);
+ src += 16;
+
+ /* q3 */
+ tmp0_r = q7_r_in - q2_r_in;
+ tmp0_r += q3_r_in;
+ tmp0_r -= p4_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+ ST8x1_UB(q3, src);
+ src += 16;
+
+ /* q4 */
+ tmp0_r = q7_r_in - q3_r_in;
+ tmp0_r += q4_r_in;
+ tmp0_r -= p3_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+ ST8x1_UB(q4, src);
+ src += 16;
+
+ /* q5 */
+ tmp0_r = q7_r_in - q4_r_in;
+ tmp0_r += q5_r_in;
+ tmp0_r -= p2_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+ ST8x1_UB(q5, src);
+ src += 16;
+
+ /* q6 */
+ tmp0_r = q7_r_in - q5_r_in;
+ tmp0_r += q6_r_in;
+ tmp0_r -= p1_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)r_out, (v16i8)r_out);
+ q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+ ST8x1_UB(q6, src);
+
+ return 0;
+ }
+}
+
+void vp9_lpf_vertical_16_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr) {+ uint8_t early_exit = 0;
+ DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]);
+ uint8_t *filter48 = &transposed_input[16 * 16];
+
+ vp9_transpose_16x8_to_8x16(src - 8, pitch, transposed_input, 16);
+
+ early_exit = vp9_vt_lpf_t4_and_t8_8w((transposed_input + 16 * 8),
+ &filter48[0], src, pitch, b_limit_ptr,
+ limit_ptr, thresh_ptr);
+
+ if (0 == early_exit) {+ early_exit = vp9_vt_lpf_t16_8w((transposed_input + 16 * 8), src, pitch,
+ &filter48[0]);
+
+ if (0 == early_exit) {+ vp9_transpose_8x16_to_16x8(transposed_input, 16, src - 8, pitch);
+ }
+ }
+}
+
+int32_t vp9_vt_lpf_t4_and_t8_16w(uint8_t *src, uint8_t *filter48,
+ uint8_t *src_org, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr) {+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+ v16u8 flat, mask, hev, thresh, b_limit, limit;
+ v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+ v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+ v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+ v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+ v16i8 zero = { 0 };+ v8i16 vec0, vec1, vec2, vec3, vec4, vec5;
+
+ /* load vector elements */
+ LD_UB8(src - (4 * 16), 16, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+ b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+ limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+ /* mask and hev */
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ /* flat4 */
+ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+ /* filter4 */
+ VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+ if (__msa_test_bz_v(flat)) {+ ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+ ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec4, vec5);
+
+ src_org -= 2;
+ ST4x8_UB(vec2, vec3, src_org, pitch);
+ src_org += 8 * pitch;
+ ST4x8_UB(vec4, vec5, src_org, pitch);
+
+ return 1;
+ } else {+ ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+ zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+ q3_r);
+ VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+ p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+ ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+ ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+ VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+ p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+ /* convert 16 bit output data into 8 bit */
+ PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+ p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+ p0_filt8_r, q0_filt8_r);
+ PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+ q2_filt8_r);
+
+ /* store pixel values */
+ p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+ p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+ p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+ q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+ q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+ q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+ ST_UB4(p2_out, p1_out, p0_out, q0_out, filter48, 16);
+ filter48 += (4 * 16);
+ ST_UB2(q1_out, q2_out, filter48, 16);
+ filter48 += (2 * 16);
+ ST_UB(flat, filter48);
+
+ return 0;
+ }
+}
+
+int32_t vp9_vt_lpf_t16_16w(uint8_t *src, uint8_t *src_org, int32_t pitch,
+ uint8_t *filter48) {+ v16u8 flat, flat2, filter8;
+ v16i8 zero = { 0 };+ v16u8 p7, p6, p5, p4, p3, p2, p1, p0, q0, q1, q2, q3, q4, q5, q6, q7;
+ v8u16 p7_r_in, p6_r_in, p5_r_in, p4_r_in, p3_r_in, p2_r_in, p1_r_in, p0_r_in;
+ v8u16 q7_r_in, q6_r_in, q5_r_in, q4_r_in, q3_r_in, q2_r_in, q1_r_in, q0_r_in;
+ v8u16 p7_l_in, p6_l_in, p5_l_in, p4_l_in, p3_l_in, p2_l_in, p1_l_in, p0_l_in;
+ v8u16 q7_l_in, q6_l_in, q5_l_in, q4_l_in, q3_l_in, q2_l_in, q1_l_in, q0_l_in;
+ v8u16 tmp0_r, tmp1_r, tmp0_l, tmp1_l;
+ v8i16 l_out, r_out;
+
+ flat = LD_UB(filter48 + 6 * 16);
+
+ LD_UB8((src - 8 * 16), 16, p7, p6, p5, p4, p3, p2, p1, p0);
+ LD_UB8(src, 16, q0, q1, q2, q3, q4, q5, q6, q7);
+
+ VP9_FLAT5(p7, p6, p5, p4, p0, q0, q4, q5, q6, q7, flat, flat2);
+
+ if (__msa_test_bz_v(flat2)) {+ v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+
+ LD_UB4(filter48, 16, p2, p1, p0, q0);
+ LD_UB2(filter48 + 4 * 16, 16, q1, q2);
+
+ ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+ ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec6, vec7);
+ ILVRL_B2_SH(q2, q1, vec2, vec5);
+
+ src_org -= 3;
+ ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src_org, pitch);
+ ST2x4_UB(vec2, 0, (src_org + 4), pitch);
+ src_org += (4 * pitch);
+ ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src_org, pitch);
+ ST2x4_UB(vec2, 4, (src_org + 4), pitch);
+ src_org += (4 * pitch);
+ ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src_org, pitch);
+ ST2x4_UB(vec5, 0, (src_org + 4), pitch);
+ src_org += (4 * pitch);
+ ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src_org, pitch);
+ ST2x4_UB(vec5, 4, (src_org + 4), pitch);
+
+ return 1;
+ } else {+ src -= 7 * 16;
+
+ ILVR_B8_UH(zero, p7, zero, p6, zero, p5, zero, p4, zero, p3, zero, p2,
+ zero, p1, zero, p0, p7_r_in, p6_r_in, p5_r_in, p4_r_in,
+ p3_r_in, p2_r_in, p1_r_in, p0_r_in);
+ q0_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q0);
+
+ tmp0_r = p7_r_in << 3;
+ tmp0_r -= p7_r_in;
+ tmp0_r += p6_r_in;
+ tmp0_r += q0_r_in;
+ tmp1_r = p6_r_in + p5_r_in;
+ tmp1_r += p4_r_in;
+ tmp1_r += p3_r_in;
+ tmp1_r += p2_r_in;
+ tmp1_r += p1_r_in;
+ tmp1_r += p0_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+
+ ILVL_B4_UH(zero, p7, zero, p6, zero, p5, zero, p4, p7_l_in, p6_l_in,
+ p5_l_in, p4_l_in);
+ ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l_in, p2_l_in,
+ p1_l_in, p0_l_in);
+ q0_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q0);
+
+ tmp0_l = p7_l_in << 3;
+ tmp0_l -= p7_l_in;
+ tmp0_l += p6_l_in;
+ tmp0_l += q0_l_in;
+ tmp1_l = p6_l_in + p5_l_in;
+ tmp1_l += p4_l_in;
+ tmp1_l += p3_l_in;
+ tmp1_l += p2_l_in;
+ tmp1_l += p1_l_in;
+ tmp1_l += p0_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ p6 = __msa_bmnz_v(p6, (v16u8)r_out, flat2);
+ ST_UB(p6, src);
+ src += 16;
+
+ /* p5 */
+ q1_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q1);
+ tmp0_r = p5_r_in - p6_r_in;
+ tmp0_r += q1_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ q1_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q1);
+ tmp0_l = p5_l_in - p6_l_in;
+ tmp0_l += q1_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ p5 = __msa_bmnz_v(p5, (v16u8)r_out, flat2);
+ ST_UB(p5, src);
+ src += 16;
+
+ /* p4 */
+ q2_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q2);
+ tmp0_r = p4_r_in - p5_r_in;
+ tmp0_r += q2_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ q2_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q2);
+ tmp0_l = p4_l_in - p5_l_in;
+ tmp0_l += q2_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ p4 = __msa_bmnz_v(p4, (v16u8)r_out, flat2);
+ ST_UB(p4, src);
+ src += 16;
+
+ /* p3 */
+ q3_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q3);
+ tmp0_r = p3_r_in - p4_r_in;
+ tmp0_r += q3_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ q3_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q3);
+ tmp0_l = p3_l_in - p4_l_in;
+ tmp0_l += q3_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ p3 = __msa_bmnz_v(p3, (v16u8)r_out, flat2);
+ ST_UB(p3, src);
+ src += 16;
+
+ /* p2 */
+ q4_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q4);
+ filter8 = LD_UB(filter48);
+ tmp0_r = p2_r_in - p3_r_in;
+ tmp0_r += q4_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ q4_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q4);
+ tmp0_l = p2_l_in - p3_l_in;
+ tmp0_l += q4_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += 16;
+
+ /* p1 */
+ q5_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q5);
+ filter8 = LD_UB(filter48 + 16);
+ tmp0_r = p1_r_in - p2_r_in;
+ tmp0_r += q5_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ q5_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q5);
+ tmp0_l = p1_l_in - p2_l_in;
+ tmp0_l += q5_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)(tmp1_l), 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += 16;
+
+ /* p0 */
+ q6_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q6);
+ filter8 = LD_UB(filter48 + 32);
+ tmp0_r = p0_r_in - p1_r_in;
+ tmp0_r += q6_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ q6_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q6);
+ tmp0_l = p0_l_in - p1_l_in;
+ tmp0_l += q6_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += 16;
+
+ /* q0 */
+ q7_r_in = (v8u16)__msa_ilvr_b(zero, (v16i8)q7);
+ filter8 = LD_UB(filter48 + 48);
+ tmp0_r = q7_r_in - p0_r_in;
+ tmp0_r += q0_r_in;
+ tmp0_r -= p7_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ q7_l_in = (v8u16)__msa_ilvl_b(zero, (v16i8)q7);
+ tmp0_l = q7_l_in - p0_l_in;
+ tmp0_l += q0_l_in;
+ tmp0_l -= p7_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += 16;
+
+ /* q1 */
+ filter8 = LD_UB(filter48 + 64);
+ tmp0_r = q7_r_in - q0_r_in;
+ tmp0_r += q1_r_in;
+ tmp0_r -= p6_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ tmp0_l = q7_l_in - q0_l_in;
+ tmp0_l += q1_l_in;
+ tmp0_l -= p6_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += 16;
+
+ /* q2 */
+ filter8 = LD_UB(filter48 + 80);
+ tmp0_r = q7_r_in - q1_r_in;
+ tmp0_r += q2_r_in;
+ tmp0_r -= p5_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ tmp0_l = q7_l_in - q1_l_in;
+ tmp0_l += q2_l_in;
+ tmp0_l -= p5_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ filter8 = __msa_bmnz_v(filter8, (v16u8)r_out, flat2);
+ ST_UB(filter8, src);
+ src += 16;
+
+ /* q3 */
+ tmp0_r = q7_r_in - q2_r_in;
+ tmp0_r += q3_r_in;
+ tmp0_r -= p4_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ tmp0_l = q7_l_in - q2_l_in;
+ tmp0_l += q3_l_in;
+ tmp0_l -= p4_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ q3 = __msa_bmnz_v(q3, (v16u8)r_out, flat2);
+ ST_UB(q3, src);
+ src += 16;
+
+ /* q4 */
+ tmp0_r = q7_r_in - q3_r_in;
+ tmp0_r += q4_r_in;
+ tmp0_r -= p3_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ tmp0_l = q7_l_in - q3_l_in;
+ tmp0_l += q4_l_in;
+ tmp0_l -= p3_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ q4 = __msa_bmnz_v(q4, (v16u8)r_out, flat2);
+ ST_UB(q4, src);
+ src += 16;
+
+ /* q5 */
+ tmp0_r = q7_r_in - q4_r_in;
+ tmp0_r += q5_r_in;
+ tmp0_r -= p2_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ tmp0_l = q7_l_in - q4_l_in;
+ tmp0_l += q5_l_in;
+ tmp0_l -= p2_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ q5 = __msa_bmnz_v(q5, (v16u8)r_out, flat2);
+ ST_UB(q5, src);
+ src += 16;
+
+ /* q6 */
+ tmp0_r = q7_r_in - q5_r_in;
+ tmp0_r += q6_r_in;
+ tmp0_r -= p1_r_in;
+ tmp1_r += tmp0_r;
+ r_out = __msa_srari_h((v8i16)tmp1_r, 4);
+ tmp0_l = q7_l_in - q5_l_in;
+ tmp0_l += q6_l_in;
+ tmp0_l -= p1_l_in;
+ tmp1_l += tmp0_l;
+ l_out = __msa_srari_h((v8i16)tmp1_l, 4);
+ r_out = (v8i16)__msa_pckev_b((v16i8)l_out, (v16i8)r_out);
+ q6 = __msa_bmnz_v(q6, (v16u8)r_out, flat2);
+ ST_UB(q6, src);
+
+ return 0;
+ }
+}
+
+void vp9_lpf_vertical_16_dual_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr) {+ uint8_t early_exit = 0;
+ DECLARE_ALIGNED(32, uint8_t, transposed_input[16 * 24]);
+ uint8_t *filter48 = &transposed_input[16 * 16];
+
+ vp9_transpose_16x16((src - 8), pitch, &transposed_input[0], 16);
+
+ early_exit = vp9_vt_lpf_t4_and_t8_16w((transposed_input + 16 * 8),
+ &filter48[0], src, pitch, b_limit_ptr,
+ limit_ptr, thresh_ptr);
+
+ if (0 == early_exit) {+ early_exit = vp9_vt_lpf_t16_16w((transposed_input + 16 * 8), src, pitch,
+ &filter48[0]);
+
+ if (0 == early_exit) {+ vp9_transpose_16x16(transposed_input, 16, (src - 8), pitch);
+ }
+ }
+}
--- /dev/null
+++ b/vpx_dsp/mips/loopfilter_4_msa.c
@@ -1,0 +1,152 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/loopfilter_msa.h"
+
+void vp9_lpf_horizontal_4_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr,
+ int32_t count) {+ uint64_t p1_d, p0_d, q0_d, q1_d;
+ v16u8 mask, hev, flat, thresh, b_limit, limit;
+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0, p1_out, p0_out, q0_out, q1_out;
+
+ (void)count;
+
+ /* load vector elements */
+ LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+ b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+ limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+ p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+ p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+ q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+ q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+ SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
+}
+
+void vp9_lpf_horizontal_4_dual_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit0_ptr,
+ const uint8_t *limit0_ptr,
+ const uint8_t *thresh0_ptr,
+ const uint8_t *b_limit1_ptr,
+ const uint8_t *limit1_ptr,
+ const uint8_t *thresh1_ptr) {+ v16u8 mask, hev, flat, thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+
+ /* load vector elements */
+ LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+ thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+ thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+ b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+ b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+ b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+ limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+ limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+ limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
+ hev, mask, flat);
+ VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+
+ ST_UB4(p1, p0, q0, q1, (src - 2 * pitch), pitch);
+}
+
+void vp9_lpf_vertical_4_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr,
+ int32_t count) {+ v16u8 mask, hev, flat, limit, thresh, b_limit;
+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v8i16 vec0, vec1, vec2, vec3;
+
+ (void)count;
+
+ LD_UB8((src - 4), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+ b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+ limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+ TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+ ILVR_B2_SH(p0, p1, q1, q0, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+
+ src -= 2;
+ ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+ src += 4 * pitch;
+ ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+}
+
+void vp9_lpf_vertical_4_dual_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit0_ptr,
+ const uint8_t *limit0_ptr,
+ const uint8_t *thresh0_ptr,
+ const uint8_t *b_limit1_ptr,
+ const uint8_t *limit1_ptr,
+ const uint8_t *thresh1_ptr) {+ v16u8 mask, hev, flat;
+ v16u8 thresh0, b_limit0, limit0, thresh1, b_limit1, limit1;
+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
+ v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
+ v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
+
+ LD_UB8(src - 4, pitch, row0, row1, row2, row3, row4, row5, row6, row7);
+ LD_UB8(src - 4 + (8 * pitch), pitch,
+ row8, row9, row10, row11, row12, row13, row14, row15);
+
+ TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
+ row8, row9, row10, row11, row12, row13, row14, row15,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh0 = (v16u8)__msa_fill_b(*thresh0_ptr);
+ thresh1 = (v16u8)__msa_fill_b(*thresh1_ptr);
+ thresh0 = (v16u8)__msa_ilvr_d((v2i64)thresh1, (v2i64)thresh0);
+
+ b_limit0 = (v16u8)__msa_fill_b(*b_limit0_ptr);
+ b_limit1 = (v16u8)__msa_fill_b(*b_limit1_ptr);
+ b_limit0 = (v16u8)__msa_ilvr_d((v2i64)b_limit1, (v2i64)b_limit0);
+
+ limit0 = (v16u8)__msa_fill_b(*limit0_ptr);
+ limit1 = (v16u8)__msa_fill_b(*limit1_ptr);
+ limit0 = (v16u8)__msa_ilvr_d((v2i64)limit1, (v2i64)limit0);
+
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
+ hev, mask, flat);
+ VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1, p0, q0, q1);
+ ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+ ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3);
+ ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
+ ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5);
+
+ src -= 2;
+
+ ST4x8_UB(tmp2, tmp3, src, pitch);
+ src += (8 * pitch);
+ ST4x8_UB(tmp4, tmp5, src, pitch);
+}
--- /dev/null
+++ b/vpx_dsp/mips/loopfilter_8_msa.c
@@ -1,0 +1,348 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "vpx_dsp/mips/loopfilter_msa.h"
+
+void vp9_lpf_horizontal_8_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr,
+ int32_t count) {+ uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
+ v16u8 mask, hev, flat, thresh, b_limit, limit;
+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+ v8i16 p2_filter8, p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8;
+ v8u16 p3_r, p2_r, p1_r, p0_r, q3_r, q2_r, q1_r, q0_r;
+ v16i8 zero = { 0 };+
+ (void)count;
+
+ /* load vector elements */
+ LD_UB8((src - 4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+ b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+ limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+ VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+ flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+ if (__msa_test_bz_v(flat)) {+ p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+ p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+ q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+ q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+ SD4(p1_d, p0_d, q0_d, q1_d, (src - 2 * pitch), pitch);
+ } else {+ ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+ zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
+ q2_r, q3_r);
+ VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filter8,
+ p1_filter8, p0_filter8, q0_filter8, q1_filter8, q2_filter8);
+
+ /* convert 16 bit output data into 8 bit */
+ PCKEV_B4_SH(zero, p2_filter8, zero, p1_filter8, zero, p0_filter8,
+ zero, q0_filter8, p2_filter8, p1_filter8, p0_filter8,
+ q0_filter8);
+ PCKEV_B2_SH(zero, q1_filter8, zero, q2_filter8, q1_filter8, q2_filter8);
+
+ /* store pixel values */
+ p2_out = __msa_bmnz_v(p2, (v16u8)p2_filter8, flat);
+ p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filter8, flat);
+ p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filter8, flat);
+ q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filter8, flat);
+ q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filter8, flat);
+ q2_out = __msa_bmnz_v(q2, (v16u8)q2_filter8, flat);
+
+ p2_d = __msa_copy_u_d((v2i64)p2_out, 0);
+ p1_d = __msa_copy_u_d((v2i64)p1_out, 0);
+ p0_d = __msa_copy_u_d((v2i64)p0_out, 0);
+ q0_d = __msa_copy_u_d((v2i64)q0_out, 0);
+ q1_d = __msa_copy_u_d((v2i64)q1_out, 0);
+ q2_d = __msa_copy_u_d((v2i64)q2_out, 0);
+
+ src -= 3 * pitch;
+
+ SD4(p2_d, p1_d, p0_d, q0_d, src, pitch);
+ src += (4 * pitch);
+ SD(q1_d, src);
+ src += pitch;
+ SD(q2_d, src);
+ }
+}
+
+void vp9_lpf_horizontal_8_dual_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *b_limit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1) {+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v16u8 p2_out, p1_out, p0_out, q0_out, q1_out, q2_out;
+ v16u8 flat, mask, hev, tmp, thresh, b_limit, limit;
+ v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+ v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+ v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+ v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+ v16u8 zero = { 0 };+
+ /* load vector elements */
+ LD_UB8(src - (4 * pitch), pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh0);
+ tmp = (v16u8)__msa_fill_b(*thresh1);
+ thresh = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)thresh);
+
+ b_limit = (v16u8)__msa_fill_b(*b_limit0);
+ tmp = (v16u8)__msa_fill_b(*b_limit1);
+ b_limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)b_limit);
+
+ limit = (v16u8)__msa_fill_b(*limit0);
+ tmp = (v16u8)__msa_fill_b(*limit1);
+ limit = (v16u8)__msa_ilvr_d((v2i64)tmp, (v2i64)limit);
+
+ /* mask and hev */
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+ VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+ if (__msa_test_bz_v(flat)) {+ ST_UB4(p1_out, p0_out, q0_out, q1_out, (src - 2 * pitch), pitch);
+ } else {+ ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+ zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r,
+ q2_r, q3_r);
+ VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+ p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+ ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+ ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+ VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+ p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+ /* convert 16 bit output data into 8 bit */
+ PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+ p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+ p0_filt8_r, q0_filt8_r);
+ PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+ q2_filt8_r);
+
+ /* store pixel values */
+ p2_out = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+ p1_out = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+ p0_out = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+ q0_out = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+ q1_out = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+ q2_out = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+ src -= 3 * pitch;
+
+ ST_UB4(p2_out, p1_out, p0_out, q0_out, src, pitch);
+ src += (4 * pitch);
+ ST_UB2(q1_out, q2_out, src, pitch);
+ src += (2 * pitch);
+ }
+}
+
+void vp9_lpf_vertical_8_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit_ptr,
+ const uint8_t *limit_ptr,
+ const uint8_t *thresh_ptr,
+ int32_t count) {+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v16u8 p1_out, p0_out, q0_out, q1_out;
+ v16u8 flat, mask, hev, thresh, b_limit, limit;
+ v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+ v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+ v16u8 zero = { 0 };+ v8i16 vec0, vec1, vec2, vec3, vec4;
+
+ (void)count;
+
+ /* load vector elements */
+ LD_UB8(src - 4, pitch, p3, p2, p1, p0, q0, q1, q2, q3);
+
+ TRANSPOSE8x8_UB_UB(p3, p2, p1, p0, q0, q1, q2, q3,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh_ptr);
+ b_limit = (v16u8)__msa_fill_b(*b_limit_ptr);
+ limit = (v16u8)__msa_fill_b(*limit_ptr);
+
+ /* mask and hev */
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ /* flat4 */
+ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+ /* filter4 */
+ VP9_LPF_FILTER4_8W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+ flat = (v16u8)__msa_ilvr_d((v2i64)zero, (v2i64)flat);
+
+ if (__msa_test_bz_v(flat)) {+ /* Store 4 pixels p1-_q1 */
+ ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+
+ src -= 2;
+ ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+ src += 4 * pitch;
+ ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+ } else {+ ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+ zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+ q3_r);
+ VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+ p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+ /* convert 16 bit output data into 8 bit */
+ PCKEV_B4_SH(p2_filt8_r, p2_filt8_r, p1_filt8_r, p1_filt8_r, p0_filt8_r,
+ p0_filt8_r, q0_filt8_r, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+ p0_filt8_r, q0_filt8_r);
+ PCKEV_B2_SH(q1_filt8_r, q1_filt8_r, q2_filt8_r, q2_filt8_r, q1_filt8_r,
+ q2_filt8_r);
+
+ /* store pixel values */
+ p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+ p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+ p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+ q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+ q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+ q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+ /* Store 6 pixels p2-_q2 */
+ ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+ vec4 = (v8i16)__msa_ilvr_b((v16i8)q2, (v16i8)q1);
+
+ src -= 3;
+ ST4x4_UB(vec2, vec2, 0, 1, 2, 3, src, pitch);
+ ST2x4_UB(vec4, 0, src + 4, pitch);
+ src += (4 * pitch);
+ ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+ ST2x4_UB(vec4, 4, src + 4, pitch);
+ }
+}
+
+void vp9_lpf_vertical_8_dual_msa(uint8_t *src, int32_t pitch,
+ const uint8_t *b_limit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *b_limit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1) {+ uint8_t *temp_src;
+ v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
+ v16u8 p1_out, p0_out, q0_out, q1_out;
+ v16u8 flat, mask, hev, thresh, b_limit, limit;
+ v16u8 row4, row5, row6, row7, row12, row13, row14, row15;
+ v8u16 p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r;
+ v8u16 p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l;
+ v8i16 p2_filt8_r, p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r;
+ v8i16 p2_filt8_l, p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l;
+ v16u8 zero = { 0 };+ v8i16 vec0, vec1, vec2, vec3, vec4, vec5, vec6, vec7;
+
+ temp_src = src - 4;
+
+ LD_UB8(temp_src, pitch, p0, p1, p2, p3, row4, row5, row6, row7);
+ temp_src += (8 * pitch);
+ LD_UB8(temp_src, pitch, q3, q2, q1, q0, row12, row13, row14, row15);
+
+ /* transpose 16x8 matrix into 8x16 */
+ TRANSPOSE16x8_UB_UB(p0, p1, p2, p3, row4, row5, row6, row7,
+ q3, q2, q1, q0, row12, row13, row14, row15,
+ p3, p2, p1, p0, q0, q1, q2, q3);
+
+ thresh = (v16u8)__msa_fill_b(*thresh0);
+ vec0 = (v8i16)__msa_fill_b(*thresh1);
+ thresh = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)thresh);
+
+ b_limit = (v16u8)__msa_fill_b(*b_limit0);
+ vec0 = (v8i16)__msa_fill_b(*b_limit1);
+ b_limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)b_limit);
+
+ limit = (v16u8)__msa_fill_b(*limit0);
+ vec0 = (v8i16)__msa_fill_b(*limit1);
+ limit = (v16u8)__msa_ilvr_d((v2i64)vec0, (v2i64)limit);
+
+ /* mask and hev */
+ LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
+ hev, mask, flat);
+ /* flat4 */
+ VP9_FLAT4(p3, p2, p0, q0, q2, q3, flat);
+ /* filter4 */
+ VP9_LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev, p1_out, p0_out, q0_out, q1_out);
+
+ if (__msa_test_bz_v(flat)) {+ ILVR_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec2, vec3);
+ ILVL_B2_SH(p0_out, p1_out, q1_out, q0_out, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec4, vec5);
+
+ src -= 2;
+ ST4x8_UB(vec2, vec3, src, pitch);
+ src += 8 * pitch;
+ ST4x8_UB(vec4, vec5, src, pitch);
+ } else {+ ILVR_B8_UH(zero, p3, zero, p2, zero, p1, zero, p0, zero, q0, zero, q1,
+ zero, q2, zero, q3, p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r,
+ q3_r);
+ VP9_FILTER8(p3_r, p2_r, p1_r, p0_r, q0_r, q1_r, q2_r, q3_r, p2_filt8_r,
+ p1_filt8_r, p0_filt8_r, q0_filt8_r, q1_filt8_r, q2_filt8_r);
+
+ ILVL_B4_UH(zero, p3, zero, p2, zero, p1, zero, p0, p3_l, p2_l, p1_l, p0_l);
+ ILVL_B4_UH(zero, q0, zero, q1, zero, q2, zero, q3, q0_l, q1_l, q2_l, q3_l);
+
+ /* filter8 */
+ VP9_FILTER8(p3_l, p2_l, p1_l, p0_l, q0_l, q1_l, q2_l, q3_l, p2_filt8_l,
+ p1_filt8_l, p0_filt8_l, q0_filt8_l, q1_filt8_l, q2_filt8_l);
+
+ /* convert 16 bit output data into 8 bit */
+ PCKEV_B4_SH(p2_filt8_l, p2_filt8_r, p1_filt8_l, p1_filt8_r, p0_filt8_l,
+ p0_filt8_r, q0_filt8_l, q0_filt8_r, p2_filt8_r, p1_filt8_r,
+ p0_filt8_r, q0_filt8_r);
+ PCKEV_B2_SH(q1_filt8_l, q1_filt8_r, q2_filt8_l, q2_filt8_r, q1_filt8_r,
+ q2_filt8_r);
+
+ /* store pixel values */
+ p2 = __msa_bmnz_v(p2, (v16u8)p2_filt8_r, flat);
+ p1 = __msa_bmnz_v(p1_out, (v16u8)p1_filt8_r, flat);
+ p0 = __msa_bmnz_v(p0_out, (v16u8)p0_filt8_r, flat);
+ q0 = __msa_bmnz_v(q0_out, (v16u8)q0_filt8_r, flat);
+ q1 = __msa_bmnz_v(q1_out, (v16u8)q1_filt8_r, flat);
+ q2 = __msa_bmnz_v(q2, (v16u8)q2_filt8_r, flat);
+
+ ILVR_B2_SH(p1, p2, q0, p0, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec3, vec4);
+ ILVL_B2_SH(p1, p2, q0, p0, vec0, vec1);
+ ILVRL_H2_SH(vec1, vec0, vec6, vec7);
+ ILVRL_B2_SH(q2, q1, vec2, vec5);
+
+ src -= 3;
+ ST4x4_UB(vec3, vec3, 0, 1, 2, 3, src, pitch);
+ ST2x4_UB(vec2, 0, src + 4, pitch);
+ src += (4 * pitch);
+ ST4x4_UB(vec4, vec4, 0, 1, 2, 3, src, pitch);
+ ST2x4_UB(vec2, 4, src + 4, pitch);
+ src += (4 * pitch);
+ ST4x4_UB(vec6, vec6, 0, 1, 2, 3, src, pitch);
+ ST2x4_UB(vec5, 0, src + 4, pitch);
+ src += (4 * pitch);
+ ST4x4_UB(vec7, vec7, 0, 1, 2, 3, src, pitch);
+ ST2x4_UB(vec5, 4, src + 4, pitch);
+ }
+}
--- /dev/null
+++ b/vpx_dsp/mips/loopfilter_msa.h
@@ -1,0 +1,246 @@
+/*
+ * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VPX_DSP_LOOPFILTER_MSA_H_
+#define VPX_DSP_LOOPFILTER_MSA_H_
+
+#include "vpx_dsp/mips/macros_msa.h"
+
+#define VP9_LPF_FILTER4_8W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \
+ p1_out, p0_out, q0_out, q1_out) { \+ v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign; \
+ v16i8 filt, filt1, filt2, cnst4b, cnst3b; \
+ v8i16 q0_sub_p0_r, filt_r, cnst3h; \
+ \
+ p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \
+ p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \
+ q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \
+ q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \
+ \
+ filt = __msa_subs_s_b(p1_m, q1_m); \
+ filt = filt & (v16i8)hev_in; \
+ q0_sub_p0 = q0_m - p0_m; \
+ filt_sign = __msa_clti_s_b(filt, 0); \
+ \
+ cnst3h = __msa_ldi_h(3); \
+ q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0); \
+ q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h); \
+ filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt); \
+ filt_r += q0_sub_p0_r; \
+ filt_r = __msa_sat_s_h(filt_r, 7); \
+ \
+ /* combine left and right part */ \
+ filt = __msa_pckev_b((v16i8)filt_r, (v16i8)filt_r); \
+ \
+ filt = filt & (v16i8)mask_in; \
+ cnst4b = __msa_ldi_b(4); \
+ filt1 = __msa_adds_s_b(filt, cnst4b); \
+ filt1 >>= 3; \
+ \
+ cnst3b = __msa_ldi_b(3); \
+ filt2 = __msa_adds_s_b(filt, cnst3b); \
+ filt2 >>= 3; \
+ \
+ q0_m = __msa_subs_s_b(q0_m, filt1); \
+ q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \
+ p0_m = __msa_adds_s_b(p0_m, filt2); \
+ p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \
+ \
+ filt = __msa_srari_b(filt1, 1); \
+ hev_in = __msa_xori_b((v16u8)hev_in, 0xff); \
+ filt = filt & (v16i8)hev_in; \
+ \
+ q1_m = __msa_subs_s_b(q1_m, filt); \
+ q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \
+ p1_m = __msa_adds_s_b(p1_m, filt); \
+ p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \
+}
+
+#define VP9_LPF_FILTER4_4W(p1_in, p0_in, q0_in, q1_in, mask_in, hev_in, \
+ p1_out, p0_out, q0_out, q1_out) { \+ v16i8 p1_m, p0_m, q0_m, q1_m, q0_sub_p0, filt_sign; \
+ v16i8 filt, filt1, filt2, cnst4b, cnst3b; \
+ v8i16 q0_sub_p0_r, q0_sub_p0_l, filt_l, filt_r, cnst3h; \
+ \
+ p1_m = (v16i8)__msa_xori_b(p1_in, 0x80); \
+ p0_m = (v16i8)__msa_xori_b(p0_in, 0x80); \
+ q0_m = (v16i8)__msa_xori_b(q0_in, 0x80); \
+ q1_m = (v16i8)__msa_xori_b(q1_in, 0x80); \
+ \
+ filt = __msa_subs_s_b(p1_m, q1_m); \
+ \
+ filt = filt & (v16i8)hev_in; \
+ \
+ q0_sub_p0 = q0_m - p0_m; \
+ filt_sign = __msa_clti_s_b(filt, 0); \
+ \
+ cnst3h = __msa_ldi_h(3); \
+ q0_sub_p0_r = (v8i16)__msa_ilvr_b(q0_sub_p0, q0_sub_p0); \
+ q0_sub_p0_r = __msa_dotp_s_h((v16i8)q0_sub_p0_r, (v16i8)cnst3h); \
+ filt_r = (v8i16)__msa_ilvr_b(filt_sign, filt); \
+ filt_r += q0_sub_p0_r; \
+ filt_r = __msa_sat_s_h(filt_r, 7); \
+ \
+ q0_sub_p0_l = (v8i16)__msa_ilvl_b(q0_sub_p0, q0_sub_p0); \
+ q0_sub_p0_l = __msa_dotp_s_h((v16i8)q0_sub_p0_l, (v16i8)cnst3h); \
+ filt_l = (v8i16)__msa_ilvl_b(filt_sign, filt); \
+ filt_l += q0_sub_p0_l; \
+ filt_l = __msa_sat_s_h(filt_l, 7); \
+ \
+ filt = __msa_pckev_b((v16i8)filt_l, (v16i8)filt_r); \
+ filt = filt & (v16i8)mask_in; \
+ \
+ cnst4b = __msa_ldi_b(4); \
+ filt1 = __msa_adds_s_b(filt, cnst4b); \
+ filt1 >>= 3; \
+ \
+ cnst3b = __msa_ldi_b(3); \
+ filt2 = __msa_adds_s_b(filt, cnst3b); \
+ filt2 >>= 3; \
+ \
+ q0_m = __msa_subs_s_b(q0_m, filt1); \
+ q0_out = __msa_xori_b((v16u8)q0_m, 0x80); \
+ p0_m = __msa_adds_s_b(p0_m, filt2); \
+ p0_out = __msa_xori_b((v16u8)p0_m, 0x80); \
+ \
+ filt = __msa_srari_b(filt1, 1); \
+ hev_in = __msa_xori_b((v16u8)hev_in, 0xff); \
+ filt = filt & (v16i8)hev_in; \
+ \
+ q1_m = __msa_subs_s_b(q1_m, filt); \
+ q1_out = __msa_xori_b((v16u8)q1_m, 0x80); \
+ p1_m = __msa_adds_s_b(p1_m, filt); \
+ p1_out = __msa_xori_b((v16u8)p1_m, 0x80); \
+}
+
+#define VP9_FLAT4(p3_in, p2_in, p0_in, q0_in, q2_in, q3_in, flat_out) { \+ v16u8 tmp, p2_a_sub_p0, q2_a_sub_q0, p3_a_sub_p0, q3_a_sub_q0; \
+ v16u8 zero_in = { 0 }; \+ \
+ tmp = __msa_ori_b(zero_in, 1); \
+ p2_a_sub_p0 = __msa_asub_u_b(p2_in, p0_in); \
+ q2_a_sub_q0 = __msa_asub_u_b(q2_in, q0_in); \
+ p3_a_sub_p0 = __msa_asub_u_b(p3_in, p0_in); \
+ q3_a_sub_q0 = __msa_asub_u_b(q3_in, q0_in); \
+ \
+ p2_a_sub_p0 = __msa_max_u_b(p2_a_sub_p0, q2_a_sub_q0); \
+ flat_out = __msa_max_u_b(p2_a_sub_p0, flat_out); \
+ p3_a_sub_p0 = __msa_max_u_b(p3_a_sub_p0, q3_a_sub_q0); \
+ flat_out = __msa_max_u_b(p3_a_sub_p0, flat_out); \
+ \
+ flat_out = (tmp < (v16u8)flat_out); \
+ flat_out = __msa_xori_b(flat_out, 0xff); \
+ flat_out = flat_out & (mask); \
+}
+
+#define VP9_FLAT5(p7_in, p6_in, p5_in, p4_in, p0_in, q0_in, q4_in, \
+ q5_in, q6_in, q7_in, flat_in, flat2_out) { \+ v16u8 tmp, zero_in = { 0 }; \+ v16u8 p4_a_sub_p0, q4_a_sub_q0, p5_a_sub_p0, q5_a_sub_q0; \
+ v16u8 p6_a_sub_p0, q6_a_sub_q0, p7_a_sub_p0, q7_a_sub_q0; \
+ \
+ tmp = __msa_ori_b(zero_in, 1); \
+ p4_a_sub_p0 = __msa_asub_u_b(p4_in, p0_in); \
+ q4_a_sub_q0 = __msa_asub_u_b(q4_in, q0_in); \
+ p5_a_sub_p0 = __msa_asub_u_b(p5_in, p0_in); \
+ q5_a_sub_q0 = __msa_asub_u_b(q5_in, q0_in); \
+ p6_a_sub_p0 = __msa_asub_u_b(p6_in, p0_in); \
+ q6_a_sub_q0 = __msa_asub_u_b(q6_in, q0_in); \
+ p7_a_sub_p0 = __msa_asub_u_b(p7_in, p0_in); \
+ q7_a_sub_q0 = __msa_asub_u_b(q7_in, q0_in); \
+ \
+ p4_a_sub_p0 = __msa_max_u_b(p4_a_sub_p0, q4_a_sub_q0); \
+ flat2_out = __msa_max_u_b(p5_a_sub_p0, q5_a_sub_q0); \
+ flat2_out = __msa_max_u_b(p4_a_sub_p0, flat2_out); \
+ p6_a_sub_p0 = __msa_max_u_b(p6_a_sub_p0, q6_a_sub_q0); \
+ flat2_out = __msa_max_u_b(p6_a_sub_p0, flat2_out); \
+ p7_a_sub_p0 = __msa_max_u_b(p7_a_sub_p0, q7_a_sub_q0); \
+ flat2_out = __msa_max_u_b(p7_a_sub_p0, flat2_out); \
+ \
+ flat2_out = (tmp < (v16u8)flat2_out); \
+ flat2_out = __msa_xori_b(flat2_out, 0xff); \
+ flat2_out = flat2_out & flat_in; \
+}
+
+#define VP9_FILTER8(p3_in, p2_in, p1_in, p0_in, \
+ q0_in, q1_in, q2_in, q3_in, \
+ p2_filt8_out, p1_filt8_out, p0_filt8_out, \
+ q0_filt8_out, q1_filt8_out, q2_filt8_out) { \+ v8u16 tmp0, tmp1, tmp2; \
+ \
+ tmp2 = p2_in + p1_in + p0_in; \
+ tmp0 = p3_in << 1; \
+ \
+ tmp0 = tmp0 + tmp2 + q0_in; \
+ tmp1 = tmp0 + p3_in + p2_in; \
+ p2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
+ \
+ tmp1 = tmp0 + p1_in + q1_in; \
+ p1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
+ \
+ tmp1 = q2_in + q1_in + q0_in; \
+ tmp2 = tmp2 + tmp1; \
+ tmp0 = tmp2 + (p0_in); \
+ tmp0 = tmp0 + (p3_in); \
+ p0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp0, 3); \
+ \
+ tmp0 = q2_in + q3_in; \
+ tmp0 = p0_in + tmp1 + tmp0; \
+ tmp1 = q3_in + q3_in; \
+ tmp1 = tmp1 + tmp0; \
+ q2_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
+ \
+ tmp0 = tmp2 + q3_in; \
+ tmp1 = tmp0 + q0_in; \
+ q0_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
+ \
+ tmp1 = tmp0 - p2_in; \
+ tmp0 = q1_in + q3_in; \
+ tmp1 = tmp0 + tmp1; \
+ q1_filt8_out = (v8i16)__msa_srari_h((v8i16)tmp1, 3); \
+}
+
+#define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in, \
+ q0_in, q1_in, q2_in, q3_in, \
+ limit_in, b_limit_in, thresh_in, \
+ hev_out, mask_out, flat_out) { \+ v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m; \
+ v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m; \
+ \
+ /* absolute subtraction of pixel values */ \
+ p3_asub_p2_m = __msa_asub_u_b(p3_in, p2_in); \
+ p2_asub_p1_m = __msa_asub_u_b(p2_in, p1_in); \
+ p1_asub_p0_m = __msa_asub_u_b(p1_in, p0_in); \
+ q1_asub_q0_m = __msa_asub_u_b(q1_in, q0_in); \
+ q2_asub_q1_m = __msa_asub_u_b(q2_in, q1_in); \
+ q3_asub_q2_m = __msa_asub_u_b(q3_in, q2_in); \
+ p0_asub_q0_m = __msa_asub_u_b(p0_in, q0_in); \
+ p1_asub_q1_m = __msa_asub_u_b(p1_in, q1_in); \
+ \
+ /* calculation of hev */ \
+ flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m); \
+ hev_out = thresh_in < (v16u8)flat_out; \
+ \
+ /* calculation of mask */ \
+ p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m); \
+ p1_asub_q1_m >>= 1; \
+ p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m); \
+ \
+ mask_out = b_limit_in < p0_asub_q0_m; \
+ mask_out = __msa_max_u_b(flat_out, mask_out); \
+ p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m); \
+ mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out); \
+ q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m); \
+ mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out); \
+ \
+ mask_out = limit_in < (v16u8)mask_out; \
+ mask_out = __msa_xori_b(mask_out, 0xff); \
+}
+#endif /* VPX_DSP_LOOPFILTER_MSA_H_ */
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -13,6 +13,36 @@
DSP_SRCS-$(HAVE_MSA) += mips/macros_msa.h
+# loop filters
+DSP_SRCS-yes += loopfilter.c
+
+DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64) += x86/loopfilter_sse2.c
+DSP_SRCS-$(HAVE_AVX2) += x86/loopfilter_avx2.c
+DSP_SRCS-$(HAVE_MMX) += x86/loopfilter_mmx.asm
+
+DSP_SRCS-$(HAVE_NEON) += arm/loopfilter_neon.c
+ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes += arm/loopfilter_mb_neon$(ASM)
+DSP_SRCS-yes += arm/loopfilter_16_neon$(ASM)
+DSP_SRCS-yes += arm/loopfilter_8_neon$(ASM)
+DSP_SRCS-yes += arm/loopfilter_4_neon$(ASM)
+else
+ifeq ($(HAVE_NEON),yes)
+DSP_SRCS-yes += arm/loopfilter_16_neon.c
+DSP_SRCS-yes += arm/loopfilter_8_neon.c
+DSP_SRCS-yes += arm/loopfilter_4_neon.c
+endif # HAVE_NEON
+endif # HAVE_NEON_ASM
+
+DSP_SRCS-$(HAVE_MSA) += mips/loopfilter_msa.h
+DSP_SRCS-$(HAVE_MSA) += mips/loopfilter_16_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/loopfilter_8_msa.c
+DSP_SRCS-$(HAVE_MSA) += mips/loopfilter_4_msa.c
+
+ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
+DSP_SRCS-$(HAVE_SSE2) += x86/highbd_loopfilter_sse2.c
+endif # CONFIG_VP9_HIGHBITDEPTH
+
ifeq ($(CONFIG_ENCODERS),yes)
DSP_SRCS-yes += sad.c
DSP_SRCS-yes += subtract.c
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -35,6 +35,82 @@
$avx_x86_64 = $avx2_x86_64 = '';
}
+#
+# Loopfilter
+#
+add_proto qw/void vp9_lpf_vertical_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vp9_lpf_vertical_16 sse2 neon_asm msa/;
+$vp9_lpf_vertical_16_neon_asm=vp9_lpf_vertical_16_neon;
+
+add_proto qw/void vp9_lpf_vertical_16_dual/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
+specialize qw/vp9_lpf_vertical_16_dual sse2 neon_asm msa/;
+$vp9_lpf_vertical_16_dual_neon_asm=vp9_lpf_vertical_16_dual_neon;
+
+add_proto qw/void vp9_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
+specialize qw/vp9_lpf_vertical_8 sse2 neon msa/;
+
+add_proto qw/void vp9_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/vp9_lpf_vertical_8_dual sse2 neon_asm msa/;
+$vp9_lpf_vertical_8_dual_neon_asm=vp9_lpf_vertical_8_dual_neon;
+
+add_proto qw/void vp9_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
+specialize qw/vp9_lpf_vertical_4 mmx neon msa/;
+
+add_proto qw/void vp9_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/vp9_lpf_vertical_4_dual sse2 neon msa/;
+
+add_proto qw/void vp9_lpf_horizontal_16/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
+specialize qw/vp9_lpf_horizontal_16 sse2 avx2 neon_asm msa/;
+$vp9_lpf_horizontal_16_neon_asm=vp9_lpf_horizontal_16_neon;
+
+add_proto qw/void vp9_lpf_horizontal_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
+specialize qw/vp9_lpf_horizontal_8 sse2 neon msa/;
+
+add_proto qw/void vp9_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/vp9_lpf_horizontal_8_dual sse2 neon_asm msa/;
+$vp9_lpf_horizontal_8_dual_neon_asm=vp9_lpf_horizontal_8_dual_neon;
+
+add_proto qw/void vp9_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count";
+specialize qw/vp9_lpf_horizontal_4 mmx neon msa/;
+
+add_proto qw/void vp9_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
+specialize qw/vp9_lpf_horizontal_4_dual sse2 neon msa/;
+
+if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {+ add_proto qw/void vp9_highbd_lpf_vertical_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+ specialize qw/vp9_highbd_lpf_vertical_16 sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_vertical_16_dual/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd";
+ specialize qw/vp9_highbd_lpf_vertical_16_dual sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_vertical_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
+ specialize qw/vp9_highbd_lpf_vertical_8 sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_vertical_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+ specialize qw/vp9_highbd_lpf_vertical_8_dual sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_vertical_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
+ specialize qw/vp9_highbd_lpf_vertical_4 sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_vertical_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+ specialize qw/vp9_highbd_lpf_vertical_4_dual sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_horizontal_16/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
+ specialize qw/vp9_highbd_lpf_horizontal_16 sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_horizontal_8/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
+ specialize qw/vp9_highbd_lpf_horizontal_8 sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_horizontal_8_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+ specialize qw/vp9_highbd_lpf_horizontal_8_dual sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_horizontal_4/, "uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int count, int bd";
+ specialize qw/vp9_highbd_lpf_horizontal_4 sse2/;
+
+ add_proto qw/void vp9_highbd_lpf_horizontal_4_dual/, "uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd";
+ specialize qw/vp9_highbd_lpf_horizontal_4_dual sse2/;
+} # CONFIG_VP9_HIGHBITDEPTH
+
if (vpx_config("CONFIG_ENCODERS") eq "yes") {#
# Block subtraction
--- /dev/null
+++ b/vpx_dsp/x86/highbd_loopfilter_sse2.c
@@ -1,0 +1,1215 @@
+/*
+ * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h> // SSE2
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_loopfilter.h"
+#include "vpx_ports/emmintrin_compat.h"
+
+static INLINE __m128i signed_char_clamp_bd_sse2(__m128i value, int bd) {+ __m128i ubounded;
+ __m128i lbounded;
+ __m128i retval;
+
+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i one = _mm_set1_epi16(1);
+ __m128i t80, max, min;
+
+ if (bd == 8) {+ t80 = _mm_set1_epi16(0x80);
+ max = _mm_subs_epi16(
+ _mm_subs_epi16(_mm_slli_epi16(one, 8), one), t80);
+ } else if (bd == 10) {+ t80 = _mm_set1_epi16(0x200);
+ max = _mm_subs_epi16(
+ _mm_subs_epi16(_mm_slli_epi16(one, 10), one), t80);
+ } else { // bd == 12+ t80 = _mm_set1_epi16(0x800);
+ max = _mm_subs_epi16(
+ _mm_subs_epi16(_mm_slli_epi16(one, 12), one), t80);
+ }
+
+ min = _mm_subs_epi16(zero, t80);
+
+ ubounded = _mm_cmpgt_epi16(value, max);
+ lbounded = _mm_cmplt_epi16(value, min);
+ retval = _mm_andnot_si128(_mm_or_si128(ubounded, lbounded), value);
+ ubounded = _mm_and_si128(ubounded, max);
+ lbounded = _mm_and_si128(lbounded, min);
+ retval = _mm_or_si128(retval, ubounded);
+ retval = _mm_or_si128(retval, lbounded);
+ return retval;
+}
+
+// TODO(debargha, peter): Break up large functions into smaller ones
+// in this file.
+static void highbd_mb_lpf_horizontal_edge_w_sse2_8(uint16_t *s,
+ int p,
+ const uint8_t *_blimit,
+ const uint8_t *_limit,
+ const uint8_t *_thresh,
+ int bd) {+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i one = _mm_set1_epi16(1);
+ __m128i blimit, limit, thresh;
+ __m128i q7, p7, q6, p6, q5, p5, q4, p4, q3, p3, q2, p2, q1, p1, q0, p0;
+ __m128i mask, hev, flat, flat2, abs_p1p0, abs_q1q0;
+ __m128i ps1, qs1, ps0, qs0;
+ __m128i abs_p0q0, abs_p1q1, ffff, work;
+ __m128i filt, work_a, filter1, filter2;
+ __m128i flat2_q6, flat2_p6, flat2_q5, flat2_p5, flat2_q4, flat2_p4;
+ __m128i flat2_q3, flat2_p3, flat2_q2, flat2_p2, flat2_q1, flat2_p1;
+ __m128i flat2_q0, flat2_p0;
+ __m128i flat_q2, flat_p2, flat_q1, flat_p1, flat_q0, flat_p0;
+ __m128i pixelFilter_p, pixelFilter_q;
+ __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+ __m128i sum_p7, sum_q7, sum_p3, sum_q3;
+ __m128i t4, t3, t80, t1;
+ __m128i eight, four;
+
+ if (bd == 8) {+ blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+ limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+ thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+ } else if (bd == 10) {+ blimit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+ limit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+ thresh = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+ } else { // bd == 12+ blimit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+ limit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+ thresh = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+ }
+
+ q4 = _mm_load_si128((__m128i *)(s + 4 * p));
+ p4 = _mm_load_si128((__m128i *)(s - 5 * p));
+ q3 = _mm_load_si128((__m128i *)(s + 3 * p));
+ p3 = _mm_load_si128((__m128i *)(s - 4 * p));
+ q2 = _mm_load_si128((__m128i *)(s + 2 * p));
+ p2 = _mm_load_si128((__m128i *)(s - 3 * p));
+ q1 = _mm_load_si128((__m128i *)(s + 1 * p));
+ p1 = _mm_load_si128((__m128i *)(s - 2 * p));
+ q0 = _mm_load_si128((__m128i *)(s + 0 * p));
+ p0 = _mm_load_si128((__m128i *)(s - 1 * p));
+
+ // highbd_filter_mask
+ abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0), _mm_subs_epu16(p0, p1));
+ abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0), _mm_subs_epu16(q0, q1));
+
+ ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+
+ abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0), _mm_subs_epu16(q0, p0));
+ abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1), _mm_subs_epu16(q1, p1));
+
+ // highbd_hev_mask (in C code this is actually called from highbd_filter4)
+ flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu16(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+ abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0); // abs(p0 - q0) * 2
+ abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1); // abs(p1 - q1) / 2
+ mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+ mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p1, p0),
+ _mm_subs_epu16(p0, p1)),
+ _mm_or_si128(_mm_subs_epu16(q1, q0),
+ _mm_subs_epu16(q0, q1)));
+ mask = _mm_max_epi16(work, mask);
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+ _mm_subs_epu16(p1, p2)),
+ _mm_or_si128(_mm_subs_epu16(q2, q1),
+ _mm_subs_epu16(q1, q2)));
+ mask = _mm_max_epi16(work, mask);
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p2),
+ _mm_subs_epu16(p2, p3)),
+ _mm_or_si128(_mm_subs_epu16(q3, q2),
+ _mm_subs_epu16(q2, q3)));
+ mask = _mm_max_epi16(work, mask);
+
+ mask = _mm_subs_epu16(mask, limit);
+ mask = _mm_cmpeq_epi16(mask, zero); // return ~mask
+
+ // lp filter
+ // highbd_filter4
+ t4 = _mm_set1_epi16(4);
+ t3 = _mm_set1_epi16(3);
+ if (bd == 8)
+ t80 = _mm_set1_epi16(0x80);
+ else if (bd == 10)
+ t80 = _mm_set1_epi16(0x200);
+ else // bd == 12
+ t80 = _mm_set1_epi16(0x800);
+
+ t1 = _mm_set1_epi16(0x1);
+
+ ps1 = _mm_subs_epi16(p1, t80);
+ qs1 = _mm_subs_epi16(q1, t80);
+ ps0 = _mm_subs_epi16(p0, t80);
+ qs0 = _mm_subs_epi16(q0, t80);
+
+ filt = _mm_and_si128(
+ signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd), hev);
+ work_a = _mm_subs_epi16(qs0, ps0);
+ filt = _mm_adds_epi16(filt, work_a);
+ filt = _mm_adds_epi16(filt, work_a);
+ filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+ filt = _mm_and_si128(filt, mask);
+ filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
+ filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
+
+ // Filter1 >> 3
+ filter1 = _mm_srai_epi16(filter1, 0x3);
+ filter2 = _mm_srai_epi16(filter2, 0x3);
+
+ qs0 = _mm_adds_epi16(
+ signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd),
+ t80);
+ ps0 = _mm_adds_epi16(
+ signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd),
+ t80);
+ filt = _mm_adds_epi16(filter1, t1);
+ filt = _mm_srai_epi16(filt, 1);
+ filt = _mm_andnot_si128(hev, filt);
+ qs1 = _mm_adds_epi16(
+ signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd),
+ t80);
+ ps1 = _mm_adds_epi16(
+ signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd),
+ t80);
+
+ // end highbd_filter4
+ // loopfilter done
+
+ // highbd_flat_mask4
+ flat = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p0),
+ _mm_subs_epu16(p0, p2)),
+ _mm_or_si128(_mm_subs_epu16(p3, p0),
+ _mm_subs_epu16(p0, p3)));
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(q2, q0),
+ _mm_subs_epu16(q0, q2)),
+ _mm_or_si128(_mm_subs_epu16(q3, q0),
+ _mm_subs_epu16(q0, q3)));
+ flat = _mm_max_epi16(work, flat);
+ work = _mm_max_epi16(abs_p1p0, abs_q1q0);
+ flat = _mm_max_epi16(work, flat);
+
+ if (bd == 8)
+ flat = _mm_subs_epu16(flat, one);
+ else if (bd == 10)
+ flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
+ else // bd == 12
+ flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
+
+ flat = _mm_cmpeq_epi16(flat, zero);
+ // end flat_mask4
+
+ // flat & mask = flat && mask (as used in filter8)
+ // (because, in both vars, each block of 16 either all 1s or all 0s)
+ flat = _mm_and_si128(flat, mask);
+
+ p5 = _mm_load_si128((__m128i *)(s - 6 * p));
+ q5 = _mm_load_si128((__m128i *)(s + 5 * p));
+ p6 = _mm_load_si128((__m128i *)(s - 7 * p));
+ q6 = _mm_load_si128((__m128i *)(s + 6 * p));
+ p7 = _mm_load_si128((__m128i *)(s - 8 * p));
+ q7 = _mm_load_si128((__m128i *)(s + 7 * p));
+
+ // highbd_flat_mask5 (arguments passed in are p0, q0, p4-p7, q4-q7
+ // but referred to as p0-p4 & q0-q4 in fn)
+ flat2 = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p4, p0),
+ _mm_subs_epu16(p0, p4)),
+ _mm_or_si128(_mm_subs_epu16(q4, q0),
+ _mm_subs_epu16(q0, q4)));
+
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p5, p0),
+ _mm_subs_epu16(p0, p5)),
+ _mm_or_si128(_mm_subs_epu16(q5, q0),
+ _mm_subs_epu16(q0, q5)));
+ flat2 = _mm_max_epi16(work, flat2);
+
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p6, p0),
+ _mm_subs_epu16(p0, p6)),
+ _mm_or_si128(_mm_subs_epu16(q6, q0),
+ _mm_subs_epu16(q0, q6)));
+ flat2 = _mm_max_epi16(work, flat2);
+
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p7, p0),
+ _mm_subs_epu16(p0, p7)),
+ _mm_or_si128(_mm_subs_epu16(q7, q0),
+ _mm_subs_epu16(q0, q7)));
+ flat2 = _mm_max_epi16(work, flat2);
+
+ if (bd == 8)
+ flat2 = _mm_subs_epu16(flat2, one);
+ else if (bd == 10)
+ flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 2));
+ else // bd == 12
+ flat2 = _mm_subs_epu16(flat2, _mm_slli_epi16(one, 4));
+
+ flat2 = _mm_cmpeq_epi16(flat2, zero);
+ flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
+ // end highbd_flat_mask5
+
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // flat and wide flat calculations
+ eight = _mm_set1_epi16(8);
+ four = _mm_set1_epi16(4);
+
+ pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6, p5),
+ _mm_add_epi16(p4, p3));
+ pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6, q5),
+ _mm_add_epi16(q4, q3));
+
+ pixetFilter_p2p1p0 = _mm_add_epi16(p0, _mm_add_epi16(p2, p1));
+ pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+ pixetFilter_q2q1q0 = _mm_add_epi16(q0, _mm_add_epi16(q2, q1));
+ pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+ pixelFilter_p = _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
+ pixelFilter_q));
+ pixetFilter_p2p1p0 = _mm_add_epi16(four,
+ _mm_add_epi16(pixetFilter_p2p1p0,
+ pixetFilter_q2q1q0));
+ flat2_p0 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(p7, p0)), 4);
+ flat2_q0 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(q7, q0)), 4);
+ flat_p0 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(p3, p0)), 3);
+ flat_q0 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(q3, q0)), 3);
+
+ sum_p7 = _mm_add_epi16(p7, p7);
+ sum_q7 = _mm_add_epi16(q7, q7);
+ sum_p3 = _mm_add_epi16(p3, p3);
+ sum_q3 = _mm_add_epi16(q3, q3);
+
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6);
+ flat2_p1 = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1)), 4);
+ flat2_q1 = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1)), 4);
+
+ pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2);
+ pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2);
+ flat_p1 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(sum_p3, p1)), 3);
+ flat_q1 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+ _mm_add_epi16(sum_q3, q1)), 3);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7);
+ sum_q7 = _mm_add_epi16(sum_q7, q7);
+ sum_p3 = _mm_add_epi16(sum_p3, p3);
+ sum_q3 = _mm_add_epi16(sum_q3, q3);
+
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5);
+ flat2_p2 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p2)), 4);
+ flat2_q2 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q2)), 4);
+
+ pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1);
+ pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1);
+ flat_p2 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(sum_p3, p2)), 3);
+ flat_q2 = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+ _mm_add_epi16(sum_q3, q2)), 3);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7);
+ sum_q7 = _mm_add_epi16(sum_q7, q7);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4);
+ flat2_p3 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p3)), 4);
+ flat2_q3 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q3)), 4);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7);
+ sum_q7 = _mm_add_epi16(sum_q7, q7);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3);
+ flat2_p4 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p4)), 4);
+ flat2_q4 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q4)), 4);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7);
+ sum_q7 = _mm_add_epi16(sum_q7, q7);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2);
+ flat2_p5 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p5)), 4);
+ flat2_q5 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q5)), 4);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7);
+ sum_q7 = _mm_add_epi16(sum_q7, q7);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1);
+ flat2_p6 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p6)), 4);
+ flat2_q6 = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q6)), 4);
+
+ // wide flat
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ // highbd_filter8
+ p2 = _mm_andnot_si128(flat, p2);
+ // p2 remains unchanged if !(flat && mask)
+ flat_p2 = _mm_and_si128(flat, flat_p2);
+ // when (flat && mask)
+ p2 = _mm_or_si128(p2, flat_p2); // full list of p2 values
+ q2 = _mm_andnot_si128(flat, q2);
+ flat_q2 = _mm_and_si128(flat, flat_q2);
+ q2 = _mm_or_si128(q2, flat_q2); // full list of q2 values
+
+ ps1 = _mm_andnot_si128(flat, ps1);
+ // p1 takes the value assigned to in in filter4 if !(flat && mask)
+ flat_p1 = _mm_and_si128(flat, flat_p1);
+ // when (flat && mask)
+ p1 = _mm_or_si128(ps1, flat_p1); // full list of p1 values
+ qs1 = _mm_andnot_si128(flat, qs1);
+ flat_q1 = _mm_and_si128(flat, flat_q1);
+ q1 = _mm_or_si128(qs1, flat_q1); // full list of q1 values
+
+ ps0 = _mm_andnot_si128(flat, ps0);
+ // p0 takes the value assigned to in in filter4 if !(flat && mask)
+ flat_p0 = _mm_and_si128(flat, flat_p0);
+ // when (flat && mask)
+ p0 = _mm_or_si128(ps0, flat_p0); // full list of p0 values
+ qs0 = _mm_andnot_si128(flat, qs0);
+ flat_q0 = _mm_and_si128(flat, flat_q0);
+ q0 = _mm_or_si128(qs0, flat_q0); // full list of q0 values
+ // end highbd_filter8
+
+ // highbd_filter16
+ p6 = _mm_andnot_si128(flat2, p6);
+ // p6 remains unchanged if !(flat2 && flat && mask)
+ flat2_p6 = _mm_and_si128(flat2, flat2_p6);
+ // get values for when (flat2 && flat && mask)
+ p6 = _mm_or_si128(p6, flat2_p6); // full list of p6 values
+ q6 = _mm_andnot_si128(flat2, q6);
+ // q6 remains unchanged if !(flat2 && flat && mask)
+ flat2_q6 = _mm_and_si128(flat2, flat2_q6);
+ // get values for when (flat2 && flat && mask)
+ q6 = _mm_or_si128(q6, flat2_q6); // full list of q6 values
+ _mm_store_si128((__m128i *)(s - 7 * p), p6);
+ _mm_store_si128((__m128i *)(s + 6 * p), q6);
+
+ p5 = _mm_andnot_si128(flat2, p5);
+ // p5 remains unchanged if !(flat2 && flat && mask)
+ flat2_p5 = _mm_and_si128(flat2, flat2_p5);
+ // get values for when (flat2 && flat && mask)
+ p5 = _mm_or_si128(p5, flat2_p5);
+ // full list of p5 values
+ q5 = _mm_andnot_si128(flat2, q5);
+ // q5 remains unchanged if !(flat2 && flat && mask)
+ flat2_q5 = _mm_and_si128(flat2, flat2_q5);
+ // get values for when (flat2 && flat && mask)
+ q5 = _mm_or_si128(q5, flat2_q5);
+ // full list of q5 values
+ _mm_store_si128((__m128i *)(s - 6 * p), p5);
+ _mm_store_si128((__m128i *)(s + 5 * p), q5);
+
+ p4 = _mm_andnot_si128(flat2, p4);
+ // p4 remains unchanged if !(flat2 && flat && mask)
+ flat2_p4 = _mm_and_si128(flat2, flat2_p4);
+ // get values for when (flat2 && flat && mask)
+ p4 = _mm_or_si128(p4, flat2_p4); // full list of p4 values
+ q4 = _mm_andnot_si128(flat2, q4);
+ // q4 remains unchanged if !(flat2 && flat && mask)
+ flat2_q4 = _mm_and_si128(flat2, flat2_q4);
+ // get values for when (flat2 && flat && mask)
+ q4 = _mm_or_si128(q4, flat2_q4); // full list of q4 values
+ _mm_store_si128((__m128i *)(s - 5 * p), p4);
+ _mm_store_si128((__m128i *)(s + 4 * p), q4);
+
+ p3 = _mm_andnot_si128(flat2, p3);
+ // p3 takes value from highbd_filter8 if !(flat2 && flat && mask)
+ flat2_p3 = _mm_and_si128(flat2, flat2_p3);
+ // get values for when (flat2 && flat && mask)
+ p3 = _mm_or_si128(p3, flat2_p3); // full list of p3 values
+ q3 = _mm_andnot_si128(flat2, q3);
+ // q3 takes value from highbd_filter8 if !(flat2 && flat && mask)
+ flat2_q3 = _mm_and_si128(flat2, flat2_q3);
+ // get values for when (flat2 && flat && mask)
+ q3 = _mm_or_si128(q3, flat2_q3); // full list of q3 values
+ _mm_store_si128((__m128i *)(s - 4 * p), p3);
+ _mm_store_si128((__m128i *)(s + 3 * p), q3);
+
+ p2 = _mm_andnot_si128(flat2, p2);
+ // p2 takes value from highbd_filter8 if !(flat2 && flat && mask)
+ flat2_p2 = _mm_and_si128(flat2, flat2_p2);
+ // get values for when (flat2 && flat && mask)
+ p2 = _mm_or_si128(p2, flat2_p2);
+ // full list of p2 values
+ q2 = _mm_andnot_si128(flat2, q2);
+ // q2 takes value from highbd_filter8 if !(flat2 && flat && mask)
+ flat2_q2 = _mm_and_si128(flat2, flat2_q2);
+ // get values for when (flat2 && flat && mask)
+ q2 = _mm_or_si128(q2, flat2_q2); // full list of q2 values
+ _mm_store_si128((__m128i *)(s - 3 * p), p2);
+ _mm_store_si128((__m128i *)(s + 2 * p), q2);
+
+ p1 = _mm_andnot_si128(flat2, p1);
+ // p1 takes value from highbd_filter8 if !(flat2 && flat && mask)
+ flat2_p1 = _mm_and_si128(flat2, flat2_p1);
+ // get values for when (flat2 && flat && mask)
+ p1 = _mm_or_si128(p1, flat2_p1); // full list of p1 values
+ q1 = _mm_andnot_si128(flat2, q1);
+ // q1 takes value from highbd_filter8 if !(flat2 && flat && mask)
+ flat2_q1 = _mm_and_si128(flat2, flat2_q1);
+ // get values for when (flat2 && flat && mask)
+ q1 = _mm_or_si128(q1, flat2_q1); // full list of q1 values
+ _mm_store_si128((__m128i *)(s - 2 * p), p1);
+ _mm_store_si128((__m128i *)(s + 1 * p), q1);
+
+ p0 = _mm_andnot_si128(flat2, p0);
+ // p0 takes value from highbd_filter8 if !(flat2 && flat && mask)
+ flat2_p0 = _mm_and_si128(flat2, flat2_p0);
+ // get values for when (flat2 && flat && mask)
+ p0 = _mm_or_si128(p0, flat2_p0); // full list of p0 values
+ q0 = _mm_andnot_si128(flat2, q0);
+ // q0 takes value from highbd_filter8 if !(flat2 && flat && mask)
+ flat2_q0 = _mm_and_si128(flat2, flat2_q0);
+ // get values for when (flat2 && flat && mask)
+ q0 = _mm_or_si128(q0, flat2_q0); // full list of q0 values
+ _mm_store_si128((__m128i *)(s - 1 * p), p0);
+ _mm_store_si128((__m128i *)(s - 0 * p), q0);
+}
+
+static void highbd_mb_lpf_horizontal_edge_w_sse2_16(uint16_t *s,
+ int p,
+ const uint8_t *_blimit,
+ const uint8_t *_limit,
+ const uint8_t *_thresh,
+ int bd) {+ highbd_mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh, bd);
+ highbd_mb_lpf_horizontal_edge_w_sse2_8(s + 8, p, _blimit, _limit, _thresh,
+ bd);
+}
+
+// TODO(yunqingwang): remove count and call these 2 functions(8 or 16) directly.
+void vp9_highbd_lpf_horizontal_16_sse2(uint16_t *s, int p,
+ const uint8_t *_blimit,
+ const uint8_t *_limit,
+ const uint8_t *_thresh,
+ int count, int bd) {+ if (count == 1)
+ highbd_mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh, bd);
+ else
+ highbd_mb_lpf_horizontal_edge_w_sse2_16(s, p, _blimit, _limit, _thresh, bd);
+}
+
+void vp9_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p,
+ const uint8_t *_blimit,
+ const uint8_t *_limit,
+ const uint8_t *_thresh,
+ int count, int bd) {+ DECLARE_ALIGNED(16, uint16_t, flat_op2[16]);
+ DECLARE_ALIGNED(16, uint16_t, flat_op1[16]);
+ DECLARE_ALIGNED(16, uint16_t, flat_op0[16]);
+ DECLARE_ALIGNED(16, uint16_t, flat_oq2[16]);
+ DECLARE_ALIGNED(16, uint16_t, flat_oq1[16]);
+ DECLARE_ALIGNED(16, uint16_t, flat_oq0[16]);
+ const __m128i zero = _mm_set1_epi16(0);
+ __m128i blimit, limit, thresh;
+ __m128i mask, hev, flat;
+ __m128i p3 = _mm_load_si128((__m128i *)(s - 4 * p));
+ __m128i q3 = _mm_load_si128((__m128i *)(s + 3 * p));
+ __m128i p2 = _mm_load_si128((__m128i *)(s - 3 * p));
+ __m128i q2 = _mm_load_si128((__m128i *)(s + 2 * p));
+ __m128i p1 = _mm_load_si128((__m128i *)(s - 2 * p));
+ __m128i q1 = _mm_load_si128((__m128i *)(s + 1 * p));
+ __m128i p0 = _mm_load_si128((__m128i *)(s - 1 * p));
+ __m128i q0 = _mm_load_si128((__m128i *)(s + 0 * p));
+ const __m128i one = _mm_set1_epi16(1);
+ const __m128i ffff = _mm_cmpeq_epi16(one, one);
+ __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+ const __m128i four = _mm_set1_epi16(4);
+ __m128i workp_a, workp_b, workp_shft;
+
+ const __m128i t4 = _mm_set1_epi16(4);
+ const __m128i t3 = _mm_set1_epi16(3);
+ __m128i t80;
+ const __m128i t1 = _mm_set1_epi16(0x1);
+ __m128i ps1, ps0, qs0, qs1;
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+
+ (void)count;
+
+ if (bd == 8) {+ blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+ limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+ thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+ t80 = _mm_set1_epi16(0x80);
+ } else if (bd == 10) {+ blimit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+ limit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+ thresh = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+ t80 = _mm_set1_epi16(0x200);
+ } else { // bd == 12+ blimit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+ limit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+ thresh = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+ t80 = _mm_set1_epi16(0x800);
+ }
+
+ ps1 = _mm_subs_epi16(p1, t80);
+ ps0 = _mm_subs_epi16(p0, t80);
+ qs0 = _mm_subs_epi16(q0, t80);
+ qs1 = _mm_subs_epi16(q1, t80);
+
+ // filter_mask and hev_mask
+ abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0),
+ _mm_subs_epu16(p0, p1));
+ abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0),
+ _mm_subs_epu16(q0, q1));
+
+ abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0),
+ _mm_subs_epu16(q0, p0));
+ abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1),
+ _mm_subs_epu16(q1, p1));
+ flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu16(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+ abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+ mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ // So taking maximums continues to work:
+ mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+ mask = _mm_max_epi16(abs_p1p0, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ mask = _mm_max_epi16(abs_q1q0, mask);
+ // mask |= (abs(q1 - q0) > limit) * -1;
+
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+ _mm_subs_epu16(p1, p2)),
+ _mm_or_si128(_mm_subs_epu16(q2, q1),
+ _mm_subs_epu16(q1, q2)));
+ mask = _mm_max_epi16(work, mask);
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p2),
+ _mm_subs_epu16(p2, p3)),
+ _mm_or_si128(_mm_subs_epu16(q3, q2),
+ _mm_subs_epu16(q2, q3)));
+ mask = _mm_max_epi16(work, mask);
+ mask = _mm_subs_epu16(mask, limit);
+ mask = _mm_cmpeq_epi16(mask, zero);
+
+ // flat_mask4
+ flat = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p0),
+ _mm_subs_epu16(p0, p2)),
+ _mm_or_si128(_mm_subs_epu16(q2, q0),
+ _mm_subs_epu16(q0, q2)));
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p3, p0),
+ _mm_subs_epu16(p0, p3)),
+ _mm_or_si128(_mm_subs_epu16(q3, q0),
+ _mm_subs_epu16(q0, q3)));
+ flat = _mm_max_epi16(work, flat);
+ flat = _mm_max_epi16(abs_p1p0, flat);
+ flat = _mm_max_epi16(abs_q1q0, flat);
+
+ if (bd == 8)
+ flat = _mm_subs_epu16(flat, one);
+ else if (bd == 10)
+ flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 2));
+ else // bd == 12
+ flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, 4));
+
+ flat = _mm_cmpeq_epi16(flat, zero);
+ flat = _mm_and_si128(flat, mask); // flat & mask
+
+ // Added before shift for rounding part of ROUND_POWER_OF_TWO
+
+ workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+ workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_store_si128((__m128i *)&flat_op2[0], workp_shft);
+
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_store_si128((__m128i *)&flat_op1[0], workp_shft);
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_store_si128((__m128i *)&flat_op0[0], workp_shft);
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_store_si128((__m128i *)&flat_oq0[0], workp_shft);
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_store_si128((__m128i *)&flat_oq1[0], workp_shft);
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_store_si128((__m128i *)&flat_oq2[0], workp_shft);
+
+ // lp filter
+ filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
+ filt = _mm_and_si128(filt, hev);
+ work_a = _mm_subs_epi16(qs0, ps0);
+ filt = _mm_adds_epi16(filt, work_a);
+ filt = _mm_adds_epi16(filt, work_a);
+ filt = _mm_adds_epi16(filt, work_a);
+ // (vp9_filter + 3 * (qs0 - ps0)) & mask
+ filt = signed_char_clamp_bd_sse2(filt, bd);
+ filt = _mm_and_si128(filt, mask);
+
+ filter1 = _mm_adds_epi16(filt, t4);
+ filter2 = _mm_adds_epi16(filt, t3);
+
+ // Filter1 >> 3
+ filter1 = signed_char_clamp_bd_sse2(filter1, bd);
+ filter1 = _mm_srai_epi16(filter1, 3);
+
+ // Filter2 >> 3
+ filter2 = signed_char_clamp_bd_sse2(filter2, bd);
+ filter2 = _mm_srai_epi16(filter2, 3);
+
+ // filt >> 1
+ filt = _mm_adds_epi16(filter1, t1);
+ filt = _mm_srai_epi16(filt, 1);
+ // filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev;
+ filt = _mm_andnot_si128(hev, filt);
+
+ work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd);
+ work_a = _mm_adds_epi16(work_a, t80);
+ q0 = _mm_load_si128((__m128i *)flat_oq0);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q0 = _mm_and_si128(flat, q0);
+ q0 = _mm_or_si128(work_a, q0);
+
+ work_a = signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd);
+ work_a = _mm_adds_epi16(work_a, t80);
+ q1 = _mm_load_si128((__m128i *)flat_oq1);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q1 = _mm_and_si128(flat, q1);
+ q1 = _mm_or_si128(work_a, q1);
+
+ work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+ q2 = _mm_load_si128((__m128i *)flat_oq2);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q2 = _mm_and_si128(flat, q2);
+ q2 = _mm_or_si128(work_a, q2);
+
+ work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd);
+ work_a = _mm_adds_epi16(work_a, t80);
+ p0 = _mm_load_si128((__m128i *)flat_op0);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p0 = _mm_and_si128(flat, p0);
+ p0 = _mm_or_si128(work_a, p0);
+
+ work_a = signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd);
+ work_a = _mm_adds_epi16(work_a, t80);
+ p1 = _mm_load_si128((__m128i *)flat_op1);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p1 = _mm_and_si128(flat, p1);
+ p1 = _mm_or_si128(work_a, p1);
+
+ work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+ p2 = _mm_load_si128((__m128i *)flat_op2);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p2 = _mm_and_si128(flat, p2);
+ p2 = _mm_or_si128(work_a, p2);
+
+ _mm_store_si128((__m128i *)(s - 3 * p), p2);
+ _mm_store_si128((__m128i *)(s - 2 * p), p1);
+ _mm_store_si128((__m128i *)(s - 1 * p), p0);
+ _mm_store_si128((__m128i *)(s + 0 * p), q0);
+ _mm_store_si128((__m128i *)(s + 1 * p), q1);
+ _mm_store_si128((__m128i *)(s + 2 * p), q2);
+}
+
+void vp9_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int p,
+ const uint8_t *_blimit0,
+ const uint8_t *_limit0,
+ const uint8_t *_thresh0,
+ const uint8_t *_blimit1,
+ const uint8_t *_limit1,
+ const uint8_t *_thresh1,
+ int bd) {+ vp9_highbd_lpf_horizontal_8_sse2(s, p, _blimit0, _limit0, _thresh0, 1, bd);
+ vp9_highbd_lpf_horizontal_8_sse2(s + 8, p, _blimit1, _limit1, _thresh1,
+ 1, bd);
+}
+
+void vp9_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
+ const uint8_t *_blimit,
+ const uint8_t *_limit,
+ const uint8_t *_thresh,
+ int count, int bd) {+ const __m128i zero = _mm_set1_epi16(0);
+ __m128i blimit, limit, thresh;
+ __m128i mask, hev, flat;
+ __m128i p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+ __m128i p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+ __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+ __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+ __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+ __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+ __m128i q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+ __m128i q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+ const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu16(p1, p0),
+ _mm_subs_epu16(p0, p1));
+ const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu16(q1, q0),
+ _mm_subs_epu16(q0, q1));
+ const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+ const __m128i one = _mm_set1_epi16(1);
+ __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu16(p0, q0),
+ _mm_subs_epu16(q0, p0));
+ __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu16(p1, q1),
+ _mm_subs_epu16(q1, p1));
+ __m128i work;
+ const __m128i t4 = _mm_set1_epi16(4);
+ const __m128i t3 = _mm_set1_epi16(3);
+ __m128i t80;
+ __m128i tff80;
+ __m128i tffe0;
+ __m128i t1f;
+ // equivalent to shifting 0x1f left by bitdepth - 8
+ // and setting new bits to 1
+ const __m128i t1 = _mm_set1_epi16(0x1);
+ __m128i t7f;
+ // equivalent to shifting 0x7f left by bitdepth - 8
+ // and setting new bits to 1
+ __m128i ps1, ps0, qs0, qs1;
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+
+ (void)count;
+
+ if (bd == 8) {+ blimit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero);
+ limit = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero);
+ thresh = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero);
+ t80 = _mm_set1_epi16(0x80);
+ tff80 = _mm_set1_epi16(0xff80);
+ tffe0 = _mm_set1_epi16(0xffe0);
+ t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 8);
+ t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 8);
+ } else if (bd == 10) {+ blimit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 2);
+ limit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 2);
+ thresh = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 2);
+ t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 2);
+ tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 2);
+ tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 2);
+ t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 6);
+ t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 6);
+ } else { // bd == 12+ blimit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit), zero), 4);
+ limit = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit), zero), 4);
+ thresh = _mm_slli_epi16(
+ _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh), zero), 4);
+ t80 = _mm_slli_epi16(_mm_set1_epi16(0x80), 4);
+ tff80 = _mm_slli_epi16(_mm_set1_epi16(0xff80), 4);
+ tffe0 = _mm_slli_epi16(_mm_set1_epi16(0xffe0), 4);
+ t1f = _mm_srli_epi16(_mm_set1_epi16(0x1fff), 4);
+ t7f = _mm_srli_epi16(_mm_set1_epi16(0x7fff), 4);
+ }
+
+ ps1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 2 * p)), t80);
+ ps0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s - 1 * p)), t80);
+ qs0 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 0 * p)), t80);
+ qs1 = _mm_subs_epi16(_mm_loadu_si128((__m128i *)(s + 1 * p)), t80);
+
+ // filter_mask and hev_mask
+ flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu16(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi16(hev, zero), ffff);
+
+ abs_p0q0 =_mm_adds_epu16(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+ mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ // So taking maximums continues to work:
+ mask = _mm_and_si128(mask, _mm_adds_epu16(limit, one));
+ mask = _mm_max_epi16(flat, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(p2, p1),
+ _mm_subs_epu16(p1, p2)),
+ _mm_or_si128(_mm_subs_epu16(p3, p2),
+ _mm_subs_epu16(p2, p3)));
+ mask = _mm_max_epi16(work, mask);
+ work = _mm_max_epi16(_mm_or_si128(_mm_subs_epu16(q2, q1),
+ _mm_subs_epu16(q1, q2)),
+ _mm_or_si128(_mm_subs_epu16(q3, q2),
+ _mm_subs_epu16(q2, q3)));
+ mask = _mm_max_epi16(work, mask);
+ mask = _mm_subs_epu16(mask, limit);
+ mask = _mm_cmpeq_epi16(mask, zero);
+
+ // filter4
+ filt = signed_char_clamp_bd_sse2(_mm_subs_epi16(ps1, qs1), bd);
+ filt = _mm_and_si128(filt, hev);
+ work_a = _mm_subs_epi16(qs0, ps0);
+ filt = _mm_adds_epi16(filt, work_a);
+ filt = _mm_adds_epi16(filt, work_a);
+ filt = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, work_a), bd);
+
+ // (vp9_filter + 3 * (qs0 - ps0)) & mask
+ filt = _mm_and_si128(filt, mask);
+
+ filter1 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t4), bd);
+ filter2 = signed_char_clamp_bd_sse2(_mm_adds_epi16(filt, t3), bd);
+
+ // Filter1 >> 3
+ work_a = _mm_cmpgt_epi16(zero, filter1); // get the values that are <0
+ filter1 = _mm_srli_epi16(filter1, 3);
+ work_a = _mm_and_si128(work_a, tffe0); // sign bits for the values < 0
+ filter1 = _mm_and_si128(filter1, t1f); // clamp the range
+ filter1 = _mm_or_si128(filter1, work_a); // reinsert the sign bits
+
+ // Filter2 >> 3
+ work_a = _mm_cmpgt_epi16(zero, filter2);
+ filter2 = _mm_srli_epi16(filter2, 3);
+ work_a = _mm_and_si128(work_a, tffe0);
+ filter2 = _mm_and_si128(filter2, t1f);
+ filter2 = _mm_or_si128(filter2, work_a);
+
+ // filt >> 1
+ filt = _mm_adds_epi16(filter1, t1);
+ work_a = _mm_cmpgt_epi16(zero, filt);
+ filt = _mm_srli_epi16(filt, 1);
+ work_a = _mm_and_si128(work_a, tff80);
+ filt = _mm_and_si128(filt, t7f);
+ filt = _mm_or_si128(filt, work_a);
+
+ filt = _mm_andnot_si128(hev, filt);
+
+ q0 = _mm_adds_epi16(
+ signed_char_clamp_bd_sse2(_mm_subs_epi16(qs0, filter1), bd), t80);
+ q1 = _mm_adds_epi16(
+ signed_char_clamp_bd_sse2(_mm_subs_epi16(qs1, filt), bd), t80);
+ p0 = _mm_adds_epi16(
+ signed_char_clamp_bd_sse2(_mm_adds_epi16(ps0, filter2), bd), t80);
+ p1 = _mm_adds_epi16(
+ signed_char_clamp_bd_sse2(_mm_adds_epi16(ps1, filt), bd), t80);
+
+ _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+ _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+ _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+ _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+}
+
+void vp9_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int p,
+ const uint8_t *_blimit0,
+ const uint8_t *_limit0,
+ const uint8_t *_thresh0,
+ const uint8_t *_blimit1,
+ const uint8_t *_limit1,
+ const uint8_t *_thresh1,
+ int bd) {+ vp9_highbd_lpf_horizontal_4_sse2(s, p, _blimit0, _limit0, _thresh0, 1, bd);
+ vp9_highbd_lpf_horizontal_4_sse2(s + 8, p, _blimit1, _limit1, _thresh1, 1,
+ bd);
+}
+
+static INLINE void highbd_transpose(uint16_t *src[], int in_p,
+ uint16_t *dst[], int out_p,
+ int num_8x8_to_transpose) {+ int idx8x8 = 0;
+ __m128i p0, p1, p2, p3, p4, p5, p6, p7, x0, x1, x2, x3, x4, x5, x6, x7;
+ do {+ uint16_t *in = src[idx8x8];
+ uint16_t *out = dst[idx8x8];
+
+ p0 = _mm_loadu_si128((__m128i *)(in + 0*in_p)); // 00 01 02 03 04 05 06 07
+ p1 = _mm_loadu_si128((__m128i *)(in + 1*in_p)); // 10 11 12 13 14 15 16 17
+ p2 = _mm_loadu_si128((__m128i *)(in + 2*in_p)); // 20 21 22 23 24 25 26 27
+ p3 = _mm_loadu_si128((__m128i *)(in + 3*in_p)); // 30 31 32 33 34 35 36 37
+ p4 = _mm_loadu_si128((__m128i *)(in + 4*in_p)); // 40 41 42 43 44 45 46 47
+ p5 = _mm_loadu_si128((__m128i *)(in + 5*in_p)); // 50 51 52 53 54 55 56 57
+ p6 = _mm_loadu_si128((__m128i *)(in + 6*in_p)); // 60 61 62 63 64 65 66 67
+ p7 = _mm_loadu_si128((__m128i *)(in + 7*in_p)); // 70 71 72 73 74 75 76 77
+ // 00 10 01 11 02 12 03 13
+ x0 = _mm_unpacklo_epi16(p0, p1);
+ // 20 30 21 31 22 32 23 33
+ x1 = _mm_unpacklo_epi16(p2, p3);
+ // 40 50 41 51 42 52 43 53
+ x2 = _mm_unpacklo_epi16(p4, p5);
+ // 60 70 61 71 62 72 63 73
+ x3 = _mm_unpacklo_epi16(p6, p7);
+ // 00 10 20 30 01 11 21 31
+ x4 = _mm_unpacklo_epi32(x0, x1);
+ // 40 50 60 70 41 51 61 71
+ x5 = _mm_unpacklo_epi32(x2, x3);
+ // 00 10 20 30 40 50 60 70
+ x6 = _mm_unpacklo_epi64(x4, x5);
+ // 01 11 21 31 41 51 61 71
+ x7 = _mm_unpackhi_epi64(x4, x5);
+
+ _mm_storeu_si128((__m128i *)(out + 0*out_p), x6);
+ // 00 10 20 30 40 50 60 70
+ _mm_storeu_si128((__m128i *)(out + 1*out_p), x7);
+ // 01 11 21 31 41 51 61 71
+
+ // 02 12 22 32 03 13 23 33
+ x4 = _mm_unpackhi_epi32(x0, x1);
+ // 42 52 62 72 43 53 63 73
+ x5 = _mm_unpackhi_epi32(x2, x3);
+ // 02 12 22 32 42 52 62 72
+ x6 = _mm_unpacklo_epi64(x4, x5);
+ // 03 13 23 33 43 53 63 73
+ x7 = _mm_unpackhi_epi64(x4, x5);
+
+ _mm_storeu_si128((__m128i *)(out + 2*out_p), x6);
+ // 02 12 22 32 42 52 62 72
+ _mm_storeu_si128((__m128i *)(out + 3*out_p), x7);
+ // 03 13 23 33 43 53 63 73
+
+ // 04 14 05 15 06 16 07 17
+ x0 = _mm_unpackhi_epi16(p0, p1);
+ // 24 34 25 35 26 36 27 37
+ x1 = _mm_unpackhi_epi16(p2, p3);
+ // 44 54 45 55 46 56 47 57
+ x2 = _mm_unpackhi_epi16(p4, p5);
+ // 64 74 65 75 66 76 67 77
+ x3 = _mm_unpackhi_epi16(p6, p7);
+ // 04 14 24 34 05 15 25 35
+ x4 = _mm_unpacklo_epi32(x0, x1);
+ // 44 54 64 74 45 55 65 75
+ x5 = _mm_unpacklo_epi32(x2, x3);
+ // 04 14 24 34 44 54 64 74
+ x6 = _mm_unpacklo_epi64(x4, x5);
+ // 05 15 25 35 45 55 65 75
+ x7 = _mm_unpackhi_epi64(x4, x5);
+
+ _mm_storeu_si128((__m128i *)(out + 4*out_p), x6);
+ // 04 14 24 34 44 54 64 74
+ _mm_storeu_si128((__m128i *)(out + 5*out_p), x7);
+ // 05 15 25 35 45 55 65 75
+
+ // 06 16 26 36 07 17 27 37
+ x4 = _mm_unpackhi_epi32(x0, x1);
+ // 46 56 66 76 47 57 67 77
+ x5 = _mm_unpackhi_epi32(x2, x3);
+ // 06 16 26 36 46 56 66 76
+ x6 = _mm_unpacklo_epi64(x4, x5);
+ // 07 17 27 37 47 57 67 77
+ x7 = _mm_unpackhi_epi64(x4, x5);
+
+ _mm_storeu_si128((__m128i *)(out + 6*out_p), x6);
+ // 06 16 26 36 46 56 66 76
+ _mm_storeu_si128((__m128i *)(out + 7*out_p), x7);
+ // 07 17 27 37 47 57 67 77
+ } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+static INLINE void highbd_transpose8x16(uint16_t *in0, uint16_t *in1,
+ int in_p, uint16_t *out, int out_p) {+ uint16_t *src0[1];
+ uint16_t *src1[1];
+ uint16_t *dest0[1];
+ uint16_t *dest1[1];
+ src0[0] = in0;
+ src1[0] = in1;
+ dest0[0] = out;
+ dest1[0] = out + 8;
+ highbd_transpose(src0, in_p, dest0, out_p, 1);
+ highbd_transpose(src1, in_p, dest1, out_p, 1);
+}
+
+void vp9_highbd_lpf_vertical_4_sse2(uint16_t *s, int p,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int count, int bd) {+ DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
+ uint16_t *src[1];
+ uint16_t *dst[1];
+ (void)count;
+
+ // Transpose 8x8
+ src[0] = s - 4;
+ dst[0] = t_dst;
+
+ highbd_transpose(src, p, dst, 8, 1);
+
+ // Loop filtering
+ vp9_highbd_lpf_horizontal_4_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1,
+ bd);
+
+ src[0] = t_dst;
+ dst[0] = s - 4;
+
+ // Transpose back
+ highbd_transpose(src, 8, dst, p, 1);
+}
+
+void vp9_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int p,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1,
+ int bd) {+ DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
+ uint16_t *src[2];
+ uint16_t *dst[2];
+
+ // Transpose 8x16
+ highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+ // Loop filtering
+ vp9_highbd_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
+ thresh0, blimit1, limit1, thresh1, bd);
+ src[0] = t_dst;
+ src[1] = t_dst + 8;
+ dst[0] = s - 4;
+ dst[1] = s - 4 + p * 8;
+
+ // Transpose back
+ highbd_transpose(src, 16, dst, p, 2);
+}
+
+void vp9_highbd_lpf_vertical_8_sse2(uint16_t *s, int p,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int count, int bd) {+ DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 8]);
+ uint16_t *src[1];
+ uint16_t *dst[1];
+ (void)count;
+
+ // Transpose 8x8
+ src[0] = s - 4;
+ dst[0] = t_dst;
+
+ highbd_transpose(src, p, dst, 8, 1);
+
+ // Loop filtering
+ vp9_highbd_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1,
+ bd);
+
+ src[0] = t_dst;
+ dst[0] = s - 4;
+
+ // Transpose back
+ highbd_transpose(src, 8, dst, p, 1);
+}
+
+void vp9_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int p,
+ const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1,
+ int bd) {+ DECLARE_ALIGNED(16, uint16_t, t_dst[16 * 8]);
+ uint16_t *src[2];
+ uint16_t *dst[2];
+
+ // Transpose 8x16
+ highbd_transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+ // Loop filtering
+ vp9_highbd_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0,
+ thresh0, blimit1, limit1, thresh1, bd);
+ src[0] = t_dst;
+ src[1] = t_dst + 8;
+
+ dst[0] = s - 4;
+ dst[1] = s - 4 + p * 8;
+
+ // Transpose back
+ highbd_transpose(src, 16, dst, p, 2);
+}
+
+void vp9_highbd_lpf_vertical_16_sse2(uint16_t *s, int p,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int bd) {+ DECLARE_ALIGNED(16, uint16_t, t_dst[8 * 16]);
+ uint16_t *src[2];
+ uint16_t *dst[2];
+
+ src[0] = s - 8;
+ src[1] = s;
+ dst[0] = t_dst;
+ dst[1] = t_dst + 8 * 8;
+
+ // Transpose 16x8
+ highbd_transpose(src, p, dst, 8, 2);
+
+ // Loop filtering
+ highbd_mb_lpf_horizontal_edge_w_sse2_8(t_dst + 8 * 8, 8, blimit, limit,
+ thresh, bd);
+ src[0] = t_dst;
+ src[1] = t_dst + 8 * 8;
+ dst[0] = s - 8;
+ dst[1] = s;
+
+ // Transpose back
+ highbd_transpose(src, 8, dst, p, 2);
+}
+
+void vp9_highbd_lpf_vertical_16_dual_sse2(uint16_t *s,
+ int p,
+ const uint8_t *blimit,
+ const uint8_t *limit,
+ const uint8_t *thresh,
+ int bd) {+ DECLARE_ALIGNED(16, uint16_t, t_dst[256]);
+
+ // Transpose 16x16
+ highbd_transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
+ highbd_transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
+
+ // Loop filtering
+ highbd_mb_lpf_horizontal_edge_w_sse2_16(t_dst + 8 * 16, 16, blimit, limit,
+ thresh, bd);
+
+ // Transpose back
+ highbd_transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
+ highbd_transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
+}
--- /dev/null
+++ b/vpx_dsp/x86/loopfilter_avx2.c
@@ -1,0 +1,986 @@
+/*
+ * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h> /* AVX2 */
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+
+static void mb_lpf_horizontal_edge_w_avx2_8(unsigned char *s, int p,
+ const unsigned char *_blimit, const unsigned char *_limit,
+ const unsigned char *_thresh) {+ __m128i mask, hev, flat, flat2;
+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i one = _mm_set1_epi8(1);
+ __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+ __m128i abs_p1p0;
+
+ const __m128i thresh = _mm_broadcastb_epi8(
+ _mm_cvtsi32_si128((int) _thresh[0]));
+ const __m128i limit = _mm_broadcastb_epi8(
+ _mm_cvtsi32_si128((int) _limit[0]));
+ const __m128i blimit = _mm_broadcastb_epi8(
+ _mm_cvtsi32_si128((int) _blimit[0]));
+
+ q4p4 = _mm_loadl_epi64((__m128i *) (s - 5 * p));
+ q4p4 = _mm_castps_si128(
+ _mm_loadh_pi(_mm_castsi128_ps(q4p4), (__m64 *) (s + 4 * p)));
+ q3p3 = _mm_loadl_epi64((__m128i *) (s - 4 * p));
+ q3p3 = _mm_castps_si128(
+ _mm_loadh_pi(_mm_castsi128_ps(q3p3), (__m64 *) (s + 3 * p)));
+ q2p2 = _mm_loadl_epi64((__m128i *) (s - 3 * p));
+ q2p2 = _mm_castps_si128(
+ _mm_loadh_pi(_mm_castsi128_ps(q2p2), (__m64 *) (s + 2 * p)));
+ q1p1 = _mm_loadl_epi64((__m128i *) (s - 2 * p));
+ q1p1 = _mm_castps_si128(
+ _mm_loadh_pi(_mm_castsi128_ps(q1p1), (__m64 *) (s + 1 * p)));
+ p1q1 = _mm_shuffle_epi32(q1p1, 78);
+ q0p0 = _mm_loadl_epi64((__m128i *) (s - 1 * p));
+ q0p0 = _mm_castps_si128(
+ _mm_loadh_pi(_mm_castsi128_ps(q0p0), (__m64 *) (s - 0 * p)));
+ p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+ {+ __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+ abs_p1p0 = _mm_or_si128(_mm_subs_epu8(q1p1, q0p0),
+ _mm_subs_epu8(q0p0, q1p1));
+ abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+ fe = _mm_set1_epi8(0xfe);
+ ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+ abs_p0q0 = _mm_or_si128(_mm_subs_epu8(q0p0, p0q0),
+ _mm_subs_epu8(p0q0, q0p0));
+ abs_p1q1 = _mm_or_si128(_mm_subs_epu8(q1p1, p1q1),
+ _mm_subs_epu8(p1q1, q1p1));
+ flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu8(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+ abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+ mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ mask = _mm_max_epu8(abs_p1p0, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(q2p2, q1p1),
+ _mm_subs_epu8(q1p1, q2p2)),
+ _mm_or_si128(_mm_subs_epu8(q3p3, q2p2),
+ _mm_subs_epu8(q2p2, q3p3)));
+ mask = _mm_max_epu8(work, mask);
+ mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+ mask = _mm_subs_epu8(mask, limit);
+ mask = _mm_cmpeq_epi8(mask, zero);
+ }
+
+ // lp filter
+ {+ const __m128i t4 = _mm_set1_epi8(4);
+ const __m128i t3 = _mm_set1_epi8(3);
+ const __m128i t80 = _mm_set1_epi8(0x80);
+ const __m128i t1 = _mm_set1_epi16(0x1);
+ __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+ __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+ __m128i qs0 = _mm_xor_si128(p0q0, t80);
+ __m128i qs1 = _mm_xor_si128(p1q1, t80);
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+ __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+ __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+ filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+ work_a = _mm_subs_epi8(qs0, qs0ps0);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ /* (vp9_filter + 3 * (qs0 - ps0)) & mask */
+ filt = _mm_and_si128(filt, mask);
+
+ filter1 = _mm_adds_epi8(filt, t4);
+ filter2 = _mm_adds_epi8(filt, t3);
+
+ filter1 = _mm_unpacklo_epi8(zero, filter1);
+ filter1 = _mm_srai_epi16(filter1, 0xB);
+ filter2 = _mm_unpacklo_epi8(zero, filter2);
+ filter2 = _mm_srai_epi16(filter2, 0xB);
+
+ /* Filter1 >> 3 */
+ filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+ qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
+
+ /* filt >> 1 */
+ filt = _mm_adds_epi16(filter1, t1);
+ filt = _mm_srai_epi16(filt, 1);
+ filt = _mm_andnot_si128(
+ _mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8), filt);
+ filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+ qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
+ // loopfilter done
+
+ {+ __m128i work;
+ flat = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(q2p2, q0p0),
+ _mm_subs_epu8(q0p0, q2p2)),
+ _mm_or_si128(_mm_subs_epu8(q3p3, q0p0),
+ _mm_subs_epu8(q0p0, q3p3)));
+ flat = _mm_max_epu8(abs_p1p0, flat);
+ flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+ flat = _mm_subs_epu8(flat, one);
+ flat = _mm_cmpeq_epi8(flat, zero);
+ flat = _mm_and_si128(flat, mask);
+
+ q5p5 = _mm_loadl_epi64((__m128i *) (s - 6 * p));
+ q5p5 = _mm_castps_si128(
+ _mm_loadh_pi(_mm_castsi128_ps(q5p5),
+ (__m64 *) (s + 5 * p)));
+
+ q6p6 = _mm_loadl_epi64((__m128i *) (s - 7 * p));
+ q6p6 = _mm_castps_si128(
+ _mm_loadh_pi(_mm_castsi128_ps(q6p6),
+ (__m64 *) (s + 6 * p)));
+
+ flat2 = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(q4p4, q0p0),
+ _mm_subs_epu8(q0p0, q4p4)),
+ _mm_or_si128(_mm_subs_epu8(q5p5, q0p0),
+ _mm_subs_epu8(q0p0, q5p5)));
+
+ q7p7 = _mm_loadl_epi64((__m128i *) (s - 8 * p));
+ q7p7 = _mm_castps_si128(
+ _mm_loadh_pi(_mm_castsi128_ps(q7p7),
+ (__m64 *) (s + 7 * p)));
+
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(q6p6, q0p0),
+ _mm_subs_epu8(q0p0, q6p6)),
+ _mm_or_si128(_mm_subs_epu8(q7p7, q0p0),
+ _mm_subs_epu8(q0p0, q7p7)));
+
+ flat2 = _mm_max_epu8(work, flat2);
+ flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+ flat2 = _mm_subs_epu8(flat2, one);
+ flat2 = _mm_cmpeq_epi8(flat2, zero);
+ flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
+ }
+
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // flat and wide flat calculations
+ {+ const __m128i eight = _mm_set1_epi16(8);
+ const __m128i four = _mm_set1_epi16(4);
+ __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+ __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+ __m128i pixelFilter_p, pixelFilter_q;
+ __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+ __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+ p7_16 = _mm_unpacklo_epi8(q7p7, zero);
+ p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+ p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+ p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+ p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+ p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+ p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+ p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+ q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+ q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+ q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+ q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+ q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+ q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+ q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+ q7_16 = _mm_unpackhi_epi8(q7p7, zero);
+
+ pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+ _mm_add_epi16(p4_16, p3_16));
+ pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+ _mm_add_epi16(q4_16, q3_16));
+
+ pixetFilter_p2p1p0 = _mm_add_epi16(p0_16,
+ _mm_add_epi16(p2_16, p1_16));
+ pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+ pixetFilter_q2q1q0 = _mm_add_epi16(q0_16,
+ _mm_add_epi16(q2_16, q1_16));
+ pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+ pixelFilter_p = _mm_add_epi16(eight,
+ _mm_add_epi16(pixelFilter_p, pixelFilter_q));
+ pixetFilter_p2p1p0 = _mm_add_epi16(four,
+ _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(p7_16, p0_16)),
+ 4);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(q7_16, q0_16)),
+ 4);
+ flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(p3_16, p0_16)), 3);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(q3_16, q0_16)), 3);
+
+ flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(p7_16, p7_16);
+ sum_q7 = _mm_add_epi16(q7_16, q7_16);
+ sum_p3 = _mm_add_epi16(p3_16, p3_16);
+ sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p1_16)),
+ 4);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q1_16)),
+ 4);
+ flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+ pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+ pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(sum_p3, p1_16)), 3);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixetFilter_q2q1q0,
+ _mm_add_epi16(sum_q3, q1_16)), 3);
+ flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+ sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p2_16)),
+ 4);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q2_16)),
+ 4);
+ flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+ pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+ pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(sum_p3, p2_16)), 3);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixetFilter_q2q1q0,
+ _mm_add_epi16(sum_q3, q2_16)), 3);
+ flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p3_16)),
+ 4);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q3_16)),
+ 4);
+ flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p4_16)),
+ 4);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q4_16)),
+ 4);
+ flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p5_16)),
+ 4);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q5_16)),
+ 4);
+ flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+ res_p = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_p, _mm_add_epi16(sum_p7, p6_16)),
+ 4);
+ res_q = _mm_srli_epi16(
+ _mm_add_epi16(pixelFilter_q, _mm_add_epi16(sum_q7, q6_16)),
+ 4);
+ flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+ }
+ // wide flat
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ flat = _mm_shuffle_epi32(flat, 68);
+ flat2 = _mm_shuffle_epi32(flat2, 68);
+
+ q2p2 = _mm_andnot_si128(flat, q2p2);
+ flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+ q2p2 = _mm_or_si128(q2p2, flat_q2p2);
+
+ qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+ flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+ q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+ qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+ flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+ q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+ q6p6 = _mm_andnot_si128(flat2, q6p6);
+ flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+ q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+ _mm_storel_epi64((__m128i *) (s - 7 * p), q6p6);
+ _mm_storeh_pi((__m64 *) (s + 6 * p), _mm_castsi128_ps(q6p6));
+
+ q5p5 = _mm_andnot_si128(flat2, q5p5);
+ flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+ q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+ _mm_storel_epi64((__m128i *) (s - 6 * p), q5p5);
+ _mm_storeh_pi((__m64 *) (s + 5 * p), _mm_castsi128_ps(q5p5));
+
+ q4p4 = _mm_andnot_si128(flat2, q4p4);
+ flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+ q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+ _mm_storel_epi64((__m128i *) (s - 5 * p), q4p4);
+ _mm_storeh_pi((__m64 *) (s + 4 * p), _mm_castsi128_ps(q4p4));
+
+ q3p3 = _mm_andnot_si128(flat2, q3p3);
+ flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+ q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+ _mm_storel_epi64((__m128i *) (s - 4 * p), q3p3);
+ _mm_storeh_pi((__m64 *) (s + 3 * p), _mm_castsi128_ps(q3p3));
+
+ q2p2 = _mm_andnot_si128(flat2, q2p2);
+ flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+ q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+ _mm_storel_epi64((__m128i *) (s - 3 * p), q2p2);
+ _mm_storeh_pi((__m64 *) (s + 2 * p), _mm_castsi128_ps(q2p2));
+
+ q1p1 = _mm_andnot_si128(flat2, q1p1);
+ flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+ q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+ _mm_storel_epi64((__m128i *) (s - 2 * p), q1p1);
+ _mm_storeh_pi((__m64 *) (s + 1 * p), _mm_castsi128_ps(q1p1));
+
+ q0p0 = _mm_andnot_si128(flat2, q0p0);
+ flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+ q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+ _mm_storel_epi64((__m128i *) (s - 1 * p), q0p0);
+ _mm_storeh_pi((__m64 *) (s - 0 * p), _mm_castsi128_ps(q0p0));
+ }
+}
+
+DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = {+ 0, 128, 1, 128, 2, 128, 3, 128, 4, 128, 5, 128, 6, 128, 7, 128,
+ 8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128
+};
+
+static void mb_lpf_horizontal_edge_w_avx2_16(unsigned char *s, int p,
+ const unsigned char *_blimit, const unsigned char *_limit,
+ const unsigned char *_thresh) {+ __m128i mask, hev, flat, flat2;
+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i one = _mm_set1_epi8(1);
+ __m128i p7, p6, p5;
+ __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+ __m128i q5, q6, q7;
+ __m256i p256_7, q256_7, p256_6, q256_6, p256_5, q256_5, p256_4,
+ q256_4, p256_3, q256_3, p256_2, q256_2, p256_1, q256_1,
+ p256_0, q256_0;
+
+ const __m128i thresh = _mm_broadcastb_epi8(
+ _mm_cvtsi32_si128((int) _thresh[0]));
+ const __m128i limit = _mm_broadcastb_epi8(
+ _mm_cvtsi32_si128((int) _limit[0]));
+ const __m128i blimit = _mm_broadcastb_epi8(
+ _mm_cvtsi32_si128((int) _blimit[0]));
+
+ p256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 5 * p)));
+ p256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 4 * p)));
+ p256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 3 * p)));
+ p256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 2 * p)));
+ p256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 1 * p)));
+ q256_0 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 0 * p)));
+ q256_1 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s + 1 * p)));
+ q256_2 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s + 2 * p)));
+ q256_3 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s + 3 * p)));
+ q256_4 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s + 4 * p)));
+
+ p4 = _mm256_castsi256_si128(p256_4);
+ p3 = _mm256_castsi256_si128(p256_3);
+ p2 = _mm256_castsi256_si128(p256_2);
+ p1 = _mm256_castsi256_si128(p256_1);
+ p0 = _mm256_castsi256_si128(p256_0);
+ q0 = _mm256_castsi256_si128(q256_0);
+ q1 = _mm256_castsi256_si128(q256_1);
+ q2 = _mm256_castsi256_si128(q256_2);
+ q3 = _mm256_castsi256_si128(q256_3);
+ q4 = _mm256_castsi256_si128(q256_4);
+
+ {+ const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+ _mm_subs_epu8(p0, p1));
+ const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+ _mm_subs_epu8(q0, q1));
+ const __m128i fe = _mm_set1_epi8(0xfe);
+ const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+ __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+ _mm_subs_epu8(q0, p0));
+ __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+ _mm_subs_epu8(q1, p1));
+ __m128i work;
+ flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu8(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+ abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+ mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ mask = _mm_max_epu8(flat, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+ _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
+ mask = _mm_max_epu8(work, mask);
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
+ _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
+ mask = _mm_max_epu8(work, mask);
+ mask = _mm_subs_epu8(mask, limit);
+ mask = _mm_cmpeq_epi8(mask, zero);
+ }
+
+ // lp filter
+ {+ const __m128i t4 = _mm_set1_epi8(4);
+ const __m128i t3 = _mm_set1_epi8(3);
+ const __m128i t80 = _mm_set1_epi8(0x80);
+ const __m128i te0 = _mm_set1_epi8(0xe0);
+ const __m128i t1f = _mm_set1_epi8(0x1f);
+ const __m128i t1 = _mm_set1_epi8(0x1);
+ const __m128i t7f = _mm_set1_epi8(0x7f);
+
+ __m128i ps1 = _mm_xor_si128(p1, t80);
+ __m128i ps0 = _mm_xor_si128(p0, t80);
+ __m128i qs0 = _mm_xor_si128(q0, t80);
+ __m128i qs1 = _mm_xor_si128(q1, t80);
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+ __m128i flat2_p6, flat2_p5, flat2_p4, flat2_p3, flat2_p2, flat2_p1,
+ flat2_p0, flat2_q0, flat2_q1, flat2_q2, flat2_q3, flat2_q4,
+ flat2_q5, flat2_q6, flat_p2, flat_p1, flat_p0, flat_q0, flat_q1,
+ flat_q2;
+
+ filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+ work_a = _mm_subs_epi8(qs0, ps0);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ /* (vp9_filter + 3 * (qs0 - ps0)) & mask */
+ filt = _mm_and_si128(filt, mask);
+
+ filter1 = _mm_adds_epi8(filt, t4);
+ filter2 = _mm_adds_epi8(filt, t3);
+
+ /* Filter1 >> 3 */
+ work_a = _mm_cmpgt_epi8(zero, filter1);
+ filter1 = _mm_srli_epi16(filter1, 3);
+ work_a = _mm_and_si128(work_a, te0);
+ filter1 = _mm_and_si128(filter1, t1f);
+ filter1 = _mm_or_si128(filter1, work_a);
+ qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+
+ /* Filter2 >> 3 */
+ work_a = _mm_cmpgt_epi8(zero, filter2);
+ filter2 = _mm_srli_epi16(filter2, 3);
+ work_a = _mm_and_si128(work_a, te0);
+ filter2 = _mm_and_si128(filter2, t1f);
+ filter2 = _mm_or_si128(filter2, work_a);
+ ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+
+ /* filt >> 1 */
+ filt = _mm_adds_epi8(filter1, t1);
+ work_a = _mm_cmpgt_epi8(zero, filt);
+ filt = _mm_srli_epi16(filt, 1);
+ work_a = _mm_and_si128(work_a, t80);
+ filt = _mm_and_si128(filt, t7f);
+ filt = _mm_or_si128(filt, work_a);
+ filt = _mm_andnot_si128(hev, filt);
+ ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+ qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+ // loopfilter done
+
+ {+ __m128i work;
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
+ _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
+ flat = _mm_max_epu8(work, flat);
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
+ _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
+ flat = _mm_max_epu8(work, flat);
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(p4, p0), _mm_subs_epu8(p0, p4)),
+ _mm_or_si128(_mm_subs_epu8(q4, q0), _mm_subs_epu8(q0, q4)));
+ flat = _mm_subs_epu8(flat, one);
+ flat = _mm_cmpeq_epi8(flat, zero);
+ flat = _mm_and_si128(flat, mask);
+
+ p256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 6 * p)));
+ q256_5 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s + 5 * p)));
+ p5 = _mm256_castsi256_si128(p256_5);
+ q5 = _mm256_castsi256_si128(q256_5);
+ flat2 = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(p5, p0), _mm_subs_epu8(p0, p5)),
+ _mm_or_si128(_mm_subs_epu8(q5, q0), _mm_subs_epu8(q0, q5)));
+
+ flat2 = _mm_max_epu8(work, flat2);
+ p256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 7 * p)));
+ q256_6 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s + 6 * p)));
+ p6 = _mm256_castsi256_si128(p256_6);
+ q6 = _mm256_castsi256_si128(q256_6);
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(p6, p0), _mm_subs_epu8(p0, p6)),
+ _mm_or_si128(_mm_subs_epu8(q6, q0), _mm_subs_epu8(q0, q6)));
+
+ flat2 = _mm_max_epu8(work, flat2);
+
+ p256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s - 8 * p)));
+ q256_7 = _mm256_castpd_si256(_mm256_broadcast_pd(
+ (__m128d const *)(s + 7 * p)));
+ p7 = _mm256_castsi256_si128(p256_7);
+ q7 = _mm256_castsi256_si128(q256_7);
+ work = _mm_max_epu8(
+ _mm_or_si128(_mm_subs_epu8(p7, p0), _mm_subs_epu8(p0, p7)),
+ _mm_or_si128(_mm_subs_epu8(q7, q0), _mm_subs_epu8(q0, q7)));
+
+ flat2 = _mm_max_epu8(work, flat2);
+ flat2 = _mm_subs_epu8(flat2, one);
+ flat2 = _mm_cmpeq_epi8(flat2, zero);
+ flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
+ }
+
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // flat and wide flat calculations
+ {+ const __m256i eight = _mm256_set1_epi16(8);
+ const __m256i four = _mm256_set1_epi16(4);
+ __m256i pixelFilter_p, pixelFilter_q, pixetFilter_p2p1p0,
+ pixetFilter_q2q1q0, sum_p7, sum_q7, sum_p3, sum_q3, res_p,
+ res_q;
+
+ const __m256i filter = _mm256_load_si256(
+ (__m256i const *)filt_loopfilter_avx2);
+ p256_7 = _mm256_shuffle_epi8(p256_7, filter);
+ p256_6 = _mm256_shuffle_epi8(p256_6, filter);
+ p256_5 = _mm256_shuffle_epi8(p256_5, filter);
+ p256_4 = _mm256_shuffle_epi8(p256_4, filter);
+ p256_3 = _mm256_shuffle_epi8(p256_3, filter);
+ p256_2 = _mm256_shuffle_epi8(p256_2, filter);
+ p256_1 = _mm256_shuffle_epi8(p256_1, filter);
+ p256_0 = _mm256_shuffle_epi8(p256_0, filter);
+ q256_0 = _mm256_shuffle_epi8(q256_0, filter);
+ q256_1 = _mm256_shuffle_epi8(q256_1, filter);
+ q256_2 = _mm256_shuffle_epi8(q256_2, filter);
+ q256_3 = _mm256_shuffle_epi8(q256_3, filter);
+ q256_4 = _mm256_shuffle_epi8(q256_4, filter);
+ q256_5 = _mm256_shuffle_epi8(q256_5, filter);
+ q256_6 = _mm256_shuffle_epi8(q256_6, filter);
+ q256_7 = _mm256_shuffle_epi8(q256_7, filter);
+
+ pixelFilter_p = _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_5),
+ _mm256_add_epi16(p256_4, p256_3));
+ pixelFilter_q = _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_5),
+ _mm256_add_epi16(q256_4, q256_3));
+
+ pixetFilter_p2p1p0 = _mm256_add_epi16(p256_0,
+ _mm256_add_epi16(p256_2, p256_1));
+ pixelFilter_p = _mm256_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+ pixetFilter_q2q1q0 = _mm256_add_epi16(q256_0,
+ _mm256_add_epi16(q256_2, q256_1));
+ pixelFilter_q = _mm256_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+
+ pixelFilter_p = _mm256_add_epi16(eight,
+ _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
+
+ pixetFilter_p2p1p0 = _mm256_add_epi16(four,
+ _mm256_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_p,
+ _mm256_add_epi16(p256_7, p256_0)), 4);
+
+ flat2_p0 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_p,
+ _mm256_add_epi16(q256_7, q256_0)), 4);
+
+ flat2_q0 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixetFilter_p2p1p0,
+ _mm256_add_epi16(p256_3, p256_0)), 3);
+
+ flat_p0 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixetFilter_p2p1p0,
+ _mm256_add_epi16(q256_3, q256_0)), 3);
+
+ flat_q0 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ sum_p7 = _mm256_add_epi16(p256_7, p256_7);
+
+ sum_q7 = _mm256_add_epi16(q256_7, q256_7);
+
+ sum_p3 = _mm256_add_epi16(p256_3, p256_3);
+
+ sum_q3 = _mm256_add_epi16(q256_3, q256_3);
+
+ pixelFilter_q = _mm256_sub_epi16(pixelFilter_p, p256_6);
+
+ pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_6);
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_p,
+ _mm256_add_epi16(sum_p7, p256_1)), 4);
+
+ flat2_p1 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_q,
+ _mm256_add_epi16(sum_q7, q256_1)), 4);
+
+ flat2_q1 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_p2p1p0, p256_2);
+
+ pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_2);
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixetFilter_p2p1p0,
+ _mm256_add_epi16(sum_p3, p256_1)), 3);
+
+ flat_p1 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixetFilter_q2q1q0,
+ _mm256_add_epi16(sum_q3, q256_1)), 3);
+
+ flat_q1 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+ sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+ sum_p3 = _mm256_add_epi16(sum_p3, p256_3);
+
+ sum_q3 = _mm256_add_epi16(sum_q3, q256_3);
+
+ pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_5);
+
+ pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_5);
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_p,
+ _mm256_add_epi16(sum_p7, p256_2)), 4);
+
+ flat2_p2 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_q,
+ _mm256_add_epi16(sum_q7, q256_2)), 4);
+
+ flat2_q2 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ pixetFilter_p2p1p0 = _mm256_sub_epi16(pixetFilter_p2p1p0, q256_1);
+
+ pixetFilter_q2q1q0 = _mm256_sub_epi16(pixetFilter_q2q1q0, p256_1);
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixetFilter_p2p1p0,
+ _mm256_add_epi16(sum_p3, p256_2)), 3);
+
+ flat_p2 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixetFilter_q2q1q0,
+ _mm256_add_epi16(sum_q3, q256_2)), 3);
+
+ flat_q2 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+ sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+ pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_4);
+
+ pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_4);
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_p,
+ _mm256_add_epi16(sum_p7, p256_3)), 4);
+
+ flat2_p3 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_q,
+ _mm256_add_epi16(sum_q7, q256_3)), 4);
+
+ flat2_q3 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+ sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+ pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_3);
+
+ pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_3);
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_p,
+ _mm256_add_epi16(sum_p7, p256_4)), 4);
+
+ flat2_p4 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_q,
+ _mm256_add_epi16(sum_q7, q256_4)), 4);
+
+ flat2_q4 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+ sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+ pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_2);
+
+ pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_2);
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_p,
+ _mm256_add_epi16(sum_p7, p256_5)), 4);
+
+ flat2_p5 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_q,
+ _mm256_add_epi16(sum_q7, q256_5)), 4);
+
+ flat2_q5 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+
+ sum_p7 = _mm256_add_epi16(sum_p7, p256_7);
+
+ sum_q7 = _mm256_add_epi16(sum_q7, q256_7);
+
+ pixelFilter_p = _mm256_sub_epi16(pixelFilter_p, q256_1);
+
+ pixelFilter_q = _mm256_sub_epi16(pixelFilter_q, p256_1);
+
+ res_p = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_p,
+ _mm256_add_epi16(sum_p7, p256_6)), 4);
+
+ flat2_p6 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p),
+ 168));
+
+ res_q = _mm256_srli_epi16(
+ _mm256_add_epi16(pixelFilter_q,
+ _mm256_add_epi16(sum_q7, q256_6)), 4);
+
+ flat2_q6 = _mm256_castsi256_si128(
+ _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q),
+ 168));
+ }
+
+ // wide flat
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ p2 = _mm_andnot_si128(flat, p2);
+ flat_p2 = _mm_and_si128(flat, flat_p2);
+ p2 = _mm_or_si128(flat_p2, p2);
+
+ p1 = _mm_andnot_si128(flat, ps1);
+ flat_p1 = _mm_and_si128(flat, flat_p1);
+ p1 = _mm_or_si128(flat_p1, p1);
+
+ p0 = _mm_andnot_si128(flat, ps0);
+ flat_p0 = _mm_and_si128(flat, flat_p0);
+ p0 = _mm_or_si128(flat_p0, p0);
+
+ q0 = _mm_andnot_si128(flat, qs0);
+ flat_q0 = _mm_and_si128(flat, flat_q0);
+ q0 = _mm_or_si128(flat_q0, q0);
+
+ q1 = _mm_andnot_si128(flat, qs1);
+ flat_q1 = _mm_and_si128(flat, flat_q1);
+ q1 = _mm_or_si128(flat_q1, q1);
+
+ q2 = _mm_andnot_si128(flat, q2);
+ flat_q2 = _mm_and_si128(flat, flat_q2);
+ q2 = _mm_or_si128(flat_q2, q2);
+
+ p6 = _mm_andnot_si128(flat2, p6);
+ flat2_p6 = _mm_and_si128(flat2, flat2_p6);
+ p6 = _mm_or_si128(flat2_p6, p6);
+ _mm_storeu_si128((__m128i *) (s - 7 * p), p6);
+
+ p5 = _mm_andnot_si128(flat2, p5);
+ flat2_p5 = _mm_and_si128(flat2, flat2_p5);
+ p5 = _mm_or_si128(flat2_p5, p5);
+ _mm_storeu_si128((__m128i *) (s - 6 * p), p5);
+
+ p4 = _mm_andnot_si128(flat2, p4);
+ flat2_p4 = _mm_and_si128(flat2, flat2_p4);
+ p4 = _mm_or_si128(flat2_p4, p4);
+ _mm_storeu_si128((__m128i *) (s - 5 * p), p4);
+
+ p3 = _mm_andnot_si128(flat2, p3);
+ flat2_p3 = _mm_and_si128(flat2, flat2_p3);
+ p3 = _mm_or_si128(flat2_p3, p3);
+ _mm_storeu_si128((__m128i *) (s - 4 * p), p3);
+
+ p2 = _mm_andnot_si128(flat2, p2);
+ flat2_p2 = _mm_and_si128(flat2, flat2_p2);
+ p2 = _mm_or_si128(flat2_p2, p2);
+ _mm_storeu_si128((__m128i *) (s - 3 * p), p2);
+
+ p1 = _mm_andnot_si128(flat2, p1);
+ flat2_p1 = _mm_and_si128(flat2, flat2_p1);
+ p1 = _mm_or_si128(flat2_p1, p1);
+ _mm_storeu_si128((__m128i *) (s - 2 * p), p1);
+
+ p0 = _mm_andnot_si128(flat2, p0);
+ flat2_p0 = _mm_and_si128(flat2, flat2_p0);
+ p0 = _mm_or_si128(flat2_p0, p0);
+ _mm_storeu_si128((__m128i *) (s - 1 * p), p0);
+
+ q0 = _mm_andnot_si128(flat2, q0);
+ flat2_q0 = _mm_and_si128(flat2, flat2_q0);
+ q0 = _mm_or_si128(flat2_q0, q0);
+ _mm_storeu_si128((__m128i *) (s - 0 * p), q0);
+
+ q1 = _mm_andnot_si128(flat2, q1);
+ flat2_q1 = _mm_and_si128(flat2, flat2_q1);
+ q1 = _mm_or_si128(flat2_q1, q1);
+ _mm_storeu_si128((__m128i *) (s + 1 * p), q1);
+
+ q2 = _mm_andnot_si128(flat2, q2);
+ flat2_q2 = _mm_and_si128(flat2, flat2_q2);
+ q2 = _mm_or_si128(flat2_q2, q2);
+ _mm_storeu_si128((__m128i *) (s + 2 * p), q2);
+
+ q3 = _mm_andnot_si128(flat2, q3);
+ flat2_q3 = _mm_and_si128(flat2, flat2_q3);
+ q3 = _mm_or_si128(flat2_q3, q3);
+ _mm_storeu_si128((__m128i *) (s + 3 * p), q3);
+
+ q4 = _mm_andnot_si128(flat2, q4);
+ flat2_q4 = _mm_and_si128(flat2, flat2_q4);
+ q4 = _mm_or_si128(flat2_q4, q4);
+ _mm_storeu_si128((__m128i *) (s + 4 * p), q4);
+
+ q5 = _mm_andnot_si128(flat2, q5);
+ flat2_q5 = _mm_and_si128(flat2, flat2_q5);
+ q5 = _mm_or_si128(flat2_q5, q5);
+ _mm_storeu_si128((__m128i *) (s + 5 * p), q5);
+
+ q6 = _mm_andnot_si128(flat2, q6);
+ flat2_q6 = _mm_and_si128(flat2, flat2_q6);
+ q6 = _mm_or_si128(flat2_q6, q6);
+ _mm_storeu_si128((__m128i *) (s + 6 * p), q6);
+ }
+}
+
+void vp9_lpf_horizontal_16_avx2(unsigned char *s, int p,
+ const unsigned char *_blimit, const unsigned char *_limit,
+ const unsigned char *_thresh, int count) {+ if (count == 1)
+ mb_lpf_horizontal_edge_w_avx2_8(s, p, _blimit, _limit, _thresh);
+ else
+ mb_lpf_horizontal_edge_w_avx2_16(s, p, _blimit, _limit, _thresh);
+}
--- /dev/null
+++ b/vpx_dsp/x86/loopfilter_mmx.asm
@@ -1,0 +1,611 @@
+;
+; Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+
+;void vp9_lpf_horizontal_4_mmx
+;(
+; unsigned char *src_ptr,
+; int src_pixel_step,
+; const char *blimit,
+; const char *limit,
+; const char *thresh,
+; int count
+;)
+global sym(vp9_lpf_horizontal_4_mmx) PRIVATE
+sym(vp9_lpf_horizontal_4_mmx):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 6
+ GET_GOT rbx
+ push rsi
+ push rdi
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 32 ; reserve 32 bytes
+ %define t0 [rsp + 0] ;__declspec(align(16)) char t0[8];
+ %define t1 [rsp + 16] ;__declspec(align(16)) char t1[8];
+
+ mov rsi, arg(0) ;src_ptr
+ movsxd rax, dword ptr arg(1) ;src_pixel_step ; destination pitch?
+
+ movsxd rcx, dword ptr arg(5) ;count
+.next8_h:
+ mov rdx, arg(3) ;limit
+ movq mm7, [rdx]
+ mov rdi, rsi ; rdi points to row +1 for indirect addressing
+ add rdi, rax
+
+ ; calculate breakout conditions
+ movq mm2, [rdi+2*rax] ; q3
+ movq mm1, [rsi+2*rax] ; q2
+ movq mm6, mm1 ; q2
+ psubusb mm1, mm2 ; q2-=q3
+ psubusb mm2, mm6 ; q3-=q2
+ por mm1, mm2 ; abs(q3-q2)
+ psubusb mm1, mm7 ;
+
+
+ movq mm4, [rsi+rax] ; q1
+ movq mm3, mm4 ; q1
+ psubusb mm4, mm6 ; q1-=q2
+ psubusb mm6, mm3 ; q2-=q1
+ por mm4, mm6 ; abs(q2-q1)
+
+ psubusb mm4, mm7
+ por mm1, mm4
+
+ movq mm4, [rsi] ; q0
+ movq mm0, mm4 ; q0
+ psubusb mm4, mm3 ; q0-=q1
+ psubusb mm3, mm0 ; q1-=q0
+ por mm4, mm3 ; abs(q0-q1)
+ movq t0, mm4 ; save to t0
+ psubusb mm4, mm7
+ por mm1, mm4
+
+
+ neg rax ; negate pitch to deal with above border
+
+ movq mm2, [rsi+4*rax] ; p3
+ movq mm4, [rdi+4*rax] ; p2
+ movq mm5, mm4 ; p2
+ psubusb mm4, mm2 ; p2-=p3
+ psubusb mm2, mm5 ; p3-=p2
+ por mm4, mm2 ; abs(p3 - p2)
+ psubusb mm4, mm7
+ por mm1, mm4
+
+
+ movq mm4, [rsi+2*rax] ; p1
+ movq mm3, mm4 ; p1
+ psubusb mm4, mm5 ; p1-=p2
+ psubusb mm5, mm3 ; p2-=p1
+ por mm4, mm5 ; abs(p2 - p1)
+ psubusb mm4, mm7
+ por mm1, mm4
+
+ movq mm2, mm3 ; p1
+
+ movq mm4, [rsi+rax] ; p0
+ movq mm5, mm4 ; p0
+ psubusb mm4, mm3 ; p0-=p1
+ psubusb mm3, mm5 ; p1-=p0
+ por mm4, mm3 ; abs(p1 - p0)
+ movq t1, mm4 ; save to t1
+ psubusb mm4, mm7
+ por mm1, mm4
+
+ movq mm3, [rdi] ; q1
+ movq mm4, mm3 ; q1
+ psubusb mm3, mm2 ; q1-=p1
+ psubusb mm2, mm4 ; p1-=q1
+ por mm2, mm3 ; abs(p1-q1)
+ pand mm2, [GLOBAL(tfe)] ; set lsb of each byte to zero
+ psrlw mm2, 1 ; abs(p1-q1)/2
+
+ movq mm6, mm5 ; p0
+ movq mm3, [rsi] ; q0
+ psubusb mm5, mm3 ; p0-=q0
+ psubusb mm3, mm6 ; q0-=p0
+ por mm5, mm3 ; abs(p0 - q0)
+ paddusb mm5, mm5 ; abs(p0-q0)*2
+ paddusb mm5, mm2 ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+ mov rdx, arg(2) ;blimit ; get blimit
+ movq mm7, [rdx] ; blimit
+
+ psubusb mm5, mm7 ; abs (p0 - q0) *2 + abs(p1-q1)/2 > blimit
+ por mm1, mm5
+ pxor mm5, mm5
+ pcmpeqb mm1, mm5 ; mask mm1
+
+ ; calculate high edge variance
+ mov rdx, arg(4) ;thresh ; get thresh
+ movq mm7, [rdx] ;
+ movq mm4, t0 ; get abs (q1 - q0)
+ psubusb mm4, mm7
+ movq mm3, t1 ; get abs (p1 - p0)
+ psubusb mm3, mm7
+ paddb mm4, mm3 ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+
+ pcmpeqb mm4, mm5
+
+ pcmpeqb mm5, mm5
+ pxor mm4, mm5
+
+
+ ; start work on filters
+ movq mm2, [rsi+2*rax] ; p1
+ movq mm7, [rdi] ; q1
+ pxor mm2, [GLOBAL(t80)] ; p1 offset to convert to signed values
+ pxor mm7, [GLOBAL(t80)] ; q1 offset to convert to signed values
+ psubsb mm2, mm7 ; p1 - q1
+ pand mm2, mm4 ; high var mask (hvm)(p1 - q1)
+ pxor mm6, [GLOBAL(t80)] ; offset to convert to signed values
+ pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
+ movq mm3, mm0 ; q0
+ psubsb mm0, mm6 ; q0 - p0
+ paddsb mm2, mm0 ; 1 * (q0 - p0) + hvm(p1 - q1)
+ paddsb mm2, mm0 ; 2 * (q0 - p0) + hvm(p1 - q1)
+ paddsb mm2, mm0 ; 3 * (q0 - p0) + hvm(p1 - q1)
+ pand mm1, mm2 ; mask filter values we don't care about
+ movq mm2, mm1
+ paddsb mm1, [GLOBAL(t4)] ; 3* (q0 - p0) + hvm(p1 - q1) + 4
+ paddsb mm2, [GLOBAL(t3)] ; 3* (q0 - p0) + hvm(p1 - q1) + 3
+
+ pxor mm0, mm0 ;
+ pxor mm5, mm5
+ punpcklbw mm0, mm2 ;
+ punpckhbw mm5, mm2 ;
+ psraw mm0, 11 ;
+ psraw mm5, 11
+ packsswb mm0, mm5
+ movq mm2, mm0 ; (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
+
+ pxor mm0, mm0 ; 0
+ movq mm5, mm1 ; abcdefgh
+ punpcklbw mm0, mm1 ; e0f0g0h0
+ psraw mm0, 11 ; sign extended shift right by 3
+ pxor mm1, mm1 ; 0
+ punpckhbw mm1, mm5 ; a0b0c0d0
+ psraw mm1, 11 ; sign extended shift right by 3
+ movq mm5, mm0 ; save results
+
+ packsswb mm0, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
+ paddsw mm5, [GLOBAL(ones)]
+ paddsw mm1, [GLOBAL(ones)]
+ psraw mm5, 1 ; partial shifted one more time for 2nd tap
+ psraw mm1, 1 ; partial shifted one more time for 2nd tap
+ packsswb mm5, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
+ pandn mm4, mm5 ; high edge variance additive
+
+ paddsb mm6, mm2 ; p0+= p0 add
+ pxor mm6, [GLOBAL(t80)] ; unoffset
+ movq [rsi+rax], mm6 ; write back
+
+ movq mm6, [rsi+2*rax] ; p1
+ pxor mm6, [GLOBAL(t80)] ; reoffset
+ paddsb mm6, mm4 ; p1+= p1 add
+ pxor mm6, [GLOBAL(t80)] ; unoffset
+ movq [rsi+2*rax], mm6 ; write back
+
+ psubsb mm3, mm0 ; q0-= q0 add
+ pxor mm3, [GLOBAL(t80)] ; unoffset
+ movq [rsi], mm3 ; write back
+
+ psubsb mm7, mm4 ; q1-= q1 add
+ pxor mm7, [GLOBAL(t80)] ; unoffset
+ movq [rdi], mm7 ; write back
+
+ add rsi,8
+ neg rax
+ dec rcx
+ jnz .next8_h
+
+ add rsp, 32
+ pop rsp
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_GOT
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+
+;void vp9_lpf_vertical_4_mmx
+;(
+; unsigned char *src_ptr,
+; int src_pixel_step,
+; const char *blimit,
+; const char *limit,
+; const char *thresh,
+; int count
+;)
+global sym(vp9_lpf_vertical_4_mmx) PRIVATE
+sym(vp9_lpf_vertical_4_mmx):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 6
+ GET_GOT rbx
+ push rsi
+ push rdi
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 64 ; reserve 64 bytes
+ %define t0 [rsp + 0] ;__declspec(align(16)) char t0[8];
+ %define t1 [rsp + 16] ;__declspec(align(16)) char t1[8];
+ %define srct [rsp + 32] ;__declspec(align(16)) char srct[32];
+
+ mov rsi, arg(0) ;src_ptr
+ movsxd rax, dword ptr arg(1) ;src_pixel_step ; destination pitch?
+
+ lea rsi, [rsi + rax*4 - 4]
+
+ movsxd rcx, dword ptr arg(5) ;count
+.next8_v:
+ mov rdi, rsi ; rdi points to row +1 for indirect addressing
+ add rdi, rax
+
+
+ ;transpose
+ movq mm6, [rsi+2*rax] ; 67 66 65 64 63 62 61 60
+ movq mm7, mm6 ; 77 76 75 74 73 72 71 70
+
+ punpckhbw mm7, [rdi+2*rax] ; 77 67 76 66 75 65 74 64
+ punpcklbw mm6, [rdi+2*rax] ; 73 63 72 62 71 61 70 60
+
+ movq mm4, [rsi] ; 47 46 45 44 43 42 41 40
+ movq mm5, mm4 ; 47 46 45 44 43 42 41 40
+
+ punpckhbw mm5, [rsi+rax] ; 57 47 56 46 55 45 54 44
+ punpcklbw mm4, [rsi+rax] ; 53 43 52 42 51 41 50 40
+
+ movq mm3, mm5 ; 57 47 56 46 55 45 54 44
+ punpckhwd mm5, mm7 ; 77 67 57 47 76 66 56 46
+
+ punpcklwd mm3, mm7 ; 75 65 55 45 74 64 54 44
+ movq mm2, mm4 ; 53 43 52 42 51 41 50 40
+
+ punpckhwd mm4, mm6 ; 73 63 53 43 72 62 52 42
+ punpcklwd mm2, mm6 ; 71 61 51 41 70 60 50 40
+
+ neg rax
+ movq mm6, [rsi+rax*2] ; 27 26 25 24 23 22 21 20
+
+ movq mm1, mm6 ; 27 26 25 24 23 22 21 20
+ punpckhbw mm6, [rsi+rax] ; 37 27 36 36 35 25 34 24
+
+ punpcklbw mm1, [rsi+rax] ; 33 23 32 22 31 21 30 20
+ movq mm7, [rsi+rax*4]; ; 07 06 05 04 03 02 01 00
+
+ punpckhbw mm7, [rdi+rax*4] ; 17 07 16 06 15 05 14 04
+ movq mm0, mm7 ; 17 07 16 06 15 05 14 04
+
+ punpckhwd mm7, mm6 ; 37 27 17 07 36 26 16 06
+ punpcklwd mm0, mm6 ; 35 25 15 05 34 24 14 04
+
+ movq mm6, mm7 ; 37 27 17 07 36 26 16 06
+ punpckhdq mm7, mm5 ; 77 67 57 47 37 27 17 07 = q3
+
+ punpckldq mm6, mm5 ; 76 66 56 46 36 26 16 06 = q2
+
+ movq mm5, mm6 ; 76 66 56 46 36 26 16 06
+ psubusb mm5, mm7 ; q2-q3
+
+ psubusb mm7, mm6 ; q3-q2
+ por mm7, mm5; ; mm7=abs (q3-q2)
+
+ movq mm5, mm0 ; 35 25 15 05 34 24 14 04
+ punpckhdq mm5, mm3 ; 75 65 55 45 35 25 15 05 = q1
+
+ punpckldq mm0, mm3 ; 74 64 54 44 34 24 15 04 = q0
+ movq mm3, mm5 ; 75 65 55 45 35 25 15 05 = q1
+
+ psubusb mm3, mm6 ; q1-q2
+ psubusb mm6, mm5 ; q2-q1
+
+ por mm6, mm3 ; mm6=abs(q2-q1)
+ lea rdx, srct
+
+ movq [rdx+24], mm5 ; save q1
+ movq [rdx+16], mm0 ; save q0
+
+ movq mm3, [rsi+rax*4] ; 07 06 05 04 03 02 01 00
+ punpcklbw mm3, [rdi+rax*4] ; 13 03 12 02 11 01 10 00
+
+ movq mm0, mm3 ; 13 03 12 02 11 01 10 00
+ punpcklwd mm0, mm1 ; 31 21 11 01 30 20 10 00
+
+ punpckhwd mm3, mm1 ; 33 23 13 03 32 22 12 02
+ movq mm1, mm0 ; 31 21 11 01 30 20 10 00
+
+ punpckldq mm0, mm2 ; 70 60 50 40 30 20 10 00 =p3
+ punpckhdq mm1, mm2 ; 71 61 51 41 31 21 11 01 =p2
+
+ movq mm2, mm1 ; 71 61 51 41 31 21 11 01 =p2
+ psubusb mm2, mm0 ; p2-p3
+
+ psubusb mm0, mm1 ; p3-p2
+ por mm0, mm2 ; mm0=abs(p3-p2)
+
+ movq mm2, mm3 ; 33 23 13 03 32 22 12 02
+ punpckldq mm2, mm4 ; 72 62 52 42 32 22 12 02 = p1
+
+ punpckhdq mm3, mm4 ; 73 63 53 43 33 23 13 03 = p0
+ movq [rdx+8], mm3 ; save p0
+
+ movq [rdx], mm2 ; save p1
+ movq mm5, mm2 ; mm5 = p1
+
+ psubusb mm2, mm1 ; p1-p2
+ psubusb mm1, mm5 ; p2-p1
+
+ por mm1, mm2 ; mm1=abs(p2-p1)
+ mov rdx, arg(3) ;limit
+
+ movq mm4, [rdx] ; mm4 = limit
+ psubusb mm7, mm4
+
+ psubusb mm0, mm4
+ psubusb mm1, mm4
+
+ psubusb mm6, mm4
+ por mm7, mm6
+
+ por mm0, mm1
+ por mm0, mm7 ; abs(q3-q2) > limit || abs(p3-p2) > limit ||abs(p2-p1) > limit || abs(q2-q1) > limit
+
+ movq mm1, mm5 ; p1
+
+ movq mm7, mm3 ; mm3=mm7=p0
+ psubusb mm7, mm5 ; p0 - p1
+
+ psubusb mm5, mm3 ; p1 - p0
+ por mm5, mm7 ; abs(p1-p0)
+
+ movq t0, mm5 ; save abs(p1-p0)
+ lea rdx, srct
+
+ psubusb mm5, mm4
+ por mm0, mm5 ; mm0=mask
+
+ movq mm5, [rdx+16] ; mm5=q0
+ movq mm7, [rdx+24] ; mm7=q1
+
+ movq mm6, mm5 ; mm6=q0
+ movq mm2, mm7 ; q1
+ psubusb mm5, mm7 ; q0-q1
+
+ psubusb mm7, mm6 ; q1-q0
+ por mm7, mm5 ; abs(q1-q0)
+
+ movq t1, mm7 ; save abs(q1-q0)
+ psubusb mm7, mm4
+
+ por mm0, mm7 ; mask
+
+ movq mm5, mm2 ; q1
+ psubusb mm5, mm1 ; q1-=p1
+ psubusb mm1, mm2 ; p1-=q1
+ por mm5, mm1 ; abs(p1-q1)
+ pand mm5, [GLOBAL(tfe)] ; set lsb of each byte to zero
+ psrlw mm5, 1 ; abs(p1-q1)/2
+
+ mov rdx, arg(2) ;blimit ;
+
+ movq mm4, [rdx] ;blimit
+ movq mm1, mm3 ; mm1=mm3=p0
+
+ movq mm7, mm6 ; mm7=mm6=q0
+ psubusb mm1, mm7 ; p0-q0
+
+ psubusb mm7, mm3 ; q0-p0
+ por mm1, mm7 ; abs(q0-p0)
+ paddusb mm1, mm1 ; abs(q0-p0)*2
+ paddusb mm1, mm5 ; abs (p0 - q0) *2 + abs(p1-q1)/2
+
+ psubusb mm1, mm4 ; abs (p0 - q0) *2 + abs(p1-q1)/2 > blimit
+ por mm1, mm0; ; mask
+
+ pxor mm0, mm0
+ pcmpeqb mm1, mm0
+
+ ; calculate high edge variance
+ mov rdx, arg(4) ;thresh ; get thresh
+ movq mm7, [rdx]
+ ;
+ movq mm4, t0 ; get abs (q1 - q0)
+ psubusb mm4, mm7
+
+ movq mm3, t1 ; get abs (p1 - p0)
+ psubusb mm3, mm7
+
+ por mm4, mm3 ; abs(q1 - q0) > thresh || abs(p1 - p0) > thresh
+ pcmpeqb mm4, mm0
+
+ pcmpeqb mm0, mm0
+ pxor mm4, mm0
+
+
+
+ ; start work on filters
+ lea rdx, srct
+
+ movq mm2, [rdx] ; p1
+ movq mm7, [rdx+24] ; q1
+
+ movq mm6, [rdx+8] ; p0
+ movq mm0, [rdx+16] ; q0
+
+ pxor mm2, [GLOBAL(t80)] ; p1 offset to convert to signed values
+ pxor mm7, [GLOBAL(t80)] ; q1 offset to convert to signed values
+
+ psubsb mm2, mm7 ; p1 - q1
+ pand mm2, mm4 ; high var mask (hvm)(p1 - q1)
+
+ pxor mm6, [GLOBAL(t80)] ; offset to convert to signed values
+ pxor mm0, [GLOBAL(t80)] ; offset to convert to signed values
+
+ movq mm3, mm0 ; q0
+ psubsb mm0, mm6 ; q0 - p0
+
+ paddsb mm2, mm0 ; 1 * (q0 - p0) + hvm(p1 - q1)
+ paddsb mm2, mm0 ; 2 * (q0 - p0) + hvm(p1 - q1)
+
+ paddsb mm2, mm0 ; 3 * (q0 - p0) + hvm(p1 - q1)
+ pand mm1, mm2 ; mask filter values we don't care about
+
+ movq mm2, mm1
+ paddsb mm1, [GLOBAL(t4)] ; 3* (q0 - p0) + hvm(p1 - q1) + 4
+
+ paddsb mm2, [GLOBAL(t3)] ; 3* (q0 - p0) + hvm(p1 - q1) + 3
+ pxor mm0, mm0 ;
+
+ pxor mm5, mm5
+ punpcklbw mm0, mm2 ;
+
+ punpckhbw mm5, mm2 ;
+ psraw mm0, 11 ;
+
+ psraw mm5, 11
+ packsswb mm0, mm5
+
+ movq mm2, mm0 ; (3* (q0 - p0) + hvm(p1 - q1) + 3) >> 3;
+
+ pxor mm0, mm0 ; 0
+ movq mm5, mm1 ; abcdefgh
+
+ punpcklbw mm0, mm1 ; e0f0g0h0
+ psraw mm0, 11 ; sign extended shift right by 3
+
+ pxor mm1, mm1 ; 0
+ punpckhbw mm1, mm5 ; a0b0c0d0
+
+ psraw mm1, 11 ; sign extended shift right by 3
+ movq mm5, mm0 ; save results
+
+ packsswb mm0, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>3
+ paddsw mm5, [GLOBAL(ones)]
+
+ paddsw mm1, [GLOBAL(ones)]
+ psraw mm5, 1 ; partial shifted one more time for 2nd tap
+
+ psraw mm1, 1 ; partial shifted one more time for 2nd tap
+ packsswb mm5, mm1 ; (3* (q0 - p0) + hvm(p1 - q1) + 4) >>4
+
+ pandn mm4, mm5 ; high edge variance additive
+
+ paddsb mm6, mm2 ; p0+= p0 add
+ pxor mm6, [GLOBAL(t80)] ; unoffset
+
+ ; mm6=p0 ;
+ movq mm1, [rdx] ; p1
+ pxor mm1, [GLOBAL(t80)] ; reoffset
+
+ paddsb mm1, mm4 ; p1+= p1 add
+ pxor mm1, [GLOBAL(t80)] ; unoffset
+ ; mm6 = p0 mm1 = p1
+
+ psubsb mm3, mm0 ; q0-= q0 add
+ pxor mm3, [GLOBAL(t80)] ; unoffset
+
+ ; mm3 = q0
+ psubsb mm7, mm4 ; q1-= q1 add
+ pxor mm7, [GLOBAL(t80)] ; unoffset
+ ; mm7 = q1
+
+ ; transpose and write back
+ ; mm1 = 72 62 52 42 32 22 12 02
+ ; mm6 = 73 63 53 43 33 23 13 03
+ ; mm3 = 74 64 54 44 34 24 14 04
+ ; mm7 = 75 65 55 45 35 25 15 05
+
+ movq mm2, mm1 ; 72 62 52 42 32 22 12 02
+ punpcklbw mm2, mm6 ; 33 32 23 22 13 12 03 02
+
+ movq mm4, mm3 ; 74 64 54 44 34 24 14 04
+ punpckhbw mm1, mm6 ; 73 72 63 62 53 52 43 42
+
+ punpcklbw mm4, mm7 ; 35 34 25 24 15 14 05 04
+ punpckhbw mm3, mm7 ; 75 74 65 64 55 54 45 44
+
+ movq mm6, mm2 ; 33 32 23 22 13 12 03 02
+ punpcklwd mm2, mm4 ; 15 14 13 12 05 04 03 02
+
+ punpckhwd mm6, mm4 ; 35 34 33 32 25 24 23 22
+ movq mm5, mm1 ; 73 72 63 62 53 52 43 42
+
+ punpcklwd mm1, mm3 ; 55 54 53 52 45 44 43 42
+ punpckhwd mm5, mm3 ; 75 74 73 72 65 64 63 62
+
+
+ ; mm2 = 15 14 13 12 05 04 03 02
+ ; mm6 = 35 34 33 32 25 24 23 22
+ ; mm5 = 55 54 53 52 45 44 43 42
+ ; mm1 = 75 74 73 72 65 64 63 62
+
+
+
+ movd [rsi+rax*4+2], mm2
+ psrlq mm2, 32
+
+ movd [rdi+rax*4+2], mm2
+ movd [rsi+rax*2+2], mm6
+
+ psrlq mm6, 32
+ movd [rsi+rax+2],mm6
+
+ movd [rsi+2], mm1
+ psrlq mm1, 32
+
+ movd [rdi+2], mm1
+ neg rax
+
+ movd [rdi+rax+2],mm5
+ psrlq mm5, 32
+
+ movd [rdi+rax*2+2], mm5
+
+ lea rsi, [rsi+rax*8]
+ dec rcx
+ jnz .next8_v
+
+ add rsp, 64
+ pop rsp
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_GOT
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+SECTION_RODATA
+align 16
+tfe:
+ times 8 db 0xfe
+align 16
+t80:
+ times 8 db 0x80
+align 16
+t3:
+ times 8 db 0x03
+align 16
+t4:
+ times 8 db 0x04
+align 16
+ones:
+ times 4 dw 0x0001
--- /dev/null
+++ b/vpx_dsp/x86/loopfilter_sse2.c
@@ -1,0 +1,1587 @@
+/*
+ * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ * Use of this source code is governed by a BSD-style license
+ * that can be found in the LICENSE file in the root of the source
+ * tree. An additional intellectual property rights grant can be found
+ * in the file PATENTS. All contributing project authors may
+ * be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h> // SSE2
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vpx_ports/emmintrin_compat.h"
+
+static INLINE __m128i abs_diff(__m128i a, __m128i b) {+ return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a));
+}
+
+static void mb_lpf_horizontal_edge_w_sse2_8(unsigned char *s,
+ int p,
+ const unsigned char *_blimit,
+ const unsigned char *_limit,
+ const unsigned char *_thresh) {+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i one = _mm_set1_epi8(1);
+ const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+ const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+ const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+ __m128i mask, hev, flat, flat2;
+ __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0, p0q0, p1q1;
+ __m128i abs_p1p0;
+
+ q4p4 = _mm_loadl_epi64((__m128i *)(s - 5 * p));
+ q4p4 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q4p4),
+ (__m64 *)(s + 4 * p)));
+ q3p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+ q3p3 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q3p3),
+ (__m64 *)(s + 3 * p)));
+ q2p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+ q2p2 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q2p2),
+ (__m64 *)(s + 2 * p)));
+ q1p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+ q1p1 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q1p1),
+ (__m64 *)(s + 1 * p)));
+ p1q1 = _mm_shuffle_epi32(q1p1, 78);
+ q0p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+ q0p0 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q0p0),
+ (__m64 *)(s - 0 * p)));
+ p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+ {+ __m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
+ abs_p1p0 = abs_diff(q1p1, q0p0);
+ abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+ fe = _mm_set1_epi8(0xfe);
+ ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+ abs_p0q0 = abs_diff(q0p0, p0q0);
+ abs_p1q1 = abs_diff(q1p1, p1q1);
+ flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu8(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+ abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+ mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ mask = _mm_max_epu8(abs_p1p0, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+
+ work = _mm_max_epu8(abs_diff(q2p2, q1p1),
+ abs_diff(q3p3, q2p2));
+ mask = _mm_max_epu8(work, mask);
+ mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+ mask = _mm_subs_epu8(mask, limit);
+ mask = _mm_cmpeq_epi8(mask, zero);
+ }
+
+ // lp filter
+ {+ const __m128i t4 = _mm_set1_epi8(4);
+ const __m128i t3 = _mm_set1_epi8(3);
+ const __m128i t80 = _mm_set1_epi8(0x80);
+ const __m128i t1 = _mm_set1_epi16(0x1);
+ __m128i qs1ps1 = _mm_xor_si128(q1p1, t80);
+ __m128i qs0ps0 = _mm_xor_si128(q0p0, t80);
+ __m128i qs0 = _mm_xor_si128(p0q0, t80);
+ __m128i qs1 = _mm_xor_si128(p1q1, t80);
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+ __m128i flat2_q6p6, flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+ __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+ filt = _mm_and_si128(_mm_subs_epi8(qs1ps1, qs1), hev);
+ work_a = _mm_subs_epi8(qs0, qs0ps0);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ // (vp9_filter + 3 * (qs0 - ps0)) & mask
+ filt = _mm_and_si128(filt, mask);
+
+ filter1 = _mm_adds_epi8(filt, t4);
+ filter2 = _mm_adds_epi8(filt, t3);
+
+ filter1 = _mm_unpacklo_epi8(zero, filter1);
+ filter1 = _mm_srai_epi16(filter1, 0xB);
+ filter2 = _mm_unpacklo_epi8(zero, filter2);
+ filter2 = _mm_srai_epi16(filter2, 0xB);
+
+ // Filter1 >> 3
+ filt = _mm_packs_epi16(filter2, _mm_subs_epi16(zero, filter1));
+ qs0ps0 = _mm_xor_si128(_mm_adds_epi8(qs0ps0, filt), t80);
+
+ // filt >> 1
+ filt = _mm_adds_epi16(filter1, t1);
+ filt = _mm_srai_epi16(filt, 1);
+ filt = _mm_andnot_si128(_mm_srai_epi16(_mm_unpacklo_epi8(zero, hev), 0x8),
+ filt);
+ filt = _mm_packs_epi16(filt, _mm_subs_epi16(zero, filt));
+ qs1ps1 = _mm_xor_si128(_mm_adds_epi8(qs1ps1, filt), t80);
+ // loopfilter done
+
+ {+ __m128i work;
+ flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
+ flat = _mm_max_epu8(abs_p1p0, flat);
+ flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+ flat = _mm_subs_epu8(flat, one);
+ flat = _mm_cmpeq_epi8(flat, zero);
+ flat = _mm_and_si128(flat, mask);
+
+ q5p5 = _mm_loadl_epi64((__m128i *)(s - 6 * p));
+ q5p5 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q5p5),
+ (__m64 *)(s + 5 * p)));
+
+ q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
+ q6p6 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q6p6),
+ (__m64 *)(s + 6 * p)));
+ flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0));
+
+ q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
+ q7p7 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q7p7),
+ (__m64 *)(s + 7 * p)));
+ work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0));
+ flat2 = _mm_max_epu8(work, flat2);
+ flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+ flat2 = _mm_subs_epu8(flat2, one);
+ flat2 = _mm_cmpeq_epi8(flat2, zero);
+ flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
+ }
+
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // flat and wide flat calculations
+ {+ const __m128i eight = _mm_set1_epi16(8);
+ const __m128i four = _mm_set1_epi16(4);
+ __m128i p7_16, p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+ __m128i q7_16, q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+ __m128i pixelFilter_p, pixelFilter_q;
+ __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+ __m128i sum_p7, sum_q7, sum_p3, sum_q3, res_p, res_q;
+
+ p7_16 = _mm_unpacklo_epi8(q7p7, zero);;
+ p6_16 = _mm_unpacklo_epi8(q6p6, zero);
+ p5_16 = _mm_unpacklo_epi8(q5p5, zero);
+ p4_16 = _mm_unpacklo_epi8(q4p4, zero);
+ p3_16 = _mm_unpacklo_epi8(q3p3, zero);
+ p2_16 = _mm_unpacklo_epi8(q2p2, zero);
+ p1_16 = _mm_unpacklo_epi8(q1p1, zero);
+ p0_16 = _mm_unpacklo_epi8(q0p0, zero);
+ q0_16 = _mm_unpackhi_epi8(q0p0, zero);
+ q1_16 = _mm_unpackhi_epi8(q1p1, zero);
+ q2_16 = _mm_unpackhi_epi8(q2p2, zero);
+ q3_16 = _mm_unpackhi_epi8(q3p3, zero);
+ q4_16 = _mm_unpackhi_epi8(q4p4, zero);
+ q5_16 = _mm_unpackhi_epi8(q5p5, zero);
+ q6_16 = _mm_unpackhi_epi8(q6p6, zero);
+ q7_16 = _mm_unpackhi_epi8(q7p7, zero);
+
+ pixelFilter_p = _mm_add_epi16(_mm_add_epi16(p6_16, p5_16),
+ _mm_add_epi16(p4_16, p3_16));
+ pixelFilter_q = _mm_add_epi16(_mm_add_epi16(q6_16, q5_16),
+ _mm_add_epi16(q4_16, q3_16));
+
+ pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
+ pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+ pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
+ pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+ pixelFilter_p = _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p,
+ pixelFilter_q));
+ pixetFilter_p2p1p0 = _mm_add_epi16(four,
+ _mm_add_epi16(pixetFilter_p2p1p0,
+ pixetFilter_q2q1q0));
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(p7_16, p0_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(q7_16, q0_16)), 4);
+ flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(p3_16, p0_16)), 3);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(q3_16, q0_16)), 3);
+
+ flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(p7_16, p7_16);
+ sum_q7 = _mm_add_epi16(q7_16, q7_16);
+ sum_p3 = _mm_add_epi16(p3_16, p3_16);
+ sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p6_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q6_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p1_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q1_16)), 4);
+ flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+ pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+ pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(sum_p3, p1_16)), 3);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+ _mm_add_epi16(sum_q3, q1_16)), 3);
+ flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+ sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p5_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p2_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q2_16)), 4);
+ flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+ pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+ pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixetFilter_p2p1p0,
+ _mm_add_epi16(sum_p3, p2_16)), 3);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixetFilter_q2q1q0,
+ _mm_add_epi16(sum_q3, q2_16)), 3);
+ flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p3_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q3_16)), 4);
+ flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p4_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q4_16)), 4);
+ flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p5_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q5_16)), 4);
+ flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+ sum_p7 = _mm_add_epi16(sum_p7, p7_16);
+ sum_q7 = _mm_add_epi16(sum_q7, q7_16);
+ pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+ pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+ res_p = _mm_srli_epi16(_mm_add_epi16(pixelFilter_p,
+ _mm_add_epi16(sum_p7, p6_16)), 4);
+ res_q = _mm_srli_epi16(_mm_add_epi16(pixelFilter_q,
+ _mm_add_epi16(sum_q7, q6_16)), 4);
+ flat2_q6p6 = _mm_packus_epi16(res_p, res_q);
+ }
+ // wide flat
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+ flat = _mm_shuffle_epi32(flat, 68);
+ flat2 = _mm_shuffle_epi32(flat2, 68);
+
+ q2p2 = _mm_andnot_si128(flat, q2p2);
+ flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+ q2p2 = _mm_or_si128(q2p2, flat_q2p2);
+
+ qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+ flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+ q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+ qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+ flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+ q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+ q6p6 = _mm_andnot_si128(flat2, q6p6);
+ flat2_q6p6 = _mm_and_si128(flat2, flat2_q6p6);
+ q6p6 = _mm_or_si128(q6p6, flat2_q6p6);
+ _mm_storel_epi64((__m128i *)(s - 7 * p), q6p6);
+ _mm_storeh_pi((__m64 *)(s + 6 * p), _mm_castsi128_ps(q6p6));
+
+ q5p5 = _mm_andnot_si128(flat2, q5p5);
+ flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+ q5p5 = _mm_or_si128(q5p5, flat2_q5p5);
+ _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
+ _mm_storeh_pi((__m64 *)(s + 5 * p), _mm_castsi128_ps(q5p5));
+
+ q4p4 = _mm_andnot_si128(flat2, q4p4);
+ flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+ q4p4 = _mm_or_si128(q4p4, flat2_q4p4);
+ _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
+ _mm_storeh_pi((__m64 *)(s + 4 * p), _mm_castsi128_ps(q4p4));
+
+ q3p3 = _mm_andnot_si128(flat2, q3p3);
+ flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+ q3p3 = _mm_or_si128(q3p3, flat2_q3p3);
+ _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
+ _mm_storeh_pi((__m64 *)(s + 3 * p), _mm_castsi128_ps(q3p3));
+
+ q2p2 = _mm_andnot_si128(flat2, q2p2);
+ flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+ q2p2 = _mm_or_si128(q2p2, flat2_q2p2);
+ _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
+ _mm_storeh_pi((__m64 *)(s + 2 * p), _mm_castsi128_ps(q2p2));
+
+ q1p1 = _mm_andnot_si128(flat2, q1p1);
+ flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+ q1p1 = _mm_or_si128(q1p1, flat2_q1p1);
+ _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
+ _mm_storeh_pi((__m64 *)(s + 1 * p), _mm_castsi128_ps(q1p1));
+
+ q0p0 = _mm_andnot_si128(flat2, q0p0);
+ flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+ q0p0 = _mm_or_si128(q0p0, flat2_q0p0);
+ _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
+ _mm_storeh_pi((__m64 *)(s - 0 * p), _mm_castsi128_ps(q0p0));
+ }
+}
+
+static INLINE __m128i filter_add2_sub2(const __m128i *const total,
+ const __m128i *const a1,
+ const __m128i *const a2,
+ const __m128i *const s1,
+ const __m128i *const s2) {+ __m128i x = _mm_add_epi16(*a1, *total);
+ x = _mm_add_epi16(_mm_sub_epi16(x, _mm_add_epi16(*s1, *s2)), *a2);
+ return x;
+}
+
+static INLINE __m128i filter8_mask(const __m128i *const flat,
+ const __m128i *const other_filt,
+ const __m128i *const f8_lo,
+ const __m128i *const f8_hi) {+ const __m128i f8 = _mm_packus_epi16(_mm_srli_epi16(*f8_lo, 3),
+ _mm_srli_epi16(*f8_hi, 3));
+ const __m128i result = _mm_and_si128(*flat, f8);
+ return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+static INLINE __m128i filter16_mask(const __m128i *const flat,
+ const __m128i *const other_filt,
+ const __m128i *const f_lo,
+ const __m128i *const f_hi) {+ const __m128i f = _mm_packus_epi16(_mm_srli_epi16(*f_lo, 4),
+ _mm_srli_epi16(*f_hi, 4));
+ const __m128i result = _mm_and_si128(*flat, f);
+ return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+static void mb_lpf_horizontal_edge_w_sse2_16(unsigned char *s,
+ int p,
+ const unsigned char *_blimit,
+ const unsigned char *_limit,
+ const unsigned char *_thresh) {+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i one = _mm_set1_epi8(1);
+ const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+ const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+ const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+ __m128i mask, hev, flat, flat2;
+ __m128i p7, p6, p5;
+ __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+ __m128i q5, q6, q7;
+
+ __m128i op2, op1, op0, oq0, oq1, oq2;
+
+ __m128i max_abs_p1p0q1q0;
+
+ p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
+ p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
+ p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
+ p4 = _mm_loadu_si128((__m128i *)(s - 5 * p));
+ p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+ p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+ p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+ p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+ q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+ q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+ q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+ q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+ q4 = _mm_loadu_si128((__m128i *)(s + 4 * p));
+ q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
+ q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
+ q7 = _mm_loadu_si128((__m128i *)(s + 7 * p));
+
+ {+ const __m128i abs_p1p0 = abs_diff(p1, p0);
+ const __m128i abs_q1q0 = abs_diff(q1, q0);
+ const __m128i fe = _mm_set1_epi8(0xfe);
+ const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+ __m128i abs_p0q0 = abs_diff(p0, q0);
+ __m128i abs_p1q1 = abs_diff(p1, q1);
+ __m128i work;
+ max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+ abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+ mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+ work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2));
+ mask = _mm_max_epu8(work, mask);
+ work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2));
+ mask = _mm_max_epu8(work, mask);
+ mask = _mm_subs_epu8(mask, limit);
+ mask = _mm_cmpeq_epi8(mask, zero);
+ }
+
+ {+ __m128i work;
+ work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
+ flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
+ work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0));
+ flat = _mm_max_epu8(work, flat);
+ work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0));
+ flat = _mm_subs_epu8(flat, one);
+ flat = _mm_cmpeq_epi8(flat, zero);
+ flat = _mm_and_si128(flat, mask);
+ flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0));
+ flat2 = _mm_max_epu8(work, flat2);
+ work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0));
+ flat2 = _mm_max_epu8(work, flat2);
+ work = _mm_max_epu8(abs_diff(p7, p0), abs_diff(q7, q0));
+ flat2 = _mm_max_epu8(work, flat2);
+ flat2 = _mm_subs_epu8(flat2, one);
+ flat2 = _mm_cmpeq_epi8(flat2, zero);
+ flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
+ }
+
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // filter4
+ {+ const __m128i t4 = _mm_set1_epi8(4);
+ const __m128i t3 = _mm_set1_epi8(3);
+ const __m128i t80 = _mm_set1_epi8(0x80);
+ const __m128i te0 = _mm_set1_epi8(0xe0);
+ const __m128i t1f = _mm_set1_epi8(0x1f);
+ const __m128i t1 = _mm_set1_epi8(0x1);
+ const __m128i t7f = _mm_set1_epi8(0x7f);
+ const __m128i ff = _mm_cmpeq_epi8(t4, t4);
+
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+
+ op1 = _mm_xor_si128(p1, t80);
+ op0 = _mm_xor_si128(p0, t80);
+ oq0 = _mm_xor_si128(q0, t80);
+ oq1 = _mm_xor_si128(q1, t80);
+
+ hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+ filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+ work_a = _mm_subs_epi8(oq0, op0);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ // (vp9_filter + 3 * (qs0 - ps0)) & mask
+ filt = _mm_and_si128(filt, mask);
+ filter1 = _mm_adds_epi8(filt, t4);
+ filter2 = _mm_adds_epi8(filt, t3);
+
+ // Filter1 >> 3
+ work_a = _mm_cmpgt_epi8(zero, filter1);
+ filter1 = _mm_srli_epi16(filter1, 3);
+ work_a = _mm_and_si128(work_a, te0);
+ filter1 = _mm_and_si128(filter1, t1f);
+ filter1 = _mm_or_si128(filter1, work_a);
+ oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
+
+ // Filter2 >> 3
+ work_a = _mm_cmpgt_epi8(zero, filter2);
+ filter2 = _mm_srli_epi16(filter2, 3);
+ work_a = _mm_and_si128(work_a, te0);
+ filter2 = _mm_and_si128(filter2, t1f);
+ filter2 = _mm_or_si128(filter2, work_a);
+ op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
+
+ // filt >> 1
+ filt = _mm_adds_epi8(filter1, t1);
+ work_a = _mm_cmpgt_epi8(zero, filt);
+ filt = _mm_srli_epi16(filt, 1);
+ work_a = _mm_and_si128(work_a, t80);
+ filt = _mm_and_si128(filt, t7f);
+ filt = _mm_or_si128(filt, work_a);
+ filt = _mm_andnot_si128(hev, filt);
+ op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+ oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
+ // loopfilter done
+
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // filter8
+ {+ const __m128i four = _mm_set1_epi16(4);
+ const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+ const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+ const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+ const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+ const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+ const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+ const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+ const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+
+ const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+ const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+ const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+ const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+ const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+ const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+ const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+ const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+ __m128i f8_lo, f8_hi;
+
+ f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four),
+ _mm_add_epi16(p3_lo, p2_lo));
+ f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo),
+ _mm_add_epi16(p2_lo, p1_lo));
+ f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
+
+ f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four),
+ _mm_add_epi16(p3_hi, p2_hi));
+ f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi),
+ _mm_add_epi16(p2_hi, p1_hi));
+ f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
+
+ op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi);
+
+ f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo);
+ f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi);
+ op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi);
+
+ f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo);
+ f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi);
+ op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi);
+
+ f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo);
+ f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi);
+ oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi);
+
+ f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo);
+ f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi);
+ oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi);
+
+ f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo);
+ f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi);
+ oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi);
+ }
+
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // wide flat calculations
+ {+ const __m128i eight = _mm_set1_epi16(8);
+ const __m128i p7_lo = _mm_unpacklo_epi8(p7, zero);
+ const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero);
+ const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero);
+ const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero);
+ const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+ const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+ const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+ const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+ const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+ const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+ const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+ const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+ const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero);
+ const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero);
+ const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero);
+ const __m128i q7_lo = _mm_unpacklo_epi8(q7, zero);
+
+ const __m128i p7_hi = _mm_unpackhi_epi8(p7, zero);
+ const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero);
+ const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero);
+ const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero);
+ const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+ const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+ const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+ const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+ const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+ const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+ const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+ const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+ const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero);
+ const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero);
+ const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero);
+ const __m128i q7_hi = _mm_unpackhi_epi8(q7, zero);
+
+ __m128i f_lo;
+ __m128i f_hi;
+
+ f_lo = _mm_sub_epi16(_mm_slli_epi16(p7_lo, 3), p7_lo); // p7 * 7
+ f_lo = _mm_add_epi16(_mm_slli_epi16(p6_lo, 1),
+ _mm_add_epi16(p4_lo, f_lo));
+ f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo),
+ _mm_add_epi16(p2_lo, p1_lo));
+ f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo);
+ f_lo = _mm_add_epi16(_mm_add_epi16(p5_lo, eight), f_lo);
+
+ f_hi = _mm_sub_epi16(_mm_slli_epi16(p7_hi, 3), p7_hi); // p7 * 7
+ f_hi = _mm_add_epi16(_mm_slli_epi16(p6_hi, 1),
+ _mm_add_epi16(p4_hi, f_hi));
+ f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi),
+ _mm_add_epi16(p2_hi, p1_hi));
+ f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi);
+ f_hi = _mm_add_epi16(_mm_add_epi16(p5_hi, eight), f_hi);
+
+ p6 = filter16_mask(&flat2, &p6, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s - 7 * p), p6);
+
+ f_lo = filter_add2_sub2(&f_lo, &q1_lo, &p5_lo, &p6_lo, &p7_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q1_hi, &p5_hi, &p6_hi, &p7_hi);
+ p5 = filter16_mask(&flat2, &p5, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
+
+ f_lo = filter_add2_sub2(&f_lo, &q2_lo, &p4_lo, &p5_lo, &p7_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q2_hi, &p4_hi, &p5_hi, &p7_hi);
+ p4 = filter16_mask(&flat2, &p4, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
+
+ f_lo = filter_add2_sub2(&f_lo, &q3_lo, &p3_lo, &p4_lo, &p7_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q3_hi, &p3_hi, &p4_hi, &p7_hi);
+ p3 = filter16_mask(&flat2, &p3, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
+
+ f_lo = filter_add2_sub2(&f_lo, &q4_lo, &p2_lo, &p3_lo, &p7_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q4_hi, &p2_hi, &p3_hi, &p7_hi);
+ op2 = filter16_mask(&flat2, &op2, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
+
+ f_lo = filter_add2_sub2(&f_lo, &q5_lo, &p1_lo, &p2_lo, &p7_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q5_hi, &p1_hi, &p2_hi, &p7_hi);
+ op1 = filter16_mask(&flat2, &op1, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+
+ f_lo = filter_add2_sub2(&f_lo, &q6_lo, &p0_lo, &p1_lo, &p7_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q6_hi, &p0_hi, &p1_hi, &p7_hi);
+ op0 = filter16_mask(&flat2, &op0, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+
+ f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q0_lo, &p0_lo, &p7_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q0_hi, &p0_hi, &p7_hi);
+ oq0 = filter16_mask(&flat2, &oq0, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+
+ f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q1_lo, &p6_lo, &q0_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q1_hi, &p6_hi, &q0_hi);
+ oq1 = filter16_mask(&flat2, &oq1, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+
+ f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q2_lo, &p5_lo, &q1_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q2_hi, &p5_hi, &q1_hi);
+ oq2 = filter16_mask(&flat2, &oq2, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
+
+ f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q3_lo, &p4_lo, &q2_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q3_hi, &p4_hi, &q2_hi);
+ q3 = filter16_mask(&flat2, &q3, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
+
+ f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q4_lo, &p3_lo, &q3_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q4_hi, &p3_hi, &q3_hi);
+ q4 = filter16_mask(&flat2, &q4, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
+
+ f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q5_lo, &p2_lo, &q4_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q5_hi, &p2_hi, &q4_hi);
+ q5 = filter16_mask(&flat2, &q5, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
+
+ f_lo = filter_add2_sub2(&f_lo, &q7_lo, &q6_lo, &p1_lo, &q5_lo);
+ f_hi = filter_add2_sub2(&f_hi, &q7_hi, &q6_hi, &p1_hi, &q5_hi);
+ q6 = filter16_mask(&flat2, &q6, &f_lo, &f_hi);
+ _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
+ }
+ // wide flat
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ }
+}
+
+// TODO(yunqingwang): remove count and call these 2 functions(8 or 16) directly.
+void vp9_lpf_horizontal_16_sse2(unsigned char *s, int p,
+ const unsigned char *_blimit,
+ const unsigned char *_limit,
+ const unsigned char *_thresh, int count) {+ if (count == 1)
+ mb_lpf_horizontal_edge_w_sse2_8(s, p, _blimit, _limit, _thresh);
+ else
+ mb_lpf_horizontal_edge_w_sse2_16(s, p, _blimit, _limit, _thresh);
+}
+
+void vp9_lpf_horizontal_8_sse2(unsigned char *s, int p,
+ const unsigned char *_blimit,
+ const unsigned char *_limit,
+ const unsigned char *_thresh, int count) {+ DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+ const __m128i limit = _mm_load_si128((const __m128i *)_limit);
+ const __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+ __m128i mask, hev, flat;
+ __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+ __m128i q3p3, q2p2, q1p1, q0p0, p1q1, p0q0;
+
+ (void)count;
+
+ q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)),
+ _mm_loadl_epi64((__m128i *)(s + 3 * p)));
+ q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
+ _mm_loadl_epi64((__m128i *)(s + 2 * p)));
+ q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+ _mm_loadl_epi64((__m128i *)(s + 1 * p)));
+ q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+ _mm_loadl_epi64((__m128i *)(s - 0 * p)));
+ p1q1 = _mm_shuffle_epi32(q1p1, 78);
+ p0q0 = _mm_shuffle_epi32(q0p0, 78);
+
+ {+ // filter_mask and hev_mask
+ const __m128i one = _mm_set1_epi8(1);
+ const __m128i fe = _mm_set1_epi8(0xfe);
+ const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+ __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+ abs_p1p0 = abs_diff(q1p1, q0p0);
+ abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+
+ abs_p0q0 = abs_diff(q0p0, p0q0);
+ abs_p1q1 = abs_diff(q1p1, p1q1);
+ flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu8(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+ abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+ mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ mask = _mm_max_epu8(abs_p1p0, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+
+ work = _mm_max_epu8(abs_diff(q2p2, q1p1),
+ abs_diff(q3p3, q2p2));
+ mask = _mm_max_epu8(work, mask);
+ mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+ mask = _mm_subs_epu8(mask, limit);
+ mask = _mm_cmpeq_epi8(mask, zero);
+
+ // flat_mask4
+
+ flat = _mm_max_epu8(abs_diff(q2p2, q0p0),
+ abs_diff(q3p3, q0p0));
+ flat = _mm_max_epu8(abs_p1p0, flat);
+ flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+ flat = _mm_subs_epu8(flat, one);
+ flat = _mm_cmpeq_epi8(flat, zero);
+ flat = _mm_and_si128(flat, mask);
+ }
+
+ {+ const __m128i four = _mm_set1_epi16(4);
+ unsigned char *src = s;
+ {+ __m128i workp_a, workp_b, workp_shft;
+ p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+ p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+ p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+ p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+ q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+ q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+ q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+ q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
+
+ workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+ workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_op2[0],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_op1[0],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_op0[0],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_oq0[0],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_oq1[0],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_oq2[0],
+ _mm_packus_epi16(workp_shft, workp_shft));
+ }
+ }
+ // lp filter
+ {+ const __m128i t4 = _mm_set1_epi8(4);
+ const __m128i t3 = _mm_set1_epi8(3);
+ const __m128i t80 = _mm_set1_epi8(0x80);
+ const __m128i t1 = _mm_set1_epi8(0x1);
+ const __m128i ps1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+ t80);
+ const __m128i ps0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+ t80);
+ const __m128i qs0 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 0 * p)),
+ t80);
+ const __m128i qs1 = _mm_xor_si128(_mm_loadl_epi64((__m128i *)(s + 1 * p)),
+ t80);
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+
+ filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+ work_a = _mm_subs_epi8(qs0, ps0);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ // (vp9_filter + 3 * (qs0 - ps0)) & mask
+ filt = _mm_and_si128(filt, mask);
+
+ filter1 = _mm_adds_epi8(filt, t4);
+ filter2 = _mm_adds_epi8(filt, t3);
+
+ // Filter1 >> 3
+ filter1 = _mm_unpacklo_epi8(zero, filter1);
+ filter1 = _mm_srai_epi16(filter1, 11);
+ filter1 = _mm_packs_epi16(filter1, filter1);
+
+ // Filter2 >> 3
+ filter2 = _mm_unpacklo_epi8(zero, filter2);
+ filter2 = _mm_srai_epi16(filter2, 11);
+ filter2 = _mm_packs_epi16(filter2, zero);
+
+ // filt >> 1
+ filt = _mm_adds_epi8(filter1, t1);
+ filt = _mm_unpacklo_epi8(zero, filt);
+ filt = _mm_srai_epi16(filt, 9);
+ filt = _mm_packs_epi16(filt, zero);
+
+ filt = _mm_andnot_si128(hev, filt);
+
+ work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+ q0 = _mm_loadl_epi64((__m128i *)flat_oq0);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q0 = _mm_and_si128(flat, q0);
+ q0 = _mm_or_si128(work_a, q0);
+
+ work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+ q1 = _mm_loadl_epi64((__m128i *)flat_oq1);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q1 = _mm_and_si128(flat, q1);
+ q1 = _mm_or_si128(work_a, q1);
+
+ work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+ q2 = _mm_loadl_epi64((__m128i *)flat_oq2);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q2 = _mm_and_si128(flat, q2);
+ q2 = _mm_or_si128(work_a, q2);
+
+ work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+ p0 = _mm_loadl_epi64((__m128i *)flat_op0);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p0 = _mm_and_si128(flat, p0);
+ p0 = _mm_or_si128(work_a, p0);
+
+ work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+ p1 = _mm_loadl_epi64((__m128i *)flat_op1);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p1 = _mm_and_si128(flat, p1);
+ p1 = _mm_or_si128(work_a, p1);
+
+ work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+ p2 = _mm_loadl_epi64((__m128i *)flat_op2);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p2 = _mm_and_si128(flat, p2);
+ p2 = _mm_or_si128(work_a, p2);
+
+ _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
+ _mm_storel_epi64((__m128i *)(s - 2 * p), p1);
+ _mm_storel_epi64((__m128i *)(s - 1 * p), p0);
+ _mm_storel_epi64((__m128i *)(s + 0 * p), q0);
+ _mm_storel_epi64((__m128i *)(s + 1 * p), q1);
+ _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
+ }
+}
+
+void vp9_lpf_horizontal_8_dual_sse2(uint8_t *s, int p,
+ const uint8_t *_blimit0,
+ const uint8_t *_limit0,
+ const uint8_t *_thresh0,
+ const uint8_t *_blimit1,
+ const uint8_t *_limit1,
+ const uint8_t *_thresh1) {+ DECLARE_ALIGNED(16, unsigned char, flat_op2[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_op1[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_op0[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_oq2[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_oq1[16]);
+ DECLARE_ALIGNED(16, unsigned char, flat_oq0[16]);
+ const __m128i zero = _mm_set1_epi16(0);
+ const __m128i blimit =
+ _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
+ _mm_load_si128((const __m128i *)_blimit1));
+ const __m128i limit =
+ _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
+ _mm_load_si128((const __m128i *)_limit1));
+ const __m128i thresh =
+ _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
+ _mm_load_si128((const __m128i *)_thresh1));
+
+ __m128i mask, hev, flat;
+ __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+
+ p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+ p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+ p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+ p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+ q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+ q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+ q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+ q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+ {+ const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+ _mm_subs_epu8(p0, p1));
+ const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+ _mm_subs_epu8(q0, q1));
+ const __m128i one = _mm_set1_epi8(1);
+ const __m128i fe = _mm_set1_epi8(0xfe);
+ const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+ __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+ _mm_subs_epu8(q0, p0));
+ __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+ _mm_subs_epu8(q1, p1));
+ __m128i work;
+
+ // filter_mask and hev_mask
+ flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu8(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+ abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+ mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ mask = _mm_max_epu8(flat, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+ work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
+ _mm_subs_epu8(p1, p2)),
+ _mm_or_si128(_mm_subs_epu8(p3, p2),
+ _mm_subs_epu8(p2, p3)));
+ mask = _mm_max_epu8(work, mask);
+ work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
+ _mm_subs_epu8(q1, q2)),
+ _mm_or_si128(_mm_subs_epu8(q3, q2),
+ _mm_subs_epu8(q2, q3)));
+ mask = _mm_max_epu8(work, mask);
+ mask = _mm_subs_epu8(mask, limit);
+ mask = _mm_cmpeq_epi8(mask, zero);
+
+ // flat_mask4
+ work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
+ _mm_subs_epu8(p0, p2)),
+ _mm_or_si128(_mm_subs_epu8(q2, q0),
+ _mm_subs_epu8(q0, q2)));
+ flat = _mm_max_epu8(work, flat);
+ work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
+ _mm_subs_epu8(p0, p3)),
+ _mm_or_si128(_mm_subs_epu8(q3, q0),
+ _mm_subs_epu8(q0, q3)));
+ flat = _mm_max_epu8(work, flat);
+ flat = _mm_subs_epu8(flat, one);
+ flat = _mm_cmpeq_epi8(flat, zero);
+ flat = _mm_and_si128(flat, mask);
+ }
+ {+ const __m128i four = _mm_set1_epi16(4);
+ unsigned char *src = s;
+ int i = 0;
+
+ do {+ __m128i workp_a, workp_b, workp_shft;
+ p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 4 * p)), zero);
+ p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 3 * p)), zero);
+ p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 2 * p)), zero);
+ p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 1 * p)), zero);
+ q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src - 0 * p)), zero);
+ q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 1 * p)), zero);
+ q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 2 * p)), zero);
+ q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(src + 3 * p)), zero);
+
+ workp_a = _mm_add_epi16(_mm_add_epi16(p3, p3), _mm_add_epi16(p2, p1));
+ workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0);
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, p2), p3);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_op2[i * 8],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_b = _mm_add_epi16(_mm_add_epi16(q0, q1), p1);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_op1[i * 8],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q2);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1), p0);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_op0[i * 8],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0), q0);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_oq0[i * 8],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0), q1);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_oq1[i * 8],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1), q3);
+ workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1), q2);
+ workp_shft = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+ _mm_storel_epi64((__m128i *)&flat_oq2[i * 8],
+ _mm_packus_epi16(workp_shft, workp_shft));
+
+ src += 8;
+ } while (++i < 2);
+ }
+ // lp filter
+ {+ const __m128i t4 = _mm_set1_epi8(4);
+ const __m128i t3 = _mm_set1_epi8(3);
+ const __m128i t80 = _mm_set1_epi8(0x80);
+ const __m128i te0 = _mm_set1_epi8(0xe0);
+ const __m128i t1f = _mm_set1_epi8(0x1f);
+ const __m128i t1 = _mm_set1_epi8(0x1);
+ const __m128i t7f = _mm_set1_epi8(0x7f);
+
+ const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
+ t80);
+ const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
+ t80);
+ const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
+ t80);
+ const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
+ t80);
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+
+ filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+ work_a = _mm_subs_epi8(qs0, ps0);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ // (vp9_filter + 3 * (qs0 - ps0)) & mask
+ filt = _mm_and_si128(filt, mask);
+
+ filter1 = _mm_adds_epi8(filt, t4);
+ filter2 = _mm_adds_epi8(filt, t3);
+
+ // Filter1 >> 3
+ work_a = _mm_cmpgt_epi8(zero, filter1);
+ filter1 = _mm_srli_epi16(filter1, 3);
+ work_a = _mm_and_si128(work_a, te0);
+ filter1 = _mm_and_si128(filter1, t1f);
+ filter1 = _mm_or_si128(filter1, work_a);
+
+ // Filter2 >> 3
+ work_a = _mm_cmpgt_epi8(zero, filter2);
+ filter2 = _mm_srli_epi16(filter2, 3);
+ work_a = _mm_and_si128(work_a, te0);
+ filter2 = _mm_and_si128(filter2, t1f);
+ filter2 = _mm_or_si128(filter2, work_a);
+
+ // filt >> 1
+ filt = _mm_adds_epi8(filter1, t1);
+ work_a = _mm_cmpgt_epi8(zero, filt);
+ filt = _mm_srli_epi16(filt, 1);
+ work_a = _mm_and_si128(work_a, t80);
+ filt = _mm_and_si128(filt, t7f);
+ filt = _mm_or_si128(filt, work_a);
+
+ filt = _mm_andnot_si128(hev, filt);
+
+ work_a = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+ q0 = _mm_load_si128((__m128i *)flat_oq0);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q0 = _mm_and_si128(flat, q0);
+ q0 = _mm_or_si128(work_a, q0);
+
+ work_a = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+ q1 = _mm_load_si128((__m128i *)flat_oq1);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q1 = _mm_and_si128(flat, q1);
+ q1 = _mm_or_si128(work_a, q1);
+
+ work_a = _mm_loadu_si128((__m128i *)(s + 2 * p));
+ q2 = _mm_load_si128((__m128i *)flat_oq2);
+ work_a = _mm_andnot_si128(flat, work_a);
+ q2 = _mm_and_si128(flat, q2);
+ q2 = _mm_or_si128(work_a, q2);
+
+ work_a = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+ p0 = _mm_load_si128((__m128i *)flat_op0);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p0 = _mm_and_si128(flat, p0);
+ p0 = _mm_or_si128(work_a, p0);
+
+ work_a = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+ p1 = _mm_load_si128((__m128i *)flat_op1);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p1 = _mm_and_si128(flat, p1);
+ p1 = _mm_or_si128(work_a, p1);
+
+ work_a = _mm_loadu_si128((__m128i *)(s - 3 * p));
+ p2 = _mm_load_si128((__m128i *)flat_op2);
+ work_a = _mm_andnot_si128(flat, work_a);
+ p2 = _mm_and_si128(flat, p2);
+ p2 = _mm_or_si128(work_a, p2);
+
+ _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+ _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+ _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+ _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+ _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+ _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+ }
+}
+
+void vp9_lpf_horizontal_4_dual_sse2(unsigned char *s, int p,
+ const unsigned char *_blimit0,
+ const unsigned char *_limit0,
+ const unsigned char *_thresh0,
+ const unsigned char *_blimit1,
+ const unsigned char *_limit1,
+ const unsigned char *_thresh1) {+ const __m128i blimit =
+ _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_blimit0),
+ _mm_load_si128((const __m128i *)_blimit1));
+ const __m128i limit =
+ _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_limit0),
+ _mm_load_si128((const __m128i *)_limit1));
+ const __m128i thresh =
+ _mm_unpacklo_epi64(_mm_load_si128((const __m128i *)_thresh0),
+ _mm_load_si128((const __m128i *)_thresh1));
+ const __m128i zero = _mm_set1_epi16(0);
+ __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+ __m128i mask, hev, flat;
+
+ p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+ p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+ p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+ p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+ q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+ q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+ q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+ q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+
+ // filter_mask and hev_mask
+ {+ const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
+ _mm_subs_epu8(p0, p1));
+ const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
+ _mm_subs_epu8(q0, q1));
+ const __m128i fe = _mm_set1_epi8(0xfe);
+ const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+ __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
+ _mm_subs_epu8(q0, p0));
+ __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
+ _mm_subs_epu8(q1, p1));
+ __m128i work;
+
+ flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+ hev = _mm_subs_epu8(flat, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+ abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
+ abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+ mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
+ mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+ // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
+ mask = _mm_max_epu8(flat, mask);
+ // mask |= (abs(p1 - p0) > limit) * -1;
+ // mask |= (abs(q1 - q0) > limit) * -1;
+ work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
+ _mm_subs_epu8(p1, p2)),
+ _mm_or_si128(_mm_subs_epu8(p3, p2),
+ _mm_subs_epu8(p2, p3)));
+ mask = _mm_max_epu8(work, mask);
+ work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
+ _mm_subs_epu8(q1, q2)),
+ _mm_or_si128(_mm_subs_epu8(q3, q2),
+ _mm_subs_epu8(q2, q3)));
+ mask = _mm_max_epu8(work, mask);
+ mask = _mm_subs_epu8(mask, limit);
+ mask = _mm_cmpeq_epi8(mask, zero);
+ }
+
+ // filter4
+ {+ const __m128i t4 = _mm_set1_epi8(4);
+ const __m128i t3 = _mm_set1_epi8(3);
+ const __m128i t80 = _mm_set1_epi8(0x80);
+ const __m128i te0 = _mm_set1_epi8(0xe0);
+ const __m128i t1f = _mm_set1_epi8(0x1f);
+ const __m128i t1 = _mm_set1_epi8(0x1);
+ const __m128i t7f = _mm_set1_epi8(0x7f);
+
+ const __m128i ps1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 2 * p)),
+ t80);
+ const __m128i ps0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s - 1 * p)),
+ t80);
+ const __m128i qs0 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 0 * p)),
+ t80);
+ const __m128i qs1 = _mm_xor_si128(_mm_loadu_si128((__m128i *)(s + 1 * p)),
+ t80);
+ __m128i filt;
+ __m128i work_a;
+ __m128i filter1, filter2;
+
+ filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+ work_a = _mm_subs_epi8(qs0, ps0);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ filt = _mm_adds_epi8(filt, work_a);
+ // (vp9_filter + 3 * (qs0 - ps0)) & mask
+ filt = _mm_and_si128(filt, mask);
+
+ filter1 = _mm_adds_epi8(filt, t4);
+ filter2 = _mm_adds_epi8(filt, t3);
+
+ // Filter1 >> 3
+ work_a = _mm_cmpgt_epi8(zero, filter1);
+ filter1 = _mm_srli_epi16(filter1, 3);
+ work_a = _mm_and_si128(work_a, te0);
+ filter1 = _mm_and_si128(filter1, t1f);
+ filter1 = _mm_or_si128(filter1, work_a);
+
+ // Filter2 >> 3
+ work_a = _mm_cmpgt_epi8(zero, filter2);
+ filter2 = _mm_srli_epi16(filter2, 3);
+ work_a = _mm_and_si128(work_a, te0);
+ filter2 = _mm_and_si128(filter2, t1f);
+ filter2 = _mm_or_si128(filter2, work_a);
+
+ // filt >> 1
+ filt = _mm_adds_epi8(filter1, t1);
+ work_a = _mm_cmpgt_epi8(zero, filt);
+ filt = _mm_srli_epi16(filt, 1);
+ work_a = _mm_and_si128(work_a, t80);
+ filt = _mm_and_si128(filt, t7f);
+ filt = _mm_or_si128(filt, work_a);
+
+ filt = _mm_andnot_si128(hev, filt);
+
+ q0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+ q1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+ p0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+ p1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+
+ _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+ _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+ _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+ _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+ }
+}
+
+static INLINE void transpose8x16(unsigned char *in0, unsigned char *in1,
+ int in_p, unsigned char *out, int out_p) {+ __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+ __m128i x8, x9, x10, x11, x12, x13, x14, x15;
+
+ // 2-way interleave w/hoisting of unpacks
+ x0 = _mm_loadl_epi64((__m128i *)in0); // 1
+ x1 = _mm_loadl_epi64((__m128i *)(in0 + in_p)); // 3
+ x0 = _mm_unpacklo_epi8(x0, x1); // 1
+
+ x2 = _mm_loadl_epi64((__m128i *)(in0 + 2 * in_p)); // 5
+ x3 = _mm_loadl_epi64((__m128i *)(in0 + 3*in_p)); // 7
+ x1 = _mm_unpacklo_epi8(x2, x3); // 2
+
+ x4 = _mm_loadl_epi64((__m128i *)(in0 + 4*in_p)); // 9
+ x5 = _mm_loadl_epi64((__m128i *)(in0 + 5*in_p)); // 11
+ x2 = _mm_unpacklo_epi8(x4, x5); // 3
+
+ x6 = _mm_loadl_epi64((__m128i *)(in0 + 6*in_p)); // 13
+ x7 = _mm_loadl_epi64((__m128i *)(in0 + 7*in_p)); // 15
+ x3 = _mm_unpacklo_epi8(x6, x7); // 4
+ x4 = _mm_unpacklo_epi16(x0, x1); // 9
+
+ x8 = _mm_loadl_epi64((__m128i *)in1); // 2
+ x9 = _mm_loadl_epi64((__m128i *)(in1 + in_p)); // 4
+ x8 = _mm_unpacklo_epi8(x8, x9); // 5
+ x5 = _mm_unpacklo_epi16(x2, x3); // 10
+
+ x10 = _mm_loadl_epi64((__m128i *)(in1 + 2 * in_p)); // 6
+ x11 = _mm_loadl_epi64((__m128i *)(in1 + 3*in_p)); // 8
+ x9 = _mm_unpacklo_epi8(x10, x11); // 6
+
+ x12 = _mm_loadl_epi64((__m128i *)(in1 + 4*in_p)); // 10
+ x13 = _mm_loadl_epi64((__m128i *)(in1 + 5*in_p)); // 12
+ x10 = _mm_unpacklo_epi8(x12, x13); // 7
+ x12 = _mm_unpacklo_epi16(x8, x9); // 11
+
+ x14 = _mm_loadl_epi64((__m128i *)(in1 + 6*in_p)); // 14
+ x15 = _mm_loadl_epi64((__m128i *)(in1 + 7*in_p)); // 16
+ x11 = _mm_unpacklo_epi8(x14, x15); // 8
+ x13 = _mm_unpacklo_epi16(x10, x11); // 12
+
+ x6 = _mm_unpacklo_epi32(x4, x5); // 13
+ x7 = _mm_unpackhi_epi32(x4, x5); // 14
+ x14 = _mm_unpacklo_epi32(x12, x13); // 15
+ x15 = _mm_unpackhi_epi32(x12, x13); // 16
+
+ // Store first 4-line result
+ _mm_storeu_si128((__m128i *)out, _mm_unpacklo_epi64(x6, x14));
+ _mm_storeu_si128((__m128i *)(out + out_p), _mm_unpackhi_epi64(x6, x14));
+ _mm_storeu_si128((__m128i *)(out + 2 * out_p), _mm_unpacklo_epi64(x7, x15));
+ _mm_storeu_si128((__m128i *)(out + 3 * out_p), _mm_unpackhi_epi64(x7, x15));
+
+ x4 = _mm_unpackhi_epi16(x0, x1);
+ x5 = _mm_unpackhi_epi16(x2, x3);
+ x12 = _mm_unpackhi_epi16(x8, x9);
+ x13 = _mm_unpackhi_epi16(x10, x11);
+
+ x6 = _mm_unpacklo_epi32(x4, x5);
+ x7 = _mm_unpackhi_epi32(x4, x5);
+ x14 = _mm_unpacklo_epi32(x12, x13);
+ x15 = _mm_unpackhi_epi32(x12, x13);
+
+ // Store second 4-line result
+ _mm_storeu_si128((__m128i *)(out + 4 * out_p), _mm_unpacklo_epi64(x6, x14));
+ _mm_storeu_si128((__m128i *)(out + 5 * out_p), _mm_unpackhi_epi64(x6, x14));
+ _mm_storeu_si128((__m128i *)(out + 6 * out_p), _mm_unpacklo_epi64(x7, x15));
+ _mm_storeu_si128((__m128i *)(out + 7 * out_p), _mm_unpackhi_epi64(x7, x15));
+}
+
+static INLINE void transpose(unsigned char *src[], int in_p,
+ unsigned char *dst[], int out_p,
+ int num_8x8_to_transpose) {+ int idx8x8 = 0;
+ __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+ do {+ unsigned char *in = src[idx8x8];
+ unsigned char *out = dst[idx8x8];
+
+ x0 = _mm_loadl_epi64((__m128i *)(in + 0*in_p)); // 00 01 02 03 04 05 06 07
+ x1 = _mm_loadl_epi64((__m128i *)(in + 1*in_p)); // 10 11 12 13 14 15 16 17
+ // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+ x0 = _mm_unpacklo_epi8(x0, x1);
+
+ x2 = _mm_loadl_epi64((__m128i *)(in + 2*in_p)); // 20 21 22 23 24 25 26 27
+ x3 = _mm_loadl_epi64((__m128i *)(in + 3*in_p)); // 30 31 32 33 34 35 36 37
+ // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+ x1 = _mm_unpacklo_epi8(x2, x3);
+
+ x4 = _mm_loadl_epi64((__m128i *)(in + 4*in_p)); // 40 41 42 43 44 45 46 47
+ x5 = _mm_loadl_epi64((__m128i *)(in + 5*in_p)); // 50 51 52 53 54 55 56 57
+ // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+ x2 = _mm_unpacklo_epi8(x4, x5);
+
+ x6 = _mm_loadl_epi64((__m128i *)(in + 6*in_p)); // 60 61 62 63 64 65 66 67
+ x7 = _mm_loadl_epi64((__m128i *)(in + 7*in_p)); // 70 71 72 73 74 75 76 77
+ // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+ x3 = _mm_unpacklo_epi8(x6, x7);
+
+ // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+ x4 = _mm_unpacklo_epi16(x0, x1);
+ // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+ x5 = _mm_unpacklo_epi16(x2, x3);
+ // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+ x6 = _mm_unpacklo_epi32(x4, x5);
+ _mm_storel_pd((double *)(out + 0*out_p),
+ _mm_castsi128_pd(x6)); // 00 10 20 30 40 50 60 70
+ _mm_storeh_pd((double *)(out + 1*out_p),
+ _mm_castsi128_pd(x6)); // 01 11 21 31 41 51 61 71
+ // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+ x7 = _mm_unpackhi_epi32(x4, x5);
+ _mm_storel_pd((double *)(out + 2*out_p),
+ _mm_castsi128_pd(x7)); // 02 12 22 32 42 52 62 72
+ _mm_storeh_pd((double *)(out + 3*out_p),
+ _mm_castsi128_pd(x7)); // 03 13 23 33 43 53 63 73
+
+ // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+ x4 = _mm_unpackhi_epi16(x0, x1);
+ // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77
+ x5 = _mm_unpackhi_epi16(x2, x3);
+ // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75
+ x6 = _mm_unpacklo_epi32(x4, x5);
+ _mm_storel_pd((double *)(out + 4*out_p),
+ _mm_castsi128_pd(x6)); // 04 14 24 34 44 54 64 74
+ _mm_storeh_pd((double *)(out + 5*out_p),
+ _mm_castsi128_pd(x6)); // 05 15 25 35 45 55 65 75
+ // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77
+ x7 = _mm_unpackhi_epi32(x4, x5);
+
+ _mm_storel_pd((double *)(out + 6*out_p),
+ _mm_castsi128_pd(x7)); // 06 16 26 36 46 56 66 76
+ _mm_storeh_pd((double *)(out + 7*out_p),
+ _mm_castsi128_pd(x7)); // 07 17 27 37 47 57 67 77
+ } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+void vp9_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1) {+ DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+ unsigned char *src[2];
+ unsigned char *dst[2];
+
+ // Transpose 8x16
+ transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+ // Loop filtering
+ vp9_lpf_horizontal_4_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
+ blimit1, limit1, thresh1);
+ src[0] = t_dst;
+ src[1] = t_dst + 8;
+ dst[0] = s - 4;
+ dst[1] = s - 4 + p * 8;
+
+ // Transpose back
+ transpose(src, 16, dst, p, 2);
+}
+
+void vp9_lpf_vertical_8_sse2(unsigned char *s, int p,
+ const unsigned char *blimit,
+ const unsigned char *limit,
+ const unsigned char *thresh, int count) {+ DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 8]);
+ unsigned char *src[1];
+ unsigned char *dst[1];
+ (void)count;
+
+ // Transpose 8x8
+ src[0] = s - 4;
+ dst[0] = t_dst;
+
+ transpose(src, p, dst, 8, 1);
+
+ // Loop filtering
+ vp9_lpf_horizontal_8_sse2(t_dst + 4 * 8, 8, blimit, limit, thresh, 1);
+
+ src[0] = t_dst;
+ dst[0] = s - 4;
+
+ // Transpose back
+ transpose(src, 8, dst, p, 1);
+}
+
+void vp9_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *blimit0,
+ const uint8_t *limit0,
+ const uint8_t *thresh0,
+ const uint8_t *blimit1,
+ const uint8_t *limit1,
+ const uint8_t *thresh1) {+ DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+ unsigned char *src[2];
+ unsigned char *dst[2];
+
+ // Transpose 8x16
+ transpose8x16(s - 4, s - 4 + p * 8, p, t_dst, 16);
+
+ // Loop filtering
+ vp9_lpf_horizontal_8_dual_sse2(t_dst + 4 * 16, 16, blimit0, limit0, thresh0,
+ blimit1, limit1, thresh1);
+ src[0] = t_dst;
+ src[1] = t_dst + 8;
+
+ dst[0] = s - 4;
+ dst[1] = s - 4 + p * 8;
+
+ // Transpose back
+ transpose(src, 16, dst, p, 2);
+}
+
+void vp9_lpf_vertical_16_sse2(unsigned char *s, int p,
+ const unsigned char *blimit,
+ const unsigned char *limit,
+ const unsigned char *thresh) {+ DECLARE_ALIGNED(8, unsigned char, t_dst[8 * 16]);
+ unsigned char *src[2];
+ unsigned char *dst[2];
+
+ src[0] = s - 8;
+ src[1] = s;
+ dst[0] = t_dst;
+ dst[1] = t_dst + 8 * 8;
+
+ // Transpose 16x8
+ transpose(src, p, dst, 8, 2);
+
+ // Loop filtering
+ mb_lpf_horizontal_edge_w_sse2_8(t_dst + 8 * 8, 8, blimit, limit, thresh);
+
+ src[0] = t_dst;
+ src[1] = t_dst + 8 * 8;
+ dst[0] = s - 8;
+ dst[1] = s;
+
+ // Transpose back
+ transpose(src, 8, dst, p, 2);
+}
+
+void vp9_lpf_vertical_16_dual_sse2(unsigned char *s, int p,
+ const uint8_t *blimit, const uint8_t *limit,
+ const uint8_t *thresh) {+ DECLARE_ALIGNED(16, unsigned char, t_dst[256]);
+
+ // Transpose 16x16
+ transpose8x16(s - 8, s - 8 + 8 * p, p, t_dst, 16);
+ transpose8x16(s, s + 8 * p, p, t_dst + 8 * 16, 16);
+
+ // Loop filtering
+ mb_lpf_horizontal_edge_w_sse2_16(t_dst + 8 * 16, 16, blimit, limit,
+ thresh);
+
+ // Transpose back
+ transpose8x16(t_dst, t_dst + 8 * 16, 16, s - 8, p);
+ transpose8x16(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * p, p);
+}