ref: 36363304903286369a63b0498d506a13fe335606
parent: 0fe4371cc0871b713acf09fea671347ed2d4f98c
author: Linfeng Zhang <[email protected]>
date: Tue Jan 30 11:01:41 EST 2018
Add vp9_highbd_iht4x4_16_add_neon() BUG=webm:1403 Change-Id: Id9833e985fb70958cf4bde38f8e6303ed83c12f9
--- a/test/dct_test.cc
+++ b/test/dct_test.cc
@@ -753,26 +753,66 @@
#if !CONFIG_EMULATE_HARDWARE
#if HAVE_NEON
-INSTANTIATE_TEST_CASE_P(
- NEON, TransHT,
- ::testing::Values(
- make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 0,
- VPX_BITS_8, 1),
- make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 1,
- VPX_BITS_8, 1),
- make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 2,
- VPX_BITS_8, 1),
- make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 3,
- VPX_BITS_8, 1),
- make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 0,
- VPX_BITS_8, 1),
- make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 1,
- VPX_BITS_8, 1),
- make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 2,
- VPX_BITS_8, 1),
- make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 3,
- VPX_BITS_8, 1)));
+const DctParam neon_ht_tests[] = {+#if CONFIG_VP9_HIGHBITDEPTH
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 0,
+ VPX_BITS_8, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 1,
+ VPX_BITS_8, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 2,
+ VPX_BITS_8, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 3,
+ VPX_BITS_8, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 0,
+ VPX_BITS_10, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 1,
+ VPX_BITS_10, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 2,
+ VPX_BITS_10, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 3,
+ VPX_BITS_10, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 0,
+ VPX_BITS_12, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 1,
+ VPX_BITS_12, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 2,
+ VPX_BITS_12, 2),
+ make_tuple(&vp9_highbd_fht4x4_c,
+ &highbd_iht_wrapper<vp9_highbd_iht4x4_16_add_neon>, 4, 3,
+ VPX_BITS_12, 2),
+#endif // CONFIG_VP9_HIGHBITDEPTH
+ make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 0,
+ VPX_BITS_8, 1),
+ make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 1,
+ VPX_BITS_8, 1),
+ make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 2,
+ VPX_BITS_8, 1),
+ make_tuple(&vp9_fht8x8_c, &iht_wrapper<vp9_iht8x8_64_add_neon>, 8, 3,
+ VPX_BITS_8, 1),
+
+ make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 0,
+ VPX_BITS_8, 1),
+ make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 1,
+ VPX_BITS_8, 1),
+ make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 2,
+ VPX_BITS_8, 1),
+ make_tuple(&vp9_fht4x4_c, &iht_wrapper<vp9_iht4x4_16_add_neon>, 4, 3,
+ VPX_BITS_8, 1)
+};
+
+INSTANTIATE_TEST_CASE_P(NEON, TransHT, ::testing::ValuesIn(neon_ht_tests));
#endif // HAVE_NEON
#if HAVE_SSE2
--- /dev/null
+++ b/vp9/common/arm/neon/vp9_highbd_iht4x4_add_neon.c
@@ -1,0 +1,162 @@
+/*
+ * Copyright (c) 2018 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 <assert.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+#include "vp9/common/arm/neon/vp9_iht_neon.h"
+#include "vpx_dsp/arm/highbd_idct_neon.h"
+#include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/arm/mem_neon.h"
+#include "vpx_dsp/txfm_common.h"
+
+static INLINE void highbd_iadst4(int32x4_t *const io) {+ const int32_t sinpis[4] = { sinpi_1_9, sinpi_2_9, sinpi_3_9, sinpi_4_9 };+ const int32x4_t sinpi = vld1q_s32(sinpis);
+ int32x4_t s[8];
+
+ s[0] = vmulq_lane_s32(io[0], vget_low_s32(sinpi), 0);
+ s[1] = vmulq_lane_s32(io[0], vget_low_s32(sinpi), 1);
+ s[2] = vmulq_lane_s32(io[1], vget_high_s32(sinpi), 0);
+ s[3] = vmulq_lane_s32(io[2], vget_high_s32(sinpi), 1);
+ s[4] = vmulq_lane_s32(io[2], vget_low_s32(sinpi), 0);
+ s[5] = vmulq_lane_s32(io[3], vget_low_s32(sinpi), 1);
+ s[6] = vmulq_lane_s32(io[3], vget_high_s32(sinpi), 1);
+ s[7] = vsubq_s32(io[0], io[2]);
+ s[7] = vaddq_s32(s[7], io[3]);
+
+ s[0] = vaddq_s32(s[0], s[3]);
+ s[0] = vaddq_s32(s[0], s[5]);
+ s[1] = vsubq_s32(s[1], s[4]);
+ s[1] = vsubq_s32(s[1], s[6]);
+ s[3] = s[2];
+ s[2] = vmulq_lane_s32(s[7], vget_high_s32(sinpi), 0);
+
+ io[0] = vaddq_s32(s[0], s[3]);
+ io[1] = vaddq_s32(s[1], s[3]);
+ io[2] = s[2];
+ io[3] = vaddq_s32(s[0], s[1]);
+ io[3] = vsubq_s32(io[3], s[3]);
+ io[0] = vrshrq_n_s32(io[0], DCT_CONST_BITS);
+ io[1] = vrshrq_n_s32(io[1], DCT_CONST_BITS);
+ io[2] = vrshrq_n_s32(io[2], DCT_CONST_BITS);
+ io[3] = vrshrq_n_s32(io[3], DCT_CONST_BITS);
+}
+
+void vp9_highbd_iht4x4_16_add_neon(const tran_low_t *input, uint16_t *dest,
+ int stride, int tx_type, int bd) {+ const int16x8_t max = vdupq_n_s16((1 << bd) - 1);
+ int16x8_t a[2];
+ int32x4_t c[4];
+
+ c[0] = vld1q_s32(input);
+ c[1] = vld1q_s32(input + 4);
+ c[2] = vld1q_s32(input + 8);
+ c[3] = vld1q_s32(input + 12);
+
+ if (bd == 8) {+ a[0] = vcombine_s16(vmovn_s32(c[0]), vmovn_s32(c[1]));
+ a[1] = vcombine_s16(vmovn_s32(c[2]), vmovn_s32(c[3]));
+ transpose_s16_4x4q(&a[0], &a[1]);
+
+ switch (tx_type) {+ case DCT_DCT:
+ idct4x4_16_kernel_bd8(a);
+ a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
+ transpose_s16_4x4q(&a[0], &a[1]);
+ idct4x4_16_kernel_bd8(a);
+ a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
+ break;
+
+ case ADST_DCT:
+ idct4x4_16_kernel_bd8(a);
+ a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
+ transpose_s16_4x4q(&a[0], &a[1]);
+ iadst4(a);
+ break;
+
+ case DCT_ADST:
+ iadst4(a);
+ transpose_s16_4x4q(&a[0], &a[1]);
+ idct4x4_16_kernel_bd8(a);
+ a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
+ break;
+
+ default:
+ assert(tx_type == ADST_ADST);
+ iadst4(a);
+ transpose_s16_4x4q(&a[0], &a[1]);
+ iadst4(a);
+ break;
+ }
+ a[0] = vrshrq_n_s16(a[0], 4);
+ a[1] = vrshrq_n_s16(a[1], 4);
+ } else {+ switch (tx_type) {+ case DCT_DCT: {+ const int32x4_t cospis = vld1q_s32(kCospi32);
+
+ if (bd == 10) {+ idct4x4_16_kernel_bd10(cospis, c);
+ idct4x4_16_kernel_bd10(cospis, c);
+ } else {+ idct4x4_16_kernel_bd12(cospis, c);
+ idct4x4_16_kernel_bd12(cospis, c);
+ }
+ break;
+ }
+
+ case ADST_DCT: {+ const int32x4_t cospis = vld1q_s32(kCospi32);
+
+ if (bd == 10) {+ idct4x4_16_kernel_bd10(cospis, c);
+ } else {+ idct4x4_16_kernel_bd12(cospis, c);
+ }
+ transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]);
+ highbd_iadst4(c);
+ break;
+ }
+
+ case DCT_ADST: {+ const int32x4_t cospis = vld1q_s32(kCospi32);
+
+ if (bd == 10) {+ transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]);
+ highbd_iadst4(c);
+ idct4x4_16_kernel_bd10(cospis, c);
+ } else {+ transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]);
+ highbd_iadst4(c);
+ idct4x4_16_kernel_bd12(cospis, c);
+ }
+ break;
+ }
+
+ default: {+ assert(tx_type == ADST_ADST);
+ transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]);
+ highbd_iadst4(c);
+ transpose_s32_4x4(&c[0], &c[1], &c[2], &c[3]);
+ highbd_iadst4(c);
+ break;
+ }
+ }
+ a[0] = vcombine_s16(vqrshrn_n_s32(c[0], 4), vqrshrn_n_s32(c[1], 4));
+ a[1] = vcombine_s16(vqrshrn_n_s32(c[2], 4), vqrshrn_n_s32(c[3], 4));
+ }
+
+ highbd_idct4x4_1_add_kernel1(&dest, stride, a[0], max);
+ highbd_idct4x4_1_add_kernel1(&dest, stride, a[1], max);
+}
--- a/vp9/common/arm/neon/vp9_iht4x4_add_neon.c
+++ b/vp9/common/arm/neon/vp9_iht4x4_add_neon.c
@@ -14,50 +14,11 @@
#include "./vp9_rtcd.h"
#include "./vpx_config.h"
#include "vp9/common/vp9_common.h"
+#include "vp9/common/arm/neon/vp9_iht_neon.h"
#include "vpx_dsp/arm/idct_neon.h"
#include "vpx_dsp/arm/mem_neon.h"
#include "vpx_dsp/txfm_common.h"
-static INLINE void iadst4(int16x8_t *const io) {- const int32x4_t c3 = vdupq_n_s32(sinpi_3_9);
- int16x4_t c[5], x[4];
- int32x4_t s[8], output[4];
-
- c[1] = vdup_n_s16(sinpi_1_9);
- c[2] = vdup_n_s16(sinpi_2_9);
- c[3] = vdup_n_s16(sinpi_3_9);
- c[4] = vdup_n_s16(sinpi_4_9);
-
- x[0] = vget_low_s16(io[0]);
- x[1] = vget_low_s16(io[1]);
- x[2] = vget_high_s16(io[0]);
- x[3] = vget_high_s16(io[1]);
-
- s[0] = vmull_s16(c[1], x[0]);
- s[1] = vmull_s16(c[2], x[0]);
- s[2] = vmull_s16(c[3], x[1]);
- s[3] = vmull_s16(c[4], x[2]);
- s[4] = vmull_s16(c[1], x[2]);
- s[5] = vmull_s16(c[2], x[3]);
- s[6] = vmull_s16(c[4], x[3]);
- s[7] = vaddl_s16(x[0], x[3]);
- s[7] = vsubw_s16(s[7], x[2]);
-
- s[0] = vaddq_s32(s[0], s[3]);
- s[0] = vaddq_s32(s[0], s[5]);
- s[1] = vsubq_s32(s[1], s[4]);
- s[1] = vsubq_s32(s[1], s[6]);
- s[3] = s[2];
- s[2] = vmulq_s32(c3, s[7]);
-
- output[0] = vaddq_s32(s[0], s[3]);
- output[1] = vaddq_s32(s[1], s[3]);
- output[2] = s[2];
- output[3] = vaddq_s32(s[0], s[1]);
- output[3] = vsubq_s32(output[3], s[3]);
- dct_const_round_shift_low_8_dual(output, &io[0], &io[1]);
-}
-
void vp9_iht4x4_16_add_neon(const tran_low_t *input, uint8_t *dest, int stride,
int tx_type) {int16x8_t a[2];
@@ -72,7 +33,7 @@
transpose_s16_4x4q(&a[0], &a[1]);
switch (tx_type) {- case 0: // DCT_DCT
+ case DCT_DCT:
idct4x4_16_kernel_bd8(a);
a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
transpose_s16_4x4q(&a[0], &a[1]);
@@ -80,7 +41,7 @@
a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
break;
- case 1: // ADST_DCT
+ case ADST_DCT:
idct4x4_16_kernel_bd8(a);
a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
transpose_s16_4x4q(&a[0], &a[1]);
@@ -87,7 +48,7 @@
iadst4(a);
break;
- case 2: // DCT_ADST
+ case DCT_ADST:
iadst4(a);
transpose_s16_4x4q(&a[0], &a[1]);
idct4x4_16_kernel_bd8(a);
@@ -94,13 +55,12 @@
a[1] = vcombine_s16(vget_high_s16(a[1]), vget_low_s16(a[1]));
break;
- case 3: // ADST_ADST
+ default:
+ assert(tx_type == ADST_ADST);
iadst4(a);
transpose_s16_4x4q(&a[0], &a[1]);
iadst4(a);
break;
-
- default: assert(0); break;
}
a[0] = vrshrq_n_s16(a[0], 4);
--- /dev/null
+++ b/vp9/common/arm/neon/vp9_iht_neon.h
@@ -1,0 +1,63 @@
+/*
+ * Copyright (c) 2018 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_ARM_NEON_VP9_IHT_NEON_H_
+#define VP9_COMMON_ARM_NEON_VP9_IHT_NEON_H_
+
+#include <arm_neon.h>
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+#include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/arm/mem_neon.h"
+#include "vpx_dsp/txfm_common.h"
+
+static INLINE void iadst4(int16x8_t *const io) {+ const int32x4_t c3 = vdupq_n_s32(sinpi_3_9);
+ int16x4_t c[5], x[4];
+ int32x4_t s[8], output[4];
+
+ c[1] = vdup_n_s16(sinpi_1_9);
+ c[2] = vdup_n_s16(sinpi_2_9);
+ c[3] = vdup_n_s16(sinpi_3_9);
+ c[4] = vdup_n_s16(sinpi_4_9);
+
+ x[0] = vget_low_s16(io[0]);
+ x[1] = vget_low_s16(io[1]);
+ x[2] = vget_high_s16(io[0]);
+ x[3] = vget_high_s16(io[1]);
+
+ s[0] = vmull_s16(c[1], x[0]);
+ s[1] = vmull_s16(c[2], x[0]);
+ s[2] = vmull_s16(c[3], x[1]);
+ s[3] = vmull_s16(c[4], x[2]);
+ s[4] = vmull_s16(c[1], x[2]);
+ s[5] = vmull_s16(c[2], x[3]);
+ s[6] = vmull_s16(c[4], x[3]);
+ s[7] = vaddl_s16(x[0], x[3]);
+ s[7] = vsubw_s16(s[7], x[2]);
+
+ s[0] = vaddq_s32(s[0], s[3]);
+ s[0] = vaddq_s32(s[0], s[5]);
+ s[1] = vsubq_s32(s[1], s[4]);
+ s[1] = vsubq_s32(s[1], s[6]);
+ s[3] = s[2];
+ s[2] = vmulq_s32(c3, s[7]);
+
+ output[0] = vaddq_s32(s[0], s[3]);
+ output[1] = vaddq_s32(s[1], s[3]);
+ output[2] = s[2];
+ output[3] = vaddq_s32(s[0], s[1]);
+ output[3] = vsubq_s32(output[3], s[3]);
+ dct_const_round_shift_low_8_dual(output, &io[0], &io[1]);
+}
+
+#endif // VP9_COMMON_ARM_NEON_VP9_IHT_NEON_H_
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -103,7 +103,7 @@
add_proto qw/void vp9_highbd_iht16x16_256_add/, "const tran_low_t *input, uint16_t *output, int pitch, int tx_type, int bd";
if (vpx_config("CONFIG_EMULATE_HARDWARE") ne "yes") {- specialize qw/vp9_highbd_iht4x4_16_add sse4_1/;
+ specialize qw/vp9_highbd_iht4x4_16_add neon sse4_1/;
specialize qw/vp9_highbd_iht8x8_64_add sse4_1/;
specialize qw/vp9_highbd_iht16x16_256_add sse4_1/;
}
--- a/vp9/common/x86/vp9_idct_intrin_sse2.c
+++ b/vp9/common/x86/vp9_idct_intrin_sse2.c
@@ -10,8 +10,6 @@
#include "./vp9_rtcd.h"
#include "vpx_dsp/x86/inv_txfm_sse2.h"
-#include "vpx_dsp/x86/txfm_common_sse2.h"
-#include "vpx_ports/mem.h"
void vp9_iht4x4_16_add_sse2(const tran_low_t *input, uint8_t *dest, int stride,
int tx_type) {--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -70,6 +70,7 @@
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_idct_intrin_sse2.c
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht4x4_add_neon.c
VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht8x8_add_neon.c
+VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_iht_neon.h
ifeq ($(CONFIG_VP9_POSTPROC),yes)
VP9_COMMON_SRCS-$(HAVE_MSA) += common/mips/msa/vp9_mfqe_msa.c
@@ -81,6 +82,7 @@
VP9_COMMON_SRCS-$(HAVE_DSPR2) += common/mips/dspr2/vp9_itrans8_dspr2.c
VP9_COMMON_SRCS-$(HAVE_DSPR2) += common/mips/dspr2/vp9_itrans16_dspr2.c
else
+VP9_COMMON_SRCS-$(HAVE_NEON) += common/arm/neon/vp9_highbd_iht4x4_add_neon.c
VP9_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/vp9_highbd_iht4x4_add_sse4.c
VP9_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/vp9_highbd_iht8x8_add_sse4.c
VP9_COMMON_SRCS-$(HAVE_SSE4_1) += common/x86/vp9_highbd_iht16x16_add_sse4.c
--- a/vpx_dsp/arm/highbd_idct4x4_add_neon.c
+++ b/vpx_dsp/arm/highbd_idct4x4_add_neon.c
@@ -11,27 +11,10 @@
#include <arm_neon.h>
#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/arm/highbd_idct_neon.h"
#include "vpx_dsp/arm/idct_neon.h"
#include "vpx_dsp/inv_txfm.h"
-static INLINE void highbd_idct4x4_1_add_kernel1(uint16_t **dest,
- const int stride,
- const int16x8_t res,
- const int16x8_t max) {- const uint16x4_t a0 = vld1_u16(*dest);
- const uint16x4_t a1 = vld1_u16(*dest + stride);
- const int16x8_t a = vreinterpretq_s16_u16(vcombine_u16(a0, a1));
- // Note: In some profile tests, res is quite close to +/-32767.
- // We use saturating addition.
- const int16x8_t b = vqaddq_s16(res, a);
- const int16x8_t c = vminq_s16(b, max);
- const uint16x8_t d = vqshluq_n_s16(c, 0);
- vst1_u16(*dest, vget_low_u16(d));
- *dest += stride;
- vst1_u16(*dest, vget_high_u16(d));
- *dest += stride;
-}
-
// res is in reverse row order
static INLINE void highbd_idct4x4_1_add_kernel2(uint16_t **dest,
const int stride,
@@ -63,67 +46,6 @@
highbd_idct4x4_1_add_kernel1(&dest, stride, dc, max);
highbd_idct4x4_1_add_kernel1(&dest, stride, dc, max);
-}
-
-static INLINE void idct4x4_16_kernel_bd10(const int32x4_t cospis,
- int32x4_t *const a) {- int32x4_t b0, b1, b2, b3;
-
- transpose_s32_4x4(&a[0], &a[1], &a[2], &a[3]);
- b0 = vaddq_s32(a[0], a[2]);
- b1 = vsubq_s32(a[0], a[2]);
- b0 = vmulq_lane_s32(b0, vget_high_s32(cospis), 0);
- b1 = vmulq_lane_s32(b1, vget_high_s32(cospis), 0);
- b2 = vmulq_lane_s32(a[1], vget_high_s32(cospis), 1);
- b3 = vmulq_lane_s32(a[1], vget_low_s32(cospis), 1);
- b2 = vmlsq_lane_s32(b2, a[3], vget_low_s32(cospis), 1);
- b3 = vmlaq_lane_s32(b3, a[3], vget_high_s32(cospis), 1);
- b0 = vrshrq_n_s32(b0, DCT_CONST_BITS);
- b1 = vrshrq_n_s32(b1, DCT_CONST_BITS);
- b2 = vrshrq_n_s32(b2, DCT_CONST_BITS);
- b3 = vrshrq_n_s32(b3, DCT_CONST_BITS);
- a[0] = vaddq_s32(b0, b3);
- a[1] = vaddq_s32(b1, b2);
- a[2] = vsubq_s32(b1, b2);
- a[3] = vsubq_s32(b0, b3);
-}
-
-static INLINE void idct4x4_16_kernel_bd12(const int32x4_t cospis,
- int32x4_t *const a) {- int32x4_t b0, b1, b2, b3;
- int64x2_t c[12];
-
- transpose_s32_4x4(&a[0], &a[1], &a[2], &a[3]);
- b0 = vaddq_s32(a[0], a[2]);
- b1 = vsubq_s32(a[0], a[2]);
- c[0] = vmull_lane_s32(vget_low_s32(b0), vget_high_s32(cospis), 0);
- c[1] = vmull_lane_s32(vget_high_s32(b0), vget_high_s32(cospis), 0);
- c[2] = vmull_lane_s32(vget_low_s32(b1), vget_high_s32(cospis), 0);
- c[3] = vmull_lane_s32(vget_high_s32(b1), vget_high_s32(cospis), 0);
- c[4] = vmull_lane_s32(vget_low_s32(a[1]), vget_high_s32(cospis), 1);
- c[5] = vmull_lane_s32(vget_high_s32(a[1]), vget_high_s32(cospis), 1);
- c[6] = vmull_lane_s32(vget_low_s32(a[1]), vget_low_s32(cospis), 1);
- c[7] = vmull_lane_s32(vget_high_s32(a[1]), vget_low_s32(cospis), 1);
- c[8] = vmull_lane_s32(vget_low_s32(a[3]), vget_low_s32(cospis), 1);
- c[9] = vmull_lane_s32(vget_high_s32(a[3]), vget_low_s32(cospis), 1);
- c[10] = vmull_lane_s32(vget_low_s32(a[3]), vget_high_s32(cospis), 1);
- c[11] = vmull_lane_s32(vget_high_s32(a[3]), vget_high_s32(cospis), 1);
- c[4] = vsubq_s64(c[4], c[8]);
- c[5] = vsubq_s64(c[5], c[9]);
- c[6] = vaddq_s64(c[6], c[10]);
- c[7] = vaddq_s64(c[7], c[11]);
- b0 = vcombine_s32(vrshrn_n_s64(c[0], DCT_CONST_BITS),
- vrshrn_n_s64(c[1], DCT_CONST_BITS));
- b1 = vcombine_s32(vrshrn_n_s64(c[2], DCT_CONST_BITS),
- vrshrn_n_s64(c[3], DCT_CONST_BITS));
- b2 = vcombine_s32(vrshrn_n_s64(c[4], DCT_CONST_BITS),
- vrshrn_n_s64(c[5], DCT_CONST_BITS));
- b3 = vcombine_s32(vrshrn_n_s64(c[6], DCT_CONST_BITS),
- vrshrn_n_s64(c[7], DCT_CONST_BITS));
- a[0] = vaddq_s32(b0, b3);
- a[1] = vaddq_s32(b1, b2);
- a[2] = vsubq_s32(b1, b2);
- a[3] = vsubq_s32(b0, b3);
}
void vpx_highbd_idct4x4_16_add_neon(const tran_low_t *input, uint16_t *dest,
--- /dev/null
+++ b/vpx_dsp/arm/highbd_idct_neon.h
@@ -1,0 +1,99 @@
+/*
+ * Copyright (c) 2018 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_ARM_HIGHBD_IDCT_NEON_H_
+#define VPX_DSP_ARM_HIGHBD_IDCT_NEON_H_
+
+#include <arm_neon.h>
+
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/inv_txfm.h"
+
+static INLINE void highbd_idct4x4_1_add_kernel1(uint16_t **dest,
+ const int stride,
+ const int16x8_t res,
+ const int16x8_t max) {+ const uint16x4_t a0 = vld1_u16(*dest);
+ const uint16x4_t a1 = vld1_u16(*dest + stride);
+ const int16x8_t a = vreinterpretq_s16_u16(vcombine_u16(a0, a1));
+ // Note: In some profile tests, res is quite close to +/-32767.
+ // We use saturating addition.
+ const int16x8_t b = vqaddq_s16(res, a);
+ const int16x8_t c = vminq_s16(b, max);
+ const uint16x8_t d = vqshluq_n_s16(c, 0);
+ vst1_u16(*dest, vget_low_u16(d));
+ *dest += stride;
+ vst1_u16(*dest, vget_high_u16(d));
+ *dest += stride;
+}
+
+static INLINE void idct4x4_16_kernel_bd10(const int32x4_t cospis,
+ int32x4_t *const a) {+ int32x4_t b0, b1, b2, b3;
+
+ transpose_s32_4x4(&a[0], &a[1], &a[2], &a[3]);
+ b0 = vaddq_s32(a[0], a[2]);
+ b1 = vsubq_s32(a[0], a[2]);
+ b0 = vmulq_lane_s32(b0, vget_high_s32(cospis), 0);
+ b1 = vmulq_lane_s32(b1, vget_high_s32(cospis), 0);
+ b2 = vmulq_lane_s32(a[1], vget_high_s32(cospis), 1);
+ b3 = vmulq_lane_s32(a[1], vget_low_s32(cospis), 1);
+ b2 = vmlsq_lane_s32(b2, a[3], vget_low_s32(cospis), 1);
+ b3 = vmlaq_lane_s32(b3, a[3], vget_high_s32(cospis), 1);
+ b0 = vrshrq_n_s32(b0, DCT_CONST_BITS);
+ b1 = vrshrq_n_s32(b1, DCT_CONST_BITS);
+ b2 = vrshrq_n_s32(b2, DCT_CONST_BITS);
+ b3 = vrshrq_n_s32(b3, DCT_CONST_BITS);
+ a[0] = vaddq_s32(b0, b3);
+ a[1] = vaddq_s32(b1, b2);
+ a[2] = vsubq_s32(b1, b2);
+ a[3] = vsubq_s32(b0, b3);
+}
+
+static INLINE void idct4x4_16_kernel_bd12(const int32x4_t cospis,
+ int32x4_t *const a) {+ int32x4_t b0, b1, b2, b3;
+ int64x2_t c[12];
+
+ transpose_s32_4x4(&a[0], &a[1], &a[2], &a[3]);
+ b0 = vaddq_s32(a[0], a[2]);
+ b1 = vsubq_s32(a[0], a[2]);
+ c[0] = vmull_lane_s32(vget_low_s32(b0), vget_high_s32(cospis), 0);
+ c[1] = vmull_lane_s32(vget_high_s32(b0), vget_high_s32(cospis), 0);
+ c[2] = vmull_lane_s32(vget_low_s32(b1), vget_high_s32(cospis), 0);
+ c[3] = vmull_lane_s32(vget_high_s32(b1), vget_high_s32(cospis), 0);
+ c[4] = vmull_lane_s32(vget_low_s32(a[1]), vget_high_s32(cospis), 1);
+ c[5] = vmull_lane_s32(vget_high_s32(a[1]), vget_high_s32(cospis), 1);
+ c[6] = vmull_lane_s32(vget_low_s32(a[1]), vget_low_s32(cospis), 1);
+ c[7] = vmull_lane_s32(vget_high_s32(a[1]), vget_low_s32(cospis), 1);
+ c[8] = vmull_lane_s32(vget_low_s32(a[3]), vget_low_s32(cospis), 1);
+ c[9] = vmull_lane_s32(vget_high_s32(a[3]), vget_low_s32(cospis), 1);
+ c[10] = vmull_lane_s32(vget_low_s32(a[3]), vget_high_s32(cospis), 1);
+ c[11] = vmull_lane_s32(vget_high_s32(a[3]), vget_high_s32(cospis), 1);
+ c[4] = vsubq_s64(c[4], c[8]);
+ c[5] = vsubq_s64(c[5], c[9]);
+ c[6] = vaddq_s64(c[6], c[10]);
+ c[7] = vaddq_s64(c[7], c[11]);
+ b0 = vcombine_s32(vrshrn_n_s64(c[0], DCT_CONST_BITS),
+ vrshrn_n_s64(c[1], DCT_CONST_BITS));
+ b1 = vcombine_s32(vrshrn_n_s64(c[2], DCT_CONST_BITS),
+ vrshrn_n_s64(c[3], DCT_CONST_BITS));
+ b2 = vcombine_s32(vrshrn_n_s64(c[4], DCT_CONST_BITS),
+ vrshrn_n_s64(c[5], DCT_CONST_BITS));
+ b3 = vcombine_s32(vrshrn_n_s64(c[6], DCT_CONST_BITS),
+ vrshrn_n_s64(c[7], DCT_CONST_BITS));
+ a[0] = vaddq_s32(b0, b3);
+ a[1] = vaddq_s32(b1, b2);
+ a[2] = vsubq_s32(b1, b2);
+ a[3] = vsubq_s32(b0, b3);
+}
+
+#endif // VPX_DSP_ARM_HIGHBD_IDCT_NEON_H_
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -242,6 +242,7 @@
DSP_SRCS-$(HAVE_NEON) += arm/highbd_idct32x32_34_add_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/highbd_idct32x32_135_add_neon.c
DSP_SRCS-$(HAVE_NEON) += arm/highbd_idct32x32_1024_add_neon.c
+DSP_SRCS-$(HAVE_NEON) += arm/highbd_idct_neon.h
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_inv_txfm_sse2.h
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_idct4x4_add_sse2.c
DSP_SRCS-$(HAVE_SSE2) += x86/highbd_idct8x8_add_sse2.c