ref: fa472129334e197f0a977fd8ccd0bf2a145d0935
parent: 9cc8995f46e423002e43eb6dab4a3fc140b9cb0b
author: Scott LaVarnway <[email protected]>
date: Wed Aug 12 12:14:53 EDT 2015
VPX: removed step checks from neon convolve code The check is handled by the predictor table. Change-Id: I42479f843e77a2d40cdcdfc9e2e6c48a05a36561
--- a/vpx_dsp/arm/vpx_convolve8_avg_neon.c
+++ b/vpx_dsp/arm/vpx_convolve8_avg_neon.c
@@ -9,6 +9,7 @@
*/
#include <arm_neon.h>
+#include <assert.h>
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
@@ -15,17 +16,6 @@
#include "vpx/vpx_integer.h"
#include "vpx_ports/mem.h"
-void vpx_convolve8_avg_horiz_c(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);
-void vpx_convolve8_avg_vert_c(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);
-
static INLINE int32x4_t MULTIPLY_BY_Q0(
int16x4_t dsrc0,
int16x4_t dsrc1,
@@ -82,12 +72,7 @@
uint16x4x2_t d0x2u16, d1x2u16;
uint32x4x2_t q0x2u32;
- if (x_step_q4 != 16) {- vpx_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride,
- filter_x, x_step_q4,
- filter_y, y_step_q4, w, h);
- return;
-}
+ assert(x_step_q4 == 16);
q0s16 = vld1q_s16(filter_x);
@@ -271,12 +256,7 @@
uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
int32x4_t q1s32, q2s32, q14s32, q15s32;
- if (y_step_q4 != 16) {- vpx_convolve8_avg_vert_c(src, src_stride, dst, dst_stride,
- filter_x, x_step_q4,
- filter_y, y_step_q4, w, h);
- return;
- }
+ assert(y_step_q4 == 16);
src -= src_stride * 3;
q0s16 = vld1q_s16(filter_y);
--- a/vpx_dsp/arm/vpx_convolve8_avg_neon_asm.asm
+++ b/vpx_dsp/arm/vpx_convolve8_avg_neon_asm.asm
@@ -19,8 +19,6 @@
EXPORT |vpx_convolve8_avg_horiz_neon|
EXPORT |vpx_convolve8_avg_vert_neon|
- IMPORT |vpx_convolve8_avg_horiz_c|
- IMPORT |vpx_convolve8_avg_vert_c|
ARM
REQUIRE8
PRESERVE8
@@ -52,10 +50,6 @@
; sp[]int h
|vpx_convolve8_avg_horiz_neon| PROC
- ldr r12, [sp, #4] ; x_step_q4
- cmp r12, #16
- bne vpx_convolve8_avg_horiz_c
-
push {r4-r10, lr}sub r0, r0, #3 ; adjust for taps
@@ -184,10 +178,6 @@
ENDP
|vpx_convolve8_avg_vert_neon| PROC
- ldr r12, [sp, #12]
- cmp r12, #16
- bne vpx_convolve8_avg_vert_c
-
push {r4-r8, lr}; adjust for taps
--- a/vpx_dsp/arm/vpx_convolve8_neon.c
+++ b/vpx_dsp/arm/vpx_convolve8_neon.c
@@ -9,6 +9,7 @@
*/
#include <arm_neon.h>
+#include <assert.h>
#include "./vpx_config.h"
#include "./vpx_dsp_rtcd.h"
@@ -15,17 +16,6 @@
#include "vpx/vpx_integer.h"
#include "vpx_ports/mem.h"
-void vpx_convolve8_horiz_c(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);
-void vpx_convolve8_vert_c(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);
-
static INLINE int32x4_t MULTIPLY_BY_Q0(
int16x4_t dsrc0,
int16x4_t dsrc1,
@@ -82,12 +72,7 @@
uint16x4x2_t d0x2u16, d1x2u16;
uint32x4x2_t q0x2u32;
- if (x_step_q4 != 16) {- vpx_convolve8_horiz_c(src, src_stride, dst, dst_stride,
- filter_x, x_step_q4,
- filter_y, y_step_q4, w, h);
- return;
- }
+ assert(x_step_q4 == 16);
q0s16 = vld1q_s16(filter_x);
@@ -255,12 +240,7 @@
uint16x8_t q1u16, q2u16, q8u16, q9u16, q10u16, q11u16, q12u16, q13u16;
int32x4_t q1s32, q2s32, q14s32, q15s32;
- if (y_step_q4 != 16) {- vpx_convolve8_vert_c(src, src_stride, dst, dst_stride,
- filter_x, x_step_q4,
- filter_y, y_step_q4, w, h);
- return;
- }
+ assert(y_step_q4 == 16);
src -= src_stride * 3;
q0s16 = vld1q_s16(filter_y);
--- a/vpx_dsp/arm/vpx_convolve8_neon_asm.asm
+++ b/vpx_dsp/arm/vpx_convolve8_neon_asm.asm
@@ -19,8 +19,6 @@
EXPORT |vpx_convolve8_horiz_neon|
EXPORT |vpx_convolve8_vert_neon|
- IMPORT |vpx_convolve8_horiz_c|
- IMPORT |vpx_convolve8_vert_c|
ARM
REQUIRE8
PRESERVE8
@@ -52,10 +50,6 @@
; sp[]int h
|vpx_convolve8_horiz_neon| PROC
- ldr r12, [sp, #4] ; x_step_q4
- cmp r12, #16
- bne vpx_convolve8_horiz_c
-
push {r4-r10, lr}sub r0, r0, #3 ; adjust for taps
@@ -173,10 +167,6 @@
ENDP
|vpx_convolve8_vert_neon| PROC
- ldr r12, [sp, #12]
- cmp r12, #16
- bne vpx_convolve8_vert_c
-
push {r4-r8, lr}; adjust for taps
--- a/vpx_dsp/arm/vpx_convolve_neon.c
+++ b/vpx_dsp/arm/vpx_convolve_neon.c
@@ -8,6 +8,8 @@
* be found in the AUTHORS file in the root of the source tree.
*/
+#include <assert.h>
+
#include "./vpx_dsp_rtcd.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_ports/mem.h"
@@ -25,14 +27,8 @@
// Account for the vertical phase needing 3 lines prior and 4 lines post
int intermediate_height = h + 7;
- if (x_step_q4 != 16 || y_step_q4 != 16) {- vpx_convolve8_c(src, src_stride,
- dst, dst_stride,
- filter_x, x_step_q4,
- filter_y, y_step_q4,
- w, h);
- return;
- }
+ assert(y_step_q4 == 16);
+ assert(x_step_q4 == 16);
/* Filter starting 3 lines back. The neon implementation will ignore the
* given height and filter a multiple of 4 lines. Since this goes in to
@@ -59,14 +55,8 @@
DECLARE_ALIGNED(8, uint8_t, temp[64 * 72]);
int intermediate_height = h + 7;
- if (x_step_q4 != 16 || y_step_q4 != 16) {- vpx_convolve8_avg_c(src, src_stride,
- dst, dst_stride,
- filter_x, x_step_q4,
- filter_y, y_step_q4,
- w, h);
- return;
- }
+ assert(y_step_q4 == 16);
+ assert(x_step_q4 == 16);
/* This implementation has the same issues as above. In addition, we only want
* to average the values after both passes.