Commit 7a483cff authored by Angie Chiang's avatar Angie Chiang

Turn on SIMD optimization for dual_filter

Let aom_convolve8_### SIMD implementation support any block width.
Turn on SIMD optimization when interpolation filter types on two
directions are different.

This will reduce 30% of encoding time when dual_filter and ext_interp
both on.

Change-Id: I539dbb2737f01835034b7269656a15b2058fa3cc
parent b968d46a
...@@ -41,12 +41,19 @@ typedef void filter8_1dfunction(const uint8_t *src_ptr, ptrdiff_t src_pitch, ...@@ -41,12 +41,19 @@ typedef void filter8_1dfunction(const uint8_t *src_ptr, ptrdiff_t src_pitch,
dst += 16; \ dst += 16; \
w -= 16; \ w -= 16; \
} \ } \
if (w == 8) { \ while (w >= 8) { \
aom_filter_block1d8_##dir##8_##avg##opt(src_start, src_stride, dst, \ aom_filter_block1d8_##dir##8_##avg##opt(src_start, src_stride, dst, \
dst_stride, h, filter); \ dst_stride, h, filter); \
} else if (w == 4) { \ src += 8; \
dst += 8; \
w -= 8; \
} \
while (w >= 4) { \
aom_filter_block1d4_##dir##8_##avg##opt(src_start, src_stride, dst, \ aom_filter_block1d4_##dir##8_##avg##opt(src_start, src_stride, dst, \
dst_stride, h, filter); \ dst_stride, h, filter); \
src += 4; \
dst += 4; \
w -= 4; \
} \ } \
} else { \ } else { \
while (w >= 16) { \ while (w >= 16) { \
...@@ -56,14 +63,25 @@ typedef void filter8_1dfunction(const uint8_t *src_ptr, ptrdiff_t src_pitch, ...@@ -56,14 +63,25 @@ typedef void filter8_1dfunction(const uint8_t *src_ptr, ptrdiff_t src_pitch,
dst += 16; \ dst += 16; \
w -= 16; \ w -= 16; \
} \ } \
if (w == 8) { \ while (w >= 8) { \
aom_filter_block1d8_##dir##2_##avg##opt(src, src_stride, dst, \ aom_filter_block1d8_##dir##2_##avg##opt(src, src_stride, dst, \
dst_stride, h, filter); \ dst_stride, h, filter); \
} else if (w == 4) { \ src += 8; \
dst += 8; \
w -= 8; \
} \
while (w >= 4) { \
aom_filter_block1d4_##dir##2_##avg##opt(src, src_stride, dst, \ aom_filter_block1d4_##dir##2_##avg##opt(src, src_stride, dst, \
dst_stride, h, filter); \ dst_stride, h, filter); \
src += 4; \
dst += 4; \
w -= 4; \
} \ } \
} \ } \
if (w) { \
aom_convolve8_##name##_c(src, src_stride, dst, dst_stride, filter_x, \
x_step_q4, filter_y, y_step_q4, w, h); \
} \
} }
#define FUN_CONV_2D(avg, opt) \ #define FUN_CONV_2D(avg, opt) \
......
...@@ -12,6 +12,7 @@ ...@@ -12,6 +12,7 @@
#include <assert.h> #include <assert.h>
#include <string.h> #include <string.h>
#include "./aom_dsp_rtcd.h"
#include "./av1_rtcd.h" #include "./av1_rtcd.h"
#include "av1/common/convolve.h" #include "av1/common/convolve.h"
#include "av1/common/filter.h" #include "av1/common/filter.h"
...@@ -104,6 +105,45 @@ static void convolve_copy(const uint8_t *src, int src_stride, uint8_t *dst, ...@@ -104,6 +105,45 @@ static void convolve_copy(const uint8_t *src, int src_stride, uint8_t *dst,
} }
} }
void av1_convolve_horiz_facade(const uint8_t *src, int src_stride, uint8_t *dst,
int dst_stride, int w, int h,
const InterpFilterParams filter_params,
const int subpel_x_q4, int x_step_q4, int avg) {
if (filter_params.taps == SUBPEL_TAPS) {
const int16_t *filter_x =
av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
if (avg == 0)
aom_convolve8_horiz(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
NULL, -1, w, h);
else
aom_convolve8_avg_horiz(src, src_stride, dst, dst_stride, filter_x,
x_step_q4, NULL, -1, w, h);
} else {
av1_convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
subpel_x_q4, x_step_q4, avg);
}
}
void av1_convolve_vert_facade(const uint8_t *src, int src_stride, uint8_t *dst,
int dst_stride, int w, int h,
const InterpFilterParams filter_params,
const int subpel_y_q4, int y_step_q4, int avg) {
if (filter_params.taps == SUBPEL_TAPS) {
const int16_t *filter_y =
av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
if (avg == 0) {
aom_convolve8_vert(src, src_stride, dst, dst_stride, NULL, -1, filter_y,
y_step_q4, w, h);
} else {
aom_convolve8_avg_vert(src, src_stride, dst, dst_stride, NULL, -1,
filter_y, y_step_q4, w, h);
}
} else {
av1_convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
subpel_y_q4, y_step_q4, avg);
}
}
void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst, void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
int dst_stride, int w, int h, int dst_stride, int w, int h,
#if CONFIG_DUAL_FILTER #if CONFIG_DUAL_FILTER
...@@ -146,11 +186,12 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst, ...@@ -146,11 +186,12 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
av1_convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params, av1_convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
subpel_y_q4, y_step_q4, ref_idx); subpel_y_q4, y_step_q4, ref_idx);
} else { } else {
// temp's size is set to (maximum possible intermediate height or width) * // temp's size is set to a 256 aligned value to facilitate SIMD
// MAX_SB_SIZE // implementation. The value is greater than (maximum possible intermediate
uint8_t temp[((((MAX_SB_SIZE - 1) * MAX_STEP + 15) >> SUBPEL_BITS) + // height or width) * MAX_SB_SIZE
MAX_FILTER_TAP) * DECLARE_ALIGNED(16, uint8_t,
MAX_SB_SIZE]; temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
int filter_size; int filter_size;
InterpFilterParams filter_params; InterpFilterParams filter_params;
#if CONFIG_DUAL_FILTER #if CONFIG_DUAL_FILTER
...@@ -171,7 +212,7 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst, ...@@ -171,7 +212,7 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
// complexity // complexity
if (filter_params_y.taps < filter_params_x.taps) { if (filter_params_y.taps < filter_params_x.taps) {
int intermediate_width; int intermediate_width;
int temp_stride; int temp_stride = max_intermediate_size;
#if CONFIG_DUAL_FILTER #if CONFIG_DUAL_FILTER
filter_params = filter_params_y; filter_params = filter_params_y;
filter_size = filter_params_x.taps; filter_size = filter_params_x.taps;
...@@ -181,13 +222,13 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst, ...@@ -181,13 +222,13 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
#endif #endif
intermediate_width = intermediate_width =
(((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size; (((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size;
temp_stride = intermediate_width; assert(intermediate_width <= max_intermediate_size);
assert(filter_params.taps <= MAX_FILTER_TAP); assert(filter_params.taps <= MAX_FILTER_TAP);
av1_convolve_vert(src - (filter_size / 2 - 1), src_stride, temp, av1_convolve_vert_facade(src - (filter_size / 2 - 1), src_stride, temp,
temp_stride, intermediate_width, h, filter_params, temp_stride, intermediate_width, h,
subpel_y_q4, y_step_q4, 0); filter_params, subpel_y_q4, y_step_q4, 0);
#if CONFIG_DUAL_FILTER #if CONFIG_DUAL_FILTER
filter_params = filter_params_x; filter_params = filter_params_x;
...@@ -196,14 +237,14 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst, ...@@ -196,14 +237,14 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
#endif #endif
assert(filter_params.taps <= MAX_FILTER_TAP); assert(filter_params.taps <= MAX_FILTER_TAP);
av1_convolve_horiz(temp + (filter_size / 2 - 1), temp_stride, dst, av1_convolve_horiz_facade(temp + (filter_size / 2 - 1), temp_stride, dst,
dst_stride, w, h, filter_params, subpel_x_q4, dst_stride, w, h, filter_params, subpel_x_q4,
x_step_q4, ref_idx); x_step_q4, ref_idx);
} else } else
#endif #endif
{ {
int intermediate_height; int intermediate_height;
int temp_stride = w; int temp_stride = MAX_SB_SIZE;
#if CONFIG_DUAL_FILTER #if CONFIG_DUAL_FILTER
filter_params = filter_params_x; filter_params = filter_params_x;
filter_size = filter_params_y.taps; filter_size = filter_params_y.taps;
...@@ -213,12 +254,15 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst, ...@@ -213,12 +254,15 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
#endif #endif
intermediate_height = intermediate_height =
(((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size; (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size;
assert(intermediate_height <= max_intermediate_size);
(void)max_intermediate_size;
assert(filter_params.taps <= MAX_FILTER_TAP); assert(filter_params.taps <= MAX_FILTER_TAP);
av1_convolve_horiz(src - src_stride * (filter_size / 2 - 1), src_stride, av1_convolve_horiz_facade(src - src_stride * (filter_size / 2 - 1),
temp, temp_stride, w, intermediate_height, src_stride, temp, temp_stride, w,
filter_params, subpel_x_q4, x_step_q4, 0); intermediate_height, filter_params, subpel_x_q4,
x_step_q4, 0);
#if CONFIG_DUAL_FILTER #if CONFIG_DUAL_FILTER
filter_params = filter_params_y; filter_params = filter_params_y;
...@@ -227,9 +271,9 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst, ...@@ -227,9 +271,9 @@ void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
#endif #endif
assert(filter_params.taps <= MAX_FILTER_TAP); assert(filter_params.taps <= MAX_FILTER_TAP);
av1_convolve_vert(temp + temp_stride * (filter_size / 2 - 1), temp_stride, av1_convolve_vert_facade(temp + temp_stride * (filter_size / 2 - 1),
dst, dst_stride, w, h, filter_params, subpel_y_q4, temp_stride, dst, dst_stride, w, h,
y_step_q4, ref_idx); filter_params, subpel_y_q4, y_step_q4, ref_idx);
} }
} }
} }
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment