Skip to content
Snippets Groups Projects
Commit 8878fa4f authored by Angie Chiang's avatar Angie Chiang
Browse files

convolve8 sse2 test 

This experiment shows that when frame size is 64x64
vpx_highbd_convolve8_sse2 and vpx_convolve8_sse2's speed are similar.
However when frame size becomes 1024x1024
vpx_highbd_convolve8_sse2 is around 50% slower than vpx_convolve8_sse2
we think the bottleneck is from memory IO

VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_8_64
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_8_64 (17 ms)
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_16_64
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_16_64 (42 ms)
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_32_64
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_32_64 (139 ms)
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_64_64
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_64_64 (499 ms)

VP10ConvolveTest.vpx_convolve8_sse2_speed_l_8_64
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_8_64 (16 ms)
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_16_64
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_16_64 (40 ms)
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_32_64
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_32_64 (130 ms)
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_64_64
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_64_64 (485 ms)

VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_8_1024
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_8_1024 (32 ms)
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_16_1024
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_16_1024 (61 ms)
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_32_1024
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_32_1024 (196 ms)
VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_64_1024

VP10ConvolveTest.vpx_highbd_convolve8_sse2_speed_64_1024 (694 ms)
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_8_1024
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_8_1024 (21 ms)
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_16_1024
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_16_1024 (44 ms)
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_32_1024
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_32_1024 (138 ms)
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_64_1024
VP10ConvolveTest.vpx_convolve8_sse2_speed_l_64_1024 (491 ms)

Change-Id: I3131a031e0380e8eae748cfcccc6cbb961d05943
parent 961668c9
No related branches found
No related tags found
Hide whitespace changes
Inline Side-by-side
test/vp10_convolve_test.cc
+ 129
− 0
View file @ 8878fa4f
......@@ -5,6 +5,7 @@
#include "vp10/common/filter.h"
#include "vp10/common/vp10_convolve.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_ports/mem.h"
using libvpx_test::ACMRandom;
......@@ -247,4 +248,132 @@ TEST(VP10ConvolveTest, vp10_highbd_convolve_avg) {
EXPECT_EQ(dst[0], ROUND_POWER_OF_TWO(dst0[0] + dst1[0], 1));
}
#endif // CONFIG_VP9_HIGHBITDEPTH
#define CONVOLVE_SPEED_TEST 0
#if CONVOLVE_SPEED_TEST
#define highbd_convolve_speed(func, block_size, frame_size) \
TEST(VP10ConvolveTest, func##_speed_##block_size##_##frame_size) { \
ACMRandom rnd(ACMRandom::DeterministicSeed()); \
INTERP_FILTER interp_filter = EIGHTTAP; \
InterpFilterParams filter_params = \
vp10_get_interp_filter_params(interp_filter); \
ptrdiff_t filter_size = filter_params.tap; \
int filter_center = filter_size / 2 - 1; \
DECLARE_ALIGNED(16, uint16_t, \
src[(frame_size + 7) * (frame_size + 7)]) = {0}; \
int src_stride = frame_size + 7; \
DECLARE_ALIGNED(16, uint16_t, dst[frame_size * frame_size]) = {0}; \
int dst_stride = frame_size; \
int x_step_q4 = 16; \
int y_step_q4 = 16; \
int subpel_x_q4 = 8; \
int subpel_y_q4 = 6; \
int bd = 10; \
\
int w = block_size; \
int h = block_size; \
\
const int16_t* filter_x = \
vp10_get_interp_filter_kernel(filter_params, subpel_x_q4); \
const int16_t* filter_y = \
vp10_get_interp_filter_kernel(filter_params, subpel_y_q4); \
\
for (int i = 0; i < src_stride * src_stride; i++) { \
src[i] = rnd.Rand16() % (1 << bd); \
} \
\
int offset = filter_center * src_stride + filter_center; \
int row_offset = 0; \
int col_offset = 0; \
for (int i = 0; i < 100000; i++) { \
int src_total_offset = offset + col_offset * src_stride + row_offset; \
int dst_total_offset = col_offset * dst_stride + row_offset; \
func(CONVERT_TO_BYTEPTR(src + src_total_offset), src_stride, \
CONVERT_TO_BYTEPTR(dst + dst_total_offset), dst_stride, filter_x, \
x_step_q4, filter_y, y_step_q4, w, h, bd); \
if (offset + w + w < frame_size) { \
row_offset += w; \
} else { \
row_offset = 0; \
col_offset += h; \
} \
if (col_offset + h >= frame_size) { \
col_offset = 0; \
} \
} \
}
#define lowbd_convolve_speed(func, block_size, frame_size) \
TEST(VP10ConvolveTest, func##_speed_l_##block_size##_##frame_size) { \
ACMRandom rnd(ACMRandom::DeterministicSeed()); \
INTERP_FILTER interp_filter = EIGHTTAP; \
InterpFilterParams filter_params = \
vp10_get_interp_filter_params(interp_filter); \
ptrdiff_t filter_size = filter_params.tap; \
int filter_center = filter_size / 2 - 1; \
DECLARE_ALIGNED(16, uint8_t, src[(frame_size + 7) * (frame_size + 7)]); \
int src_stride = frame_size + 7; \
DECLARE_ALIGNED(16, uint8_t, dst[frame_size * frame_size]); \
int dst_stride = frame_size; \
int x_step_q4 = 16; \
int y_step_q4 = 16; \
int subpel_x_q4 = 8; \
int subpel_y_q4 = 6; \
int bd = 8; \
\
int w = block_size; \
int h = block_size; \
\
const int16_t* filter_x = \
vp10_get_interp_filter_kernel(filter_params, subpel_x_q4); \
const int16_t* filter_y = \
vp10_get_interp_filter_kernel(filter_params, subpel_y_q4); \
\
for (int i = 0; i < src_stride * src_stride; i++) { \
src[i] = rnd.Rand16() % (1 << bd); \
} \
\
int offset = filter_center * src_stride + filter_center; \
int row_offset = 0; \
int col_offset = 0; \
for (int i = 0; i < 100000; i++) { \
func(src + offset, src_stride, dst, dst_stride, filter_x, x_step_q4, \
filter_y, y_step_q4, w, h); \
if (offset + w + w < frame_size) { \
row_offset += w; \
} else { \
row_offset = 0; \
col_offset += h; \
} \
if (col_offset + h >= frame_size) { \
col_offset = 0; \
} \
} \
}
// This experiment shows that when frame size is 64x64
// vpx_highbd_convolve8_sse2 and vpx_convolve8_sse2's speed are similar.
// However when frame size becomes 1024x1024
// vpx_highbd_convolve8_sse2 is around 50% slower than vpx_convolve8_sse2
// we think the bottleneck is from memory IO
highbd_convolve_speed(vpx_highbd_convolve8_sse2, 8, 64);
highbd_convolve_speed(vpx_highbd_convolve8_sse2, 16, 64);
highbd_convolve_speed(vpx_highbd_convolve8_sse2, 32, 64);
highbd_convolve_speed(vpx_highbd_convolve8_sse2, 64, 64);
lowbd_convolve_speed(vpx_convolve8_sse2, 8, 64);
lowbd_convolve_speed(vpx_convolve8_sse2, 16, 64);
lowbd_convolve_speed(vpx_convolve8_sse2, 32, 64);
lowbd_convolve_speed(vpx_convolve8_sse2, 64, 64);
highbd_convolve_speed(vpx_highbd_convolve8_sse2, 8, 1024);
highbd_convolve_speed(vpx_highbd_convolve8_sse2, 16, 1024);
highbd_convolve_speed(vpx_highbd_convolve8_sse2, 32, 1024);
highbd_convolve_speed(vpx_highbd_convolve8_sse2, 64, 1024);
lowbd_convolve_speed(vpx_convolve8_sse2, 8, 1024);
lowbd_convolve_speed(vpx_convolve8_sse2, 16, 1024);
lowbd_convolve_speed(vpx_convolve8_sse2, 32, 1024);
lowbd_convolve_speed(vpx_convolve8_sse2, 64, 1024);
#endif // CONVOLVE_SPEED_TEST
} // namespace
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