generic-library/vpx
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

sad neon: avg for 64x[32,64]

Browse Source 
BUG=webm:1425

Change-Id: Id84d97807a6a0fbcc889c4dfe11929d54f85493d
This commit is contained in:
Johann 2017-07-06 09:16:02 -07:00
parent 6ae8f8dbe8
commit e4e08556db
3 changed files with 55 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
test/sad_test.cc
View File

@ -658,6 +658,8 @@ const SadMxNParam neon_tests[] = {
INSTANTIATE_TEST_CASE_P(NEON, SADTest, ::testing::ValuesIn(neon_tests));
const SadMxNAvgParam avg_neon_tests[] = {
SadMxNAvgParam(64, 64, &vpx_sad64x64_avg_neon),
SadMxNAvgParam(64, 32, &vpx_sad64x32_avg_neon),
SadMxNAvgParam(32, 64, &vpx_sad32x64_avg_neon),
SadMxNAvgParam(32, 32, &vpx_sad32x32_avg_neon),
SadMxNAvgParam(32, 16, &vpx_sad32x16_avg_neon),

51
vpx_dsp/arm/sad_neon.c
View File

@ -297,11 +297,62 @@ static INLINE uint32x4_t sad64x(const uint8_t *a, int a_stride,
}
}
static INLINE uint32x4_t sad64x_avg(const uint8_t *a, int a_stride,
const uint8_t *b, int b_stride,
const uint8_t *c, const int height) {
int i;
uint16x8_t abs_0 = vdupq_n_u16(0);
uint16x8_t abs_1 = vdupq_n_u16(0);
for (i = 0; i < height; ++i) {
const uint8x16_t a_0 = vld1q_u8(a);
const uint8x16_t a_1 = vld1q_u8(a + 16);
const uint8x16_t a_2 = vld1q_u8(a + 32);
const uint8x16_t a_3 = vld1q_u8(a + 48);
const uint8x16_t b_0 = vld1q_u8(b);
const uint8x16_t b_1 = vld1q_u8(b + 16);
const uint8x16_t b_2 = vld1q_u8(b + 32);
const uint8x16_t b_3 = vld1q_u8(b + 48);
const uint8x16_t c_0 = vld1q_u8(c);
const uint8x16_t c_1 = vld1q_u8(c + 16);
const uint8x16_t c_2 = vld1q_u8(c + 32);
const uint8x16_t c_3 = vld1q_u8(c + 48);
const uint8x16_t avg_0 = vrhaddq_u8(b_0, c_0);
const uint8x16_t avg_1 = vrhaddq_u8(b_1, c_1);
const uint8x16_t avg_2 = vrhaddq_u8(b_2, c_2);
const uint8x16_t avg_3 = vrhaddq_u8(b_3, c_3);
a += a_stride;
b += b_stride;
c += 64;
abs_0 = vabal_u8(abs_0, vget_low_u8(a_0), vget_low_u8(avg_0));
abs_0 = vabal_u8(abs_0, vget_high_u8(a_0), vget_high_u8(avg_0));
abs_0 = vabal_u8(abs_0, vget_low_u8(a_1), vget_low_u8(avg_1));
abs_0 = vabal_u8(abs_0, vget_high_u8(a_1), vget_high_u8(avg_1));
abs_1 = vabal_u8(abs_1, vget_low_u8(a_2), vget_low_u8(avg_2));
abs_1 = vabal_u8(abs_1, vget_high_u8(a_2), vget_high_u8(avg_2));
abs_1 = vabal_u8(abs_1, vget_low_u8(a_3), vget_low_u8(avg_3));
abs_1 = vabal_u8(abs_1, vget_high_u8(a_3), vget_high_u8(avg_3));
}
{
const uint32x4_t sum = vpaddlq_u16(abs_0);
return vpadalq_u16(sum, abs_1);
}
}
#define sad64xN(n) \
uint32_t vpx_sad64x##n##_neon(const uint8_t *src, int src_stride, \
const uint8_t *ref, int ref_stride) { \
const uint32x4_t abs = sad64x(src, src_stride, ref, ref_stride, n); \
return horizontal_add_32x4(abs); \
} \
\
uint32_t vpx_sad64x##n##_avg_neon(const uint8_t *src, int src_stride, \
const uint8_t *ref, int ref_stride, \
const uint8_t *second_pred) { \
const uint32x4_t abs = \
sad64x_avg(src, src_stride, ref, ref_stride, second_pred, n); \
return horizontal_add_32x4(abs); \
}
sad64xN(32);

4
vpx_dsp/vpx_dsp_rtcd_defs.pl
View File

@ -780,10 +780,10 @@ if (vpx_config("CONFIG_VP9_ENCODER") eq "yes") {
} # CONFIG_VP9_ENCODER
add_proto qw/unsigned int vpx_sad64x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
specialize qw/vpx_sad64x64_avg avx2 msa sse2 vsx/;
specialize qw/vpx_sad64x64_avg neon avx2 msa sse2 vsx/;
add_proto qw/unsigned int vpx_sad64x32_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
specialize qw/vpx_sad64x32_avg avx2 msa sse2 vsx/;
specialize qw/vpx_sad64x32_avg neon avx2 msa sse2 vsx/;
add_proto qw/unsigned int vpx_sad32x64_avg/, "const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred";
specialize qw/vpx_sad32x64_avg neon avx2 msa sse2 vsx/;