hashirama/uvg266
1
0
Fork 0
mirror of https://github.com/ultravideo/uvg266.git synced 2024-11-27 19:24:06 +00:00
Code Issues Actions Projects Releases Wiki Activity

Create constant width SAD versions for 8 and 16 pixels

Browse source
This commit is contained in:
Pauli Oikkonen 2019-01-15 19:32:25 +02:00
parent a13fc51003
commit d2db0086e1
1 changed files with 70 additions and 3 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

73
src/strategies/sse41/picture-sse41.c
View file

@ -27,10 +27,65 @@
#include "kvazaar.h"
#include "strategyselector.h"
uint32_t kvz_reg_sad_sse41(const kvz_pixel * const data1, const kvz_pixel * const data2,
const int32_t width, const int32_t height, const uint32_t stride1,
static uint32_t reg_sad_w8(const kvz_pixel * const data1, const kvz_pixel * const data2,
const int32_t height, const uint32_t stride1,
const uint32_t stride2)
{
__m128i sse_inc = _mm_setzero_si128();
uint64_t result = 0;
int32_t y;
const int32_t height_xmm_bytes = height & ~1;
const int32_t height_parity = height & 1;
for (y = 0; y < height_xmm_bytes; y += 2) {
__m128d a_d = _mm_setzero_pd();
__m128d b_d = _mm_setzero_pd();
a_d = _mm_loadl_pd(a_d, (const double *)(data1 + (y + 0) * stride1));
b_d = _mm_loadl_pd(b_d, (const double *)(data2 + (y + 0) * stride2));
a_d = _mm_loadh_pd(a_d, (const double *)(data1 + (y + 1) * stride1));
b_d = _mm_loadh_pd(b_d, (const double *)(data2 + (y + 1) * stride2));
__m128i a = _mm_castpd_si128(a_d);
__m128i b = _mm_castpd_si128(b_d);
__m128i curr_sads = _mm_sad_epu8(a, b);
sse_inc = _mm_add_epi64(sse_inc, curr_sads);
}
if (height_parity) {
__m64 a = *(__m64 *)(data1 + y * stride1);
__m64 b = *(__m64 *)(data2 + y * stride2);
__m64 sads = _mm_sad_pu8(a, b);
result = (uint64_t)sads;
}
__m128i sse_inc_2 = _mm_shuffle_epi32(sse_inc, _MM_SHUFFLE(1, 0, 3, 2));
__m128i sad = _mm_add_epi64 (sse_inc, sse_inc_2);
result += _mm_cvtsi128_si32(sad);
return result;
}
static uint32_t reg_sad_w16(const kvz_pixel * const data1, const kvz_pixel * const data2,
const int32_t height, const uint32_t stride1,
const uint32_t stride2)
{
__m128i sse_inc = _mm_setzero_si128();
int32_t y;
for (y = 0; y < height; y++) {
__m128i a = _mm_loadu_si128((__m128i const*) &data1[y * stride1]);
__m128i b = _mm_loadu_si128((__m128i const*) &data2[y * stride2]);
__m128i curr_sads = _mm_sad_epu8(a, b);
sse_inc = _mm_add_epi64(sse_inc, curr_sads);
}
__m128i sse_inc_2 = _mm_shuffle_epi32(sse_inc, _MM_SHUFFLE(1, 0, 3, 2));
__m128i sad = _mm_add_epi64 (sse_inc, sse_inc_2);
return _mm_cvtsi128_si32(sad);
}
static uint32_t reg_sad_arbitrary(const kvz_pixel * const data1, const kvz_pixel * const data2,
const int32_t width, const int32_t height, const uint32_t stride1,
const uint32_t stride2)
{
int32_t y, x;
__m128i sse_inc = _mm_setzero_si128();
@ -67,6 +122,18 @@ uint32_t kvz_reg_sad_sse41(const kvz_pixel * const data1, const kvz_pixel * cons
return _mm_cvtsi128_si32(sad);
}
uint32_t kvz_reg_sad_sse41(const kvz_pixel * const data1, const kvz_pixel * const data2,
const int32_t width, const int32_t height, const uint32_t stride1,
const uint32_t stride2)
{
if (width == 8)
return reg_sad_w8(data1, data2, height, stride1, stride2);
if (width == 16)
return reg_sad_w16(data1, data2, height, stride1, stride2);
else
return reg_sad_arbitrary(data1, data2, width, height, stride1, stride2);
}
#endif //COMPILE_INTEL_SSE41