| 1 |
/* { dg-do compile } */ |
| 2 |
/* { dg-options "-O3 -march=x86-64 -mtune=skylake -Wno-attributes" } */ |
| 3 |
|
| 4 |
#include <x86intrin.h> |
| 5 |
#include <stdint.h> |
| 6 |
|
| 7 |
__attribute__((always_inline, target("avx2"))) |
| 8 |
static __m256i |
| 9 |
load8bit_4x4_avx2(const uint8_t *const src, const uint32_t stride) |
| 10 |
{ |
| 11 |
__m128i src01, src23 = _mm_setzero_si128(); |
| 12 |
src01 = _mm_cvtsi32_si128(*(int32_t*)(src + 0 * stride)); |
| 13 |
src23 = _mm_insert_epi32(src23, *(int32_t *)(src + 3 * stride), 1); |
| 14 |
return _mm256_setr_m128i(src01, src23); |
| 15 |
} |
| 16 |
|
| 17 |
__attribute__ ((noinline, noipa, target("avx2"))) |
| 18 |
uint32_t |
| 19 |
compute4x_m_sad_avx2_intrin(uint8_t *src, uint32_t src_stride, |
| 20 |
uint8_t *ref, uint32_t ref_stride, |
| 21 |
uint32_t height) |
| 22 |
{ |
| 23 |
__m128i xmm0; |
| 24 |
__m256i ymm = _mm256_setzero_si256(); |
| 25 |
uint32_t y; |
| 26 |
|
| 27 |
for (y = 0; y < height; y += 4) { |
| 28 |
const __m256i src0123 = load8bit_4x4_avx2(src, src_stride); |
| 29 |
const __m256i ref0123 = load8bit_4x4_avx2(ref, ref_stride); |
| 30 |
ymm = _mm256_add_epi32(ymm, _mm256_sad_epu8(src0123, ref0123)); |
| 31 |
src += src_stride << 2; |
| 32 |
ref += ref_stride << 2; |
| 33 |
} |
| 34 |
|
| 35 |
xmm0 = _mm_add_epi32(_mm256_castsi256_si128(ymm), |
| 36 |
_mm256_extracti128_si256(ymm, 1)); |
| 37 |
|
| 38 |
return (uint32_t)_mm_cvtsi128_si32(xmm0); |
| 39 |
} |
| 40 |
|
| 41 |
/* Expect assembly like: |
| 42 |
|
| 43 |
vextracti128 $0x1, %ymm3, %xmm3 |
| 44 |
vpaddd %xmm3, %xmm0, %xmm0 |
| 45 |
vmovd %xmm0, %eax |
| 46 |
vzeroupper |
| 47 |
|
| 48 |
rather than: |
| 49 |
|
| 50 |
vzeroupper |
| 51 |
vextracti128 $0x1, %ymm3, %xmm3 |
| 52 |
vpaddd %xmm3, %xmm0, %xmm0 |
| 53 |
vmovd %xmm0, %eax |
| 54 |
|
| 55 |
*/ |
| 56 |
|
| 57 |
/* { dg-final { scan-assembler "\[ \t\]+vextracti128\[ \t\]+\[^\n\]+\n\[ \t\]+vpaddd\[ \t\]+\[^\n\]+\n\[ \t\]+vmovd\[ \t\]+\[^\n\]+\n\[ \t\]+vzeroupper" } } */ |