diff options
Diffstat (limited to '')
| -rw-r--r-- | third_party/aom/aom_dsp/x86/synonyms_avx2.h | 25 |
1 files changed, 25 insertions, 0 deletions
diff --git a/third_party/aom/aom_dsp/x86/synonyms_avx2.h b/third_party/aom/aom_dsp/x86/synonyms_avx2.h index b729e5f410..7548d4d4f4 100644 --- a/third_party/aom/aom_dsp/x86/synonyms_avx2.h +++ b/third_party/aom/aom_dsp/x86/synonyms_avx2.h @@ -43,6 +43,16 @@ static INLINE void yy_storeu_256(void *const a, const __m256i v) { _mm256_storeu_si256((__m256i *)a, v); } +// Fill an AVX register using an interleaved pair of values, ie. set the +// 16 channels to {a, b} repeated 8 times, using the same channel ordering +// as when a register is stored to / loaded from memory. +// +// This is useful for rearranging filter kernels for use with the _mm_madd_epi16 +// instruction +static INLINE __m256i yy_set2_epi16(int16_t a, int16_t b) { + return _mm256_setr_epi16(a, b, a, b, a, b, a, b, a, b, a, b, a, b, a, b); +} + // The _mm256_set1_epi64x() intrinsic is undefined for some Visual Studio // compilers. The following function is equivalent to _mm256_set1_epi64x() // acting on a 32-bit integer. @@ -61,11 +71,26 @@ static INLINE __m256i yy_set_m128i(__m128i hi, __m128i lo) { return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1); } +#define GCC_VERSION (__GNUC__ * 10000 \ + + __GNUC_MINOR__ * 100 \ + + __GNUC_PATCHLEVEL__) + +// _mm256_loadu2_m128i has been introduced in GCC 10.1 +#if !defined(__clang__) && GCC_VERSION < 101000 +static INLINE __m256i yy_loadu2_128(const void *hi, const void *lo) { + __m128i mhi = _mm_loadu_si128((const __m128i *)(hi)); + __m128i mlo = _mm_loadu_si128((const __m128i *)(lo)); + return _mm256_set_m128i(mhi, mlo); +} +#else static INLINE __m256i yy_loadu2_128(const void *hi, const void *lo) { __m128i mhi = _mm_loadu_si128((const __m128i *)(hi)); __m128i mlo = _mm_loadu_si128((const __m128i *)(lo)); return yy_set_m128i(mhi, mlo); } +#endif + +#undef GCC_VERSION static INLINE void yy_storeu2_128(void *hi, void *lo, const __m256i a) { _mm_storeu_si128((__m128i *)hi, _mm256_extracti128_si256(a, 1)); |