diff options
Diffstat (limited to 'intgemm/intrinsics.h')
| -rw-r--r-- | intgemm/intrinsics.h | 13 |
1 files changed, 13 insertions, 0 deletions
diff --git a/intgemm/intrinsics.h b/intgemm/intrinsics.h index 67b36fc..9f370cd 100644 --- a/intgemm/intrinsics.h +++ b/intgemm/intrinsics.h @@ -9,6 +9,9 @@ #ifdef INTGEMM_COMPILER_SUPPORTS_AVX2 #include <immintrin.h> #endif +#ifdef INTGEMM_WORMHOLE +#include <wasm_simd128.h> +#endif #include <cstdint> @@ -92,10 +95,20 @@ template <> INTGEMM_SSE2 inline __m128 loadu_ps(const float* mem_addr) { return _mm_loadu_ps(mem_addr); } INTGEMM_SSE2 static inline __m128i madd_epi16(__m128i first, __m128i second) { +// https://bugzilla.mozilla.org/show_bug.cgi?id=1672160 +#ifdef INTGEMM_WORMHOLE + return wasm_v8x16_shuffle(first, second, 31, 0, 30, 2, 29, 4, 28, 6, 27, 8, 26, 10, 25, 12, 24, 2 /* PMADDWD */); +#else return _mm_madd_epi16(first, second); +#endif } INTGEMM_SSSE3 static inline __m128i maddubs_epi16(__m128i first, __m128i second) { +// https://bugzilla.mozilla.org/show_bug.cgi?id=1672160 +#ifdef INTGEMM_WORMHOLE + return wasm_v8x16_shuffle(first, second, 31, 0, 30, 2, 29, 4, 28, 6, 27, 8, 26, 10, 25, 12, 24, 1 /* PMADDUBSW */); +#else return _mm_maddubs_epi16(first, second); +#endif } /* * Missing max_epi8 for SSE2 |