diff options
Diffstat (limited to 'contrib/llvm-project/clang/lib/Headers/avx512bf16intrin.h')
| -rw-r--r-- | contrib/llvm-project/clang/lib/Headers/avx512bf16intrin.h | 44 |
1 files changed, 24 insertions, 20 deletions
diff --git a/contrib/llvm-project/clang/lib/Headers/avx512bf16intrin.h b/contrib/llvm-project/clang/lib/Headers/avx512bf16intrin.h index d1d87e72f147..b28d2e243f2c 100644 --- a/contrib/llvm-project/clang/lib/Headers/avx512bf16intrin.h +++ b/contrib/llvm-project/clang/lib/Headers/avx512bf16intrin.h @@ -10,18 +10,21 @@ #error "Never use <avx512bf16intrin.h> directly; include <immintrin.h> instead." #endif +#ifdef __SSE2__ + #ifndef __AVX512BF16INTRIN_H #define __AVX512BF16INTRIN_H -typedef short __m512bh __attribute__((__vector_size__(64), __aligned__(64))); -typedef short __m256bh __attribute__((__vector_size__(32), __aligned__(32))); -typedef unsigned short __bfloat16; +typedef __bf16 __v32bf __attribute__((__vector_size__(64), __aligned__(64))); +typedef __bf16 __m512bh __attribute__((__vector_size__(64), __aligned__(64))); +typedef __bf16 __bfloat16 __attribute__((deprecated("use __bf16 instead"))); #define __DEFAULT_FN_ATTRS512 \ - __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"), \ + __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16,evex512"), \ __min_vector_width__(512))) #define __DEFAULT_FN_ATTRS \ - __attribute__((__always_inline__, __nodebug__, __target__("avx512bf16"))) + __attribute__((__always_inline__, __nodebug__, \ + __target__("avx512bf16,no-evex512"))) /// Convert One BF16 Data to One Single Float Data. /// @@ -33,7 +36,7 @@ typedef unsigned short __bfloat16; /// A bfloat data. /// \returns A float data whose sign field and exponent field keep unchanged, /// and fraction field is extended to 23 bits. -static __inline__ float __DEFAULT_FN_ATTRS _mm_cvtsbh_ss(__bfloat16 __A) { +static __inline__ float __DEFAULT_FN_ATTRS _mm_cvtsbh_ss(__bf16 __A) { return __builtin_ia32_cvtsbf162ss_32(__A); } @@ -74,9 +77,9 @@ _mm512_cvtne2ps_pbh(__m512 __A, __m512 __B) { /// conversion of __B, and higher 256 bits come from conversion of __A. static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask_cvtne2ps_pbh(__m512bh __W, __mmask32 __U, __m512 __A, __m512 __B) { - return (__m512bh)__builtin_ia32_selectw_512((__mmask32)__U, - (__v32hi)_mm512_cvtne2ps_pbh(__A, __B), - (__v32hi)__W); + return (__m512bh)__builtin_ia32_selectpbf_512((__mmask32)__U, + (__v32bf)_mm512_cvtne2ps_pbh(__A, __B), + (__v32bf)__W); } /// Convert Two Packed Single Data to One Packed BF16 Data. @@ -96,9 +99,9 @@ _mm512_mask_cvtne2ps_pbh(__m512bh __W, __mmask32 __U, __m512 __A, __m512 __B) { /// conversion of __B, and higher 256 bits come from conversion of __A. static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtne2ps_pbh(__mmask32 __U, __m512 __A, __m512 __B) { - return (__m512bh)__builtin_ia32_selectw_512((__mmask32)__U, - (__v32hi)_mm512_cvtne2ps_pbh(__A, __B), - (__v32hi)_mm512_setzero_si512()); + return (__m512bh)__builtin_ia32_selectpbf_512((__mmask32)__U, + (__v32bf)_mm512_cvtne2ps_pbh(__A, __B), + (__v32bf)_mm512_setzero_si512()); } /// Convert Packed Single Data to Packed BF16 Data. @@ -113,7 +116,7 @@ _mm512_maskz_cvtne2ps_pbh(__mmask32 __U, __m512 __A, __m512 __B) { static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_cvtneps_pbh(__m512 __A) { return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, - (__v16hi)_mm256_undefined_si256(), + (__v16bf)_mm256_undefined_si256(), (__mmask16)-1); } @@ -134,7 +137,7 @@ _mm512_cvtneps_pbh(__m512 __A) { static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_mask_cvtneps_pbh(__m256bh __W, __mmask16 __U, __m512 __A) { return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, - (__v16hi)__W, + (__v16bf)__W, (__mmask16)__U); } @@ -153,7 +156,7 @@ _mm512_mask_cvtneps_pbh(__m256bh __W, __mmask16 __U, __m512 __A) { static __inline__ __m256bh __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtneps_pbh(__mmask16 __U, __m512 __A) { return (__m256bh)__builtin_ia32_cvtneps2bf16_512_mask((__v16sf)__A, - (__v16hi)_mm256_setzero_si256(), + (__v16bf)_mm256_setzero_si256(), (__mmask16)__U); } @@ -174,8 +177,8 @@ _mm512_maskz_cvtneps_pbh(__mmask16 __U, __m512 __A) { static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_dpbf16_ps(__m512 __D, __m512bh __A, __m512bh __B) { return (__m512)__builtin_ia32_dpbf16ps_512((__v16sf) __D, - (__v16si) __A, - (__v16si) __B); + (__v32bf) __A, + (__v32bf) __B); } /// Dot Product of BF16 Pairs Accumulated into Packed Single Precision. @@ -232,7 +235,7 @@ _mm512_maskz_dpbf16_ps(__mmask16 __U, __m512 __D, __m512bh __A, __m512bh __B) { /// /// \param __A /// A 256-bit vector of [16 x bfloat]. -/// \returns A 512-bit vector of [16 x float] come from convertion of __A +/// \returns A 512-bit vector of [16 x float] come from conversion of __A static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_cvtpbh_ps(__m256bh __A) { return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32( (__m512i)_mm512_cvtepi16_epi32((__m256i)__A), 16)); @@ -247,7 +250,7 @@ static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_cvtpbh_ps(__m256bh __A) { /// bit is not set. /// \param __A /// A 256-bit vector of [16 x bfloat]. -/// \returns A 512-bit vector of [16 x float] come from convertion of __A +/// \returns A 512-bit vector of [16 x float] come from conversion of __A static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtpbh_ps(__mmask16 __U, __m256bh __A) { return _mm512_castsi512_ps((__m512i)_mm512_slli_epi32( @@ -265,7 +268,7 @@ _mm512_maskz_cvtpbh_ps(__mmask16 __U, __m256bh __A) { /// A 16-bit mask. /// \param __A /// A 256-bit vector of [16 x bfloat]. -/// \returns A 512-bit vector of [16 x float] come from convertion of __A +/// \returns A 512-bit vector of [16 x float] come from conversion of __A static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_cvtpbh_ps(__m512 __S, __mmask16 __U, __m256bh __A) { return _mm512_castsi512_ps((__m512i)_mm512_mask_slli_epi32( @@ -277,3 +280,4 @@ _mm512_mask_cvtpbh_ps(__m512 __S, __mmask16 __U, __m256bh __A) { #undef __DEFAULT_FN_ATTRS512 #endif +#endif |