diff options
| author | Michael Zuckerman <Michael.zuckerman@intel.com> | 2016-05-10 07:30:58 +0000 |
|---|---|---|
| committer | Michael Zuckerman <Michael.zuckerman@intel.com> | 2016-05-10 07:30:58 +0000 |
| commit | 7360d8a9cc771dbc030b300f478c92eb5b5e8637 (patch) | |
| tree | 5b5a6bbee3ea463496ddd8ab8b500d262c95a20a /clang/lib | |
| parent | 15d1e64b2b1cdece177102074f03ee161fed91fb (diff) | |
| download | bcm5719-llvm-7360d8a9cc771dbc030b300f478c92eb5b5e8637.tar.gz bcm5719-llvm-7360d8a9cc771dbc030b300f478c92eb5b5e8637.zip | |
[Clang][AVX512] completing missing intrinsics [roundscale, ceil, floor]
Differential Revision: http://reviews.llvm.org/D20070
llvm-svn: 269022
Diffstat (limited to 'clang/lib')
| -rw-r--r-- | clang/lib/Headers/avx512fintrin.h | 62 |
1 files changed, 62 insertions, 0 deletions
diff --git a/clang/lib/Headers/avx512fintrin.h b/clang/lib/Headers/avx512fintrin.h index 5a09b0141d9..6107790852a 100644 --- a/clang/lib/Headers/avx512fintrin.h +++ b/clang/lib/Headers/avx512fintrin.h @@ -1307,6 +1307,16 @@ _mm512_floor_ps(__m512 __A) _MM_FROUND_CUR_DIRECTION); } +static __inline__ __m512 __DEFAULT_FN_ATTRS +_mm512_mask_floor_ps (__m512 __W, __mmask16 __U, __m512 __A) +{ + return (__m512) __builtin_ia32_rndscaleps_mask ((__v16sf) __A, + _MM_FROUND_FLOOR, + (__v16sf) __W, __U, + _MM_FROUND_CUR_DIRECTION); +} + + static __inline __m512d __DEFAULT_FN_ATTRS _mm512_floor_pd(__m512d __A) { @@ -1316,6 +1326,25 @@ _mm512_floor_pd(__m512d __A) _MM_FROUND_CUR_DIRECTION); } +static __inline__ __m512d __DEFAULT_FN_ATTRS +_mm512_mask_floor_pd (__m512d __W, __mmask8 __U, __m512d __A) +{ + return (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __A, + _MM_FROUND_FLOOR, + (__v8df) __W, __U, + _MM_FROUND_CUR_DIRECTION); +} + + +static __inline__ __m512 __DEFAULT_FN_ATTRS +_mm512_mask_ceil_ps (__m512 __W, __mmask16 __U, __m512 __A) +{ + return (__m512) __builtin_ia32_rndscaleps_mask ((__v16sf) __A, + _MM_FROUND_CEIL, + (__v16sf) __W, __U, + _MM_FROUND_CUR_DIRECTION); +} + static __inline __m512 __DEFAULT_FN_ATTRS _mm512_ceil_ps(__m512 __A) { @@ -1334,6 +1363,15 @@ _mm512_ceil_pd(__m512d __A) _MM_FROUND_CUR_DIRECTION); } +static __inline__ __m512d __DEFAULT_FN_ATTRS +_mm512_mask_ceil_pd (__m512d __W, __mmask8 __U, __m512d __A) +{ + return (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __A, + _MM_FROUND_CEIL, + (__v8df) __W, __U, + _MM_FROUND_CUR_DIRECTION); +} + static __inline __m512i __DEFAULT_FN_ATTRS _mm512_abs_epi64(__m512i __A) { @@ -1839,10 +1877,34 @@ _mm512_maskz_div_ps(__mmask16 __U, __m512 __A, __m512 __B) { (__m512)__builtin_ia32_rndscaleps_mask((__v16sf)(A), (B), (__v16sf)(A), \ -1, _MM_FROUND_CUR_DIRECTION); }) +#define _mm512_mask_roundscale_ps( __A, __B, __C, __imm) __extension__ ({\ + (__m512) __builtin_ia32_rndscaleps_mask ((__v16sf) __C, __imm,\ + (__v16sf) __A,(__mmask16) __B,\ + _MM_FROUND_CUR_DIRECTION);\ +}) + +#define _mm512_maskz_roundscale_ps( __A, __B, __imm) __extension__ ({\ + (__m512) __builtin_ia32_rndscaleps_mask ((__v16sf) __B, __imm,\ + (__v16sf) _mm512_setzero_ps (),\ + (__mmask16) __A, _MM_FROUND_CUR_DIRECTION);\ +}) + #define _mm512_roundscale_pd(A, B) __extension__ ({ \ (__m512d)__builtin_ia32_rndscalepd_mask((__v8df)(A), (B), (__v8df)(A), \ -1, _MM_FROUND_CUR_DIRECTION); }) +#define _mm512_mask_roundscale_pd(__A, __B, __C, __imm) __extension__ ({\ + (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __C, __imm,\ + (__v8df) __A, (__mmask8) __B,\ + _MM_FROUND_CUR_DIRECTION);\ +}) + +#define _mm512_maskz_roundscale_pd(__A, __B, __imm) __extension__ ({\ + (__m512d) __builtin_ia32_rndscalepd_mask ((__v8df) __B, __imm,\ + (__v8df) _mm512_setzero_pd (),\ + (__mmask8) __A, _MM_FROUND_CUR_DIRECTION);\ +}) + #define _mm512_fmadd_round_pd(A, B, C, R) __extension__ ({ \ (__m512d) __builtin_ia32_vfmaddpd512_mask ((__v8df) (A), \ (__v8df) (B), (__v8df) (C), \ |
