diff options
author | Michael Zuckerman <Michael.zuckerman@intel.com> | 2016-04-14 06:48:09 +0000 |
---|---|---|
committer | Michael Zuckerman <Michael.zuckerman@intel.com> | 2016-04-14 06:48:09 +0000 |
commit | d871531687b062862234a3346b50f1824e27ed3b (patch) | |
tree | e9570d4d503fbe4783f757efa02d8f142571ccd7 /clang/lib | |
parent | 3d7224c2a7f9a5e78e3c2a7a890547791d6ef8e2 (diff) | |
download | bcm5719-llvm-d871531687b062862234a3346b50f1824e27ed3b.tar.gz bcm5719-llvm-d871531687b062862234a3346b50f1824e27ed3b.zip |
[Clang][AVX512][Builtin] Adding intrinsics of vpmovus{d|q}{b|w|d}{128|256|512} instruction set
Differential Revision: http://reviews.llvm.org/D19050
llvm-svn: 266278
Diffstat (limited to 'clang/lib')
-rw-r--r-- | clang/lib/Headers/avx512fintrin.h | 148 | ||||
-rw-r--r-- | clang/lib/Headers/avx512vlintrin.h | 294 |
2 files changed, 442 insertions, 0 deletions
diff --git a/clang/lib/Headers/avx512fintrin.h b/clang/lib/Headers/avx512fintrin.h index 38d49af9ae8..2699127bc67 100644 --- a/clang/lib/Headers/avx512fintrin.h +++ b/clang/lib/Headers/avx512fintrin.h @@ -5765,6 +5765,154 @@ _mm512_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m512i __A) __builtin_ia32_pmovsqw512mem_mask ((__v8hi *) __P, (__v8di) __A, __M); } +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_cvtusepi32_epi8 (__m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb512_mask ((__v16si) __A, + (__v16qi) _mm_undefined_si128 (), + (__mmask16) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi32_epi8 (__m128i __O, __mmask16 __M, __m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb512_mask ((__v16si) __A, + (__v16qi) __O, + __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_maskz_cvtusepi32_epi8 (__mmask16 __M, __m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb512_mask ((__v16si) __A, + (__v16qi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask16 __M, __m512i __A) +{ + __builtin_ia32_pmovusdb512mem_mask ((__v16qi *) __P, (__v16si) __A, __M); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_cvtusepi32_epi16 (__m512i __A) +{ + return (__m256i) __builtin_ia32_pmovusdw512_mask ((__v16si) __A, + (__v16hi) _mm256_undefined_si256 (), + (__mmask16) -1); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi32_epi16 (__m256i __O, __mmask16 __M, __m512i __A) +{ + return (__m256i) __builtin_ia32_pmovusdw512_mask ((__v16si) __A, + (__v16hi) __O, + __M); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_maskz_cvtusepi32_epi16 (__mmask16 __M, __m512i __A) +{ + return (__m256i) __builtin_ia32_pmovusdw512_mask ((__v16si) __A, + (__v16hi) _mm256_setzero_si256 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi32_storeu_epi16 (void *__P, __mmask16 __M, __m512i __A) +{ + __builtin_ia32_pmovusdw512mem_mask ((__v16hi*) __P, (__v16si) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_cvtusepi64_epi8 (__m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb512_mask ((__v8di) __A, + (__v16qi) _mm_undefined_si128 (), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb512_mask ((__v8di) __A, + (__v16qi) __O, + __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_maskz_cvtusepi64_epi8 (__mmask8 __M, __m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb512_mask ((__v8di) __A, + (__v16qi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m512i __A) +{ + __builtin_ia32_pmovusqb512mem_mask ((__v16qi *) __P, (__v8di) __A, __M); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_cvtusepi64_epi32 (__m512i __A) +{ + return (__m256i) __builtin_ia32_pmovusqd512_mask ((__v8di) __A, + (__v8si) _mm256_undefined_si256 (), + (__mmask8) -1); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi64_epi32 (__m256i __O, __mmask8 __M, __m512i __A) +{ + return (__m256i) __builtin_ia32_pmovusqd512_mask ((__v8di) __A, + (__v8si) __O, __M); +} + +static __inline__ __m256i __DEFAULT_FN_ATTRS +_mm512_maskz_cvtusepi64_epi32 (__mmask8 __M, __m512i __A) +{ + return (__m256i) __builtin_ia32_pmovusqd512_mask ((__v8di) __A, + (__v8si) _mm256_setzero_si256 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi64_storeu_epi32 (void* __P, __mmask8 __M, __m512i __A) +{ + __builtin_ia32_pmovusqd512mem_mask ((__v8si*) __P, (__v8di) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_cvtusepi64_epi16 (__m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw512_mask ((__v8di) __A, + (__v8hi) _mm_undefined_si128 (), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw512_mask ((__v8di) __A, + (__v8hi) __O, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm512_maskz_cvtusepi64_epi16 (__mmask8 __M, __m512i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw512_mask ((__v8di) __A, + (__v8hi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm512_mask_cvtusepi64_storeu_epi16 (void *__P, __mmask8 __M, __m512i __A) +{ + __builtin_ia32_pmovusqw512mem_mask ((__v8hi*) __P, (__v8di) __A, __M); +} + #undef __DEFAULT_FN_ATTRS #endif // __AVX512FINTRIN_H diff --git a/clang/lib/Headers/avx512vlintrin.h b/clang/lib/Headers/avx512vlintrin.h index 51eadfb156e..df00884fec3 100644 --- a/clang/lib/Headers/avx512vlintrin.h +++ b/clang/lib/Headers/avx512vlintrin.h @@ -8218,6 +8218,300 @@ _mm256_mask_cvtsepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) __builtin_ia32_pmovsqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); } +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_cvtusepi32_epi8 (__m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, + (__v16qi)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, + (__v16qi) __O, + __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_maskz_cvtusepi32_epi8 (__mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb128_mask ((__v4si) __A, + (__v16qi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) +{ + __builtin_ia32_pmovusdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_cvtusepi32_epi8 (__m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, + (__v16qi)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi32_epi8 (__m128i __O, __mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, + (__v16qi) __O, + __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_maskz_cvtusepi32_epi8 (__mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusdb256_mask ((__v8si) __A, + (__v16qi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) +{ + __builtin_ia32_pmovusdb256mem_mask ((__v16qi*) __P, (__v8si) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_cvtusepi32_epi16 (__m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, + (__v8hi)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, + (__v8hi) __O, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_maskz_cvtusepi32_epi16 (__mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusdw128_mask ((__v4si) __A, + (__v8hi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) +{ + __builtin_ia32_pmovusdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_cvtusepi32_epi16 (__m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, + (__v8hi) _mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi32_epi16 (__m128i __O, __mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, + (__v8hi) __O, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_maskz_cvtusepi32_epi16 (__mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusdw256_mask ((__v8si) __A, + (__v8hi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) +{ + __builtin_ia32_pmovusdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_cvtusepi64_epi8 (__m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, + (__v16qi)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, + (__v16qi) __O, + __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_maskz_cvtusepi64_epi8 (__mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb128_mask ((__v2di) __A, + (__v16qi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) +{ + __builtin_ia32_pmovusqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_cvtusepi64_epi8 (__m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, + (__v16qi)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi64_epi8 (__m128i __O, __mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, + (__v16qi) __O, + __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_maskz_cvtusepi64_epi8 (__mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqb256_mask ((__v4di) __A, + (__v16qi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A) +{ + __builtin_ia32_pmovusqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_cvtusepi64_epi32 (__m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, + (__v4si)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, + (__v4si) __O, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_maskz_cvtusepi64_epi32 (__mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqd128_mask ((__v2di) __A, + (__v4si) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A) +{ + __builtin_ia32_pmovusqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_cvtusepi64_epi32 (__m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, + (__v4si)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi64_epi32 (__m128i __O, __mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, + (__v4si) __O, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_maskz_cvtusepi64_epi32 (__mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqd256_mask ((__v4di) __A, + (__v4si) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A) +{ + __builtin_ia32_pmovusqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_cvtusepi64_epi16 (__m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, + (__v8hi)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, + (__v8hi) __O, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm_maskz_cvtusepi64_epi16 (__mmask8 __M, __m128i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw128_mask ((__v2di) __A, + (__v8hi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A) +{ + __builtin_ia32_pmovusqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_cvtusepi64_epi16 (__m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, + (__v8hi)_mm_undefined_si128(), + (__mmask8) -1); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi64_epi16 (__m128i __O, __mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, + (__v8hi) __O, __M); +} + +static __inline__ __m128i __DEFAULT_FN_ATTRS +_mm256_maskz_cvtusepi64_epi16 (__mmask8 __M, __m256i __A) +{ + return (__m128i) __builtin_ia32_pmovusqw256_mask ((__v4di) __A, + (__v8hi) _mm_setzero_si128 (), + __M); +} + +static __inline__ void __DEFAULT_FN_ATTRS +_mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A) +{ + return __builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M); +} + #undef __DEFAULT_FN_ATTRS #undef __DEFAULT_FN_ATTRS_BOTH |