summaryrefslogtreecommitdiffstats
path: root/contrib/llvm/tools/clang/lib/Headers/avx512erintrin.h
diff options
context:
space:
mode:
Diffstat (limited to 'contrib/llvm/tools/clang/lib/Headers/avx512erintrin.h')
-rw-r--r--contrib/llvm/tools/clang/lib/Headers/avx512erintrin.h326
1 files changed, 250 insertions, 76 deletions
diff --git a/contrib/llvm/tools/clang/lib/Headers/avx512erintrin.h b/contrib/llvm/tools/clang/lib/Headers/avx512erintrin.h
index 1a5ea15..57c61aa 100644
--- a/contrib/llvm/tools/clang/lib/Headers/avx512erintrin.h
+++ b/contrib/llvm/tools/clang/lib/Headers/avx512erintrin.h
@@ -28,85 +28,259 @@
#define __AVX512ERINTRIN_H
+// exp2a23
+#define _mm512_exp2a23_round_pd(A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_exp2pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)_mm512_setzero_pd(), \
+ (__mmask8)-1, (R)); })
+
+#define _mm512_mask_exp2a23_round_pd(S, M, A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_exp2pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)(__m512d)(S), \
+ (__mmask8)(M), (R)); })
+
+#define _mm512_maskz_exp2a23_round_pd(M, A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_exp2pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)_mm512_setzero_pd(), \
+ (__mmask8)(M), (R)); })
+
+#define _mm512_exp2a23_pd(A) \
+ _mm512_exp2a23_round_pd((A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_mask_exp2a23_pd(S, M, A) \
+ _mm512_mask_exp2a23_round_pd((S), (M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_maskz_exp2a23_pd(M, A) \
+ _mm512_maskz_exp2a23_round_pd((M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_exp2a23_round_ps(A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_exp2ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)_mm512_setzero_ps(), \
+ (__mmask8)-1, (R)); })
+
+#define _mm512_mask_exp2a23_round_ps(S, M, A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_exp2ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)(__m512)(S), \
+ (__mmask8)(M), (R)); })
+
+#define _mm512_maskz_exp2a23_round_ps(M, A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_exp2ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)_mm512_setzero_ps(), \
+ (__mmask8)(M), (R)); })
+
+#define _mm512_exp2a23_ps(A) \
+ _mm512_exp2a23_round_ps((A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_mask_exp2a23_ps(S, M, A) \
+ _mm512_mask_exp2a23_round_ps((S), (M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_maskz_exp2a23_ps(M, A) \
+ _mm512_maskz_exp2a23_round_ps((M), (A), _MM_FROUND_CUR_DIRECTION)
+
// rsqrt28
-static __inline__ __m512d __attribute__((__always_inline__, __nodebug__))
-_mm512_rsqrt28_round_pd (__m512d __A, int __R)
-{
- return (__m512d)__builtin_ia32_rsqrt28pd_mask ((__v8df)__A,
- (__v8df)_mm512_setzero_pd(),
- (__mmask8)-1,
- __R);
-}
-static __inline__ __m512 __attribute__((__always_inline__, __nodebug__))
-_mm512_rsqrt28_round_ps(__m512 __A, int __R)
-{
- return (__m512)__builtin_ia32_rsqrt28ps_mask ((__v16sf)__A,
- (__v16sf)_mm512_setzero_ps(),
- (__mmask16)-1,
- __R);
-}
-
-static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
-_mm_rsqrt28_round_ss(__m128 __A, __m128 __B, int __R)
-{
- return (__m128) __builtin_ia32_rsqrt28ss_mask ((__v4sf) __A,
- (__v4sf) __B,
- (__v4sf)
- _mm_setzero_ps (),
- (__mmask8) -1,
- __R);
-}
-
-static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
-_mm_rsqrt28_round_sd (__m128d __A, __m128d __B, int __R)
-{
- return (__m128d) __builtin_ia32_rsqrt28sd_mask ((__v2df) __A,
- (__v2df) __B,
- (__v2df)
- _mm_setzero_pd (),
- (__mmask8) -1,
- __R);
-}
+#define _mm512_rsqrt28_round_pd(A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_rsqrt28pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)_mm512_setzero_pd(), \
+ (__mmask8)-1, (R)); })
+
+#define _mm512_mask_rsqrt28_round_pd(S, M, A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_rsqrt28pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)(__m512d)(S), \
+ (__mmask8)(M), (R)); })
+
+#define _mm512_maskz_rsqrt28_round_pd(M, A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_rsqrt28pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)_mm512_setzero_pd(), \
+ (__mmask8)(M), (R)); })
+
+#define _mm512_rsqrt28_pd(A) \
+ _mm512_rsqrt28_round_pd((A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_mask_rsqrt28_pd(S, M, A) \
+ _mm512_mask_rsqrt28_round_pd((S), (M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_maskz_rsqrt28_pd(M, A) \
+ _mm512_maskz_rsqrt28_round_pd((M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_rsqrt28_round_ps(A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_rsqrt28ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)_mm512_setzero_ps(), \
+ (__mmask16)-1, (R)); })
+
+#define _mm512_mask_rsqrt28_round_ps(S, M, A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_rsqrt28ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)(__m512)(S), \
+ (__mmask16)(M), (R)); })
+
+#define _mm512_maskz_rsqrt28_round_ps(M, A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_rsqrt28ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)_mm512_setzero_ps(), \
+ (__mmask16)(M), (R)); })
+
+#define _mm512_rsqrt28_ps(A) \
+ _mm512_rsqrt28_round_ps((A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_mask_rsqrt28_ps(S, M, A) \
+ _mm512_mask_rsqrt28_round_ps((S), (M), A, _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_maskz_rsqrt28_ps(M, A) \
+ _mm512_maskz_rsqrt28_round_ps((M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_rsqrt28_round_ss(A, B, R) __extension__ ({ \
+ (__m128)__builtin_ia32_rsqrt28ss_mask((__v4sf)(__m128)(A), \
+ (__v4sf)(__m128)(B), \
+ (__v4sf)_mm_setzero_ps(), \
+ (__mmask8)-1, (R)); })
+
+#define _mm_mask_rsqrt28_round_ss(S, M, A, B, R) __extension__ ({ \
+ (__m128)__builtin_ia32_rsqrt28ss_mask((__v4sf)(__m128)(A), \
+ (__v4sf)(__m128)(B), \
+ (__v4sf)(__m128)(S), \
+ (__mmask8)(M), (R)); })
+
+#define _mm_maskz_rsqrt28_round_ss(M, A, B, R) __extension__ ({ \
+ (__m128)__builtin_ia32_rsqrt28ss_mask((__v4sf)(__m128)(A), \
+ (__v4sf)(__m128)(B), \
+ (__v4sf)_mm_setzero_ps(), \
+ (__mmask8)(M), (R)); })
+
+#define _mm_rsqrt28_ss(A, B) \
+ _mm_rsqrt28_round_ss((A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_mask_rsqrt28_ss(S, M, A, B) \
+ _mm_mask_rsqrt28_round_ss((S), (M), (A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_maskz_rsqrt28_ss(M, A, B) \
+ _mm_maskz_rsqrt28_round_ss((M), (A), (B), _MM_FROUND_CUR_DIRECTION)
+#define _mm_rsqrt28_round_sd(A, B, R) __extension__ ({ \
+ (__m128d)__builtin_ia32_rsqrt28sd_mask((__v2df)(__m128d)(A), \
+ (__v2df)(__m128d)(B), \
+ (__v2df)_mm_setzero_pd(), \
+ (__mmask8)-1, (R)); })
+
+#define _mm_mask_rsqrt28_round_sd(S, M, A, B, R) __extension__ ({ \
+ (__m128d)__builtin_ia32_rsqrt28sd_mask((__v2df)(__m128d)(A), \
+ (__v2df)(__m128d)(B), \
+ (__v2df)(__m128d)(S), \
+ (__mmask8)(M), (R)); })
+
+#define _mm_maskz_rsqrt28_round_sd(M, A, B, R) __extension__ ({ \
+ (__m128d)__builtin_ia32_rsqrt28sd_mask((__v2df)(__m128d)(A), \
+ (__v2df)(__m128d)(B), \
+ (__v2df)_mm_setzero_pd(), \
+ (__mmask8)(M), (R)); })
+
+#define _mm_rsqrt28_sd(A, B) \
+ _mm_rsqrt28_round_sd((A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_mask_rsqrt28_sd(S, M, A, B) \
+ _mm_mask_rsqrt28_round_sd((S), (M), (A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_maskz_rsqrt28_sd(M, A, B) \
+ _mm_mask_rsqrt28_round_sd((M), (A), (B), _MM_FROUND_CUR_DIRECTION)
// rcp28
-static __inline__ __m512d __attribute__((__always_inline__, __nodebug__))
-_mm512_rcp28_round_pd (__m512d __A, int __R)
-{
- return (__m512d)__builtin_ia32_rcp28pd_mask ((__v8df)__A,
- (__v8df)_mm512_setzero_pd(),
- (__mmask8)-1,
- __R);
-}
-
-static __inline__ __m512 __attribute__((__always_inline__, __nodebug__))
-_mm512_rcp28_round_ps (__m512 __A, int __R)
-{
- return (__m512)__builtin_ia32_rcp28ps_mask ((__v16sf)__A,
- (__v16sf)_mm512_setzero_ps (),
- (__mmask16)-1,
- __R);
-}
-
-static __inline__ __m128 __attribute__((__always_inline__, __nodebug__))
-_mm_rcp28_round_ss (__m128 __A, __m128 __B, int __R)
-{
- return (__m128) __builtin_ia32_rcp28ss_mask ((__v4sf) __A,
- (__v4sf) __B,
- (__v4sf)
- _mm_setzero_ps (),
- (__mmask8) -1,
- __R);
-}
-static __inline__ __m128d __attribute__((__always_inline__, __nodebug__))
-_mm_rcp28_round_sd (__m128d __A, __m128d __B, int __R)
-{
- return (__m128d) __builtin_ia32_rcp28sd_mask ((__v2df) __A,
- (__v2df) __B,
- (__v2df)
- _mm_setzero_pd (),
- (__mmask8) -1,
- __R);
-}
+#define _mm512_rcp28_round_pd(A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_rcp28pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)_mm512_setzero_pd(), \
+ (__mmask8)-1, (R)); })
+
+#define _mm512_mask_rcp28_round_pd(S, M, A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_rcp28pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)(__m512d)(S), \
+ (__mmask8)(M), (R)); })
+
+#define _mm512_maskz_rcp28_round_pd(M, A, R) __extension__ ({ \
+ (__m512d)__builtin_ia32_rcp28pd_mask((__v8df)(__m512d)(A), \
+ (__v8df)_mm512_setzero_pd(), \
+ (__mmask8)(M), (R)); })
+
+#define _mm512_rcp28_pd(A) \
+ _mm512_rcp28_round_pd((A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_mask_rcp28_pd(S, M, A) \
+ _mm512_mask_rcp28_round_pd((S), (M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_maskz_rcp28_pd(M, A) \
+ _mm512_maskz_rcp28_round_pd((M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_rcp28_round_ps(A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_rcp28ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)_mm512_setzero_ps(), \
+ (__mmask16)-1, (R)); })
+
+#define _mm512_mask_rcp28_round_ps(S, M, A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_rcp28ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)(__m512)(S), \
+ (__mmask16)(M), (R)); })
+
+#define _mm512_maskz_rcp28_round_ps(M, A, R) __extension__ ({ \
+ (__m512)__builtin_ia32_rcp28ps_mask((__v16sf)(__m512)(A), \
+ (__v16sf)_mm512_setzero_ps(), \
+ (__mmask16)(M), (R)); })
+
+#define _mm512_rcp28_ps(A) \
+ _mm512_rcp28_round_ps((A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_mask_rcp28_ps(S, M, A) \
+ _mm512_mask_rcp28_round_ps((S), (M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm512_maskz_rcp28_ps(M, A) \
+ _mm512_maskz_rcp28_round_ps((M), (A), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_rcp28_round_ss(A, B, R) __extension__ ({ \
+ (__m128)__builtin_ia32_rcp28ss_mask((__v4sf)(__m128)(A), \
+ (__v4sf)(__m128)(B), \
+ (__v4sf)_mm_setzero_ps(), \
+ (__mmask8)-1, (R)); })
+
+#define _mm_mask_rcp28_round_ss(S, M, A, B, R) __extension__ ({ \
+ (__m128)__builtin_ia32_rcp28ss_mask((__v4sf)(__m128)(A), \
+ (__v4sf)(__m128)(B), \
+ (__v4sf)(__m128)(S), \
+ (__mmask8)(M), (R)); })
+
+#define _mm_maskz_rcp28_round_ss(M, A, B, R) __extension__ ({ \
+ (__m128)__builtin_ia32_rcp28ss_mask((__v4sf)(__m128)(A), \
+ (__v4sf)(__m128)(B), \
+ (__v4sf)_mm_setzero_ps(), \
+ (__mmask8)(M), (R)); })
+
+#define _mm_rcp28_ss(A, B) \
+ _mm_rcp28_round_ss((A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_mask_rcp28_ss(S, M, A, B) \
+ _mm_mask_rcp28_round_ss((S), (M), (A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_maskz_rcp28_ss(M, A, B) \
+ _mm_maskz_rcp28_round_ss((M), (A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_rcp28_round_sd(A, B, R) __extension__ ({ \
+ (__m128d)__builtin_ia32_rcp28sd_mask((__v2df)(__m128d)(A), \
+ (__v2df)(__m128d)(B), \
+ (__v2df)_mm_setzero_pd(), \
+ (__mmask8)-1, (R)); })
+
+#define _mm_mask_rcp28_round_sd(S, M, A, B, R) __extension__ ({ \
+ (__m128d)__builtin_ia32_rcp28sd_mask((__v2df)(__m128d)(A), \
+ (__v2df)(__m128d)(B), \
+ (__v2df)(__m128d)(S), \
+ (__mmask8)(M), (R)); })
+
+#define _mm_maskz_rcp28_round_sd(M, A, B, R) __extension__ ({ \
+ (__m128d)__builtin_ia32_rcp28sd_mask((__v2df)(__m128d)(A), \
+ (__v2df)(__m128d)(B), \
+ (__v2df)_mm_setzero_pd(), \
+ (__mmask8)(M), (R)); })
+
+#define _mm_rcp28_sd(A, B) \
+ _mm_rcp28_round_sd((A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_mask_rcp28_sd(S, M, A, B) \
+ _mm_mask_rcp28_round_sd((S), (M), (A), (B), _MM_FROUND_CUR_DIRECTION)
+
+#define _mm_maskz_rcp28_sd(M, A, B) \
+ _mm_maskz_rcp28_round_sd((M), (A), (B), _MM_FROUND_CUR_DIRECTION)
#endif // __AVX512ERINTRIN_H
OpenPOWER on IntegriCloud