On Fri, Apr 7, 2017 at 4:52 PM, Jakub Jelinek <[email protected]> wrote:
> Hi!
>
> This patch is slightly larger, so I haven't included it in the patch I've
> sent a few minutes ago.
>
> I've looked at godbolt for what ICC generates for these and picked sequences
> that generate approx. as good code as that. For
> min_epi64/max_epi64/min_epu64/max_epu64 there is a slight complication that
> in AVX512F there is only _mm512_{min,max}_ep{i,u}64 but not the _mm256_ or
> _mm_ ones, so we need to perform 512-bit operations all the time rather than
> perform extractions, 256-bit operation, further extractions and then 128-bit
> operations.
>
> Seems we need to teach our permutation code further instructions, e.g.
> typedef long long V __attribute__((vector_size (64)));
> typedef int W __attribute__((vector_size (64)));
> W f0 (W x) {
> return __builtin_shuffle (x, (W) { 8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2,
> 3, 4, 5, 6, 7 });
> }
> V f1 (V x) {
> return __builtin_shuffle (x, (V) { 4, 5, 6, 7, 0, 1, 2, 3 });
> }
> generate unnecessarily bad code (could use vpshufi64x2 instruction),
> guess that can be resolved for GCC8.
>
> Tested with
> make -j272 -k check-gcc RUNTESTFLAGS='--target_board=unix\{-m32,-m64\}
> i386.exp'
> on KNL, will bootstrap/regtest on my Haswell-E next, ok for trunk
> if that passes?
>
> It is not a regression, on the other side it really shouldn't affect any
> code that is not using those intrinsics.
>
> 2017-04-07 Jakub Jelinek <[email protected]>
>
> PR target/80324
> * config/i386/avx512fintrin.h (_mm512_reduce_add_epi32,
> _mm512_reduce_mul_epi32, _mm512_reduce_and_epi32,
> _mm512_reduce_or_epi32, _mm512_mask_reduce_add_epi32,
> _mm512_mask_reduce_mul_epi32, _mm512_mask_reduce_and_epi32,
> _mm512_mask_reduce_or_epi32, _mm512_reduce_min_epi32,
> _mm512_reduce_max_epi32, _mm512_reduce_min_epu32,
> _mm512_reduce_max_epu32, _mm512_mask_reduce_min_epi32,
> _mm512_mask_reduce_max_epi32, _mm512_mask_reduce_min_epu32,
> _mm512_mask_reduce_max_epu32, _mm512_reduce_add_ps,
> _mm512_reduce_mul_ps, _mm512_mask_reduce_add_ps,
> _mm512_mask_reduce_mul_ps, _mm512_reduce_min_ps, _mm512_reduce_max_ps,
> _mm512_mask_reduce_min_ps, _mm512_mask_reduce_max_ps,
> _mm512_reduce_add_epi64, _mm512_reduce_mul_epi64,
> _mm512_reduce_and_epi64, _mm512_reduce_or_epi64,
> _mm512_mask_reduce_add_epi64, _mm512_mask_reduce_mul_epi64,
> _mm512_mask_reduce_and_epi64, _mm512_mask_reduce_or_epi64,
> _mm512_reduce_min_epi64, _mm512_reduce_max_epi64,
> _mm512_mask_reduce_min_epi64, _mm512_mask_reduce_max_epi64,
> _mm512_reduce_min_epu64, _mm512_reduce_max_epu64,
> _mm512_mask_reduce_min_epu64, _mm512_mask_reduce_max_epu64,
> _mm512_reduce_add_pd, _mm512_reduce_mul_pd, _mm512_mask_reduce_add_pd,
> _mm512_mask_reduce_mul_pd, _mm512_reduce_min_pd, _mm512_reduce_max_pd,
> _mm512_mask_reduce_min_pd, _mm512_mask_reduce_max_pd): New intrinsics.
>
> * gcc.target/i386/avx512f-reduce-op-1.c: New test.
LGTM, but please wait for Kirill's opinion on the implementation.
Thanks,
Uros.
> --- gcc/config/i386/avx512fintrin.h.jj 2017-04-07 12:25:13.065643755 +0200
> +++ gcc/config/i386/avx512fintrin.h 2017-04-07 16:34:37.976974227 +0200
> @@ -13282,6 +13282,470 @@ _mm512_cmpgt_epu64_mask (__m512i __A, __
> (__mmask8) -1);
> }
>
> +#undef __MM512_REDUCE_OP
> +#define __MM512_REDUCE_OP(op) \
> + __v8si __T1 = (__v8si) _mm512_extracti64x4_epi64 (__A, 1); \
> + __v8si __T2 = (__v8si) _mm512_extracti64x4_epi64 (__A, 0); \
> + __m256i __T3 = (__m256i) (__T1 op __T2); \
> + __v4si __T4 = (__v4si) _mm256_extracti128_si256 (__T3, 1); \
> + __v4si __T5 = (__v4si) _mm256_extracti128_si256 (__T3, 0); \
> + __v4si __T6 = __T4 op __T5; \
> + __v4si __T7 = __builtin_shuffle (__T6, (__v4si) { 2, 3, 0, 1 }); \
> + __v4si __T8 = __T6 op __T7; \
> + return __T8[0] op __T8[1]
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_add_epi32 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (+);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_mul_epi32 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (*);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_and_epi32 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (&);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_or_epi32 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (|);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_add_epi32 (__mmask16 __U, __m512i __A)
> +{
> + __A = _mm512_maskz_mov_epi32 (__U, __A);
> + __MM512_REDUCE_OP (+);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_mul_epi32 (__mmask16 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi32 (_mm512_set1_epi32 (1), __U, __A);
> + __MM512_REDUCE_OP (*);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_and_epi32 (__mmask16 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi32 (_mm512_set1_epi32 (~0), __U, __A);
> + __MM512_REDUCE_OP (&);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_or_epi32 (__mmask16 __U, __m512i __A)
> +{
> + __A = _mm512_maskz_mov_epi32 (__U, __A);
> + __MM512_REDUCE_OP (|);
> +}
> +
> +#undef __MM512_REDUCE_OP
> +#define __MM512_REDUCE_OP(op) \
> + __m256i __T1 = (__m256i) _mm512_extracti64x4_epi64 (__A, 1); \
> + __m256i __T2 = (__m256i) _mm512_extracti64x4_epi64 (__A, 0); \
> + __m256i __T3 = _mm256_##op (__T1, __T2); \
> + __m128i __T4 = (__m128i) _mm256_extracti128_si256 (__T3, 1); \
> + __m128i __T5 = (__m128i) _mm256_extracti128_si256 (__T3, 0); \
> + __m128i __T6 = _mm_##op (__T4, __T5);
> \
> + __m128i __T7 = (__m128i) __builtin_shuffle ((__v4si) __T6, \
> + (__v4si) { 2, 3, 0, 1 }); \
> + __m128i __T8 = _mm_##op (__T6, __T7);
> \
> + __m128i __T9 = (__m128i) __builtin_shuffle ((__v4si) __T8, \
> + (__v4si) { 1, 0, 1, 0 }); \
> + __v4si __T10 = (__v4si) _mm_##op (__T8, __T9); \
> + return __T10[0]
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_min_epi32 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (min_epi32);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_max_epi32 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (max_epi32);
> +}
> +
> +extern __inline unsigned int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_min_epu32 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (min_epu32);
> +}
> +
> +extern __inline unsigned int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_max_epu32 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (max_epu32);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_min_epi32 (__mmask16 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi32 (_mm512_set1_epi32 (__INT_MAX__), __U, __A);
> + __MM512_REDUCE_OP (min_epi32);
> +}
> +
> +extern __inline int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_max_epi32 (__mmask16 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi32 (_mm512_set1_epi32 (-__INT_MAX__ - 1), __U,
> __A);
> + __MM512_REDUCE_OP (max_epi32);
> +}
> +
> +extern __inline unsigned int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_min_epu32 (__mmask16 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi32 (_mm512_set1_epi32 (~0), __U, __A);
> + __MM512_REDUCE_OP (min_epu32);
> +}
> +
> +extern __inline unsigned int
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_max_epu32 (__mmask16 __U, __m512i __A)
> +{
> + __A = _mm512_maskz_mov_epi32 (__U, __A);
> + __MM512_REDUCE_OP (max_epu32);
> +}
> +
> +#undef __MM512_REDUCE_OP
> +#define __MM512_REDUCE_OP(op) \
> + __m256 __T1 = (__m256) _mm512_extractf64x4_pd ((__m512d) __A, 1); \
> + __m256 __T2 = (__m256) _mm512_extractf64x4_pd ((__m512d) __A, 0); \
> + __m256 __T3 = __T1 op __T2; \
> + __m128 __T4 = _mm256_extractf128_ps (__T3, 1); \
> + __m128 __T5 = _mm256_extractf128_ps (__T3, 0); \
> + __m128 __T6 = __T4 op __T5; \
> + __m128 __T7 = __builtin_shuffle (__T6, (__v4si) { 2, 3, 0, 1 }); \
> + __m128 __T8 = __T6 op __T7; \
> + return __T8[0] op __T8[1]
> +
> +extern __inline float
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_add_ps (__m512 __A)
> +{
> + __MM512_REDUCE_OP (+);
> +}
> +
> +extern __inline float
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_mul_ps (__m512 __A)
> +{
> + __MM512_REDUCE_OP (*);
> +}
> +
> +extern __inline float
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_add_ps (__mmask16 __U, __m512 __A)
> +{
> + __A = _mm512_maskz_mov_ps (__U, __A);
> + __MM512_REDUCE_OP (+);
> +}
> +
> +extern __inline float
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_mul_ps (__mmask16 __U, __m512 __A)
> +{
> + __A = _mm512_mask_mov_ps (_mm512_set1_ps (1.0f), __U, __A);
> + __MM512_REDUCE_OP (*);
> +}
> +
> +#undef __MM512_REDUCE_OP
> +#define __MM512_REDUCE_OP(op) \
> + __m256 __T1 = (__m256) _mm512_extractf64x4_pd ((__m512d) __A, 1); \
> + __m256 __T2 = (__m256) _mm512_extractf64x4_pd ((__m512d) __A, 0); \
> + __m256 __T3 = _mm256_##op (__T1, __T2); \
> + __m128 __T4 = _mm256_extractf128_ps (__T3, 1); \
> + __m128 __T5 = _mm256_extractf128_ps (__T3, 0); \
> + __m128 __T6 = _mm_##op (__T4, __T5); \
> + __m128 __T7 = __builtin_shuffle (__T6, (__v4si) { 2, 3, 0, 1 }); \
> + __m128 __T8 = _mm_##op (__T6, __T7); \
> + __m128 __T9 = __builtin_shuffle (__T8, (__v4si) { 1, 0, 1, 0 }); \
> + __m128 __T10 = _mm_##op (__T8, __T9);
> \
> + return __T10[0]
> +
> +extern __inline float
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_min_ps (__m512 __A)
> +{
> + __MM512_REDUCE_OP (min_ps);
> +}
> +
> +extern __inline float
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_max_ps (__m512 __A)
> +{
> + __MM512_REDUCE_OP (max_ps);
> +}
> +
> +extern __inline float
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_min_ps (__mmask16 __U, __m512 __A)
> +{
> + __A = _mm512_mask_mov_ps (_mm512_set1_ps (__builtin_inff ()), __U, __A);
> + __MM512_REDUCE_OP (min_ps);
> +}
> +
> +extern __inline float
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_max_ps (__mmask16 __U, __m512 __A)
> +{
> + __A = _mm512_mask_mov_ps (_mm512_set1_ps (-__builtin_inff ()), __U, __A);
> + __MM512_REDUCE_OP (max_ps);
> +}
> +
> +#undef __MM512_REDUCE_OP
> +#define __MM512_REDUCE_OP(op) \
> + __v4di __T1 = (__v4di) _mm512_extracti64x4_epi64 (__A, 1); \
> + __v4di __T2 = (__v4di) _mm512_extracti64x4_epi64 (__A, 0); \
> + __m256i __T3 = (__m256i) (__T1 op __T2); \
> + __v2di __T4 = (__v2di) _mm256_extracti128_si256 (__T3, 1); \
> + __v2di __T5 = (__v2di) _mm256_extracti128_si256 (__T3, 0); \
> + __v2di __T6 = __T4 op __T5; \
> + return __T6[0] op __T6[1]
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_add_epi64 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (+);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_mul_epi64 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (*);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_and_epi64 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (&);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_or_epi64 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (|);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_add_epi64 (__mmask8 __U, __m512i __A)
> +{
> + __A = _mm512_maskz_mov_epi64 (__U, __A);
> + __MM512_REDUCE_OP (+);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_mul_epi64 (__mmask8 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi64 (_mm512_set1_epi64 (1LL), __U, __A);
> + __MM512_REDUCE_OP (*);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_and_epi64 (__mmask8 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi64 (_mm512_set1_epi64 (~0LL), __U, __A);
> + __MM512_REDUCE_OP (&);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_or_epi64 (__mmask8 __U, __m512i __A)
> +{
> + __A = _mm512_maskz_mov_epi64 (__U, __A);
> + __MM512_REDUCE_OP (|);
> +}
> +
> +#undef __MM512_REDUCE_OP
> +#define __MM512_REDUCE_OP(op) \
> + __m512i __T1 = _mm512_shuffle_i64x2 (__A, __A, 0x4e);
> \
> + __m512i __T2 = _mm512_##op (__A, __T1); \
> + __m512i __T3 \
> + = (__m512i) __builtin_shuffle ((__v8di) __T2, \
> + (__v8di) { 2, 3, 0, 1, 6, 7, 4, 5 });\
> + __m512i __T4 = _mm512_##op (__T2, __T3); \
> + __m512i __T5 \
> + = (__m512i) __builtin_shuffle ((__v8di) __T4, \
> + (__v8di) { 1, 0, 3, 2, 5, 4, 7, 6 });\
> + __v8di __T6 = (__v8di) _mm512_##op (__T4, __T5); \
> + return __T6[0]
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_min_epi64 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (min_epi64);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_max_epi64 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (max_epi64);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_min_epi64 (__mmask8 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi64 (_mm512_set1_epi64 (__LONG_LONG_MAX__),
> + __U, __A);
> + __MM512_REDUCE_OP (min_epi64);
> +}
> +
> +extern __inline long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_max_epi64 (__mmask8 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi64 (_mm512_set1_epi64 (-__LONG_LONG_MAX__ - 1),
> + __U, __A);
> + __MM512_REDUCE_OP (max_epi64);
> +}
> +
> +extern __inline unsigned long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_min_epu64 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (min_epu64);
> +}
> +
> +extern __inline unsigned long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_max_epu64 (__m512i __A)
> +{
> + __MM512_REDUCE_OP (max_epu64);
> +}
> +
> +extern __inline unsigned long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_min_epu64 (__mmask8 __U, __m512i __A)
> +{
> + __A = _mm512_mask_mov_epi64 (_mm512_set1_epi64 (~0LL), __U, __A);
> + __MM512_REDUCE_OP (min_epu64);
> +}
> +
> +extern __inline unsigned long long
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_max_epu64 (__mmask8 __U, __m512i __A)
> +{
> + __A = _mm512_maskz_mov_epi64 (__U, __A);
> + __MM512_REDUCE_OP (max_epu64);
> +}
> +
> +#undef __MM512_REDUCE_OP
> +#define __MM512_REDUCE_OP(op) \
> + __m256d __T1 = (__m256d) _mm512_extractf64x4_pd (__A, 1); \
> + __m256d __T2 = (__m256d) _mm512_extractf64x4_pd (__A, 0); \
> + __m256d __T3 = __T1 op __T2; \
> + __m128d __T4 = _mm256_extractf128_pd (__T3, 1); \
> + __m128d __T5 = _mm256_extractf128_pd (__T3, 0); \
> + __m128d __T6 = __T4 op __T5; \
> + return __T6[0] op __T6[1]
> +
> +extern __inline double
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_add_pd (__m512d __A)
> +{
> + __MM512_REDUCE_OP (+);
> +}
> +
> +extern __inline double
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_mul_pd (__m512d __A)
> +{
> + __MM512_REDUCE_OP (*);
> +}
> +
> +extern __inline double
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_add_pd (__mmask8 __U, __m512d __A)
> +{
> + __A = _mm512_maskz_mov_pd (__U, __A);
> + __MM512_REDUCE_OP (+);
> +}
> +
> +extern __inline double
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_mul_pd (__mmask8 __U, __m512d __A)
> +{
> + __A = _mm512_mask_mov_pd (_mm512_set1_pd (1.0), __U, __A);
> + __MM512_REDUCE_OP (*);
> +}
> +
> +#undef __MM512_REDUCE_OP
> +#define __MM512_REDUCE_OP(op) \
> + __m256d __T1 = (__m256d) _mm512_extractf64x4_pd (__A, 1); \
> + __m256d __T2 = (__m256d) _mm512_extractf64x4_pd (__A, 0); \
> + __m256d __T3 = _mm256_##op (__T1, __T2); \
> + __m128d __T4 = _mm256_extractf128_pd (__T3, 1); \
> + __m128d __T5 = _mm256_extractf128_pd (__T3, 0); \
> + __m128d __T6 = _mm_##op (__T4, __T5);
> \
> + __m128d __T7 = (__m128d) __builtin_shuffle (__T6, (__v2di) { 1, 0 });
> \
> + __m128d __T8 = _mm_##op (__T6, __T7);
> \
> + return __T8[0]
> +
> +extern __inline double
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_min_pd (__m512d __A)
> +{
> + __MM512_REDUCE_OP (min_pd);
> +}
> +
> +extern __inline double
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_reduce_max_pd (__m512d __A)
> +{
> + __MM512_REDUCE_OP (max_pd);
> +}
> +
> +extern __inline double
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_min_pd (__mmask8 __U, __m512d __A)
> +{
> + __A = _mm512_mask_mov_pd (_mm512_set1_pd (__builtin_inf ()), __U, __A);
> + __MM512_REDUCE_OP (min_pd);
> +}
> +
> +extern __inline double
> +__attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
> +_mm512_mask_reduce_max_pd (__mmask8 __U, __m512d __A)
> +{
> + __A = _mm512_mask_mov_pd (_mm512_set1_pd (-__builtin_inf ()), __U, __A);
> + __MM512_REDUCE_OP (max_pd);
> +}
> +
> +#undef __MM512_REDUCE_OP
> +
> #ifdef __DISABLE_AVX512F__
> #undef __DISABLE_AVX512F__
> #pragma GCC pop_options
> --- gcc/testsuite/gcc.target/i386/avx512f-reduce-op-1.c.jj 2017-04-07
> 12:25:19.578556015 +0200
> +++ gcc/testsuite/gcc.target/i386/avx512f-reduce-op-1.c 2017-04-07
> 16:34:37.972974281 +0200
> @@ -0,0 +1,410 @@
> +/* { dg-do run } */
> +/* { dg-options "-O2 -mavx512f" } */
> +/* { dg-require-effective-target avx512f } */
> +
> +#define AVX512F
> +
> +#include "avx512f-helper.h"
> +
> +__attribute__((noinline, noclone)) int
> +test_reduce_add_epi32 (__m512i a)
> +{
> + return _mm512_reduce_add_epi32 (a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_reduce_mul_epi32 (__m512i a)
> +{
> + return _mm512_reduce_mul_epi32 (a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_reduce_and_epi32 (__m512i a)
> +{
> + return _mm512_reduce_and_epi32 (a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_reduce_or_epi32 (__m512i a)
> +{
> + return _mm512_reduce_or_epi32 (a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_mask_reduce_add_epi32 (__mmask16 u, __m512i a)
> +{
> + return _mm512_mask_reduce_add_epi32 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_mask_reduce_mul_epi32 (__mmask16 u, __m512i a)
> +{
> + return _mm512_mask_reduce_mul_epi32 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_mask_reduce_and_epi32 (__mmask16 u, __m512i a)
> +{
> + return _mm512_mask_reduce_and_epi32 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_mask_reduce_or_epi32 (__mmask16 u, __m512i a)
> +{
> + return _mm512_mask_reduce_or_epi32 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_reduce_min_epi32 (__m512i a)
> +{
> + return _mm512_reduce_min_epi32 (a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_reduce_max_epi32 (__m512i a)
> +{
> + return _mm512_reduce_max_epi32 (a);
> +}
> +
> +__attribute__((noinline, noclone)) unsigned int
> +test_reduce_min_epu32 (__m512i a)
> +{
> + return _mm512_reduce_min_epu32 (a);
> +}
> +
> +__attribute__((noinline, noclone)) unsigned int
> +test_reduce_max_epu32 (__m512i a)
> +{
> + return _mm512_reduce_max_epu32 (a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_mask_reduce_min_epi32 (__mmask16 u, __m512i a)
> +{
> + return _mm512_mask_reduce_min_epi32 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) int
> +test_mask_reduce_max_epi32 (__mmask16 u, __m512i a)
> +{
> + return _mm512_mask_reduce_max_epi32 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) unsigned int
> +test_mask_reduce_min_epu32 (__mmask16 u, __m512i a)
> +{
> + return _mm512_mask_reduce_min_epu32 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) unsigned int
> +test_mask_reduce_max_epu32 (__mmask16 u, __m512i a)
> +{
> + return _mm512_mask_reduce_max_epu32 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) float
> +test_reduce_add_ps (__m512 a)
> +{
> + return _mm512_reduce_add_ps (a);
> +}
> +
> +__attribute__((noinline, noclone)) float
> +test_reduce_mul_ps (__m512 a)
> +{
> + return _mm512_reduce_mul_ps (a);
> +}
> +
> +__attribute__((noinline, noclone)) float
> +test_mask_reduce_add_ps (__mmask16 u, __m512 a)
> +{
> + return _mm512_mask_reduce_add_ps (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) float
> +test_mask_reduce_mul_ps (__mmask16 u, __m512 a)
> +{
> + return _mm512_mask_reduce_mul_ps (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) float
> +test_reduce_min_ps (__m512 a)
> +{
> + return _mm512_reduce_min_ps (a);
> +}
> +
> +__attribute__((noinline, noclone)) float
> +test_reduce_max_ps (__m512 a)
> +{
> + return _mm512_reduce_max_ps (a);
> +}
> +
> +__attribute__((noinline, noclone)) float
> +test_mask_reduce_min_ps (__mmask16 u, __m512 a)
> +{
> + return _mm512_mask_reduce_min_ps (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) float
> +test_mask_reduce_max_ps (__mmask16 u, __m512 a)
> +{
> + return _mm512_mask_reduce_max_ps (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_reduce_add_epi64 (__m512i a)
> +{
> + return _mm512_reduce_add_epi64 (a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_reduce_mul_epi64 (__m512i a)
> +{
> + return _mm512_reduce_mul_epi64 (a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_reduce_and_epi64 (__m512i a)
> +{
> + return _mm512_reduce_and_epi64 (a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_reduce_or_epi64 (__m512i a)
> +{
> + return _mm512_reduce_or_epi64 (a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_mask_reduce_add_epi64 (__mmask8 u, __m512i a)
> +{
> + return _mm512_mask_reduce_add_epi64 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_mask_reduce_mul_epi64 (__mmask8 u, __m512i a)
> +{
> + return _mm512_mask_reduce_mul_epi64 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_mask_reduce_and_epi64 (__mmask8 u, __m512i a)
> +{
> + return _mm512_mask_reduce_and_epi64 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_mask_reduce_or_epi64 (__mmask8 u, __m512i a)
> +{
> + return _mm512_mask_reduce_or_epi64 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_reduce_min_epi64 (__m512i a)
> +{
> + return _mm512_reduce_min_epi64 (a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_reduce_max_epi64 (__m512i a)
> +{
> + return _mm512_reduce_max_epi64 (a);
> +}
> +
> +__attribute__((noinline, noclone)) unsigned long long
> +test_reduce_min_epu64 (__m512i a)
> +{
> + return _mm512_reduce_min_epu64 (a);
> +}
> +
> +__attribute__((noinline, noclone)) unsigned long long
> +test_reduce_max_epu64 (__m512i a)
> +{
> + return _mm512_reduce_max_epu64 (a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_mask_reduce_min_epi64 (__mmask8 u, __m512i a)
> +{
> + return _mm512_mask_reduce_min_epi64 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) long long
> +test_mask_reduce_max_epi64 (__mmask8 u, __m512i a)
> +{
> + return _mm512_mask_reduce_max_epi64 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) unsigned long long
> +test_mask_reduce_min_epu64 (__mmask8 u, __m512i a)
> +{
> + return _mm512_mask_reduce_min_epu64 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) unsigned long long
> +test_mask_reduce_max_epu64 (__mmask8 u, __m512i a)
> +{
> + return _mm512_mask_reduce_max_epu64 (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) double
> +test_reduce_add_pd (__m512d a)
> +{
> + return _mm512_reduce_add_pd (a);
> +}
> +
> +__attribute__((noinline, noclone)) double
> +test_reduce_mul_pd (__m512d a)
> +{
> + return _mm512_reduce_mul_pd (a);
> +}
> +
> +__attribute__((noinline, noclone)) double
> +test_mask_reduce_add_pd (__mmask8 u, __m512d a)
> +{
> + return _mm512_mask_reduce_add_pd (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) double
> +test_mask_reduce_mul_pd (__mmask8 u, __m512d a)
> +{
> + return _mm512_mask_reduce_mul_pd (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) double
> +test_reduce_min_pd (__m512d a)
> +{
> + return _mm512_reduce_min_pd (a);
> +}
> +
> +__attribute__((noinline, noclone)) double
> +test_reduce_max_pd (__m512d a)
> +{
> + return _mm512_reduce_max_pd (a);
> +}
> +
> +__attribute__((noinline, noclone)) double
> +test_mask_reduce_min_pd (__mmask8 u, __m512d a)
> +{
> + return _mm512_mask_reduce_min_pd (u, a);
> +}
> +
> +__attribute__((noinline, noclone)) double
> +test_mask_reduce_max_pd (__mmask8 u, __m512d a)
> +{
> + return _mm512_mask_reduce_max_pd (u, a);
> +}
> +
> +#define TESTOP(opname, op, type, suffix, neutral) \
> + do { \
> + type r1 = _mm512_reduce_##opname##_##suffix (v.x); \
> + type r2 = test_reduce_##opname##_##suffix (v.x); \
> + type r3 = neutral; \
> + if (r1 != r2) \
> + __builtin_abort (); \
> + for (int i = 0; i < SIZE; i++) \
> + r3 = r3 op v.a[i]; \
> + if (r1 != r3) \
> + __builtin_abort (); \
> + type r4 = _mm512_mask_reduce_##opname##_##suffix (MASK_VALUE, v.x);
> \
> + type r5 = test_mask_reduce_##opname##_##suffix (MASK_VALUE, v.x); \
> + if (r4 != r5) \
> + __builtin_abort (); \
> + r3 = neutral; \
> + for (int i = 0; i < SIZE; i++) \
> + if (MASK_VALUE & (1 << i)) \
> + r3 = r3 op v.a[i]; \
> + if (r4 != r3) \
> + __builtin_abort (); \
> + type r6 = _mm512_mask_reduce_##opname##_##suffix (0, v.x); \
> + type r7 = test_mask_reduce_##opname##_##suffix (0, v.x); \
> + if (r6 != r7 || r6 != neutral) \
> + __builtin_abort (); \
> + } while (0)
> +
> +#define SIZE (AVX512F_LEN / 32)
> +#include "avx512f-mask-type.h"
> +
> +#define TEST_EPI32(c1, c2, c3, c4, c5, c6, c7, c8, \
> + c9, c10, c11, c12, c13, c14, c15, c16) \
> + do { \
> + UNION_TYPE (AVX512F_LEN, i_d) v; \
> + v.x = _mm512_set_epi32 (c1, c2, c3, c4, c5, c6, c7, c8, \
> + c9, c10, c11, c12, c13, c14, c15, c16); \
> + TESTOP (add, +, int, epi32, 0); \
> + TESTOP (mul, *, int, epi32, 1); \
> + TESTOP (and, &, int, epi32, ~0); \
> + TESTOP (or, |, int, epi32, 0); \
> + TESTOP (min, < v.a[i] ? r3 :, int, epi32, __INT_MAX__); \
> + TESTOP (max, > v.a[i] ? r3 :, int, epi32, -__INT_MAX__ - 1); \
> + TESTOP (min, < (unsigned) v.a[i] ? r3 :, unsigned, epu32, ~0U); \
> + TESTOP (max, > (unsigned) v.a[i] ? r3 :, unsigned, epu32, 0); \
> + } while (0)
> +
> +#define TEST_PS(c1, c2, c3, c4, c5, c6, c7, c8, \
> + c9, c10, c11, c12, c13, c14, c15, c16) \
> + do { \
> + UNION_TYPE (AVX512F_LEN, ) v; \
> + v.x = _mm512_set_ps (c1, c2, c3, c4, c5, c6, c7, c8, \
> + c9, c10, c11, c12, c13, c14, c15, c16); \
> + TESTOP (add, +, float, ps, 0.0f); \
> + TESTOP (mul, *, float, ps, 1.0f); \
> + TESTOP (min, < v.a[i] ? r3 :, float, ps, __builtin_inff ()); \
> + TESTOP (max, > v.a[i] ? r3 :, float, ps, -__builtin_inff ()); \
> + } while (0)
> +
> +static void
> +test_epi32_ps (void)
> +{
> + TEST_EPI32 (1, 2, 3, 4, 5, 6, 6, 5, 4, 3, 2, 1, 7, 6, 5, 4);
> + TEST_EPI32 (-1, 15, -1, 7, -1, 7, -1, -1, 6, 6, -1, -1, -1, -1, 7, 6);
> + TEST_PS (1, 2, 3, 4, 5, 6, 6, 5, 4, 3, 2, 1, 7, 6, 5, 4);
> + TEST_PS (1.25f, 2.25f, -0.25f, 4.0f, -2.0f, 4.0f, -3.0f, 2.0f,
> + -0.5f, -1.0f, 1.0f, -1.0f, 1.0f, 1.0f, 2.0f, 4.0f);
> +}
> +
> +#undef SIZE
> +#define SIZE (AVX512F_LEN / 64)
> +#include "avx512f-mask-type.h"
> +
> +#define TEST_EPI64(c1, c2, c3, c4, c5, c6, c7, c8) \
> + do { \
> + UNION_TYPE (AVX512F_LEN, i_q) v; \
> + v.x = _mm512_set_epi64 (c1, c2, c3, c4, c5, c6, c7, c8); \
> + TESTOP (add, +, long long, epi64, 0); \
> + TESTOP (mul, *, long long, epi64, 1); \
> + TESTOP (and, &, long long, epi64, ~0LL); \
> + TESTOP (or, |, long long, epi64, 0); \
> + TESTOP (min, < v.a[i] ? r3 :, long long, epi64, __LONG_LONG_MAX__);
> \
> + TESTOP (max, > v.a[i] ? r3 :, long long, epi64, \
> + -__LONG_LONG_MAX__ - 1); \
> + TESTOP (min, < (unsigned long long) v.a[i] ? r3 :, \
> + unsigned long long, epu64, ~0ULL); \
> + TESTOP (max, > (unsigned long long) v.a[i] ? r3 :, \
> + unsigned long long, epu64, 0); \
> + } while (0)
> +
> +#define TEST_PD(c1, c2, c3, c4, c5, c6, c7, c8) \
> + do { \
> + UNION_TYPE (AVX512F_LEN, d) v; \
> + v.x = _mm512_set_pd (c1, c2, c3, c4, c5, c6, c7, c8); \
> + TESTOP (add, +, double, pd, 0.0); \
> + TESTOP (mul, *, double, pd, 1.0); \
> + TESTOP (min, < v.a[i] ? r3 :, double, pd, __builtin_inf ()); \
> + TESTOP (max, > v.a[i] ? r3 :, double, pd, -__builtin_inf ()); \
> + } while (0)
> +
> +static void
> +test_epi64_pd (void)
> +{
> + TEST_EPI64 (1, 2, 3, 4, 5, 6, 6, 5);
> + TEST_EPI64 (-1, 15, -1, 7, -1, 7, -1, -1);
> + TEST_PD (1, 2, 3, 4, 5, 6, 6, 5);
> + TEST_PD (1.25f, 2.25f, -0.25f, 4.0f, -2.0f, 4.0f, -3.0f, 2.0f);
> +}
> +
> +void
> +test_512 (void)
> +{
> + test_epi32_ps ();
> + test_epi64_pd ();
> +}
>
> Jakub
