diff options
Diffstat (limited to 'src/vppinfra')
| -rw-r--r-- | src/vppinfra/vector.h | 17 | ||||
| -rw-r--r-- | src/vppinfra/vector_avx2.h | 80 | ||||
| -rw-r--r-- | src/vppinfra/vector_avx512.h | 53 | ||||
| -rw-r--r-- | src/vppinfra/vector_sse42.h | 4 |
4 files changed, 149 insertions, 5 deletions
diff --git a/src/vppinfra/vector.h b/src/vppinfra/vector.h index 094cc85e22b..fcff5e79d95 100644 --- a/src/vppinfra/vector.h +++ b/src/vppinfra/vector.h @@ -58,8 +58,11 @@ #define CLIB_HAVE_VEC128 #endif -#if defined (__AVX__) +#if defined (__AVX2__) #define CLIB_HAVE_VEC256 +#if defined (__clang__) && __clang_major__ < 4 +#undef CLIB_HAVE_VEC256 +#endif #endif #if defined (__AVX512F__) @@ -179,7 +182,7 @@ t##s##x##c##_sub (t##s##x##c v1, t##s##x##c v2) \ foreach_vec #undef _ -/* this macro generate _splat inline funcitons for each scalar vector type */ +/* this macro generate _splat inline functions for each scalar vector type */ #define _(t, s, c) \ static_always_inline t##s##x##c \ t##s##x##c##_splat (t##s x) \ @@ -192,13 +195,21 @@ t##s##x##c##_splat (t##s x) \ \ return r; \ } - foreach_int_vec foreach_uint_vec + foreach_vec128i foreach_vec128u #undef _ #if defined (__SSE4_2__) && __GNUC__ >= 4 #include <vppinfra/vector_sse42.h> #endif +#if defined (__AVX2__) +#include <vppinfra/vector_avx2.h> +#endif + +#if defined (__AVX512F__) +#include <vppinfra/vector_avx512.h> +#endif + #if defined (__ALTIVEC__) #include <vppinfra/vector_altivec.h> #endif diff --git a/src/vppinfra/vector_avx2.h b/src/vppinfra/vector_avx2.h new file mode 100644 index 00000000000..ad7e7d4dea8 --- /dev/null +++ b/src/vppinfra/vector_avx2.h @@ -0,0 +1,80 @@ +/* + * Copyright (c) 2018 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef included_vector_avx2_h +#define included_vector_avx2_h + +#include <vppinfra/clib.h> +#include <x86intrin.h> + +#define foreach_avx2_vec256i \ + _(i,8,32,epi8) _(i,16,16,epi16) _(i,32,8,epi32) _(i,64,4,epi64x) +#define foreach_avx2_vec256u \ + _(u,8,32,epi8) _(u,16,16,epi16) _(u,32,8,epi32) _(u,64,4,epi64x) +#define foreach_avx2_vec256f \ + _(f,32,8,ps) _(f,64,4,pd) + +/* splat, load_unaligned, store_unaligned, is_all_zero, is_all_equal */ +#define _(t, s, c, i) \ +static_always_inline t##s##x##c \ +t##s##x##c##_splat (t##s x) \ +{ return (t##s##x##c) _mm256_set1_##i (x); } \ +\ +static_always_inline t##s##x##c \ +t##s##x##c##_load_unaligned (void *p) \ +{ return (t##s##x##c) _mm256_loadu_si256 (p); } \ +\ +static_always_inline void \ +t##s##x##c##_store_unaligned (t##s##x##c v, void *p) \ +{ _mm256_storeu_si256 ((__m256i *) p, (__m256i) v); } \ +\ +static_always_inline int \ +t##s##x##c##_is_all_zero (t##s##x##c x) \ +{ return _mm256_testz_si256 ((__m256i) x, (__m256i) x); } \ +\ +static_always_inline int \ +t##s##x##c##_is_all_equal (t##s##x##c v, t##s x) \ +{ return t##s##x##c##_is_all_zero (v != t##s##x##c##_splat (x)); }; \ +\ + +foreach_avx2_vec256i foreach_avx2_vec256u +#undef _ + always_inline u32x8 +u32x8_permute (u32x8 v, u32x8 idx) +{ + return (u32x8) _mm256_permutevar8x32_epi32 ((__m256i) v, (__m256i) idx); +} + +always_inline u32x4 +u32x8_extract_lo (u32x8 v) +{ + return (u32x4) _mm256_extracti128_si256 ((__m256i) v, 0); +} + +always_inline u32x4 +u32x8_extract_hi (u32x8 v) +{ + return (u32x4) _mm256_extracti128_si256 ((__m256i) v, 1); +} + +#endif /* included_vector_avx2_h */ + +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vppinfra/vector_avx512.h b/src/vppinfra/vector_avx512.h new file mode 100644 index 00000000000..ac4c09b8375 --- /dev/null +++ b/src/vppinfra/vector_avx512.h @@ -0,0 +1,53 @@ +/* + * Copyright (c) 2015 Cisco and/or its affiliates. + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at: + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +#ifndef included_vector_avx512_h +#define included_vector_avx512_h + +#include <vppinfra/clib.h> +#include <x86intrin.h> + +#define foreach_avx512_vec512i \ + _(i,8,64,epi8) _(i,16,32,epi16) _(i,32,16,epi32) _(i,64,8,epi64) +#define foreach_avx512_vec512u \ + _(u,8,64,epi8) _(u,16,32,epi16) _(u,32,16,epi32) _(u,64,8,epi64) +#define foreach_avx512_vec512f \ + _(f,32,8,ps) _(f,64,4,pd) + +/* splat, load_unaligned, store_unaligned */ +#define _(t, s, c, i) \ +static_always_inline t##s##x##c \ +t##s##x##c##_splat (t##s x) \ +{ return (t##s##x##c) _mm512_set1_##i (x); } \ +\ +static_always_inline t##s##x##c \ +t##s##x##c##_load_unaligned (void *p) \ +{ return (t##s##x##c) _mm512_loadu_si512 (p); } \ +\ +static_always_inline void \ +t##s##x##c##_store_unaligned (t##s##x##c v, void *p) \ +{ _mm512_storeu_si512 ((__m512i *) p, (__m512i) v); } \ +\ + +foreach_avx512_vec512i foreach_avx512_vec512u +#undef _ +#endif /* included_vector_avx512_h */ +/* + * fd.io coding-style-patch-verification: ON + * + * Local Variables: + * eval: (c-set-style "gnu") + * End: + */ diff --git a/src/vppinfra/vector_sse42.h b/src/vppinfra/vector_sse42.h index dab22deff7c..cf7f158b873 100644 --- a/src/vppinfra/vector_sse42.h +++ b/src/vppinfra/vector_sse42.h @@ -200,9 +200,9 @@ u64x2_write_hi (u64x2 x, u64 * a) /* Unaligned loads/stores. */ #define _(t) \ - always_inline void t##_store_unaligned (t x, t * a) \ + always_inline void t##_store_unaligned (t x, void * a) \ { _mm_storeu_si128 ((__m128i *) a, (__m128i) x); } \ - always_inline t t##_load_unaligned (t * a) \ + always_inline t t##_load_unaligned (void * a) \ { return (t) _mm_loadu_si128 ((__m128i *) a); } _(u8x16) _(u16x8) _(u32x4) _(u64x2) _(i8x16) _(i16x8) _(i32x4) _(i64x2) |