floraachy/OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Optimize the performance of rot by using universal intrinsics

Browse Source
This commit is contained in:
Qiyu8
2020-11-11 14:33:12 +08:00
parent ff16329cb7
commit 8c0b206d4c
6 changed files with 174 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
kernel/simd/intrin_avx.h
View File

@@ -12,6 +12,8 @@ typedef __m256d v_f64;
***************************/
#define v_add_f32 _mm256_add_ps
#define v_add_f64 _mm256_add_pd
#define v_sub_f32 _mm256_sub_ps
#define v_sub_f64 _mm256_sub_pd
#define v_mul_f32 _mm256_mul_ps
#define v_mul_f64 _mm256_mul_pd
@@ -19,12 +21,20 @@ typedef __m256d v_f64;
// multiply and add, a*b + c
#define v_muladd_f32 _mm256_fmadd_ps
#define v_muladd_f64 _mm256_fmadd_pd
// multiply and subtract, a*b - c
#define v_mulsub_f32 _mm256_fmsub_ps
#define v_mulsub_f64 _mm256_fmsub_pd
#else
// multiply and add, a*b + c
BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
{ return v_add_f32(v_mul_f32(a, b), c); }
BLAS_FINLINE v_f64 v_muladd_f64(v_f64 a, v_f64 b, v_f64 c)
{ return v_add_f64(v_mul_f64(a, b), c); }
// multiply and subtract, a*b - c
BLAS_FINLINE v_f32 v_mulsub_f32(v_f32 a, v_f32 b, v_f32 c)
{ return v_sub_f32(v_mul_f32(a, b), c); }
BLAS_FINLINE v_f64 v_mulsub_f64(v_f64 a, v_f64 b, v_f64 c)
{ return v_sub_f64(v_mul_f64(a, b), c); }
#endif // !HAVE_FMA3
// Horizontal add: Calculates the sum of all vector elements.

5
kernel/simd/intrin_avx512.h
View File

@@ -12,11 +12,16 @@ typedef __m512d v_f64;
***************************/
#define v_add_f32 _mm512_add_ps
#define v_add_f64 _mm512_add_pd
#define v_sub_f32 _mm512_sub_ps
#define v_sub_f64 _mm512_sub_pd
#define v_mul_f32 _mm512_mul_ps
#define v_mul_f64 _mm512_mul_pd
// multiply and add, a*b + c
#define v_muladd_f32 _mm512_fmadd_ps
#define v_muladd_f64 _mm512_fmadd_pd
// multiply and subtract, a*b - c
#define v_mulsub_f32 _mm512_fmsub_ps
#define v_mulsub_f64 _mm512_fmsub_pd
BLAS_FINLINE float v_sum_f32(v_f32 a)
{
__m512 h64 = _mm512_shuffle_f32x4(a, a, _MM_SHUFFLE(3, 2, 3, 2));

10
kernel/simd/intrin_neon.h
View File

@@ -18,6 +18,8 @@ typedef float32x4_t v_f32;
***************************/
#define v_add_f32 vaddq_f32
#define v_add_f64 vaddq_f64
#define v_sub_f32 vsubq_f32
#define v_sub_f64 vsubq_f64
#define v_mul_f32 vmulq_f32
#define v_mul_f64 vmulq_f64
@@ -26,16 +28,24 @@ typedef float32x4_t v_f32;
// multiply and add, a*b + c
BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
{ return vfmaq_f32(c, a, b); }
// multiply and subtract, a*b - c
BLAS_FINLINE v_f32 v_mulsub_f32(v_f32 a, v_f32 b, v_f32 c)
{ return vfmaq_f32(vnegq_f32(c), a, b); }
#else
// multiply and add, a*b + c
BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
{ return vmlaq_f32(c, a, b); }
// multiply and subtract, a*b - c
BLAS_FINLINE v_f32 v_mulsub_f32(v_f32 a, v_f32 b, v_f32 c)
{ return vmlaq_f32(vnegq_f32(c), a, b); }
#endif
// FUSED F64
#if V_SIMD_F64
BLAS_FINLINE v_f64 v_muladd_f64(v_f64 a, v_f64 b, v_f64 c)
{ return vfmaq_f64(c, a, b); }
BLAS_FINLINE v_f64 v_mulsub_f64(v_f64 a, v_f64 b, v_f64 c)
{ return vfmaq_f64(vnegq_f64(c), a, b); }
#endif
// Horizontal add: Calculates the sum of all vector elements.

13
kernel/simd/intrin_sse.h
View File

@@ -12,22 +12,35 @@ typedef __m128d v_f64;
***************************/
#define v_add_f32 _mm_add_ps
#define v_add_f64 _mm_add_pd
#define v_sub_f32 _mm_sub_ps
#define v_sub_f64 _mm_sub_pd
#define v_mul_f32 _mm_mul_ps
#define v_mul_f64 _mm_mul_pd
#ifdef HAVE_FMA3
// multiply and add, a*b + c
#define v_muladd_f32 _mm_fmadd_ps
#define v_muladd_f64 _mm_fmadd_pd
// multiply and subtract, a*b - c
#define v_mulsub_f32 _mm_fmsub_ps
#define v_mulsub_f64 _mm_fmsub_pd
#elif defined(HAVE_FMA4)
// multiply and add, a*b + c
#define v_muladd_f32 _mm_macc_ps
#define v_muladd_f64 _mm_macc_pd
// multiply and subtract, a*b - c
#define v_mulsub_f32 _mm_msub_ps
#define v_mulsub_f64 _mm_msub_pd
#else
// multiply and add, a*b + c
BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 c)
{ return v_add_f32(v_mul_f32(a, b), c); }
BLAS_FINLINE v_f64 v_muladd_f64(v_f64 a, v_f64 b, v_f64 c)
{ return v_add_f64(v_mul_f64(a, b), c); }
// multiply and subtract, a*b - c
BLAS_FINLINE v_f32 v_mulsub_f32(v_f32 a, v_f32 b, v_f32 c)
{ return v_sub_f32(v_mul_f32(a, b), c); }
BLAS_FINLINE v_f64 v_mulsub_f64(v_f64 a, v_f64 b, v_f64 c)
{ return v_sub_f64(v_mul_f64(a, b), c); }
#endif // HAVE_FMA3
// Horizontal add: Calculates the sum of all vector elements.