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

Optimize the performance of dot by using universal intrinsics in X86/ARM

Browse Source
This commit is contained in:
Qiyu8
2020-09-28 20:36:53 +08:00
parent 0d98ce202c
commit 1b1a757f5f
7 changed files with 177 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
kernel/simd/intrin_sse.h
View File

@@ -1,13 +1,13 @@
#define V_SIMD 128
#define V_SIMD_F64 1
/*
Data Type
*/
/***************************
* Data Type
***************************/
typedef __m128 v_f32;
#define v_nlanes_f32 4
/*
arithmetic
*/
/***************************
* Arithmetic
***************************/
#define v_add_f32 _mm_add_ps
#define v_mul_f32 _mm_mul_ps
#ifdef HAVE_FMA3
@@ -21,10 +21,26 @@ arithmetic
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); }
#endif // HAVE_FMA3
/*
memory
*/
// Horizontal add: Calculates the sum of all vector elements.
BLAS_FINLINE float v_sum_f32(__m128 a)
{
#ifdef HAVE_SSE3
__m128 sum_halves = _mm_hadd_ps(a, a);
return _mm_cvtss_f32(_mm_hadd_ps(sum_halves, sum_halves));
#else
__m128 t1 = _mm_movehl_ps(a, a);
__m128 t2 = _mm_add_ps(a, t1);
__m128 t3 = _mm_shuffle_ps(t2, t2, 1);
__m128 t4 = _mm_add_ss(t2, t3);
return _mm_cvtss_f32(t4);
#endif
}
/***************************
* memory
***************************/
// unaligned load
#define v_loadu_f32 _mm_loadu_ps
#define v_storeu_f32 _mm_storeu_ps
#define v_setall_f32(VAL) _mm_set1_ps(VAL)
#define v_setall_f32(VAL) _mm_set1_ps(VAL)
#define v_zero_f32 _mm_setzero_ps