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

Add double precision operations

This commit is contained in:
Martin Kroeker 2020-10-14 18:10:45 +02:00 committed by GitHub
parent a897bc3bd2
commit c9c3ae07af
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 45 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

48
kernel/simd/intrin_sse.h
View File

@ -3,25 +3,59 @@
/*************************** /***************************
* Data Type * Data Type
***************************/ ***************************/
#ifdef DOUBLE
typedef __m128d v_f32;
#else
typedef __m128 v_f32; typedef __m128 v_f32;
#endif
#define v_nlanes_f32 4 #define v_nlanes_f32 4
/*************************** /***************************
* Arithmetic * Arithmetic
***************************/ ***************************/
#ifdef DOUBLE
#define v_add_f32 _mm_add_pd
#define v_mul_f32 _mm_mul_pd
#else
#define v_add_f32 _mm_add_ps #define v_add_f32 _mm_add_ps
#define v_mul_f32 _mm_mul_ps #define v_mul_f32 _mm_mul_ps
#endif
#ifdef HAVE_FMA3 #ifdef HAVE_FMA3
// multiply and add, a*b + c // multiply and add, a*b + c
#define v_muladd_f32 _mm_fmadd_ps #ifdef DOUBLE
#define v_muladd_f32 _mm_fmadd_pd
#else
#define v_muladd_f32 _mm_fmadd_ps
#endif
#elif defined(HAVE_FMA4) #elif defined(HAVE_FMA4)
// multiply and add, a*b + c // multiply and add, a*b + c
#define v_muladd_f32 _mm_macc_ps #ifdef DOUBLE
#define v_muladd_f32 _mm_macc_pd
#else
#define v_muladd_f32 _mm_macc_ps
#endif
#else #else
// multiply and add, a*b + c // multiply and add, a*b + c
BLAS_FINLINE v_f32 v_muladd_f32(v_f32 a, v_f32 b, v_f32 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); } { return v_add_f32(v_mul_f32(a, b), c); }
#endif // HAVE_FMA3 #endif // HAVE_FMA3
// Horizontal add: Calculates the sum of all vector elements.
#ifdef DOUBLE
BLAS_FINLINE double v_sum_f32(__m128d a)
{
#ifdef HAVE_SSE3
__m128d sum_halves = _mm_hadd_pd(a, a);
return _mm_cvtsd_f64(_mm_hadd_pd(sum_halves, sum_halves));
#else
__m128d t1 = _mm_movehl_pd(a, a);
__m128d t2 = _mm_add_pd(a, t1);
__m128d t3 = _mm_shuffle_pd(t2, t2, 1);
__m128d t4 = _mm_add_ss(t2, t3);
return _mm_cvtsd_f64(t4);
#endif
}
#else
// Horizontal add: Calculates the sum of all vector elements. // Horizontal add: Calculates the sum of all vector elements.
BLAS_FINLINE float v_sum_f32(__m128 a) BLAS_FINLINE float v_sum_f32(__m128 a)
{ {
@ -36,11 +70,19 @@ BLAS_FINLINE float v_sum_f32(__m128 a)
return _mm_cvtss_f32(t4); return _mm_cvtss_f32(t4);
#endif #endif
} }
#endif
/*************************** /***************************
* memory * memory
***************************/ ***************************/
// unaligned load // unaligned load
#ifdef DOUBLE
#define v_loadu_f32 _mm_loadu_pd
#define v_storeu_f32 _mm_storeu_pd
#define v_setall_f32(VAL) _mm_set1_pd(VAL)
#define v_zero_f32 _mm_setzero_pd
#else
#define v_loadu_f32 _mm_loadu_ps #define v_loadu_f32 _mm_loadu_ps
#define v_storeu_f32 _mm_storeu_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 #define v_zero_f32 _mm_setzero_ps
#endif