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

dgemm: Use the skylakex beta function also for haswell 

it's more efficient for certain tall/skinny matrices
This commit is contained in:
Arjan van de Ven 2018-12-16 22:59:02 +00:00
parent e8ca5a59a9
commit c43331ad0a
2 changed files with 13 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
kernel/x86_64/KERNEL.HASWELL
View File

@ -45,6 +45,7 @@ SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
DTRMMKERNEL = dtrmm_kernel_4x8_haswell.c DTRMMKERNEL = dtrmm_kernel_4x8_haswell.c
DGEMMKERNEL = dgemm_kernel_4x8_haswell.S DGEMMKERNEL = dgemm_kernel_4x8_haswell.S
DGEMM_BETA = dgemm_beta_skylakex.c
DGEMMINCOPY = ../generic/gemm_ncopy_4.c DGEMMINCOPY = ../generic/gemm_ncopy_4.c
DGEMMITCOPY = ../generic/gemm_tcopy_4.c DGEMMITCOPY = ../generic/gemm_tcopy_4.c
DGEMMONCOPY = ../generic/gemm_ncopy_8.c DGEMMONCOPY = ../generic/gemm_ncopy_8.c

16
kernel/x86_64/dgemm_beta_skylakex.c
View File

@ -61,17 +61,17 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
c_offset = c; c_offset = c;
if (beta == ZERO){ if (beta == ZERO){
__m512d z_zero;
z_zero = _mm512_setzero_pd();
j = n; j = n;
do { do {
c_offset1 = c_offset; c_offset1 = c_offset;
c_offset += ldc; c_offset += ldc;
i = m; i = m;
#ifdef __AVX2__
#ifdef __AVX512CD__
while (i >= 32) { while (i >= 32) {
__m512d z_zero = _mm512_setzero_pd();
_mm512_storeu_pd(c_offset1, z_zero); _mm512_storeu_pd(c_offset1, z_zero);
_mm512_storeu_pd(c_offset1 + 8, z_zero); _mm512_storeu_pd(c_offset1 + 8, z_zero);
_mm512_storeu_pd(c_offset1 + 16, z_zero); _mm512_storeu_pd(c_offset1 + 16, z_zero);
@ -79,12 +79,20 @@ int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT beta,
c_offset1 += 32; c_offset1 += 32;
i -= 32; i -= 32;
} }
#endif
while (i >= 8) { while (i >= 8) {
#ifdef __AVX512CD__
__m512d z_zero = _mm512_setzero_pd();
_mm512_storeu_pd(c_offset1, z_zero); _mm512_storeu_pd(c_offset1, z_zero);
#else
__m256d y_zero = _mm256_setzero_pd();
_mm256_storeu_pd(c_offset1, y_zero);
_mm256_storeu_pd(c_offset1 + 4, y_zero);
#endif
c_offset1 += 8; c_offset1 += 8;
i -= 8; i -= 8;
} }
#endif
while (i > 0) { while (i > 0) {
*c_offset1 = ZERO; *c_offset1 = ZERO;
c_offset1 ++; c_offset1 ++;