diff options
| author | Martin Kroeker <martin@ruby.chemie.uni-freiburg.de> | 2019-08-09 07:55:20 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2019-08-09 07:55:20 +0200 |
| commit | dccff2e7859cf0dc18e37f5f696dfbf3cca221a8 (patch) | |
| tree | 187aa5a3cd6bea9e2eca846afdbe449facc9f38a /kernel | |
| parent | 5c3458a6e78f6c7bd7365372e15adc82c5f3badb (diff) | |
| parent | 2dfb804cb943ac12035fe51859d109daca76b4f4 (diff) | |
| download | openblas-dccff2e7859cf0dc18e37f5f696dfbf3cca221a8.tar.gz openblas-dccff2e7859cf0dc18e37f5f696dfbf3cca221a8.tar.bz2 openblas-dccff2e7859cf0dc18e37f5f696dfbf3cca221a8.zip | |
Merge pull request #2206 from martin-frbg/zen-dtrmm
Replace vpermpd with vpermilpd in the Haswell DTRMM kernel
Diffstat (limited to 'kernel')
| -rw-r--r-- | kernel/x86_64/dtrmm_kernel_4x8_haswell.c | 24 |
1 files changed, 10 insertions, 14 deletions
diff --git a/kernel/x86_64/dtrmm_kernel_4x8_haswell.c b/kernel/x86_64/dtrmm_kernel_4x8_haswell.c index 651736b89..2acdc4615 100644 --- a/kernel/x86_64/dtrmm_kernel_4x8_haswell.c +++ b/kernel/x86_64/dtrmm_kernel_4x8_haswell.c @@ -33,7 +33,7 @@ static void dtrmm_kernel_4x8( BLASLONG n, FLOAT *alpha ,FLOAT *a, FLOAT *b, FLOA " vfmadd231pd %%ymm0 , %%ymm1 , %%ymm4 \n\t" " vfmadd231pd %%ymm0 , %%ymm2 , %%ymm8 \n\t" - " vpermpd $0xb1 , %%ymm0 , %%ymm0 \n\t" + " vpermilpd $0x05 , %%ymm0 , %%ymm0 \n\t" " vfmadd231pd %%ymm0 , %%ymm1 , %%ymm5 \n\t" " vfmadd231pd %%ymm0 , %%ymm2 , %%ymm9 \n\t" @@ -41,7 +41,7 @@ static void dtrmm_kernel_4x8( BLASLONG n, FLOAT *alpha ,FLOAT *a, FLOAT *b, FLOA " vfmadd231pd %%ymm0 , %%ymm1 , %%ymm6 \n\t" " vfmadd231pd %%ymm0 , %%ymm2 , %%ymm10 \n\t" - " vpermpd $0xb1 , %%ymm0 , %%ymm0 \n\t" + " vpermilpd $0x05 , %%ymm0 , %%ymm0 \n\t" " vfmadd231pd %%ymm0 , %%ymm1 , %%ymm7 \n\t" " vfmadd231pd %%ymm0 , %%ymm2 , %%ymm11 \n\t" @@ -62,18 +62,16 @@ static void dtrmm_kernel_4x8( BLASLONG n, FLOAT *alpha ,FLOAT *a, FLOAT *b, FLOA " vmulpd %%ymm0 , %%ymm10, %%ymm10 \n\t" " vmulpd %%ymm0 , %%ymm11, %%ymm11 \n\t" - " vpermpd $0xb1 , %%ymm5 , %%ymm5 \n\t" - " vpermpd $0xb1 , %%ymm7 , %%ymm7 \n\t" + " vpermilpd $0x05 , %%ymm5 , %%ymm5 \n\t" + " vpermilpd $0x05 , %%ymm7 , %%ymm7 \n\t" " vblendpd $0x0a , %%ymm5 , %%ymm4 , %%ymm0 \n\t" " vblendpd $0x05 , %%ymm5 , %%ymm4 , %%ymm1 \n\t" " vblendpd $0x0a , %%ymm7 , %%ymm6 , %%ymm2 \n\t" " vblendpd $0x05 , %%ymm7 , %%ymm6 , %%ymm3 \n\t" - " vpermpd $0x1b , %%ymm2 , %%ymm2 \n\t" - " vpermpd $0x1b , %%ymm3 , %%ymm3 \n\t" - " vpermpd $0xb1 , %%ymm2 , %%ymm2 \n\t" - " vpermpd $0xb1 , %%ymm3 , %%ymm3 \n\t" + " vperm2f128 $0x01 , %%ymm2 , %%ymm2 , %%ymm2 \n\t" + " vperm2f128 $0x01 , %%ymm3 , %%ymm3 , %%ymm3 \n\t" " vblendpd $0x03 , %%ymm0 , %%ymm2 , %%ymm4 \n\t" " vblendpd $0x03 , %%ymm1 , %%ymm3 , %%ymm5 \n\t" @@ -85,18 +83,16 @@ static void dtrmm_kernel_4x8( BLASLONG n, FLOAT *alpha ,FLOAT *a, FLOAT *b, FLOA " vmovups %%ymm6 , (%7) \n\t" " vmovups %%ymm7 , (%8) \n\t" - " vpermpd $0xb1 , %%ymm9 , %%ymm9 \n\t" - " vpermpd $0xb1 , %%ymm11, %%ymm11 \n\t" + " vpermilpd $0x05 , %%ymm9 , %%ymm9 \n\t" + " vpermilpd $0x05 , %%ymm11, %%ymm11 \n\t" " vblendpd $0x0a , %%ymm9 , %%ymm8 , %%ymm0 \n\t" " vblendpd $0x05 , %%ymm9 , %%ymm8 , %%ymm1 \n\t" " vblendpd $0x0a , %%ymm11, %%ymm10, %%ymm2 \n\t" " vblendpd $0x05 , %%ymm11, %%ymm10, %%ymm3 \n\t" - " vpermpd $0x1b , %%ymm2 , %%ymm2 \n\t" - " vpermpd $0x1b , %%ymm3 , %%ymm3 \n\t" - " vpermpd $0xb1 , %%ymm2 , %%ymm2 \n\t" - " vpermpd $0xb1 , %%ymm3 , %%ymm3 \n\t" + " vperm2f128 $0x01 , %%ymm2 , %%ymm2 , %%ymm2 \n\t" + " vperm2f128 $0x01 , %%ymm3 , %%ymm3 , %%ymm3 \n\t" " vblendpd $0x03 , %%ymm0 , %%ymm2 , %%ymm4 \n\t" " vblendpd $0x03 , %%ymm1 , %%ymm3 , %%ymm5 \n\t" |