diff options
| author | Martin Kroeker <martin@ruby.chemie.uni-freiburg.de> | 2019-07-22 08:28:16 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2019-07-22 08:28:16 +0200 |
| commit | 28e96458e5a4b2d8039ed16048a07892a7c960bf (patch) | |
| tree | fa3c04366ced9bb2c25844949458e3260a8044ee | |
| parent | abea977ded8729c6dcfcfbee51a18eceef8d8440 (diff) | |
| download | openblas-28e96458e5a4b2d8039ed16048a07892a7c960bf.tar.gz openblas-28e96458e5a4b2d8039ed16048a07892a7c960bf.tar.bz2 openblas-28e96458e5a4b2d8039ed16048a07892a7c960bf.zip | |
Replace vpermpd with vpermilpd
to improve performance on Zen/Zen2 (as demonstrated by wjc404 in #2180)
| -rw-r--r-- | kernel/x86_64/zdot_microk_haswell-2.c | 24 |
1 files changed, 16 insertions, 8 deletions
diff --git a/kernel/x86_64/zdot_microk_haswell-2.c b/kernel/x86_64/zdot_microk_haswell-2.c index 9f2fc2c1d..4eade7bfd 100644 --- a/kernel/x86_64/zdot_microk_haswell-2.c +++ b/kernel/x86_64/zdot_microk_haswell-2.c @@ -66,13 +66,17 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot) "vfmadd231pd %%ymm8 , %%ymm12, %%ymm0 \n\t" // x_r * y_r, x_i * y_i "vfmadd231pd %%ymm9 , %%ymm13, %%ymm1 \n\t" // x_r * y_r, x_i * y_i - "vpermpd $0xb1 , %%ymm12, %%ymm12 \n\t" - "vpermpd $0xb1 , %%ymm13, %%ymm13 \n\t" + "vpermilpd $0x05 , %%ymm12, %%ymm12 \n\t" + "vpermilpd $0x05 , %%ymm13, %%ymm13 \n\t" +// "vpermpd $0xb1 , %%ymm12, %%ymm12 \n\t" +// "vpermpd $0xb1 , %%ymm13, %%ymm13 \n\t" "vfmadd231pd %%ymm10, %%ymm14, %%ymm2 \n\t" // x_r * y_r, x_i * y_i "vfmadd231pd %%ymm11, %%ymm15, %%ymm3 \n\t" // x_r * y_r, x_i * y_i - "vpermpd $0xb1 , %%ymm14, %%ymm14 \n\t" - "vpermpd $0xb1 , %%ymm15, %%ymm15 \n\t" + "vpermilpd $0x05 , %%ymm14, %%ymm14 \n\t" + "vpermilpd $0x05 , %%ymm15, %%ymm15 \n\t" +// "vpermpd $0xb1 , %%ymm14, %%ymm14 \n\t" +// "vpermpd $0xb1 , %%ymm15, %%ymm15 \n\t" "vfmadd231pd %%ymm8 , %%ymm12, %%ymm4 \n\t" // x_r * y_i, x_i * y_r "addq $16 , %0 \n\t" @@ -151,13 +155,17 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot) "vfmadd231pd %%ymm8 , %%ymm12, %%ymm0 \n\t" // x_r * y_r, x_i * y_i "vfmadd231pd %%ymm9 , %%ymm13, %%ymm1 \n\t" // x_r * y_r, x_i * y_i - "vpermpd $0xb1 , %%ymm12, %%ymm12 \n\t" - "vpermpd $0xb1 , %%ymm13, %%ymm13 \n\t" + "vpermilpd $0x05 , %%ymm12, %%ymm12 \n\t" + "vpermilpd $0x05 , %%ymm13, %%ymm13 \n\t" +// "vpermpd $0xb1 , %%ymm12, %%ymm12 \n\t" +// "vpermpd $0xb1 , %%ymm13, %%ymm13 \n\t" "vfmadd231pd %%ymm10, %%ymm14, %%ymm2 \n\t" // x_r * y_r, x_i * y_i "vfmadd231pd %%ymm11, %%ymm15, %%ymm3 \n\t" // x_r * y_r, x_i * y_i - "vpermpd $0xb1 , %%ymm14, %%ymm14 \n\t" - "vpermpd $0xb1 , %%ymm15, %%ymm15 \n\t" + "vpermilpd $0x05 , %%ymm14, %%ymm14 \n\t" + "vpermilpd $0x05 , %%ymm15, %%ymm15 \n\t" +// "vpermpd $0xb1 , %%ymm14, %%ymm14 \n\t" +// "vpermpd $0xb1 , %%ymm15, %%ymm15 \n\t" "vfmadd231pd %%ymm8 , %%ymm12, %%ymm4 \n\t" // x_r * y_i, x_i * y_r "addq $16 , %0 \n\t" |