powerpc: Optimized SHGEMM kernel for POWER10

This patch introduces new optimized version of SHGEMM kernel
using power10 Matrix-Multiply Assist (MMA) feature introduced in
POWER ISA v3.1. This patch makes use of new POWER10 compute instructions
for matrix multiplication operation.

Tested on simulator and there are no new test failures.
This commit is contained in:
Rajalakshmi Srinivasaraghavan 2020-06-25 22:19:08 -05:00
parent dfe819f3bd
commit d23419accc
7 changed files with 1137 additions and 69 deletions

View File

@ -39,24 +39,24 @@
#include <stdio.h>
#include "common.h"
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
BLASLONG i, j;
FLOAT *aoffset;
FLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
FLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
FLOAT *aoffset9, *aoffset10, *aoffset11, *aoffset12;
FLOAT *aoffset13, *aoffset14, *aoffset15, *aoffset16;
IFLOAT *aoffset;
IFLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
IFLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
IFLOAT *aoffset9, *aoffset10, *aoffset11, *aoffset12;
IFLOAT *aoffset13, *aoffset14, *aoffset15, *aoffset16;
FLOAT *boffset;
FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
IFLOAT *boffset;
IFLOAT ctemp01, ctemp02, ctemp03, ctemp04;
IFLOAT ctemp05, ctemp06, ctemp07, ctemp08;
IFLOAT ctemp09, ctemp10, ctemp11, ctemp12;
IFLOAT ctemp13, ctemp14, ctemp15, ctemp16;
IFLOAT ctemp17, ctemp18, ctemp19, ctemp20;
IFLOAT ctemp21, ctemp22, ctemp23, ctemp24;
IFLOAT ctemp25, ctemp26, ctemp27, ctemp28;
IFLOAT ctemp29, ctemp30, ctemp31, ctemp32;
aoffset = a;
boffset = b;

View File

@ -39,30 +39,30 @@
#include <stdio.h>
#include "common.h"
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
BLASLONG i, j;
FLOAT *aoffset;
FLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
FLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
IFLOAT *aoffset;
IFLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
IFLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
FLOAT *boffset;
FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
FLOAT ctemp33, ctemp34, ctemp35, ctemp36;
FLOAT ctemp37, ctemp38, ctemp39, ctemp40;
FLOAT ctemp41, ctemp42, ctemp43, ctemp44;
FLOAT ctemp45, ctemp46, ctemp47, ctemp48;
FLOAT ctemp49, ctemp50, ctemp51, ctemp52;
FLOAT ctemp53, ctemp54, ctemp55, ctemp56;
FLOAT ctemp57, ctemp58, ctemp59, ctemp60;
FLOAT ctemp61, ctemp62, ctemp63, ctemp64;
IFLOAT *boffset;
IFLOAT ctemp01, ctemp02, ctemp03, ctemp04;
IFLOAT ctemp05, ctemp06, ctemp07, ctemp08;
IFLOAT ctemp09, ctemp10, ctemp11, ctemp12;
IFLOAT ctemp13, ctemp14, ctemp15, ctemp16;
IFLOAT ctemp17, ctemp18, ctemp19, ctemp20;
IFLOAT ctemp21, ctemp22, ctemp23, ctemp24;
IFLOAT ctemp25, ctemp26, ctemp27, ctemp28;
IFLOAT ctemp29, ctemp30, ctemp31, ctemp32;
IFLOAT ctemp33, ctemp34, ctemp35, ctemp36;
IFLOAT ctemp37, ctemp38, ctemp39, ctemp40;
IFLOAT ctemp41, ctemp42, ctemp43, ctemp44;
IFLOAT ctemp45, ctemp46, ctemp47, ctemp48;
IFLOAT ctemp49, ctemp50, ctemp51, ctemp52;
IFLOAT ctemp53, ctemp54, ctemp55, ctemp56;
IFLOAT ctemp57, ctemp58, ctemp59, ctemp60;
IFLOAT ctemp61, ctemp62, ctemp63, ctemp64;
aoffset = a;

View File

@ -39,22 +39,22 @@
#include <stdio.h>
#include "common.h"
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
BLASLONG i, j;
FLOAT *aoffset;
FLOAT *aoffset1, *aoffset2;
FLOAT *boffset;
IFLOAT *aoffset;
IFLOAT *aoffset1, *aoffset2;
IFLOAT *boffset;
FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
IFLOAT ctemp01, ctemp02, ctemp03, ctemp04;
IFLOAT ctemp05, ctemp06, ctemp07, ctemp08;
IFLOAT ctemp09, ctemp10, ctemp11, ctemp12;
IFLOAT ctemp13, ctemp14, ctemp15, ctemp16;
IFLOAT ctemp17, ctemp18, ctemp19, ctemp20;
IFLOAT ctemp21, ctemp22, ctemp23, ctemp24;
IFLOAT ctemp25, ctemp26, ctemp27, ctemp28;
IFLOAT ctemp29, ctemp30, ctemp31, ctemp32;
aoffset = a;
boffset = b;

View File

@ -39,32 +39,32 @@
#include <stdio.h>
#include "common.h"
int CNAME(BLASLONG m, BLASLONG n, FLOAT *a, BLASLONG lda, FLOAT *b){
int CNAME(BLASLONG m, BLASLONG n, IFLOAT *a, BLASLONG lda, IFLOAT *b){
BLASLONG i, j;
FLOAT *aoffset;
FLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
FLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
IFLOAT *aoffset;
IFLOAT *aoffset1, *aoffset2, *aoffset3, *aoffset4;
IFLOAT *aoffset5, *aoffset6, *aoffset7, *aoffset8;
FLOAT *boffset, *boffset1, *boffset2, *boffset3, *boffset4;
IFLOAT *boffset, *boffset1, *boffset2, *boffset3, *boffset4;
FLOAT ctemp01, ctemp02, ctemp03, ctemp04;
FLOAT ctemp05, ctemp06, ctemp07, ctemp08;
FLOAT ctemp09, ctemp10, ctemp11, ctemp12;
FLOAT ctemp13, ctemp14, ctemp15, ctemp16;
FLOAT ctemp17, ctemp18, ctemp19, ctemp20;
FLOAT ctemp21, ctemp22, ctemp23, ctemp24;
FLOAT ctemp25, ctemp26, ctemp27, ctemp28;
FLOAT ctemp29, ctemp30, ctemp31, ctemp32;
FLOAT ctemp33, ctemp34, ctemp35, ctemp36;
FLOAT ctemp37, ctemp38, ctemp39, ctemp40;
FLOAT ctemp41, ctemp42, ctemp43, ctemp44;
FLOAT ctemp45, ctemp46, ctemp47, ctemp48;
FLOAT ctemp49, ctemp50, ctemp51, ctemp52;
FLOAT ctemp53, ctemp54, ctemp55, ctemp56;
FLOAT ctemp57, ctemp58, ctemp59, ctemp60;
FLOAT ctemp61, ctemp62, ctemp63, ctemp64;
IFLOAT ctemp01, ctemp02, ctemp03, ctemp04;
IFLOAT ctemp05, ctemp06, ctemp07, ctemp08;
IFLOAT ctemp09, ctemp10, ctemp11, ctemp12;
IFLOAT ctemp13, ctemp14, ctemp15, ctemp16;
IFLOAT ctemp17, ctemp18, ctemp19, ctemp20;
IFLOAT ctemp21, ctemp22, ctemp23, ctemp24;
IFLOAT ctemp25, ctemp26, ctemp27, ctemp28;
IFLOAT ctemp29, ctemp30, ctemp31, ctemp32;
IFLOAT ctemp33, ctemp34, ctemp35, ctemp36;
IFLOAT ctemp37, ctemp38, ctemp39, ctemp40;
IFLOAT ctemp41, ctemp42, ctemp43, ctemp44;
IFLOAT ctemp45, ctemp46, ctemp47, ctemp48;
IFLOAT ctemp49, ctemp50, ctemp51, ctemp52;
IFLOAT ctemp53, ctemp54, ctemp55, ctemp56;
IFLOAT ctemp57, ctemp58, ctemp59, ctemp60;
IFLOAT ctemp61, ctemp62, ctemp63, ctemp64;
aoffset = a;
boffset = b;

View File

@ -7,6 +7,17 @@ else
#CGEMM_BETA = ../generic/zgemm_beta.c
#ZGEMM_BETA = ../generic/zgemm_beta.c
SHGEMM_BETA = ../generic/gemm_beta.c
SHGEMMKERNEL = shgemm_kernel_power10.c
SHGEMMINCOPY = ../generic/gemm_ncopy_16.c
SHGEMMITCOPY = ../generic/gemm_tcopy_16.c
SHGEMMONCOPY = ../generic/gemm_ncopy_8.c
SHGEMMOTCOPY = ../generic/gemm_tcopy_8.c
SHGEMMINCOPYOBJ = shgemm_incopy$(TSUFFIX).$(SUFFIX)
SHGEMMITCOPYOBJ = shgemm_itcopy$(TSUFFIX).$(SUFFIX)
SHGEMMONCOPYOBJ = shgemm_oncopy$(TSUFFIX).$(SUFFIX)
SHGEMMOTCOPYOBJ = shgemm_otcopy$(TSUFFIX).$(SUFFIX)
STRMMKERNEL = sgemm_kernel_power10.c
DTRMMKERNEL = dgemm_kernel_power10.c
CTRMMKERNEL = cgemm_kernel_power10.S

File diff suppressed because it is too large Load Diff

13
param.h
View File

@ -2297,6 +2297,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#endif
#if defined(POWER10)
#undef SHGEMM_DEFAULT_UNROLL_N
#undef SHGEMM_DEFAULT_UNROLL_M
#undef SHGEMM_DEFAULT_P
#undef SHGEMM_DEFAULT_R
#undef SHGEMM_DEFAULT_Q
#define SHGEMM_DEFAULT_UNROLL_M 16
#define SHGEMM_DEFAULT_UNROLL_N 8
#define SHGEMM_DEFAULT_P 832
#define SHGEMM_DEFAULT_Q 1026
#define SHGEMM_DEFAULT_R 4096
#endif
#if defined(SPARC) && defined(V7)
#define SNUMOPT 4