Include shgemm in benchtest

This patch is to enable benchtest for half precision gemm
when BUILD_HALF is set during make.

2 files changed, 26 insertions, 7 deletions

diff --git a/benchmark/Makefile b/benchmark/Makefile
index 90d903ad7..53f422be4 100644
--- a/benchmark/Makefile
+++ b/benchmark/Makefile
@@ -49,6 +49,12 @@ else
 GOTO_LAPACK_TARGETS=

 endif

 

+ifeq ($(BUILD_HALF),1)

+GOTO_HALF_TARGETS=shgemm.goto

+else

+GOTO_HALF_TARGETS=

+endif

+

 ifeq ($(OSNAME), WINNT)

 

 goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \

@@ -91,7 +97,7 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        sgetri.goto dgetri.goto cgetri.goto zgetri.goto \

        spotrf.goto dpotrf.goto cpotrf.goto zpotrf.goto \

        ssymm.goto dsymm.goto csymm.goto zsymm.goto \

-       saxpby.goto daxpby.goto caxpby.goto zaxpby.goto

+       saxpby.goto daxpby.goto caxpby.goto zaxpby.goto $(GOTO_HALF_TARGETS)

 

 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \

        scholesky.acml dcholesky.acml ccholesky.acml zcholesky.acml \

@@ -264,7 +270,7 @@ goto :: sgemm.goto dgemm.goto cgemm.goto zgemm.goto \
        samin.goto damin.goto camin.goto zamin.goto \

        smin.goto dmin.goto \

        saxpby.goto daxpby.goto caxpby.goto zaxpby.goto \

-       snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS)

+       snrm2.goto dnrm2.goto scnrm2.goto dznrm2.goto $(GOTO_LAPACK_TARGETS) $(GOTO_HALF_TARGETS)

 

 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \

        scholesky.acml dcholesky.acml ccholesky.acml zcholesky.acml \

@@ -614,6 +620,11 @@ zcholesky.essl : zcholesky.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBESSL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)

 

 ##################################### Sgemm ####################################################

+ifeq ($(BUILD_HALF),1)

+shgemm.goto : shgemm.$(SUFFIX) ../$(LIBNAME)

+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm

+endif

+

 sgemm.goto : sgemm.$(SUFFIX) ../$(LIBNAME)

 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm

 

@@ -2916,6 +2927,11 @@ ccholesky.$(SUFFIX) : cholesky.c
 zcholesky.$(SUFFIX) : cholesky.c

 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^

 

+ifeq ($(BUILD_HALF),1)

+shgemm.$(SUFFIX) : gemm.c

+	$(CC) $(CFLAGS) -c -DHALF -UCOMPLEX -UDOUBLE -o $(@F) $^

+endif

+

 sgemm.$(SUFFIX) : gemm.c

 	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^

 

diff --git a/benchmark/gemm.c b/benchmark/gemm.c
index dd016a7c3..d2235330b 100644
--- a/benchmark/gemm.c
+++ b/benchmark/gemm.c
@@ -39,6 +39,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #ifdef DOUBLE
 #define GEMM   BLASFUNC(dgemm)
+#elif defined(HALF)
+#define GEMM   BLASFUNC(shgemm)
 #else
 #define GEMM   BLASFUNC(sgemm)
 #endif
@@ -120,7 +122,8 @@ static void *huge_malloc(BLASLONG size){
 
 int main(int argc, char *argv[]){
 
-  FLOAT *a, *b, *c;
+  IFLOAT *a, *b;
+  FLOAT *c;
   FLOAT alpha[] = {1.0, 0.0};
   FLOAT beta [] = {0.0, 0.0};
   char transa = 'N';
@@ -184,10 +187,10 @@ int main(int argc, char *argv[]){
     k = to;
   }
 
-  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * m * k * COMPSIZE)) == NULL) {
+  if (( a = (IFLOAT *)malloc(sizeof(IFLOAT) * m * k * COMPSIZE)) == NULL) {
     fprintf(stderr,"Out of Memory!!\n");exit(1);
   }
-  if (( b = (FLOAT *)malloc(sizeof(FLOAT) * k * n * COMPSIZE)) == NULL) {
+  if (( b = (IFLOAT *)malloc(sizeof(IFLOAT) * k * n * COMPSIZE)) == NULL) {
     fprintf(stderr,"Out of Memory!!\n");exit(1);
   }
   if (( c = (FLOAT *)malloc(sizeof(FLOAT) * m * n * COMPSIZE)) == NULL) {
@@ -199,10 +202,10 @@ int main(int argc, char *argv[]){
 #endif
 
   for (i = 0; i < m * k * COMPSIZE; i++) {
-    a[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+    a[i] = ((IFLOAT) rand() / (IFLOAT) RAND_MAX) - 0.5;
   }
   for (i = 0; i < k * n * COMPSIZE; i++) {
-    b[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+    b[i] = ((IFLOAT) rand() / (IFLOAT) RAND_MAX) - 0.5;
   }
   for (i = 0; i < m * n * COMPSIZE; i++) {
     c[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;