Merge pull request #53 from xianyi/develop

rebase

.gitignore

@@ -70,6 +70,7 @@ test/SBLAT2.SUMM
 test/SBLAT3.SUMM
 test/ZBLAT2.SUMM
 test/ZBLAT3.SUMM
+test/SHBLAT3.SUMM
 test/cblat1
 test/cblat2
 test/cblat3
@@ -79,6 +80,7 @@ test/dblat3
 test/sblat1
 test/sblat2
 test/sblat3
+test/test_shgemm
 test/zblat1
 test/zblat2
 test/zblat3

CMakeLists.txt

@@ -86,10 +86,13 @@ if (NOT NO_LAPACK)
   list(APPEND SUBDIRS lapack)
 endif ()
 
+if (NOT DEFINED BUILD_HALF)
+ set (BUILD_HALF false)
+endif ()
 # set which float types we want to build for
 if (NOT DEFINED BUILD_SINGLE AND NOT DEFINED BUILD_DOUBLE AND NOT DEFINED BUILD_COMPLEX AND NOT DEFINED BUILD_COMPLEX16)
   # if none are defined, build for all
-  set(BUILD_HALF true)
+#  set(BUILD_HALF true)
   set(BUILD_SINGLE true)
   set(BUILD_DOUBLE true)
   set(BUILD_COMPLEX true)
@@ -121,7 +124,7 @@ if (BUILD_COMPLEX16)
   list(APPEND FLOAT_TYPES "ZCOMPLEX") # defines COMPLEX and DOUBLE
 endif ()
 
-if (BUILD_SINGLE OR BUILD_HALF)
+if (BUILD_HALF)
   message(STATUS "Building Half Precision")
   list(APPEND FLOAT_TYPES "HALF") # defines nothing
 endif ()

Makefile.rule

@@ -273,6 +273,9 @@ COMMON_PROF = -pg
 #
 # CPP_THREAD_SAFETY_TEST = 1
 
+
+# If you want to enable the experimental BFLOAT16 support
+# BUILD_HALF = 1
 #
 #  End of user configuration
 #

Makefile.system

@@ -1124,6 +1124,10 @@ ifeq ($(USE_TLS), 1)
 CCOMMON_OPT += -DUSE_TLS
 endif
 
+ifeq ($(BUILD_HALF), 1)
+CCOMMON_OPT += -DBUILD_HALF
+endif
+
 CCOMMON_OPT += -DVERSION=\"$(VERSION)\"
 
 ifndef SYMBOLPREFIX
@@ -1395,6 +1399,7 @@ export KERNELDIR
 export FUNCTION_PROFILE
 export TARGET_CORE
 export NO_AVX512
+export BUILD_HALF
 
 export SHGEMM_UNROLL_M
 export SHGEMM_UNROLL_N

cmake/kernel.cmake

@@ -113,6 +113,7 @@ macro(SetDefaultL1)
   set(ZSUMKERNEL zsum.S)
   set(QSUMKERNEL sum.S)
   set(XSUMKERNEL zsum.S)
+if (BUILD_HALF)
   set(SHAMINKERNEL ../arm/amin.c)
   set(SHAMAXKERNEL ../arm/amax.c)
   set(SHMAXKERNEL ../arm/max.c)
@@ -131,6 +132,7 @@ macro(SetDefaultL1)
   set(SHNRM2KERNEL ../arm/nrm2.c)
   set(SHSUMKERNEL ../arm/sum.c)
   set(SHSWAPKERNEL ../arm/swap.c)
+endif ()
 endmacro ()
 
 macro(SetDefaultL2)
@@ -179,10 +181,11 @@ macro(SetDefaultL2)
   set(XHEMV_L_KERNEL ../generic/zhemv_k.c)
   set(XHEMV_V_KERNEL ../generic/zhemv_k.c)
   set(XHEMV_M_KERNEL ../generic/zhemv_k.c)
+if (BUILD_HALF)
   set(SHGEMVNKERNEL ../arm/gemv_n.c)
   set(SHGEMVTKERNEL ../arm/gemv_t.c)
   set(SHGERKERNEL ../generic/ger.c)
-
+endif ()
 endmacro ()
 
 macro(SetDefaultL3)
@@ -190,6 +193,7 @@ macro(SetDefaultL3)
   set(DGEADD_KERNEL ../generic/geadd.c)
   set(CGEADD_KERNEL ../generic/zgeadd.c)
   set(ZGEADD_KERNEL ../generic/zgeadd.c)
+if (BUILD_HALF)
   set(SHGEADD_KERNEL ../generic/geadd.c)
   set(SHGEMMKERNEL ../generic/gemmkernel_2x2.c)
   set(SHGEMM_BETA  ../generic/gemm_beta.c)
@@ -201,6 +205,6 @@ macro(SetDefaultL3)
   set(SHGEMMITCOPYOBJ shgemm_itcopy.o)
   set(SHGEMMONCOPYOBJ shgemm_oncopy.o)
   set(SHGEMMOTCOPYOBJ shgemm_otcopy.o)
-
+endif ()
 
 endmacro ()

common_param.h

@@ -47,7 +47,7 @@ typedef struct {
   int dtb_entries;
   int offsetA, offsetB, align;
 
-#if 1
+#ifdef BUILD_HALF
   int shgemm_p, shgemm_q, shgemm_r;
   int shgemm_unroll_m, shgemm_unroll_n, shgemm_unroll_mn;
 
@@ -1002,12 +1002,14 @@ extern gotoblas_t *gotoblas;
 
 #define HAVE_EX_L2	gotoblas -> exclusive_cache
 
+#ifdef BUILD_HALF
 #define	SHGEMM_P		gotoblas -> shgemm_p
 #define	SHGEMM_Q		gotoblas -> shgemm_q
 #define	SHGEMM_R		gotoblas -> shgemm_r
 #define	SHGEMM_UNROLL_M	gotoblas -> shgemm_unroll_m
 #define	SHGEMM_UNROLL_N	gotoblas -> shgemm_unroll_n
 #define	SHGEMM_UNROLL_MN	gotoblas -> shgemm_unroll_mn
+#endif
 
 #define	SGEMM_P		gotoblas -> sgemm_p
 #define	SGEMM_Q		gotoblas -> sgemm_q
@@ -1086,6 +1088,7 @@ extern gotoblas_t *gotoblas;
 #define HAVE_EX_L2	0
 #endif
 
+#ifdef BUILD_HALF
 #define	SHGEMM_P		SHGEMM_DEFAULT_P
 #define	SHGEMM_Q		SHGEMM_DEFAULT_Q
 #define	SHGEMM_R		SHGEMM_DEFAULT_R
@@ -1096,6 +1099,7 @@ extern gotoblas_t *gotoblas;
 #else
 #define SHGEMM_UNROLL_MN	MAX((SHGEMM_UNROLL_M), (SHGEMM_UNROLL_N))
 #endif
+#endif
 
 #define	SGEMM_P		SGEMM_DEFAULT_P
 #define	SGEMM_Q		SGEMM_DEFAULT_Q
@@ -1330,31 +1334,31 @@ extern gotoblas_t *gotoblas;
 #endif
 
 #ifndef SHGEMM_DEFAULT_R
-#define SHGEMM_DEFAULT_R (((BUFFER_SIZE - ((SHGEMM_DEFAULT_P * SHGEMM_DEFAULT_Q *  4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SHGEMM_DEFAULT_Q *  4) - 15) & ~15)
+#define SHGEMM_DEFAULT_R (((BUFFER_SIZE - ((SHGEMM_DEFAULT_P * SHGEMM_DEFAULT_Q *  4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SHGEMM_DEFAULT_Q *  4) - 15) & ~15UL)
 #endif
 
 #ifndef SGEMM_DEFAULT_R
-#define SGEMM_DEFAULT_R (((BUFFER_SIZE - ((SGEMM_DEFAULT_P * SGEMM_DEFAULT_Q *  4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SGEMM_DEFAULT_Q *  4) - 15) & ~15)
+#define SGEMM_DEFAULT_R (((BUFFER_SIZE - ((SGEMM_DEFAULT_P * SGEMM_DEFAULT_Q *  4 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (SGEMM_DEFAULT_Q *  4) - 15) & ~15UL)
 #endif
 
 #ifndef DGEMM_DEFAULT_R
-#define DGEMM_DEFAULT_R (((BUFFER_SIZE - ((DGEMM_DEFAULT_P * DGEMM_DEFAULT_Q *  8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (DGEMM_DEFAULT_Q *  8) - 15) & ~15)
+#define DGEMM_DEFAULT_R (((BUFFER_SIZE - ((DGEMM_DEFAULT_P * DGEMM_DEFAULT_Q *  8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (DGEMM_DEFAULT_Q *  8) - 15) & ~15UL)
 #endif
 
 #ifndef QGEMM_DEFAULT_R
-#define QGEMM_DEFAULT_R (((BUFFER_SIZE - ((QGEMM_DEFAULT_P * QGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (QGEMM_DEFAULT_Q * 16) - 15) & ~15)
+#define QGEMM_DEFAULT_R (((BUFFER_SIZE - ((QGEMM_DEFAULT_P * QGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (QGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
 #endif
 
 #ifndef CGEMM_DEFAULT_R
-#define CGEMM_DEFAULT_R (((BUFFER_SIZE - ((CGEMM_DEFAULT_P * CGEMM_DEFAULT_Q *  8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (CGEMM_DEFAULT_Q *  8) - 15) & ~15)
+#define CGEMM_DEFAULT_R (((BUFFER_SIZE - ((CGEMM_DEFAULT_P * CGEMM_DEFAULT_Q *  8 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (CGEMM_DEFAULT_Q *  8) - 15) & ~15UL)
 #endif
 
 #ifndef ZGEMM_DEFAULT_R
-#define ZGEMM_DEFAULT_R (((BUFFER_SIZE - ((ZGEMM_DEFAULT_P * ZGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (ZGEMM_DEFAULT_Q * 16) - 15) & ~15)
+#define ZGEMM_DEFAULT_R (((BUFFER_SIZE - ((ZGEMM_DEFAULT_P * ZGEMM_DEFAULT_Q * 16 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (ZGEMM_DEFAULT_Q * 16) - 15) & ~15UL)
 #endif
 
 #ifndef XGEMM_DEFAULT_R
-#define XGEMM_DEFAULT_R (((BUFFER_SIZE - ((XGEMM_DEFAULT_P * XGEMM_DEFAULT_Q * 32 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (XGEMM_DEFAULT_Q * 32) - 15) & ~15)
+#define XGEMM_DEFAULT_R (((BUFFER_SIZE - ((XGEMM_DEFAULT_P * XGEMM_DEFAULT_Q * 32 + GEMM_DEFAULT_OFFSET_A + GEMM_DEFAULT_ALIGN) & ~GEMM_DEFAULT_ALIGN)) / (XGEMM_DEFAULT_Q * 32) - 15) & ~15UL)
 #endif
 
 #ifndef SNUMOPT

driver/level3/Makefile

@@ -19,7 +19,10 @@ ifeq ($(ARCH), MIPS)
 USE_GEMM3M = 1
 endif
 
+ifeq ($(BUILD_HALF),1)
 SHBLASOBJS       += shgemm_nn.$(SUFFIX) shgemm_nt.$(SUFFIX) shgemm_tn.$(SUFFIX) shgemm_tt.$(SUFFIX)
+endif
+
 SBLASOBJS	+= \
 	sgemm_nn.$(SUFFIX) sgemm_nt.$(SUFFIX) sgemm_tn.$(SUFFIX) sgemm_tt.$(SUFFIX) \
 	strmm_LNUU.$(SUFFIX) strmm_LNUN.$(SUFFIX) strmm_LNLU.$(SUFFIX) strmm_LNLN.$(SUFFIX) \
@@ -204,8 +207,9 @@ COMMONOBJS  += gemm_thread_m.$(SUFFIX) gemm_thread_n.$(SUFFIX) gemm_thread_mn.$(
 COMMONOBJS  += syrk_thread.$(SUFFIX)
 
 ifndef USE_SIMPLE_THREADED_LEVEL3
-
+ifeq ($(BUILD_HALF),1)
 SHBLASOBJS    += shgemm_thread_nn.$(SUFFIX) shgemm_thread_nt.$(SUFFIX) shgemm_thread_tn.$(SUFFIX) shgemm_thread_tt.$(SUFFIX)
+endif
 SBLASOBJS    += sgemm_thread_nn.$(SUFFIX) sgemm_thread_nt.$(SUFFIX) sgemm_thread_tn.$(SUFFIX) sgemm_thread_tt.$(SUFFIX)
 DBLASOBJS    += dgemm_thread_nn.$(SUFFIX) dgemm_thread_nt.$(SUFFIX) dgemm_thread_tn.$(SUFFIX) dgemm_thread_tt.$(SUFFIX)
 QBLASOBJS    += qgemm_thread_nn.$(SUFFIX) qgemm_thread_nt.$(SUFFIX) qgemm_thread_tn.$(SUFFIX) qgemm_thread_tt.$(SUFFIX)

driver/others/blas_server.c

@@ -272,7 +272,7 @@ static void legacy_exec(void *func, int mode, blas_arg_t *args, void *sb){
       }
 }
 
-#if defined(OS_LINUX) && !defined(NO_AFFINITY)
+#if defined(OS_LINUX) && !defined(NO_AFFINITY) 
 int gotoblas_set_affinity(int);
 int gotoblas_set_affinity2(int);
 int get_node(void);
@@ -281,6 +281,8 @@ int get_node(void);
 static int increased_threads = 0;
 
 #ifdef OS_LINUX
+extern int openblas_get_num_threads(void);  
+
 int openblas_setaffinity(int thread_idx, size_t cpusetsize, cpu_set_t* cpu_set) {
   const int active_threads = openblas_get_num_threads();
 
@@ -602,7 +604,7 @@ int blas_thread_init(void){
       if(ret!=0){
 	struct rlimit rlim;
         const char *msg = strerror(ret);
-        fprintf(STDERR, "OpenBLAS blas_thread_init: pthread_create failed for thread %ld of %ld: %s\n", i+1,blas_num_threads,msg);
+        fprintf(STDERR, "OpenBLAS blas_thread_init: pthread_create failed for thread %ld of %d: %s\n", i+1,blas_num_threads,msg);
 #ifdef RLIMIT_NPROC
         if(0 == getrlimit(RLIMIT_NPROC, &rlim)) {
           fprintf(STDERR, "OpenBLAS blas_thread_init: RLIMIT_NPROC "

driver/others/memory.c

@@ -2070,7 +2070,7 @@ if (!release->address) return;
   if (munmap(release -> address, BUFFER_SIZE)) {
       int errsv=errno;
        perror("OpenBLAS : munmap failed:");
-       printf("error code=%d,\trelease->address=%lx\n",errsv,release->address);
+       printf("error code=%d,\trelease->address=%p\n",errsv,release->address);
   }
 }
 

exports/Makefile

@@ -30,6 +30,10 @@ ifndef BUILD_LAPACK_DEPRECATED
 BUILD_LAPACK_DEPRECATED = 0
 endif
 
+ifndef BUILD_HALF
+BUILD_HALF = 0
+endif
+
 ifeq ($(OSNAME), WINNT)
 ifeq ($(F_COMPILER), GFORTRAN)
 ifndef ONLY_CBLAS
@@ -234,23 +238,23 @@ static : ../$(LIBNAME)
 	rm -f goto.$(SUFFIX)
 
 osx.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol osx $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
+	perl ./gensymbol osx $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > $(@F)
 
 aix.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol aix $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
+	perl ./gensymbol aix $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > $(@F)
 
 objcopy.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol objcopy $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
+	perl ./gensymbol objcopy $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > $(@F)
 
 objconv.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol objconv $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
+	perl ./gensymbol objconv $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > $(@F)
 
 test : linktest.c
 	$(CC) $(CFLAGS) $(LDFLAGS) -w -o linktest linktest.c ../$(LIBSONAME) -lm && echo OK.
 	rm -f linktest
 
 linktest.c : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol linktest  $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > linktest.c
+	perl ./gensymbol linktest  $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) $(BUILD_HALF) > linktest.c
 
 clean ::
 	@rm -f *.def *.dylib __.SYMDEF* *.renamed

exports/gensymbol

@@ -30,7 +30,7 @@
     icamax,icamin,idamax,idamin,idmax,idmin,isamax,isamin,ismax,ismin,
     izamax,izamin,lsame,samax,samin,sasum,saxpy,scabs1,scamax,
     scamin,scasum,scnrm2,scopy,sdot,sdsdot,sgbmv,sgemm,sgemv,sger,
-    shgemm, smax,smin,snrm2,
+    smax,smin,snrm2,
     srot,srotg,srotm,srotmg,ssbmv,sscal,sspmv,sspr2,sspr,sswap,
     ssymm,ssymv,ssyr2,ssyr2k,ssyr,ssyrk,stbmv,stbsv,stpmv,stpsv,
     strmm,strmv,strsm,strsv,zaxpy,zcopy,zdotc,zdotu,zdrot,
@@ -51,6 +51,7 @@
     zimatcopy,
 );
 
+@halfblasobjs = (shgemm);
 @cblasobjs = (
     cblas_caxpy, cblas_ccopy, cblas_cdotc, cblas_cdotu, cblas_cgbmv, cblas_cgemm, cblas_cgemv,
     cblas_cgerc, cblas_cgeru, cblas_chbmv, cblas_chemm, cblas_chemv, cblas_cher2, cblas_cher2k,
@@ -67,7 +68,7 @@
     cblas_isamax, cblas_izamax,
     cblas_sasum, cblas_saxpy,
     cblas_scasum, cblas_scnrm2, cblas_scopy, cblas_sdot, cblas_sdsdot, cblas_sgbmv, cblas_sgemm,
-    cblas_sgemv, cblas_sger, cblas_shgemm, cblas_snrm2, cblas_srot, cblas_srotg,
+    cblas_sgemv, cblas_sger, cblas_snrm2, cblas_srot, cblas_srotg,
     cblas_srotm, cblas_srotmg, cblas_ssbmv, cblas_sscal, cblas_sspmv, cblas_sspr2, cblas_sspr,
     cblas_sswap, cblas_ssymm, cblas_ssymv, cblas_ssyr2, cblas_ssyr2k, cblas_ssyr, cblas_ssyrk,
     cblas_stbmv, cblas_stbsv, cblas_stpmv, cblas_stpsv, cblas_strmm, cblas_strmv, cblas_strsm,
@@ -83,6 +84,8 @@
     cblas_sgeadd, cblas_dgeadd,cblas_cgeadd, cblas_zgeadd
 );
 
+@halfcblasobjs = (cblas_shgemm);
+
 @exblasobjs = (
     qamax,qamin,qasum,qaxpy,qcabs1,qcopy,qdot,qgbmv,qgemm,
     qgemv,qger,qmax,qmin,
@@ -3454,6 +3457,10 @@ use File::Spec;
 use File::Basename;
 my $dirname = File::Spec->catfile(dirname(dirname(File::Spec->rel2abs(__FILE__))), "lapack-netlib");
 
+if ($ARGV[12] == 1) {
+	@blasobjs = (@blasobjs, @halfblasobjs);
+	@cblasobjs = (@cblasobjs, @halfcblasobjs);
+}
 if ($ARGV[8] == 1) {
     #ONLY_CBLAS=1
     @underscore_objs = (@misc_underscore_objs);

interface/Makefile

@@ -46,7 +46,9 @@ SBLAS3OBJS    = \
 		somatcopy.$(SUFFIX) simatcopy.$(SUFFIX)\
 		sgeadd.$(SUFFIX)
 
+ifeq ($(BUILD_HALF),1)
 SHBLAS3OBJS    = shgemm.$(SUFFIX)
+endif
 
 DBLAS1OBJS    = \
 		daxpy.$(SUFFIX) dswap.$(SUFFIX) \
@@ -278,7 +280,9 @@ CSBLAS3OBJS   = \
 	cblas_ssyrk.$(SUFFIX) cblas_ssyr2k.$(SUFFIX) cblas_somatcopy.$(SUFFIX)  cblas_simatcopy.$(SUFFIX)\
 	cblas_sgeadd.$(SUFFIX)
 
+ifeq ($(BUILD_HALF),1)
 CSHBLAS3OBJS = cblas_shgemm.$(SUFFIX)
+endif
 
 CDBLAS1OBJS   = \
 	cblas_idamax.$(SUFFIX) cblas_idamin.$(SUFFIX) cblas_dasum.$(SUFFIX) cblas_daxpy.$(SUFFIX) \
@@ -1214,8 +1218,10 @@ zhpr2.$(SUFFIX) zhpr2.$(PSUFFIX) : zhpr2.c
 xhpr2.$(SUFFIX) xhpr2.$(PSUFFIX) : zhpr2.c
 	$(CC) -c $(CFLAGS) $< -o $(@F)
 
+ifeq ($(BUILD_HALF),1)
 shgemm.$(SUFFIX) shgemm.$(PSUFFIX) : gemm.c ../param.h
 	$(CC) -c $(CFLAGS) $< -o $(@F)
+endif
 
 sgemm.$(SUFFIX) sgemm.$(PSUFFIX) : gemm.c ../param.h
 	$(CC) -c $(CFLAGS) $< -o $(@F)
@@ -1778,8 +1784,10 @@ cblas_zhemv.$(SUFFIX) cblas_zhemv.$(PSUFFIX) : zhemv.c
 cblas_sgemm.$(SUFFIX) cblas_sgemm.$(PSUFFIX) : gemm.c ../param.h
 	$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
 
+ifeq ($(BUILD_HALF),1)
 cblas_shgemm.$(SUFFIX) cblas_shgemm.$(PSUFFIX) : gemm.c ../param.h
 	$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)
+endif
 
 cblas_dgemm.$(SUFFIX) cblas_dgemm.$(PSUFFIX) : gemm.c ../param.h
 	$(CC) -DCBLAS -c $(CFLAGS) $< -o $(@F)

kernel/CMakeLists.txt

@@ -137,7 +137,11 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
     foreach (float_type SINGLE DOUBLE HALF)
       string(SUBSTRING ${float_type} 0 1 float_char)
       if (${float_type} STREQUAL "HALF")
-	set (float_char "SH")
+        if (NOT ${BUILD_HALF})
+	  continue ()
+        else ()
+	  set (float_char "SH")
+      endif ()
       endif ()
       GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMMKERNEL}" "" "gemm_kernel" false "" "" false ${float_type})
     endforeach()

kernel/Makefile.L3

@@ -59,7 +59,8 @@ ifeq ($(CORE), Z14)
 USE_TRMM = 1
 endif
 
-#ifndef SHGEMMKERNEL
+ifeq ($(BUILD_HALF), 1)
+ifndef SHGEMMKERNEL
 SHGEMM_BETA = ../generic/gemm_beta.c
 SHGEMMKERNEL    = ../generic/gemmkernel_2x2.c
 SHGEMMINCOPY    = ../generic/gemm_ncopy_2.c
@@ -70,12 +71,13 @@ SHGEMMINCOPYOBJ =  shgemm_incopy$(TSUFFIX).$(SUFFIX)
 SHGEMMITCOPYOBJ =  shgemm_itcopy$(TSUFFIX).$(SUFFIX)
 SHGEMMONCOPYOBJ =  shgemm_oncopy$(TSUFFIX).$(SUFFIX)
 SHGEMMOTCOPYOBJ =  shgemm_otcopy$(TSUFFIX).$(SUFFIX)
-#endif
+endif
 
 SHKERNELOBJS	+= \
 	shgemm_kernel$(TSUFFIX).$(SUFFIX) \
 	$(SHGEMMINCOPYOBJ) $(SHGEMMITCOPYOBJ) \
 	$(SHGEMMONCOPYOBJ) $(SHGEMMOTCOPYOBJ)
+endif
 
 SKERNELOBJS	+= \
 	sgemm_kernel$(TSUFFIX).$(SUFFIX) \
@@ -110,7 +112,9 @@ XKERNELOBJS	+= \
 	$(XGEMMINCOPYOBJ) $(XGEMMITCOPYOBJ) \
 	$(XGEMMONCOPYOBJ) $(XGEMMOTCOPYOBJ)
 
+ifeq ($(BUILD_HALF),1)
 SHBLASOBJS      += $(SHKERNELOBJS)
+endif
 SBLASOBJS	+= $(SKERNELOBJS)
 DBLASOBJS	+= $(DKERNELOBJS)
 QBLASOBJS	+= $(QKERNELOBJS)
@@ -118,7 +122,10 @@ CBLASOBJS	+= $(CKERNELOBJS)
 ZBLASOBJS	+= $(ZKERNELOBJS)
 XBLASOBJS	+= $(XKERNELOBJS)
 
+ifeq ($(BUILD_HALF),1)
 SHBLASOBJS += shgemm_beta$(TSUFFIX).$(SUFFIX)
+endif
+
 SBLASOBJS	+= \
 	sgemm_beta$(TSUFFIX).$(SUFFIX) \
 	strmm_kernel_LN$(TSUFFIX).$(SUFFIX) strmm_kernel_LT$(TSUFFIX).$(SUFFIX) \
@@ -408,11 +415,13 @@ ZBLASOBJS += \
 	zimatcopy_k_ctc$(TSUFFIX).$(SUFFIX) zimatcopy_k_rtc$(TSUFFIX).$(SUFFIX) \
 	zgeadd_k$(TSUFFIX).$(SUFFIX) 
 
-
+ifeq ($(BUILD_HALF), 1)
 SHGEMMINCOPYOBJ_P = $(SHGEMMINCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 SHGEMMITCOPYOBJ_P = $(SHGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 SHGEMMONCOPYOBJ_P = $(SHGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 SHGEMMOTCOPYOBJ_P = $(SHGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
+endif
+
 SGEMMINCOPYOBJ_P = $(SGEMMINCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 SGEMMITCOPYOBJ_P = $(SGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 SGEMMONCOPYOBJ_P = $(SGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
@@ -438,8 +447,10 @@ XGEMMITCOPYOBJ_P = $(XGEMMITCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 XGEMMONCOPYOBJ_P = $(XGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 XGEMMOTCOPYOBJ_P = $(XGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
 
+ifeq ($(BUILD_HALF),1)
 $(KDIR)shgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SHGEMM_BETA)
 	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
 
 $(KDIR)sgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_BETA)
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
@@ -459,10 +470,14 @@ $(KDIR)zgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMM_BETA)
 $(KDIR)xgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(XGEMM_BETA)
 	$(CC) $(CFLAGS) -c -DXDOUBLE -DCOMPLEX $< -o $@
 
+
+ifeq ($(BUILD_HALF), 1)
+
 $(KDIR)$(SHGEMMONCOPYOBJ) : $(KERNELDIR)/$(SHGEMMONCOPY)
 	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
 
 $(KDIR)$(SHGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SHGEMMOTCOPY)
+
 ifeq ($(OS), AIX)
 	$(CC) $(CFLAGS) -E -DHALF -UDOUBLE -UCOMPLEX $< -o shgemmotcopy.s
 	m4 shgemmotcopy.s > shgemmotcopy_nomacros.s
@@ -487,6 +502,7 @@ else
 	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
 endif
 
+endif
 endif
 
 $(KDIR)$(SGEMMONCOPYOBJ) : $(KERNELDIR)/$(SGEMMONCOPY)
@@ -646,6 +662,8 @@ else
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 endif
 
+ifeq ($(BUILD_HALF), 1)
+
 $(KDIR)shgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SHGEMMKERNEL) $(SHGEMMDEPEND)
 ifeq ($(OS), AIX)
 	$(CC) $(CFLAGS) -E -DHALF -UDOUBLE -UCOMPLEX  $< -o shgemm_kernel$(TSUFFIX).s
@@ -655,6 +673,7 @@ ifeq ($(OS), AIX)
 else
 	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
 endif
+endif
 
 $(KDIR)dgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMMKERNEL) $(DGEMMDEPEND)
 ifeq ($(OS), AIX)
@@ -2272,8 +2291,10 @@ $(KDIR)xtrsm_oltncopy$(TSUFFIX).$(SUFFIX) : generic/ztrsm_ltcopy_$(XGEMM_UNROLL_
 $(KDIR)sgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMM_BETA)
 	$(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 
+ifeq ($(BUILD_HALF),1)
 $(KDIR)shgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SHGEMM_BETA)
 	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
 
 $(KDIR)dgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(DGEMM_BETA)
 	$(CC) $(PFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
@@ -2290,6 +2311,8 @@ $(KDIR)zgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(ZGEMM_BETA)
 $(KDIR)xgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(XGEMM_BETA)
 	$(CC) $(PFLAGS) -c -DXDOUBLE -DCOMPLEX $< -o $@
 
+
+ifeq ($(BUILD_HALF), 1)
 $(SHGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMONCOPY)
 	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
 
@@ -2304,6 +2327,8 @@ $(SHGEMMITCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMITCOPY)
 	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
 
 endif
+endif
+
 $(SGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SGEMMONCOPY)
 	$(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 
@@ -2408,8 +2433,11 @@ endif
 
 endif
 
+
+ifeq ($(BUILD_HALF), 1)
 $(KDIR)shgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SHGEMMKERNEL) $(SHGEMMDEPEND)
 	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+endif
 
 $(KDIR)sgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL) $(SGEMMDEPEND)
 	$(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@

kernel/setparam-ref.c

@@ -53,6 +53,7 @@ gotoblas_t TABLE_NAME = {
 
   GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN,
 
+#ifdef BUILD_HALF
   0, 0, 0,
   SHGEMM_DEFAULT_UNROLL_M, SHGEMM_DEFAULT_UNROLL_N,
 #ifdef SHGEMM_DEFAULT_UNROLL_MN
@@ -109,7 +110,7 @@ gotoblas_t TABLE_NAME = {
 #else
   NULL,NULL,
 #endif
-
+#endif
 
   0, 0, 0,
   SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N,
@@ -706,19 +707,25 @@ gotoblas_t TABLE_NAME = {
 
 #if defined(ARCH_ARM64)
 static void init_parameter(void) {
+#if defined(BUILD_HALF)
   TABLE_NAME.shgemm_p = SHGEMM_DEFAULT_P;
+#endif
   TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
   TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
   TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
   TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
 
+#if defined(BUILD_HALF)
   TABLE_NAME.shgemm_q = SHGEMM_DEFAULT_Q;
+#endif
   TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
   TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
   TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
   TABLE_NAME.zgemm_q = ZGEMM_DEFAULT_Q;
 
+#if defined(BUILD_HALF)
   TABLE_NAME.shgemm_r = SHGEMM_DEFAULT_R;
+#endif
   TABLE_NAME.sgemm_r = SGEMM_DEFAULT_R;
   TABLE_NAME.dgemm_r = DGEMM_DEFAULT_R;
   TABLE_NAME.cgemm_r = CGEMM_DEFAULT_R;
@@ -782,20 +789,26 @@ static void init_parameter(void) {
 #if defined(ARCH_POWER)
 static void init_parameter(void) {
 
+#ifdef BUILD_HALF
   TABLE_NAME.shgemm_p = SHGEMM_DEFAULT_P;
+#endif
   TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
   TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
   TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
   TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
 
+#ifdef BUILD_HALF
   TABLE_NAME.shgemm_r = SHGEMM_DEFAULT_R;
+#endif
   TABLE_NAME.sgemm_r = SGEMM_DEFAULT_R;
   TABLE_NAME.dgemm_r = DGEMM_DEFAULT_R;
   TABLE_NAME.cgemm_r = CGEMM_DEFAULT_R;
   TABLE_NAME.zgemm_r = ZGEMM_DEFAULT_R;
 
 
+#ifdef BUILD_HALF
   TABLE_NAME.shgemm_q = SHGEMM_DEFAULT_Q;
+#endif
   TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
   TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
   TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
@@ -805,20 +818,26 @@ static void init_parameter(void) {
 
 #if defined(ARCH_ZARCH)
 static void init_parameter(void) {
+#ifdef BUILD_HALF
 	TABLE_NAME.shgemm_p = SHGEMM_DEFAULT_P;
+#endif
 	TABLE_NAME.sgemm_p = SGEMM_DEFAULT_P;
 	TABLE_NAME.dgemm_p = DGEMM_DEFAULT_P;
 	TABLE_NAME.cgemm_p = CGEMM_DEFAULT_P;
 	TABLE_NAME.zgemm_p = ZGEMM_DEFAULT_P;
 
+#ifdef BUILD_HALF
 	TABLE_NAME.shgemm_r = SHGEMM_DEFAULT_R;
+#endif
 	TABLE_NAME.sgemm_r = SGEMM_DEFAULT_R;
 	TABLE_NAME.dgemm_r = DGEMM_DEFAULT_R;
 	TABLE_NAME.cgemm_r = CGEMM_DEFAULT_R;
 	TABLE_NAME.zgemm_r = ZGEMM_DEFAULT_R;
 
 
+#ifdef BUILD_HALF
 	TABLE_NAME.shgemm_q = SHGEMM_DEFAULT_Q;
+#endif
 	TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
 	TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
 	TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;
@@ -958,9 +977,11 @@ static void init_parameter(void) {
   (void) l2; /* dirty trick to suppress unused variable warning for targets */
              /* where the GEMM unrolling parameters do not depend on l2 */
   
+#ifdef BUILD_HALF
   TABLE_NAME.shgemm_p = SHGEMM_DEFAULT_P;
   TABLE_NAME.shgemm_r = SHGEMM_DEFAULT_R;
   TABLE_NAME.shgemm_q = SHGEMM_DEFAULT_Q;
+#endif
   TABLE_NAME.sgemm_q = SGEMM_DEFAULT_Q;
   TABLE_NAME.dgemm_q = DGEMM_DEFAULT_Q;
   TABLE_NAME.cgemm_q = CGEMM_DEFAULT_Q;

test/Makefile

@@ -64,9 +64,17 @@ endif
 endif
 endif
 
+ifeq ($(BUILD_HALF),1)
+level3 : test_shgemm sblat3 dblat3 cblat3 zblat3
+else
 level3 : sblat3 dblat3 cblat3 zblat3
+endif
 ifndef CROSS
 	rm -f ?BLAT3.SUMM
+ifeq ($(BUILD_HALF),1)
+	OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 ./test_shgemm > SHBLAT3.SUMM
+	@$(GREP) -q FATAL SHBLAT3.SUMM && cat SHBLAT3.SUMM || exit 0
+endif
 	OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 ./sblat3 < ./sblat3.dat
 	@$(GREP) -q FATAL SBLAT3.SUMM && cat SBLAT3.SUMM || exit 0
 	OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 ./dblat3 < ./dblat3.dat
@@ -78,6 +86,10 @@ ifndef CROSS
 ifdef SMP
 	rm -f ?BLAT3.SUMM
 ifeq ($(USE_OPENMP), 1)
+ifeq ($(BUILD_HALF),1)
+	OMP_NUM_THREADS=2 ./test_shgemm > SHBLAT3.SUMM
+	@$(GREP) -q FATAL SHBLAT3.SUMM && cat SHBLAT3.SUMM || exit 0
+endif
 	OMP_NUM_THREADS=2 ./sblat3 < ./sblat3.dat
 	@$(GREP) -q FATAL SBLAT3.SUMM && cat SBLAT3.SUMM || exit 0
 	OMP_NUM_THREADS=2 ./dblat3 < ./dblat3.dat
@@ -87,6 +99,10 @@ ifeq ($(USE_OPENMP), 1)
 	OMP_NUM_THREADS=2 ./zblat3 < ./zblat3.dat
 	@$(GREP) -q FATAL ZBLAT3.SUMM && cat ZBLAT3.SUMM || exit 0
 else
+ifeq ($(BUILD_HALF),1)
+	OPENBLAS_NUM_THREADS=2 ./test_shgemm > SHBLAT3.SUMM
+	@$(GREP) -q FATAL SHBLAT3.SUMM && cat SHBLAT3.SUMM || exit 0
+endif
 	OPENBLAS_NUM_THREADS=2 ./sblat3 < ./sblat3.dat
 	@$(GREP) -q FATAL SBLAT3.SUMM && cat SBLAT3.SUMM || exit 0
 	OPENBLAS_NUM_THREADS=2 ./dblat3 < ./dblat3.dat
@@ -165,6 +181,11 @@ zblat2 : zblat2.$(SUFFIX) ../$(LIBNAME)
 sblat3 : sblat3.$(SUFFIX) ../$(LIBNAME)
 	$(FC) $(FLDFLAGS) -o sblat3 sblat3.$(SUFFIX) ../$(LIBNAME) $(EXTRALIB) $(CEXTRALIB)
 
+ifeq ($(BUILD_HALF),1)
+test_shgemm : compare_sgemm_shgemm.c ../$(LIBNAME)
+	$(FC) $(FLDFLAGS) -o test_shgemm compare_sgemm_shgemm.c ../$(LIBNAME) $(EXTRALIB) $(CEXTRALIB)
+endif
+
 dblat3 : dblat3.$(SUFFIX) ../$(LIBNAME)
 	$(FC) $(FLDFLAGS) -o dblat3 dblat3.$(SUFFIX) ../$(LIBNAME) $(EXTRALIB) $(CEXTRALIB)
 
@@ -187,7 +208,7 @@ clean:
 	@rm -f *.$(SUFFIX) *.$(PSUFFIX) gmon.$(SUFFIX)ut *.SUMM *.cxml *.exe *.pdb *.dwf \
 	sblat1 dblat1 cblat1 zblat1 \
 	sblat2 dblat2 cblat2 zblat2 \
-	sblat3 dblat3 cblat3 zblat3 \
+	test_shgemm sblat3 dblat3 cblat3 zblat3 \
 	sblat1p dblat1p cblat1p zblat1p \
 	sblat2p dblat2p cblat2p zblat2p \
 	sblat3p dblat3p cblat3p zblat3p \

test/compare_sgemm_shgemm.c

@@ -26,7 +26,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *****************************************************************************/
 #include <stdio.h>
 #include <stdint.h>
-#include "common.h"
+#include "../common.h"
 #define SGEMM   BLASFUNC(sgemm)
 #define SHGEMM   BLASFUNC(shgemm)
 typedef union
@@ -52,7 +52,7 @@ main (int argc, char *argv[])
   int m, n, k;
   int i, j, l;
   int ret = 0;
-  int loop = 20;
+  int loop = 100;
   char transA = 'N', transB = 'N';
   float alpha = 1.0, beta = 0.0;
   char transa = 'N';
@@ -71,8 +71,8 @@ main (int argc, char *argv[])
 	{
 	  for (int i = 0; i < m; i++)
 	    {
-	      A[j * k + i] = j * 9.0;
-	      B[j * k + i] = i * 2.0;
+	      A[j * k + i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) + 0.5;
+	      B[j * k + i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) + 0.5;
 	      C[j * k + i] = 0;
 	      AA[j * k + i].v = *(uint32_t *) & A[j * k + i] >> 16;
 	      BB[j * k + i].v = *(uint32_t *) & B[j * k + i] >> 16;
@@ -85,11 +85,12 @@ main (int argc, char *argv[])
 	       &m, BB, &k, &beta, CC, &m);
 
       for (i = 0; i < n; i++)
-	for (j = 0; j < m; j++)
-	  for (l = 0; l < k; l++)
-	    if (CC[i * m + j] != C[i * m + j])
-	      ret++;
+        for (j = 0; j < m; j++)
+          for (l = 0; l < k; l++)
+            if (fabs(CC[i * m + j]-C[i * m + j]) > 1.0)
+              ret++;
     }
-  fprintf (stderr, "Return code: %d\n", ret);
+  if (ret != 0)
+    fprintf (stderr, "FATAL ERROR SHGEMM - Return code: %d\n", ret);
   return ret;
 }