Rename "HALF" precision functions (sh prefix) to "BFLOAT16" with "sb" prefix

[martin-frbg]

supersedes #2771 , fixes #2767

[RajalakshmiSR]

@martin-frbg
make BUILD_BFLOAT16 fails on POWER9.
I have attached changes from my sandbox (common.h , interface/Makefile, test/Makefile) for reference. I could still not get sbgemm.o (from gemm.c)compiled from interface/Makefile.
In particular, the following 2 corresponding sbgemm lines are missing compared to BUILD_HALF (interface/Makefile)

cc -c -g -DMAX_STACK_ALLOC=2048 -fopenmp -Wall -m64 -DF_INTERFACE_GFORT -fPIC -DSMP_SERVER -DUSE_OPENMP -DNO_WARMUP -DMAX_CPU_NUMBER=160 -DMAX_PARALLEL_NUMBER=1 -DBUILD_HALF -DBUILD_SINGLE -DBUILD_DOUBLE -DBUILD_COMPLEX -DBUILD_COMPLEX16 -DVERSION="0.3.10.dev" -Ofast -mvsx -fno-fast-math -mcpu=power9 -mtune=power9 -DUSE_OPENMP -fopenmp -UASMNAME -UASMFNAME -UNAME -UCNAME -UCHAR_NAME -UCHAR_CNAME -DASMNAME=shgemm -DASMFNAME=shgemm_ -DNAME=shgemm_ -DCNAME=shgemm -DCHAR_NAME="shgemm_" -DCHAR_CNAME="shgemm" -DNO_AFFINITY -I.. -I. -DHALF -UDOUBLE -UCOMPLEX gemm.c -o shgemm.o
cc -DCBLAS -c -g -DMAX_STACK_ALLOC=2048 -fopenmp -Wall -m64 -DF_INTERFACE_GFORT -fPIC -DSMP_SERVER -DUSE_OPENMP -DNO_WARMUP -DMAX_CPU_NUMBER=160 -DMAX_PARALLEL_NUMBER=1 -DBUILD_HALF -DBUILD_SINGLE -DBUILD_DOUBLE -DBUILD_COMPLEX -DBUILD_COMPLEX16 -DVERSION="0.3.10.dev" -Ofast -mvsx -fno-fast-math -mcpu=power9 -mtune=power9 -DUSE_OPENMP -fopenmp -UASMNAME -UASMFNAME -UNAME -UCNAME -UCHAR_NAME -UCHAR_CNAME -DASMNAME=cblas_shgemm -DASMFNAME=cblas_shgemm_ -DNAME=cblas_shgemm_ -DCNAME=cblas_shgemm -DCHAR_NAME="cblas_shgemm_" -DCHAR_CNAME="cblas_shgemm" -DNO_AFFINITY -I.. -I. -DHALF -UDOUBLE -UCOMPLEX gemm.c -o cblas_shgemm.o

I will debug further tomorrow.

---

`diff --git a/common.h b/common.h
index 89eeb197..a3ef99b5 100644
--- a/common.h
+++ b/common.h
@@ -257,7 +257,7 @@ typedef long BLASLONG;
typedef unsigned long BLASULONG;
#endif

-#ifndef BFLOAT16
+#ifndef bfloat16
#include <stdint.h>
typedef uint16_t bfloat16;
#define BFLOAT16CONVERSION 1
diff --git a/interface/Makefile b/interface/Makefile
index a35d5327..38728561 100644
--- a/interface/Makefile
+++ b/interface/Makefile
@@ -283,8 +283,8 @@ CSBLAS3OBJS =
cblas_sgeadd.$(SUFFIX)

ifeq ($(BUILD_BFLOAT16),1)
-CBHBLAS1OBJS = cblas_sbdot.$(SUFFIX)
-CBHBLAS3OBJS = cblas_sbgemm.$(SUFFIX)
+CSBBLAS1OBJS = cblas_sbdot.$(SUFFIX)
+CSBBLAS3OBJS = cblas_sbgemm.$(SUFFIX)
CBHEXTOBJS = cblas_sbstobf16.$(SUFFIX) cblas_sbdtobf16.$(SUFFIX) cblas_sbf16tos.$(SUFFIX) cblas_dbf16tod.$(SUFFIX)
endif

@@ -535,19 +535,19 @@ endif
clean ::
@rm -f functable.h

-level1 : $(BEXTOBJS) $(SHBLAS1OBJS) $(SBLAS1OBJS) $(DBLAS1OBJS) $(QBLAS1OBJS) $(CBLAS1OBJS) $(ZBLAS1OBJS) $(XBLAS1OBJS)
+level1 : $(BEXTOBJS) $(SBBLAS1OBJS) $(SBLAS1OBJS) $(DBLAS1OBJS) $(QBLAS1OBJS) $(CBLAS1OBJS) $(ZBLAS1OBJS) $(XBLAS1OBJS)
$(AR) $(ARFLAGS) -ru $(TOPDIR)/$(LIBNAME) $^

level2 : $(SBLAS2OBJS) $(DBLAS2OBJS) $(QBLAS2OBJS) $(CBLAS2OBJS) $(ZBLAS2OBJS) $(XBLAS2OBJS)
$(AR) $(ARFLAGS) -ru $(TOPDIR)/$(LIBNAME) $^

-level3 : $(SHBLAS3OBJS) $(SBLAS3OBJS) $(DBLAS3OBJS) $(QBLAS3OBJS) $(CBLAS3OBJS) $(ZBLAS3OBJS) $(XBLAS3OBJS)
+level3 : $(SBBLAS3OBJS) $(SBLAS3OBJS) $(DBLAS3OBJS) $(QBLAS3OBJS) $(CBLAS3OBJS) $(ZBLAS3OBJS) $(XBLAS3OBJS)
$(AR) $(ARFLAGS) -ru $(TOPDIR)/$(LIBNAME) $^

aux : $(CBAUXOBJS)
$(AR) $(ARFLAGS) -ru $(TOPDIR)/$(LIBNAME) $^

-$(CSHBLASOBJS) $(CSHBLASOBJS_P) $(CSBLASOBJS) $(CSBLASOBJS_P) $(CDBLASOBJS) $(CDBLASOBJS_P) $(CQBLASOBJS) $(CQBLASOBJS_P)
+$(CSBBLASOBJS) $(CSBBLASOBJS_P) $(CSBLASOBJS) $(CSBLASOBJS_P) $(CDBLASOBJS) $(CDBLASOBJS_P) $(CQBLASOBJS) $(CQBLASOBJS_P)
$(CCBLASOBJS) $(CCBLASOBJS_P) $(CZBLASOBJS) $(CZBLASOBJS_P) $(CXBLASOBJS) $(CXBLASOBJS_P) $(CBAUXOBJS_P) : override CFLAGS += -DCBLAS

srot.$(SUFFIX) srot.$(PSUFFIX) : rot.c
diff --git a/test/Makefile b/test/Makefile
index 06fb7fe8..21234338 100644
--- a/test/Makefile
+++ b/test/Makefile
@@ -214,11 +214,9 @@ endif

-#ifeq ($(BUILD_BFLOAT16),1)
-#level3 : test_sbgemm sblat3 dblat3 cblat3 zblat3
-#else
-#level3 : sblat3 dblat3 cblat3 zblat3
-#endif
+ifeq ($(BUILD_BFLOAT16),1)
+level3 : test_sbgemm
+endif

ifndef CROSS
rm -f ?BLAT3.SUMM

[martin-frbg]

Sorry, think the main cause of this breakage is similar oversight in Makefile.tail. I'm creating a PR to fix this and add a P9/BFLOAT16 build to Travis. (Also need to kill that ridiculous "build matrix" in test/Makefile)