Fix build issues with bfloat16

This patch fixes compilation errors due to recent renaming from SH to SB
with BUILD_BFLOAT16.

cblas.h

@@ -384,9 +384,9 @@ void cblas_zgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint
 
 /*** BFLOAT16 and INT8 extensions ***/
 /* convert float array to BFLOAT16 array by rounding */
-void   cblas_shstobf16(OPENBLAS_CONST blasint n, OPENBLAS_CONST float  *in, OPENBLAS_CONST blasint incin, bfloat16 *out, OPENBLAS_CONST blasint incout);
+void   cblas_sbstobf16(OPENBLAS_CONST blasint n, OPENBLAS_CONST float  *in, OPENBLAS_CONST blasint incin, bfloat16 *out, OPENBLAS_CONST blasint incout);
 /* convert double array to BFLOAT16 array by rounding */
-void   cblas_shdtobf16(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *in, OPENBLAS_CONST blasint incin, bfloat16 *out, OPENBLAS_CONST blasint incout);
+void   cblas_sbdtobf16(OPENBLAS_CONST blasint n, OPENBLAS_CONST double *in, OPENBLAS_CONST blasint incin, bfloat16 *out, OPENBLAS_CONST blasint incout);
 /* convert BFLOAT16 array to float array */
 void   cblas_sbf16tos(OPENBLAS_CONST blasint n, OPENBLAS_CONST bfloat16 *in, OPENBLAS_CONST blasint incin, float  *out, OPENBLAS_CONST blasint incout);
 /* convert BFLOAT16 array to double array */

common_interface.h

@@ -55,8 +55,8 @@ double BLASFUNC(ddot)  (blasint *, double *, blasint *, double *, blasint *);
 xdouble BLASFUNC(qdot)  (blasint *, xdouble *, blasint *, xdouble *, blasint *);
 
 float  BLASFUNC(sbdot)     (blasint *, bfloat16 *, blasint *, bfloat16 *, blasint *);
-void   BLASFUNC(shstobf16) (blasint *, float *,    blasint *, bfloat16 *, blasint *);
-void   BLASFUNC(shdtobf16) (blasint *, double *,   blasint *, bfloat16 *, blasint *);
+void   BLASFUNC(sbstobf16) (blasint *, float *,    blasint *, bfloat16 *, blasint *);
+void   BLASFUNC(sbdtobf16) (blasint *, double *,   blasint *, bfloat16 *, blasint *);
 void   BLASFUNC(sbf16tos)  (blasint *, bfloat16 *, blasint *, float *,    blasint *);
 void   BLASFUNC(dbf16tod)  (blasint *, bfloat16 *, blasint *, double *,   blasint *);
 

common_level1.h

@@ -48,8 +48,8 @@ double  ddot_k(BLASLONG, double  *, BLASLONG, double  *, BLASLONG);
 xdouble qdot_k(BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
 float  sbdot_k(BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
 
-void   shstobf16_k(BLASLONG, float    *, BLASLONG, bfloat16 *, BLASLONG);
-void   shdtobf16_k(BLASLONG, double   *, BLASLONG, bfloat16 *, BLASLONG);
+void   sbstobf16_k(BLASLONG, float    *, BLASLONG, bfloat16 *, BLASLONG);
+void   sbdtobf16_k(BLASLONG, double   *, BLASLONG, bfloat16 *, BLASLONG);
 void   sbf16tos_k (BLASLONG, bfloat16 *, BLASLONG, float    *, BLASLONG);
 void   dbf16tod_k (BLASLONG, bfloat16 *, BLASLONG, double   *, BLASLONG);
 

common_macro.h

@@ -646,9 +646,9 @@
 
 #elif defined(BFLOAT16)
 
-#define  D_TO_BF16_K    SHDTOBF16_K
+#define  D_TO_BF16_K    SBDTOBF16_K
 #define  D_BF16_TO_K    DBF16TOD_K
-#define  S_TO_BF16_K    SHSTOBF16_K
+#define  S_TO_BF16_K    SBSTOBF16_K
 #define  S_BF16_TO_K    SBF16TOS_K
 
 #define	AMAX_K			SAMAX_K

driver/level3/Makefile

@@ -20,7 +20,7 @@ USE_GEMM3M = 1
 endif
 
 ifeq ($(BUILD_BFLOAT16),1)
-SHBLASOBJS       += sbgemm_nn.$(SUFFIX) sbgemm_nt.$(SUFFIX) sbgemm_tn.$(SUFFIX) sbgemm_tt.$(SUFFIX)
+SBBLASOBJS       += sbgemm_nn.$(SUFFIX) sbgemm_nt.$(SUFFIX) sbgemm_tn.$(SUFFIX) sbgemm_tt.$(SUFFIX)
 endif
 
 SBLASOBJS	+= \
@@ -208,7 +208,7 @@ COMMONOBJS  += syrk_thread.$(SUFFIX)
 
 ifndef USE_SIMPLE_THREADED_LEVEL3
 ifeq ($(BUILD_BFLOAT16),1)
-SHBLASOBJS    += sbgemm_thread_nn.$(SUFFIX) sbgemm_thread_nt.$(SUFFIX) sbgemm_thread_tn.$(SUFFIX) sbgemm_thread_tt.$(SUFFIX)
+SBBLASOBJS    += sbgemm_thread_nn.$(SUFFIX) sbgemm_thread_nt.$(SUFFIX) sbgemm_thread_tn.$(SUFFIX) sbgemm_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)

exports/gensymbol

@@ -51,7 +51,7 @@
     zgeadd, dzsum);
 
 @cblasobjs = (lsame, xerbla);
-@halfblasobjs = (sbgemm, sbdot, shstobf16, shdtobf16, sbf16tos, dbf16tod);
+@halfblasobjs = (sbgemm, sbdot, sbstobf16, sbdtobf16, sbf16tos, dbf16tod);
 @cblasobjsc = (
     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,
@@ -94,7 +94,7 @@
 
 @cblasobjs = (  cblas_xerbla );
 
-@halfcblasobjs = (cblas_sbgemm, cblas_sbdot, cblas_shstobf16, cblas_shdtobf16, cblas_sbf16tos, cblas_dbf16tod);
+@halfcblasobjs = (cblas_sbgemm, cblas_sbdot, cblas_sbstobf16, cblas_sbdtobf16, cblas_sbf16tos, cblas_dbf16tod);
 
 @exblasobjs = (
     qamax,qamin,qasum,qaxpy,qcabs1,qcopy,qdot,qgbmv,qgemm,

interface/Makefile

@@ -775,9 +775,9 @@ dsdot.$(SUFFIX) dsdot.$(PSUFFIX) : dsdot.c
 ifeq ($(BUILD_BFLOAT16),1)
 sbdot.$(SUFFIX) sbdot.$(PSUFFIX) : bf16dot.c
 	$(CC) $(CFLAGS) -c $< -o $(@F)
-shstobf16.$(SUFFIX) shstobf16.$(PSUFFIX) : tobf16.c
+sbstobf16.$(SUFFIX) sbstobf16.$(PSUFFIX) : tobf16.c
 	$(CC) $(CFLAGS) -DSINGLE_PREC -UDOUBLE_PREC -c $< -o $(@F)
-shdtobf16.$(SUFFIX) shdtobf16.$(PSUFFIX) : tobf16.c
+sbdtobf16.$(SUFFIX) sbdtobf16.$(PSUFFIX) : tobf16.c
 	$(CC) $(CFLAGS) -USINGLE_PREC -DDOUBLE_PREC -c $< -o $(@F)
 sbf16tos.$(SUFFIX)  sbf16tos.$(PSUFFIX) : bf16to.c
 	$(CC) $(CFLAGS) -DSINGLE_PREC -UDOUBLE_PREC -c $< -o $(@F)
@@ -1526,9 +1526,9 @@ cblas_dsdot.$(SUFFIX) cblas_dsdot.$(PSUFFIX) : dsdot.c
 ifeq ($(BUILD_BFLOAT16),1)
 cblas_sbdot.$(SUFFIX) cblas_sbdot.$(PSUFFIX) : bf16dot.c
 	$(CC) $(CFLAGS) -DCBLAS -c $< -o $(@F)
-cblas_shstobf16.$(SUFFIX) cblas_shstobf16.$(PSUFFIX) : tobf16.c
+cblas_sbstobf16.$(SUFFIX) cblas_sbstobf16.$(PSUFFIX) : tobf16.c
 	$(CC) $(CFLAGS) -DCBLAS -DSINGLE_PREC -UDOUBLE_PREC -c $< -o $(@F)
-cblas_shdtobf16.$(SUFFIX) cblas_shdtobf16.$(PSUFFIX) : tobf16.c
+cblas_sbdtobf16.$(SUFFIX) cblas_sbdtobf16.$(PSUFFIX) : tobf16.c
 	$(CC) $(CFLAGS) -DCBLAS -USINGLE_PREC -DDOUBLE_PREC -c $< -o $(@F)
 cblas_sbf16tos.$(SUFFIX)  cblas_sbf16tos.$(PSUFFIX) : bf16to.c
 	$(CC) $(CFLAGS) -DCBLAS -DSINGLE_PREC -UDOUBLE_PREC -c $< -o $(@F)

kernel/Makefile.L1

@@ -531,11 +531,11 @@ XBLASOBJS	+= \
 	xscal_k$(TSUFFIX).$(SUFFIX) xswap_k$(TSUFFIX).$(SUFFIX) xsum_k$(TSUFFIX).$(SUFFIX)
 
 ifeq ($(BUILD_BFLOAT16),1)
-SHBLASOBJS	+= \
+SBBLASOBJS	+= \
 	 sbdot_k$(TSUFFIX).$(SUFFIX)
-SHEXTOBJS        += \
+SBEXTOBJS        += \
 	 sbstobf16_k$(TSUFFIX).$(SUFFIX) sbdtobf16_k$(TSUFFIX).$(SUFFIX)
-SHEXTOBJS        += \
+SBEXTOBJS        += \
 	 sbf16tos_k$(TSUFFIX).$(SUFFIX) dbf16tod_k$(TSUFFIX).$(SUFFIX)
 endif
 

kernel/Makefile.L3

@@ -94,7 +94,7 @@ SBGEMMONCOPYOBJ =  sbgemm_oncopy$(TSUFFIX).$(SUFFIX)
 SBGEMMOTCOPYOBJ =  sbgemm_otcopy$(TSUFFIX).$(SUFFIX)
 endif
 
-SHKERNELOBJS	+= \
+SBKERNELOBJS	+= \
 	sbgemm_kernel$(TSUFFIX).$(SUFFIX) \
 	$(SBGEMMINCOPYOBJ) $(SBGEMMITCOPYOBJ) \
 	$(SBGEMMONCOPYOBJ) $(SBGEMMOTCOPYOBJ)
@@ -150,7 +150,7 @@ XKERNELOBJS	+= \
 	$(XGEMMONCOPYOBJ) $(XGEMMOTCOPYOBJ)
 
 ifeq ($(BUILD_BFLOAT16),1)
-SHBLASOBJS      += $(SHKERNELOBJS)
+SBBLASOBJS      += $(SBKERNELOBJS)
 endif
 SBLASOBJS	+= $(SKERNELOBJS)
 DBLASOBJS	+= $(DKERNELOBJS)
@@ -160,7 +160,7 @@ ZBLASOBJS	+= $(ZKERNELOBJS)
 XBLASOBJS	+= $(XKERNELOBJS)
 
 ifeq ($(BUILD_BFLOAT16),1)
-SHBLASOBJS += sbgemm_beta$(TSUFFIX).$(SUFFIX)
+SBBLASOBJS += sbgemm_beta$(TSUFFIX).$(SUFFIX)
 endif
 
 ifneq "$(or $(BUILD_SINGLE),$(BUILD_DOUBLE))" ""

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