Merge pull request #84 from xianyi/develop

rebase

Author: martin-frbg

Makefile.rule

@@ -277,5 +277,10 @@ COMMON_PROF = -pg
 # If you want to enable the experimental BFLOAT16 support
 # BUILD_HALF = 1
 #
+# the below is not yet configurable, use cmake if you need to build only select types
+BUILD_SINGLE = 1
+BUILD_DOUBLE = 1
+BUILD_COMPLEX = 1
+BUILD_COMPLEX16 = 1
 #  End of user configuration
 #

Makefile.system

@@ -295,6 +295,7 @@ endif
 ifeq ($(C_COMPILER), GCC)
 GCCVERSIONGTEQ4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 4)
 GCCVERSIONGT4 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 4)
+GCCVERSIONEQ5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` = 5)
 GCCVERSIONGT5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 5)
 GCCVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 7)
 GCCVERSIONGTEQ9 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 9)
@@ -593,35 +594,33 @@ endif
 ifeq ($(ARCH), zarch)
 DYNAMIC_CORE = ZARCH_GENERIC
 
-# if the compiler accepts -march=arch11 or -march=z13 and can compile a file
-# with z13-specific inline assembly, then we can include support for Z13.
-# note: -march=z13 is equivalent to -march=arch11 yet some compiler releases
-# only support one or the other.
-# note: LLVM version 6.x supported -march=z13 yet could not handle vector
-# registers in inline assembly, so the check for supporting the -march flag is
-# not enough.
-ZARCH_TEST_COMPILE=-c $(TOPDIR)/kernel/zarch/damin_z13.c -I$(TOPDIR) -o /dev/null > /dev/null 2> /dev/null
-ZARCH_CC_SUPPORTS_ARCH11=$(shell $(CC) -march=arch11 $(ZARCH_TEST_COMPILE) && echo 1)
-ZARCH_CC_SUPPORTS_Z13=$(shell $(CC) -march=z13 $(ZARCH_TEST_COMPILE) && echo 1)
-
-ifeq ($(or $(ZARCH_CC_SUPPORTS_ARCH11), $(ZARCH_CC_SUPPORTS_Z13)), 1)
+# Z13 is supported since gcc-5.2, gcc-6, and in RHEL 7.3 and newer
+ifeq ($(GCCVERSIONGT5), 1)
+	ZARCH_SUPPORT_Z13 := 1
+else ifeq ($(GCCVERSIONEQ5), 1)
+ifeq ($(GCCMINORVERSIONGTEQ2), 1)
+	ZARCH_SUPPORT_Z13 := 1
+endif
+endif
+
+ifeq ($(wildcard /etc/redhat-release), /etc/redhat-release)
+ifeq ($(shell source /etc/os-release ; expr $$VERSION_ID \>= "7.3"), 1)
+	ZARCH_SUPPORT_Z13 := 1
+endif
+endif
+
+ifeq ($(ZARCH_SUPPORT_Z13), 1)
 DYNAMIC_CORE += Z13
-CCOMMON_OPT += -DDYN_Z13
 else
-$(info OpenBLAS: Not building Z13 kernels because the compiler $(CC) does not support it)
+$(info OpenBLAS: Not building Z13 kernels because gcc is older than 5.2 or 6.x)
 endif
 
-# as above for z13, check for -march=arch12 and z14 support in the compiler.
-ZARCH_CC_SUPPORTS_ARCH12=$(shell $(CC) -march=arch12 $(ZARCH_TEST_COMPILE) && echo 1)
-ZARCH_CC_SUPPORTS_Z14=$(shell $(CC) -march=z14 $(ZARCH_TEST_COMPILE) && echo 1)
-ifeq ($(or $(ZARCH_CC_SUPPORTS_ARCH12), $(ZARCH_CC_SUPPORTS_Z14)), 1)
+ifeq ($(GCCVERSIONGTEQ7), 1)
 DYNAMIC_CORE += Z14
-CCOMMON_OPT += -DDYN_Z14
 else
-$(info OpenBLAS: Not building Z14 kernels because the compiler $(CC) does not support it)
+$(info OpenBLAS: Not building Z14 kernels because gcc is older than 7.x)
+endif
 endif
-
-endif # ARCH zarch
 
 ifeq ($(ARCH), power)
 DYNAMIC_CORE = POWER6
@@ -1223,6 +1222,18 @@ endif
 ifeq ($(BUILD_HALF), 1)
 CCOMMON_OPT += -DBUILD_HALF
 endif
+ifeq ($(BUILD_SINGLE), 1)
+CCOMMON_OPT += -DBUILD_SINGLE
+endif
+ifeq ($(BUILD_DOUBLE), 1)
+CCOMMON_OPT += -DBUILD_DOUBLE
+endif
+ifeq ($(BUILD_COMPLEX), 1)
+CCOMMON_OPT += -DBUILD_COMPLEX
+endif
+ifeq ($(BUILD_COMPLEX16), 1)
+CCOMMON_OPT += -DBUILD_COMPLEX16
+endif
 
 CCOMMON_OPT += -DVERSION=\"$(VERSION)\"
 

Makefile.tail

@@ -5,13 +5,14 @@ QBLASOBJS_P = $(QBLASOBJS:.$(SUFFIX)=.$(PSUFFIX))
 CBLASOBJS_P = $(CBLASOBJS:.$(SUFFIX)=.$(PSUFFIX))
 ZBLASOBJS_P = $(ZBLASOBJS:.$(SUFFIX)=.$(PSUFFIX))
 XBLASOBJS_P = $(XBLASOBJS:.$(SUFFIX)=.$(PSUFFIX))
+SHEXTOBJS_P = $(SHEXTOBJS:.$(SUFFIX)=.$(PSUFFIX))
 
 COMMONOBJS_P = $(COMMONOBJS:.$(SUFFIX)=.$(PSUFFIX))
 
 HPLOBJS_P   = $(HPLOBJS:.$(SUFFIX)=.$(PSUFFIX))
 
-BLASOBJS    = $(SHBLASOBJS)  $(SBLASOBJS)   $(DBLASOBJS)   $(CBLASOBJS)   $(ZBLASOBJS)
-BLASOBJS_P  = $(SHBLASOBJS_P) $(SBLASOBJS_P) $(DBLASOBJS_P) $(CBLASOBJS_P) $(ZBLASOBJS_P)
+BLASOBJS    = $(SHEXTOBJS) $(SHBLASOBJS)  $(SBLASOBJS)   $(DBLASOBJS)   $(CBLASOBJS)   $(ZBLASOBJS)
+BLASOBJS_P  = $(SHEXTOBJS_P) $(SHBLASOBJS_P) $(SBLASOBJS_P) $(DBLASOBJS_P) $(CBLASOBJS_P) $(ZBLASOBJS_P)
 
 ifdef EXPRECISION
 BLASOBJS   += $(QBLASOBJS)   $(XBLASOBJS)
@@ -30,6 +31,7 @@ $(QBLASOBJS) $(QBLASOBJS_P) : override CFLAGS += -DXDOUBLE -UCOMPLEX
 $(CBLASOBJS) $(CBLASOBJS_P) : override CFLAGS += -UDOUBLE  -DCOMPLEX
 $(ZBLASOBJS) $(ZBLASOBJS_P) : override CFLAGS += -DDOUBLE  -DCOMPLEX
 $(XBLASOBJS) $(XBLASOBJS_P) : override CFLAGS += -DXDOUBLE -DCOMPLEX
+$(SHEXTOBJS) $(SHEXTOBJS_P) : override CFLAGS += -DHALF -UDOUBLE  -UCOMPLEX
 
 $(SHBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 $(SBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
@@ -38,6 +40,7 @@ $(QBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 $(CBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 $(ZBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 $(XBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
+$(SHEXTOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 
 libs    :: $(BLASOBJS) $(COMMONOBJS)
 	$(AR) $(ARFLAGS) -ru $(TOPDIR)/$(LIBNAME) $^

cblas.h

@@ -382,6 +382,17 @@ void cblas_cgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint
 void cblas_zgeadd(OPENBLAS_CONST enum CBLAS_ORDER CORDER,OPENBLAS_CONST blasint crows, OPENBLAS_CONST blasint ccols, OPENBLAS_CONST double *calpha, double *a, OPENBLAS_CONST blasint clda, OPENBLAS_CONST double *cbeta, 
 		  double *c, OPENBLAS_CONST blasint cldc); 
 
+/*** 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);
+/* 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);
+/* 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 */
+void   cblas_dbf16tod(OPENBLAS_CONST blasint n, OPENBLAS_CONST bfloat16 *in, OPENBLAS_CONST blasint incin, double *out, OPENBLAS_CONST blasint incout);
+/* dot production of BFLOAT16 input arrays, and output as float */
+float  cblas_shdot(OPENBLAS_CONST blasint n, OPENBLAS_CONST bfloat16 *x, OPENBLAS_CONST blasint incx, OPENBLAS_CONST bfloat16 *y, OPENBLAS_CONST blasint incy);
 
 #ifdef __cplusplus
 }

cmake/kernel.cmake

@@ -126,12 +126,14 @@ if (BUILD_HALF)
   set(SHAXPYKERNEL ../arm/axpy.c)
   set(SHAXPBYKERNEL ../arm/axpby.c)
   set(SHCOPYKERNEL ../arm/copy.c)
-  set(SHDOTKERNEL ../arm/dot.c)
+  set(SHDOTKERNEL ../x86_64/shdot.c)
   set(SHROTKERNEL ../arm/rot.c)
   set(SHSCALKERNEL ../arm/scal.c)
   set(SHNRM2KERNEL ../arm/nrm2.c)
   set(SHSUMKERNEL ../arm/sum.c)
   set(SHSWAPKERNEL ../arm/swap.c)
+  set(TOBF16KERNEL ../x86_64/tobf16.c)
+  set(BF16TOKERNEL ../x86_64/bf16to.c)
 endif ()
 endmacro ()
 

cmake/system.cmake

@@ -393,6 +393,18 @@ set(REVISION "-r${OpenBLAS_VERSION}")
 set(MAJOR_VERSION ${OpenBLAS_MAJOR_VERSION})
 
 set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${CCOMMON_OPT}")
+if (BUILD_SINGLE)
+	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_SINGLE")
+endif()
+if (BUILD_DOUBLE)
+	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_DOUBLE")
+endif()
+if (BUILD_COMPLEX)
+	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX")
+endif()
+if (BUILD_COMPLEX16)
+	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX16")
+endif()
 if(NOT MSVC)
 set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CCOMMON_OPT}")
 endif()

common.h

@@ -258,7 +258,8 @@ typedef unsigned long BLASULONG;
 #endif
 
 #ifndef BFLOAT16
-typedef unsigned short bfloat16;
+#include <stdint.h>
+typedef uint16_t bfloat16;
 #define HALFCONVERSION 1
 #endif
 

common_interface.h

@@ -54,6 +54,11 @@ double BLASFUNC(dsdot) (blasint *, float  *, blasint *, float  *, blasint *);
 double BLASFUNC(ddot)  (blasint *, double *, blasint *, double *, blasint *);
 xdouble BLASFUNC(qdot)  (blasint *, xdouble *, blasint *, xdouble *, blasint *);
 
+float  BLASFUNC(shdot)     (blasint *, bfloat16 *, blasint *, bfloat16 *, blasint *);
+void   BLASFUNC(shstobf16) (blasint *, float *,    blasint *, bfloat16 *, blasint *);
+void   BLASFUNC(shdtobf16) (blasint *, double *,   blasint *, bfloat16 *, blasint *);
+void   BLASFUNC(sbf16tos)  (blasint *, bfloat16 *, blasint *, float *,    blasint *);
+void   BLASFUNC(dbf16tod)  (blasint *, bfloat16 *, blasint *, double *,   blasint *);
 
 #ifdef RETURN_BY_STRUCT
 typedef struct {

common_level1.h

@@ -46,6 +46,12 @@ float   sdot_k(BLASLONG, float   *, BLASLONG, float   *, BLASLONG);
 double dsdot_k(BLASLONG, float   *, BLASLONG, float *, BLASLONG);
 double  ddot_k(BLASLONG, double  *, BLASLONG, double  *, BLASLONG);
 xdouble qdot_k(BLASLONG, xdouble *, BLASLONG, xdouble *, BLASLONG);
+float  shdot_k(BLASLONG, bfloat16 *, BLASLONG, bfloat16 *, BLASLONG);
+
+void   shstobf16_k(BLASLONG, float    *, BLASLONG, bfloat16 *, BLASLONG);
+void   shdtobf16_k(BLASLONG, double   *, BLASLONG, bfloat16 *, BLASLONG);
+void   sbf16tos_k (BLASLONG, bfloat16 *, BLASLONG, float    *, BLASLONG);
+void   dbf16tod_k (BLASLONG, bfloat16 *, BLASLONG, double   *, BLASLONG);
 
 openblas_complex_float cdotc_k (BLASLONG, float  *, BLASLONG, float  *, BLASLONG);
 openblas_complex_float cdotu_k (BLASLONG, float  *, BLASLONG, float  *, BLASLONG);

common_macro.h

@@ -646,6 +646,11 @@
 
 #elif defined(HALF)
 
+#define  D_TO_BF16_K    SHDTOBF16_K
+#define  D_BF16_TO_K    DBF16TOD_K
+#define  S_TO_BF16_K    SHSTOBF16_K
+#define  S_BF16_TO_K    SBF16TOS_K
+
 #define	AMAX_K			SAMAX_K
 #define	AMIN_K			SAMIN_K
 #define	MAX_K			SMAX_K
@@ -657,6 +662,7 @@
 #define	ASUM_K			SASUM_K
 #define	DOTU_K			SDOTU_K
 #define	DOTC_K			SDOTC_K
+#define BF16_DOT_K      SHDOT_K
 #define	AXPYU_K			SAXPYU_K
 #define	AXPYC_K			SAXPYC_K
 #define AXPBY_K     SAXPBY_K