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

CMakeLists.txt

@@ -29,10 +29,8 @@ option(NO_AFFINITY "Disable support for CPU affinity masks to avoid binding proc
 else()
 set(NO_AFFINITY 1)
 endif()
-option(BUILD_SINGLE "Single precision" OFF)
-option(BUILD_DOUBLE "Double precision" OFF)
-option(BUILD_COMPLEX "Single precision" OFF)
-option(BUILD_COMPLEX16 "Single precision" OFF)
+option(CPP_THREAD_SAFETY_TEST "Run a massively parallel DGEMM test to confirm thread safety of the library (requires OpenMP and about 1.3GB of RAM)" OFF)
+option(CPP_THREAD_SAFETY_GEMV "Run a massively parallel DGEMV test to confirm thread safety of the library (requires OpenMP)" OFF)
 
 # Add a prefix or suffix to all exported symbol names in the shared library.
 # Avoids conflicts with other BLAS libraries, especially when using
@@ -91,13 +89,13 @@ if (NOT NO_LAPACK)
   list(APPEND SUBDIRS lapack)
 endif ()
 
-if (NOT DEFINED BUILD_HALF)
- set (BUILD_HALF false)
+if (NOT DEFINED BUILD_BFLOAT16)
+ set (BUILD_BFLOAT16 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_BFLOAT16 true)
   set(BUILD_SINGLE true)
   set(BUILD_DOUBLE true)
   set(BUILD_COMPLEX true)
@@ -110,33 +108,28 @@ endif()
 
 set(FLOAT_TYPES "")
 if (BUILD_SINGLE)
-  message(STATUS "Building Songle Precision")
-  list(APPEND FLOAT_TYPES "SINGLE") 
-  #    set(CCOMMON_OPT "${CCOMMON_OPT} -DBUILD_SINGLE=1")
+  message(STATUS "Building Single Precision")
+  list(APPEND FLOAT_TYPES "SINGLE") # defines nothing
 endif ()
 
 if (BUILD_DOUBLE)
   message(STATUS "Building Double Precision")
-  list(APPEND FLOAT_TYPES "DOUBLE")
-  #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_DOUBLE=1")
+  list(APPEND FLOAT_TYPES "DOUBLE") # defines DOUBLE
 endif ()
 
 if (BUILD_COMPLEX)
   message(STATUS "Building Complex Precision")
-  list(APPEND FLOAT_TYPES "COMPLEX")
-  #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX=1")
-endif () 
+  list(APPEND FLOAT_TYPES "COMPLEX") # defines COMPLEX
+endif ()
 
 if (BUILD_COMPLEX16)
   message(STATUS "Building Double Complex Precision")
-  list(APPEND FLOAT_TYPES "ZCOMPLEX") 
-  #set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX16=1")
+  list(APPEND FLOAT_TYPES "ZCOMPLEX") # defines COMPLEX and DOUBLE
 endif ()
 
-if (BUILD_HALF)
+if (BUILD_BFLOAT16)
   message(STATUS "Building Half Precision")
-  list(APPEND FLOAT_TYPES "HALF")
-  set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_HALF")
+  list(APPEND FLOAT_TYPES "BFLOAT16") # defines nothing
 endif ()
 
 if (NOT DEFINED CORE OR "${CORE}" STREQUAL "UNKNOWN")
@@ -243,6 +236,9 @@ if (NOT MSVC AND NOT NOFORTRAN)
     add_subdirectory(ctest)
   endif()
   add_subdirectory(lapack-netlib/TESTING)
+	  if (CPP_THREAD_SAFETY_TEST OR CPP_THREAD_SAFETY_GEMV)
+	  	add_subdirectory(cpp_thread_test)
+    	  endif()
 endif()
 
 set_target_properties(${OpenBLAS_LIBNAME} PROPERTIES

Makefile.rule

@@ -272,17 +272,33 @@ COMMON_PROF = -pg
 # work at all.
 #
 # CPP_THREAD_SAFETY_TEST = 1
+#
+# use this to run only the less memory-hungry GEMV test
+# CPP_THREAD_SAFETY_GEMV = 1
 
 
 # If you want to enable the experimental BFLOAT16 support
-# BUILD_HALF = 1
-#
-# Select if you need to build only select types
-# BUILD_SINGLE = 1
-# BUILD_DOUBLE = 1
-# BUILD_COMPLEX = 1
-# BUILD_COMPLEX16 = 1
-#
-#
+# BUILD_BFLOAT16 = 1
+
+
+# Set the thread number threshold beyond which the job array for the threaded level3 BLAS
+# will be allocated on the heap rather than the stack. (This array alone requires 
+# NUM_THREADS*NUM_THREADS*128 bytes of memory so should not pose a problem at low cpu
+# counts, but obviously it is not the only item that ends up on the stack.
+# The default value of 32 ensures that the overall requirement is compatible
+# with the default 1MB stacksize imposed by having the Java VM loaded without use
+# of its -Xss parameter.
+# The value of 160 formerly used from about version 0.2.7 until 0.3.10 is easily compatible
+# with the common Linux stacksize of 8MB but will cause crashes with unwary use of the java
+# VM e.g. in Octave or with the java-based libhdfs in numpy or scipy code
+# BLAS3_MEM_ALLOC_THRESHOLD = 160
+
+
+
+# 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

@@ -1232,8 +1232,8 @@ ifeq ($(USE_TLS), 1)
 CCOMMON_OPT += -DUSE_TLS
 endif
 
-ifeq ($(BUILD_HALF), 1)
-CCOMMON_OPT += -DBUILD_HALF
+ifeq ($(BUILD_BFLOAT16), 1)
+CCOMMON_OPT += -DBUILD_BFLOAT16
 endif
 ifeq ($(BUILD_SINGLE), 1)
 CCOMMON_OPT += -DBUILD_SINGLE=1
@@ -1521,10 +1521,10 @@ export KERNELDIR
 export FUNCTION_PROFILE
 export TARGET_CORE
 export NO_AVX512
-export BUILD_HALF
+export BUILD_BFLOAT16
 
-export SHGEMM_UNROLL_M
-export SHGEMM_UNROLL_N
+export SBGEMM_UNROLL_M
+export SBGEMM_UNROLL_N
 export SGEMM_UNROLL_M
 export SGEMM_UNROLL_N
 export DGEMM_UNROLL_M

Makefile.tail

@@ -24,14 +24,14 @@ BLASOBJS   += $(QBLASOBJS)   $(XBLASOBJS)
 BLASOBJS_P += $(QBLASOBJS_P) $(XBLASOBJS_P)
 endif
 
-$(SHBLASOBJS) $(SHBLASOBJS_P) : override CFLAGS += -DHALF -UDOUBLE  -UCOMPLEX
+$(SHBLASOBJS) $(SHBLASOBJS_P) : override CFLAGS += -DBFLOAT16 -UDOUBLE  -UCOMPLEX
 $(SBLASOBJS) $(SBLASOBJS_P) : override CFLAGS += -UDOUBLE  -UCOMPLEX
 $(DBLASOBJS) $(DBLASOBJS_P) : override CFLAGS += -DDOUBLE  -UCOMPLEX
 $(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) $(SHEXTOBJS_P) : override CFLAGS += -DHALF -UDOUBLE  -UCOMPLEX
+$(SHEXTOBJS) $(SHEXTOBJS_P) : override CFLAGS += -DBFLOAT16 -UDOUBLE  -UCOMPLEX
 
 $(SHBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)
 $(SBLASOBJS_P) : override CFLAGS += -DPROFILE $(COMMON_PROF)

benchmark/Makefile

@@ -49,8 +49,8 @@ else
 GOTO_LAPACK_TARGETS=
 endif
 
-ifeq ($(BUILD_HALF),1)
-GOTO_HALF_TARGETS=shgemm.goto
+ifeq ($(BUILD_BFLOAT16),1)
+GOTO_HALF_TARGETS=sbgemm.goto
 else
 GOTO_HALF_TARGETS=
 endif
@@ -620,8 +620,8 @@ zcholesky.essl : zcholesky.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBESSL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
 ##################################### Sgemm ####################################################
-ifeq ($(BUILD_HALF),1)
-shgemm.goto : shgemm.$(SUFFIX) ../$(LIBNAME)
+ifeq ($(BUILD_BFLOAT16),1)
+sbgemm.goto : sbgemm.$(SUFFIX) ../$(LIBNAME)
 	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB) -lm
 endif
 
@@ -2927,8 +2927,8 @@ ccholesky.$(SUFFIX) : cholesky.c
 zcholesky.$(SUFFIX) : cholesky.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
-ifeq ($(BUILD_HALF),1)
-shgemm.$(SUFFIX) : gemm.c
+ifeq ($(BUILD_BFLOAT16),1)
+sbgemm.$(SUFFIX) : gemm.c
 	$(CC) $(CFLAGS) -c -DHALF -UCOMPLEX -UDOUBLE -o $(@F) $^
 endif
 

benchmark/gemm.c

@@ -40,7 +40,7 @@ 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)
+#define GEMM   BLASFUNC(sbgemm)
 #else
 #define GEMM   BLASFUNC(sgemm)
 #endif

cblas.h

@@ -392,7 +392,7 @@ void   cblas_sbf16tos(OPENBLAS_CONST blasint n, OPENBLAS_CONST bfloat16 *in, OPE
 /* 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);
+float  cblas_sbdot(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

@@ -113,7 +113,7 @@ macro(SetDefaultL1)
   set(ZSUMKERNEL zsum.S)
   set(QSUMKERNEL sum.S)
   set(XSUMKERNEL zsum.S)
-if (BUILD_HALF)
+if (BUILD_BFLOAT16)
   set(SHAMINKERNEL ../arm/amin.c)
   set(SHAMAXKERNEL ../arm/amax.c)
   set(SHMAXKERNEL ../arm/max.c)
@@ -126,7 +126,7 @@ if (BUILD_HALF)
   set(SHAXPYKERNEL ../arm/axpy.c)
   set(SHAXPBYKERNEL ../arm/axpby.c)
   set(SHCOPYKERNEL ../arm/copy.c)
-  set(SHDOTKERNEL ../x86_64/shdot.c)
+  set(SBDOTKERNEL ../x86_64/sbdot.c)
   set(SHROTKERNEL ../arm/rot.c)
   set(SHSCALKERNEL ../arm/scal.c)
   set(SHNRM2KERNEL ../arm/nrm2.c)
@@ -183,9 +183,9 @@ 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)
+if (BUILD_BFLOAT16)
+  set(SBGEMVNKERNEL ../arm/gemv_n.c)
+  set(SBGEMVTKERNEL ../arm/gemv_t.c)
   set(SHGERKERNEL ../generic/ger.c)
 endif ()
 endmacro ()
@@ -195,18 +195,18 @@ macro(SetDefaultL3)
   set(DGEADD_KERNEL ../generic/geadd.c)
   set(CGEADD_KERNEL ../generic/zgeadd.c)
   set(ZGEADD_KERNEL ../generic/zgeadd.c)
-if (BUILD_HALF)
+if (BUILD_BFLOAT16)
   set(SHGEADD_KERNEL ../generic/geadd.c)
-  set(SHGEMMKERNEL ../generic/gemmkernel_2x2.c)
-  set(SHGEMM_BETA  ../generic/gemm_beta.c)
-  set(SHGEMMINCOPY ../generic/gemm_ncopy_2.c)
-  set(SHGEMMITCOPY ../generic/gemm_tcopy_2.c)
-  set(SHGEMMONCOPY ../generic/gemm_ncopy_2.c)
-  set(SHGEMMOTCOPY ../generic/gemm_tcopy_2.c)
-  set(SHGEMMINCOPYOBJ shgemm_incopy.o)
-  set(SHGEMMITCOPYOBJ shgemm_itcopy.o)
-  set(SHGEMMONCOPYOBJ shgemm_oncopy.o)
-  set(SHGEMMOTCOPYOBJ shgemm_otcopy.o)
+  set(SBGEMMKERNEL ../generic/gemmkernel_2x2.c)
+  set(SBGEMM_BETA  ../generic/gemm_beta.c)
+  set(SBGEMMINCOPY ../generic/gemm_ncopy_2.c)
+  set(SBGEMMITCOPY ../generic/gemm_tcopy_2.c)
+  set(SBGEMMONCOPY ../generic/gemm_ncopy_2.c)
+  set(SBGEMMOTCOPY ../generic/gemm_tcopy_2.c)
+  set(SBGEMMINCOPYOBJ sbgemm_incopy.o)
+  set(SBGEMMITCOPYOBJ sbgemm_itcopy.o)
+  set(SBGEMMONCOPYOBJ sbgemm_oncopy.o)
+  set(SBGEMMOTCOPYOBJ sbgemm_otcopy.o)
 endif ()
 
 endmacro ()

cmake/prebuild.cmake

@@ -16,8 +16,8 @@
 # HAVE_SSE2
 # HAVE_SSE3
 # MAKE
-# SHGEMM_UNROLL_M
-# SHGEMM_UNROLL_N
+# SBGEMM_UNROLL_M
+# SBGEMM_UNROLL_N
 # SGEMM_UNROLL_M
 # SGEMM_UNROLL_N
 # DGEMM_UNROLL_M
@@ -471,8 +471,8 @@ endif ()
     set(ZGEMM_UNROLL_N 2)
     set(SYMV_P 8)
   endif()
-  set(SHGEMM_UNROLL_M 8)
-  set(SHGEMM_UNROLL_N 4)
+  set(SBGEMM_UNROLL_M 8)
+  set(SBGEMM_UNROLL_N 4)
 
   # Or should this actually be NUM_CORES?
   if (${NUM_THREADS} GREATER 0)

cmake/system.cmake

@@ -70,6 +70,9 @@ if (DEFINED TARGET)
       set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -mavx2")
     endif()
   endif()
+  if (DEFINED HAVE_SSE3)
+    set (KERNEL_DEFINITIONS "${KERNEL_DEFINITIONS} -msse3")
+  endif()
 endif()
 
 if (DEFINED TARGET)
@@ -323,7 +326,13 @@ else ()
 set(CCOMMON_OPT "${CCOMMON_OPT} -DMAX_STACK_ALLOC=2048")
 endif ()
 endif ()
-
+if (NOT ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
+if (DEFINED BLAS3_MEM_ALLOC_THRESHOLD)
+if (NOT ${BLAS3_MEM_ALLOC_THRESHOLD} EQUAL 32)
+set(CCOMMON_OPT "${CCOMMON_OPT} -DBLAS3_MEM_ALLOC_THRESHOLD=${BLAS3_MEM_ALLOC_THRESHOLD}")
+endif()
+endif()
+endif()
 if (DEFINED LIBNAMESUFFIX)
   set(LIBPREFIX "libopenblas_${LIBNAMESUFFIX}")
 else ()
@@ -401,20 +410,16 @@ if (NOT BUILD_SINGLE AND NOT BUILD_DOUBLE AND NOT BUILD_COMPLEX AND NOT BUILD_CO
 	set (BUILD_COMPLEX16 ON)
 endif()
 if (BUILD_SINGLE)
-	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_SINGLE=1")
-	set(CCOMMON_OPT "${CCOMMON_OPT} -DBUILD_SINGLE=1")
+	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_SINGLE")
 endif()
 if (BUILD_DOUBLE)
-	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_DOUBLE=1")
-	set(CCOMMON_OPT "${CCOMMON_OPT} -DBUILD_SINGLE=1")
+	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_DOUBLE")
 endif()
 if (BUILD_COMPLEX)
-	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX=1")
-	set(CCOMMON_OPT "${CCOMMON_OPT} -DBUILD_COMPLEX=1")
+	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX")
 endif()
 if (BUILD_COMPLEX16)
-	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX16=1")
-	set(CCOMMON_OPT "${CCOMMON_OPT} -DBUILD_COMPLEX16=1")
+	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -DBUILD_COMPLEX16")
 endif()
 if(NOT MSVC)
 set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} ${CCOMMON_OPT}")
@@ -588,8 +593,8 @@ endif ()
 #export FUNCTION_PROFILE
 #export TARGET_CORE
 #
-#export SHGEMM_UNROLL_M
-#export SHGEMM_UNROLL_N
+#export SBGEMM_UNROLL_M
+#export SBGEMM_UNROLL_N
 #export SGEMM_UNROLL_M
 #export SGEMM_UNROLL_N
 #export DGEMM_UNROLL_M

cmake/utils.cmake

@@ -211,7 +211,7 @@ function(GenerateNamedObjects sources_in)
   if (complex_only)
     list(REMOVE_ITEM float_list "SINGLE")
     list(REMOVE_ITEM float_list "DOUBLE")
-    list(REMOVE_ITEM float_list "HALF")
+    list(REMOVE_ITEM float_list "BFLOAT16")
   elseif (real_only)
     list(REMOVE_ITEM float_list "COMPLEX")
     list(REMOVE_ITEM float_list "ZCOMPLEX")
@@ -225,8 +225,8 @@ function(GenerateNamedObjects sources_in)
       if (NOT no_float_type)
         string(SUBSTRING ${float_type} 0 1 float_char)
         string(TOLOWER ${float_char} float_char)
-	if (${float_type} STREQUAL "HALF")
-		set (float_char "sh")
+	if (${float_type} STREQUAL "BFLOAT16")
+		set (float_char "sb")
 	endif ()
       endif ()
 
@@ -262,8 +262,8 @@ function(GenerateNamedObjects sources_in)
       if (${float_type} STREQUAL "DOUBLE" OR ${float_type} STREQUAL "ZCOMPLEX")
         list(APPEND obj_defines "DOUBLE")
       endif ()
-      if (${float_type} STREQUAL "HALF")
-        list(APPEND obj_defines "HALF")
+      if (${float_type} STREQUAL "BFLOAT16")
+        list(APPEND obj_defines "BFLOAT16")
       endif ()
       if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX")
         list(APPEND obj_defines "COMPLEX")