Change "HALF" and "sh" to "BFLOAT16" and "sb"

Author: martin-frbg

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")