From 283fe7f8c741a1020ac326b42d63309666952c18 Mon Sep 17 00:00:00 2001
From: CentOS Sources <bugs@centos.org>
Date: Fri, 30 Oct 2020 06:11:18 +0000
Subject: [PATCH] import openblas-0.3.12-1.el8

---
 .gitignore                                    |   2 +-
 .openblas.metadata                            |   2 +-
 SOURCES/openblas-0.3.10-concurrency.patch     |  54 ---
 SOURCES/openblas-0.3.10-lapacke-macro.patch   |  13 -
 ...s-0.3.10-zarch-gcc-version-detection.patch |  57 ---
 ...s.patch => openblas-0.3.12-asmflags.patch} | 394 ++++++++++--------
 ...me.patch => openblas-0.3.12-libname.patch} |   4 +-
 ...oopt.patch => openblas-0.3.12-noopt.patch} |  17 +-
 ...h => openblas-0.3.12-power-optimize.patch} |  18 +-
 ...ch => openblas-0.3.12-system-lapack.patch} |   0
 ...ests.patch => openblas-0.3.12-tests.patch} |   6 +-
 SPECS/openblas.spec                           |  29 +-
 12 files changed, 262 insertions(+), 334 deletions(-)
 delete mode 100644 SOURCES/openblas-0.3.10-concurrency.patch
 delete mode 100644 SOURCES/openblas-0.3.10-lapacke-macro.patch
 delete mode 100644 SOURCES/openblas-0.3.10-zarch-gcc-version-detection.patch
 rename SOURCES/{openblas-0.3.10-asmflags.patch => openblas-0.3.12-asmflags.patch} (97%)
 rename SOURCES/{openblas-0.3.10-libname.patch => openblas-0.3.12-libname.patch} (92%)
 rename SOURCES/{openblas-0.3.10-noopt.patch => openblas-0.3.12-noopt.patch} (85%)
 rename SOURCES/{openblas-0.3.10-power-optimize.patch => openblas-0.3.12-power-optimize.patch} (67%)
 rename SOURCES/{openblas-0.3.10-system-lapack.patch => openblas-0.3.12-system-lapack.patch} (100%)
 rename SOURCES/{openblas-0.3.10-tests.patch => openblas-0.3.12-tests.patch} (89%)

diff --git a/.gitignore b/.gitignore
index 464fe71..f697ae0 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1 +1 @@
-SOURCES/v0.3.10.tar.gz
+SOURCES/v0.3.12.tar.gz
diff --git a/.openblas.metadata b/.openblas.metadata
index d457f79..058300c 100644
--- a/.openblas.metadata
+++ b/.openblas.metadata
@@ -1 +1 @@
-cbe3fdd0e6ee235debc611d76976dac62f3ddc1c SOURCES/v0.3.10.tar.gz
+ae647fed597ae891a7f122b9ddc6b15d4b7e0656 SOURCES/v0.3.12.tar.gz
diff --git a/SOURCES/openblas-0.3.10-concurrency.patch b/SOURCES/openblas-0.3.10-concurrency.patch
deleted file mode 100644
index 124ec0e..0000000
--- a/SOURCES/openblas-0.3.10-concurrency.patch
+++ /dev/null
@@ -1,54 +0,0 @@
-diff --git a/cpp_thread_test/cpp_thread_safety_common.h b/cpp_thread_test/cpp_thread_safety_common.h
-index 60ab5bb..8005369 100644
---- a/cpp_thread_test/cpp_thread_safety_common.h
-+++ b/cpp_thread_test/cpp_thread_safety_common.h
-@@ -5,6 +5,14 @@ inline void pauser(){
-     std::getline(std::cin, dummy);
- }
- 
-+void FailIfThreadsAreZero(uint32_t numConcurrentThreads) {
-+	if(numConcurrentThreads == 0) {
-+		std::cout<<"ERROR: Invalid parameter 0 for number of concurrent calls into OpenBLAS!"<<std::endl;
-+		std::cout<<"CBLAS DGEMV thread safety test FAILED!"<<std::endl;
-+		exit(-1);
-+	}
-+}
-+
- void FillMatrices(std::vector<std::vector<double>>& matBlock, std::mt19937_64& PRNG, std::uniform_real_distribution<double>& rngdist, const blasint randomMatSize, const uint32_t numConcurrentThreads, const uint32_t numMat){
- 	for(uint32_t i=0; i<numMat; i++){
- 		for(uint32_t j = 0; j < static_cast<uint32_t>(randomMatSize*randomMatSize); j++){
-diff --git a/cpp_thread_test/dgemm_thread_safety.cpp b/cpp_thread_test/dgemm_thread_safety.cpp
-index 1c52875..104c64f 100644
---- a/cpp_thread_test/dgemm_thread_safety.cpp
-+++ b/cpp_thread_test/dgemm_thread_safety.cpp
-@@ -46,6 +46,8 @@ int main(int argc, char* argv[]){
- 	std::cout<<"Number of concurrent calls into OpenBLAS : "<<numConcurrentThreads<<'\n';
- 	std::cout<<"Number of testing rounds : "<<numTestRounds<<'\n';
- 	std::cout<<"This test will need "<<(static_cast<uint64_t>(randomMatSize*randomMatSize)*numConcurrentThreads*3*8)/static_cast<double>(1024*1024)<<" MiB of RAM\n"<<std::endl;
-+
-+	FailIfThreadsAreZero(numConcurrentThreads);
- 	
- 	std::cout<<"Initializing random number generator..."<<std::flush;
- 	std::mt19937_64 PRNG = InitPRNG();
-diff --git a/cpp_thread_test/dgemv_thread_safety.cpp b/cpp_thread_test/dgemv_thread_safety.cpp
-index 5411fec..20ea381 100644
---- a/cpp_thread_test/dgemv_thread_safety.cpp
-+++ b/cpp_thread_test/dgemv_thread_safety.cpp
-@@ -18,7 +18,7 @@ int main(int argc, char* argv[]){
- 	uint32_t maxHwThreads = omp_get_max_threads();
- 	
- 	if (maxHwThreads < 52)
--		numConcurrentThreads = maxHwThreads -4;
-+		numConcurrentThreads = maxHwThreads;
- 	
- 	if (argc > 4){
- 		std::cout<<"ERROR: too many arguments for thread safety tester"<<std::endl;
-@@ -47,6 +47,8 @@ int main(int argc, char* argv[]){
- 	std::cout<<"Number of concurrent calls into OpenBLAS : "<<numConcurrentThreads<<'\n';
- 	std::cout<<"Number of testing rounds : "<<numTestRounds<<'\n';
- 	std::cout<<"This test will need "<<((static_cast<uint64_t>(randomMatSize*randomMatSize)*numConcurrentThreads*8)+(static_cast<uint64_t>(randomMatSize)*numConcurrentThreads*8*2))/static_cast<double>(1024*1024)<<" MiB of RAM\n"<<std::endl;
-+
-+	FailIfThreadsAreZero(numConcurrentThreads);
- 	
- 	std::cout<<"Initializing random number generator..."<<std::flush;
- 	std::mt19937_64 PRNG = InitPRNG();
diff --git a/SOURCES/openblas-0.3.10-lapacke-macro.patch b/SOURCES/openblas-0.3.10-lapacke-macro.patch
deleted file mode 100644
index cc32eb0..0000000
--- a/SOURCES/openblas-0.3.10-lapacke-macro.patch
+++ /dev/null
@@ -1,13 +0,0 @@
-diff --git a/lapack-netlib/LAPACKE/src/lapacke_zgesvdq.c b/lapack-netlib/LAPACKE/src/lapacke_zgesvdq.c
-index f58a5c4..4928b1b 100644
---- a/lapack-netlib/LAPACKE/src/lapacke_zgesvdq.c
-+++ b/lapack-netlib/LAPACKE/src/lapacke_zgesvdq.c
-@@ -71,7 +71,7 @@ lapack_int LAPACKE_zgesvdq( int matrix_layout, char joba, char jobp,
-         goto exit_level_0;
-     }
-     liwork = iwork_query;
--    lcwork = LAPACK_C2INT(cwork_query);
-+    lcwork = LAPACK_Z2INT(cwork_query);
-     lrwork = (lapack_int)rwork_query;
-     /* Allocate memory for work arrays */
-     iwork = (lapack_int*)LAPACKE_malloc( sizeof(lapack_int) * liwork );
diff --git a/SOURCES/openblas-0.3.10-zarch-gcc-version-detection.patch b/SOURCES/openblas-0.3.10-zarch-gcc-version-detection.patch
deleted file mode 100644
index e3ad2e0..0000000
--- a/SOURCES/openblas-0.3.10-zarch-gcc-version-detection.patch
+++ /dev/null
@@ -1,57 +0,0 @@
-diff --git a/Makefile.system b/Makefile.system
-index c947a19..cbf419a 100644
---- a/Makefile.system
-+++ b/Makefile.system
-@@ -282,9 +282,11 @@ 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)
-+GCCMINORVERSIONGTEQ2 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 2)
- GCCMINORVERSIONGTEQ7 := $(shell expr `$(CC) -dumpversion | cut -f2 -d.` \>= 7)
- endif
- 
-@@ -570,20 +572,27 @@ ifeq ($(ARCH), zarch)
- DYNAMIC_CORE = ZARCH_GENERIC
- 
- # Z13 is supported since gcc-5.2, gcc-6, and in RHEL 7.3 and newer
--GCC_GE_52 := $(subst 0,,$(shell expr `$(CC) -dumpversion` \>= "5.2"))
-+ifeq ($(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)
--RHEL_WITH_Z13 := $(subst 0,,$(shell source /etc/os-release ; expr $$VERSION_ID \>= "7.3"))
-+ifeq ($(shell source /etc/os-release ; expr $$VERSION_ID \>= "7.3"), 1)
-+	ZARCH_SUPPORT_Z13 := 1
-+endif
- endif
- 
--ifeq ($(or $(GCC_GE_52),$(RHEL_WITH_Z13)), 1)
-+ifeq ($(ZARCH_SUPPORT_Z13), 1)
- DYNAMIC_CORE += Z13
- else
- $(info OpenBLAS: Not building Z13 kernels because gcc is older than 5.2 or 6.x)
- endif
- 
--GCC_MAJOR_GE_7 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \>= 7)
--ifeq ($(GCC_MAJOR_GE_7), 1)
-+ifeq ($(GCCVERSIONGTEQ7), 1)
- DYNAMIC_CORE += Z14
- else
- $(info OpenBLAS: Not building Z14 kernels because gcc is older than 7.x)
-@@ -597,7 +606,6 @@ ifneq ($(C_COMPILER), GCC)
- DYNAMIC_CORE += POWER9
- endif
- ifeq ($(C_COMPILER), GCC)
--GCCVERSIONGT5 := $(shell expr `$(CC) -dumpversion | cut -f1 -d.` \> 5)
- ifeq ($(GCCVERSIONGT5), 1)
- DYNAMIC_CORE += POWER9
- else
diff --git a/SOURCES/openblas-0.3.10-asmflags.patch b/SOURCES/openblas-0.3.12-asmflags.patch
similarity index 97%
rename from SOURCES/openblas-0.3.10-asmflags.patch
rename to SOURCES/openblas-0.3.12-asmflags.patch
index 16dd8db..1f767a7 100644
--- a/SOURCES/openblas-0.3.10-asmflags.patch
+++ b/SOURCES/openblas-0.3.12-asmflags.patch
@@ -1,8 +1,8 @@
 diff --git a/kernel/Makefile b/kernel/Makefile
-index 9b468a6..87df5fa 100644
+index e52781c6..c925837a 100644
 --- a/kernel/Makefile
 +++ b/kernel/Makefile
-@@ -113,25 +113,32 @@ endif
+@@ -151,25 +151,32 @@ endif
  all : libs
  
  scabs1.$(SUFFIX): $(KERNELDIR)/$(SCABS_KERNEL)
@@ -42,7 +42,7 @@ index 9b468a6..87df5fa 100644
  endif
  
  setparam$(TSUFFIX).c : setparam-ref.c
-@@ -142,25 +149,32 @@ kernel$(TSUFFIX).h : $(KERNEL_INTERFACE)
+@@ -180,25 +187,32 @@ kernel$(TSUFFIX).h : $(KERNEL_INTERFACE)
  
  
  cpuid.$(SUFFIX): $(KERNELDIR)/cpuid.S
@@ -83,10 +83,10 @@ index 9b468a6..87df5fa 100644
  #ifdef DYNAMIC_ARCH
  clean ::
 diff --git a/kernel/Makefile.L1 b/kernel/Makefile.L1
-index 9707032..7835f0d 100644
+index 7ad94118..3487a9e3 100644
 --- a/kernel/Makefile.L1
 +++ b/kernel/Makefile.L1
-@@ -522,339 +522,443 @@ XBLASOBJS	+= \
+@@ -545,217 +545,282 @@ endif
  
  
  $(KDIR)samax_k$(TSUFFIX).$(SUFFIX)  $(KDIR)samax_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SAMAXKERNEL)
@@ -432,6 +432,11 @@ index 9707032..7835f0d 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UCOMPLEX -DXDOUBLE $< -o $@
  
+ ifeq ($(BUILD_BFLOAT16),1)
+ $(KDIR)sbdot_k$(TSUFFIX).$(SUFFIX) $(KDIR)sbdot_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SBDOTKERNEL)
+@@ -771,126 +836,165 @@ $(KDIR)dbf16tod_k$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(BF16TOKERNEL)
+ endif
+ 
  $(KDIR)sdot_k$(TSUFFIX).$(SUFFIX) $(KDIR)sdot_k$(TPSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SDOTKERNEL)
 -	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -635,12 +640,12 @@ index 9707032..7835f0d 100644
  
  
 diff --git a/kernel/Makefile.L2 b/kernel/Makefile.L2
-index 2aeb8f0..e7c49d2 100644
+index 79399c34..6d605f82 100644
 --- a/kernel/Makefile.L2
 +++ b/kernel/Makefile.L2
-@@ -220,209 +220,278 @@ XBLASOBJS	+= \
- 	xgeru_k$(TSUFFIX).$(SUFFIX) xgerc_k$(TSUFFIX).$(SUFFIX) xgerv_k$(TSUFFIX).$(SUFFIX) xgerd_k$(TSUFFIX).$(SUFFIX)
+@@ -236,251 +236,320 @@ XBLASOBJS	+= \
  
+ ifneq "$(or $(BUILD_SINGLE), $(BUILD_DOUBLE), $(BUILD_COMPLEX))" ""
  $(KDIR)sgemv_n$(TSUFFIX).$(SUFFIX)  $(KDIR)sgemv_n$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SGEMVNKERNEL) $(TOPDIR)/common.h $(GEMVDEP)
 -	$(CC) -c $(CFLAGS) -UDOUBLE -UCOMPLEX  -UTRANS $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -650,7 +655,9 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -UDOUBLE -UCOMPLEX  -DTRANS  $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UDOUBLE -UCOMPLEX  -DTRANS  $< -o $@
+ endif
  
+ ifneq "$(or $(BUILD_DOUBLE),$(BUILD_COMPLEX16))" ""
  $(KDIR)dgemv_n$(TSUFFIX).$(SUFFIX)  $(KDIR)dgemv_n$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(DGEMVNKERNEL) $(TOPDIR)/common.h $(GEMVDEP)
 -	$(CC) -c $(CFLAGS) -DDOUBLE -UCOMPLEX  -UTRANS $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -660,6 +667,7 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -DDOUBLE -UCOMPLEX  -DTRANS $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DDOUBLE -UCOMPLEX  -DTRANS $< -o $@
+ endif
  
  $(KDIR)qgemv_n$(TSUFFIX).$(SUFFIX)  $(KDIR)qgemv_n$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(QGEMVNKERNEL)
 -	$(CC) -c $(CFLAGS) -DXDOUBLE -UCOMPLEX -UTRANS $< -o $@
@@ -671,6 +679,8 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DXDOUBLE -UCOMPLEX -DTRANS $< -o $@
  
+ 
+ ifneq "$(or $(BUILD_COMPLEX),$(BUILD_COMPLEX16))" ""
  $(KDIR)cgemv_n$(TSUFFIX).$(SUFFIX) $(KDIR)cgemv_n$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(CGEMVNKERNEL) $(TOPDIR)/common.h $(GEMVDEP)
 -	$(CC) -c $(CFLAGS) -UDOUBLE -DCOMPLEX -UTRANS -UCONJ -UXCONJ $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -710,6 +720,10 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -UDOUBLE -DCOMPLEX -DTRANS -DCONJ -DXCONJ $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UDOUBLE -DCOMPLEX -DTRANS -DCONJ -DXCONJ $< -o $@
+ endif
+ 
+ 
+ ifeq ($(BUILD_COMPLEX16),1)
  
  $(KDIR)zgemv_n$(TSUFFIX).$(SUFFIX) $(KDIR)zgemv_n$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(ZGEMVNKERNEL) $(TOPDIR)/common.h $(GEMVDEP)
 -	$(CC) -c $(CFLAGS) -DDOUBLE -DCOMPLEX -UTRANS -UCONJ -UXCONJ $< -o $@
@@ -750,6 +764,7 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -DDOUBLE -DCOMPLEX -DTRANS -DCONJ -DXCONJ $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DDOUBLE -DCOMPLEX -DTRANS -DCONJ -DXCONJ $< -o $@
+ endif
  
  $(KDIR)xgemv_n$(TSUFFIX).$(SUFFIX) $(KDIR)xgemv_n$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(XGEMVNKERNEL)
 -	$(CC) -c $(CFLAGS) -DXDOUBLE -DCOMPLEX -UTRANS -UCONJ -UXCONJ $< -o $@
@@ -791,6 +806,9 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DXDOUBLE -DCOMPLEX -DTRANS -DCONJ -DXCONJ $< -o $@
  
+ 
+ ifeq ($(BUILD_SINGLE),1)
+ 
  $(KDIR)ssymv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)ssymv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SSYMV_U_KERNEL) $(SSYMV_U_PARAM)
 -	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE -ULOWER $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -800,6 +818,10 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE -DLOWER $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UCOMPLEX -UDOUBLE -DLOWER $< -o $@
+ endif
+ 
+ 
+ ifeq ($(BUILD_DOUBLE),1)
  
  $(KDIR)dsymv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)dsymv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(DSYMV_U_KERNEL)  $(DSYMV_U_PARAM)
 -	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE -ULOWER $< -o $@
@@ -810,6 +832,7 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE -DLOWER $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UCOMPLEX -DDOUBLE -DLOWER $< -o $@
+ endif
  
  $(KDIR)qsymv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)qsymv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(QSYMV_U_KERNEL)
 -	$(CC) -c $(CFLAGS) -UCOMPLEX -DXDOUBLE -ULOWER $< -o $@
@@ -821,6 +844,8 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UCOMPLEX -DXDOUBLE -DLOWER $< -o $@
  
+ ifeq ($(BUILD_COMPLEX),1)
+ 
  $(KDIR)csymv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)csymv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(CSYMV_U_KERNEL)   $(CSYMV_U_PARAM)
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -UDOUBLE -ULOWER $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -830,6 +855,9 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -UDOUBLE -DLOWER $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -UDOUBLE -DLOWER $< -o $@
+ endif
+ 
+ ifeq ($(BUILD_COMPLEX16),1)
  
  $(KDIR)zsymv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)zsymv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(ZSYMV_U_KERNEL)   $(ZSYMV_U_PARAM)
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE -ULOWER $< -o $@
@@ -840,6 +868,7 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE -DLOWER $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -DDOUBLE -DLOWER $< -o $@
+ endif
  
  $(KDIR)xsymv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)xsymv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XSYMV_U_KERNEL)
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE -ULOWER $< -o $@
@@ -851,21 +880,29 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -DXDOUBLE -DLOWER $< -o $@
  
+ ifeq ($(BUILD_SINGLE),1)
+ 
  $(KDIR)sger_k$(TSUFFIX).$(SUFFIX)  $(KDIR)sger_k$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SGERKERNEL) $(SGERPARAM)
 -	$(CC) -c $(CFLAGS) -UDOUBLE $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UDOUBLE $< -o $@
+ endif
+ 
+ ifeq ($(BUILD_DOUBLE),1)
  
  $(KDIR)dger_k$(TSUFFIX).$(SUFFIX)  $(KDIR)dger_k$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(DGERKERNEL) $(DGERPARAM)
 -	$(CC) -c $(CFLAGS) -DDOUBLE $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DDOUBLE $< -o $@
+ endif
  
  $(KDIR)qger_k$(TSUFFIX).$(SUFFIX)  $(KDIR)qger_k$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(QGERKERNEL) $(QGERPARAM)
 -	$(CC) -c $(CFLAGS) -DXDOUBLE $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DXDOUBLE $< -o $@
  
+ ifeq ($(BUILD_COMPLEX),1)
+ 
  $(KDIR)cgeru_k$(TSUFFIX).$(SUFFIX)  $(KDIR)cgeru_k$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(CGERUKERNEL) $(CGERPARAM)
 -	$(CC) -c $(CFLAGS) -UDOUBLE -UCONJ $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -885,6 +922,9 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -UDOUBLE -DCONJ -DXCONJ $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UDOUBLE -DCONJ -DXCONJ $< -o $@
+ endif
+ 
+ ifeq ($(BUILD_COMPLEX16),1)
  
  $(KDIR)zgeru_k$(TSUFFIX).$(SUFFIX)  $(KDIR)zgeru_k$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(ZGERUKERNEL) $(ZGERPARAM)
 -	$(CC) -c $(CFLAGS) -DDOUBLE -UCONJ $< -o $@
@@ -905,6 +945,7 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -DDOUBLE -DCONJ -DXCONJ $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DDOUBLE -DCONJ -DXCONJ $< -o $@
+ endif
  
  $(KDIR)xgeru_k$(TSUFFIX).$(SUFFIX)  $(KDIR)xgeru_k$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XGERUKERNEL) $(XGERPARAM)
 -	$(CC) -c $(CFLAGS) -DXDOUBLE -UCONJ $< -o $@
@@ -926,6 +967,8 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DXDOUBLE -DCONJ -DXCONJ $< -o $@
  
+ ifeq ($(BUILD_COMPLEX),1)
+ 
  $(KDIR)chemv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)chemv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(CHEMV_U_KERNEL) $(CHEMV_U_PARAM)
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -UDOUBLE -ULOWER -DHEMV $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -945,6 +988,9 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -UDOUBLE -DLOWER -DHEMV -DHEMVREV $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -UDOUBLE -DLOWER -DHEMV -DHEMVREV $< -o $@
+ endif
+ 
+ ifeq ($(BUILD_COMPLEX16),1)
  
  $(KDIR)zhemv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)zhemv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(ZHEMV_U_KERNEL) $(ZHEMV_U_PARAM)
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE -ULOWER -DHEMV $< -o $@
@@ -965,7 +1011,7 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE -DLOWER -DHEMV -DHEMVREV $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -DDOUBLE -DLOWER -DHEMV -DHEMVREV $< -o $@
- 
+ endif
  $(KDIR)xhemv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)xhemv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XHEMV_U_KERNEL)
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE -ULOWER -DHEMV $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -986,17 +1032,18 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -DXDOUBLE -DLOWER -DHEMV -DHEMVREV $< -o $@
  
+ 
 diff --git a/kernel/Makefile.L3 b/kernel/Makefile.L3
-index da6c5fd..c3f92fb 100644
+index 2ba593c2..110674b2 100644
 --- a/kernel/Makefile.L3
 +++ b/kernel/Makefile.L3
-@@ -449,119 +449,150 @@ XGEMMOTCOPYOBJ_P = $(XGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
+@@ -526,119 +526,150 @@ XGEMMOTCOPYOBJ_P = $(XGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
  
- ifeq ($(BUILD_HALF),1)
- $(KDIR)shgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SHGEMM_BETA)
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ ifeq ($(BUILD_BFLOAT16),1)
+ $(KDIR)sbgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SBGEMM_BETA)
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  endif
  
  $(KDIR)sgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_BETA)
@@ -1030,51 +1077,55 @@ index da6c5fd..c3f92fb 100644
 +	$(CC) $$CFLAGS -c -DXDOUBLE -DCOMPLEX $< -o $@
  
  
- ifeq ($(BUILD_HALF), 1)
+ ifeq ($(BUILD_BFLOAT16), 1)
  
- $(KDIR)$(SHGEMMONCOPYOBJ) : $(KERNELDIR)/$(SHGEMMONCOPY)
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ $(KDIR)$(SBGEMMONCOPYOBJ) : $(KERNELDIR)/$(SBGEMMONCOPY)
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  
- $(KDIR)$(SHGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SHGEMMOTCOPY)
+ $(KDIR)$(SBGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SBGEMMOTCOPY)
  
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DHALF -UDOUBLE -UCOMPLEX $< -o shgemmotcopy.s
+-	$(CC) $(CFLAGS) -S -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o - > sbgemmotcopy.s
+-	m4 sbgemmotcopy.s > sbgemmotcopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX sbgemmotcopy_nomacros.s -o $@
+-	rm sbgemmotcopy.s sbgemmotcopy_nomacros.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DHALF -UDOUBLE -UCOMPLEX $< -o shgemmotcopy.s
- 	m4 shgemmotcopy.s > shgemmotcopy_nomacros.s
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX shgemmotcopy_nomacros.s -o $@
++	$(CC) $$CFLAGS -E -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o shgemmotcopy.s
++	m4 shgemmotcopy.s > shgemmotcopy_nomacros.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX shgemmotcopy_nomacros.s -o $@
- 	rm shgemmotcopy.s shgemmotcopy_nomacros.s
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX shgemmotcopy_nomacros.s -o $@
++	rm shgemmotcopy.s shgemmotcopy_nomacros.s
  else
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  endif
  
- ifneq ($(SHGEMM_UNROLL_M), $(SHGEMM_UNROLL_N))
+ ifneq ($(SBGEMM_UNROLL_M), $(SBGEMM_UNROLL_N))
  
- $(KDIR)$(SHGEMMINCOPYOBJ) : $(KERNELDIR)/$(SHGEMMINCOPY)
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ $(KDIR)$(SBGEMMINCOPYOBJ) : $(KERNELDIR)/$(SBGEMMINCOPY)
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  
- $(KDIR)$(SHGEMMITCOPYOBJ) : $(KERNELDIR)/$(SHGEMMITCOPY)
+ $(KDIR)$(SBGEMMITCOPYOBJ) : $(KERNELDIR)/$(SBGEMMITCOPY)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DHALF -UDOUBLE -UCOMPLEX $< -o shgemmitcopy.s
+-	$(CC) $(CFLAGS) -S -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o - > sbgemmitcopy.s
+-	m4 sbgemmitcopy.s > sbgemmitcopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX sbgemmitcopy_nomacros.s -o $@
+-	rm sbgemmitcopy.s sbgemmitcopy_nomacros.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DHALF -UDOUBLE -UCOMPLEX $< -o shgemmitcopy.s
- 	m4 shgemmitcopy.s > shgemmitcopy_nomacros.s
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX shgemmitcopy_nomacros.s -o $@
++	$(CC) $$CFLAGS -E -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o shgemmitcopy.s
++	m4 shgemmitcopy.s > shgemmitcopy_nomacros.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX shgemmitcopy_nomacros.s -o $@
- 	rm shgemmitcopy.s shgemmitcopy_nomacros.s
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX shgemmitcopy_nomacros.s -o $@
++	rm shgemmitcopy.s shgemmitcopy_nomacros.s
  else
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  endif
  
  endif
@@ -1087,9 +1138,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)$(SGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SGEMMOTCOPY)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -UDOUBLE -UCOMPLEX $< -o sgemmotcopy.s	
+-	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX $< -o - > sgemmotcopy.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -UDOUBLE -UCOMPLEX $< -o sgemmotcopy.s	
++	$(CC) $$CFLAGS -E -UDOUBLE -UCOMPLEX $< -o - > sgemmotcopy.s
  	m4 sgemmotcopy.s > sgemmotcopy_nomacros.s
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX sgemmotcopy_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1111,9 +1162,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)$(SGEMMITCOPYOBJ) : $(KERNELDIR)/$(SGEMMITCOPY)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -UDOUBLE -UCOMPLEX $< -o sgemmitcopy.s	
+-	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX $< -o - > sgemmitcopy.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -UDOUBLE -UCOMPLEX $< -o sgemmitcopy.s	
++	$(CC) $$CFLAGS -E -UDOUBLE -UCOMPLEX $< -o - > sgemmitcopy.s
  	m4 sgemmitcopy.s > sgemmitcopy_nomacros.s
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX sgemmitcopy_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1129,9 +1180,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)$(DGEMMONCOPYOBJ) : $(KERNELDIR)/$(DGEMMONCOPY)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DDOUBLE -UCOMPLEX $< -o dgemm_ncopy.s
+-	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_ncopy.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o dgemm_ncopy.s
++	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o - > dgemm_ncopy.s
  	m4 dgemm_ncopy.s > dgemm_ncopy_nomacros.s
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_ncopy_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1157,9 +1208,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)$(DGEMMITCOPYOBJ) : $(KERNELDIR)/$(DGEMMITCOPY)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DDOUBLE -UCOMPLEX $< -o dgemm_itcopy.s
+-	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_itcopy.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o dgemm_itcopy.s
++	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o - > dgemm_itcopy.s
  	m4 dgemm_itcopy.s > dgemm_itcopy_nomacros.s
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_itcopy_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1172,7 +1223,7 @@ index da6c5fd..c3f92fb 100644
  endif
  
  endif
-@@ -569,65 +600,81 @@ endif
+@@ -646,65 +677,81 @@ endif
  ifdef EXPRECISION
  
  $(KDIR)$(QGEMMONCOPYOBJ) : $(KERNELDIR)/$(QGEMMONCOPY)
@@ -1220,9 +1271,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)$(CGEMMITCOPYOBJ) : $(KERNELDIR)/$(CGEMMITCOPY)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -UDOUBLE -UCOMPLEX -E $< -o cgemm_itcopy.s
+-	$(CC) $(CFLAGS) -UDOUBLE -UCOMPLEX -S $< -o - > cgemm_itcopy.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -UDOUBLE -UCOMPLEX -E $< -o cgemm_itcopy.s
++	$(CC) $$CFLAGS -UDOUBLE -UCOMPLEX -E $< -o - > cgemm_itcopy.s
  	m4 cgemm_itcopy.s > cgemm_itcopy_nomacros.s
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX cgemm_itcopy_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1255,9 +1306,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)$(ZGEMMITCOPYOBJ) : $(KERNELDIR)/$(ZGEMMITCOPY)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DDOUBLE -UCOMPLEX $< -o zgemm_itcopy.s
+-	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > zgemm_itcopy.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o zgemm_itcopy.s
++	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o - > zgemm_itcopy.s
  	m4 zgemm_itcopy.s > zgemm_itcopy_nomacros.s
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX zgemm_itcopy_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1270,7 +1321,7 @@ index da6c5fd..c3f92fb 100644
  endif
  
  endif
-@@ -635,18 +682,22 @@ endif
+@@ -712,18 +759,22 @@ endif
  ifdef EXPRECISION
  
  $(KDIR)$(XGEMMONCOPYOBJ) : $(KERNELDIR)/$(XGEMMONCOPY)
@@ -1297,13 +1348,13 @@ index da6c5fd..c3f92fb 100644
  
  endif
  
-@@ -654,1780 +705,2344 @@ endif
+@@ -731,12 +782,15 @@ endif
  
  $(KDIR)sgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL) $(SGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -UDOUBLE -UCOMPLEX  $< -o sgemm_kernel$(TSUFFIX).s
+-	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX  $< -o - > sgemm_kernel$(TSUFFIX).s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -UDOUBLE -UCOMPLEX  $< -o sgemm_kernel$(TSUFFIX).s
++	$(CC) $$CFLAGS -E -UDOUBLE -UCOMPLEX  $< -o - > sgemm_kernel$(TSUFFIX).s
  	m4 sgemm_kernel$(TSUFFIX).s > sgemm_kernel$(TSUFFIX)_nomacros.s
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX sgemm_kernel$(TSUFFIX)_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1315,30 +1366,33 @@ index da6c5fd..c3f92fb 100644
 +	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
  endif
  
- ifeq ($(BUILD_HALF), 1)
+ ifdef USE_DIRECT_SGEMM
+@@ -750,1768 +804,2329 @@ ifeq ($(BUILD_BFLOAT16), 1)
  
- $(KDIR)shgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SHGEMMKERNEL) $(SHGEMMDEPEND)
+ $(KDIR)sbgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SBGEMMKERNEL) $(SBGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DHALF -UDOUBLE -UCOMPLEX  $< -o shgemm_kernel$(TSUFFIX).s
+-	$(CC) $(CFLAGS) -S -DBFLOAT16 -UDOUBLE -UCOMPLEX  $< -o - > sbgemm_kernel$(TSUFFIX).s
+-	m4 sbgemm_kernel$(TSUFFIX).s > sbgemm_kernel$(TSUFFIX)_nomacros.s
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX sbgemm_kernel$(TSUFFIX)_nomacros.s -o $@
+-	rm sbgemm_kernel$(TSUFFIX).s sbgemm_kernel$(TSUFFIX)_nomacros.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DHALF -UDOUBLE -UCOMPLEX  $< -o shgemm_kernel$(TSUFFIX).s
- 	m4 shgemm_kernel$(TSUFFIX).s > shgemm_kernel$(TSUFFIX)_nomacros.s
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX shgemm_kernel$(TSUFFIX)_nomacros.s -o $@
++	$(CC) $$CFLAGS -E -DHALF -UDOUBLE -UCOMPLEX  $< -o - > shgemm_kernel$(TSUFFIX).s
++	m4 shgemm_kernel$(TSUFFIX).s > shgemm_kernel$(TSUFFIX)_nomacros.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX shgemm_kernel$(TSUFFIX)_nomacros.s -o $@
- 	rm shgemm_kernel$(TSUFFIX).s shgemm_kernel$(TSUFFIX)_nomacros.s
++	rm shgemm_kernel$(TSUFFIX).s shgemm_kernel$(TSUFFIX)_nomacros.s
  else
--	$(CC) $(CFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  endif
  endif
  
  $(KDIR)dgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMMKERNEL) $(DGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DDOUBLE -UCOMPLEX $< -o dgemm_kernel$(TSUFFIX).s
+-	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_kernel$(TSUFFIX).s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o dgemm_kernel$(TSUFFIX).s
++	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o - > dgemm_kernel$(TSUFFIX).s
  	m4 dgemm_kernel$(TSUFFIX).s > dgemm_kernel$(TSUFFIX)_nomacros.s
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_kernel$(TSUFFIX)_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1357,9 +1411,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)cgemm_kernel_n$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -UDOUBLE -DCOMPLEX -DNN $< -o cgemm_kernel_n.s
+-	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DNN $< -o - > cgemm_kernel_n.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DNN $< -o cgemm_kernel_n.s
++	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DNN $< -o - > cgemm_kernel_n.s
  	m4 cgemm_kernel_n.s > cgemm_kernel_n_nomacros.s
 -	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNN cgemm_kernel_n_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1373,9 +1427,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)cgemm_kernel_l$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -UDOUBLE -DCOMPLEX -DCN $< -o cgemm_kernel_l.s
+-	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DCN $< -o - > cgemm_kernel_l.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DCN $< -o cgemm_kernel_l.s
++	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DCN $< -o - > cgemm_kernel_l.s
  	m4 cgemm_kernel_l.s > cgemm_kernel_l_nomacros.s
 -	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DCN cgemm_kernel_l_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1389,9 +1443,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)cgemm_kernel_r$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -UDOUBLE -DCOMPLEX -DNC  $< -o cgemm_kernel_r.s
+-	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DNC  $< -o - > cgemm_kernel_r.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DNC  $< -o cgemm_kernel_r.s
++	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DNC  $< -o - > cgemm_kernel_r.s
  	m4 cgemm_kernel_r.s > cgemm_kernel_r_nomacros.s
 -	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNC cgemm_kernel_r_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1405,9 +1459,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)cgemm_kernel_b$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -UDOUBLE -DCOMPLEX -DCC $< -o cgemm_kernel_b.s
+-	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DCC $< -o - > cgemm_kernel_b.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DCC $< -o cgemm_kernel_b.s
++	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DCC $< -o - > cgemm_kernel_b.s
  	m4 cgemm_kernel_b.s > cgemm_kernel_b_nomacros.s
 -	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DCC cgemm_kernel_b_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1421,9 +1475,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)zgemm_kernel_n$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DDOUBLE -DCOMPLEX -DNN $< -o zgemm_kernel_n.s
+-	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DNN $< -o - > zgemm_kernel_n.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DNN $< -o zgemm_kernel_n.s
++	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DNN $< -o - > zgemm_kernel_n.s
  	m4 zgemm_kernel_n.s > zgemm_kernel_n_nomacros.s
 -	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DNN zgemm_kernel_n_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1437,9 +1491,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)zgemm_kernel_l$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DDOUBLE -DCOMPLEX -DCN $< -o zgemm_kernel_l.s
+-	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DCN $< -o - > zgemm_kernel_l.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DCN $< -o zgemm_kernel_l.s
++	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DCN $< -o - > zgemm_kernel_l.s
  	m4 zgemm_kernel_l.s > zgemm_kernel_l_nomacros.s
 -	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DCN zgemm_kernel_l_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1453,9 +1507,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)zgemm_kernel_r$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DDOUBLE -DCOMPLEX -DNC $< -o zgemm_kernel_r.s
+-	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DNC $< -o - > zgemm_kernel_r.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DNC $< -o zgemm_kernel_r.s
++	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DNC $< -o - > zgemm_kernel_r.s
  	m4 zgemm_kernel_r.s > zgemm_kernel_r_nomacros.s
 -	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DNC zgemm_kernel_r_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1469,9 +1523,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)zgemm_kernel_b$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DDOUBLE -DCOMPLEX -DCC $< -o zgemm_kernel_b.s
+-	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DCC $< -o - > zgemm_kernel_b.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DCC $< -o zgemm_kernel_b.s
++	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DCC $< -o - > zgemm_kernel_b.s
  	m4 zgemm_kernel_b.s > zgemm_kernel_b_nomacros.s
 -	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DCC zgemm_kernel_b_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1507,9 +1561,9 @@ index da6c5fd..c3f92fb 100644
  ifdef USE_TRMM
  $(KDIR)strmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o strmmkernel_ln.s	
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > strmmkernel_ln.s	
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o strmmkernel_ln.s	
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > strmmkernel_ln.s	
  	m4 strmmkernel_ln.s > strmmkernel_ln_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA strmmkernel_ln_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1523,9 +1577,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)strmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o strmmkernel_lt.s	
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > strmmkernel_lt.s	
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o strmmkernel_lt.s	
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > strmmkernel_lt.s	
  	m4 strmmkernel_lt.s > strmmkernel_lt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA strmmkernel_lt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1539,9 +1593,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)strmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o strmmkernel_rn.s	
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > strmmkernel_rn.s	
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o strmmkernel_rn.s	
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > strmmkernel_rn.s	
  	m4 strmmkernel_rn.s > strmmkernel_rn_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA strmmkernel_rn_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1555,9 +1609,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)strmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
  	m4 strmm_kernel_rt.s > strmm_kernel_rt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1571,9 +1625,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)dtrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o dtrmm_kernel_ln.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > dtrmm_kernel_ln.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o dtrmm_kernel_ln.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > dtrmm_kernel_ln.s
  	m4 dtrmm_kernel_ln.s > dtrmm_kernel_ln_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA dtrmm_kernel_ln_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1587,9 +1641,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)dtrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o dtrmm_kernel_lt.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > dtrmm_kernel_lt.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o dtrmm_kernel_lt.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > dtrmm_kernel_lt.s
  	m4 dtrmm_kernel_lt.s > dtrmm_kernel_lt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA dtrmm_kernel_lt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1603,9 +1657,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)dtrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o dtrmm_kernel_rn.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > dtrmm_kernel_rn.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o dtrmm_kernel_rn.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > dtrmm_kernel_rn.s
  	m4 dtrmm_kernel_rn.s > dtrmm_kernel_rn_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA dtrmm_kernel_rn_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1619,9 +1673,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)dtrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o dtrmm_kernel_rt.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > dtrmm_kernel_rt.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o dtrmm_kernel_rt.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > dtrmm_kernel_rt.s
  	m4 dtrmm_kernel_rt.s > dtrmm_kernel_rt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA dtrmm_kernel_rt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1655,9 +1709,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ctrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN  $< -o ctrmm_kernel_ln.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN  $< -o - > ctrmm_kernel_ln.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN  $< -o ctrmm_kernel_ln.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN  $< -o - > ctrmm_kernel_ln.s
  	m4 ctrmm_kernel_ln.s > ctrmm_kernel_ln_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN ctrmm_kernel_ln_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1671,9 +1725,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ctrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o ctrmm_kernel_lt.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_lt.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o ctrmm_kernel_lt.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_lt.s
  	m4 ctrmm_kernel_lt.s > ctrmm_kernel_lt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN ctrmm_kernel_lt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1687,9 +1741,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ctrmm_kernel_LR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o ctrmm_kernel_lr.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lr.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o ctrmm_kernel_lr.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lr.s
  	m4 ctrmm_kernel_lr.s > ctrmm_kernel_lr_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN  ctrmm_kernel_lr_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1703,9 +1757,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ctrmm_kernel_LC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o ctrmm_kernel_lc.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lc.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o ctrmm_kernel_lc.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lc.s
  	m4 ctrmm_kernel_lc.s > ctrmm_kernel_lc_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN ctrmm_kernel_lc_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1719,9 +1773,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ctrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o ctrmm_kernel_rn.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rn.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o ctrmm_kernel_rn.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rn.s
  	m4 ctrmm_kernel_rn.s > ctrmm_kernel_rn_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN ctrmm_kernel_rn_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1735,9 +1789,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ctrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o ctrmm_kernel_rt.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rt.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o ctrmm_kernel_rt.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rt.s
  	m4 ctrmm_kernel_rt.s > ctrmm_kernel_rt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN ctrmm_kernel_rt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1751,9 +1805,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ctrmm_kernel_RR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o ctrmm_kernel_rr.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_rr.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o ctrmm_kernel_rr.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_rr.s
  	m4 ctrmm_kernel_rr.s > ctrmm_kernel_rr_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC ctrmm_kernel_rr_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1767,9 +1821,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ctrmm_kernel_RC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o ctrmm_kernel_RC.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_RC.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o ctrmm_kernel_RC.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_RC.s
  	m4 ctrmm_kernel_RC.s > ctrmm_kernel_RC_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC ctrmm_kernel_RC_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1783,9 +1837,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ztrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o ztrmm_kernel_ln.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_ln.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o ztrmm_kernel_ln.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_ln.s
  	m4 ztrmm_kernel_ln.s > ztrmm_kernel_ln_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN ztrmm_kernel_ln_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1799,9 +1853,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ztrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o ztrmm_kernel_lt.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_lt.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o ztrmm_kernel_lt.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_lt.s
  	m4 ztrmm_kernel_lt.s > ztrmm_kernel_lt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN ztrmm_kernel_lt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1815,9 +1869,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ztrmm_kernel_LR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o ztrmm_kernel_lr.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lr.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o ztrmm_kernel_lr.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lr.s
  	m4 ztrmm_kernel_lr.s > ztrmm_kernel_lr_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN ztrmm_kernel_lr_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1831,9 +1885,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ztrmm_kernel_LC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o ztrmm_kernel_lc.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lc.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o ztrmm_kernel_lc.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lc.s
  	m4 ztrmm_kernel_lc.s >ztrmm_kernel_lc_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN ztrmm_kernel_lc_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1847,9 +1901,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ztrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o ztrmm_kernel_rn.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rn.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o ztrmm_kernel_rn.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rn.s
  	m4 ztrmm_kernel_rn.s > ztrmm_kernel_rn_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN ztrmm_kernel_rn_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1863,9 +1917,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ztrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o ztrmm_kernel_rt.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rt.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o ztrmm_kernel_rt.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rt.s
  	m4 ztrmm_kernel_rt.s > ztrmm_kernel_rt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN ztrmm_kernel_rt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1879,9 +1933,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ztrmm_kernel_RR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o ztrmm_kernel_rr.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rr.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o ztrmm_kernel_rr.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rr.s
  	m4 ztrmm_kernel_rr.s > ztrmm_kernel_rr_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC ztrmm_kernel_rr_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1895,9 +1949,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)ztrmm_kernel_RC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o ztrmm_kernel_rc.s
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rc.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o ztrmm_kernel_rc.s
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rc.s
  	m4 ztrmm_kernel_rc.s > ztrmm_kernel_rc_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC ztrmm_kernel_rc_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -1927,9 +1981,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)strmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
  	m4 strmm_kernel_rt.s > strmm_kernel_rt_nomacros.s
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -2147,9 +2201,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)dtrsm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRSMKERNEL_LT) $(DTRSMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o dtrsm_kernel_lt.s
+-	$(CC) $(CFLAGS) -S -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o - > dtrsm_kernel_lt.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o dtrsm_kernel_lt.s
++	$(CC) $$CFLAGS -E -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o - > dtrsm_kernel_lt.s
  	m4 dtrsm_kernel_lt.s > dtrsm_kernel_lt_nomacros.s
 -	$(CC) -c $(CFLAGS) -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ dtrsm_kernel_lt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -3998,11 +4052,11 @@ index da6c5fd..c3f92fb 100644
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
 +	$(CC) $$PFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
  
- ifeq ($(BUILD_HALF),1)
- $(KDIR)shgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SHGEMM_BETA)
--	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ ifeq ($(BUILD_BFLOAT16),1)
+ $(KDIR)sbgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SBGEMM_BETA)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  endif
  
  $(KDIR)dgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(DGEMM_BETA)
@@ -4031,27 +4085,27 @@ index da6c5fd..c3f92fb 100644
 +	$(CC) $$PFLAGS -c -DXDOUBLE -DCOMPLEX $< -o $@
  
  
- ifeq ($(BUILD_HALF), 1)
- $(SHGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMONCOPY)
--	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ ifeq ($(BUILD_BFLOAT16), 1)
+ $(SBGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SBGEMMONCOPY)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  
- $(SHGEMMOTCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMOTCOPY)
--	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ $(SBGEMMOTCOPYOBJ_P) : $(KERNELDIR)/$(SBGEMMOTCOPY)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  
- ifneq ($(SHGEMM_UNROLL_M), $(SHGEMM_UNROLL_N))
- $(SHGEMMINCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMINCOPY)
--	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ ifneq ($(SBGEMM_UNROLL_M), $(SBGEMM_UNROLL_N))
+ $(SBGEMMINCOPYOBJ_P) : $(KERNELDIR)/$(SBGEMMINCOPY)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  
- $(SHGEMMITCOPYOBJ_P) : $(KERNELDIR)/$(SHGEMMITCOPY)
--	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ $(SBGEMMITCOPYOBJ_P) : $(KERNELDIR)/$(SBGEMMITCOPY)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  
  endif
  endif
@@ -4080,7 +4134,7 @@ index da6c5fd..c3f92fb 100644
  
  endif
  
- $(D<GEMMONCOPYOBJ_P) : $(KERNELDIR)/$(DGEMMONCOPY)
+ $(DGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(DGEMMONCOPY)
 -	$(CC) $(PFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
 +	$(CC) $$PFLAGS -c -DDOUBLE -UCOMPLEX $< -o $@
@@ -4206,13 +4260,13 @@ index da6c5fd..c3f92fb 100644
  
  endif
  
-@@ -2436,1301 +3051,1732 @@ endif
+@@ -2520,1301 +3135,1732 @@ endif
  
- ifeq ($(BUILD_HALF), 1)
- $(KDIR)shgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SHGEMMKERNEL) $(SHGEMMDEPEND)
--	$(CC) $(PFLAGS) -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
+ ifeq ($(BUILD_BFLOAT16), 1)
+ $(KDIR)sbgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SBGEMMKERNEL) $(SBGEMMDEPEND)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -c -DHALF -UDOUBLE -UCOMPLEX $< -o $@
++	$(CC) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  endif
  
  $(KDIR)sgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL) $(SGEMMDEPEND)
@@ -4242,9 +4296,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)cgemm_kernel_r$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
  ifeq ($(OS), AIX)
--	$(CC) $(PFLAGS) -E -UDOUBLE -DCOMPLEX -DNC $< -o cgemm_kernel_r.s
+-	$(CC) $(PFLAGS) -S -UDOUBLE -DCOMPLEX -DNC $< -o - > cgemm_kernel_r.s
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -E -UDOUBLE -DCOMPLEX -DNC $< -o cgemm_kernel_r.s
++	$(CC) $$PFLAGS -E -UDOUBLE -DCOMPLEX -DNC $< -o - > cgemm_kernel_r.s
  	m4 cgemm_kernel_r.s > cgemm_kernel_r_nomacros.s
 -	$(CC) $(PFLAGS) -c -UDOUBLE -DCOMPLEX -DNC cgemm_kernel_r_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
@@ -4318,9 +4372,9 @@ index da6c5fd..c3f92fb 100644
  
  $(KDIR)strmm_kernel_RT$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
  ifeq ($(OS), AIX)
--	$(CC) $(CFLAGS) -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o strmm_kernel_rt.s	
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
  	m4 strmmkernel_rn.s > strmm_kernel_rt_nomacros.s
 -	$(CC) $(PFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
@@ -6370,7 +6424,7 @@ index da6c5fd..c3f92fb 100644
  
  
  #####  BLAS extensions ######
-@@ -3740,112 +4786,128 @@ DOMATCOPY_CN = ../arm/omatcopy_cn.c
+@@ -3824,112 +4870,128 @@ DOMATCOPY_CN = ../arm/omatcopy_cn.c
  endif
  
  $(KDIR)domatcopy_k_cn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DOMATCOPY_CN)
@@ -6515,7 +6569,7 @@ index da6c5fd..c3f92fb 100644
  
  
  ifndef COMATCOPY_CN
-@@ -3853,112 +4915,128 @@ COMATCOPY_CN = ../arm/zomatcopy_cn.c
+@@ -3937,112 +4999,128 @@ COMATCOPY_CN = ../arm/zomatcopy_cn.c
  endif
  
  $(KDIR)comatcopy_k_cn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(COMATCOPY_CN)
@@ -6660,7 +6714,7 @@ index da6c5fd..c3f92fb 100644
  
  
  
-@@ -3967,112 +5045,128 @@ ZOMATCOPY_CN = ../arm/zomatcopy_cn.c
+@@ -4051,112 +5129,128 @@ ZOMATCOPY_CN = ../arm/zomatcopy_cn.c
  endif
  
  $(KDIR)zomatcopy_k_cn$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZOMATCOPY_CN)
@@ -6805,7 +6859,7 @@ index da6c5fd..c3f92fb 100644
  
  
  ifndef SGEADD_K
-@@ -4080,26 +5174,30 @@ SGEADD_K = ../generic/geadd.c
+@@ -4164,26 +5258,30 @@ SGEADD_K = ../generic/geadd.c
  endif
  
  $(KDIR)sgeadd_k$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEADD_K)
@@ -6841,7 +6895,7 @@ index da6c5fd..c3f92fb 100644
 +	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -UROWM $< -o $@
  
 diff --git a/kernel/Makefile.LA b/kernel/Makefile.LA
-index 8834271..d9c6739 100644
+index 88342718..d9c67390 100644
 --- a/kernel/Makefile.LA
 +++ b/kernel/Makefile.LA
 @@ -14,38 +14,50 @@ XBLASOBJS	+= xneg_tcopy$(TSUFFIX).$(SUFFIX) xlaswp_ncopy$(TSUFFIX).$(SUFFIX)
diff --git a/SOURCES/openblas-0.3.10-libname.patch b/SOURCES/openblas-0.3.12-libname.patch
similarity index 92%
rename from SOURCES/openblas-0.3.10-libname.patch
rename to SOURCES/openblas-0.3.12-libname.patch
index 5787ab6..d75de48 100644
--- a/SOURCES/openblas-0.3.10-libname.patch
+++ b/SOURCES/openblas-0.3.12-libname.patch
@@ -1,8 +1,8 @@
 diff --git a/Makefile.system b/Makefile.system
-index 47d8eec..c947a19 100644
+index 30d8f4cc..803219a8 100644
 --- a/Makefile.system
 +++ b/Makefile.system
-@@ -1331,16 +1331,16 @@ ifndef SMP
+@@ -1425,16 +1425,16 @@ ifndef SMP
  LIBNAME		= $(LIBPREFIX)_$(LIBCORE)$(REVISION).$(LIBSUFFIX)
  LIBNAME_P	= $(LIBPREFIX)_$(LIBCORE)$(REVISION)_p.$(LIBSUFFIX)
  else
diff --git a/SOURCES/openblas-0.3.10-noopt.patch b/SOURCES/openblas-0.3.12-noopt.patch
similarity index 85%
rename from SOURCES/openblas-0.3.10-noopt.patch
rename to SOURCES/openblas-0.3.12-noopt.patch
index f331b8c..499571b 100644
--- a/SOURCES/openblas-0.3.10-noopt.patch
+++ b/SOURCES/openblas-0.3.12-noopt.patch
@@ -1,5 +1,5 @@
 diff --git a/lapack-netlib/INSTALL/Makefile b/lapack-netlib/INSTALL/Makefile
-index 1007c1b..348b8d6 100644
+index 1007c1bc..348b8d60 100644
 --- a/lapack-netlib/INSTALL/Makefile
 +++ b/lapack-netlib/INSTALL/Makefile
 @@ -46,5 +46,5 @@ cleanexe:
@@ -11,10 +11,10 @@ index 1007c1b..348b8d6 100644
 +slamch.o: slamch.f ; $(FC) $(FFLAGS) -c -o $@ $<
 +dlamch.o: dlamch.f ; $(FC) $(FFLAGS) -c -o $@ $<
 diff --git a/lapack-netlib/SRC/Makefile b/lapack-netlib/SRC/Makefile
-index 9f79e20..cf47a02 100644
+index 83baac87..bc42e82d 100644
 --- a/lapack-netlib/SRC/Makefile
 +++ b/lapack-netlib/SRC/Makefile
-@@ -613,9 +613,9 @@ cleanobj:
+@@ -643,9 +643,9 @@ cleanobj:
  cleanlib:
  	rm -f $(LAPACKLIB)
  
@@ -31,14 +31,17 @@ index 9f79e20..cf47a02 100644
 +cla_wwaddw.o: cla_wwaddw.f ; $(FC) $(FFLAGS) -c -o $@ $<
 +zla_wwaddw.o: zla_wwaddw.f ; $(FC) $(FFLAGS) -c -o $@ $<
 diff --git a/lapack-netlib/TESTING/MATGEN/Makefile b/lapack-netlib/TESTING/MATGEN/Makefile
-index 87432fd..bd484e6 100644
+index e21ebd6c..8582e944 100644
 --- a/lapack-netlib/TESTING/MATGEN/Makefile
 +++ b/lapack-netlib/TESTING/MATGEN/Makefile
-@@ -97,5 +97,5 @@ cleanobj:
- cleanlib:
+@@ -110,8 +110,8 @@ cleanlib:
  	rm -f $(TMGLIB)
  
+ ifeq ($(filter $(BUILD_SINGLE) $(BUILD_COMPLEX),1),)
 -slaran.o: slaran.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
--dlaran.o: dlaran.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
 +slaran.o: slaran.f ; $(FC) $(FFLAGS) -c -o $@ $<
+ endif
+ ifeq ($(filter $(BUILD_DOUBLE) $(BUILD_COMPLEX16),1),)
+-dlaran.o: dlaran.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
 +dlaran.o: dlaran.f ; $(FC) $(FFLAGS) -c -o $@ $<
+ endif
diff --git a/SOURCES/openblas-0.3.10-power-optimize.patch b/SOURCES/openblas-0.3.12-power-optimize.patch
similarity index 67%
rename from SOURCES/openblas-0.3.10-power-optimize.patch
rename to SOURCES/openblas-0.3.12-power-optimize.patch
index b2d9766..5d82b6f 100644
--- a/SOURCES/openblas-0.3.10-power-optimize.patch
+++ b/SOURCES/openblas-0.3.12-power-optimize.patch
@@ -1,5 +1,5 @@
 diff --git a/kernel/power/drot.c b/kernel/power/drot.c
-index baeb542..daeae9f 100644
+index 951c2f9c..98788c0f 100644
 --- a/kernel/power/drot.c
 +++ b/kernel/power/drot.c
 @@ -37,8 +37,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
@@ -8,11 +8,11 @@ index baeb542..daeae9f 100644
  
 -#pragma GCC optimize "O1"
 -
- #if defined(POWER8) || defined(POWER9)
+ #if defined(POWER8) || defined(POWER9) || defined(POWER10)
+ #if defined(__VEC__) || defined(__ALTIVEC__)
  #include "drot_microk_power8.c"
- #endif
 diff --git a/kernel/power/srot.c b/kernel/power/srot.c
-index 6af813c..0a172d5 100644
+index a53342f6..73b477d9 100644
 --- a/kernel/power/srot.c
 +++ b/kernel/power/srot.c
 @@ -37,8 +37,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
@@ -21,11 +21,11 @@ index 6af813c..0a172d5 100644
  
 -#pragma GCC optimize "O1"
 -
- #if defined(POWER8) || defined(POWER9)
+ #if defined(POWER8) || defined(POWER9) || defined(POWER10)
+ #if defined(__VEC__) || defined(__ALTIVEC__)
  #include "srot_microk_power8.c"
- #endif
 diff --git a/kernel/power/zscal.c b/kernel/power/zscal.c
-index a1b441d..3c80ba0 100644
+index 5526f4d6..3c2efbfa 100644
 --- a/kernel/power/zscal.c
 +++ b/kernel/power/zscal.c
 @@ -36,8 +36,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
@@ -34,6 +34,6 @@ index a1b441d..3c80ba0 100644
  
 -#pragma GCC optimize "O1"
 -
- #if defined(POWER8) || defined(POWER9)
+ #if defined(POWER8) || defined(POWER9) || defined(POWER10)
+ #if defined(__VEC__) || defined(__ALTIVEC__)
  #if defined(DOUBLE)
- #include "zscal_microk_power8.c"
diff --git a/SOURCES/openblas-0.3.10-system-lapack.patch b/SOURCES/openblas-0.3.12-system-lapack.patch
similarity index 100%
rename from SOURCES/openblas-0.3.10-system-lapack.patch
rename to SOURCES/openblas-0.3.12-system-lapack.patch
diff --git a/SOURCES/openblas-0.3.10-tests.patch b/SOURCES/openblas-0.3.12-tests.patch
similarity index 89%
rename from SOURCES/openblas-0.3.10-tests.patch
rename to SOURCES/openblas-0.3.12-tests.patch
index 22ed290..0bd9864 100644
--- a/SOURCES/openblas-0.3.10-tests.patch
+++ b/SOURCES/openblas-0.3.12-tests.patch
@@ -1,8 +1,8 @@
 diff --git a/Makefile b/Makefile
-index ae8f7de..61f325c 100644
+index a9af62a2..2f0f11c0 100644
 --- a/Makefile
 +++ b/Makefile
-@@ -133,13 +133,13 @@ tests :
+@@ -138,13 +138,13 @@ tests :
  ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
  	touch $(LIBNAME)
  ifndef NO_FBLAS
@@ -11,7 +11,7 @@ index ae8f7de..61f325c 100644
  endif
 -	$(MAKE) -C utest all
 +	$(MAKE) -C utest FC="$(FC)" CC="$(CC)" COMMON_OPT="$(COMMON_OPT)" FCOMMON_OPT="$(FCOMMON_OPT)" all
- ifndef NO_CBLAS
+ ifneq ($(NO_CBLAS), 1)
 -	$(MAKE) -C ctest all
 +	$(MAKE) -C ctest FC="$(FC)" CC="$(CC)" COMMON_OPT="$(COMMON_OPT)" FCOMMON_OPT="$(FCOMMON_OPT)" all
  ifeq ($(CPP_THREAD_SAFETY_TEST), 1)
diff --git a/SPECS/openblas.spec b/SPECS/openblas.spec
index 73e99b8..d1d590f 100644
--- a/SPECS/openblas.spec
+++ b/SPECS/openblas.spec
@@ -14,31 +14,25 @@
 # "obsoleted" features are still kept in the spec.
 
 Name:           openblas
-Version:        0.3.10
-Release:        2%{?dist}
+Version:        0.3.12
+Release:        1%{?dist}
 Summary:        An optimized BLAS library based on GotoBLAS2
 Group:          Development/Libraries
 License:        BSD
 URL:            https://github.com/xianyi/OpenBLAS/
 Source0:        https://github.com/xianyi/OpenBLAS/archive/v%{version}.tar.gz
 # Use system lapack
-Patch0:         openblas-0.3.10-system-lapack.patch
+Patch0:         openblas-0.3.12-system-lapack.patch
 # Drop extra p from threaded library name
-Patch1:         openblas-0.3.10-libname.patch
+Patch1:         openblas-0.3.12-libname.patch
 # Supply the proper flags to the test makefile
-Patch2:         openblas-0.3.10-tests.patch
+Patch2:         openblas-0.3.12-tests.patch
 # Enable optimizations for all LAPACK sources
-Patch3:         openblas-0.3.10-noopt.patch
+Patch3:         openblas-0.3.12-noopt.patch
 # Pass ASMFLAGS to assembler compiler
-Patch4:         openblas-0.3.10-asmflags.patch
+Patch4:         openblas-0.3.12-asmflags.patch
 # Remove optimization pragmas on ppc64le
-Patch5:         openblas-0.3.10-power-optimize.patch
-# https://github.com/xianyi/OpenBLAS/pull/2669
-Patch6:         openblas-0.3.10-zarch-gcc-version-detection.patch
-# https://github.com/xianyi/OpenBLAS/pull/2672
-Patch7:         openblas-0.3.10-concurrency.patch
-# https://github.com/Reference-LAPACK/lapack/pull/458
-Patch8:         openblas-0.3.10-lapacke-macro.patch
+Patch5:         openblas-0.3.12-power-optimize.patch
 
 BuildRequires:  gcc
 BuildRequires:  gcc-gfortran
@@ -239,9 +233,6 @@ cd OpenBLAS-%{version}
 %patch3 -p1 -b .noopt
 %patch4 -p1 -b .asmflags
 %patch5 -p1 -b .power-optimize
-%patch6 -p1 -b .zarch-gcc-version-detection
-%patch7 -p1 -b .concurrency
-%patch8 -p1 -b .lapacke-macro
 
 # Fix source permissions
 find -name \*.f -exec chmod 644 {} \;
@@ -684,6 +675,10 @@ rm -rf %{buildroot}%{_libdir}/pkgconfig
 %endif
 
 %changelog
+* Wed Oct 28 2020 Nikola Forró <nforro@redhat.com> - 0.3.12-1
+- Rebase to version 0.3.12
+  related: #1847435
+
 * Wed Oct 21 2020 Nikola Forró <nforro@redhat.com> - 0.3.10-2
 - Fix macro used in LAPACKE_zgesvdq
   related: #1847435