diff --git a/.gitignore b/.gitignore
index d55f61c..f697ae0 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1 +1 @@
-SOURCES/v0.3.3.tar.gz
+SOURCES/v0.3.12.tar.gz
diff --git a/.openblas.metadata b/.openblas.metadata
index ac9f8a7..058300c 100644
--- a/.openblas.metadata
+++ b/.openblas.metadata
@@ -1 +1 @@
-bff159c528c1a860cee4976114d224da32d302a2 SOURCES/v0.3.3.tar.gz
+ae647fed597ae891a7f122b9ddc6b15d4b7e0656 SOURCES/v0.3.12.tar.gz
diff --git a/SOURCES/openblas-0.2.15-constructor.patch b/SOURCES/openblas-0.2.15-constructor.patch
deleted file mode 100644
index 6b6a092..0000000
--- a/SOURCES/openblas-0.2.15-constructor.patch
+++ /dev/null
@@ -1,19 +0,0 @@
-diff -up OpenBLAS-0.2.15/driver/others/memory.c.priority OpenBLAS-0.2.15/driver/others/memory.c
---- OpenBLAS-0.2.15/driver/others/memory.c.priority	2015-10-27 21:44:50.000000000 +0100
-+++ OpenBLAS-0.2.15/driver/others/memory.c	2016-01-13 21:12:01.862225898 +0100
-@@ -146,8 +146,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF
- #define CONSTRUCTOR	__attribute__ ((constructor))
- #define DESTRUCTOR	__attribute__ ((destructor))
- #else
-+#if __GNUC__ && INIT_PRIORITY && ((GCC_VERSION >= 40300) || (CLANG_VERSION >= 20900))
- #define CONSTRUCTOR	__attribute__ ((constructor(101)))
- #define DESTRUCTOR	__attribute__ ((destructor(101)))
-+#elif __GNUC__ && INIT_PRIORITY
-+#define CONSTRUCTOR	__attribute__ ((constructor))
-+#define DESTRUCTOR	__attribute__ ((destructor))
-+#else
-+#define CONSTRUCTOR
-+#define DESTRUCTOR
- #endif
- 
- #ifdef DYNAMIC_ARCH
diff --git a/SOURCES/openblas-0.2.20-asmflags.patch b/SOURCES/openblas-0.3.12-asmflags.patch
similarity index 91%
rename from SOURCES/openblas-0.2.20-asmflags.patch
rename to SOURCES/openblas-0.3.12-asmflags.patch
index 239812f..1f767a7 100644
--- a/SOURCES/openblas-0.2.20-asmflags.patch
+++ b/SOURCES/openblas-0.3.12-asmflags.patch
@@ -1,8 +1,8 @@
 diff --git a/kernel/Makefile b/kernel/Makefile
-index a0a8fcd..df0669b 100644
+index e52781c6..c925837a 100644
 --- a/kernel/Makefile
 +++ b/kernel/Makefile
-@@ -73,22 +73,28 @@ endif
+@@ -151,25 +151,32 @@ endif
  all : libs
  
  scabs1.$(SUFFIX): $(KERNELDIR)/$(SCABS_KERNEL)
@@ -31,13 +31,18 @@ index a0a8fcd..df0669b 100644
 +	$(CC) -c $$CFLAGS -DF_INTERFACE $< -o $(@F)
  
  setparam$(TSUFFIX).$(SUFFIX): setparam$(TSUFFIX).c kernel$(TSUFFIX).h
+ ifeq ($(USE_GEMM3M), 1)
+-	$(CC) -c $(CFLAGS) -DUSE_GEMM3M $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) -c $$CFLAGS -DUSE_GEMM3M $< -o $@
+ else
 -	$(CC) -c $(CFLAGS) $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS $< -o $@
+ endif
  
  setparam$(TSUFFIX).c : setparam-ref.c
- 	sed 's/TS/$(TSUFFIX)/g' $< > $(@F)
-@@ -98,25 +104,32 @@ kernel$(TSUFFIX).h : $(KERNEL_INTERFACE)
+@@ -180,25 +187,32 @@ kernel$(TSUFFIX).h : $(KERNEL_INTERFACE)
  
  
  cpuid.$(SUFFIX): $(KERNELDIR)/cpuid.S
@@ -78,10 +83,10 @@ index a0a8fcd..df0669b 100644
  #ifdef DYNAMIC_ARCH
  clean ::
 diff --git a/kernel/Makefile.L1 b/kernel/Makefile.L1
-index a8f9cf0..5ebfc9c 100644
+index 7ad94118..3487a9e3 100644
 --- a/kernel/Makefile.L1
 +++ b/kernel/Makefile.L1
-@@ -495,319 +495,417 @@ XBLASOBJS	+= \
+@@ -545,217 +545,282 @@ endif
  
  
  $(KDIR)samax_k$(TSUFFIX).$(SUFFIX)  $(KDIR)samax_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SAMAXKERNEL)
@@ -279,7 +284,7 @@ index a8f9cf0..5ebfc9c 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -UCOMPLEX -DXDOUBLE -UUSE_ABS -DUSE_MIN $< -o $@
  
- 
+ ### ASUM ###
  $(KDIR)sasum_k$(TSUFFIX).$(SUFFIX)   $(KDIR)sasum_k$(TPSUFFIX).$(PSUFFIX)   : $(KERNELDIR)/$(SASUMKERNEL)
 -	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -310,6 +315,38 @@ index a8f9cf0..5ebfc9c 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -DXDOUBLE $< -o $@
  
+ ### SUM ###
+ $(KDIR)ssum_k$(TSUFFIX).$(SUFFIX)   $(KDIR)ssum_k$(TPSUFFIX).$(PSUFFIX)   : $(KERNELDIR)/$(SSUMKERNEL)
+-	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) -c $$CFLAGS -UCOMPLEX -UDOUBLE $< -o $@
+ 
+ $(KDIR)dsum_k$(TSUFFIX).$(SUFFIX)   $(KDIR)dsum_k$(TPSUFFIX).$(PSUFFIX)   : $(KERNELDIR)/$(DSUMKERNEL)
+-	$(CC) -c $(CFLAGS) -UCOMPLEX -DDOUBLE $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) -c $$CFLAGS -UCOMPLEX -DDOUBLE $< -o $@
+ 
+ $(KDIR)qsum_k$(TSUFFIX).$(SUFFIX)   $(KDIR)qsum_k$(TPSUFFIX).$(PSUFFIX)   : $(KERNELDIR)/$(QSUMKERNEL)
+-	$(CC) -c $(CFLAGS) -UCOMPLEX -DXDOUBLE $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) -c $$CFLAGS -UCOMPLEX -DXDOUBLE $< -o $@
+ 
+ $(KDIR)csum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)csum_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(CSUMKERNEL)
+-	$(CC) -c $(CFLAGS) -DCOMPLEX -UDOUBLE $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) -c $$CFLAGS -DCOMPLEX -UDOUBLE $< -o $@
+ 
+ $(KDIR)zsum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)zsum_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(ZSUMKERNEL)
+-	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) -c $$CFLAGS -DCOMPLEX -DDOUBLE $< -o $@
+ 
+ $(KDIR)xsum_k$(TSUFFIX).$(SUFFIX)  $(KDIR)xsum_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XSUMKERNEL)
+-	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) -c $$CFLAGS -DCOMPLEX -DXDOUBLE $< -o $@
+ 
+ ### AXPY ###
  $(KDIR)saxpy_k$(TSUFFIX).$(SUFFIX)  $(KDIR)saxpy_k$(TPSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(SAXPYKERNEL)
 -	$(CC) -c $(CFLAGS) -UCOMPLEX -UDOUBLE $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -395,6 +432,11 @@ index a8f9cf0..5ebfc9c 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)"; \
@@ -598,12 +640,12 @@ index a8f9cf0..5ebfc9c 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)"; \
@@ -613,7 +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)"; \
@@ -623,6 +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 $@
@@ -634,6 +679,8 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.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)"; \
@@ -673,6 +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 $@
@@ -713,6 +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 $@
@@ -754,6 +806,9 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.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)"; \
@@ -763,6 +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 $@
@@ -773,6 +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 $@
@@ -784,6 +844,8 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.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)"; \
@@ -793,6 +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 $@
@@ -803,6 +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 $@
@@ -814,21 +880,29 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.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)"; \
@@ -848,6 +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 $@
@@ -868,6 +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 $@
@@ -889,6 +967,8 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.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)"; \
@@ -908,6 +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 $@
@@ -928,7 +1011,7 @@ index 2aeb8f0..e7c49d2 100644
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -DDOUBLE -DLOWER -DHEMV -DHEMVREV $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -DDOUBLE -DLOWER -DHEMV -DHEMVREV $< -o $@
- 
+ endif
  $(KDIR)xhemv_U$(TSUFFIX).$(SUFFIX)  $(KDIR)xhemv_U$(TSUFFIX).$(PSUFFIX)  : $(KERNELDIR)/$(XHEMV_U_KERNEL)
 -	$(CC) -c $(CFLAGS) -DCOMPLEX -DXDOUBLE -ULOWER -DHEMV $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
@@ -949,12 +1032,19 @@ index 2aeb8f0..e7c49d2 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DCOMPLEX -DXDOUBLE -DLOWER -DHEMV -DHEMVREV $< -o $@
  
+ 
 diff --git a/kernel/Makefile.L3 b/kernel/Makefile.L3
-index 0664263..de28ab3 100644
+index 2ba593c2..110674b2 100644
 --- a/kernel/Makefile.L3
 +++ b/kernel/Makefile.L3
-@@ -403,2897 +403,3837 @@ XGEMMONCOPYOBJ_P = $(XGEMMONCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
- XGEMMOTCOPYOBJ_P = $(XGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
+@@ -526,119 +526,150 @@ XGEMMOTCOPYOBJ_P = $(XGEMMOTCOPYOBJ:.$(SUFFIX)=.$(PSUFFIX))
+ 
+ ifeq ($(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 -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ endif
  
  $(KDIR)sgemm_beta$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMM_BETA)
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
@@ -986,15 +1076,82 @@ index 0664263..de28ab3 100644
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DXDOUBLE -DCOMPLEX $< -o $@
  
+ 
+ ifeq ($(BUILD_BFLOAT16), 1)
+ 
+ $(KDIR)$(SBGEMMONCOPYOBJ) : $(KERNELDIR)/$(SBGEMMONCOPY)
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ 
+ $(KDIR)$(SBGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SBGEMMOTCOPY)
+ 
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o - > sbgemmotcopy.s
+-	m4 sbgemmotcopy.s > sbgemmotcopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX sbgemmotcopy_nomacros.s -o $@
+-	rm sbgemmotcopy.s sbgemmotcopy_nomacros.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o shgemmotcopy.s
++	m4 shgemmotcopy.s > shgemmotcopy_nomacros.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX shgemmotcopy_nomacros.s -o $@
++	rm shgemmotcopy.s shgemmotcopy_nomacros.s
+ else
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ endif
+ 
+ ifneq ($(SBGEMM_UNROLL_M), $(SBGEMM_UNROLL_N))
+ 
+ $(KDIR)$(SBGEMMINCOPYOBJ) : $(KERNELDIR)/$(SBGEMMINCOPY)
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ 
+ $(KDIR)$(SBGEMMITCOPYOBJ) : $(KERNELDIR)/$(SBGEMMITCOPY)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o - > sbgemmitcopy.s
+-	m4 sbgemmitcopy.s > sbgemmitcopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX sbgemmitcopy_nomacros.s -o $@
+-	rm sbgemmitcopy.s sbgemmitcopy_nomacros.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o shgemmitcopy.s
++	m4 shgemmitcopy.s > shgemmitcopy_nomacros.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX shgemmitcopy_nomacros.s -o $@
++	rm shgemmitcopy.s shgemmitcopy_nomacros.s
+ else
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ endif
+ 
+ endif
+ endif
+ 
  $(KDIR)$(SGEMMONCOPYOBJ) : $(KERNELDIR)/$(SGEMMONCOPY)
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
  
  $(KDIR)$(SGEMMOTCOPYOBJ) : $(KERNELDIR)/$(SGEMMOTCOPY)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX $< -o - > sgemmotcopy.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -UDOUBLE -UCOMPLEX $< -o - > sgemmotcopy.s
+ 	m4 sgemmotcopy.s > sgemmotcopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX sgemmotcopy_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX sgemmotcopy_nomacros.s -o $@
+ 	rm sgemmotcopy.s sgemmotcopy_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
+ endif	
+ 
  
  ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
  
@@ -1004,16 +1161,38 @@ index 0664263..de28ab3 100644
 +	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
  
  $(KDIR)$(SGEMMITCOPYOBJ) : $(KERNELDIR)/$(SGEMMITCOPY)
--	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX $< -o - > sgemmitcopy.s
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
-+	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
- 
++	$(CC) $$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)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX sgemmitcopy_nomacros.s -o $@
+ 	rm sgemmitcopy.s sgemmitcopy_nomacros.s
+ else
+-	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@ 
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@ 
+ endif
+ 	
  endif
  
  $(KDIR)$(DGEMMONCOPYOBJ) : $(KERNELDIR)/$(DGEMMONCOPY)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_ncopy.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o - > dgemm_ncopy.s
+ 	m4 dgemm_ncopy.s > dgemm_ncopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_ncopy_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX dgemm_ncopy_nomacros.s -o $@
+ 	rm dgemm_ncopy.s dgemm_ncopy_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX $< -o $@
+ endif
  
  $(KDIR)$(DGEMMOTCOPYOBJ) : $(KERNELDIR)/$(DGEMMOTCOPY)
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
@@ -1028,12 +1207,23 @@ index 0664263..de28ab3 100644
 +	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX $< -o $@
  
  $(KDIR)$(DGEMMITCOPYOBJ) : $(KERNELDIR)/$(DGEMMITCOPY)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_itcopy.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o - > dgemm_itcopy.s
+ 	m4 dgemm_itcopy.s > dgemm_itcopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_itcopy_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX dgemm_itcopy_nomacros.s -o $@
+ 	rm dgemm_itcopy.s dgemm_itcopy_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX $< -o $@
- 
  endif
  
+ endif
+@@ -646,65 +677,81 @@ endif
  ifdef EXPRECISION
  
  $(KDIR)$(QGEMMONCOPYOBJ) : $(KERNELDIR)/$(QGEMMONCOPY)
@@ -1080,9 +1270,20 @@ index 0664263..de28ab3 100644
 +	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
  
  $(KDIR)$(CGEMMITCOPYOBJ) : $(KERNELDIR)/$(CGEMMITCOPY)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -UDOUBLE -UCOMPLEX -S $< -o - > cgemm_itcopy.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -UDOUBLE -UCOMPLEX -E $< -o - > cgemm_itcopy.s
+ 	m4 cgemm_itcopy.s > cgemm_itcopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX cgemm_itcopy_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX cgemm_itcopy_nomacros.s -o $@
+ 	rm cgemm_itcopy.s cgemm_itcopy_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
+ endif
  
  endif
  
@@ -1104,12 +1305,23 @@ index 0664263..de28ab3 100644
 +	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX $< -o $@
  
  $(KDIR)$(ZGEMMITCOPYOBJ) : $(KERNELDIR)/$(ZGEMMITCOPY)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > zgemm_itcopy.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o - > zgemm_itcopy.s
+ 	m4 zgemm_itcopy.s > zgemm_itcopy_nomacros.s
+-	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX zgemm_itcopy_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX zgemm_itcopy_nomacros.s -o $@
+ 	rm zgemm_itcopy.s zgemm_itcopy_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX $< -o $@
- 
  endif
  
+ endif
+@@ -712,18 +759,22 @@ endif
  ifdef EXPRECISION
  
  $(KDIR)$(XGEMMONCOPYOBJ) : $(KERNELDIR)/$(XGEMMONCOPY)
@@ -1136,17 +1348,61 @@ index 0664263..de28ab3 100644
  
  endif
  
- endif
+@@ -731,12 +782,15 @@ endif
  
  $(KDIR)sgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL) $(SGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -UDOUBLE -UCOMPLEX  $< -o - > sgemm_kernel$(TSUFFIX).s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -UDOUBLE -UCOMPLEX  $< -o - > sgemm_kernel$(TSUFFIX).s
+ 	m4 sgemm_kernel$(TSUFFIX).s > sgemm_kernel$(TSUFFIX)_nomacros.s
+-	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX sgemm_kernel$(TSUFFIX)_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX sgemm_kernel$(TSUFFIX)_nomacros.s -o $@
+ 	rm sgemm_kernel$(TSUFFIX).s sgemm_kernel$(TSUFFIX)_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -UDOUBLE -UCOMPLEX $< -o $@
+ endif
+ 
+ ifdef USE_DIRECT_SGEMM
+@@ -750,1768 +804,2329 @@ ifeq ($(BUILD_BFLOAT16), 1)
+ 
+ $(KDIR)sbgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SBGEMMKERNEL) $(SBGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DBFLOAT16 -UDOUBLE -UCOMPLEX  $< -o - > sbgemm_kernel$(TSUFFIX).s
+-	m4 sbgemm_kernel$(TSUFFIX).s > sbgemm_kernel$(TSUFFIX)_nomacros.s
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX sbgemm_kernel$(TSUFFIX)_nomacros.s -o $@
+-	rm sbgemm_kernel$(TSUFFIX).s sbgemm_kernel$(TSUFFIX)_nomacros.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DHALF -UDOUBLE -UCOMPLEX  $< -o - > shgemm_kernel$(TSUFFIX).s
++	m4 shgemm_kernel$(TSUFFIX).s > shgemm_kernel$(TSUFFIX)_nomacros.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DHALF -UDOUBLE -UCOMPLEX shgemm_kernel$(TSUFFIX)_nomacros.s -o $@
++	rm shgemm_kernel$(TSUFFIX).s shgemm_kernel$(TSUFFIX)_nomacros.s
+ else
+-	$(CC) $(CFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ endif
+ endif
  
  $(KDIR)dgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMMKERNEL) $(DGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DDOUBLE -UCOMPLEX $< -o - > dgemm_kernel$(TSUFFIX).s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DDOUBLE -UCOMPLEX $< -o - > dgemm_kernel$(TSUFFIX).s
+ 	m4 dgemm_kernel$(TSUFFIX).s > dgemm_kernel$(TSUFFIX)_nomacros.s
+-	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX dgemm_kernel$(TSUFFIX)_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX dgemm_kernel$(TSUFFIX)_nomacros.s -o $@
+ 	rm dgemm_kernel$(TSUFFIX).s dgemm_kernel$(TSUFFIX)_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DDOUBLE -UCOMPLEX $< -o $@
+ endif
  
  $(KDIR)qgemm_kernel$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(QGEMMKERNEL) $(QGEMMDEPEND)
 -	$(CC) $(CFLAGS) -c -DXDOUBLE -UCOMPLEX $< -o $@
@@ -1154,44 +1410,132 @@ index 0664263..de28ab3 100644
 +	$(CC) $$CFLAGS -c -DXDOUBLE -UCOMPLEX $< -o $@
  
  $(KDIR)cgemm_kernel_n$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DNN $< -o - > cgemm_kernel_n.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DNN $< -o - > cgemm_kernel_n.s
+ 	m4 cgemm_kernel_n.s > cgemm_kernel_n_nomacros.s
+-	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNN cgemm_kernel_n_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DNN cgemm_kernel_n_nomacros.s -o $@
+ 	rm cgemm_kernel_n.s cgemm_kernel_n_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DNN $< -o $@
+ endif
  
  $(KDIR)cgemm_kernel_l$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DCN $< -o - > cgemm_kernel_l.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DCN $< -o - > cgemm_kernel_l.s
+ 	m4 cgemm_kernel_l.s > cgemm_kernel_l_nomacros.s
+-	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DCN cgemm_kernel_l_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DCN cgemm_kernel_l_nomacros.s -o $@
+ 	rm cgemm_kernel_l.s cgemm_kernel_l_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DCN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DCN $< -o $@
+ endif
  
  $(KDIR)cgemm_kernel_r$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DNC  $< -o - > cgemm_kernel_r.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DNC  $< -o - > cgemm_kernel_r.s
+ 	m4 cgemm_kernel_r.s > cgemm_kernel_r_nomacros.s
+-	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNC cgemm_kernel_r_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DNC cgemm_kernel_r_nomacros.s -o $@
+ 	rm cgemm_kernel_r.s cgemm_kernel_r_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNC $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DNC $< -o $@
+ endif
  
  $(KDIR)cgemm_kernel_b$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -UDOUBLE -DCOMPLEX -DCC $< -o - > cgemm_kernel_b.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -UDOUBLE -DCOMPLEX -DCC $< -o - > cgemm_kernel_b.s
+ 	m4 cgemm_kernel_b.s > cgemm_kernel_b_nomacros.s
+-	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DCC cgemm_kernel_b_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DCC cgemm_kernel_b_nomacros.s -o $@
+ 	rm cgemm_kernel_b.s cgemm_kernel_b_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DCC $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DCC $< -o $@
+ endif
  
  $(KDIR)zgemm_kernel_n$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DNN $< -o - > zgemm_kernel_n.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DNN $< -o - > zgemm_kernel_n.s
+ 	m4 zgemm_kernel_n.s > zgemm_kernel_n_nomacros.s
+-	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DNN zgemm_kernel_n_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -DNN zgemm_kernel_n_nomacros.s -o $@
+ 	rm zgemm_kernel_n.s zgemm_kernel_n_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -DNN $< -o $@
+ endif
  
  $(KDIR)zgemm_kernel_l$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DCN $< -o - > zgemm_kernel_l.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DCN $< -o - > zgemm_kernel_l.s
+ 	m4 zgemm_kernel_l.s > zgemm_kernel_l_nomacros.s
+-	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DCN zgemm_kernel_l_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -DCN zgemm_kernel_l_nomacros.s -o $@
+ 	rm zgemm_kernel_l.s zgemm_kernel_l_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DCN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -DCN $< -o $@
+ endif
  
  $(KDIR)zgemm_kernel_r$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DNC $< -o - > zgemm_kernel_r.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DNC $< -o - > zgemm_kernel_r.s
+ 	m4 zgemm_kernel_r.s > zgemm_kernel_r_nomacros.s
+-	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DNC zgemm_kernel_r_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -DNC zgemm_kernel_r_nomacros.s -o $@
+ 	rm zgemm_kernel_r.s zgemm_kernel_r_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DNC $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -DNC $< -o $@
+ endif
  
  $(KDIR)zgemm_kernel_b$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZGEMMKERNEL) $(ZGEMMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DDOUBLE -DCOMPLEX -DCC $< -o - > zgemm_kernel_b.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DDOUBLE -DCOMPLEX -DCC $< -o - > zgemm_kernel_b.s
+ 	m4 zgemm_kernel_b.s > zgemm_kernel_b_nomacros.s
+-	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DCC zgemm_kernel_b_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -DCC zgemm_kernel_b_nomacros.s -o $@
+ 	rm zgemm_kernel_b.s zgemm_kernel_b_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DDOUBLE -DCOMPLEX -DCC $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DDOUBLE -DCOMPLEX -DCC $< -o $@
+ endif
  
  $(KDIR)xgemm_kernel_n$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(XGEMMKERNEL) $(XGEMMDEPEND)
 -	$(CC) $(CFLAGS) -c -DXDOUBLE -DCOMPLEX -DNN $< -o $@
@@ -1216,44 +1560,132 @@ index 0664263..de28ab3 100644
  
  ifdef USE_TRMM
  $(KDIR)strmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > strmmkernel_ln.s	
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > strmmkernel_ln.s	
+ 	m4 strmmkernel_ln.s > strmmkernel_ln_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA strmmkernel_ln_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA strmmkernel_ln_nomacros.s -o $@
+ 	rm strmmkernel_ln.s strmmkernel_ln_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o $@
+ endif
  
  $(KDIR)strmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > strmmkernel_lt.s	
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > strmmkernel_lt.s	
+ 	m4 strmmkernel_lt.s > strmmkernel_lt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA strmmkernel_lt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA strmmkernel_lt_nomacros.s -o $@
+ 	rm strmmkernel_lt.s strmmkernel_lt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o $@
+ endif
  
  $(KDIR)strmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > strmmkernel_rn.s	
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > strmmkernel_rn.s	
+ 	m4 strmmkernel_rn.s > strmmkernel_rn_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA strmmkernel_rn_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA strmmkernel_rn_nomacros.s -o $@
+ 	rm strmmkernel_rn.s strmmkernel_rn_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o $@
+ endif
  
  $(KDIR)strmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(STRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
+ 	m4 strmm_kernel_rt.s > strmm_kernel_rt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
+ 	rm strmm_kernel_rt.s strmm_kernel_rt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
+ endif
  
  $(KDIR)dtrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > dtrmm_kernel_ln.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o - > dtrmm_kernel_ln.s
+ 	m4 dtrmm_kernel_ln.s > dtrmm_kernel_ln_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA dtrmm_kernel_ln_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA dtrmm_kernel_ln_nomacros.s -o $@
+ 	rm dtrmm_kernel_ln.s dtrmm_kernel_ln_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o $@
+ endif
  
  $(KDIR)dtrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > dtrmm_kernel_lt.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o - > dtrmm_kernel_lt.s
+ 	m4 dtrmm_kernel_lt.s > dtrmm_kernel_lt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA dtrmm_kernel_lt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA dtrmm_kernel_lt_nomacros.s -o $@
+ 	rm dtrmm_kernel_lt.s dtrmm_kernel_lt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -DTRANSA $< -o $@
+ endif
  
  $(KDIR)dtrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > dtrmm_kernel_rn.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o - > dtrmm_kernel_rn.s
+ 	m4 dtrmm_kernel_rn.s > dtrmm_kernel_rn_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA dtrmm_kernel_rn_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA dtrmm_kernel_rn_nomacros.s -o $@
+ 	rm dtrmm_kernel_rn.s dtrmm_kernel_rn_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o $@
+ endif
  
  $(KDIR)dtrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > dtrmm_kernel_rt.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > dtrmm_kernel_rt.s
+ 	m4 dtrmm_kernel_rt.s > dtrmm_kernel_rt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA dtrmm_kernel_rt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA dtrmm_kernel_rt_nomacros.s -o $@
+ 	rm dtrmm_kernel_rt.s dtrmm_kernel_rt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
+ endif
  
  $(KDIR)qtrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(QGEMMKERNEL)
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DXDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o $@
@@ -1276,84 +1708,261 @@ index 0664263..de28ab3 100644
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DXDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
  
  $(KDIR)ctrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN  $< -o - > ctrmm_kernel_ln.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN  $< -o - > ctrmm_kernel_ln.s
+ 	m4 ctrmm_kernel_ln.s > ctrmm_kernel_ln_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN ctrmm_kernel_ln_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN ctrmm_kernel_ln_nomacros.s -o $@
+ 	rm ctrmm_kernel_ln.s ctrmm_kernel_ln_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o $@
+ endif
  
  $(KDIR)ctrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_lt.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_lt.s
+ 	m4 ctrmm_kernel_lt.s > ctrmm_kernel_lt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN ctrmm_kernel_lt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN ctrmm_kernel_lt_nomacros.s -o $@
+ 	rm ctrmm_kernel_lt.s ctrmm_kernel_lt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o $@
+ endif
  
  $(KDIR)ctrmm_kernel_LR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lr.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lr.s
+ 	m4 ctrmm_kernel_lr.s > ctrmm_kernel_lr_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN  ctrmm_kernel_lr_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN  ctrmm_kernel_lr_nomacros.s -o $@
+ 	rm ctrmm_kernel_lr.s ctrmm_kernel_lr_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o $@
+ endif
  
  $(KDIR)ctrmm_kernel_LC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lc.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ctrmm_kernel_lc.s
+ 	m4 ctrmm_kernel_lc.s > ctrmm_kernel_lc_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN ctrmm_kernel_lc_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN ctrmm_kernel_lc_nomacros.s -o $@
+ 	rm ctrmm_kernel_lc_nomacros.s ctrmm_kernel_lc.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o $@
+ endif
  
  $(KDIR)ctrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rn.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rn.s
+ 	m4 ctrmm_kernel_rn.s > ctrmm_kernel_rn_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN ctrmm_kernel_rn_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN ctrmm_kernel_rn_nomacros.s -o $@
+ 	rm ctrmm_kernel_rn.s ctrmm_kernel_rn_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o $@
+ endif
  
  $(KDIR)ctrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rt.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ctrmm_kernel_rt.s
+ 	m4 ctrmm_kernel_rt.s > ctrmm_kernel_rt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN ctrmm_kernel_rt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN ctrmm_kernel_rt_nomacros.s -o $@
+ 	rm ctrmm_kernel_rt.s ctrmm_kernel_rt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o $@
+ endif
  
  $(KDIR)ctrmm_kernel_RR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_rr.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_rr.s
+ 	m4 ctrmm_kernel_rr.s > ctrmm_kernel_rr_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC ctrmm_kernel_rr_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC ctrmm_kernel_rr_nomacros.s -o $@
+ 	rm ctrmm_kernel_rr.s ctrmm_kernel_rr_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o $@
+ endif
  
  $(KDIR)ctrmm_kernel_RC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(CTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_RC.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ctrmm_kernel_RC.s
+ 	m4 ctrmm_kernel_RC.s > ctrmm_kernel_RC_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC ctrmm_kernel_RC_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC ctrmm_kernel_RC_nomacros.s -o $@
+ 	rm ctrmm_kernel_RC.s ctrmm_kernel_RC_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o $@
+ endif
  
  $(KDIR)ztrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_ln.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_ln.s
+ 	m4 ztrmm_kernel_ln.s > ztrmm_kernel_ln_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN ztrmm_kernel_ln_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN ztrmm_kernel_ln_nomacros.s -o $@
+ 	rm ztrmm_kernel_ln.s ztrmm_kernel_ln_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -UCONJ -DNN $< -o $@
+ endif
  
  $(KDIR)ztrmm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_lt.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_lt.s
+ 	m4 ztrmm_kernel_lt.s > ztrmm_kernel_lt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN ztrmm_kernel_lt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN ztrmm_kernel_lt_nomacros.s -o $@
+ 	rm ztrmm_kernel_lt.s ztrmm_kernel_lt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -UCONJ -DNN $< -o $@
+ endif
  
  $(KDIR)ztrmm_kernel_LR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lr.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lr.s
+ 	m4 ztrmm_kernel_lr.s > ztrmm_kernel_lr_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN ztrmm_kernel_lr_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN ztrmm_kernel_lr_nomacros.s -o $@
+ 	rm ztrmm_kernel_lr.s ztrmm_kernel_lr_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -UTRANSA -DCONJ -DCN $< -o $@
+ endif
  
  $(KDIR)ztrmm_kernel_LC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lc.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o - > ztrmm_kernel_lc.s
+ 	m4 ztrmm_kernel_lc.s >ztrmm_kernel_lc_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN ztrmm_kernel_lc_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN ztrmm_kernel_lc_nomacros.s -o $@
+ 	rm ztrmm_kernel_lc.s ztrmm_kernel_lc_nomacros.s 
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -DLEFT -DTRANSA -DCONJ -DCN $< -o $@
+ endif
  
  $(KDIR)ztrmm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rn.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rn.s
+ 	m4 ztrmm_kernel_rn.s > ztrmm_kernel_rn_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN ztrmm_kernel_rn_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN ztrmm_kernel_rn_nomacros.s -o $@
+ 	rm ztrmm_kernel_rn.s ztrmm_kernel_rn_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -UCONJ -DNN $< -o $@
+ endif
  
  $(KDIR)ztrmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rt.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o - > ztrmm_kernel_rt.s
+ 	m4 ztrmm_kernel_rt.s > ztrmm_kernel_rt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN ztrmm_kernel_rt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN ztrmm_kernel_rt_nomacros.s -o $@
+ 	rm ztrmm_kernel_rt.s ztrmm_kernel_rt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -UCONJ -DNN $< -o $@
+ endif
  
  $(KDIR)ztrmm_kernel_RR$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rr.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rr.s
+ 	m4 ztrmm_kernel_rr.s > ztrmm_kernel_rr_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC ztrmm_kernel_rr_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC ztrmm_kernel_rr_nomacros.s -o $@
+ 	rm ztrmm_kernel_rr.s ztrmm_kernel_rr_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -UTRANSA -DCONJ -DNC $< -o $@
+ endif
  
  $(KDIR)ztrmm_kernel_RC$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(ZTRMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rc.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o - > ztrmm_kernel_rc.s
+ 	m4 ztrmm_kernel_rc.s > ztrmm_kernel_rc_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC ztrmm_kernel_rc_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC ztrmm_kernel_rc_nomacros.s -o $@
+ 	rm ztrmm_kernel_rc.s ztrmm_kernel_rc_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -DDOUBLE -DCOMPLEX -ULEFT -DTRANSA -DCONJ -DNC $< -o $@
+ endif
+ 
  else
  $(KDIR)strmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o $@
@@ -1371,9 +1980,20 @@ index 0664263..de28ab3 100644
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o $@
  
  $(KDIR)strmm_kernel_RT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
+ 	m4 strmm_kernel_rt.s > strmm_kernel_rt_nomacros.s
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
+ 	rm strmm_kernel_rt.s strmm_kernel_rt_nomacros.s
+ else
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
+ endif
  
  $(KDIR)dtrmm_kernel_LN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DGEMMKERNEL)
 -	$(CC) $(CFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o $@
@@ -1580,9 +2200,20 @@ index 0664263..de28ab3 100644
 +	$(CC) -c $$CFLAGS -DTRSMKERNEL -UCOMPLEX -DDOUBLE -DUPPER -DLN -UCONJ $< -o $@
  
  $(KDIR)dtrsm_kernel_LT$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRSMKERNEL_LT) $(DTRSMDEPEND)
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o - > dtrsm_kernel_lt.s
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o - > dtrsm_kernel_lt.s
+ 	m4 dtrsm_kernel_lt.s > dtrsm_kernel_lt_nomacros.s
+-	$(CC) -c $(CFLAGS) -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ dtrsm_kernel_lt_nomacros.s -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) -c $$CFLAGS -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ dtrsm_kernel_lt_nomacros.s -o $@
+ 	rm dtrsm_kernel_lt.s dtrsm_kernel_lt_nomacros.s
+ else
 -	$(CC) -c $(CFLAGS) -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o $@
 +	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
 +	$(CC) -c $$CFLAGS -DTRSMKERNEL -UCOMPLEX -DDOUBLE -UUPPER -DLT -UCONJ $< -o $@
+ endif
  
  $(KDIR)dtrsm_kernel_RN$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(DTRSMKERNEL_RN) $(DTRSMDEPEND)
 -	$(CC) -c $(CFLAGS) -DTRSMKERNEL -UCOMPLEX -DDOUBLE -DUPPER -DRN -UCONJ $< -o $@
@@ -3421,6 +4052,13 @@ index 0664263..de28ab3 100644
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
 +	$(CC) $$PFLAGS -c -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) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ endif
+ 
  $(KDIR)dgemm_beta$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(DGEMM_BETA)
 -	$(CC) $(PFLAGS) -c -DDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
@@ -3446,6 +4084,32 @@ index 0664263..de28ab3 100644
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
 +	$(CC) $$PFLAGS -c -DXDOUBLE -DCOMPLEX $< -o $@
  
+ 
+ ifeq ($(BUILD_BFLOAT16), 1)
+ $(SBGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SBGEMMONCOPY)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
++	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
++	$(CC) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ 
+ $(SBGEMMOTCOPYOBJ_P) : $(KERNELDIR)/$(SBGEMMOTCOPY)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
++	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
++	$(CC) $$PFLAGS -c -DBFLOAT16 -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) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ 
+ $(SBGEMMITCOPYOBJ_P) : $(KERNELDIR)/$(SBGEMMITCOPY)
+-	$(CC) $(PFLAGS) -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
++	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
++	$(CC) $$PFLAGS -c -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
+ 
+ endif
+ endif
+ 
  $(SGEMMONCOPYOBJ_P) : $(KERNELDIR)/$(SGEMMONCOPY)
 -	$(CC) $(PFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
@@ -3596,6 +4260,13 @@ index 0664263..de28ab3 100644
  
  endif
  
+@@ -2520,1301 +3135,1732 @@ endif
+ 
+ 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 -DBFLOAT16 -UDOUBLE -UCOMPLEX $< -o $@
  endif
  
  $(KDIR)sgemm_kernel$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL) $(SGEMMDEPEND)
@@ -3624,9 +4295,20 @@ index 0664263..de28ab3 100644
 +	$(CC) $$PFLAGS -c -UDOUBLE -DCOMPLEX -DCN $< -o $@
  
  $(KDIR)cgemm_kernel_r$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
--	$(CC) $(PFLAGS) -c -UDOUBLE -DCOMPLEX -DNC $< -o $@
+ ifeq ($(OS), AIX)
+-	$(CC) $(PFLAGS) -S -UDOUBLE -DCOMPLEX -DNC $< -o - > cgemm_kernel_r.s
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -c -UDOUBLE -DCOMPLEX -DNC $< -o $@
++	$(CC) $$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)"; \
++	$(CC) $$PFLAGS -c -UDOUBLE -DCOMPLEX -DNC cgemm_kernel_r_nomacros.s -o $@
+ 	rm cgemm_kernel_r.s cgemm_kernel_r_nomacros.s 
+ else
+-	$(CC) $(CFLAGS) -c -UDOUBLE -DCOMPLEX -DNC $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -UDOUBLE -DCOMPLEX -DNC $< -o $@
+ endif
  
  $(KDIR)cgemm_kernel_b$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(CGEMMKERNEL) $(CGEMMDEPEND)
 -	$(CC) $(PFLAGS) -c -UDOUBLE -DCOMPLEX -DCC $< -o $@
@@ -3689,9 +4371,20 @@ index 0664263..de28ab3 100644
 +	$(CC) $$PFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -UTRANSA $< -o $@
  
  $(KDIR)strmm_kernel_RT$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(SGEMMKERNEL)
--	$(CC) $(PFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
+ ifeq ($(OS), AIX)
+-	$(CC) $(CFLAGS) -S -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -E -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o - > strmm_kernel_rt.s	
+ 	m4 strmmkernel_rn.s > strmm_kernel_rt_nomacros.s
+-	$(CC) $(PFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
 +	[[ "$<" == *.S ]] && PFLAGS="$(PFLAGS) $(ASMFLAGS)" || PFLAGS="$(PFLAGS)"; \
-+	$(CC) $$PFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
++	$(CC) $$PFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA strmm_kernel_rt_nomacros.s -o $@
+ 	rm strmm_kernel_rt.s strmm_kernel_rt_nomacros.s
+ else
+-	$(CC) $(CFLAGS) -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
++	[[ "$<" == *.S ]] && CFLAGS="$(CFLAGS) $(ASMFLAGS)" || CFLAGS="$(CFLAGS)"; \
++	$(CC) $$CFLAGS -c -DTRMMKERNEL -UDOUBLE -UCOMPLEX -ULEFT -DTRANSA $< -o $@
+ endif
  
  $(KDIR)dtrmm_kernel_LN$(TSUFFIX).$(PSUFFIX) : $(KERNELDIR)/$(DGEMMKERNEL)
 -	$(CC) $(PFLAGS) -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA $< -o $@
@@ -5731,7 +6424,7 @@ index 0664263..de28ab3 100644
  
  
  #####  BLAS extensions ######
-@@ -3303,112 +4243,128 @@ 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)
@@ -5876,7 +6569,7 @@ index 0664263..de28ab3 100644
  
  
  ifndef COMATCOPY_CN
-@@ -3416,112 +4372,128 @@ 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)
@@ -6021,7 +6714,7 @@ index 0664263..de28ab3 100644
  
  
  
-@@ -3530,112 +4502,128 @@ 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)
@@ -6166,7 +6859,7 @@ index 0664263..de28ab3 100644
  
  
  ifndef SGEADD_K
-@@ -3643,26 +4631,30 @@ SGEADD_K = ../generic/geadd.c
+@@ -4164,26 +5258,30 @@ SGEADD_K = ../generic/geadd.c
  endif
  
  $(KDIR)sgeadd_k$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEADD_K)
@@ -6202,7 +6895,7 @@ index 0664263..de28ab3 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-0.2.5-libname.patch b/SOURCES/openblas-0.3.12-libname.patch
similarity index 74%
rename from SOURCES/openblas-0.2.5-libname.patch
rename to SOURCES/openblas-0.3.12-libname.patch
index e30ab8b..d75de48 100644
--- a/SOURCES/openblas-0.2.5-libname.patch
+++ b/SOURCES/openblas-0.3.12-libname.patch
@@ -1,7 +1,8 @@
-diff -up OpenBLAS-0.2.5/Makefile.system.orig OpenBLAS-0.2.5/Makefile.system
---- OpenBLAS-0.2.5/Makefile.system.orig	2012-11-27 01:24:53.000000000 +0200
-+++ OpenBLAS-0.2.5/Makefile.system	2012-12-24 16:13:57.316689688 +0200
-@@ -758,16 +758,16 @@ ifndef SMP
+diff --git a/Makefile.system b/Makefile.system
+index 30d8f4cc..803219a8 100644
+--- a/Makefile.system
++++ b/Makefile.system
+@@ -1425,16 +1425,16 @@ ifndef SMP
  LIBNAME		= $(LIBPREFIX)_$(LIBCORE)$(REVISION).$(LIBSUFFIX)
  LIBNAME_P	= $(LIBPREFIX)_$(LIBCORE)$(REVISION)_p.$(LIBSUFFIX)
  else
diff --git a/SOURCES/openblas-0.3.12-noopt.patch b/SOURCES/openblas-0.3.12-noopt.patch
new file mode 100644
index 0000000..499571b
--- /dev/null
+++ b/SOURCES/openblas-0.3.12-noopt.patch
@@ -0,0 +1,47 @@
+diff --git a/lapack-netlib/INSTALL/Makefile b/lapack-netlib/INSTALL/Makefile
+index 1007c1bc..348b8d60 100644
+--- a/lapack-netlib/INSTALL/Makefile
++++ b/lapack-netlib/INSTALL/Makefile
+@@ -46,5 +46,5 @@ cleanexe:
+ cleantest:
+ 	rm -f core
+ 
+-slamch.o: slamch.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
+-dlamch.o: dlamch.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
++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 83baac87..bc42e82d 100644
+--- a/lapack-netlib/SRC/Makefile
++++ b/lapack-netlib/SRC/Makefile
+@@ -643,9 +643,9 @@ cleanobj:
+ cleanlib:
+ 	rm -f $(LAPACKLIB)
+ 
+-slaruv.o: slaruv.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
+-dlaruv.o: dlaruv.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
+-sla_wwaddw.o: sla_wwaddw.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
+-dla_wwaddw.o: dla_wwaddw.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
+-cla_wwaddw.o: cla_wwaddw.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
+-zla_wwaddw.o: zla_wwaddw.f ; $(FC) $(FFLAGS_NOOPT) -c -o $@ $<
++slaruv.o: slaruv.f ; $(FC) $(FFLAGS) -c -o $@ $<
++dlaruv.o: dlaruv.f ; $(FC) $(FFLAGS) -c -o $@ $<
++sla_wwaddw.o: sla_wwaddw.f ; $(FC) $(FFLAGS) -c -o $@ $<
++dla_wwaddw.o: dla_wwaddw.f ; $(FC) $(FFLAGS) -c -o $@ $<
++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 e21ebd6c..8582e944 100644
+--- a/lapack-netlib/TESTING/MATGEN/Makefile
++++ b/lapack-netlib/TESTING/MATGEN/Makefile
+@@ -110,8 +110,8 @@ cleanlib:
+ 	rm -f $(TMGLIB)
+ 
+ ifeq ($(filter $(BUILD_SINGLE) $(BUILD_COMPLEX),1),)
+-slaran.o: slaran.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-0.2.20-power-optimize.patch b/SOURCES/openblas-0.3.12-power-optimize.patch
similarity index 67%
rename from SOURCES/openblas-0.2.20-power-optimize.patch
rename to SOURCES/openblas-0.3.12-power-optimize.patch
index 5d14162..5d82b6f 100644
--- a/SOURCES/openblas-0.2.20-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 3e10748..9d2b8c0 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 3e10748..9d2b8c0 100644
  
 -#pragma GCC optimize "O1"
 -
- #if defined(POWER8)
+ #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 d2910ff..18f5362 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 d2910ff..18f5362 100644
  
 -#pragma GCC optimize "O1"
 -
- #if defined(POWER8)
+ #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 14d677f..79f7271 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 14d677f..79f7271 100644
  
 -#pragma GCC optimize "O1"
 -
- #if defined(POWER8)
+ #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-0.2.15-system_lapack.patch b/SOURCES/openblas-0.3.12-system-lapack.patch
similarity index 52%
rename from SOURCES/openblas-0.2.15-system_lapack.patch
rename to SOURCES/openblas-0.3.12-system-lapack.patch
index 4b843a9..9797677 100644
--- a/SOURCES/openblas-0.2.15-system_lapack.patch
+++ b/SOURCES/openblas-0.3.12-system-lapack.patch
@@ -1,20 +1,27 @@
-diff -up OpenBLAS-0.2.15/Makefile.system_lapack OpenBLAS-0.2.15/Makefile
---- OpenBLAS-0.2.15/Makefile.system_lapack	2015-10-27 13:44:50.000000000 -0700
-+++ OpenBLAS-0.2.15/Makefile	2015-10-28 09:14:56.696685503 -0700
-@@ -16,11 +16,7 @@ BLASDIRS += reference
+diff --git a/Makefile b/Makefile
+index e113026..ae8f7de 100644
+--- a/Makefile
++++ b/Makefile
+@@ -12,9 +12,6 @@ BLASDIRS += reference
  endif
  
  SUBDIRS	= $(BLASDIRS)
 -ifneq ($(NO_LAPACK), 1)
 -SUBDIRS	+= lapack
 -endif
--
+ 
+ RELA =
+ ifeq ($(BUILD_RELAPACK), 1)
+@@ -32,8 +29,6 @@ export NOFORTRAN
+ export NO_LAPACK
+ endif
+ 
 -LAPACK_NOOPT := $(filter-out -O0 -O1 -O2 -O3 -Ofast,$(LAPACK_FFLAGS))
-+SUBDIRS += lapack
+-
+ SUBDIRS_ALL = $(SUBDIRS) test ctest utest exports benchmark ../laswp ../bench cpp_thread_test
  
- SUBDIRS_ALL = $(SUBDIRS) test ctest utest exports benchmark ../laswp ../bench
- 
-@@ -211,57 +207,8 @@ hpl_p :
+ .PHONY : all libs netlib $(RELA) test ctest shared install
+@@ -235,76 +230,8 @@ hpl_p :
  	fi; \
  	done
  
@@ -23,7 +30,7 @@ diff -up OpenBLAS-0.2.15/Makefile.system_lapack OpenBLAS-0.2.15/Makefile
 -
 -else
 -netlib : lapack_prebuild
--ifndef NOFORTRAN
+-ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
 -	@$(MAKE) -C $(NETLIB_LAPACK_DIR) lapacklib
 -	@$(MAKE) -C $(NETLIB_LAPACK_DIR) tmglib
 -endif
@@ -32,34 +39,49 @@ diff -up OpenBLAS-0.2.15/Makefile.system_lapack OpenBLAS-0.2.15/Makefile
 -endif
 -endif
 -
+-ifeq ($(NO_LAPACK), 1)
+-re_lapack :
+-
+-else
+-re_lapack :
+-	@$(MAKE) -C relapack
+-endif
+-
 -prof_lapack : lapack_prebuild
 -	@$(MAKE) -C $(NETLIB_LAPACK_DIR) lapack_prof
 -
 -lapack_prebuild :
--ifndef NOFORTRAN
--	-@echo "FORTRAN     = $(FC)" > $(NETLIB_LAPACK_DIR)/make.inc
--	-@echo "OPTS        = $(LAPACK_FFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
+-	-@echo "FC          = $(FC)" > $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "FFLAGS      = $(LAPACK_FFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "FFLAGS_DRV  = $(LAPACK_FFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "POPTS       = $(LAPACK_FPFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
--	-@echo "NOOPT       = -O0 $(LAPACK_NOOPT)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "FFLAGS_NOOPT       = -O0 $(LAPACK_NOOPT)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "PNOOPT      = $(LAPACK_FPFLAGS) -O0" >> $(NETLIB_LAPACK_DIR)/make.inc
--	-@echo "LOADOPTS    = $(FFLAGS) $(EXTRALIB)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "LDFLAGS     = $(FFLAGS) $(EXTRALIB)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "CC          = $(CC)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "override CFLAGS      = $(LAPACK_CFLAGS)" >> $(NETLIB_LAPACK_DIR)/make.inc
--	-@echo "ARCH        = $(AR)" >> $(NETLIB_LAPACK_DIR)/make.inc
--	-@echo "ARCHFLAGS   = -ru" >> $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "AR          = $(AR)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "ARFLAGS     = $(ARFLAGS) -ru" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "RANLIB      = $(RANLIB)" >> $(NETLIB_LAPACK_DIR)/make.inc
--	-@echo "LAPACKLIB   = ../$(LIBNAME)" >> $(NETLIB_LAPACK_DIR)/make.inc
--	-@echo "TMGLIB      = ../$(LIBNAME)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "LAPACKLIB   = ../../$(LIBNAME)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "TMGLIB      = ../../../$(LIBNAME)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "BLASLIB     = ../../../$(LIBNAME)" >> $(NETLIB_LAPACK_DIR)/make.inc
--	-@echo "LAPACKELIB  = ../$(LIBNAME)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-	-@echo "LAPACKELIB  = ../../../$(LIBNAME)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "LAPACKLIB_P = ../$(LIBNAME_P)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "SUFFIX      = $(SUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "PSUFFIX     = $(PSUFFIX)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "CEXTRALIB   = $(EXTRALIB)" >> $(NETLIB_LAPACK_DIR)/make.inc
--ifeq ($(FC), gfortran)
+-ifeq ($(F_COMPILER), GFORTRAN)
 -	-@echo "TIMER       = INT_ETIME" >> $(NETLIB_LAPACK_DIR)/make.inc
 -ifdef SMP
+-ifeq ($(OSNAME), WINNT)
+-	-@echo "LOADER      = $(FC)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-else ifeq ($(OSNAME), Haiku)
+-	-@echo "LOADER      = $(FC)" >> $(NETLIB_LAPACK_DIR)/make.inc
+-else
 -	-@echo "LOADER      = $(FC) -pthread" >> $(NETLIB_LAPACK_DIR)/make.inc
+-endif
 -else
 -	-@echo "LOADER      = $(FC)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -endif
@@ -67,17 +89,22 @@ diff -up OpenBLAS-0.2.15/Makefile.system_lapack OpenBLAS-0.2.15/Makefile
 -	-@echo "TIMER       = NONE" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@echo "LOADER      = $(FC)" >> $(NETLIB_LAPACK_DIR)/make.inc
 -endif
+-ifeq ($(BUILD_LAPACK_DEPRECATED), 1)
+-	-@echo "BUILD_DEPRECATED      = 1" >> $(NETLIB_LAPACK_DIR)/make.inc
+-endif
+-	-@echo "LAPACKE_WITH_TMG      = 1" >> $(NETLIB_LAPACK_DIR)/make.inc
 -	-@cat  make.inc >> $(NETLIB_LAPACK_DIR)/make.inc
 -endif
 +	@$(MAKE) -C $(NETLIB_LAPACK_DIR)
  
  large.tgz :
- ifndef NOFORTRAN
-diff -up OpenBLAS-0.2.15/Makefile.system.system_lapack OpenBLAS-0.2.15/Makefile.system
---- OpenBLAS-0.2.15/Makefile.system.system_lapack	2015-10-27 13:44:50.000000000 -0700
-+++ OpenBLAS-0.2.15/Makefile.system	2015-10-28 09:14:39.994350500 -0700
-@@ -9,7 +9,7 @@ ifndef TOPDIR
- TOPDIR = .
+ ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
+diff --git a/Makefile.system b/Makefile.system
+index 8d78b42..47d8eec 100644
+--- a/Makefile.system
++++ b/Makefile.system
+@@ -31,7 +31,7 @@ else ifeq ($(ARCH), zarch)
+ override ARCH=zarch
  endif
  
 -NETLIB_LAPACK_DIR = $(TOPDIR)/lapack-netlib
diff --git a/SOURCES/openblas-0.3.3-tests.patch b/SOURCES/openblas-0.3.12-tests.patch
similarity index 66%
rename from SOURCES/openblas-0.3.3-tests.patch
rename to SOURCES/openblas-0.3.12-tests.patch
index ff3bd37..0bd9864 100644
--- a/SOURCES/openblas-0.3.3-tests.patch
+++ b/SOURCES/openblas-0.3.12-tests.patch
@@ -1,19 +1,22 @@
 diff --git a/Makefile b/Makefile
-index d99521b..01bba2f 100644
+index a9af62a2..2f0f11c0 100644
 --- a/Makefile
 +++ b/Makefile
-@@ -122,11 +122,11 @@ tests :
+@@ -138,13 +138,13 @@ tests :
  ifeq ($(NOFORTRAN), $(filter 0,$(NOFORTRAN)))
  	touch $(LIBNAME)
  ifndef NO_FBLAS
 -	$(MAKE) -C test all
--	$(MAKE) -C utest all
 +	$(MAKE) -C test FC="$(FC)" CC="$(CC)" COMMON_OPT="$(COMMON_OPT)" FCOMMON_OPT="$(FCOMMON_OPT)" all
-+	$(MAKE) -C utest FC="$(FC)" CC="$(CC)" COMMON_OPT="$(COMMON_OPT)" FCOMMON_OPT="$(FCOMMON_OPT)" all
  endif
- ifndef NO_CBLAS
+-	$(MAKE) -C utest all
++	$(MAKE) -C utest FC="$(FC)" CC="$(CC)" COMMON_OPT="$(COMMON_OPT)" FCOMMON_OPT="$(FCOMMON_OPT)" all
+ ifneq ($(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)
+-	$(MAKE) -C cpp_thread_test all
++	$(MAKE) -C cpp_thread_test FC="$(FC)" CC="$(CC)" COMMON_OPT="$(COMMON_OPT)" FCOMMON_OPT="$(FCOMMON_OPT)" all
+ endif
  endif
  endif
- 
diff --git a/SOURCES/openblas-0.3.3-izamax-s390x.patch b/SOURCES/openblas-0.3.3-izamax-s390x.patch
deleted file mode 100644
index b41cee9..0000000
--- a/SOURCES/openblas-0.3.3-izamax-s390x.patch
+++ /dev/null
@@ -1,13 +0,0 @@
-diff --git a/kernel/zarch/izamax.c b/kernel/zarch/izamax.c
-index 216c341..6cde691 100644
---- a/kernel/zarch/izamax.c
-+++ b/kernel/zarch/izamax.c
-@@ -185,7 +185,7 @@ static BLASLONG   ziamax_kernel_16_TUNED(BLASLONG n, FLOAT *x, FLOAT *maxf) {
-             "vsteg  %%v6,%[maxf],0  \n\t"
-             "vmnlg  %%v1,%%v5,%%v7 \n\t"
-             "vlgvg  %[index],%%v1,0  \n\t"
--            "j 3    \n\t"
-+            "j 3f   \n\t"
-             "2:     \n\t"
-             "wfchdb %%v16,%%v26,%%v6      \n\t"
-             "vsel   %%v1,%%v5,%%v7,%%v16  \n\t"
diff --git a/SOURCES/openblas-0.3.3-noopt.patch b/SOURCES/openblas-0.3.3-noopt.patch
deleted file mode 100644
index ec0b143..0000000
--- a/SOURCES/openblas-0.3.3-noopt.patch
+++ /dev/null
@@ -1,37 +0,0 @@
-diff --git a/lapack-netlib/INSTALL/Makefile b/lapack-netlib/INSTALL/Makefile
-index 150a061..abf0843 100644
---- a/lapack-netlib/INSTALL/Makefile
-+++ b/lapack-netlib/INSTALL/Makefile
-@@ -45,6 +45,3 @@ cleantest:
- .SUFFIXES: .o .f
- .f.o:
- 	$(FORTRAN) $(OPTS) -c -o $@ $<
--
--slamch.o: slamch.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
--dlamch.o: dlamch.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
-diff --git a/lapack-netlib/SRC/Makefile b/lapack-netlib/SRC/Makefile
-index 531cb51..94051a1 100644
---- a/lapack-netlib/SRC/Makefile
-+++ b/lapack-netlib/SRC/Makefile
-@@ -603,10 +603,3 @@ clean:
- 
- .F.o:
- 	$(FORTRAN) $(OPTS) -c $< -o $@
--	
--slaruv.o: slaruv.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
--dlaruv.o: dlaruv.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
--sla_wwaddw.o: sla_wwaddw.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
--dla_wwaddw.o: dla_wwaddw.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
--cla_wwaddw.o: cla_wwaddw.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
--zla_wwaddw.o: zla_wwaddw.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
-diff --git a/lapack-netlib/TESTING/MATGEN/Makefile b/lapack-netlib/TESTING/MATGEN/Makefile
-index e20004c..f168821 100644
---- a/lapack-netlib/TESTING/MATGEN/Makefile
-+++ b/lapack-netlib/TESTING/MATGEN/Makefile
-@@ -95,6 +95,3 @@ cleanlib:
- 
- .f.o:
- 	$(FORTRAN) $(OPTS) -c -o $@ $<
--
--slaran.o: slaran.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
--dlaran.o: dlaran.f ; $(FORTRAN) $(NOOPT) -c -o $@ $<
diff --git a/SOURCES/openblas-0.3.3-power9.patch b/SOURCES/openblas-0.3.3-power9.patch
deleted file mode 100644
index 4b3d046..0000000
--- a/SOURCES/openblas-0.3.3-power9.patch
+++ /dev/null
@@ -1,57 +0,0 @@
-diff --git a/cpuid_power.c b/cpuid_power.c
-index 6c7baef..388ea18 100644
---- a/cpuid_power.c
-+++ b/cpuid_power.c
-@@ -56,6 +56,7 @@
- #define CPUTYPE_CELL       6
- #define CPUTYPE_PPCG4	   7
- #define CPUTYPE_POWER8     8
-+#define CPUTYPE_POWER9     9
- 
- char *cpuname[] = {
-   "UNKNOWN",
-@@ -66,7 +67,8 @@ char *cpuname[] = {
-   "POWER6",
-   "CELL",
-   "PPCG4",
--  "POWER8"
-+  "POWER8",
-+  "POWER9"
- };
- 
- char *lowercpuname[] = {
-@@ -78,7 +80,8 @@ char *lowercpuname[] = {
-   "power6",
-   "cell",
-   "ppcg4",
--  "power8"
-+  "power8",
-+  "power9"
- };
- 
- char *corename[] = {
-@@ -90,6 +93,7 @@ char *corename[] = {
-   "POWER6",
-   "CELL",
-   "PPCG4",
-+  "POWER8",
-   "POWER8"
- };
- 
-@@ -120,6 +124,7 @@ int detect(void){
-   if (!strncasecmp(p, "POWER6", 6)) return CPUTYPE_POWER6;
-   if (!strncasecmp(p, "POWER7", 6)) return CPUTYPE_POWER6;
-   if (!strncasecmp(p, "POWER8", 6)) return CPUTYPE_POWER8;
-+  if (!strncasecmp(p, "POWER9", 6)) return CPUTYPE_POWER8;
-   if (!strncasecmp(p, "Cell",   4)) return CPUTYPE_CELL;
-   if (!strncasecmp(p, "7447",   4)) return CPUTYPE_PPCG4;
- 
-@@ -148,7 +153,7 @@ int id;
- id = __asm __volatile("mfpvr %0" : "=r"(id));
- switch ( id >> 16 ) {
-   case 0x4e: // POWER9
--    return  return CPUTYPE_POWER8;
-+    return CPUTYPE_POWER8;
-     break;
-   case 0x4d:
-   case 0x4b: // POWER8/8E 
diff --git a/SPECS/openblas.spec b/SPECS/openblas.spec
index ff30d9d..d1d590f 100644
--- a/SPECS/openblas.spec
+++ b/SPECS/openblas.spec
@@ -1,6 +1,6 @@
 %bcond_with system_lapack
 # Version of bundled lapack
-%global lapackver 3.7.0
+%global lapackver 3.9.0
 
 # DO NOT "CLEAN UP" OR MODIFY THIS SPEC FILE WITHOUT ASKING THE
 # MAINTAINER FIRST!
@@ -14,31 +14,25 @@
 # "obsoleted" features are still kept in the spec.
 
 Name:           openblas
-Version:        0.3.3
-Release:        5%{?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-0.2.15-system_lapack.patch
+Patch0:         openblas-0.3.12-system-lapack.patch
 # Drop extra p from threaded library name
-Patch1:         openblas-0.2.5-libname.patch
-# Don't use constructor priorities on too old architectures
-Patch2:         openblas-0.2.15-constructor.patch
+Patch1:         openblas-0.3.12-libname.patch
 # Supply the proper flags to the test makefile
-Patch3:         openblas-0.3.3-tests.patch
+Patch2:         openblas-0.3.12-tests.patch
 # Enable optimizations for all LAPACK sources
-Patch4:         openblas-0.3.3-noopt.patch
+Patch3:         openblas-0.3.12-noopt.patch
 # Pass ASMFLAGS to assembler compiler
-Patch5:         openblas-0.2.20-asmflags.patch
+Patch4:         openblas-0.3.12-asmflags.patch
 # Remove optimization pragmas on ppc64le
-Patch6:         openblas-0.2.20-power-optimize.patch
-# Fix izamax on s390x
-Patch7:         openblas-0.3.3-izamax-s390x.patch
-# Detect POWER9 as POWER8
-Patch8:         openblas-0.3.3-power9.patch
+Patch5:         openblas-0.3.12-power-optimize.patch
 
 BuildRequires:  gcc
 BuildRequires:  gcc-gfortran
@@ -232,18 +226,13 @@ This package contains the static libraries.
 tar zxf %{SOURCE0}
 cd OpenBLAS-%{version}
 %if %{with system_lapack}
-%patch0 -p1 -b .system_lapack
+%patch0 -p1 -b .system-lapack
 %endif
 %patch1 -p1 -b .libname
-%if 0%{?rhel} == 5
-%patch2 -p1 -b .constructor
-%endif
-%patch3 -p1 -b .tests
-%patch4 -p1 -b .noopt
-%patch5 -p1 -b .asmflags
-%patch6 -p1 -b .power-optimize
-%patch7 -p1 -b .izamax-s390x
-%patch8 -p1 -b .power9
+%patch2 -p1 -b .tests
+%patch3 -p1 -b .noopt
+%patch4 -p1 -b .asmflags
+%patch5 -p1 -b .power-optimize
 
 # Fix source permissions
 find -name \*.f -exec chmod 644 {} \;
@@ -382,7 +371,7 @@ TARGET="TARGET=POWER8 DYNAMIC_ARCH=0"
 TARGET="TARGET=ARMV8 DYNAMIC_ARCH=0"
 %endif
 %ifarch s390x
-TARGET="TARGET=Z13 DYNAMIC_ARCH=0"
+TARGET="TARGET=ZARCH_GENERIC DYNAMIC_ARCH=1"
 %endif
 
 %if 0%{?rhel} == 5
@@ -461,9 +450,6 @@ suffix="_power8"
 %ifarch aarch64
 suffix="_armv8"
 %endif
-%ifarch s390x
-suffix="_z13"
-%endif
 slibname=`basename %{buildroot}%{_libdir}/libopenblas${suffix}-*.so .so`
 mv %{buildroot}%{_libdir}/${slibname}.a %{buildroot}%{_libdir}/lib%{name}.a
 if [[ "$suffix" != "" ]]; then
@@ -689,6 +675,18 @@ 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
+
+* Tue Oct 20 2020 Nikola Forró <nforro@redhat.com> - 0.3.10-1
+- Rebase to version 0.3.10
+  resolves: #1847435
+
 * Fri Nov 22 2019 Nikola Forró <nforro@redhat.com> - 0.3.3-5
 - Detect POWER9 as POWER8
   related: #1752241