Merge branch 'develop'

CONTRIBUTORS.md

@@ -114,5 +114,8 @@ In chronological order:
 * carlkl <https://github.com/carlkl>
   * [2013-12-13] Fixed LAPACKE building bug on Windows
 
+* Isaac Dunham <https://github.com/idunham>
+  * [2014-08-03] Fixed link error on Linux/musl
+
 * [Your name or handle] <[email or website]>
   * [Date] [Brief summary of your changes]

Changelog.txt

@@ -1,4 +1,19 @@
 OpenBLAS ChangeLog
+====================================================================
+Version 0.2.10
+18-Aug-2014
+common:
+	* Added some benchmark codes.
+	* Fix link error on Linux/musl.(Thanks Isaac Dunham)
+
+x86/x86-64:
+	* Improved s/c/zgemm performance for Intel Haswell.
+	* Improved s/d/c/zgemv performance.
+	* Support the big numa machine.(EXPERIMENT)
+
+ARM:
+	* Fix detection when cpuinfo uses "Processor". (Thanks Isaiah)
+
 ====================================================================
 Version 0.2.10
 16-Jul-2014

Makefile

@@ -247,7 +247,7 @@ ifndef NOFORTRAN
 	-@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 ($(FC), gfortran)
 	-@echo "TIMER       = INT_ETIME" >> $(NETLIB_LAPACK_DIR)/make.inc
 ifdef SMP
 	-@echo "LOADER      = $(FC) -pthread" >> $(NETLIB_LAPACK_DIR)/make.inc

Makefile.install

@@ -79,7 +79,8 @@ endif
 ifeq ($(OSNAME), Darwin)
 	@-cp $(LIBDYNNAME) $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
 	@-install_name_tool -id $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)/$(LIBDYNNAME) $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)/$(LIBDYNNAME)
-	@-ln -fs $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)/$(LIBDYNNAME) $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)/$(LIBPREFIX).dylib
+	@cd $(DESTDIR)$(OPENBLAS_LIBRARY_DIR) ; \
+	ln -fs $(LIBDYNNAME) $(LIBPREFIX).dylib
 endif
 ifeq ($(OSNAME), WINNT)
 	@-cp $(LIBDLLNAME) $(OPENBLAS_BINARY_DIR)

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.2.10
+VERSION = 0.2.11
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library
@@ -95,6 +95,9 @@ NO_WARMUP = 1
 # If you want to disable CPU/Memory affinity on Linux.
 NO_AFFINITY = 1
 
+# if you are compiling for Linux and you have more than 16 numa nodes or more than 256 cpus
+# BIGNUMA = 1
+
 # Don't use AVX kernel on Sandy Bridge. It is compatible with old compilers
 # and OS. However, the performance is low.
 # NO_AVX = 1

Makefile.system

@@ -803,6 +803,10 @@ ifeq ($(USE_OPENMP), 1)
 CCOMMON_OPT	+= -DUSE_OPENMP
 endif
 
+ifeq ($(BIGNUMA), 1)
+CCOMMON_OPT	+= -DBIGNUMA
+endif
+
 endif
 
 ifeq ($(NO_WARMUP), 1)

benchmark/Makefile

@@ -2,12 +2,12 @@ TOPDIR	= ..
 include $(TOPDIR)/Makefile.system
 
 # ACML standard
-ACML=/opt/acml5.3.1/gfortran64_mp/lib
-LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
+#ACML=/opt/acml5.3.1/gfortran64_mp/lib
+#LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
 
 # ACML custom
-#ACML=/opt/pb/acml-5-3-1-gfortran-64bit/gfortran64_fma4_mp/lib
-#LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
+ACML=/opt/pb/acml-5-3-1-gfortran-64bit/gfortran64_fma4_mp/lib
+LIBACML = -fopenmp $(ACML)/libacml_mp.a -lgfortran -lm
 
 # Atlas Ubuntu 
 #ATLAS=/usr/lib/atlas-base
@@ -18,12 +18,12 @@ ATLAS=/usr/lib64/atlas
 LIBATLAS = -fopenmp $(ATLAS)/liblapack.a  $(ATLAS)/libptcblas.a  $(ATLAS)/libptf77blas.a  $(ATLAS)/libatlas.a -lgfortran -lm
 
 # Intel standard
-MKL=/opt/intel/mkl/lib/intel64
-LIBMKL = -L$(MKL) -lmkl_intel_lp64 -lmkl_gnu_thread -lmkl_core -lgomp -lpthread -lm
+# MKL=/opt/intel/mkl/lib/intel64
+# LIBMKL = -L$(MKL) -lmkl_intel_lp64 -lmkl_gnu_thread -lmkl_core -lgomp -lpthread -lm
 
 # Intel custom
-#MKL=/home/saar/intel_mkl
-#LIBMKL = -L$(MKL) -lmkl_intel_lp64 -lmkl_gnu_thread -lmkl_core -lgomp -lpthread -lm
+MKL=/home/saar/intel_mkl
+LIBMKL = -L$(MKL) -lmkl_intel_lp64 -lmkl_gnu_thread -lmkl_core -lgomp -lpthread -lm
 
 
 
@@ -34,9 +34,15 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.goto \
        strsm.goto dtrsm.goto ctrsm.goto ztrsm.goto \
        ssyrk.goto dsyrk.goto csyrk.goto zsyrk.goto \
        ssyr2k.goto dsyr2k.goto csyr2k.goto zsyr2k.goto \
+       sger.goto dger.goto \
+       ssymv.goto dsymv.goto \
        chemm.goto zhemm.goto \
        cherk.goto zherk.goto \
        cher2k.goto zher2k.goto \
+       sgemv.goto dgemv.goto cgemv.goto zgemv.goto \
+       sgeev.goto dgeev.goto cgeev.goto zgeev.goto \
+       sgetri.goto dgetri.goto cgetri.goto zgetri.goto \
+       spotrf.goto dpotrf.goto cpotrf.goto zpotrf.goto \
        ssymm.goto dsymm.goto csymm.goto zsymm.goto
 
 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
@@ -46,9 +52,15 @@ acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        strsm.acml dtrsm.acml ctrsm.acml ztrsm.acml \
        ssyrk.acml dsyrk.acml csyrk.acml zsyrk.acml \
        ssyr2k.acml dsyr2k.acml csyr2k.acml zsyr2k.acml \
+       sger.acml dger.acml \
+       ssymv.acml dsymv.acml \
        chemm.acml zhemm.acml \
        cherk.acml zherk.acml \
        cher2k.acml zher2k.acml \
+       sgemv.acml dgemv.acml cgemv.acml zgemv.acml \
+       sgeev.acml dgeev.acml cgeev.acml zgeev.acml \
+       sgetri.acml dgetri.acml cgetri.acml zgetri.acml \
+       spotrf.acml dpotrf.acml cpotrf.acml zpotrf.acml \
        ssymm.acml dsymm.acml csymm.acml zsymm.acml
 
 atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
@@ -58,9 +70,16 @@ atlas :: slinpack.atlas dlinpack.atlas clinpack.atlas zlinpack.atlas \
        strsm.atlas dtrsm.atlas ctrsm.atlas ztrsm.atlas \
        ssyrk.atlas dsyrk.atlas csyrk.atlas zsyrk.atlas \
        ssyr2k.atlas dsyr2k.atlas csyr2k.atlas zsyr2k.atlas \
+       sger.atlas dger.atlas \
+       ssymv.atlas dsymv.atlas \
+       chemm.acml zhemm.acml \
        chemm.atlas zhemm.atlas \
        cherk.atlas zherk.atlas \
        cher2k.atlas zher2k.atlas \
+       sgemv.atlas dgemv.atlas cgemv.atlas zgemv.atlas \
+       sgeev.atlas dgeev.atlas cgeev.atlas zgeev.atlas \
+       sgetri.atlas dgetri.atlas cgetri.atlas zgetri.atlas \
+       spotrf.atlas dpotrf.atlas cpotrf.atlas zpotrf.atlas \
        ssymm.atlas dsymm.atlas csymm.atlas zsymm.atlas
 
 mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
@@ -70,9 +89,15 @@ mkl :: slinpack.mkl dlinpack.mkl clinpack.mkl zlinpack.mkl \
        strsm.mkl dtrsm.mkl ctrsm.mkl ztrsm.mkl \
        ssyrk.mkl dsyrk.mkl csyrk.mkl zsyrk.mkl \
        ssyr2k.mkl dsyr2k.mkl csyr2k.mkl zsyr2k.mkl \
+       sger.mkl dger.mkl \
+       ssymv.mkl dsymv.mkl \
        chemm.mkl zhemm.mkl \
        cherk.mkl zherk.mkl \
        cher2k.mkl zher2k.mkl \
+       sgemv.mkl dgemv.mkl cgemv.mkl zgemv.mkl \
+       sgeev.mkl dgeev.mkl cgeev.mkl zgeev.mkl \
+       sgetri.mkl dgetri.mkl cgetri.mkl zgetri.mkl \
+       spotrf.mkl dpotrf.mkl cpotrf.mkl zpotrf.mkl \
        ssymm.mkl dsymm.mkl csymm.mkl zsymm.mkl
 
 all :: goto atlas acml mkl
@@ -601,6 +626,277 @@ zher2k.atlas : zher2k.$(SUFFIX)
 zher2k.mkl : zher2k.$(SUFFIX)
 	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
 
+##################################### Sgemv ####################################################
+sgemv.goto : sgemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+sgemv.acml : sgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sgemv.atlas : sgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sgemv.mkl : sgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Dgemv ####################################################
+dgemv.goto : dgemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+dgemv.acml : dgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dgemv.atlas : dgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dgemv.mkl : dgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Cgemv ####################################################
+
+cgemv.goto : cgemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+cgemv.acml : cgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cgemv.atlas : cgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cgemv.mkl : cgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Zgemv ####################################################
+
+zgemv.goto : zgemv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+zgemv.acml : zgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zgemv.atlas : zgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zgemv.mkl : zgemv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Sger ####################################################
+sger.goto : sger.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+sger.acml : sger.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sger.atlas : sger.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sger.mkl : sger.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Dger ####################################################
+dger.goto : dger.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+dger.acml : dger.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dger.atlas : dger.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dger.mkl : dger.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Ssymv ####################################################
+ssymv.goto : ssymv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+ssymv.acml : ssymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+ssymv.atlas : ssymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+ssymv.mkl : ssymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Dsymv ####################################################
+dsymv.goto : dsymv.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+dsymv.acml : dsymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dsymv.atlas : dsymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dsymv.mkl : dsymv.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Sgeev ####################################################
+sgeev.goto : sgeev.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+sgeev.acml : sgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sgeev.atlas : sgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sgeev.mkl : sgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Dgeev ####################################################
+dgeev.goto : dgeev.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+dgeev.acml : dgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dgeev.atlas : dgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dgeev.mkl : dgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Cgeev ####################################################
+
+cgeev.goto : cgeev.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+cgeev.acml : cgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cgeev.atlas : cgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cgeev.mkl : cgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Zgeev ####################################################
+
+zgeev.goto : zgeev.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+zgeev.acml : zgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zgeev.atlas : zgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zgeev.mkl : zgeev.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+
+##################################### Sgetri ####################################################
+sgetri.goto : sgetri.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+sgetri.acml : sgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sgetri.atlas : sgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+sgetri.mkl : sgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Dgetri ####################################################
+dgetri.goto : dgetri.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+dgetri.acml : dgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dgetri.atlas : dgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dgetri.mkl : dgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Cgetri ####################################################
+
+cgetri.goto : cgetri.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+cgetri.acml : cgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cgetri.atlas : cgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cgetri.mkl : cgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Zgetri ####################################################
+
+zgetri.goto : zgetri.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+zgetri.acml : zgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zgetri.atlas : zgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zgetri.mkl : zgetri.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+
+##################################### Spotrf ####################################################
+spotrf.goto : spotrf.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+spotrf.acml : spotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+spotrf.atlas : spotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+spotrf.mkl : spotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Dpotrf ####################################################
+dpotrf.goto : dpotrf.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+dpotrf.acml : dpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dpotrf.atlas : dpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+dpotrf.mkl : dpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Cpotrf ####################################################
+
+cpotrf.goto : cpotrf.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+cpotrf.acml : cpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cpotrf.atlas : cpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+cpotrf.mkl : cpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+##################################### Zpotrf ####################################################
+
+zpotrf.goto : zpotrf.$(SUFFIX) ../$(LIBNAME)
+	$(CC) $(CFLAGS) -o $(@F) $^ $(CEXTRALIB) $(EXTRALIB) -lm
+
+zpotrf.acml : zpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBACML) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zpotrf.atlas : zpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBATLAS) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+zpotrf.mkl : zpotrf.$(SUFFIX)
+	-$(CC) $(CFLAGS) -o $(@F) $^ $(LIBMKL) $(CEXTRALIB) $(EXTRALIB) $(FEXTRALIB)
+
+
 ###################################################################################################
 
 slinpack.$(SUFFIX) : linpack.c
@@ -717,6 +1013,67 @@ cher2k.$(SUFFIX) : her2k.c
 zher2k.$(SUFFIX) : her2k.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
+sgemv.$(SUFFIX) : gemv.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+
+dgemv.$(SUFFIX) : gemv.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
+cgemv.$(SUFFIX) : gemv.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
+
+zgemv.$(SUFFIX) : gemv.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
+
+sger.$(SUFFIX) : ger.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+
+dger.$(SUFFIX) : ger.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
+ssymv.$(SUFFIX) : symv.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+
+dsymv.$(SUFFIX) : symv.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
+sgeev.$(SUFFIX) : geev.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+
+dgeev.$(SUFFIX) : geev.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
+cgeev.$(SUFFIX) : geev.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
+
+zgeev.$(SUFFIX) : geev.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
+
+sgetri.$(SUFFIX) : getri.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+
+dgetri.$(SUFFIX) : getri.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
+cgetri.$(SUFFIX) : getri.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
+
+zgetri.$(SUFFIX) : getri.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
+
+spotrf.$(SUFFIX) : potrf.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -UDOUBLE -o $(@F) $^
+
+dpotrf.$(SUFFIX) : potrf.c
+	$(CC) $(CFLAGS) -c -UCOMPLEX -DDOUBLE -o $(@F) $^
+
+cpotrf.$(SUFFIX) : potrf.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -UDOUBLE -o $(@F) $^
+
+zpotrf.$(SUFFIX) : potrf.c
+	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
+
+
 
 
 

benchmark/cholesky.c

@@ -119,7 +119,11 @@ static __inline double getmflops(int ratio, int m, double secs){
 
 int MAIN__(int argc, char *argv[]){
 
+#ifndef COMPLEX
   char *trans[] = {"T", "N"};
+#else
+  char *trans[] = {"C", "N"};
+#endif
   char *uplo[]  = {"U", "L"};
   FLOAT alpha[] = {1.0, 0.0};
   FLOAT beta [] = {0.0, 0.0};

benchmark/geev.c

@@ -0,0 +1,260 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+
+#undef GEEV
+
+#ifndef COMPLEX
+#ifdef XDOUBLE
+#define GEEV   BLASFUNC(qgeev)
+#elif defined(DOUBLE)
+#define GEEV   BLASFUNC(dgeev)
+#else
+#define GEEV   BLASFUNC(sgeev)
+#endif
+#else
+#ifdef XDOUBLE
+#define GEEV   BLASFUNC(xgeev)
+#elif defined(DOUBLE)
+#define GEEV   BLASFUNC(zgeev)
+#else
+#define GEEV   BLASFUNC(cgeev)
+#endif
+#endif
+
+#ifndef COMPLEX
+extern void GEEV( char* jobvl, char* jobvr, blasint* n, FLOAT* a,
+                blasint* lda, FLOAT* wr, FLOAT* wi, FLOAT* vl, blasint* ldvl,
+                FLOAT* vr, blasint* ldvr, FLOAT* work, blasint* lwork, blasint* info );
+#else
+extern void GEEV( char* jobvl, char* jobvr, blasint* n, FLOAT* a,
+                blasint* lda, FLOAT* wr, FLOAT* vl, blasint* ldvl,
+                FLOAT* vr, blasint* ldvr, FLOAT* work, blasint* lwork, FLOAT *rwork, blasint* info );
+#endif
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+  FLOAT *a,*vl,*vr,*wi,*wr,*work,*rwork;
+  FLOAT wkopt[4];
+  char job='V';
+  char *p;
+
+  blasint m, i, j, info,lwork;
+  double factor = 26.33;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  if ((p = getenv("OPENBLAS_JOB")))  job=*p;
+
+  if ( job == 'N' ) factor = 10.0;
+
+  fprintf(stderr, "From : %3d  To : %3d Step = %3d Job=%c\n", from, to, step,job);
+
+  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( vl = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( vr = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( wr = (FLOAT *)malloc(sizeof(FLOAT) * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( wi = (FLOAT *)malloc(sizeof(FLOAT) * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( rwork = (FLOAT *)malloc(sizeof(FLOAT) * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+    for(j = 0; j < to; j++){
+      for(i = 0; i < to * COMPSIZE; i++){
+	a[i + j * to * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      }
+    }
+
+
+    lwork = -1;
+    m=to;
+#ifndef COMPLEX
+    GEEV (&job, &job, &m, a, &m, wr, wi, vl, &m, vr, &m, wkopt, &lwork, &info);
+#else
+    GEEV (&job, &job, &m, a, &m, wr, vl, &m, vr, &m, wkopt, &lwork,rwork, &info);
+#endif
+
+  lwork = (blasint)wkopt[0];
+  if (( work = (FLOAT *)malloc(sizeof(FLOAT) * lwork * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE           FLops           Time          Lwork\n");
+
+  for(m = from; m <= to; m += step){
+
+    fprintf(stderr, " %6d : ", (int)m);
+    gettimeofday( &start, (struct timezone *)0);
+
+    lwork = -1;
+#ifndef COMPLEX
+    GEEV (&job, &job, &m, a, &m, wr, wi, vl, &m, vr, &m, wkopt, &lwork, &info);
+#else
+    GEEV (&job, &job, &m, a, &m, wr, vl, &m, vr, &m, wkopt, &lwork,rwork, &info);
+#endif
+
+    lwork = (blasint)wkopt[0];
+#ifndef COMPLEX
+    GEEV (&job, &job, &m, a, &m, wr, wi, vl, &m, vr, &m, work, &lwork, &info);
+#else
+    GEEV (&job, &job, &m, a, &m, wr, vl, &m, vr, &m, work, &lwork,rwork, &info);
+#endif
+
+    gettimeofday( &stop, (struct timezone *)0);
+
+    if (info) {
+      fprintf(stderr, "failed to compute eigenvalues .. %d\n", info);
+      exit(1);
+    }
+
+    time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+    fprintf(stderr,
+	    " %10.2f MFlops : %10.2f Sec : %d\n",
+	    COMPSIZE * COMPSIZE * factor * (double)m * (double)m * (double)m / time1 * 1.e-6,time1,lwork);
+
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

benchmark/gemv.c

@@ -0,0 +1,229 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+
+#undef GEMV
+
+#ifndef COMPLEX
+
+#ifdef DOUBLE
+#define GEMV   BLASFUNC(dgemv)
+#else
+#define GEMV   BLASFUNC(sgemv)
+#endif
+
+#else
+
+#ifdef DOUBLE
+#define GEMV   BLASFUNC(zgemv)
+#else
+#define GEMV   BLASFUNC(cgemv)
+#endif
+
+#endif
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+  FLOAT *a, *x, *y;
+  FLOAT alpha[] = {1.0, 1.0};
+  FLOAT beta [] = {1.0, 1.0};
+  char trans='N';
+  blasint m, i, j;
+  blasint inc_x=1,inc_y=1;
+  blasint n=0;
+  int has_param_n = 0;
+  int loops = 1;
+  int l;
+  char *p;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1,timeg;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  if ((p = getenv("OPENBLAS_LOOPS")))  loops = atoi(p);
+  if ((p = getenv("OPENBLAS_INCX")))   inc_x = atoi(p);
+  if ((p = getenv("OPENBLAS_INCY")))   inc_y = atoi(p);
+  if ((p = getenv("OPENBLAS_TRANS")))  trans=*p;
+  if ((p = getenv("OPENBLAS_PARAM_N"))) {
+	  n = atoi(p);
+	  if ((n>0) && (n<=to)) has_param_n = 1;
+  }
+
+  if ( has_param_n == 1 )
+    fprintf(stderr, "From : %3d  To : %3d Step = %3d Trans = '%c' N = %d Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,trans,n,inc_x,inc_y,loops);
+  else
+    fprintf(stderr, "From : %3d  To : %3d Step = %3d Trans = '%c' Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,trans,inc_x,inc_y,loops);
+
+  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE       Flops\n");
+
+  for(m = from; m <= to; m += step)
+  {
+
+   timeg=0;
+
+   if ( has_param_n == 0 ) n = m;
+
+   fprintf(stderr, " %6dx%d : ", (int)m,(int)n);
+
+   for(j = 0; j < m; j++){
+      		for(i = 0; i < n * COMPSIZE; i++){
+			a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      		}
+   }
+
+
+    for (l=0; l<loops; l++)
+    {
+
+   	for(i = 0; i < n * COMPSIZE * abs(inc_x); i++){
+			x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+
+   	for(i = 0; i < n * COMPSIZE * abs(inc_y); i++){
+			y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+    	gettimeofday( &start, (struct timezone *)0);
+
+    	GEMV (&trans, &m, &n, alpha, a, &m, x, &inc_x, beta, y, &inc_y );
+
+    	gettimeofday( &stop, (struct timezone *)0);
+
+    	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+	timeg += time1;
+
+    }
+
+    timeg /= loops;
+
+    fprintf(stderr,
+	    " %10.2f MFlops\n",
+	    COMPSIZE * COMPSIZE * 2. * (double)m * (double)n / timeg * 1.e-6);
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

benchmark/ger.c

@@ -0,0 +1,218 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+
+#undef GER
+
+
+#ifdef DOUBLE
+#define GER   BLASFUNC(dger)
+#else
+#define GER   BLASFUNC(sger)
+#endif
+
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+  FLOAT *a, *x, *y;
+  FLOAT alpha[] = {1.0, 1.0};
+  blasint m, i, j;
+  blasint inc_x=1,inc_y=1;
+  blasint n=0;
+  int has_param_n = 0;
+  int loops = 1;
+  int l;
+  char *p;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1,timeg;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  if ((p = getenv("OPENBLAS_LOOPS")))  loops = atoi(p);
+  if ((p = getenv("OPENBLAS_INCX")))   inc_x = atoi(p);
+  if ((p = getenv("OPENBLAS_INCY")))   inc_y = atoi(p);
+  if ((p = getenv("OPENBLAS_PARAM_N"))) {
+	  n = atoi(p);
+	  if ((n>0) && (n<=to)) has_param_n = 1;
+  }
+
+  if ( has_param_n == 1 )
+    fprintf(stderr, "From : %3d  To : %3d Step = %3d N = %d Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,n,inc_x,inc_y,loops);
+  else
+    fprintf(stderr, "From : %3d  To : %3d Step = %3d Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,inc_x,inc_y,loops);
+
+  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE       Flops\n");
+
+  for(m = from; m <= to; m += step)
+  {
+
+   timeg=0;
+
+   if ( has_param_n == 0 ) n = m;
+
+   fprintf(stderr, " %6dx%d : ", (int)m,(int)n);
+
+   for(j = 0; j < m; j++){
+      		for(i = 0; i < n * COMPSIZE; i++){
+			a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      		}
+   }
+
+
+   for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){
+			x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   }
+
+   for(i = 0; i < n * COMPSIZE * abs(inc_y); i++){
+			y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   }
+
+    for (l=0; l<loops; l++)
+    {
+
+    	gettimeofday( &start, (struct timezone *)0);
+
+    	GER (&m, &n, alpha, x, &inc_x, y, &inc_y, a , &m);
+
+    	gettimeofday( &stop, (struct timezone *)0);
+
+    	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+	timeg += time1;
+
+    }
+
+    timeg /= loops;
+
+    fprintf(stderr,
+	    " %10.2f MFlops\n",
+	    COMPSIZE * COMPSIZE * 2. * (double)m * (double)n / timeg * 1.e-6);
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+

benchmark/getri.c

@@ -0,0 +1,234 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+#undef GETRF
+#undef GETRI
+
+#ifndef COMPLEX
+#ifdef XDOUBLE
+#define GETRF   BLASFUNC(qgetrf)
+#define GETRI   BLASFUNC(qgetri)
+#elif defined(DOUBLE)
+#define GETRF   BLASFUNC(dgetrf)
+#define GETRI   BLASFUNC(dgetri)
+#else
+#define GETRF   BLASFUNC(sgetrf)
+#define GETRI   BLASFUNC(sgetri)
+#endif
+#else
+#ifdef XDOUBLE
+#define GETRF   BLASFUNC(xgetrf)
+#define GETRI   BLASFUNC(xgetri)
+#elif defined(DOUBLE)
+#define GETRF   BLASFUNC(zgetrf)
+#define GETRI   BLASFUNC(zgetri)
+#else
+#define GETRF   BLASFUNC(cgetrf)
+#define GETRI   BLASFUNC(cgetri)
+#endif
+#endif
+
+extern void GETRI(blasint *m, FLOAT *a, blasint *lda, blasint *ipiv, FLOAT *work, blasint *lwork, blasint *info);
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+  FLOAT *a,*work;
+  FLOAT wkopt[4];
+  blasint *ipiv;
+  blasint m, i, j, info,lwork;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+
+  fprintf(stderr, "From : %3d  To : %3d Step = %3d\n", from, to, step);
+
+  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( ipiv = (blasint *)malloc(sizeof(blasint) * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+
+
+    for(j = 0; j < to; j++){
+      for(i = 0; i < to * COMPSIZE; i++){
+	a[i + j * to * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      }
+    }
+
+
+    lwork = -1;
+    m=to;
+
+  GETRI(&m, a, &m, ipiv, wkopt, &lwork, &info);
+
+  lwork = (blasint)wkopt[0];
+  if (( work = (FLOAT *)malloc(sizeof(FLOAT) * lwork * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE           FLops           Time          Lwork\n");
+
+  for(m = from; m <= to; m += step){
+
+    fprintf(stderr, " %6d : ", (int)m);
+
+    GETRF (&m, &m, a, &m, ipiv, &info);
+
+    if (info) {
+      fprintf(stderr, "Matrix is not singular .. %d\n", info);
+      exit(1);
+    }
+
+    gettimeofday( &start, (struct timezone *)0);
+
+    lwork = -1;
+    GETRI(&m, a, &m, ipiv, wkopt, &lwork, &info);
+
+    lwork = (blasint)wkopt[0];
+    GETRI(&m, a, &m, ipiv, work, &lwork, &info);
+    gettimeofday( &stop, (struct timezone *)0);
+
+    if (info) {
+      fprintf(stderr, "failed compute inverse matrix .. %d\n", info);
+      exit(1);
+    }
+
+    time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+    fprintf(stderr,
+	    " %10.2f MFlops : %10.2f Sec : %d\n",
+	    COMPSIZE * COMPSIZE * (4.0/3.0 * (double)m * (double)m *(double)m - (double)m *(double)m + 5.0/3.0* (double)m) / time1 * 1.e-6,time1,lwork);
+
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

benchmark/plot-filter.sh

@@ -0,0 +1,60 @@
+#!/bin/sh
+# **********************************************************************************
+# Copyright (c) 2014, The OpenBLAS Project
+# All rights reserved.
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+# 1. Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+# 2. Redistributions in binary form must reproduce the above copyright
+# notice, this list of conditions and the following disclaimer in
+# the documentation and/or other materials provided with the
+# distribution.
+# 3. Neither the name of the OpenBLAS project nor the names of
+# its contributors may be used to endorse or promote products
+# derived from this software without specific prior written permission.
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+# ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+# USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+# **********************************************************************************
+
+# ************************************************************************
+# sample filter for data output from benchmark programs
+#
+# usage example: 
+#   ./dgemm.goto 2>&1|./plotfilter.sh >OpenBLAS
+# ************************************************************************
+
+if [ $# -eq 1 ]
+then
+	arg1=$1
+else
+	arg1=0
+fi
+
+case $arg1 in
+
+L)
+	# Linpack Benchmark
+	awk '/MFlops/ { print $1,int($8) }'|tail --lines=+2
+	;;
+
+C)
+	# Cholesky Benchmark
+	awk '/MFlops/ { print $3,int($9) }'|tail --lines=+2
+	;;
+
+
+*)
+	awk '/MFlops/ { print $1,int($3) }'|tail --lines=+2
+	;;
+esac
+

benchmark/plot-header

@@ -0,0 +1,42 @@
+# **********************************************************************************
+# Copyright (c) 2014, The OpenBLAS Project
+# All rights reserved.
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+# 1. Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+# 2. Redistributions in binary form must reproduce the above copyright
+# notice, this list of conditions and the following disclaimer in
+# the documentation and/or other materials provided with the
+# distribution.
+# 3. Neither the name of the OpenBLAS project nor the names of
+# its contributors may be used to endorse or promote products
+# derived from this software without specific prior written permission.
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+# ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+# USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+# **********************************************************************************
+
+set term x11 font sans;
+set ylabel "MFlops";
+set xlabel "Size";
+set grid xtics;
+set grid ytics;
+set key left;
+set timestamp "generated on %Y-%m-%d by `whoami`"
+set title "Dtrsm\nUPLO=U TRANS=N SIDE=L\nBulldozer 1 Thread"
+plot 'OpenBLAS' smooth bezier, 'ACML' smooth bezier, 'MKL' smooth bezier;
+set output "print.png";
+show title;
+show plot;
+show output;
+
+

benchmark/potrf.c

@@ -0,0 +1,282 @@
+/*********************************************************************/
+/* Copyright 2009, 2010 The University of Texas at Austin.           */
+/* All rights reserved.                                              */
+/*                                                                   */
+/* Redistribution and use in source and binary forms, with or        */
+/* without modification, are permitted provided that the following   */
+/* conditions are met:                                               */
+/*                                                                   */
+/*   1. Redistributions of source code must retain the above         */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer.                                                  */
+/*                                                                   */
+/*   2. Redistributions in binary form must reproduce the above      */
+/*      copyright notice, this list of conditions and the following  */
+/*      disclaimer in the documentation and/or other materials       */
+/*      provided with the distribution.                              */
+/*                                                                   */
+/*    THIS  SOFTWARE IS PROVIDED  BY THE  UNIVERSITY OF  TEXAS AT    */
+/*    AUSTIN  ``AS IS''  AND ANY  EXPRESS OR  IMPLIED WARRANTIES,    */
+/*    INCLUDING, BUT  NOT LIMITED  TO, THE IMPLIED  WARRANTIES OF    */
+/*    MERCHANTABILITY  AND FITNESS FOR  A PARTICULAR  PURPOSE ARE    */
+/*    DISCLAIMED.  IN  NO EVENT SHALL THE UNIVERSITY  OF TEXAS AT    */
+/*    AUSTIN OR CONTRIBUTORS BE  LIABLE FOR ANY DIRECT, INDIRECT,    */
+/*    INCIDENTAL,  SPECIAL, EXEMPLARY,  OR  CONSEQUENTIAL DAMAGES    */
+/*    (INCLUDING, BUT  NOT LIMITED TO,  PROCUREMENT OF SUBSTITUTE    */
+/*    GOODS  OR  SERVICES; LOSS  OF  USE,  DATA,  OR PROFITS;  OR    */
+/*    BUSINESS INTERRUPTION) HOWEVER CAUSED  AND ON ANY THEORY OF    */
+/*    LIABILITY, WHETHER  IN CONTRACT, STRICT  LIABILITY, OR TORT    */
+/*    (INCLUDING NEGLIGENCE OR OTHERWISE)  ARISING IN ANY WAY OUT    */
+/*    OF  THE  USE OF  THIS  SOFTWARE,  EVEN  IF ADVISED  OF  THE    */
+/*    POSSIBILITY OF SUCH DAMAGE.                                    */
+/*                                                                   */
+/* The views and conclusions contained in the software and           */
+/* documentation are those of the authors and should not be          */
+/* interpreted as representing official policies, either expressed   */
+/* or implied, of The University of Texas at Austin.                 */
+/*********************************************************************/
+
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+double fabs(double);
+
+#undef POTRF
+
+#ifndef COMPLEX
+#ifdef XDOUBLE
+#define POTRF   BLASFUNC(qpotrf)
+#define POTRS   BLASFUNC(qpotrs)
+#define POTRI   BLASFUNC(qpotri)
+#define SYRK    BLASFUNC(qsyrk)
+#elif defined(DOUBLE)
+#define POTRF   BLASFUNC(dpotrf)
+#define POTRS   BLASFUNC(dpotrs)
+#define POTRI   BLASFUNC(dpotri)
+#define SYRK    BLASFUNC(dsyrk)
+#else
+#define POTRF   BLASFUNC(spotrf)
+#define POTRS   BLASFUNC(spotrs)
+#define POTRI   BLASFUNC(spotri)
+#define SYRK    BLASFUNC(ssyrk)
+#endif
+#else
+#ifdef XDOUBLE
+#define POTRF   BLASFUNC(xpotrf)
+#define POTRS   BLASFUNC(xpotrs)
+#define POTRI   BLASFUNC(xpotri)
+#define SYRK    BLASFUNC(xherk)
+#elif defined(DOUBLE)
+#define POTRF   BLASFUNC(zpotrf)
+#define POTRS   BLASFUNC(zpotrs)
+#define POTRI   BLASFUNC(zpotri)
+#define SYRK    BLASFUNC(zherk)
+#else
+#define POTRF   BLASFUNC(cpotrf)
+#define POTRS   BLASFUNC(cpotrs)
+#define POTRI   BLASFUNC(cpotri)
+#define SYRK    BLASFUNC(cherk)
+#endif
+#endif
+
+// extern void POTRI(char *uplo, blasint *m, FLOAT *a, blasint *lda, blasint *info);
+// extern void POTRS(char *uplo, blasint *m, blasint *n, FLOAT *a, blasint *lda, FLOAT *b, blasint *ldb, blasint *info);
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+#ifndef COMPLEX
+  char *trans[] = {"T", "N"};
+#else
+  char *trans[] = {"C", "N"};
+#endif
+  char *uplo[]  = {"U", "L"};
+  FLOAT alpha[] = {1.0, 0.0};
+  FLOAT beta [] = {0.0, 0.0};
+
+  FLOAT *a, *b;
+
+  char *p;
+  char btest = 'F';
+
+  blasint m, i, j, info, uplos=0;
+  double flops;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  if ((p = getenv("OPENBLAS_UPLO")))
+	if (*p == 'L') uplos=1;
+
+  if ((p = getenv("OPENBLAS_TEST"))) btest=*p;
+
+  fprintf(stderr, "From : %3d  To : %3d Step = %3d Uplo = %c\n", from, to, step,*uplo[uplos]);
+
+  if (( a    = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( b    = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  for(m = from; m <= to; m += step){
+
+#ifndef COMPLEX
+      if (uplos & 1) {
+	for (j = 0; j < m; j++) {
+	  for(i = 0; i < j; i++)     a[i + j * m] = 0.;
+	                             a[j + j * m] = ((double) rand() / (double) RAND_MAX) + 8.;
+	  for(i = j + 1; i < m; i++) a[i + j * m] = ((double) rand() / (double) RAND_MAX) - 0.5;
+	}
+      } else {
+	for (j = 0; j < m; j++) {
+	  for(i = 0; i < j; i++)     a[i + j * m] = ((double) rand() / (double) RAND_MAX) - 0.5;
+	                             a[j + j * m] = ((double) rand() / (double) RAND_MAX) + 8.;
+	  for(i = j + 1; i < m; i++) a[i + j * m] = 0.;
+	}
+      }
+#else
+      if (uplos & 1) {
+	for (j = 0; j < m; j++) {
+	  for(i = 0; i < j; i++) {
+	    a[(i + j * m) * 2 + 0] = 0.;
+	    a[(i + j * m) * 2 + 1] = 0.;
+	  }
+
+	  a[(j + j * m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) + 8.;
+	  a[(j + j * m) * 2 + 1] = 0.;
+
+	  for(i = j + 1; i < m; i++) {
+	    a[(i + j * m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) - 0.5;
+	    a[(i + j * m) * 2 + 1] = ((double) rand() / (double) RAND_MAX) - 0.5;
+	  }
+	}
+      } else {
+	for (j = 0; j < m; j++) {
+	  for(i = 0; i < j; i++) {
+	    a[(i + j * m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) - 0.5;
+	    a[(i + j * m) * 2 + 1] = ((double) rand() / (double) RAND_MAX) - 0.5;
+	  }
+
+	  a[(j + j * m) * 2 + 0] = ((double) rand() / (double) RAND_MAX) + 8.;
+	  a[(j + j * m) * 2 + 1] = 0.;
+
+	  for(i = j + 1; i < m; i++) {
+	    a[(i + j * m) * 2 + 0] = 0.;
+	    a[(i + j * m) * 2 + 1] = 0.;
+	  }
+	}
+      }
+#endif
+
+      SYRK(uplo[uplos], trans[uplos], &m, &m, alpha, a, &m, beta, b, &m);
+
+      gettimeofday( &start, (struct timezone *)0);
+
+      POTRF(uplo[uplos], &m, b, &m, &info);
+
+      gettimeofday( &stop, (struct timezone *)0);
+
+      if (info != 0) {
+	fprintf(stderr, "Potrf info = %d\n", info);
+	exit(1);
+      }
+
+      time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+      flops = COMPSIZE * COMPSIZE * (1.0/3.0 * (double)m * (double)m *(double)m +1.0/2.0* (double)m *(double)m + 1.0/6.0* (double)m) / time1 * 1.e-6;
+
+      if ( btest == 'S' )
+      {
+	
+ 	for(j = 0; j < to; j++){
+      		for(i = 0; i < to * COMPSIZE; i++){
+        		a[i + j * to * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      		}
+    	}
+
+      	gettimeofday( &start, (struct timezone *)0);
+
+      	POTRS(uplo[uplos], &m, &m, b, &m, a, &m,  &info);
+
+      	gettimeofday( &stop, (struct timezone *)0);
+
+      	if (info != 0) {
+		fprintf(stderr, "Potrs info = %d\n", info);
+		exit(1);
+        }
+        time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+        flops = COMPSIZE * COMPSIZE * (2.0 * (double)m * (double)m *(double)m ) / time1 * 1.e-6;
+
+      }
+	
+      if ( btest == 'I' )
+      {
+	
+      	gettimeofday( &start, (struct timezone *)0);
+
+      	POTRI(uplo[uplos], &m, b, &m, &info);
+
+      	gettimeofday( &stop, (struct timezone *)0);
+
+      	if (info != 0) {
+		fprintf(stderr, "Potri info = %d\n", info);
+		exit(1);
+        }
+
+        time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+        flops = COMPSIZE * COMPSIZE * (2.0/3.0 * (double)m * (double)m *(double)m +1.0/2.0* (double)m *(double)m + 5.0/6.0* (double)m) / time1 * 1.e-6;
+      }
+	
+      fprintf(stderr, "%8d : %10.2f MFlops : %10.3f Sec : Test=%c\n",m,flops ,time1,btest);
+
+
+  }
+
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));
+

benchmark/symv.c

@@ -0,0 +1,218 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdio.h>
+#include <stdlib.h>
+#ifdef __CYGWIN32__
+#include <sys/time.h>
+#endif
+#include "common.h"
+
+
+#undef SYMV
+
+#ifndef COMPLEX
+
+#ifdef DOUBLE
+#define SYMV   BLASFUNC(dsymv)
+#else
+#define SYMV   BLASFUNC(ssymv)
+#endif
+
+#else
+
+#ifdef DOUBLE
+#define SYMV   BLASFUNC(zsymv)
+#else
+#define SYMV   BLASFUNC(csymv)
+#endif
+
+#endif
+
+#if defined(__WIN32__) || defined(__WIN64__)
+
+#ifndef DELTA_EPOCH_IN_MICROSECS
+#define DELTA_EPOCH_IN_MICROSECS 11644473600000000ULL
+#endif
+
+int gettimeofday(struct timeval *tv, void *tz){
+
+  FILETIME ft;
+  unsigned __int64 tmpres = 0;
+  static int tzflag;
+
+  if (NULL != tv)
+    {
+      GetSystemTimeAsFileTime(&ft);
+
+      tmpres |= ft.dwHighDateTime;
+      tmpres <<= 32;
+      tmpres |= ft.dwLowDateTime;
+
+      /*converting file time to unix epoch*/
+      tmpres /= 10;  /*convert into microseconds*/
+      tmpres -= DELTA_EPOCH_IN_MICROSECS;
+      tv->tv_sec = (long)(tmpres / 1000000UL);
+      tv->tv_usec = (long)(tmpres % 1000000UL);
+    }
+
+  return 0;
+}
+
+#endif
+
+#if !defined(__WIN32__) && !defined(__WIN64__) && !defined(__CYGWIN32__) && 0
+
+static void *huge_malloc(BLASLONG size){
+  int shmid;
+  void *address;
+
+#ifndef SHM_HUGETLB
+#define SHM_HUGETLB 04000
+#endif
+
+  if ((shmid =shmget(IPC_PRIVATE,
+		     (size + HUGE_PAGESIZE) & ~(HUGE_PAGESIZE - 1),
+		     SHM_HUGETLB | IPC_CREAT |0600)) < 0) {
+    printf( "Memory allocation failed(shmget).\n");
+    exit(1);
+  }
+
+  address = shmat(shmid, NULL, SHM_RND);
+
+  if ((BLASLONG)address == -1){
+    printf( "Memory allocation failed(shmat).\n");
+    exit(1);
+  }
+
+  shmctl(shmid, IPC_RMID, 0);
+
+  return address;
+}
+
+#define malloc huge_malloc
+
+#endif
+
+int MAIN__(int argc, char *argv[]){
+
+  FLOAT *a, *x, *y;
+  FLOAT alpha[] = {1.0, 1.0};
+  FLOAT beta [] = {1.0, 1.0};
+  char uplo='L';
+  blasint m, i, j;
+  blasint inc_x=1,inc_y=1;
+  int loops = 1;
+  int l;
+  char *p;
+
+  int from =   1;
+  int to   = 200;
+  int step =   1;
+
+  struct timeval start, stop;
+  double time1,timeg;
+
+  argc--;argv++;
+
+  if (argc > 0) { from     = atol(*argv);		argc--; argv++;}
+  if (argc > 0) { to       = MAX(atol(*argv), from);	argc--; argv++;}
+  if (argc > 0) { step     = atol(*argv);		argc--; argv++;}
+
+  if ((p = getenv("OPENBLAS_LOOPS")))  loops = atoi(p);
+  if ((p = getenv("OPENBLAS_INCX")))   inc_x = atoi(p);
+  if ((p = getenv("OPENBLAS_INCY")))   inc_y = atoi(p);
+  if ((p = getenv("OPENBLAS_UPLO")))  uplo=*p;
+
+  fprintf(stderr, "From : %3d  To : %3d Step = %3d Uplo = '%c' Inc_x = %d Inc_y = %d Loops = %d\n", from, to, step,uplo,inc_x,inc_y,loops);
+
+  if (( a = (FLOAT *)malloc(sizeof(FLOAT) * to * to * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( x = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_x) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+  if (( y = (FLOAT *)malloc(sizeof(FLOAT) * to * abs(inc_y) * COMPSIZE)) == NULL){
+    fprintf(stderr,"Out of Memory!!\n");exit(1);
+  }
+
+#ifdef linux
+  srandom(getpid());
+#endif
+
+  fprintf(stderr, "   SIZE       Flops\n");
+
+  for(m = from; m <= to; m += step)
+  {
+
+   timeg=0;
+
+   fprintf(stderr, " %6dx%d : ", (int)m,(int)m);
+
+   for(j = 0; j < m; j++){
+      		for(i = 0; i < m * COMPSIZE; i++){
+			a[i + j * m * COMPSIZE] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+      		}
+   }
+
+
+    for (l=0; l<loops; l++)
+    {
+
+   	for(i = 0; i < m * COMPSIZE * abs(inc_x); i++){
+			x[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+
+   	for(i = 0; i < m * COMPSIZE * abs(inc_y); i++){
+			y[i] = ((FLOAT) rand() / (FLOAT) RAND_MAX) - 0.5;
+   	}
+    	gettimeofday( &start, (struct timezone *)0);
+
+    	SYMV (&uplo, &m, alpha, a, &m, x, &inc_x, beta, y, &inc_y );
+
+    	gettimeofday( &stop, (struct timezone *)0);
+
+    	time1 = (double)(stop.tv_sec - start.tv_sec) + (double)((stop.tv_usec - start.tv_usec)) * 1.e-6;
+
+	timeg += time1;
+
+    }
+
+    timeg /= loops;
+
+    fprintf(stderr,
+	    " %10.2f MFlops\n",
+	    COMPSIZE * COMPSIZE * 2. * (double)m * (double)m / timeg * 1.e-6);
+
+  }
+
+  return 0;
+}
+
+void main(int argc, char *argv[]) __attribute__((weak, alias("MAIN__")));

common.h

@@ -524,10 +524,21 @@ static __inline void blas_unlock(volatile BLASULONG *address){
   *address = 0;
 }
 
+
+#ifdef OS_WINDOWS
 static __inline int readenv_atoi(char *env) {
   env_var_t p;
   return readenv(p,env) ? 0 : atoi(p);
 }
+#else
+static __inline int readenv_atoi(char *env) {
+  char *p;
+  if (( p = getenv(env) ))
+  	return (atoi(p));
+  else
+	return(0);
+}
+#endif
 
 
 #if !defined(XDOUBLE) || !defined(QUAD_PRECISION)

common_interface.h

@@ -679,13 +679,6 @@ int BLASFUNC(cgesv)(blasint *, blasint *, float  *, blasint *, blasint *, float
 int BLASFUNC(zgesv)(blasint *, blasint *, double *, blasint *, blasint *, double*, blasint *, blasint *);
 int BLASFUNC(xgesv)(blasint *, blasint *, xdouble *, blasint *, blasint *, xdouble*, blasint *, blasint *);
 
-int BLASFUNC(sgesvd)(char *, char *, blasint *, blasint *, float   *, blasint *, float   *, float   *, blasint *, float   *, blasint *, float   *, blasint *, blasint *);
-int BLASFUNC(dgesvd)(char *, char *, blasint *, blasint *, double  *, blasint *, double  *, double  *, blasint *, double  *, blasint *, double  *, blasint *, blasint *);
-int BLASFUNC(qgesvd)(char *, char *, blasint *, blasint *, xdouble *, blasint *, xdouble *, xdouble *, blasint *, xdouble *, blasint *, xdouble *, blasint *, blasint *);
-int BLASFUNC(cgesvd)(char *, char *, blasint *, blasint *, float   *, blasint *, float   *, float   *, blasint *, float   *, blasint *, float   *, blasint *, blasint *);
-int BLASFUNC(zgesvd)(char *, char *, blasint *, blasint *, double  *, blasint *, double  *, double  *, blasint *, double  *, blasint *, double  *, blasint *, blasint *);
-int BLASFUNC(xgesvd)(char *, char *, blasint *, blasint *, xdouble *, blasint *, xdouble *, xdouble *, blasint *, xdouble *, blasint *, xdouble *, blasint *, blasint *);
-
 int BLASFUNC(spotf2)(char *, blasint *, float  *, blasint *, blasint *);
 int BLASFUNC(dpotf2)(char *, blasint *, double *, blasint *, blasint *);
 int BLASFUNC(qpotf2)(char *, blasint *, xdouble *, blasint *, blasint *);
@@ -735,19 +728,6 @@ int BLASFUNC(ctrtri)(char *, char *, blasint *, float  *, blasint *, blasint *);
 int BLASFUNC(ztrtri)(char *, char *, blasint *, double *, blasint *, blasint *);
 int BLASFUNC(xtrtri)(char *, char *, blasint *, xdouble *, blasint *, blasint *);
 
-int BLASFUNC(spotri)(char *, blasint *, float  *, blasint *, blasint *);
-int BLASFUNC(dpotri)(char *, blasint *, double *, blasint *, blasint *);
-int BLASFUNC(qpotri)(char *, blasint *, xdouble *, blasint *, blasint *);
-int BLASFUNC(cpotri)(char *, blasint *, float  *, blasint *, blasint *);
-int BLASFUNC(zpotri)(char *, blasint *, double *, blasint *, blasint *);
-int BLASFUNC(xpotri)(char *, blasint *, xdouble *, blasint *, blasint *);
-
-int BLASFUNC(slarf)(char *, blasint *, blasint *, float *, blasint *, float *, float *, blasint *, float *);
-int BLASFUNC(dlarf)(char *, blasint *, blasint *, double *, blasint *, double *, double *, blasint *, double *);
-int BLASFUNC(qlarf)(char *, blasint *, blasint *, xdouble *, blasint *, xdouble *, xdouble *, blasint *, xdouble *);
-int BLASFUNC(clarf)(char *, blasint *, blasint *, float *, blasint *, float *, float *, blasint *, float *);
-int BLASFUNC(zlarf)(char *, blasint *, blasint *, double *, blasint *, double *, double *, blasint *, double *);
-int BLASFUNC(xlarf)(char *, blasint *, blasint *, xdouble *, blasint *, xdouble *, xdouble *, blasint *, xdouble *);
 
 FLOATRET  BLASFUNC(slamch)(char *);
 double    BLASFUNC(dlamch)(char *);

common_param.h

@@ -919,14 +919,22 @@ extern gotoblas_t *gotoblas;
 #define	SGEMM_R		SGEMM_DEFAULT_R
 #define SGEMM_UNROLL_M	SGEMM_DEFAULT_UNROLL_M
 #define SGEMM_UNROLL_N	SGEMM_DEFAULT_UNROLL_N
+#ifdef  SGEMM_DEFAULT_UNROLL_MN
+#define SGEMM_UNROLL_MN	SGEMM_DEFAULT_UNROLL_MN
+#else
 #define SGEMM_UNROLL_MN	MAX((SGEMM_UNROLL_M), (SGEMM_UNROLL_N))
+#endif
 
 #define	DGEMM_P		DGEMM_DEFAULT_P
 #define	DGEMM_Q		DGEMM_DEFAULT_Q
 #define	DGEMM_R		DGEMM_DEFAULT_R
 #define DGEMM_UNROLL_M	DGEMM_DEFAULT_UNROLL_M
 #define DGEMM_UNROLL_N	DGEMM_DEFAULT_UNROLL_N
+#ifdef  DGEMM_DEFAULT_UNROLL_MN
+#define DGEMM_UNROLL_MN	DGEMM_DEFAULT_UNROLL_MN
+#else
 #define DGEMM_UNROLL_MN	MAX((DGEMM_UNROLL_M), (DGEMM_UNROLL_N))
+#endif
 
 #define	QGEMM_P		QGEMM_DEFAULT_P
 #define	QGEMM_Q		QGEMM_DEFAULT_Q
@@ -940,14 +948,22 @@ extern gotoblas_t *gotoblas;
 #define	CGEMM_R		CGEMM_DEFAULT_R
 #define CGEMM_UNROLL_M	CGEMM_DEFAULT_UNROLL_M
 #define CGEMM_UNROLL_N	CGEMM_DEFAULT_UNROLL_N
+#ifdef  CGEMM_DEFAULT_UNROLL_MN
+#define CGEMM_UNROLL_MN	CGEMM_DEFAULT_UNROLL_MN
+#else
 #define CGEMM_UNROLL_MN	MAX((CGEMM_UNROLL_M), (CGEMM_UNROLL_N))
+#endif
 
 #define	ZGEMM_P		ZGEMM_DEFAULT_P
 #define	ZGEMM_Q		ZGEMM_DEFAULT_Q
 #define	ZGEMM_R		ZGEMM_DEFAULT_R
 #define ZGEMM_UNROLL_M	ZGEMM_DEFAULT_UNROLL_M
 #define ZGEMM_UNROLL_N	ZGEMM_DEFAULT_UNROLL_N
+#ifdef  ZGEMM_DEFAULT_UNROLL_MN
+#define ZGEMM_UNROLL_MN	ZGEMM_DEFAULT_UNROLL_MN
+#else
 #define ZGEMM_UNROLL_MN	MAX((ZGEMM_UNROLL_M), (ZGEMM_UNROLL_N))
+#endif
 
 #define	XGEMM_P		XGEMM_DEFAULT_P
 #define	XGEMM_Q		XGEMM_DEFAULT_Q

cpuid_arm.c

@@ -90,7 +90,7 @@ int detect(void)
 	while (fgets(buffer, sizeof(buffer), infile))
 	{
 
-		if (!strncmp("model name", buffer, 10))
+		if ((!strncmp("model name", buffer, 10)) || (!strncmp("Processor", buffer, 9)))
 		{
 			p = strchr(buffer, ':') + 2;
 			break;

driver/level3/level3.c

@@ -333,16 +333,10 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
 
-#if ( defined(BULLDOZER) || defined(PILEDRIVER) || defined(HASWELL) ) && defined(ARCH_X86_64) && !defined(XDOUBLE) && !defined(COMPLEX)
-                if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N;
-                else
-                        if (min_jj >= 3*GEMM_UNROLL_N) min_jj = 3*GEMM_UNROLL_N;
-                        else
-                                if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
-#else
+        if (min_jj >= 3*GEMM_UNROLL_N) min_jj = 3*GEMM_UNROLL_N;
+        else
+          if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
-        if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
-#endif
 
 
 	START_RPCC();

driver/level3/level3_syrk_threaded.c

@@ -538,10 +538,10 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
   mask  = MAX(QGEMM_UNROLL_M, QGEMM_UNROLL_N) - 1;
 #elif defined(DOUBLE)
   mode  =  BLAS_DOUBLE  | BLAS_REAL;
-  mask  = MAX(DGEMM_UNROLL_M, DGEMM_UNROLL_N) - 1;
+  mask  = DGEMM_UNROLL_MN - 1;
 #else
   mode  =  BLAS_SINGLE  | BLAS_REAL;
-  mask  = MAX(SGEMM_UNROLL_M, SGEMM_UNROLL_N) - 1;
+  mask  = SGEMM_UNROLL_MN - 1;
 #endif
 #else
 #ifdef XDOUBLE
@@ -549,10 +549,10 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
   mask  = MAX(XGEMM_UNROLL_M, XGEMM_UNROLL_N) - 1;
 #elif defined(DOUBLE)
   mode  =  BLAS_DOUBLE  | BLAS_COMPLEX;
-  mask  = MAX(ZGEMM_UNROLL_M, ZGEMM_UNROLL_N) - 1;
+  mask  = ZGEMM_UNROLL_MN - 1;
 #else
   mode  =  BLAS_SINGLE  | BLAS_COMPLEX;
-  mask  = MAX(CGEMM_UNROLL_M, CGEMM_UNROLL_N) - 1;
+  mask  = CGEMM_UNROLL_MN - 1;
 #endif
 #endif
 

driver/level3/level3_thread.c

@@ -367,16 +367,10 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
       for(jjs = xxx; jjs < MIN(n_to, xxx + div_n); jjs += min_jj){
 	min_jj = MIN(n_to, xxx + div_n) - jjs;
 
-#if ( defined(BULLDOZER) || defined(PILEDRIVER) || defined(HASWELL) ) && defined(ARCH_X86_64) && !defined(XDOUBLE) && !defined(COMPLEX)
-		if (min_jj >= 6*GEMM_UNROLL_N) min_jj = 6*GEMM_UNROLL_N;
-		else
-			if (min_jj >= 3*GEMM_UNROLL_N) min_jj = 3*GEMM_UNROLL_N;
-			else
-				if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
-#else
+	if (min_jj >= 3*GEMM_UNROLL_N) min_jj = 3*GEMM_UNROLL_N;
+	else
+		if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
-#endif
 
 	START_RPCC();
 

driver/level3/syrk_thread.c

@@ -57,10 +57,10 @@ int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (
 
     switch (mode & BLAS_PREC) {
     case BLAS_SINGLE:
-      mask = MAX(SGEMM_UNROLL_M, SGEMM_UNROLL_N) - 1;
+      mask = SGEMM_UNROLL_MN - 1;
       break;
     case BLAS_DOUBLE:
-      mask = MAX(DGEMM_UNROLL_M, DGEMM_UNROLL_N) - 1;
+      mask = DGEMM_UNROLL_MN - 1;
       break;
 #ifdef EXPRECISION
     case BLAS_XDOUBLE:
@@ -71,10 +71,10 @@ int CNAME(int mode, blas_arg_t *arg, BLASLONG *range_m, BLASLONG *range_n, int (
   } else {
     switch (mode & BLAS_PREC) {
     case BLAS_SINGLE:
-      mask = MAX(CGEMM_UNROLL_M, CGEMM_UNROLL_N) - 1;
+      mask = CGEMM_UNROLL_MN - 1;
       break;
     case BLAS_DOUBLE:
-      mask = MAX(ZGEMM_UNROLL_M, ZGEMM_UNROLL_N) - 1;
+      mask = ZGEMM_UNROLL_MN - 1;
       break;
 #ifdef EXPRECISION
     case BLAS_XDOUBLE:

driver/level3/trmm_L.c

@@ -135,7 +135,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
     for(jjs = js; jjs < js + min_j; jjs += min_jj){
       min_jj = min_j + js - jjs;
-      if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+      if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+      else
+      	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
       START_RPCC();
 
@@ -199,7 +201,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+        if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+        else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 	START_RPCC();
 
@@ -288,7 +292,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
     for(jjs = js; jjs < js + min_j; jjs += min_jj){
       min_jj = min_j + js - jjs;
-      if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+      if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+      else
+        if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
       START_RPCC();
 
@@ -352,7 +358,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+        if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+        else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 	START_RPCC();
 

driver/level3/trmm_R.c

@@ -119,7 +119,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = 0; jjs < ls - js; jjs += min_jj){
 	min_jj = ls - js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	GEMM_ONCOPY(min_l, min_jj, a + (ls + (js + jjs) * lda) * COMPSIZE, lda, sb + min_l * jjs * COMPSIZE);
@@ -137,7 +139,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = 0; jjs < min_l; jjs += min_jj){
 	min_jj = min_l - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	TRMM_OLNCOPY(min_l, min_jj, a, lda, ls, ls + jjs, sb + min_l * (ls - js + jjs) * COMPSIZE);
@@ -188,7 +192,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	GEMM_ONCOPY(min_l, min_jj, a + (ls + jjs * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE);
@@ -239,7 +245,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = 0; jjs < min_l; jjs += min_jj){
 	min_jj = min_l - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	TRMM_OUNCOPY(min_l, min_jj, a, lda, ls, ls + jjs, sb + min_l * jjs * COMPSIZE);
@@ -258,7 +266,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = 0; jjs < js - ls - min_l; jjs += min_jj){
 	min_jj = js - ls - min_l - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	GEMM_ONCOPY(min_l, min_jj, a + (ls + (ls + min_l + jjs) * lda) * COMPSIZE, lda,
@@ -313,7 +323,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	GEMM_ONCOPY(min_l, min_jj, a + (ls + (jjs - min_j) * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE);

driver/level3/trsm_L.c

@@ -128,7 +128,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 	GEMM_ONCOPY(min_l, min_jj, b + (ls + jjs * ldb) * COMPSIZE, ldb, sb + min_l * (jjs - js) * COMPSIZE);
 
@@ -194,7 +196,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 	GEMM_ONCOPY(min_l, min_jj, b + (ls - min_l + jjs * ldb) * COMPSIZE, ldb, sb + min_l * (jjs - js) * COMPSIZE);
 

driver/level3/trsm_R.c

@@ -123,7 +123,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	GEMM_ONCOPY(min_l, min_jj, a + (ls + jjs * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE);
@@ -177,7 +179,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = 0; jjs < min_j - min_l - ls + js; jjs += min_jj){
 	min_jj = min_j - min_l - ls + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
       GEMM_ONCOPY (min_l, min_jj, a + (ls + (ls + min_l + jjs) * lda) * COMPSIZE, lda,
@@ -238,7 +242,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = js; jjs < js + min_j; jjs += min_jj){
 	min_jj = min_j + js - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	GEMM_ONCOPY(min_l, min_jj, a + (ls + (jjs - min_j) * lda) * COMPSIZE, lda, sb + min_l * (jjs - js) * COMPSIZE);
@@ -297,7 +303,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, FLOAT *sa, FLO
 
       for(jjs = 0; jjs < min_j - js + ls; jjs += min_jj){
 	min_jj = min_j - js + ls - jjs;
-	if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
+	if (min_jj > GEMM_UNROLL_N*3) min_jj = GEMM_UNROLL_N*3;
+	else
+	  if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 #ifndef TRANSA
 	GEMM_ONCOPY (min_l, min_jj, a + (ls + (js - min_j + jjs) * lda) * COMPSIZE, lda,

driver/others/init.c

@@ -85,8 +85,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <unistd.h>
 #include <string.h>
 
+#if defined(BIGNUMA)
+// max number of nodes as defined in numa.h
+// max cpus as defined in sched.h
+#define MAX_NODES	128
+#define MAX_CPUS	CPU_SETSIZE
+#else
 #define MAX_NODES	16
 #define MAX_CPUS	256
+#endif
+
 #define NCPUBITS        (8*sizeof(unsigned long))
 #define MAX_BITMASK_LEN (MAX_CPUS/NCPUBITS)
 #define CPUELT(cpu)	((cpu) / NCPUBITS)
@@ -544,16 +552,26 @@ static inline int is_dead(int id) {
 
   return shmctl(id, IPC_STAT, &ds);
 }
+
 static void open_shmem(void) {
 
   int try = 0;
 
   do {
 
+#if defined(BIGNUMA)
+    // raised to 32768, enough for 128 nodes and 1024 cups
+    shmid = shmget(SH_MAGIC, 32768, 0666);
+#else
     shmid = shmget(SH_MAGIC, 4096, 0666);
+#endif
 
     if (shmid == -1) {
+#if defined(BIGNUMA)
+      shmid = shmget(SH_MAGIC, 32768, IPC_CREAT | 0666);
+#else
       shmid = shmget(SH_MAGIC, 4096, IPC_CREAT | 0666);
+#endif
     }
 
     try ++;
@@ -847,7 +865,7 @@ void gotoblas_set_affinity2(int threads) {};
 
 void gotoblas_affinity_reschedule(void) {};
 
-int get_num_procs(void) { return get_nprocs(); }
+int get_num_procs(void) { return sysconf(_SC_NPROCESSORS_ONLN); }
 
 int get_num_nodes(void) { return 1; }
 

driver/others/memory.c

@@ -162,7 +162,7 @@ int get_num_procs(void);
 #else
 int get_num_procs(void) {
   static int nums = 0;
-  if (!nums) nums = get_nprocs();
+  if (!nums) nums = sysconf(_SC_NPROCESSORS_ONLN);
   return nums;
 }
 #endif

interface/gemm.c

@@ -405,11 +405,11 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 
 #ifndef COMPLEX
   double MNK = (double) args.m * (double) args.n * (double) args.k;
-  if ( MNK <= (1024.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
+  if ( MNK <= (16.0 * 1024.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
 	nthreads_max = 1;
   else
   {
-  	if ( MNK <= (65536.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
+  	if ( MNK <= (2.0 * 65536.0 * (double) GEMM_MULTITHREAD_THRESHOLD) )
 	{
 		nthreads_max = 4;
 		if ( args.m < 16 * GEMM_MULTITHREAD_THRESHOLD )

interface/gemv.c

@@ -211,7 +211,18 @@ void CNAME(enum CBLAS_ORDER order,
   buffer = (FLOAT *)blas_memory_alloc(1);
 
 #ifdef SMP
-  nthreads = num_cpu_avail(2);
+
+  int  nthreads_max = num_cpu_avail(2);
+  int  nthreads_avail = nthreads_max;
+
+  double MNK = (double) m * (double) n;
+  if ( MNK <= (500.0 * 100.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
+        nthreads_max = 1;
+
+  if ( nthreads_max > nthreads_avail )
+        nthreads = nthreads_avail;
+  else
+        nthreads = nthreads_max;
 
   if (nthreads == 1) {
 #endif

interface/zgemv.c

@@ -233,7 +233,19 @@ void CNAME(enum CBLAS_ORDER order,
   buffer = (FLOAT *)blas_memory_alloc(1);
 
 #ifdef SMP
-  nthreads = num_cpu_avail(2);
+
+  int  nthreads_max = num_cpu_avail(2);
+  int  nthreads_avail = nthreads_max;
+
+  double MNK = (double) m * (double) n;
+  if ( MNK <= (80.0 * 20.0  * (double) GEMM_MULTITHREAD_THRESHOLD)  )
+        nthreads_max = 1;
+
+  if ( nthreads_max > nthreads_avail )
+        nthreads = nthreads_avail;
+  else
+        nthreads = nthreads_max;
+
 
   if (nthreads == 1) {
 #endif

kernel/setparam-ref.c

@@ -54,7 +54,14 @@ gotoblas_t TABLE_NAME = {
   GEMM_DEFAULT_OFFSET_A, GEMM_DEFAULT_OFFSET_B, GEMM_DEFAULT_ALIGN,
 
   0, 0, 0,
-  SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N, MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
+  SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N,
+#ifdef SGEMM_DEFAULT_UNROLL_MN
+ SGEMM_DEFAULT_UNROLL_MN,
+#else
+ MAX(SGEMM_DEFAULT_UNROLL_M, SGEMM_DEFAULT_UNROLL_N),
+#endif
+
+
 #ifdef HAVE_EXCLUSIVE_CACHE
   1,
 #else
@@ -110,7 +117,12 @@ gotoblas_t TABLE_NAME = {
 #endif
 
   0, 0, 0,
-  DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N, MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
+  DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N,
+#ifdef DGEMM_DEFAULT_UNROLL_MN
+ DGEMM_DEFAULT_UNROLL_MN,
+#else
+ MAX(DGEMM_DEFAULT_UNROLL_M, DGEMM_DEFAULT_UNROLL_N),
+#endif
 
   damax_kTS,  damin_kTS,  dmax_kTS,  dmin_kTS,
   idamax_kTS, idamin_kTS, idmax_kTS, idmin_kTS,
@@ -214,7 +226,12 @@ gotoblas_t TABLE_NAME = {
 #endif
 
   0, 0, 0,
-  CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N, MAX(CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N),
+  CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N,
+#ifdef CGEMM_DEFAULT_UNROLL_MN
+ CGEMM_DEFAULT_UNROLL_MN,
+#else
+ MAX(CGEMM_DEFAULT_UNROLL_M, CGEMM_DEFAULT_UNROLL_N),
+#endif
 
   camax_kTS, camin_kTS, icamax_kTS, icamin_kTS,
   cnrm2_kTS, casum_kTS, ccopy_kTS,
@@ -307,7 +324,12 @@ gotoblas_t TABLE_NAME = {
 #endif
 
   0, 0, 0,
-  ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N, MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
+  ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N,
+#ifdef ZGEMM_DEFAULT_UNROLL_MN
+ ZGEMM_DEFAULT_UNROLL_MN,
+#else
+ MAX(ZGEMM_DEFAULT_UNROLL_M, ZGEMM_DEFAULT_UNROLL_N),
+#endif
 
   zamax_kTS, zamin_kTS, izamax_kTS, izamin_kTS,
   znrm2_kTS, zasum_kTS, zcopy_kTS,

kernel/x86_64/KERNEL

@@ -373,11 +373,11 @@ endif
 GEMVDEP = ../l2param.h
 
 ifndef SGEMVNKERNEL
-SGEMVNKERNEL = ../arm/gemv_n.c
+SGEMVNKERNEL = sgemv_n.c
 endif
 
 ifndef SGEMVTKERNEL
-SGEMVTKERNEL = ../arm/gemv_t.c
+SGEMVTKERNEL = sgemv_t.c
 endif
 
 ifndef DGEMVNKERNEL

kernel/x86_64/KERNEL.BULLDOZER

@@ -1,6 +1,8 @@
+SGEMVNKERNEL = sgemv_n.c
+SGEMVTKERNEL = sgemv_t.c
 
 ZGEMVNKERNEL = zgemv_n_dup.S
-ZGEMVTKERNEL = zgemv_t.S
+ZGEMVTKERNEL = zgemv_t.c
 
 DGEMVNKERNEL = dgemv_n_bulldozer.S
 DGEMVTKERNEL = dgemv_t_bulldozer.S

kernel/x86_64/KERNEL.HASWELL

@@ -1,3 +1,14 @@
+SGEMVNKERNEL = sgemv_n.c
+SGEMVTKERNEL = sgemv_t.c
+
+DGEMVNKERNEL = dgemv_n.c
+DGEMVTKERNEL = dgemv_t.c
+
+ZGEMVNKERNEL = zgemv_n.c
+ZGEMVTKERNEL = zgemv_t.c
+
+CGEMVNKERNEL = cgemv_n.c
+CGEMVTKERNEL = cgemv_t.c
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_haswell.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c

kernel/x86_64/KERNEL.NEHALEM

@@ -1,3 +1,5 @@
+SGEMVNKERNEL = sgemv_n.c
+SGEMVTKERNEL = sgemv_t.c
 
 SGEMMKERNEL    =  gemm_kernel_4x8_nehalem.S
 SGEMMINCOPY    =  gemm_ncopy_4.S

kernel/x86_64/KERNEL.PILEDRIVER

@@ -1,3 +1,5 @@
+SGEMVNKERNEL = sgemv_n.c
+SGEMVTKERNEL = sgemv_t.c
 
 ZGEMVNKERNEL = zgemv_n_dup.S
 ZGEMVTKERNEL = zgemv_t.S

kernel/x86_64/KERNEL.SANDYBRIDGE

@@ -1,3 +1,8 @@
+SGEMVNKERNEL = sgemv_n.c
+SGEMVTKERNEL = sgemv_t.c
+
+ZGEMVNKERNEL = zgemv_n.c
+
 
 SGEMMKERNEL    =  sgemm_kernel_16x4_sandy.S
 SGEMMINCOPY    =  ../generic/gemm_ncopy_16.c

kernel/x86_64/cgemm_kernel_8x2_sandy.S

@@ -25,6 +25,32 @@ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
 USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 **********************************************************************************/
 
+/*********************************************************************
+* 2014/07/29 Saar
+*        BLASTEST               : OK
+*        CTEST                  : OK
+*        TEST                   : OK
+*
+* 2013/10/28 Saar
+* Parameter:
+*       CGEMM_DEFAULT_UNROLL_N  2
+*       CGEMM_DEFAULT_UNROLL_M  8
+*       CGEMM_DEFAULT_P         768
+*       CGEMM_DEFAULT_Q         512
+*       A_PR1                   512
+*       B_PR1                   512
+*
+* 2014/07/29 Saar
+* Performance at 6192x6192x6192:
+*       1 thread:       49 GFLOPS       (MKL:   52)
+*       2 threads:      99 GFLOPS       (MKL:  102)
+*       3 threads:     148 GFLOPS       (MKL:  150)
+*       4 threads:     195 GFLOPS       (MKL:  194)
+*       8 threads:     354 GFLOPS       (MKL:  317)
+*
+*
+*********************************************************************/
+
 
 #define ASSEMBLER
 #include "common.h"
@@ -192,22 +218,108 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 /***************************************************************************************************************************/
 
-.macro KERNEL8x2_SUB
+.macro KERNEL8x2_1
 
         vmovups         -16 * SIZE(AO, %rax, SIZE), %ymm0
         vbroadcastss         -8 * SIZE(BO, BI, SIZE), %ymm4
-        VFMADDPS_YR(        %ymm8,%ymm4,%ymm0  )
         vmovups          -8 * SIZE(AO, %rax, SIZE), %ymm1
-        VFMADDPS_YR(        %ymm12,%ymm4,%ymm1 )
         vbroadcastss         -7 * SIZE(BO, BI, SIZE), %ymm5
+	prefetcht0	A_PR1(AO, %rax, SIZE)
+
+        VFMADDPS_YR(        %ymm8,%ymm4,%ymm0  )
+        vbroadcastss         -6 * SIZE(BO, BI, SIZE), %ymm6
         VFMADDPS_YI(        %ymm9,%ymm5,%ymm0  )
+        vbroadcastss         -5 * SIZE(BO, BI, SIZE), %ymm7
+        VFMADDPS_YR(        %ymm12,%ymm4,%ymm1 )
         VFMADDPS_YI(        %ymm13,%ymm5,%ymm1 )
-        vbroadcastss         -6 * SIZE(BO, BI, SIZE), %ymm6
+
+
+        VFMADDPS_YR(        %ymm10,%ymm6,%ymm0 )
+        vbroadcastss         -4 * SIZE(BO, BI, SIZE), %ymm4
+        VFMADDPS_YI(        %ymm11,%ymm7,%ymm0 )
+        vbroadcastss         -3 * SIZE(BO, BI, SIZE), %ymm5
+        VFMADDPS_YR(        %ymm14,%ymm6,%ymm1 )
+        vmovups           0 * SIZE(AO, %rax, SIZE), %ymm0
+        VFMADDPS_YI(        %ymm15,%ymm7,%ymm1 )
+
+        vmovups           8 * SIZE(AO, %rax, SIZE), %ymm1
+	prefetcht0	A_PR1+64(AO, %rax, SIZE)
+
+        VFMADDPS_YR(        %ymm8,%ymm4,%ymm0  )
+        vbroadcastss         -2 * SIZE(BO, BI, SIZE), %ymm6
+        VFMADDPS_YI(        %ymm9,%ymm5,%ymm0  )
+        vbroadcastss         -1 * SIZE(BO, BI, SIZE), %ymm7
+        VFMADDPS_YR(        %ymm12,%ymm4,%ymm1 )
+        VFMADDPS_YI(        %ymm13,%ymm5,%ymm1 )
+
+
+        VFMADDPS_YR(        %ymm10,%ymm6,%ymm0 )
+        vbroadcastss          0 * SIZE(BO, BI, SIZE), %ymm4
+        VFMADDPS_YI(        %ymm11,%ymm7,%ymm0 )
+        vbroadcastss          1 * SIZE(BO, BI, SIZE), %ymm5
+        VFMADDPS_YR(        %ymm14,%ymm6,%ymm1 )
+        vmovups          16 * SIZE(AO, %rax, SIZE), %ymm0
+        VFMADDPS_YI(        %ymm15,%ymm7,%ymm1 )
+
+        vmovups          24 * SIZE(AO, %rax, SIZE), %ymm1
+	prefetcht0	A_PR1+128(AO, %rax, SIZE)
+
+        VFMADDPS_YR(        %ymm8,%ymm4,%ymm0  )
+        vbroadcastss          2 * SIZE(BO, BI, SIZE), %ymm6
+        VFMADDPS_YI(        %ymm9,%ymm5,%ymm0  )
+        vbroadcastss          3 * SIZE(BO, BI, SIZE), %ymm7
+        VFMADDPS_YR(        %ymm12,%ymm4,%ymm1 )
+        VFMADDPS_YI(        %ymm13,%ymm5,%ymm1 )
+
+
         VFMADDPS_YR(        %ymm10,%ymm6,%ymm0 )
+        vbroadcastss          4 * SIZE(BO, BI, SIZE), %ymm4
+        VFMADDPS_YI(        %ymm11,%ymm7,%ymm0 )
+        vbroadcastss          5 * SIZE(BO, BI, SIZE), %ymm5
         VFMADDPS_YR(        %ymm14,%ymm6,%ymm1 )
+        vmovups          32 * SIZE(AO, %rax, SIZE), %ymm0
+        VFMADDPS_YI(        %ymm15,%ymm7,%ymm1 )
+
+        vmovups          40 * SIZE(AO, %rax, SIZE), %ymm1
+	prefetcht0	A_PR1+192(AO, %rax, SIZE)
+
+        VFMADDPS_YR(        %ymm8,%ymm4,%ymm0  )
+        vbroadcastss          6 * SIZE(BO, BI, SIZE), %ymm6
+        VFMADDPS_YI(        %ymm9,%ymm5,%ymm0  )
+        vbroadcastss          7 * SIZE(BO, BI, SIZE), %ymm7
+        VFMADDPS_YR(        %ymm12,%ymm4,%ymm1 )
+        VFMADDPS_YI(        %ymm13,%ymm5,%ymm1 )
+
+        VFMADDPS_YR(        %ymm10,%ymm6,%ymm0 )
+        VFMADDPS_YI(        %ymm11,%ymm7,%ymm0 )
+        addq    $ 16, BI                           
+        VFMADDPS_YR(        %ymm14,%ymm6,%ymm1 )
+        VFMADDPS_YI(        %ymm15,%ymm7,%ymm1 )
+
+        addq    $ 64, %rax                         
+.endm
+
+
+.macro KERNEL8x2_SUB
+
+        vmovups         -16 * SIZE(AO, %rax, SIZE), %ymm0
+        vmovups          -8 * SIZE(AO, %rax, SIZE), %ymm1
+        vbroadcastss         -8 * SIZE(BO, BI, SIZE), %ymm4
+        vbroadcastss         -7 * SIZE(BO, BI, SIZE), %ymm5
+
+        VFMADDPS_YR(        %ymm8,%ymm4,%ymm0  )
+        vbroadcastss         -6 * SIZE(BO, BI, SIZE), %ymm6
+        VFMADDPS_YI(        %ymm9,%ymm5,%ymm0  )
         vbroadcastss         -5 * SIZE(BO, BI, SIZE), %ymm7
+        VFMADDPS_YR(        %ymm12,%ymm4,%ymm1 )
+        VFMADDPS_YI(        %ymm13,%ymm5,%ymm1 )
+
+
+        VFMADDPS_YR(        %ymm10,%ymm6,%ymm0 )
         VFMADDPS_YI(        %ymm11,%ymm7,%ymm0 )
+        VFMADDPS_YR(        %ymm14,%ymm6,%ymm1 )
         VFMADDPS_YI(        %ymm15,%ymm7,%ymm1 )
+
         addq    $ 4 , BI                           
         addq    $ 16, %rax                         
 .endm
@@ -984,47 +1096,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .L2_8_12:
 
-	prefetcht0	A_PR1(AO,%rax,SIZE)
 	prefetcht0	B_PR1(BO,BI,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
+	KERNEL8x2_1
 
-	prefetcht0	A_PR1(AO,%rax,SIZE)
 	prefetcht0	B_PR1(BO,BI,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
+	KERNEL8x2_1
 
 	je	.L2_8_16
 
-	prefetcht0	A_PR1(AO,%rax,SIZE)
 	prefetcht0	B_PR1(BO,BI,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
+	KERNEL8x2_1
 
-	prefetcht0	A_PR1(AO,%rax,SIZE)
 	prefetcht0	B_PR1(BO,BI,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
-	prefetcht0	A_PR1(AO,%rax,SIZE)
-	KERNEL8x2_SUB
+	KERNEL8x2_1
 
 	je	.L2_8_16
 
@@ -1152,7 +1236,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .L2_4_12:
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL4x2_SUB
 	KERNEL4x2_SUB
 	prefetcht0	A_PR1(AO,%rax,SIZE)
@@ -1160,7 +1243,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	KERNEL4x2_SUB
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL4x2_SUB
 	KERNEL4x2_SUB
 	prefetcht0	A_PR1(AO,%rax,SIZE)
@@ -1170,7 +1252,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	je	.L2_4_16
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL4x2_SUB
 	KERNEL4x2_SUB
 	prefetcht0	A_PR1(AO,%rax,SIZE)
@@ -1178,7 +1259,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	KERNEL4x2_SUB
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL4x2_SUB
 	KERNEL4x2_SUB
 	prefetcht0	A_PR1(AO,%rax,SIZE)
@@ -1305,14 +1385,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .L2_4_22:
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
@@ -1321,14 +1399,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	je	.L2_4_26
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
 	KERNEL2x2_SUB
@@ -1507,13 +1583,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .L2_4_42:
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
 
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
@@ -1522,13 +1596,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	je	.L2_4_46
 
 	prefetcht0	A_PR1(AO,%rax,SIZE)
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
 
-	prefetcht0	B_PR1(BO,BI,SIZE)
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB
 	KERNEL1x2_SUB

kernel/x86_64/cgemv_n.c

@@ -0,0 +1,255 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdlib.h>
+#include <stdio.h>
+#include "common.h"
+
+#if defined(HASWELL)
+#include "cgemv_n_microk_haswell-2.c"
+#endif
+
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void cgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] - a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] - a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] - a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] + a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] + a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] + a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
+#endif
+	}
+}
+	
+#endif
+
+static void cgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+#endif
+
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<2*n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
+{
+	BLASLONG i;
+	FLOAT temp_r;
+	FLOAT temp_i;
+	for ( i=0; i<n; i++ )
+	{
+#if !defined(XCONJ) 
+		temp_r = alpha_r * src[0] - alpha_i * src[1];
+		temp_i = alpha_r * src[1] + alpha_i * src[0];
+#else
+		temp_r =  alpha_r * src[0] + alpha_i * src[1];
+		temp_i = -alpha_r * src[1] + alpha_i * src[0];
+#endif
+
+		*dest += temp_r;
+		*(dest+1) += temp_i;
+
+		src+=2;
+		dest += inc_dest;
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[8],*ybuffer;
+
+
+#if 0
+printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
+#endif
+
+	if ( m < 1 ) return(0);
+	if ( n < 1 ) return(0);
+
+	ybuffer = buffer;
+	
+	inc_x *= 2;
+	inc_y *= 2;
+	lda   *= 2;
+
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[2] = x_ptr[0];
+			xbuffer[3] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[4] = x_ptr[0];
+			xbuffer[5] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[6] = x_ptr[0];
+			xbuffer[7] = x_ptr[1];
+			x_ptr += inc_x;	
+
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			cgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
+			x_ptr += inc_x;	
+			cgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
+		a     += 2 * NB;
+		y_ptr += NB * inc_y;
+	}
+
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+			temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+
+#if !defined(XCONJ) 
+		y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+		y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+		y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+		y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+		y_ptr += inc_y;
+		a+=2;
+		j++;
+	}
+	return(0);
+}
+
+

kernel/x86_64/cgemv_n_microk_haswell-2.c

@@ -0,0 +1,137 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastss	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastss	 4(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastss	 8(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastss	12(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastss	16(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastss	20(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastss	24(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastss	28(%2), %%ymm7                  \n\t"  // imag part x3
+
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+	"prefetcht0      320(%4,%0,4)			\n\t"
+	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
+	"vmovups      32(%4,%0,4), %%ymm9	        \n\t" // 4 complex values form a0
+
+	"prefetcht0      320(%5,%0,4)			\n\t"
+	"vmovups	(%5,%0,4), %%ymm10              \n\t" // 4 complex values form a1
+	"vmovups      32(%5,%0,4), %%ymm11              \n\t" // 4 complex values form a1
+
+	"vmulps      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulps      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulps      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulps      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%6,%0,4)			\n\t"
+	"vmovups	(%6,%0,4), %%ymm8	        \n\t" // 4 complex values form a2
+	"vmovups      32(%6,%0,4), %%ymm9	        \n\t" // 4 complex values form a2
+
+	"vfmadd231ps      %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm11, %%ymm2, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm11, %%ymm3, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%7,%0,4)			\n\t"
+	"vmovups	(%7,%0,4), %%ymm10              \n\t" // 4 complex values form a3
+	"vmovups      32(%7,%0,4), %%ymm11              \n\t" // 4 complex values form a3
+
+	"vfmadd231ps      %%ymm8 , %%ymm4, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm8 , %%ymm5, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm9 , %%ymm4, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm9 , %%ymm5, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"vfmadd231ps      %%ymm10, %%ymm6, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231ps      %%ymm10, %%ymm7, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231ps      %%ymm11, %%ymm6, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231ps      %%ymm11, %%ymm7, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      320(%3,%0,4)			\n\t"
+	"vmovups	  (%3,%0,4),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,4),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddps         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddps         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,4)		        \n\t" // 4 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,4)		        \n\t"	
+
+        "addq		$16, %0	  	 	        \n\t"
+	"subq	        $8 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/cgemv_t.c

@@ -0,0 +1,265 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(HASWELL)
+#include "cgemv_t_microk_haswell-2.c"
+#endif
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void cgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp_r0 = 0.0;
+	FLOAT temp_r1 = 0.0;
+	FLOAT temp_r2 = 0.0;
+	FLOAT temp_r3 = 0.0;
+	FLOAT temp_i0 = 0.0;
+	FLOAT temp_i1 = 0.0;
+	FLOAT temp_i2 = 0.0;
+	FLOAT temp_i3 = 0.0;
+
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r0 += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   - a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   - a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   - a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];		
+#else
+		temp_r0 += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   + a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   + a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   + a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];		
+#endif
+	}
+	y[0] = temp_r0;
+	y[1] = temp_i0;
+	y[2] = temp_r1;
+	y[3] = temp_i1;
+	y[4] = temp_r2;
+	y[5] = temp_i2;
+	y[6] = temp_r3;
+	y[7] = temp_i3;
+}
+	
+#endif
+
+static void cgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp_r = 0.0;
+	FLOAT temp_i = 0.0;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] + a0[i+1]*x[i];		
+#else
+		temp_r += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] - a0[i+1]*x[i];		
+#endif
+	}
+	*y      = temp_r;
+	*(y+1)  = temp_i;
+}
+
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest     = *src;
+                *(dest+1) = *(src+1);
+                dest+=2;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[8];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[8],*xbuffer;
+
+        inc_x *= 2;
+        inc_y *= 2;
+        lda   *= 2;
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			cgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] - alpha_i * ybuffer[3];
+			y_ptr[1] += alpha_r * ybuffer[3] + alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] - alpha_i * ybuffer[5];
+			y_ptr[1] += alpha_r * ybuffer[5] + alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] - alpha_i * ybuffer[7];
+			y_ptr[1] += alpha_r * ybuffer[7] + alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] + alpha_i * ybuffer[3];
+			y_ptr[1] -= alpha_r * ybuffer[3] - alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] + alpha_i * ybuffer[5];
+			y_ptr[1] -= alpha_r * ybuffer[5] - alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] + alpha_i * ybuffer[7];
+			y_ptr[1] -= alpha_r * ybuffer[7] - alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#endif
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			cgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#endif
+
+		}
+		a += 2* NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+
+	x_ptr = x;
+	copy_x(m3,x_ptr,xbuffer,inc_x);
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < m3*2; i+=2 )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[i] * xbuffer[i]   - a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] + a_ptr[i+1] * xbuffer[i];
+#else
+			temp_r += a_ptr[i] * xbuffer[i]   + a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] - a_ptr[i+1] * xbuffer[i];
+#endif
+		}
+		a_ptr += lda;
+
+#if !defined(XCONJ) 
+                y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+                y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+                y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+                y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+

kernel/x86_64/cgemv_t_microk_haswell-2.c

@@ -0,0 +1,171 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary froms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary from must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void cgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
+	"vxorps		%%ymm10, %%ymm10, %%ymm10	\n\t" // temp
+	"vxorps		%%ymm11, %%ymm11, %%ymm11	\n\t" // temp
+	"vxorps		%%ymm12, %%ymm12, %%ymm12	\n\t" // temp
+	"vxorps		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorps		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorps		%%ymm15, %%ymm15, %%ymm15	\n\t"
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+        "prefetcht0      192(%4,%0,4)                   \n\t"
+	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
+        "prefetcht0      192(%5,%0,4)                   \n\t"
+	"vmovups	(%5,%0,4), %%ymm5               \n\t" // 4 complex values from a1
+
+        "prefetcht0      192(%2,%0,4)                   \n\t"
+	"vmovups	    (%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+	
+        "prefetcht0      192(%6,%0,4)                   \n\t"
+	"vmovups	(%6,%0,4), %%ymm6	        \n\t" // 4 complex values from a2
+        "prefetcht0      192(%7,%0,4)                   \n\t"
+	"vmovups	(%7,%0,4), %%ymm7               \n\t" // 4 complex values from a3
+
+	"vfmadd231ps      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+	"vmovups       32(%4,%0,4), %%ymm4	        \n\t" // 2 complex values from a0
+	"vmovups       32(%5,%0,4), %%ymm5              \n\t" // 2 complex values from a1
+
+	"vmovups	  32(%2,%0,4)  , %%ymm6		\n\t" // 4 complex values from x
+	"vpermilps        $0xb1, %%ymm6, %%ymm7		\n\t" // exchange real and imap parts
+	"vblendps $0x55, %%ymm6, %%ymm7, %%ymm0         \n\t" // only the real parts
+	"vblendps $0x55, %%ymm7, %%ymm6, %%ymm1         \n\t" // only the imag parts
+
+	"vmovups       32(%6,%0,4), %%ymm6	        \n\t" // 2 complex values from a2
+	"vmovups       32(%7,%0,4), %%ymm7              \n\t" // 2 complex values from a3
+
+	"vfmadd231ps      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231ps      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231ps      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+        "addq		$16 , %0	  	 	        \n\t"
+	"subq	        $8  , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilps      $0xb1 , %%ymm9 , %%ymm9                \n\t"
+        "vpermilps      $0xb1 , %%ymm11, %%ymm11               \n\t"
+        "vpermilps      $0xb1 , %%ymm13, %%ymm13               \n\t"
+        "vpermilps      $0xb1 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubps      %%ymm9 , %%ymm8, %%ymm8               \n\t" 
+        "vaddsubps      %%ymm11, %%ymm10, %%ymm10             \n\t"
+        "vaddsubps      %%ymm13, %%ymm12, %%ymm12             \n\t"
+        "vaddsubps      %%ymm15, %%ymm14, %%ymm14             \n\t"
+#else
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm10, %%ymm10               \n\t"
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubps      %%ymm8 , %%ymm9 , %%ymm8              \n\t"
+        "vaddsubps      %%ymm10, %%ymm11, %%ymm10             \n\t"
+        "vaddsubps      %%ymm12, %%ymm13, %%ymm12             \n\t"
+        "vaddsubps      %%ymm14, %%ymm15, %%ymm14             \n\t"
+        "vpermilps      $0xb1 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilps      $0xb1 , %%ymm10, %%ymm10               \n\t"
+        "vpermilps      $0xb1 , %%ymm12, %%ymm12               \n\t"
+        "vpermilps      $0xb1 , %%ymm14, %%ymm14               \n\t"
+#endif
+
+	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
+	"vextractf128   $1, %%ymm10, %%xmm11	      	      \n\t"
+	"vextractf128   $1, %%ymm12, %%xmm13		      \n\t"
+	"vextractf128   $1, %%ymm14, %%xmm15		      \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddps		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddps		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+	"vshufpd        $0x1, %%xmm8 , %%xmm8 , %%xmm9   \n\t"
+	"vshufpd        $0x1, %%xmm10, %%xmm10, %%xmm11  \n\t"
+	"vshufpd        $0x1, %%xmm12, %%xmm12, %%xmm13  \n\t"
+	"vshufpd        $0x1, %%xmm14, %%xmm14, %%xmm15  \n\t"
+
+	"vaddps		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddps		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddps		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddps		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+	"vmovsd	%%xmm8 ,   (%3)			\n\t"
+	"vmovsd	%%xmm10,  8(%3)			\n\t"
+	"vmovsd	%%xmm12, 16(%3)			\n\t"
+	"vmovsd	%%xmm14, 24(%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/dgemv_n.c

@@ -0,0 +1,206 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+
+#if defined(HASWELL)
+#include "dgemv_n_microk_haswell-2.c"
+#endif
+
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void dgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+
+	for ( i=0; i< n; i+=4 )
+	{
+		y[i] += a0[i]*x[0] + a1[i]*x[1] + a2[i]*x[2] + a3[i]*x[3];		
+		y[i+1] += a0[i+1]*x[0] + a1[i+1]*x[1] + a2[i+1]*x[2] + a3[i+1]*x[3];		
+		y[i+2] += a0[i+2]*x[0] + a1[i+2]*x[1] + a2[i+2]*x[2] + a3[i+2]*x[3];		
+		y[i+3] += a0[i+3]*x[0] + a1[i+3]*x[1] + a2[i+3]*x[2] + a3[i+3]*x[3];		
+	}
+}
+	
+#endif
+
+static void dgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		y[i] += a0[i]*x[0];		
+		y[i+1] += a0[i+1]*x[0];		
+		y[i+2] += a0[i+2]*x[0];		
+		y[i+3] += a0[i+3]*x[0];		
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
+{
+	BLASLONG i;
+	if ( inc_dest == 1 )
+	{
+		for ( i=0; i<n; i+=4 )
+		{
+			dest[i] += src[i];
+			dest[i+1] += src[i+1];
+			dest[i+2] += src[i+2];
+			dest[i+3] += src[i+3];
+		}
+
+	}
+	else
+	{
+		for ( i=0; i<n; i++ )
+		{
+			*dest += *src;
+			src++;
+			dest += inc_dest;
+		}
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[4],*ybuffer;
+
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
+	ybuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+			xbuffer[0] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[1] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[2] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[3] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			dgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			xbuffer[0] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			dgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y);
+		a     += NB;
+		y_ptr += NB * inc_y;
+	}
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+			temp += a_ptr[0] * x_ptr[0];
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+		y_ptr[0] += alpha * temp;
+		y_ptr += inc_y;
+		a++;
+		j++;
+	}
+	return(0);
+}
+
+

kernel/x86_64/dgemv_n_microk_haswell-2.c

@@ -0,0 +1,89 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void dgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void dgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+	"vbroadcastsd    (%2), %%ymm12	 \n\t"	// x0 
+	"vbroadcastsd   8(%2), %%ymm13	 \n\t"	// x1 
+	"vbroadcastsd  16(%2), %%ymm14	 \n\t"	// x2 
+	"vbroadcastsd  24(%2), %%ymm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 192(%3,%0,8)		 \n\t"
+	"vmovups	(%3,%0,8), %%ymm4	 \n\t"	// 4 * y
+	"vmovups      32(%3,%0,8), %%ymm5	 \n\t"	// 4 * y
+
+	"prefetcht0	 192(%4,%0,8)		       \n\t"
+	"vfmadd231pd   (%4,%0,8), %%ymm12, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%4,%0,8), %%ymm12, %%ymm5      \n\t" 
+	"prefetcht0	 192(%5,%0,8)		       \n\t"
+	"vfmadd231pd   (%5,%0,8), %%ymm13, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%5,%0,8), %%ymm13, %%ymm5      \n\t" 
+	"prefetcht0	 192(%6,%0,8)		       \n\t"
+	"vfmadd231pd   (%6,%0,8), %%ymm14, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%6,%0,8), %%ymm14, %%ymm5      \n\t" 
+	"prefetcht0	 192(%7,%0,8)		       \n\t"
+	"vfmadd231pd   (%7,%0,8), %%ymm15, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%7,%0,8), %%ymm15, %%ymm5      \n\t" 
+
+	"vmovups  %%ymm4,   (%3,%0,8)		      \n\t"	// 4 * y
+	"vmovups  %%ymm5, 32(%3,%0,8)		      \n\t"	// 4 * y
+
+        "addq		$8 , %0	  	 	      \n\t"
+	"subq	        $8 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/dgemv_t.c

@@ -0,0 +1,191 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(HASWELL)
+#include "dgemv_t_microk_haswell-2.c"
+#endif
+
+#define NBMAX 2048
+
+#ifndef HAVE_KERNEL_16x4
+
+static void dgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp0 = 0.0;
+	FLOAT temp1 = 0.0;
+	FLOAT temp2 = 0.0;
+	FLOAT temp3 = 0.0;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		temp0 += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];		
+		temp1 += a1[i]*x[i] + a1[i+1]*x[i+1] + a1[i+2]*x[i+2] + a1[i+3]*x[i+3];		
+		temp2 += a2[i]*x[i] + a2[i+1]*x[i+1] + a2[i+2]*x[i+2] + a2[i+3]*x[i+3];		
+		temp3 += a3[i]*x[i] + a3[i+1]*x[i+1] + a3[i+2]*x[i+2] + a3[i+3]*x[i+3];		
+	}
+	y[0] = temp0;
+	y[1] = temp1;
+	y[2] = temp2;
+	y[3] = temp3;
+}
+	
+#endif
+
+static void dgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp = 0.0;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		temp += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];		
+	}
+	*y = temp;
+}
+	
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest = *src;
+                dest++;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[4],*xbuffer;
+
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			dgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+			*y_ptr += ybuffer[0]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[1]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[2]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[3]*alpha;
+			y_ptr  += inc_y;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			dgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+			*y_ptr += ybuffer[0]*alpha;
+			y_ptr  += inc_y;
+
+		}
+		a += NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+	x_ptr = x;
+	for ( i=0; i< m3; i++ )
+	{
+		xbuffer[i] = *x_ptr;
+		x_ptr += inc_x;
+	}
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp = 0.0;
+		for( i = 0; i < m3; i++ )
+		{
+			temp += a_ptr[i] * xbuffer[i];
+		}
+		a_ptr += lda;
+		y_ptr[0] += alpha * temp;
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+

kernel/x86_64/dgemv_t_microk_haswell-2.c

@@ -0,0 +1,107 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void dgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void dgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			         \n\t"
+	"vxorpd		%%ymm4 , %%ymm4, %%ymm4  \n\t"
+	"vxorpd		%%ymm5 , %%ymm5, %%ymm5  \n\t"
+	"vxorpd		%%ymm6 , %%ymm6, %%ymm6  \n\t"
+	"vxorpd		%%ymm7 , %%ymm7, %%ymm7  \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 384(%2,%0,8)		 \n\t"
+	"vmovups	(%2,%0,8), %%ymm12       \n\t"	// 4 * x
+	"vmovups      32(%2,%0,8), %%ymm13       \n\t"	// 4 * x
+
+	"prefetcht0	 384(%4,%0,8)		       \n\t"
+	"vfmadd231pd   (%4,%0,8), %%ymm12, %%ymm4      \n\t" 
+	"vfmadd231pd   (%5,%0,8), %%ymm12, %%ymm5      \n\t" 
+	"prefetcht0	 384(%5,%0,8)		       \n\t"
+	"vfmadd231pd 32(%4,%0,8), %%ymm13, %%ymm4      \n\t" 
+	"vfmadd231pd 32(%5,%0,8), %%ymm13, %%ymm5      \n\t" 
+	"prefetcht0	 384(%6,%0,8)		       \n\t"
+	"vfmadd231pd   (%6,%0,8), %%ymm12, %%ymm6      \n\t" 
+	"vfmadd231pd   (%7,%0,8), %%ymm12, %%ymm7      \n\t" 
+	"prefetcht0	 384(%7,%0,8)		       \n\t"
+	"vfmadd231pd 32(%6,%0,8), %%ymm13, %%ymm6      \n\t" 
+	"vfmadd231pd 32(%7,%0,8), %%ymm13, %%ymm7      \n\t" 
+
+        "addq		$8 , %0	  	 	      \n\t"
+	"subq	        $8 , %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vextractf128   $1 , %%ymm4, %%xmm12	      \n\t"
+	"vextractf128   $1 , %%ymm5, %%xmm13	      \n\t"
+	"vextractf128   $1 , %%ymm6, %%xmm14	      \n\t"
+	"vextractf128   $1 , %%ymm7, %%xmm15	      \n\t"
+
+	"vaddpd		%%xmm4, %%xmm12, %%xmm4       \n\t"
+	"vaddpd		%%xmm5, %%xmm13, %%xmm5       \n\t"
+	"vaddpd		%%xmm6, %%xmm14, %%xmm6       \n\t"
+	"vaddpd		%%xmm7, %%xmm15, %%xmm7       \n\t"
+
+        "vhaddpd        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddpd        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddpd        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddpd        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vmovsd         %%xmm4,    (%3)         \n\t"
+        "vmovsd         %%xmm5,   8(%3)         \n\t"
+        "vmovsd         %%xmm6,  16(%3)         \n\t"
+        "vmovsd         %%xmm7,  24(%3)         \n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_n.c

@@ -0,0 +1,212 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sgemv_n_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "sgemv_n_microk_haswell-2.c"
+#elif defined(SANDYBRIDGE)
+#include "sgemv_n_microk_sandy-2.c"
+#elif defined(NEHALEM)
+#include "sgemv_n_microk_nehalem-2.c"
+#endif
+
+
+#define NBMAX 4096
+
+#ifndef HAVE_KERNEL_16x4
+
+static void sgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+
+	for ( i=0; i< n; i+=4 )
+	{
+		y[i] += a0[i]*x[0] + a1[i]*x[1] + a2[i]*x[2] + a3[i]*x[3];		
+		y[i+1] += a0[i+1]*x[0] + a1[i+1]*x[1] + a2[i+1]*x[2] + a3[i+1]*x[3];		
+		y[i+2] += a0[i+2]*x[0] + a1[i+2]*x[1] + a2[i+2]*x[2] + a3[i+2]*x[3];		
+		y[i+3] += a0[i+3]*x[0] + a1[i+3]*x[1] + a2[i+3]*x[2] + a3[i+3]*x[3];		
+	}
+}
+	
+#endif
+
+static void sgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		y[i] += a0[i]*x[0];		
+		y[i+1] += a0[i+1]*x[0];		
+		y[i+2] += a0[i+2]*x[0];		
+		y[i+3] += a0[i+3]*x[0];		
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
+{
+	BLASLONG i;
+	if ( inc_dest == 1 )
+	{
+		for ( i=0; i<n; i+=4 )
+		{
+			dest[i] += src[i];
+			dest[i+1] += src[i+1];
+			dest[i+2] += src[i+2];
+			dest[i+3] += src[i+3];
+		}
+
+	}
+	else
+	{
+		for ( i=0; i<n; i++ )
+		{
+			*dest += *src;
+			src++;
+			dest += inc_dest;
+		}
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[4],*ybuffer;
+
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
+	ybuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+			xbuffer[0] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[1] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[2] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			xbuffer[3] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			sgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			xbuffer[0] = alpha * x_ptr[0];
+			x_ptr += inc_x;	
+			sgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y);
+		a     += NB;
+		y_ptr += NB * inc_y;
+	}
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+			temp += a_ptr[0] * x_ptr[0];
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+		y_ptr[0] += alpha * temp;
+		y_ptr += inc_y;
+		a++;
+		j++;
+	}
+	return(0);
+}
+
+

kernel/x86_64/sgemv_n_avx.c

@@ -0,0 +1,218 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sgemv_n_microk_bulldozer.c"
+#elif defined(HASWELL)
+#include "sgemv_n_microk_haswell.c"
+#else
+#include "sgemv_n_microk_sandy.c"
+#endif
+
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest = *src;
+		dest++;
+		src += inc_src;
+	}
+}
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest += *src;
+		src++;
+		dest += inc_dest;
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG register m2;
+	BLASLONG register n2;
+	FLOAT *xbuffer,*ybuffer;
+	xbuffer = buffer;
+	ybuffer = xbuffer + 2048 + 256;
+	
+	n1 = n / 512 ;
+	n2 = n % 512 ;
+
+	m1 = m / 64;
+	m2 = m % 64;
+
+	y_ptr = y;
+	x_ptr = x;
+
+	for (j=0; j<n1; j++)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(512,x_ptr,xbuffer,inc_x);
+
+		a_ptr = a + j * 512 * lda;
+		y_ptr = y;
+
+		for(i = 0; i<m1; i++ )
+		{
+			sgemv_kernel_64(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(64,ybuffer,y_ptr,inc_y);
+			y_ptr += 64 * inc_y;
+			a_ptr += 64;			
+
+		}
+
+		if ( m2 & 32 )
+		{
+			sgemv_kernel_32(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(32,ybuffer,y_ptr,inc_y);
+			y_ptr += 32 * inc_y;
+			a_ptr += 32;			
+
+		}
+
+		if ( m2 & 16 )
+		{
+			sgemv_kernel_16(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(16,ybuffer,y_ptr,inc_y);
+			y_ptr += 16 * inc_y;
+			a_ptr += 16;			
+		}
+		if ( m2 & 8 )
+		{
+			sgemv_kernel_8(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(8,ybuffer,y_ptr,inc_y);
+			y_ptr += 8 * inc_y;
+			a_ptr += 8;			
+		}
+		if ( m2 & 4 )
+		{
+			sgemv_kernel_4(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(4,ybuffer,y_ptr,inc_y);
+			y_ptr += 4 * inc_y;
+			a_ptr += 4;			
+		}
+		if ( m2 & 2 )
+		{
+			sgemv_kernel_2(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(2,ybuffer,y_ptr,inc_y);
+			y_ptr += 2 * inc_y;
+			a_ptr += 2;			
+		}
+		if ( m2 & 1 )
+		{
+			sgemv_kernel_1(512,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(1,ybuffer,y_ptr,inc_y);
+		}
+		x_ptr += 512 * inc_x;
+
+	}
+
+	if ( n2 > 0 )
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(n2,x_ptr,xbuffer,inc_x);
+
+		a_ptr = a + n1 * 512 * lda;
+		y_ptr = y;
+
+		for(i = 0; i<m1; i++ )
+		{
+			sgemv_kernel_64(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(64,ybuffer,y_ptr,inc_y);
+			y_ptr += 64 * inc_y;
+			a_ptr += 64;			
+
+		}
+
+		if ( m2 & 32 )
+		{
+			sgemv_kernel_32(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(32,ybuffer,y_ptr,inc_y);
+			y_ptr += 32 * inc_y;
+			a_ptr += 32;			
+
+		}
+		if ( m2 & 16 )
+		{
+			sgemv_kernel_16(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(16,ybuffer,y_ptr,inc_y);
+			y_ptr += 16 * inc_y;
+			a_ptr += 16;			
+		}
+		if ( m2 & 8 )
+		{
+			sgemv_kernel_8(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(8,ybuffer,y_ptr,inc_y);
+			y_ptr += 8 * inc_y;
+			a_ptr += 8;			
+		}
+		if ( m2 & 4 )
+		{
+			sgemv_kernel_4(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(4,ybuffer,y_ptr,inc_y);
+			y_ptr += 4 * inc_y;
+			a_ptr += 4;			
+		}
+		if ( m2 & 2 )
+		{
+			sgemv_kernel_2(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(2,ybuffer,y_ptr,inc_y);
+			y_ptr += 2 * inc_y;
+			a_ptr += 2;			
+		}
+		if ( m2 & 1 )
+		{
+			sgemv_kernel_1(n2,alpha,a_ptr,lda,xbuffer,ybuffer);
+			add_y(1,ybuffer,y_ptr,inc_y);
+		}
+
+
+	}
+	return(0);
+}
+
+

kernel/x86_64/sgemv_n_microk_bulldozer-2.c

@@ -0,0 +1,99 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vbroadcastss    (%2), %%xmm12	 \n\t"	// x0 
+	"vbroadcastss   4(%2), %%xmm13	 \n\t"	// x1 
+	"vbroadcastss   8(%2), %%xmm14	 \n\t"	// x2 
+	"vbroadcastss  12(%2), %%xmm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vmovups	(%3,%0,4), %%xmm4	 \n\t"	// 4 * y
+	"vmovups      16(%3,%0,4), %%xmm5	 \n\t"	// 4 * y
+	"vmovups      32(%3,%0,4), %%xmm6	 \n\t"	// 4 * y
+	"vmovups      48(%3,%0,4), %%xmm7	 \n\t"	// 4 * y
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%4,%0,4), %%xmm12, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%4,%0,4), %%xmm12, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 32(%4,%0,4), %%xmm12, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%4,%0,4), %%xmm12, %%xmm7 \n\t" 
+	"prefetcht0	 192(%5,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%5,%0,4), %%xmm13, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm13, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 32(%5,%0,4), %%xmm13, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%5,%0,4), %%xmm13, %%xmm7 \n\t" 
+	"prefetcht0	 192(%6,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%6,%0,4), %%xmm14, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%6,%0,4), %%xmm14, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 32(%6,%0,4), %%xmm14, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%6,%0,4), %%xmm14, %%xmm7 \n\t" 
+	"prefetcht0	 192(%7,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%7,%0,4), %%xmm15, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%7,%0,4), %%xmm15, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 32(%7,%0,4), %%xmm15, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%7,%0,4), %%xmm15, %%xmm7 \n\t" 
+
+	"vmovups  %%xmm4,   (%3,%0,4)		      \n\t"	// 4 * y
+	"vmovups  %%xmm5, 16(%3,%0,4)		      \n\t"	// 4 * y
+	"vmovups  %%xmm6, 32(%3,%0,4)		      \n\t"	// 4 * y
+	"vmovups  %%xmm7, 48(%3,%0,4)		      \n\t"	// 4 * y
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_n_microk_bulldozer.c

@@ -0,0 +1,451 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+static void  sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*3;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
+	"vxorps		%%ymm14, %%ymm14, %%ymm14\n\t"	// set to zero
+	"vxorps		%%ymm15, %%ymm15, %%ymm15\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%ymm8 ,   0*4(%%rsi), %%ymm0, %%ymm8 \n\t" // multiply a and c and add to temp
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%ymm9 ,   8*4(%%rsi), %%ymm0, %%ymm9 \n\t" // multiply a and c and add to temp
+	"prefetcht0   128(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%ymm10,  16*4(%%rsi), %%ymm0, %%ymm10\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm11,  24*4(%%rsi), %%ymm0, %%ymm11\n\t" // multiply a and c and add to temp
+	"prefetcht0   192(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%ymm12,  32*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm13,  40*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm14,  48*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm15,  56*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
+	"vmulps		%%ymm14, %%ymm1,  %%ymm14\n\t"  // scale by alpha
+	"vmulps		%%ymm15, %%ymm1,  %%ymm15\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm12, 32*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13, 40*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm14, 48*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm15, 56*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*3;
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%xmm8 , %%xmm8 , %%xmm8 \n\t"	// set to zero
+	"vxorps		%%xmm9 , %%xmm9 , %%xmm9 \n\t"	// set to zero
+	"vxorps		%%xmm10, %%xmm10, %%xmm10\n\t"	// set to zero
+	"vxorps		%%xmm11, %%xmm11, %%xmm11\n\t"	// set to zero
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+	"vxorps		%%xmm14, %%xmm14, %%xmm14\n\t"	// set to zero
+	"vxorps		%%xmm15, %%xmm15, %%xmm15\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%xmm8 ,   0*4(%%rsi), %%xmm0, %%xmm8 \n\t" // multiply a and c and add to temp
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+	"vfmaddps %%xmm9 ,   4*4(%%rsi), %%xmm0, %%xmm9 \n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm10,   8*4(%%rsi), %%xmm0, %%xmm10\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm11,  12*4(%%rsi), %%xmm0, %%xmm11\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm12,  16*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm13,  20*4(%%rsi), %%xmm0, %%xmm13\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm14,  24*4(%%rsi), %%xmm0, %%xmm14\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm15,  28*4(%%rsi), %%xmm0, %%xmm15\n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%xmm8 , %%xmm1,  %%xmm8 \n\t"  // scale by alpha
+	"vmulps		%%xmm9 , %%xmm1,  %%xmm9 \n\t"  // scale by alpha
+	"vmulps		%%xmm10, %%xmm1,  %%xmm10\n\t"  // scale by alpha
+	"vmulps		%%xmm11, %%xmm1,  %%xmm11\n\t"  // scale by alpha
+	"vmulps		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+	"vmulps		%%xmm13, %%xmm1,  %%xmm13\n\t"  // scale by alpha
+	"vmulps		%%xmm14, %%xmm1,  %%xmm14\n\t"  // scale by alpha
+	"vmulps		%%xmm15, %%xmm1,  %%xmm15\n\t"  // scale by alpha
+
+	"vmovups	%%xmm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm9 ,  4*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm10,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm11, 12*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm12, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm13, 20*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm14, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%xmm15, 28*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	);
+
+} 
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	float *pre = a + lda*3;
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+
+	"vfmaddps %%ymm12,   0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%ymm13,   8*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
+
+	"vmovups	%%ymm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13,  8*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+static void  sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vfmaddps %%ymm12,   0*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+
+	"vmovups	%%ymm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+static void  sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vfmaddps %%xmm12,   0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovups	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+static void  sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vfmaddss %%xmm12,   0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
+	"vfmaddss %%xmm13,   1*4(%%rsi), %%xmm0, %%xmm13\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+	"vmulss		%%xmm13, %%xmm1,  %%xmm13\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovss 	%%xmm13,    4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vfmaddss %%xmm12,   0*4(%%rsi), %%xmm0, %%xmm12\n\t" // multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_n_microk_haswell-2.c

@@ -0,0 +1,88 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+	"vbroadcastss    (%2), %%ymm12	 \n\t"	// x0 
+	"vbroadcastss   4(%2), %%ymm13	 \n\t"	// x1 
+	"vbroadcastss   8(%2), %%ymm14	 \n\t"	// x2 
+	"vbroadcastss  12(%2), %%ymm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vmovups	(%3,%0,4), %%ymm4	 \n\t"	// 8 * y
+	"vmovups      32(%3,%0,4), %%ymm5	 \n\t"	// 8 * y
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+	"vfmadd231ps   (%4,%0,4), %%ymm12, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%4,%0,4), %%ymm12, %%ymm5      \n\t" 
+	"prefetcht0	 192(%5,%0,4)		       \n\t"
+	"vfmadd231ps   (%5,%0,4), %%ymm13, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5      \n\t" 
+	"prefetcht0	 192(%6,%0,4)		       \n\t"
+	"vfmadd231ps   (%6,%0,4), %%ymm14, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%6,%0,4), %%ymm14, %%ymm5      \n\t" 
+	"prefetcht0	 192(%7,%0,4)		       \n\t"
+	"vfmadd231ps   (%7,%0,4), %%ymm15, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%7,%0,4), %%ymm15, %%ymm5      \n\t" 
+
+	"vmovups  %%ymm4,   (%3,%0,4)		      \n\t"	// 8 * y
+	"vmovups  %%ymm5, 32(%3,%0,4)		      \n\t"	// 8 * y
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_n_microk_haswell.c

@@ -0,0 +1,461 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+static void  sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*2;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
+	"vxorps		%%ymm14, %%ymm14, %%ymm14\n\t"	// set to zero
+	"vxorps		%%ymm15, %%ymm15, %%ymm15\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"vfmadd231ps   0*4(%%rsi), %%ymm0, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vfmadd231ps   8*4(%%rsi), %%ymm0, %%ymm9 \n\t" // multiply a and c and add to temp
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+	"vfmadd231ps  16*4(%%rsi), %%ymm0, %%ymm10\n\t" // multiply a and c and add to temp
+	"vfmadd231ps  24*4(%%rsi), %%ymm0, %%ymm11\n\t" // multiply a and c and add to temp
+	"prefetcht0  128(%%r8)\n\t"			// Prefetch
+	"vfmadd231ps  32*4(%%rsi), %%ymm0, %%ymm12\n\t" // multiply a and c and add to temp
+	"vfmadd231ps  40*4(%%rsi), %%ymm0, %%ymm13\n\t" // multiply a and c and add to temp
+	"prefetcht0  192(%%r8)\n\t"			// Prefetch
+	"vfmadd231ps  48*4(%%rsi), %%ymm0, %%ymm14\n\t" // multiply a and c and add to temp
+	"vfmadd231ps  56*4(%%rsi), %%ymm0, %%ymm15\n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
+	"vmulps		%%ymm14, %%ymm1,  %%ymm14\n\t"  // scale by alpha
+	"vmulps		%%ymm15, %%ymm1,  %%ymm15\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm12, 32*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13, 40*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm14, 48*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm15, 56*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*3;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"vmulps  16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
+	"vmulps  24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
+
+
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+
+} 
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	float *pre = a + lda*3;
+	
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+} 
+
+
+static void  sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+	
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+} 
+
+
+static void  sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vaddps %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovups	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+static void  sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vmulps   1*4(%%rsi), %%xmm0, %%xmm5 \n\t" 		// multiply a and c and add to temp
+
+	"vaddps %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+	"vaddps %%xmm13, %%xmm5, %%xmm13     \n\t" 		// multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+	"vmulss		%%xmm13, %%xmm1,  %%xmm13\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovss 	%%xmm13,    4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vmulss   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vaddss %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_n_microk_nehalem-2.c

@@ -0,0 +1,144 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"movss    (%2), %%xmm12	 \n\t"	// x0 
+	"movss   4(%2), %%xmm13	 \n\t"	// x1 
+	"movss   8(%2), %%xmm14	 \n\t"	// x2 
+	"movss  12(%2), %%xmm15	 \n\t"	// x3 
+	"shufps $0,  %%xmm12, %%xmm12\n\t"	
+	"shufps $0,  %%xmm13, %%xmm13\n\t"	
+	"shufps $0,  %%xmm14, %%xmm14\n\t"	
+	"shufps $0,  %%xmm15, %%xmm15\n\t"	
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"movups	       (%3,%0,4), %%xmm4	 \n\t"	// 4 * y
+	"movups      16(%3,%0,4), %%xmm5	 \n\t"	// 4 * y
+	"movups      32(%3,%0,4), %%xmm6	 \n\t"	// 4 * y
+	"movups      48(%3,%0,4), %%xmm7	 \n\t"	// 4 * y
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+
+	"movups             (%4,%0,4), %%xmm8          \n\t" 
+	"movups           16(%4,%0,4), %%xmm9          \n\t" 
+	"movups           32(%4,%0,4), %%xmm10         \n\t" 
+	"movups           48(%4,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm4		       \n\t"
+	"mulps		%%xmm12, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"mulps		%%xmm12, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm6		       \n\t"
+	"mulps		%%xmm12, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+	"prefetcht0	 192(%5,%0,4)		       \n\t"
+
+	"movups             (%5,%0,4), %%xmm8          \n\t" 
+	"movups           16(%5,%0,4), %%xmm9          \n\t" 
+	"movups           32(%5,%0,4), %%xmm10         \n\t" 
+	"movups           48(%5,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm13, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm4		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"mulps		%%xmm13, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm6		       \n\t"
+	"mulps		%%xmm13, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+	"prefetcht0	 192(%6,%0,4)		       \n\t"
+
+	"movups             (%6,%0,4), %%xmm8          \n\t" 
+	"movups           16(%6,%0,4), %%xmm9          \n\t" 
+	"movups           32(%6,%0,4), %%xmm10         \n\t" 
+	"movups           48(%6,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm14, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm4		       \n\t"
+	"mulps		%%xmm14, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm6		       \n\t"
+	"mulps		%%xmm14, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+	"prefetcht0	 192(%7,%0,4)		       \n\t"
+
+	"movups             (%7,%0,4), %%xmm8          \n\t" 
+	"movups           16(%7,%0,4), %%xmm9          \n\t" 
+	"movups           32(%7,%0,4), %%xmm10         \n\t" 
+	"movups           48(%7,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm15, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm4		       \n\t"
+	"mulps		%%xmm15, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"mulps		%%xmm15, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm6		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+
+	"movups  %%xmm4,   (%3,%0,4)		      \n\t"	// 4 * y
+	"movups  %%xmm5, 16(%3,%0,4)		      \n\t"	// 4 * y
+	"movups  %%xmm6, 32(%3,%0,4)		      \n\t"	// 4 * y
+	"movups  %%xmm7, 48(%3,%0,4)		      \n\t"	// 4 * y
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_n_microk_sandy-2.c

@@ -0,0 +1,97 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+	"vbroadcastss    (%2), %%ymm12	 \n\t"	// x0 
+	"vbroadcastss   4(%2), %%ymm13	 \n\t"	// x1 
+	"vbroadcastss   8(%2), %%ymm14	 \n\t"	// x2 
+	"vbroadcastss  12(%2), %%ymm15	 \n\t"	// x3 
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vmovups	(%3,%0,4), %%ymm4	 \n\t"	// 8 * y
+	"vmovups      32(%3,%0,4), %%ymm5	 \n\t"	// 8 * y
+
+	"prefetcht0	 192(%4,%0,4)		       \n\t"
+	"vmulps   (%4,%0,4), %%ymm12, %%ymm8      \n\t" 
+	"vaddps	  %%ymm4, %%ymm8 , %%ymm4	  \n\t"
+	"vmulps 32(%4,%0,4), %%ymm12, %%ymm9      \n\t" 
+	"vaddps	  %%ymm5, %%ymm9 , %%ymm5	  \n\t"
+	"prefetcht0	 192(%5,%0,4)		       \n\t"
+	"vmulps   (%5,%0,4), %%ymm13, %%ymm10     \n\t" 
+	"vaddps	  %%ymm4, %%ymm10, %%ymm4	  \n\t"
+	"vmulps 32(%5,%0,4), %%ymm13, %%ymm11     \n\t" 
+	"vaddps	  %%ymm5, %%ymm11, %%ymm5	  \n\t"
+	"prefetcht0	 192(%6,%0,4)		       \n\t"
+	"vmulps   (%6,%0,4), %%ymm14, %%ymm8      \n\t" 
+	"vaddps	  %%ymm4, %%ymm8 , %%ymm4	  \n\t"
+	"vmulps 32(%6,%0,4), %%ymm14, %%ymm9      \n\t" 
+	"vaddps	  %%ymm5, %%ymm9 , %%ymm5	  \n\t"
+	"prefetcht0	 192(%7,%0,4)		       \n\t"
+	"vmulps   (%7,%0,4), %%ymm15, %%ymm10     \n\t" 
+	"vaddps	  %%ymm4, %%ymm10, %%ymm4	  \n\t"
+	"vmulps 32(%7,%0,4), %%ymm15, %%ymm11     \n\t" 
+	"vaddps	  %%ymm5, %%ymm11, %%ymm5	  \n\t"
+
+	"vmovups  %%ymm4,   (%3,%0,4)		      \n\t"	// 8 * y
+	"vmovups  %%ymm5, 32(%3,%0,4)		      \n\t"	// 8 * y
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_n_microk_sandy.c

@@ -0,0 +1,473 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+static void  sgemv_kernel_64( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*2;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	"vxorps		%%ymm12, %%ymm12, %%ymm12\n\t"	// set to zero
+	"vxorps		%%ymm13, %%ymm13, %%ymm13\n\t"	// set to zero
+	"vxorps		%%ymm14, %%ymm14, %%ymm14\n\t"	// set to zero
+	"vxorps		%%ymm15, %%ymm15, %%ymm15\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+	"vmulps  16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
+	"vmulps  24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+	"prefetcht0  128(%%r8)\n\t"			// Prefetch
+	"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
+
+	"prefetcht0  192(%%r8)\n\t"			// Prefetch
+	"vmulps  32*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps  40*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"vmulps  48*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
+	"vmulps  56*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm12, %%ymm4, %%ymm12\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm13, %%ymm5, %%ymm13\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm14, %%ymm6, %%ymm14\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm15, %%ymm7, %%ymm15\n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+	"vmulps		%%ymm12, %%ymm1,  %%ymm12\n\t"  // scale by alpha
+	"vmulps		%%ymm13, %%ymm1,  %%ymm13\n\t"  // scale by alpha
+	"vmulps		%%ymm14, %%ymm1,  %%ymm14\n\t"  // scale by alpha
+	"vmulps		%%ymm15, %%ymm1,  %%ymm15\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm12, 32*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm13, 40*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm14, 48*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm15, 56*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_32( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	float *pre = a + lda*3;
+
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	"vxorps		%%ymm10, %%ymm10, %%ymm10\n\t"	// set to zero
+	"vxorps		%%ymm11, %%ymm11, %%ymm11\n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+	"vmulps  16*4(%%rsi), %%ymm0, %%ymm6 \n\t" // multiply a and c and add to temp
+	"vmulps  24*4(%%rsi), %%ymm0, %%ymm7 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm10, %%ymm6, %%ymm10\n\t" // multiply a and c and add to temp
+	"vaddps %%ymm11, %%ymm7, %%ymm11\n\t" // multiply a and c and add to temp
+
+
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+	"vmulps		%%ymm10, %%ymm1,  %%ymm10\n\t"  // scale by alpha
+	"vmulps		%%ymm11, %%ymm1,  %%ymm11\n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm10, 16*4(%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm11, 24*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+
+} 
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	float *pre = a + lda*3;
+	
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+	"movq		%6,	 %%r8\n\t"		// address for prefetch
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+	"prefetcht0   64(%%r8)\n\t"			// Prefetch
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	"vxorps		%%ymm9 , %%ymm9 , %%ymm9 \n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+	"nop					 \n\t"
+	"leaq     (%%r8 , %%rcx, 4), %%r8	 \n\t"	// add lda to pointer for prefetch
+
+	"prefetcht0	(%%r8)\n\t"			// Prefetch
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vmulps   8*4(%%rsi), %%ymm0, %%ymm5 \n\t" // multiply a and c and add to temp
+
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm9 , %%ymm5, %%ymm9 \n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmulps		%%ymm9 , %%ymm1,  %%ymm9 \n\t"  // scale by alpha
+
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovups	%%ymm9 ,  8*4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y),      // 5
+	  "m" (pre)	// 6
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+} 
+
+
+static void  sgemv_kernel_8( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+	
+	__asm__  __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%ymm1\n\t"		// alpha -> ymm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%ymm8 , %%ymm8 , %%ymm8 \n\t"	// set to zero
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%ymm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%ymm0, %%ymm4 \n\t" // multiply a and c and add to temp
+	"vaddps %%ymm8 , %%ymm4, %%ymm8 \n\t" // multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%ymm8 , %%ymm1,  %%ymm8 \n\t"  // scale by alpha
+	"vmovups	%%ymm8 ,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "memory"
+	);
+
+
+} 
+
+
+static void  sgemv_kernel_4( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vbroadcastss   %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vbroadcastss	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vaddps %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulps		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovups	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+static void  sgemv_kernel_2( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+
+	"vmulps   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vmulps   1*4(%%rsi), %%xmm0, %%xmm5 \n\t" 		// multiply a and c and add to temp
+
+	"vaddps %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+	"vaddps %%xmm13, %%xmm5, %%xmm13     \n\t" 		// multiply a and c and add to temp
+
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+	"vmulss		%%xmm13, %%xmm1,  %%xmm13\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+	"vmovss 	%%xmm13,    4(%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+
+static void  sgemv_kernel_1( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+
+	".L01LOOP%=:				 \n\t"
+	"vmovss      	(%%rdi),   %%xmm0	 \n\t"	// load values of c
+        "addq		$4     ,   %%rdi	 \n\t"  // increment pointer of c 
+
+	"vmulss   0*4(%%rsi), %%xmm0, %%xmm4 \n\t" 		// multiply a and c and add to temp
+	"vaddss %%xmm12, %%xmm4, %%xmm12     \n\t" 		// multiply a and c and add to temp
+
+	"leaq     (%%rsi, %%rcx, 4), %%rsi	 \n\t"	// add lda to pointer of a
+
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vmulss		%%xmm12, %%xmm1,  %%xmm12\n\t"  // scale by alpha
+
+	"vmovss 	%%xmm12,     (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_t.c

@@ -0,0 +1,197 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sgemv_t_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "sgemv_t_microk_haswell-2.c"
+#elif defined(SANDYBRIDGE)
+#include "sgemv_t_microk_sandy-2.c"
+#elif defined(NEHALEM)
+#include "sgemv_t_microk_nehalem-2.c"
+#endif
+
+#define NBMAX 4096
+
+#ifndef HAVE_KERNEL_16x4
+
+static void sgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp0 = 0.0;
+	FLOAT temp1 = 0.0;
+	FLOAT temp2 = 0.0;
+	FLOAT temp3 = 0.0;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		temp0 += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];		
+		temp1 += a1[i]*x[i] + a1[i+1]*x[i+1] + a1[i+2]*x[i+2] + a1[i+3]*x[i+3];		
+		temp2 += a2[i]*x[i] + a2[i+1]*x[i+1] + a2[i+2]*x[i+2] + a2[i+3]*x[i+3];		
+		temp3 += a3[i]*x[i] + a3[i+1]*x[i+1] + a3[i+2]*x[i+2] + a3[i+3]*x[i+3];		
+	}
+	y[0] = temp0;
+	y[1] = temp1;
+	y[2] = temp2;
+	y[3] = temp3;
+}
+	
+#endif
+
+static void sgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp = 0.0;
+
+	for ( i=0; i< n; i+=4 )
+	{
+		temp += a0[i]*x[i] + a0[i+1]*x[i+1] + a0[i+2]*x[i+2] + a0[i+3]*x[i+3];		
+	}
+	*y = temp;
+}
+	
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest = *src;
+                dest++;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[4],*xbuffer;
+
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			sgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+			*y_ptr += ybuffer[0]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[1]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[2]*alpha;
+			y_ptr  += inc_y;
+			*y_ptr += ybuffer[3]*alpha;
+			y_ptr  += inc_y;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			sgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+			*y_ptr += ybuffer[0]*alpha;
+			y_ptr  += inc_y;
+
+		}
+		a += NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+	x_ptr = x;
+	for ( i=0; i< m3; i++ )
+	{
+		xbuffer[i] = *x_ptr;
+		x_ptr += inc_x;
+	}
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp = 0.0;
+		for( i = 0; i < m3; i++ )
+		{
+			temp += a_ptr[i] * xbuffer[i];
+		}
+		a_ptr += lda;
+		y_ptr[0] += alpha * temp;
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+

kernel/x86_64/sgemv_t_avx.c

@@ -0,0 +1,232 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "sgemv_t_microk_bulldozer.c"
+#elif defined(HASWELL)
+#include "sgemv_t_microk_haswell.c"
+#else
+#include "sgemv_t_microk_sandy.c"
+#endif
+
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+	BLASLONG i;
+	for ( i=0; i<n; i++ )
+	{
+		*dest = *src;
+		dest++;
+		src += inc_src;
+	}
+}
+
+static void  sgemv_kernel_1( BLASLONG n, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, FLOAT *y)
+{
+
+	FLOAT register temp0 = 0.0;
+	BLASLONG i;
+	for ( i=0; i<n ; i++)
+	{
+		temp0 += a[i] * x[i];
+	}
+	temp0 *= alpha ;
+	*y += temp0;
+}
+
+
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *a_ptrl;
+	BLASLONG m1;
+	BLASLONG register m2;
+	FLOAT *xbuffer;
+	xbuffer = buffer;
+	BLASLONG register Mblock;
+
+	m1 = m / 1024 ;
+	m2 = m % 1024 ;
+
+	x_ptr = x;
+	a_ptr = a;
+
+	for (j=0; j<m1; j++)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(1024,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_16(1024,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += 1024;	
+		x_ptr += 1024 * inc_x;
+	}
+
+	if ( m2 == 0 ) return(0);
+
+	Mblock = 512;
+	while ( Mblock >= 16 )
+	{
+	  if ( m2 & Mblock)
+	  {
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(Mblock,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_16(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += Mblock;	
+		x_ptr += Mblock * inc_x;
+
+
+ 	  }
+	  Mblock /= 2;
+
+	}
+
+        if ( m2 & Mblock)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(Mblock,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += Mblock;	
+		x_ptr += Mblock * inc_x;
+
+
+ 	}
+	Mblock /= 2;
+
+
+        if ( m2 & Mblock)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(Mblock,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += Mblock;	
+		x_ptr += Mblock * inc_x;
+
+
+ 	}
+	Mblock /= 2;
+
+        if ( m2 & Mblock)
+	{
+
+		if ( inc_x == 1 )
+			xbuffer = x_ptr;
+		else
+			copy_x(Mblock,x_ptr,xbuffer,inc_x);
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+		a_ptr += Mblock;	
+		x_ptr += Mblock * inc_x;
+
+
+ 	}
+	Mblock /= 2;
+
+        if ( m2 & Mblock)
+	{
+
+		xbuffer = x_ptr;
+
+		y_ptr = y;
+		a_ptrl = a_ptr;
+
+		for(i = 0; i<n; i++ )
+		{
+			sgemv_kernel_1(Mblock,alpha,a_ptrl,lda,xbuffer,y_ptr);
+			y_ptr += inc_y;
+			a_ptrl += lda;
+		}
+
+
+ 	}
+
+	return(0);
+}
+
+

kernel/x86_64/sgemv_t_microk_bulldozer-2.c

@@ -0,0 +1,109 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vxorps		%%xmm4, %%xmm4, %%xmm4	 \n\t"
+	"vxorps		%%xmm5, %%xmm5, %%xmm5	 \n\t"
+	"vxorps		%%xmm6, %%xmm6, %%xmm6	 \n\t"
+	"vxorps		%%xmm7, %%xmm7, %%xmm7	 \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+        "vmovups        (%2,%0,4), %%xmm12       \n\t"  // 4 * x
+
+	"prefetcht0	 384(%4,%0,4)		       \n\t"
+	"vfmaddps %%xmm4,   (%4,%0,4), %%xmm12, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5,   (%5,%0,4), %%xmm12, %%xmm5 \n\t" 
+        "vmovups      16(%2,%0,4), %%xmm13       \n\t"  // 4 * x
+	"vfmaddps %%xmm6,   (%6,%0,4), %%xmm12, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7,   (%7,%0,4), %%xmm12, %%xmm7 \n\t" 
+	"prefetcht0	 384(%5,%0,4)		       \n\t"
+	"vfmaddps %%xmm4, 16(%4,%0,4), %%xmm13, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 16(%5,%0,4), %%xmm13, %%xmm5 \n\t" 
+        "vmovups      32(%2,%0,4), %%xmm14       \n\t"  // 4 * x
+	"vfmaddps %%xmm6, 16(%6,%0,4), %%xmm13, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 16(%7,%0,4), %%xmm13, %%xmm7 \n\t" 
+	"prefetcht0	 384(%6,%0,4)		       \n\t"
+	"vfmaddps %%xmm4, 32(%4,%0,4), %%xmm14, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 32(%5,%0,4), %%xmm14, %%xmm5 \n\t" 
+        "vmovups      48(%2,%0,4), %%xmm15       \n\t"  // 4 * x
+	"vfmaddps %%xmm6, 32(%6,%0,4), %%xmm14, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 32(%7,%0,4), %%xmm14, %%xmm7 \n\t" 
+	"prefetcht0	 384(%7,%0,4)		       \n\t"
+	"vfmaddps %%xmm4, 48(%4,%0,4), %%xmm15, %%xmm4 \n\t" 
+	"vfmaddps %%xmm5, 48(%5,%0,4), %%xmm15, %%xmm5 \n\t" 
+	"vfmaddps %%xmm6, 48(%6,%0,4), %%xmm15, %%xmm6 \n\t" 
+	"vfmaddps %%xmm7, 48(%7,%0,4), %%xmm15, %%xmm7 \n\t" 
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+	"vhaddps        %%xmm5, %%xmm5, %%xmm5	\n\t"
+	"vhaddps        %%xmm6, %%xmm6, %%xmm6	\n\t"
+	"vhaddps        %%xmm7, %%xmm7, %%xmm7	\n\t"
+
+	"vhaddps        %%xmm4, %%xmm4, %%xmm4	\n\t"
+	"vhaddps        %%xmm5, %%xmm5, %%xmm5	\n\t"
+	"vhaddps        %%xmm6, %%xmm6, %%xmm6	\n\t"
+	"vhaddps        %%xmm7, %%xmm7, %%xmm7	\n\t"
+
+	"vmovss		%%xmm4,    (%3)		\n\t"
+	"vmovss		%%xmm5,   4(%3)		\n\t"
+	"vmovss		%%xmm6,   8(%3)		\n\t"
+	"vmovss		%%xmm7,  12(%3)		\n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", "%xmm7", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_t_microk_bulldozer.c

@@ -0,0 +1,99 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	//n = n / 16;
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"leaq	(, %%rcx,4), %%rcx       \n\t"		// scale lda by size of float
+	"leaq	(%%rsi,%%rcx,1), %%r8    \n\t"		// pointer to next line
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+	"vxorps		%%xmm14, %%xmm14, %%xmm14\n\t"	// set to zero
+	"vxorps		%%xmm15, %%xmm15, %%xmm15\n\t"	// set to zero
+
+	"sarq		$4, %%rax		 \n\t"	// n = n / 16
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	// "prefetcht0	512(%%rsi)		 \n\t"
+	"prefetcht0	(%%r8)		 	 \n\t" //prefetch next line of a
+	"vmovups	(%%rsi), %%xmm4		 \n\t"
+	"vmovups     4*4(%%rsi), %%xmm5		 \n\t"
+	"vmovups     8*4(%%rsi), %%xmm6		 \n\t"
+	"vmovups    12*4(%%rsi), %%xmm7		 \n\t"
+
+	"vfmaddps %%xmm12,   0*4(%%rdi), %%xmm4, %%xmm12\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm13,   4*4(%%rdi), %%xmm5, %%xmm13\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm14,   8*4(%%rdi), %%xmm6, %%xmm14\n\t" // multiply a and c and add to temp
+	"vfmaddps %%xmm15,  12*4(%%rdi), %%xmm7, %%xmm15\n\t" // multiply a and c and add to temp
+
+        "addq		$16*4    ,   %%r8	 \n\t"  // increment prefetch pointer 
+        "addq		$16*4    ,   %%rsi	 \n\t"  // increment pointer of a 
+        "addq		$16*4    ,   %%rdi	 \n\t"  // increment pointer of c 
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vaddps		%%xmm12, %%xmm14, %%xmm12\n\t"	
+	"vaddps		%%xmm13, %%xmm15, %%xmm13\n\t"	
+	"vaddps		%%xmm12, %%xmm13, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+
+	"vfmaddss	(%%rdx), %%xmm12, %%xmm1, %%xmm12\n\t"
+	"vmovss		%%xmm12, (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+

kernel/x86_64/sgemv_t_microk_haswell-2.c

@@ -0,0 +1,112 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			         \n\t"
+	"vxorps		%%ymm4 , %%ymm4, %%ymm4  \n\t"
+	"vxorps		%%ymm5 , %%ymm5, %%ymm5  \n\t"
+	"vxorps		%%ymm6 , %%ymm6, %%ymm6  \n\t"
+	"vxorps		%%ymm7 , %%ymm7, %%ymm7  \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 384(%2,%0,4)		 \n\t"
+	"vmovups	(%2,%0,4), %%ymm12       \n\t"	// 8 * x
+	"vmovups      32(%2,%0,4), %%ymm13       \n\t"	// 8 * x
+
+	"prefetcht0	 384(%4,%0,4)		       \n\t"
+	"vfmadd231ps   (%4,%0,4), %%ymm12, %%ymm4      \n\t" 
+	"vfmadd231ps   (%5,%0,4), %%ymm12, %%ymm5      \n\t" 
+	"prefetcht0	 384(%5,%0,4)		       \n\t"
+	"vfmadd231ps 32(%4,%0,4), %%ymm13, %%ymm4      \n\t" 
+	"vfmadd231ps 32(%5,%0,4), %%ymm13, %%ymm5      \n\t" 
+	"prefetcht0	 384(%6,%0,4)		       \n\t"
+	"vfmadd231ps   (%6,%0,4), %%ymm12, %%ymm6      \n\t" 
+	"vfmadd231ps   (%7,%0,4), %%ymm12, %%ymm7      \n\t" 
+	"prefetcht0	 384(%7,%0,4)		       \n\t"
+	"vfmadd231ps 32(%6,%0,4), %%ymm13, %%ymm6      \n\t" 
+	"vfmadd231ps 32(%7,%0,4), %%ymm13, %%ymm7      \n\t" 
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"vextractf128   $1 , %%ymm4, %%xmm12	      \n\t"
+	"vextractf128   $1 , %%ymm5, %%xmm13	      \n\t"
+	"vextractf128   $1 , %%ymm6, %%xmm14	      \n\t"
+	"vextractf128   $1 , %%ymm7, %%xmm15	      \n\t"
+
+	"vaddps		%%xmm4, %%xmm12, %%xmm4       \n\t"
+	"vaddps		%%xmm5, %%xmm13, %%xmm5       \n\t"
+	"vaddps		%%xmm6, %%xmm14, %%xmm6       \n\t"
+	"vaddps		%%xmm7, %%xmm15, %%xmm7       \n\t"
+
+        "vhaddps        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddps        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddps        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddps        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vhaddps        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddps        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddps        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddps        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vmovss         %%xmm4,    (%3)         \n\t"
+        "vmovss         %%xmm5,   4(%3)         \n\t"
+        "vmovss         %%xmm6,   8(%3)         \n\t"
+        "vmovss         %%xmm7,  12(%3)         \n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_t_microk_haswell.c

@@ -0,0 +1,100 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	//n = n / 16;
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"leaq	(, %%rcx,4), %%rcx       \n\t"		// scale lda by size of float
+	"leaq	(%%rsi,%%rcx,1), %%r8    \n\t"		// pointer to next line
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+	"vxorps		%%xmm14, %%xmm14, %%xmm14\n\t"	// set to zero
+	"vxorps		%%xmm15, %%xmm15, %%xmm15\n\t"	// set to zero
+
+	"sarq		$4, %%rax		 \n\t"	// n = n / 16
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	// "prefetcht0	512(%%rsi)		 \n\t"
+	"prefetcht0	(%%r8)		 	 \n\t" //prefetch next line of a
+	"vmovups	(%%rsi), %%xmm4		 \n\t"
+	"vmovups     4*4(%%rsi), %%xmm5		 \n\t"
+	"vmovups     8*4(%%rsi), %%xmm6		 \n\t"
+	"vmovups    12*4(%%rsi), %%xmm7		 \n\t"
+
+	"vfmadd231ps   0*4(%%rdi), %%xmm4, %%xmm12\n\t" // multiply a and c and add to temp
+	"vfmadd231ps   4*4(%%rdi), %%xmm5, %%xmm13\n\t" // multiply a and c and add to temp
+	"vfmadd231ps   8*4(%%rdi), %%xmm6, %%xmm14\n\t" // multiply a and c and add to temp
+	"vfmadd231ps  12*4(%%rdi), %%xmm7, %%xmm15\n\t" // multiply a and c and add to temp
+
+        "addq		$16*4    ,   %%r8	 \n\t"  // increment prefetch pointer 
+        "addq		$16*4    ,   %%rsi	 \n\t"  // increment pointer of a 
+        "addq		$16*4    ,   %%rdi	 \n\t"  // increment pointer of c 
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vaddps		%%xmm12, %%xmm14, %%xmm12\n\t"	
+	"vaddps		%%xmm13, %%xmm15, %%xmm13\n\t"	
+	"vaddps		%%xmm12, %%xmm13, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+
+	"vmulss	        %%xmm12, %%xmm1, %%xmm12\n\t"
+	"vaddss		(%%rdx), %%xmm12,%%xmm12\n\t"
+	"vmovss		%%xmm12, (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+

kernel/x86_64/sgemv_t_microk_nehalem-2.c

@@ -0,0 +1,159 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"xorps		%%xmm0 , %%xmm0	         \n\t"
+	"xorps		%%xmm1 , %%xmm1	         \n\t"
+	"xorps		%%xmm2 , %%xmm2	         \n\t"
+	"xorps		%%xmm3 , %%xmm3	         \n\t"
+	"xorps		%%xmm4 , %%xmm4	         \n\t"
+	"xorps		%%xmm5 , %%xmm5	         \n\t"
+	"xorps		%%xmm6 , %%xmm6	         \n\t"
+	"xorps		%%xmm7 , %%xmm7	         \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 384(%2,%0,4)		       \n\t"
+	"movups	       (%2,%0,4), %%xmm12        \n\t"	// 4 * x
+	"movups      16(%2,%0,4), %%xmm13        \n\t"	// 4 * x
+	"movups             (%4,%0,4), %%xmm8          \n\t" 
+	"movups      32(%2,%0,4), %%xmm14        \n\t"	// 4 * x
+	"movups      48(%2,%0,4), %%xmm15        \n\t"	// 4 * x
+
+	"prefetcht0	 384(%4,%0,4)		       \n\t"
+
+	"movups           16(%4,%0,4), %%xmm9          \n\t" 
+	"movups           32(%4,%0,4), %%xmm10         \n\t" 
+	"movups           48(%4,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm0		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm4		       \n\t"
+	"movups             (%5,%0,4), %%xmm8          \n\t" 
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm0		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm4		       \n\t"
+
+	"prefetcht0	 384(%5,%0,4)		       \n\t"
+
+	"movups           16(%5,%0,4), %%xmm9          \n\t" 
+	"movups           32(%5,%0,4), %%xmm10         \n\t" 
+	"movups           48(%5,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm1		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm5		       \n\t"
+	"movups             (%6,%0,4), %%xmm8          \n\t" 
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm1		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm5		       \n\t"
+
+	"prefetcht0	 384(%6,%0,4)		       \n\t"
+
+	"movups           16(%6,%0,4), %%xmm9          \n\t" 
+	"movups           32(%6,%0,4), %%xmm10         \n\t" 
+	"movups           48(%6,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm2		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm6		       \n\t"
+	"movups             (%7,%0,4), %%xmm8          \n\t" 
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm2		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm6		       \n\t"
+
+	"prefetcht0	 384(%7,%0,4)		       \n\t"
+
+	"movups           16(%7,%0,4), %%xmm9          \n\t" 
+	"movups           32(%7,%0,4), %%xmm10         \n\t" 
+	"movups           48(%7,%0,4), %%xmm11         \n\t" 
+	"mulps		%%xmm12, %%xmm8		       \n\t"
+	"addps		%%xmm8 , %%xmm3		       \n\t"
+	"mulps		%%xmm13, %%xmm9		       \n\t"
+	"addps		%%xmm9 , %%xmm7		       \n\t"
+	"mulps		%%xmm14, %%xmm10	       \n\t"
+	"addps		%%xmm10, %%xmm3		       \n\t"
+	"mulps		%%xmm15, %%xmm11	       \n\t"
+	"addps		%%xmm11, %%xmm7		       \n\t"
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+	"addps	       %%xmm0, %%xmm4		      \n\t"
+	"addps	       %%xmm1, %%xmm5		      \n\t"
+	"addps	       %%xmm2, %%xmm6		      \n\t"
+	"addps	       %%xmm3, %%xmm7		      \n\t"
+
+        "haddps        %%xmm4, %%xmm4  \n\t"
+        "haddps        %%xmm5, %%xmm5  \n\t"
+        "haddps        %%xmm6, %%xmm6  \n\t"
+        "haddps        %%xmm7, %%xmm7  \n\t"
+
+        "haddps        %%xmm4, %%xmm4  \n\t"
+        "haddps        %%xmm5, %%xmm5  \n\t"
+        "haddps        %%xmm6, %%xmm6  \n\t"
+        "haddps        %%xmm7, %%xmm7  \n\t"
+
+        "movss         %%xmm4,    (%3)         \n\t"
+        "movss         %%xmm5,   4(%3)         \n\t"
+        "movss         %%xmm6,   8(%3)         \n\t"
+        "movss         %%xmm7,  12(%3)         \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_t_microk_sandy-2.c

@@ -0,0 +1,132 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void sgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+        "vxorps         %%ymm0 , %%ymm0, %%ymm0  \n\t"
+        "vxorps         %%ymm1 , %%ymm1, %%ymm1  \n\t"
+        "vxorps         %%ymm2 , %%ymm2, %%ymm2  \n\t"
+        "vxorps         %%ymm3 , %%ymm3, %%ymm3  \n\t"
+        "vxorps         %%ymm4 , %%ymm4, %%ymm4  \n\t"
+        "vxorps         %%ymm5 , %%ymm5, %%ymm5  \n\t"
+        "vxorps         %%ymm6 , %%ymm6, %%ymm6  \n\t"
+        "vxorps         %%ymm7 , %%ymm7, %%ymm7  \n\t"
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	"prefetcht0	 384(%2,%0,4)		       \n\t"
+        "vmovups        (%2,%0,4), %%ymm12       \n\t"  // 8 * x
+        "vmovups      32(%2,%0,4), %%ymm13       \n\t"  // 8 * x
+
+	"prefetcht0	 384(%4,%0,4)		       \n\t"
+	"vmulps   (%4,%0,4), %%ymm12, %%ymm8      \n\t" 
+	"vaddps	  %%ymm4, %%ymm8 , %%ymm4	  \n\t"
+	"vmulps 32(%4,%0,4), %%ymm13, %%ymm9      \n\t" 
+	"vaddps	  %%ymm0, %%ymm9 , %%ymm0	  \n\t"
+	"prefetcht0	 384(%5,%0,4)		       \n\t"
+	"vmulps   (%5,%0,4), %%ymm12, %%ymm10     \n\t" 
+	"vaddps	  %%ymm1, %%ymm10, %%ymm1	  \n\t"
+	"vmulps 32(%5,%0,4), %%ymm13, %%ymm11     \n\t" 
+	"vaddps	  %%ymm5, %%ymm11, %%ymm5	  \n\t"
+	"prefetcht0	 384(%6,%0,4)		       \n\t"
+	"vmulps   (%6,%0,4), %%ymm12, %%ymm8      \n\t" 
+	"vaddps	  %%ymm6, %%ymm8 , %%ymm6	  \n\t"
+	"vmulps 32(%6,%0,4), %%ymm13, %%ymm9      \n\t" 
+	"vaddps	  %%ymm2, %%ymm9 , %%ymm2	  \n\t"
+	"prefetcht0	 384(%7,%0,4)		       \n\t"
+	"vmulps   (%7,%0,4), %%ymm12, %%ymm10     \n\t" 
+	"vaddps	  %%ymm7, %%ymm10, %%ymm7	  \n\t"
+	"vmulps 32(%7,%0,4), %%ymm13, %%ymm11     \n\t" 
+	"vaddps	  %%ymm3, %%ymm11, %%ymm3	  \n\t"
+
+        "addq		$16, %0	  	 	      \n\t"
+	"subq	        $16, %1			      \n\t"		
+	"jnz		.L01LOOP%=		      \n\t"
+
+        "vaddps         %%ymm4, %%ymm0, %%ymm4       \n\t"
+        "vaddps         %%ymm5, %%ymm1, %%ymm5       \n\t"
+        "vaddps         %%ymm6, %%ymm2, %%ymm6       \n\t"
+        "vaddps         %%ymm7, %%ymm3, %%ymm7       \n\t"
+
+        "vextractf128   $1 , %%ymm4, %%xmm12          \n\t"
+        "vextractf128   $1 , %%ymm5, %%xmm13          \n\t"
+        "vextractf128   $1 , %%ymm6, %%xmm14          \n\t"
+        "vextractf128   $1 , %%ymm7, %%xmm15          \n\t"
+
+        "vaddps         %%xmm4, %%xmm12, %%xmm4       \n\t"
+        "vaddps         %%xmm5, %%xmm13, %%xmm5       \n\t"
+        "vaddps         %%xmm6, %%xmm14, %%xmm6       \n\t"
+        "vaddps         %%xmm7, %%xmm15, %%xmm7       \n\t"
+
+        "vhaddps        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddps        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddps        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddps        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vhaddps        %%xmm4, %%xmm4, %%xmm4  \n\t"
+        "vhaddps        %%xmm5, %%xmm5, %%xmm5  \n\t"
+        "vhaddps        %%xmm6, %%xmm6, %%xmm6  \n\t"
+        "vhaddps        %%xmm7, %%xmm7, %%xmm7  \n\t"
+
+        "vmovss         %%xmm4,    (%3)         \n\t"
+        "vmovss         %%xmm5,   4(%3)         \n\t"
+        "vmovss         %%xmm6,   8(%3)         \n\t"
+        "vmovss         %%xmm7,  12(%3)         \n\t"
+
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/sgemv_t_microk_sandy.c

@@ -0,0 +1,106 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+static void  sgemv_kernel_16( long n, float alpha, float *a, long lda, float *x, float *y)
+{
+
+	//n = n / 16;
+
+	__asm__ __volatile__
+	(
+	"movq 	        %0, 	 %%rax\n\t"		// n -> rax
+	"vmovss         %1, 	 %%xmm1\n\t"		// alpha -> xmm1
+	"movq		%2,	 %%rsi\n\t"		// adress of a -> rsi
+	"movq	        %3,	 %%rcx\n\t"		// value of lda > rcx
+	"movq		%4,	 %%rdi\n\t"		// adress of x -> rdi
+	"movq		%5,	 %%rdx\n\t"		// adress of y -> rdx
+
+	"leaq	(, %%rcx,4), %%rcx       \n\t"		// scale lda by size of float
+	"leaq	(%%rsi,%%rcx,1), %%r8    \n\t"		// pointer to next line
+
+	"vxorps		%%xmm12, %%xmm12, %%xmm12\n\t"	// set to zero
+	"vxorps		%%xmm13, %%xmm13, %%xmm13\n\t"	// set to zero
+	"vxorps		%%xmm14, %%xmm14, %%xmm14\n\t"	// set to zero
+	"vxorps		%%xmm15, %%xmm15, %%xmm15\n\t"	// set to zero
+
+	"sarq		$4, %%rax		 \n\t"	// n = n / 16
+
+	".align 16				 \n\t"
+	".L01LOOP%=:				 \n\t"
+	// "prefetcht0	512(%%rsi)		 \n\t"
+	"prefetcht0	(%%r8)		 	 \n\t" //prefetch next line of a
+	"vmovups	(%%rsi), %%xmm4		 \n\t"
+	"vmovups     4*4(%%rsi), %%xmm5		 \n\t"
+	"vmovups     8*4(%%rsi), %%xmm6		 \n\t"
+	"vmovups    12*4(%%rsi), %%xmm7		 \n\t"
+
+	"vmulps      0*4(%%rdi), %%xmm4, %%xmm8 \n\t" // multiply a and c and add to temp
+	"vmulps      4*4(%%rdi), %%xmm5, %%xmm9 \n\t" // multiply a and c and add to temp
+	"vmulps      8*4(%%rdi), %%xmm6, %%xmm10\n\t" // multiply a and c and add to temp
+	"vmulps     12*4(%%rdi), %%xmm7, %%xmm11\n\t" // multiply a and c and add to temp
+
+	"vaddps		%%xmm12, %%xmm8 , %%xmm12\n\t"	
+	"vaddps		%%xmm13, %%xmm9 , %%xmm13\n\t"	
+	"vaddps		%%xmm14, %%xmm10, %%xmm14\n\t"	
+	"vaddps		%%xmm15, %%xmm11, %%xmm15\n\t"	
+
+        "addq		$16*4    ,   %%r8	 \n\t"  // increment prefetch pointer 
+        "addq		$16*4    ,   %%rsi	 \n\t"  // increment pointer of a 
+        "addq		$16*4    ,   %%rdi	 \n\t"  // increment pointer of c 
+	"dec		%%rax			 \n\t"  // n = n -1
+	"jnz		.L01LOOP%=		 \n\t"
+
+	"vaddps		%%xmm12, %%xmm14, %%xmm12\n\t"	
+	"vaddps		%%xmm13, %%xmm15, %%xmm13\n\t"	
+	"vaddps		%%xmm12, %%xmm13, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+	"vhaddps	%%xmm12, %%xmm12, %%xmm12\n\t"	
+
+	"vmulss		%%xmm12, %%xmm1, %%xmm12 \n\t"
+	"vaddss	       (%%rdx), %%xmm12, %%xmm12\n\t"
+	"vmovss		%%xmm12, (%%rdx)	 \n\t"  // store temp -> y
+
+	:
+        :
+          "m" (n),	// 0	
+	  "m" (alpha),  // 1
+	  "m" (a),      // 2
+          "m" (lda),    // 3
+          "m" (x),      // 4
+          "m" (y)       // 5
+	: "%rax", "%rcx", "%rdx", "%rsi", "%rdi", "%r8", "cc",
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+
+

kernel/x86_64/zgemv_n.c

@@ -0,0 +1,258 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include <stdlib.h>
+#include <stdio.h>
+#include "common.h"
+
+#if defined(HASWELL)
+#include "zgemv_n_microk_haswell-2.c"
+#elif defined(SANDYBRIDGE)
+#include "zgemv_n_microk_sandy-2.c"
+#endif
+
+
+
+#define NBMAX 1024
+
+#ifndef HAVE_KERNEL_16x4
+
+static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] - a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] + a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] - a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] + a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] - a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] + a3[i+1] * x[6];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+		y[i]   += a1[i]*x[2] + a1[i+1] * x[3];
+		y[i+1] += a1[i]*x[3] - a1[i+1] * x[2];
+		y[i]   += a2[i]*x[4] + a2[i+1] * x[5];
+		y[i+1] += a2[i]*x[5] - a2[i+1] * x[4];
+		y[i]   += a3[i]*x[6] + a3[i+1] * x[7];
+		y[i+1] += a3[i]*x[7] - a3[i+1] * x[6];
+#endif
+	}
+}
+	
+#endif
+
+static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		y[i]   += a0[i]*x[0] - a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] + a0[i+1] * x[0];
+#else 
+		y[i]   += a0[i]*x[0] + a0[i+1] * x[1];
+		y[i+1] += a0[i]*x[1] - a0[i+1] * x[0];
+#endif
+
+	}
+}
+	
+
+static void zero_y(BLASLONG n, FLOAT *dest)
+{
+	BLASLONG i;
+	for ( i=0; i<2*n; i++ )
+	{
+		*dest = 0.0;
+		dest++;
+	}
+}
+
+
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT alpha_r, FLOAT alpha_i)
+{
+	BLASLONG i;
+	FLOAT temp_r;
+	FLOAT temp_i;
+	for ( i=0; i<n; i++ )
+	{
+#if !defined(XCONJ) 
+		temp_r = alpha_r * src[0] - alpha_i * src[1];
+		temp_i = alpha_r * src[1] + alpha_i * src[0];
+#else
+		temp_r =  alpha_r * src[0] + alpha_i * src[1];
+		temp_i = -alpha_r * src[1] + alpha_i * src[0];
+#endif
+
+		*dest += temp_r;
+		*(dest+1) += temp_i;
+
+		src+=2;
+		dest += inc_dest;
+	}
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r,FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[4];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT xbuffer[8],*ybuffer;
+
+
+#if 0
+printf("%s %d %d %.16f %.16f %d %d %d\n","zgemv_n",m,n,alpha_r,alpha_i,lda,inc_x,inc_y);
+#endif
+
+	if ( m < 1 ) return(0);
+	if ( n < 1 ) return(0);
+
+	ybuffer = buffer;
+	
+	inc_x *= 2;
+	inc_y *= 2;
+	lda   *= 2;
+
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+	y_ptr = y;
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		a_ptr = a;
+		x_ptr = x;
+		zero_y(NB,ybuffer);
+		for( i = 0; i < n1 ; i++)
+		{
+
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[2] = x_ptr[0];
+			xbuffer[3] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[4] = x_ptr[0];
+			xbuffer[5] = x_ptr[1];
+			x_ptr += inc_x;	
+			xbuffer[6] = x_ptr[0];
+			xbuffer[7] = x_ptr[1];
+			x_ptr += inc_x;	
+
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			xbuffer[0] = x_ptr[0];
+			xbuffer[1] = x_ptr[1];
+			x_ptr += inc_x;	
+			zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+		}
+		add_y(NB,ybuffer,y_ptr,inc_y,alpha_r,alpha_i);
+		a     += 2 * NB;
+		y_ptr += NB * inc_y;
+	}
+
+	j=0;
+	while ( j < (m % 16))
+	{
+		a_ptr = a;
+		x_ptr = x;
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < n; i++ )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+			temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+			temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+
+			a_ptr += lda;
+			x_ptr += inc_x;
+		}
+
+#if !defined(XCONJ) 
+		y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+		y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+		y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+		y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+		y_ptr += inc_y;
+		a+=2;
+		j++;
+	}
+	return(0);
+}
+
+

kernel/x86_64/zgemv_n_microk_haswell-2.c

@@ -0,0 +1,137 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastsd	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastsd	 8(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastsd	16(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastsd	24(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastsd	32(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastsd	40(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastsd	48(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastsd	56(%2), %%ymm7                  \n\t"  // imag part x3
+
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+	"prefetcht0      192(%4,%0,8)			\n\t"
+	"vmovups	(%4,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmovups      32(%4,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"prefetcht0      192(%5,%0,8)			\n\t"
+	"vmovups	(%5,%0,8), %%ymm10              \n\t" // 2 complex values form a1
+	"vmovups      32(%5,%0,8), %%ymm11              \n\t" // 2 complex values form a1
+
+	"vmulpd      %%ymm8 , %%ymm0, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vmulpd      %%ymm8 , %%ymm1, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vmulpd      %%ymm9 , %%ymm0, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vmulpd      %%ymm9 , %%ymm1, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      192(%6,%0,8)			\n\t"
+	"vmovups	(%6,%0,8), %%ymm8	        \n\t" // 2 complex values form a2
+	"vmovups      32(%6,%0,8), %%ymm9	        \n\t" // 2 complex values form a2
+
+	"vfmadd231pd      %%ymm10, %%ymm2, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231pd      %%ymm10, %%ymm3, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231pd      %%ymm11, %%ymm2, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231pd      %%ymm11, %%ymm3, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      192(%7,%0,8)			\n\t"
+	"vmovups	(%7,%0,8), %%ymm10              \n\t" // 2 complex values form a3
+	"vmovups      32(%7,%0,8), %%ymm11              \n\t" // 2 complex values form a3
+
+	"vfmadd231pd      %%ymm8 , %%ymm4, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231pd      %%ymm8 , %%ymm5, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231pd      %%ymm9 , %%ymm4, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231pd      %%ymm9 , %%ymm5, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"vfmadd231pd      %%ymm10, %%ymm6, %%ymm12      \n\t" // a_r[0] * x_r , a_i[0] * x_r, a_r[1] * x_r, a_i[1] * x_r
+	"vfmadd231pd      %%ymm10, %%ymm7, %%ymm13      \n\t" // a_r[0] * x_i , a_i[0] * x_i, a_r[1] * x_i, a_i[1] * x_i
+	"vfmadd231pd      %%ymm11, %%ymm6, %%ymm14      \n\t" // a_r[2] * x_r , a_i[2] * x_r, a_r[3] * x_r, a_i[3] * x_r
+	"vfmadd231pd      %%ymm11, %%ymm7, %%ymm15      \n\t" // a_r[2] * x_i , a_i[2] * x_i, a_r[3] * x_i, a_i[3] * x_i
+
+	"prefetcht0      192(%3,%0,8)			\n\t"
+	"vmovups	  (%3,%0,8),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,8),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
+        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubpd      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddpd         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddpd         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,8)		        \n\t" // 2 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,8)		        \n\t"	
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/zgemv_n_microk_sandy-2.c

@@ -0,0 +1,149 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vbroadcastsd	  (%2), %%ymm0                  \n\t"  // real part x0
+	"vbroadcastsd	 8(%2), %%ymm1                  \n\t"  // imag part x0
+	"vbroadcastsd	16(%2), %%ymm2                  \n\t"  // real part x1
+	"vbroadcastsd	24(%2), %%ymm3                  \n\t"  // imag part x1
+	"vbroadcastsd	32(%2), %%ymm4                  \n\t"  // real part x2
+	"vbroadcastsd	40(%2), %%ymm5                  \n\t"  // imag part x2
+	"vbroadcastsd	48(%2), %%ymm6                  \n\t"  // real part x3
+	"vbroadcastsd	56(%2), %%ymm7                  \n\t"  // imag part x3
+
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+
+        "prefetcht0      256(%4,%0,8)                   \n\t"
+	"vmovups	(%4,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmovups      32(%4,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"vmulpd		  %%ymm8 , %%ymm0 , %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm1 , %%ymm13	\n\t"
+        "prefetcht0      256(%5,%0,8)                   \n\t"
+	"vmulpd		  %%ymm9 , %%ymm0 , %%ymm14	\n\t"
+	"vmovups	(%5,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmulpd		  %%ymm9 , %%ymm1 , %%ymm15	\n\t"
+	"vmovups      32(%5,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"vmulpd		  %%ymm8 , %%ymm2 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm3 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+        "prefetcht0      256(%6,%0,8)                   \n\t"
+	"vmulpd		  %%ymm9 , %%ymm2 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmovups	(%6,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmulpd		  %%ymm9 , %%ymm3 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+	"vmovups      32(%6,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"vmulpd		  %%ymm8 , %%ymm4 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm5 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+        "prefetcht0      256(%7,%0,8)                   \n\t"
+	"vmulpd		  %%ymm9 , %%ymm4 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmovups	(%7,%0,8), %%ymm8	        \n\t" // 2 complex values form a0
+	"vmulpd		  %%ymm9 , %%ymm5 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+	"vmovups      32(%7,%0,8), %%ymm9	        \n\t" // 2 complex values form a0
+
+	"vmulpd		  %%ymm8 , %%ymm6 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm12, %%ymm10, %%ymm12	\n\t"
+	"vmulpd		  %%ymm8 , %%ymm7 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm13, %%ymm11, %%ymm13	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm6 , %%ymm10	\n\t"
+	"vaddpd		  %%ymm14, %%ymm10, %%ymm14	\n\t"
+	"vmulpd		  %%ymm9 , %%ymm7 , %%ymm11	\n\t"
+	"vaddpd		  %%ymm15, %%ymm11, %%ymm15	\n\t"
+
+	"prefetcht0      256(%3,%0,8)			\n\t"
+	"vmovups	  (%3,%0,8),  %%ymm10           \n\t"
+	"vmovups	32(%3,%0,8),  %%ymm11           \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
+        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm9              \n\t"
+#else
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubpd      %%ymm12, %%ymm13, %%ymm8              \n\t"
+        "vaddsubpd      %%ymm14, %%ymm15, %%ymm9              \n\t"
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
+#endif
+
+        "vaddpd         %%ymm8, %%ymm10, %%ymm12              \n\t"
+        "vaddpd         %%ymm9, %%ymm11, %%ymm13              \n\t"
+
+	"vmovups  %%ymm12,   (%3,%0,8)		        \n\t" // 2 complex values to y	
+	"vmovups  %%ymm13, 32(%3,%0,8)		        \n\t"	
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/zgemv_t.c

@@ -0,0 +1,272 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+
+#include "common.h"
+
+
+#if defined(BULLDOZER) || defined(PILEDRIVER)
+#include "zgemv_t_microk_bulldozer-2.c"
+#elif defined(HASWELL)
+#include "zgemv_t_microk_haswell-2.c"
+#endif
+
+
+#define NBMAX 1028
+
+#ifndef HAVE_KERNEL_16x4
+
+static void zgemv_kernel_16x4(BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0,*a1,*a2,*a3;
+	a0 = ap[0];
+	a1 = ap[1];
+	a2 = ap[2];
+	a3 = ap[3];
+	FLOAT temp_r0 = 0.0;
+	FLOAT temp_r1 = 0.0;
+	FLOAT temp_r2 = 0.0;
+	FLOAT temp_r3 = 0.0;
+	FLOAT temp_i0 = 0.0;
+	FLOAT temp_i1 = 0.0;
+	FLOAT temp_i2 = 0.0;
+	FLOAT temp_i3 = 0.0;
+
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r0 += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] + a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   - a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] + a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   - a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] + a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   - a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] + a3[i+1]*x[i];		
+#else
+		temp_r0 += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i0 += a0[i]*x[i+1] - a0[i+1]*x[i];		
+		temp_r1 += a1[i]*x[i]   + a1[i+1]*x[i+1];		
+		temp_i1 += a1[i]*x[i+1] - a1[i+1]*x[i];		
+		temp_r2 += a2[i]*x[i]   + a2[i+1]*x[i+1];		
+		temp_i2 += a2[i]*x[i+1] - a2[i+1]*x[i];		
+		temp_r3 += a3[i]*x[i]   + a3[i+1]*x[i+1];		
+		temp_i3 += a3[i]*x[i+1] - a3[i+1]*x[i];		
+#endif
+	}
+	y[0] = temp_r0;
+	y[1] = temp_i0;
+	y[2] = temp_r1;
+	y[3] = temp_i1;
+	y[4] = temp_r2;
+	y[5] = temp_i2;
+	y[6] = temp_r3;
+	y[7] = temp_i3;
+}
+	
+#endif
+
+static void zgemv_kernel_16x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
+{
+	BLASLONG i;
+	FLOAT *a0;
+	a0 = ap;
+	FLOAT temp_r = 0.0;
+	FLOAT temp_i = 0.0;
+
+	for ( i=0; i< 2*n; i+=2 )
+	{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+		temp_r += a0[i]*x[i]   - a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] + a0[i+1]*x[i];		
+#else
+		temp_r += a0[i]*x[i]   + a0[i+1]*x[i+1];		
+		temp_i += a0[i]*x[i+1] - a0[i+1]*x[i];		
+#endif
+	}
+	*y      = temp_r;
+	*(y+1)  = temp_i;
+}
+	
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src)
+{
+        BLASLONG i;
+        for ( i=0; i<n; i++ )
+        {
+                *dest     = *src;
+                *(dest+1) = *(src+1);
+                dest+=2;
+                src += inc_src;
+        }
+}
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer)
+{
+	BLASLONG i;
+	BLASLONG j;
+	FLOAT *a_ptr;
+	FLOAT *x_ptr;
+	FLOAT *y_ptr;
+	FLOAT *ap[8];
+	BLASLONG n1;
+	BLASLONG m1;
+	BLASLONG m2;
+	BLASLONG n2;
+	FLOAT ybuffer[8],*xbuffer;
+
+        if ( m < 1 ) return(0);
+        if ( n < 1 ) return(0);
+
+        inc_x *= 2;
+        inc_y *= 2;
+        lda   *= 2;
+
+	xbuffer = buffer;
+	
+	n1 = n / 4 ;
+	n2 = n % 4 ;
+	
+	m1 = m - ( m % 16 );
+	m2 = (m % NBMAX) - (m % 16) ;
+	
+
+	BLASLONG NB = NBMAX;
+
+	while ( NB == NBMAX )
+	{
+		
+		m1 -= NB;
+		if ( m1 < 0)
+		{
+			if ( m2 == 0 ) break;	
+			NB = m2;
+		}
+		
+		y_ptr = y;
+		a_ptr = a;
+		x_ptr = x;
+		copy_x(NB,x_ptr,xbuffer,inc_x);
+		for( i = 0; i < n1 ; i++)
+		{
+			ap[0] = a_ptr;
+			ap[1] = a_ptr + lda;
+			ap[2] = ap[1] + lda;
+			ap[3] = ap[2] + lda;
+			zgemv_kernel_16x4(NB,ap,xbuffer,ybuffer);
+			a_ptr += 4 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] - alpha_i * ybuffer[3];
+			y_ptr[1] += alpha_r * ybuffer[3] + alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] - alpha_i * ybuffer[5];
+			y_ptr[1] += alpha_r * ybuffer[5] + alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] - alpha_i * ybuffer[7];
+			y_ptr[1] += alpha_r * ybuffer[7] + alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[2] + alpha_i * ybuffer[3];
+			y_ptr[1] -= alpha_r * ybuffer[3] - alpha_i * ybuffer[2];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[4] + alpha_i * ybuffer[5];
+			y_ptr[1] -= alpha_r * ybuffer[5] - alpha_i * ybuffer[4];
+			y_ptr  += inc_y;
+			y_ptr[0] += alpha_r * ybuffer[6] + alpha_i * ybuffer[7];
+			y_ptr[1] -= alpha_r * ybuffer[7] - alpha_i * ybuffer[6];
+			y_ptr  += inc_y;
+#endif
+		}
+
+		for( i = 0; i < n2 ; i++)
+		{
+			zgemv_kernel_16x1(NB,a_ptr,xbuffer,ybuffer);
+			a_ptr += 1 * lda;
+
+#if !defined(XCONJ)
+			y_ptr[0] += alpha_r * ybuffer[0] - alpha_i * ybuffer[1];
+			y_ptr[1] += alpha_r * ybuffer[1] + alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#else
+			y_ptr[0] += alpha_r * ybuffer[0] + alpha_i * ybuffer[1];
+			y_ptr[1] -= alpha_r * ybuffer[1] - alpha_i * ybuffer[0];
+			y_ptr  += inc_y;
+#endif
+
+		}
+		a += 2* NB;
+		x += NB * inc_x;	
+	}
+
+	BLASLONG m3 = m % 16;
+	if ( m3 == 0 ) return(0);
+
+	x_ptr = x;
+	copy_x(m3,x_ptr,xbuffer,inc_x);
+	j=0;
+	a_ptr = a;
+	y_ptr = y;
+	while ( j < n)
+	{
+		FLOAT temp_r = 0.0;
+		FLOAT temp_i = 0.0;
+		for( i = 0; i < m3*2; i+=2 )
+		{
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+			temp_r += a_ptr[i] * xbuffer[i]   - a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] + a_ptr[i+1] * xbuffer[i];
+#else
+			temp_r += a_ptr[i] * xbuffer[i]   + a_ptr[i+1] * xbuffer[i+1];
+			temp_i += a_ptr[i] * xbuffer[i+1] - a_ptr[i+1] * xbuffer[i];
+#endif
+		}
+		a_ptr += lda;
+
+#if !defined(XCONJ) 
+                y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+                y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+                y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+                y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+		y_ptr += inc_y;
+		j++;
+	}
+	return(0);
+}
+
+

kernel/x86_64/zgemv_t_microk_bulldozer-2.c

@@ -0,0 +1,180 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary froms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary from must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorpd		%%xmm8 , %%xmm8 , %%xmm8 	\n\t" // temp
+	"vxorpd		%%xmm9 , %%xmm9 , %%xmm9 	\n\t" // temp
+	"vxorpd		%%xmm10, %%xmm10, %%xmm10	\n\t" // temp
+	"vxorpd		%%xmm11, %%xmm11, %%xmm11	\n\t" // temp
+	"vxorpd		%%xmm12, %%xmm12, %%xmm12	\n\t" // temp
+	"vxorpd		%%xmm13, %%xmm13, %%xmm13	\n\t"
+	"vxorpd		%%xmm14, %%xmm14, %%xmm14	\n\t"
+	"vxorpd		%%xmm15, %%xmm15, %%xmm15	\n\t"
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+
+	"vmovddup	   (%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	  8(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+
+	"prefetcht0  192(%4,%0,8)                       \n\t"
+	"vmovups	(%4,%0,8), %%xmm4	        \n\t" // 1 complex values from a0
+	"prefetcht0  192(%5,%0,8)                       \n\t"
+	"vmovups	(%5,%0,8), %%xmm5               \n\t" // 1 complex values from a1
+	"prefetcht0  192(%6,%0,8)                       \n\t"
+	"vmovups	(%6,%0,8), %%xmm6	        \n\t" // 1 complex values from a2
+	"prefetcht0  192(%7,%0,8)                       \n\t"
+	"vmovups	(%7,%0,8), %%xmm7               \n\t" // 1 complex values from a3
+
+	"vfmaddpd   %%xmm8 ,   %%xmm4 , %%xmm0, %%xmm8       \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm9 ,   %%xmm4 , %%xmm1, %%xmm9       \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm10,   %%xmm5 , %%xmm0, %%xmm10      \n\t" // ar0*xr0,al0*xr0
+	"vfmaddpd   %%xmm11,   %%xmm5 , %%xmm1, %%xmm11      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm12,   %%xmm6 , %%xmm0, %%xmm12      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm13,   %%xmm6 , %%xmm1, %%xmm13      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm14,   %%xmm7 , %%xmm0, %%xmm14      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm15,   %%xmm7 , %%xmm1, %%xmm15      \n\t" // ar0*xl0,al0*xl0 
+
+	"vmovddup	 16(%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	 24(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+
+	"vmovups      16(%4,%0,8), %%xmm4	        \n\t" // 1 complex values from a0
+	"vmovups      16(%5,%0,8), %%xmm5               \n\t" // 1 complex values from a1
+	"vmovups      16(%6,%0,8), %%xmm6	        \n\t" // 1 complex values from a2
+	"vmovups      16(%7,%0,8), %%xmm7               \n\t" // 1 complex values from a3
+
+	"vfmaddpd   %%xmm8 ,   %%xmm4 , %%xmm0, %%xmm8       \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm9 ,   %%xmm4 , %%xmm1, %%xmm9       \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm10,   %%xmm5 , %%xmm0, %%xmm10      \n\t" // ar0*xr0,al0*xr0
+	"vfmaddpd   %%xmm11,   %%xmm5 , %%xmm1, %%xmm11      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm12,   %%xmm6 , %%xmm0, %%xmm12      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm13,   %%xmm6 , %%xmm1, %%xmm13      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm14,   %%xmm7 , %%xmm0, %%xmm14      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm15,   %%xmm7 , %%xmm1, %%xmm15      \n\t" // ar0*xl0,al0*xl0 
+
+	"vmovddup	 32(%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	 40(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+
+	"vmovups      32(%4,%0,8), %%xmm4	        \n\t" // 1 complex values from a0
+	"vmovups      32(%5,%0,8), %%xmm5               \n\t" // 1 complex values from a1
+	"vmovups      32(%6,%0,8), %%xmm6	        \n\t" // 1 complex values from a2
+	"vmovups      32(%7,%0,8), %%xmm7               \n\t" // 1 complex values from a3
+
+	"vfmaddpd   %%xmm8 ,   %%xmm4 , %%xmm0, %%xmm8       \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm9 ,   %%xmm4 , %%xmm1, %%xmm9       \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm10,   %%xmm5 , %%xmm0, %%xmm10      \n\t" // ar0*xr0,al0*xr0
+	"vfmaddpd   %%xmm11,   %%xmm5 , %%xmm1, %%xmm11      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm12,   %%xmm6 , %%xmm0, %%xmm12      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm13,   %%xmm6 , %%xmm1, %%xmm13      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm14,   %%xmm7 , %%xmm0, %%xmm14      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm15,   %%xmm7 , %%xmm1, %%xmm15      \n\t" // ar0*xl0,al0*xl0 
+
+	"vmovddup	 48(%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	 56(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+
+	"vmovups      48(%4,%0,8), %%xmm4	        \n\t" // 1 complex values from a0
+	"vmovups      48(%5,%0,8), %%xmm5               \n\t" // 1 complex values from a1
+	"vmovups      48(%6,%0,8), %%xmm6	        \n\t" // 1 complex values from a2
+	"vmovups      48(%7,%0,8), %%xmm7               \n\t" // 1 complex values from a3
+
+	"vfmaddpd   %%xmm8 ,   %%xmm4 , %%xmm0, %%xmm8       \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm9 ,   %%xmm4 , %%xmm1, %%xmm9       \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm10,   %%xmm5 , %%xmm0, %%xmm10      \n\t" // ar0*xr0,al0*xr0
+	"vfmaddpd   %%xmm11,   %%xmm5 , %%xmm1, %%xmm11      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm12,   %%xmm6 , %%xmm0, %%xmm12      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm13,   %%xmm6 , %%xmm1, %%xmm13      \n\t" // ar0*xl0,al0*xl0 
+	"vfmaddpd   %%xmm14,   %%xmm7 , %%xmm0, %%xmm14      \n\t" // ar0*xr0,al0*xr0 
+	"vfmaddpd   %%xmm15,   %%xmm7 , %%xmm1, %%xmm15      \n\t" // ar0*xl0,al0*xl0 
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x1 , %%xmm9 , %%xmm9                \n\t"
+        "vpermilpd      $0x1 , %%xmm11, %%xmm11               \n\t"
+        "vpermilpd      $0x1 , %%xmm13, %%xmm13               \n\t"
+        "vpermilpd      $0x1 , %%xmm15, %%xmm15               \n\t"
+        "vaddsubpd      %%xmm9 , %%xmm8, %%xmm8               \n\t" 
+        "vaddsubpd      %%xmm11, %%xmm10, %%xmm10             \n\t"
+        "vaddsubpd      %%xmm13, %%xmm12, %%xmm12             \n\t"
+        "vaddsubpd      %%xmm15, %%xmm14, %%xmm14             \n\t"
+#else
+        "vpermilpd      $0x1 , %%xmm8 , %%xmm8                \n\t"
+        "vpermilpd      $0x1 , %%xmm10, %%xmm10               \n\t"
+        "vpermilpd      $0x1 , %%xmm12, %%xmm12               \n\t"
+        "vpermilpd      $0x1 , %%xmm14, %%xmm14               \n\t"
+        "vaddsubpd      %%xmm8 , %%xmm9 , %%xmm8              \n\t"
+        "vaddsubpd      %%xmm10, %%xmm11, %%xmm10             \n\t"
+        "vaddsubpd      %%xmm12, %%xmm13, %%xmm12             \n\t"
+        "vaddsubpd      %%xmm14, %%xmm15, %%xmm14             \n\t"
+        "vpermilpd      $0x1 , %%xmm8 , %%xmm8                \n\t"
+        "vpermilpd      $0x1 , %%xmm10, %%xmm10               \n\t"
+        "vpermilpd      $0x1 , %%xmm12, %%xmm12               \n\t"
+        "vpermilpd      $0x1 , %%xmm14, %%xmm14               \n\t"
+#endif
+
+
+	"vmovups	%%xmm8 ,   (%3)			\n\t"
+	"vmovups	%%xmm10, 16(%3)			\n\t"
+	"vmovups	%%xmm12, 32(%3)			\n\t"
+	"vmovups	%%xmm14, 48(%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

kernel/x86_64/zgemv_t_microk_haswell-2.c

@@ -0,0 +1,162 @@
+/***************************************************************************
+Copyright (c) 2014, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary froms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary from must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#define HAVE_KERNEL_16x4 1
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y) __attribute__ ((noinline));
+
+static void zgemv_kernel_16x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vzeroupper			 \n\t"
+
+	"vxorpd		%%ymm8 , %%ymm8 , %%ymm8 	\n\t" // temp
+	"vxorpd		%%ymm9 , %%ymm9 , %%ymm9 	\n\t" // temp
+	"vxorpd		%%ymm10, %%ymm10, %%ymm10	\n\t" // temp
+	"vxorpd		%%ymm11, %%ymm11, %%ymm11	\n\t" // temp
+	"vxorpd		%%ymm12, %%ymm12, %%ymm12	\n\t" // temp
+	"vxorpd		%%ymm13, %%ymm13, %%ymm13	\n\t"
+	"vxorpd		%%ymm14, %%ymm14, %%ymm14	\n\t"
+	"vxorpd		%%ymm15, %%ymm15, %%ymm15	\n\t"
+
+	".align 16				        \n\t"
+	".L01LOOP%=:				        \n\t"
+
+        "prefetcht0      192(%2,%0,8)                   \n\t"
+	"vmovddup	   (%2,%0,8), %%xmm0            \n\t"  // real value from x0
+        "prefetcht0      192(%4,%0,8)                   \n\t"
+	"vmovups	(%5,%0,8), %%ymm5               \n\t" // 2 complex values from a1
+	"vmovddup	  8(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+	"vmovups	(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
+        "prefetcht0      192(%5,%0,8)                   \n\t"
+	"vmovddup	 16(%2,%0,8), %%xmm2            \n\t"  // real value from x1
+        "prefetcht0      192(%6,%0,8)                   \n\t"
+	"vmovups	(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
+	"vmovddup	 24(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
+        "prefetcht0      192(%7,%0,8)                   \n\t"
+	"vmovups	(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
+	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
+	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
+
+	"vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+	"vmovups       32(%4,%0,8), %%ymm4	        \n\t" // 2 complex values from a0
+	"vmovups       32(%5,%0,8), %%ymm5              \n\t" // 2 complex values from a1
+	"vmovddup	 32(%2,%0,8), %%xmm0            \n\t"  // real value from x0
+	"vmovddup	 40(%2,%0,8), %%xmm1            \n\t"  // imag value from x0
+	"vmovddup	 48(%2,%0,8), %%xmm2            \n\t"  // real value from x1
+	"vmovddup	 56(%2,%0,8), %%xmm3            \n\t"  // imag value from x1
+	"vmovups       32(%6,%0,8), %%ymm6	        \n\t" // 2 complex values from a2
+	"vmovups       32(%7,%0,8), %%ymm7               \n\t" // 2 complex values from a3
+	"vinsertf128	 $1, %%xmm2, %%ymm0 , %%ymm0	\n\t"  // real values from x0 and x1
+	"vinsertf128	 $1, %%xmm3, %%ymm1 , %%ymm1	\n\t"  // imag values from x0 and x1
+
+	"vfmadd231pd      %%ymm4 , %%ymm0, %%ymm8       \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm4 , %%ymm1, %%ymm9       \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm5 , %%ymm0, %%ymm10      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm5 , %%ymm1, %%ymm11      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm6 , %%ymm0, %%ymm12      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm6 , %%ymm1, %%ymm13      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+	"vfmadd231pd      %%ymm7 , %%ymm0, %%ymm14      \n\t" // ar0*xr0,al0*xr0,ar1*xr1,al1*xr1 
+	"vfmadd231pd      %%ymm7 , %%ymm1, %%ymm15      \n\t" // ar0*xl0,al0*xl0,ar1*xl1,al1*xl1 
+
+        "addq		$8 , %0	  	 	        \n\t"
+	"subq	        $4 , %1			        \n\t"		
+	"jnz		.L01LOOP%=		        \n\t"
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        "vpermilpd      $0x5 , %%ymm9 , %%ymm9                \n\t"
+        "vpermilpd      $0x5 , %%ymm11, %%ymm11               \n\t"
+        "vpermilpd      $0x5 , %%ymm13, %%ymm13               \n\t"
+        "vpermilpd      $0x5 , %%ymm15, %%ymm15               \n\t"
+        "vaddsubpd      %%ymm9 , %%ymm8, %%ymm8               \n\t" 
+        "vaddsubpd      %%ymm11, %%ymm10, %%ymm10             \n\t"
+        "vaddsubpd      %%ymm13, %%ymm12, %%ymm12             \n\t"
+        "vaddsubpd      %%ymm15, %%ymm14, %%ymm14             \n\t"
+#else
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm10, %%ymm10               \n\t"
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+        "vaddsubpd      %%ymm8 , %%ymm9 , %%ymm8              \n\t"
+        "vaddsubpd      %%ymm10, %%ymm11, %%ymm10             \n\t"
+        "vaddsubpd      %%ymm12, %%ymm13, %%ymm12             \n\t"
+        "vaddsubpd      %%ymm14, %%ymm15, %%ymm14             \n\t"
+        "vpermilpd      $0x5 , %%ymm8 , %%ymm8                \n\t"
+        "vpermilpd      $0x5 , %%ymm10, %%ymm10               \n\t"
+        "vpermilpd      $0x5 , %%ymm12, %%ymm12               \n\t"
+        "vpermilpd      $0x5 , %%ymm14, %%ymm14               \n\t"
+#endif
+
+	"vextractf128   $1, %%ymm8 , %%xmm9		      \n\t"
+	"vextractf128   $1, %%ymm10, %%xmm11	      	      \n\t"
+	"vextractf128   $1, %%ymm12, %%xmm13		      \n\t"
+	"vextractf128   $1, %%ymm14, %%xmm15		      \n\t"
+
+	"vaddpd		%%xmm8 , %%xmm9 , %%xmm8       \n\t"
+	"vaddpd		%%xmm10, %%xmm11, %%xmm10      \n\t"
+	"vaddpd		%%xmm12, %%xmm13, %%xmm12      \n\t"
+	"vaddpd		%%xmm14, %%xmm15, %%xmm14      \n\t"
+
+	"vmovups	%%xmm8 ,   (%3)			\n\t"
+	"vmovups	%%xmm10, 16(%3)			\n\t"
+	"vmovups	%%xmm12, 32(%3)			\n\t"
+	"vmovups	%%xmm14, 48(%3)			\n\t"
+
+	"vzeroupper			 \n\t"
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (ap[2]),  // 6
+          "r" (ap[3])   // 7
+	: "cc", 
+	  "%xmm0", "%xmm1", "%xmm2", "%xmm3", 
+	  "%xmm4", "%xmm5", "%xmm6", "%xmm7", 
+	  "%xmm8", "%xmm9", "%xmm10", "%xmm11", 
+	  "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+	  "memory"
+	);
+
+} 
+
+

lapack-netlib/TESTING/dstest.in

@@ -1,8 +1,8 @@
 Data file for testing DSGESV/DSPOSV LAPACK routines
 12                                      Number of values of M
 0 1 2 13 17 45 78 91 101 119 120 132    values of M (row dimension)
-4                                       Number of values of NRHS
-1 2 14 16                               Values of NRHS (number of right hand sides)
+6                                       Number of values of NRHS
+1 2 14 15 16 13                         Values of NRHS (number of right hand sides)
 30.0                                    Threshold value of test ratio
 T                                       Put T to test the driver routine
 T                                       Put T to test the error exits

param.h

@@ -1134,9 +1134,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define QGEMM_DEFAULT_P 504
 #define QGEMM_DEFAULT_R qgemm_r
 
-#define CGEMM_DEFAULT_P 384
-//#define CGEMM_DEFAULT_R cgemm_r
-#define CGEMM_DEFAULT_R 1024
+#define CGEMM_DEFAULT_P 768
+#define CGEMM_DEFAULT_R cgemm_r
+//#define CGEMM_DEFAULT_R 1024
 
 #define ZGEMM_DEFAULT_P 512
 #define ZGEMM_DEFAULT_R zgemm_r
@@ -1148,7 +1148,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SGEMM_DEFAULT_Q 384
 #define DGEMM_DEFAULT_Q 256
 #define QGEMM_DEFAULT_Q 128
-#define CGEMM_DEFAULT_Q 192
+#define CGEMM_DEFAULT_Q 512
 #define ZGEMM_DEFAULT_Q 192
 #define XGEMM_DEFAULT_Q 128
 
@@ -1206,6 +1206,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ZGEMM_DEFAULT_UNROLL_N 2
 #define XGEMM_DEFAULT_UNROLL_N 1
 
+#define DGEMM_DEFAULT_UNROLL_MN 16
 #endif
 
 #ifdef ARCH_X86
@@ -1236,10 +1237,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CGEMM_DEFAULT_P 384
 #define ZGEMM_DEFAULT_P 256
 
-#define SGEMM_DEFAULT_Q 384
 #ifdef WINDOWS_ABI
+#define SGEMM_DEFAULT_Q 320
 #define DGEMM_DEFAULT_Q 128
 #else
+#define SGEMM_DEFAULT_Q 384
 #define DGEMM_DEFAULT_Q 256
 #endif
 #define CGEMM_DEFAULT_Q 192