Release 0.2.16 rc1

On Mac OS X, it should use .align 4 (equal to .align 16 on Linux).
I didn't get the performance benefit from .align. Thus, I deleted it.

.travis.yml

@@ -24,7 +24,11 @@ before_install:
  - if [[ "$TARGET_BOX" == "WIN64" ]]; then sudo apt-get install -qq binutils-mingw-w64-x86-64 gcc-mingw-w64-x86-64 gfortran-mingw-w64-x86-64; fi
  - if [[ "$TARGET_BOX" == "LINUX32" ]]; then sudo apt-get install -qq  gcc-multilib gfortran-multilib; fi
 
-script: make QUIET_MAKE=1 DYNAMIC_ARCH=1 TARGET=NEHALEM NUM_THREADS=32 $BTYPE
+script: 
+ - make QUIET_MAKE=1 DYNAMIC_ARCH=1 TARGET=NEHALEM NUM_THREADS=32 $BTYPE
+ - if [ "$TARGET_BOX" == "LINUX32" ] || [ "$TARGET_BOX" == "LINUX64" ]; then make -C test DYNAMIC_ARCH=1 TARGET=NEHALEM NUM_THREADS=32 $BTYPE; fi
+ - if [ "$TARGET_BOX" == "LINUX32" ] || [ "$TARGET_BOX" == "LINUX64" ]; then make -C ctest DYNAMIC_ARCH=1 TARGET=NEHALEM NUM_THREADS=32 $BTYPE; fi
+ - if [ "$TARGET_BOX" == "LINUX32" ] || [ "$TARGET_BOX" == "LINUX64" ]; then make -C utest DYNAMIC_ARCH=1 TARGET=NEHALEM NUM_THREADS=32 $BTYPE; fi
 
 # whitelist
 branches:

CMakeLists.txt

@@ -6,7 +6,7 @@ cmake_minimum_required(VERSION 2.8.4)
 project(OpenBLAS)
 set(OpenBLAS_MAJOR_VERSION 0)
 set(OpenBLAS_MINOR_VERSION 2)
-set(OpenBLAS_PATCH_VERSION 16.dev)
+set(OpenBLAS_PATCH_VERSION 16.rc1)
 set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
 
 enable_language(ASM)
@@ -54,10 +54,6 @@ if (NOT DYNAMIC_ARCH)
   list(APPEND BLASDIRS kernel)
 endif ()
 
-if (DEFINED UTEST_CHECK)
-  set(SANITY_CHECK 1)
-endif ()
-
 if (DEFINED SANITY_CHECK)
   list(APPEND BLASDIRS reference)
 endif ()
@@ -110,6 +106,10 @@ if (${NO_STATIC} AND ${NO_SHARED})
   message(FATAL_ERROR "Neither static nor shared are enabled.")
 endif ()
 
+#Set default output directory
+set( CMAKE_LIBRARY_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/lib )
+set( CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/lib )
+
 # get obj vars into format that add_library likes: $<TARGET_OBJS:objlib> (see http://www.cmake.org/cmake/help/v3.0/command/add_library.html)
 set(TARGET_OBJS "")
 foreach (SUBDIR ${SUBDIRS})
@@ -139,6 +139,17 @@ add_library(${OpenBLAS_LIBNAME} SHARED ${LA_SOURCES} ${LAPACKE_SOURCES} ${TARGET
 
 include("${CMAKE_SOURCE_DIR}/cmake/export.cmake")
 
+# Set output for libopenblas
+set_target_properties( ${OpenBLAS_LIBNAME} PROPERTIES RUNTIME_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/lib)
+foreach (OUTPUTCONFIG ${CMAKE_CONFIGURATION_TYPES})
+  string( TOUPPER ${OUTPUTCONFIG} OUTPUTCONFIG )
+  set_target_properties( ${OpenBLAS_LIBNAME} PROPERTIES RUNTIME_OUTPUT_DIRECTORY_${OUTPUTCONFIG} ${PROJECT_BINARY_DIR}/lib)
+  set_target_properties( ${OpenBLAS_LIBNAME} PROPERTIES LIBRARY_OUTPUT_DIRECTORY_${OUTPUTCONFIG} ${PROJECT_BINARY_DIR}/lib)
+  set_target_properties( ${OpenBLAS_LIBNAME} PROPERTIES ARCHIVE_OUTPUT_DIRECTORY_${OUTPUTCONFIG} ${PROJECT_BINARY_DIR}/lib)
+endforeach()
+
+enable_testing()
+add_subdirectory(utest)
 
 if(NOT MSVC)
 #only build shared library for MSVC
@@ -152,7 +163,6 @@ target_link_libraries(${OpenBLAS_LIBNAME}_static pthread)
 endif()
 
 #build test and ctest
-enable_testing()
 add_subdirectory(test)
 if(NOT NO_CBLAS)
 add_subdirectory(ctest)

CONTRIBUTORS.md

@@ -121,6 +121,17 @@ In chronological order:
   * [2014-10-10] trmm and sgemm kernels (optimized for APM's X-Gene 1).
                  ARMv8 support.
 
+* Jerome Robert <jeromerobert@gmx.com>
+  * [2015-01-01] Speed-up small `ger` and `gemv` using stack allocation (bug #478)
+  * [2015-12-23] `stack_check` in `gemv.c` (bug #722)
+  * [2015-12-28] Allow to force the number of parallel make job
+  * [2015-12-28] Fix detection of AMD E2-3200 detection
+  * [2015-12-31] Let `make MAX_STACK_ALLOC=0` do what expected
+  * [2016-01-19] Disable multi-threading in `ger` and `swap` for small matrices (bug #731)
+  * [2016-01-24] Use `GEMM_MULTITHREAD_THRESHOLD` as a number of ops (bug #742)
+  * [2016-01-26] Let `openblas_get_num_threads` return the number of active threads (bug #760)
+  * [2016-01-30] Speed-up small `zger`, `zgemv`, `ztrmv` using stack allocation (bug #727)
+
 * Dan Kortschak
   * [2015-01-07] Added test for drotmg bug #484.
 

Makefile

@@ -7,10 +7,6 @@ ifneq ($(DYNAMIC_ARCH), 1)
 BLASDIRS += kernel
 endif
 
-ifdef UTEST_CHECK
-SANITY_CHECK = 1
-endif
-
 ifdef SANITY_CHECK
 BLASDIRS += reference
 endif
@@ -85,7 +81,7 @@ endif
 
 shared :
 ifndef NO_SHARED
-ifeq ($(OSNAME), Linux)
+ifeq ($(OSNAME), $(filter $(OSNAME),Linux SunOS))
 	@$(MAKE) -C exports so
 	@-ln -fs $(LIBSONAME) $(LIBPREFIX).so
 	@-ln -fs $(LIBSONAME) $(LIBPREFIX).so.$(MAJOR_VERSION)
@@ -117,10 +113,8 @@ ifndef CROSS
 	touch $(LIBNAME)
 ifndef NO_FBLAS
 	$(MAKE) -C test all
-ifdef UTEST_CHECK
 	$(MAKE) -C utest all
 endif
-endif
 ifndef NO_CBLAS
 	$(MAKE) -C ctest all
 endif
@@ -263,6 +257,9 @@ endif
 else
 	-@echo "TIMER       = NONE" >> $(NETLIB_LAPACK_DIR)/make.inc
 	-@echo "LOADER      = $(FC)" >> $(NETLIB_LAPACK_DIR)/make.inc
+endif
+ifeq ($(BUILD_LAPACK_DEPRECATED), 1)
+	-@echo "BUILD_DEPRECATED      = 1" >> $(NETLIB_LAPACK_DIR)/make.inc
 endif
 	-@cat  make.inc >> $(NETLIB_LAPACK_DIR)/make.inc
 endif

Makefile.arm

@@ -11,8 +11,8 @@ endif
 
 ifeq ($(CORE), ARMV7)
 ifeq ($(OSNAME), Android)
-CCOMMON_OPT += -marm -mfpu=neon  -mfloat-abi=hard -march=armv7-a
-FCOMMON_OPT += -marm -mfpu=neon  -mfloat-abi=hard -march=armv7-a
+CCOMMON_OPT += -marm -mfpu=neon  -mfloat-abi=hard -march=armv7-a -Wl,--no-warn-mismatch
+FCOMMON_OPT += -marm -mfpu=neon  -mfloat-abi=hard -march=armv7-a -Wl,--no-warn-mismatch
 else
 CCOMMON_OPT += -marm -mfpu=vfpv3  -mfloat-abi=hard -march=armv7-a
 FCOMMON_OPT += -marm -mfpu=vfpv3  -mfloat-abi=hard -march=armv7-a
@@ -29,5 +29,3 @@ ifeq ($(CORE), ARMV5)
 CCOMMON_OPT += -marm -march=armv5
 FCOMMON_OPT += -marm -march=armv5
 endif
-
-

Makefile.install

@@ -29,7 +29,7 @@ install : 	lib.grd
 #for inc
 	@echo \#ifndef OPENBLAS_CONFIG_H > $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/openblas_config.h
 	@echo \#define OPENBLAS_CONFIG_H >> $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/openblas_config.h
-	@awk 'NF {print $$1, "OPENBLAS_"$$2, $$3}' config_last.h >> $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/openblas_config.h
+	@$(AWK) 'NF {print $$1, "OPENBLAS_"$$2, $$3}' config_last.h >> $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/openblas_config.h
 	@echo \#define OPENBLAS_VERSION \" OpenBLAS $(VERSION) \" >> $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/openblas_config.h
 	@cat openblas_config_template.h >> $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/openblas_config.h
 	@echo \#endif  \/\* OPENBLAS_CONFIG_H \*\/ >> $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/openblas_config.h
@@ -48,10 +48,10 @@ endif
 
 ifndef NO_LAPACKE
 	@echo Copying LAPACKE header files to $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)
-	@-install -pm644 $(NETLIB_LAPACK_DIR)/lapacke/include/lapacke.h $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke.h
-	@-install -pm644 $(NETLIB_LAPACK_DIR)/lapacke/include/lapacke_config.h $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_config.h
-	@-install -pm644 $(NETLIB_LAPACK_DIR)/lapacke/include/lapacke_mangling_with_flags.h $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_mangling.h
-	@-install -pm644 $(NETLIB_LAPACK_DIR)/lapacke/include/lapacke_utils.h $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_utils.h
+	@-install -pm644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke.h $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke.h
+	@-install -pm644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke_config.h $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_config.h
+	@-install -pm644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke_mangling_with_flags.h $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_mangling.h
+	@-install -pm644 $(NETLIB_LAPACK_DIR)/LAPACKE/include/lapacke_utils.h $(DESTDIR)$(OPENBLAS_INCLUDE_DIR)/lapacke_utils.h
 endif
 
 #for install static library
@@ -64,7 +64,7 @@ endif
 #for install shared library
 ifndef NO_SHARED
 	@echo Copying the shared library to $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
-ifeq ($(OSNAME), Linux)
+ifeq ($(OSNAME), $(filter $(OSNAME),Linux SunOS))
 	@install -pm755 $(LIBSONAME) $(DESTDIR)$(OPENBLAS_LIBRARY_DIR)
 	@cd $(DESTDIR)$(OPENBLAS_LIBRARY_DIR) ; \
 	ln -fs $(LIBSONAME) $(LIBPREFIX).so ; \

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.2.16.dev
+VERSION = 0.2.16.rc1
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library
@@ -79,8 +79,11 @@ VERSION = 0.2.16.dev
 # If you don't need LAPACKE (C Interface to LAPACK), please comment it in.
 # NO_LAPACKE = 1
 
+# Build LAPACK Deprecated functions since LAPACK 3.6.0
+# BUILD_LAPACK_DEPRECATED = 1
+
 # If you want to use legacy threaded Level 3 implementation.
- USE_SIMPLE_THREADED_LEVEL3 = 1
+# USE_SIMPLE_THREADED_LEVEL3 = 1
 
 # If you want to drive whole 64bit region by BLAS. Not all Fortran
 # compiler supports this. It's safe to keep comment it out if you
@@ -108,6 +111,10 @@ NO_AFFINITY = 1
 # Don't use parallel make.
 # NO_PARALLEL_MAKE = 1
 
+# Force number of make jobs. The default is the number of logical CPU of the host.
+# This is particularly useful when using distcc
+# MAKE_NB_JOBS = 2
+
 # If you would like to know minute performance report of GotoBLAS.
 # FUNCTION_PROFILE = 1
 
@@ -138,10 +145,6 @@ NO_AFFINITY = 1
 # slow (Not implemented yet).
 # SANITY_CHECK = 1
 
-# Run testcases in utest/ . When you enable UTEST_CHECK, it would enable
-# SANITY_CHECK to compare the result with reference BLAS.
-# UTEST_CHECK = 1
-
 # The installation directory.
 # PREFIX = /opt/OpenBLAS
 
@@ -159,10 +162,11 @@ COMMON_PROF = -pg
 # Build Debug version
 # DEBUG = 1
 
-# Improve GEMV and GER for small matrices by stack allocation.
-# For details, https://github.com/xianyi/OpenBLAS/pull/482
+# Set maximum stack allocation.
+# The default value is 2048. 0 disable stack allocation a may reduce GER and GEMV
+# performance. For details, https://github.com/xianyi/OpenBLAS/pull/482
 #
- MAX_STACK_ALLOC=2048
+# MAX_STACK_ALLOC = 0
 
 # Add a prefix or suffix to all exported symbol names in the shared library.
 # Avoid conflicts with other BLAS libraries, especially when using

Makefile.system

@@ -139,6 +139,10 @@ NO_PARALLEL_MAKE=0
 endif
 GETARCH_FLAGS	+= -DNO_PARALLEL_MAKE=$(NO_PARALLEL_MAKE)
 
+ifdef MAKE_NB_JOBS
+GETARCH_FLAGS += -DMAKE_NB_JOBS=$(MAKE_NB_JOBS)
+endif
+
 ifeq ($(HOSTCC), loongcc)
 GETARCH_FLAGS  += -static
 endif
@@ -292,12 +296,14 @@ endif
 ifneq ($(OSNAME), WINNT)
 ifneq ($(OSNAME), CYGWIN_NT)
 ifneq ($(OSNAME), Interix)
+ifneq ($(OSNAME), Android)
 ifdef SMP
 EXTRALIB   += -lpthread
 endif
 endif
 endif
 endif
+endif
 
 # ifeq logical or
 ifeq ($(OSNAME), $(filter $(OSNAME),WINNT CYGWIN_NT Interix))
@@ -324,7 +330,8 @@ ifdef SANITY_CHECK
 CCOMMON_OPT	+= -DSANITY_CHECK -DREFNAME=$(*F)f$(BU)
 endif
 
-ifdef MAX_STACK_ALLOC
+MAX_STACK_ALLOC ?= 2048
+ifneq ($(MAX_STACK_ALLOC), 0)
 CCOMMON_OPT	+= -DMAX_STACK_ALLOC=$(MAX_STACK_ALLOC)
 endif
 
@@ -374,7 +381,7 @@ FCOMMON_OPT	+= -m128bit-long-double
 endif
 ifeq ($(C_COMPILER), CLANG)
 EXPRECISION	= 1
-CCOMMON_OPT	+= -DEXPRECISION 
+CCOMMON_OPT	+= -DEXPRECISION
 FCOMMON_OPT	+= -m128bit-long-double
 endif
 endif
@@ -388,7 +395,7 @@ endif
 
 ifeq ($(USE_OPENMP), 1)
 
-#check 
+#check
 ifeq ($(USE_THREAD), 0)
 $(error OpenBLAS: Cannot set both USE_OPENMP=1 and USE_THREAD=0. The USE_THREAD=0 is only for building single thread version.)
 endif
@@ -952,17 +959,18 @@ ifeq ($(OSNAME), SunOS)
 TAR	= gtar
 PATCH	= gpatch
 GREP	= ggrep
+AWK	= nawk
 else
 TAR	= tar
 PATCH	= patch
 GREP	= grep
+AWK	= awk
 endif
 
 ifndef MD5SUM
 MD5SUM	= md5sum
 endif
 
-AWK	= awk
 
 REVISION = -r$(VERSION)
 MAJOR_VERSION = $(word 1,$(subst ., ,$(VERSION)))
@@ -980,12 +988,8 @@ COMMON_OPT = -O2
 endif
 
 ifndef FCOMMON_OPT
-ifeq ($(OSNAME), WINNT)
-FCOMMON_OPT = -O0 
-else
 FCOMMON_OPT = -O2 -frecursive
 endif
-endif
 
 
 
@@ -1183,4 +1187,3 @@ SUNPATH		= /opt/sunstudio12.1
 else
 SUNPATH		= /opt/SUNWspro
 endif
-

README.md

@@ -75,10 +75,11 @@ Please read GotoBLAS_01Readme.txt
 
 #### ARM64:
 - **ARMV8**: Experimental
+- **ARM Cortex-A57**: Experimental
 
 ### Support OS:
 - **GNU/Linux**
-- **MingWin/Windows**: Please read <https://github.com/xianyi/OpenBLAS/wiki/How-to-use-OpenBLAS-in-Microsoft-Visual-Studio>.
+- **MingWin or Visual Studio(CMake)/Windows**: Please read <https://github.com/xianyi/OpenBLAS/wiki/How-to-use-OpenBLAS-in-Microsoft-Visual-Studio>.
 - **Darwin/Mac OS X**: Experimental. Although GotoBLAS2 supports Darwin, we are the beginner on Mac OS X.
 - **FreeBSD**: Supported by community. We didn't test the library on this OS.
 

appveyor.yml

@@ -39,4 +39,6 @@ before_build:
   - cmake -G "Visual Studio 12 Win64" .
 
 test_script:
-  - echo Build OK!
+  - echo Running Test
+  - cd c:\projects\OpenBLAS\utest
+  - openblas_utest

benchmark/Makefile

@@ -166,7 +166,8 @@ goto :: slinpack.goto dlinpack.goto clinpack.goto zlinpack.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
+       ssymm.goto dsymm.goto csymm.goto zsymm.goto \
+       smallscaling
 
 acml :: slinpack.acml dlinpack.acml clinpack.acml zlinpack.acml \
        scholesky.acml dcholesky.acml ccholesky.acml zcholesky.acml \
@@ -2132,6 +2133,8 @@ cgemm3m.$(SUFFIX) : gemm3m.c
 zgemm3m.$(SUFFIX) : gemm3m.c
 	$(CC) $(CFLAGS) -c -DCOMPLEX -DDOUBLE -o $(@F) $^
 
+smallscaling: smallscaling.c ../$(LIBNAME)
+	$(CC) $(CFLAGS) -lpthread -fopenmp -lm -o $(@F) $^
 
 clean ::
 	@rm -f *.goto *.mkl *.acml *.atlas *.veclib

benchmark/scripts/SCIPY/dsyrk.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python
+
+import os
+import sys
+import time
+import numpy
+from numpy import zeros
+from numpy.random import randn
+from scipy.linalg import blas
+
+
+def run_dsyrk(N, l):
+
+    A = randn(N, N).astype('float64', order='F')
+    C = zeros((N, N), dtype='float64', order='F')
+
+    start = time.time()
+    for i in range(0, l):
+        blas.dsyrk(1.0, A, c=C, overwrite_c=True)
+    end = time.time()
+
+    timediff = (end - start)
+    mflops = (N * N * N) * l / timediff
+    mflops *= 1e-6
+
+    size = "%dx%d" % (N, N)
+    print("%14s :\t%20f MFlops\t%20f sec" % (size, mflops, timediff))
+
+
+if __name__ == "__main__":
+    N = 128
+    NMAX = 2048
+    NINC = 128
+    LOOPS = 1
+
+    z = 0
+    for arg in sys.argv:
+        if z == 1:
+            N = int(arg)
+        elif z == 2:
+            NMAX = int(arg)
+        elif z == 3:
+            NINC = int(arg)
+        elif z == 4:
+            LOOPS = int(arg)
+
+        z = z + 1
+
+    if 'OPENBLAS_LOOPS' in os.environ:
+        p = os.environ['OPENBLAS_LOOPS']
+        if p:
+            LOOPS = int(p)
+
+    print("From: %d To: %d Step=%d Loops=%d" % (N, NMAX, NINC, LOOPS))
+    print("\tSIZE\t\t\tFlops\t\t\t\t\tTime")
+
+    for i in range(N, NMAX + NINC, NINC):
+        run_dsyrk(i, LOOPS)

benchmark/scripts/SCIPY/ssyrk.py

@@ -0,0 +1,58 @@
+#!/usr/bin/env python
+
+import os
+import sys
+import time
+import numpy
+from numpy import zeros
+from numpy.random import randn
+from scipy.linalg import blas
+
+
+def run_ssyrk(N, l):
+
+    A = randn(N, N).astype('float32', order='F')
+    C = zeros((N, N), dtype='float32', order='F')
+
+    start = time.time()
+    for i in range(0, l):
+        blas.ssyrk(1.0, A, c=C, overwrite_c=True)
+    end = time.time()
+
+    timediff = (end - start)
+    mflops = (N * N * N) * l / timediff
+    mflops *= 1e-6
+
+    size = "%dx%d" % (N, N)
+    print("%14s :\t%20f MFlops\t%20f sec" % (size, mflops, timediff))
+
+
+if __name__ == "__main__":
+    N = 128
+    NMAX = 2048
+    NINC = 128
+    LOOPS = 1
+
+    z = 0
+    for arg in sys.argv:
+        if z == 1:
+            N = int(arg)
+        elif z == 2:
+            NMAX = int(arg)
+        elif z == 3:
+            NINC = int(arg)
+        elif z == 4:
+            LOOPS = int(arg)
+
+        z = z + 1
+
+    if 'OPENBLAS_LOOPS' in os.environ:
+        p = os.environ['OPENBLAS_LOOPS']
+        if p:
+            LOOPS = int(p)
+
+    print("From: %d To: %d Step=%d Loops=%d" % (N, NMAX, NINC, LOOPS))
+    print("\tSIZE\t\t\tFlops\t\t\t\t\tTime")
+
+    for i in range(N, NMAX + NINC, NINC):
+        run_ssyrk(i, LOOPS)

benchmark/smallscaling.c

@@ -0,0 +1,191 @@
+// run with OPENBLAS_NUM_THREADS=1 and OMP_NUM_THREADS=n
+#include <math.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <time.h>
+#include <cblas.h>
+#include <omp.h>
+#define MIN_SIZE 5
+#define MAX_SIZE 60
+#define NB_SIZE 10
+
+// number of loop for a 1x1 matrix. Lower it if the test is
+// too slow on you computer.
+#define NLOOP 2e7
+
+typedef struct {
+    int matrix_size;
+    int n_loop;
+    void (* bench_func)();
+    void (* blas_func)();
+    void * (* create_matrix)(int size);
+} BenchParam;
+
+void * s_create_matrix(int size) {
+    float * r = malloc(size * sizeof(double));
+    for(int i = 0; i < size; i++)
+        r[i] = 1e3 * i / size;
+    return r;
+}
+
+void * c_create_matrix(int size) {
+    float * r = malloc(size * 2 * sizeof(double));
+    for(int i = 0; i < 2 * size; i++)
+        r[i] = 1e3 * i / size;
+    return r;
+}
+
+void * z_create_matrix(int size) {
+    double * r = malloc(size * 2 * sizeof(double));
+    for(int i = 0; i < 2 * size; i++)
+        r[i] = 1e3 * i / size;
+    return r;
+}
+
+void * d_create_matrix(int size) {
+    double * r = malloc(size * sizeof(double));
+    for(int i = 0; i < size; i++)
+        r[i] = 1e3 * i / size;
+    return r;
+}
+
+void trmv_bench(BenchParam * param)
+{
+    int i, n;
+    int size = param->matrix_size;
+    n = param->n_loop / size;
+    int one = 1;
+    void * A = param->create_matrix(size * size);
+    void * y = param->create_matrix(size);
+    for(i = 0; i < n; i++) {
+        param->blas_func("U", "N", "N", &size, A, &size, y, &one);
+    }
+    free(A);
+    free(y);
+}
+
+void gemv_bench(BenchParam * param)
+{
+    int i, n;
+    int size = param->matrix_size;
+    n = param->n_loop / size;
+    double v = 1.01;
+    int one = 1;
+    void * A = param->create_matrix(size * size);
+    void * y = param->create_matrix(size);
+    for(i = 0; i < n; i++) {
+        param->blas_func("N", &size, &size, &v, A, &size, y, &one, &v, y, &one);
+    }
+    free(A);
+    free(y);
+}
+
+void ger_bench(BenchParam * param) {
+    int i, n;
+    int size = param->matrix_size;
+    n = param->n_loop / size;
+    double v = 1.01;
+    int one = 1;
+    void * A = param->create_matrix(size * size);
+    void * y = param->create_matrix(size);
+    for(i = 0; i < n; i++) {
+        param->blas_func(&size, &size, &v, y, &one, y, &one, A, &size);
+    }
+    free(A);
+    free(y);
+}
+
+#ifndef _WIN32
+void * pthread_func_wrapper(void * param) {
+    ((BenchParam *)param)->bench_func(param);
+    pthread_exit(NULL);
+}
+#endif
+
+#define NB_TESTS 5
+void * TESTS[4 * NB_TESTS] = {
+    trmv_bench, ztrmv_, z_create_matrix, "ztrmv",
+    gemv_bench, dgemv_, d_create_matrix, "dgemv",
+    gemv_bench, zgemv_, z_create_matrix, "zgemv",
+    ger_bench, dger_, d_create_matrix, "dger",
+    ger_bench, zgerc_, z_create_matrix, "zgerc",
+};
+
+inline static double delta_time(struct timespec tick) {
+    struct timespec tock;
+	clock_gettime(CLOCK_MONOTONIC, &tock);
+	return (tock.tv_sec - tick.tv_sec) + (tock.tv_nsec - tick.tv_nsec) / 1e9;
+}
+
+double pthread_bench(BenchParam * param, int nb_threads)
+{
+#ifdef _WIN32
+    return 0;
+#else
+    BenchParam threaded_param = *param;
+    pthread_t threads[nb_threads];
+    int t, rc;
+    struct timespec tick;
+    threaded_param.n_loop /= nb_threads;
+    clock_gettime(CLOCK_MONOTONIC, &tick);
+    for(t=0; t<nb_threads; t++){
+        rc = pthread_create(&threads[t], NULL, pthread_func_wrapper, &threaded_param);
+        if (rc){
+            printf("ERROR; return code from pthread_create() is %d\n", rc);
+            exit(-1);
+        }
+    }
+    for(t=0; t<nb_threads; t++){
+        pthread_join(threads[t], NULL);
+    }
+	return delta_time(tick);
+#endif
+}
+
+double seq_bench(BenchParam * param) {
+    struct timespec tick;
+    clock_gettime(CLOCK_MONOTONIC, &tick);
+    param->bench_func(param);
+    return delta_time(tick);
+}
+
+double omp_bench(BenchParam * param) {
+    BenchParam threaded_param = *param;
+    struct timespec tick;
+    int t;
+    int nb_threads = omp_get_max_threads();
+    threaded_param.n_loop /= nb_threads;
+    clock_gettime(CLOCK_MONOTONIC, &tick);
+    #pragma omp parallel for
+    for(t = 0; t < nb_threads; t ++){
+        param->bench_func(&threaded_param);
+    }
+    return delta_time(tick);
+}
+
+int main(int argc, char * argv[]) {
+    double inc_factor = exp(log((double)MAX_SIZE / MIN_SIZE) / NB_SIZE);
+    BenchParam param;
+    int test_id;
+    printf ("Running on %d threads\n", omp_get_max_threads());
+    for(test_id = 0; test_id < NB_TESTS; test_id ++) {
+        double size = MIN_SIZE;
+        param.bench_func = TESTS[test_id * 4];
+        param.blas_func = TESTS[test_id * 4 + 1];
+        param.create_matrix = TESTS[test_id * 4 + 2];
+        printf("\nBenchmark of %s\n", (char*)TESTS[test_id * 4 + 3]);
+        param.n_loop = NLOOP;
+        while(size <= MAX_SIZE) {
+            param.matrix_size = (int)(size + 0.5);
+            double seq_time = seq_bench(&param);
+            double omp_time = omp_bench(&param);
+            double pthread_time = pthread_bench(&param, omp_get_max_threads());
+            printf("matrix size %d, sequential %gs, openmp %gs, speedup %g, "
+                   "pthread %gs, speedup %g\n",
+                   param.matrix_size, seq_time,
+                   omp_time, seq_time / omp_time,
+                   pthread_time, seq_time / pthread_time);
+            size *= inc_factor;
+        }
+    }
+}

cmake/arch.cmake

@@ -14,12 +14,12 @@ if (${ARCH} STREQUAL "x86" OR ${ARCH} STREQUAL "x86_64")
   if (NOT NO_EXPRECISION)
     if (${F_COMPILER} MATCHES "GFORTRAN")
       # N.B. I'm not sure if CMake differentiates between GCC and LSB -hpa
-      if (${CMAKE_C_COMPILER} STREQUAL "GNU" OR ${CMAKE_C_COMPILER} STREQUAL "LSB")
+      if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU" OR ${CMAKE_C_COMPILER_ID} STREQUAL "LSB")
         set(EXPRECISION	1)
         set(CCOMMON_OPT "${CCOMMON_OPT} -DEXPRECISION -m128bit-long-double")
         set(FCOMMON_OPT	"${FCOMMON_OPT} -m128bit-long-double")
       endif ()
-      if (${CMAKE_C_COMPILER} STREQUAL "Clang")
+      if (${CMAKE_C_COMPILER_ID} STREQUAL "Clang")
         set(EXPRECISION	1)
         set(CCOMMON_OPT "${CCOMMON_OPT} -DEXPRECISION")
         set(FCOMMON_OPT	"${FCOMMON_OPT} -m128bit-long-double")
@@ -28,35 +28,35 @@ if (${ARCH} STREQUAL "x86" OR ${ARCH} STREQUAL "x86_64")
   endif ()
 endif ()
 
-if (${CMAKE_C_COMPILER} STREQUAL "Intel")
+if (${CMAKE_C_COMPILER_ID} STREQUAL "Intel")
   set(CCOMMON_OPT "${CCOMMON_OPT} -wd981")
 endif ()
 
 if (USE_OPENMP)
 
-  if (${CMAKE_C_COMPILER} STREQUAL "GNU" OR ${CMAKE_C_COMPILER} STREQUAL "LSB")
+  if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU" OR ${CMAKE_C_COMPILER_ID} STREQUAL "LSB")
     set(CCOMMON_OPT "${CCOMMON_OPT} -fopenmp")
   endif ()
 
-  if (${CMAKE_C_COMPILER} STREQUAL "Clang")
+  if (${CMAKE_C_COMPILER_ID} STREQUAL "Clang")
     message(WARNING "Clang doesn't support OpenMP yet.")
     set(CCOMMON_OPT "${CCOMMON_OPT} -fopenmp")
   endif ()
 
-  if (${CMAKE_C_COMPILER} STREQUAL "Intel")
+  if (${CMAKE_C_COMPILER_ID} STREQUAL "Intel")
     set(CCOMMON_OPT "${CCOMMON_OPT} -openmp")
   endif ()
 
-  if (${CMAKE_C_COMPILER} STREQUAL "PGI")
+  if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI")
     set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
   endif ()
 
-  if (${CMAKE_C_COMPILER} STREQUAL "OPEN64")
+  if (${CMAKE_C_COMPILER_ID} STREQUAL "OPEN64")
     set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
     set(CEXTRALIB "${CEXTRALIB} -lstdc++")
   endif ()
 
-  if (${CMAKE_C_COMPILER} STREQUAL "PATHSCALE")
+  if (${CMAKE_C_COMPILER_ID} STREQUAL "PATHSCALE")
     set(CCOMMON_OPT "${CCOMMON_OPT} -mp")
   endif ()
 endif ()
@@ -87,7 +87,7 @@ if (${ARCH} STREQUAL "ia64")
   set(BINARY_DEFINED 1)
 
   if (${F_COMPILER} MATCHES "GFORTRAN")
-    if (${CMAKE_C_COMPILER} STREQUAL "GNU")
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
       # EXPRECISION	= 1
       # CCOMMON_OPT	+= -DEXPRECISION
     endif ()

cmake/lapack.cmake

@@ -48,18 +48,18 @@ set(SLASRC
   sgbbrd.f sgbcon.f sgbequ.f sgbrfs.f sgbsv.f
   sgbsvx.f sgbtf2.f sgbtrf.f sgbtrs.f sgebak.f sgebal.f sgebd2.f
   sgebrd.f sgecon.f sgeequ.f sgees.f  sgeesx.f sgeev.f  sgeevx.f
-  sgegs.f  sgegv.f  sgehd2.f sgehrd.f sgelq2.f sgelqf.f
-  sgels.f  sgelsd.f sgelss.f sgelsx.f sgelsy.f sgeql2.f sgeqlf.f
-  sgeqp3.f sgeqpf.f sgeqr2.f sgeqr2p.f sgeqrf.f sgeqrfp.f sgerfs.f
+  DEPRECATED/sgegs.f  DEPRECATED/sgegv.f  sgehd2.f sgehrd.f sgelq2.f sgelqf.f
+  sgels.f  sgelsd.f sgelss.f DEPRECATED/sgelsx.f sgelsy.f sgeql2.f sgeqlf.f
+  sgeqp3.f DEPRECATED/sgeqpf.f sgeqr2.f sgeqr2p.f sgeqrf.f sgeqrfp.f sgerfs.f
   sgerq2.f sgerqf.f sgesc2.f sgesdd.f  sgesvd.f sgesvx.f
   sgetc2.f sgetri.f
   sggbak.f sggbal.f sgges.f  sggesx.f sggev.f  sggevx.f
   sggglm.f sgghrd.f sgglse.f sggqrf.f
-  sggrqf.f sggsvd.f sggsvp.f sgtcon.f sgtrfs.f sgtsv.f
+  sggrqf.f DEPRECATED/sggsvd.f DEPRECATED/sggsvp.f sgtcon.f sgtrfs.f sgtsv.f
   sgtsvx.f sgttrf.f sgttrs.f sgtts2.f shgeqz.f
   shsein.f shseqr.f slabrd.f slacon.f slacn2.f
   slaein.f slaexc.f slag2.f  slags2.f slagtm.f slagv2.f slahqr.f
-  slahrd.f slahr2.f slaic1.f slaln2.f slals0.f slalsa.f slalsd.f
+  DEPRECATED/slahrd.f slahr2.f slaic1.f slaln2.f slals0.f slalsa.f slalsd.f
   slangb.f slange.f slangt.f slanhs.f slansb.f slansp.f
   slansy.f slantb.f slantp.f slantr.f slanv2.f
   slapll.f slapmt.f
@@ -69,7 +69,7 @@ set(SLASRC
   slarf.f  slarfb.f slarfg.f slarfgp.f slarft.f slarfx.f slargv.f
   slarrv.f slartv.f
   slarz.f  slarzb.f slarzt.f slasy2.f slasyf.f slasyf_rook.f
-  slatbs.f slatdf.f slatps.f slatrd.f slatrs.f slatrz.f slatzm.f
+  slatbs.f slatdf.f slatps.f slatrd.f slatrs.f slatrz.f DEPRECATED/slatzm.f
   sopgtr.f sopmtr.f sorg2l.f sorg2r.f
   sorgbr.f sorghr.f sorgl2.f sorglq.f sorgql.f sorgqr.f sorgr2.f
   sorgrq.f sorgtr.f sorm2l.f sorm2r.f
@@ -97,7 +97,7 @@ set(SLASRC
   stgsja.f stgsna.f stgsy2.f stgsyl.f stpcon.f stprfs.f stptri.f
   stptrs.f
   strcon.f strevc.f strexc.f strrfs.f strsen.f strsna.f strsyl.f
-  strtrs.f stzrqf.f stzrzf.f sstemr.f
+  strtrs.f DEPRECATED/stzrqf.f stzrzf.f sstemr.f
   slansf.f spftrf.f spftri.f spftrs.f ssfrk.f stfsm.f stftri.f stfttp.f
   stfttr.f stpttf.f stpttr.f strttf.f strttp.f
   sgejsv.f  sgesvj.f  sgsvj0.f  sgsvj1.f
@@ -114,14 +114,14 @@ set(CLASRC
   cbdsqr.f cgbbrd.f cgbcon.f cgbequ.f cgbrfs.f cgbsv.f  cgbsvx.f
   cgbtf2.f cgbtrf.f cgbtrs.f cgebak.f cgebal.f cgebd2.f cgebrd.f
   cgecon.f cgeequ.f cgees.f  cgeesx.f cgeev.f  cgeevx.f
-  cgegs.f  cgegv.f  cgehd2.f cgehrd.f cgelq2.f cgelqf.f
-  cgels.f  cgelsd.f cgelss.f cgelsx.f cgelsy.f cgeql2.f cgeqlf.f cgeqp3.f
-  cgeqpf.f cgeqr2.f cgeqr2p.f cgeqrf.f cgeqrfp.f cgerfs.f
+  DEPRECATED/cgegs.f  DEPRECATED/cgegv.f  cgehd2.f cgehrd.f cgelq2.f cgelqf.f
+  cgels.f  cgelsd.f cgelss.f DEPRECATED/cgelsx.f cgelsy.f cgeql2.f cgeqlf.f cgeqp3.f
+  DEPRECATED/cgeqpf.f cgeqr2.f cgeqr2p.f cgeqrf.f cgeqrfp.f cgerfs.f
   cgerq2.f cgerqf.f cgesc2.f cgesdd.f  cgesvd.f
   cgesvx.f cgetc2.f cgetri.f
   cggbak.f cggbal.f cgges.f  cggesx.f cggev.f  cggevx.f cggglm.f
   cgghrd.f cgglse.f cggqrf.f cggrqf.f
-  cggsvd.f cggsvp.f
+  DEPRECATED/cggsvd.f DEPRECATED/cggsvp.f
   cgtcon.f cgtrfs.f cgtsv.f  cgtsvx.f cgttrf.f cgttrs.f cgtts2.f chbev.f
   chbevd.f chbevx.f chbgst.f chbgv.f  chbgvd.f chbgvx.f chbtrd.f
   checon.f cheev.f  cheevd.f cheevr.f cheevx.f chegs2.f chegst.f
@@ -138,7 +138,7 @@ set(CLASRC
   claed0.f claed7.f claed8.f
   claein.f claesy.f claev2.f clags2.f clagtm.f
   clahef.f clahef_rook.f clahqr.f
-  clahrd.f clahr2.f claic1.f clals0.f clalsa.f clalsd.f clangb.f clange.f clangt.f
+  DEPRECATED/clahrd.f clahr2.f claic1.f clals0.f clalsa.f clalsd.f clangb.f clange.f clangt.f
   clanhb.f clanhe.f
   clanhp.f clanhs.f clanht.f clansb.f clansp.f clansy.f clantb.f
   clantp.f clantr.f clapll.f clapmt.f clarcm.f claqgb.f claqge.f
@@ -149,7 +149,7 @@ set(CLASRC
   clarfx.f clargv.f clarnv.f clarrv.f clartg.f clartv.f
   clarz.f  clarzb.f clarzt.f clascl.f claset.f clasr.f  classq.f
   clasyf.f clasyf_rook.f clatbs.f clatdf.f clatps.f clatrd.f clatrs.f clatrz.f
-  clatzm.f cpbcon.f cpbequ.f cpbrfs.f cpbstf.f cpbsv.f
+  DEPRECATED/clatzm.f cpbcon.f cpbequ.f cpbrfs.f cpbstf.f cpbsv.f
   cpbsvx.f cpbtf2.f cpbtrf.f cpbtrs.f cpocon.f cpoequ.f cporfs.f
   cposv.f  cposvx.f cpstrf.f cpstf2.f
   cppcon.f cppequ.f cpprfs.f cppsv.f  cppsvx.f cpptrf.f cpptri.f cpptrs.f
@@ -166,7 +166,7 @@ set(CLASRC
   ctgexc.f ctgsen.f ctgsja.f ctgsna.f ctgsy2.f ctgsyl.f ctpcon.f
   ctprfs.f ctptri.f
   ctptrs.f ctrcon.f ctrevc.f ctrexc.f ctrrfs.f ctrsen.f ctrsna.f
-  ctrsyl.f ctrtrs.f ctzrqf.f ctzrzf.f cung2l.f cung2r.f
+  ctrsyl.f ctrtrs.f DEPRECATED/ctzrqf.f ctzrzf.f cung2l.f cung2r.f
   cungbr.f cunghr.f cungl2.f cunglq.f cungql.f cungqr.f cungr2.f
   cungrq.f cungtr.f cunm2l.f cunm2r.f cunmbr.f cunmhr.f cunml2.f
   cunmlq.f cunmql.f cunmqr.f cunmr2.f cunmr3.f cunmrq.f cunmrz.f
@@ -186,18 +186,18 @@ set(DLASRC
   dgbbrd.f dgbcon.f dgbequ.f dgbrfs.f dgbsv.f
   dgbsvx.f dgbtf2.f dgbtrf.f dgbtrs.f dgebak.f dgebal.f dgebd2.f
   dgebrd.f dgecon.f dgeequ.f dgees.f  dgeesx.f dgeev.f  dgeevx.f
-  dgegs.f  dgegv.f  dgehd2.f dgehrd.f dgelq2.f dgelqf.f
-  dgels.f  dgelsd.f dgelss.f dgelsx.f dgelsy.f dgeql2.f dgeqlf.f
-  dgeqp3.f dgeqpf.f dgeqr2.f dgeqr2p.f dgeqrf.f dgeqrfp.f dgerfs.f
+  DEPRECATED/dgegs.f  DEPRECATED/dgegv.f  dgehd2.f dgehrd.f dgelq2.f dgelqf.f
+  dgels.f  dgelsd.f dgelss.f DEPRECATED/dgelsx.f dgelsy.f dgeql2.f dgeqlf.f
+  dgeqp3.f DEPRECATED/dgeqpf.f dgeqr2.f dgeqr2p.f dgeqrf.f dgeqrfp.f dgerfs.f
   dgerq2.f dgerqf.f dgesc2.f dgesdd.f  dgesvd.f dgesvx.f
   dgetc2.f dgetri.f
   dggbak.f dggbal.f dgges.f  dggesx.f dggev.f  dggevx.f
   dggglm.f dgghrd.f dgglse.f dggqrf.f
-  dggrqf.f dggsvd.f dggsvp.f dgtcon.f dgtrfs.f dgtsv.f
+  dggrqf.f DEPRECATED/dggsvd.f DEPRECATED/dggsvp.f dgtcon.f dgtrfs.f dgtsv.f
   dgtsvx.f dgttrf.f dgttrs.f dgtts2.f dhgeqz.f
   dhsein.f dhseqr.f dlabrd.f dlacon.f dlacn2.f
   dlaein.f dlaexc.f dlag2.f  dlags2.f dlagtm.f dlagv2.f dlahqr.f
-  dlahrd.f dlahr2.f dlaic1.f dlaln2.f dlals0.f dlalsa.f dlalsd.f
+  DEPRECATED/dlahrd.f dlahr2.f dlaic1.f dlaln2.f dlals0.f dlalsa.f dlalsd.f
   dlangb.f dlange.f dlangt.f dlanhs.f dlansb.f dlansp.f
   dlansy.f dlantb.f dlantp.f dlantr.f dlanv2.f
   dlapll.f dlapmt.f
@@ -207,7 +207,7 @@ set(DLASRC
   dlarf.f  dlarfb.f dlarfg.f dlarfgp.f dlarft.f dlarfx.f
   dlargv.f dlarrv.f dlartv.f
   dlarz.f  dlarzb.f dlarzt.f dlasy2.f dlasyf.f dlasyf_rook.f
-  dlatbs.f dlatdf.f dlatps.f dlatrd.f dlatrs.f dlatrz.f dlatzm.f
+  dlatbs.f dlatdf.f dlatps.f dlatrd.f dlatrs.f dlatrz.f DEPRECATED/dlatzm.f
   dopgtr.f dopmtr.f dorg2l.f dorg2r.f
   dorgbr.f dorghr.f dorgl2.f dorglq.f dorgql.f dorgqr.f dorgr2.f
   dorgrq.f dorgtr.f dorm2l.f dorm2r.f
@@ -235,7 +235,7 @@ set(DLASRC
   dtgsja.f dtgsna.f dtgsy2.f dtgsyl.f dtpcon.f dtprfs.f dtptri.f
   dtptrs.f
   dtrcon.f dtrevc.f dtrexc.f dtrrfs.f dtrsen.f dtrsna.f dtrsyl.f
-  dtrtrs.f dtzrqf.f dtzrzf.f dstemr.f
+  dtrtrs.f DEPRECATED/dtzrqf.f dtzrzf.f dstemr.f
   dsgesv.f dsposv.f dlag2s.f slag2d.f dlat2s.f
   dlansf.f dpftrf.f dpftri.f dpftrs.f dsfrk.f dtfsm.f dtftri.f dtfttp.f
   dtfttr.f dtpttf.f dtpttr.f dtrttf.f dtrttp.f
@@ -251,14 +251,14 @@ set(ZLASRC
   zbdsqr.f zgbbrd.f zgbcon.f zgbequ.f zgbrfs.f zgbsv.f  zgbsvx.f
   zgbtf2.f zgbtrf.f zgbtrs.f zgebak.f zgebal.f zgebd2.f zgebrd.f
   zgecon.f zgeequ.f zgees.f  zgeesx.f zgeev.f  zgeevx.f
-  zgegs.f  zgegv.f  zgehd2.f zgehrd.f zgelq2.f zgelqf.f
-  zgels.f  zgelsd.f zgelss.f zgelsx.f zgelsy.f zgeql2.f zgeqlf.f zgeqp3.f
-  zgeqpf.f zgeqr2.f zgeqr2p.f zgeqrf.f zgeqrfp.f zgerfs.f zgerq2.f zgerqf.f
+  DEPRECATED/zgegs.f  DEPRECATED/zgegv.f  zgehd2.f zgehrd.f zgelq2.f zgelqf.f
+  zgels.f  zgelsd.f zgelss.f DEPRECATED/zgelsx.f zgelsy.f zgeql2.f zgeqlf.f zgeqp3.f
+  DEPRECATED/zgeqpf.f zgeqr2.f zgeqr2p.f zgeqrf.f zgeqrfp.f zgerfs.f zgerq2.f zgerqf.f
   zgesc2.f zgesdd.f zgesvd.f zgesvx.f zgetc2.f
   zgetri.f
   zggbak.f zggbal.f zgges.f  zggesx.f zggev.f  zggevx.f zggglm.f
   zgghrd.f zgglse.f zggqrf.f zggrqf.f
-  zggsvd.f zggsvp.f
+  DEPRECATED/zggsvd.f DEPRECATED/zggsvp.f
   zgtcon.f zgtrfs.f zgtsv.f  zgtsvx.f zgttrf.f zgttrs.f zgtts2.f zhbev.f
   zhbevd.f zhbevx.f zhbgst.f zhbgv.f  zhbgvd.f zhbgvx.f zhbtrd.f
   zhecon.f zheev.f  zheevd.f zheevr.f zheevx.f zhegs2.f zhegst.f
@@ -275,7 +275,7 @@ set(ZLASRC
   zlaed0.f zlaed7.f zlaed8.f
   zlaein.f zlaesy.f zlaev2.f zlags2.f zlagtm.f
   zlahef.f zlahef_rook.f zlahqr.f
-  zlahrd.f zlahr2.f zlaic1.f zlals0.f zlalsa.f zlalsd.f zlangb.f zlange.f
+  DEPRECATED/zlahrd.f zlahr2.f zlaic1.f zlals0.f zlalsa.f zlalsd.f zlangb.f zlange.f
   zlangt.f zlanhb.f
   zlanhe.f
   zlanhp.f zlanhs.f zlanht.f zlansb.f zlansp.f zlansy.f zlantb.f
@@ -288,7 +288,7 @@ set(ZLASRC
   zlarfx.f zlargv.f zlarnv.f zlarrv.f zlartg.f zlartv.f
   zlarz.f  zlarzb.f zlarzt.f zlascl.f zlaset.f zlasr.f
   zlassq.f zlasyf.f zlasyf_rook.f
-  zlatbs.f zlatdf.f zlatps.f zlatrd.f zlatrs.f zlatrz.f zlatzm.f
+  zlatbs.f zlatdf.f zlatps.f zlatrd.f zlatrs.f zlatrz.f DEPRECATED/zlatzm.f
   zpbcon.f zpbequ.f zpbrfs.f zpbstf.f zpbsv.f
   zpbsvx.f zpbtf2.f zpbtrf.f zpbtrs.f zpocon.f zpoequ.f zporfs.f
   zposv.f  zposvx.f zpotrs.f zpstrf.f zpstf2.f
@@ -306,7 +306,7 @@ set(ZLASRC
   ztgexc.f ztgsen.f ztgsja.f ztgsna.f ztgsy2.f ztgsyl.f ztpcon.f
   ztprfs.f ztptri.f
   ztptrs.f ztrcon.f ztrevc.f ztrexc.f ztrrfs.f ztrsen.f ztrsna.f
-  ztrsyl.f ztrtrs.f ztzrqf.f ztzrzf.f zung2l.f
+  ztrsyl.f ztrtrs.f DEPRECATED/ztzrqf.f ztzrzf.f zung2l.f
   zung2r.f zungbr.f zunghr.f zungl2.f zunglq.f zungql.f zungqr.f zungr2.f
   zungrq.f zungtr.f zunm2l.f zunm2r.f zunmbr.f zunmhr.f zunml2.f
   zunmlq.f zunmql.f zunmqr.f zunmr2.f zunmr3.f zunmrq.f zunmrz.f

cmake/lapacke.cmake

@@ -2038,6 +2038,59 @@ set(MATGEN
   lapacke_zlagsy_work.c
 )
 
+set(Utils_SRC
+lapacke_cgb_nancheck.c  lapacke_dpf_nancheck.c         lapacke_ssy_trans.c
+lapacke_cgb_trans.c     lapacke_dpf_trans.c            lapacke_stb_nancheck.c
+lapacke_cge_nancheck.c  lapacke_dpo_nancheck.c         lapacke_stb_trans.c
+lapacke_cge_trans.c     lapacke_dpo_trans.c            lapacke_stf_nancheck.c
+lapacke_cgg_nancheck.c  lapacke_dpp_nancheck.c         lapacke_stf_trans.c
+lapacke_cgg_trans.c     lapacke_dpp_trans.c            lapacke_stp_nancheck.c
+lapacke_cgt_nancheck.c  lapacke_dpt_nancheck.c         lapacke_stp_trans.c
+lapacke_chb_nancheck.c  lapacke_dsb_nancheck.c         lapacke_str_nancheck.c
+lapacke_chb_trans.c     lapacke_dsb_trans.c            lapacke_str_trans.c
+lapacke_che_nancheck.c  lapacke_dsp_nancheck.c         lapacke_xerbla.c
+lapacke_che_trans.c     lapacke_dsp_trans.c            lapacke_zgb_nancheck.c
+lapacke_chp_nancheck.c  lapacke_dst_nancheck.c         lapacke_zgb_trans.c
+lapacke_chp_trans.c     lapacke_dsy_nancheck.c         lapacke_zge_nancheck.c
+lapacke_chs_nancheck.c  lapacke_dsy_trans.c            lapacke_zge_trans.c
+lapacke_chs_trans.c     lapacke_dtb_nancheck.c         lapacke_zgg_nancheck.c
+lapacke_c_nancheck.c    lapacke_dtb_trans.c            lapacke_zgg_trans.c
+lapacke_cpb_nancheck.c  lapacke_dtf_nancheck.c         lapacke_zgt_nancheck.c
+lapacke_cpb_trans.c     lapacke_dtf_trans.c            lapacke_zhb_nancheck.c
+lapacke_cpf_nancheck.c  lapacke_dtp_nancheck.c         lapacke_zhb_trans.c
+lapacke_cpf_trans.c     lapacke_dtp_trans.c            lapacke_zhe_nancheck.c
+lapacke_cpo_nancheck.c  lapacke_dtr_nancheck.c         lapacke_zhe_trans.c
+lapacke_cpo_trans.c     lapacke_dtr_trans.c            lapacke_zhp_nancheck.c
+lapacke_cpp_nancheck.c  lapacke_lsame.c                lapacke_zhp_trans.c
+lapacke_cpp_trans.c     lapacke_make_complex_double.c  lapacke_zhs_nancheck.c
+lapacke_cpt_nancheck.c  lapacke_make_complex_float.c   lapacke_zhs_trans.c
+lapacke_csp_nancheck.c  lapacke_sgb_nancheck.c         lapacke_z_nancheck.c
+lapacke_csp_trans.c     lapacke_sgb_trans.c            lapacke_zpb_nancheck.c
+lapacke_cst_nancheck.c  lapacke_sge_nancheck.c         lapacke_zpb_trans.c
+lapacke_csy_nancheck.c  lapacke_sge_trans.c            lapacke_zpf_nancheck.c
+lapacke_csy_trans.c     lapacke_sgg_nancheck.c         lapacke_zpf_trans.c
+lapacke_ctb_nancheck.c  lapacke_sgg_trans.c            lapacke_zpo_nancheck.c
+lapacke_ctb_trans.c     lapacke_sgt_nancheck.c         lapacke_zpo_trans.c
+lapacke_ctf_nancheck.c  lapacke_shs_nancheck.c         lapacke_zpp_nancheck.c
+lapacke_ctf_trans.c     lapacke_shs_trans.c            lapacke_zpp_trans.c
+lapacke_ctp_nancheck.c  lapacke_s_nancheck.c           lapacke_zpt_nancheck.c
+lapacke_ctp_trans.c     lapacke_spb_nancheck.c         lapacke_zsp_nancheck.c
+lapacke_ctr_nancheck.c  lapacke_spb_trans.c            lapacke_zsp_trans.c
+lapacke_ctr_trans.c     lapacke_spf_nancheck.c         lapacke_zst_nancheck.c
+lapacke_dgb_nancheck.c  lapacke_spf_trans.c            lapacke_zsy_nancheck.c
+lapacke_dgb_trans.c     lapacke_spo_nancheck.c         lapacke_zsy_trans.c
+lapacke_dge_nancheck.c  lapacke_spo_trans.c            lapacke_ztb_nancheck.c
+lapacke_dge_trans.c     lapacke_spp_nancheck.c         lapacke_ztb_trans.c
+lapacke_dgg_nancheck.c  lapacke_spp_trans.c            lapacke_ztf_nancheck.c
+lapacke_dgg_trans.c     lapacke_spt_nancheck.c         lapacke_ztf_trans.c
+lapacke_dgt_nancheck.c  lapacke_ssb_nancheck.c         lapacke_ztp_nancheck.c
+lapacke_dhs_nancheck.c  lapacke_ssb_trans.c            lapacke_ztp_trans.c
+lapacke_dhs_trans.c     lapacke_ssp_nancheck.c         lapacke_ztr_nancheck.c
+lapacke_d_nancheck.c    lapacke_ssp_trans.c            lapacke_ztr_trans.c
+lapacke_dpb_nancheck.c  lapacke_sst_nancheck.c
+lapacke_dpb_trans.c     lapacke_ssy_nancheck.c
+)
+
 set(LAPACKE_REL_SRC "")
 if (BUILD_SINGLE)
   list(APPEND LAPACKE_REL_SRC ${SSRC})
@@ -2058,10 +2111,14 @@ endif ()
 # add lapack-netlib folder to the sources
 set(LAPACKE_SOURCES "")
 foreach (LAE_FILE ${LAPACKE_REL_SRC})
-  list(APPEND LAPACKE_SOURCES "${NETLIB_LAPACK_DIR}/lapacke/src/${LAE_FILE}")
+  list(APPEND LAPACKE_SOURCES "${NETLIB_LAPACK_DIR}/LAPACKE/src/${LAE_FILE}")
 endforeach ()
 
-set(lapacke_include_dir "${NETLIB_LAPACK_DIR}/lapacke/include")
+foreach (Utils_FILE ${Utils_SRC})
+  list(APPEND LAPACKE_SOURCES "${NETLIB_LAPACK_DIR}/LAPACKE/utils/${Utils_FILE}")
+endforeach ()
+
+set(lapacke_include_dir "${NETLIB_LAPACK_DIR}/LAPACKE/include")
 execute_process(COMMAND ${CMAKE_COMMAND} -E copy "${lapacke_include_dir}/lapacke_mangling_with_flags.h" "${lapacke_include_dir}/lapacke_mangling.h")
 include_directories(${lapacke_include_dir})
 set_source_files_properties(${LAPACKE_SOURCES} PROPERTIES COMPILE_FLAGS "${LAPACK_CFLAGS}")

common.h

@@ -93,7 +93,7 @@ extern "C" {
 #include <sched.h>
 #endif
 
-#if defined(OS_DARWIN) || defined(OS_FREEBSD) || defined(OS_NETBSD)
+#if defined(OS_DARWIN) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(OS_ANDROID)
 #include <sched.h>
 #endif
 
@@ -411,7 +411,7 @@ please https://github.com/xianyi/OpenBLAS/issues/246
 #ifndef ASSEMBLER
 #ifdef OS_WINDOWS
 typedef char env_var_t[MAX_PATH];
-#define readenv(p, n) GetEnvironmentVariable((n), (p), sizeof(p))
+#define readenv(p, n) GetEnvironmentVariable((LPCTSTR)(n), (LPTSTR)(p), sizeof(p))
 #else
 typedef char* env_var_t;
 #define readenv(p, n) ((p)=getenv(n))
@@ -727,6 +727,7 @@ typedef struct {
 #endif
 
 #ifndef ASSEMBLER
+#include "common_stackalloc.h"
 #if 0
 #include "symcopy.h"
 #endif

common_param.h

@@ -1194,8 +1194,6 @@ extern gotoblas_t *gotoblas;
 #define XGEMM_DEFAULT_UNROLL_N 2
 #endif
 
-#define GEMM_THREAD gemm_thread_m
-
 #ifndef GEMM_THREAD
 #define GEMM_THREAD gemm_thread_n
 #endif

common_stackalloc.h

@@ -0,0 +1,73 @@
+/*******************************************************************************
+Copyright (c) 2016, 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 STACK_ALLOC_PROTECT
+#ifdef STACK_ALLOC_PROTECT
+// Try to detect stack smashing
+#include <assert.h>
+#define STACK_ALLOC_PROTECT_SET volatile int stack_check = 0x7fc01234;
+#define STACK_ALLOC_PROTECT_CHECK assert(stack_check == 0x7fc01234);
+#else
+#define STACK_ALLOC_PROTECT_SET
+#define STACK_ALLOC_PROTECT_CHECK
+#endif
+
+#if defined(MAX_STACK_ALLOC) && MAX_STACK_ALLOC > 0
+
+/*
+ * Allocate a buffer on the stack if the size is smaller than MAX_STACK_ALLOC.
+ * Stack allocation is much faster than blas_memory_alloc or malloc, particularly
+ * when OpenBLAS is used from a multi-threaded application.
+ * SIZE must be carefully chosen to be:
+ * - as small as possible to maximize the number of stack allocation
+ * - large enough to support all architectures and kernel
+ * Chosing a too small SIZE will lead to a stack smashing.
+ */
+#define STACK_ALLOC(SIZE, TYPE, BUFFER)                                    \
+  /* make it volatile because some function (ex: dgemv_n.S) */             \
+  /* do not restore all register */                                        \
+  volatile int stack_alloc_size = SIZE;                                    \
+  if(stack_alloc_size > MAX_STACK_ALLOC / sizeof(TYPE))                    \
+    stack_alloc_size = 0;                                                  \
+  STACK_ALLOC_PROTECT_SET                                                  \
+  TYPE stack_buffer[stack_alloc_size] __attribute__((aligned(0x20)));    \
+  BUFFER = stack_alloc_size ? stack_buffer : (TYPE *)blas_memory_alloc(1);
+#else
+  //Original OpenBLAS/GotoBLAS codes.
+  #define STACK_ALLOC(SIZE, TYPE, BUFFER) BUFFER = (TYPE *)blas_memory_alloc(1)
+#endif
+
+
+#if defined(MAX_STACK_ALLOC) && MAX_STACK_ALLOC > 0
+#define STACK_FREE(BUFFER)    \
+  STACK_ALLOC_PROTECT_CHECK   \
+  if(!stack_alloc_size)       \
+    blas_memory_free(BUFFER);
+#else
+#define STACK_FREE(BUFFER) blas_memory_free(BUFFER)
+#endif
+

common_x86.h

@@ -41,6 +41,10 @@
 
 #ifndef ASSEMBLER
 
+#ifdef C_MSVC
+#include <intrin.h>
+#endif
+
 #define MB
 #define WMB
 
@@ -170,12 +174,13 @@ static __inline int blas_quickdivide(unsigned int x, unsigned int y){
 
   if (y <= 1) return x;
 
+#if defined(_MSC_VER) && !defined(__clang__)
+  result = x/y;
+  return result;
+#else
+
   y = blas_quick_divide_table[y];
 
-#if defined(_MSC_VER) && !defined(__clang__)
-  (void*)result;
-  return x*y;
-#else
   __asm__ __volatile__  ("mull %0" :"=d" (result) :"a"(x), "0" (y));
 
   return result;

cpuid_arm.c

@@ -115,6 +115,9 @@ int detect(void)
 	  if (strstr(p, "0xc0f")) {
 	    return CPU_CORTEXA15;
 	  }
+	  if (strstr(p, "0xd07")) {
+	    return CPU_ARMV7;  //ARMV8 on 32-bit
+	  }
 
 	}
 
@@ -158,6 +161,27 @@ int detect(void)
 
 
 	}
+
+  	p = (char *) NULL ;
+  	infile = fopen("/proc/cpuinfo", "r");
+
+	while (fgets(buffer, sizeof(buffer), infile))
+	{
+
+		if ((!strncmp("CPU architecture", buffer, 16)))
+		{
+			p = strchr(buffer, ':') + 2;
+			break;
+      		}
+  	}
+  	fclose(infile);
+  	if(p != NULL) {
+	  if (strstr(p, "8")) {
+	    return CPU_ARMV7;  //ARMV8 on 32-bit
+	  }
+
+	}
+
 #endif
 
 	return CPU_UNKNOWN;

cpuid_arm64.c

@@ -191,6 +191,8 @@ void get_cpuconfig(void)
 			printf("#define L2_SIZE 2097152\n");
 			printf("#define L2_LINESIZE 64\n");
 			printf("#define L2_ASSOCIATIVE 16\n");
+    			printf("#define DTB_DEFAULT_ENTRIES 64\n");
+    			printf("#define DTB_SIZE 4096\n");
 			break;
 	}
 }

cpuid_x86.c

@@ -1172,6 +1172,9 @@ int get_cpuname(void){
 #endif
           else
 	    return CPUTYPE_NEHALEM;
+	case 13:
+	  // Avoton
+	    return CPUTYPE_NEHALEM;
         }
         break;
       case 5:
@@ -1229,6 +1232,7 @@ int get_cpuname(void){
       case  2:
 	return CPUTYPE_OPTERON;
       case  1:
+      case  3:
       case 10:
 	return CPUTYPE_BARCELONA;
       case  6:
@@ -1674,6 +1678,9 @@ int get_coretype(void){
 #endif
           else
 	    return CORE_NEHALEM;
+	case 13:
+	  // Avoton
+	    return CORE_NEHALEM;
         }
         break;
       case 5:

ctest/c_cblat3.f

@@ -1365,8 +1365,9 @@
 *
   150 CONTINUE
       WRITE( NOUT, FMT = 9996 )SNAME
-      CALL CPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
-     $      M, N, ALPHA, LDA, LDB)
+      IF( TRACE )
+     $   CALL CPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
+     $         M, N, ALPHA, LDA, LDB)
 *
   160 CONTINUE
       RETURN

ctest/c_cblat3_3m.f

@@ -1365,8 +1365,9 @@
 *
   150 CONTINUE
       WRITE( NOUT, FMT = 9996 )SNAME
-      CALL CPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
-     $      M, N, ALPHA, LDA, LDB)
+      IF( TRACE )
+     $   CALL CPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
+     $         M, N, ALPHA, LDA, LDB)
 *
   160 CONTINUE
       RETURN

ctest/c_dblat3.f

@@ -1335,8 +1335,9 @@
 *
   150 CONTINUE
       WRITE( NOUT, FMT = 9996 )SNAME
-      CALL DPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
-     $      M, N, ALPHA, LDA, LDB)
+      IF( TRACE )
+     $   CALL DPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
+     $         M, N, ALPHA, LDA, LDB)
 *
   160 CONTINUE
       RETURN

ctest/c_sblat3.f

@@ -1339,8 +1339,9 @@
 *
   150 CONTINUE
       WRITE( NOUT, FMT = 9996 )SNAME
-      CALL SPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
-     $      M, N, ALPHA, LDA, LDB)
+      IF( TRACE )
+     $   CALL SPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
+     $         M, N, ALPHA, LDA, LDB)
 *
   160 CONTINUE
       RETURN

ctest/c_zblat2.f

@@ -1350,7 +1350,7 @@
 *
 *                       Call the subroutine.
 *
-                        IF( SNAME( 4: 5 ).EQ.'mv' )THEN
+                        IF( SNAME( 10: 11 ).EQ.'mv' )THEN
                            IF( FULL )THEN
                               IF( TRACE )
      $                           WRITE( NTRA, FMT = 9993 )NC, SNAME,
@@ -1376,7 +1376,7 @@
                               CALL CZTPMV( IORDER, UPLO, TRANS, DIAG,
      $                                    N, AA, XX, INCX )
                            END IF
-                        ELSE IF( SNAME( 4: 5 ).EQ.'sv' )THEN
+                        ELSE IF( SNAME( 10: 11 ).EQ.'sv' )THEN
                            IF( FULL )THEN
                               IF( TRACE )
      $                           WRITE( NTRA, FMT = 9993 )NC, SNAME,
@@ -1465,7 +1465,7 @@
                         END IF
 *
                         IF( .NOT.NULL )THEN
-                           IF( SNAME( 4: 5 ).EQ.'mv' )THEN
+                           IF( SNAME( 10: 11 ).EQ.'mv' )THEN
 *
 *                             Check the result.
 *
@@ -1473,7 +1473,7 @@
      $                                    INCX, ZERO, Z, INCX, XT, G,
      $                                    XX, EPS, ERR, FATAL, NOUT,
      $                                    .TRUE. )
-                           ELSE IF( SNAME( 4: 5 ).EQ.'sv' )THEN
+                           ELSE IF( SNAME( 10: 11 ).EQ.'sv' )THEN
 *
 *                             Compute approximation to original vector.
 *
@@ -1611,7 +1611,7 @@
 *     .. Common blocks ..
       COMMON             /INFOC/INFOT, NOUTC, OK
 *     .. Executable Statements ..
-      CONJ = SNAME( 5: 5 ).EQ.'c'
+      CONJ = SNAME( 11: 11 ).EQ.'c'
 *     Define the number of arguments.
       NARGS = 9
 *

ctest/c_zblat3.f

@@ -1366,8 +1366,9 @@
 *
   150 CONTINUE
       WRITE( NOUT, FMT = 9996 )SNAME
-      CALL ZPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
-     $      M, N, ALPHA, LDA, LDB)
+      IF( TRACE )
+     $   CALL ZPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
+     $         M, N, ALPHA, LDA, LDB)
 *
   160 CONTINUE
       RETURN

ctest/c_zblat3_3m.f

@@ -1366,8 +1366,9 @@
 *
   150 CONTINUE
       WRITE( NOUT, FMT = 9996 )SNAME
-      CALL ZPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
-     $      M, N, ALPHA, LDA, LDB)
+      IF( TRACE )
+     $   CALL ZPRCN3( NTRA, NC, SNAME, IORDER, SIDE, UPLO, TRANSA, DIAG,
+     $         M, N, ALPHA, LDA, LDB)
 *
   160 CONTINUE
       RETURN

driver/level2/trmv_thread.c

@@ -119,7 +119,7 @@ static int trmv_kernel(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n, F
 #endif
 
     x = buffer;
-    buffer += ((COMPSIZE * args -> m + 1023) & ~1023);
+    buffer += ((COMPSIZE * args -> m + 3) & ~3);
   }
 
 #ifndef TRANS
@@ -403,7 +403,7 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG incx, FLOAT *bu
 
   if (num_cpu) {
     queue[0].sa = NULL;
-    queue[0].sb = buffer + num_cpu * (((m + 255) & ~255) + 16) * COMPSIZE;
+    queue[0].sb = buffer + num_cpu * (((m + 3) & ~3) + 16) * COMPSIZE;
 
     queue[num_cpu - 1].next = NULL;
 

driver/level2/ztrmv_U.c

@@ -56,7 +56,7 @@ int CNAME(BLASLONG m, FLOAT *a, BLASLONG lda, FLOAT *b, BLASLONG incb, FLOAT *bu
 
   if (incb != 1) {
     B = buffer;
-    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 4095) & ~4095);
+    gemvbuffer = (FLOAT *)(((BLASLONG)buffer + m * sizeof(FLOAT) * 2 + 15) & ~15);
     COPY_K(m, b, incb, buffer, 1);
   }
 

driver/level3/CMakeLists.txt

@@ -48,8 +48,7 @@ foreach (float_type ${FLOAT_TYPES})
     # TRANS needs to be set/unset when CONJ is set/unset, so can't use it as a combination
     GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK" 3 "herk_N" false ${float_type})
     GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK;TRANS;CONJ" 3 "herk_C" false ${float_type})
-    GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK;THREADED_LEVEL3" 3 "herk_thread_N" false ${float_type})
-    GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK;THREADED_LEVEL3;TRANS;CONJ" 3 "herk_thread_C" false ${float_type})
+
     # Need to set CONJ for trmm and trsm
     GenerateCombinationObjects("trmm_L.c" "UPPER;UNIT" "L;N" "CONJ" 0 "trmm_LR" false ${float_type})
     GenerateCombinationObjects("trmm_L.c" "UPPER;UNIT" "L;N" "TRANSA;CONJ" 0 "trmm_LC" false ${float_type})
@@ -72,6 +71,10 @@ foreach (float_type ${FLOAT_TYPES})
     GenerateNamedObjects("zher2k_k.c" "HER2K;LOWER;TRANS;CONJ" "her2k_LC" false "" "" false ${float_type})
 
     if (SMP AND NOT USE_SIMPLE_THREADED_LEVEL3)
+      #herk
+      GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK;THREADED_LEVEL3" 3 "herk_thread_N" false ${float_type})
+      GenerateCombinationObjects("zherk_k.c" "LOWER" "U" "HERK;THREADED_LEVEL3;TRANS;CONJ" 3 "herk_thread_C" false ${float_type})
+
       #hemm
       GenerateCombinationObjects("zhemm_k.c" "LOWER" "U" "NN;THREADED_LEVEL3" 0 "hemm_thread_L" false ${float_type})
       GenerateCombinationObjects("zhemm_k.c" "LOWER" "U" "NC;RSIDE;THREADED_LEVEL3" 0 "hemm_thread_R" false ${float_type})
@@ -96,6 +99,17 @@ foreach (float_type ${FLOAT_TYPES})
 	endif()
       endif ()
     endforeach ()
+    
+    # for gemm3m
+    if(USE_GEMM3M)
+      foreach (GEMM_DEFINE ${GEMM_DEFINES})
+	string(TOLOWER ${GEMM_DEFINE} GEMM_DEFINE_LC)
+	GenerateNamedObjects("gemm3m.c" "${GEMM_DEFINE}" "gemm3m_${GEMM_DEFINE_LC}" false "" "" false ${float_type})
+	if (SMP AND NOT USE_SIMPLE_THREADED_LEVEL3)
+	  GenerateNamedObjects("gemm3m.c" "${GEMM_DEFINE};THREADED_LEVEL3" "gemm3m_thread_${GEMM_DEFINE_LC}" false "" "" false ${float_type})
+	endif ()
+      endforeach ()
+    endif()
   endif ()
 endforeach ()
 

driver/level3/level3.c

@@ -335,7 +335,9 @@ int CNAME(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
         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 >= 2*GEMM_UNROLL_N) min_jj = 2*GEMM_UNROLL_N;
+        	else
+          		if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 
 

driver/level3/level3_thread.c

@@ -367,7 +367,9 @@ static int inner_thread(blas_arg_t *args, BLASLONG *range_m, BLASLONG *range_n,
 
 	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 >= 2*GEMM_UNROLL_N) min_jj = 2*GEMM_UNROLL_N;
+		else
+			if (min_jj > GEMM_UNROLL_N) min_jj = GEMM_UNROLL_N;
 
 
 	START_RPCC();

driver/others/blas_server.c

@@ -70,7 +70,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 /*********************************************************************/
 
 #include "common.h"
-#if defined(OS_LINUX) || defined(OS_NETBSD) || defined(OS_DARWIN) || defined(OS_ANDROID)
+#if defined(OS_LINUX) || defined(OS_NETBSD) || defined(OS_DARWIN) || defined(OS_ANDROID) || defined(OS_SUNOS)
 #include <dlfcn.h>
 #include <signal.h>
 #include <sys/resource.h>
@@ -576,10 +576,12 @@ int blas_thread_init(void){
 	struct rlimit rlim;
         const char *msg = strerror(ret);
         fprintf(STDERR, "OpenBLAS blas_thread_init: pthread_create: %s\n", msg);
+#ifdef RLIMIT_NPROC
         if(0 == getrlimit(RLIMIT_NPROC, &rlim)) {
           fprintf(STDERR, "OpenBLAS blas_thread_init: RLIMIT_NPROC "
                   "%ld current, %ld max\n", (long)(rlim.rlim_cur), (long)(rlim.rlim_max));
         }
+#endif
         if(0 != raise(SIGINT)) {
           fprintf(STDERR, "OpenBLAS blas_thread_init: calling exit(3)\n");
           exit(EXIT_FAILURE);

driver/others/dynamic.c

@@ -261,6 +261,11 @@ static gotoblas_t *get_coretype(void){
 	    return &gotoblas_NEHALEM; //OS doesn't support AVX. Use old kernels.
 	  }
 	}
+	//Intel Avoton
+	if (model == 13) { 
+	  openblas_warning(FALLBACK_VERBOSE, NEHALEM_FALLBACK); 
+	  return &gotoblas_NEHALEM;
+	}	
 	return NULL;
       case 5:
 	//Intel Broadwell

driver/others/memory.c

@@ -144,7 +144,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #if defined(_MSC_VER) && !defined(__clang__)
 #define CONSTRUCTOR __cdecl
 #define DESTRUCTOR __cdecl
-#elif defined(OS_DARWIN) && defined(C_GCC)
+#elif (defined(OS_DARWIN) || defined(OS_SUNOS)) && defined(C_GCC)
 #define CONSTRUCTOR	__attribute__ ((constructor))
 #define DESTRUCTOR	__attribute__ ((destructor))
 #else
@@ -169,7 +169,7 @@ void goto_set_num_threads(int num_threads) {};
 
 #else
 
-#ifdef OS_LINUX
+#if defined(OS_LINUX) || defined(OS_SUNOS)
 #ifndef NO_AFFINITY
 int get_num_procs(void);
 #else
@@ -357,7 +357,9 @@ int openblas_get_num_threads(void) {
 #ifndef SMP
   return 1;
 #else
-  return blas_get_cpu_number();
+  // init blas_cpu_number if needed
+  blas_get_cpu_number();
+  return blas_cpu_number;
 #endif
 }
 

exports/Makefile

@@ -26,6 +26,10 @@ ifndef ONLY_CBLAS
 ONLY_CBLAS	= 0
 endif
 
+ifndef BUILD_LAPACK_DEPRECATED
+BUILD_LAPACK_DEPRECATED = 0
+endif
+
 ifeq ($(OSNAME), WINNT)
 ifeq ($(F_COMPILER), GFORTRAN)
 ifndef ONLY_CBLAS
@@ -92,17 +96,17 @@ dll  : ../$(LIBDLLNAME)
 	-Wl,--whole-archive ../$(LIBNAME) -Wl,--no-whole-archive $(FEXTRALIB) $(EXTRALIB)
 
 libopenblas.def : gensymbol
-	perl ./gensymbol win2k    $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" > $(@F)
+	perl ./gensymbol win2k    $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
 
 libgoto_hpl.def : gensymbol
-	perl ./gensymbol win2khpl $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" > $(@F)
+	perl ./gensymbol win2khpl $(ARCH) dummy $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
 
 ifeq (, $(SYMBOLPREFIX)$(SYMBOLSUFFIX))
 $(LIBDYNNAME) : ../$(LIBNAME) osx.def
 else
-../$(LIBNAME).renamed : ../$(LIBNAME) objconv.def
-	$(OBJCONV) @objconv.def ../$(LIBNAME) ../$(LIBNAME).renamed
-$(LIBDYNNAME) : ../$(LIBNAME).renamed osx.def
+../$(LIBNAME).osx.renamed : ../$(LIBNAME) objconv.def
+	$(OBJCONV) @objconv.def ../$(LIBNAME) ../$(LIBNAME).osx.renamed
+$(LIBDYNNAME) : ../$(LIBNAME).osx.renamed osx.def
 endif
 ifeq ($(NOFORTRAN), $(filter $(NOFORTRAN),1 2))
 #only build without Fortran
@@ -114,7 +118,7 @@ endif
 dllinit.$(SUFFIX) : dllinit.c
 	$(CC) $(CFLAGS) -c -o $(@F) -s $<
 
-ifeq ($(OSNAME), Linux)
+ifeq ($(OSNAME), $(filter $(OSNAME),Linux SunOS))
 
 so : ../$(LIBSONAME)
 
@@ -205,26 +209,26 @@ static : ../$(LIBNAME)
 	rm -f goto.$(SUFFIX)
 
 osx.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol osx $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" > $(@F)
+	perl ./gensymbol osx $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
 
 aix.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol aix $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" > $(@F)
+	perl ./gensymbol aix $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
 
 objcopy.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol objcopy $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" > $(@F)
+	perl ./gensymbol objcopy $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
 
 objconv.def : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol objconv $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" > $(@F)
+	perl ./gensymbol objconv $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS)  $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > $(@F)
 
 test : linktest.c
 	$(CC) $(CFLAGS) $(LDFLAGS) -w -o linktest linktest.c ../$(LIBSONAME) -lm && echo OK.
 	rm -f linktest
 
 linktest.c : gensymbol ../Makefile.system ../getarch.c
-	perl ./gensymbol linktest  $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" > linktest.c
+	perl ./gensymbol linktest  $(ARCH) $(BU) $(EXPRECISION) $(NO_CBLAS) $(NO_LAPACK) $(NO_LAPACKE) $(NEED2UNDERSCORES) $(ONLY_CBLAS) "$(SYMBOLPREFIX)" "$(SYMBOLSUFFIX)" $(BUILD_LAPACK_DEPRECATED) > linktest.c
 
 clean ::
-	@rm -f *.def *.dylib __.SYMDEF*
+	@rm -f *.def *.dylib __.SYMDEF* *.renamed
 
 include ../Makefile.tail
 

exports/gensymbol

@@ -548,7 +548,6 @@
 		slatmt,
 		sorm22,
 		spotrf2,
-		xerbla,
 		zgejsv,
 		zgesvdx,
 		zgesvj,
@@ -591,6 +590,13 @@
                 dlagsy, dsysvxx, sporfsx, slatms, zlatms, zherfsx, csysvxx,
                 );
 
+@lapack_deprecated_objs = (
+    cgegs,   cggsvd,  ctzrqf,  dgeqpf,  dlatzm,  sgelsx,  slahrd,  zgegv,   zggsvp,
+    cgegv,   cggsvp,  dgegs,   dggsvd,  dtzrqf,  sgeqpf,  slatzm,  zgelsx,  zlahrd,
+    cgelsx,  clahrd,  dgegv,   dggsvp,  sgegs,   sggsvd,  stzrqf,  zgeqpf,  zlatzm,
+    cgeqpf,  clatzm,  dgelsx,  dlahrd,  sgegv,   sggsvp,  zgegs,   zggsvd,  ztzrqf,
+    );
+
 @lapackeobjs = (
 		# LAPACK C interface routines.
 		#
@@ -2985,6 +2991,11 @@ if ($ARGV[8] == 1) {
 	@need_2underscore_objs = (@lapack_embeded_underscore_objs);
     };
 
+    if ($ARGV[11] == 1){
+	#BUILD_LAPACK_DEPRECATED=1
+	@underscore_objs =(@underscore_objs, @lapack_deprecated_objs);
+    }
+
 } else {
     @underscore_objs = (@blasobjs, @lapackobjs, @misc_underscore_objs);
 }

f_check

@@ -1,5 +1,7 @@
 #!/usr/bin/perl
 
+$hostos   = `uname -s | sed -e s/\-.*//`;    chop($hostos);
+
 #
 # 1. Not specified
 #   1.1 Automatically detect, then check compiler
@@ -272,8 +274,9 @@ if ($link ne "") {
 	}
 
 	if ($flags =~ /^\-Y/) {
+	    next if ($hostos eq 'SunOS');
 	    $linker_L .= "-Wl,". $flags . " ";
-	    }
+        }
 
 	if ($flags =~ /^\-rpath\@/) {
 	    $flags =~ s/\@/\,/g;

getarch.c

@@ -86,7 +86,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <sys/types.h>
 #include <sys/sysctl.h>
 #endif
-#ifdef linux
+#if defined(linux) || defined(__sun__)
 #include <sys/sysinfo.h>
 #include <unistd.h>
 #endif
@@ -906,7 +906,7 @@ static int get_num_cores(void) {
   size_t len;
 #endif
 
-#ifdef linux
+#if defined(linux) || defined(__sun__)
   //returns the number of processors which are currently online
   return sysconf(_SC_NPROCESSORS_ONLN);
 
@@ -998,7 +998,9 @@ int main(int argc, char *argv[]){
 #endif
 #endif
 
-#if NO_PARALLEL_MAKE==1
+#ifdef MAKE_NB_JOBS
+    printf("MAKE += -j %d\n", MAKE_NB_JOBS);
+#elif NO_PARALLEL_MAKE==1
     printf("MAKE += -j 1\n");
 #else
 #ifndef OS_WINDOWS

interface/gemv.c

@@ -79,11 +79,9 @@ void NAME(char *TRANS, blasint *M, blasint *N,
   FLOAT alpha = *ALPHA;
   FLOAT beta  = *BETA;
   FLOAT *buffer;
+  int buffer_size;
 #ifdef SMP
   int nthreads;
-  int nthreads_max;
-  int nthreads_avail;
-  double MNK;
 #endif
 
   int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG,  FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
@@ -134,13 +132,10 @@ void CNAME(enum CBLAS_ORDER order,
 
   FLOAT *buffer;
   blasint lenx, leny;
-  int trans;
+  int trans, buffer_size;
   blasint info, t;
 #ifdef SMP
   int nthreads;
-  int nthreads_max;
-  int nthreads_avail;
-  double MNK;
 #endif
 
   int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT *, BLASLONG,  FLOAT * , BLASLONG, FLOAT *, BLASLONG, FLOAT *) = {
@@ -215,43 +210,20 @@ void CNAME(enum CBLAS_ORDER order,
   if (incx < 0) x -= (lenx - 1) * incx;
   if (incy < 0) y -= (leny - 1) * incy;
 
-#ifdef MAX_STACK_ALLOC
-  // make it volatile because some gemv implementation (ex: dgemv_n.S)
-  // do not restore all register
-  volatile int stack_alloc_size = 0;
-  //for gemv_n and gemv_t, try to allocate on stack
-  stack_alloc_size = m + n;
-#ifdef ALIGNED_ACCESS
-  stack_alloc_size += 3;
-#endif
-  if(stack_alloc_size < 128)
-    //dgemv_n.S require a 128 bytes buffer
-    stack_alloc_size = 128;
-
-  if(stack_alloc_size > MAX_STACK_ALLOC / sizeof(FLOAT))
-    stack_alloc_size = 0;
-
-  FLOAT stack_buffer[stack_alloc_size];
-  buffer = stack_alloc_size ? stack_buffer : (FLOAT *)blas_memory_alloc(1);
-  //  printf("stack_alloc_size=%d\n", stack_alloc_size);
-#else
-  //Original OpenBLAS/GotoBLAS codes.
-  buffer = (FLOAT *)blas_memory_alloc(1);
+  buffer_size = m + n + 128 / sizeof(FLOAT);
+#ifdef WINDOWS_ABI
+  buffer_size += 160 / sizeof(FLOAT) ;
 #endif
+  // for alignment
+  buffer_size = (buffer_size + 3) & ~3;
+  STACK_ALLOC(buffer_size, FLOAT, buffer);
 
 #ifdef SMP
 
-  nthreads_max = num_cpu_avail(2);
-  nthreads_avail = nthreads_max;
-
-  MNK = (double) m * (double) n;
-  if ( MNK <= (24.0 * 24.0  * (double) (GEMM_MULTITHREAD_THRESHOLD*GEMM_MULTITHREAD_THRESHOLD) )  )
-        nthreads_max = 1;
-
-  if ( nthreads_max > nthreads_avail )
-        nthreads = nthreads_avail;
+  if ( 1L * m * n < 2304L * GEMM_MULTITHREAD_THRESHOLD )
+    nthreads = 1;
   else
-        nthreads = nthreads_max;
+    nthreads = num_cpu_avail(2);
 
   if (nthreads == 1) {
 #endif
@@ -266,14 +238,7 @@ void CNAME(enum CBLAS_ORDER order,
   }
 #endif
 
-#ifdef MAX_STACK_ALLOC
-  if(!stack_alloc_size){
-    blas_memory_free(buffer);
-  }
-#else
-    blas_memory_free(buffer);
-#endif
-  
+  STACK_FREE(buffer);
   FUNCTION_PROFILE_END(1, m * n + m + n,  2 * m * n);
 
   IDEBUG_END;

interface/ger.c

@@ -171,19 +171,14 @@ void CNAME(enum CBLAS_ORDER order,
   if (incy < 0) y -= (n - 1) * incy;
   if (incx < 0) x -= (m - 1) * incx;
 
-#ifdef MAX_STACK_ALLOC
-  volatile int stack_alloc_size = m;
-  if(stack_alloc_size > MAX_STACK_ALLOC / sizeof(FLOAT))
-      stack_alloc_size = 0;
-  FLOAT stack_buffer[stack_alloc_size];
-  buffer = stack_alloc_size ? stack_buffer : (FLOAT *)blas_memory_alloc(1);
-#else
-  buffer = (FLOAT *)blas_memory_alloc(1);
-#endif
+  STACK_ALLOC(m, FLOAT, buffer);
 
 #ifdef SMPTEST
-  nthreads = num_cpu_avail(2);
-
+  // Threshold chosen so that speed-up is > 1 on a Xeon E5-2630
+  if(1L * m * n > 2048L * GEMM_MULTITHREAD_THRESHOLD)
+    nthreads = num_cpu_avail(2);
+  else
+    nthreads = 1;
 
   if (nthreads == 1) {
 #endif
@@ -198,11 +193,7 @@ void CNAME(enum CBLAS_ORDER order,
   }
 #endif
 
-#ifdef MAX_STACK_ALLOC
-  if(!stack_alloc_size)
-#endif
-    blas_memory_free(buffer);
-
+  STACK_FREE(buffer);
   FUNCTION_PROFILE_END(1, m * n + m + n, 2 * m * n);
 
   IDEBUG_END;

interface/rotg.c

@@ -95,7 +95,7 @@ void CNAME(FLOAT *DA, FLOAT *DB, FLOAT *C, FLOAT *S){
     s = db / r;
     z = ONE;
     if (ada > adb) z = s;
-    if ((ada < adb) && (c != ZERO)) z = ONE / c;
+    if ((ada <= adb) && (c != ZERO)) z = ONE / c;
 
     *C = c;
     *S = s;

interface/swap.c

@@ -77,12 +77,13 @@ void CNAME(blasint n, FLOAT *x, blasint incx, FLOAT *y, blasint incy){
   if (incy < 0) y -= (n - 1) * incy;
 
 #ifdef SMP
-  nthreads = num_cpu_avail(1);
 
   //disable multi-thread when incx==0 or incy==0
   //In that case, the threads would be dependent.
-  if (incx == 0 || incy == 0)
-	  nthreads = 1;
+  if (incx == 0 || incy == 0 || n < 2097152 * GEMM_MULTITHREAD_THRESHOLD / sizeof(FLOAT))
+    nthreads = 1;
+  else
+    nthreads = num_cpu_avail(1);
 
   if (nthreads == 1) {
 #endif

interface/zgemv.c

@@ -77,11 +77,9 @@ void NAME(char *TRANS, blasint *M, blasint *N,
   blasint incy = *INCY;
 
   FLOAT *buffer;
+  int buffer_size;
 #ifdef SMP
   int nthreads;
-  int nthreads_max;
-  int nthreads_avail;
-  double MNK;
 #endif
 
   int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG,
@@ -144,13 +142,10 @@ void CNAME(enum CBLAS_ORDER order,
 
   FLOAT *buffer;
   blasint    lenx, leny;
-  int trans;
+  int trans, buffer_size;
   blasint info, t;
 #ifdef SMP
   int nthreads;
-  int nthreads_max;
-  int nthreads_avail;
-  double MNK;
 #endif
 
   int (*gemv[])(BLASLONG, BLASLONG, BLASLONG, FLOAT, FLOAT, FLOAT *, BLASLONG,
@@ -236,22 +231,26 @@ void CNAME(enum CBLAS_ORDER order,
   if (incx < 0) x -= (lenx - 1) * incx * 2;
   if (incy < 0) y -= (leny - 1) * incy * 2;
 
-  buffer = (FLOAT *)blas_memory_alloc(1);
+  buffer_size = 2 * (m + n) + 128 / sizeof(FLOAT);
+#ifdef WINDOWS_ABI
+  buffer_size += 160 / sizeof(FLOAT) ;
+#endif
+  // for alignment
+  buffer_size = (buffer_size + 3) & ~3;
+  STACK_ALLOC(buffer_size, FLOAT, buffer);
+
+#if defined(ARCH_X86_64) && defined(MAX_STACK_ALLOC) && MAX_STACK_ALLOC > 0
+  // cgemv_t.S return NaN if there are NaN or Inf in the buffer (see bug #746)
+  if(trans && stack_alloc_size)
+    memset(buffer, 0, MIN(BUFFER_SIZE, sizeof(FLOAT) * buffer_size));
+#endif
 
 #ifdef SMP
 
-  nthreads_max = num_cpu_avail(2);
-  nthreads_avail = nthreads_max;
-
-  MNK = (double) m * (double) n;
-  if ( MNK <= ( 256.0  * (double) (GEMM_MULTITHREAD_THRESHOLD * GEMM_MULTITHREAD_THRESHOLD)  ))
-        nthreads_max = 1;
-
-  if ( nthreads_max > nthreads_avail )
-        nthreads = nthreads_avail;
+  if ( 1L * m * n < 1024L * GEMM_MULTITHREAD_THRESHOLD )
+    nthreads = 1;
   else
-        nthreads = nthreads_max;
-
+    nthreads = num_cpu_avail(2);
 
   if (nthreads == 1) {
 #endif
@@ -267,7 +266,7 @@ void CNAME(enum CBLAS_ORDER order,
   }
 #endif
 
-  blas_memory_free(buffer);
+  STACK_FREE(buffer);
 
   FUNCTION_PROFILE_END(4, m * n + m + n,  2 * m * n);
 

interface/zger.c

@@ -210,10 +210,14 @@ void CNAME(enum CBLAS_ORDER order,
   if (incy < 0) y -= (n - 1) * incy * 2;
   if (incx < 0) x -= (m - 1) * incx * 2;
 
-  buffer = (FLOAT *)blas_memory_alloc(1);
+  STACK_ALLOC(2 * m, FLOAT, buffer);
 
 #ifdef SMPTEST
-  nthreads = num_cpu_avail(2);
+  // Threshold chosen so that speed-up is > 1 on a Xeon E5-2630
+  if(1L * m * n > 36L * sizeof(FLOAT) * sizeof(FLOAT) * GEMM_MULTITHREAD_THRESHOLD)
+    nthreads = num_cpu_avail(2);
+  else
+    nthreads = 1;
 
   if (nthreads == 1) {
 #endif
@@ -245,7 +249,7 @@ void CNAME(enum CBLAS_ORDER order,
   }
 #endif
 
-  blas_memory_free(buffer);
+  STACK_FREE(buffer);
 
   FUNCTION_PROFILE_END(4, m * n + m + n, 2 * m * n);
 

interface/ztrmv.c

@@ -107,7 +107,7 @@ void NAME(char *UPLO, char *TRANS, char *DIAG,
   blasint info;
   int uplo;
   int unit;
-  int trans;
+  int trans, buffer_size;
   FLOAT *buffer;
 #ifdef SMP
   int nthreads;
@@ -154,7 +154,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 	   enum CBLAS_TRANSPOSE TransA, enum CBLAS_DIAG Diag,
 	   blasint n, FLOAT  *a, blasint lda, FLOAT  *x, blasint incx) {
 
-  int trans, uplo, unit;
+  int trans, uplo, unit, buffer_size;
   blasint info;
   FLOAT *buffer;
 #ifdef SMP
@@ -227,11 +227,28 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
 
   if (incx < 0 ) x -= (n - 1) * incx * 2;
 
-  buffer = (FLOAT *)blas_memory_alloc(1);
+#ifdef SMP
+  // Calibrated on a Xeon E5-2630
+  if(1L * n * n > 36L * sizeof(FLOAT) * sizeof(FLOAT) * GEMM_MULTITHREAD_THRESHOLD) {
+    nthreads = num_cpu_avail(2);
+    if(nthreads > 2 && 1L * n * n < 64L * sizeof(FLOAT) * sizeof(FLOAT) * GEMM_MULTITHREAD_THRESHOLD)
+      nthreads = 2;
+  } else
+      nthreads = 1;
+
+  if(nthreads > 1) {
+    buffer_size = n > 16 ? 0 : n * 4 + 40;
+  }
+  else
+#endif
+  {
+    buffer_size = ((n - 1) / DTB_ENTRIES) * 2 * DTB_ENTRIES + 32 / sizeof(FLOAT);
+    if(incx != 1)
+      buffer_size += n * 2;
+  }
+  STACK_ALLOC(buffer_size, FLOAT, buffer);
 
 #ifdef SMP
-  nthreads = num_cpu_avail(2);
-
   if (nthreads == 1) {
 #endif
 
@@ -245,7 +262,7 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_UPLO Uplo,
   }
 #endif
 
-  blas_memory_free(buffer);
+  STACK_FREE(buffer);
 
   FUNCTION_PROFILE_END(4, n * n / 2 + n,  n * n);
 

kernel/CMakeLists.txt

@@ -227,6 +227,28 @@ foreach (float_type ${FLOAT_TYPES})
   GenerateNamedObjects("generic/ztrsm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trsm_oltucopy" false "" "" false ${float_type})
   GenerateNamedObjects("generic/ztrsm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trsm_oltncopy" false "" "" false ${float_type})
 
+  #gemm3m
+  if (USE_GEMM3M)
+    GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMM3MKERNEL}" "NN" "gemm3m_kernel" false "" "" false ${float_type})
+
+    GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA" "gemm3m_oncopyb" false "" "" false ${float_type})
+    GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA;REAL_ONLY" "gemm3m_oncopyr" false "" "" false ${float_type})
+    GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA;IMAGE_ONLY" "gemm3m_oncopyi" false "" "" false ${float_type})
+
+    GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA" "gemm3m_otcopyb" false "" "" false ${float_type})
+    GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA;REAL_ONLY" "gemm3m_otcopyr" false "" "" false ${float_type})
+    GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_N}.c" "USE_ALPHA;IMAGE_ONLY" "gemm3m_otcopyi" false "" "" false ${float_type})
+
+    GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY" "gemm3m_incopyb" false "" "" false ${float_type})
+    GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY;REAL_ONLY" "gemm3m_incopyr" false "" "" false ${float_type})
+    GenerateNamedObjects("generic/zgemm3m_ncopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY;IMAGE_ONLY" "gemm3m_incopyi" false "" "" false ${float_type})
+
+    GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY" "gemm3m_itcopyb" false "" "" false ${float_type})
+    GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY;REAL_ONLY" "gemm3m_itcopyr" false "" "" false ${float_type})
+    GenerateNamedObjects("generic/zgemm3m_tcopy_${${float_char}GEMM3M_UNROLL_M}.c" "ICOPY;IMAGE_ONLY" "gemm3m_itcopyi" false "" "" false ${float_type})
+
+  endif()
+
   else ()  #For real
     GenerateCombinationObjects("${KERNELDIR}/${TRMM_KERNEL}" "LEFT;TRANSA" "R;N" "TRMMKERNEL" 2 "trmm_kernel" false ${float_type})
 

kernel/arm/KERNEL.ARMV6

@@ -1,26 +1,4 @@
-SGEMVNKERNEL = ../arm/gemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-CGEMVNKERNEL = ../arm/zgemv_n.c
-CGEMVTKERNEL = ../arm/zgemv_t.c
 
-DGEMVNKERNEL = ../arm/gemv_n.c
-DGEMVTKERNEL = ../arm/gemv_t.c
-
-CTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
-CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
-CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
-CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-
-#ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
-#ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
-#ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
-#ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-
-
-#STRMMKERNEL	= ../generic/trmmkernel_2x2.c
-#SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
-#SGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
-#SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
 
 
 ###############################################################################
@@ -96,19 +74,19 @@ DSWAPKERNEL  = swap_vfp.S
 CSWAPKERNEL  = swap_vfp.S
 ZSWAPKERNEL  = swap_vfp.S
 
-# BAD SGEMVNKERNEL = gemv_n_vfp.S
-# BAD DGEMVNKERNEL = gemv_n_vfp.S
-# CGEMVNKERNEL = cgemv_n_vfp.S
+SGEMVNKERNEL = gemv_n_vfp.S
+DGEMVNKERNEL = gemv_n_vfp.S
+CGEMVNKERNEL = cgemv_n_vfp.S
 ZGEMVNKERNEL = zgemv_n_vfp.S
 
-# BAD SGEMVTKERNEL = gemv_t_vfp.S
-# BAD DGEMVTKERNEL = gemv_t_vfp.S
-# CGEMVTKERNEL = cgemv_t_vfp.S
+SGEMVTKERNEL = gemv_t_vfp.S
+DGEMVTKERNEL = gemv_t_vfp.S
+CGEMVTKERNEL = cgemv_t_vfp.S
 ZGEMVTKERNEL = zgemv_t_vfp.S
 
 STRMMKERNEL	= strmm_kernel_4x2_vfp.S
 DTRMMKERNEL	= dtrmm_kernel_4x2_vfp.S
-#CTRMMKERNEL	= ctrmm_kernel_2x2_vfp.S
+CTRMMKERNEL	= ctrmm_kernel_2x2_vfp.S
 ZTRMMKERNEL	= ztrmm_kernel_2x2_vfp.S
 
 SGEMMKERNEL    = sgemm_kernel_4x2_vfp.S
@@ -131,9 +109,9 @@ DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
 DGEMMONCOPYOBJ = dgemm_oncopy.o
 DGEMMOTCOPYOBJ = dgemm_otcopy.o
 
-#CGEMMKERNEL    = cgemm_kernel_2x2_vfp.S
-#CGEMMONCOPY    = cgemm_ncopy_2_vfp.S
-#CGEMMOTCOPY    = cgemm_tcopy_2_vfp.S
+CGEMMKERNEL    = cgemm_kernel_2x2_vfp.S
+CGEMMONCOPY    = cgemm_ncopy_2_vfp.S
+CGEMMOTCOPY    = cgemm_tcopy_2_vfp.S
 CGEMMONCOPYOBJ = cgemm_oncopy.o
 CGEMMOTCOPYOBJ = cgemm_otcopy.o
 

kernel/arm/KERNEL.ARMV7

@@ -1,8 +1,3 @@
-SGEMVNKERNEL = ../arm/gemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-CGEMVNKERNEL = ../arm/zgemv_n.c
-CGEMVTKERNEL = ../arm/zgemv_t.c
-
 
 #################################################################################
 SAMAXKERNEL  = iamax_vfp.S
@@ -77,14 +72,14 @@ DSCALKERNEL  = scal.c
 CSCALKERNEL  = zscal.c
 ZSCALKERNEL  = zscal.c
 
-# BAD SGEMVNKERNEL = gemv_n_vfp.S
-DGEMVNKERNEL = gemv_n_vfp.S
-#CGEMVNKERNEL = cgemv_n_vfp.S
+SGEMVNKERNEL = gemv_n_vfpv3.S
+DGEMVNKERNEL = gemv_n_vfpv3.S
+CGEMVNKERNEL = cgemv_n_vfp.S
 ZGEMVNKERNEL = zgemv_n_vfp.S
 
-# BAD SGEMVTKERNEL = gemv_t_vfp.S
+SGEMVTKERNEL = gemv_t_vfp.S
 DGEMVTKERNEL = gemv_t_vfp.S
-#CGEMVTKERNEL = cgemv_t_vfp.S
+CGEMVTKERNEL = cgemv_t_vfp.S
 ZGEMVTKERNEL = zgemv_t_vfp.S
 
 STRMMKERNEL  =  strmm_kernel_4x4_vfpv3.S
@@ -92,24 +87,15 @@ DTRMMKERNEL  =  dtrmm_kernel_4x4_vfpv3.S
 CTRMMKERNEL  =  ctrmm_kernel_2x2_vfpv3.S
 ZTRMMKERNEL  =  ztrmm_kernel_2x2_vfpv3.S
 
-#SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
 SGEMMKERNEL    =  sgemm_kernel_4x4_vfpv3.S
-SGEMMINCOPY    =
-SGEMMITCOPY    =
 SGEMMONCOPY    =  sgemm_ncopy_4_vfp.S
 SGEMMOTCOPY    =  sgemm_tcopy_4_vfp.S
-SGEMMINCOPYOBJ =
-SGEMMITCOPYOBJ =
 SGEMMONCOPYOBJ =  sgemm_oncopy.o
 SGEMMOTCOPYOBJ =  sgemm_otcopy.o
 
 DGEMMKERNEL    =  dgemm_kernel_4x4_vfpv3.S
-DGEMMINCOPY    =
-DGEMMITCOPY    =
 DGEMMONCOPY    =  dgemm_ncopy_4_vfp.S
 DGEMMOTCOPY    =  dgemm_tcopy_4_vfp.S
-DGEMMINCOPYOBJ =
-DGEMMITCOPYOBJ =
 DGEMMONCOPYOBJ = dgemm_oncopy.o
 DGEMMOTCOPYOBJ = dgemm_otcopy.o
 

kernel/arm/asum_vfp.S

@@ -367,12 +367,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	.align 5
 
-#if defined(DOUBLE)
-	vsub.f64                d0 , d0 , d0
-	vsub.f64                d1 , d1 , d1
-#else
-	vsub.f32                s0 , s0 , s0
-	vsub.f32                s1 , s1 , s1
+        movs    r12, #0                                          // clear floating point register
+        vmov    s0, r12
+        vmov    s1, r12
+#if     defined(DOUBLE)
+        vcvt.f64.f32    d0, s0
+        vcvt.f64.f32    d1, s1
 #endif
 
 	cmp	N, #0

kernel/arm/cdot_vfp.S

@@ -185,14 +185,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r4, fp, #128
 	vstm	r4, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0                                          // clear floating point register
+        vmov    s0, r4
+	vmov	s1, s0
+	vmov	s2, s0
+	vmov	s3, s0
+
 	mov	Y, OLD_Y
 	ldr	INC_Y, OLD_INC_Y
 
-	vsub.f32                s0 , s0 , s0
-	vsub.f32                s1 , s1 , s1
-	vsub.f32                s2 , s2 , s2
-	vsub.f32                s3 , s3 , s3
-
 	cmp	N, #0
 	ble	cdot_kernel_L999
 

kernel/arm/cgemm_kernel_2x2_vfp.S

@@ -57,6 +57,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -138,7 +142,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -340,7 +344,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s12, s8
 	vmov.f32		s13, s8
@@ -514,7 +518,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -681,7 +685,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 
 .endm
@@ -822,6 +826,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 4 * 2
 	str	r3, LDC

kernel/arm/cgemm_kernel_2x2_vfpv3.S

@@ -73,6 +73,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -147,7 +151,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -368,7 +372,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -550,7 +554,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -730,7 +734,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s24, s16
 	vmov.f32		s25, s16
@@ -879,6 +883,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 4 * 2
 	str	r3, LDC

kernel/arm/cgemv_n_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define ALPHA_I [fp, #-236]
 #define ALPHA_R [fp, #-244]
 
@@ -117,7 +121,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .macro INIT_F4
 
 	pld	[ YO, #Y_PRE ]
-        vsub.f32                s8 , s8 , s8
+	flds			s8 , FP_ZERO
         vmov.f32                s9 , s8
         vmov.f32                s10, s8
         vmov.f32                s11, s8
@@ -220,7 +224,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-        vsub.f32                s8 , s8 , s8
+	flds			s8 , FP_ZERO
         vmov.f32                s9 , s8
 
 .endm
@@ -267,7 +271,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S4
 
-        vsub.f32                s8 , s8 , s8
+	flds			s8 , FP_ZERO
         vmov.f32                s9 , s8
         vmov.f32                s10, s8
         vmov.f32                s11, s8
@@ -384,7 +388,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-        vsub.f32                s8 , s8 , s8
+	flds			s8 , FP_ZERO
         vmov.f32                s9 , s8
 
 .endm
@@ -448,6 +452,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	OLD_M, #0
 	ble	cgemvn_kernel_L999
 

kernel/arm/cgemv_t_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define N	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
@@ -116,10 +120,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F2
 
-	vsub.f32	s12, s12, s12
-	vsub.f32	s13, s13, s13
-	vsub.f32	s14, s14, s14
-	vsub.f32	s15, s15, s15
+	flds		s12, FP_ZERO
+	vmov.f32	s13, s12
+	vmov.f32	s14, s12
+	vmov.f32	s15, s12
 
 .endm
 
@@ -172,8 +176,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f32	s12, s12, s12
-	vsub.f32	s13, s13, s13
+	flds		s12, FP_ZERO
+	vmov.f32	s13, s12
 
 .endm
 
@@ -215,10 +219,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S2
 
-	vsub.f32	s12, s12, s12
-	vsub.f32	s13, s13, s13
-	vsub.f32	s14, s14, s14
-	vsub.f32	s15, s15, s15
+	flds		s12, FP_ZERO
+	vmov.f32	s13, s12
+	vmov.f32	s14, s12
+	vmov.f32	s15, s12
 
 .endm
 
@@ -281,8 +285,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f32	s12, s12, s12
-	vsub.f32	s13, s13, s13
+	flds		s12, FP_ZERO
+	vmov.f32	s13, s12
 
 .endm
 
@@ -345,6 +349,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	M, #0
 	ble	cgemvt_kernel_L999
 

kernel/arm/ctrmm_kernel_2x2_vfp.S

@@ -59,6 +59,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-232]
+#define FP_ZERO_0 [fp, #-232]
+#define FP_ZERO_1 [fp, #-228]
+
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -136,7 +141,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -301,10 +306,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	flds		s0, ALPHA_R
 	flds		s1, ALPHA_I
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
-	vsub.f32	s6, s6, s6
-	vsub.f32	s7, s7, s7
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
+	vmov.f32	s6, s4
+	vmov.f32	s7, s4
 
 	FMAC_R1 s4 , s0 , s8
 	FMAC_I1 s5 , s0 , s9
@@ -318,10 +323,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	fstmias CO1, { s4 - s7 }
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
-	vsub.f32	s6, s6, s6
-	vsub.f32	s7, s7, s7
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
+	vmov.f32	s6, s4
+	vmov.f32	s7, s4
 
 	FMAC_R1 s4 , s0 , s12
 	FMAC_I1 s5 , s0 , s13
@@ -343,7 +348,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s12, s8
 	vmov.f32		s13, s8
@@ -490,8 +495,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	flds		s0, ALPHA_R
 	flds		s1, ALPHA_I
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
 
 	FMAC_R1 s4 , s0 , s8
 	FMAC_I1 s5 , s0 , s9
@@ -500,8 +505,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	fstmias CO1, { s4 - s5 }
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
 
 	FMAC_R1 s4 , s0 , s12
 	FMAC_I1 s5 , s0 , s13
@@ -519,7 +524,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -663,10 +668,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	flds		s0, ALPHA_R
 	flds		s1, ALPHA_I
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
-	vsub.f32	s6, s6, s6
-	vsub.f32	s7, s7, s7
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
+	vmov.f32	s6, s4
+	vmov.f32	s7, s4
 
 	FMAC_R1 s4 , s0 , s8
 	FMAC_I1 s5 , s0 , s9
@@ -689,7 +694,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 
 .endm
@@ -795,8 +800,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	flds		s0, ALPHA_R
 	flds		s1, ALPHA_I
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
 
 	FMAC_R1 s4 , s0 , s8
 	FMAC_I1 s5 , s0 , s9
@@ -831,6 +836,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 4 * 2
 	str	r3, LDC

kernel/arm/ctrmm_kernel_2x2_vfpv3.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -134,7 +138,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -351,7 +355,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -529,7 +533,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -706,7 +710,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s24, s16
 	vmov.f32		s25, s16
@@ -852,6 +856,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 4 * 2
 	str	r3, LDC

kernel/arm/ddot_vfp.S

@@ -31,6 +31,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 * 	 CTEST			: OK
 * 	 TEST			: OK
 *
+* 2016/01/23 Saar
+*        Bugfix for Refs #750 and #740
 **************************************************************************************/
 
 #define ASSEMBLER
@@ -152,8 +154,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	mov	Y, OLD_Y
 	ldr	INC_Y, OLD_INC_Y
 
-	vsub.f64                d0 , d0 , d0
-	vsub.f64                d1 , d1 , d1
+        movs    r4, #0                                          // clear floating point register
+        vmov    s0, r4
+        vmov    s1, r4
+        vcvt.f64.f32    d0, s0
+        vcvt.f64.f32    d1, s1
+
 
 	cmp	N, #0
 	ble	ddot_kernel_L999

kernel/arm/dgemm_kernel_4x2_vfp.S

@@ -56,8 +56,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
+
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
@@ -85,7 +90,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -173,7 +178,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -233,7 +238,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d12, d8
 
 .endm
@@ -283,7 +288,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -338,7 +343,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -380,7 +385,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 
 .endm
 
@@ -433,6 +438,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 8
 	str	r3, LDC

kernel/arm/dgemm_kernel_4x4_vfpv3.S

@@ -73,6 +73,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
 #define B	[fp, #4 ]
@@ -102,7 +106,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -376,7 +380,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -470,7 +474,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d20, d16
 	vmov.f64		d24, d16
 	vmov.f64		d28, d16
@@ -533,7 +537,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -617,7 +621,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -678,7 +682,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d20, d16
 
 .endm
@@ -723,7 +727,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -782,7 +786,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 
 .endm
@@ -826,7 +830,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 
 .endm
 
@@ -880,6 +884,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	str	OLD_A, A
 	vstr	OLD_ALPHA, ALPHA
 
+	movs	r4, #0
+	str	r4, FP_ZERO
+	str	r4, FP_ZERO_1
+
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 

kernel/arm/dtrmm_kernel_4x2_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-232]
+#define FP_ZERO_0 [fp, #-232]
+#define FP_ZERO_1 [fp, #-228]
+
 #define ALPHA	[fp, #-276 ]
 
 #define B	[fp, #4 ]
@@ -90,7 +94,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -165,7 +169,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -220,7 +224,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d12, d8
 
 .endm
@@ -268,7 +272,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -318,7 +322,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -357,7 +361,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 
 .endm
 
@@ -409,6 +413,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 8
 	str	r3, LDC

kernel/arm/dtrmm_kernel_4x4_vfpv3.S

@@ -59,6 +59,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
+
 #define ALPHA	[fp, #-276 ]
 
 #define B	[fp, #4 ]
@@ -89,7 +94,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -386,7 +391,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -468,7 +473,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d20, d16
 	vmov.f64		d24, d16
 	vmov.f64		d28, d16
@@ -527,7 +532,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -601,7 +606,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -656,7 +661,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d20, d16
 
 .endm
@@ -699,7 +704,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -753,7 +758,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 
 .endm
@@ -794,7 +799,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 
 .endm
 
@@ -850,6 +855,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 8
 	str	r3, LDC

kernel/arm/gemv_n_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define M	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
@@ -79,7 +83,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	pld     [ YO , #Y_PRE ]
 	pld     [ YO , #Y_PRE+32 ]
 
-	vsub.f64	d8 , d8 , d8
+	fldd		d8 , FP_ZERO
 	vmov.f64	d9  , d8
 	vmov.f64	d10 , d8
 	vmov.f64	d11 , d8
@@ -158,7 +162,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f64	d12 , d12 , d12
+	fldd		d12 , FP_ZERO
 
 .endm
 
@@ -185,7 +189,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S4
 
-	vsub.f64	d12 , d12 , d12
+	fldd		d12 , FP_ZERO
 	vmov.f64	d13 , d12
 	vmov.f64	d14 , d12
 	vmov.f64	d15 , d12
@@ -245,7 +249,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f64	d12 , d12 , d12
+	fldd		d12 , FP_ZERO
 
 .endm
 
@@ -279,7 +283,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	pld     [ YO , #Y_PRE ]
 
-	vsub.f32	s8 , s8 , s8
+	flds		s8 , FP_ZERO
 	vmov.f32	s9  , s8
 	vmov.f32	s10 , s8
 	vmov.f32	s11 , s8
@@ -357,7 +361,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f32	s12 , s12 , s12
+	flds		s12 , FP_ZERO
 
 .endm
 
@@ -384,7 +388,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S4
 
-	vsub.f32	s12 , s12 , s12
+	flds		s12 , FP_ZERO
 	vmov.f32	s13 , s12
 	vmov.f32	s14 , s12
 	vmov.f32	s15 , s12
@@ -445,7 +449,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f32	s12 , s12 , s12
+	flds		s12 , FP_ZERO
 
 .endm
 
@@ -494,6 +498,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	OLD_M, #0
 	ble	gemvn_kernel_L999
 

kernel/arm/gemv_n_vfpv3.S

@@ -62,6 +62,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define M	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 
 #define X_PRE	64
 #define Y_PRE	0
@@ -79,7 +83,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	pld     [ YO , #Y_PRE ]
 	pld     [ YO , #Y_PRE+32 ]
 
-	vsub.f64	d24 , d24 , d24
+	fldd		d24 , FP_ZERO
 	vmov.f64	d25 , d24
 	vmov.f64	d26 , d24
 	vmov.f64	d27 , d24
@@ -147,7 +151,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f64	d24 , d24 , d24
+	fldd		d24 , FP_ZERO
 
 .endm
 
@@ -175,7 +179,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S8
 
-	vsub.f64	d24 , d24 , d24
+	fldd		d24 , FP_ZERO
 	vmov.f64	d25 , d24
 	vmov.f64	d26 , d24
 	vmov.f64	d27 , d24
@@ -269,7 +273,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f64	d24 , d24 , d24
+	fldd		d24 , FP_ZERO
 
 .endm
 
@@ -302,7 +306,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	pld     [ YO , #Y_PRE ]
 
-	vsub.f32	s24 , s24 , s24
+	flds		s24 , FP_ZERO
 	vmov.f32	s25 , s24
 	vmov.f32	s26 , s24
 	vmov.f32	s27 , s24
@@ -368,7 +372,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f32	s24 , s24 , s24
+	flds		s24 , FP_ZERO
 
 .endm
 
@@ -396,7 +400,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S8
 
-	vsub.f32	s24 , s24 , s24
+	flds		s24 , FP_ZERO
 	vmov.f32	s25 , s24
 	vmov.f32	s26 , s24
 	vmov.f32	s27 , s24
@@ -489,7 +493,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f32	s24 , s24 , s24
+	flds		s24 , FP_ZERO
 
 .endm
 
@@ -538,6 +542,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s31 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	OLD_M, #0
 	ble	gemvn_kernel_L999
 

kernel/arm/gemv_t_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define N	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
@@ -75,8 +79,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F2
 
-	vsub.f64	d2 , d2 , d2
-	vsub.f64	d3 , d3 , d3
+	fldd		d2, FP_ZERO
+	vmov.f64	d3 , d2 
 
 .endm
 
@@ -123,7 +127,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f64	d2 , d2 , d2
+	fldd		d2, FP_ZERO
+	vmov.f64	d3 , d2 
 
 .endm
 
@@ -160,8 +165,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S2
 
-	vsub.f64	d2 , d2 , d2
-	vsub.f64	d3 , d3 , d3
+	fldd		d2, FP_ZERO
+	vmov.f64	d3 , d2 
 
 .endm
 
@@ -224,7 +229,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f64	d2 , d2 , d2
+	fldd		d2, FP_ZERO
+	vmov.f64	d3 , d2 
 
 .endm
 
@@ -276,8 +282,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F2
 
-	vsub.f32	s2 , s2 , s2
-	vsub.f32	s3 , s3 , s3
+	flds		s2 , FP_ZERO
+	vmov.f32	s3 , s2 
+
 
 .endm
 
@@ -321,7 +328,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f32	s2 , s2 , s2
+	flds		s2 , FP_ZERO
 
 .endm
 
@@ -356,8 +363,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S2
 
-	vsub.f32	s2 , s2 , s2
-	vsub.f32	s3 , s3 , s3
+	flds		s2 , FP_ZERO
+	vmov.f32	s3 , s2 
 
 .endm
 
@@ -418,7 +425,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f32	s2 , s2 , s2
+	flds		s2 , FP_ZERO
 
 .endm
 
@@ -488,6 +495,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	M, #0
 	ble	gemvt_kernel_L999
 

kernel/arm/iamax_vfp.S

@@ -341,11 +341,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	.align 5
 	push    {r4}
 
-#if defined(DOUBLE)
-	vsub.f64                d0 , d0 , d0
-#else
-	vsub.f32                s0 , s0 , s0
+        movs    r12, #0                                          // clear floating point register
+        vmov    s0, r12
+#if     defined(DOUBLE)
+        vcvt.f64.f32    d0, s0
 #endif
+
 	mov	INDEX, #0
 
 	cmp	N, #0

kernel/arm/nrm2_vfp.S

@@ -409,12 +409,20 @@ KERNEL_S1_END_\@:
 
 #if	defined(DOUBLE)
 
+znrm2_zero:
+	.word	0x00000000
+	.word	0x00000000
+
+
 znrm2_one:
 	.word	0x00000000
 	.word	0x3ff00000
 
 #else
 
+cnrm2_zero:
+	.word	0x00000000
+
 cnrm2_one:
 	.word	0x3f800000
 
@@ -424,12 +432,20 @@ cnrm2_one:
 
 #if	defined(DOUBLE)
 
+dnrm2_zero:
+	.word	0x00000000
+	.word	0x00000000
+
+
 dnrm2_one:
 	.word	0x00000000
 	.word	0x3ff00000
 
 #else
 
+snrm2_zero:
+	.word	0x00000000
+
 snrm2_one:
 	.word	0x3f800000
 
@@ -446,12 +462,12 @@ nrm2_begin:
 #if defined(COMPLEX)
 
 #if defined(DOUBLE)
-	vsub.f64                d0 , d0 , d0		// scale=0.0
+	vldr.64			d0 , znrm2_zero	
 	vldr.64			d1 , znrm2_one		// ssq=1.0
 	vmov.f64		d7 , d1			// value 1.0
 	vmov.f64		d6 , d0			// value 0.0
 #else
-	vsub.f32                s0 , s0 , s0		// scale=0.0
+	vldr.32			s0 , cnrm2_zero
 	vldr.32			s1 , cnrm2_one		// ssq=1.0
 	vmov.f32		s7 , s1			// value 1.0
 	vmov.f32		s6 , s0			// value 0.0
@@ -460,12 +476,12 @@ nrm2_begin:
 #else
 
 #if defined(DOUBLE)
-	vsub.f64                d0 , d0 , d0		// scale=0.0
+	vldr.64			d0 , dnrm2_zero	
 	vldr.64			d1 , dnrm2_one		// ssq=1.0
 	vmov.f64		d7 , d1			// value 1.0
 	vmov.f64		d6 , d0			// value 0.0
 #else
-	vsub.f32                s0 , s0 , s0		// scale=0.0
+	vldr.32			s0 , snrm2_zero
 	vldr.32			s1 , snrm2_one		// ssq=1.0
 	vmov.f32		s7 , s1			// value 1.0
 	vmov.f32		s6 , s0			// value 0.0

kernel/arm/nrm2_vfpv3.S

@@ -405,12 +405,15 @@ KERNEL_S1_END_\@:
 	.align 5
 
 #if defined(DOUBLE)
-	vsub.f64                d0 , d0 , d0		// scale=0.0
+	movs			r12 , #0
+	vmov.f32		s0 , r12		// scale=0.0
+	vcvt.f64.f32    d0, s0
 	vmov.f64		d1 , #1.0		// ssq=1.0
 	vmov.f64		d7 , d1			// value 1.0
 	vmov.f64		d6 , d0			// value 0.0
 #else
-	vsub.f32                s0 , s0 , s0		// scale=0.0
+	movs			r12 , #0
+	vmov.f32		s0 , r12		// scale=0.0
 	vmov.f32		s1 , #1.0		// ssq=1.0
 	vmov.f32		s7 , s1			// value 1.0
 	vmov.f32		s6 , s0			// value 0.0

kernel/arm/sdot_vfp.S

@@ -31,6 +31,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 * 	 CTEST			: OK (no test for dsdot)
 * 	 TEST			: OK (no test for dsdot)
 *
+* 2016/01/23 Saar
+*	 Bugfix	for Refs #750 and #740
 **************************************************************************************/
 
 #define ASSEMBLER
@@ -240,16 +242,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	mov	Y, OLD_Y
 	ldr	INC_Y, OLD_INC_Y
 
+	movs	r4, #0						// clear floating point register
+	vmov	s0, r4
+	vmov	s1, r4
 #if	defined(DSDOT)
-
-	vsub.f64                d0 , d0 , d0
-	vsub.f64                d1 , d1 , d1
-
-#else
-
-	vsub.f32                s0 , s0 , s0
-	vsub.f32                s1 , s1 , s1
-
+	vcvt.f64.f32    d0, s0
+	vcvt.f64.f32    d1, s1
 #endif
 
 	cmp	N, #0

kernel/arm/sgemm_kernel_4x2_vfp.S

@@ -56,6 +56,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
 #define B	[fp, #4 ]
@@ -85,7 +89,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -161,7 +165,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s12, s8
 	vmov.f32		s13, s8
@@ -221,7 +225,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s12, s8
 
 .endm
@@ -271,7 +275,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -326,7 +330,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9 , s8
 
 .endm
@@ -368,7 +372,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 
 .endm
 
@@ -421,6 +425,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/sgemm_kernel_4x4_vfpv3.S

@@ -73,7 +73,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
-#define ALPHA	[fp, #-280]
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, #-240]
+#define FP_ZERO_1 [fp, #-236]
+
+#define ALPHA   [fp, #-280]
 
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
@@ -102,7 +106,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -349,7 +353,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -443,7 +447,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s20, s16
 	vmov.f32		s24, s16
 	vmov.f32		s28, s16
@@ -506,7 +510,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -590,7 +594,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -651,7 +655,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s20, s16
 
 .endm
@@ -696,7 +700,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -755,7 +759,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 
 .endm
@@ -799,7 +803,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 
 .endm
 
@@ -856,6 +860,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/strmm_kernel_4x2_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-232]
+#define FP_ZERO_0 [fp, #-232]
+#define FP_ZERO_1 [fp, #-228]
+
 #define ALPHA	[fp, #-276 ]
 
 #define B	[fp, #4 ]
@@ -90,7 +94,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -156,7 +160,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s12, s8
 	vmov.f32		s13, s8
@@ -211,7 +215,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s12, s8
 
 .endm
@@ -259,7 +263,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -309,7 +313,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 
 .endm
@@ -348,7 +352,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 
 .endm
 
@@ -400,6 +404,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/strmm_kernel_4x4_vfpv3.S

@@ -58,6 +58,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
 #define B	[fp, #4 ]
@@ -88,7 +92,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -322,7 +326,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -405,7 +409,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s20, s16
 	vmov.f32		s24, s16
 	vmov.f32		s28, s16
@@ -464,7 +468,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -538,7 +542,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -593,7 +597,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s20, s16
 
 .endm
@@ -636,7 +640,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -690,7 +694,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 
 .endm
@@ -731,7 +735,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 
 .endm
 
@@ -787,6 +791,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/zdot_vfp.S

@@ -187,13 +187,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r4, fp, #128
 	vstm	r4, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0                                          // clear floating point register
+        vmov    s0, r4
+        vcvt.f64.f32    d0, s0
+        vcvt.f64.f32    d1, s0
+        vcvt.f64.f32    d2, s0
+        vcvt.f64.f32    d3, s0
+
 	mov	Y, OLD_Y
 	ldr	INC_Y, OLD_INC_Y
 
-	vsub.f64                d0 , d0 , d0
-	vsub.f64                d1 , d1 , d1
-	vsub.f64                d2 , d2 , d2
-	vsub.f64                d3 , d3 , d3
 
 	cmp	N, #0
 	ble	zdot_kernel_L999

kernel/arm/zgemm_kernel_2x2_vfp.S

@@ -57,6 +57,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -131,7 +135,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -383,7 +387,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -557,7 +561,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -724,7 +728,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -869,6 +873,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC

kernel/arm/zgemm_kernel_2x2_vfpv3.S

@@ -73,6 +73,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -147,7 +151,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -404,7 +408,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -586,7 +590,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -766,7 +770,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d24, d16
 	vmov.f64		d25, d16
@@ -915,6 +919,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC

kernel/arm/zgemv_n_vfp.S

@@ -59,6 +59,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
+
 #define ALPHA_I [fp, #-236]
 #define ALPHA_R [fp, #-244]
 
@@ -117,7 +122,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .macro INIT_F4
 
 	pld	[ YO, #Y_PRE ]
-        vsub.f64                d8 , d8 , d8
+	fldd			d8, FP_ZERO
         vmov.f64                d9 , d8
         vmov.f64                d10, d8
         vmov.f64                d11, d8
@@ -222,7 +227,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-        vsub.f64                d8 , d8 , d8
+	fldd			d8, FP_ZERO
         vmov.f64                d9 , d8
 
 .endm
@@ -269,7 +274,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S4
 
-        vsub.f64                d8 , d8 , d8
+	fldd			d8, FP_ZERO
         vmov.f64                d9 , d8
         vmov.f64                d10, d8
         vmov.f64                d11, d8
@@ -386,7 +391,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-        vsub.f64                d8 , d8 , d8
+	fldd			d8, FP_ZERO
         vmov.f64                d9 , d8
 
 .endm
@@ -450,6 +455,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	OLD_M, #0
 	ble	zgemvn_kernel_L999
 

kernel/arm/zgemv_t_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define N	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
@@ -117,10 +121,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F2
 
-	vsub.f64	d12, d12, d12
-	vsub.f64	d13, d13, d13
-	vsub.f64	d14, d14, d14
-	vsub.f64	d15, d15, d15
+	fldd		d12, FP_ZERO
+	vmov.f64	d13, d12
+	vmov.f64	d14, d12
+	vmov.f64	d15, d12
 
 .endm
 
@@ -173,8 +177,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f64	d12, d12, d12
-	vsub.f64	d13, d13, d13
+	fldd		d12, FP_ZERO
+	vmov.f64	d13, d12
 
 .endm
 
@@ -216,10 +220,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S2
 
-	vsub.f64	d12, d12, d12
-	vsub.f64	d13, d13, d13
-	vsub.f64	d14, d14, d14
-	vsub.f64	d15, d15, d15
+	fldd		d12, FP_ZERO
+	vmov.f64	d13, d12
+	vmov.f64	d14, d12
+	vmov.f64	d15, d12
 
 .endm
 
@@ -282,8 +286,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f64	d12, d12, d12
-	vsub.f64	d13, d13, d13
+	fldd		d12, FP_ZERO
+	vmov.f64	d13, d12
 
 .endm
 
@@ -346,6 +350,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	M, #0
 	ble	zgemvt_kernel_L999
 

kernel/arm/ztrmm_kernel_2x2_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-232]
+#define FP_ZERO_0 [fp, #-232]
+#define FP_ZERO_1 [fp, #-228]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -140,7 +144,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -356,10 +360,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fldd		d0, ALPHA_R
 	fldd		d1, ALPHA_I
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
-	vsub.f64	d6, d6 , d6
-	vsub.f64	d7, d7 , d7
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
+	vmov.f64	d6 , d4
+	vmov.f64	d7 , d4
 
 	FMAC_R1 d4 , d0 , d8
 	FMAC_I1 d5 , d0 , d9
@@ -373,10 +377,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	fstmiad CO1, { d4 - d7 }
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
-	vsub.f64	d6, d6 , d6
-	vsub.f64	d7, d7 , d7
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
+	vmov.f64	d6 , d4
+	vmov.f64	d7 , d4
 
 	FMAC_R1 d4 , d0 , d12
 	FMAC_I1 d5 , d0 , d13
@@ -398,7 +402,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -545,8 +549,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fldd		d0, ALPHA_R
 	fldd		d1, ALPHA_I
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
 
 	FMAC_R1 d4 , d0 , d8
 	FMAC_I1 d5 , d0 , d9
@@ -555,8 +559,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	fstmiad CO1, { d4 - d5 }
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
 
 	FMAC_R1 d4 , d0 , d12
 	FMAC_I1 d5 , d0 , d13
@@ -574,7 +578,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -718,10 +722,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fldd		d0, ALPHA_R
 	fldd		d1, ALPHA_I
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
-	vsub.f64	d6, d6 , d6
-	vsub.f64	d7, d7 , d7
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
+	vmov.f64	d6 , d4
+	vmov.f64	d7 , d4
 
 	FMAC_R1 d4 , d0 , d8
 	FMAC_I1 d5 , d0 , d9
@@ -744,7 +748,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -850,8 +854,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fldd		d0, ALPHA_R
 	fldd		d1, ALPHA_I
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
 
 	FMAC_R1 d4 , d0 , d8
 	FMAC_I1 d5 , d0 , d9
@@ -888,6 +892,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC

kernel/arm/ztrmm_kernel_2x2_vfpv3.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -134,7 +138,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -388,7 +392,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -566,7 +570,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -743,7 +747,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d24, d16
 	vmov.f64		d25, d16
@@ -889,6 +893,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC

kernel/arm64/zaxpy.S

@@ -147,12 +147,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fmla	v4.4s, v3.4s, v17.4s
 #endif
 
-#if !defined(CONJ)
+
 	fmla	v5.4s, v2.4s, v17.4s
-#else
-	fmls	v5.4s, v2.4s, v17.4s
-#endif
+#if !defined(CONJ)
 	fmla	v5.4s, v3.4s, v16.4s
+#else
+	fmls	v5.4s, v3.4s, v16.4s
+#endif
 
 	st2	{v4.4s, v5.4s}, [Y], #32
 #else // DOUBLE
@@ -165,12 +166,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #else
 	fmla	v4.2d, v3.2d, v17.2d
 #endif
-#if !defined(CONJ)
+
 	fmla	v5.2d, v2.2d, v17.2d
-#else
-	fmls	v5.2d, v2.2d, v17.2d
-#endif
+#if !defined(CONJ)
 	fmla	v5.2d, v3.2d, v16.2d
+#else
+	fmls	v5.2d, v3.2d, v16.2d
+#endif
 
 	st2	{v4.2d, v5.2d}, [Y], #32
 
@@ -183,13 +185,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #else
 	fmla	v20.2d, v19.2d, v17.2d
 #endif
-#if !defined(CONJ)
-	fmla	v21.2d, v18.2d, v17.2d
-#else
-	fmls	v21.2d, v18.2d, v17.2d
-#endif
-	fmla	v21.2d, v19.2d, v16.2d
 
+	fmla	v21.2d, v18.2d, v17.2d
+#if !defined(CONJ)
+	fmla	v21.2d, v19.2d, v16.2d
+#else
+	fmls	v21.2d, v19.2d, v16.2d
+#endif
 	st2	{v20.2d, v21.2d}, [Y], #32
 #endif
 	PRFM	PLDL1KEEP, [X, #512]

kernel/generic/zgemmkernel_2x2.c

@@ -797,7 +797,7 @@ int CNAME(BLASLONG bm,BLASLONG bn,BLASLONG bk,FLOAT alphar,FLOAT alphai,FLOAT* b
                   res1 = res1-load0*load3;
 #endif
 #if   defined(RN) || defined(RT) || defined(CN) || defined(CT)
-				  load0 = ptrba[2*0+0];
+                  load0 = ptrba[2*0+0];
                   load1 = ptrbb[2*0+0];
                   res0 = res0+load0*load1;
                   load2 = ptrba[2*0+1];

kernel/x86_64/KERNEL.BULLDOZER

@@ -69,24 +69,24 @@ ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_barcelona.S
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
 
-STRSMKERNEL_LN  =  ../generic/trsm_kernel_LN.c
-STRSMKERNEL_LT  =  ../generic/trsm_kernel_LT.c
-STRSMKERNEL_RN  =  ../generic/trsm_kernel_RN.c
-STRSMKERNEL_RT  =  ../generic/trsm_kernel_RT.c
+STRSMKERNEL_LN  = strsm_kernel_LN_bulldozer.c 
+STRSMKERNEL_LT  = strsm_kernel_LT_bulldozer.c 
+STRSMKERNEL_RN  = strsm_kernel_RN_bulldozer.c 
+STRSMKERNEL_RT  = strsm_kernel_RT_bulldozer.c
 
-DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LN  = dtrsm_kernel_LN_bulldozer.c
 DTRSMKERNEL_LT  = dtrsm_kernel_LT_8x2_bulldozer.S
 DTRSMKERNEL_RN  = dtrsm_kernel_RN_8x2_bulldozer.S
-DTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+DTRSMKERNEL_RT  = dtrsm_kernel_RT_bulldozer.c
 
-CTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
-CTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
-CTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
-CTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+CTRSMKERNEL_LN  = ctrsm_kernel_LN_bulldozer.c
+CTRSMKERNEL_LT  = ctrsm_kernel_LT_bulldozer.c
+CTRSMKERNEL_RN  = ctrsm_kernel_RN_bulldozer.c
+CTRSMKERNEL_RT  = ctrsm_kernel_RT_bulldozer.c
 
-ZTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
-ZTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
-ZTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
-ZTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+ZTRSMKERNEL_LN  = ztrsm_kernel_LN_bulldozer.c
+ZTRSMKERNEL_LT  = ztrsm_kernel_LT_bulldozer.c
+ZTRSMKERNEL_RN  = ztrsm_kernel_RN_bulldozer.c
+ZTRSMKERNEL_RT  = ztrsm_kernel_RT_bulldozer.c
 
 

kernel/x86_64/KERNEL.HASWELL

@@ -80,7 +80,7 @@ STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
 
 DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
 DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
-DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RN	= dtrsm_kernel_RN_haswell.c
 DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 
 CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c

kernel/x86_64/KERNEL.PILEDRIVER

@@ -66,25 +66,23 @@ ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_barcelona.S
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
 
-STRSMKERNEL_LN  =  ../generic/trsm_kernel_LN.c
-STRSMKERNEL_LT  =  ../generic/trsm_kernel_LT.c
-STRSMKERNEL_RN  =  ../generic/trsm_kernel_RN.c
-STRSMKERNEL_RT  =  ../generic/trsm_kernel_RT.c
+STRSMKERNEL_LN  = strsm_kernel_LN_bulldozer.c 
+STRSMKERNEL_LT  = strsm_kernel_LT_bulldozer.c 
+STRSMKERNEL_RN  = strsm_kernel_RN_bulldozer.c 
+STRSMKERNEL_RT  = strsm_kernel_RT_bulldozer.c
 
-
-DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LN  = dtrsm_kernel_LN_bulldozer.c
 DTRSMKERNEL_LT  = dtrsm_kernel_LT_8x2_bulldozer.S
 DTRSMKERNEL_RN  = dtrsm_kernel_RN_8x2_bulldozer.S
-DTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+DTRSMKERNEL_RT  = dtrsm_kernel_RT_bulldozer.c
 
-CTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
-CTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
-CTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
-CTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
-
-ZTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
-ZTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
-ZTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
-ZTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+CTRSMKERNEL_LN  = ctrsm_kernel_LN_bulldozer.c
+CTRSMKERNEL_LT  = ctrsm_kernel_LT_bulldozer.c
+CTRSMKERNEL_RN  = ctrsm_kernel_RN_bulldozer.c
+CTRSMKERNEL_RT  = ctrsm_kernel_RT_bulldozer.c
 
+ZTRSMKERNEL_LN  = ztrsm_kernel_LN_bulldozer.c
+ZTRSMKERNEL_LT  = ztrsm_kernel_LT_bulldozer.c
+ZTRSMKERNEL_RN  = ztrsm_kernel_RN_bulldozer.c
+ZTRSMKERNEL_RT  = ztrsm_kernel_RT_bulldozer.c
 

kernel/x86_64/KERNEL.STEAMROLLER

@@ -72,25 +72,23 @@ ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)
 CGEMM3MKERNEL    =  zgemm3m_kernel_8x4_barcelona.S
 ZGEMM3MKERNEL    =  zgemm3m_kernel_4x4_barcelona.S
 
-STRSMKERNEL_LN  =  ../generic/trsm_kernel_LN.c
-STRSMKERNEL_LT  =  ../generic/trsm_kernel_LT.c
-STRSMKERNEL_RN  =  ../generic/trsm_kernel_RN.c
-STRSMKERNEL_RT  =  ../generic/trsm_kernel_RT.c
+STRSMKERNEL_LN  = strsm_kernel_LN_bulldozer.c 
+STRSMKERNEL_LT  = strsm_kernel_LT_bulldozer.c 
+STRSMKERNEL_RN  = strsm_kernel_RN_bulldozer.c 
+STRSMKERNEL_RT  = strsm_kernel_RT_bulldozer.c
 
-
-DTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LN  = dtrsm_kernel_LN_bulldozer.c
 DTRSMKERNEL_LT  = dtrsm_kernel_LT_8x2_bulldozer.S
 DTRSMKERNEL_RN  = dtrsm_kernel_RN_8x2_bulldozer.S
-DTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+DTRSMKERNEL_RT  = dtrsm_kernel_RT_bulldozer.c
 
-CTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
-CTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
-CTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
-CTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
-
-ZTRSMKERNEL_LN  = ../generic/trsm_kernel_LN.c
-ZTRSMKERNEL_LT  = ../generic/trsm_kernel_LT.c
-ZTRSMKERNEL_RN  = ../generic/trsm_kernel_RN.c
-ZTRSMKERNEL_RT  = ../generic/trsm_kernel_RT.c
+CTRSMKERNEL_LN  = ctrsm_kernel_LN_bulldozer.c
+CTRSMKERNEL_LT  = ctrsm_kernel_LT_bulldozer.c
+CTRSMKERNEL_RN  = ctrsm_kernel_RN_bulldozer.c
+CTRSMKERNEL_RT  = ctrsm_kernel_RT_bulldozer.c
 
+ZTRSMKERNEL_LN  = ztrsm_kernel_LN_bulldozer.c
+ZTRSMKERNEL_LT  = ztrsm_kernel_LT_bulldozer.c
+ZTRSMKERNEL_RN  = ztrsm_kernel_RN_bulldozer.c
+ZTRSMKERNEL_RT  = ztrsm_kernel_RT_bulldozer.c
 

kernel/x86_64/cgemv_n_microk_haswell-4.c

@@ -51,7 +51,7 @@ static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"cmpq		$0 , %1				\n\t"
 	"je		2f			        \n\t"
 
-	".align 16				        \n\t"
+	//	".align 16				        \n\t"
 	"1:				        \n\t"
 	"prefetcht0      320(%4,%0,4)			\n\t"
 	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
@@ -202,7 +202,7 @@ static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"cmpq		$0 , %1				\n\t"
 	"je		2f			        \n\t"
 
-	".align 16				        \n\t"
+	//	".align 16				        \n\t"
 	"1:				        \n\t"
 	"prefetcht0      320(%4,%0,4)			\n\t"
 	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
@@ -322,7 +322,7 @@ static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 	"cmpq		$0 , %1				\n\t"
 	"je		2f			        \n\t"
 
-	".align 16				        \n\t"
+	//	".align 16				        \n\t"
 	"1:				        \n\t"
 	"prefetcht0      320(%4,%0,4)			\n\t"
 	"vmovups	(%4,%0,4), %%ymm8	        \n\t" // 4 complex values form a0
@@ -454,7 +454,7 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT a
 	"cmpq		$0 , %1				\n\t"
 	"je		2f			        \n\t"
 
-	".align 16				        \n\t"
+	//	".align 16				        \n\t"
 	"1:				        \n\t"
 	"vmovups	(%2,%0,4), %%ymm8	        \n\t" // 4 complex values from src
 	"vmovups      32(%2,%0,4), %%ymm9	        \n\t" 

kernel/x86_64/cgemv_t_microk_haswell-4.c

@@ -76,7 +76,7 @@ static void cgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"cmpq           $0, %1                         \n\t"
         "je             3f                      \n\t"
 
-	".align 16				        \n\t"
+	//	".align 16				        \n\t"
 	"1:				        \n\t"
         "prefetcht0      192(%4,%0,4)                   \n\t"
 	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
@@ -292,7 +292,7 @@ static void cgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"cmpq           $0, %1                         \n\t"
         "je             3f                      \n\t"
 
-	".align 16				        \n\t"
+	//	".align 16				        \n\t"
 	"1:				        \n\t"
         "prefetcht0      192(%4,%0,4)                   \n\t"
 	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0
@@ -446,7 +446,7 @@ static void cgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *
 	"cmpq           $0, %1                         \n\t"
         "je             3f                      \n\t"
 
-	".align 16				        \n\t"
+	//	".align 16				        \n\t"
 	"1:				        \n\t"
         "prefetcht0      192(%4,%0,4)                   \n\t"
 	"vmovups	(%4,%0,4), %%ymm4	        \n\t" // 4 complex values from a0

kernel/x86_64/ctrsm_kernel_LN_bulldozer.c

@@ -0,0 +1,472 @@
+/*********************************************************************/
+/* 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 "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ctrsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ctrsm_LN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss        (%3,%1,4), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vbroadcastss       4(%3,%1,4), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vbroadcastss       8(%3,%1,4), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vbroadcastss      12(%3,%1,4), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+
+	"	vfnmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufps		$0xb1 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufps		$0xb1 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufps		$0xb1 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufps		$0xb1 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorps		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddps		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddps		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddps		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddps		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (m - 1) * m;
+  b += (m - 1) * n;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a -= m;
+    b -= 2 * n;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+  a += (m - 1) * m * 2;
+  b += (m - 1) * n * 2;
+
+  for (i = m - 1; i >= 0; i--) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -=   cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= - cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a -= m * 2;
+    b -= 4 * n;
+  }
+
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LN : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = m + offset;
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	if (m & i) {
+	  aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	  cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	  if (k - kk > 0) {
+	    GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + i             * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+	  }
+
+	  solve(i, GEMM_UNROLL_N,
+		aa + (kk - i) * i             * COMPSIZE,
+		b  + (kk - i) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  kk -= i;
+	}
+      }
+    }
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+    if (i > 0) {
+      aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+      cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+      do {
+
+#ifdef CONJ
+	if (k - kk > 0) {
+	  GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+		      ZERO,
+#endif
+		      aa + GEMM_UNROLL_M * kk * COMPSIZE,
+		      b +  GEMM_UNROLL_N * kk * COMPSIZE,
+		      cc,
+		      ldc);
+	}
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+	      b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+	ctrsm_LN_solve_opt(k-kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b +  GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+			   aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE, b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+	      b  + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+	aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	cc -= GEMM_UNROLL_M     * COMPSIZE;
+	kk -= GEMM_UNROLL_M;
+	i --;
+      } while (i > 0);
+    }
+
+    b += GEMM_UNROLL_N * k * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = m + offset;
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  for (i = 1; i < GEMM_UNROLL_M; i *= 2){
+	    if (m & i) {
+	      aa = a + ((m & ~(i - 1)) - i) * k * COMPSIZE;
+	      cc = c + ((m & ~(i - 1)) - i)     * COMPSIZE;
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - i) * i * COMPSIZE,
+		    b  + (kk - i) * j * COMPSIZE,
+		    cc, ldc);
+
+	      kk -= i;
+	    }
+	  }
+	}
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+	  aa = a + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M) * k * COMPSIZE;
+	  cc = c + ((m & ~(GEMM_UNROLL_M - 1)) - GEMM_UNROLL_M)     * COMPSIZE;
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b +  j             * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - GEMM_UNROLL_M) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_M) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa -= GEMM_UNROLL_M * k * COMPSIZE;
+	    cc -= GEMM_UNROLL_M     * COMPSIZE;
+	    kk -= GEMM_UNROLL_M;
+	    i --;
+	  } while (i > 0);
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ctrsm_kernel_LT_bulldozer.c

@@ -0,0 +1,455 @@
+/*********************************************************************/
+/* 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 "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_L
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ctrsm_LT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ctrsm_LT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss        (%3,%1,4), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vbroadcastss       4(%3,%1,4), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vbroadcastss       8(%3,%1,4), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vbroadcastss      12(%3,%1,4), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+
+	"	vfnmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufps		$0xb1 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufps		$0xb1 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufps		$0xb1 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufps		$0xb1 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorps		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddps		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddps		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddps		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddps		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < m; i++) {
+
+    aa = *(a + i);
+
+    for (j = 0; j < n; j ++) {
+      bb = *(c + i + j * ldc);
+      bb *= aa;
+      *b             = bb;
+      *(c + i + j * ldc) = bb;
+      b ++;
+
+      for (k = i + 1; k < m; k ++){
+	*(c + k + j * ldc) -= bb * *(a + k);
+      }
+
+    }
+    a += m;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < m; i++) {
+
+    aa1 = *(a + i * 2 + 0);
+    aa2 = *(a + i * 2 + 1);
+
+    for (j = 0; j < n; j ++) {
+      bb1 = *(c + i * 2 + 0 + j * ldc);
+      bb2 = *(c + i * 2 + 1 + j * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 = aa1 * bb1 + aa2 * bb2;
+      cc2 = aa1 * bb2 - aa2 * bb1;
+#endif
+
+      *(b + 0) = cc1;
+      *(b + 1) = cc2;
+      *(c + i * 2 + 0 + j * ldc) = cc1;
+      *(c + i * 2 + 1 + j * ldc) = cc2;
+      b += 2;
+
+      for (k = i + 1; k < m; k ++){
+#ifndef CONJ
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) - cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#else
+	*(c + k * 2 + 0 + j * ldc) -= cc1 * *(a + k * 2 + 0) + cc2 * *(a + k * 2 + 1);
+	*(c + k * 2 + 1 + j * ldc) -= -cc1 * *(a + k * 2 + 1) + cc2 * *(a + k * 2 + 0);
+#endif
+      }
+
+    }
+    a += m * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM KERNEL LT : m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  while (j > 0) {
+
+    kk = offset;
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    while (i > 0) {
+
+#ifdef CONJ
+
+	if (kk > 0) {
+	  GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+		      ZERO,
+#endif
+		      aa, b, cc, ldc);
+	}
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+	ctrsm_LT_solve_opt(kk, aa, b, cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b  + kk * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+      aa += GEMM_UNROLL_M * k * COMPSIZE;
+      cc += GEMM_UNROLL_M     * COMPSIZE;
+      kk += GEMM_UNROLL_M;
+      i --;
+    }
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	  kk += i;
+	}
+	i >>= 1;
+      }
+    }
+
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	kk = offset;
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  kk += GEMM_UNROLL_M;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      kk += i;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ctrsm_kernel_RN_bulldozer.c

@@ -0,0 +1,454 @@
+/*********************************************************************/
+/* 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 "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ctrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ctrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss        (%3,%1,4), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vbroadcastss       4(%3,%1,4), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vbroadcastss       8(%3,%1,4), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vbroadcastss      12(%3,%1,4), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+
+	"	vfnmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufps		$0xb1 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufps		$0xb1 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufps		$0xb1 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufps		$0xb1 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorps		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddps		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddps		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddps		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddps		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa, bb;
+
+  int i, j, k;
+
+  for (i = 0; i < n; i++) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a  = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = i + 1; k < n; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b += n;
+  }
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  for (i = 0; i < n; i++) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1 + aa2 * bb2;
+      cc2 = -aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = i + 1; k < n; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= - cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b += n * 2;
+  }
+}
+
+#endif
+
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k, FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+  BLASLONG i, j, jj;
+
+#if 0
+  fprintf(stderr, "TRSM RN KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  jj = 0;
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+  kk = -offset;
+
+  while (j > 0) {
+
+    aa = a;
+    cc = c;
+
+    i = (m >> GEMM_UNROLL_M_SHIFT);
+
+    if (i > 0) {
+      do {
+
+#ifndef CONJ
+
+        ctrsm_RN_solve_opt(kk, aa, b, cc, ldc, aa + kk * GEMM_UNROLL_M * COMPSIZE, b  + kk * GEMM_UNROLL_N * COMPSIZE);
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#else
+
+        if (kk > 0) {
+          GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+                      ZERO,
+#endif
+                      aa, b, cc, ldc);
+        }
+
+	solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+	      aa + kk * GEMM_UNROLL_M * COMPSIZE,
+	      b  + kk * GEMM_UNROLL_N * COMPSIZE,
+	      cc, ldc);
+#endif
+
+	aa += GEMM_UNROLL_M * k * COMPSIZE;
+	cc += GEMM_UNROLL_M     * COMPSIZE;
+	i --;
+      } while (i > 0);
+    }
+
+
+    if (m & (GEMM_UNROLL_M - 1)) {
+      i = (GEMM_UNROLL_M >> 1);
+      while (i > 0) {
+	if (m & i) {
+	    if (kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa, b, cc, ldc);
+	    }
+	  solve(i, GEMM_UNROLL_N,
+		aa + kk * i             * COMPSIZE,
+		b  + kk * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+	  aa += i * k * COMPSIZE;
+	  cc += i     * COMPSIZE;
+	}
+	i >>= 1;
+      }
+    }
+
+    kk += GEMM_UNROLL_N;
+    b += GEMM_UNROLL_N * k   * COMPSIZE;
+    c += GEMM_UNROLL_N * ldc * COMPSIZE;
+    j --;
+    jj += GEMM_UNROLL_M;
+  }
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = (GEMM_UNROLL_N >> 1);
+    while (j > 0) {
+      if (n & j) {
+
+	aa = a;
+	cc = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+
+	while (i > 0) {
+	  if (kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, j, kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa,
+			b,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, j,
+		aa + kk * GEMM_UNROLL_M * COMPSIZE,
+		b  + kk * j             * COMPSIZE, cc, ldc);
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  while (i > 0) {
+	    if (m & i) {
+	      if (kk > 0) {
+		GEMM_KERNEL(i, j, kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa,
+			    b,
+			    cc,
+			    ldc);
+	      }
+
+	      solve(i, j,
+		    aa + kk * i * COMPSIZE,
+		    b  + kk * j * COMPSIZE, cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+	      }
+	    i >>= 1;
+	  }
+	}
+
+	b += j * k   * COMPSIZE;
+	c += j * ldc * COMPSIZE;
+	kk += j;
+      }
+      j >>= 1;
+    }
+  }
+
+  return 0;
+}

kernel/x86_64/ctrsm_kernel_RT_bulldozer.c

@@ -0,0 +1,482 @@
+/*********************************************************************/
+/* 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 "common.h"
+
+static FLOAT dm1 = -1.;
+
+#ifdef CONJ
+#define GEMM_KERNEL   GEMM_KERNEL_R
+#else
+#define GEMM_KERNEL   GEMM_KERNEL_N
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 1
+#define GEMM_UNROLL_M_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 2
+#define GEMM_UNROLL_M_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 4
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 6
+#define GEMM_UNROLL_M_SHIFT 2
+#endif
+
+
+#if GEMM_DEFAULT_UNROLL_M == 8
+#define GEMM_UNROLL_M_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_M == 16
+#define GEMM_UNROLL_M_SHIFT 4
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 1
+#define GEMM_UNROLL_N_SHIFT 0
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 2
+#define GEMM_UNROLL_N_SHIFT 1
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 4
+#define GEMM_UNROLL_N_SHIFT 2
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 8
+#define GEMM_UNROLL_N_SHIFT 3
+#endif
+
+#if GEMM_DEFAULT_UNROLL_N == 16
+#define GEMM_UNROLL_N_SHIFT 4
+#endif
+
+
+
+#ifndef CONJ
+
+static void ctrsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)  __attribute__ ((noinline));
+
+static void ctrsm_RT_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, FLOAT *as, FLOAT *bs)
+{
+
+	FLOAT *c1 = c + ldc*2 ;
+	BLASLONG n1 = n * 4;
+	BLASLONG i=0;
+
+        __asm__  __volatile__
+        (
+	"	vzeroupper							\n\t"
+	"	prefetcht0	(%4)						\n\t"
+	"	prefetcht0	(%5)						\n\t"
+	"	vxorps	%%xmm8 , %%xmm8 , %%xmm8				\n\t"
+	"	vxorps	%%xmm9 , %%xmm9 , %%xmm9				\n\t"
+	"	vxorps	%%xmm10, %%xmm10, %%xmm10				\n\t"
+	"	vxorps	%%xmm11, %%xmm11, %%xmm11				\n\t"
+	"	vxorps	%%xmm12, %%xmm12, %%xmm12				\n\t"
+	"	vxorps	%%xmm13, %%xmm13, %%xmm13				\n\t"
+	"	vxorps	%%xmm14, %%xmm14, %%xmm14				\n\t"
+	"	vxorps	%%xmm15, %%xmm15, %%xmm15				\n\t"
+
+	"	cmpq	       $0, %0						\n\t"
+	"	je	       3f						\n\t"
+
+	"	.align 16							\n\t"
+	"1:									\n\t"
+
+	"	vbroadcastss        (%3,%1,4), %%xmm0				\n\t"	// b0 real, b0 real
+	"	vbroadcastss       4(%3,%1,4), %%xmm1				\n\t"	// b0 imag, b0 imag
+	"	vbroadcastss       8(%3,%1,4), %%xmm2				\n\t"	// b1 real, b1 real
+	"	vbroadcastss      12(%3,%1,4), %%xmm3				\n\t"	// b1 imag, b1 imag
+
+	"	vmovups         (%2,%1,8), %%xmm4				\n\t"	// a0 real , a0 imag
+	"	vmovups       16(%2,%1,8), %%xmm5				\n\t"	// a1 real , a1 imag
+
+	"	vfnmaddps	%%xmm8 , %%xmm0 , %%xmm4 , %%xmm8		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm9 , %%xmm1 , %%xmm4 , %%xmm9		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm10, %%xmm0 , %%xmm5 , %%xmm10		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm11, %%xmm1 , %%xmm5 , %%xmm11		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm12, %%xmm2 , %%xmm4 , %%xmm12		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm13, %%xmm3 , %%xmm4 , %%xmm13		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	vfnmaddps	%%xmm14, %%xmm2 , %%xmm5 , %%xmm14		\n\t"   // a_real * b_real , a_imag * b_real
+	"	vfnmaddps	%%xmm15, %%xmm3 , %%xmm5 , %%xmm15		\n\t"   // a_real * b_imag , a_imag * b_imag
+
+	"	addq		$4, %1						\n\t"
+	"	cmpq		%1, %0						\n\t"
+
+	"	jnz		1b						\n\t"
+
+
+	"2:									\n\t"
+
+	"	vshufps		$0xb1 , %%xmm9  , %%xmm9, %%xmm9		\n\t"
+	"	vshufps		$0xb1 , %%xmm11 , %%xmm11 , %%xmm11		\n\t"
+	"	vshufps		$0xb1 , %%xmm13 , %%xmm13 , %%xmm13		\n\t"
+	"	vshufps		$0xb1 , %%xmm15 , %%xmm15 , %%xmm15		\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm9 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm11, %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm13, %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm15, %%xmm14		\n\t"
+
+	"	vxorps		%%xmm7 , %%xmm7 , %%xmm7			\n\t"
+
+	"	vaddsubps		%%xmm8 , %%xmm7 , %%xmm8		\n\t"
+	"	vaddsubps		%%xmm10, %%xmm7 , %%xmm10		\n\t"
+	"	vaddsubps		%%xmm12, %%xmm7 , %%xmm12		\n\t"
+	"	vaddsubps		%%xmm14, %%xmm7 , %%xmm14		\n\t"
+
+	"	vmovups		  (%4) , %%xmm0					\n\t"
+	"	vmovups		16(%4) , %%xmm1					\n\t"
+
+	"	vmovups		  (%5) , %%xmm4					\n\t"
+	"	vmovups		16(%5) , %%xmm5					\n\t"
+
+	"	vaddps		%%xmm0 , %%xmm8 , %%xmm8			\n\t"
+	"	vaddps		%%xmm1 , %%xmm10, %%xmm10			\n\t"
+	"	vaddps		%%xmm4 , %%xmm12, %%xmm12			\n\t"
+	"	vaddps		%%xmm5 , %%xmm14, %%xmm14			\n\t"
+
+	"	vmovups		%%xmm8  ,  (%4)					\n\t"
+	"	vmovups		%%xmm10 ,16(%4)					\n\t"
+
+	"	vmovups		%%xmm12 ,  (%5)					\n\t"
+	"	vmovups		%%xmm14 ,16(%5)					\n\t"
+
+	"3:									\n\t"	
+
+	"	vzeroupper							\n\t"
+
+        :
+        :
+          "r" (n1),     // 0    
+          "a" (i),      // 1    
+          "r" (a),      // 2
+          "r" (b),      // 3
+          "r" (c),      // 4
+          "r" (c1),     // 5
+          "r" (as),     // 6
+          "r" (bs)      // 7
+        : "cc",
+          "%xmm0", "%xmm1", "%xmm2", "%xmm3",
+          "%xmm4", "%xmm5", "%xmm6", "%xmm7",
+          "%xmm8", "%xmm9", "%xmm10", "%xmm11",
+          "%xmm12", "%xmm13", "%xmm14", "%xmm15",
+          "memory"
+        );
+
+}
+
+#endif
+
+
+
+#ifndef COMPLEX
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa,  bb;
+
+  int i, j, k;
+
+  a += (n - 1) * m;
+  b += (n - 1) * n;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb = *(b + i);
+
+    for (j = 0; j < m; j ++) {
+      aa = *(c + j + i * ldc);
+      aa *= bb;
+      *a   = aa;
+      *(c + j + i * ldc) = aa;
+      a ++;
+
+      for (k = 0; k < i; k ++){
+	*(c + j + k * ldc) -= aa * *(b + k);
+      }
+
+    }
+    b -= n;
+    a -= 2 * m;
+  }
+
+}
+
+#else
+
+static inline void solve(BLASLONG m, BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc) {
+
+  FLOAT aa1, aa2;
+  FLOAT bb1, bb2;
+  FLOAT cc1, cc2;
+
+  int i, j, k;
+
+  ldc *= 2;
+
+  a += (n - 1) * m * 2;
+  b += (n - 1) * n * 2;
+
+  for (i = n - 1; i >= 0; i--) {
+
+    bb1 = *(b + i * 2 + 0);
+    bb2 = *(b + i * 2 + 1);
+
+    for (j = 0; j < m; j ++) {
+
+      aa1 = *(c + j * 2 + 0 + i * ldc);
+      aa2 = *(c + j * 2 + 1 + i * ldc);
+
+#ifndef CONJ
+      cc1 = aa1 * bb1 - aa2 * bb2;
+      cc2 = aa1 * bb2 + aa2 * bb1;
+#else
+      cc1 =  aa1 * bb1  + aa2 * bb2;
+      cc2 = - aa1 * bb2 + aa2 * bb1;
+#endif
+
+      *(a + 0) = cc1;
+      *(a + 1) = cc2;
+
+      *(c + j * 2 + 0 + i * ldc) = cc1;
+      *(c + j * 2 + 1 + i * ldc) = cc2;
+      a += 2;
+
+      for (k = 0; k < i; k ++){
+#ifndef CONJ
+	*(c + j * 2 + 0 + k * ldc) -= cc1 * *(b + k * 2 + 0) - cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -= cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#else
+	*(c + j * 2 + 0 + k * ldc) -=   cc1 * *(b + k * 2 + 0) + cc2 * *(b + k * 2 + 1);
+	*(c + j * 2 + 1 + k * ldc) -=  -cc1 * *(b + k * 2 + 1) + cc2 * *(b + k * 2 + 0);
+#endif
+      }
+
+    }
+    b -= n * 2;
+    a -= 4 * m;
+  }
+
+}
+
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG k,  FLOAT dummy1,
+#ifdef COMPLEX
+	   FLOAT dummy2,
+#endif
+	   FLOAT *a, FLOAT *b, FLOAT *c, BLASLONG ldc, BLASLONG offset){
+
+  BLASLONG i, j;
+  FLOAT *aa, *cc;
+  BLASLONG  kk;
+
+#if 0
+  fprintf(stderr, "TRSM RT KERNEL m = %3ld  n = %3ld  k = %3ld offset = %3ld\n",
+	  m, n, k, offset);
+#endif
+
+  kk = n - offset;
+  c += n * ldc * COMPSIZE;
+  b += n * k   * COMPSIZE;
+
+  if (n & (GEMM_UNROLL_N - 1)) {
+
+    j = 1;
+    while (j < GEMM_UNROLL_N) {
+      if (n & j) {
+
+	aa  = a;
+	b -= j * k  * COMPSIZE;
+	c -= j * ldc* COMPSIZE;
+	cc  = c;
+
+	i = (m >> GEMM_UNROLL_M_SHIFT);
+	if (i > 0) {
+
+	  do {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(GEMM_UNROLL_M, j, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			  b  +  j            * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(GEMM_UNROLL_M, j,
+		  aa + (kk - j) * GEMM_UNROLL_M * COMPSIZE,
+		  b  + (kk - j) * j             * COMPSIZE,
+		  cc, ldc);
+
+	    aa += GEMM_UNROLL_M * k * COMPSIZE;
+	    cc += GEMM_UNROLL_M     * COMPSIZE;
+	    i --;
+	  } while (i > 0);
+	}
+
+	if (m & (GEMM_UNROLL_M - 1)) {
+	  i = (GEMM_UNROLL_M >> 1);
+	  do {
+	    if (m & i) {
+
+	      if (k - kk > 0) {
+		GEMM_KERNEL(i, j, k - kk, dm1,
+#ifdef COMPLEX
+			    ZERO,
+#endif
+			    aa + i * kk * COMPSIZE,
+			    b  + j * kk * COMPSIZE,
+			    cc, ldc);
+	      }
+
+	      solve(i, j,
+		    aa + (kk - j) * i * COMPSIZE,
+		    b  + (kk - j) * j * COMPSIZE,
+		    cc, ldc);
+
+	      aa += i * k * COMPSIZE;
+	      cc += i     * COMPSIZE;
+
+	    }
+	    i >>= 1;
+	  } while (i > 0);
+	}
+	kk -= j;
+      }
+      j <<= 1;
+    }
+  }
+
+  j = (n >> GEMM_UNROLL_N_SHIFT);
+
+  if (j > 0) {
+
+    do {
+      aa  = a;
+      b -= GEMM_UNROLL_N * k   * COMPSIZE;
+      c -= GEMM_UNROLL_N * ldc * COMPSIZE;
+      cc  = c;
+
+      i = (m >> GEMM_UNROLL_M_SHIFT);
+      if (i > 0) {
+	do {
+
+#ifndef CONJ
+
+	  ctrsm_RT_solve_opt(k-kk, aa + GEMM_UNROLL_M * kk * COMPSIZE, b  + GEMM_UNROLL_N * kk * COMPSIZE, cc, ldc,
+                             aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE, b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE);
+ 
+	  solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+		aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE,
+		b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+#else
+	  if (k - kk > 0) {
+	    GEMM_KERNEL(GEMM_UNROLL_M, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			ZERO,
+#endif
+			aa + GEMM_UNROLL_M * kk * COMPSIZE,
+			b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			cc,
+			ldc);
+	  }
+
+	  solve(GEMM_UNROLL_M, GEMM_UNROLL_N,
+		aa + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_M * COMPSIZE,
+		b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		cc, ldc);
+
+#endif
+
+	  aa += GEMM_UNROLL_M * k * COMPSIZE;
+	  cc += GEMM_UNROLL_M     * COMPSIZE;
+	  i --;
+	} while (i > 0);
+      }
+
+      if (m & (GEMM_UNROLL_M - 1)) {
+	i = (GEMM_UNROLL_M >> 1);
+	do {
+	  if (m & i) {
+	    if (k - kk > 0) {
+	      GEMM_KERNEL(i, GEMM_UNROLL_N, k - kk, dm1,
+#ifdef COMPLEX
+			  ZERO,
+#endif
+			  aa + i             * kk * COMPSIZE,
+			  b  + GEMM_UNROLL_N * kk * COMPSIZE,
+			  cc,
+			  ldc);
+	    }
+
+	    solve(i, GEMM_UNROLL_N,
+		  aa + (kk - GEMM_UNROLL_N) * i             * COMPSIZE,
+		  b  + (kk - GEMM_UNROLL_N) * GEMM_UNROLL_N * COMPSIZE,
+		  cc, ldc);
+
+	    aa += i * k * COMPSIZE;
+	    cc += i     * COMPSIZE;
+	  }
+	  i >>= 1;
+	} while (i > 0);
+      }
+
+      kk -= GEMM_UNROLL_N;
+      j --;
+    } while (j > 0);
+  }
+
+  return 0;
+}
+
+

kernel/x86_64/dgemv_n_4.c

@@ -82,7 +82,7 @@ static void dgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"shufpd $0,  %%xmm12, %%xmm12    \n\t"	
 	"shufpd $0,  %%xmm13, %%xmm13    \n\t"	
 
-	".align 16				       \n\t"
+	//	".align 16				       \n\t"
 	"1:				       \n\t"
 	"movups	       (%3,%0,8), %%xmm4	       \n\t"	// 2 * y
 	"movups	     16(%3,%0,8), %%xmm5	       \n\t"	// 2 * y
@@ -129,7 +129,7 @@ static void dgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 
 #endif
 
-#ifndef HAVE_KERNEL_4x2
+#ifndef HAVE_KERNEL_4x1
 
 static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
 
@@ -144,7 +144,7 @@ static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT *a
         "mulsd          (%5), %%xmm12            \n\t"  // alpha 
         "shufpd $0,  %%xmm12, %%xmm12            \n\t"
 
-        ".align 16                               \n\t"
+	//        ".align 16                               \n\t"
         "1:                             \n\t"
         "movups       (%4,%0,8), %%xmm8          \n\t"  // 2 * a
         "movups     16(%4,%0,8), %%xmm9          \n\t"  // 2 * a

kernel/x86_64/dgemv_n_microk_haswell-4.c

@@ -52,7 +52,7 @@ static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"subq	        $4 , %1			      \n\t"		
 	"jz		2f		      \n\t"
 
-	".align 16				 \n\t"
+	//		".align 16				 \n\t"
 	"1:				 \n\t"
 
 	"vmulpd        %%ymm0 , %%ymm12, %%ymm4      \n\t" 
@@ -114,3 +114,78 @@ static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 } 
 
 
+#define HAVE_KERNEL_4x2
+
+static void dgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha) __attribute__ ((noinline));
+
+static void dgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT *alpha)
+{
+
+	BLASLONG register i = 0;
+
+	__asm__  __volatile__
+	(
+	"vbroadcastsd    (%2), %%ymm12	 \n\t"	// x0 
+	"vbroadcastsd   8(%2), %%ymm13	 \n\t"	// x1 
+
+	"vmovups	(%4,%0,8), %%ymm0	 \n\t"
+	"vmovups	(%5,%0,8), %%ymm1	 \n\t"
+
+	"vbroadcastsd    (%6), %%ymm6 	 \n\t"	// alpha 
+
+        "addq		$4 , %0	  	 	      \n\t"
+	"subq	        $4 , %1			      \n\t"		
+	"jz		2f		      \n\t"
+
+	"1:				 \n\t"
+
+	"vmulpd        %%ymm0 , %%ymm12, %%ymm4      \n\t" 
+	"vmulpd        %%ymm1 , %%ymm13, %%ymm5      \n\t" 
+	"vmovups	(%4,%0,8), %%ymm0	 \n\t"
+	"vmovups	(%5,%0,8), %%ymm1	 \n\t"
+
+	"vmovups	-32(%3,%0,8), %%ymm8	       \n\t"	// 4 * y
+	"vaddpd		 %%ymm4 , %%ymm5 , %%ymm4      \n\t"
+	"vfmadd231pd     %%ymm6 , %%ymm4 , %%ymm8      \n\t"
+
+	"vmovups         %%ymm8,   -32(%3,%0,8)	      \n\t"	// 4 * y
+
+        "addq		$4 , %0	  	 	      \n\t"
+	"subq	        $4 , %1			      \n\t"		
+	"jnz		1b		      \n\t"
+	
+
+	"2:				 \n\t"
+
+	"vmulpd        %%ymm0 , %%ymm12, %%ymm4      \n\t" 
+	"vmulpd        %%ymm1 , %%ymm13, %%ymm5      \n\t" 
+
+
+	"vmovups	-32(%3,%0,8), %%ymm8	       \n\t"	// 4 * y
+	"vaddpd		 %%ymm4 , %%ymm5 , %%ymm4      \n\t"
+	"vfmadd231pd     %%ymm6 , %%ymm4 , %%ymm8      \n\t"
+
+	"vmovups  %%ymm8,   -32(%3,%0,8)	      \n\t"	// 4 * y
+
+
+	"vzeroupper			              \n\t"
+
+
+	:
+        : 
+          "r" (i),	// 0	
+	  "r" (n),  	// 1
+          "r" (x),      // 2
+          "r" (y),      // 3
+          "r" (ap[0]),  // 4
+          "r" (ap[1]),  // 5
+          "r" (alpha)   // 6
+	: "cc", 
+	  "%xmm0", "%xmm1", 
+	  "%xmm4", "%xmm5", 
+	  "%xmm6", 
+	  "%xmm8", 
+	  "%xmm12", "%xmm13",
+	  "memory"
+	);
+}

kernel/x86_64/dgemv_t_4.c

@@ -95,7 +95,7 @@ static void dgemv_kernel_4x2(BLASLONG n, FLOAT *ap0, FLOAT *ap1, FLOAT *x, FLOAT
 	"cmpq	$0, %1				\n\t"
 	"je	3f			\n\t"
 
-        ".align 16                              \n\t"
+	//        ".align 16                              \n\t"
         "1:                            \n\t"
 
 	"movups  (%5,%0,8) , %%xmm14		\n\t" // x
@@ -171,7 +171,7 @@ static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 	"cmpq	$0, %1				\n\t"
 	"je	3f			\n\t"
 
-        ".align 16                              \n\t"
+	//        ".align 16                              \n\t"
         "1:                            \n\t"
 
 	"movups    (%3,%0,8) , %%xmm12		\n\t"
@@ -245,7 +245,7 @@ static void add_y(BLASLONG n, FLOAT da , FLOAT *src, FLOAT *dest, BLASLONG inc_d
 	"movsd	 (%2) , %%xmm10                 \n\t"
 	"shufpd  $0 , %%xmm10 , %%xmm10		\n\t"
 
-        ".align 16                              \n\t"
+	//        ".align 16                              \n\t"
         "1:                            \n\t"
 
 	"movups  (%3,%0,8) , %%xmm12		\n\t"

kernel/x86_64/dgemv_t_microk_haswell-4.c

@@ -59,7 +59,7 @@ static void dgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
         "cmpq           $0, %1                         \n\t"
         "je             3f                      \n\t"
 
-	".align 16				 \n\t"
+	// ".align 16				 \n\t"
 	"1:				 \n\t"
 	// "prefetcht0	 384(%2,%0,8)		 \n\t"
 	"vmovups	(%2,%0,8), %%ymm12       \n\t"	// 4 * x