Revert "Fix 32bit HASWELL builds"

Do not try to build the fork utest when NO_CBLAS=1

.travis.yml

@@ -7,6 +7,7 @@ language: c
 jobs:
   include:
     - &test-ubuntu
+      os: linux
       stage: test
       compiler: gcc
       addons:
@@ -57,7 +58,8 @@ jobs:
         - TARGET_BOX=LINUX32
         - BTYPE="BINARY=32"
 
-    - stage: test
+    - os: linux
+      stage: test
       compiler: gcc
       addons:
         apt:
@@ -77,6 +79,7 @@ jobs:
     # which is slower than container-based infrastructure used for jobs
     # that don't require sudo.
     - &test-alpine
+      os: linux
       stage: test
       dist: trusty
       sudo: true
@@ -120,6 +123,7 @@ jobs:
         - BTYPE="BINARY=64 NO_AFFINITY=1 USE_OPENMP=0 NO_LAPACK=0 TARGET=core2"
 
     - &test-cmake
+      os: linux
       stage: test
       compiler: clang
       addons:
@@ -147,6 +151,23 @@ jobs:
       env:
         - CMAKE=1
 
+    - &test-macos
+      os: osx
+      stage: test
+      osx_image: xcode8
+      before_script:
+        - COMMON_FLAGS="DYNAMIC_ARCH=1 TARGET=NEHALEM NUM_THREADS=32"
+        - brew update
+        - brew install gcc # for gfortran
+      script:
+        - travis_wait 45 make QUIET_MAKE=1 $COMMON_FLAGS $BTYPE
+      env:
+        - BTYPE="BINARY=64 INTERFACE64=1"
+
+    - <<: *test-macos
+      env:
+        - BTYPE="BINARY=32"
+
 # whitelist
 branches:
   only:

Makefile

@@ -91,11 +91,7 @@ ifeq ($(OSNAME), $(filter $(OSNAME),Linux SunOS Android))
 	@ln -fs $(LIBSONAME) $(LIBPREFIX).so
 	@ln -fs $(LIBSONAME) $(LIBPREFIX).so.$(MAJOR_VERSION)
 endif
-ifeq ($(OSNAME), FreeBSD)
-	@$(MAKE) -C exports so
-	@ln -fs $(LIBSONAME) $(LIBPREFIX).so
-endif
-ifeq ($(OSNAME), NetBSD)
+ifeq ($(OSNAME), $(filter $(OSNAME),FreeBSD OpenBSD NetBSD DragonFly))
 	@$(MAKE) -C exports so
 	@ln -fs $(LIBSONAME) $(LIBPREFIX).so
 endif

Makefile.install

@@ -72,12 +72,7 @@ ifeq ($(OSNAME), $(filter $(OSNAME),Linux SunOS Android))
 	ln -fs $(LIBSONAME) $(LIBPREFIX).so ; \
 	ln -fs $(LIBSONAME) $(LIBPREFIX).so.$(MAJOR_VERSION)
 endif
-ifeq ($(OSNAME), FreeBSD)
-	@cp $(LIBSONAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
-	@cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \
-	ln -fs $(LIBSONAME) $(LIBPREFIX).so
-endif
-ifeq ($(OSNAME), NetBSD)
+ifeq ($(OSNAME), $(filter $(OSNAME),FreeBSD OpenBSD NetBSD DragonFly))
 	@cp $(LIBSONAME) "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)"
 	@cd "$(DESTDIR)$(OPENBLAS_LIBRARY_DIR)" ; \
 	ln -fs $(LIBSONAME) $(LIBPREFIX).so
@@ -115,7 +110,7 @@ endif
 
 ifndef NO_SHARED
 #ifeq logical or
-ifeq ($(OSNAME), $(filter $(OSNAME),Linux FreeBSD NetBSD))
+ifeq ($(OSNAME), $(filter $(OSNAME),Linux FreeBSD NetBSD OpenBSD DragonFly))
 	@echo "SET(OpenBLAS_LIBRARIES ${OPENBLAS_LIBRARY_DIR}/$(LIBPREFIX).so)" >> "$(DESTDIR)$(OPENBLAS_CMAKE_DIR)/$(OPENBLAS_CMAKE_CONFIG)"
 endif
 ifeq ($(OSNAME), $(filter $(OSNAME),WINNT CYGWIN_NT))

Makefile.rule

@@ -100,7 +100,7 @@ BUILD_LAPACK_DEPRECATED = 1
 NO_WARMUP = 1
 
 # If you want to disable CPU/Memory affinity on Linux.
-#NO_AFFINITY = 1
+NO_AFFINITY = 1
 
 # if you are compiling for Linux and you have more than 16 numa nodes or more than 256 cpus
 # BIGNUMA = 1

Makefile.system

@@ -17,15 +17,24 @@ NETLIB_LAPACK_DIR = $(TOPDIR)/lapack-netlib
 #   http://stackoverflow.com/questions/4029274/mingw-and-make-variables
 # - Default value is 'cc' which is not always a valid command (e.g. MinGW).
 ifeq ($(origin CC),default)
+
+# Check if $(CC) refers to a valid command and set the value to gcc if not
+ifneq ($(findstring cmd.exe,$(SHELL)),)
+ifeq ($(shell where $(CC) 2>NUL),)
 CC = gcc
-# Change the default compile to clang on Mac OSX.
-# http://stackoverflow.com/questions/714100/os-detecting-makefile
-UNAME_S := $(shell uname -s)
-ifeq ($(UNAME_S),Darwin)
-     CC = clang
-#     EXTRALIB += -Wl,-no_compact_unwind
-endif
 endif
+else # POSIX-ish
+ifeq ($(shell command -v $(CC) 2>/dev/null),)
+ifeq ($(shell uname -s),Darwin)
+CC = clang
+# EXTRALIB += -Wl,-no_compact_unwind
+else
+CC = gcc
+endif # Darwin
+endif # CC exists
+endif # Shell is sane
+
+endif # CC is set to default
 
 # Default Fortran compiler (FC) is selected by f_check.
 
@@ -230,7 +239,7 @@ endif
 MD5SUM = md5 -r
 endif
 
-ifeq ($(OSNAME), FreeBSD)
+ifneq (,$(findstring $(OSNAME), FreeBSD OpenBSD DragonFly))
 MD5SUM = md5 -r
 endif
 

README.md

@@ -5,175 +5,219 @@
 Travis CI: [![Build Status](https://travis-ci.org/xianyi/OpenBLAS.svg?branch=develop)](https://travis-ci.org/xianyi/OpenBLAS)
 
 AppVeyor: [![Build status](https://ci.appveyor.com/api/projects/status/09sohd35n8nkkx64/branch/develop?svg=true)](https://ci.appveyor.com/project/xianyi/openblas/branch/develop)
+
 ## Introduction
+
 OpenBLAS is an optimized BLAS library based on GotoBLAS2 1.13 BSD version.
 
-Please read the documents on OpenBLAS wiki pages <http://github.com/xianyi/OpenBLAS/wiki>.
+Please read the documentation on the OpenBLAS wiki pages: <http://github.com/xianyi/OpenBLAS/wiki>.
 
 ## Binary Packages
-We provide binary packages for the following platform.
+
+We provide official binary packages for the following platform:
 
   * Windows x86/x86_64
 
 You can download them from [file hosting on sourceforge.net](https://sourceforge.net/projects/openblas/files/).
 
 ## Installation from Source
-Download from project homepage. http://xianyi.github.com/OpenBLAS/
 
-Or, check out codes from git://github.com/xianyi/OpenBLAS.git
+Download from project homepage, http://xianyi.github.com/OpenBLAS/, or check out the code
+using Git from https://github.com/xianyi/OpenBLAS.git.
+
+### Dependencies
+
+Building OpenBLAS requires the following to be installed:
+
+* GNU Make
+* A C compiler, e.g. GCC or Clang
+* A Fortran compiler (optional, for LAPACK)
+* IBM MASS (optional, see below)
+
 ### Normal compile
-  * type "make" to detect the CPU automatically.
-  or
-  * type "make TARGET=xxx" to set target CPU, e.g. "make TARGET=NEHALEM". The full target list is in file TargetList.txt.
+
+Simply invoking `make` (or `gmake` on BSD) will detect the CPU automatically.
+To set a specific target CPU, use `make TARGET=xxx`, e.g. `make TARGET=NEHALEM`.
+The full target list is in the file `TargetList.txt`.
 
 ### Cross compile
-Please set CC and FC with the cross toolchains. Then, set HOSTCC with your host C compiler. At last, set TARGET explicitly.
+
+Set `CC` and `FC` to point to the cross toolchains, and set `HOSTCC` to your host C compiler.
+The target must be specified explicitly when cross compiling.
 
 Examples:
 
-On X86 box, compile this library for loongson3a CPU.
+* On an x86 box, compile this library for a loongson3a CPU:
+  ```sh
+  make BINARY=64 CC=mips64el-unknown-linux-gnu-gcc FC=mips64el-unknown-linux-gnu-gfortran HOSTCC=gcc TARGET=LOONGSON3A
+  ```
 
-    make BINARY=64 CC=mips64el-unknown-linux-gnu-gcc FC=mips64el-unknown-linux-gnu-gfortran HOSTCC=gcc TARGET=LOONGSON3A
-
-On X86 box, compile this library for loongson3a CPU with loongcc (based on Open64) compiler.
-
-    make CC=loongcc FC=loongf95 HOSTCC=gcc TARGET=LOONGSON3A CROSS=1 CROSS_SUFFIX=mips64el-st-linux-gnu-   NO_LAPACKE=1 NO_SHARED=1 BINARY=32
+* On an x86 box, compile this library for a loongson3a CPU with loongcc (based on Open64) compiler:
+  ```sh
+  make CC=loongcc FC=loongf95 HOSTCC=gcc TARGET=LOONGSON3A CROSS=1 CROSS_SUFFIX=mips64el-st-linux-gnu-   NO_LAPACKE=1 NO_SHARED=1 BINARY=32
+  ```
 
 ### Debug version
 
-    make DEBUG=1
+A debug version can be built using `make DEBUG=1`.
 
-### Compile with MASS Support on Power CPU (Optional dependency)
+### Compile with MASS support on Power CPU (optional)
 
-[IBM MASS](http://www-01.ibm.com/software/awdtools/mass/linux/mass-linux.html) library consists of a set of mathematical functions for C, C++, and
-Fortran-language applications that are tuned for optimum performance on POWER architectures. OpenBLAS with MASS requires 64-bit, little-endian OS on POWER.
-The library can be installed as below -
+The [IBM MASS](http://www-01.ibm.com/software/awdtools/mass/linux/mass-linux.html) library
+consists of a set of mathematical functions for C, C++, and Fortran applications that are
+are tuned for optimum performance on POWER architectures.
+OpenBLAS with MASS requires a 64-bit, little-endian OS on POWER.
+The library can be installed as shown:
 
- * On Ubuntu:
+* On Ubuntu:
+  ```sh
+  wget -q http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/ubuntu/public.gpg -O- | sudo apt-key add -
+  echo "deb http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/ubuntu/ trusty main" | sudo tee /etc/apt/sources.list.d/ibm-xl-compiler-eval.list
+  sudo apt-get update
+  sudo apt-get install libxlmass-devel.8.1.5
+  ```
 
-    wget -q http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/ubuntu/public.gpg -O- | sudo apt-key add -</br>
-    echo "deb http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/ubuntu/ trusty main" | sudo tee /etc/apt/sources.list.d/ibm-xl-compiler-eval.list</br>
-    sudo apt-get update</br>
-    sudo apt-get install libxlmass-devel.8.1.5</br>
+* On RHEL/CentOS:
+  ```sh
+  wget http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/rhel7/repodata/repomd.xml.key
+  sudo rpm --import repomd.xml.key
+  wget http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/rhel7/ibm-xl-compiler-eval.repo
+  sudo cp ibm-xl-compiler-eval.repo /etc/yum.repos.d/
+  sudo yum install libxlmass-devel.8.1.5
+  ```
 
- * On RHEL/CentOS:
+After installing the MASS library, compile OpenBLAS with `USE_MASS=1`.
+For example, to compile on Power8 with MASS support: `make USE_MASS=1 TARGET=POWER8`.
 
-    wget http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/rhel7/repodata/repomd.xml.key</br>
-    sudo rpm --import repomd.xml.key</br>
-    wget http://public.dhe.ibm.com/software/server/POWER/Linux/xl-compiler/eval/ppc64le/rhel7/ibm-xl-compiler-eval.repo</br>
-    sudo cp ibm-xl-compiler-eval.repo /etc/yum.repos.d/</br>
-    sudo yum install libxlmass-devel.8.1.5</br>
+### Install to a specific directory (optional)
 
-After installing MASS library, compile openblas with USE_MASS=1.
+Use `PREFIX=` when invoking `make`, for example
 
-Example:
+```sh
+make install PREFIX=your_installation_directory
+```
 
-Compiling on Power8 with MASS support -
+The default installation directory is `/opt/OpenBLAS`.
 
-    make USE_MASS=1 TARGET=POWER8
+## Supported CPUs and Operating Systems
 
-### Install to the directory (optional)
+Please read `GotoBLAS_01Readme.txt`.
 
-Example:
+### Additional supported CPUs
 
-    make install PREFIX=your_installation_directory
+#### x86/x86-64
 
-The default directory is /opt/OpenBLAS
-
-## Support CPU & OS
-Please read GotoBLAS_01Readme.txt
-
-### Additional support CPU:
-
-#### x86/x86-64:
 - **Intel Xeon 56xx (Westmere)**: Used GotoBLAS2 Nehalem codes.
 - **Intel Sandy Bridge**: Optimized Level-3 and Level-2 BLAS with AVX on x86-64.
 - **Intel Haswell**: Optimized Level-3 and Level-2 BLAS with AVX2 and FMA  on x86-64.
 - **AMD Bobcat**: Used GotoBLAS2 Barcelona codes.
-- **AMD Bulldozer**: x86-64 ?GEMM FMA4 kernels. (Thank Werner Saar)
+- **AMD Bulldozer**: x86-64 ?GEMM FMA4 kernels. (Thanks to Werner Saar)
 - **AMD PILEDRIVER**: Uses Bulldozer codes with some optimizations.
 - **AMD STEAMROLLER**: Uses Bulldozer codes with some optimizations.
 
-#### MIPS64:
+#### MIPS64
+
 - **ICT Loongson 3A**: Optimized Level-3 BLAS and the part of Level-1,2.
 - **ICT Loongson 3B**: Experimental
 
-#### ARM:
-- **ARMV6**: Optimized BLAS for vfpv2 and vfpv3-d16 ( e.g. BCM2835, Cortex M0+ )
-- **ARMV7**: Optimized BLAS for vfpv3-d32 ( e.g. Cortex A8, A9 and A15 )
+#### ARM
 
-#### ARM64:
-- **ARMV8**: Experimental
+- **ARMv6**: Optimized BLAS for vfpv2 and vfpv3-d16 (e.g. BCM2835, Cortex M0+)
+- **ARMv7**: Optimized BLAS for vfpv3-d32 (e.g. Cortex A8, A9 and A15)
+
+#### ARM64
+
+- **ARMv8**: Experimental
 - **ARM Cortex-A57**: Experimental
 
 #### PPC/PPC64
-- **POWER8**: Optmized Level-3 BLAS and some Level-1, only with USE_OPENMP=1
 
-#### IBM zEnterprise System:
+- **POWER8**: Optmized Level-3 BLAS and some Level-1, only with `USE_OPENMP=1`
+
+#### IBM zEnterprise System
+
 - **Z13**: Optimized Level-3 BLAS and Level-1,2 (double precision)
-     
 
-### Support OS:
+### Supported OS
+
 - **GNU/Linux**
-- **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.
-- **Android**: Supported by community. Please read <https://github.com/xianyi/OpenBLAS/wiki/How-to-build-OpenBLAS-for-Android>.
+- **MinGW or Visual Studio (CMake)/Windows**: Please read <https://github.com/xianyi/OpenBLAS/wiki/How-to-use-OpenBLAS-in-Microsoft-Visual-Studio>.
+- **Darwin/macOS**: Experimental. Although GotoBLAS2 supports Darwin, we are not macOS experts.
+- **FreeBSD**: Supported by the community. We don't actively test the library on this OS.
+- **OpenBSD**: Supported by the community. We don't actively test the library on this OS.
+- **DragonFly BSD**: Supported by the community. We don't actively test the library on this OS.
+- **Android**: Supported by the community. Please read <https://github.com/xianyi/OpenBLAS/wiki/How-to-build-OpenBLAS-for-Android>.
 
-## Usages
-Link with libopenblas.a or -lopenblas for shared library.
+## Usage
 
-### Set the number of threads with environment variables.
+Statically link with `libopenblas.a` or dynamically link with `-lopenblas` if OpenBLAS was
+compiled as a shared library.
 
-Examples:
+### Setting the number of threads using environment variables
 
-    export OPENBLAS_NUM_THREADS=4
+Environment variables are used to specify a maximum number of threads.
+For example,
 
- or
+```sh
+export OPENBLAS_NUM_THREADS=4
+export GOTO_NUM_THREADS=4
+export OMP_NUM_THREADS=4
+```
 
-    export GOTO_NUM_THREADS=4
+The priorities are `OPENBLAS_NUM_THREADS` > `GOTO_NUM_THREADS` > `OMP_NUM_THREADS`.
 
- or
+If you compile this library with `USE_OPENMP=1`, you should set the `OMP_NUM_THREADS`
+environment variable; OpenBLAS ignores `OPENBLAS_NUM_THREADS` and `GOTO_NUM_THREADS` when
+compiled with `USE_OPENMP=1`.
 
-    export OMP_NUM_THREADS=4
+### Setting the number of threads at runtime
 
-The priorities are OPENBLAS_NUM_THREADS > GOTO_NUM_THREADS > OMP_NUM_THREADS.
+We provide the following functions to control the number of threads at runtime:
 
-If you compile this lib with USE_OPENMP=1, you should set OMP_NUM_THREADS environment variable. OpenBLAS ignores OPENBLAS_NUM_THREADS and GOTO_NUM_THREADS with USE_OPENMP=1.
+```c
+void goto_set_num_threads(int num_threads);
+void openblas_set_num_threads(int num_threads);
+```
 
-### Set the number of threads on runtime.
+If you compile this library with `USE_OPENMP=1`, you should use the above functions too.
 
-We provided the below functions to control the number of threads on runtime.
+## Reporting bugs
 
-    void goto_set_num_threads(int num_threads);
-
-    void openblas_set_num_threads(int num_threads);
-
-If you compile this lib with USE_OPENMP=1, you should use the above functions, too.
-
-## Report Bugs
-Please add a issue in https://github.com/xianyi/OpenBLAS/issues
+Please submit an issue in https://github.com/xianyi/OpenBLAS/issues.
 
 ## Contact
+
 * OpenBLAS users mailing list: https://groups.google.com/forum/#!forum/openblas-users
 * OpenBLAS developers mailing list: https://groups.google.com/forum/#!forum/openblas-dev
 
-## ChangeLog
-Please see Changelog.txt to obtain the differences between GotoBLAS2 1.13 BSD version.
+## Change log
+
+Please see Changelog.txt to view the differences between OpenBLAS and GotoBLAS2 1.13 BSD version.
 
 ## Troubleshooting
-* Please read [Faq](https://github.com/xianyi/OpenBLAS/wiki/Faq) at first.
-* Please use gcc version 4.6 and above to compile Sandy Bridge AVX kernels on Linux/MingW/BSD.
-* Please use Clang version 3.1 and above to compile the library on Sandy Bridge microarchitecture. The Clang 3.0 will generate the wrong AVX binary code.
-* The number of CPUs/Cores should less than or equal to 256. On Linux x86_64(amd64), there is experimental support for up to 1024 CPUs/Cores and 128 numa nodes if you build the library with BIGNUMA=1.
-* OpenBLAS does not set processor affinity by default. On Linux, you can enable processor affinity by commenting the line NO_AFFINITY=1 in Makefile.rule. But this may cause [the conflict with R parallel](https://stat.ethz.ch/pipermail/r-sig-hpc/2012-April/001348.html).
-* On Loongson 3A. make test would be failed because of pthread_create error. The error code is EAGAIN. However, it will be OK when you run the same testcase on shell.
+
+* Please read the [FAQ](https://github.com/xianyi/OpenBLAS/wiki/Faq) first.
+* Please use GCC version 4.6 and above to compile Sandy Bridge AVX kernels on Linux/MinGW/BSD.
+* Please use Clang version 3.1 and above to compile the library on Sandy Bridge microarchitecture.
+  Clang 3.0 will generate the wrong AVX binary code.
+* The number of CPUs/cores should less than or equal to 256. On Linux `x86_64` (`amd64`),
+  there is experimental support for up to 1024 CPUs/cores and 128 numa nodes if you build
+  the library with `BIGNUMA=1`.
+* OpenBLAS does not set processor affinity by default.
+  On Linux, you can enable processor affinity by commenting out the line `NO_AFFINITY=1` in
+  Makefile.rule. However, note that this may cause
+  [a conflict with R parallel](https://stat.ethz.ch/pipermail/r-sig-hpc/2012-April/001348.html).
+* On Loongson 3A, `make test` may fail with a `pthread_create` error (`EAGAIN`).
+  However, it will be okay when you run the same test case on the shell.
 
 ## Contributing
-1. [Check for open issues](https://github.com/xianyi/OpenBLAS/issues) or open a fresh issue to start a discussion around a feature idea or a bug.
-1. Fork the [OpenBLAS](https://github.com/xianyi/OpenBLAS) repository to start making your changes.
-1. Write a test which shows that the bug was fixed or that the feature works as expected.
-1. Send a pull request. Make sure to add yourself to `CONTRIBUTORS.md`.
+
+1. [Check for open issues](https://github.com/xianyi/OpenBLAS/issues) or open a fresh issue
+   to start a discussion around a feature idea or a bug.
+2. Fork the [OpenBLAS](https://github.com/xianyi/OpenBLAS) repository to start making your changes.
+3. Write a test which shows that the bug was fixed or that the feature works as expected.
+4. Send a pull request. Make sure to add yourself to `CONTRIBUTORS.md`.
 
 ## Donation
+
 Please read [this wiki page](https://github.com/xianyi/OpenBLAS/wiki/Donation).

USAGE.md

@@ -14,6 +14,20 @@ Please build OpenBLAS with larger `NUM_THREADS`. For example, `make
 NUM_THREADS=32` or `make NUM_THREADS=64`.  In `Makefile.system`, we will set
 `MAX_CPU_NUMBER=NUM_THREADS`.
 
+Despite its name, and due to the use of memory buffers in functions like SGEMM,
+the setting of NUM_THREADS can be relevant even for a single-threaded build 
+of OpenBLAS, if such functions get called by multiple threads of a program
+that uses OpenBLAS. In some cases, the affected code may simply crash or throw 
+a segmentation fault without displaying the above warning first.
+
+Note that the number of threads used at runtime can be altered to differ from the
+value NUM_THREADS was set to at build time. At runtime, the actual number of
+threads can be set anywhere from 1 to the build's NUM_THREADS (note however,
+that this does not change the number of memory buffers that will be allocated,
+which is set at build time). The number of threads for a process can be set by
+using the mechanisms described below.
+
+
 #### How can I use OpenBLAS in multi-threaded applications?
 
 If your application is already multi-threaded, it will conflict with OpenBLAS

benchmark/gemm.c

@@ -237,7 +237,7 @@ int main(int argc, char *argv[]){
     timeg = time1/loops;
     fprintf(stderr,
 	    " %10.2f MFlops %10.6f sec\n",
-	    COMPSIZE * COMPSIZE * 2. * (double)m * (double)m * (double)n / timeg * 1.e-6, time1);
+	    COMPSIZE * COMPSIZE * 2. * (double)k * (double)m * (double)n / timeg * 1.e-6, time1);
     
   }
 

c_check

@@ -54,6 +54,8 @@ $compiler = GCC       if ($compiler eq "");
 $os = Linux           if ($data =~ /OS_LINUX/);
 $os = FreeBSD         if ($data =~ /OS_FREEBSD/);
 $os = NetBSD          if ($data =~ /OS_NETBSD/);
+$os = OpenBSD         if ($data =~ /OS_OPENBSD/);
+$os = DragonFly       if ($data =~ /OS_DRAGONFLY/);
 $os = Darwin          if ($data =~ /OS_DARWIN/);
 $os = SunOS           if ($data =~ /OS_SUNOS/);
 $os = AIX             if ($data =~ /OS_AIX/);

cmake/system.cmake

@@ -97,7 +97,7 @@ if (NOT CMAKE_CROSSCOMPILING)
 endif()
 
 if (NOT DEFINED NUM_THREADS)
-  if (NOT NUM_CORES EQUAL 0)
+  if (DEFINED NUM_CORES AND NOT NUM_CORES EQUAL 0)
     # HT?
     set(NUM_THREADS ${NUM_CORES})
   else ()

common.h

@@ -93,7 +93,7 @@ extern "C" {
 #include <sched.h>
 #endif
 
-#if defined(OS_DARWIN) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(OS_ANDROID)
+#if defined(OS_DARWIN) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_ANDROID)
 #include <sched.h>
 #endif
 

common_x86.h

@@ -327,7 +327,7 @@ REALNAME:
 #endif
 #endif
 
-#if defined(OS_LINUX) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(__ELF__)
+#if defined(OS_LINUX) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(OS_OPENBSD) || defined(__ELF__)
 #define PROLOGUE \
 	.text; \
 	.align 16; \

common_x86_64.h

@@ -403,7 +403,7 @@ REALNAME:
 #define EPILOGUE .end
 #endif
 
-#if defined(OS_LINUX) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(__ELF__) || defined(C_PGI)
+#if defined(OS_LINUX) || defined(OS_FREEBSD) || defined(OS_NETBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(__ELF__) || defined(C_PGI)
 #define PROLOGUE \
 	.text; \
 	.align 512; \

cpuid_arm64.c

@@ -121,7 +121,7 @@ int detect(void)
 			return CPU_VULCAN;
 		else if (strstr(cpu_part, "0x0a1") && strstr(cpu_implementer, "0x43"))
 			return CPU_THUNDERX;
-		else if (strstr(cpu_part, "0xFFF") && strstr(cpu_implementer, "0x43")) /* TODO */
+		else if (strstr(cpu_part, "0x0af") && strstr(cpu_implementer, "0x43"))
 			return CPU_THUNDERX2T99;
 	}
 

ctest.c

@@ -60,6 +60,14 @@ OS_FREEBSD
 OS_NETBSD
 #endif
 
+#if defined(__OpenBSD__)
+OS_OPENBSD
+#endif
+
+#if defined(__DragonFly__)
+OS_DRAGONFLY
+#endif
+
 #if defined(__sun)
 OS_SUNOS
 #endif

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) || defined(OS_SUNOS) || defined(OS_FREEBSD)
+#if defined(OS_LINUX) || defined(OS_NETBSD) || defined(OS_DARWIN) || defined(OS_ANDROID) || defined(OS_SUNOS) || defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY)
 #include <dlfcn.h>
 #include <signal.h>
 #include <sys/resource.h>

driver/others/memory.c

@@ -108,7 +108,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include <sys/resource.h>
 #endif
 
-#if defined(OS_FREEBSD) || defined(OS_DARWIN)
+#if defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_DARWIN)
 #include <sys/sysctl.h>
 #include <sys/resource.h>
 #endif
@@ -147,9 +147,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #elif (defined(OS_DARWIN) || defined(OS_SUNOS)) && defined(C_GCC)
 #define CONSTRUCTOR	__attribute__ ((constructor))
 #define DESTRUCTOR	__attribute__ ((destructor))
-#else
+#elif __GNUC__ && INIT_PRIORITY && ((GCC_VERSION >= 40300) || (CLANG_VERSION >= 20900))
 #define CONSTRUCTOR	__attribute__ ((constructor(101)))
 #define DESTRUCTOR	__attribute__ ((destructor(101)))
+#else
+#define CONSTRUCTOR	__attribute__ ((constructor))
+#define DESTRUCTOR	__attribute__ ((destructor))
 #endif
 
 #ifdef DYNAMIC_ARCH
@@ -209,7 +212,8 @@ int ret;
   size = CPU_ALLOC_SIZE(nums);
   ret = sched_getaffinity(0,size,cpusetp);
   if (ret!=0) return nums;
-  nums = CPU_COUNT_S(size,cpusetp);
+  ret = CPU_COUNT_S(size,cpusetp);
+  if (ret > 0 && ret < nums) nums = ret;	
   CPU_FREE(cpusetp);
   return nums;
  #endif
@@ -246,7 +250,7 @@ int get_num_procs(void) {
 
 #endif
 
-#if defined(OS_FREEBSD)
+#if defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY)
 
 int get_num_procs(void) {
 
@@ -336,7 +340,7 @@ extern int openblas_goto_num_threads_env();
 extern int openblas_omp_num_threads_env();
 
 int blas_get_cpu_number(void){
-#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_DARWIN) || defined(OS_ANDROID)
+#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_DARWIN) || defined(OS_ANDROID)
   int max_num;
 #endif
   int blas_goto_num   = 0;
@@ -344,7 +348,7 @@ int blas_get_cpu_number(void){
 
   if (blas_num_threads) return blas_num_threads;
 
-#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_DARWIN) || defined(OS_ANDROID)
+#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_DARWIN) || defined(OS_ANDROID)
   max_num = get_num_procs();
 #endif
 
@@ -368,7 +372,7 @@ int blas_get_cpu_number(void){
   else if (blas_omp_num > 0) blas_num_threads = blas_omp_num;
   else blas_num_threads = MAX_CPU_NUMBER;
 
-#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_DARWIN) || defined(OS_ANDROID)
+#if defined(OS_LINUX) || defined(OS_WINDOWS) || defined(OS_FREEBSD) || defined(OS_OPENBSD) || defined(OS_DRAGONFLY) || defined(OS_DARWIN) || defined(OS_ANDROID)
   if (blas_num_threads > max_num) blas_num_threads = max_num;
 #endif
 
@@ -455,11 +459,15 @@ static void *alloc_mmap(void *address){
   }
 
   if (map_address != (void *)-1) {
+#if defined(SMP) && !defined(USE_OPENMP)
     LOCK_COMMAND(&alloc_lock);
+#endif    
     release_info[release_pos].address = map_address;
     release_info[release_pos].func    = alloc_mmap_free;
     release_pos ++;
+#if defined(SMP) && !defined(USE_OPENMP)
     UNLOCK_COMMAND(&alloc_lock);
+#endif    
   }
 
 #ifdef OS_LINUX
@@ -601,14 +609,18 @@ static void *alloc_mmap(void *address){
 #if defined(OS_LINUX) && !defined(NO_WARMUP)
   }
 #endif
-  LOCK_COMMAND(&alloc_lock);
 
   if (map_address != (void *)-1) {
+#if defined(SMP) && !defined(USE_OPENMP)
+    LOCK_COMMAND(&alloc_lock);
+#endif
     release_info[release_pos].address = map_address;
     release_info[release_pos].func    = alloc_mmap_free;
     release_pos ++;
+#if defined(SMP) && !defined(USE_OPENMP)
+    UNLOCK_COMMAND(&alloc_lock);
+#endif
   }
-  UNLOCK_COMMAND(&alloc_lock);
 
   return map_address;
 }
@@ -1007,6 +1019,11 @@ void *blas_memory_alloc(int procpos){
     NULL,
   };
   void *(**func)(void *address);
+
+#if defined(USE_OPENMP)
+  if (!memory_initialized) {
+#endif
+
   LOCK_COMMAND(&alloc_lock);
 
   if (!memory_initialized) {
@@ -1042,6 +1059,9 @@ void *blas_memory_alloc(int procpos){
 
   }
   UNLOCK_COMMAND(&alloc_lock);
+#if defined(USE_OPENMP)
+  }
+#endif
 
 #ifdef DEBUG
   printf("Alloc Start ...\n");
@@ -1056,13 +1076,17 @@ void *blas_memory_alloc(int procpos){
 
   do {
     if (!memory[position].used && (memory[position].pos == mypos)) {
+#if defined(SMP) && !defined(USE_OPENMP)
       LOCK_COMMAND(&alloc_lock);
-/*      blas_lock(&memory[position].lock);*/
-
+#else      
+      blas_lock(&memory[position].lock);
+#endif
       if (!memory[position].used) goto allocation;
-
+#if defined(SMP) && !defined(USE_OPENMP)
       UNLOCK_COMMAND(&alloc_lock);
-/*      blas_unlock(&memory[position].lock);*/
+#else
+      blas_unlock(&memory[position].lock);
+#endif      
     }
 
     position ++;
@@ -1075,15 +1099,19 @@ void *blas_memory_alloc(int procpos){
   position = 0;
 
   do {
-/*    if (!memory[position].used) { */
+#if defined(SMP) && !defined(USE_OPENMP)
       LOCK_COMMAND(&alloc_lock);
-/*      blas_lock(&memory[position].lock);*/
-
+#else
+    if (!memory[position].used) { 
+      blas_lock(&memory[position].lock);
+#endif
       if (!memory[position].used) goto allocation;
-      
+#if defined(SMP) && !defined(USE_OPENMP)
       UNLOCK_COMMAND(&alloc_lock);
-/*      blas_unlock(&memory[position].lock);*/
-/*    } */
+#else      
+      blas_unlock(&memory[position].lock);
+      }
+#endif
 
     position ++;
 
@@ -1098,9 +1126,11 @@ void *blas_memory_alloc(int procpos){
 #endif
 
   memory[position].used = 1;
-
+#if defined(SMP) && !defined(USE_OPENMP)
   UNLOCK_COMMAND(&alloc_lock);
-/*  blas_unlock(&memory[position].lock);*/
+#else
+  blas_unlock(&memory[position].lock);
+#endif
 
   if (!memory[position].addr) {
     do {
@@ -1146,9 +1176,13 @@ void *blas_memory_alloc(int procpos){
 
     } while ((BLASLONG)map_address == -1);
 
+#if defined(SMP) && !defined(USE_OPENMP)
     LOCK_COMMAND(&alloc_lock);
+#endif    
     memory[position].addr = map_address;
+#if defined(SMP) && !defined(USE_OPENMP)
     UNLOCK_COMMAND(&alloc_lock);
+#endif
 
 #ifdef DEBUG
     printf("  Mapping Succeeded. %p(%d)\n", (void *)memory[position].addr, position);
@@ -1202,8 +1236,9 @@ void blas_memory_free(void *free_area){
 #endif
 
   position = 0;
+#if defined(SMP) && !defined(USE_OPENMP)
   LOCK_COMMAND(&alloc_lock);
-
+#endif
   while ((position < NUM_BUFFERS) && (memory[position].addr != free_area))
     position++;
 
@@ -1217,7 +1252,9 @@ void blas_memory_free(void *free_area){
   WMB;
 
   memory[position].used = 0;
+#if defined(SMP) && !defined(USE_OPENMP)
   UNLOCK_COMMAND(&alloc_lock);
+#endif
 
 #ifdef DEBUG
   printf("Unmap Succeeded.\n\n");
@@ -1232,8 +1269,9 @@ void blas_memory_free(void *free_area){
   for (position = 0; position < NUM_BUFFERS; position++)
     printf("%4ld  %p : %d\n", position, memory[position].addr, memory[position].used);
 #endif
+#if defined(SMP) && !defined(USE_OPENMP)
   UNLOCK_COMMAND(&alloc_lock);
-
+#endif
   return;
 }
 

exports/Makefile

@@ -156,7 +156,7 @@ endif
 endif
 
 #http://stackoverflow.com/questions/7656425/makefile-ifeq-logical-or
-ifeq ($(OSNAME), $(filter $(OSNAME),FreeBSD NetBSD))
+ifeq ($(OSNAME), $(filter $(OSNAME),FreeBSD OpenBSD NetBSD DragonFly))
 
 so : ../$(LIBSONAME)
 

getarch.c

@@ -82,7 +82,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #ifdef OS_WINDOWS
 #include <windows.h>
 #endif
-#if defined(__FreeBSD__) || defined(__APPLE__)
+#if defined(__FreeBSD__) || defined(__OpenBSD__) || defined(__DragonFly__) || defined(__APPLE__)
 #include <sys/types.h>
 #include <sys/sysctl.h>
 #endif
@@ -1074,7 +1074,7 @@ static int get_num_cores(void) {
 
 #ifdef OS_WINDOWS
   SYSTEM_INFO sysinfo;
-#elif defined(__FreeBSD__) || defined(__APPLE__)
+#elif defined(__FreeBSD__) || defined(__OpenBSD__) || defined(__DragonFly__) || defined(__APPLE__)
   int m[2], count;
   size_t len;
 #endif
@@ -1088,7 +1088,7 @@ static int get_num_cores(void) {
   GetSystemInfo(&sysinfo);
   return sysinfo.dwNumberOfProcessors;
 
-#elif defined(__FreeBSD__) || defined(__APPLE__)
+#elif defined(__FreeBSD__) || defined(__OpenBSD__) || defined(__DragonFly__) || defined(__APPLE__)
   m[0] = CTL_HW;
   m[1] = HW_NCPU;
   len = sizeof(int);

interface/rotmg.c

@@ -64,6 +64,13 @@ void CNAME(FLOAT *dd1, FLOAT *dd2, FLOAT *dx1, FLOAT dy1, FLOAT *dparam){
 
 	FLOAT du, dp1, dp2, dq2, dq1, dh11=ZERO, dh21=ZERO, dh12=ZERO, dh22=ZERO, dflag=-ONE, dtemp;
 
+	if (*dd2 == ZERO || dy1 == ZERO)
+	{
+		dflag = -TWO;
+		dparam[0] = dflag;
+		return;
+	}
+		
 	if(*dd1 < ZERO)
 	{
 		dflag = -ONE;
@@ -76,6 +83,16 @@ void CNAME(FLOAT *dd1, FLOAT *dd2, FLOAT *dx1, FLOAT dy1, FLOAT *dparam){
 		*dd2  = ZERO;
 		*dx1  = ZERO;
 	}
+	else if ((*dd1 == ZERO || *dx1 == ZERO) && *dd2 > ZERO)
+	{
+		dflag = ONE;
+		dh12 = 1;
+		dh21 = -1;
+		*dx1 = dy1;
+		dtemp = *dd1;
+		*dd1 = *dd2;
+		*dd2 = dtemp;
+	} 
 	else
 	{
 		dp2 = *dd2 * dy1;
@@ -90,6 +107,9 @@ void CNAME(FLOAT *dd1, FLOAT *dd2, FLOAT *dx1, FLOAT dy1, FLOAT *dparam){
 		dq1 =  dp1 * *dx1;
 		if(ABS(dq1) > ABS(dq2))
 		{
+			dflag = ZERO;
+			dh11  =  ONE;
+			dh22  =  ONE;
 			dh21 = -  dy1 / *dx1;
 			dh12 =    dp2 /  dp1;
 
@@ -100,8 +120,19 @@ void CNAME(FLOAT *dd1, FLOAT *dd2, FLOAT *dx1, FLOAT dy1, FLOAT *dparam){
 				*dd1  = *dd1 / du;
 				*dd2  = *dd2 / du;
 				*dx1  = *dx1 * du;
+			} else {
+				dflag = -ONE;
 
+				dh11  = ZERO;
+				dh12  = ZERO;
+				dh21  = ZERO;
+				dh22  = ZERO;
+
+				*dd1  = ZERO;
+				*dd2  = ZERO;
+				*dx1  = ZERO;
 			}
+			
 		}
 		else
 		{
@@ -120,7 +151,9 @@ void CNAME(FLOAT *dd1, FLOAT *dd2, FLOAT *dx1, FLOAT dy1, FLOAT *dparam){
 			}
 			else
 			{
-				dflag = ONE;
+				dflag =  ONE;
+				dh21  = -ONE;
+				dh12  =  ONE;
 
 				dh11  =  dp1 /  dp2;
 				dh22  = *dx1 /  dy1;
@@ -134,76 +167,33 @@ void CNAME(FLOAT *dd1, FLOAT *dd2, FLOAT *dx1, FLOAT dy1, FLOAT *dparam){
 		}
 
 
-		if(*dd1 != ZERO)
+		while ( *dd1 <= RGAMSQ && *dd1 != ZERO)
 		{
-			if( (*dd1 <= RGAMSQ) || (*dd1 >= GAMSQ) )
-			{
-				if(dflag == ZERO)
-				{
-					dh11  =  ONE;
-					dh22  =  ONE;
-					dflag = -ONE;
-				}
-				else
-				{
-						dh21  = -ONE;
-						dh12  =  ONE;
-						dflag = -ONE;
-				}
-				if( *dd1 <= RGAMSQ )
-				{
-					while (ABS(*dd1) <= RGAMSQ) {
-						*dd1  = *dd1 * (GAM * GAM);
-						*dx1  = *dx1 / GAM;
-						dh11  = dh11 / GAM;
-						dh12  = dh12 / GAM;
-					}
-				}
-				else
-				{
-					while (ABS(*dd1) >= GAMSQ) {
-						*dd1  = *dd1 / (GAM * GAM);
-						*dx1  = *dx1 * GAM;
-						dh11  = dh11 * GAM;
-						dh12  = dh12 * GAM;
-					}
-				}
-			}
+			dflag = -ONE;
+			*dd1  = *dd1 * (GAM * GAM);
+			*dx1  = *dx1 / GAM;
+			dh11  = dh11 / GAM;
+			dh12  = dh12 / GAM;
+		}
+		while (ABS(*dd1) > GAMSQ) {
+			dflag = -ONE;
+			*dd1  = *dd1 / (GAM * GAM);
+			*dx1  = *dx1 * GAM;
+			dh11  = dh11 * GAM;
+			dh12  = dh12 * GAM;
 		}
 
-		if(*dd2 != ZERO)
-		{
-			if( (ABS(*dd2) <= RGAMSQ) || (ABS(*dd2) >= GAMSQ) )
-			{
-				if(dflag == ZERO)
-				{
-					dh11  =  ONE;
-					dh22  =  ONE;
-					dflag = -ONE;
-				}
-				else
-				{
-						dh21  = -ONE;
-						dh12  =  ONE;
-						dflag = -ONE;
-				}
-				if( ABS(*dd2) <= RGAMSQ )
-				{
-					while (ABS(*dd2) <= RGAMSQ) {
-						*dd2  = *dd2 * (GAM * GAM);
-						dh21  = dh21 / GAM;
-						dh22  = dh22 / GAM;
-					}
-				}
-				else
-				{
-					while (ABS(*dd2) >= GAMSQ) {
-						*dd2  = *dd2 / (GAM * GAM);
-						dh21  = dh21 * GAM;
-						dh22  = dh22 * GAM;
-					}
-				}
-			}
+		while (ABS(*dd2) <= RGAMSQ && *dd2 != ZERO) {
+			dflag = -ONE;
+			*dd2  = *dd2 * (GAM * GAM);
+			dh21  = dh21 / GAM;
+			dh22  = dh22 / GAM;
+		}
+		while (ABS(*dd2) > GAMSQ) {
+			dflag = -ONE;
+			*dd2  = *dd2 / (GAM * GAM);
+			dh21  = dh21 * GAM;
+			dh22  = dh22 * GAM;
 		}
 
 	}

kernel/Makefile.L3

@@ -29,10 +29,8 @@ USE_TRMM = 1
 endif
 
 ifeq ($(CORE), HASWELL)
-ifeq ($(ARCH), x86_64)
 USE_TRMM = 1
 endif
-endif
 
 ifeq ($(CORE), ZEN)
 USE_TRMM = 1

kernel/arm/KERNEL.ARMV5

@@ -49,6 +49,7 @@ SDOTKERNEL   = ../arm/dot.c
 DDOTKERNEL   = ../arm/dot.c
 CDOTKERNEL   = ../arm/zdot.c
 ZDOTKERNEL   = ../arm/zdot.c
+DSDOTKERNEL  = ../generic/dot.c
 
 SNRM2KERNEL  = ../arm/nrm2.c
 DNRM2KERNEL  = ../arm/nrm2.c

kernel/arm64/KERNEL.ARMV8

@@ -49,6 +49,7 @@ SDOTKERNEL   = dot.S
 DDOTKERNEL   = dot.S
 CDOTKERNEL   = zdot.S
 ZDOTKERNEL   = zdot.S
+DSDOTKERNEL  = dot.S
 
 SNRM2KERNEL  = nrm2.S
 DNRM2KERNEL  = nrm2.S

kernel/arm64/KERNEL.CORTEXA57

@@ -29,6 +29,7 @@ SDOTKERNEL   = dot.S
 DDOTKERNEL   = dot.S
 CDOTKERNEL   = zdot.S
 ZDOTKERNEL   = zdot.S
+DSDOTKERNEL  = dot.S
 
 SNRM2KERNEL  = nrm2.S
 DNRM2KERNEL  = nrm2.S

kernel/arm64/dot.S

@@ -74,8 +74,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #if !defined(DSDOT)
 	fmadd	DOTF, TMPX, TMPY, DOTF
 #else // DSDOT
-	fmul	TMPX, TMPX, TMPY
+	fcvt	d3, TMPY
 	fcvt	d2, TMPX
+	fmul	d2, d2, d3
 	fadd	DOTF, DOTF, d2
 #endif
 .endm
@@ -87,12 +88,14 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #if !defined(DSDOT)
 	fmla	v0.4s, v2.4s, v3.4s
 #else
-	fmul	v2.4s, v2.4s, v3.4s
-	ext	v3.16b, v2.16b, v2.16b, #8
-	fcvtl	v2.2d, v2.2s
+	fcvtl2	v5.2d, v3.4s
+	fcvtl2	v4.2d, v2.4s
 	fcvtl	v3.2d, v3.2s
+	fcvtl	v2.2d, v2.2s
+	fmul	v4.2d, v4.2d, v5.2d
+	fmul	v2.2d, v2.2d, v3.2d
+	fadd	v2.2d, v2.2d, v4.2d
 	fadd	v0.2d, v0.2d, v2.2d
-	fadd	v0.2d, v0.2d, v3.2d
 #endif
 #else //DOUBLE
 	ld1	{v2.2d, v3.2d}, [X], #32
@@ -136,8 +139,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #if !defined(DSDOT)
 	fmadd	DOTF, TMPX, TMPY, DOTF
 #else // DSDOT
-	fmul	TMPX, TMPX, TMPY
+	fcvt	d3, TMPY
 	fcvt	d2, TMPX
+	fmul	d2, d2, d3
 	fadd	DOTF, DOTF, d2
 #endif
 .endm

kernel/mips/dot.c

@@ -41,8 +41,11 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 
 	while(i < n)
 	{
-
-		dot += y[iy] * x[ix] ;
+#if defined(DSDOT)
+		dot += (double)(y[iy] * (double)x[ix] ;
+#else
+		dot += y[iy] * x[ix];
+#endif				
 		ix  += inc_x ;
 		iy  += inc_y ;
 		i++ ;

kernel/power/KERNEL.POWER8

@@ -90,14 +90,14 @@ ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 #DMINKERNEL   = ../arm/min.c
 #
 #ISAMAXKERNEL = ../arm/iamax.c
-#IDAMAXKERNEL = ../arm/iamax.c
+IDAMAXKERNEL = idamax.c
 #ICAMAXKERNEL = ../arm/izamax.c
-#IZAMAXKERNEL = ../arm/izamax.c
+IZAMAXKERNEL =  izamax.c
 #
 #ISAMINKERNEL = ../arm/iamin.c
-#IDAMINKERNEL = ../arm/iamin.c
+IDAMINKERNEL =  idamin.c
 #ICAMINKERNEL = ../arm/izamin.c
-#IZAMINKERNEL = ../arm/izamin.c
+IZAMINKERNEL = izamin.c
 #
 #ISMAXKERNEL  = ../arm/imax.c
 #IDMAXKERNEL  = ../arm/imax.c
@@ -134,7 +134,7 @@ ZNRM2KERNEL  = ../arm/znrm2.c
 SROTKERNEL   = srot.c
 DROTKERNEL   = drot.c
 #CROTKERNEL   = ../arm/zrot.c
-#ZROTKERNEL   = ../arm/zrot.c
+ZROTKERNEL   = zrot.c
 #
 SSCALKERNEL  = sscal.c
 DSCALKERNEL  = dscal.c
@@ -150,12 +150,12 @@ ZSWAPKERNEL  = zswap.c
 #SGEMVNKERNEL = ../arm/gemv_n.c
 DGEMVNKERNEL = dgemv_n.c
 #CGEMVNKERNEL = ../arm/zgemv_n.c
-#ZGEMVNKERNEL = ../arm/zgemv_n.c
+ZGEMVNKERNEL = zgemv_n_4.c
 #
 #SGEMVTKERNEL = ../arm/gemv_t.c
-#DGEMVTKERNEL = ../arm/gemv_t.c
+DGEMVTKERNEL = dgemv_t.c
 #CGEMVTKERNEL = ../arm/zgemv_t.c
-#ZGEMVTKERNEL = zgemv_t_4.c
+ZGEMVTKERNEL = zgemv_t_4.c
 
 
 #SSYMV_U_KERNEL =  ../generic/symv_k.c

kernel/power/cgemm_tcopy_8_power8.S

@@ -110,7 +110,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "cgemm_tcopy_macros_8_power8.S"
 
-#define STACKSIZE 576
+#define STACKSIZE 144
 
 
 	PROLOGUE
@@ -119,49 +119,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	addi	SP, SP, -STACKSIZE
 	li	r0, 0
 
-        std     r31,  144(SP)
-        std     r30,  152(SP)
-        std     r29,  160(SP)
-        std     r28,  168(SP)
-        std     r27,  176(SP)
-        std     r26,  184(SP)
-        std     r25,  192(SP)
-        std     r24,  200(SP)
-        std     r23,  208(SP)
-        std     r22,  216(SP)
-        std     r21,  224(SP)
-        std     r20,  232(SP)
-        std     r19,  240(SP)
-        std     r18,  248(SP)
-        std     r17,  256(SP)
-        std     r16,  264(SP)
-        std     r15,  272(SP)
-        std     r14,  280(SP)
-        addi    r11, SP, 288
-        stvx    v20, r11, r0
-        addi    r11, r11, 16
-        stvx    v21, r11, r0
-        addi    r11, r11, 16
-        stvx    v22, r11, r0
-        addi    r11, r11, 16
-        stvx    v23, r11, r0
-        addi    r11, r11, 16
-        stvx    v24, r11, r0
-        addi    r11, r11, 16
-        stvx    v25, r11, r0
-        addi    r11, r11, 16
-        stvx    v26, r11, r0
-        addi    r11, r11, 16
-        stvx    v27, r11, r0
-        addi    r11, r11, 16
-        stvx    v28, r11, r0
-        addi    r11, r11, 16
-        stvx    v29, r11, r0
-        addi    r11, r11, 16
-        stvx    v30, r11, r0
-        addi    r11, r11, 16
-        stvx    v31, r11, r0
-        li      r11, 0
+        std r14,   0(SP) 
+        std r15,   8(SP) 
+        std r16,  16(SP) 
+        std r17,  24(SP) 
+        std r18,  32(SP) 
+        std r19,  40(SP) 
+        std r20,  48(SP) 
+        std r21,  56(SP) 
+        std r22,  64(SP) 
+        std r23,  72(SP) 
+        std r24,  80(SP) 
+        std r25,  88(SP) 
+        std r26,  96(SP) 
+        std r27, 104(SP) 
+        std r28, 112(SP) 
+        std r29, 120(SP) 
+        std r30, 128(SP) 
+        std r31, 136(SP) 
 
 	cmpwi	cr0, M, 0
 	ble-	L999
@@ -203,51 +178,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 L999:
 
-	li	r3, 0
-
-        ld      r31,  144(SP)
-        ld      r30,  152(SP)
-        ld      r29,  160(SP)
-        ld      r28,  168(SP)
-        ld      r27,  176(SP)
-        ld      r26,  184(SP)
-        ld      r25,  192(SP)
-        ld      r24,  200(SP)
-        ld      r23,  208(SP)
-        ld      r22,  216(SP)
-        ld      r21,  224(SP)
-        ld      r20,  232(SP)
-        ld      r19,  240(SP)
-        ld      r18,  248(SP)
-        ld      r17,  256(SP)
-        ld      r16,  264(SP)
-        ld      r15,  272(SP)
-        ld      r14,  280(SP)
-        addi    r11, SP, 288
-        lvx     v20, r11, r3
-        addi    r11, r11, 16
-        lvx     v21, r11, r3
-        addi    r11, r11, 16
-        lvx     v22, r11, r3
-        addi    r11, r11, 16
-        lvx     v23, r11, r3
-        addi    r11, r11, 16
-        lvx     v24, r11, r3
-        addi    r11, r11, 16
-        lvx     v25, r11, r3
-        addi    r11, r11, 16
-        lvx     v26, r11, r3
-        addi    r11, r11, 16
-        lvx     v27, r11, r3
-        addi    r11, r11, 16
-        lvx     v28, r11, r3
-        addi    r11, r11, 16
-        lvx     v29, r11, r3
-        addi    r11, r11, 16
-        lvx     v30, r11, r3
-        addi    r11, r11, 16
-        lvx     v31, r11, r3
-        li      r11, 0
+        ld r14,   0(SP) 
+        ld r15,   8(SP) 
+        ld r16,  16(SP) 
+        ld r17,  24(SP) 
+        ld r18,  32(SP) 
+        ld r19,  40(SP) 
+        ld r20,  48(SP) 
+        ld r21,  56(SP) 
+        ld r22,  64(SP) 
+        ld r23,  72(SP) 
+        ld r24,  80(SP) 
+        ld r25,  88(SP) 
+        ld r26,  96(SP) 
+        ld r27, 104(SP) 
+        ld r28, 112(SP) 
+        ld r29, 120(SP) 
+        ld r30, 128(SP) 
+        ld r31, 136(SP) 
 
 	addi	SP, SP, STACKSIZE
 	blr

kernel/power/dgemm_ncopy_4_power8.S

@@ -109,81 +109,36 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "dgemm_ncopy_macros_4_power8.S"
 
-#define STACKSIZE 384
-#define STACKSIZE 576
+#define STACKSIZE 144
+
 
 	PROLOGUE
 	PROFCODE
 
 	addi	SP, SP, -STACKSIZE
-//addi SP, SP, -208
+ 
 	li	r0, 0
 
-        stfd    f14,    0(SP)
-        stfd    f15,    8(SP)
-        stfd    f16,   16(SP)
-        stfd    f17,   24(SP)
-        stfd    f18,   32(SP)
-        stfd    f19,   40(SP)
-        stfd    f20,   48(SP)
-        stfd    f21,   56(SP)
-        stfd    f22,   64(SP)
-        stfd    f23,   72(SP)
-        stfd    f24,   80(SP)
-        stfd    f25,   88(SP)
-        stfd    f26,   96(SP)
-        stfd    f27,  104(SP)
-        stfd    f28,  112(SP)
-        stfd    f29,  120(SP)
-        stfd    f30,  128(SP)
-        stfd    f31,  136(SP)
+        std r14,   0(SP) 
+        std r15,   8(SP) 
+        std r16,  16(SP) 
+        std r17,  24(SP) 
+        std r18,  32(SP) 
+        std r19,  40(SP) 
+        std r20,  48(SP) 
+        std r21,  56(SP) 
+        std r22,  64(SP) 
+        std r23,  72(SP) 
+        std r24,  80(SP) 
+        std r25,  88(SP) 
+        std r26,  96(SP) 
+        std r27, 104(SP) 
+        std r28, 112(SP) 
+        std r29, 120(SP) 
+        std r30, 128(SP) 
+        std r31, 136(SP) 
 
 
-        std     r31,  144(SP)
-        std     r30,  152(SP)
-        std     r29,  160(SP)
-        std     r28,  168(SP)
-        std     r27,  176(SP)
-        std     r26,  184(SP)
-        std     r25,  192(SP)
-        std     r24,  200(SP)
-        std     r23,  208(SP)
-        std     r22,  216(SP)
-        std     r21,  224(SP)
-        std     r20,  232(SP)
-        std     r19,  240(SP)
-        std     r18,  248(SP)
-        std     r17,  256(SP)
-        std     r16,  264(SP)
-        std     r15,  272(SP)
-        std     r14,  280(SP)
-
-addi r11,SP,288
-        stvx    v20,  r11,r0
-addi r11,r11,16
-        stvx    v21,  r11,r0
-addi r11,r11,16
-        stvx    v22,  r11,r0
-addi r11,r11,16
-        stvx    v23,  r11,r0
-addi r11,r11,16
-        stvx    v24,  r11,r0
-addi r11,r11,16
-        stvx    v25,  r11,r0
-addi r11,r11,16
-        stvx    v26,  r11,r0
-addi r11,r11,16
-        stvx    v27,  r11,r0
-addi r11,r11,16
-        stvx    v28,  r11,r0
-addi r11,r11,16
-        stvx    v29,  r11,r0
-addi r11,r11,16
-        stvx    v30,  r11,r0
-addi r11,r11,16
-        stvx    v31,  r11,r0
-li r11,0
-
 	cmpwi	cr0, M, 0
 	ble-	L999
 	cmpwi	cr0, N, 0
@@ -191,10 +146,8 @@ li r11,0
 
 	slwi	LDA, LDA, BASE_SHIFT
 
-	//li	PREA,  384
-	//li	PREB,  384
-	li	PREA,  576
-	li	PREB,  576
+	li	PREA,  384
+	li	PREB,  384 
 
 
 	li	o8,	8
@@ -210,70 +163,24 @@ li r11,0
 
 L999:
 
-	li	r3, 0
-
-        lfd     f14,    0(SP)
-        lfd     f15,    8(SP)
-        lfd     f16,   16(SP)
-        lfd     f17,   24(SP)
-        lfd     f18,   32(SP)
-        lfd     f19,   40(SP)
-        lfd     f20,   48(SP)
-        lfd     f21,   56(SP)
-        lfd     f22,   64(SP)
-        lfd     f23,   72(SP)
-        lfd     f24,   80(SP)
-        lfd     f25,   88(SP)
-        lfd     f26,   96(SP)
-        lfd     f27,  104(SP)
-        lfd     f28,  112(SP)
-        lfd     f29,  120(SP)
-        lfd     f30,  128(SP)
-        lfd     f31,  136(SP)
-
-        ld      r31,  144(SP)
-        ld      r30,  152(SP)
-        ld      r29,  160(SP)
-        ld      r28,  168(SP)
-        ld      r27,  176(SP)
-        ld      r26,  184(SP)
-        ld      r25,  192(SP)
-        ld      r24,  200(SP)
-        ld      r23,  208(SP)
-        ld      r22,  216(SP)
-        ld      r21,  224(SP)
-        ld      r20,  232(SP)
-        ld      r19,  240(SP)
-        ld      r18,  248(SP)
-        ld      r17,  256(SP)
-        ld      r16,  264(SP)
-        ld      r15,  272(SP)
-        ld      r14,  280(SP)
-addi r11,SP,288
-        lvx    v20,  r11,r3
-addi r11,r11,16
-        lvx    v21,  r11,r3
-addi r11,r11,16
-        lvx    v22,  r11,r3
-addi r11,r11,16
-        lvx    v23,  r11,r3
-addi r11,r11,16
-        lvx    v24,  r11,r3
-addi r11,r11,16
-        lvx    v25,  r11,r3
-addi r11,r11,16
-        lvx    v26,  r11,r3
-addi r11,r11,16
-        lvx    v27,  r11,r3
-addi r11,r11,16
-        lvx    v28,  r11,r3
-addi r11,r11,16
-        lvx    v29,  r11,r3
-addi r11,r11,16
-        lvx    v30,  r11,r3
-addi r11,r11,16
-        lvx    v31,  r11,r3
-li r11,0
+        ld r14,   0(SP) 
+        ld r15,   8(SP) 
+        ld r16,  16(SP) 
+        ld r17,  24(SP) 
+        ld r18,  32(SP) 
+        ld r19,  40(SP) 
+        ld r20,  48(SP) 
+        ld r21,  56(SP) 
+        ld r22,  64(SP) 
+        ld r23,  72(SP) 
+        ld r24,  80(SP) 
+        ld r25,  88(SP) 
+        ld r26,  96(SP) 
+        ld r27, 104(SP) 
+        ld r28, 112(SP) 
+        ld r29, 120(SP) 
+        ld r30, 128(SP) 
+        ld r31, 136(SP) 
 
 	addi	SP, SP, STACKSIZE
 //addi SP, SP, 208

kernel/power/dgemm_ncopy_macros_4_power8.S

@@ -41,94 +41,98 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .macro COPY_4x16
 
 	lxvd2x		vs0,	o0,	A0
-	lxvd2x		vs8,	o0,	A1
-	lxvd2x		vs24,	o0,	A3
-	lxvd2x		vs16,	o0,	A2
+	lxvd2x		vs1,	o0,	A1
+	lxvd2x		vs2,	o0,	A2
+	lxvd2x		vs3,	o0,	A3
 
-	lxvd2x		vs1,	o16,	A0
-	lxvd2x		vs9,	o16,	A1
-	lxvd2x		vs17,	o16,	A2
-	lxvd2x		vs25,	o16,	A3
+	lxvd2x		vs4,	o16,	A0
+	lxvd2x		vs5,	o16,	A1
+	lxvd2x		vs6,	o16,	A2
+	lxvd2x		vs7,	o16,	A3
 
-	lxvd2x		vs2,	o32,	A0
-	lxvd2x		vs10,	o32,	A1
-	lxvd2x		vs18,	o32,	A2
-	lxvd2x		vs26,	o32,	A3
+	xxpermdi	vs32,	vs0,	vs1,	0
+	xxpermdi	vs33,	vs2,	vs3,	0
+	xxpermdi	vs34,	vs0,	vs1,	3
+	xxpermdi	vs35,	vs2,	vs3,	3
 
-	lxvd2x		vs3,	o48,	A0
-	lxvd2x		vs11,	o48,	A1
-	lxvd2x		vs19,	o48,	A2
-	lxvd2x		vs27,	o48,	A3
+	xxpermdi	vs36,	vs4,	vs5,	0
+	xxpermdi	vs37,	vs6,	vs7,	0
+	xxpermdi	vs38,	vs4,	vs5,	3
+	xxpermdi	vs39,	vs6,	vs7,	3	
 
-	lxvd2x		vs4,	o64,	A0
-	lxvd2x		vs12,	o64,	A1
-	lxvd2x		vs20,	o64,	A2
-	lxvd2x		vs28,	o64,	A3
+	lxvd2x		vs0,	o32,	A0
+	lxvd2x		vs1,	o32,	A1
+	lxvd2x		vs2,	o32,	A2
+	lxvd2x		vs3,	o32,	A3
 
-	lxvd2x		vs5,	o80,	A0
-	lxvd2x		vs13,	o80,	A1
-	lxvd2x		vs21,	o80,	A2
-	lxvd2x		vs29,	o80,	A3
-
-	lxvd2x		vs6,	o96,	A0
-	lxvd2x		vs14,	o96,	A1
-	lxvd2x		vs22,	o96,	A2
-	lxvd2x		vs30,	o96,	A3
-
-	lxvd2x		vs7,	o112,	A0
-	lxvd2x		vs15,	o112,	A1
-	lxvd2x		vs23,	o112,	A2
-	lxvd2x		vs31,	o112,	A3
+	lxvd2x		vs4,	o48,	A0
+	lxvd2x		vs5,	o48,	A1
+	lxvd2x		vs6,	o48,	A2
+	lxvd2x		vs7,	o48,	A3
 
 
-	xxpermdi	vs32,	vs0,	vs8,	0
-	xxpermdi	vs33,	vs16,	vs24,	0
-	xxpermdi	vs34,	vs0,	vs8,	3
-	xxpermdi	vs35,	vs16,	vs24,	3
+	xxpermdi	vs40,	vs0,	vs1,	0
+	xxpermdi	vs41,	vs2,	vs3,	0
+	xxpermdi	vs42,	vs0,	vs1,	3
+	xxpermdi	vs43,	vs2,	vs3,	3
 
-	xxpermdi	vs36,	vs1,	vs9,	0
-	xxpermdi	vs37,	vs17,	vs25,	0
-	xxpermdi	vs38,	vs1,	vs9,	3
-	xxpermdi	vs39,	vs17,	vs25,	3
+	xxpermdi	vs44,	vs4,	vs5,	0
+	xxpermdi	vs45,	vs6,	vs7,	0
+	xxpermdi	vs46,	vs4,	vs5,	3
+	xxpermdi	vs47,	vs6,	vs7,	3	
 
-	xxpermdi	vs40,	vs2,	vs10,	0
-	xxpermdi	vs41,	vs18,	vs26,	0
-	xxpermdi	vs42,	vs2,	vs10,	3
-	xxpermdi	vs43,	vs18,	vs26,	3
+	lxvd2x		vs0,	o64,	A0
+	lxvd2x		vs1,	o64,	A1
+	lxvd2x		vs2,	o64,	A2
+	lxvd2x		vs3,	o64,	A3
 
-	xxpermdi	vs44,	vs3,	vs11,	0
-	xxpermdi	vs45,	vs19,	vs27,	0
-	xxpermdi	vs46,	vs3,	vs11,	3
-	xxpermdi	vs47,	vs19,	vs27,	3
+	lxvd2x		vs4,	o80,	A0
+	lxvd2x		vs5,	o80,	A1
+	lxvd2x		vs6,	o80,	A2
+	lxvd2x		vs7,	o80,	A3
 
-	xxpermdi	vs48,	vs4,	vs12,	0
-	xxpermdi	vs49,	vs20,	vs28,	0
-	xxpermdi	vs50,	vs4,	vs12,	3
-	xxpermdi	vs51,	vs20,	vs28,	3
 
-	xxpermdi	vs52,	vs5,	vs13,	0
-	xxpermdi	vs53,	vs21,	vs29,	0
-	xxpermdi	vs54,	vs5,	vs13,	3
-	xxpermdi	vs55,	vs21,	vs29,	3
+	xxpermdi	vs48,	vs0,	vs1,	0
+	xxpermdi	vs49,	vs2,	vs3,	0
+	xxpermdi	vs50,	vs0,	vs1,	3
+	xxpermdi	vs51,	vs2,	vs3,	3
+
+	xxpermdi	vs8,	vs4,	vs5,	0
+	xxpermdi	vs9,	vs6,	vs7,	0
+	xxpermdi	vs10,	vs4,	vs5,	3
+	xxpermdi	vs11,	vs6,	vs7,	3	
+
+	lxvd2x		vs0,	o96,	A0
+	lxvd2x		vs1,	o96,	A1
+	lxvd2x		vs2,	o96,	A2
+	lxvd2x		vs3,	o96,	A3
+
+ 
+	lxvd2x		vs6,	o112,	A0
+	lxvd2x		vs7,	o112,	A1
+	lxvd2x		vs12,	o112,	A2
+	lxvd2x		vs13,	o112,	A3
+
+
+	xxpermdi	vs4,	vs0,	vs1,	0
+	xxpermdi	vs5,	vs2,	vs3,	0
+	xxpermdi	vs0,	vs0,	vs1,	3
+	xxpermdi	vs2,	vs2,	vs3,	3
+
 
 	addi		A0,	A0,	128
 	addi		A1,	A1,	128
 
-	xxpermdi	vs56,	vs6,	vs14,	0
-	xxpermdi	vs57,	vs22,	vs30,	0
-	xxpermdi	vs58,	vs6,	vs14,	3
-	xxpermdi	vs59,	vs22,	vs30,	3
+	xxpermdi	vs1,	vs6,	vs7,	0
+	xxpermdi	vs3,	vs12,	vs13,	0
+	xxpermdi	vs6,	vs6,	vs7,	3
+	xxpermdi	vs12,	vs12,	vs13,	3	
+
+	dcbt		BO,	PREB
 
 	addi		A3,	A3,	128
 	addi		A2,	A2,	128
 
-	xxpermdi	vs60,	vs7,	vs15,	0
-	xxpermdi	vs61,	vs23,	vs31,	0
-	xxpermdi	vs62,	vs7,	vs15,	3
-	xxpermdi	vs63,	vs23,	vs31,	3
-
-	dcbt		BO,	PREB
-
 	stxvd2x		vs32,	o0,	BO
 	stxvd2x		vs33,	o16,	BO
 	stxvd2x		vs34,	o32,	BO
@@ -157,22 +161,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	stxvd2x		vs49,	o16,	BO
 	stxvd2x		vs50,	o32,	BO
 	stxvd2x		vs51,	o48,	BO
-	stxvd2x		vs52,	o64,	BO
-	stxvd2x		vs53,	o80,	BO
-	stxvd2x		vs54,	o96,	BO
-	stxvd2x		vs55,	o112,	BO
+	stxvd2x		vs8,	o64,	BO
+	stxvd2x		vs9,	o80,	BO
+	stxvd2x		vs10,	o96,	BO
+	stxvd2x		vs11,	o112,	BO
 	addi		BO,	BO,	128
 
 	dcbt		BO,	PREB
 
-	stxvd2x		vs56,	o0,	BO
-	stxvd2x		vs57,	o16,	BO
-	stxvd2x		vs58,	o32,	BO
-	stxvd2x		vs59,	o48,	BO
-	stxvd2x		vs60,	o64,	BO
-	stxvd2x		vs61,	o80,	BO
-	stxvd2x		vs62,	o96,	BO
-	stxvd2x		vs63,	o112,	BO
+	stxvd2x		vs4,	o0,	BO
+	stxvd2x		vs5,	o16,	BO
+	stxvd2x		vs0,	o32,	BO
+	stxvd2x		vs2,	o48,	BO
+	stxvd2x		vs1,	o64,	BO
+	stxvd2x		vs3,	o80,	BO
+	stxvd2x		vs6,	o96,	BO
+	stxvd2x		vs12,	o112,	BO
 	addi		BO,	BO,	128
 
 
@@ -199,39 +203,39 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	addi		A1,	A1,	64
 
 
-	lxvd2x		vs16,	o0,	A2
-	lxvd2x		vs17,	o16,	A2
-	lxvd2x		vs18,	o32,	A2
-	lxvd2x		vs19,	o48,	A2
+	lxvd2x		vs4,	o0,	A2
+	lxvd2x		vs5,	o16,	A2
+	lxvd2x		vs6,	o32,	A2
+	lxvd2x		vs7,	o48,	A2
 	addi		A2,	A2,	64
 
 
-	lxvd2x		vs24,	o0,	A3
-	lxvd2x		vs25,	o16,	A3
-	lxvd2x		vs26,	o32,	A3
-	lxvd2x		vs27,	o48,	A3
+	lxvd2x		vs12,	o0,	A3
+	lxvd2x		vs13,	o16,	A3
+	lxvd2x		vs50,	o32,	A3
+	lxvd2x		vs51,	o48,	A3
 	addi		A3,	A3,	64
 
 
 	xxpermdi	vs32,	vs0,	vs8,	0
-	xxpermdi	vs33,	vs16,	vs24,	0
+	xxpermdi	vs33,	vs4,	vs12,	0
 	xxpermdi	vs34,	vs0,	vs8,	3
-	xxpermdi	vs35,	vs16,	vs24,	3
+	xxpermdi	vs35,	vs4,	vs12,	3
 
 	xxpermdi	vs36,	vs1,	vs9,	0
-	xxpermdi	vs37,	vs17,	vs25,	0
+	xxpermdi	vs37,	vs5,	vs13,	0
 	xxpermdi	vs38,	vs1,	vs9,	3
-	xxpermdi	vs39,	vs17,	vs25,	3
+	xxpermdi	vs39,	vs5,	vs13,	3
 
 	xxpermdi	vs40,	vs2,	vs10,	0
-	xxpermdi	vs41,	vs18,	vs26,	0
+	xxpermdi	vs41,	vs6,	vs50,	0
 	xxpermdi	vs42,	vs2,	vs10,	3
-	xxpermdi	vs43,	vs18,	vs26,	3
+	xxpermdi	vs43,	vs6,	vs50,	3
 
 	xxpermdi	vs44,	vs3,	vs11,	0
-	xxpermdi	vs45,	vs19,	vs27,	0
+	xxpermdi	vs45,	vs7,	vs51,	0
 	xxpermdi	vs46,	vs3,	vs11,	3
-	xxpermdi	vs47,	vs19,	vs27,	3
+	xxpermdi	vs47,	vs7,	vs51,	3
 
 
 	stxvd2x		vs32,	o0,	BO
@@ -274,25 +278,25 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	addi		A1,	A1,	32
 
 
-	lxvd2x		vs16,	o0,	A2
-	lxvd2x		vs17,	o16,	A2
+	lxvd2x		vs10,	o0,	A2
+	lxvd2x		vs11,	o16,	A2
 	addi		A2,	A2,	32
 
 
-	lxvd2x		vs24,	o0,	A3
-	lxvd2x		vs25,	o16,	A3
+	lxvd2x		vs12,	o0,	A3
+	lxvd2x		vs13,	o16,	A3
 	addi		A3,	A3,	32
 
 
 	xxpermdi	vs32,	vs0,	vs8,	0
-	xxpermdi	vs33,	vs16,	vs24,	0
+	xxpermdi	vs33,	vs10,	vs12,	0
 	xxpermdi	vs34,	vs0,	vs8,	3
-	xxpermdi	vs35,	vs16,	vs24,	3
+	xxpermdi	vs35,	vs10,	vs12,	3
 
 	xxpermdi	vs36,	vs1,	vs9,	0
-	xxpermdi	vs37,	vs17,	vs25,	0
+	xxpermdi	vs37,	vs11,	vs13,	0
 	xxpermdi	vs38,	vs1,	vs9,	3
-	xxpermdi	vs39,	vs17,	vs25,	3
+	xxpermdi	vs39,	vs11,	vs13,	3
 
 
 	stxvd2x		vs32,	o0,	BO
@@ -323,18 +327,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	addi		A1,	A1,	16
 
 
-	lxvd2x		vs16,	o0,	A2
+	lxvd2x		vs9,	o0,	A2
 	addi		A2,	A2,	16
 
 
-	lxvd2x		vs24,	o0,	A3
+	lxvd2x		vs10,	o0,	A3
 	addi		A3,	A3,	16
 
 
 	xxpermdi	vs32,	vs0,	vs8,	0
-	xxpermdi	vs33,	vs16,	vs24,	0
+	xxpermdi	vs33,	vs9,	vs10,	0
 	xxpermdi	vs34,	vs0,	vs8,	3
-	xxpermdi	vs35,	vs16,	vs24,	3
+	xxpermdi	vs35,	vs9,	vs10,	3
 
 
 	stxvd2x		vs32,	o0,	BO
@@ -361,16 +365,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	addi		A1,	A1,	8
 
 
-	lxsdx		vs16,	o0,	A2
+	lxsdx		vs9,	o0,	A2
 	addi		A2,	A2,	8
 
 
-	lxsdx		vs24,	o0,	A3
+	lxsdx		vs10,	o0,	A3
 	addi		A3,	A3,	8
 
 
 	xxpermdi	vs32,	vs0,	vs8,	0
-	xxpermdi	vs33,	vs16,	vs24,	0
+	xxpermdi	vs33,	vs9,	vs10,	0
 
 
 	stxvd2x		vs32,	o0,	BO
@@ -404,8 +408,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	lxvd2x		vs11,	o48,	A1
 	lxvd2x		vs12,	o64,	A1
 	lxvd2x		vs13,	o80,	A1
-	lxvd2x		vs14,	o96,	A1
-	lxvd2x		vs15,	o112,	A1
+	lxvd2x		vs48,	o96,	A1
+	lxvd2x		vs49,	o112,	A1
 	addi		A1,	A1,	128
 
 
@@ -427,11 +431,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	xxpermdi	vs42,	vs5,	vs13,	0
 	xxpermdi	vs43,	vs5,	vs13,	3
 
-	xxpermdi	vs44,	vs6,	vs14,	0
-	xxpermdi	vs45,	vs6,	vs14,	3
+	xxpermdi	vs44,	vs6,	vs48,	0
+	xxpermdi	vs45,	vs6,	vs48,	3
 
-	xxpermdi	vs46,	vs7,	vs15,	0
-	xxpermdi	vs47,	vs7,	vs15,	3
+	xxpermdi	vs46,	vs7,	vs49,	0
+	xxpermdi	vs47,	vs7,	vs49,	3
 
 
 	stxvd2x		vs32,	o0,	BO

kernel/power/dgemm_tcopy_16_power8.S

@@ -109,61 +109,35 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "dgemm_tcopy_macros_16_power8.S"
 
-#define STACKSIZE 384
-#define STACKSIZE 576
+#define STACKSIZE 144
+ 
 
 
 	PROLOGUE
 	PROFCODE
 
 	addi	SP, SP, -STACKSIZE
-//addi SP, SP, -208
 
 	li	r0, 0
 
-        std     r31,  144(SP)
-        std     r30,  152(SP)
-        std     r29,  160(SP)
-        std     r28,  168(SP)
-        std     r27,  176(SP)
-        std     r26,  184(SP)
-        std     r25,  192(SP)
-        std     r24,  200(SP)
-        std     r23,  208(SP)
-        std     r22,  216(SP)
-        std     r21,  224(SP)
-        std     r20,  232(SP)
-        std     r19,  240(SP)
-        std     r18,  248(SP)
-        std     r17,  256(SP)
-        std     r16,  264(SP)
-        std     r15,  272(SP)
-        std     r14,  280(SP)
-addi r11,SP,288
-        stvx    v20,  r11,r0
-addi r11,r11,16
-        stvx    v21,  r11,r0
-addi r11,r11,16
-        stvx    v22,  r11,r0
-addi r11,r11,16
-        stvx    v23,  r11,r0
-addi r11,r11,16
-        stvx    v24,  r11,r0
-addi r11,r11,16
-        stvx    v25,  r11,r0
-addi r11,r11,16
-        stvx    v26,  r11,r0
-addi r11,r11,16
-        stvx    v27,  r11,r0
-addi r11,r11,16
-        stvx    v28,  r11,r0
-addi r11,r11,16
-        stvx    v29,  r11,r0
-addi r11,r11,16
-        stvx    v30,  r11,r0
-addi r11,r11,16
-        stvx    v31,  r11,r0
-li r11,0
+    std r14,0(SP) 
+    std r15,8(SP) 
+    std r16,16(SP) 
+    std r17,24(SP) 
+    std r18,32(SP) 
+    std r19,40(SP) 
+    std r20,48(SP) 
+    std r21,56(SP) 
+    std r22,64(SP) 
+    std r23,72(SP) 
+    std r24,80(SP) 
+    std r25,88(SP) 
+    std r26,96(SP) 
+    std r27,104(SP) 
+    std r28,112(SP) 
+    std r29,120(SP) 
+    std r30,128(SP) 
+    std r31,136(SP) 
 
 	cmpwi	cr0, M, 0
 	ble-	L999
@@ -198,8 +172,7 @@ li r11,0
 	add	B2, B2, B
 	add	B1, B1, B
 
-	//li	PREA,  384
-	li	PREA,  576
+	li	PREA,  384 
 	addi	PREB,  M16, 128
 
 	li	o8,	8
@@ -213,52 +186,27 @@ L999:
 
 	li	r3, 0
 
-        ld      r31,  144(SP)
-        ld      r30,  152(SP)
-        ld      r29,  160(SP)
-        ld      r28,  168(SP)
-        ld      r27,  176(SP)
-        ld      r26,  184(SP)
-        ld      r25,  192(SP)
-        ld      r24,  200(SP)
-        ld      r23,  208(SP)
-        ld      r22,  216(SP)
-        ld      r21,  224(SP)
-        ld      r20,  232(SP)
-        ld      r19,  240(SP)
-        ld      r18,  248(SP)
-        ld      r17,  256(SP)
-        ld      r16,  264(SP)
-        ld      r15,  272(SP)
-        ld      r14,  280(SP)
-addi r11,SP,288
-        lvx    v20,  r11,r3
-addi r11,r11,16
-        lvx    v21,  r11,r3
-addi r11,r11,16
-        lvx    v22,  r11,r3
-addi r11,r11,16
-        lvx    v23,  r11,r3
-addi r11,r11,16
-        lvx    v24,  r11,r3
-addi r11,r11,16
-        lvx    v25,  r11,r3
-addi r11,r11,16
-        lvx    v26,  r11,r3
-addi r11,r11,16
-        lvx    v27,  r11,r3
-addi r11,r11,16
-        lvx    v28,  r11,r3
-addi r11,r11,16
-        lvx    v29,  r11,r3
-addi r11,r11,16
-        lvx    v30,  r11,r3
-addi r11,r11,16
-        lvx    v31,  r11,r3
-li r11,0
+    ld r14,0(SP) 
+    ld r15,8(SP) 
+    ld r16,16(SP) 
+    ld r17,24(SP) 
+    ld r18,32(SP) 
+    ld r19,40(SP) 
+    ld r20,48(SP) 
+    ld r21,56(SP) 
+    ld r22,64(SP) 
+    ld r23,72(SP) 
+    ld r24,80(SP) 
+    ld r25,88(SP) 
+    ld r26,96(SP) 
+    ld r27,104(SP) 
+    ld r28,112(SP) 
+    ld r29,120(SP) 
+    ld r30,128(SP) 
+    ld r31,136(SP) 
 
 	addi	SP, SP, STACKSIZE
-//addi SP, SP, 208
+ 
 	blr
 	EPILOGUE
 

kernel/power/dgemm_tcopy_macros_16_power8.S

@@ -58,10 +58,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	lxvd2x		vs51,	o48,	A2
 	addi		A2,	A2,	64
 
-	lxvd2x		vs56,	o0,	A3
-	lxvd2x		vs57,	o16,	A3
-	lxvd2x		vs58,	o32,	A3
-	lxvd2x		vs59,	o48,	A3
+	lxvd2x		vs4,	o0,	A3
+	lxvd2x		vs5,	o16,	A3
+	lxvd2x		vs6,	o32,	A3
+	lxvd2x		vs7,	o48,	A3
 	addi		A3,	A3,	64
 
 	lxvd2x		vs36,	o0,	A0
@@ -76,16 +76,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	lxvd2x		vs47,	o48,	A1
 	addi		A1,	A1,	64
 
-	lxvd2x		vs52,	o0,	A2
-	lxvd2x		vs53,	o16,	A2
-	lxvd2x		vs54,	o32,	A2
-	lxvd2x		vs55,	o48,	A2
+	lxvd2x		vs12,	o0,	A2
+	lxvd2x		vs13,	o16,	A2
+	lxvd2x		vs2,	o32,	A2
+	lxvd2x		vs3,	o48,	A2
 	addi		A2,	A2,	64
 
-	lxvd2x		vs60,	o0,	A3
-	lxvd2x		vs61,	o16,	A3
-	lxvd2x		vs62,	o32,	A3
-	lxvd2x		vs63,	o48,	A3
+	lxvd2x		vs8,	o0,	A3
+	lxvd2x		vs9,	o16,	A3
+	lxvd2x		vs10,	o32,	A3
+	lxvd2x		vs11,	o48,	A3
 	addi		A3,	A3,	64
 
 	mr		T1,	BO
@@ -122,23 +122,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	stxvd2x		vs51,	o48,	T1
 	addi		T1,	T1,	64
 
-	stxvd2x		vs52,	o0,	T1
-	stxvd2x		vs53,	o16,	T1
-	stxvd2x		vs54,	o32,	T1
-	stxvd2x		vs55,	o48,	T1
+	stxvd2x		vs12,	o0,	T1
+	stxvd2x		vs13,	o16,	T1
+	stxvd2x		vs2,	o32,	T1
+	stxvd2x		vs3,	o48,	T1
 
 	addi		T1,	T1,	64
 
-	stxvd2x		vs56,	o0,	T1
-	stxvd2x		vs57,	o16,	T1
-	stxvd2x		vs58,	o32,	T1
-	stxvd2x		vs59,	o48,	T1
+	stxvd2x		vs4,	o0,	T1
+	stxvd2x		vs5,	o16,	T1
+	stxvd2x		vs6,	o32,	T1
+	stxvd2x		vs7,	o48,	T1
 	addi		T1,	T1,	64
 
-	stxvd2x		vs60,	o0,	T1
-	stxvd2x		vs61,	o16,	T1
-	stxvd2x		vs62,	o32,	T1
-	stxvd2x		vs63,	o48,	T1
+	stxvd2x		vs8,	o0,	T1
+	stxvd2x		vs9,	o16,	T1
+	stxvd2x		vs10,	o32,	T1
+	stxvd2x		vs11,	o48,	T1
 
 .endm
 

kernel/power/dgemv_t.c

@@ -0,0 +1,886 @@
+/***************************************************************************
+Copyright (c) 2018, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *****************************************************************************/
+
+#include "common.h"
+
+#define NBMAX 8192
+#define PREFETCH 1
+#include <altivec.h> 
+
+#define HAVE_KERNEL4x8_ASM 1
+
+
+#if defined(HAVE_KERNEL4x8_ASM)
+static void dgemv_kernel_4x8(BLASLONG n, BLASLONG lda, double *ap, double *x, double *y, double alpha) {
+
+    FLOAT *a0, *a1, *a2, *a3, *a4, *a5, *a6, *a7;
+    BLASLONG off2;
+    BLASLONG tempR;
+    __asm__(
+         
+            "sldi   %[temp],%[off], 4 \n\t" // lda * sizeof (double) *2
+            "sldi   %[off], %[off], 3 \n\t" // lda * sizeof (double)    
+            "xxlxor 34,34,34  \n\t"
+            "xxlxor 35,34,34  \n\t"
+            "add    %[a2], %[a0], %[temp]    \n\t"
+            "add    %[a1], %[a0], %[off]     \n\t"
+            "xxlxor 4,34,34 \n\t"
+            "xxlxor 5,34,34 \n\t"
+            "xxlxor 6,34,34 \n\t"
+            "xxlxor 7,34,34 \n\t"            
+            "add    %[a3], %[a2], %[off]     \n\t"
+            "add    %[a4], %[a2], %[temp]    \n\t"            
+ 
+            "xxlxor 8,34,34 \n\t"
+            "xxlxor 9,34,34 \n\t"              
+            "add    %[a5], %[a3], %[temp]    \n\t"
+            "li     %[off],0    \n\t"
+            "li     %[off2],16  \n\t" 
+  
+            "add    %[a6], %[a4], %[temp]    \n\t" 
+            "add    %[a7], %[a5], %[temp]    \n\t"
+    
+
+
+
+            "lxvd2x 32, %[x],  %[off]  \n\t"
+            "lxvd2x 36, %[a0], %[off]  \n\t"
+            "lxvd2x 38, %[a1], %[off]   \n\t"
+            "lxvd2x 40, %[a2], %[off]   \n\t"
+            "lxvd2x 42, %[a3], %[off]   \n\t"
+            "lxvd2x 44, %[a4], %[off]  \n\t"
+            "lxvd2x 46, %[a5], %[off]   \n\t"
+            "lxvd2x 48, %[a6], %[off]   \n\t"
+            "lxvd2x 50, %[a7], %[off]   \n\t"
+            "lxvd2x 33, %[x],  %[off2]  \n\t"
+            "lxvd2x 37, %[a0], %[off2]  \n\t" 
+            "lxvd2x 39, %[a1], %[off2]  \n\t"
+            "lxvd2x 41, %[a2], %[off2]  \n\t"
+            "lxvd2x 43, %[a3], %[off2]  \n\t" 
+            "lxvd2x 45, %[a4], %[off2]  \n\t" 
+            "lxvd2x 47, %[a5], %[off2]  \n\t"
+            "lxvd2x 49, %[a6], %[off2]  \n\t"
+            "lxvd2x 51, %[a7], %[off2]  \n\t"  
+#if defined(PREFETCH)    
+            "li     %[temp],896  \n\t"
+#endif    
+            "addic. %[n],%[n],-4 \n\t"
+
+            "li       %[off],32 \n\t" 
+
+
+            "ble-     2f      \n\t"
+
+            //--------------------------------------------------           
+            ".p2align   5           \n\t"
+            "1:                     \n\t"
+            "xvmaddadp   34,36,32  \n\t"
+            "xvmaddadp   35,38,32  \n\t"
+            "addi   %[off2],  %[off2],32 \n\t"
+            "lxvd2x     36, %[a0], %[off]  \n\t"
+            "lxvd2x     38, %[a1], %[off]  \n\t"
+            "xvmaddadp   4,40,32  \n\t"
+            "xvmaddadp   5,42,32  \n\t"
+            "lxvd2x     40, %[a2], %[off]  \n\t"
+            "lxvd2x     42, %[a3], %[off]  \n\t"
+            "xvmaddadp   6,44,32  \n\t"
+            "xvmaddadp   7,46,32  \n\t" 
+            "lxvd2x     44, %[a4], %[off]  \n\t"
+            "lxvd2x     46, %[a5], %[off]  \n\t"
+            "xvmaddadp   8,48,32  \n\t"
+            "xvmaddadp   9,50,32  \n\t"
+            "lxvd2x     48, %[a6], %[off]  \n\t"
+            "lxvd2x     50, %[a7], %[off]  \n\t"
+            "lxvd2x     32,  %[x], %[off]  \n\t"
+
+            "xvmaddadp  34,37,33  \n\t"
+            "xvmaddadp  35,39,33  \n\t"            
+            "lxvd2x     37, %[a0], %[off2]  \n\t"
+            "lxvd2x     39, %[a1], %[off2]  \n\t"
+            "xvmaddadp  4,41,33  \n\t"
+            "xvmaddadp  5,43,33  \n\t"            
+            "addi       %[off],  %[off],32 \n\t"
+            "lxvd2x     41, %[a2], %[off2]  \n\t"
+            "lxvd2x     43, %[a3], %[off2]  \n\t"
+            "xvmaddadp  6,45,33  \n\t"
+            "xvmaddadp  7,47,33  \n\t"            
+            "lxvd2x     45, %[a4], %[off2]  \n\t"
+            "lxvd2x     47, %[a5], %[off2]  \n\t" 
+            "xvmaddadp  8,49,33  \n\t"
+            "xvmaddadp  9,51,33  \n\t" 
+            
+            "addic. %[n],%[n],-4 \n\t"                        
+            "lxvd2x     49, %[a6], %[off2]  \n\t"
+            "lxvd2x     51, %[a7], %[off2]  \n\t" 
+            "lxvd2x     33,  %[x], %[off2]  \n\t"            
+            "ble- 2f  \n\t"
+            "xvmaddadp   34,36,32  \n\t"
+            "xvmaddadp   35,38,32  \n\t"
+            "addi   %[off2],  %[off2],32 \n\t" 
+            "lxvd2x     36, %[a0], %[off]  \n\t"
+            "lxvd2x     38, %[a1], %[off]  \n\t"
+            "xvmaddadp   4,40,32  \n\t"
+            "xvmaddadp   5,42,32  \n\t"
+            "lxvd2x     40, %[a2], %[off]  \n\t"
+            "lxvd2x     42, %[a3], %[off]  \n\t"
+            "xvmaddadp   6,44,32  \n\t"
+            "xvmaddadp   7,46,32  \n\t" 
+            "lxvd2x     44, %[a4], %[off]  \n\t"
+            "lxvd2x     46, %[a5], %[off]  \n\t"
+            "xvmaddadp   8,48,32  \n\t"
+            "xvmaddadp   9,50,32  \n\t"
+            "lxvd2x     48, %[a6], %[off]  \n\t"
+            "lxvd2x     50, %[a7], %[off]  \n\t"
+            "lxvd2x     32,  %[x], %[off]  \n\t"
+
+            "xvmaddadp  34,37,33  \n\t"
+            "xvmaddadp  35,39,33  \n\t"            
+            "lxvd2x     37, %[a0], %[off2]  \n\t"
+            "lxvd2x     39, %[a1], %[off2]  \n\t"
+            "xvmaddadp  4,41,33  \n\t"
+            "xvmaddadp  5,43,33  \n\t"            
+            "addi       %[off],  %[off],32 \n\t"
+            "lxvd2x     41, %[a2], %[off2]  \n\t"
+            "lxvd2x     43, %[a3], %[off2]  \n\t"
+            "xvmaddadp  6,45,33  \n\t"
+            "xvmaddadp  7,47,33  \n\t"            
+            "lxvd2x     45, %[a4], %[off2]  \n\t"
+            "lxvd2x     47, %[a5], %[off2]  \n\t" 
+            "xvmaddadp  8,49,33  \n\t"
+            "xvmaddadp  9,51,33  \n\t" 
+            
+            "addic. %[n],%[n],-4 \n\t"                        
+            "lxvd2x     49, %[a6], %[off2]  \n\t"
+            "lxvd2x     51, %[a7], %[off2]  \n\t" 
+            "lxvd2x     33,  %[x], %[off2]  \n\t"            
+            "ble- 2f  \n\t"
+            "xvmaddadp   34,36,32  \n\t"
+            "xvmaddadp   35,38,32  \n\t"
+#if defined(PREFETCH)            
+            "addi    %[temp],%[temp],128 \n\t"   
+#endif                                             
+            "addi   %[off2],  %[off2],32 \n\t" 
+            "lxvd2x     36, %[a0], %[off]  \n\t"
+            "lxvd2x     38, %[a1], %[off]  \n\t"
+            "xvmaddadp   4,40,32  \n\t"
+            "xvmaddadp   5,42,32  \n\t"
+            "lxvd2x     40, %[a2], %[off]  \n\t"
+            "lxvd2x     42, %[a3], %[off]  \n\t"
+            "xvmaddadp   6,44,32  \n\t"
+            "xvmaddadp   7,46,32  \n\t" 
+            "lxvd2x     44, %[a4], %[off]  \n\t"
+            "lxvd2x     46, %[a5], %[off]  \n\t"
+            "xvmaddadp   8,48,32  \n\t"
+            "xvmaddadp   9,50,32  \n\t"
+            "lxvd2x     48, %[a6], %[off]  \n\t"
+            "lxvd2x     50, %[a7], %[off]  \n\t"
+#if defined(PREFETCH)
+            "dcbt   %[temp],%[a0]  \n\t"
+#endif            
+            "lxvd2x     32,  %[x], %[off]  \n\t"
+
+            "xvmaddadp  34,37,33  \n\t"
+            "xvmaddadp  35,39,33  \n\t"            
+            "lxvd2x     37, %[a0], %[off2]  \n\t"
+            "lxvd2x     39, %[a1], %[off2]  \n\t"
+            "xvmaddadp  4,41,33  \n\t"
+            "xvmaddadp  5,43,33  \n\t"            
+#if defined(PREFETCH)
+            "dcbt   %[temp],%[a1]  \n\t"
+#endif            
+            "lxvd2x     41, %[a2], %[off2]  \n\t"
+            "addi       %[off],  %[off],32 \n\t"
+            "lxvd2x     43, %[a3], %[off2]  \n\t"
+            "xvmaddadp  6,45,33  \n\t"
+            "xvmaddadp  7,47,33  \n\t"            
+            "lxvd2x     45, %[a4], %[off2]  \n\t"
+            "lxvd2x     47, %[a5], %[off2]  \n\t" 
+            "xvmaddadp  8,49,33  \n\t"
+            "xvmaddadp  9,51,33  \n\t" 
+#if defined(PREFETCH)
+            "dcbt   %[temp],%[a3]  \n\t"
+#endif            
+            "lxvd2x     49, %[a6], %[off2]  \n\t"
+            "lxvd2x     51, %[a7], %[off2]  \n\t" 
+            
+            "lxvd2x     33,  %[x], %[off2]  \n\t"            
+            "addic. %[n],%[n],-4 \n\t"                        
+            "ble- 2f  \n\t"            
+ 
+            "addi   %[off2],  %[off2],32 \n\t" 
+#if defined(PREFETCH)
+            "dcbt   %[temp],%[a2]  \n\t"
+#endif            
+            "xvmaddadp   34,36,32  \n\t"
+            "xvmaddadp   35,38,32  \n\t"
+            "lxvd2x     36, %[a0], %[off]  \n\t"
+            "lxvd2x     38, %[a1], %[off]  \n\t"
+            "xvmaddadp   4,40,32  \n\t"
+            "xvmaddadp   5,42,32  \n\t"
+            "lxvd2x     40, %[a2], %[off]  \n\t"
+            "lxvd2x     42, %[a3], %[off]  \n\t"
+#if defined(PREFETCH)
+            "dcbt   %[temp],%[a4]  \n\t"                         
+#endif            
+            "xvmaddadp   6,44,32  \n\t"
+            "xvmaddadp   7,46,32  \n\t" 
+            "lxvd2x     44, %[a4], %[off]  \n\t"
+            "lxvd2x     46, %[a5], %[off]  \n\t"
+            "xvmaddadp   8,48,32  \n\t"
+            "xvmaddadp   9,50,32  \n\t"
+            "lxvd2x     48, %[a6], %[off]  \n\t"
+            "lxvd2x     50, %[a7], %[off]  \n\t"
+            "lxvd2x     32,  %[x], %[off]  \n\t"
+
+#if defined(PREFETCH)
+          "dcbt   %[temp],%[a5]  \n\t"
+#endif              
+            "xvmaddadp  34,37,33  \n\t"
+            "xvmaddadp  35,39,33  \n\t"            
+            "lxvd2x     37, %[a0], %[off2]  \n\t"
+            "lxvd2x     39, %[a1], %[off2]  \n\t"
+            "xvmaddadp  4,41,33  \n\t"
+            "xvmaddadp  5,43,33  \n\t"            
+            "addi       %[off],  %[off],32 \n\t"
+            "lxvd2x     41, %[a2], %[off2]  \n\t"
+            "lxvd2x     43, %[a3], %[off2]  \n\t"
+#if defined(PREFETCH)
+            "dcbt   %[temp],%[a6]  \n\t"  
+#endif            
+            "xvmaddadp  6,45,33  \n\t"
+            "xvmaddadp  7,47,33  \n\t"            
+            "lxvd2x     45, %[a4], %[off2]  \n\t"
+            "lxvd2x     47, %[a5], %[off2]  \n\t" 
+            "xvmaddadp  8,49,33  \n\t"
+            "xvmaddadp  9,51,33  \n\t" 
+            
+#if defined(PREFETCH)
+            "dcbt   %[temp],%[a7]  \n\t"  
+#endif            
+            "lxvd2x     49, %[a6], %[off2]  \n\t"
+            "addic. %[n],%[n],-4 \n\t"
+            "lxvd2x     51, %[a7], %[off2]  \n\t" 
+            "lxvd2x     33,  %[x], %[off2]  \n\t"            
+#if defined(PREFETCH)
+            "dcbt   %[temp],%[x]  \n\t" 
+#endif            
+            "bgt+ 1b  \n\t"
+            ".p2align   5           \n\t"
+            "2: \n\t"
+            //--------------------------------------------
+
+            "xvmaddadp   34,36,32  \n\t"
+            "xvmaddadp   35,38,32  \n\t"
+            "xvmaddadp   4,40,32  \n\t"
+            "xvmaddadp   5,42,32  \n\t"
+            "xvmaddadp   6,44,32  \n\t"
+            "xvmaddadp   7,46,32  \n\t" 
+            "xvmaddadp   8,48,32  \n\t"
+            "xvmaddadp   9,50,32  \n\t" 
+            "xxspltd 36, %x[alpha], 0 \n\t"
+            "xvmaddadp  34,37,33  \n\t"
+            "xvmaddadp  35,39,33  \n\t"            
+            "xvmaddadp  4,41,33  \n\t"
+            "xvmaddadp  5,43,33  \n\t"            
+            "xvmaddadp  6,45,33  \n\t"
+            "xvmaddadp  7,47,33  \n\t"            
+            "xvmaddadp  8,49,33  \n\t"
+            "xvmaddadp  9,51,33  \n\t"  
+
+            "lxvd2x 37, 0, %[y] \n\t"
+            "li     %[off2],16   \n\t"
+            "lxvd2x 38, %[off2], %[y] \n\t"
+   
+            "li     %[off2],32   \n\t"
+            "lxvd2x 39, %[off2], %[y] \n\t"
+            "li     %[off2],48   \n\t"            
+            "lxvd2x 40, %[off2], %[y] \n\t"
+
+ 
+
+            "xxmrgld 42,34,35 \n\t"
+            "xxmrghd 43,34,35 \n\t"
+
+            "xxmrgld 44,4,5 \n\t"
+            "xxmrghd 45,4,5 \n\t"
+
+            "xvadddp 42,42,43 \n\t"
+
+            "xxmrgld 46,6,7 \n\t"
+            "xxmrghd 47,6,7 \n\t"
+
+            "xvadddp 44,44,45 \n\t"
+
+            "xxmrgld 48,8,9 \n\t"
+            "xxmrghd 49,8,9 \n\t"
+
+            "xvadddp 46,46,47 \n\t"
+            
+            "xvmaddadp  37,42,36  \n\t"
+            "xvmaddadp  38,44,36  \n\t"
+            
+            "xvadddp 48,48,49 \n\t"
+
+            "xvmaddadp  39,46,36  \n\t"
+
+            "stxvd2x 37, 0, %[y] \n\t"
+            "li %[off],16 \n\t"
+            "stxvd2x 38, %[off], %[y]  \n\t"
+            "xvmaddadp  40,48,36  \n\t" 
+            "li %[off],32 \n\t"
+            "stxvd2x 39, %[off], %[y]  \n\t"
+            "stxvd2x 40, %[off2], %[y]  \n\t"     
+                 
+            : [memy] "+m" (*(const double (*)[8])y),
+            [n] "+&r" (n),
+            [a0] "=b" (a0),
+            [a1] "=&b" (a1),
+            [a2] "=&b" (a2),
+            [a3] "=&b" (a3),
+            [a4] "=&b" (a4),
+            [a5] "=&b" (a5),
+            [a6] "=&b" (a6),
+            [a7] "=&b" (a7),            
+            [off] "+&b" (lda),
+            [off2]"=&b" (off2),
+            [temp] "=&b" (tempR)
+            : [memx] "m" (*(const double (*)[n])x),
+            [mem_ap] "m" (*(const double (*)[]) ap),
+            [alpha] "d" (alpha),
+            "[a0]" (ap),
+            [x] "b" (x),
+            [y] "b" (y)
+            : "cc","vs4","vs5","vs6","vs7","vs8","vs9" ,"vs32","vs33","vs34","vs35", "vs36", "vs37", "vs38", "vs39",
+            "vs40", "vs41", "vs42", "vs43", "vs44", "vs45", "vs46", "vs47", "vs48", "vs49", "vs50", "vs51"
+            );
+    return;
+}
+#else
+static void dgemv_kernel_4x8(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha) {
+    BLASLONG i;
+#if defined(PREFETCH)  
+    BLASLONG j, c, k;
+#endif    
+    FLOAT *a0, *a1, *a2, *a3, *a4, *a5, *a6, *a7;
+    __vector double *va0, *va1, *va2, *va3, *va4, *va5, *va6, *va7, *v_x;
+    register __vector double temp0 = {0, 0};
+    register __vector double temp1 = {0, 0};
+    register __vector double temp2 = {0, 0};
+    register __vector double temp3 = {0, 0};
+    register __vector double temp4 = {0, 0};
+    register __vector double temp5 = {0, 0};
+    register __vector double temp6 = {0, 0};
+    register __vector double temp7 = {0, 0};
+
+    a0 = ap;
+    a1 = ap + lda;
+    a2 = a1 + lda;
+    a3 = a2 + lda;
+    a4 = a3 + lda;
+    a5 = a4 + lda;
+    a6 = a5 + lda;
+    a7 = a6 + lda;
+    va0 = (__vector double*) a0;
+    va1 = (__vector double*) a1;
+    va2 = (__vector double*) a2;
+    va3 = (__vector double*) a3;
+    va4 = (__vector double*) a4;
+    va5 = (__vector double*) a5;
+    va6 = (__vector double*) a6;
+    va7 = (__vector double*) a7;
+    v_x = (__vector double*) x;
+ 
+#if defined(PREFETCH)
+
+    c = n >> 1;
+
+    for (j = 0; j < c; j += 64) {
+        k = (c - j) > 64 ? 64 : (c - j);
+        __builtin_prefetch(v_x + 64);
+        __builtin_prefetch(va0 + 64);
+        __builtin_prefetch(va1 + 64);
+        __builtin_prefetch(va2 + 64);
+        __builtin_prefetch(va3 + 64);
+        __builtin_prefetch(va4 + 64);
+        __builtin_prefetch(va5 + 64);
+        __builtin_prefetch(va6 + 64);
+        __builtin_prefetch(va7 + 64); 
+         for (i = 0; i < k; i += 2) {
+#else
+        
+        for (i = 0; i < n/2; i += 2) {
+#endif
+            temp0 += v_x[i] * va0[i];
+            temp1 += v_x[i] * va1[i];
+            temp2 += v_x[i] * va2[i];
+            temp3 += v_x[i] * va3[i];
+            temp4 += v_x[i] * va4[i];
+            temp5 += v_x[i] * va5[i];
+            temp6 += v_x[i] * va6[i];
+            temp7 += v_x[i] * va7[i];
+            temp0 += v_x[i + 1] * va0[i + 1];
+            temp1 += v_x[i + 1] * va1[i + 1];
+            temp2 += v_x[i + 1] * va2[i + 1];
+            temp3 += v_x[i + 1] * va3[i + 1];
+
+            temp4 += v_x[i + 1] * va4[i + 1];
+            temp5 += v_x[i + 1] * va5[i + 1];
+            temp6 += v_x[i + 1] * va6[i + 1];
+            temp7 += v_x[i + 1] * va7[i + 1];
+        }
+#if defined(PREFETCH)
+        va0 += 64;
+        va1 += 64;
+        va2 += 64;
+        va3 += 64;
+        va4 += 64;
+        va5 += 64;
+        va6 += 64;
+        va7 += 64;
+        v_x += 64;
+
+    }
+#endif
+    y[0] += alpha * (temp0[0] + temp0[1]);
+    y[1] += alpha * (temp1[0] + temp1[1]);
+    y[2] += alpha * (temp2[0] + temp2[1]);
+    y[3] += alpha * (temp3[0] + temp3[1]);
+
+    y[4] += alpha * (temp4[0] + temp4[1]);
+    y[5] += alpha * (temp5[0] + temp5[1]);
+    y[6] += alpha * (temp6[0] + temp6[1]);
+    y[7] += alpha * (temp7[0] + temp7[1]);
+
+}
+
+#endif
+ 
+
+static void dgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha) {
+    BLASLONG i = 0;
+    FLOAT *a0, *a1, *a2, *a3;
+    a0 = ap;
+    a1 = ap + lda;
+    a2 = a1 + lda;
+    a3 = a2 + lda;
+    __vector double* va0 = (__vector double*) a0;
+    __vector double* va1 = (__vector double*) a1;
+    __vector double* va2 = (__vector double*) a2;
+    __vector double* va3 = (__vector double*) a3;
+    __vector double* v_x = (__vector double*) x;
+    register __vector double temp0 = {0, 0};
+    register __vector double temp1 = {0, 0};
+    register __vector double temp2 = {0, 0};
+    register __vector double temp3 = {0, 0};
+    register __vector double temp4 = {0, 0};
+    register __vector double temp5 = {0, 0};
+    register __vector double temp6 = {0, 0};
+    register __vector double temp7 = {0, 0};
+
+    for (i = 0; i < n / 2; i += 2) {
+        temp0 += v_x[i] * va0[i];
+        temp1 += v_x[i] * va1[i];
+        temp2 += v_x[i] * va2[i];
+        temp3 += v_x[i] * va3[i];
+        temp4 += v_x[i + 1] * va0[i + 1];
+        temp5 += v_x[i + 1] * va1[i + 1];
+        temp6 += v_x[i + 1] * va2[i + 1];
+        temp7 += v_x[i + 1] * va3[i + 1];
+    }
+
+    temp0 += temp4;
+    temp1 += temp5;
+    temp2 += temp6;
+    temp3 += temp7;
+    y[0] += alpha * (temp0[0] + temp0[1]);
+    y[1] += alpha * (temp1[0] + temp1[1]);
+    y[2] += alpha * (temp2[0] + temp2[1]);
+    y[3] += alpha * (temp3[0] + temp3[1]);
+
+}
+ 
+
+static void dgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha, BLASLONG inc_y) {
+
+    BLASLONG i;
+    FLOAT *a0, *a1;
+    a0 = ap;
+    a1 = ap + lda;
+    __vector double* va0 = (__vector double*) a0;
+    __vector double* va1 = (__vector double*) a1;
+    __vector double* v_x = (__vector double*) x;
+    __vector double temp0 = {0, 0};
+    __vector double temp1 = {0, 0};
+    for (i = 0; i < n / 2; i += 2) {
+        temp0 += v_x[i] * va0[i] + v_x[i + 1] * va0[i + 1];
+        temp1 += v_x[i] * va1[i] + v_x[i + 1] * va1[i + 1];
+    }
+
+
+
+    y[0] += alpha * (temp0[0] + temp0[1]);
+    y[inc_y] += alpha * (temp1[0] + temp1[1]);
+}
+
+static void dgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha) {
+
+    BLASLONG i;
+    FLOAT *a0;
+    a0 = ap;
+    __vector double* va0 = (__vector double*) a0;
+    __vector double* v_x = (__vector double*) x;
+    __vector double temp0 = {0, 0};
+    for (i = 0; i < n / 2; i += 2) {
+        temp0 += v_x[i] * va0[i] + v_x[i + 1] * va0[i + 1];
+    }
+
+    *y += alpha * (temp0[0] + temp0[1]);
+
+}
+
+static void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src) {
+    BLASLONG i;
+    for (i = 0; i < n; i++) {
+        *dest++ = *src;
+        src += inc_src;
+    }
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer) {
+    BLASLONG i;
+    BLASLONG j;
+    FLOAT *a_ptr;
+    FLOAT *x_ptr;
+    FLOAT *y_ptr;
+
+    BLASLONG n1;
+    BLASLONG m1;
+    BLASLONG m2;
+    BLASLONG m3;
+    BLASLONG n2;
+
+    FLOAT ybuffer[8], *xbuffer;
+
+    if (m < 1) return (0);
+    if (n < 1) return (0);
+
+    xbuffer = buffer;
+
+    n1 = n >> 3;
+    n2 = n & 7;
+
+    m3 = m & 3;
+    m1 = m - m3;
+    m2 = (m & (NBMAX - 1)) - m3;
+
+    BLASLONG NB = NBMAX;
+
+    while (NB == NBMAX) {
+
+        m1 -= NB;
+        if (m1 < 0) {
+            if (m2 == 0) break;
+            NB = m2;
+        }
+
+        y_ptr = y;
+        a_ptr = a;
+        x_ptr = x;
+
+        if (inc_x != 1)
+            copy_x(NB, x_ptr, xbuffer, inc_x);
+        else
+            xbuffer = x_ptr;
+
+        BLASLONG lda8 = lda << 3;
+
+
+        if (inc_y == 1) {
+
+            for (i = 0; i < n1; i++) {
+                 
+                dgemv_kernel_4x8(NB, lda, a_ptr, xbuffer, y_ptr, alpha);
+ 
+                y_ptr += 8;
+                a_ptr += lda8;
+#if defined(PREFETCH)                
+               __builtin_prefetch(y_ptr+64);
+#endif               
+            }
+
+        } else {
+                   
+            for (i = 0; i < n1; i++) {
+                ybuffer[0] = 0;
+                ybuffer[1] = 0;
+                ybuffer[2] = 0;
+                ybuffer[3] = 0;
+                ybuffer[4] = 0;
+                ybuffer[5] = 0;
+                ybuffer[6] = 0;
+                ybuffer[7] = 0;
+                dgemv_kernel_4x8(NB, lda, a_ptr, xbuffer, ybuffer, alpha);
+
+ 
+
+                *y_ptr += ybuffer[0];
+                y_ptr += inc_y;
+                *y_ptr += ybuffer[1];
+                y_ptr += inc_y;
+                *y_ptr += ybuffer[2];
+                y_ptr += inc_y;
+                *y_ptr += ybuffer[3];
+                y_ptr += inc_y;
+
+                *y_ptr += ybuffer[4];
+                y_ptr += inc_y;
+                *y_ptr += ybuffer[5];
+                y_ptr += inc_y;
+                *y_ptr += ybuffer[6];
+                y_ptr += inc_y;
+                *y_ptr += ybuffer[7];
+                y_ptr += inc_y;
+
+                a_ptr += lda8;
+            }
+
+        }
+
+
+        if (n2 & 4) {
+            ybuffer[0] = 0;
+            ybuffer[1] = 0;
+            ybuffer[2] = 0;
+            ybuffer[3] = 0;
+            dgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, ybuffer, alpha);
+
+            a_ptr += lda<<2;
+
+            *y_ptr += ybuffer[0];
+            y_ptr += inc_y;
+            *y_ptr += ybuffer[1];
+            y_ptr += inc_y;
+            *y_ptr += ybuffer[2];
+            y_ptr += inc_y;
+            *y_ptr += ybuffer[3];
+            y_ptr += inc_y;
+        }
+
+        if (n2 & 2) {
+            dgemv_kernel_4x2(NB, lda, a_ptr, xbuffer, y_ptr, alpha, inc_y);
+            a_ptr += lda << 1;
+            y_ptr += 2 * inc_y;
+
+        }
+
+        if (n2 & 1) {
+            dgemv_kernel_4x1(NB, a_ptr, xbuffer, y_ptr, alpha);
+            a_ptr += lda;
+            y_ptr += inc_y;
+
+        }
+
+        a += NB;
+        x += NB * inc_x;
+
+
+    }
+
+    if (m3 == 0) return (0);
+
+    x_ptr = x;
+    a_ptr = a;
+    if (m3 == 3) {
+        FLOAT xtemp0 = *x_ptr * alpha;
+        x_ptr += inc_x;
+        FLOAT xtemp1 = *x_ptr * alpha;
+        x_ptr += inc_x;
+        FLOAT xtemp2 = *x_ptr * alpha;
+
+        FLOAT *aj = a_ptr;
+        y_ptr = y;
+
+        if (lda == 3 && inc_y == 1) {
+
+            for (j = 0; j < (n & -4); j += 4) {
+
+                y_ptr[j] += aj[0] * xtemp0 + aj[1] * xtemp1 + aj[2] * xtemp2;
+                y_ptr[j + 1] += aj[3] * xtemp0 + aj[4] * xtemp1 + aj[5] * xtemp2;
+                y_ptr[j + 2] += aj[6] * xtemp0 + aj[7] * xtemp1 + aj[8] * xtemp2;
+                y_ptr[j + 3] += aj[9] * xtemp0 + aj[10] * xtemp1 + aj[11] * xtemp2;
+                aj += 12;
+            }
+
+            for (; j < n; j++) {
+                y_ptr[j] += aj[0] * xtemp0 + aj[1] * xtemp1 + aj[2] * xtemp2;
+                aj += 3;
+            }
+
+        } else {
+
+            if (inc_y == 1) {
+
+                BLASLONG register lda2 = lda << 1;
+                BLASLONG register lda4 = lda << 2;
+                BLASLONG register lda3 = lda2 + lda;
+
+                for (j = 0; j < (n & -4); j += 4) {
+
+                    y_ptr[j] += *aj * xtemp0 + *(aj + 1) * xtemp1 + *(aj + 2) * xtemp2;
+                    y_ptr[j + 1] += *(aj + lda) * xtemp0 + *(aj + lda + 1) * xtemp1 + *(aj + lda + 2) * xtemp2;
+                    y_ptr[j + 2] += *(aj + lda2) * xtemp0 + *(aj + lda2 + 1) * xtemp1 + *(aj + lda2 + 2) * xtemp2;
+                    y_ptr[j + 3] += *(aj + lda3) * xtemp0 + *(aj + lda3 + 1) * xtemp1 + *(aj + lda3 + 2) * xtemp2;
+                    aj += lda4;
+                }
+
+                for (; j < n; j++) {
+
+                    y_ptr[j] += *aj * xtemp0 + *(aj + 1) * xtemp1 + *(aj + 2) * xtemp2;
+                    aj += lda;
+                }
+
+            } else {
+
+                for (j = 0; j < n; j++) {
+                    *y_ptr += *aj * xtemp0 + *(aj + 1) * xtemp1 + *(aj + 2) * xtemp2;
+                    y_ptr += inc_y;
+                    aj += lda;
+                }
+
+            }
+
+        }
+        return (0);
+    }
+
+    if (m3 == 2) {
+        FLOAT xtemp0 = *x_ptr * alpha;
+        x_ptr += inc_x;
+        FLOAT xtemp1 = *x_ptr * alpha;
+
+        FLOAT *aj = a_ptr;
+        y_ptr = y;
+
+        if (lda == 2 && inc_y == 1) {
+
+            for (j = 0; j < (n & -4); j += 4) {
+                y_ptr[j] += aj[0] * xtemp0 + aj[1] * xtemp1;
+                y_ptr[j + 1] += aj[2] * xtemp0 + aj[3] * xtemp1;
+                y_ptr[j + 2] += aj[4] * xtemp0 + aj[5] * xtemp1;
+                y_ptr[j + 3] += aj[6] * xtemp0 + aj[7] * xtemp1;
+                aj += 8;
+
+            }
+
+            for (; j < n; j++) {
+                y_ptr[j] += aj[0] * xtemp0 + aj[1] * xtemp1;
+                aj += 2;
+            }
+
+        } else {
+            if (inc_y == 1) {
+
+                BLASLONG register lda2 = lda << 1;
+                BLASLONG register lda4 = lda << 2;
+                BLASLONG register lda3 = lda2 + lda;
+
+                for (j = 0; j < (n & -4); j += 4) {
+
+                    y_ptr[j] += *aj * xtemp0 + *(aj + 1) * xtemp1;
+                    y_ptr[j + 1] += *(aj + lda) * xtemp0 + *(aj + lda + 1) * xtemp1;
+                    y_ptr[j + 2] += *(aj + lda2) * xtemp0 + *(aj + lda2 + 1) * xtemp1;
+                    y_ptr[j + 3] += *(aj + lda3) * xtemp0 + *(aj + lda3 + 1) * xtemp1;
+                    aj += lda4;
+                }
+
+                for (; j < n; j++) {
+
+                    y_ptr[j] += *aj * xtemp0 + *(aj + 1) * xtemp1;
+                    aj += lda;
+                }
+
+            } else {
+                for (j = 0; j < n; j++) {
+                    *y_ptr += *aj * xtemp0 + *(aj + 1) * xtemp1;
+                    y_ptr += inc_y;
+                    aj += lda;
+                }
+            }
+
+        }
+        return (0);
+
+    }
+
+    FLOAT xtemp = *x_ptr * alpha;
+    FLOAT *aj = a_ptr;
+    y_ptr = y;
+    if (lda == 1 && inc_y == 1) {
+        for (j = 0; j < (n & -4); j += 4) {
+            y_ptr[j] += aj[j] * xtemp;
+            y_ptr[j + 1] += aj[j + 1] * xtemp;
+            y_ptr[j + 2] += aj[j + 2] * xtemp;
+            y_ptr[j + 3] += aj[j + 3] * xtemp;
+        }
+        for (; j < n; j++) {
+            y_ptr[j] += aj[j] * xtemp;
+        }
+
+
+    } else {
+        if (inc_y == 1) {
+
+            BLASLONG register lda2 = lda << 1;
+            BLASLONG register lda4 = lda << 2;
+            BLASLONG register lda3 = lda2 + lda;
+            for (j = 0; j < (n & -4); j += 4) {
+                y_ptr[j] += *aj * xtemp;
+                y_ptr[j + 1] += *(aj + lda) * xtemp;
+                y_ptr[j + 2] += *(aj + lda2) * xtemp;
+                y_ptr[j + 3] += *(aj + lda3) * xtemp;
+                aj += lda4;
+            }
+
+            for (; j < n; j++) {
+                y_ptr[j] += *aj * xtemp;
+                aj += lda;
+            }
+
+        } else {
+            for (j = 0; j < n; j++) {
+                *y_ptr += *aj * xtemp;
+                y_ptr += inc_y;
+                aj += lda;
+            }
+
+        }
+    }
+
+    return (0);
+
+}
+

kernel/power/idamax.c

@@ -0,0 +1,383 @@
+/***************************************************************************
+Copyright (c) 2013-2018, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *****************************************************************************/
+#include "common.h"
+#include <math.h>
+#include <altivec.h>
+#if defined(DOUBLE)
+
+#define ABS fabs
+
+#else
+
+#define ABS fabsf
+
+#endif
+
+/**
+ * Find  maximum index 
+ * Warning: requirements n>0  and n % 32 == 0
+ * @param n     
+ * @param x     pointer to the vector
+ * @param maxf  (out) maximum absolute value .( only for output )
+ * @return  index 
+ */
+static BLASLONG diamax_kernel_32(BLASLONG n, FLOAT *x, FLOAT *maxf) {
+    BLASLONG index;
+    register __vector long long start = {1,0};
+    register __vector long long temp_add_index = {2, 2}; 
+    __asm__(
+ 
+
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t" 
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"
+
+            "xxlor  40,%x[start],%x[start]  \n\t" //{ 1,0} vs40 | v8 
+            "vaddudm  9,8,%[adder]   \n\t" //{3,2} vs41 
+            "xxlxor  37,37 ,37       \n\t" //v5 v37 index_count
+            "vaddudm  10,9,%[adder]  \n\t" //{5,4} vs42
+            "xxlxor  38 ,38 ,38      \n\t"  // v6 | vs38 vec_max_index
+            "vaddudm  11,10,%[adder] \n\t" //{7,6} vs43
+            "xxlxor  39,39,39        \n\t" //   vs39 vec_max_value
+            "vaddudm 4,11,  %[adder] \n\t" // {9,8} -{8;8} vs36 | v4
+            "xxspltd   36,36,0       \n\t"
+    
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t"
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"                         
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+  
+            //jump first half forward 
+            "b 2f  \n\t"
+
+//===================================================================
+
+            ".p2align   5            \n\t"
+
+            "1: \n\t"
+            "xvcmpgtdp  2,45,44  \n\t "
+            "xvcmpgtdp  3,47,46  \n\t "
+            "xvcmpgtdp  4,49,48  \n\t "
+            "xvcmpgtdp  5,51,50  \n\t"
+
+            "xxsel    32,40,41,2 \n\t"
+            "xxsel     0,44,45,2 \n\t" 
+            "xxsel    33,42,43,3 \n\t"
+            "xxsel     1,46,47,3 \n\t" 
+            "xxsel    34,40,41,4 \n\t"
+            "xxsel    45,48,49,4 \n\t"
+            "xxsel    35,42,43,5 \n\t"
+            "xxsel    47,50,51,5 \n\t"
+
+            "xvcmpgtdp 2, 1,0     \n\t"
+            "xvcmpgtdp 3,47, 45   \n\t"
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t" 
+    
+            "xxsel     32,32,33,2 \n\t"
+            "xxsel       0 ,0,1,2 \n\t"
+            "xxsel     34,34,35,3 \n\t"
+            "xxsel     5,45,47,3 \n\t"
+ 
+          
+            //load next 64
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+    
+            // for {second 8 elements } we have to add 8 to each so that it became {from 8 to 16}
+            "vaddudm     2,2,4  \n\t" // vs34=vs34 + vs36{8,8} 
+    
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t" 
+
+            //choose bigger from first and second part
+            "xvcmpgtdp 4,5 , 0    \n\t" 
+            "xxsel     3, 0,5,4  \n\t"
+            "xxsel     33,32,34,4  \n\t"
+
+            //load next 64
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"
+
+            "vaddudm  1,1,5  \n\t"  //  get real index for first bigger  
+
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"
+
+            //compare with previous to get vec_max_index(v6 | vs38 ) and vec_max_value (vs39)   
+            "xvcmpgtdp 2, 3,39    \n\t"
+            "xxsel     39,39,3,2  \n\t"
+            "xxsel     38,38,33,2  \n\t"
+    
+            //update index += 8 
+            "vaddudm   5,5,4       \n\t"  
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t"
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+
+            //update index += 8 
+            "vaddudm   5,5,4       \n\t"  
+
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+//<-----------jump here from first load
+             "2:                  \n\t"
+    
+            "xvcmpgtdp  2,45,44  \n\t "
+            "xvcmpgtdp  3,47,46  \n\t "
+            "xvcmpgtdp  4,49,48  \n\t "
+            "xvcmpgtdp  5,51,50  \n\t"
+
+            "xxsel    32,40,41,2 \n\t"
+            "xxsel     0,44,45,2 \n\t" 
+            "xxsel    33,42,43,3 \n\t"
+            "xxsel     1,46,47,3 \n\t" 
+            "xxsel    34,40,41,4 \n\t"
+            "xxsel    45,48,49,4 \n\t"
+            "xxsel    35,42,43,5 \n\t"
+            "xxsel    47,50,51,5 \n\t"
+
+            "xvcmpgtdp 2, 1,0     \n\t"
+            "xvcmpgtdp 3,47, 45   \n\t"
+            "xxsel     32,32,33,2 \n\t"
+            "xxsel       0 ,0,1,2 \n\t"
+            "xxsel     34,34,35,3 \n\t"
+            "xxsel     5,45,47,3 \n\t"
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t" 
+            // for {second 8 elements } we have to add 8 to each so that it became {from 8 to 16}
+            "vaddudm     2,2,4  \n\t" // vs34=vs34 + vs36{8,8} 
+          
+            //load next 64
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t" 
+
+            //choose bigger from first and second part
+            "xvcmpgtdp 4,5 , 0    \n\t" 
+            "xxsel     3, 0,5,4  \n\t"
+            "xxsel     33,32,34,4  \n\t"
+
+            //load next 64
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"
+
+            "vaddudm  1,1,5  \n\t"  //  get real index for first bigger  
+
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"
+
+ 
+
+            //compare with previous to get vec_max_index(v6 | vs38 ) and vec_max_value (vs39)   
+            "xvcmpgtdp 2, 3,39    \n\t"
+            "xxsel     39,39,3,2  \n\t"
+            "xxsel     38,38,33,2  \n\t"
+    
+            //update index += 8 
+            "vaddudm   5,5,4       \n\t"  
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t"
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+
+            //update index += 8 
+            "vaddudm   5,5,4       \n\t"  
+
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+            //decrement n
+            "addic.    %[n], %[n], -32 \n\t"
+           
+            //Loop back if >0
+            "bgt+ 1b  \n\t"
+
+//==============================================================================
+
+           "xvcmpgtdp  2,45,44  \n\t "
+            "xvcmpgtdp  3,47,46  \n\t "
+            "xvcmpgtdp  4,49,48  \n\t "
+            "xvcmpgtdp  5,51,50  \n\t"
+
+            "xxsel    32,40,41,2 \n\t"
+            "xxsel     0,44,45,2 \n\t" 
+            "xxsel    33,42,43,3 \n\t"
+            "xxsel     1,46,47,3 \n\t" 
+            "xxsel    34,40,41,4 \n\t"
+            "xxsel    45,48,49,4 \n\t"
+            "xxsel    35,42,43,5 \n\t"
+            "xxsel    47,50,51,5 \n\t"
+
+            "xvcmpgtdp 2, 1,0     \n\t"
+            "xvcmpgtdp 3,47, 45   \n\t"
+  
+    
+            "xxsel     32,32,33,2 \n\t"
+            "xxsel       0 ,0,1,2 \n\t"
+            "xxsel     34,34,35,3 \n\t"
+            "xxsel     5,45,47,3 \n\t" 
+    
+            // for {second 8 elements } we have to add 8 to each so that it became {from 8 to 16}
+            "vaddudm     2,2,4  \n\t" // vs34=vs34 + vs36{8,8}  
+            //choose bigger from first and second part
+            "xvcmpgtdp 4,5 , 0    \n\t" 
+            "xxsel     3, 0,5,4   \n\t"
+            "xxsel     33,32,34,4 \n\t" 
+
+            "vaddudm  1,1,5  \n\t"  //  get real index for first bigger   
+
+            //compare with previous to get vec_max_index(v6 | vs38 ) and vec_max_value (vs39)   
+            "xvcmpgtdp 2, 3,39    \n\t"
+            "xxsel     39,39,3,2  \n\t"
+            "xxsel     38,38,33,2  \n\t"
+
+            ///////extract max value and max index from vector
+
+            "xxspltd   32,38,1     \n\t"
+            "xxspltd   40,39,1     \n\t"
+            "xvcmpeqdp.  2, 40,39  \n\t"
+    
+            //cr6 0 bit set if all true, cr6=4*6+bit_ind=24,0011at CR(BI)==1, at=10 hint that it occurs rarely
+             //0b001110=14
+            "bc 14,24, 3f  \n\t" 
+            "xvcmpgtdp  4, 40,39  \n\t"
+            "xxsel    0,39,40,4           \n\t"
+            "xxsel    1,38,32,4  \n\t"
+            "stxsdx    0,0,%[ptr_maxf]     \n\t" 
+            "b 4f    \n\t"
+
+            "3:      \n\t" 
+                //if elements value are equal then choose minimum index
+            "xxspltd  0,40,0          \n\t"
+            "vminud   0,0,6    \n\t"  //vs32 vs38
+            "xxlor 1,32,32     \n\t"
+            "stxsdx   0,0,%[ptr_maxf]  \n\t"
+          
+
+            "4:      \n\t"
+            "mfvsrd   %[index],1 \n\t"
+
+            : [maxf] "=m"(*maxf),[ptr_tmp] "+&b"(x),[index] "=r"(index), [n] "+&r"(n)
+            : [mem] "m"(*(const double (*)[n])x), [ptr_x] "b"(x), [ptr_maxf] "b"(maxf) ,
+            [i16] "b"(16), [i32] "b"(32), [i48] "b"(48),
+            [i64] "b"(64), [i80] "b"(80), [i96] "b"(96), [i112] "b"(112),
+            [start] "v"(start),  [adder] "v"(temp_add_index)
+            : "cc", "vs0", "vs1","vs2","vs3", "vs4","vs5","vs32", "vs33", "vs34", "vs35", "vs36",
+            "vs37", "vs38", "vs39", "vs40", "vs41", "vs42", "vs43", "vs44", "vs45", "vs46", "vs47", "vs48", "vs49", "vs50", "vs51"
+            );
+
+ 
+    return index;
+
+}
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
+    BLASLONG i = 0;
+    BLASLONG j = 0;
+    FLOAT maxf = 0.0;
+    BLASLONG max = 0;
+
+    if (n <= 0 || inc_x <= 0) return (max);
+
+    if (inc_x == 1) {
+
+        BLASLONG n1 = n & -32;
+        if (n1 > 0) {
+
+            max = diamax_kernel_32(n1, x, &maxf);
+
+            i = n1;
+        }
+
+        while (i < n) {
+            if (ABS(x[i]) > maxf) {
+                max = i;
+                maxf = ABS(x[i]);
+            }
+            i++;
+        }
+        return (max + 1);
+
+    } else {
+
+        BLASLONG n1 = n & -4;
+        while (j < n1) {
+
+            if (ABS(x[i]) > maxf) {
+                max = j;
+                maxf = ABS(x[i]);
+            }
+            if (ABS(x[i + inc_x]) > maxf) {
+                max = j + 1;
+                maxf = ABS(x[i + inc_x]);
+            }
+            if (ABS(x[i + 2 * inc_x]) > maxf) {
+                max = j + 2;
+                maxf = ABS(x[i + 2 * inc_x]);
+            }
+            if (ABS(x[i + 3 * inc_x]) > maxf) {
+                max = j + 3;
+                maxf = ABS(x[i + 3 * inc_x]);
+            }
+
+            i += inc_x * 4;
+
+            j += 4;
+
+        }
+
+
+        while (j < n) {
+            if (ABS(x[i]) > maxf) {
+                max = j;
+                maxf = ABS(x[i]);
+            }
+            i += inc_x;
+            j++;
+        }
+        return (max + 1);
+    }
+}

kernel/power/idamin.c

@@ -0,0 +1,384 @@
+/***************************************************************************
+Copyright (c) 2013-2018, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *****************************************************************************/
+#include "common.h"
+#include <math.h>
+
+#if defined(DOUBLE)
+
+#define ABS fabs
+
+#else
+
+#define ABS fabsf
+
+#endif
+
+/**
+ * Find  minimum index 
+ * Warning: requirements n>0  and n % 32 == 0
+ * @param n     
+ * @param x     pointer to the vector
+ * @param minf  (out) minimum absolute value .( only for output )
+ * @return minimum index 
+ */
+static BLASLONG diamin_kernel_32(BLASLONG n, FLOAT *x, FLOAT *minf) {
+    BLASLONG index;
+    register __vector long long start = {1,0};
+    register __vector long long temp_add_index = {2, 2}; 
+    __asm__(
+ 
+
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t" 
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"
+
+            "xxlor  40,%x[start],%x[start]  \n\t" //{ 1,0} vs40 | v8 
+            "vaddudm  9,8, %[adder]   \n\t" //{3,2} vs41 
+            "xxlxor  37,37 ,37        \n\t" //v5 v37 index_count
+            "vaddudm  10,9,%[adder]   \n\t" //{5,4} vs42
+            "xxlxor  38 ,38 ,38       \n\t"  // v6 | vs38 vec_min_index
+            "vaddudm  11,10,%[adder]  \n\t" //{7,6} vs43
+            "lxvdsx  39,0,%[ptr_minf] \n\t" //   vs39 vec_min_value
+            "vaddudm  4,11, %[adder]  \n\t" // {9,8} -{8;8} vs36 | v4
+            "xxspltd   36,36,0        \n\t"
+            "xvabsdp  39, 39 \n\t"             
+    
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t"
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"                         
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+  
+            //jump first half forward 
+            "b 2f  \n\t"
+
+//===================================================================
+
+            ".p2align   5            \n\t"
+
+            "1: \n\t"
+            "xvcmpgedp  2,44,45  \n\t "
+            "xvcmpgedp  3,46,47  \n\t "
+            "xvcmpgedp  4,48,49  \n\t "
+            "xvcmpgedp  5,50,51  \n\t"
+
+            "xxsel    32,40,41,2 \n\t"
+            "xxsel     0,44,45,2 \n\t" 
+            "xxsel    33,42,43,3 \n\t"
+            "xxsel     1,46,47,3 \n\t" 
+            "xxsel    34,40,41,4 \n\t"
+            "xxsel    45,48,49,4 \n\t"
+            "xxsel    35,42,43,5 \n\t"
+            "xxsel    47,50,51,5 \n\t"
+
+            "xvcmpgedp 2,0, 1     \n\t"
+            "xvcmpgedp 3, 45,47   \n\t"
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t" 
+    
+            "xxsel     32,32,33,2 \n\t"
+            "xxsel       0 ,0,1,2 \n\t"
+            "xxsel     34,34,35,3 \n\t"
+            "xxsel     5,45,47,3 \n\t"
+ 
+          
+            //load next 64
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+    
+            // for {second 8 elements } we have to add 8 to each so that it became {from 8 to 16}
+            "vaddudm     2,2,4  \n\t" // vs34=vs34 + vs36{8,8} 
+    
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t" 
+
+            //choose smaller from first and second part
+            "xvcmpgedp 4, 0,5     \n\t" 
+            "xxsel     3, 0,5,4  \n\t"
+            "xxsel     33,32,34,4  \n\t"
+
+            //load next 64
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"
+
+            "vaddudm  1,1,5  \n\t"  //  get real index for first smaller  
+
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"
+
+            //compare with previous to get vec_min_index(v6 | vs38 ) and vec_min_value (vs39)   
+            "xvcmpgedp 2,39, 3    \n\t"
+            "xxsel     39,39,3,2  \n\t"
+            "xxsel     38,38,33,2  \n\t"
+    
+            //update index += 8 
+            "vaddudm   5,5,4       \n\t"  
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t"
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+
+            //update index += 8 
+            "vaddudm   5,5,4       \n\t"  
+
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+//<-----------jump here from first load
+             "2:                  \n\t"
+    
+            "xvcmpgedp  2,44,45  \n\t "
+            "xvcmpgedp  3,46,47  \n\t "
+            "xvcmpgedp  4,48,49  \n\t "
+            "xvcmpgedp  5,50,51  \n\t"
+
+            "xxsel    32,40,41,2 \n\t"
+            "xxsel     0,44,45,2 \n\t" 
+            "xxsel    33,42,43,3 \n\t"
+            "xxsel     1,46,47,3 \n\t" 
+            "xxsel    34,40,41,4 \n\t"
+            "xxsel    45,48,49,4 \n\t"
+            "xxsel    35,42,43,5 \n\t"
+            "xxsel    47,50,51,5 \n\t"
+
+            "xvcmpgedp 2,0, 1     \n\t"
+            "xvcmpgedp 3, 45,47   \n\t"
+            "xxsel     32,32,33,2 \n\t"
+            "xxsel       0 ,0,1,2 \n\t"
+            "xxsel     34,34,35,3 \n\t"
+            "xxsel     5,45,47,3 \n\t"
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t" 
+            // for {second 8 elements } we have to add 8 to each so that it became {from 8 to 16}
+            "vaddudm     2,2,4  \n\t" // vs34=vs34 + vs36{8,8} 
+          
+            //load next 64
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t" 
+
+            //choose smaller from first and second part
+            "xvcmpgedp 4, 0,5     \n\t" 
+            "xxsel     3, 0,5,4  \n\t"
+            "xxsel     33,32,34,4  \n\t"
+
+            //load next 64
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"
+
+            "vaddudm  1,1,5  \n\t"  //  get real index for first smaller  
+
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"
+
+ 
+
+            //compare with previous to get vec_min_index(v6 | vs38 ) and vec_min_value (vs39)   
+            "xvcmpgedp 2,39, 3    \n\t"
+            "xxsel     39,39,3,2  \n\t"
+            "xxsel     38,38,33,2  \n\t"
+    
+            //update index += 8 
+            "vaddudm   5,5,4       \n\t"  
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t"
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+
+            //update index += 8 
+            "vaddudm   5,5,4       \n\t"  
+
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+            //decrement n
+            "addic.    %[n], %[n], -32 \n\t"
+           
+            //Loop back if >0
+            "bgt+ 1b  \n\t"
+
+//==============================================================================
+
+            "xvcmpgedp  2,44,45  \n\t "
+            "xvcmpgedp  3,46,47  \n\t "
+            "xvcmpgedp  4,48,49  \n\t "
+            "xvcmpgedp  5,50,51  \n\t"
+
+            "xxsel    32,40,41,2 \n\t"
+            "xxsel     0,44,45,2 \n\t" 
+            "xxsel    33,42,43,3 \n\t"
+            "xxsel     1,46,47,3 \n\t" 
+            "xxsel    34,40,41,4 \n\t"
+            "xxsel    45,48,49,4 \n\t"
+            "xxsel    35,42,43,5 \n\t"
+            "xxsel    47,50,51,5 \n\t"
+
+            "xvcmpgedp 2,0, 1     \n\t"
+            "xvcmpgedp 3, 45,47   \n\t"
+  
+    
+            "xxsel     32,32,33,2 \n\t"
+            "xxsel       0 ,0,1,2 \n\t"
+            "xxsel     34,34,35,3 \n\t"
+            "xxsel     5,45,47,3 \n\t" 
+    
+            // for {second 8 elements } we have to add 8 to each so that it became {from 8 to 16}
+            "vaddudm     2,2,4  \n\t" // vs34=vs34 + vs36{8,8}  
+            //choose smaller from first and second part
+            "xvcmpgedp 4, 0,5     \n\t" 
+            "xxsel     3, 0,5,4   \n\t"
+            "xxsel     33,32,34,4 \n\t" 
+
+            "vaddudm  1,1,5  \n\t"  //  get real index for first smaller   
+
+            //compare with previous to get vec_min_index(v6 | vs38 ) and vec_min_value (vs39)   
+            "xvcmpgedp 2,39, 3    \n\t"
+            "xxsel     39,39,3,2  \n\t"
+            "xxsel     38,38,33,2  \n\t"
+
+            ///////extract min value and min index from vector
+
+            "xxspltd   32,38,1     \n\t"
+            "xxspltd   40,39,1     \n\t"
+            "xvcmpeqdp.  2, 40,39  \n\t"
+    
+            //cr6 0 bit set if all true, cr6=4*6+bit_ind=24,0011at CR(BI)==1, at=10 hint that it occurs rarely
+             //0b001110=14
+            "bc 14,24, 3f  \n\t" 
+            "xvcmpgedp  4,39, 40  \n\t"
+            "xxsel    0,39,40,4           \n\t"
+            "xxsel    1,38,32,4  \n\t"
+            "stxsdx    0,0,%[ptr_minf]     \n\t" 
+            "b 4f    \n\t"
+
+            "3:      \n\t" 
+                //if elements value are equal then choose minimum index
+            "xxspltd  0,40,0          \n\t"
+            "vminud   0,0,6    \n\t"  //vs32 vs38
+            "xxlor 1,32,32     \n\t"
+            "stxsdx   0,0,%[ptr_minf]  \n\t"
+          
+
+            "4:      \n\t"
+            "mfvsrd   %[index],1 \n\t"
+
+            : [minf] "=m"(*minf),[ptr_tmp] "+&b"(x),[index] "=r"(index), [n] "+&r"(n)
+            : [mem] "m"(*(const double (*)[n])x), [ptr_x] "b"(x), [ptr_minf] "b"(minf) ,
+            [i16] "b"(16), [i32] "b"(32), [i48] "b"(48),
+            [i64] "b"(64), [i80] "b"(80), [i96] "b"(96), [i112] "b"(112),
+            [start] "v"(start),  [adder] "v"(temp_add_index)
+            : "cc", "vs0", "vs1","vs2","vs3", "vs4","vs5","vs32", "vs33", "vs34", "vs35", "vs36",
+            "vs37", "vs38", "vs39", "vs40", "vs41", "vs42", "vs43", "vs44", "vs45", "vs46", "vs47", "vs48", "vs49", "vs50", "vs51"
+            );
+ 
+    return index;
+
+}
+
+
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
+    BLASLONG i = 0;
+    BLASLONG j = 0; 
+    BLASLONG min = 0;
+    FLOAT minf = 0.0;
+    
+    if (n <= 0 || inc_x <= 0) return (min);
+    minf = ABS(x[0]); //index's not incremented
+    if (inc_x == 1) {
+
+        BLASLONG n1 = n & -32;
+        if (n1 > 0) {
+
+            min = diamin_kernel_32(n1, x, &minf);
+            i = n1;
+        }
+
+        while (i < n) {
+            if (ABS(x[i]) < minf) {
+                min = i;
+                minf = ABS(x[i]);
+            }
+            i++;
+        }
+        return (min + 1);
+
+    } else {
+
+        BLASLONG n1 = n & -4;
+        while (j < n1) {
+
+            if (ABS(x[i]) < minf) {
+                min = j;
+                minf = ABS(x[i]);
+            }
+            if (ABS(x[i + inc_x]) < minf) {
+                min = j + 1;
+                minf = ABS(x[i + inc_x]);
+            }
+            if (ABS(x[i + 2 * inc_x]) < minf) {
+                min = j + 2;
+                minf = ABS(x[i + 2 * inc_x]);
+            }
+            if (ABS(x[i + 3 * inc_x]) < minf) {
+                min = j + 3;
+                minf = ABS(x[i + 3 * inc_x]);
+            }
+
+            i += inc_x * 4;
+
+            j += 4;
+
+        }
+
+
+        while (j < n) {
+            if (ABS(x[i]) < minf) {
+                min = j;
+                minf = ABS(x[i]);
+            }
+            i += inc_x;
+            j++;
+        }
+        return (min + 1);
+    }
+}

kernel/power/izamax.c

@@ -0,0 +1,362 @@
+/***************************************************************************
+Copyright (c) 2017, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+ 
+
+#include "common.h"
+#include <math.h>
+ 
+#define ABS fabs 
+#define CABS1(x,i)    ABS(x[i])+ABS(x[i+1])
+
+
+
+ 
+/**
+ * Find  maximum index 
+ * Warning: requirements n>0  and n % 16 == 0
+ * @param n     
+ * @param x     pointer to the vector
+ * @param maxf  (out) maximum absolute value .( only for output )
+ * @return  index 
+ */
+static BLASLONG   ziamax_kernel_16(BLASLONG n, FLOAT *x, FLOAT *maxf) { 
+
+    BLASLONG index;
+    register __vector long long start = {1,0};
+    register __vector long long temp_add_index = {2, 2}; 
+   __asm__(
+
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t"
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"  
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"      
+
+            "xxlor  40,%x[start],%x[start]  \n\t" //{ 1,0} vs40 | v8 
+            "vaddudm  9,8,%[adder]   \n\t" //{3,2} vs41 
+            "xxlxor  37,37 ,37       \n\t" //v5 v37 index_count
+            "vaddudm  10,9,%[adder]  \n\t" //{5,4} vs42
+            "xxlxor  38 ,38 ,38      \n\t"  // v6 | vs38 vec_max_index
+            "vaddudm  11,10,%[adder] \n\t" //{7,6} vs43
+            "xxlxor  39,39,39        \n\t" //   vs39 vec_max_value is zero
+            "vaddudm 4,11,  %[adder] \n\t" // {9,8} -{8;8} vs36 | v4
+            "xxspltd   36,36,0       \n\t"
+
+
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t" 
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"   
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+            //jump first half forward 
+            "b 2f  \n\t"
+
+            ".p2align   5            \n\t"
+            "1: \n\t"
+
+ 
+            "xxmrghd    0,44,45  \n\t"
+            "xxmrgld    1,44,45  \n\t"
+            "xxmrghd    2,46,47  \n\t"
+            "xxmrgld    3,46,47  \n\t"
+            "xxmrghd    4,48,49  \n\t"
+            "xxmrgld    5,48,49  \n\t"
+            "xxmrghd    44,50,51  \n\t"
+            "xxmrgld    45,50,51  \n\t"
+
+            "xvadddp    46,  0,1 \n\t"
+            "xvadddp    47,  2,3 \n\t" 
+            "xvadddp    48,  4,5 \n\t"
+            "xvadddp    49,  44,45 \n\t"
+
+
+
+            "xvcmpgtdp  50,47,46  \n\t "
+            "xvcmpgtdp  51,49,48  \n\t "
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t"   
+
+            "xxsel    32,40,41,50 \n\t"
+            "xxsel     0,46,47,50 \n\t" 
+            "xxsel    33,42,43,51 \n\t"
+            "xxsel     1,48,49,51 \n\t"  
+
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+
+            "xvcmpgtdp  2,1,0  \n\t "             
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t"
+
+
+            "xxsel    32,32,33,2 \n\t" 
+            "xxsel    3,0,1,2 \n\t" 
+     
+            "vaddudm   0,0,5       \n\t"
+
+             //cmp with previous
+
+            "xvcmpgtdp 4,3,39     \n\t "  
+            "vaddudm   5,5,4      \n\t"     
+
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"  
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"        
+             //select with previous 
+            "xxsel     38,38,32,4 \n\t" 
+            "xxsel    39,39,3,4    \n\t" 
+
+
+
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t" 
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"   
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+
+//>>///////////////////////////////  half start
+            "2:   \n\t"
+            "xxmrghd    0,44,45  \n\t"
+            "xxmrgld    1,44,45  \n\t"
+            "xxmrghd    2,46,47  \n\t"
+            "xxmrgld    3,46,47  \n\t"
+            "xxmrghd    4,48,49  \n\t"
+            "xxmrgld    5,48,49  \n\t"
+            "xxmrghd    44,50,51  \n\t"
+            "xxmrgld    45,50,51  \n\t"
+
+            "xvadddp    46,  0,1 \n\t"
+            "xvadddp    47,  2,3 \n\t" 
+            "xvadddp    48,  4,5 \n\t"
+            "xvadddp    49,  44,45 \n\t"
+
+            "xvcmpgtdp  50,47,46  \n\t "
+            "xvcmpgtdp  51,49,48  \n\t "
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t"   
+
+            "xxsel    32,40,41,50 \n\t"
+            "xxsel     0,46,47,50 \n\t" 
+            "xxsel    33,42,43,51 \n\t"
+            "xxsel     1,48,49,51 \n\t"  
+
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+
+            "xvcmpgtdp  2,1,0  \n\t "             
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t"
+
+
+            "xxsel    32,32,33,2 \n\t" 
+            "xxsel    3,0,1,2 \n\t" 
+     
+            "vaddudm   0,0,5       \n\t"
+
+             //cmp with previous
+
+            "xvcmpgtdp 4,3,39     \n\t "  
+            "vaddudm   5,5,4      \n\t"     
+
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"  
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"        
+             //select with previous 
+            "xxsel     38,38,32,4 \n\t" 
+            "xxsel    39,39,3,4    \n\t" 
+ 
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t" 
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"   
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+
+            //decrement n
+            "addic.    %[n], %[n], -16 \n\t"
+            //Loop back if >0
+            "bgt+ 1b  \n\t"
+
+
+            "xxmrghd    0,44,45  \n\t"
+            "xxmrgld    1,44,45  \n\t"
+            "xxmrghd    2,46,47  \n\t"
+            "xxmrgld    3,46,47  \n\t"
+            "xxmrghd    4,48,49  \n\t"
+            "xxmrgld    5,48,49  \n\t"
+            "xxmrghd    44,50,51  \n\t"
+            "xxmrgld    45,50,51  \n\t"
+
+            "xvadddp    46,  0,1 \n\t"
+            "xvadddp    47,  2,3 \n\t" 
+            "xvadddp    48,  4,5 \n\t"
+            "xvadddp    49,  44,45 \n\t"
+
+
+
+            "xvcmpgtdp  50,47,46  \n\t "
+            "xvcmpgtdp  51,49,48  \n\t "
+
+            "xxsel    32,40,41,50 \n\t"
+            "xxsel     0,46,47,50 \n\t" 
+            "xxsel    33,42,43,51 \n\t"
+            "xxsel     1,48,49,51 \n\t"  
+
+            "xvcmpgtdp  2,1,0  \n\t " 
+            "xxsel    32,32,33,2 \n\t" 
+            "xxsel    3,0,1,2 \n\t" 
+     
+            "vaddudm   0,0,5       \n\t"
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t"            
+             //cmp with previous
+
+            "xvcmpgtdp 4,3,39     \n\t "  
+            "vaddudm   5,5,4      \n\t"     
+            "xxsel     38,38,32,4 \n\t" 
+            "xxsel    39,39,3,4    \n\t" 
+
+
+            ///////extract max value and max index from vector
+
+            "xxspltd   32,38,1     \n\t"
+            "xxspltd   40,39,1     \n\t"
+            "xvcmpeqdp.  2, 40,39  \n\t"
+    
+            //cr6 0 bit set if all true, cr6=4*6+bit_ind=24,0011at CR(BI)==1, at=10 hint that it occurs rarely
+             //0b001110=14
+            "bc 14,24, 3f  \n\t" 
+            "xvcmpgtdp  4, 40,39  \n\t"
+            "xxsel    0,39,40,4           \n\t"
+            "xxsel    1,38,32,4  \n\t"
+            "stxsdx    0,0,%[ptr_maxf]     \n\t" 
+            "b 4f    \n\t"
+
+            "3:      \n\t" 
+                //if elements value are equal then choose minimum index
+            "xxspltd  0,40,0          \n\t"
+            "vminud   0,0,6    \n\t"  //vs32 vs38
+            "xxlor 1,32,32     \n\t"
+            "stxsdx   0,0,%[ptr_maxf]  \n\t"
+          
+
+            "4:      \n\t"
+            "mfvsrd   %[index],1 \n\t"
+
+            : [maxf] "=m"(*maxf),[ptr_tmp] "+&b"(x),[index] "=r"(index), [n] "+&r"(n)
+            : [mem] "m"(*(const double (*)[2*n])x), [ptr_x] "b"(x), [ptr_maxf] "b"(maxf) ,
+            [i16] "b"(16), [i32] "b"(32), [i48] "b"(48),
+            [i64] "b"(64), [i80] "b"(80), [i96] "b"(96), [i112] "b"(112),
+            [start] "v"(start),  [adder] "v"(temp_add_index)
+            : "cc", "vs0", "vs1","vs2","vs3", "vs4","vs5","vs32", "vs33", "vs34", "vs35", "vs36",
+            "vs37", "vs38", "vs39", "vs40", "vs41", "vs42", "vs43", "vs44", "vs45", "vs46", "vs47", "vs48", "vs49", "vs50", "vs51"
+            );
+ 
+    return index;
+
+}
+ 
+  
+
+ 
+ 
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    BLASLONG i = 0;
+    BLASLONG ix = 0;
+    FLOAT maxf = 0;
+    BLASLONG max = 0;
+    BLASLONG inc_x2;
+
+    if (n <= 0 || inc_x <= 0) return(max);
+     
+    if (inc_x == 1) {
+
+      BLASLONG n1 = n & -16;
+      if (n1 > 0) {
+
+            max = ziamax_kernel_16(n1, x, &maxf); 
+            i = n1;
+            ix = n1 << 1;
+      }
+
+      while(i < n)
+    {
+        if( CABS1(x,ix) > maxf )
+        {
+            max = i;
+            maxf = CABS1(x,ix);
+        }
+        ix += 2;
+        i++;
+    }
+        return (max + 1);
+
+    } else {
+ 
+      inc_x2 = 2 * inc_x;
+
+    maxf = CABS1(x,0);
+    ix += inc_x2;
+    i++;
+
+    while(i < n)
+    {
+        if( CABS1(x,ix) > maxf )
+        {
+            max = i;
+            maxf = CABS1(x,ix);
+        }
+        ix += inc_x2;
+        i++;
+    }
+        return (max + 1);
+    }
+ 
+}
+
+

kernel/power/izamin.c

@@ -0,0 +1,361 @@
+/***************************************************************************
+Copyright (c) 2017, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+ 
+
+#include "common.h"
+#include <math.h>
+ 
+#define ABS fabs 
+#define CABS1(x,i)    ABS(x[i])+ABS(x[i+1])
+
+ 
+/**
+ * Find  minimum index 
+ * Warning: requirements n>0  and n % 16 == 0
+ * @param n     
+ * @param x     pointer to the vector
+ * @param minf  (out) minimum absolute value .( only for output )
+ * @return minimum index 
+ */
+static BLASLONG ziamin_kernel_16_TUNED(BLASLONG n, FLOAT *x, FLOAT *minf) { 
+
+    BLASLONG index;
+    register __vector long long start = {1,0};
+    register __vector long long temp_add_index = {2, 2}; 
+   __asm__(
+
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t"
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"  
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"      
+
+            "xxlor  40,%x[start],%x[start]  \n\t" //{ 1,0} vs40 | v8 
+            "vaddudm  9,8,%[adder]   \n\t" //{3,2} vs41 
+            "xxlxor  37,37 ,37       \n\t" //v5 v37 index_count
+            "vaddudm  10,9,%[adder]  \n\t" //{5,4} vs42
+            "xxlxor  38 ,38 ,38      \n\t"  // v6 | vs38 vec_min_index
+            "vaddudm  11,10,%[adder] \n\t" //{7,6} vs43
+            "lxvdsx  39,0,%[ptr_minf] \n\t" //   vs39 vec_min_value
+            "vaddudm 4,11,  %[adder] \n\t" // {9,8} -{8;8} vs36 | v4
+            "xxspltd   36,36,0       \n\t"
+
+
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t" 
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"   
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+            //jump first half forward 
+            "b 2f  \n\t"
+
+            ".p2align   5            \n\t"
+            "1: \n\t"
+
+ 
+            "xxmrghd    0,44,45  \n\t"
+            "xxmrgld    1,44,45  \n\t"
+            "xxmrghd    2,46,47  \n\t"
+            "xxmrgld    3,46,47  \n\t"
+            "xxmrghd    4,48,49  \n\t"
+            "xxmrgld    5,48,49  \n\t"
+            "xxmrghd    44,50,51  \n\t"
+            "xxmrgld    45,50,51  \n\t"
+
+            "xvadddp    46,  0,1 \n\t"
+            "xvadddp    47,  2,3 \n\t" 
+            "xvadddp    48,  4,5 \n\t"
+            "xvadddp    49,  44,45 \n\t"
+
+
+
+            "xvcmpgedp  50,46,47  \n\t "
+            "xvcmpgedp  51,48,49  \n\t "
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t"   
+
+            "xxsel    32,40,41,50 \n\t"
+            "xxsel     0,46,47,50 \n\t" 
+            "xxsel    33,42,43,51 \n\t"
+            "xxsel     1,48,49,51 \n\t"  
+
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+
+            "xvcmpgedp  2,0,1  \n\t "             
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t"
+
+
+            "xxsel    32,32,33,2 \n\t" 
+            "xxsel    3,0,1,2 \n\t" 
+     
+            "vaddudm   0,0,5       \n\t"
+
+             //cmp with previous
+
+            "xvcmpgedp 4,39,3     \n\t "  
+            "vaddudm   5,5,4      \n\t"     
+
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"  
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"        
+             //select with previous 
+            "xxsel     38,38,32,4 \n\t" 
+            "xxsel    39,39,3,4    \n\t" 
+
+
+
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t" 
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"   
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+
+//>>///////////////////////////////  half start
+            "2:   \n\t"
+            "xxmrghd    0,44,45  \n\t"
+            "xxmrgld    1,44,45  \n\t"
+            "xxmrghd    2,46,47  \n\t"
+            "xxmrgld    3,46,47  \n\t"
+            "xxmrghd    4,48,49  \n\t"
+            "xxmrgld    5,48,49  \n\t"
+            "xxmrghd    44,50,51  \n\t"
+            "xxmrgld    45,50,51  \n\t"
+
+            "xvadddp    46,  0,1 \n\t"
+            "xvadddp    47,  2,3 \n\t" 
+            "xvadddp    48,  4,5 \n\t"
+            "xvadddp    49,  44,45 \n\t"
+
+            "xvcmpgedp  50,46,47  \n\t "
+            "xvcmpgedp  51,48,49  \n\t "
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t"   
+
+            "xxsel    32,40,41,50 \n\t"
+            "xxsel     0,46,47,50 \n\t" 
+            "xxsel    33,42,43,51 \n\t"
+            "xxsel     1,48,49,51 \n\t"  
+
+            "lxvd2x  44,      0,%[ptr_tmp] \n\t"
+            "lxvd2x  45, %[i16],%[ptr_tmp] \n\t"
+
+            "xvcmpgedp  2,0,1  \n\t "             
+            "lxvd2x  46, %[i32],%[ptr_tmp] \n\t"
+            "lxvd2x  47, %[i48],%[ptr_tmp] \n\t"
+
+
+            "xxsel    32,32,33,2 \n\t" 
+            "xxsel    3,0,1,2 \n\t" 
+     
+            "vaddudm   0,0,5       \n\t"
+
+             //cmp with previous
+
+            "xvcmpgedp 4,39,3     \n\t "  
+            "vaddudm   5,5,4      \n\t"     
+
+            "lxvd2x  48, %[i64],%[ptr_tmp] \n\t"
+            "lxvd2x  49, %[i80],%[ptr_tmp] \n\t"  
+            "lxvd2x  50, %[i96],%[ptr_tmp] \n\t"
+            "lxvd2x  51,%[i112],%[ptr_tmp] \n\t"        
+             //select with previous 
+            "xxsel     38,38,32,4 \n\t" 
+            "xxsel    39,39,3,4    \n\t" 
+ 
+
+            "xvabsdp  44, 44 \n\t"
+            "xvabsdp  45, 45 \n\t" 
+            "xvabsdp  46, 46 \n\t"
+            "xvabsdp  47, 47 \n\t"   
+            "xvabsdp  48, 48 \n\t"
+            "xvabsdp  49, 49 \n\t"   
+            "xvabsdp  50, 50 \n\t"
+            "xvabsdp  51, 51 \n\t"
+
+
+            //decrement n
+            "addic.    %[n], %[n], -16 \n\t"
+            //Loop back if >0
+            "bgt+ 1b  \n\t"
+
+
+            "xxmrghd    0,44,45  \n\t"
+            "xxmrgld    1,44,45  \n\t"
+            "xxmrghd    2,46,47  \n\t"
+            "xxmrgld    3,46,47  \n\t"
+            "xxmrghd    4,48,49  \n\t"
+            "xxmrgld    5,48,49  \n\t"
+            "xxmrghd    44,50,51  \n\t"
+            "xxmrgld    45,50,51  \n\t"
+
+            "xvadddp    46,  0,1 \n\t"
+            "xvadddp    47,  2,3 \n\t" 
+            "xvadddp    48,  4,5 \n\t"
+            "xvadddp    49,  44,45 \n\t"
+
+
+
+            "xvcmpgedp  50,46,47  \n\t "
+            "xvcmpgedp  51,48,49  \n\t "
+
+            "xxsel    32,40,41,50 \n\t"
+            "xxsel     0,46,47,50 \n\t" 
+            "xxsel    33,42,43,51 \n\t"
+            "xxsel     1,48,49,51 \n\t"  
+
+            "xvcmpgedp  2,0,1  \n\t " 
+            "xxsel    32,32,33,2 \n\t" 
+            "xxsel    3,0,1,2 \n\t" 
+     
+            "vaddudm   0,0,5       \n\t"
+
+            "addi     %[ptr_tmp] ,%[ptr_tmp] , 128 \n\t"            
+             //cmp with previous
+
+            "xvcmpgedp 4,39,3     \n\t "  
+            "vaddudm   5,5,4      \n\t"     
+            "xxsel     38,38,32,4 \n\t" 
+            "xxsel    39,39,3,4    \n\t" 
+
+
+            ///////extract min value and min index from vector
+
+            "xxspltd   32,38,1     \n\t"
+            "xxspltd   40,39,1     \n\t"
+            "xvcmpeqdp.  2, 40,39  \n\t"
+    
+            //cr6 0 bit set if all true, cr6=4*6+bit_ind=24,0011at CR(BI)==1, at=10 hint that it occurs rarely
+             //0b001110=14
+            "bc 14,24, 3f  \n\t" 
+            "xvcmpgedp  4,39, 40  \n\t"
+            "xxsel    0,39,40,4           \n\t"
+            "xxsel    1,38,32,4  \n\t"
+            "stxsdx    0,0,%[ptr_minf]     \n\t" 
+            "b 4f    \n\t"
+
+            "3:      \n\t" 
+                //if elements value are equal then choose minimum index
+            "xxspltd  0,40,0          \n\t"
+            "vminud   0,0,6    \n\t"  //vs32 vs38
+            "xxlor 1,32,32     \n\t"
+            "stxsdx   0,0,%[ptr_minf]  \n\t"
+          
+
+            "4:      \n\t"
+            "mfvsrd   %[index],1 \n\t"
+
+            : [minf] "=m"(*minf),[ptr_tmp] "+&b"(x),[index] "=r"(index), [n] "+&r"(n)
+            : [mem] "m"(*(const double (*)[2*n])x), [ptr_x] "b"(x), [ptr_minf] "b"(minf) ,
+            [i16] "b"(16), [i32] "b"(32), [i48] "b"(48),
+            [i64] "b"(64), [i80] "b"(80), [i96] "b"(96), [i112] "b"(112),
+            [start] "v"(start),  [adder] "v"(temp_add_index)
+            : "cc", "vs0", "vs1","vs2","vs3", "vs4","vs5","vs32", "vs33", "vs34", "vs35", "vs36",
+            "vs37", "vs38", "vs39", "vs40", "vs41", "vs42", "vs43", "vs44", "vs45", "vs46", "vs47", "vs48", "vs49", "vs50", "vs51"
+            );
+ 
+    return index; 
+}
+
+ 
+
+ 
+ 
+
+BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
+{
+    BLASLONG i=0;
+    BLASLONG ix=0;
+    FLOAT minf;
+    BLASLONG min=0;
+    BLASLONG inc_x2;
+
+    if (n <= 0 || inc_x <= 0) return(min);
+    
+
+    if (inc_x == 1) {
+        minf = CABS1(x,0); //index will not be incremented
+        BLASLONG n1 = n & -16;
+        if (n1 > 0) {
+
+            min = ziamin_kernel_16_TUNED(n1, x, &minf);
+            i = n1;
+            ix = n1 << 1;
+        }
+      
+
+        while(i < n)
+        {
+            if( CABS1(x,ix) < minf )
+            {
+                min = i;
+                minf = CABS1(x,ix);
+            }
+            ix += 2;
+            i++;
+        }
+        return (min + 1);
+
+    } else {
+ 
+        inc_x2 = 2 * inc_x;
+
+        minf = CABS1(x,0);
+        ix += inc_x2;
+        i++;
+
+        while(i < n)
+        {
+            if( CABS1(x,ix) < minf )
+            {
+                min = i;
+                minf = CABS1(x,ix);
+            }
+            ix += inc_x2;
+            i++;
+        }
+        return (min + 1);
+    }
+ 
+}
+
+

kernel/power/sgemm_tcopy_16_power8.S

@@ -110,7 +110,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "sgemm_tcopy_macros_16_power8.S"
 
-#define STACKSIZE 576
+#define STACKSIZE 144
 
 	PROLOGUE
 	PROFCODE
@@ -118,49 +118,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	addi	SP, SP, -STACKSIZE
 	li	r0, 0
 
-        std     r31,  144(SP)
-        std     r30,  152(SP)
-        std     r29,  160(SP)
-        std     r28,  168(SP)
-        std     r27,  176(SP)
-        std     r26,  184(SP)
-        std     r25,  192(SP)
-        std     r24,  200(SP)
-        std     r23,  208(SP)
-        std     r22,  216(SP)
-        std     r21,  224(SP)
-        std     r20,  232(SP)
-        std     r19,  240(SP)
-        std     r18,  248(SP)
-        std     r17,  256(SP)
-        std     r16,  264(SP)
-        std     r15,  272(SP)
-        std     r14,  280(SP)
-        addi    r11 ,SP, 288
-        stvx    v20, r11, r0
-        addi    r11, r11, 16
-        stvx    v21, r11, r0
-        addi    r11, r11, 16
-        stvx    v22, r11, r0
-        addi    r11, r11, 16
-        stvx    v23, r11, r0
-        addi    r11, r11, 16
-        stvx    v24, r11, r0
-        addi    r11, r11, 16
-        stvx    v25, r11, r0
-        addi    r11, r11, 16
-        stvx    v26, r11, r0
-        addi    r11, r11, 16
-        stvx    v27, r11, r0
-        addi    r11, r11, 16
-        stvx    v28, r11, r0
-        addi    r11, r11, 16
-        stvx    v29, r11, r0
-        addi    r11, r11, 16
-        stvx    v30, r11, r0
-        addi    r11, r11, 16
-        stvx    v31, r11, r0
-        li      r11, 0
+        std r14,   0(SP) 
+        std r15,   8(SP) 
+        std r16,  16(SP) 
+        std r17,  24(SP) 
+        std r18,  32(SP) 
+        std r19,  40(SP) 
+        std r20,  48(SP) 
+        std r21,  56(SP) 
+        std r22,  64(SP) 
+        std r23,  72(SP) 
+        std r24,  80(SP) 
+        std r25,  88(SP) 
+        std r26,  96(SP) 
+        std r27, 104(SP) 
+        std r28, 112(SP) 
+        std r29, 120(SP) 
+        std r30, 128(SP) 
+        std r31, 136(SP)  
 
 	cmpwi	cr0, M, 0
 	ble-	L999
@@ -207,51 +182,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 L999:
 
-	li	r3, 0
-
-        ld      r31,  144(SP)
-        ld      r30,  152(SP)
-        ld      r29,  160(SP)
-        ld      r28,  168(SP)
-        ld      r27,  176(SP)
-        ld      r26,  184(SP)
-        ld      r25,  192(SP)
-        ld      r24,  200(SP)
-        ld      r23,  208(SP)
-        ld      r22,  216(SP)
-        ld      r21,  224(SP)
-        ld      r20,  232(SP)
-        ld      r19,  240(SP)
-        ld      r18,  248(SP)
-        ld      r17,  256(SP)
-        ld      r16,  264(SP)
-        ld      r15,  272(SP)
-        ld      r14,  280(SP)
-        addi    r11, SP, 288
-        lvx     v20, r11, r3
-        addi    r11, r11, 16
-        lvx     v21, r11, r3
-        addi    r11, r11, 16
-        lvx     v22, r11, r3
-        addi    r11, r11, 16
-        lvx     v23, r11, r3
-        addi    r11, r11, 16
-        lvx     v24, r11, r3
-        addi    r11, r11, 16
-        lvx     v25, r11, r3
-        addi    r11, r11, 16
-        lvx     v26, r11, r3
-        addi    r11, r11, 16
-        lvx     v27, r11, r3
-        addi    r11, r11, 16
-        lvx     v28, r11, r3
-        addi    r11, r11, 16
-        lvx     v29, r11, r3
-        addi    r11, r11, 16
-        lvx     v30, r11, r3
-        addi    r11, r11, 16
-        lvx     v31, r11, r3
-        li      r11, 0
+        ld r14,   0(SP) 
+        ld r15,   8(SP) 
+        ld r16,  16(SP) 
+        ld r17,  24(SP) 
+        ld r18,  32(SP) 
+        ld r19,  40(SP) 
+        ld r20,  48(SP) 
+        ld r21,  56(SP) 
+        ld r22,  64(SP) 
+        ld r23,  72(SP) 
+        ld r24,  80(SP) 
+        ld r25,  88(SP) 
+        ld r26,  96(SP) 
+        ld r27, 104(SP) 
+        ld r28, 112(SP) 
+        ld r29, 120(SP) 
+        ld r30, 128(SP) 
+        ld r31, 136(SP) 
 
 	addi	SP, SP, STACKSIZE
 	blr

kernel/power/sgemm_tcopy_8_power8.S

@@ -110,57 +110,32 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "sgemm_tcopy_macros_8_power8.S"
 
-#define STACKSIZE 576
+#define STACKSIZE 144
 
 	PROLOGUE
 	PROFCODE
 
 	addi	SP, SP, -STACKSIZE
 	li	r0, 0
+        std r14,   0(SP) 
+        std r15,   8(SP) 
+        std r16,  16(SP) 
+        std r17,  24(SP) 
+        std r18,  32(SP) 
+        std r19,  40(SP) 
+        std r20,  48(SP) 
+        std r21,  56(SP) 
+        std r22,  64(SP) 
+        std r23,  72(SP) 
+        std r24,  80(SP) 
+        std r25,  88(SP) 
+        std r26,  96(SP) 
+        std r27, 104(SP) 
+        std r28, 112(SP) 
+        std r29, 120(SP) 
+        std r30, 128(SP) 
+        std r31, 136(SP) 
 
-        std     r31,  144(SP)
-        std     r30,  152(SP)
-        std     r29,  160(SP)
-        std     r28,  168(SP)
-        std     r27,  176(SP)
-        std     r26,  184(SP)
-        std     r25,  192(SP)
-        std     r24,  200(SP)
-        std     r23,  208(SP)
-        std     r22,  216(SP)
-        std     r21,  224(SP)
-        std     r20,  232(SP)
-        std     r19,  240(SP)
-        std     r18,  248(SP)
-        std     r17,  256(SP)
-        std     r16,  264(SP)
-        std     r15,  272(SP)
-        std     r14,  280(SP)
-        addi    r11, SP, 288
-        stvx    v20, r11, r0
-        addi    r11, r11, 16
-        stvx    v21, r11, r0
-        addi    r11, r11, 16
-        stvx    v22, r11, r0
-        addi    r11, r11, 16
-        stvx    v23, r11, r0
-        addi    r11, r11, 16
-        stvx    v24, r11, r0
-        addi    r11, r11, 16
-        stvx    v25, r11, r0
-        addi    r11, r11, 16
-        stvx    v26, r11, r0
-        addi    r11, r11, 16
-        stvx    v27, r11, r0
-        addi    r11, r11, 16
-        stvx    v28, r11, r0
-        addi    r11, r11, 16
-        stvx    v29, r11, r0
-        addi    r11, r11, 16
-        stvx    v30, r11, r0
-        addi    r11, r11, 16
-        stvx    v31, r11, r0
-        li      r11, 0
 
 	cmpwi	cr0, M, 0
 	ble-	L999
@@ -202,51 +177,24 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 L999:
 
-	li	r3, 0
-
-        ld      r31,  144(SP)
-        ld      r30,  152(SP)
-        ld      r29,  160(SP)
-        ld      r28,  168(SP)
-        ld      r27,  176(SP)
-        ld      r26,  184(SP)
-        ld      r25,  192(SP)
-        ld      r24,  200(SP)
-        ld      r23,  208(SP)
-        ld      r22,  216(SP)
-        ld      r21,  224(SP)
-        ld      r20,  232(SP)
-        ld      r19,  240(SP)
-        ld      r18,  248(SP)
-        ld      r17,  256(SP)
-        ld      r16,  264(SP)
-        ld      r15,  272(SP)
-        ld      r14,  280(SP)
-        addi r11,SP,288
-        lvx    v20, r11, r3
-        addi   r11, r11, 16
-        lvx    v21, r11, r3
-        addi   r11, r11, 16
-        lvx    v22, r11, r3
-        addi   r11, r11, 16
-        lvx    v23, r11, r3
-        addi   r11, r11, 16
-        lvx    v24, r11, r3
-        addi   r11, r11, 16
-        lvx    v25, r11, r3
-        addi   r11, r11, 16
-        lvx    v26, r11, r3
-        addi   r11, r11, 16
-        lvx    v27, r11, r3
-        addi   r11, r11, 16
-        lvx    v28, r11, r3
-        addi   r11, r11, 16
-        lvx    v29, r11, r3
-        addi   r11, r11, 16
-        lvx    v30, r11, r3
-        addi   r11, r11, 16
-        lvx    v31, r11, r3
-        li     r11, 0
+        ld r14,   0(SP) 
+        ld r15,   8(SP) 
+        ld r16,  16(SP) 
+        ld r17,  24(SP) 
+        ld r18,  32(SP) 
+        ld r19,  40(SP) 
+        ld r20,  48(SP) 
+        ld r21,  56(SP) 
+        ld r22,  64(SP) 
+        ld r23,  72(SP) 
+        ld r24,  80(SP) 
+        ld r25,  88(SP) 
+        ld r26,  96(SP) 
+        ld r27, 104(SP) 
+        ld r28, 112(SP) 
+        ld r29, 120(SP) 
+        ld r30, 128(SP) 
+        ld r31, 136(SP) 
 
 	addi	SP, SP, STACKSIZE
 	blr

kernel/power/zgemm_tcopy_8_power8.S

@@ -109,8 +109,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "zgemm_tcopy_macros_8_power8.S"
 
-#define STACKSIZE 384
-#define STACKSIZE 576
+#define STACKSIZE 144
+
 
 
 	PROLOGUE
@@ -119,49 +119,25 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	addi	SP, SP, -STACKSIZE
 	li	r0, 0
 
-        std     r31,  144(SP)
-        std     r30,  152(SP)
-        std     r29,  160(SP)
-        std     r28,  168(SP)
-        std     r27,  176(SP)
-        std     r26,  184(SP)
-        std     r25,  192(SP)
-        std     r24,  200(SP)
-        std     r23,  208(SP)
-        std     r22,  216(SP)
-        std     r21,  224(SP)
-        std     r20,  232(SP)
-        std     r19,  240(SP)
-        std     r18,  248(SP)
-        std     r17,  256(SP)
-        std     r16,  264(SP)
-        std     r15,  272(SP)
-        std     r14,  280(SP)
-        addi    r11, SP ,288
-        stvx    v20, r11, r0
-        addi    r11, r11, 16
-        stvx    v21, r11, r0
-        addi    r11, r11, 16
-        stvx    v22, r11, r0
-        addi    r11, r11, 16
-        stvx    v23, r11, r0
-        addi    r11, r11, 16
-        stvx    v24, r11, r0
-        addi    r11, r11, 16
-        stvx    v25, r11, r0
-        addi    r11, r11, 16
-        stvx    v26, r11, r0
-        addi    r11, r11, 16
-        stvx    v27, r11, r0
-        addi    r11, r11, 16
-        stvx    v28, r11, r0
-        addi    r11, r11, 16
-        stvx    v29, r11, r0
-        addi    r11, r11, 16
-        stvx    v30, r11, r0
-        addi    r11, r11 ,16
-        stvx    v31, r11, r0
-        li r11,0
+        std r14,   0(SP) 
+        std r15,   8(SP) 
+        std r16,  16(SP) 
+        std r17,  24(SP) 
+        std r18,  32(SP) 
+        std r19,  40(SP) 
+        std r20,  48(SP) 
+        std r21,  56(SP) 
+        std r22,  64(SP) 
+        std r23,  72(SP) 
+        std r24,  80(SP) 
+        std r25,  88(SP) 
+        std r26,  96(SP) 
+        std r27, 104(SP) 
+        std r28, 112(SP) 
+        std r29, 120(SP) 
+        std r30, 128(SP) 
+        std r31, 136(SP)  
+
 
 	cmpwi	cr0, M, 0
 	ble-	L999
@@ -204,49 +180,24 @@ L999:
 
 	li	r3, 0
 
-        ld      r31,  144(SP)
-        ld      r30,  152(SP)
-        ld      r29,  160(SP)
-        ld      r28,  168(SP)
-        ld      r27,  176(SP)
-        ld      r26,  184(SP)
-        ld      r25,  192(SP)
-        ld      r24,  200(SP)
-        ld      r23,  208(SP)
-        ld      r22,  216(SP)
-        ld      r21,  224(SP)
-        ld      r20,  232(SP)
-        ld      r19,  240(SP)
-        ld      r18,  248(SP)
-        ld      r17,  256(SP)
-        ld      r16,  264(SP)
-        ld      r15,  272(SP)
-        ld      r14,  280(SP)
-        addi    r11, SP, 288
-        lvx     v20,  r11,r3
-        addi    r11, r11, 16
-        lvx     v21, r11, r3
-        addi    r11, r11, 16
-        lvx     v22, r11, r3
-        addi    r11, r11, 16
-        lvx     v23, r11, r3
-        addi    r11, r11, 16
-        lvx     v24, r11, r3
-        addi    r11, r11, 16
-        lvx     v25, r11, r3
-        addi    r11, r11, 16
-        lvx     v26, r11, r3
-        addi    r11, r11, 16
-        lvx     v27, r11, r3
-        addi    r11, r11, 16
-        lvx     v28, r11, r3
-        addi    r11, r11, 16
-        lvx     v29, r11, r3
-        addi    r11, r11, 16
-        lvx     v30, r11, r3
-        addi    r11, r11, 16
-        lvx     v31, r11, r3
-        li r11,0
+        ld r14,   0(SP) 
+        ld r15,   8(SP) 
+        ld r16,  16(SP) 
+        ld r17,  24(SP) 
+        ld r18,  32(SP) 
+        ld r19,  40(SP) 
+        ld r20,  48(SP) 
+        ld r21,  56(SP) 
+        ld r22,  64(SP) 
+        ld r23,  72(SP) 
+        ld r24,  80(SP) 
+        ld r25,  88(SP) 
+        ld r26,  96(SP) 
+        ld r27, 104(SP) 
+        ld r28, 112(SP) 
+        ld r29, 120(SP) 
+        ld r30, 128(SP) 
+        ld r31, 136(SP) 
 
 	addi	SP, SP, STACKSIZE
 	blr

kernel/power/zgemm_tcopy_macros_8_power8.S

@@ -72,23 +72,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	lxvd2x		vs51,	o48,	A2
 	addi		A2,	A2,	64
 
-	lxvd2x		vs52,	o0,	A2
-	lxvd2x		vs53,	o16,	A2
-	lxvd2x		vs54,	o32,	A2
-	lxvd2x		vs55,	o48,	A2
+	lxvd2x		vs2,	o0,	A2
+	lxvd2x		vs3,	o16,	A2
+	lxvd2x		vs4,	o32,	A2
+	lxvd2x		vs5,	o48,	A2
 	addi		A2,	A2,	64
 
 
-	lxvd2x		vs56,	o0,	A3
-	lxvd2x		vs57,	o16,	A3
-	lxvd2x		vs58,	o32,	A3
-	lxvd2x		vs59,	o48,	A3
+	lxvd2x		vs6,	o0,	A3
+	lxvd2x		vs7,	o16,	A3
+	lxvd2x		vs8,	o32,	A3
+	lxvd2x		vs9,	o48,	A3
 	addi		A3,	A3,	64
 
-	lxvd2x		vs60,	o0,	A3
-	lxvd2x		vs61,	o16,	A3
-	lxvd2x		vs62,	o32,	A3
-	lxvd2x		vs63,	o48,	A3
+	lxvd2x		vs10,	o0,	A3
+	lxvd2x		vs11,	o16,	A3
+	lxvd2x		vs12,	o32,	A3
+	lxvd2x		vs13,	o48,	A3
 	addi		A3,	A3,	64
 
 
@@ -126,23 +126,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	stxvd2x		vs51,	o48,	T1
 	addi		T1,	T1,	64
 
-	stxvd2x		vs52,	o0,	T1
-	stxvd2x		vs53,	o16,	T1
-	stxvd2x		vs54,	o32,	T1
-	stxvd2x		vs55,	o48,	T1
+	stxvd2x		vs2,	o0,	T1
+	stxvd2x		vs3,	o16,	T1
+	stxvd2x		vs4,	o32,	T1
+	stxvd2x		vs5,	o48,	T1
 
 	addi		T1,	T1,	64
 
-	stxvd2x		vs56,	o0,	T1
-	stxvd2x		vs57,	o16,	T1
-	stxvd2x		vs58,	o32,	T1
-	stxvd2x		vs59,	o48,	T1
+	stxvd2x		vs6,	o0,	T1
+	stxvd2x		vs7,	o16,	T1
+	stxvd2x		vs8,	o32,	T1
+	stxvd2x		vs9,	o48,	T1
 	addi		T1,	T1,	64
 
-	stxvd2x		vs60,	o0,	T1
-	stxvd2x		vs61,	o16,	T1
-	stxvd2x		vs62,	o32,	T1
-	stxvd2x		vs63,	o48,	T1
+	stxvd2x		vs10,	o0,	T1
+	stxvd2x		vs11,	o16,	T1
+	stxvd2x		vs12,	o32,	T1
+	stxvd2x		vs13,	o48,	T1
 
 .endm
 

kernel/power/zgemv_n_4.c

@@ -0,0 +1,958 @@
+/***************************************************************************
+Copyright (c) 2018, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *****************************************************************************/
+
+#include <stdlib.h>
+#include <stdio.h>
+#include "common.h"
+
+#define HAVE_KERNEL_4x4_VEC 1
+#define HAVE_KERNEL_4x2_VEC 1
+#define HAVE_KERNEL_4x1_VEC 1
+#define HAVE_KERNEL_ADDY 1
+
+#if defined(HAVE_KERNEL_4x4_VEC) || defined(HAVE_KERNEL_4x2_VEC) || defined(HAVE_KERNEL_4x1_VEC)
+#include <altivec.h> 
+#endif
+
+// 
+#define NBMAX 4096
+
+#ifdef HAVE_KERNEL_4x4_VEC_ASM
+
+#elif HAVE_KERNEL_4x4_VEC
+
+static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y) {
+  
+    FLOAT *a0, *a1, *a2, *a3;
+    a0 = ap;
+    a1 = ap + lda;
+    a2 = a1 + lda;
+    a3 = a2 + lda;
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+    register __vector double vx0_r = {x[0], x[0]};
+    register __vector double vx0_i = {-x[1], x[1]};
+    register __vector double vx1_r = {x[2], x[2]};
+    register __vector double vx1_i = {-x[3], x[3]};
+    register __vector double vx2_r = {x[4], x[4]};
+    register __vector double vx2_i = {-x[5], x[5]};
+    register __vector double vx3_r = {x[6], x[6]};
+    register __vector double vx3_i = {-x[7], x[7]};
+
+#else
+    register __vector double vx0_r = {x[0], -x[0]};
+    register __vector double vx0_i = {x[1], x[1]};
+    register __vector double vx1_r = {x[2], -x[2]};
+    register __vector double vx1_i = {x[3], x[3]};
+    register __vector double vx2_r = {x[4], -x[4]};
+    register __vector double vx2_i = {x[5], x[5]};
+    register __vector double vx3_r = {x[6], -x[6]};
+    register __vector double vx3_i = {x[7], x[7]};
+#endif
+
+    register __vector double *vy = (__vector double *) y;
+    register __vector double *vptr_a0 = (__vector double *) a0;
+    register __vector double *vptr_a1 = (__vector double *) a1;
+    register __vector double *vptr_a2 = (__vector double *) a2;
+    register __vector double *vptr_a3 = (__vector double *) a3;
+ 
+   
+    register __vector double vy_0;
+    register __vector double va0;
+    register __vector double va1;
+    register __vector double va2;
+    register __vector double va3;
+    register __vector double vy_1;
+    register __vector double va0_1;
+    register __vector double va1_1;
+    register __vector double va2_1;
+    register __vector double va3_1;
+    register __vector double vy_2;
+    register __vector double va0_2;
+    register __vector double va1_2;
+    register __vector double va2_2;
+    register __vector double va3_2;
+    register __vector double vy_3;
+    register __vector double va0_3;
+    register __vector double va1_3;
+    register __vector double va2_3;
+    register __vector double va3_3;
+      
+     BLASLONG  i = 0; 
+    while (i < n) {
+
+        vy_0 = vy[i];
+        va0 = vptr_a0[i];
+        va1 = vptr_a1[i];
+        va2 = vptr_a2[i];
+        va3 = vptr_a3[i];
+
+        vy_1 = vy[i + 1];
+        va0_1 = vptr_a0[i + 1];
+        va1_1 = vptr_a1[i + 1];
+        va2_1 = vptr_a2[i + 1];
+        va3_1 = vptr_a3[i + 1];
+
+        vy_2 = vy[i + 2];
+        va0_2 = vptr_a0[i + 2];
+        va1_2 = vptr_a1[i + 2];
+        va2_2 = vptr_a2[i + 2];
+        va3_2 = vptr_a3[i + 2];
+
+        vy_3 = vy[i + 3];
+        va0_3 = vptr_a0[i + 3];
+        va1_3 = vptr_a1[i + 3];
+        va2_3 = vptr_a2[i + 3];
+        va3_3 = vptr_a3[i + 3];
+
+        vy_0 += va0*vx0_r;
+        vy_1 += va0_1*vx0_r;
+        vy_2 += va0_2*vx0_r;
+        vy_3 += va0_3*vx0_r;
+ 
+
+        vy_0 += va1*vx1_r;
+        vy_1 += va1_1*vx1_r;
+        vy_2 += va1_2*vx1_r;
+        vy_3 += va1_3*vx1_r;
+
+        va0 = vec_xxpermdi(va0, va0, 2);
+        va0_1 = vec_xxpermdi(va0_1, va0_1, 2);
+
+
+        vy_0 += va2*vx2_r;
+        vy_1 += va2_1*vx2_r;
+        va0_2 = vec_xxpermdi(va0_2, va0_2, 2);
+        va0_3 = vec_xxpermdi(va0_3, va0_3, 2);
+        vy_2 += va2_2*vx2_r;
+        vy_3 += va2_3*vx2_r;
+
+        va1 = vec_xxpermdi(va1, va1, 2);
+        va1_1 = vec_xxpermdi(va1_1, va1_1, 2);
+
+
+        vy_0 += va3*vx3_r;
+        vy_1 += va3_1*vx3_r;
+
+        va1_2 = vec_xxpermdi(va1_2, va1_2, 2);
+        va1_3 = vec_xxpermdi(va1_3, va1_3, 2);
+
+        vy_2 += va3_2*vx3_r;
+        vy_3 += va3_3*vx3_r;
+
+        va2 = vec_xxpermdi(va2, va2, 2);
+        va2_1 = vec_xxpermdi(va2_1, va2_1, 2);
+
+
+        vy_0 += va0*vx0_i;
+        vy_1 += va0_1*vx0_i;
+
+        va2_2 = vec_xxpermdi(va2_2, va2_2, 2);
+        va2_3 = vec_xxpermdi(va2_3, va2_3, 2);
+
+        vy_2 += va0_2*vx0_i;
+        vy_3 += va0_3*vx0_i;
+
+        va3 = vec_xxpermdi(va3, va3, 2);
+        va3_1 = vec_xxpermdi(va3_1, va3_1, 2);
+
+
+        vy_0 += va1*vx1_i;
+        vy_1 += va1_1*vx1_i;
+
+        va3_2 = vec_xxpermdi(va3_2, va3_2, 2);
+        va3_3 = vec_xxpermdi(va3_3, va3_3, 2);
+
+        vy_2 += va1_2*vx1_i;
+        vy_3 += va1_3*vx1_i;
+
+        vy_0 += va2*vx2_i;
+        vy_1 += va2_1*vx2_i;
+        vy_2 += va2_2*vx2_i;
+        vy_3 += va2_3*vx2_i;
+
+        vy_0 += va3*vx3_i;
+        vy_1 += va3_1*vx3_i;
+        vy_2 += va3_2*vx3_i;
+        vy_3 += va3_3*vx3_i;
+
+        vy[i] = vy_0;
+        vy[i + 1] = vy_1;
+        vy[i + 2] = vy_2;
+        vy[i + 3] = vy_3;
+
+
+        i += 4;
+         
+       
+    }
+
+}
+#else
+
+static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y) {
+    BLASLONG i;
+    FLOAT *a0, *a1, *a2, *a3;
+    a0 = ap;
+    a1 = ap + lda;
+    a2 = a1 + lda;
+    a3 = a2 + lda;
+
+    for (i = 0; i < 2 * n; i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        y[i] += a0[i] * x[0] - a0[i + 1] * x[1];
+        y[i + 1] += a0[i] * x[1] + a0[i + 1] * x[0];
+        y[i] += a1[i] * x[2] - a1[i + 1] * x[3];
+        y[i + 1] += a1[i] * x[3] + a1[i + 1] * x[2];
+        y[i] += a2[i] * x[4] - a2[i + 1] * x[5];
+        y[i + 1] += a2[i] * x[5] + a2[i + 1] * x[4];
+        y[i] += a3[i] * x[6] - a3[i + 1] * x[7];
+        y[i + 1] += a3[i] * x[7] + a3[i + 1] * x[6];
+#else 
+        y[i] += a0[i] * x[0] + a0[i + 1] * x[1];
+        y[i + 1] += a0[i] * x[1] - a0[i + 1] * x[0];
+        y[i] += a1[i] * x[2] + a1[i + 1] * x[3];
+        y[i + 1] += a1[i] * x[3] - a1[i + 1] * x[2];
+        y[i] += a2[i] * x[4] + a2[i + 1] * x[5];
+        y[i + 1] += a2[i] * x[5] - a2[i + 1] * x[4];
+        y[i] += a3[i] * x[6] + a3[i + 1] * x[7];
+        y[i + 1] += a3[i] * x[7] - a3[i + 1] * x[6];
+#endif
+    }
+}
+
+#endif
+
+#ifdef  HAVE_KERNEL_4x2_VEC
+
+static void zgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y) {
+    BLASLONG i;
+    FLOAT *a0, *a1;
+    a0 = ap;
+    a1 = ap + lda;
+
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+    register __vector double vx0_r = {x[0], x[0]};
+    register __vector double vx0_i = {-x[1], x[1]};
+    register __vector double vx1_r = {x[2], x[2]};
+    register __vector double vx1_i = {-x[3], x[3]};
+
+#else
+    register __vector double vx0_r = {x[0], -x[0]};
+    register __vector double vx0_i = {x[1], x[1]};
+    register __vector double vx1_r = {x[2], -x[2]};
+    register __vector double vx1_i = {x[3], x[3]};
+#endif
+
+
+    register __vector double *vy = (__vector double *) y;
+    register __vector double *vptr_a0 = (__vector double *) a0;
+    register __vector double *vptr_a1 = (__vector double *) a1;
+
+    for (i = 0; i < n; i += 4) {
+
+        register __vector double vy_0 = vy[i];
+        register __vector double vy_1 = vy[i + 1];
+        register __vector double vy_2 = vy[i + 2];
+        register __vector double vy_3 = vy[i + 3];
+
+        register __vector double va0 = vptr_a0[i];
+        register __vector double va0_1 = vptr_a0[i + 1];
+        register __vector double va0_2 = vptr_a0[i + 2];
+        register __vector double va0_3 = vptr_a0[i + 3];
+
+        register __vector double va1 = vptr_a1[i];
+        register __vector double va1_1 = vptr_a1[i + 1];
+        register __vector double va1_2 = vptr_a1[i + 2];
+        register __vector double va1_3 = vptr_a1[i + 3];
+
+        vy_0 += va0*vx0_r;
+        vy_1 += va0_1*vx0_r;
+        vy_2 += va0_2*vx0_r;
+        vy_3 += va0_3*vx0_r;
+
+        va0 = vec_xxpermdi(va0, va0, 2);
+        va0_1 = vec_xxpermdi(va0_1, va0_1, 2);
+        va0_2 = vec_xxpermdi(va0_2, va0_2, 2);
+        va0_3 = vec_xxpermdi(va0_3, va0_3, 2);
+
+        vy_0 += va1*vx1_r;
+        vy_1 += va1_1*vx1_r;
+        vy_2 += va1_2*vx1_r;
+        vy_3 += va1_3*vx1_r;
+
+        va1 = vec_xxpermdi(va1, va1, 2);
+        va1_1 = vec_xxpermdi(va1_1, va1_1, 2);
+        va1_2 = vec_xxpermdi(va1_2, va1_2, 2);
+        va1_3 = vec_xxpermdi(va1_3, va1_3, 2);
+
+        vy_0 += va0*vx0_i;
+        vy_1 += va0_1*vx0_i;
+        vy_2 += va0_2*vx0_i;
+        vy_3 += va0_3*vx0_i;
+
+        vy_0 += va1*vx1_i;
+        vy_1 += va1_1*vx1_i;
+        vy_2 += va1_2*vx1_i;
+        vy_3 += va1_3*vx1_i;
+
+        vy[i] = vy_0;
+        vy[i + 1] = vy_1;
+        vy[i + 2] = vy_2;
+        vy[i + 3] = vy_3;
+
+    }
+}
+#else
+
+static void zgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y) {
+    BLASLONG i;
+    FLOAT *a0, *a1;
+    a0 = ap;
+    a1 = ap + lda;
+
+    for (i = 0; i < 2 * n; i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        y[i] += a0[i] * x[0] - a0[i + 1] * x[1];
+        y[i + 1] += a0[i] * x[1] + a0[i + 1] * x[0];
+        y[i] += a1[i] * x[2] - a1[i + 1] * x[3];
+        y[i + 1] += a1[i] * x[3] + a1[i + 1] * x[2];
+#else 
+        y[i] += a0[i] * x[0] + a0[i + 1] * x[1];
+        y[i + 1] += a0[i] * x[1] - a0[i + 1] * x[0];
+        y[i] += a1[i] * x[2] + a1[i + 1] * x[3];
+        y[i + 1] += a1[i] * x[3] - a1[i + 1] * x[2];
+#endif
+    }
+}
+
+#endif
+
+#ifdef  HAVE_KERNEL_4x1_VEC
+
+static void zgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) {
+    BLASLONG i;
+    FLOAT *a0;
+    a0 = ap;
+
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+    register __vector double vx0_r = {x[0], x[0]};
+    register __vector double vx0_i = {-x[1], x[1]};
+
+#else
+    register __vector double vx0_r = {x[0], -x[0]};
+    register __vector double vx0_i = {x[1], x[1]};
+#endif
+
+
+    register __vector double *vy = (__vector double *) y;
+    register __vector double *vptr_a0 = (__vector double *) a0;
+
+    for (i = 0; i < n; i += 4) {
+
+        register __vector double vy_0 = vy[i];
+        register __vector double vy_1 = vy[i + 1];
+        register __vector double vy_2 = vy[i + 2];
+        register __vector double vy_3 = vy[i + 3];
+
+        register __vector double va0 = vptr_a0[i];
+        register __vector double va0_1 = vptr_a0[i + 1];
+        register __vector double va0_2 = vptr_a0[i + 2];
+        register __vector double va0_3 = vptr_a0[i + 3];
+
+        vy_0 += va0*vx0_r;
+        vy_1 += va0_1*vx0_r;
+        vy_2 += va0_2*vx0_r;
+        vy_3 += va0_3*vx0_r;
+
+        va0 = vec_xxpermdi(va0, va0, 2);
+        va0_1 = vec_xxpermdi(va0_1, va0_1, 2);
+        va0_2 = vec_xxpermdi(va0_2, va0_2, 2);
+        va0_3 = vec_xxpermdi(va0_3, va0_3, 2);
+
+        vy_0 += va0*vx0_i;
+        vy_1 += va0_1*vx0_i;
+        vy_2 += va0_2*vx0_i;
+        vy_3 += va0_3*vx0_i;
+
+        vy[i] = vy_0;
+        vy[i + 1] = vy_1;
+        vy[i + 2] = vy_2;
+        vy[i + 3] = vy_3;
+
+    }
+}
+
+#else
+
+static void zgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y) {
+    BLASLONG i;
+    FLOAT *a0;
+    a0 = ap;
+
+    for (i = 0; i < 2 * n; i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        y[i] += a0[i] * x[0] - a0[i + 1] * x[1];
+        y[i + 1] += a0[i] * x[1] + a0[i + 1] * x[0];
+#else 
+        y[i] += a0[i] * x[0] + a0[i + 1] * x[1];
+        y[i + 1] += a0[i] * x[1] - a0[i + 1] * x[0];
+#endif
+
+    }
+}
+
+#endif
+
+#ifdef HAVE_KERNEL_ADDY
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+
+
+#if   !defined(XCONJ) 
+
+    register __vector double valpha_r = {alpha_r, alpha_r};
+    register __vector double valpha_i = {-alpha_i, alpha_i};
+
+#else
+    register __vector double valpha_r = {alpha_r, -alpha_r};
+    register __vector double valpha_i = {alpha_i, alpha_i};
+#endif
+
+    register __vector double *vptr_src = (__vector double *) src;
+    if (inc_dest != 2) {
+        register __vector double *vptr_y = (__vector double *) dest;
+        //note that inc_dest is already 2x. so we should add it to double*
+        register __vector double *vptr_y1 = (__vector double *) (dest + inc_dest);
+        register __vector double *vptr_y2 = (__vector double *) (dest + 2 * inc_dest);
+        register __vector double *vptr_y3 = (__vector double *) (dest + 3 * inc_dest);
+        BLASLONG dest_t = 0;
+        BLASLONG add_dest = inc_dest << 1; //inc_dest is already multiplied by 2, so for vector 4  we just multiply 2 times
+        for (i = 0; i < n; i += 4) {
+
+            register __vector double vy_0 = vptr_y[dest_t];
+            register __vector double vy_1 = vptr_y1[dest_t];
+            register __vector double vy_2 = vptr_y2[dest_t];
+            register __vector double vy_3 = vptr_y3[dest_t];
+
+            register __vector double vsrc = vptr_src[i];
+            register __vector double vsrc_1 = vptr_src[i + 1];
+            register __vector double vsrc_2 = vptr_src[i + 2];
+            register __vector double vsrc_3 = vptr_src[i + 3];
+
+            vy_0 += vsrc*valpha_r;
+            vy_1 += vsrc_1*valpha_r;
+            vy_2 += vsrc_2*valpha_r;
+            vy_3 += vsrc_3*valpha_r;
+
+            vsrc = vec_xxpermdi(vsrc, vsrc, 2);
+            vsrc_1 = vec_xxpermdi(vsrc_1, vsrc_1, 2);
+            vsrc_2 = vec_xxpermdi(vsrc_2, vsrc_2, 2);
+            vsrc_3 = vec_xxpermdi(vsrc_3, vsrc_3, 2);
+
+            vy_0 += vsrc*valpha_i;
+            vy_1 += vsrc_1*valpha_i;
+            vy_2 += vsrc_2*valpha_i;
+            vy_3 += vsrc_3*valpha_i;
+
+            vptr_y[dest_t] = vy_0;
+            vptr_y1[dest_t ] = vy_1;
+            vptr_y2[dest_t] = vy_2;
+            vptr_y3[dest_t] = vy_3;
+
+            dest_t += add_dest;
+
+        }
+
+        return;
+    } else {
+        register __vector double *vptr_y = (__vector double *) dest;
+        for (i = 0; i < n; i += 4) {
+
+            register __vector double vy_0 = vptr_y[i];
+            register __vector double vy_1 = vptr_y[i + 1];
+            register __vector double vy_2 = vptr_y[i + 2];
+            register __vector double vy_3 = vptr_y[i + 3];
+
+            register __vector double vsrc = vptr_src[i];
+            register __vector double vsrc_1 = vptr_src[i + 1];
+            register __vector double vsrc_2 = vptr_src[i + 2];
+            register __vector double vsrc_3 = vptr_src[i + 3];
+
+            vy_0 += vsrc*valpha_r;
+            vy_1 += vsrc_1*valpha_r;
+            vy_2 += vsrc_2*valpha_r;
+            vy_3 += vsrc_3*valpha_r;
+
+            vsrc = vec_xxpermdi(vsrc, vsrc, 2);
+            vsrc_1 = vec_xxpermdi(vsrc_1, vsrc_1, 2);
+            vsrc_2 = vec_xxpermdi(vsrc_2, vsrc_2, 2);
+            vsrc_3 = vec_xxpermdi(vsrc_3, vsrc_3, 2);
+
+            vy_0 += vsrc*valpha_i;
+            vy_1 += vsrc_1*valpha_i;
+            vy_2 += vsrc_2*valpha_i;
+            vy_3 += vsrc_3*valpha_i;
+
+            vptr_y[i] = vy_0;
+            vptr_y[i + 1 ] = vy_1;
+            vptr_y[i + 2] = vy_2;
+            vptr_y[i + 3] = vy_3;
+
+        }
+
+        return;
+    }
+    return;
+}
+
+#else
+
+static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+
+    if (inc_dest != 2) {
+
+        FLOAT temp_r;
+        FLOAT temp_i;
+        for (i = 0; i < n; i++) {
+#if !defined(XCONJ) 
+            temp_r = alpha_r * src[0] - alpha_i * src[1];
+            temp_i = alpha_r * src[1] + alpha_i * src[0];
+#else
+            temp_r = alpha_r * src[0] + alpha_i * src[1];
+            temp_i = -alpha_r * src[1] + alpha_i * src[0];
+#endif
+
+            *dest += temp_r;
+            *(dest + 1) += temp_i;
+
+            src += 2;
+            dest += inc_dest;
+        }
+        return;
+    }
+
+    FLOAT temp_r0;
+    FLOAT temp_i0;
+    FLOAT temp_r1;
+    FLOAT temp_i1;
+    FLOAT temp_r2;
+    FLOAT temp_i2;
+    FLOAT temp_r3;
+    FLOAT temp_i3;
+    for (i = 0; i < n; i += 4) {
+#if !defined(XCONJ) 
+        temp_r0 = alpha_r * src[0] - alpha_i * src[1];
+        temp_i0 = alpha_r * src[1] + alpha_i * src[0];
+        temp_r1 = alpha_r * src[2] - alpha_i * src[3];
+        temp_i1 = alpha_r * src[3] + alpha_i * src[2];
+        temp_r2 = alpha_r * src[4] - alpha_i * src[5];
+        temp_i2 = alpha_r * src[5] + alpha_i * src[4];
+        temp_r3 = alpha_r * src[6] - alpha_i * src[7];
+        temp_i3 = alpha_r * src[7] + alpha_i * src[6];
+#else
+        temp_r0 = alpha_r * src[0] + alpha_i * src[1];
+        temp_i0 = -alpha_r * src[1] + alpha_i * src[0];
+        temp_r1 = alpha_r * src[2] + alpha_i * src[3];
+        temp_i1 = -alpha_r * src[3] + alpha_i * src[2];
+        temp_r2 = alpha_r * src[4] + alpha_i * src[5];
+        temp_i2 = -alpha_r * src[5] + alpha_i * src[4];
+        temp_r3 = alpha_r * src[6] + alpha_i * src[7];
+        temp_i3 = -alpha_r * src[7] + alpha_i * src[6];
+#endif
+
+        dest[0] += temp_r0;
+        dest[1] += temp_i0;
+        dest[2] += temp_r1;
+        dest[3] += temp_i1;
+        dest[4] += temp_r2;
+        dest[5] += temp_i2;
+        dest[6] += temp_r3;
+        dest[7] += temp_i3;
+
+        src += 8;
+        dest += 8;
+    }
+    return;
+}
+#endif
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT * buffer) {
+    BLASLONG i;
+    BLASLONG j;
+    FLOAT *a_ptr;
+    FLOAT *x_ptr;
+    FLOAT *y_ptr;
+
+    BLASLONG n1;
+    BLASLONG m1;
+    BLASLONG m2;
+    BLASLONG m3;
+    BLASLONG n2;
+
+    FLOAT xbuffer[8], *ybuffer;
+
+    if (m < 1) return (0);
+    if (n < 1) return (0);
+
+    ybuffer = buffer;
+
+    inc_x *= 2;
+    inc_y *= 2;
+    lda *= 2;
+
+    n1 = n / 4;
+    n2 = n % 4;
+
+    m3 = m % 4;
+    m1 = m - (m % 4);
+    m2 = (m % NBMAX) - (m % 4);
+
+    y_ptr = y;
+
+    BLASLONG NB = NBMAX;
+
+    while (NB == NBMAX) {
+
+        m1 -= NB;
+        if (m1 < 0) {
+            if (m2 == 0) break;
+            NB = m2;
+        }
+
+        a_ptr = a;
+
+        x_ptr = x;
+        //zero_y(NB,ybuffer);
+        memset(ybuffer, 0, NB * 16);
+
+        if (inc_x == 2) {
+
+            for (i = 0; i < n1; i++) {
+                zgemv_kernel_4x4(NB, lda, a_ptr, x_ptr, ybuffer);
+
+                a_ptr += lda << 2;
+                x_ptr += 8;
+            }
+
+            if (n2 & 2) {
+                zgemv_kernel_4x2(NB, lda, a_ptr, x_ptr, ybuffer);
+                x_ptr += 4;
+                a_ptr += 2 * lda;
+
+            }
+
+            if (n2 & 1) {
+                zgemv_kernel_4x1(NB, a_ptr, x_ptr, ybuffer);
+                x_ptr += 2;
+                a_ptr += lda;
+
+            }
+        } else {
+
+            for (i = 0; i < n1; i++) {
+
+                xbuffer[0] = x_ptr[0];
+                xbuffer[1] = x_ptr[1];
+                x_ptr += inc_x;
+                xbuffer[2] = x_ptr[0];
+                xbuffer[3] = x_ptr[1];
+                x_ptr += inc_x;
+                xbuffer[4] = x_ptr[0];
+                xbuffer[5] = x_ptr[1];
+                x_ptr += inc_x;
+                xbuffer[6] = x_ptr[0];
+                xbuffer[7] = x_ptr[1];
+                x_ptr += inc_x;
+
+                zgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, ybuffer);
+
+                a_ptr += lda << 2;
+            }
+
+            for (i = 0; i < n2; i++) {
+                xbuffer[0] = x_ptr[0];
+                xbuffer[1] = x_ptr[1];
+                x_ptr += inc_x;
+                zgemv_kernel_4x1(NB, a_ptr, xbuffer, ybuffer);
+                a_ptr += lda;
+
+            }
+
+        }
+
+        add_y(NB, ybuffer, y_ptr, inc_y, alpha_r, alpha_i);
+        a += 2 * NB;
+        y_ptr += NB * inc_y;
+    }
+
+    if (m3 == 0) return (0);
+
+    if (m3 == 1) {
+        a_ptr = a;
+        x_ptr = x;
+        FLOAT temp_r = 0.0;
+        FLOAT temp_i = 0.0;
+
+        if (lda == 2 && inc_x == 2) {
+
+            for (i = 0; i < (n & -2); i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r += a_ptr[2] * x_ptr[2] - a_ptr[3] * x_ptr[3];
+                temp_i += a_ptr[2] * x_ptr[3] + a_ptr[3] * x_ptr[2];
+#else
+                temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r += a_ptr[2] * x_ptr[2] + a_ptr[3] * x_ptr[3];
+                temp_i += a_ptr[2] * x_ptr[3] - a_ptr[3] * x_ptr[2];
+#endif
+
+                a_ptr += 4;
+                x_ptr += 4;
+            }
+
+            for (; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+                temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+
+                a_ptr += 2;
+                x_ptr += 2;
+            }
+
+        } else {
+
+            for (i = 0; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+#else
+                temp_r += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+#endif
+
+                a_ptr += lda;
+                x_ptr += inc_x;
+            }
+
+        }
+#if !defined(XCONJ) 
+        y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+        y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+        y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+        y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+        return (0);
+    }
+
+    if (m3 == 2) {
+        a_ptr = a;
+        x_ptr = x;
+        FLOAT temp_r0 = 0.0;
+        FLOAT temp_i0 = 0.0;
+        FLOAT temp_r1 = 0.0;
+        FLOAT temp_i1 = 0.0;
+
+        if (lda == 4 && inc_x == 2) {
+
+            for (i = 0; i < (n & -2); i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+
+                temp_r0 += a_ptr[4] * x_ptr[2] - a_ptr[5] * x_ptr[3];
+                temp_i0 += a_ptr[4] * x_ptr[3] + a_ptr[5] * x_ptr[2];
+                temp_r1 += a_ptr[6] * x_ptr[2] - a_ptr[7] * x_ptr[3];
+                temp_i1 += a_ptr[6] * x_ptr[3] + a_ptr[7] * x_ptr[2];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+
+                temp_r0 += a_ptr[4] * x_ptr[2] + a_ptr[5] * x_ptr[3];
+                temp_i0 += a_ptr[4] * x_ptr[3] - a_ptr[5] * x_ptr[2];
+                temp_r1 += a_ptr[6] * x_ptr[2] + a_ptr[7] * x_ptr[3];
+                temp_i1 += a_ptr[6] * x_ptr[3] - a_ptr[7] * x_ptr[2];
+#endif
+
+                a_ptr += 8;
+                x_ptr += 4;
+            }
+
+            for (; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+#endif
+
+                a_ptr += 4;
+                x_ptr += 2;
+            }
+
+        } else {
+
+            for (i = 0; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+#endif
+
+                a_ptr += lda;
+                x_ptr += inc_x;
+            }
+
+        }
+#if !defined(XCONJ) 
+        y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+        y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
+        y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
+#else
+        y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+        y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
+        y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+#endif
+        return (0);
+    }
+
+    if (m3 == 3) {
+        a_ptr = a;
+        x_ptr = x;
+        FLOAT temp_r0 = 0.0;
+        FLOAT temp_i0 = 0.0;
+        FLOAT temp_r1 = 0.0;
+        FLOAT temp_i1 = 0.0;
+        FLOAT temp_r2 = 0.0;
+        FLOAT temp_i2 = 0.0;
+
+        if (lda == 6 && inc_x == 2) {
+
+            for (i = 0; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+                temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
+                temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+                temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
+                temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
+#endif
+
+                a_ptr += 6;
+                x_ptr += 2;
+            }
+
+        } else {
+
+            for (i = 0; i < n; i++) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+                temp_r0 += a_ptr[0] * x_ptr[0] - a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] + a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] - a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] + a_ptr[3] * x_ptr[0];
+                temp_r2 += a_ptr[4] * x_ptr[0] - a_ptr[5] * x_ptr[1];
+                temp_i2 += a_ptr[4] * x_ptr[1] + a_ptr[5] * x_ptr[0];
+#else
+                temp_r0 += a_ptr[0] * x_ptr[0] + a_ptr[1] * x_ptr[1];
+                temp_i0 += a_ptr[0] * x_ptr[1] - a_ptr[1] * x_ptr[0];
+                temp_r1 += a_ptr[2] * x_ptr[0] + a_ptr[3] * x_ptr[1];
+                temp_i1 += a_ptr[2] * x_ptr[1] - a_ptr[3] * x_ptr[0];
+                temp_r2 += a_ptr[4] * x_ptr[0] + a_ptr[5] * x_ptr[1];
+                temp_i2 += a_ptr[4] * x_ptr[1] - a_ptr[5] * x_ptr[0];
+#endif
+
+                a_ptr += lda;
+                x_ptr += inc_x;
+            }
+
+        }
+#if !defined(XCONJ) 
+        y_ptr[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+        y_ptr[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
+        y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r2 - alpha_i * temp_i2;
+        y_ptr[1] += alpha_r * temp_i2 + alpha_i * temp_r2;
+#else
+        y_ptr[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+        y_ptr[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
+        y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+        y_ptr += inc_y;
+        y_ptr[0] += alpha_r * temp_r2 + alpha_i * temp_i2;
+        y_ptr[1] -= alpha_r * temp_i2 - alpha_i * temp_r2;
+#endif
+        return (0);
+    }
+
+    return (0);
+}
+

kernel/power/zgemv_t_4.c

@@ -0,0 +1,847 @@
+/***************************************************************************
+Copyright (c) 2018, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ *****************************************************************************/
+
+#include "common.h"
+
+#define NBMAX 4096
+#define HAVE_KERNEL_4x4_VEC 1
+#define HAVE_KERNEL_4x2_VEC 1
+#define HAVE_KERNEL_4x1_VEC 1
+
+#if defined(HAVE_KERNEL_4x4_VEC) || defined(HAVE_KERNEL_4x2_VEC) || defined(HAVE_KERNEL_4x1_VEC)
+#include <altivec.h> 
+#endif
+
+#ifdef HAVE_KERNEL_4x4_VEC_ASM
+
+#elif HAVE_KERNEL_4x4_VEC
+
+static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+    FLOAT *a0, *a1, *a2, *a3;
+    a0 = ap;
+    a1 = ap + lda;
+    a2 = a1 + lda;
+    a3 = a2 + lda;
+    //p for positive(real*real,image*image) r for image (real*image,image*real)
+    register __vector double vtemp0_p = {0.0, 0.0};
+    register __vector double vtemp0_r = {0.0, 0.0};
+    register __vector double vtemp1_p = {0.0, 0.0};
+    register __vector double vtemp1_r = {0.0, 0.0};
+    register __vector double vtemp2_p = {0.0, 0.0};
+    register __vector double vtemp2_r = {0.0, 0.0};
+    register __vector double vtemp3_p = {0.0, 0.0};
+    register __vector double vtemp3_r = {0.0, 0.0};
+    i = 0;
+    n = n << 1;
+    while (i < n) {
+//        __builtin_prefetch(&x[i]);
+//        __builtin_prefetch(&a0[i]);   
+//        __builtin_prefetch(&a1[i]);
+//        __builtin_prefetch(&a2[i]);
+//        __builtin_prefetch(&a3[i]);
+        register __vector double vx_0 = *(__vector double*) (&x[i]);
+        register __vector double vx_1 = *(__vector double*) (&x[i + 2]);
+        register __vector double vx_2 = *(__vector double*) (&x[i + 4]);
+        register __vector double vx_3 = *(__vector double*) (&x[i + 6]);
+
+        register __vector double va0 = *(__vector double*) (&a0[i]);
+        register __vector double va0_1 = *(__vector double*) (&a0[i + 2]);
+        register __vector double va0_2 = *(__vector double*) (&a0[i + 4]);
+        register __vector double va0_3 = *(__vector double*) (&a0[i + 6]);
+
+        register __vector double va1 = *(__vector double*) (&a1[i]);
+        register __vector double va1_1 = *(__vector double*) (&a1[i + 2]);
+        register __vector double va1_2 = *(__vector double*) (&a1[i + 4]);
+        register __vector double va1_3 = *(__vector double*) (&a1[i + 6]);
+
+        register __vector double va2 = *(__vector double*) (&a2[i]);
+        register __vector double va2_1 = *(__vector double*) (&a2[i + 2]);
+        register __vector double va2_2 = *(__vector double*) (&a2[i + 4]);
+        register __vector double va2_3 = *(__vector double*) (&a2[i + 6]);
+
+        register __vector double va3 = *(__vector double*) (&a3[i]);
+        register __vector double va3_1 = *(__vector double*) (&a3[i + 2]);
+        register __vector double va3_2 = *(__vector double*) (&a3[i + 4]);
+        register __vector double va3_3 = *(__vector double*) (&a3[i + 6]);
+
+        register __vector double vxr_0 = vec_xxpermdi(vx_0, vx_0, 2);
+        register __vector double vxr_1 = vec_xxpermdi(vx_1, vx_1, 2);
+
+        i += 8;
+
+        vtemp0_p += vx_0*va0;
+        vtemp0_r += vxr_0*va0;
+
+        vtemp1_p += vx_0*va1;
+        vtemp1_r += vxr_0*va1;
+
+        vtemp2_p += vx_0*va2;
+        vtemp2_r += vxr_0*va2;
+
+        vtemp3_p += vx_0*va3;
+        vtemp3_r += vxr_0*va3;
+
+        vtemp0_p += vx_1*va0_1;
+        vtemp0_r += vxr_1*va0_1;
+
+        vtemp1_p += vx_1*va1_1;
+        vtemp1_r += vxr_1*va1_1;
+        vxr_0 = vec_xxpermdi(vx_2, vx_2, 2);
+        vtemp2_p += vx_1*va2_1;
+        vtemp2_r += vxr_1*va2_1;
+
+        vtemp3_p += vx_1*va3_1;
+        vtemp3_r += vxr_1*va3_1;
+
+        vtemp0_p += vx_2*va0_2;
+        vtemp0_r += vxr_0*va0_2;
+        vxr_1 = vec_xxpermdi(vx_3, vx_3, 2);
+
+        vtemp1_p += vx_2*va1_2;
+        vtemp1_r += vxr_0*va1_2;
+
+        vtemp2_p += vx_2*va2_2;
+        vtemp2_r += vxr_0*va2_2;
+
+        vtemp3_p += vx_2*va3_2;
+        vtemp3_r += vxr_0*va3_2;
+
+        vtemp0_p += vx_3*va0_3;
+        vtemp0_r += vxr_1*va0_3;
+
+        vtemp1_p += vx_3*va1_3;
+        vtemp1_r += vxr_1*va1_3;
+
+        vtemp2_p += vx_3*va2_3;
+        vtemp2_r += vxr_1*va2_3;
+
+        vtemp3_p += vx_3*va3_3;
+        vtemp3_r += vxr_1*va3_3;
+
+    }
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+
+    register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1];
+    register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1];
+
+    register FLOAT temp_r1 = vtemp1_p[0] - vtemp1_p[1];
+    register FLOAT temp_i1 = vtemp1_r[0] + vtemp1_r[1];
+
+    register FLOAT temp_r2 = vtemp2_p[0] - vtemp2_p[1];
+    register FLOAT temp_i2 = vtemp2_r[0] + vtemp2_r[1];
+
+    register FLOAT temp_r3 = vtemp3_p[0] - vtemp3_p[1];
+    register FLOAT temp_i3 = vtemp3_r[0] + vtemp3_r[1];
+
+#else
+    register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1];
+    register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1];
+
+    register FLOAT temp_r1 = vtemp1_p[0] + vtemp1_p[1];
+    register FLOAT temp_i1 = vtemp1_r[0] - vtemp1_r[1];
+
+    register FLOAT temp_r2 = vtemp2_p[0] + vtemp2_p[1];
+    register FLOAT temp_i2 = vtemp2_r[0] - vtemp2_r[1];
+
+    register FLOAT temp_r3 = vtemp3_p[0] + vtemp3_p[1];
+    register FLOAT temp_i3 = vtemp3_r[0] - vtemp3_r[1];
+
+#endif    
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
+    y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
+    y[4] += alpha_r * temp_r2 - alpha_i * temp_i2;
+    y[5] += alpha_r * temp_i2 + alpha_i * temp_r2;
+    y[6] += alpha_r * temp_r3 - alpha_i * temp_i3;
+    y[7] += alpha_r * temp_i3 + alpha_i * temp_r3;
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
+    y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+    y[4] += alpha_r * temp_r2 + alpha_i * temp_i2;
+    y[5] -= alpha_r * temp_i2 - alpha_i * temp_r2;
+    y[6] += alpha_r * temp_r3 + alpha_i * temp_i3;
+    y[7] -= alpha_r * temp_i3 - alpha_i * temp_r3;
+
+#endif
+}
+
+#else
+
+static void zgemv_kernel_4x4(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+    FLOAT *a0, *a1, *a2, *a3;
+    a0 = ap;
+    a1 = ap + lda;
+    a2 = a1 + lda;
+    a3 = a2 + lda;
+
+    FLOAT temp_r0 = 0.0;
+    FLOAT temp_r1 = 0.0;
+    FLOAT temp_r2 = 0.0;
+    FLOAT temp_r3 = 0.0;
+    FLOAT temp_i0 = 0.0;
+    FLOAT temp_i1 = 0.0;
+    FLOAT temp_i2 = 0.0;
+    FLOAT temp_i3 = 0.0;
+
+    for (i = 0; i < 2 * n; i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        temp_r0 += a0[i] * x[i] - a0[i + 1] * x[i + 1];
+        temp_i0 += a0[i] * x[i + 1] + a0[i + 1] * x[i];
+        temp_r1 += a1[i] * x[i] - a1[i + 1] * x[i + 1];
+        temp_i1 += a1[i] * x[i + 1] + a1[i + 1] * x[i];
+        temp_r2 += a2[i] * x[i] - a2[i + 1] * x[i + 1];
+        temp_i2 += a2[i] * x[i + 1] + a2[i + 1] * x[i];
+        temp_r3 += a3[i] * x[i] - a3[i + 1] * x[i + 1];
+        temp_i3 += a3[i] * x[i + 1] + a3[i + 1] * x[i];
+#else
+        temp_r0 += a0[i] * x[i] + a0[i + 1] * x[i + 1];
+        temp_i0 += a0[i] * x[i + 1] - a0[i + 1] * x[i];
+        temp_r1 += a1[i] * x[i] + a1[i + 1] * x[i + 1];
+        temp_i1 += a1[i] * x[i + 1] - a1[i + 1] * x[i];
+        temp_r2 += a2[i] * x[i] + a2[i + 1] * x[i + 1];
+        temp_i2 += a2[i] * x[i + 1] - a2[i + 1] * x[i];
+        temp_r3 += a3[i] * x[i] + a3[i + 1] * x[i + 1];
+        temp_i3 += a3[i] * x[i + 1] - a3[i + 1] * x[i];
+#endif
+    }
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
+    y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
+    y[4] += alpha_r * temp_r2 - alpha_i * temp_i2;
+    y[5] += alpha_r * temp_i2 + alpha_i * temp_r2;
+    y[6] += alpha_r * temp_r3 - alpha_i * temp_i3;
+    y[7] += alpha_r * temp_i3 + alpha_i * temp_r3;
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
+    y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+    y[4] += alpha_r * temp_r2 + alpha_i * temp_i2;
+    y[5] -= alpha_r * temp_i2 - alpha_i * temp_r2;
+    y[6] += alpha_r * temp_r3 + alpha_i * temp_i3;
+    y[7] -= alpha_r * temp_i3 - alpha_i * temp_r3;
+
+#endif
+}
+
+#endif
+
+#ifdef HAVE_KERNEL_4x2_VEC
+
+static void zgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+    FLOAT *a0, *a1;
+    a0 = ap;
+    a1 = ap + lda; 
+    //p for positive(real*real,image*image) r for image (real*image,image*real)
+    register __vector double vtemp0_p = {0.0, 0.0};
+    register __vector double vtemp0_r = {0.0, 0.0};
+    register __vector double vtemp1_p = {0.0, 0.0};
+    register __vector double vtemp1_r = {0.0, 0.0}; 
+    i = 0;
+    n = n << 1;
+    while (i < n) {
+
+        register __vector double vx_0 = *(__vector double*) (&x[i]);
+        register __vector double vx_1 = *(__vector double*) (&x[i + 2]);
+        register __vector double vx_2 = *(__vector double*) (&x[i + 4]);
+        register __vector double vx_3 = *(__vector double*) (&x[i + 6]);
+
+        register __vector double va0 = *(__vector double*) (&a0[i]);
+        register __vector double va0_1 = *(__vector double*) (&a0[i + 2]);
+        register __vector double va0_2 = *(__vector double*) (&a0[i + 4]);
+        register __vector double va0_3 = *(__vector double*) (&a0[i + 6]);
+
+        register __vector double va1 = *(__vector double*) (&a1[i]);
+        register __vector double va1_1 = *(__vector double*) (&a1[i + 2]);
+        register __vector double va1_2 = *(__vector double*) (&a1[i + 4]);
+        register __vector double va1_3 = *(__vector double*) (&a1[i + 6]);
+
+        register __vector double vxr_0 = vec_xxpermdi(vx_0, vx_0, 2);
+        register __vector double vxr_1 = vec_xxpermdi(vx_1, vx_1, 2);
+
+        i += 8;
+
+        vtemp0_p += vx_0*va0;
+        vtemp0_r += vxr_0*va0;
+
+        vtemp1_p += vx_0*va1;
+        vtemp1_r += vxr_0*va1;
+
+        vxr_0 = vec_xxpermdi(vx_2, vx_2, 2);  
+        vtemp0_p += vx_1*va0_1;
+        vtemp0_r += vxr_1*va0_1;
+
+        vtemp1_p += vx_1*va1_1;
+        vtemp1_r += vxr_1*va1_1;
+        vxr_1 = vec_xxpermdi(vx_3, vx_3, 2);
+
+        vtemp0_p += vx_2*va0_2;
+        vtemp0_r += vxr_0*va0_2;
+
+        vtemp1_p += vx_2*va1_2;
+        vtemp1_r += vxr_0*va1_2;
+
+        vtemp0_p += vx_3*va0_3;
+        vtemp0_r += vxr_1*va0_3;
+
+        vtemp1_p += vx_3*va1_3;
+        vtemp1_r += vxr_1*va1_3;
+ 
+    }
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+    register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1];
+    register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1];
+
+    register FLOAT temp_r1 = vtemp1_p[0] - vtemp1_p[1];
+    register FLOAT temp_i1 = vtemp1_r[0] + vtemp1_r[1]; 
+
+#else
+    register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1];
+    register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1];
+
+    register FLOAT temp_r1 = vtemp1_p[0] + vtemp1_p[1];
+    register FLOAT temp_i1 = vtemp1_r[0] - vtemp1_r[1];
+
+#endif    
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
+    y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
+    y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+
+#endif
+}
+
+#else
+
+static void zgemv_kernel_4x2(BLASLONG n, BLASLONG lda, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+    FLOAT *a0, *a1;
+    a0 = ap;
+    a1 = ap + lda;
+
+    FLOAT temp_r0 = 0.0;
+    FLOAT temp_r1 = 0.0;
+    FLOAT temp_i0 = 0.0;
+    FLOAT temp_i1 = 0.0;
+
+    for (i = 0; i < 2 * n; i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        temp_r0 += a0[i] * x[i] - a0[i + 1] * x[i + 1];
+        temp_i0 += a0[i] * x[i + 1] + a0[i + 1] * x[i];
+        temp_r1 += a1[i] * x[i] - a1[i + 1] * x[i + 1];
+        temp_i1 += a1[i] * x[i + 1] + a1[i + 1] * x[i];
+#else
+        temp_r0 += a0[i] * x[i] + a0[i + 1] * x[i + 1];
+        temp_i0 += a0[i] * x[i + 1] - a0[i + 1] * x[i];
+        temp_r1 += a1[i] * x[i] + a1[i + 1] * x[i + 1];
+        temp_i1 += a1[i] * x[i + 1] - a1[i + 1] * x[i];
+#endif
+    }
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 - alpha_i * temp_i1;
+    y[3] += alpha_r * temp_i1 + alpha_i * temp_r1;
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+    y[2] += alpha_r * temp_r1 + alpha_i * temp_i1;
+    y[3] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+
+#endif
+}
+
+#endif
+
+#ifdef HAVE_KERNEL_4x1_VEC
+
+static void zgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+    FLOAT *a0 ;
+    a0 = ap;  
+    //p for positive(real*real,image*image) r for image (real*image,image*real)
+    register __vector double vtemp0_p = {0.0, 0.0};
+    register __vector double vtemp0_r = {0.0, 0.0};
+    i = 0;
+    n = n << 1;
+    while (i < n) {
+
+        register __vector double vx_0 = *(__vector double*) (&x[i]);
+        register __vector double vx_1 = *(__vector double*) (&x[i + 2]);
+        register __vector double vx_2 = *(__vector double*) (&x[i + 4]);
+        register __vector double vx_3 = *(__vector double*) (&x[i + 6]);
+
+        register __vector double va0 = *(__vector double*) (&a0[i]);
+        register __vector double va0_1 = *(__vector double*) (&a0[i + 2]);
+        register __vector double va0_2 = *(__vector double*) (&a0[i + 4]);
+        register __vector double va0_3 = *(__vector double*) (&a0[i + 6]);
+       
+        register __vector double vxr_0 = vec_xxpermdi(vx_0, vx_0, 2);
+        register __vector double vxr_1 = vec_xxpermdi(vx_1, vx_1, 2);
+
+        i += 8;
+
+        vtemp0_p += vx_0*va0;
+        vtemp0_r += vxr_0*va0;
+ 
+        vxr_0 = vec_xxpermdi(vx_2, vx_2, 2);  
+        vtemp0_p += vx_1*va0_1;
+        vtemp0_r += vxr_1*va0_1;
+ 
+        vxr_1 = vec_xxpermdi(vx_3, vx_3, 2);
+
+        vtemp0_p += vx_2*va0_2;
+        vtemp0_r += vxr_0*va0_2;
+ 
+        vtemp0_p += vx_3*va0_3;
+        vtemp0_r += vxr_1*va0_3;
+ 
+    }
+
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+    register FLOAT temp_r0 = vtemp0_p[0] - vtemp0_p[1];
+    register FLOAT temp_i0 = vtemp0_r[0] + vtemp0_r[1];
+
+#else
+    register FLOAT temp_r0 = vtemp0_p[0] + vtemp0_p[1];
+    register FLOAT temp_i0 = vtemp0_r[0] - vtemp0_r[1]; 
+
+#endif    
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0; 
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0; 
+#endif
+
+}
+
+#else
+
+static void zgemv_kernel_4x1(BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y, FLOAT alpha_r, FLOAT alpha_i) {
+    BLASLONG i;
+    FLOAT *a0;
+    a0 = ap;
+
+    FLOAT temp_r0 = 0.0;
+    FLOAT temp_i0 = 0.0;
+
+    for (i = 0; i < 2 * n; i += 2) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+        temp_r0 += a0[i] * x[i] - a0[i + 1] * x[i + 1];
+        temp_i0 += a0[i] * x[i + 1] + a0[i + 1] * x[i];
+#else
+        temp_r0 += a0[i] * x[i] + a0[i + 1] * x[i + 1];
+        temp_i0 += a0[i] * x[i + 1] - a0[i + 1] * x[i];
+#endif
+    }
+
+#if !defined(XCONJ)
+
+    y[0] += alpha_r * temp_r0 - alpha_i * temp_i0;
+    y[1] += alpha_r * temp_i0 + alpha_i * temp_r0;
+
+#else
+
+    y[0] += alpha_r * temp_r0 + alpha_i * temp_i0;
+    y[1] -= alpha_r * temp_i0 - alpha_i * temp_r0;
+
+#endif
+
+}
+
+#endif
+
+static __attribute__((always_inline)) void copy_x(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_src) {
+    BLASLONG i;
+    for (i = 0; i < n; i++) {
+        *dest = *src;
+        *(dest + 1) = *(src + 1);
+        dest += 2;
+        src += inc_src;
+    }
+}
+
+int CNAME(BLASLONG m, BLASLONG n, BLASLONG dummy1, FLOAT alpha_r, FLOAT alpha_i, FLOAT *a, BLASLONG lda, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *buffer) {
+    BLASLONG i;
+    BLASLONG j;
+    FLOAT *a_ptr;
+    FLOAT *x_ptr;
+    FLOAT *y_ptr;
+
+    BLASLONG n1;
+    BLASLONG m1;
+    BLASLONG m2;
+    BLASLONG m3;
+    BLASLONG n2;
+
+    FLOAT ybuffer[8], *xbuffer;
+
+    if (m < 1) return (0);
+    if (n < 1) return (0);
+
+    inc_x <<= 1;
+    inc_y <<= 1;
+    lda <<= 1;
+
+    xbuffer = buffer;
+
+    n1 = n >> 2;
+    n2 = n & 3;
+
+    m3 = m & 3;
+    m1 = m - m3;
+    m2 = (m & (NBMAX - 1)) - m3;
+
+    BLASLONG NB = NBMAX;
+
+    while (NB == NBMAX) {
+
+        m1 -= NB;
+        if (m1 < 0) {
+            if (m2 == 0) break;
+            NB = m2;
+        }
+
+        y_ptr = y;
+        a_ptr = a;
+        x_ptr = x;
+
+        if (inc_x != 2)
+            copy_x(NB, x_ptr, xbuffer, inc_x);
+        else
+            xbuffer = x_ptr;
+
+        if (inc_y == 2) {
+
+            for (i = 0; i < n1; i++) {
+                zgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);
+                a_ptr += lda << 2;
+                y_ptr += 8;
+
+            }
+
+            if (n2 & 2) {
+                zgemv_kernel_4x2(NB, lda, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);
+                a_ptr += lda << 1;
+                y_ptr += 4;
+
+            }
+
+            if (n2 & 1) {
+                zgemv_kernel_4x1(NB, a_ptr, xbuffer, y_ptr, alpha_r, alpha_i);
+                a_ptr += lda;
+                y_ptr += 2;
+
+            }
+
+        } else {
+
+            for (i = 0; i < n1; i++) {
+                memset(ybuffer, 0, sizeof (ybuffer));
+                zgemv_kernel_4x4(NB, lda, a_ptr, xbuffer, ybuffer, alpha_r, alpha_i);
+
+                a_ptr += lda << 2;
+
+                y_ptr[0] += ybuffer[0];
+                y_ptr[1] += ybuffer[1];
+                y_ptr += inc_y;
+                y_ptr[0] += ybuffer[2];
+                y_ptr[1] += ybuffer[3];
+                y_ptr += inc_y;
+                y_ptr[0] += ybuffer[4];
+                y_ptr[1] += ybuffer[5];
+                y_ptr += inc_y;
+                y_ptr[0] += ybuffer[6];
+                y_ptr[1] += ybuffer[7];
+                y_ptr += inc_y;
+
+            }
+
+            for (i = 0; i < n2; i++) {
+                memset(ybuffer, 0, sizeof (ybuffer));
+                zgemv_kernel_4x1(NB, a_ptr, xbuffer, ybuffer, alpha_r, alpha_i);
+                a_ptr += lda;
+                y_ptr[0] += ybuffer[0];
+                y_ptr[1] += ybuffer[1];
+                y_ptr += inc_y;
+
+            }
+
+        }
+        a += 2 * NB;
+        x += NB * inc_x;
+    }
+
+    if (m3 == 0) return (0);
+
+    x_ptr = x;
+    j = 0;
+    a_ptr = a;
+    y_ptr = y;
+
+    if (m3 == 3) {
+
+        FLOAT temp_r;
+        FLOAT temp_i;
+        FLOAT x0 = x_ptr[0];
+        FLOAT x1 = x_ptr[1];
+        x_ptr += inc_x;
+        FLOAT x2 = x_ptr[0];
+        FLOAT x3 = x_ptr[1];
+        x_ptr += inc_x;
+        FLOAT x4 = x_ptr[0];
+        FLOAT x5 = x_ptr[1];
+        while (j < n) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
+            temp_r += a_ptr[4] * x4 - a_ptr[5] * x5;
+            temp_i += a_ptr[4] * x5 + a_ptr[5] * x4;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
+            temp_r += a_ptr[4] * x4 + a_ptr[5] * x5;
+            temp_i += a_ptr[4] * x5 - a_ptr[5] * x4;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j++;
+        }
+        return (0);
+    }
+
+    if (m3 == 2) {
+
+        FLOAT temp_r;
+        FLOAT temp_i;
+        FLOAT temp_r1;
+        FLOAT temp_i1;
+        FLOAT x0 = x_ptr[0];
+        FLOAT x1 = x_ptr[1];
+        x_ptr += inc_x;
+        FLOAT x2 = x_ptr[0];
+        FLOAT x3 = x_ptr[1];
+
+        while (j < (n & -2)) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
+            a_ptr += lda;
+            temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
+            temp_r1 += a_ptr[2] * x2 - a_ptr[3] * x3;
+            temp_i1 += a_ptr[2] * x3 + a_ptr[3] * x2;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
+            a_ptr += lda;
+            temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
+            temp_r1 += a_ptr[2] * x2 + a_ptr[3] * x3;
+            temp_i1 += a_ptr[2] * x3 - a_ptr[3] * x2;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+            y_ptr += inc_y;
+            y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
+            y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+            y_ptr += inc_y;
+            y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
+            y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j += 2;
+        }
+
+        while (j < n) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 - a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 + a_ptr[3] * x2;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+            temp_r += a_ptr[2] * x2 + a_ptr[3] * x3;
+            temp_i += a_ptr[2] * x3 - a_ptr[3] * x2;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j++;
+        }
+
+        return (0);
+    }
+
+    if (m3 == 1) {
+
+        FLOAT temp_r;
+        FLOAT temp_i;
+        FLOAT temp_r1;
+        FLOAT temp_i1;
+        FLOAT x0 = x_ptr[0];
+        FLOAT x1 = x_ptr[1];
+
+        while (j < (n & -2)) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+            a_ptr += lda;
+            temp_r1 = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i1 = a_ptr[0] * x1 + a_ptr[1] * x0;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+            a_ptr += lda;
+            temp_r1 = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i1 = a_ptr[0] * x1 - a_ptr[1] * x0;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+            y_ptr += inc_y;
+            y_ptr[0] += alpha_r * temp_r1 - alpha_i * temp_i1;
+            y_ptr[1] += alpha_r * temp_i1 + alpha_i * temp_r1;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+            y_ptr += inc_y;
+            y_ptr[0] += alpha_r * temp_r1 + alpha_i * temp_i1;
+            y_ptr[1] -= alpha_r * temp_i1 - alpha_i * temp_r1;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j += 2;
+        }
+
+        while (j < n) {
+#if ( !defined(CONJ) && !defined(XCONJ) ) || ( defined(CONJ) && defined(XCONJ) )
+            temp_r = a_ptr[0] * x0 - a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 + a_ptr[1] * x0;
+#else
+
+            temp_r = a_ptr[0] * x0 + a_ptr[1] * x1;
+            temp_i = a_ptr[0] * x1 - a_ptr[1] * x0;
+#endif
+
+#if !defined(XCONJ) 
+            y_ptr[0] += alpha_r * temp_r - alpha_i * temp_i;
+            y_ptr[1] += alpha_r * temp_i + alpha_i * temp_r;
+#else
+            y_ptr[0] += alpha_r * temp_r + alpha_i * temp_i;
+            y_ptr[1] -= alpha_r * temp_i - alpha_i * temp_r;
+#endif
+
+            a_ptr += lda;
+            y_ptr += inc_y;
+            j++;
+        }
+        return (0);
+    }
+
+    return (0);
+
+}
+

kernel/power/zrot.c

@@ -0,0 +1,265 @@
+/***************************************************************************
+Copyright (c) 2018, The OpenBLAS Project
+All rights reserved.
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are
+met:
+1. Redistributions of source code must retain the above copyright
+notice, this list of conditions and the following disclaimer.
+2. Redistributions in binary form must reproduce the above copyright
+notice, this list of conditions and the following disclaimer in
+the documentation and/or other materials provided with the
+distribution.
+3. Neither the name of the OpenBLAS project nor the names of
+its contributors may be used to endorse or promote products
+derived from this software without specific prior written permission.
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
+USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+*****************************************************************************/
+
+#include "common.h"
+
+static void   zrot_kernel_4(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT cosA, FLOAT sinA)
+{
+  __vector double t0;
+  __vector double t1;
+  __vector double t2;
+  __vector double t3;
+  __vector double t4;
+  __vector double t5;
+  __vector double t6;
+  __vector double t7;
+
+ __asm__
+    (
+       "xxspltd     36, %x[cos], 0 \n\t"   // load c to both dwords
+       "xxspltd     37, %x[sin], 0 \n\t"   // load s to both dwords 
+
+       "lxvd2x      32,  0,     %[x_ptr] \n\t"   // load x
+       "lxvd2x      33, %[i16], %[x_ptr] \n\t"
+       "lxvd2x      34, %[i32], %[x_ptr] \n\t"
+       "lxvd2x      35, %[i48],  %[x_ptr] \n\t" 
+
+       "lxvd2x      48, 0,      %[y_ptr]   \n\t"   // load y
+       "lxvd2x      49, %[i16], %[y_ptr] \n\t"
+       "lxvd2x      50, %[i32], %[y_ptr] \n\t"
+       "lxvd2x      51, %[i48], %[y_ptr] \n\t"  
+
+       "addi        %[x_ptr], %[x_ptr], 64  \n\t"
+       "addi        %[y_ptr], %[y_ptr], 64  \n\t"
+
+       "addic.          %[temp_n], %[temp_n], -4    \n\t"
+       "ble       2f          \n\t"
+
+       ".p2align  5           \n"
+     "1:                      \n\t"
+
+       "xvmuldp         40, 32, 36  \n\t" // c * x
+       "xvmuldp         41, 33, 36  \n\t"
+       "xvmuldp         42, 34, 36  \n\t"
+       "xvmuldp         43, 35, 36  \n\t"
+
+       "xvmuldp         %x[x0], 48, 36    \n\t" // c * y
+       "xvmuldp         %x[x1], 49, 36    \n\t"
+       "xvmuldp         %x[x2], 50, 36    \n\t"
+       "xvmuldp         %x[x3], 51, 36    \n\t"
+
+       "xvmuldp         44, 32, 37  \n\t" // s * x
+       "xvmuldp         45, 33, 37  \n\t"
+
+       "lxvd2x          32, 0,     %[x_ptr]  \n\t" // load x
+       "lxvd2x          33, %[i16],%[x_ptr]  \n\t"
+
+       "xvmuldp         46, 34, 37  \n\t"
+       "xvmuldp         47, 35, 37  \n\t"
+
+       "lxvd2x          34, %[i32], %[x_ptr] \n\t"
+       "lxvd2x          35, %[i48], %[x_ptr] \n\t"
+
+       "xvmuldp         %x[x4], 48, 37    \n\t" // s * y
+       "xvmuldp         %x[x5], 49, 37    \n\t"
+
+       "lxvd2x          48, 0,      %[y_ptr] \n\t" // load y
+       "lxvd2x          49, %[i16], %[y_ptr] \n\t"
+
+       "xvmuldp         %x[x6], 50, 37    \n\t"
+       "xvmuldp         %x[x7], 51, 37    \n\t"
+
+       "lxvd2x          50, %[i32], %[y_ptr] \n\t"
+       "lxvd2x          51, %[i48], %[y_ptr] \n\t"
+
+       "xvadddp         40, 40, %x[x4]    \n\t" // c * x + s * y
+       "xvadddp         41, 41, %x[x5]    \n\t" // c * x + s * y
+
+       "addi            %[x_ptr], %[x_ptr], -64 \n\t"
+       "addi            %[y_ptr], %[y_ptr], -64 \n\t"
+
+       "xvadddp         42, 42, %x[x6]    \n\t" // c * x + s * y
+       "xvadddp         43, 43, %x[x7]    \n\t" // c * x + s * y
+
+       "xvsubdp         %x[x0], %x[x0], 44      \n\t" // c * y - s * x
+       "xvsubdp         %x[x1], %x[x1], 45      \n\t" // c * y - s * x
+       "xvsubdp         %x[x2], %x[x2], 46      \n\t" // c * y - s * x
+       "xvsubdp         %x[x3], %x[x3], 47      \n\t" // c * y - s * x
+
+       "stxvd2x         40, 0,      %[x_ptr] \n\t" // store x
+       "stxvd2x         41, %[i16], %[x_ptr] \n\t"
+       "stxvd2x         42, %[i32], %[x_ptr] \n\t"
+       "stxvd2x         43, %[i48], %[x_ptr] \n\t"
+
+       "stxvd2x         %x[x0], 0,      %[y_ptr]   \n\t" // store y
+       "stxvd2x         %x[x1], %[i16], %[y_ptr]   \n\t"
+       "stxvd2x         %x[x2], %[i32], %[y_ptr]   \n\t"
+       "stxvd2x         %x[x3], %[i48], %[y_ptr]   \n\t"
+
+       "addi            %[x_ptr], %[x_ptr], 128 \n\t"
+       "addi            %[y_ptr], %[y_ptr], 128 \n\t"
+
+       "addic.          %[temp_n], %[temp_n], -4      \n\t"
+       "bgt+            1b          \n"
+
+     "2:                      \n\t"
+
+       "xvmuldp         40, 32, 36  \n\t" // c * x
+       "xvmuldp         41, 33, 36  \n\t"
+       "xvmuldp         42, 34, 36  \n\t"
+       "xvmuldp         43, 35, 36  \n\t"
+
+       "xvmuldp         %x[x0], 48, 36    \n\t" // c * y
+       "xvmuldp         %x[x1], 49, 36    \n\t"
+       "xvmuldp         %x[x2], 50, 36    \n\t"
+       "xvmuldp         %x[x3], 51, 36    \n\t"
+
+       "xvmuldp         44, 32, 37  \n\t" // s * x
+       "xvmuldp         45, 33, 37  \n\t"
+       "xvmuldp         46, 34, 37  \n\t"
+       "xvmuldp         47, 35, 37  \n\t"
+
+       "xvmuldp         %x[x4], 48, 37    \n\t" // s * y
+       "xvmuldp         %x[x5], 49, 37    \n\t"
+       "xvmuldp         %x[x6], 50, 37    \n\t"
+       "xvmuldp         %x[x7], 51, 37    \n\t"
+
+       "addi            %[x_ptr], %[x_ptr], -64 \n\t"
+       "addi            %[y_ptr], %[y_ptr], -64 \n\t"
+
+       "xvadddp         40, 40, %x[x4]    \n\t" // c * x + s * y
+       "xvadddp         41, 41, %x[x5]    \n\t" // c * x + s * y
+       "xvadddp         42, 42, %x[x6]    \n\t" // c * x + s * y
+       "xvadddp         43, 43, %x[x7]    \n\t" // c * x + s * y
+
+       "xvsubdp         %x[x0], %x[x0], 44      \n\t" // c * y - s * x
+       "xvsubdp         %x[x1], %x[x1], 45      \n\t" // c * y - s * x
+       "xvsubdp         %x[x2], %x[x2], 46      \n\t" // c * y - s * x
+       "xvsubdp         %x[x3], %x[x3], 47      \n\t" // c * y - s * x
+
+       "stxvd2x         40, 0,      %[x_ptr]   \n\t" // store x
+       "stxvd2x         41, %[i16], %[x_ptr] \n\t"
+       "stxvd2x         42, %[i32], %[x_ptr] \n\t"
+       "stxvd2x         43, %[i48], %[x_ptr] \n\t"
+
+       "stxvd2x         %x[x0], 0,      %[y_ptr]      \n\t" // store y
+       "stxvd2x         %x[x1], %[i16], %[y_ptr]      \n\t"
+       "stxvd2x         %x[x2], %[i32], %[y_ptr]      \n\t"
+       "stxvd2x         %x[x3], %[i48], %[y_ptr]      \n\t"
+
+  
+     :
+       [mem_x] "+m" (*(double (*)[2*n])x),
+       [mem_y] "+m" (*(double (*)[2*n])y),
+       [temp_n] "+&r" (n),     
+       [x_ptr] "+&b"(x), [y_ptr]  "+&b"(y),
+       [x0] "=wa" (t0), 
+       [x1] "=wa" (t1), 
+       [x2] "=wa" (t2), 
+       [x3] "=wa" (t3), 
+       [x4] "=wa" (t4), 
+       [x5] "=wa" (t5),  
+       [x6] "=wa" (t6), 
+       [x7] "=wa" (t7)   
+     :
+       [cos] "d" (cosA),            
+       [sin] "d" (sinA),            
+       [i16] "b" (16),  
+       [i32] "b" (32),  
+       [i48] "b" (48)         
+     :
+       "cr0",
+       "vs32","vs33","vs34","vs35","vs36","vs37",
+       "vs40","vs41","vs42","vs43","vs44","vs45","vs46","vs47",
+       "vs48","vs49","vs50","vs51"
+     );
+             
+      return;
+
+}
+
+int CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT c, FLOAT s)
+{
+    BLASLONG i=0;
+    BLASLONG ix=0,iy=0;
+    FLOAT temp[2];
+    BLASLONG inc_x2;
+    BLASLONG inc_y2;
+
+    if ( n <= 0     )  return(0); 
+
+    if ( (inc_x == 1) && (inc_y == 1) )
+    {
+
+        BLASLONG n1 = n & -4;
+        if ( n1 > 0 )
+        { 
+            zrot_kernel_4(n1, x, y, c, s);
+            i=n1; 
+            ix=2*n1; 
+        }
+
+         while(i < n)
+           {
+                temp[0]   = c*x[ix]   + s*y[ix] ;
+                temp[1]   = c*x[ix+1] + s*y[ix+1] ;
+                y[ix]     = c*y[ix]   - s*x[ix] ;
+                y[ix+1]   = c*y[ix+1] - s*x[ix+1] ;
+                x[ix]     = temp[0] ;
+                x[ix+1]   = temp[1] ;
+
+                ix += 2 ; 
+                i++ ;
+
+            }
+
+    }
+    else
+    {
+        inc_x2 = 2 * inc_x ;
+        inc_y2 = 2 * inc_y ;
+        while(i < n)
+        {
+            temp[0]   = c*x[ix]   + s*y[iy] ;
+            temp[1]   = c*x[ix+1] + s*y[iy+1] ;
+            y[iy]     = c*y[iy]   - s*x[ix] ;
+            y[iy+1]   = c*y[iy+1] - s*x[ix+1] ;
+            x[ix]     = temp[0] ;
+            x[ix+1]   = temp[1] ;
+
+            ix += inc_x2 ;
+            iy += inc_y2 ;
+            i++ ;
+
+        }
+
+    }
+    return(0);
+ 
+}
+
+ 

kernel/x86/KERNEL

@@ -169,7 +169,7 @@ ifndef ZDOTKERNEL
 ZDOTKERNEL   = ../arm/zdot.c
 endif
 
-DSDOTKERNEL   = ../arm/dot.c
+DSDOTKERNEL   = ../generic/dot.c
 
 # Bug in znrm2 assembler kernel
 ifndef ZNRM2KERNEL

kernel/x86_64/caxpy_microk_haswell-2.c

@@ -50,11 +50,11 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vmulps		(%5), %%ymm0 , %%ymm0		    \n\t"
 #endif
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
 	"vmovups        (%2,%0,4), %%ymm5                   \n\t" // 4 complex values from x
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups      32(%2,%0,4), %%ymm7                   \n\t" // 4 complex values from x
 	"vmovups      64(%2,%0,4), %%ymm9                   \n\t" // 4 complex values from x
 	"vmovups      96(%2,%0,4), %%ymm11                  \n\t" // 4 complex values from x
@@ -70,7 +70,7 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vpermilps	$0xb1 , %%ymm11, %%ymm10 	    \n\t"  // exchange real and imag part
 
 	"vfmadd213ps    (%3,%0,4), %%ymm0 , %%ymm5          \n\t"
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vfmadd213ps  32(%3,%0,4), %%ymm0 , %%ymm7          \n\t"
 	"vfmadd213ps  64(%3,%0,4), %%ymm0 , %%ymm9          \n\t"
 	"vfmadd213ps  96(%3,%0,4), %%ymm0 , %%ymm11         \n\t"
@@ -96,7 +96,7 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vfmadd231ps	%%ymm1 , %%ymm10, %%ymm15  \n\t"
 
 	"vmovups	%%ymm5 ,   (%3,%0,4)		    \n\t"
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups	%%ymm7 , 32(%3,%0,4)		    \n\t"
 	"vmovups	%%ymm9 , 64(%3,%0,4)		    \n\t"
 	"vmovups	%%ymm11, 96(%3,%0,4)		    \n\t"

kernel/x86_64/caxpy_microk_sandy-2.c

@@ -50,11 +50,11 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vmulps		(%5), %%ymm0 , %%ymm0		    \n\t"
 #endif
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
 	"vmovups        (%2,%0,4), %%ymm5                   \n\t" // 4 complex values from x
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups      32(%2,%0,4), %%ymm7                   \n\t" // 4 complex values from x
 	"vmovups      64(%2,%0,4), %%ymm9                   \n\t" // 4 complex values from x
 	"vmovups      96(%2,%0,4), %%ymm11                  \n\t" // 4 complex values from x
@@ -85,7 +85,7 @@ static void caxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vaddps         %%ymm10, %%ymm11, %%ymm11           \n\t"
 
 	"vmovups	%%ymm5 ,   (%3,%0,4)		    \n\t"
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups	%%ymm7 , 32(%3,%0,4)		    \n\t"
 	"vmovups	%%ymm9 , 64(%3,%0,4)		    \n\t"
 	"vmovups	%%ymm11, 96(%3,%0,4)		    \n\t"

kernel/x86_64/cdot_microk_haswell-2.c

@@ -46,7 +46,7 @@ static void cdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vxorps		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorps		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16			             \n\t"
+	".p2align 4			             \n\t"
 	"1:				             \n\t"
         "vmovups                  (%2,%0,4), %%ymm8          \n\t"  // 2 * x
         "vmovups                32(%2,%0,4), %%ymm9          \n\t"  // 2 * x

kernel/x86_64/cdot_microk_sandy-2.c

@@ -46,7 +46,7 @@ static void cdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vxorps		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorps		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16			             \n\t"
+	".p2align 4			             \n\t"
 	"1:				             \n\t"
         "vmovups                  (%2,%0,4), %%ymm8          \n\t"  // 2 * x
         "vmovups                32(%2,%0,4), %%ymm9          \n\t"  // 2 * x

kernel/x86_64/cscal_microk_haswell-2.c

@@ -54,7 +54,7 @@ static void cscal_kernel_16( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"subq	        $16, %0			            \n\t"		
 	"jz	2f					    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	//"prefetcht0     128(%1)				    \n\t"
@@ -156,7 +156,7 @@ static void cscal_kernel_16_zero_r( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"subq	        $16, %0			            \n\t"		
 	"jz	2f					    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	//"prefetcht0     128(%1)				    \n\t"
@@ -245,7 +245,7 @@ static void cscal_kernel_16_zero_i( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"subq	        $16, %0			            \n\t"		
 	"jz	2f					    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	//"prefetcht0     128(%1)				    \n\t"
@@ -312,7 +312,7 @@ static void cscal_kernel_16_zero( BLASLONG n, FLOAT *alpha, FLOAT *x)
 
 	"addq	$128, %1				    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	//"prefetcht0     128(%1)				    \n\t"

kernel/x86_64/daxpy.c

@@ -33,17 +33,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "daxpy_microk_nehalem-2.c"
 #elif defined(BULLDOZER)
 #include "daxpy_microk_bulldozer-2.c"
-/*
-these appear to be broken, see issue 1332
 #elif defined(STEAMROLLER) || defined(EXCAVATOR)
 #include "daxpy_microk_steamroller-2.c"
 #elif defined(PILEDRIVER)
 #include "daxpy_microk_piledriver-2.c"
 #elif defined(HASWELL) || defined(ZEN)
 #include "daxpy_microk_haswell-2.c"
-*/
-#elif defined(HASWELL) || defined(ZEN) || defined(PILEDRIVER) || defined(STEAMROLLER) || defined(EXCAVATOR)
-#include "daxpy_microk_sandy-2.c"
 #elif defined(SANDYBRIDGE)
 #include "daxpy_microk_sandy-2.c"
 #endif

kernel/x86_64/daxpy_microk_haswell-2.c

@@ -38,7 +38,7 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	(
 	"vbroadcastsd		(%4), %%ymm0		    \n\t"  // alpha	
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
         "vmovups                  (%3,%0,8), %%ymm12         \n\t"  // 4 * y

kernel/x86_64/daxpy_microk_nehalem-2.c

@@ -39,7 +39,7 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"movsd			(%4), %%xmm0		    \n\t"  // alpha	
 	"shufpd          $0,  %%xmm0, %%xmm0                \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
         // "prefetcht0      192(%2,%0,8)                       \n\t"
         // "prefetcht0      192(%3,%0,8)                       \n\t"

kernel/x86_64/daxpy_microk_sandy-2.c

@@ -50,7 +50,7 @@ static void daxpy_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"subq	        $16, %1			             \n\t"		
 	"jz		2f		             \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
 	"vmulpd		%%ymm4, %%ymm0, %%ymm4		\n\t"

kernel/x86_64/ddot.c

@@ -43,6 +43,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #include "ddot_microk_sandy-2.c"
 #endif
 
+#if !defined(DSDOT)
+#define RETURN_TYPE     FLOAT
+#else
+#define RETURN_TYPE     double
+#endif
+
 
 #ifndef HAVE_KERNEL_8
 
@@ -71,7 +77,7 @@ static void ddot_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *d)
 
 #endif
 
-FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+static FLOAT dot_compute(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 {
 	BLASLONG i=0;
 	BLASLONG ix=0,iy=0;
@@ -139,4 +145,63 @@ FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
 
 }
 
+#if defined(SMP)
+static int dot_thread_function(BLASLONG n, BLASLONG dummy0,
+        BLASLONG dummy1, FLOAT dummy2, FLOAT *x, BLASLONG inc_x, FLOAT *y,
+        BLASLONG inc_y, RETURN_TYPE *result, BLASLONG dummy3)
+{
+        *(RETURN_TYPE *)result = dot_compute(n, x, inc_x, y, inc_y);
 
+        return 0;
+}
+
+extern int blas_level1_thread_with_return_value(int mode, BLASLONG m, BLASLONG n,
+        BLASLONG k, void *alpha, void *a, BLASLONG lda, void *b, BLASLONG ldb,
+        void *c, BLASLONG ldc, int (*function)(), int nthreads);
+#endif
+
+FLOAT CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y)
+{
+#if defined(SMP)
+	int nthreads;
+	FLOAT dummy_alpha;
+#endif
+	FLOAT dot = 0.0;
+
+#if defined(SMP)
+	nthreads = num_cpu_avail(1);
+
+	if (inc_x == 0 || inc_y == 0)
+		nthreads = 1;
+
+	if (n <= 10000)
+		nthreads = 1;
+
+	if (nthreads == 1) {
+		dot = dot_compute(n, x, inc_x, y, inc_y);
+	} else {
+		int mode, i;
+		char result[MAX_CPU_NUMBER * sizeof(double) * 2];
+		RETURN_TYPE *ptr;
+
+#if !defined(DOUBLE)
+		mode = BLAS_SINGLE  | BLAS_REAL;
+#else
+		mode = BLAS_DOUBLE  | BLAS_REAL;
+#endif
+		blas_level1_thread_with_return_value(mode, n, 0, 0, &dummy_alpha,
+				   x, inc_x, y, inc_y, result, 0,
+				   ( void *)dot_thread_function, nthreads);
+
+		ptr = (RETURN_TYPE *)result;
+		for (i = 0; i < nthreads; i++) {
+			dot = dot + (*ptr);
+			ptr = (RETURN_TYPE *)(((char *)ptr) + sizeof(double) * 2);
+		}
+	}
+#else
+	dot = dot_compute(n, x, inc_x, y, inc_y);
+#endif
+
+	return dot;
+}

kernel/x86_64/ddot_microk_haswell-2.c

@@ -41,7 +41,7 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vxorpd		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorpd		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16				             \n\t"
+	".p2align 4				             \n\t"
 	"1:				             \n\t"
         "vmovups                  (%2,%0,8), %%ymm12         \n\t"  // 2 * x
         "vmovups                32(%2,%0,8), %%ymm13         \n\t"  // 2 * x

kernel/x86_64/ddot_microk_nehalem-2.c

@@ -41,7 +41,7 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"xorpd		%%xmm6, %%xmm6	             \n\t"
 	"xorpd		%%xmm7, %%xmm7	             \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
         "movups                  (%2,%0,8), %%xmm12         \n\t"  // 2 * x

kernel/x86_64/ddot_microk_sandy-2.c

@@ -41,7 +41,7 @@ static void ddot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vxorpd		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorpd		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16				             \n\t"
+	".p2align 4				             \n\t"
 	"1:				             \n\t"
         "vmovups                  (%2,%0,8), %%ymm12         \n\t"  // 2 * x
         "vmovups                32(%2,%0,8), %%ymm13         \n\t"  // 2 * x

kernel/x86_64/dgemv_n_microk_nehalem-4.c

@@ -62,7 +62,7 @@ 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"
+	".p2align 4				 \n\t"
 	"1:				 \n\t"
 
 	"xorpd           %%xmm4 , %%xmm4	 \n\t"

kernel/x86_64/dger_microk_sandy-2.c

@@ -53,7 +53,7 @@ static void dger_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"subq	        $8, %1			             \n\t"		
 	"jz		2f		             \n\t"
 
-	".align 8				            \n\t"
+	".p2align 3				            \n\t"
 	"1:				            \n\t"
 
 	"vmulpd		%%xmm4, %%xmm0, %%xmm4		\n\t"

kernel/x86_64/dscal.c

@@ -99,7 +99,7 @@ static void dscal_kernel_inc_8(BLASLONG n, FLOAT *alpha, FLOAT *x, BLASLONG inc_
 
 	"leaq		(%1,%4,4), %2		            \n\t"
 
-        ".align 16                                          \n\t"
+        ".p2align 4                                          \n\t"
 
         "1:                                                 \n\t"
 	"movsd	(%1)     , %%xmm4			    \n\t"

kernel/x86_64/dscal_microk_haswell-2.c

@@ -58,7 +58,7 @@ static void dscal_kernel_8( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"subq	        $1 , %0			            \n\t"		
 	"jz		2f		             	    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 	// "prefetcht0     640(%1)				    \n\t" 
 
@@ -156,7 +156,7 @@ static void dscal_kernel_8_zero( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"cmpq 	$0, %0					    \n\t"
 	"je	2f					    \n\t" 
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	"vmovups	%%xmm0  ,-128(%1)		    \n\t"

kernel/x86_64/dscal_microk_sandy-2.c

@@ -58,7 +58,7 @@ static void dscal_kernel_8( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"subq	        $1 , %0			            \n\t"		
 	"jz		2f		             	    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 	"prefetcht0     640(%1)				    \n\t" 
 
@@ -156,7 +156,7 @@ static void dscal_kernel_8_zero( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"cmpq 	$0, %0					    \n\t"
 	"je	2f					    \n\t" 
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	"vmovups	%%xmm0  ,-128(%1)		    \n\t"

kernel/x86_64/dsymv_L_microk_haswell-2.c

@@ -44,7 +44,7 @@ static void dsymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FL
 	"vbroadcastsd 16(%8),    %%ymm6	             \n\t"	// temp1[1]
 	"vbroadcastsd 24(%8),    %%ymm7	             \n\t"	// temp1[1]
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,8), %%ymm9	           \n\t"  // 2 * y

kernel/x86_64/dsymv_L_microk_nehalem-2.c

@@ -47,7 +47,7 @@ static void dsymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FL
 	"shufpd $0,  %%xmm6, %%xmm6                  \n\t"
 	"shufpd $0,  %%xmm7, %%xmm7                  \n\t"
 
-	".align 16		  		       \n\t"
+	".p2align 4		  		       \n\t"
 	"1:				       \n\t"
 	"movups	            (%4,%0,8), %%xmm12	       \n\t"	// 2 * a
 	"movups	            (%2,%0,8), %%xmm8	       \n\t"	// 2 * x

kernel/x86_64/dsymv_L_microk_sandy-2.c

@@ -44,7 +44,7 @@ static void dsymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FL
 	"vbroadcastsd 16(%8),    %%ymm6	             \n\t"	// temp1[1]
 	"vbroadcastsd 24(%8),    %%ymm7	             \n\t"	// temp1[1]
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,8), %%ymm9	           \n\t"  // 2 * y

kernel/x86_64/dsymv_U_microk_haswell-2.c

@@ -46,7 +46,7 @@ static void dsymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT
 	"vbroadcastsd 24(%8),    %%ymm7	             \n\t"	// temp1[1]
 	"xorq           %0,%0                        \n\t"
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,8), %%ymm9	           \n\t"  // 2 * y

kernel/x86_64/dsymv_U_microk_nehalem-2.c

@@ -50,7 +50,7 @@ static void dsymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT
 
 	"xorq		%0,%0			     \n\t"
 
-	".align 16		  		       \n\t"
+	".p2align 4		  		       \n\t"
 	"1:				       \n\t"
 	"movups	            (%4,%0,8), %%xmm12	       \n\t"	// 2 * a
 	"movups	            (%2,%0,8), %%xmm8	       \n\t"	// 2 * x

kernel/x86_64/dsymv_U_microk_sandy-2.c

@@ -46,7 +46,7 @@ static void dsymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT
 	"vbroadcastsd 24(%8),    %%ymm7	             \n\t"	// temp1[1]
 	"xorq           %0,%0                        \n\t"
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,8), %%ymm9	           \n\t"  // 2 * y

kernel/x86_64/dtrmm_kernel_4x8_haswell.c

@@ -24,7 +24,7 @@ static void dtrmm_kernel_4x8( BLASLONG n, FLOAT *alpha ,FLOAT *a, FLOAT *b, FLOA
 		"	cmp $0, %1						\n\t"
 		"	jz 2f							\n\t"
 
-		"	.align 16						\n\t"
+		"	.p2align 4						\n\t"
 		"1:								\n\t"
 		"	vmovups   	(%2,%0,4) , %%ymm0			\n\t"
 		"	vmovups   	(%3,%0,8) , %%ymm1			\n\t"

kernel/x86_64/dtrsm_kernel_RN_haswell.c

@@ -128,7 +128,7 @@ static void dtrsm_RN_solve_opt(BLASLONG n, FLOAT *a, FLOAT *b, FLOAT *c, BLASLON
 	"	cmpq		%1, %0						\n\t"
 	"	je	       21f						\n\t"
 
-	"	.align 16							\n\t"
+	"	.p2align 4							\n\t"
 	"1:									\n\t"
 
 	"	vmovups         (%2,%1,4), %%ymm4				\n\t"	// read a

kernel/x86_64/saxpy_microk_haswell-2.c

@@ -38,7 +38,7 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	(
 	"vbroadcastss		(%4), %%ymm0		    \n\t"  // alpha	
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
         "vmovups                  (%3,%0,4), %%ymm12         \n\t"  // 8 * y

kernel/x86_64/saxpy_microk_nehalem-2.c

@@ -39,7 +39,7 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"movss			(%4), %%xmm0		    \n\t"  // alpha	
 	"shufps          $0,  %%xmm0, %%xmm0                \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
         // "prefetcht0      192(%2,%0,4)                       \n\t"
         // "prefetcht0      192(%3,%0,4)                       \n\t"

kernel/x86_64/saxpy_microk_sandy-2.c

@@ -50,7 +50,7 @@ static void saxpy_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"subq	        $32, %1			             \n\t"		
 	"jz		2f		             \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
 	"vmulps		%%ymm4, %%ymm0, %%ymm4		\n\t"

kernel/x86_64/sdot_microk_haswell-2.c

@@ -41,7 +41,7 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vxorps		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorps		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16				             \n\t"
+	".p2align 4				             \n\t"
 	"1:				             \n\t"
         "vmovups                  (%2,%0,4), %%ymm12         \n\t"  // 2 * x
         "vmovups                32(%2,%0,4), %%ymm13         \n\t"  // 2 * x

kernel/x86_64/sdot_microk_nehalem-2.c

@@ -41,7 +41,7 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"xorps		%%xmm6, %%xmm6	             \n\t"
 	"xorps		%%xmm7, %%xmm7	             \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
         "movups                  (%2,%0,4), %%xmm12         \n\t"  // 4 * x
         "movups                  (%3,%0,4), %%xmm8          \n\t"  // 4 * x

kernel/x86_64/sdot_microk_sandy-2.c

@@ -41,7 +41,7 @@ static void sdot_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vxorps		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorps		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16				             \n\t"
+	".p2align 4				             \n\t"
 	"1:				             \n\t"
         "vmovups                  (%2,%0,4), %%ymm12         \n\t"  // 2 * x
         "vmovups                32(%2,%0,4), %%ymm13         \n\t"  // 2 * x

kernel/x86_64/sgemv_n_microk_nehalem-4.c

@@ -59,18 +59,18 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
 	"shufps $0,  %%xmm6 , %%xmm6 \n\t"	
 
 
-	".align 16				 \n\t"
+	".p2align 4				 \n\t"
 	"1:				 \n\t"
 	"xorps           %%xmm4 , %%xmm4	 \n\t"
 	"xorps           %%xmm5 , %%xmm5	 \n\t"
 	"movups             (%3,%0,4), %%xmm7          \n\t" // 4 * y
 
-	".align 2				       \n\t"
+	".p2align 1				       \n\t"
 	"movups             (%4,%0,4), %%xmm8          \n\t" 
 	"movups             (%5,%0,4), %%xmm9          \n\t" 
 	"movups             (%6,%0,4), %%xmm10         \n\t" 
 	"movups             (%7,%0,4), %%xmm11         \n\t" 
-	".align 2				       \n\t"
+	".p2align 1				       \n\t"
 	"mulps		%%xmm12, %%xmm8		       \n\t"
 	"mulps		%%xmm13, %%xmm9		       \n\t"
 	"mulps		%%xmm14, %%xmm10	       \n\t"
@@ -84,7 +84,7 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
 	"movups             (%5,%8,4), %%xmm9          \n\t" 
 	"movups             (%6,%8,4), %%xmm10         \n\t" 
 	"movups             (%7,%8,4), %%xmm11         \n\t" 
-	".align 2				       \n\t"
+	".p2align 1				       \n\t"
 	"mulps		%%xmm0 , %%xmm8		       \n\t"
 	"mulps		%%xmm1 , %%xmm9		       \n\t"
 	"mulps		%%xmm2 , %%xmm10	       \n\t"
@@ -154,7 +154,7 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
 	"movss    (%8), %%xmm6	     \n\t"	// alpha 
 	"shufps $0,  %%xmm6 , %%xmm6 \n\t"	
 
-	".align 16				 \n\t"
+	".p2align 4				 \n\t"
 	"1:				 \n\t"
 	"xorps           %%xmm4 , %%xmm4	 \n\t"
 	"movups	       (%3,%0,4), %%xmm7	 \n\t"	// 4 * y

kernel/x86_64/sgemv_n_microk_sandy-4.c

@@ -129,7 +129,7 @@ static void sgemv_kernel_4x8( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, BLASLO
         "je             4f                 \n\t"
 
 
-	".align 16				 \n\t"
+	".p2align 4				 \n\t"
 	"1:				 \n\t"
 	"vxorps	  %%ymm4 , %%ymm4 , %%ymm4        \n\t"
 	"vxorps	  %%ymm5 , %%ymm5 , %%ymm5        \n\t"
@@ -299,7 +299,7 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y, FLOAT
         "je             4f                 \n\t"
 
 
-	".align 16				 \n\t"
+	".p2align 4				 \n\t"
 	"1:				 \n\t"
 	"vxorps	  %%ymm4 , %%ymm4 , %%ymm4        \n\t"
 	"vxorps	  %%ymm5 , %%ymm5 , %%ymm5        \n\t"

kernel/x86_64/sgemv_t_microk_nehalem-4.c

@@ -40,7 +40,7 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"xorps		%%xmm6 , %%xmm6	         \n\t"
 	"xorps		%%xmm7 , %%xmm7	         \n\t"
 
-	".align 16				 \n\t"
+	".p2align 4				 \n\t"
 	"1:				 \n\t"
 
 	"movups	       (%2,%0,4), %%xmm12              \n\t"   // 4 * x

kernel/x86_64/sgemv_t_microk_sandy-4.c

@@ -85,7 +85,7 @@ static void sgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
         "je             4f                      \n\t"
 
 
-	".align 16				 \n\t"
+	".p2align 4				 \n\t"
 	"1:				 \n\t"
 	"prefetcht0	 384(%2,%0,4)		       \n\t"
         "vmovups        (%2,%0,4), %%ymm12       \n\t"  // 8 * x

kernel/x86_64/sger_microk_sandy-2.c

@@ -53,7 +53,7 @@ static void sger_kernel_16( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"subq	        $16, %1			             \n\t"		
 	"jz		2f		             \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
 	"vmulps		%%xmm4, %%xmm0, %%xmm4		\n\t"

kernel/x86_64/ssymv_L_microk_haswell-2.c

@@ -43,7 +43,7 @@ static void ssymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FL
 	"vbroadcastss  8(%8),    %%xmm6	             \n\t"	// temp1[1]
 	"vbroadcastss 12(%8),    %%xmm7	             \n\t"	// temp1[1]
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,4), %%xmm9	           \n\t"  // 2 * y

kernel/x86_64/ssymv_L_microk_nehalem-2.c

@@ -47,7 +47,7 @@ static void ssymv_kernel_4x4(BLASLONG from, BLASLONG to,  FLOAT **a, FLOAT *x, F
 	"shufps $0,  %%xmm6, %%xmm6                  \n\t"
 	"shufps $0,  %%xmm7, %%xmm7                  \n\t"
 
-	".align 16		  		       \n\t"
+	".p2align 4		  		       \n\t"
 	"1:				       \n\t"
 	"movups	            (%2,%0,4), %%xmm8	       \n\t"	// 4 * x
 	"movups	            (%3,%0,4), %%xmm9         \n\t"	// 4 * y

kernel/x86_64/ssymv_L_microk_sandy-2.c

@@ -45,7 +45,7 @@ static void ssymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FL
 	"vbroadcastss  8(%8),    %%xmm6	             \n\t"	// temp1[1]
 	"vbroadcastss 12(%8),    %%xmm7	             \n\t"	// temp1[1]
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,4), %%xmm9	           \n\t"  // 2 * y
@@ -143,7 +143,7 @@ static void ssymv_kernel_4x4(BLASLONG from, BLASLONG to, FLOAT **a, FLOAT *x, FL
 	"vbroadcastss  8(%8),    %%ymm6	             \n\t"	// temp1[1]
 	"vbroadcastss 12(%8),    %%ymm7	             \n\t"	// temp1[1]
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,4), %%ymm9	           \n\t"  // 2 * y

kernel/x86_64/ssymv_U_microk_haswell-2.c

@@ -46,7 +46,7 @@ static void ssymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT
 	"vbroadcastss 12(%8),    %%ymm7	             \n\t"	// temp1[1]
 	"xorq           %0,%0                        \n\t"
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,4), %%ymm9	           \n\t"  // 2 * y

kernel/x86_64/ssymv_U_microk_nehalem-2.c

@@ -50,7 +50,7 @@ static void ssymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT
 
 	"xorq		%0,%0			     \n\t"
 
-	".align 16		  		       \n\t"
+	".p2align 4		  		       \n\t"
 	"1:				       \n\t"
 	"movups	            (%2,%0,4), %%xmm8	       \n\t"	// 4 * x
 	"movups	            (%3,%0,4), %%xmm9         \n\t"	// 4 * y

kernel/x86_64/ssymv_U_microk_sandy-2.c

@@ -46,7 +46,7 @@ static void ssymv_kernel_4x4(BLASLONG n, FLOAT *a0, FLOAT *a1, FLOAT *a2, FLOAT
 	"vbroadcastss 12(%8),    %%ymm7	             \n\t"	// temp1[1]
 	"xorq           %0,%0                        \n\t"
 
-	".align 16				     \n\t"
+	".p2align 4				     \n\t"
 	"1:				     \n\t"
 
 	"vmovups	(%3,%0,4), %%ymm9	           \n\t"  // 2 * y

kernel/x86_64/zaxpy_microk_haswell-2.c

@@ -50,11 +50,11 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vmulpd		(%5), %%ymm0 , %%ymm0		    \n\t"
 #endif
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
 	"vmovups        (%2,%0,8), %%ymm5                   \n\t" // 2 complex values from x
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups      32(%2,%0,8), %%ymm7                   \n\t" // 2 complex values from x
 	"vmovups      64(%2,%0,8), %%ymm9                   \n\t" // 2 complex values from x
 	"vmovups      96(%2,%0,8), %%ymm11                  \n\t" // 2 complex values from x
@@ -70,7 +70,7 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vpermilpd	$0x5 , %%ymm11, %%ymm10 	    \n\t"  // exchange real and imag part
 
 	"vfmadd213pd    (%3,%0,8), %%ymm0 , %%ymm5          \n\t"
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vfmadd213pd  32(%3,%0,8), %%ymm0 , %%ymm7          \n\t"
 	"vfmadd213pd  64(%3,%0,8), %%ymm0 , %%ymm9          \n\t"
 	"vfmadd213pd  96(%3,%0,8), %%ymm0 , %%ymm11         \n\t"
@@ -96,7 +96,7 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vfmadd231pd	%%ymm1 , %%ymm10, %%ymm15  \n\t"
 
 	"vmovups	%%ymm5 ,   (%3,%0,8)		    \n\t"
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups	%%ymm7 , 32(%3,%0,8)		    \n\t"
 	"vmovups	%%ymm9 , 64(%3,%0,8)		    \n\t"
 	"vmovups	%%ymm11, 96(%3,%0,8)		    \n\t"

kernel/x86_64/zaxpy_microk_sandy-2.c

@@ -54,11 +54,11 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vmulpd		(%5), %%ymm0 , %%ymm0		    \n\t"
 #endif
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
 	"vmovups        (%2,%0,8), %%ymm5                   \n\t" // 4 complex values from x
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups      32(%2,%0,8), %%ymm7                   \n\t" // 4 complex values from x
 	"vmovups      64(%2,%0,8), %%ymm9                   \n\t" // 4 complex values from x
 	"vmovups      96(%2,%0,8), %%ymm11                  \n\t" // 4 complex values from x
@@ -89,7 +89,7 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vaddpd         %%ymm10, %%ymm11, %%ymm11           \n\t"
 
 	"vmovups	%%ymm5 ,   (%3,%0,8)		    \n\t"
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups	%%ymm7 , 32(%3,%0,8)		    \n\t"
 	"vmovups	%%ymm9 , 64(%3,%0,8)		    \n\t"
 	"vmovups	%%ymm11, 96(%3,%0,8)		    \n\t"
@@ -127,13 +127,13 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vmulpd		(%5), %%ymm0 , %%ymm0		    \n\t"
 #endif
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            \n\t"
 
 	"prefetcht0	512(%2,%0,8)			    \n\t"
 	"prefetcht0	576(%2,%0,8)			    \n\t"
 	"vmovups        (%2,%0,8), %%ymm5                   \n\t" // 4 complex values from x
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups      32(%2,%0,8), %%ymm7                   \n\t" // 4 complex values from x
 	"vmovups      64(%2,%0,8), %%ymm9                   \n\t" // 4 complex values from x
 	"vmovups      96(%2,%0,8), %%ymm11                  \n\t" // 4 complex values from x
@@ -166,7 +166,7 @@ static void zaxpy_kernel_4( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *alpha)
 	"vaddpd         %%ymm10, %%ymm11, %%ymm11           \n\t"
 
 	"vmovups	%%ymm5 ,   (%3,%0,8)		    \n\t"
-	".align 2					    \n\t"
+	".p2align 1					    \n\t"
 	"vmovups	%%ymm7 , 32(%3,%0,8)		    \n\t"
 	"vmovups	%%ymm9 , 64(%3,%0,8)		    \n\t"
 	"vmovups	%%ymm11, 96(%3,%0,8)		    \n\t"

kernel/x86_64/zdot_microk_haswell-2.c

@@ -50,7 +50,7 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vxorpd		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorpd		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16			             \n\t"
+	".p2align 4			             \n\t"
 	"1:				             \n\t"
         "vmovups                  (%2,%0,8), %%ymm8          \n\t"  // 2 * x
         "vmovups                32(%2,%0,8), %%ymm9          \n\t"  // 2 * x
@@ -131,7 +131,7 @@ static void zdot_kernel_8( BLASLONG n, FLOAT *x, FLOAT *y, FLOAT *dot)
 	"vxorpd		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorpd		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16			             \n\t"
+	".p2align 4			             \n\t"
 	"1:				             \n\t"
 	"prefetcht0	512(%2,%0,8)		     \n\t"
         "vmovups                  (%2,%0,8), %%ymm8          \n\t"  // 2 * x

kernel/x86_64/zdot_microk_sandy-2.c

@@ -49,7 +49,7 @@ if ( n < 1280 )
 	"vxorpd		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorpd		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16			             \n\t"
+	".p2align 4			             \n\t"
 	"1:				             \n\t"
         "vmovups                  (%2,%0,8), %%ymm8          \n\t"  // 2 * x
         "vmovups                32(%2,%0,8), %%ymm9          \n\t"  // 2 * x
@@ -137,7 +137,7 @@ if ( n < 1280 )
 	"vxorpd		%%ymm6, %%ymm6, %%ymm6	             \n\t"
 	"vxorpd		%%ymm7, %%ymm7, %%ymm7	             \n\t"
 
-	".align 16			             \n\t"
+	".p2align 4			             \n\t"
 	"1:				             \n\t"
 	"prefetcht0	512(%2,%0,8)		     \n\t"
         "vmovups                  (%2,%0,8), %%ymm8          \n\t"  // 2 * x

kernel/x86_64/zgemv_n_microk_sandy-4.c

@@ -47,7 +47,7 @@ static void zgemv_kernel_4x4( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vbroadcastsd	56(%2), %%ymm7                  \n\t"  // imag part x3
 
 
-	".align 16				        \n\t"
+	".p2align 4				        \n\t"
 	"1:				        \n\t"
 
         //"prefetcht0      256(%4,%0,8)                   \n\t"
@@ -164,7 +164,7 @@ static void zgemv_kernel_4x2( BLASLONG n, FLOAT **ap, FLOAT *x, FLOAT *y)
 	"vbroadcastsd	16(%2), %%ymm2                  \n\t"  // real part x1
 	"vbroadcastsd	24(%2), %%ymm3                  \n\t"  // imag part x1
 
-	".align 16				        \n\t"
+	".p2align 4				        \n\t"
 	"1:				        \n\t"
 
         // "prefetcht0      256(%4,%0,8)                   \n\t"
@@ -253,7 +253,7 @@ static void zgemv_kernel_4x1( BLASLONG n, FLOAT *ap, FLOAT *x, FLOAT *y)
 	"vbroadcastsd	  (%2), %%ymm0                  \n\t"  // real part x0
 	"vbroadcastsd	 8(%2), %%ymm1                  \n\t"  // imag part x0
 
-	".align 16				        \n\t"
+	".p2align 4				        \n\t"
 	"1:				        \n\t"
 
         // "prefetcht0      256(%4,%0,8)                   \n\t"
@@ -355,7 +355,7 @@ static void add_y(BLASLONG n, FLOAT *src, FLOAT *dest, BLASLONG inc_dest,FLOAT a
 	"vbroadcastsd	  (%4), %%ymm0                  \n\t"  // alpha_r
 	"vbroadcastsd	  (%5), %%ymm1                  \n\t"  // alpha_i
 
-	".align 16				        \n\t"
+	".p2align 4				        \n\t"
 	"1:				        \n\t"
 	// "prefetcht0      192(%2,%0,8)			\n\t"
 	"vmovups	(%2,%0,8), %%ymm8	        \n\t" // 2 complex values from src

kernel/x86_64/zscal_microk_haswell-2.c

@@ -54,7 +54,7 @@ static void zscal_kernel_8( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"subq	        $8 , %0			            \n\t"		
 	"jz	2f					    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	//"prefetcht0     128(%1)				    \n\t"
@@ -156,7 +156,7 @@ static void zscal_kernel_8_zero_r( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"subq	        $8 , %0			            \n\t"		
 	"jz	2f					    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	//"prefetcht0     128(%1)				    \n\t"
@@ -245,7 +245,7 @@ static void zscal_kernel_8_zero_i( BLASLONG n, FLOAT *alpha, FLOAT *x)
 	"subq	        $8 , %0			            \n\t"		
 	"jz	2f					    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	//"prefetcht0     128(%1)				    \n\t"
@@ -312,7 +312,7 @@ static void zscal_kernel_8_zero( BLASLONG n, FLOAT *alpha, FLOAT *x)
 
 	"addq	$128, %1				    \n\t"
 
-	".align 16				            \n\t"
+	".p2align 4				            \n\t"
 	"1:				            	    \n\t"
 
 	//"prefetcht0     128(%1)				    \n\t"

kernel/zarch/KERNEL.Z13

@@ -49,6 +49,7 @@ SDOTKERNEL   = ../arm/dot.c
 DDOTKERNEL   = ddot.c
 CDOTKERNEL   = ../arm/zdot.c
 ZDOTKERNEL   = zdot.c
+DSDOTKERNEL  = ../generic/dot.c
 
 SNRM2KERNEL  = ../arm/nrm2.c
 DNRM2KERNEL  = ../arm/nrm2.c

kernel/zarch/KERNEL.ZARCH_GENERIC

@@ -49,6 +49,7 @@ SDOTKERNEL   = ../arm/dot.c
 DDOTKERNEL   = ../arm/dot.c
 CDOTKERNEL   = ../arm/zdot.c
 ZDOTKERNEL   = ../arm/zdot.c
+DSDOTKERNEL  = ../generic/dot.c
 
 SNRM2KERNEL  = ../arm/nrm2.c
 DNRM2KERNEL  = ../arm/nrm2.c

kernel/zarch/dscal.c

@@ -27,8 +27,54 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
+#ifdef Z13_A
+static void   dscal_kernel_32( BLASLONG n, FLOAT  da , FLOAT *x )
+{
 
-
+          
+             __asm__ ("pfd    2, 0(%[x_ptr])   \n\t"
+                      "lgdr   %%r0,%[alpha]    \n\t"
+                      "vlvgp  %%v0,%%r0,%%r0   \n\t"
+                      "srlg   %[n],%[n],4 \n\t"
+                      "vlr    %%v1,%%v0        \n\t"
+                      "vlm    %%v16,%%v23, 0(%[x_ptr])          \n\t"
+                      "la     %[x_ptr], 128(%[x_ptr])     \n\t"
+                      "aghik  %[n], %[n], -1             \n\t"
+                      "jle     2f     \n\t"
+                       ".align 16 \n\t"
+                      "1:          \n\t"
+                      "vfmdb  %%v24, %%v16, %%v0          \n\t"
+                      "vfmdb  %%v25, %%v17, %%v0          \n\t"
+                      "vfmdb  %%v26, %%v18, %%v0          \n\t"
+                      "vfmdb  %%v27, %%v19, %%v1          \n\t"
+                      "vlm     %%v16,%%v19, 0(%[x_ptr])         \n\t"
+                      "vfmdb  %%v28, %%v20, %%v0          \n\t"
+                      "vfmdb  %%v29, %%v21, %%v1          \n\t"
+                      "vfmdb  %%v30, %%v22, %%v0          \n\t"
+                      "vfmdb  %%v31, %%v23, %%v1          \n\t"
+                      "vlm     %%v20,%%v23, 64(%[x_ptr])         \n\t"
+                      "lay    %[x_ptr], -128(%[x_ptr])    \n\t"
+                      "vstm   %%v24,%%v31, 0(%[x_ptr])          \n\t"
+                      "la     %[x_ptr],256(%[x_ptr])      \n\t"
+                      "brctg %[n],1b     \n\t"
+                      "2:            \n\t"
+                      "vfmdb  %%v24, %%v16, %%v0          \n\t"
+                      "vfmdb  %%v25, %%v17, %%v1          \n\t"
+                      "vfmdb  %%v26, %%v18, %%v0          \n\t"
+                      "vfmdb  %%v27, %%v19, %%v1          \n\t"
+                      "lay    %[x_ptr] , -128(%[x_ptr])   \n\t"
+                      "vfmdb  %%v28, %%v20, %%v0          \n\t"
+                      "vfmdb  %%v29, %%v21, %%v1          \n\t"
+                      "vfmdb  %%v30, %%v22, %%v0          \n\t"
+                      "vfmdb  %%v31, %%v23, %%v1          \n\t"
+                      "vstm   %%v24,%%v31, 0(%[x_ptr])         \n\t"
+                      : [mem] "+m" (*(double (*)[n])x) ,[x_ptr] "+&a"(x),[n] "+&r"(n)
+                                       : [alpha] "f"(da)
+                                       :"cc" ,  "r0","v0","v1","v16","v17","v18","v19","v20","v21",
+                                       "v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
+                 );
+ }
+#else
 static void   dscal_kernel_32( BLASLONG n, FLOAT  da , FLOAT *x )
 {
 
@@ -71,7 +117,7 @@ static void   dscal_kernel_32( BLASLONG n, FLOAT  da , FLOAT *x )
                  );
 
  }
-
+#endif
 static void   dscal_kernel_32_zero( BLASLONG n,  FLOAT *x )
 {
    
@@ -214,6 +260,4 @@ int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da, FLOAT *x, BLAS
     }
     return 0;
 
-}
-
-
+}

kernel/zarch/zaxpy.c

@@ -28,126 +28,134 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #include "common.h"
 
+ 
 static void  zaxpy_kernel_8(BLASLONG n, FLOAT *x, FLOAT *y, FLOAT da_r,FLOAT da_i) {
-    __asm__ ("pfd   1, 0(%[x_tmp]) \n\t"
-            "pfd    2, 0(%[y_tmp]) \n\t"
-            "lgdr   %%r1,%[alpha_r]    \n\t"
-            "vlvgp  %%v28,%%r1,%%r1    \n\t"
-            "lgdr   %%r1,%[alpha_i]    \n\t"
-            "vlvgp  %%v29,%%r1,%%r1    \n\t"
-            "sllg   %[tmp],%[tmp],4    \n\t"
-            "xgr    %%r1,%%r1          \n\t"         
 
+    BLASLONG tempR1 ;
+    __asm__ ("pfd   1, 0(%[x_tmp]) \n\t"
+             "pfd    2, 0(%[y_tmp]) \n\t" 
+#if !defined(CONJ)
+            "lgdr   %[t1],%[alpha_r]    \n\t" 
+            "vlvgp  %%v28,%[t1],%[t1]   \n\t" //load both from disjoint          
+            "lgdr   %[t1],%[alpha_i]    \n\t"  
+            "vlvgp  %%v29,%[t1],%[t1]   \n\t" //load both from disjoint   
+            "vflcdb %%v29,%%v29       \n\t" //complement both
+            "vlvgg  %%v29,%[t1],1     \n\t" //restore 2nd  so that  {-alpha_i, alpha_i}   
+
+#else
+            "lgdr   %[t1],%[alpha_i]    \n\t"  
+            "vlvgp  %%v29,%[t1],%[t1]   \n\t" //load both from disjoint        
+            "lgdr   %[t1],%[alpha_r]    \n\t" 
+            "vlvgp  %%v28,%[t1],%[t1]   \n\t" //load both from disjoint    
+            "vflcdb %%v28,%%v28       \n\t" //complement both
+            "vlvgg  %%v28,%[t1],0     \n\t" //restore 1st  so that  {alpha_r,-alpha_r}   
+#endif           
+                               
+            "xgr     %[t1],%[t1]  \n\t" 
+            "sllg   %[tmp],%[tmp],4    \n\t" 
+            "vl   %%v30 ,  0(%[t1],%[y_tmp]) \n\t" 
+            "vl   %%v31 , 16(%[t1],%[y_tmp]) \n\t" 
+            "vl   %%v6 , 32(%[t1],%[y_tmp]) \n\t" 
+            "vl   %%v7 , 48(%[t1],%[y_tmp]) \n\t" 
+            "vl   %%v20 ,  0(%[t1],%[x_tmp]) \n\t" 
+            "vl   %%v21 , 16(%[t1],%[x_tmp]) \n\t" 
+            "vl   %%v22 , 32(%[t1],%[x_tmp]) \n\t" 
+            "vl   %%v23 , 48(%[t1],%[x_tmp]) \n\t"                         
+            "lay  %[tmp],-64 (%[tmp]) \n\t" //tmp-=64 so that t1+64 can break tmp condition
+            "j 2f \n\t"
             ".align 16 \n\t"
             "1:     \n\t"
-            "pfd    1, 256(%%r1,%[x_tmp]) \n\t"
-            "pfd    2, 256(%%r1,%[y_tmp]) \n\t"
-            "vleg   %%v16 ,  0(%%r1,%[y_tmp]),0 \n\t"
-            "vleg   %%v17 ,  8(%%r1,%[y_tmp]),0 \n\t"
-            "vleg   %%v16 , 16(%%r1,%[y_tmp]),1 \n\t"
-            "vleg   %%v17 , 24(%%r1,%[y_tmp]),1 \n\t"
-            "vleg   %%v18 , 32(%%r1,%[y_tmp]),0 \n\t"
-            "vleg   %%v19 , 40(%%r1,%[y_tmp]),0 \n\t"
-            "vleg   %%v18 , 48(%%r1,%[y_tmp]),1 \n\t"
-            "vleg   %%v19 , 56(%%r1,%[y_tmp]),1 \n\t"
-            "vleg   %%v24 ,  0(%%r1,%[x_tmp]),0 \n\t"
-            "vleg   %%v25 ,  8(%%r1,%[x_tmp]),0 \n\t"
-            "vleg   %%v24 , 16(%%r1,%[x_tmp]),1 \n\t"
-            "vleg   %%v25 , 24(%%r1,%[x_tmp]),1 \n\t"
-            "vleg   %%v26 , 32(%%r1,%[x_tmp]),0 \n\t"
-            "vleg   %%v27 , 40(%%r1,%[x_tmp]),0 \n\t"
-            "vleg   %%v26 , 48(%%r1,%[x_tmp]),1 \n\t"
-            "vleg   %%v27 , 56(%%r1,%[x_tmp]),1 \n\t"
-#if !defined(CONJ)
-            "vfmsdb %%v16,  %%v25, %%v29,%%v16  \n\t"
-            "vfmadb %%v17,  %%v24, %%v29, %%v17 \n\t"
-            "vfmsdb %%v18,  %%v27, %%v29, %%v18 \n\t"
-            "vfmadb %%v19,  %%v26, %%v29, %%v19 \n\t"
+  
+            "vpdi   %%v24 , %%v20, %%v20, 4     \n\t"
+            "vpdi   %%v25 , %%v21, %%v21, 4     \n\t"
+            "vpdi   %%v26 , %%v22, %%v22, 4     \n\t"
+            "vpdi   %%v27 , %%v23, %%v23, 4     \n\t" 
+            "vfmadb %%v16,  %%v20, %%v28, %%v16 \n\t"
+            "vfmadb %%v17,  %%v21, %%v28, %%v17 \n\t"
+            "vfmadb %%v18,  %%v22, %%v28, %%v18 \n\t"
+            "vfmadb %%v19,  %%v23, %%v28, %%v19 \n\t"
+            "vl     %%v30,  64(%[t1],%[y_tmp])  \n\t" 
+            "vl     %%v31,  80(%[t1],%[y_tmp])  \n\t" 
+            "vl     %%v6 ,  96(%[t1],%[y_tmp])  \n\t" 
+            "vl     %%v7 , 112(%[t1],%[y_tmp])  \n\t" 
+            "vfmadb %%v16,  %%v24, %%v29, %%v16 \n\t"
+            "vfmadb %%v17,  %%v25, %%v29, %%v17 \n\t" 
+            "vfmadb %%v18,  %%v26, %%v29, %%v18 \n\t"
+            "vfmadb %%v19,  %%v27, %%v29, %%v19 \n\t"
+            "vl     %%v20 , 64(%[t1],%[x_tmp])  \n\t" 
+            "vl     %%v21 , 80(%[t1],%[x_tmp])  \n\t"      
+            "vl     %%v22 , 96(%[t1],%[x_tmp])  \n\t" 
+            "vl     %%v23 ,112(%[t1],%[x_tmp])  \n\t" 
 
-            "vfmsdb %%v16, %%v24, %%v28 ,%%v16  \n\t"
-            "vfmadb %%v17, %%v25, %%v28, %%v17  \n\t"
-            "vfmsdb %%v18, %%v26, %%v28, %%v18  \n\t"
-            "vfmadb %%v19, %%v27, %%v28, %%v19  \n\t"
-#else
-            "vfmadb %%v16, %%v25, %%v29, %%v16  \n\t"
-            "vfmsdb %%v17, %%v25, %%v28, %%v17  \n\t"
-            "vfmadb %%v18, %%v27, %%v29, %%v18  \n\t"
-            "vfmsdb %%v19, %%v27, %%v28, %%v19  \n\t"
-            "vfmadb %%v16, %%v24, %%v28, %%v16  \n\t"
-            "vfmsdb %%v17, %%v24, %%v29, %%v17  \n\t"
-            "vfmadb %%v18, %%v26, %%v28, %%v18  \n\t"
-            "vfmsdb %%v19, %%v26, %%v29, %%v19  \n\t"
+            "vst    %%v16 ,  0(%[t1],%[y_tmp])  \n\t" 
+            "vst    %%v17 , 16(%[t1],%[y_tmp])  \n\t" 
+            "vst    %%v18 , 32(%[t1],%[y_tmp])  \n\t" 
+            "vst    %%v19 , 48(%[t1],%[y_tmp])  \n\t"   
+    
+            "la     %[t1],64(%[t1] ) \n\t" 
+            "2:  \n\t"
+            "pfd    1, 256(%[t1],%[x_tmp])  \n\t"
+            "pfd    2, 256(%[t1],%[y_tmp])  \n\t"  
+            "vpdi   %%v24 , %%v20, %%v20, 4     \n\t"
+            "vpdi   %%v25 , %%v21, %%v21, 4     \n\t"
+            "vpdi   %%v26 , %%v22, %%v22, 4     \n\t"
+            "vpdi   %%v27 , %%v23, %%v23, 4     \n\t" 
 
-#endif 
-            "vsteg %%v16 ,  0(%%r1,%[y_tmp]),0  \n\t"
-            "vsteg %%v17 ,  8(%%r1,%[y_tmp]),0  \n\t"
-            "vsteg %%v16 , 16(%%r1,%[y_tmp]),1  \n\t"
-            "vsteg %%v17 , 24(%%r1,%[y_tmp]),1  \n\t"
+            "vfmadb %%v30,  %%v20, %%v28, %%v30 \n\t"
+            "vfmadb %%v31,  %%v21, %%v28, %%v31 \n\t"
+            "vfmadb %%v6,  %%v22, %%v28, %%v6   \n\t"
+            "vfmadb %%v7,  %%v23, %%v28, %%v7   \n\t"
+            "vl     %%v16,  64(%[t1],%[y_tmp])  \n\t" 
+            "vl     %%v17,  80(%[t1],%[y_tmp])  \n\t" 
+            "vl     %%v18,  96(%[t1],%[y_tmp])  \n\t" 
+            "vl     %%v19, 112(%[t1],%[y_tmp])  \n\t" 
+            "vfmadb %%v30,  %%v24, %%v29, %%v30 \n\t"
+            "vfmadb %%v31,  %%v25, %%v29, %%v31 \n\t"
+            "vfmadb %%v6,  %%v26, %%v29, %%v6   \n\t"
+            "vfmadb %%v7,  %%v27, %%v29, %%v7   \n\t"
 
-            "vsteg %%v18 , 32(%%r1,%[y_tmp]),0  \n\t"
-            "vsteg %%v19 , 40(%%r1,%[y_tmp]),0  \n\t"
-            "vsteg %%v18 , 48(%%r1,%[y_tmp]),1  \n\t"
-            "vsteg %%v19 , 56(%%r1,%[y_tmp]),1  \n\t"
+            "vl     %%v20 , 64(%[t1],%[x_tmp])  \n\t" 
+            "vl     %%v21 , 80(%[t1],%[x_tmp])  \n\t"  
+            "vl     %%v22 , 96(%[t1],%[x_tmp])  \n\t" 
+            "vl     %%v23 ,112(%[t1],%[x_tmp])  \n\t" 
 
-            "vleg %%v20 , 64(%%r1,%[y_tmp]),0   \n\t"
-            "vleg %%v21 , 72(%%r1,%[y_tmp]),0   \n\t"
-            "vleg %%v20 , 80(%%r1,%[y_tmp]),1   \n\t"
-            "vleg %%v21 , 88(%%r1,%[y_tmp]),1   \n\t"
+            "vst    %%v30 ,  0(%[t1],%[y_tmp])  \n\t" 
+            "vst    %%v31 , 16(%[t1],%[y_tmp])  \n\t" 
+            "vst    %%v6 ,  32(%[t1],%[y_tmp])  \n\t" 
+            "vst    %%v7 ,  48(%[t1],%[y_tmp])  \n\t"  
+ 
+            "la     %[t1],64(%[t1] ) \n\t"
+          
 
-            "vleg %%v22 ,  96(%%r1,%[y_tmp]),0  \n\t"
-            "vleg %%v23 , 104(%%r1,%[y_tmp]),0  \n\t"
-            "vleg %%v22 , 112(%%r1,%[y_tmp]),1  \n\t"
-            "vleg %%v23 , 120(%%r1,%[y_tmp]),1  \n\t"
+             "clgrjl %[t1],%[tmp],1b         \n\t"   
+//----------------------------------------------------------------------
+            "vfmadb %%v16,  %%v20, %%v28, %%v16 \n\t"
+            "vfmadb %%v17,  %%v21, %%v28, %%v17 \n\t"
+            "vfmadb %%v18,  %%v22, %%v28, %%v18 \n\t"
+            "vfmadb %%v19,  %%v23, %%v28, %%v19 \n\t" 
+            "vpdi   %%v24 , %%v20, %%v20, 4     \n\t"
+            "vpdi   %%v25 , %%v21, %%v21, 4     \n\t" 
+            "vpdi   %%v26 , %%v22, %%v22, 4     \n\t"
+            "vpdi   %%v27 , %%v23, %%v23, 4     \n\t"             
+            "vfmadb %%v16,  %%v24, %%v29, %%v16 \n\t"
+            "vfmadb %%v17,  %%v25, %%v29, %%v17 \n\t"
+            "vfmadb %%v18,  %%v26, %%v29, %%v18 \n\t"
+            "vfmadb %%v19,  %%v27, %%v29, %%v19 \n\t"
 
-            "vleg %%v24 , 64(%%r1,%[x_tmp]),0   \n\t"
-            "vleg %%v25 , 72(%%r1,%[x_tmp]),0   \n\t"
-            "vleg %%v24 , 80(%%r1,%[x_tmp]),1   \n\t"
-            "vleg %%v25 , 88(%%r1,%[x_tmp]),1   \n\t"
+            "vst   %%v16 ,  0(%[t1],%[y_tmp])   \n\t" 
+            "vst   %%v17 , 16(%[t1],%[y_tmp])   \n\t" 
+            "vst   %%v18 , 32(%[t1],%[y_tmp])   \n\t" 
+            "vst   %%v19 , 48(%[t1],%[y_tmp])   \n\t"   
 
-            "vleg %%v26 ,  96(%%r1,%[x_tmp]),0  \n\t"
-            "vleg %%v27 , 104(%%r1,%[x_tmp]),0  \n\t"
-            "vleg %%v26 , 112(%%r1,%[x_tmp]),1  \n\t"
-            "vleg %%v27 , 120(%%r1,%[x_tmp]),1  \n\t"
-#if !defined(CONJ)
-            "vfmsdb %%v20,  %%v25, %%v29,%%v20   \n\t"
-            "vfmadb %%v21,  %%v24, %%v29, %%v21  \n\t"
-            "vfmsdb %%v22,  %%v27, %%v29, %%v22  \n\t"
-            "vfmadb %%v23,  %%v26, %%v29, %%v23  \n\t"
-
-            "vfmsdb %%v20, %%v24, %%v28 ,%%v20   \n\t"
-            "vfmadb %%v21, %%v25, %%v28, %%v21   \n\t"
-            "vfmsdb %%v22, %%v26, %%v28, %%v22   \n\t"
-            "vfmadb %%v23, %%v27, %%v28, %%v23   \n\t"
-#else
-            "vfmadb %%v20, %%v25, %%v29, %%v20   \n\t"
-            "vfmsdb %%v21, %%v25, %%v28, %%v21   \n\t"
-            "vfmadb %%v22, %%v27, %%v29, %%v22   \n\t"
-            "vfmsdb %%v23, %%v27, %%v28, %%v23   \n\t"
-            "vfmadb %%v20, %%v24, %%v28, %%v20   \n\t"
-            "vfmsdb %%v21, %%v24, %%v29, %%v21   \n\t"
-            "vfmadb %%v22, %%v26, %%v28, %%v22   \n\t"
-            "vfmsdb %%v23, %%v26, %%v29, %%v23   \n\t"
-#endif 
-            "vsteg %%v20 , 64(%%r1,%[y_tmp]),0   \n\t"
-            "vsteg %%v21 , 72(%%r1,%[y_tmp]),0   \n\t"
-            "vsteg %%v20 , 80(%%r1,%[y_tmp]),1   \n\t"
-            "vsteg %%v21 , 88(%%r1,%[y_tmp]),1   \n\t"
-
-            "vsteg %%v22 ,  96(%%r1,%[y_tmp]),0  \n\t"
-            "vsteg %%v23 , 104(%%r1,%[y_tmp]),0  \n\t"
-            "vsteg %%v22 , 112(%%r1,%[y_tmp]),1  \n\t"
-            "vsteg %%v23 , 120(%%r1,%[y_tmp]),1  \n\t"
-
-            "la     %%r1,128(%%r1) \n\t"
-            "clgrjl %%r1,%[tmp],1b        \n\t"   
-            : [mem_y] "+m" (*(double (*)[2*n])y),[tmp]"+&r"(n)
+            : [mem_y] "+m" (*(double (*)[2*n])y),[tmp]"+&r"(n) ,  [t1] "=&a" (tempR1) 
             : [mem_x] "m" (*(const double (*)[2*n])x), [x_tmp] "a"(x), [y_tmp] "a"(y), [alpha_r] "f"(da_r),[alpha_i] "f"(da_i)
-            : "cc",  "r1","v16",
-            "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29"
+            : "cc",  "v6","v7", "v16",
+            "v17","v18","v19","v20","v21","v22","v23","v24","v25","v26","v27","v28","v29","v30","v31"
             );
 
 }
 
+
 int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT da_r, FLOAT da_i, FLOAT *x, BLASLONG inc_x, FLOAT *y, BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2) {
     BLASLONG i = 0;
     BLASLONG ix = 0, iy = 0;