==============================================
Using GCC Vector-Extensions for  Micro-Kernels                            [TOC:2]
==============================================

Tar-Ball for this Session
=========================
The tar-ball __session5.tgz__ contains the files:

---- SHELL ---------------------------------------------------------------------
tar tfvz session5.tgz
--------------------------------------------------------------------------------

:links: session5.tgz -> http://www.mathematik.uni-ulm.de/~lehn/test_ublas/session5.tgz

Compiling for FMA or AVX
========================

- If you have FMA (i.e. AVX2 with fused mult-add) compile with

  `g++ -Wall -DNDEBUG -mfma -Ofast -I ../../boost_1_60_0/ -std=c++11 -DHAVE_GCCVEC -DBS_D_NR=12 -DBS_D_NC=4092 matprod.cc`

- If you have AVX but not FMA with

  `g++ -Wall -DNDEBUG -mavx -Ofast -I ../../boost_1_60_0/ -std=c++11 -DHAVE_GCCVEC matprod.cc`


Compile and Run Benchmark
=========================
The machine this page gets generated only has AVX.  Also note that the GCC vector
extensions need a recent GCC compiler.  Here we will use GCC 5.3:

---- SHELL (path=session5,hostname=heim) ---------------------------------------
g++-5.3 -Wall -DNDEBUG -mavx -Ofast -I ../../boost_1_60_0/ -std=c++11 -DHAVE_GCCVEC -DM_MAX=500 matprod.cc
./a.out
--------------------------------------------------------------------------------


Micro-Kernel
============
:import: session5/gccvec.hpp

Modified `gemm.hpp`
===================
:import: session5/gccvec.hpp

Benchmark Results for GEMM: Single-Threaded
===========================================

Resuts
------

---- SHELL (path=session5,hostname=heim) ---------------------------------------
g++-5.3 -Wall -DNDEBUG -mavx -Ofast -I ../../boost_1_60_0/ -std=c++11 -DHAVE_GCCVEC matprod.cc
./a.out  > report.session5
cat report.session5
gnuplot plot.session5.mflops
gnuplot plot.session5.time
gnuplot plot.session5.time_log
--------------------------------------------------------------------------------


MFLOPS
------
---- IMAGE ----------------------------
session5/bench.session5.mflops.svg
---------------------------------------


Time
----

---- IMAGE --------------------------
session5/bench.session5.time.svg
-------------------------------------

Time with Logarithmic scale
---------------------------

---- IMAGE ------------------------------
session5/bench.session5.time_log.svg
-----------------------------------------


Benchmark Results for GEMM: Multi-Threaded
==========================================

Resuts
------

---- SHELL (path=session5,hostname=heim) ---------------------------------------
g++-5.3 -Wall -DNDEBUG -mavx -Ofast -I ../../boost_1_60_0/ -std=c++11 -DHAVE_GCCVEC -fopenmp matprod.cc
./a.out  > report.mt.session5
cat report.mt.session5
gnuplot plot.mt.session5.mflops
gnuplot plot.mt.session5.time
gnuplot plot.mt.session5.time_log
--------------------------------------------------------------------------------


MFLOPS
------
---- IMAGE ----------------------------
session5/bench.mt.session5.mflops.svg
---------------------------------------


Time
----

---- IMAGE --------------------------
session5/bench.mt.session5.time.svg
-------------------------------------

Time with Logarithmic scale
---------------------------

---- IMAGE ------------------------------
session5/bench.mt.session5.time_log.svg
-----------------------------------------



:navigate: back  -> doc:session4/page01
           next  -> doc:session6/page01