CERN Workshop on Numerical Computing, May 2014

Agner Fog

Methods for vectorization of mathematical software and performance measurements

Files:

Test different vector implementations of mathematical functions

Unpack mathtest.zip to a separate folder.
Compile and run testvectormath.cpp with appropriate command line parameters, e.g.
g++ -Ofast -mavx testvectormath.cpp Run several times with ./a.out to see if you can get consistent timings.

Things you can try to vary:

Compile for different instruction sets, e.g. -msse2
Try different vector sizes by changing VECTOR_TYPE in testvectormath.cpp
Try different math libraries by changing MAT_IMPLEMENT. If you want to use the Intel SVML library then add /opt/intel/lib/intel64/libsvml.a on the g++ command line.
Try different mathematical functions by changing MAT_FUNCTION
Measure with performance monitor counters by setting e.g. #define USEPMC 0x40000001.
See below how to set up the performance monitor counters.

Test with performance monitor counters

Unpack testp.zip to a separete folder. Then unpack DriverSrcLinux.zip and PMCTest.zip. Install the driver as explained in testp.pdf. Make all scripts executable with chmod a+x *.sh

To test a small piece of C++ code: Edit PMCTestB.cpp, and run ./c64.sh

To test a small piece of asm code: Edit PMCTestB64.nasm, and run ./a64.sh

To enable performance monitor counters for use in another program: Run ./c64.sh, then run ./a.out startcounters
Now you can use the performance monitor counters in testvectormath.cpp or in your own program. You can specify different performance monitor counters in PMCTestB.cpp or on the command line.

To measure latency and throughput of all machine instructions, unpack TestScripts.zip and run some of the scripts therein. You may need to update nasm as described in testp.pdf.

Powerpoint from my talk

link