Contains the linear algebra kernels in C and BLAS equivalents for both OpenCL and CUDA
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
CUDA
OpenCL
polybench-c-3.2
README

README

Required software:


* PPCG and dependent packages
* Autotuner 
* OpenCL or Cuda SDK 5.5
* AMD CLBLAS 2.2.0 https://github.com/clMathLibraries/clBLAS
* CuBLAS 5.5

OpenCl Blas kernels
OpneCL_Dir = {Benchmark}/Opencl
* modify the OpenCL_SDK and BLAS_SDK varibles according to your machine installation in {OpenCL_Dir}/common.mk
* to build, execute make blas command in each of the benchmark directories 
* run the resulting executable


CUDA Blas Kernels
CUDA_DIR = {Benchmark}/CUDA
* make sure nvcc compiler is installed
* to build, execute make blas command in each of the benchmark directories
* run the resulting executable

PPCG generated kernels
Polybench_DIR = {Benchmark}/polybench-c-3.2
* Setup PPCG_DIR, OpenCL_DIR, AUTOTUNE_DIR, PENCIL_UTIL and POLYBENCH variables in {Polybench_DIR}/common.mk
* Add ${PENCIL_UTIL}/runtime/src/.libs directory to the LD_LIBRARY_PATH environment variable
* to build and run, execute make opencl or make cuda in each of the benchmark directories
* run make autotune_cuda or autotune_opencl to launch autotuning. Autotuning parameters can be changed in expolre-params.py in each benchmark directory. Note that some of the parameters could be invalid on a particular machine.
* Setup the TARGET variable in {Polybench_DIR}/commom.mk accordingly
* Best configuration found after autotuning will be printed in {benchmark}_autotune.log
* Update these configurations in the Makefile inside individual directory and do make opencl/make cuda, to execute the benchmark with the best configuration