OpenCL-based data-parallel primitives (including gather, scatter, scan and split) on heterogenous processors (GPUs, CPUs and 1st generation MICs).
This project uses CMake for compilation.
cd opencl
mkdir -p build && cd build
cmake -DOpenCL_LIBRARY="PATH_TO_OPENCL_LIBRARY" .. (the address of libOpenCL.so)
make -j
If CUDA is installed, libOpenCL.so is usually located at cuda/targets/x86_64-linux/lib. The OpenCL drivers for Intel CPUs and MICs should be installed manually if running the code on CPUs and MICs.
./test_access : test the performance of column-major order, row-major order and mixed order sequential access patters
./test_access_wg DATA_NUM : test the work-group-wise sequential access efficiency on specified number of input data
./test_bandwidth : test the sequential bandwidth with the Stream Benchmark (copy, scalar, addition and triad operations)
./test_gather DATA_NUM : test the performance of multi-pass gather on specified number of data
./test_scatter DATA_NUM : test the performance of multi-pass scatter on specified number of data
./test_scan_local : test the performance of local scan schemes
./test_scan_global : test the performance of global scan schemes
./test_split : test the performance of split
This project uses CMake for compilation.
cd cuda
mkdir -p build && cd build
cmake ..
make -j
./test_bandwidth : test the sequential bandwidth with the Stream Benchmark (copy, scalar, addition and triad operations)
./test_gather DATA_NUM : test the performance of multi-pass gather on specified number of data
./test_scatter DATA_NUM : test the performance of multi-pass scatter on specified number of data
./test_scan : test the performance of CUB scan
./test_split : test the performance of GPU multi-split
This project uses CMake for compilation.
cd cuda
mkdir -p build && cd build
cmake ..
make -j
./test_bandwidth_CPU : test the sequential bandwidth with the Stream Benchmark (copy and scalar)
./test_gather_scatter_CPU DATA_NUM : test the performance of gather and scatter on specified number of data
./test_scan_CPU : test the performance of OpenMP-based SSA, RTS scan and TBB scan