Skip to content

Latest commit

 

History

History
104 lines (76 loc) · 3.79 KB

README.md

File metadata and controls

104 lines (76 loc) · 3.79 KB

QCT oneAPI Experiments - Polybench

  • Variants
    • CUDA @ GPU CUDA
    • CUDA -> SYCL using DPCT @ CPU DPCT_CPU
    • CUDA -> SYCL using DPCT @ GPU DPCT_GPU
    • sycl-bench SYCL @ CPU SYCL_CPU
    • sycl-bench SYCL @ GPU SYCL_GPU

Running experiments

CUDA @ GPU CUDA

Requires CUDA toolkit

  1. make_all_cuda.sh
    • Build all CUDA benchmarks using nvcc

    Makefile is at ../CUDA/utilities/common.mk, target cuda

  2. run_all_cuda.sh
    • Run all CUDA benchmarks

    Outputs are written to output.cuda.txt in the testcase folders

CUDA -> SYCL using DPCT @ CPU DPCT_CPU

Requires oneAPI DPC++ compiler & DPCT

  1. dpct_convert_all.sh

    • Generates SYCL code from CUDA code using DPCT

    Converted code are already in the repo, so this step is not required.

  2. make_all_dpct_cpu.sh

    • Build all DPCT converted SYCL code targeting Intel CPU with icpx

    Makefile is at ../CUDA/utilities/common.mk, target dpct_cpu

  3. run_all_dpct_cpu.sh

    • Run all DPCT benchmarks built with CPU target

    Run 5 times by default. Override using REPEAT=x ./run_all_dpct_cpu.sh

    Outputs are written to output.dpct_cpu.txt in the testcase folders

CUDA -> SYCL using DPCT @ GPU DPCT_GPU

Requires intel/llvm built with CUDA & DPCT

  1. dpct_convert_all.sh

    • Generates SYCL code from CUDA code using DPCT

    Converted code are already in the repo, so this step is not required.

  2. make_all_dpct_gpu.sh

    • Build all DPCT converted SYCL code targeting NVIDIA GPU with intel/llvm's clang++

    Makefile is at ../CUDA/utilities/common.mk, target dpct_cpu

    Note: change --cuda-gpu-arch= if you are not using Volta architecture GPU

  3. run_all_dpct_gpu.sh

    • Run all DPCT benchmarks built with GPU target

    Run 5 times by default. Override using REPEAT=x ./run_all_dpct_cpu.sh

    Outputs are written to output.dpct_gpu.txt in the testcase folders

sycl-bench SYCL @ CPU SYCL_CPU

Requires oneAPI DPC++ compiler

Make sure the sycl-bench submodule is cloned. If not, run git submodule update --init

  1. make_all_sycl_cpu.sh
    • Build all SYCL-Bench polybench codes targeting Intel CPU with icpx
  2. run_all_sycl_cpu.sh
    • Run all SYCL polybench benchmarks built with CPU target

    Run 5 times by default. Override using REPEAT=x ./run_all_dpct_cpu.sh Outputs are written to output.sycl_cpu.txt in sycl-bench/build

sycl-bench SYCL @ CPU SYCL_GPU

Requires oneAPI DPC++ compiler

Make sure the sycl-bench submodule is cloned. If not, run git submodule update --init

  1. make_all_sycl_gpu.sh
    • Build all SYCL-Bench polybench codes targeting NVIDIA GPU with intel/llvm's clang++

    Note: change --cuda-gpu-arch= if you are not using Volta architecture GPU

  2. run_all_sycl_gpu.sh
    • Run all SYCL polybench benchmarks built with GPU target

    Run 5 times by default. Override using REPEAT=x ./run_all_dpct_cpu.sh Outputs are written to output.sycl_gpu.txt in sycl-bench/build

Generating reports

Tip: to remove old experiments, use git clean -fxdn, and remove the -n option to perform deletion

Recommendation: rename results(_sycl-bench).csv to results(_sycl-bench)-[host].csv before commiting

CUDA, DPCT_CPU and DPCT_GPU

  • Run python3 parse_log.py
    • Outputs to results.csv

SYCL_CPU and SYCL_GPU

  • Run python3 parse_log_sycl-bench.py
    • Outputs to results_sycl-bench.csv

Changing dataset sizes

  1. Edit ../CUDA/utilities/common.mk and add -D<SIZE>_DATASET to compile args
    • Dataset sizes are: MINI_DATASET, SMALL_DATASET, STANDARD_DATASET(default), LARGE_DATASET, EXTRALARGE_DATASET
  2. Edit gen_sizes.py and set the DATASET variable
  3. Run gen_sizes.py to export the sizes of the testcases to sycl-bench