minhhn2910 / cuda-mixed-precision Goto Github PK

Mixed precision between FP32 and FP16x2 in CUDA programs This repo contains all benchmarks for mixed precision tuning between FP16 and FP32 in CUDA program (and some extension to FP32-FP64, microbenchmarks). The below is the function of each subfolder (and how to run benchmarks in them). There will be further README file in each benchmark to guide the execution and performace measurement if they are not standard:

1. include/ contains the most important part: the operator overload headers and the approximate math library. Without this, the FP16 benchmarks cannot compile.

1. Nvidia_example/benchmarks/: Each subfolder with the postfix :_mixed is the mixed precision version. The makefiles are provided. The run_and_compare.sh script is used to measure performance + error with the respective input parameters. For this category, there are 3 benchmarks:
   • Nvidia_example/benchmarks/bilateralFilter_mixed
   • Nvidia_example/benchmarks/nbody_mixed
   • Nvidia_example/benchmarks/matmul_sharedmem/matrixMul_mixed_final
2. axbench/: There are two benchmarks. For this benchmarks suite, we use the compile.sh script inside each benchmark. If your GPU architecture is different from Volta, please change the arch flag for correct compiling. The run_and_compare.sh script now takes a speed parameter, which is the number of approximated blocks. It ranges from 0 to 2048. e.g. ./run_and_compare.sh 1024
   • inversek2j_mixed
   • newton-raph_mixed
3. parboil/. There are 2 benchmarks. To compile them, run compile_mixed.sh or build_mixed.sh. To change the approximation degree, please go to the subfolder cuda_mixed inside each benchmark to change the SPEED definition at the beginning of the code. In the code, there will be comment mentioning the range of the SPEED value. Again, ./run_and_compare.sh is used to run and show the output error.
   • mri-q
   • cp
4. rodinia/. There are 2 benchmarks. To compile them, use the respective make file inside the subfolders. The method to change the approximation degree is similar to parboil.
   • kmeans_mixed
   • lavaMD_mixed

1. Nvidia_example/benchmarks/BlackScholes_mixed contains the mixed precision method for blackscholes. The same compile and run method as Nvidia_example/in the previous section. This subfolder also contains the search script on python.
2. lulesh/lulesh_new/lulesh_mixed LULESH example. The makefile and run.sh are provided. The SPEED parameter need to be changed.
3. funarc/funarc_mixed_fp3264/ contains the arclength example.
4. microbenchmarkSFU/benchmarks : This contains the microbenchmarks and compile scripts to benchmark the math functions in CUDA. There will be a README file explains how to rerun the microbenchmarks

The input data are provided in this repo, there is no need to generate input (except for kmeans, where the randomly generated input is too large). You can generate input for kmeans by going to the rodinia/data/kmeans/inpuGen/, type make ; ./datagen 640000 -f.

Todo: This is a mirror of the private developing bitbucket repo, we update the code when each release come out. This version is release v2, contains automation scripts and math library v2. Gradually update README files as I dont have enough time during deadline.