This is a repository aimed at providing GPU parallel codes with different parallel APIs for the NAS Parallel Benchmarks (NPB) from a C/C++ version (NPB-CPP). You can also contribute with this project, writing issues and pull requests. 😄
🔉News: Parametrization support for configuring number of threads per block and CUDA parallelism optimizations. 📅25/Jul/2021
🔉News: CUDA versions for pseudo-applications added and IS improved. 📅11/Feb/2021
🔉News: Paper published in the journal Software: Practice and Experience (SPE). 📅29/Nov/2021
DOI - Araujo, G.; Griebler, D.; Rockenbach, D. A.; Danelutto, M.; Fernandes, L. G.; NAS Parallel Benchmarks with CUDA and beyond, Software: Practice and Experience (SPE), 2021.
DOI - Araujo, G.; Griebler, D.; Danelutto, M.; Fernandes, L. G.; Efficient NAS Benchmark Kernels with CUDA. 28th Euromicro International Conference on Parallel, Distributed and Network-based Processing (PDP), Västerås, 2020.
The parallel CUDA version was implemented from the serial version of NPB-CPP.
==================================================================
NAS Parallel Benchmarks code contributors with CUDA are:
Dalvan Griebler: dalvan.griebler@pucrs.br
Gabriell Araujo: gabriell.araujo@edu.pucrs.br
==================================================================
Each directory is independent and contains its own implemented version:
Five kernels
- IS - Integer Sort
- EP - Embarrassingly Parallel
- CG - Conjugate Gradient
- MG - Multi-Grid
- FT - discrete 3D fast Fourier Transform
Three pseudo-application
- SP - Scalar Penta-diagonal solver
- BT - Block Tri-diagonal solver
- LU - Lower-Upper Gauss-Seidel solver
Warning: our tests were made with GCC and CUDA
Go inside the directory CUDA directory and execute:
make _BENCHMARK CLASS=_VERSION
_BENCHMARKs are:
CG, EP, FT, IS, MG, BT, LU, and SP
_VERSIONs are:
-
Class S: small for quick test purposes
-
Class W: workstation size (a 90's workstation; now likely too small)
-
Classes A, B, C: standard test problems; ~4X size increase going from one class to the next
-
Classes D, E, F: large test problems; ~16X size increase from each of the previous Classes
Command example:
make ep CLASS=B
NPB-GPU has additional timers for profiling purpose. To activate these timers, create a dummy file 'timer.flag' in the main directory of the NPB version (e.g. CUDA/timer.flag).
NPB-GPU allows configuring the number of threads per block of each GPU kernel in the benchmarks. The user can specify the number of threads per block by editing the file gpu.config in the directory /config/. If no file is specified, all GPU kernels are executed using the warp size of the GPU as the number of threads per block.
Syntax of the gpu.config file:
<benchmark-name>_THREADS_PER_BLOCK_<gpu-kernel-name> = <interger-value>
Configuring CG benchmark as example:
CG_THREADS_PER_BLOCK_ON_KERNEL_ONE = 32
CG_THREADS_PER_BLOCK_ON_KERNEL_TWO = 128
CG_THREADS_PER_BLOCK_ON_KERNEL_THREE = 64
CG_THREADS_PER_BLOCK_ON_KERNEL_FOUR = 256
CG_THREADS_PER_BLOCK_ON_KERNEL_FIVE = 32
CG_THREADS_PER_BLOCK_ON_KERNEL_SIX = 64
CG_THREADS_PER_BLOCK_ON_KERNEL_SEVEN = 128
CG_THREADS_PER_BLOCK_ON_KERNEL_EIGHT = 64
CG_THREADS_PER_BLOCK_ON_KERNEL_NINE = 512
CG_THREADS_PER_BLOCK_ON_KERNEL_TEN = 512
CG_THREADS_PER_BLOCK_ON_KERNEL_ELEVEN = 1024
The NPB-GPU also allows changing the GPU device by providing the following syntax in the gpu.config file:
GPU_DEVICE = <interger-value>
At first, the compilation flags were set to GPU with CUDA compute capability version of 6.1 (GPU used by the original authors of this repository for the research purposes was GeForce GTX 1080 Ti)
This error shows, that GPU code was not compiled correctly. In order to fix it, change directory to /CUDA/config and edit the "make.def" file. Then, edit this line, by replacing number 61 with corresponding CUDA compute capability:
COMPUTE_CAPABILITY = -gencode arch=compute_61,code=sm_61
One can check the compute capability of their GPU, by checking the table on Wikipedia:
https://en.wikipedia.org/wiki/CUDA