search

GMAP / NPB-GPU

NAS Parallel Benchmarks for evaluating GPU and APIs
Other
18 stars 7 forks source link

choosing GPU_DEVICE from gpu.config doesn't work #10

Open odiakun opened 4 months ago

odiakun commented 4 months ago

For LU, BT and SP algorithms, device 0 is selected every time, even after a change of GPU_DEVICE in the gpu.config file.

gabriellaraujocoding commented 4 months ago

Hello, @odiakun.

First of all, thank you for using the NPB-GPU.

Second, are you recompiling the benchmark after modifying the gpu.config file? After making changes to the gpu.config file, it's essential to recompile the benchmark for the modifications to be applied and take effect.

Third, have you downloaded the latest version of NPB-GPU? We updated NPB-GPU a few days ago to automatically collect the compute capability according to the GPU_DEVICE in the gpu.config file.

In the latest version of NPB-GPU, when you compile a benchmark, the terminal will show a message if a GPU device is not valid. For example:

The GPU id 5 is not valid. Using GPU 0 as default.
Updated COMPUTE_CAPABILITY (GPU of id 0) in ../config/make.def to -gencode arch=compute_89,code=sm_89

I have a few questions for you:

  1. Is the GPU device being correctly selected in the other benchmarks CG, EP, FT, IS, and MG?
  2. When we compile a benchmark, it generates a file named npbparams.hpp in the directory of the benchmark you are compiling. Can you share the npbparams.hpp file generated?
  3. Can you share the output of the compilation and execution?
  4. What are your GPUs and CUDA versions?
  5. If the GPU_DEVICE occurs several times in the gpu.config file, NPB-GPU collects only the first occurrence. You provided the GPU device info in the first lines of the gpu.config file, right?
gabriellaraujocoding commented 4 months ago

Could you modify the gpu.config file and recompile the benchmark?

dalvangriebler commented 1 month ago

@odiakun is the issue solved?

odiakun commented 1 month ago

Hello,Due to the limited scope of my Master Thesis and the time within I had to finish it,  I could not afford to apply the changes you suggested. Although my thesis got a high grade and I’ll publish an article based on that, so when it’s official I’ll let you know. I’m really grateful for your help! Also further examination of the topic still can be made, so I’m sure I won’t forget your suggestions in the future work. Best regards,Oksana Diakun. Wiadomość napisana przez Dalvan Griebler @.***> w dniu 21.09.2024, o godz. 00:17: @odiakun is the issue solved?

—Reply to this email directly, view it on GitHub, or unsubscribe.You are receiving this because you were mentioned.Message ID: @.***>