TRNS CPU GPU

[trinayan]

Hi,
The Chai paper mentions that SC , PAD and TRNS support only GPU execution in the OPENCL-D benchmarks. I did not try the OpenCL-D ones but the CUDA-D SC and PAD version is CPU-GPU executing together. I was wondering if it is also possible to do this for TRNS of CUDA-D as well or it is not possible at all? Thanks

Best,
Trinayan

[el1goluj]

Hi Trinayan,

As stated in the Chai paper, inter-worker synchronization between CPU and GPU workers is not possible without system-wide atomics. For this reason, in the paper we use GPU-only version for PAD, SC and TRNS when we compare OpenCL-D and OpenCL-U versions (Figure 2).
PAD and SC -D versions in the repository support CPU+GPU out-of-place implementations that do not require CPU-GPU inter-worker synchronization. The input arrays are divided into two parts, each of which is assigned to CPU or GPU. These versions are not directly comparable to PAD and SC -U versions, which are in-place.
TRNS uses multiple depending cycles for concurrency (see Sung et al., Innovative Parallel Computing, 2012). Elements in a depending cycle are scattered across the whole matrix, making the elements hard to be collected in -D version. Thus, the input array cannot easily split into two parts as in PAD and SC.

Juan

[trinayan]

Hi,

Thanks a lot for this information. Now I understand clearly. Is it also possible to generate larger input sets for Bezier Surface in a manner similar to the other benchmarks ?

Best,
Trinayan

[el1goluj]

Yes, for BS you can use:
-m : input size in both dimensions (default=3)
-n : output resolution in both dimensions (default=300)

[robers97]

As stated in the Chai paper, inter-worker synchronization between CPU and GPU workers is not possible without system-wide atomics. For this reason, in the paper we use GPU-only version for PAD, SC and TRNS when we compare OpenCL-D and OpenCL-U versions (Figure 2).

TRANS appears to be a good focus to test implementations of system wide atomics wither that is in software: https://docs.nvidia.com/cuda/pascal-tuning-guide/index.html or hardware (TBD). I see you have the CUDA_8_0 flag in the code, but haven't gotten to the point where that compiles though I'm running CUDA 9.2.

Can we revisit this together? I'm willing to contribute back.

[el1goluj]

Hi robers97,

Could you be more specific about your question?
The CUDA-U version uses system-wide atomics. It is tested with CUDA 8.0 (first CUDA version with system-wide atomics), thus it should work with CUDA 9.2. You will need Pascal or Volta GPU.
Yes, TRNS is a good benchmark to test implementations of system-wide atomics. Actually all CUDA-U benchmarks use them, and might be useful for you.

Thanks,
Juan