This repository was archived by the owner on Mar 20, 2023. It is now read-only.
Set by default the number of warps #700
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…umber and update the related documentation
kotsaloscv
approved these changes
Nov 29, 2021
|
Logfiles from GitLab pipeline #27115 (:white_check_mark:) have been uploaded here! Status and direct links: |
pramodk
approved these changes
Nov 29, 2021
olupton
pushed a commit
that referenced
this pull request
Nov 30, 2021
…umber and update the related documentation (#700)
olupton
added a commit
that referenced
this pull request
Dec 23, 2021
Summary of changes: - Support OpenMP target offload when NMODL and GPU support are enabled. (#693, #704, #705, #707, #708, #716, #719) - Use sensible defaults for the --nwarp parameter, improving the performance of the Hines solver with --cell-permute=2 on GPU. (#700, #710, #718) - Use a Boost memory pool, if Boost is available, to reduce the number of independent CUDA unified memory allocations used for Random123 stream objects. This speeds up initialisation of models using Random123, and also makes it feasible to use NSight Compute on models using Random123 and for NSight Systems to profile initialisation. (#702, #703) - Use -cuda when compiling with NVHPC and OpenACC or OpenMP, as recommended on the NVIDIA forums. (#721) - Do not compile for compute capability 6.0 by default, as this is not supported by NVHPC with OpenMP target offload. - Add new GitLab CI tests so we test CoreNEURON + NMODL with both OpenACC and OpenMP. (#698, #717) - Add CUDA runtime header search path explicitly, so we don't rely on it being implicit in our NVHPC localrc. - Cleanup unused code. (#711) Co-authored-by: Pramod Kumbhar <[email protected]> Co-authored-by: Ioannis Magkanaris <[email protected]> Co-authored-by: Christos Kotsalos <[email protected]> Co-authored-by: Nicolas Cornu <[email protected]>
pramodk
pushed a commit
to neuronsimulator/nrn
that referenced
this pull request
Nov 2, 2022
Summary of changes: - Support OpenMP target offload when NMODL and GPU support are enabled. (BlueBrain/CoreNeuron#693, BlueBrain/CoreNeuron#704, BlueBrain/CoreNeuron#705, BlueBrain/CoreNeuron#707, BlueBrain/CoreNeuron#708, BlueBrain/CoreNeuron#716, BlueBrain/CoreNeuron#719) - Use sensible defaults for the --nwarp parameter, improving the performance of the Hines solver with --cell-permute=2 on GPU. (BlueBrain/CoreNeuron#700, BlueBrain/CoreNeuron#710, BlueBrain/CoreNeuron#718) - Use a Boost memory pool, if Boost is available, to reduce the number of independent CUDA unified memory allocations used for Random123 stream objects. This speeds up initialisation of models using Random123, and also makes it feasible to use NSight Compute on models using Random123 and for NSight Systems to profile initialisation. (BlueBrain/CoreNeuron#702, BlueBrain/CoreNeuron#703) - Use -cuda when compiling with NVHPC and OpenACC or OpenMP, as recommended on the NVIDIA forums. (BlueBrain/CoreNeuron#721) - Do not compile for compute capability 6.0 by default, as this is not supported by NVHPC with OpenMP target offload. - Add new GitLab CI tests so we test CoreNEURON + NMODL with both OpenACC and OpenMP. (BlueBrain/CoreNeuron#698, BlueBrain/CoreNeuron#717) - Add CUDA runtime header search path explicitly, so we don't rely on it being implicit in our NVHPC localrc. - Cleanup unused code. (BlueBrain/CoreNeuron#711) Co-authored-by: Pramod Kumbhar <[email protected]> Co-authored-by: Ioannis Magkanaris <[email protected]> Co-authored-by: Christos Kotsalos <[email protected]> Co-authored-by: Nicolas Cornu <[email protected]> CoreNEURON Repo SHA: BlueBrain/CoreNeuron@423ae6c
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Set by default the number of warps to execute in a large reasonable number and update the related documentation
Description
Cell permute 2 algorithm is distributing the cells in groups based on the
nwarpoption based to the CLI. Since it used to be 0 as default all the cells were executed using one warp on GPU with the--cell-permute 2option and there was no interleaving of the cells in the way they are sorted in memory.This generated suboptimal performance.
A quick solution provided is setting the
nwarpto a reasonable large number that might not generate the best load balancing but should be enough to hide large latencies introduced by memory accesses.TODO: