Skip to content

Speedup bernoulli_scalar_cuda_kernel with grid-stride loop #21300

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
syed-ahmed wants to merge 4 commits into from
Closed

Speedup bernoulli_scalar_cuda_kernel with grid-stride loop #21300

syed-ahmed wants to merge 4 commits into from

Conversation

@syed-ahmed
Copy link
Collaborator

@syed-ahmed syed-ahmed commented Jun 3, 2019 •
edited by ezyang
Loading

Stack from ghstack:

Resubmit of #20626

Effective Bandwidth Benchmark

Float Type

Before:

bernoulli, size, elements 65536 forward 5.810260772705078e-06 bandwidth (GB/s) 45.117424200902754
bernoulli, size, elements 131072 forward 5.700588226318359e-06 bandwidth (GB/s) 91.97085970522794
bernoulli, size, elements 262144 forward 7.650852203369141e-06 bandwidth (GB/s) 137.0534905298847
bernoulli, size, elements 524288 forward 1.1038780212402343e-05 bandwidth (GB/s) 189.98041084682507
bernoulli, size, elements 1048576 forward 1.817464828491211e-05 bandwidth (GB/s) 230.77772588765578
bernoulli, size, elements 2097152 forward 3.152847290039063e-05 bandwidth (GB/s) 266.06451972800966
bernoulli, size, elements 4194304 forward 5.8722496032714846e-05 bandwidth (GB/s) 285.7033868358262
bernoulli, size, elements 8388608 forward 0.0001120924949645996 bandwidth (GB/s) 299.3459286511284
bernoulli, size, elements 16777216 forward 0.0002196049690246582 bandwidth (GB/s) 305.58900510336235
bernoulli, size, elements 33554432 forward 0.0004137754440307617 bandwidth (GB/s) 324.3733525907877

After:

bernoulli, size, elements 65536 forward 5.7387351989746094e-06 bandwidth (GB/s) 45.679751881013715
bernoulli, size, elements 131072 forward 5.600452423095703e-06 bandwidth (GB/s) 93.61529397837378
bernoulli, size, elements 262144 forward 6.201267242431641e-06 bandwidth (GB/s) 169.09060019623223
bernoulli, size, elements 524288 forward 6.272792816162109e-06 bandwidth (GB/s) 334.3250863629039
bernoulli, size, elements 1048576 forward 8.275508880615235e-06 bandwidth (GB/s) 506.83336342310577
bernoulli, size, elements 2097152 forward 1.2857913970947266e-05 bandwidth (GB/s) 652.4081603714445
bernoulli, size, elements 4194304 forward 2.348184585571289e-05 bandwidth (GB/s) 714.4760298270282
bernoulli, size, elements 8388608 forward 4.356622695922851e-05 bandwidth (GB/s) 770.1936647257047
bernoulli, size, elements 16777216 forward 8.656024932861328e-05 bandwidth (GB/s) 775.2850126994326
bernoulli, size, elements 33554432 forward 0.0001675891876220703 bandwidth (GB/s) 800.8734328534002

Double Type

Before:

bernoulli, size, elements 65536 forward 5.733966827392578e-06 bandwidth (GB/s) 45.717739200665285
bernoulli, size, elements 131072 forward 6.6208839416503905e-06 bandwidth (GB/s) 79.18700956254952
bernoulli, size, elements 262144 forward 1.0859966278076171e-05 bandwidth (GB/s) 96.55425929975851
bernoulli, size, elements 524288 forward 1.7333030700683594e-05 bandwidth (GB/s) 120.99165092445668
bernoulli, size, elements 1048576 forward 3.1557083129882816e-05 bandwidth (GB/s) 132.91165038090057
bernoulli, size, elements 2097152 forward 5.902767181396485e-05 bandwidth (GB/s) 142.11314358523305
bernoulli, size, elements 4194304 forward 0.00011337995529174805 bandwidth (GB/s) 147.9733869785806
bernoulli, size, elements 8388608 forward 0.00022054195404052734 bandwidth (GB/s) 152.14534643070206
bernoulli, size, elements 16777216 forward 0.0004380941390991211 bandwidth (GB/s) 153.18366079491483
bernoulli, size, elements 33554432 forward 0.0008704972267150879 bandwidth (GB/s) 154.1851299245198

After:

bernoulli, size, elements 65536 forward 5.877017974853515e-06 bandwidth (GB/s) 44.60493418969575
bernoulli, size, elements 131072 forward 5.819797515869141e-06 bandwidth (GB/s) 90.08698302138468
bernoulli, size, elements 262144 forward 6.091594696044922e-06 bandwidth (GB/s) 172.1348928025049
bernoulli, size, elements 524288 forward 8.232593536376953e-06 bandwidth (GB/s) 254.73770698546193
bernoulli, size, elements 1048576 forward 1.3000965118408203e-05 bandwidth (GB/s) 322.6148183461581
bernoulli, size, elements 2097152 forward 2.2871494293212892e-05 bandwidth (GB/s) 366.7713133413114
bernoulli, size, elements 4194304 forward 4.316329956054687e-05 bandwidth (GB/s) 388.69169342501107
bernoulli, size, elements 8388608 forward 8.46099853515625e-05 bandwidth (GB/s) 396.5776835981966
bernoulli, size, elements 16777216 forward 0.00016601085662841796 bandwidth (GB/s) 404.2438269577137
bernoulli, size, elements 33554432 forward 0.00031869888305664063 bandwidth (GB/s) 421.14276244936264

Differential Revision: D15632935

@pytorchbot pytorchbot added module: cuda Related to torch.cuda, and CUDA support in general module: operators labels Jun 3, 2019
This was referenced Jun 3, 2019
Closed
Closed
Closed
Closed
syed-ahmed added 3 commits June 3, 2019 12:03
@syed-ahmed
Speedup bernoulli_scalar_cuda_kernel with grid-stride loop
5b9ff4f
@syed-ahmed
Update on "Speedup bernoulli_scalar_cuda_kernel with grid-stride loop"
cdb036b 
Speedup bernoulli_scalar_cuda_kernel with grid-stride loop

gh-metadata: pytorch pytorch 21300 gh/syed-ahmed/12/head
@syed-ahmed
Update on "Speedup bernoulli_scalar_cuda_kernel with grid-stride loop"
d12f716 
Speedup bernoulli_scalar_cuda_kernel with grid-stride loop

gh-metadata: pytorch pytorch 21300 gh/syed-ahmed/12/head
@syed-ahmed syed-ahmed requested a review from ezyang June 4, 2019 05:53
@syed-ahmed syed-ahmed mentioned this pull request Jun 4, 2019
Closed
@syed-ahmed
Update on "Speedup bernoulli_scalar_cuda_kernel with grid-stride loop"
fe3f3fc 
Speedup bernoulli_scalar_cuda_kernel with grid-stride loop

gh-metadata: pytorch pytorch 21300 gh/syed-ahmed/12/head
@zou3519 zou3519 deleted the gh/syed-ahmed/12/head branch June 5, 2019 02:16
@facebook-github-bot
Copy link
Contributor

facebook-github-bot commented Jun 5, 2019

@ezyang merged this pull request in eadac84.

zdevito pushed a commit to zdevito/ATen that referenced this pull request Jun 5, 2019
@syed-ahmed @facebook-github-bot
Speedup bernoulli_scalar_cuda_kernel with grid-stride loop (#21300)
a1d7d07 
Summary:
Pull Request resolved: pytorch/pytorch#21300
ghimport-source-id: c314c28cb693b554d6f24de235c11ba24ed6bf61

Reviewed By: jerryzh168

Differential Revision: D15632935

Pulled By: ezyang

fbshipit-source-id: 9bb24f17d78151bf50942905c967bdcfe1ff00cb
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@ezyang ezyang ezyang approved these changes

Assignees

No one assigned

Labels

Merged module: cuda Related to torch.cuda, and CUDA support in general open source

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

6 participants

@syed-ahmed @facebook-github-bot @ezyang @pytorchbot @mruberry