[CodeGen] Support transform.oneflow.apply_patterns Op in MLIR #10255

howin98 · 2023-05-11T07:43:16Z

iree中通过 transform.iree.apply_patterns 这个op灵活的将预定义的一些pattern在transform dialect的变换环节作用于指定的op。如，在transform.sequeue中做完tiling之后，做一些规范化的操作：

  transform.structured.tile_to_forall_op %parallel_linalg_ops num_threads [1, 4, 32]
    ( mapping = [#gpu.thread<z>, #gpu.thread<y>, #gpu.thread<x>] )

  // Canonicalizations.
  transform.iree.apply_patterns %variant_op
    { canonicalization, tiling_canonicalization, licm, cse } : (!pdl.operation) -> ()

该pr中在oneflow::transform_dialect命名空间引入类似的op，并完成一个canonicalization的支持。具体实例见：https://github.com/Oneflow-Inc/oneflow/pull/10255/files#diff-a36168c1a81d37cf7f56ff09c02af22e9e3bcc39902ccdb8c7f21805d309e72fR1

除此之外，该pr中对transform dialect interpret等从llvm主仓库迁移过来的设施进行剪裁和重构。

…-Inc/oneflow into support-transform-apply-patterns

github-actions · 2023-05-11T08:03:10Z

Code got formatted by CI. Please request CI again if you still want to have this PR merged. If the PR is from a forked repo, please download the patch files from the GitHub Actions web page and apply them locally.

github-actions · 2023-05-11T09:02:16Z

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10255/

github-actions · 2023-05-11T09:19:12Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3090 

❌ OneFlow resnet50 time: 43.0ms (= 4298.9ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.2ms (= 5716.4ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.33 (= 57.2ms / 43.0ms)

OneFlow resnet50 time: 26.2ms (= 2619.7ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.1ms (= 3709.7ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.42 (= 37.1ms / 26.2ms)

OneFlow resnet50 time: 19.7ms (= 3931.8ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.6ms (= 7128.1ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.81 (= 35.6ms / 19.7ms)

OneFlow resnet50 time: 18.1ms (= 3620.5ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 31.2ms (= 6241.2ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.72 (= 31.2ms / 18.1ms)

OneFlow resnet50 time: 17.9ms (= 3580.0ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.1ms (= 5828.9ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.63 (= 29.1ms / 17.9ms)

OneFlow swin dataloader time: 0.202s (= 40.411s / 200, num_workers=1)
PyTorch swin dataloader time: 0.129s (= 25.758s / 200, num_workers=1)
Relative speed: 0.637 (= 0.129s / 0.202s)

OneFlow swin dataloader time: 0.057s (= 11.331s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.590s / 200, num_workers=4)
Relative speed: 0.582 (= 0.033s / 0.057s)

OneFlow swin dataloader time: 0.033s (= 6.542s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.347s / 200, num_workers=8)
Relative speed: 0.512 (= 0.017s / 0.033s)

❌ OneFlow resnet50 time: 48.8ms (= 4879.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 64.3ms (= 6432.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.32 (= 64.3ms / 48.8ms)

OneFlow resnet50 time: 37.1ms (= 3711.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 45.3ms (= 4528.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.22 (= 45.3ms / 37.1ms)

OneFlow resnet50 time: 28.6ms (= 5726.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 39.4ms (= 7873.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.37 (= 39.4ms / 28.6ms)

OneFlow resnet50 time: 25.6ms (= 5120.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.7ms (= 7739.1ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.51 (= 38.7ms / 25.6ms)

OneFlow resnet50 time: 24.1ms (= 4822.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.4ms (= 7282.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.51 (= 36.4ms / 24.1ms)

github-actions · 2023-05-11T09:19:29Z

CI failed when running job: cuda-module. PR label automerge has been removed

github-actions · 2023-05-11T10:41:18Z

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10255/

github-actions · 2023-05-11T10:47:59Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3090 

❌ OneFlow resnet50 time: 43.0ms (= 4302.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.2ms (= 5723.6ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.33 (= 57.2ms / 43.0ms)

OneFlow resnet50 time: 26.2ms (= 2621.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.4ms (= 3739.3ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.43 (= 37.4ms / 26.2ms)

OneFlow resnet50 time: 18.7ms (= 3748.8ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.2ms (= 7040.9ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.88 (= 35.2ms / 18.7ms)

OneFlow resnet50 time: 17.7ms (= 3548.4ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 32.1ms (= 6425.3ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.81 (= 32.1ms / 17.7ms)

OneFlow resnet50 time: 16.7ms (= 3344.0ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 29.3ms (= 5863.3ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.75 (= 29.3ms / 16.7ms)

OneFlow swin dataloader time: 0.199s (= 39.724s / 200, num_workers=1)
PyTorch swin dataloader time: 0.129s (= 25.891s / 200, num_workers=1)
Relative speed: 0.652 (= 0.129s / 0.199s)

OneFlow swin dataloader time: 0.054s (= 10.806s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.576s / 200, num_workers=4)
Relative speed: 0.609 (= 0.033s / 0.054s)

OneFlow swin dataloader time: 0.031s (= 6.284s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.376s / 200, num_workers=8)
Relative speed: 0.537 (= 0.017s / 0.031s)

❌ OneFlow resnet50 time: 48.6ms (= 4859.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 66.2ms (= 6617.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.36 (= 66.2ms / 48.6ms)

OneFlow resnet50 time: 36.4ms (= 3644.8ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 45.5ms (= 4546.5ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.25 (= 45.5ms / 36.4ms)

OneFlow resnet50 time: 28.2ms (= 5644.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.8ms (= 7753.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.37 (= 38.8ms / 28.2ms)

OneFlow resnet50 time: 25.9ms (= 5182.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.6ms (= 7728.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.49 (= 38.6ms / 25.9ms)

OneFlow resnet50 time: 24.7ms (= 4930.1ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.0ms (= 7198.6ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.46 (= 36.0ms / 24.7ms)

github-actions · 2023-05-11T13:23:31Z

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10255/

github-actions · 2023-05-11T13:30:33Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3090 

❌ OneFlow resnet50 time: 43.2ms (= 4323.1ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 64.8ms (= 6483.9ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.50 (= 64.8ms / 43.2ms)

OneFlow resnet50 time: 26.3ms (= 2627.2ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 44.3ms (= 4431.9ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.69 (= 44.3ms / 26.3ms)

OneFlow resnet50 time: 18.4ms (= 3688.4ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 38.6ms (= 7720.5ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 2.09 (= 38.6ms / 18.4ms)

OneFlow resnet50 time: 16.9ms (= 3375.5ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 34.7ms (= 6932.6ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 2.05 (= 34.7ms / 16.9ms)

OneFlow resnet50 time: 15.6ms (= 3128.3ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 28.3ms (= 5657.0ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.81 (= 28.3ms / 15.6ms)

OneFlow swin dataloader time: 0.202s (= 40.344s / 200, num_workers=1)
PyTorch swin dataloader time: 0.131s (= 26.125s / 200, num_workers=1)
Relative speed: 0.648 (= 0.131s / 0.202s)

OneFlow swin dataloader time: 0.056s (= 11.199s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.557s / 200, num_workers=4)
Relative speed: 0.585 (= 0.033s / 0.056s)

OneFlow swin dataloader time: 0.032s (= 6.374s / 200, num_workers=8)
PyTorch swin dataloader time: 0.016s (= 3.294s / 200, num_workers=8)
Relative speed: 0.517 (= 0.016s / 0.032s)

❌ OneFlow resnet50 time: 48.6ms (= 4862.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 68.3ms (= 6828.9ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.40 (= 68.3ms / 48.6ms)

OneFlow resnet50 time: 37.5ms (= 3749.6ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 44.8ms (= 4476.2ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.19 (= 44.8ms / 37.5ms)

OneFlow resnet50 time: 28.9ms (= 5771.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.5ms (= 7699.6ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.33 (= 38.5ms / 28.9ms)

OneFlow resnet50 time: 25.5ms (= 5090.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 38.6ms (= 7715.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.52 (= 38.6ms / 25.5ms)

OneFlow resnet50 time: 23.7ms (= 4742.6ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 35.9ms (= 7177.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.51 (= 35.9ms / 23.7ms)

github-actions · 2023-05-12T02:37:55Z

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10255/

github-actions · 2023-05-12T02:42:54Z

Speed stats:

GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.7ms (= 4370.4ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 61.6ms (= 6160.3ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.41 (= 61.6ms / 43.7ms)

OneFlow resnet50 time: 26.2ms (= 2622.6ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.8ms (= 3775.9ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.44 (= 37.8ms / 26.2ms)

OneFlow resnet50 time: 18.9ms (= 3786.9ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.0ms (= 7009.8ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.85 (= 35.0ms / 18.9ms)

OneFlow resnet50 time: 18.9ms (= 3773.7ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 33.9ms (= 6777.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.80 (= 33.9ms / 18.9ms)

OneFlow resnet50 time: 17.4ms (= 3484.1ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 28.2ms (= 5645.8ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.62 (= 28.2ms / 17.4ms)

OneFlow swin dataloader time: 0.203s (= 40.578s / 200, num_workers=1)
PyTorch swin dataloader time: 0.129s (= 25.798s / 200, num_workers=1)
Relative speed: 0.636 (= 0.129s / 0.203s)

OneFlow swin dataloader time: 0.054s (= 10.770s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.572s / 200, num_workers=4)
Relative speed: 0.610 (= 0.033s / 0.054s)

OneFlow swin dataloader time: 0.030s (= 6.067s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.341s / 200, num_workers=8)
Relative speed: 0.551 (= 0.017s / 0.030s)

❌ OneFlow resnet50 time: 47.6ms (= 4762.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 65.2ms (= 6518.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.37 (= 65.2ms / 47.6ms)

OneFlow resnet50 time: 30.9ms (= 3087.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 43.8ms (= 4383.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.42 (= 43.8ms / 30.9ms)

OneFlow resnet50 time: 24.2ms (= 4830.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 40.5ms (= 8093.0ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.68 (= 40.5ms / 24.2ms)

OneFlow resnet50 time: 22.2ms (= 4437.5ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 36.8ms (= 7355.6ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.66 (= 36.8ms / 22.2ms)

OneFlow resnet50 time: 21.1ms (= 4228.6ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 34.1ms (= 6825.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.61 (= 34.1ms / 21.1ms)

init

c69d3fa

howin98 added feature automerge ir labels May 11, 2023

howin98 added 3 commits May 11, 2023 15:47

fix

6d1f0ef

fix

56b698a

fix

4911ab0

howin98 marked this pull request as ready for review May 11, 2023 07:58

howin98 requested review from daquexian, jackalcooper, hjchen2 and BBuf as code owners May 11, 2023 07:58

Merge branch 'master' into support-transform-apply-patterns

f1d0006

howin98 requested a review from oneflow-ci-bot May 11, 2023 07:58

fix

787401f

jackalcooper approved these changes May 11, 2023

View reviewed changes

howin98 and others added 3 commits May 11, 2023 16:00

Merge branch 'support-transform-apply-patterns' of github.com:Oneflow…

4b847eb

…-Inc/oneflow into support-transform-apply-patterns

fix

4fe35fd

auto format by CI

5a1b9ac

github-actions bot removed the automerge label May 11, 2023

Merge branch 'master' into support-transform-apply-patterns

9018718

Merge branch 'master' into support-transform-apply-patterns

3f020cb

howin98 added the automerge label May 11, 2023

howin98 enabled auto-merge (squash) May 11, 2023 12:20

mosout approved these changes May 12, 2023

View reviewed changes

Merge branch 'master' into support-transform-apply-patterns

82a5580

howin98 merged commit 1af9915 into master May 12, 2023

howin98 deleted the support-transform-apply-patterns branch May 12, 2023 03:08

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[CodeGen] Support transform.oneflow.apply_patterns Op in MLIR #10255

[CodeGen] Support transform.oneflow.apply_patterns Op in MLIR #10255

howin98 commented May 11, 2023 •

edited

Loading

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 12, 2023

github-actions bot commented May 12, 2023

[CodeGen] Support transform.oneflow.apply_patterns Op in MLIR #10255

[CodeGen] Support transform.oneflow.apply_patterns Op in MLIR #10255

Conversation

howin98 commented May 11, 2023 • edited Loading

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 11, 2023

github-actions bot commented May 12, 2023

github-actions bot commented May 12, 2023

howin98 commented May 11, 2023 •

edited

Loading