Skip to content
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

Speedup bincount #10308

Merged
merged 6 commits into from
Aug 2, 2023
Merged

Speedup bincount #10308

merged 6 commits into from
Aug 2, 2023

Conversation

marigoold
Copy link
Contributor

before:

  OP         Args                Lib   KT(GPU)    BW(GPU)                                          KT(1 CPU)   ET(1 CPU)   KT(32 CPU)   ET(32 CPU)
 ──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  bincount   ones(4096).int()    OF    4814.5     cast: , reduce_min: , reduce_max: , bincount:    685.9       771.7       692.0        772.1
  bincount   ones(65536).int()   OF    545205.9   cast: , reduce_min: , reduce_max: , bincount:    1458.3      1547.9      1508.3       1596.0
  bincount   ones(4096).int()    PT    183.2      -                                                18.3        21.0        21.8         24.9
  bincount   ones(65536).int()   PT    310.7      -                                                158.6       161.2       190.0        192.8

after:

  OP         Args                Lib   KT(GPU)   BW(GPU)                                          KT(1 CPU)   ET(1 CPU)   KT(32 CPU)   ET(32 CPU)
 ─────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
  bincount   ones(4096).int()    OF    360.2     cast: , reduce_min: , reduce_max: , bincount:    617.1       688.0       624.6        694.3
  bincount   ones(65536).int()   OF    623.3     cast: , reduce_min: , reduce_max: , bincount:    1279.0      1345.1      1554.4       1633.6
  bincount   ones(4096).int()    PT    188.3     -                                                16.8        19.4        17.6         20.2
  bincount   ones(65536).int()   PT    317.5     -                                                158.9       161.4       188.1        191.3

@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10308/

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.8ms (= 4375.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.0ms (= 5696.5ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.30 (= 57.0ms / 43.8ms)

OneFlow resnet50 time: 25.8ms (= 2583.9ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.5ms (= 3747.6ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.45 (= 37.5ms / 25.8ms)

OneFlow resnet50 time: 18.9ms (= 3772.1ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 34.7ms (= 6948.8ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.84 (= 34.7ms / 18.9ms)

OneFlow resnet50 time: 18.8ms (= 3766.8ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 32.3ms (= 6452.4ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.71 (= 32.3ms / 18.8ms)

OneFlow resnet50 time: 17.4ms (= 3489.3ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 28.8ms (= 5767.1ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.65 (= 28.8ms / 17.4ms)

OneFlow swin dataloader time: 0.203s (= 40.556s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.590s / 200, num_workers=1)
Relative speed: 0.631 (= 0.128s / 0.203s)

OneFlow swin dataloader time: 0.054s (= 10.728s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.650s / 200, num_workers=4)
Relative speed: 0.620 (= 0.033s / 0.054s)

OneFlow swin dataloader time: 0.030s (= 6.042s / 200, num_workers=8)
PyTorch swin dataloader time: 0.017s (= 3.366s / 200, num_workers=8)
Relative speed: 0.557 (= 0.017s / 0.030s)

❌ OneFlow resnet50 time: 47.3ms (= 4732.9ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 64.3ms (= 6425.5ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.36 (= 64.3ms / 47.3ms)

OneFlow resnet50 time: 31.7ms (= 3172.8ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 44.0ms (= 4402.1ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.39 (= 44.0ms / 31.7ms)

OneFlow resnet50 time: 24.3ms (= 4850.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 40.8ms (= 8163.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.68 (= 40.8ms / 24.3ms)

OneFlow resnet50 time: 22.7ms (= 4530.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 37.3ms (= 7462.1ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.65 (= 37.3ms / 22.7ms)

OneFlow resnet50 time: 22.1ms (= 4411.1ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 35.3ms (= 7057.1ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.60 (= 35.3ms / 22.1ms)

@marigoold marigoold added the op label Jul 28, 2023
@github-actions
Copy link
Contributor

CI failed when running job: Build llvm15. PR label automerge has been removed

BBuf
BBuf approved these changes Jul 28, 2023
View reviewed changes
Copy link
Contributor

@BBuf BBuf left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

lgtm, thank you!

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

Speed stats:

@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10308/

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.5ms (= 4350.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 61.6ms (= 6159.9ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.42 (= 61.6ms / 43.5ms)

OneFlow resnet50 time: 26.5ms (= 2646.2ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.9ms (= 3786.6ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.43 (= 37.9ms / 26.5ms)

OneFlow resnet50 time: 20.0ms (= 4007.5ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 34.2ms (= 6839.0ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.71 (= 34.2ms / 20.0ms)

OneFlow resnet50 time: 19.7ms (= 3948.1ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 31.0ms (= 6203.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.57 (= 31.0ms / 19.7ms)

OneFlow resnet50 time: 17.8ms (= 3566.2ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 32.4ms (= 6472.2ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.81 (= 32.4ms / 17.8ms)

OneFlow swin dataloader time: 0.201s (= 40.254s / 200, num_workers=1)
PyTorch swin dataloader time: 0.130s (= 25.909s / 200, num_workers=1)
Relative speed: 0.644 (= 0.130s / 0.201s)

OneFlow swin dataloader time: 0.054s (= 10.754s / 200, num_workers=4)
PyTorch swin dataloader time: 0.032s (= 6.459s / 200, num_workers=4)
Relative speed: 0.601 (= 0.032s / 0.054s)

OneFlow swin dataloader time: 0.030s (= 6.055s / 200, num_workers=8)
PyTorch swin dataloader time: 0.016s (= 3.290s / 200, num_workers=8)
Relative speed: 0.543 (= 0.016s / 0.030s)

❌ OneFlow resnet50 time: 47.7ms (= 4770.5ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 64.2ms (= 6421.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.35 (= 64.2ms / 47.7ms)

OneFlow resnet50 time: 30.7ms (= 3066.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 44.3ms (= 4433.7ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.45 (= 44.3ms / 30.7ms)

OneFlow resnet50 time: 24.4ms (= 4875.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 41.0ms (= 8199.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.68 (= 41.0ms / 24.4ms)

OneFlow resnet50 time: 22.3ms (= 4467.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 37.6ms (= 7525.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.68 (= 37.6ms / 22.3ms)

OneFlow resnet50 time: 21.4ms (= 4281.6ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 34.2ms (= 6831.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.60 (= 34.2ms / 21.4ms)

@github-actions
Copy link
Contributor

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

8 similar comments
@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

github-actions bot commented Aug 2, 2023

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

@github-actions
Copy link
Contributor

github-actions bot commented Aug 2, 2023

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

@github-actions
Copy link
Contributor

github-actions bot commented Aug 2, 2023

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10308/

@github-actions
Copy link
Contributor

github-actions bot commented Aug 2, 2023

Speed stats: 
GPU Name: NVIDIA GeForce RTX 3080 Ti 

❌ OneFlow resnet50 time: 43.6ms (= 4364.7ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 57.0ms (= 5699.8ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.31 (= 57.0ms / 43.6ms)

OneFlow resnet50 time: 26.2ms (= 2620.6ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 37.9ms (= 3788.0ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.45 (= 37.9ms / 26.2ms)

OneFlow resnet50 time: 20.8ms (= 4158.2ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 35.6ms (= 7127.1ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.71 (= 35.6ms / 20.8ms)

OneFlow resnet50 time: 18.7ms (= 3749.5ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 32.3ms (= 6458.7ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.72 (= 32.3ms / 18.7ms)

OneFlow resnet50 time: 17.9ms (= 3585.6ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 31.7ms (= 6332.9ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.77 (= 31.7ms / 17.9ms)

OneFlow swin dataloader time: 0.201s (= 40.253s / 200, num_workers=1)
PyTorch swin dataloader time: 0.128s (= 25.664s / 200, num_workers=1)
Relative speed: 0.638 (= 0.128s / 0.201s)

OneFlow swin dataloader time: 0.055s (= 10.931s / 200, num_workers=4)
PyTorch swin dataloader time: 0.033s (= 6.595s / 200, num_workers=4)
Relative speed: 0.603 (= 0.033s / 0.055s)

OneFlow swin dataloader time: 0.030s (= 5.971s / 200, num_workers=8)
PyTorch swin dataloader time: 0.016s (= 3.274s / 200, num_workers=8)
Relative speed: 0.548 (= 0.016s / 0.030s)

❌ OneFlow resnet50 time: 47.7ms (= 4766.0ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 65.3ms (= 6532.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.37 (= 65.3ms / 47.7ms)

OneFlow resnet50 time: 30.8ms (= 3083.3ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 44.7ms (= 4465.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.45 (= 44.7ms / 30.8ms)

OneFlow resnet50 time: 23.9ms (= 4779.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 41.5ms (= 8303.8ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.74 (= 41.5ms / 23.9ms)

OneFlow resnet50 time: 22.5ms (= 4509.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 37.8ms (= 7567.5ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.68 (= 37.8ms / 22.5ms)

OneFlow resnet50 time: 21.3ms (= 4268.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 34.8ms (= 6950.9ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.63 (= 34.8ms / 21.3ms)

@marigoold marigoold merged commit b664acc into master Aug 2, 2023
@marigoold marigoold deleted the speedup_bincount branch August 2, 2023 06:29
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@BBuf BBuf BBuf approved these changes

@daquexian daquexian Awaiting requested review from daquexian

@liujuncheng liujuncheng Awaiting requested review from liujuncheng liujuncheng is a code owner

Assignees
No one assigned
Labels
api enhancement op
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
2 participants
@marigoold @BBuf