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

Add onednn softmax backward #8075

Merged
merged 22 commits into from
Jun 1, 2022
Merged

Add onednn softmax backward #8075

merged 22 commits into from
Jun 1, 2022

Conversation

luqiang-guo
Copy link
Contributor

No description provided.

daquexian
daquexian previously approved these changes Apr 26, 2022
View reviewed changes
@daquexian daquexian dismissed their stale review May 27, 2022 01:35

need change based on master

@github-actions
Copy link
Contributor

Speed stats:

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 130.1ms (= 13010.9ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 147.4ms (= 14737.6ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.13 (= 147.4ms / 130.1ms)

OneFlow resnet50 time: 77.4ms (= 7739.7ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 84.7ms (= 8466.5ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.09 (= 84.7ms / 77.4ms)

OneFlow resnet50 time: 52.1ms (= 10429.1ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 61.5ms (= 12297.4ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.18 (= 61.5ms / 52.1ms)

OneFlow resnet50 time: 42.6ms (= 8515.7ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 42.3ms (= 8460.4ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 0.99 (= 42.3ms / 42.6ms)

OneFlow resnet50 time: 38.1ms (= 7612.3ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 36.8ms (= 7352.8ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 0.97 (= 36.8ms / 38.1ms)

OneFlow swin dataloader time: 0.260s (= 52.083s / 200, num_workers=1)
PyTorch swin dataloader time: 0.160s (= 31.946s / 200, num_workers=1)
Relative speed: 0.613 (= 0.160s / 0.260s)

OneFlow swin dataloader time: 0.064s (= 12.896s / 200, num_workers=4)
PyTorch swin dataloader time: 0.040s (= 8.075s / 200, num_workers=4)
Relative speed: 0.626 (= 0.040s / 0.064s)

OneFlow swin dataloader time: 0.037s (= 7.485s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.478s / 200, num_workers=8)
Relative speed: 0.598 (= 0.022s / 0.037s)

❌ OneFlow resnet50 time: 146.6ms (= 14656.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 172.8ms (= 17284.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.18 (= 172.8ms / 146.6ms)

OneFlow resnet50 time: 96.1ms (= 9614.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 112.0ms (= 11198.2ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.16 (= 112.0ms / 96.1ms)

OneFlow resnet50 time: 72.0ms (= 14406.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 88.6ms (= 17713.0ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.23 (= 88.6ms / 72.0ms)

OneFlow resnet50 time: 58.1ms (= 11612.5ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 75.7ms (= 15137.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.30 (= 75.7ms / 58.1ms)

OneFlow resnet50 time: 54.8ms (= 10956.4ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 74.2ms (= 14848.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.36 (= 74.2ms / 54.8ms)

@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8075/

@luqiang-guo luqiang-guo requested review from oneflow-ci-bot and removed request for oneflow-ci-bot May 31, 2022 12:21
@github-actions
Copy link
Contributor

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8075/

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 130.1ms (= 13013.8ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 142.9ms (= 14295.0ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.10 (= 142.9ms / 130.1ms)

OneFlow resnet50 time: 77.0ms (= 7700.9ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 85.0ms (= 8495.5ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.10 (= 85.0ms / 77.0ms)

OneFlow resnet50 time: 52.8ms (= 10565.7ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 62.8ms (= 12562.3ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.19 (= 62.8ms / 52.8ms)

OneFlow resnet50 time: 43.1ms (= 8627.0ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 41.5ms (= 8309.0ms / 200, input_shape=[2, 3, 224, 224])
❌ Relative speed: 0.96 (= 41.5ms / 43.1ms)

OneFlow resnet50 time: 36.4ms (= 7273.3ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 35.9ms (= 7175.0ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 0.99 (= 35.9ms / 36.4ms)

OneFlow swin dataloader time: 0.384s (= 76.864s / 200, num_workers=1)
PyTorch swin dataloader time: 0.152s (= 30.378s / 200, num_workers=1)
Relative speed: 0.395 (= 0.152s / 0.384s)

OneFlow swin dataloader time: 0.101s (= 20.279s / 200, num_workers=4)
PyTorch swin dataloader time: 0.044s (= 8.736s / 200, num_workers=4)
Relative speed: 0.431 (= 0.044s / 0.101s)

OneFlow swin dataloader time: 0.056s (= 11.116s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.461s / 200, num_workers=8)
Relative speed: 0.401 (= 0.022s / 0.056s)

❌ OneFlow resnet50 time: 146.0ms (= 14597.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 172.1ms (= 17213.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.18 (= 172.1ms / 146.0ms)

OneFlow resnet50 time: 95.9ms (= 9586.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 113.4ms (= 11342.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.18 (= 113.4ms / 95.9ms)

OneFlow resnet50 time: 71.2ms (= 14243.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 87.9ms (= 17584.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.23 (= 87.9ms / 71.2ms)

OneFlow resnet50 time: 61.7ms (= 12342.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 73.8ms (= 14760.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.20 (= 73.8ms / 61.7ms)

OneFlow resnet50 time: 54.2ms (= 10846.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 76.3ms (= 15254.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.41 (= 76.3ms / 54.2ms)

@github-actions
Copy link
Contributor

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

@github-actions
Copy link
Contributor

github-actions bot commented Jun 1, 2022

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8075/

@github-actions
Copy link
Contributor

github-actions bot commented Jun 1, 2022

Speed stats: 
GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 130.2ms (= 13016.6ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 142.2ms (= 14216.8ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.09 (= 142.2ms / 130.2ms)

OneFlow resnet50 time: 76.7ms (= 7674.6ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 87.2ms (= 8722.1ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.14 (= 87.2ms / 76.7ms)

OneFlow resnet50 time: 52.3ms (= 10466.2ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 56.5ms (= 11304.4ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.08 (= 56.5ms / 52.3ms)

OneFlow resnet50 time: 42.0ms (= 8390.8ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 46.0ms (= 9199.3ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.10 (= 46.0ms / 42.0ms)

OneFlow resnet50 time: 36.5ms (= 7295.3ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 37.3ms (= 7454.9ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.02 (= 37.3ms / 36.5ms)

OneFlow swin dataloader time: 0.245s (= 48.986s / 200, num_workers=1)
PyTorch swin dataloader time: 0.150s (= 29.958s / 200, num_workers=1)
Relative speed: 0.612 (= 0.150s / 0.245s)

OneFlow swin dataloader time: 0.065s (= 12.989s / 200, num_workers=4)
PyTorch swin dataloader time: 0.042s (= 8.362s / 200, num_workers=4)
Relative speed: 0.644 (= 0.042s / 0.065s)

OneFlow swin dataloader time: 0.036s (= 7.224s / 200, num_workers=8)
PyTorch swin dataloader time: 0.021s (= 4.252s / 200, num_workers=8)
Relative speed: 0.589 (= 0.021s / 0.036s)

❌ OneFlow resnet50 time: 146.2ms (= 14619.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 173.8ms (= 17376.6ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.19 (= 173.8ms / 146.2ms)

OneFlow resnet50 time: 96.1ms (= 9607.7ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 111.5ms (= 11148.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.16 (= 111.5ms / 96.1ms)

OneFlow resnet50 time: 71.4ms (= 14280.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 88.3ms (= 17652.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.24 (= 88.3ms / 71.4ms)

OneFlow resnet50 time: 59.2ms (= 11843.9ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 76.9ms (= 15384.3ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.30 (= 76.9ms / 59.2ms)

OneFlow resnet50 time: 53.6ms (= 10717.8ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 71.4ms (= 14272.7ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.33 (= 71.4ms / 53.6ms)

@luqiang-guo luqiang-guo removed the request for review from oneflow-ci-bot June 1, 2022 03:13
@github-actions
Copy link
Contributor

github-actions bot commented Jun 1, 2022

Speed stats:

@github-actions
Copy link
Contributor

github-actions bot commented Jun 1, 2022

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/8075/

@github-actions
Copy link
Contributor

github-actions bot commented Jun 1, 2022

Speed stats: 
GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 130.3ms (= 13034.5ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 145.0ms (= 14505.0ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.11 (= 145.0ms / 130.3ms)

OneFlow resnet50 time: 81.0ms (= 8098.7ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 87.4ms (= 8743.5ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.08 (= 87.4ms / 81.0ms)

OneFlow resnet50 time: 54.1ms (= 10821.1ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 60.4ms (= 12088.9ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.12 (= 60.4ms / 54.1ms)

OneFlow resnet50 time: 42.9ms (= 8577.6ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 42.3ms (= 8457.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 0.99 (= 42.3ms / 42.9ms)

OneFlow resnet50 time: 35.0ms (= 7000.7ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 38.4ms (= 7687.7ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.10 (= 38.4ms / 35.0ms)

OneFlow swin dataloader time: 0.244s (= 48.868s / 200, num_workers=1)
PyTorch swin dataloader time: 0.150s (= 29.925s / 200, num_workers=1)
Relative speed: 0.612 (= 0.150s / 0.244s)

OneFlow swin dataloader time: 0.069s (= 13.884s / 200, num_workers=4)
PyTorch swin dataloader time: 0.041s (= 8.241s / 200, num_workers=4)
Relative speed: 0.594 (= 0.041s / 0.069s)

OneFlow swin dataloader time: 0.036s (= 7.221s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.414s / 200, num_workers=8)
Relative speed: 0.611 (= 0.022s / 0.036s)

❌ OneFlow resnet50 time: 146.6ms (= 14662.4ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 172.3ms (= 17229.0ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.18 (= 172.3ms / 146.6ms)

OneFlow resnet50 time: 96.9ms (= 9687.9ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 114.5ms (= 11449.0ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.18 (= 114.5ms / 96.9ms)

OneFlow resnet50 time: 72.9ms (= 14582.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 97.8ms (= 19552.9ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.34 (= 97.8ms / 72.9ms)

OneFlow resnet50 time: 60.0ms (= 11996.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 74.6ms (= 14922.2ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.24 (= 74.6ms / 60.0ms)

OneFlow resnet50 time: 55.6ms (= 11116.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 72.3ms (= 14461.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.30 (= 72.3ms / 55.6ms)

@mergify mergify bot merged commit 7cc955e into master Jun 1, 2022
@mergify mergify bot deleted the dev_add_onednn_softmax_backward branch June 1, 2022 09:38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@daquexian daquexian daquexian approved these changes

@guo-ran guo-ran guo-ran approved these changes

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

@oneflow-ci-bot oneflow-ci-bot Awaiting requested review from oneflow-ci-bot

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

Successfully merging this pull request may close these issues.
3 participants
@luqiang-guo @daquexian @guo-ran