release np array in main thread to eliminate the time-consuming gil request #10050

daquexian · 2023-03-28T05:42:49Z

No description provided.

…equest Signed-off-by: daquexian <daquexian566@gmail.com>

Signed-off-by: daquexian <daquexian566@gmail.com>

github-actions · 2023-03-28T13:10:43Z

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

github-actions · 2023-03-28T15:13:37Z

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

github-actions · 2023-03-28T16:30:53Z

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

github-actions · 2023-03-28T23:40:02Z

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

Signed-off-by: daquexian <daquexian566@gmail.com>

github-actions · 2023-03-29T02:19:18Z

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 · 2023-03-29T06:34:15Z

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

Signed-off-by: daquexian <daquexian566@gmail.com>

This reverts commit 4d23a13.

daquexian · 2023-03-30T15:11:28Z

CI 问题似乎是 numpy 的 bug 引起的，在本地使用和 CI 同样的 1.17 版本可以复现，使用 1.21 就无法复现，现在试一下升级 CI 中的 numpy 版本

Signed-off-by: daquexian <daquexian566@gmail.com>

This reverts commit 137ab19.

github-actions · 2023-03-30T16:28:58Z

Speed stats:

GPU Name: GeForce GTX 1080 

❌ OneFlow resnet50 time: 141.1ms (= 14111.8ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 146.3ms (= 14629.3ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 1.04 (= 146.3ms / 141.1ms)

OneFlow resnet50 time: 82.3ms (= 8226.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 91.2ms (= 9123.9ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.11 (= 91.2ms / 82.3ms)

OneFlow resnet50 time: 50.9ms (= 10171.6ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 73.7ms (= 14734.9ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.45 (= 73.7ms / 50.9ms)

OneFlow resnet50 time: 33.7ms (= 6745.7ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 65.8ms (= 13152.8ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.95 (= 65.8ms / 33.7ms)

OneFlow resnet50 time: 25.9ms (= 5177.2ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 64.3ms (= 12850.1ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 2.48 (= 64.3ms / 25.9ms)

OneFlow swin dataloader time: 0.238s (= 47.685s / 200, num_workers=1)
PyTorch swin dataloader time: 0.149s (= 29.802s / 200, num_workers=1)
Relative speed: 0.625 (= 0.149s / 0.238s)

OneFlow swin dataloader time: 0.067s (= 13.478s / 200, num_workers=4)
PyTorch swin dataloader time: 0.041s (= 8.296s / 200, num_workers=4)
Relative speed: 0.616 (= 0.041s / 0.067s)

OneFlow swin dataloader time: 0.041s (= 8.272s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.389s / 200, num_workers=8)
Relative speed: 0.531 (= 0.022s / 0.041s)

❌ OneFlow resnet50 time: 153.1ms (= 15311.7ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 165.3ms (= 16528.9ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 1.08 (= 165.3ms / 153.1ms)

OneFlow resnet50 time: 93.2ms (= 9318.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 103.0ms (= 10303.5ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.11 (= 103.0ms / 93.2ms)

OneFlow resnet50 time: 60.6ms (= 12123.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 78.7ms (= 15748.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.30 (= 78.7ms / 60.6ms)

OneFlow resnet50 time: 42.9ms (= 8587.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 80.3ms (= 16061.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.87 (= 80.3ms / 42.9ms)

OneFlow resnet50 time: 35.3ms (= 7060.2ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 67.0ms (= 13401.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.90 (= 67.0ms / 35.3ms)

github-actions · 2023-03-30T16:32:16Z

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10050/

Signed-off-by: daquexian <daquexian566@gmail.com>

github-actions · 2023-04-01T00:30:44Z

Speed stats:

GPU Name: GeForce GTX 1080 

❌ OneFlow resnet50 time: 141.1ms (= 14114.5ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 145.9ms (= 14589.4ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 1.03 (= 145.9ms / 141.1ms)

OneFlow resnet50 time: 81.3ms (= 8128.9ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 93.0ms (= 9299.2ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.14 (= 93.0ms / 81.3ms)

OneFlow resnet50 time: 50.9ms (= 10177.8ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 73.0ms (= 14596.0ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.43 (= 73.0ms / 50.9ms)

OneFlow resnet50 time: 33.4ms (= 6678.7ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 64.8ms (= 12958.5ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.94 (= 64.8ms / 33.4ms)

OneFlow resnet50 time: 26.0ms (= 5209.0ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 64.3ms (= 12869.8ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 2.47 (= 64.3ms / 26.0ms)

OneFlow swin dataloader time: 0.240s (= 48.024s / 200, num_workers=1)
PyTorch swin dataloader time: 0.150s (= 30.056s / 200, num_workers=1)
Relative speed: 0.626 (= 0.150s / 0.240s)

OneFlow swin dataloader time: 0.068s (= 13.671s / 200, num_workers=4)
PyTorch swin dataloader time: 0.041s (= 8.247s / 200, num_workers=4)
Relative speed: 0.603 (= 0.041s / 0.068s)

OneFlow swin dataloader time: 0.043s (= 8.581s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.374s / 200, num_workers=8)
Relative speed: 0.510 (= 0.022s / 0.043s)

❌ OneFlow resnet50 time: 152.5ms (= 15248.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 163.3ms (= 16331.6ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 1.07 (= 163.3ms / 152.5ms)

OneFlow resnet50 time: 92.6ms (= 9261.8ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 103.1ms (= 10307.4ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.11 (= 103.1ms / 92.6ms)

OneFlow resnet50 time: 61.3ms (= 12251.3ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 79.2ms (= 15846.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.29 (= 79.2ms / 61.3ms)

OneFlow resnet50 time: 42.4ms (= 8470.4ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 77.5ms (= 15493.1ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.83 (= 77.5ms / 42.4ms)

OneFlow resnet50 time: 36.4ms (= 7280.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 67.5ms (= 13498.0ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.85 (= 67.5ms / 36.4ms)

github-actions · 2023-04-01T00:32:45Z

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10050/

github-actions · 2023-04-01T00:38:43Z

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

github-actions · 2023-04-01T02:09:07Z

Speed stats:

GPU Name: GeForce GTX 1080 

❌ OneFlow resnet50 time: 141.1ms (= 14105.1ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 146.4ms (= 14641.7ms / 100, input_shape=[16, 3, 224, 224])
❌ Relative speed: 1.04 (= 146.4ms / 141.1ms)

OneFlow resnet50 time: 81.4ms (= 8137.4ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 93.1ms (= 9314.7ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.14 (= 93.1ms / 81.4ms)

OneFlow resnet50 time: 50.9ms (= 10178.2ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 71.1ms (= 14214.7ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.40 (= 71.1ms / 50.9ms)

OneFlow resnet50 time: 33.5ms (= 6707.1ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 64.3ms (= 12862.6ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.92 (= 64.3ms / 33.5ms)

OneFlow resnet50 time: 25.3ms (= 5058.2ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 65.0ms (= 12990.6ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 2.57 (= 65.0ms / 25.3ms)

OneFlow swin dataloader time: 0.240s (= 47.912s / 200, num_workers=1)
PyTorch swin dataloader time: 0.150s (= 30.004s / 200, num_workers=1)
Relative speed: 0.626 (= 0.150s / 0.240s)

OneFlow swin dataloader time: 0.068s (= 13.591s / 200, num_workers=4)
PyTorch swin dataloader time: 0.041s (= 8.266s / 200, num_workers=4)
Relative speed: 0.608 (= 0.041s / 0.068s)

OneFlow swin dataloader time: 0.043s (= 8.509s / 200, num_workers=8)
PyTorch swin dataloader time: 0.023s (= 4.538s / 200, num_workers=8)
Relative speed: 0.533 (= 0.023s / 0.043s)

❌ OneFlow resnet50 time: 152.6ms (= 15262.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 164.1ms (= 16407.3ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 1.08 (= 164.1ms / 152.6ms)

OneFlow resnet50 time: 92.7ms (= 9273.8ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 103.9ms (= 10392.1ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.12 (= 103.9ms / 92.7ms)

OneFlow resnet50 time: 60.7ms (= 12130.1ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 79.2ms (= 15831.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.31 (= 79.2ms / 60.7ms)

OneFlow resnet50 time: 42.4ms (= 8475.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 73.6ms (= 14728.8ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.74 (= 73.6ms / 42.4ms)

OneFlow resnet50 time: 36.9ms (= 7384.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 69.6ms (= 13916.5ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.88 (= 69.6ms / 36.9ms)

github-actions · 2023-04-01T02:21:17Z

View latest API docs preview at: https://staging.oneflow.info/docs/Oneflow-Inc/oneflow/pr/10050/

release np array in main thread to eliminate the time-consuming gil r…

235d38e

…equest Signed-off-by: daquexian <daquexian566@gmail.com>

daquexian added enhancement automerge eager labels Mar 28, 2023

add mutex

575eb5e

Signed-off-by: daquexian <daquexian566@gmail.com>

liujuncheng approved these changes Mar 28, 2023

View reviewed changes

clackhan approved these changes Mar 28, 2023

View reviewed changes

Merge branch 'master' into remove_gil_when_rel_instr

f65806b

liujuncheng enabled auto-merge (squash) March 28, 2023 07:40

liujuncheng requested a review from oneflow-ci-bot March 28, 2023 07:40

Merge branch 'master' into remove_gil_when_rel_instr

b56f719

github-actions bot removed the automerge label Mar 28, 2023

daquexian added the automerge label Mar 28, 2023

Merge branch 'master' into remove_gil_when_rel_instr

adccdbd

github-actions bot removed the automerge label Mar 28, 2023

liujuncheng added the automerge label Mar 28, 2023

Merge branch 'master' into remove_gil_when_rel_instr

4b6e476

github-actions bot removed the automerge label Mar 28, 2023

daquexian and others added 2 commits March 29, 2023 10:17

run remaining pending tasks in destructor

b0a48cc

Signed-off-by: daquexian <daquexian566@gmail.com>

auto format by CI

cbad400

liujuncheng requested review from oneflow-ci-bot and removed request for oneflow-ci-bot March 29, 2023 02:36

daquexian added the automerge label Mar 29, 2023

Merge branch 'master' into remove_gil_when_rel_instr

4308bf0

github-actions bot removed the automerge label Mar 29, 2023

sleep before c1 to debug

137ab19

Signed-off-by: daquexian <daquexian566@gmail.com>

daquexian requested a review from jackalcooper as a code owner March 29, 2023 08:25

daquexian added 2 commits March 29, 2023 19:52

debug with ssh action

4d23a13

Signed-off-by: daquexian <daquexian566@gmail.com>

Revert "debug with ssh action"

94fe891

This reverts commit 4d23a13.

daquexian added 2 commits March 30, 2023 23:14

increase numpy version for python3.7

6c92b99

Signed-off-by: daquexian <daquexian566@gmail.com>

Revert "sleep before c1 to debug"

741b080

This reverts commit 137ab19.

daquexian force-pushed the remove_gil_when_rel_instr branch from e1691b5 to 741b080 Compare March 30, 2023 15:14

python3.6 -> python3.7

fb871b6

Signed-off-by: daquexian <daquexian566@gmail.com>

jackalcooper approved these changes Mar 31, 2023

View reviewed changes

daquexian added the automerge label Mar 31, 2023

mergify bot added 2 commits March 31, 2023 02:36

Merge branch 'master' into remove_gil_when_rel_instr

289889f

Merge branch 'master' into remove_gil_when_rel_instr

ab0d968

github-actions bot removed the automerge label Apr 1, 2023

liujuncheng merged commit b7aac72 into master Apr 1, 2023

liujuncheng deleted the remove_gil_when_rel_instr branch April 1, 2023 02:50

yz-chen18 restored the remove_gil_when_rel_instr branch May 8, 2023 05:41

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

release np array in main thread to eliminate the time-consuming gil request #10050

release np array in main thread to eliminate the time-consuming gil request #10050

daquexian commented Mar 28, 2023

github-actions bot commented Mar 28, 2023

github-actions bot commented Mar 28, 2023

github-actions bot commented Mar 28, 2023

github-actions bot commented Mar 28, 2023

github-actions bot commented Mar 29, 2023

github-actions bot commented Mar 29, 2023

daquexian commented Mar 30, 2023

github-actions bot commented Mar 30, 2023

github-actions bot commented Mar 30, 2023

github-actions bot commented Apr 1, 2023

github-actions bot commented Apr 1, 2023

github-actions bot commented Apr 1, 2023

github-actions bot commented Apr 1, 2023

github-actions bot commented Apr 1, 2023

release np array in main thread to eliminate the time-consuming gil request #10050

release np array in main thread to eliminate the time-consuming gil request #10050

Conversation

daquexian commented Mar 28, 2023

github-actions bot commented Mar 28, 2023

github-actions bot commented Mar 28, 2023

github-actions bot commented Mar 28, 2023

github-actions bot commented Mar 28, 2023

github-actions bot commented Mar 29, 2023

github-actions bot commented Mar 29, 2023

daquexian commented Mar 30, 2023

github-actions bot commented Mar 30, 2023

github-actions bot commented Mar 30, 2023

github-actions bot commented Apr 1, 2023

github-actions bot commented Apr 1, 2023

github-actions bot commented Apr 1, 2023

github-actions bot commented Apr 1, 2023

github-actions bot commented Apr 1, 2023