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 CUBLAS Matmul BiasAdd Grad Kernel #8063

Merged
merged 17 commits into from
Apr 29, 2022
Merged

Add CUBLAS Matmul BiasAdd Grad Kernel #8063

merged 17 commits into from
Apr 29, 2022

Conversation

MARD1NO
Copy link
Contributor

@MARD1NO MARD1NO commented Apr 20, 2022

No description provided.

@MARD1NO MARD1NO added WIP work in progress embedding labels Apr 20, 2022
@MARD1NO MARD1NO changed the title Add CUBLAS Matmul BiasAdd Grad Kernel[WIP] Add CUBLAS Matmul BiasAdd Grad Kernel Apr 22, 2022
@MARD1NO MARD1NO marked this pull request as ready for review April 22, 2022 05:51
@MARD1NO MARD1NO added enhancement op and removed WIP work in progress labels Apr 22, 2022
liujuncheng
liujuncheng reviewed Apr 26, 2022
View reviewed changes
oneflow/user/kernels/cublas_fused_matmul_bias_add_grad.cu Outdated
#include "oneflow/core/ep/include/primitive/memcpy.h"
#include "oneflow/core/ep/cuda/cuda_device.h"
// CUBLAS_AUX_EPILOGUE only support in cuda11.4 or higher version, in cuda11.4 it need static link.
#if CUDA_VERSION >= 11040
Copy link
Collaborator

Choose a reason for hiding this comment

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

https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html#cublas-11.4.2

Copy link
Contributor Author

Choose a reason for hiding this comment

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

已修改为11042

oneflow/user/kernels/cublas_fused_matmul_bias_add_grad.cu Outdated
#include "oneflow/user/kernels/cublas_fused_mlp_util.cuh"
#include "oneflow/core/ep/include/primitive/memcpy.h"
#include "oneflow/core/ep/cuda/cuda_device.h"
// CUBLAS_AUX_EPILOGUE only support in cuda11.4 or higher version, in cuda11.4 it need static link.
Copy link
Collaborator

Choose a reason for hiding this comment

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

Suggested change
// CUBLAS_AUX_EPILOGUE only support in cuda11.4 or higher version, in cuda11.4 it need static link.
// CUBLASLT_EPILOGUE_BGRADB only support in cuda11.4 or higher version, in cuda11.4 it need static link.

?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

已修改

@MARD1NO MARD1NO requested a review from oneflow-ci-bot April 28, 2022 00:09
@MARD1NO MARD1NO requested review from oneflow-ci-bot and removed request for oneflow-ci-bot April 28, 2022 09:20
@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/8063/

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 129.3ms (= 12931.4ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 142.6ms (= 14258.4ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.10 (= 142.6ms / 129.3ms)

OneFlow resnet50 time: 83.7ms (= 8370.1ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 84.9ms (= 8489.5ms / 100, input_shape=[8, 3, 224, 224])
❌ Relative speed: 1.01 (= 84.9ms / 83.7ms)

OneFlow resnet50 time: 51.3ms (= 10264.5ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 58.9ms (= 11785.6ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.15 (= 58.9ms / 51.3ms)

OneFlow resnet50 time: 44.1ms (= 8810.0ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 50.3ms (= 10068.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.14 (= 50.3ms / 44.1ms)

OneFlow resnet50 time: 35.8ms (= 7168.8ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 37.8ms (= 7559.7ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.05 (= 37.8ms / 35.8ms)

OneFlow swin dataloader time: 0.255s (= 51.032s / 200, num_workers=1)
PyTorch swin dataloader time: 0.153s (= 30.605s / 200, num_workers=1)
Relative speed: 0.600 (= 0.153s / 0.255s)

OneFlow swin dataloader time: 0.068s (= 13.550s / 200, num_workers=4)
PyTorch swin dataloader time: 0.044s (= 8.700s / 200, num_workers=4)
Relative speed: 0.642 (= 0.044s / 0.068s)

OneFlow swin dataloader time: 0.035s (= 7.069s / 200, num_workers=8)
PyTorch swin dataloader time: 0.021s (= 4.282s / 200, num_workers=8)
Relative speed: 0.606 (= 0.021s / 0.035s)

❌ OneFlow resnet50 time: 145.2ms (= 14520.5ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 169.4ms (= 16938.8ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.17 (= 169.4ms / 145.2ms)

OneFlow resnet50 time: 96.2ms (= 9621.3ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 109.0ms (= 10902.7ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.13 (= 109.0ms / 96.2ms)

OneFlow resnet50 time: 75.8ms (= 15166.0ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 88.7ms (= 17731.5ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
❌ Relative speed: 1.17 (= 88.7ms / 75.8ms)

OneFlow resnet50 time: 63.1ms (= 12610.0ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 75.3ms (= 15068.2ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.19 (= 75.3ms / 63.1ms)

OneFlow resnet50 time: 57.5ms (= 11502.3ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 69.4ms (= 13881.9ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.21 (= 69.4ms / 57.5ms)

@github-actions
Copy link
Contributor

Speed stats: 
GPU Name: NVIDIA GeForce GTX 1080 

❌ OneFlow resnet50 time: 129.2ms (= 12920.3ms / 100, input_shape=[16, 3, 224, 224])
PyTorch resnet50 time: 142.8ms (= 14284.8ms / 100, input_shape=[16, 3, 224, 224])
✔️ Relative speed: 1.11 (= 142.8ms / 129.2ms)

OneFlow resnet50 time: 80.1ms (= 8009.5ms / 100, input_shape=[8, 3, 224, 224])
PyTorch resnet50 time: 85.3ms (= 8528.6ms / 100, input_shape=[8, 3, 224, 224])
✔️ Relative speed: 1.06 (= 85.3ms / 80.1ms)

OneFlow resnet50 time: 50.4ms (= 10081.4ms / 200, input_shape=[4, 3, 224, 224])
PyTorch resnet50 time: 59.3ms (= 11862.5ms / 200, input_shape=[4, 3, 224, 224])
✔️ Relative speed: 1.18 (= 59.3ms / 50.4ms)

OneFlow resnet50 time: 41.3ms (= 8252.9ms / 200, input_shape=[2, 3, 224, 224])
PyTorch resnet50 time: 44.0ms (= 8802.1ms / 200, input_shape=[2, 3, 224, 224])
✔️ Relative speed: 1.07 (= 44.0ms / 41.3ms)

OneFlow resnet50 time: 38.9ms (= 7774.9ms / 200, input_shape=[1, 3, 224, 224])
PyTorch resnet50 time: 42.2ms (= 8443.6ms / 200, input_shape=[1, 3, 224, 224])
✔️ Relative speed: 1.09 (= 42.2ms / 38.9ms)

OneFlow swin dataloader time: 0.254s (= 50.713s / 200, num_workers=1)
PyTorch swin dataloader time: 0.149s (= 29.898s / 200, num_workers=1)
Relative speed: 0.590 (= 0.149s / 0.254s)

OneFlow swin dataloader time: 0.066s (= 13.253s / 200, num_workers=4)
PyTorch swin dataloader time: 0.042s (= 8.441s / 200, num_workers=4)
Relative speed: 0.637 (= 0.042s / 0.066s)

OneFlow swin dataloader time: 0.037s (= 7.456s / 200, num_workers=8)
PyTorch swin dataloader time: 0.022s (= 4.333s / 200, num_workers=8)
Relative speed: 0.581 (= 0.022s / 0.037s)

❌ OneFlow resnet50 time: 145.3ms (= 14530.2ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 169.6ms (= 16961.1ms / 100, input_shape=[16, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.17 (= 169.6ms / 145.3ms)

OneFlow resnet50 time: 96.7ms (= 9671.6ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 109.2ms (= 10915.7ms / 100, input_shape=[8, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.13 (= 109.2ms / 96.7ms)

OneFlow resnet50 time: 75.7ms (= 15145.7ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 98.9ms (= 19770.2ms / 200, input_shape=[4, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.31 (= 98.9ms / 75.7ms)

OneFlow resnet50 time: 65.5ms (= 13095.5ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 81.2ms (= 16241.7ms / 200, input_shape=[2, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.24 (= 81.2ms / 65.5ms)

OneFlow resnet50 time: 57.4ms (= 11479.7ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
PyTorch resnet50 time: 69.1ms (= 13829.7ms / 200, input_shape=[1, 3, 224, 224], ddp, world size=2)
✔️ Relative speed: 1.20 (= 69.1ms / 57.4ms)

@mergify mergify bot merged commit baee1b0 into master Apr 29, 2022
@mergify mergify bot deleted the dev_cublas_bias_weight_grad branch April 29, 2022 06:13
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@liujuncheng liujuncheng liujuncheng approved these changes

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

@hjchen2 hjchen2 Awaiting requested review from hjchen2 hjchen2 is a code owner

@BBuf BBuf Awaiting requested review from BBuf

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

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

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

Successfully merging this pull request may close these issues.
3 participants
@MARD1NO @liujuncheng @guo-ran