depthwise_conv2d_implicit_gemm slower than nn.Conv2d

[wdmwhh]

🐛 Describe the bug
Calling depthwise_conv2d_implicit_gemm.DepthWiseConv2dImplicitGEMM, on CUDA, is orders of magnitude slower than calling torch.nn.Conv2d.

I have installed it according to README.

cc: @DingXiaoH
Versions
torch 1.8.2+cuda11.1
cuda-11.1.1 + cudnn-8.1.1
both A100 and V100

[wdmwhh]

Here I add the speed test dwblocks_speed.py.

test on
python 3.7.11 + torch 1.8.2 + cuda-11.1.1 + cudnn-8.1.1 + V100

[wdmwhh]

10x slower: depthwise_conv2d_implicit_gemm.DepthWiseConv2dImplicitGEMM takes 0.01943465073903402s while nn.Conv2d takes 0.0012518405914306641s.

[DingXiaoH]

Hi, I checked the code and found no "synchronized()" so that the time recorded may not be the actual running time on GPU. I would suggest you follow the speed test script of Swin (https://github.com/microsoft/Swin-Transformer/blob/main/main.py#L287)

[wdmwhh]

The test code is a small replication of the phenomenon (depthwise_conv2d_implicit_gemm slower), which occurred in training a large model.

[wdmwhh]

The code that adds torch.cuda.synchronize() before calling time.time() gives rather close time to the original code.

[xiaocenxiaocen]

This implementation is not suited for small batch sizes. In this case the batch size is 1, so the cutlass implmentation is slower than pytorch. You can try megengine instead.

[wdmwhh]

Thanks for your reply. It help me a lot.

[YanShuang17]

I meet the same question.

I trained ATSS detector with ReoLKNet31B and batch_size 1(2080Ti GPU, 11 GB memory..., and 'use_checkpoint' seems to be not compatible with DDP):

• when use torch.nn.Conv2d(), training time is about 1.00s per iteration.
• when use DepthWiseConv2dImplicitGEMM, training time is about 4.87s per iteration.

[EddieEduardo]

Hi, I encountered with the same problem.
When using nn.Conv2d, the running time of the model is just ~0.5s,
while using the DepthWiseConv2dImplicitGEMM, the time is ~6s.
The batchsize is set to 1 owing to the memory (RTX3060, 1 single GPU, 12G).

[DingXiaoH]

Thank you for sharing the results. As explained by @xiaocenxiaocen , our implementation is designed to pursue high throughput. Larger the batch size, higher the throughput.