Tensor cores not being used for conv2d operations. (Tesla T4)
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 9k
- Forks
- 1.5k
- Avg merge
- 2d 4h
- Merged PRs (30d)
- 3
Description
The network is optimized using O2 optimization on -->
Cuda 10.0
PyTorch 1.0.0
CudNN 7.6.3
All conv2d operations have input output channels which are a multiple of 8. I did profiling of the code using the nvprof from apex
I have checked the outputs and the operations are outputting fp16 tensors (tensor.dtype in pytorch). Despite this the calculation is not on tensor cores and hence I am not able to observe any speedups that comes with FP16. Can someone please help me out? Any inputs will be really helpful. TIA.
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reviewing the reported PyTorch 1.0.0, cuDNN 7.6.3, and CUDA 10.0 setup, the nvprof profiling results, and the attached final_test.xlsx. Determine why the fp16 conv2d operations on the Tesla T4 are not using tensor cores; done means the cause is identified and the expected tensor-core behavior is documented or corrected.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100