Padding support for wave transfer

[EnricoDeg]

Proposed changes

Summary:

• Add padding support for wave transfer with transpose:
  • if loading index is in the padding region, read data at index 0 (always valid) to be able to use global load with transpose at wave level
  • before writing to lds, set register data to 0 if loading index was in the padding region
• there are still some validity restrictions with transpose which are checked before dispatching the kernel (specific for wave transfer):
  • for 16 bit types, each 8x8 subtile must be fully in the valid or the padding region
  • for 8 bit types, each 8x16 subtile must be fully in the valid or the padding region
• New test cases added for gemm universal to check new validity restrictions

Wave transfer can now be applied when both the vector size for loading from Vmem and the vector size for storing to LDS are equal to 8.

Next step: integrate wave transfer in convolution when it maps to explicit gemm (for default convolution, the thread transfer will still be used)

Checklist

Please put an x into the boxes that apply. You can also fill these out after creating the PR. If you're not sure, please don't hesitate to ask.

• I have added tests relevant to the introduced functionality, and the unit tests are passing locally
• I have added the test to REGRESSION_TESTS list defined at the top of CMakeLists.txt in tests/CMakeLists.txt, IF the test takes more than 30 seconds to run.
• I have added inline documentation which enables the maintainers with understanding the motivation
• I have removed the stale documentation which is no longer relevant after this pull request
• (If this change is user-facing) I have added release notes which provide the end users with a brief summary of the improvement from this pull request
• I have run clang-format on all changed files
• Any dependent changes have been merged

Discussion

If this is a relatively large or complex change, feel free to start a discussion by explaining why you chose the solution you did and what alternatives you considered

[krithalith]

Can we add comments here to indicate the names of the template parameters? It can be a bit hard to tell with this many bools in a row. Same for the CTranspose version.

[EnricoDeg]

Done

[krithalith]

Looks good! I had one small comment and also I was wondering if we still need to force threadTileTransfer for the convolution implementations. It seems that we still set this to true for all of them, with the exception of a small handful of special Fwd instances without CTranspose.

[ErwinTerpstra]

Nice improvements! I had some small questions and comments, but nothing major. I also have to admit I didn't fully grok the changes in the tensor slice transfer, so couldn't comment on that too much.

[EnricoDeg]

Looks good! I had one small comment and also I was wondering if we still need to force threadTileTransfer for the convolution implementations. It seems that we still set this to true for all of them, with the exception of a small handful of special Fwd instances without CTranspose.

In order to have better support in convolution, we need to change the handling of grid descriptors like it was done in conv fwd: create M,K grid descriptors on host and then modify them on the device to be K0,M,K1 for thread transfer and something more complicated for wave transfer. This is already work in progress for conv bwd