Add block ptr test for dot product with transpose

[alexbaden]

Adds the end-to-end dot product on block ptr testing to the block load unit test (maybe it should be renamed test_block_ptr.py?). Adds additional shapes, A transpose, and B transpose. The cold runtime (no cache) is approximately 1 minute on PVC 1100 in my environment. I picked the block shapes somewhat randomly, trying to balance breadth and runtime.

This somewhat duplicates tutorial 10 but allows us to run many more combinations in shorter time. I added this because #4463 is passing CI but has a few bugs that are not being caught by existing unit tests, including tutorials.

[whitneywhtsang]

I added this because #4463 is passing CI but has a few bugs that are not being caught by existing unit tests, including tutorials.

Is it because some configs/shapes are missing from tutorial 10 or is it because autotune would hide failures?

[alexbaden]

Is it because some configs/shapes are missing from tutorial 10 or is it because autotune would hide failures?

Both reasons - plus it is more expensive because the overall shape is bigger. Here we make the overall shape smaller which dramatically improves runtime, which is fine because we only care about large enough to have different configurations in a single block vs tutorial 10 where we want to see performance and need bigger shapes to really get good throughput.

[whitneywhtsang]

IMO we can remove all comments about the kernel, as it adds maintenance cost, one could refer to tutorial 3 or 10 to understand how to write gemm in Triton.

[whitneywhtsang]

Another thought is we can add test_op in tutorial 10, similar to how it is done in tutorial 6.

[alexbaden]

Let's go with this approach for now, I can move it to tutorial 10 (or enable the tensor descriptor version of tutorial 10 and then add test_op there, if upstream does not do that first). After all theoretically block_ptr won't be around much longer...

[whitneywhtsang]

Let's go with this approach for now, I can move it to tutorial 10 (or enable the tensor descriptor version of tutorial 10 and then add test_op there, if upstream does not do that first). After all theoretically block_ptr won't be around much longer...

Sure, I am fine to do this first.

[whitneywhtsang]

IMO we can remove all comments about the kernel, as it adds maintenance cost, one could refer to tutorial 3 or 10 to understand how to write gemm in Triton.

@alexbaden WDYT?

[alexbaden]

I think it's ok as is - long term we should either remove it, since block_ptr will be deprecated, or fold it into tutorial 10 (but probably remove it).
For now though this coverage is important as it lets us know the block load lowering has no regressions - we will still rely on that even if we move completely to tensor descriptor.

[alexbaden]

Synced with Whitney offline - going to merge this now and then either merge it into tutorial 10 (if block ptr sticks around) or probably delete it alongside tutorial 10 when block ptr goes away.