[SM6.10][Bugfix][Exec] Final test tweaks for preview

[V-FEXrt]

This is the final test change PR going into the initial preview build. We'll continue to add tests after the build that can be pulled from main or the preview branch.

In this PR are the following changes

• OuterProduct smoke test was removed as it requires an OuterProductOptimal Layout, which requires a bit more test harness work to verify. Instead of delaying the preview for it we are punting the test for now. It will quickly be ready after the preview release
• Groupshared operations require that stride be the number of "row elements" previously we were setting the "row bytes". This has been fixed
• The 2D->1D index calculation was incorrect but hidden by the fact that all test matrices are square. This has been fixed

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[tex3d]

I think it would be better to always pass STRIDE in bytes (and side chat brought up OFFSET, so same issue), then divide STRIDE and OFFSET by ELEM_SIZE for any groupshared array based operations. That's assuming ELEM_SIZE is represents the native HLSL scalar size, which is different than the element size for packed elements!

[xiaolin-ji]

Just to be clear, so for groupshared operation, the input offset to load&store is element count, not the byte offset?

For coordToByteOffset(), since coord.xy is (row, column), it should change to (coord.x * N_DIM + coord.y) * ELEM_SIZE for row major offset calculation.

[V-FEXrt]

Just to be clear, so for groupshared operation, the input offset to load&store is element count, not the byte offset?

Yes! The spec needs to be clarified on this point. I'll be making that change after preview processes calm down

For coordToByteOffset(), since coord.xy is (row, column), it should change to (coord.x * N_DIM + coord.y) * ELEM_SIZE for row major offset calculation.

Per the other thread, I'm still not sure we agree on this point. I am happy to revisit it post-preview though just to really confirm. Since the smoke test matrices are all square this shouldn't effect the current tests even if it's wrong as is. We'll be adding non-square tests very soon so we'll be forced to get it right at that point regardless

[xiaolin-ji]

Just to be clear, so for groupshared operation, the input offset to load&store is element count, not the byte offset?

Yes! The spec needs to be clarified on this point. I'll be making that change after preview processes calm down

For coordToByteOffset(), since coord.xy is (row, column), it should change to (coord.x * N_DIM + coord.y) * ELEM_SIZE for row major offset calculation.

Per the other thread, I'm still not sure we agree on this point. I am happy to revisit it post-preview though just to really confirm. Since the smoke test matrices are all square this shouldn't effect the current tests even if it's wrong as is. We'll be adding non-square tests very soon so we'll be forced to get it right at that point regardless

thanks confirm the offset to group share operation.
For coordToByteOffset(), I just leave the comment at another thread, yes, N_DIM and M_DIM is same, but actually, we need spec to be clear, current coord.x is coordinate within the row or just row index.
if coord.x is coordinate within the row, and coord.y is coordinate within the column, then it is (coord.y * DIM + coord.x) * ELEM_SIZE
if coord.x is row and coord.y is column, then it should be (coord.x * DIM + coord.y) * ELEM_SIZE