[INTERPRETER] Support padding_option of tl.load

[tongyuantongyu]

This would make debugging kernels using block pointer easier.

[Jokeren]

Please also add the interpreter annotation to test_block_ptr_matmul_no_scf

[oliverdutton]

Hello, love block pointers.

Why is the mask value forced as an enum in padding_option rather than through the other spec?

For example, using block pointers for writing a softmax kernel I want to load with masking generating -inf?

Can we support other but only allowing a scalar value?

Similarly, if I ran a tl.min you've want inf. Or a cumprod (not a great idea I know) it'd be 1 for null op.

[Jokeren]

We tired to mimick the tensor map functionality in cuda. See CUtensorMapFloatOOBfill

[oliverdutton]

I see, and this is a fast path with only two options. Much faster than the explicit checking and compare that other would do?

would you be open to a ‘use padding_mode’ or ‘other’ version of this when you can specify either (but not both)? [or could you infer zero later on and specialise to the backend, such as this NVIDIA op]?

[Jokeren]

I see, and this is a fast path with only two options. Much faster than the explicit checking and compare that other would do?

On GPU, yes.

would you be open to a ‘use padding_mode’ or ‘other’ version of this when you can specify either (but not both)? [or could you infer zero later on and specialise to the backend, such as this NVIDIA op]?

TMA is temporarily disabled in Triton, so there's only subtle difference between those two options. Using the padding mode will be slightly faster in practice at this moment.

But once TMA is back, please use the padding mode.