[Frontend] Print out config when autotune fails

[htyu]

Printing out config when autotune fails under TRITON_DEBUG. This will allow kernel developers to narrow down a problematic config and remove it to unblock themselves. This could also help compiler developers focus on debugging that config.

[Jokeren]

Can you use TRITON_PRINT_AUTOTUNING?

[htyu]

Can you use TRITON_PRINT_AUTOTUNING?

Sometimes TRITON_PRINT_AUTOTUNING isn't enough. For example, in the present of runtime failure or llvm failure that cannot captured by the compiler driver, the process simply quits, and users don't get a message about which autotune config went wrong.

[bertmaher]

Maybe instead of gating this functionality on TRITON_DEBUG it should be added to TRITON_PRINT_AUTOTUNING I feel like if you're using the print-autotuning flag, you're probably debugging something about autotuning, and it might be nicer to have a single flag for that instead of two, particularly since TRITON_DEBUG also gates some debugging features in codegen (device-side asserts and stuff)

[htyu]

Maybe instead of gating this functionality on TRITON_DEBUG it should be added to TRITON_PRINT_AUTOTUNING I feel like if you're using the print-autotuning flag, you're probably debugging something about autotuning, and it might be nicer to have a single flag for that instead of two, particularly since TRITON_DEBUG also gates some debugging features in codegen (device-side asserts and stuff)

Sounds good. Swithed to using TRITON_PRINT_AUTOTUNING.

[bertmaher]

Lgtm! I’ll wait a bit to merge just in case anyone else wants to weigh in