[INTERPRETER] Use acc's dtype as computation dtype in tl.dot

[tongyuantongyu]

I found with TRITON_INTERPRET=1, tl.dot almost always gives wrong result for INT8 input, due to np.matmul use inputs' dtype as result dtype by default:

>>> a = np.array([[10, 20]], np.int8)
>>> np.matmul(a, a.transpose())
array([[-12]], dtype=int8)
>>> np.matmul(a, a.transpose(), dtype=np.int32)
array([[500]])

[jlebar]

Thank you! Probably we should add a test, or fix an existing one so it catches this bug?

[Jokeren]

Yes, please add a test case

[tongyuantongyu]

Done. It seems related test cases were deliberately skipped due to this deficiency.

[Jokeren]

Thanks

[tongyuantongyu]

Side note: I found a 3xtf32 vs tf32x3 mismatch when running related tests on my end:

https://github.com/openai/triton/blob/c9800840489fbe04619ad2d4e0e46b868c589d48/python/triton/runtime/interpreter.py#L73
https://github.com/openai/triton/blob/c9800840489fbe04619ad2d4e0e46b868c589d48/python/test/unit/language/test_core.py#L2919

Both are from commit 47a35b6, not sure why this was not caught by CI.

[Jokeren]

allowed_dot_input_precisions

It doesn't quite matter. allowed_dot_input_precisions is just a placeholder, not used

[Jokeren]

allowed_dot_input_precisions feel free to fix the typo in this PR directly. Maybe the name was changed.