[OPENCL] Always use convert_T for type conversion

[tqchen]

This PR changes the Cast in OpenCL to always relying on convert_T to get closer to the spec and more reliable.

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @echuraev, @elvin-n _{See #10317 for details}

_{Generated by tvm-bot}

[echuraev]

This test has covered a bug in TVM which was fixed. Why did you change the compute function? Can we keep it as it was?

[tqchen]

i also migrated the original test to the latest TensorIR so the test is future proof.

Interestingly TensorIR becomes smarter :) so the original condition get simplified and becomes i == 15 here. So I modified the test to ensure we cover most behavior we intend to cover around type casting

[tqchen]

Unfortunately relying on string pattern was too reliable given the set of possible transformations here. So we need to adapt the testcase as we go. If there is a new test cases in TVMScript, i think we can also put it here

[echuraev]

so the original condition get simplified and becomes i == 15 here

Do you mean that this condition:

tvm.tir.all(
    *[
        i // block_size == tvm.tir.const(3, "int32"),
        i % block_size == tvm.tir.const(3, "int32"),
    ]
),

Was transformed to i == 15?

[echuraev]

I took a look at the git history and the original problem which should be covered by this test was that left and right parts of condition should have the same data type. You can see detailed description in #11021. And next this test was added in #11038.

By decreasing the number of conditions in tvm.tir.all, you prevent compiler to generate code such as lcond && rcond. Am I right? In this case, the original problem won't be tested. Sorry if I missed something.

[tqchen]

Sorry I wasn't being clear, right

tvm.tir.all(
    *[
        i // block_size == tvm.tir.const(3, "int32"),
        i % block_size == tvm.tir.const(3, "int32"),
    ]
)

In TensorIR schedule the ^ condition get transformed to i == 15, as indeed that makes sense.

So we might need a different condition to test the chains.

[tqchen]

OK updated to a case where simplification won't happen

[echuraev]

LGTM! Thank you :)

[tqchen]

@tvm-bot rerun

[echuraev]

@tvm-bot rerun