[TensorIR][M1c] LowerInitBlock

[Hzfengsy]

This PR is the first part of the stage M1c, TensorIR upstreaming plan (#7527) on lowering init block.

We just use if branch to replace the init block. e.g.

Before

with tir.block([16, 16, tir.reduce_axis(16)]) as [i, j ,k]:
    with tir.init():
        C[i, j] = tir.float32(0)
    C[i, j] += A[i, k] * B[j, k]

After

with tir.block([16, 16, tir.reduce_axis(16)]) as [i, j ,k]:
    if k == 0:
        C[i, j] = tir.float32(0)
    C[i, j] += A[i, k] * B[j, k]

Please note that we only provide the default behavior for lowering the init part, which is not performent. We will provide primitives for users to manually "decompose" init part.

cc @tqchen @junrushao1994 @jroesch @comaniac @jcf94

[tqchen]

minor nits, overall LGTM

[comaniac]

Overall LGTM.
One suggestion: Since we have defined a clear stage (i.e., M1c) for this series of PRs, it would be better for all PR submitters to use the unified PR titles (for exampe: [TensorIR][M1c] ...) so that they can be easily queried.

[comaniac]

Could you clarify the meaning of the word "realize" here?

[junrushao]

What about "concretize"? My understanding is that the init-block is something "conceptual" for initializing "zero point" of a reduction, and this pass is used to make it concrete, lowering to "if-then-else" statement

[comaniac]

I got your point, and I feel "LowerInitBlock" itself is the best name to describe what it does, so maybe the simplest way is manually inlining this function so that we won't be bothered by its name, or just give it a general helper name to let people know this is just the implementation of lowering init block.

[junrushao]

This helper is something a bit standalone (making the init block to If-Then-Else), and inlining this function might make the logic in the function above a bit vague, so what about we just call it "DoLowering"?

[comaniac]

DoLowering sounds good to me.

[junrushao]

Just a note here: per discussion with @Hzfengsy, we need to remove "()" here in future PRs, so that the script parser API is consistent across PrimFunc and IRModule

[tqchen]

@Hzfengsy please address last comment by @comaniac (DoLowering) and we can merge this in

[tqchen]

Thanks @comaniac @junrushao1994 for reviewing. Thanks @Hzfengsy