Support blocked dot operand layout conversion to linear layout

[binarman]

Goal

Support FMA/Blocked dot operand layout in linear layout converter.

DoD

Implemented LinearLayout converter and related tests are implemented and passing, switch to Linear layout converter do not break python tests.

Existing Linear Layout converters for dot operands

Nvidia MMA dot operand
AMD MFMA dot operand

Both of these examples are fully functional, but I want to refactor MFMA and WMMA converters in similar to Nvidia MMA fashion soon. Please try to follow MMA style.

In progress PRs

WMMA dot operand

Legacy SharedLayout->dotOperand converter examples

• Shared Layout -> Blocked/FMA dot operand
• Shared Layout -> MFMA dot operand
• Shared Layout -> WMMA dot operand

Implementation details

1. Implement linear converter similar to WMMA/MFMA (I am going to refactor these converter a little soon) in LinearLayoutConversions.cpp; Add appropriate call in DotOperandEncodingAttr::toLinearLayout function
2. Add blocked layout in MemoryOpToLLVM.cpp:isSupportedDotOpLayout to enable shared->blocked dotOperand conversion with LL; check all test_core.py tests work.
3. Enable LL converter in ConvertLayoutOpToLLVM.cpp:transferWithinBlock to enable ttg.convert_layout operation conversion with LL. Maybe you need to rework some other places as well, we will clarify this during implementation.
4. Implement ctest tests in LinearLayoutConversionsTest.cpp and few lit tests to specifically verify ttg.convert_layout operation conversion

[binarman]

+cc @simonidaa
please, consider this task next

[Jokeren]

+@lezcano Maybe you can offload this to AMD? Or there's something you've already done but not pushed yet?

[Jokeren]

@binarman btw, something related, we plan to completely abandon the getCSwizzleOffset method and use allocShape in memory descriptor to determine if prefetch has occurred.

[binarman]

Thank you @Jokeren! this is good to know, I will consider removing getCSwizzleOffset from our converters.

[minjang]

Very good to hear for this work!

(For my experimental project, using LL for non-GPUs, yeah, this was a missing feature. I hacked badly just to make it passable for unit tests, but looking forward to seeing this support.)

[lezcano]

I did not start with this, so feel free to take over.

[binarman]

@lezcano told me this task is blocking him, so I am going to implement it asap,
@simonidaa I will look for something else, but still related to LL for you

[binarman]

I've made a draft PR: #5469
Along the way I understood that current FMA implementation does not align with the rest of layouts, so I decided to rework it. Currently some tests are failing probably will finish this work tomorrow.