[Dataflow] Packed systolic matmul test caes

[jiahanxie353]

Description

This PR tries to add a test case for packed systolic using the Dataflow module

Proposed Solutions

I have added a test case called tests/dataflow/test_packed_systolic.py; but I'm still trying to figure out the indexing problems on this line, which raises:

  File "allo/allo/ir/builder.py", line 1232, in build_Subscript
    return ASTTransformer.build_bit_operation(ctx, node, val=val, idx=idx)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "allo/allo/ir/builder.py", line 1152, in build_bit_operation
    value = build_stmt(ctx, node.value)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "allo/allo/ir/builder.py", line 73, in __call__
    res = method(ctx, node, **kwargs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "allo/allo/ir/builder.py", line 1228, in build_Subscript
    return ASTTransformer.build_memory_access(ctx, node, val=val, idx=idx)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "allo/allo/ir/builder.py", line 1117, in build_memory_access
    val.results[idx], value.result, ivs, affine_attr, ip=ctx.get_ip()
    ^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'results'

Examples

Example is shown in the tests/dataflow/test_packed_systolic.py. Is there any obvious error in the example?

Checklist

• PR's title starts with a category (e.g. [Bugfix], [IR], [Builder], etc)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage (It would be better to provide ~2 different test cases to test the robustness of your code)
• Code is well-documented

[chhzh123]

@jiahanxie353 Thanks for contributing! Please follow this guideline to format your code.

[jiahanxie353]

@jiahanxie353 Thanks for contributing! Please follow this guideline to format your code.

thanks for getting back @chhzh123 !
can you maybe send another link to the guideline? the one you sent was the homepage of allo

[chhzh123]

https://cornell-zhang.github.io/allo/developer/index.html#id1

[jiahanxie353]

https://cornell-zhang.github.io/allo/developer/index.html#id1

thanks! I believe I have passed all the code-format testing

any advice on fixing the issue? I wonder why I can't index into an int16[][] type like this: Z_packed[i - 1, j - 1][k * 8 : (k + 1) * 8]

[chhzh123]

Maybe you can try separate the memory access -- first load Z_packed[i - 1, j - 1], increment the bits of [k * 8 : (k + 1) * 8], and then store back the new value. You can check out a data packing example here.

[chhzh123]

Can you test a larger systolic array? I tested your code with L, D = 4, 4 -- the effective SA size of 2x2, and it failed

[chhzh123]

What is 1 here? Maybe create a constant variable?

[jiahanxie353]

I removed L, D and only kept M, N, K for clearance. Please let me know if you prefer to keep L, D

[chhzh123]

Can you also add PP=4, 8 cases?

[jiahanxie353]

Sure, I added PP=4.
Now I'm hard-coding each basic element size is int8, which forces allo_type == int8 * PP. I will make it parametric and add different PP-allo_type combinations

[jiahanxie353]

This line looks suspicious.
X_packed has type allo_type[M, K // PP]. And for example, when M, N, K = 4, 4, 2, X_packed type is allo_type[4, 1]. And if we have a for-loop: for k in range(2). Suppose k = 1, not only it didn't go out of bound when accessing X_packed array, the results are all correct...

maybe we should look into this...