[MLIR] Incorrect result for RuntimeVerifiableOpInterface on MemRef::ReinterpretCastOpInterface

[chentong319]

The upper bound check is wrong.
Simple example:

    %reinterpret_cast = memref.reinterpret_cast %arg0 to offset: [0], sizes: [12, 64], strides: [64, 1] : memref<768xf32> to memref<12x64xf32>

Obvious, 768 = 12x64. But the result contains assert(false, ...).

I did some investigation. The generated upper bounds for base and result are 668 and 732 . The indices used to compute the upper bound are [768] and [12, 64]. I think that the indices of last element should be [767] and [11, 63]. In general, the index of a dimension starts from 0 and ends at size-1.

[joker-eph]

Can you provide a complete reproducer?

[chentong319]

Source code:

func.func @main_graph(%arg0: memref<768xf32>) -> (memref<12x64xf32>) attributes {llvm.emit_c_interface} {
  %reinterpret_cast = memref.reinterpret_cast %arg0 to offset: [0], sizes: [12, 64], strides: [64, 1] : memref<768xf32> to memref<12x64xf32>
  return %reinterpret_cast : memref<12x64xf32>
}

Command: mlir-opt error.mlir --generate-runtime-verification

Output:

  func.func @main_graph(%arg0: memref<768xf32>) -> memref<12x64xf32> attributes {llvm.emit_c_interface} {
    %reinterpret_cast = memref.reinterpret_cast %arg0 to offset: [0], sizes: [12, 64], strides: [64, 1] : memref<768xf32> to memref<12x64xf32>
    %base_buffer, %offset, %sizes, %strides = memref.extract_strided_metadata %arg0 : memref<768xf32> -> memref<f32>, index, index, index
    %c0 = arith.constant 0 : index
    %c768 = arith.constant 768 : index
    %base_buffer_0, %offset_1, %sizes_2:2, %strides_3:2 = memref.extract_strided_metadata %reinterpret_cast : memref<12x64xf32> -> memref<f32>, index, index, index, index, index
    %c0_4 = arith.constant 0 : index
    %c832 = arith.constant 832 : index
    %true = arith.constant true
    %false = arith.constant false
    cf.assert %false, "ERROR: Runtime op verification failed\0A%0 = \22memref.reinterpret_cast\22(%arg0) <{operandSegmentSizes = array<i32: 1, 0, 0, 0>, static_offsets = array<i64: 0>, static_sizes = array<i64: 12, 64>, static_strides = array<i64: 64, 1>}> : (memref<768xf32>) -> memref<12x64xf32>\0A^ result of reinterpret_cast is out-of-bounds of the base memref\0ALocation: loc(\22error.mlir\22:2:23)"
    return %reinterpret_cast : memref<12x64xf32>
  }

The upper bounds are 768 and 832, causing the assertion error: cf.assert %false .

[ethanuppal]

Hi - this would be my first time contributing to LLVM, but since it's marked as a "good first issue", I'd like to take a stab. It seems the relevant part of the code is here?

llvm-project/mlir/lib/Dialect/MemRef/Transforms/RuntimeOpVerification.cpp

Line 234 in bce2498

auto assertCond = builder.createOrFold<arith::AndIOp>(loc, geLow, leHigh);

[joker-eph]

Seems like you find the right function @ethanuppal !

[ethanuppal]

Hopefully I can start working on a fix once I can make LLVM compile!

[ethanuppal]

Just an update: I'm still interested in working on this, but I still can't get LLVM to build from source. I'll keep trying.

[ethanuppal]

Since I'm still having troubles building LLVM from source, it might be better to have someone else take a look at this :(

[hmalgewatta]

Hi! I would like to work on this. This will be my first time contributing too.

[ethanuppal]

Yay, glad @hmalgewatta was able to get a PR. In the meantime, I've solved my building issues :)

[chentong319]

@hmalgewatta Could you review this issue?