Error: 'migraphx.mlir.as.underlying.shape' op memory layout dimension 12 is not a multiple of logical dimension 5; this indicates invalid strides

[pfultz2]

Following up on #2195. So this happens with non-packed output buffers like <4x5x24xf16, 288x24x1> which happens when concatting a buffer of 4x5x24 and 4x7x24 together which has a final buffer as <4x12x24xf16, 288x24x1>. So the strides are valid. Here is the mlir module:

module {
  func.func @mlir_dot_sigmoid(%arg0: !migraphx.shaped<4x5x16xf16, 80x16x1>, %arg1: !migraphx.shaped<4x16x24xf16, 384x24x1>) -> !migraphx.shaped<4x5x24xf16, 288x24x1> attributes {arch = "gfx1201", kernel = "mixr", num_cu = 32 : i64} {
    %0 = migraphx.dot %arg0, %arg1 : <4x5x16xf16, 80x16x1>, <4x16x24xf16, 384x24x1> -> <4x5x24xf16, 120x24x1>
    %1 = migraphx.sigmoid %0 : <4x5x24xf16, 120x24x1> -> <4x5x24xf16, 288x24x1>
    return %1 : !migraphx.shaped<4x5x24xf16, 288x24x1>
  }
}

And here is the error:

Invalid MLIR created: Error: 'migraphx.mlir.as.underlying.shape' op memory layout dimension 12 is not a multiple of logical dimension 5; this indicates invalid strides
Note: see current operation: %10 = "migraphx.mlir.as.underlying.shape"(%9) : (tensor<4x5x24xf16>) -> !migraphx.shaped<4x5x24xf16, 288x24x1>

[pfultz2]

It seems this error comes from this check here:

https://github.com/ROCm/rocMLIR/blame/ffbacc7d379dc1dc02ec96e2e5319be3cf7ca09a/mlir/lib/Conversion/MIGraphXToTosa/MIGraphXToTosa.cpp#L1446

This check was added in #2198. Is this check necessary? @justinrosner

[pfultz2]

I see, it looks like @dhernandez0 was asking about what would happen with this scenario which lead to adding the checks:

https://github.com/ROCm/rocMLIR/pull/2198/changes#r2668184953

But we do need support for these kind of strides.