Conversion/TorchToTosa/basic.mlir failing on ARM, due to different random numbers in torch.aten.uniform

[dbabokin]

After 4dd213b by @justin-ngo-arm, I see Conversion/TorchToTosa/basic.mlir test is failing on my ARM-based MacbookPro for torch.aten.uniform check. This line doesn't match:

torch-mlir/test/Conversion/TorchToTosa/basic.mlir

Line 2266 in 7058f45

// CHECK: %[[VAL_4:.*]] = "tosa.const"() <{value = dense<{{\[\[}}1.00007045, 2.18384027, 7.80044794, 5.12785149], [5.79490519, 2.97063255, 1.42340159, 7.10978221], [7.11366796, 9.41223621, 4.45151854, 5.67474747]]> : tensor<3x4xf32>}> : () -> tensor<3x4xf32>

Instead I have the following line in my output:

%0 = "tosa.const"() <{value = dense<[[1.0002023, 1.76529205, 6.41217375, 9.0245018], [9.71160125, 2.70720792, 5.63478279, 4.5820756], [3.36615539, 7.69161224, 1.8059299, 6.04350948]]> : tensor<3x4xf32>}> : () -> tensor<3x4xf32>

Which is also uniform distribution, but with different random numbers. The test works fine for me on x86 Linux, though.

I'm not sure if the test needs to be fixed or torch.aten.uniform itself to be more reproducible across platforms.

[justin-ngo-arm]

Hi @dbabokin , I've fixed the test and will tag you in the PR review process. Thank you for catching this issue.

[dbabokin]

@justin-ngo-arm Thanks for blazingly fast response!

[justin-ngo-arm]

@dbabokin I've tagged you in a comment in this recently merged PR - #3855 - which includes the fix for the issue. Please let me know if this resolves your problem and we can close the issue out. Thank you!

[dbabokin]

I've just checked and the issue is fixed. Thank you!