[lower_to_mlir] Lower reduce op from tt-forge to mlir

[dgolubovicTT]

Lowering reduce sum and mean ops from tt-forge to mlir. Adding end to end tests for those ops. Adding method for pcc comparison with torch golden instead of using torch.allclose since it fails on reduce ops when data format is float32.

I checked on metal reduce tests (tests/ttnn/unit_
tests/operations/test_mean.py::test_mean) and using bfloat16 gives output that matches on fourth decimal with golden, while float32 has differences on third and fourth decimal.

[sdjordjevicTT]

Should we use this new function for verifying the output in other tests also?

[nvukobratTT]

I agree, can we update the rest of the test/mlir?

Also, please update pybuda/test/mlir/mnist/test_inference.py. At the moment, it has the similar issues as other ops. However, the current implementation doesn't detect it as [False] interpreted as valid assert case. In sum, code should look something like this:

def test_mnist_inference():
    inputs = [torch.rand(1, 784)]

    framework_model = MNISTLinear()
    fw_out = framework_model(*inputs)

    compiled_model = pybuda.compile(framework_model, sample_inputs=inputs)
    co_out = compiled_model(*[i.to("tt") for i in inputs])
    co_out = co_out[0].to("cpu")

    torch.allclose(fw_out, co_out) # This will do a proper failure. Should be replaced with compare_with_golden_pcc for valid verification. 

[dgolubovicTT]

I tried running this test but it fails due to data type error in ttnn: ttnn.linear: Tensor must be of type {DataType::BFLOAT16, DataType::BFLOAT8_B, DataType::BFLOAT4_B}, but got DataType::FLOAT32

[nvukobratTT]

I agree, can we update the rest of the test/mlir?

Also, please update pybuda/test/mlir/mnist/test_inference.py. At the moment, it has the similar issues as other ops. However, the current implementation doesn't detect it as [False] interpreted as valid assert case. In sum, code should look something like this:

def test_mnist_inference():
    inputs = [torch.rand(1, 784)]

    framework_model = MNISTLinear()
    fw_out = framework_model(*inputs)

    compiled_model = pybuda.compile(framework_model, sample_inputs=inputs)
    co_out = compiled_model(*[i.to("tt") for i in inputs])
    co_out = co_out[0].to("cpu")

    torch.allclose(fw_out, co_out) # This will do a proper failure. Should be replaced with compare_with_golden_pcc for valid verification.