Sayan Saha

[BugFix] : Update padding computation for TFL->TOSA for transpose_conv.

11

This change fixes the padding computation for translating `tfl.transpose_conv` to `tosa.transpose_conv2d` as per the padding definition change for the padding attribute in TOSA spec in https://git.mlplatform.org/tosa/specification.git/commit/?id=eda7b126d3914e9461cf014439b3571b9e6a9c41 For more info, please...

[Tosa] : Add support for negative indices in index.tensor and index.Tensor_hacked_twin for TorchToTosa lowering.

5

1. Negative indices for tensor indexing is handled by wrapping around the index values by checking their values at run time. Without the fix, there was a runtime error. 2....

Support convolution with `valid` padding.

1

Convolution created with `valid` padding produces the `aten.convolution` op in the following fashion: ``` module { func.func @main(%arg0: !torch.vtensor) -> !torch.vtensor attributes {torch.assume_strict_symbolic_shapes} { %false = torch.constant.bool false %int1 =...

[TorchToTosa] Refactoring to separate construction of legal/illegal ops and conversion patterns.

3

This PR refactors TorchToTosa to separate the construction of legal/illegal ops and conversion patterns in their own functions: 1. populateTorchToTosaConversionLegalOps -- populate any ops that are legal after the conversion...

Assertion when lowering from Torch IR for `AvgPool2d` when `kernel` is a tuple of a single int

14

ExportedProgram for ``` class AvgPool2dFloatModule(torch.nn.Module): def __init__(self): super().__init__() self.ap2d = torch.nn.AvgPool2d( kernel_size=6, ) def forward(self, x): return self.ap2d(x) ``` produces the call to `AvgPool2d` as `torch.ops.aten.avg_pool2d.default(x, [6, 6], [6, 6])`....

There are two types of failures when running the python based tests on Windows: 1. End-to-end execution tests are [failing](https://github.com/sahas3/torch-mlir/actions/runs/13413554894/job/37468876174#step:10:75) due to the `signal.SIGALRM` (used for timeout handling in the...

Wrap indices to be within max range for as_strided + index_select

1

Found these issues when trying to lower the ExportedProgram generated from decomposing `nn.GRU` op in torch. - Added a new e2e test representing the root cause which involves a special...

[tosa] : Use QuantizedType signed info for legalization.

3

This is a follow-up to https://github.com/llvm/llvm-project/pull/169387 to use the signed info from quantized type correctly.

[linalg] : Use (-)realmax instead of (-)inf to avoid usage of non-finites.

4

This change replaces usage of non-finite value `inf` with finite value `realmax` for init value of various max/min operations -- no change in semantics of the ops.