Gaurav Shukla

Add decomposition for `aten.native_batch_norm_backward` op

10

This commit adds decomposition for the `aten.native_batch_norm_backward` op. Signed-Off-by: Gaurav Shukla

[TORCH] Add decomposition of `aten.linear` op

9

This commit adds decomposition of `aten.linear` op. TODO: remove existing implementation and add a simple test case. Also handle the (3D,2D) case of matmul. Signed-Off-by: Gaurav Shukla

static shapes propagation: add `torch-shape-refinement-pipeline` after canonicalization

2

We can simplify the IR with the help of canonicalization/folding and the result of those simplifications can be used to compute few dynamic shapes at the compile time. For example:...

[TORCH] Fold `aten.contiguous` op

2

This commit adds folding of `aten.contiguous` op. Since we do not take memory format into account currently, the `aten.contiguous` op can be folded away. Signed-Off-by: Gaurav Shukla

Lowering `torch.scatter` op

3

This issue is created in order to figure out the best possible approach to lower `torch.scatter` op. https://pytorch.org/docs/stable/generated/torch.Tensor.scatter_.html#torch.Tensor.scatter_

[TORCH] Add canonicalization of `aten.transpose.int` op

2

This commit adds a canonicalization of `aten.tranpose.int` op when its operand is `aten.permute` op. `aten.permute` + `aten.transpose.int` -> `aten.permute` Signed-Off-by: Gaurav Shukla

bf16 result mismatch for Conv2D op

30

1. Following is the Conv2d pytorch module. ``` import torch import torch.nn as nn class op_conv2d(nn.Module): def __init__(self): super().__init__() self.layers = nn.Sequential( nn.Conv2d(8, 10, (3, 5), stride=(2, 1), padding=(4, 2),...

[MLIR][ONNX] Add support for onnx.scan op

1

This commit lowers onnx.scan op to torch.prim.Loop op and adds the lowering in the onnx pipeline.

This commit adds support for onnx.ScatterND op in the onnx pipeline.