#0: Port Matmul, Conv, and AllGather to compute_ouput_specs #15978
Conversation
sminakov-tt
requested review from
TT-BrianLiu,
bbradelTT,
yugaoTT,
asandhupatlaTT,
mywoodstock,
sankarmanoj-tt,
pavlejosipovic,
shwetankTT,
SeanNijjar,
cfjchu,
ayerofieiev-tt and
dmakoviichuk-tt
as code owners
| return {create_device_tensor(output_shape, output_dtype, Layout::ROW_MAJOR, input_tensor.device(), out_mem_config)}; | ||
| return {TensorSpec( | ||
| output_shape.logical_shape(), | ||
| TensorLayout::fromLegacyPaddedShape( |
There was a problem hiding this comment.
ouch.. means we will have to make another pass to get rid of this constructor
There was a problem hiding this comment.
Yes, getting rid of fromLegacyPaddedShape is not trivial and will require help from op writers. It's a part of the planned refactoring to remove all Shape/LegacyShape from ops
| } // namespace detail | ||
|
|
There was a problem hiding this comment.
Reason for moving it out of detail namespae?
There was a problem hiding this comment.
I moved default_create_output_tensors out of detail namespace, so I can use it in matmul.
| auto output_tensor_shape = ttnn::Shape(tt::tt_metal::LegacyShape({1, 1, padded_shape_w, padded_shape_c}, output_padding)); | ||
| return {output_tensor_shape.value}; | ||
| } | ||
| auto output_shape = tt::tt_metal::LegacyShape({1, 1, padded_shape_w, padded_shape_c}, output_padding); |
There was a problem hiding this comment.
removal of the use of LegacyShape is on the OP owners?
There was a problem hiding this comment.
Yes. Removal of Shape/LegacyShape within each individual OP is on the OP's owners. It may involve modifying the operation to avoid unnecessary paddings.
Ticket
Problem description
We're continuing to port ops from compute_output_shapes to the new compute_output_specs
What's changed
Ported matmul, all_gather, all_gather_matmul, conv2d, and halo ops to use compute_output_specs
Checklist