implement transpose operator for MKLDNN #19955
Conversation
pytorchbot
added
caffe2
module: pybind
|
@bddppq We have upgrade the ideep transpose so handle more than just 4d mkl-dnn tensor. We were not sure where we should use it to support torch.transpose. torch.transpose(tensor, dim0, dim1) doesn't change the data buffer layout until contiguous() is called. But mkl-dnn tensor is not supposed to be contiguous tensor. So this PR only contain the Caffe2 transpose op for your reference. Do you have any suggestion how it can be used to speedup Pytorch op? |
|
rebased. pls help review. |
Matrix multiplication with bias $Y = A * B + Bias$, where `A` has size (M x K) or (Batch_size x M x k), `B` has size (K x N), `Bias` has size (n), and `Y` will have a size (M x N) or (Batch_size x M x N). To transpose `A` or `B` before multiplication, pass 1 to the `trans_a` and/or `trans_b` arguments, which separate the first and second dimensions of the respective matrices using `axis_a` and `axis_b`. This patch depend on pytorch#19955. Change-Id: Iea33d609f814a062d4e2b1b48a96e8a1da9f1072
Signed-off-by: Gu, Jinghui <jinghui.gu@intel.com>
VitalyFedyunin
added
the
module: mkldnn
remove extra parenthesis in transpose_op_test.py
|
@pytorchbot rebase this please |
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@ezyang is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
facebook-github-bot
left a comment
facebook-github-bot
left a comment
There was a problem hiding this comment.
@bddppq has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.
|
Could you add transpose aten op in a follow up diff? :-) |
|
Sure. We will work on it. |
8 participants
implement transpose operator for MKLDNN