Optimizer: optimize transposes in variety of circumstances #3509
Conversation
|
now removes nop transposes, consecutive transposes that cancel out, and transposes-into-Gemm (which can just be a parameter to Gemm) |
torch/csrc/jit/passes/peephole.cpp
Outdated
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
|
Thanks, this looks great! I wrote one minor comment about generalizing transpose/permutation fusion, but everything looks fine. Before we merge, we're going to need a test. Do you know how to operate the accept test machinery? |
|
@pytorchbot test this please |
|
generalized logic, cleaned up code and comments, and corrected several bugs related to mutating shared references and to resource deallocation. |
f591005 to
b4a776b
Compare
houseroad
left a comment
There was a problem hiding this comment.
Thanks, this is great!
BTW, what happens to our CI?
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
|
Looks great! As for the tests - it seems that the ordering of ops in backward pass is not deterministic (https://travis-ci.org/pytorch/pytorch/jobs/299841429). It's not related to this diff, but @zdevito and @ezyang might be interested in figuring out a better way to unittest graphs. |
This comment was marked as off-topic.
This comment was marked as off-topic.
Sorry, something went wrong.
6435199 to
b60bb9e
Compare
|
Are we going to land it (I guess the only missing part is tests) or do you plan to put it in the separate transforms library? |
|
I would say no reason not to land it - and also some tests are here ezyang/onnx-pytorch#40 |
- No-op transposes - Consecutive transposes (fuse them) - Transposes into Gemm (fuse them into transA/transB parameter)
|
@pytorchbot test this please |
* Optimizer: optimize transposes in variety of circumstances (#3509) * Optimizer: Optimize transposes in variety of circumstances - No-op transposes - Consecutive transposes (fuse them) - Transposes into Gemm (fuse them into transA/transB parameter) * touch up out of date comment * Backporting optimizer changes
* Optimizer: optimize transposes in variety of circumstances (#3509) * Optimizer: Optimize transposes in variety of circumstances - No-op transposes - Consecutive transposes (fuse them) - Transposes into Gemm (fuse them into transA/transB parameter) * touch up out of date comment * Backporting optimizer changes
* Optimizer: optimize transposes in variety of circumstances (pytorch#3509) * Optimizer: Optimize transposes in variety of circumstances - No-op transposes - Consecutive transposes (fuse them) - Transposes into Gemm (fuse them into transA/transB parameter) * touch up out of date comment * Backporting optimizer changes
* Optimizer: optimize transposes in variety of circumstances (pytorch#3509) * Optimizer: Optimize transposes in variety of circumstances - No-op transposes - Consecutive transposes (fuse them) - Transposes into Gemm (fuse them into transA/transB parameter) * touch up out of date comment * Backporting optimizer changes
This is a first pass to get feedback