Fix dim order in op_permute #6432
Conversation
* hardtanh * permute Change-Id: Ia1802bdc37d365af382835b3c14174d841892927
Permute vector needs to be permuted when dim_order != (0, 1, 2, 3) Change-Id: I2a35c6852376f9a57deeedd4fc38bda870e453a4 Signed-off-by: Oscar Andersson <[email protected]> Signed-off-by: Erik Lundell <[email protected]>
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6432
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New FailureAs of commit b86ae33 with merge base 0aa802d ( NEW FAILURE - The following job has failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
Erik-Lundell
added
ciflow/trunk
partner: arm
| @@ -40,8 +88,33 @@ def define_node( | |||
| ) | |||
| return | |||
|
|
|||
| # The permutation vector describes a permutation P in default Pytorch dim_order. | |||
| # For rank 4, the default dim_order NCHW. | |||
| # E.g. (2,3,0,1) -> permute (n,c,h,w) to (w,c,n,h) | |||
There was a problem hiding this comment.
| # E.g. (2,3,0,1) -> permute (n,c,h,w) to (w,c,n,h) | |
| # E.g. (2,3,0,1) -> permute (n,c,h,w) to (h,w,n,c) |
There was a problem hiding this comment.
Your suggestion is correct, messed this up.
| P = permutation_vector_to_matrix(permutation_vector) | ||
|
|
||
| # The complete transformation is S * P * S_inverse. | ||
| transformation_matrix = S.matmul(P.matmul(S_inverse)) |
There was a problem hiding this comment.
This is very cool, and I think I get what it is doing, but can you help me understand the math here.
There was a problem hiding this comment.
Sure, the problem I am addressing is how a permutation vector expressed in nchw should be modified to work for nhwc. The key idea is to view the dim order as a coordinate system (the index (1,2,3,4) in nchw-coordinates is the same as (1,3,4,2) in nhwc) and to realize that both transformations between the systems and a permutation ops are linear operations that can be described by matrices.
For each nhwc-index in the incoming tensor, the full transformation of the permute op can then be described by:
- translate from nhwc to nchw
- do the permutation in nchw coordinates
- translate back to nhwc
Since all steps are permutation transformations, we can combine them into one single permutation transformation, which can then be expressed by a permutation vector - this is the vector we were looking for.
| ArmTester( | ||
| module, | ||
| example_inputs=test_data, | ||
| compile_spec=common.get_tosa_compile_spec(), |
There was a problem hiding this comment.
can we test permute to nhwc here at least for u85?
There was a problem hiding this comment.
Good point, right now we are only testing the case when we already are in nhwc for rank 4 tensors.
|
@digantdesai has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator. |
No description provided.