[Relay][Topi][AutoTVM] Winograd support for Conv3D #5186
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…eight_transform operator.
|
@merrymercy @icemelon9 @comaniac @masahi this PR is likely most relevant to you guys, can you take a look and let me know what you think? |
FrozenGene
reviewed
Apr 1, 2020
icemelon
requested changes
Apr 3, 2020
masahi
reviewed
Apr 4, 2020
masahi
reviewed
Apr 4, 2020
masahi
reviewed
Apr 4, 2020
masahi
reviewed
Apr 4, 2020
masahi
approved these changes
Apr 5, 2020
icemelon
approved these changes
Apr 5, 2020
|
Thanks @jwfromm @masahi @FrozenGene. This is now merged. |
trevor-m
pushed a commit
to trevor-m/tvm
that referenced
this pull request
Apr 16, 2020
* Functional conv3d winograd working. * Formatted python code. * registered conv3d winograd compute and started adding relay without_weight_transform operator. * Add topi testing for conv3d winograd. * Format file. * small tweak to unrolling to prevent build sticking. * Refactoring convolution ops in relay. * Refactored relay convolutions. * Bug fixes. * Fixed static bug in convolution. * Added conv3d alter op layout and related support. * Bug fixes and testing done. * Fix a few autotvm bugs. * Drop silly debug print. * Removed debug_skip_region. * Add variant of conv3d_winograd that doesn't transform depth. * initial infrastructure done for depthless conv. * Fix no_depth schedule bugs. * automatic topi switching between depth and depthless winograd. * Fixed bug in schedule. * lint fixes. * Removed indents in convolution.cc * missed a few indents oops. * fixed flop count. * One more small tweak. * Change kernel pack inner axes order. * Style changes. * Comment fixes.
zhiics
pushed a commit
to neo-ai/tvm
that referenced
this pull request
Apr 17, 2020
* Functional conv3d winograd working. * Formatted python code. * registered conv3d winograd compute and started adding relay without_weight_transform operator. * Add topi testing for conv3d winograd. * Format file. * small tweak to unrolling to prevent build sticking. * Refactoring convolution ops in relay. * Refactored relay convolutions. * Bug fixes. * Fixed static bug in convolution. * Added conv3d alter op layout and related support. * Bug fixes and testing done. * Fix a few autotvm bugs. * Drop silly debug print. * Removed debug_skip_region. * Add variant of conv3d_winograd that doesn't transform depth. * initial infrastructure done for depthless conv. * Fix no_depth schedule bugs. * automatic topi switching between depth and depthless winograd. * Fixed bug in schedule. * lint fixes. * Removed indents in convolution.cc * missed a few indents oops. * fixed flop count. * One more small tweak. * Change kernel pack inner axes order. * Style changes. * Comment fixes.
dpankratz
pushed a commit
to dpankratz/incubator-tvm
that referenced
this pull request
Apr 24, 2020
* Functional conv3d winograd working. * Formatted python code. * registered conv3d winograd compute and started adding relay without_weight_transform operator. * Add topi testing for conv3d winograd. * Format file. * small tweak to unrolling to prevent build sticking. * Refactoring convolution ops in relay. * Refactored relay convolutions. * Bug fixes. * Fixed static bug in convolution. * Added conv3d alter op layout and related support. * Bug fixes and testing done. * Fix a few autotvm bugs. * Drop silly debug print. * Removed debug_skip_region. * Add variant of conv3d_winograd that doesn't transform depth. * initial infrastructure done for depthless conv. * Fix no_depth schedule bugs. * automatic topi switching between depth and depthless winograd. * Fixed bug in schedule. * lint fixes. * Removed indents in convolution.cc * missed a few indents oops. * fixed flop count. * One more small tweak. * Change kernel pack inner axes order. * Style changes. * Comment fixes.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR adds winograd functions and schedules to Conv3D in Topi. I also add an alter op layout pass to pretransform weights along with the associated relay functions. While working writing the relay operators, I noticed quite a bit of inconsistency across various convolution operators and decided to organize them and reduce code duplication. Now, make functions for all convolution variants are shared where possible so that each ND variant doesn't need its own. I also moved all type relationship templates to
convolution.hrather than have them split across two files.Note that unlike conv2d, it is quite common for conv3d to have non-uniform kernel sizes (8x3x3) for example. Thus, this PR includes two variations of conv3d_winograd, one for NxNxN kernels that transforms all axes and one for kernels where KD!=KH,KW where only height and width are transformed.
I've found that for some workloads this new winograd function can be up to 4X faster than direct convolution.