| Age | Commit message (Collapse) | Author |
|
* Overhaul `transposeParameterBlock` to support `inout` params.
* Small bug fixes.
* Bug fix on differentiable intrinsic specialization.
* Fixes.
* Run autodiff tests on CPU.
* Clean up.
* More bug fixes.,
* Add test coverage on inout param.
* Fix language server hinting for transcribed mutable params.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Support `out` parameters in backward differentiation.
* Fixes.
* Fix cleanup.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
(#2617)
|
|
* Add transposition logic for constructor opcodes.
* Fix.
* Add language server regression test.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Added switch-case support; fixed non-diff parameter transposition
* Made region propagation much more robust. Partial loop unzip implementation
* WIP: Added most loop handling code, and a test. Still untested
* Added CFG Normalization pass + CFG Reversal Pass + Loop Unzipping + most loop transcription
* Add single-iter-loop test.
* proj files
* removed comments
* Update reverse-loop.slang
* Removed out-of-date code
* Disabled IR validation during constructSSA phase of normalizeCFG. constructSSA now reuses sharedBuilder
* Moved normalizeCFG() call to prepareFuncForBackwardDiff()
|
|
* Unify UpdateField and UpdateElement with access chain.
* Fix warnings.
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Reimplement address elimination pass.
* Fix error.
* Update test references.
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
|
|
|
|
|
|
* Make backward differentiation work with generics.
* Fix.
* Another fix.
* More fix.
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Split bwd_diff op into separate ops for primal and propagate func.
* Fix.
* Download swiftshader with github actions instead of curl on linux.
* Fix github action.
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Initial multi-block implementation
* Implemented multi-block reverse-mode (without loops)
* Added logic to remove block-level decorations to avoid confusing IR simplification passes
* Fixed issues with block-level decorations during IR simplification by removing them prior to simplification.
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
* Further unify the autodiff passes.
* Fix clang compilation error.
* Rename ForwardDerivativeTranscriber->ForwardDiffTranscriber.
* Remove unused fields from Transcriber classes.
* More small cleanups.
* Cleanup.
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
function. (#2569)
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Added initial support for nested calls
* removed comments
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
* Add vector arithmetic test. Make gradient accumulation work for any IRLoad
* Added support for general vector types, and split transposition into transpose & materialize to allow emitting the fully accumulated gradient for complex types.
* Several bug fixes + finished up support for vector & struct types + removed prop pass
* minor fixes (int/uint casts)
* Removed IRConstruct
* Added some type casts to prevent warnings
* minor fix for unused variable
|
|
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
|
|
* Added partial implementation for reverse-mode
* Fixing several compile and runtime errors.
* Fixed several issues with reverse-mode passes.
* Fixed more issues. Basic reverse-mode tests passing
Co-authored-by: Edward Liu <shiqiu1105@gmail.com>
|
|
* Initial refactor
* Refactor passes tests
* Removed Differential Bottom references from the IR side
|