| Age | Commit message (Collapse) | Author |
|---|
|
auto-diff results (#5394)
* Various AD enhancements
* Fix issue with pt-loop test
* Update pt-loop.slang
* More fixes for perf. Final minimal context test now passes.
* Fix issue with loop-elimination pass not running after dce
* Try fix wgpu test by removing select operator
* Disable wgpu
* Delete out.wgsl
* Remove comments
* Update slang-ir-util.cpp
* Fix header relative paths for slang-embed
* Disbale wgpu for a few other tests
* Better way of determining which params to ignore for side-effects
* Update slang-ir-dce.cpp
* Fix issue with circular reference from previous AD pass being left behind for the next AD pass
* Update slang-ir-dce.cpp
|
|
|
|
* Initial work to support custom derivatives for member methods of differentiable types
* Support custom derivatives of member functions of differentiable types
- Also adds support for declaring custom derivatives via extensions.
* Fix
* move defs
* Update slang-check-decl.cpp
* Create diff-member-func-custom-derivative.slang.expected.txt
* Update slang-check-decl.cpp
* Fix for static custom derivatives
* Fix diagnostics for [PreferRecompute]
* Add backward custom derivative tests
|
|
* Fix diagnostics for [PreferRecompute]
* Update dont-warn-on-simple-prefer-recompute.slang
* Update slang-ir-autodiff.cpp
* Update dont-warn-on-simple-prefer-recompute.slang
* Update warn-on-prefer-recompute-side-effects.slang
|
|
|
|
* initial diff-ref-type interface
* Initial support for `IDifferentiablePtrType`
* Fix unused vars
* More tests + fix switch case fallthrough.
* Update slang-ir-autodiff.cpp
* Update diff-ptr-type-loop.slang
* Add optimization to allow more complex pair types
* Update slang-ir-autodiff-primal-hoist.cpp
* Update diff-ptr-type-loop.slang
* Update slang-ir-autodiff-primal-hoist.cpp
* More fixes to address reviews
* Update slang-check-expr.cpp
* Optimizations + rename `differentiableRefInterfaceType` -> `differentiablePtrInterfaceType`
* Move pair logic to ir-builder, unify the type dictionaries.
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
* Transferring source locations when creating phi instructions
* Tracking for simple variables
* Deriving source locations for loop counters
* Printing checkpoint structure breakdown
* More readable output format
* Special behavior for loop counters
* Writing report to file
* Add slangc option to enable checkpoint reports
* Display types of checkpointed fields
* Message in case there are no checkpointing contexts
* Catch source locations for function calls
* Source cleanup
* Fix compilation warnings
* Remove stray dump()
* Provide the report through diagnostic notes
* Add missing path for sourceLoc during unzip pass
* Add tests for reporting intermediates
* Include more transfer cases for source locations
* Fix ordering in address elimination
* Fill in more holes with source location transfer
* Remove debugging line
* Reverting changes to diagnostic sink
* Simplify address elimination using source location RAII contexts
* Eliminating manual source loc transfers in forward transcription
* Fix local var adaptation to use RAII location setter
* Simplify primal hoisting logic for source location transfer
* Simplify unzipping with RAII location scopes
* Simplify transpose logic
* Cleaning up for rev.cpp
* Reverting spacing changes
* Fix mistake with source loc RAII instantiation
* Fix formatting issues
|
|
side effects (#4707)
* Adds a warning for using prefer-recompute on methods that contain side effects
* Rename `SideEffects` -> `SideEffectBehavior`
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
differentiable args (#4901)
* Explicitly detach derivative when forming a non-differentiable struct out of differentiable args
This fixes an issue where initializer lists get optimized out and lose information about non-differentiability.
There are 2 places where this could have been fixed:
1. When coercing initializer-list exprs, we can check for non-differentiable aggregate types and use a detach derivative on all the args.
2. Add an extra case in the peephole optimization step that adds detach-derivative when simplifying a make-struct of a non-differentiable type.
Even though solution 2 is more elegant, this PR goes with solution 1 simply to avoid having to use a differentiable-type-conformance-context that is used in the auto-diff IR passes to check for differentiability.
* Change test name + add expected vals
|
|
Fixes: #4899
Fixes invalid test results since `{...}` was differentiating the constructor of `myvector` when it should not (see #4877). This change modifies the test so it is correct so other PRs may be merged if indirectly/directly fixing the old use-case for this test.
|
|
* Fix `tests\autodiff\reverse-while-loop-3.slang` test
Fixes: #4876
Fix `tests\autodiff\reverse-while-loop-3.slang` test to have correct listed output values
* set default to 1 (different val)
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
Addresses #4698 as one approach to diagnose the potential problem.
Emit warnings when a user marks a parameter as `inout` but never writes to it in the function. A new intrinsic function `unmodified(out T)` has been added to explicitly indicate that an `inout` variable will not be modified in the function.
This is only one way to address the specific validation error in #4698. In general it seems that DXC does some more extensive checks on actual struct fields (as opposed to observing arbitrary struct writes), so that will be the next step.
|
|
* Fix auto-diff synthesized method naming conventions
* Update tests; remove unused var
|
|
* Support parameter block in metal shader objects.
* Ingore parameter block tests on devices without tier2 argument buffer.
* Fix warning.
* Fix texture subscript test.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Handle out/inout functions with separate consideration
* Fixing bug with passing aliasable instructions
* Handle autodiff functions (fwd and rev) in warning system
* Handling interface methods
* Handling ref parameters like out/inout
* Temporary fix to remaining bugs
* Refactoring methods and tests
* Recursive check for empty structs
* Using default initializable interface in tests
* Resolving CI fail
|
|
generic types (#4602)
* Fix issue with synthesizing `Differential` type for self-differential generic types
The problem was that we were using the type that was performing the lookup for `.Differential` which can have substitutions based on the local context where the decl is being referenced.
We need to synthesize the type local to the decl itself
* Update auto-differential-type-generic.slang
|
|
`[DerivativeMember(<diff-member>)]` (#4525)
* Add diagnostic for missing diff-member associations
+ Automatically create diff member associations if differential type is the same as the primal type.
+ Move diff-member attribute checking to conformance-checking phase to avoid circularity issues.
Fixes #4103
* Update slang-check-decl.cpp
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
derivative data-flow (#4526)
Global parameters are by-default not differentiable (even if they are of a differentiable type), because our auto-diff passes do not touch anything outside of function bodies.
The solution is to use wrapper objects with differentiable getter/setter methods (and we should provide a few such objects in the stdlib).
Fixes: #3289
This is a potentially breaking change: User code that was previously working with global variables of a differentiable type will now throw an error (previously the gradient would be dropped without warning). The solution is to use `detach()` to keep same behavior as before or rewrite the access using differentiable getter/setter methods.
|
|
* Fix and enable tests for metal.
* Fix.
* Fix.
* Fix tests.
* Fix warnings.
* Fix.
---------
Co-authored-by: Yong He <yonghe@Yongs-Mac-mini.local>
|
|
non-CUDA/Torch targets (#4364)
* Remove `IRHLSLExportDecoration` and `IRKeepAliveDecoration` for non-CUDA/Torch targets
* Update hlsl-torch-cross-compile.slang
|
|
|
|
Handles a corner case where the first block after the condition on the true-side is another condition. This would currently result in an invalid reverse graph, where the reverse version of the true-block is the merge point for two different branching insts (the reverse version of the loop as well as the second condition).
This patch simply adds a blank block when constructing the reverse-loop (similar to critical edge breaking) so that each branch inst in the reversed loop has a unique merge block.
|
|
* Fix race-condition and visual artifacts issues
In PerformanceProfiler::getProfiler() we return a static object
for the profiler implementation, this is not thread-safe, so change
it to thead_local.
There is still some visual artifacts when using slang as the shading
language. We don't know the root cause yet, but found out it's related
to our loop inversion algorithm. So stage this feature for now, and turn
it into an internal option and default off. We will re-enable it after
more investigation on this optimization.
File an new issue 4151 to track it.
* Add '-loop-inversion' to the few tests
|
|
* Fix unzipping logic for inout non-diff parameters and adjust tests
+ Removed `-g0` from `struct-this-parameter.slang` test. Works correctly with the new unzipping logic.
+ Removed `-g0` from `was/warped-sampling-1d.slang` test. Works correctly with DX12 & CS_5_1. CS_5_0 appears to run into an FXC compiler bug with detecting infinite loops where there don't appear to be any.
* Update slang-ir-autodiff-unzip.h
* Update warped-sampling-1d.slang
|
|
* Avoid synthesis for when types can be used as their own differenial
+ Add test
* Add missing files..
* Fix issue with method synthesis for self-differential types
+ Add a generic test
* Fix
* Fix issue with out-of-date type resolution cache.
Witness tables created during the conformance checking phase not being taken into account during the decl type resolution phase because the epoch is not updated after conformance checking.
This leads to certain complex associated-type lookup chains (such as the one in tests/compute/assoctype-nested-lookup) not resolving properly and causing errors.
* Delete self-differential-type-synthesis-extension.slang
* Quick fix to repopulate stdlib cache for deferred stdlib loading
* Update slang-check-decl.cpp
|
|
* Fix compile failures when using debug symbol.
* Various fixes.
* Fix intrinsic.
* Fix test.
|
|
* Added diagnostics & built-in type lowering for `[CUDAKernel]` functions
This PR adds
- Diagnostics for non-void return from a cuda kernel entry point
- Diagnostics for using differentiable types in a differentiable cuda kernel entry point
- Logic for converting built-in types (float3, float3x3, etc..) to portable struct types and unpacks the parameter back into a built-in type on the CUDA side. This is because built-in types have different implementations in CUDA & CPP targets, which causes signature mis-match when linking.
* Fix error codes
* Add ability to lower structs and arrays that contain built-in types.
+ Added tests
+ Fix issue where the host-side was not marshalling data to lowered types.
* Update slang-ir-pytorch-cpp-binding.cpp
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
|
|
* Fix a bug in fwd-diff for cross product
* Also add a test for the reverse-mode AD
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
* Handle case where struct is created with `no_diff`-wrapped operands.
* Add test to reproduce crash upon initializing a struct with `no_diff`-wrapped operand type
* Add expected result for test
|
|
* Support visibility control and default to `internal`.
* Fix wip.
* Fixes.
* Fix.
* Fix test.
* Add legacy language detection and compatibility for existing code.
* Add doc.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* wip: clean up IArithmetic
* wip.
* Cleanup builtin arithmetic interfaces.
* Fix.
* Fixes.
* Fix.
* Fix.
* Fix.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
definition. (#3316)
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
(#3312)
* Fix ICE when lowering an associatedtype declref from an derived interface.
* Fixes.
* Fix test.
* Fix GLSL/SPIRV image subscript swizzle store regression.
* Fix.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Add more diagnostics on invalid custom derivative use.
* cleanup.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Fix crash when writing to `no_diff` out parameter.
* Fix.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
* Fix generic specialization bug.
* Update test.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
|
|
* Various fixes
* Remove unused parameter
* Update slang-ir-loop-unroll.cpp
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
* Support `constref` parameters passing.
* Fix.
* Fix.
* Add test and diagnostic on mix use of __constref and no_diff.
* check for [constref] on differentiable member method.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
unreachable. (#3241)
Also added a test for this.
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
|
|
* Update user-guide with new slangpy features
* More polishing of new slangpy docs
* Update a1-02-slangpy.md
* Only require contiguity for vector element types
* Added `loadOnce/storeOnce` and subscript operations
* Added docs, `DiffTensorView.dims()` & `DiffTensorView.stride(uint)`
* Add constructors, remove storeOnce/loadOnce test
* Adjusted intrinsic definitions
|
|
* Make dynamic cast transparent through `IRAttributedType`.
* Add [CUDAXxx] variant of attributes.
* Support marshaling of vector types.
* Wrap cuda kernels in `extern "C"` block.
---------
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
`SelectExpr` (#3219)
* Redirect destination-driven lowering logic to regular lowering for select expr
* Add test
|
|
exporting type information (#3209)
* Initial: add a DiffTensor impl
* Auto-binding and diff tensor implementations now work
* Refactored diff-tensor implementation + added py-export for struct types
* Cleanup
* Update slang-ir-pytorch-cpp-binding.cpp
* Updated test names
* Update autodiff-data-flow.slang.expected
* Add more versions of load/store & default generic args for DiffTensorView.
* Add diagnostic for default generic arg and more tests
* Add more `[AutoPyBind]` tests
|
|
predecessors (#3211)
* Add test case for while loop
* Fix loop inversion issue that caused ordinary blocks with multiple predecessors
The original version can leave the CFG in an invalid state with `e4` not being a break block or merge point, but having multiple predecessors.
The updated version creates a separate jump block for each break instruction to avoid this issue.
* Fixup tests
|
|
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
method. (#3192)
* Add a unit test
* Fix.
---------
Co-authored-by: Lifan Wu <lifanw@nvidia.com>
Co-authored-by: Yong He <yhe@nvidia.com>
|
|
unnecessary loop state (#3165)
* Extend the unit tests for MxLayeredMaterial
* Add breaking loop test
* Fix subtle corner-case with vars getting hoisted out of the loop creating unnecessary loop state
* remove whitespace changes
* Create loop-init.slang.expected.txt
* Add filecheck tests to ensure correct loop state
* Update comment
---------
Co-authored-by: winmad <winmad.wlf@gmail.com>
Co-authored-by: Yong He <yonghe@outlook.com>
|
