| Age | Commit message (Collapse) | Author |
|---|
|
Close #6176.
If the struct has a `no_diff` member, it should not be its Differential
type. We miss this check.
|
|
* Make interface types non c-style.
* Make Optional<T> work with autodiff and existential types.
* Fix.
* patch behind slang 2026.
* Fix warnings.
* cleanup.
* Fix tests.
* Fix.
* Fix com interface lowering.
* Add comment to test.
* regenerate command line reference
* Add test for passing `none` to autodiff function.
* Fix recording of `getDynamicObjectRTTIBytes`.
* Fix nested Optional types.
---------
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
(#6696)
* Initial loop analysis pass
* More changes for a single-pass implication propagation
* Update slang-ir-autodiff-loop-analysis.cpp
* Cleanup + new system for loop analysis
* Fixup bugs in loop analysis
* Remove some relation types to simplify the analysis. Add test
* Remove unused
* Address comments
* Fix issue with continue loops
* Update reverse-loop-exit-value-inference-1.slang
* Update reverse-continue-loop.slang
|
|
# Make `IRWitnessTable` Hoistable
## Intention of the PR
This commit makes `IRWitnessTable` Hoistable so that we can avoid duplicated `IRWitnessTable`.
## Problems
This commit tries to address the following issues arise after turning `IRWitnessTable` into Hoistable:
1. A Hoistable instance is immutable.
2. When tries to create a duplicated child, you will get a previously created instance of `IRWitnessTable`, instead of a new one.
3. We don't actually want to hoist `IRWitnessTable`.
4. There can be only one instance of Hoistable and it cannot appear as childs multiple times.
5. Different import/export mangled names were used for the same Witness-table when its type is "enum" interface.
## Implementation
### Solution for "1. A Hoistable instance is immutable."
`IRWitnessTable::setConcreteType()` is removed, because when an `IRInst` is Hoistable, it is treated as immutable. Any `IRInst::setXXX()` methods don't work anymore.
There were two places calling `setConcreteType()` and their logic had to change little bit.
`DeclLoweringVisitor::visitInheritanceDecl()` in `source/slang/slang-lower-to-ir.cpp` was calling `setConcreteType()`. It had a little strange logic around `lowerType()`. The `IRWitnessTable` was added with `context->setGlobalValue()` first and its `concreteType` was changed later. This commit works around in a way that it sets the parent of `IRWitnessTable` temporarily and reset it with the correct `IRWitnessTable`. Without this logic, it went into an infinite recursion.
`AutoDiffPass::fillDifferentialTypeImplementation()` in `source/slang/slang-ir-autodiff.cpp` was calling `setConcreteType()`. It was changing the concreteType of `innerResult.diffWitness`. This commit creates a new `IRWitnessTable` and copies its `IRWitnessTableEntry`.
### Solution for "2. When tries to create a duplicated child, you will get a previously created instance of IRWitnessTable, instead of a new one"
After a call to `IRBuilder::createWitnessTable()`, this commit checks if the returned `IRWitnessTable` is a brand new or not. If it is not a new one, we have to avoid adding the decorations and children.
This commit decides when to add decorations and children based on whether `IRWitnessTable` has any of decorations or children already. It doesn't seem like a proper way to check. But when I tried, it was difficult to find a bottleneck point where the decorations and children are added to `IRWitnessTable` first time. Note that we are not trying to find when `IRWitnessTable` is created for the first time; we need to find if the decorations and children were added once.
It might be fine to have duplicated `IRWitnessTableEntry` in most of the cases, but I noticed that it fails an assertion check when `shouldDeepCloneWitnessTable()` returns false in `cloneWitnessTableImpl()`.
### Solution for "3. We don't actually want to hoist IRWitnessTable."
The reason why this commit makes `IRWitnessTable` is to prevent the duplicated instances of `IRInst`. But we don't really want to "Hoist" them.
When an `IRWitnessTable` gets Hoisted out, it causes unexpected problems and the specialization process fails due to the missing `IRWitnessTable` in the input.
This commit prevent from hoisting `IRWitnessTable` in `_replaceInstUsesWith()`. The way this is implemented feel little hack but we discussed on Slack and decided to go with this. One of the proper approaches could be to add a new flag in `IROpFlags` and have a new one like `kIROpFlag_Deduplicate`, which is different from just `kIROpFlag_Hoistable`.
### Solution for "4. There can be only one instance of Hoistable and it cannot appear as childs multiple times."
When `IRWitnessTable` is Hoistable, there can be only a unique set of instances. And we cannot have an instance as a duplicated childs. It is because `IRInst` has only one set of `IRInst* next` and `IRInst* prev`.
Before this commit, an instance of `IRGeneral` could have duplicated instances of `IRWitnessTable`. As an example, `IInteger` interface inherits two other interfaces, `IArithmetic` and `ILogical`. And they both inherits from `IComparable`.
```
interface IInteger : IArithmetic, ILogical {}
interface IArithmetic : IComparable {}
interface ILogical : IComparable
```
When we specialize it in `specializeGenericImpl()`, an `IRBlock` gets the following list of children:
- IRWitnessTable for IComparable,
- IRWitnessTable for IArithmetic,
- IRWitnessTable for IComparable,
- IRWitnessTable for ILogical,
For the cloning during the specialize, "IRWitnessTable for `IComparable`" must be cloned before the cloning of "IRWitnessTable for `IArithmetic`". Because "IRWitnessTable for `IArithmetic`" refers "IRWitnessTable for `IComparable`" as its `IRWitnessTableEntry`. The order they appear in the `IRBlock` as children decides which instances will be cloned first. And "IRWitnessTable for `IComparable`" must appear before "IRWitnessTable for `IArithmetic`".
Note that "IRWitnessTable for `IComparable`" appears twice, The first one was added for "IRWitnessTable for `IArithmetic`". And the second one is added for "IRWitnessTable for `ILogical`".
With this commit "IRWitnessTable for `IComparable`" can appear as a child only once in `IRBlock`. So it causes an error if it gets the following list:
- IRWitnessTable for IArithmetic,
- IRWitnessTable for IComparable,
- IRWitnessTable for ILogical,
In order to resolve the problem, "IRWitnessTable for `IComparable`" must appear before both "IRWitnessTable for `IArithmetic`" and "IRWitnessTable for `ILogical`" as following:
- IRWitnessTable for IComparable,
- IRWitnessTable for IArithmetic,
- IRWitnessTable for ILogical,
To address the problem, the instances of `IRWitnessTable` is always added to the end of the children list. If it is already added to the list, we don't move. This works out because the AST tree is built based on the dependencies.
### Solution for "5. Different import/export mangled names were used for the same Witness-table when its type is "enum" interface."
This issue was found while testing with Falcor tests where it uses Conformance-type feature of Slang.
We are using different import and export mangled names for a same Witness-table when the witness-table is for "Enum" interface.
The way we simplify the implementation of "Enum" causes a problem when it comes to generate export/import for the witness-table. And the exact repro step is still unclear.
There were two suggested solutions for the problem and this PR adopted the first option for now. Maybe we want to improve it with the second option later.
option 1, when we produce mangled names for those witness-table, we can use a mangled name with the underlying "int" type instead of the name of the enum type. In this way, all witness-tables for enum types whose underlying type is same will get the same mangled name. It will allow us to deduplicate the witness-table during the linking.
option 2, we can preserve type info for enum type when generating IR. We can still erase all other uses of the type info of enum types for now. But when we generate the witness-table, instead of filling the conforming type operand to IntType, we fill it as EnumType(IntType) where EnumType is a new global IROp code to represent all enum types (like InterfaceType/StructType). This way the operands for the two witness-tables will be different.
"option 1" is more quick and dirty and "option 2" is more proper way to address it.
I should go with "option 1" and improve it with "option 2" approach later.
|
|
Co-authored-by: Ellie Hermaszewska <ellieh@nvidia.com>
|
|
|
|
* Fix crash when swizzling non-differentiable types
* Update slang-ir-autodiff-fwd.cpp
|
|
* Fix lowering of associated types in generic interfaces.
* Update diff-assoctype-generic-interface.slang
* Fix-up lowering of differentiable witnesses for implicit ops
* Update slang-ir-autodiff-transcriber-base.cpp
* Fix issue with differentiating type-packs
|
|
* Fix a bug with hoisting 'IRVar' insts that are used outside the loop
- We introduce a 'CheckpointObject' inst and use that to split loop state insts into two pieces (one for within-loop uses and one for outside-loop uses.
- This allows the two kinds of uses to be handled separately by the hoisting mechanism
- CheckpointObject is then lowered to a no-op after hoisting is complete.
* Update slang-ir-autodiff-primal-hoist.cpp
* Update slang-ir-autodiff-primal-hoist.cpp
|
|
* Fix `UseGraph::isTrivial()` test.
* Fix.
* Fix.
* Refactor `UseGraph` and `UseChain`
* Update slang-ir-autodiff-primal-hoist.cpp
* Update all auto-diff locations that handle pointers to treat user pointers as regular values
* Update test to use direct-SPIRV only
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
Co-authored-by: Yong He <yonghe@outlook.com>
Co-authored-by: Ellie Hermaszewska <ellieh@nvidia.com>
|
|
* Fix DCE for calls to functions that have associations
* Update slang-ir-util.cpp
* Update slang-ir-util.cpp
|
|
* SP004: implement initialize list translation to ctor
- We synthesize a member-wise constructor for each struct follow
the rules described in SP004.
- Add logic to translate the initialize list to constructor invoke
- Add cuda-host decoration for the synthesized constructor
- Remove the default constructor when we have a valid member init constructor
- Disable -zero-initialize option, will re-implement it in followup (#6109).
- Fix the overload lookup issue
When creating invoke expression for ctor, we need to call
ResolveInvoke() to find us the best candidates, however
the existing lookup logic could find us the base constructor
for child struct, we should eliminate this case by providing
the LookupOptions::IgnoreInheritance to lookup, this requires
us to create a subcontext on SemanticsVisitor to indicate that
we only want to use this option on looking the constructor.
- Do not implicit initialize a struct that doesn't have explicit default
constructor.
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
|
|
|
|
|
|
|
|
|
|
differentiable functions (#5866)
* Overhauled the auto-diff system for dynamic dispatch
* More fixes
* remove intermediate dumps
* Update slang-ast-type.h
* More fixes + add a workaround for existential no-diff
* Update reverse-control-flow-3.slang
* remove dumps
* remove more dumps
* Delete working-reverse-control-flow-3.hlsl
* Cleanup comments + unused variables
* More comment cleanup
* Add support for lowering `DiffPairType(TypePack)` & `MakePair(MakeValuePack, MakeValuePack)`
* Fix array of issues in Falcor tests.
* Update slang-ir-autodiff-pairs.cpp
* More fixes for Falcor image tests
* Small fixups.
---------
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
* Correct IR generation for no-diff pointer type
Close #5805
There is an issue on checking whether a pointer type parameter
is no_diff, we should first check whether this parameter is
an Attribute type first, then check the data type.
In the back-propagate pass, for the pointer type parameter, we should
load this parameter to a temp variable, then pass it to the primal
function call. Otherwise, the temp variable will no be initialized,
which will cause the following calculation wrong.
|
|
|
|
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
(#5702)
This closes #5613.
Co-authored-by: Ellie Hermaszewska <ellieh@nvidia.com>
|
|
Closes https://github.com/shader-slang/slang/issues/5263
|
|
* Update Slang-RHI/slang
This brings in new fixes for WebGPU.
In particular, the "use_dxc" toggle is now used, which should enable these tests to run on
WebGPU, if f16 is otherwise supported:
- `tests/language-feature/generics/variadic-0.slang`
- `tests/language-feature/generics/tuple.slang`
This closes #5605.
* Disable tests/autodiff/float-cast.slang for wgpu
This test was previously not running for WebGPU because it required the 'half' render
feature, and Slang-RHI was previously not reporting it.
With the Slang-RHI update, the test now runs on WebGPU.
It now fails because the test is using 'double' which is just not supported on WebGPU.
Thus this commit disables the test.
|
|
can't be automatically inferred (#5630)
* [AD] Add support for resolving custom derivatives where generic parameters can't be automatically inferred
* Fix failing tests
* Update custom-derivative-generic.slang
|
|
Some tests are now passing and are enabled.
Other tests are still failing, but are given comments categorizing the failures.
Tests in the 'Not supported in WGSL' category are also removed from the expected failures
list. (Though they are still kept disabled for WebGPU, of course.)
This closes #5519.
|
|
* Add `IDifferentiablePtrType` support for arrays
- Also fixes an issue with spirv-emit of constructors that contain references to global params
* Fix GLSL legalization for arrays of resource types
|
|
* Transfer offset decorations when legalizing laid-out structs
Struct legalization removing fields not representable in memory should transfer all
decorations in case the struct has already had offsets calculated.
Closes #5264.
* format code
---------
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
Co-authored-by: Yong He <yonghe@outlook.com>
|
|
This closes issue #5505.
|
|
regression. (#5508)
* Fix IntVal unification logic to insert type casts.
* Fix regression.
|
|
* Fix WGSL parameter block binding.
* Re-enable tests.
* Update failure list.
* Fix entrypoint parameters.
* Update tests.
* Enable stat-var test.
|
|
* [WGSL] make sure switch has a default label.
* Various WGSL fixes.
* Update rhi submodule commit
* format code
* Remove unnecessary DISABLE_TEST directive on not applicable test.
* Matrix comp mul + `select`.
* Legalize binary ops for wgsl.
---------
Co-authored-by: slangbot <186143334+slangbot@users.noreply.github.com>
|
|
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>
|
