Support SV_DispatchGrid semantic in a nested record #6931
Conversation
The SV_DispatchGrid DXIL metadata for a node input record was not generated in cases where: - the field with the SV_DispatchGrid semantic was in a nested record - the field with the SV_DispatchGrid semantic was in a record field - the field with the SV_DispatchGrid semantic was inherited from a base record - in any combinations of the above Added FindDispatchGridSemantic() to be used by the AddHLSLNodeRecordTypeInfo() function, and added a test case. Fixes microsoft#6928
There was a problem hiding this comment.
We should also make sure we have error detection in Sema for the case of multiple fields in a single or nested structure definition having SV_DispatchGrid applied. Since the CodeGen can only handle a single declaration, we should error if a structure has more than one annotated field.
tools/clang/lib/CodeGen/CGHLSLMS.cpp
Outdated
| // Collect any non-virtual bases. | ||
| SmallVector<const CXXRecordDecl *, 4> Bases; | ||
| for (const CXXBaseSpecifier &Base : RD->bases()) { | ||
| if (!Base.isVirtual() && !Base.getType()->isDependentType()) |
There was a problem hiding this comment.
I assume it isn't actually possible in HLSL to have virtual base classes right? The language certainly shouldn't allow them, so I think this check is probably unnecessary.
There was a problem hiding this comment.
No, virtual base classes aren't possible - but the dependent type is, so we do need that check here. This condition just follows the style and layout of similar checks elsewhere - the isVirtual() check could safely be removed.
There was a problem hiding this comment.
Yea, I think we should remove the isVirtual check since that will always be false and I can't imagine any world where we'd be adding virtual inheritance to HLSL...
|
The presence of more than one SV_DispatchGrid field is already detected in Sema, and covers the cases of base classes and nested records. |
Remove the check for a virtual base class from the code in FindDispatchGridSemantic() as virtual classes can't appear in HLSL code.
| [NumThreads(32,1,1)] | ||
| void node5(DispatchNodeInputRecord<Record5> input) {} | ||
| // CHECK: , i32 1, ![[SVDG_5:[0-9]+]] | ||
| // CHECK: [[SVDG_5]] = !{i32 20, i32 5, i32 2} |
There was a problem hiding this comment.
What about some test cases with templates?
Something like:
template <typename T>
struct Base {
T DG : SV_DispatchGrid;
};
struct Derived1 : Base<uint3> {
int4 x;
};
[Shader("node")]
[NodeLaunch("broadcasting")]
[NodeMaxDispatchGrid(32,16,1)]
[NumThreads(32,1,1)]
void node6(DispatchNodeInputRecord<Derived1 > input) {}
template <typename T>
struct Derived2 : Base<T> {
T Y;
};
[Shader("node")]
[NodeLaunch("broadcasting")]
[NodeMaxDispatchGrid(32,16,1)]
[NumThreads(32,1,1)]
void node7(DispatchNodeInputRecord<Derived2<uint2> > input) {}
template <typename T>
struct Derived3 {
Derived2<T> V;
};
[Shader("node")]
[NodeLaunch("broadcasting")]
[NodeMaxDispatchGrid(32,16,1)]
[NumThreads(32,1,1)]
void node8(DispatchNodeInputRecord< Derived3 <uint3> > input) {}There was a problem hiding this comment.
Great idea! I've updated the test to include these cases.
Added test cases to cover nested SV_DispatchGrid used in records using templates.
The SV_DispatchGrid DXIL metadata for a node input record was not generated in cases where:
Added FindDispatchGridSemantic() to be used by the AddHLSLNodeRecordTypeInfo() function, and added a test case.
Fixes #6928