Support SV_DispatchGrid semantic in a nested record #6931

tcorringham · 2024-09-24T15:31:44Z

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.

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

llvm-beanz

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.

llvm-beanz · 2024-12-02T20:22:44Z

tools/clang/lib/CodeGen/CGHLSLMS.cpp

+  // Collect any non-virtual bases.
+  SmallVector<const CXXRecordDecl *, 4> Bases;
+  for (const CXXBaseSpecifier &Base : RD->bases()) {
+    if (!Base.isVirtual() && !Base.getType()->isDependentType())


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.

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.

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...

tcorringham · 2024-12-03T11:23:21Z

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.

llvm-beanz · 2024-12-11T22:22:16Z

tools/clang/test/HLSLFileCheck/hlsl/workgraph/nested_sv_dispatchgrid.hlsl

+[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}


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) {}

Great idea! I've updated the test to include these cases.

Added test cases to cover nested SV_DispatchGrid used in records using templates.

damyanp · 2025-01-30T22:12:07Z

@llvm-beanz - could you have a look at this with some urgency please (we want it in the upcoming release) when you're back?

pow2clk

Looks good! I think the parts I'm most concerned about are the repeated checks of parent classes and that I might not understand the motive for sorting base offsets. The other stuff is minor.

pow2clk · 2025-01-30T21:56:56Z

tools/clang/lib/CodeGen/CGHLSLMS.cpp

+        return true;
+    }
+    // Otherwise check this field for the SV_DispatchGrid semantic annotation
+    for (const hlsl::UnusualAnnotation *it : Field->getUnusualAnnotations()) {


could/should this be an else case? I don't think we can apply SV_DispatchGrid to a record?

According to the way semantics normally work, you could do this:

struct MyVec3 { uint3 value; }; struct Record { MyVec3 DispatchGrid : SV_DispatchGrid; };

Other code here won't handle this though.

Currently the checks that are implemented in Sema will error if the SV_DispatchGrid is applied to a record as in the example above.
So, either there could be an else or Sema (and code here) is wrong.
If SV_DispatchGrid can be applied to a record then we would have to check that the first/only(?) field is a suitable type, and deal with nested records/inheritance - which sounds fairly horrible.

pow2clk · 2025-01-30T22:26:31Z

tools/clang/lib/CodeGen/CGHLSLMS.cpp

+          SDGRec.NumComponents = 1;
+          SDGRec.ByteOffset = (unsigned)FieldOffset.getQuantity();
+          if (const llvm::VectorType *VT = dyn_cast<llvm::VectorType>(FTy)) {
+            SDGRec.NumComponents = VT->getNumElements();


These used to assert it was an integer. was that invalid?

The types are checked in Sema so we won't get this far if the type is wrong.

pow2clk · 2025-01-30T22:42:22Z

tools/clang/lib/CodeGen/CGHLSLMS.cpp

+          const llvm::Type *FTy = CGM.getTypes().ConvertType(Field->getType());
+          const llvm::Type *ElTy = FTy;
+          SDGRec.NumComponents = 1;
+          SDGRec.ByteOffset = (unsigned)FieldOffset.getQuantity();


The removal of the struct check is because of the above that iterates into any records?

Yes, we iterate into records above.

pow2clk · 2025-01-30T22:47:29Z

tools/clang/lib/CodeGen/CGHLSLMS.cpp

+                   [&](const CXXRecordDecl *L, const CXXRecordDecl *R) {
+                     return Layout.getBaseClassOffset(L) <
+                            Layout.getBaseClassOffset(R);
+                   });


What does ordering the bases give us? I could imagine this might be intended to prevent duplicate testing, but If there are multiple levels of bases, it seems this results in checking the highest parent base once for each descendent.

We need to order the bases so we can keep track of the offset from the start of the input record.
EDIT: I have only now realized there can't be more than one concrete base in HLSL, so the sort step is not necessary after all.

If a record type appears more than once then it can't contain a field with SV_DispatchGrid as that would be errored in Sema, so we could track which records we have already checked and skip checking the fields of any we have seen before. However, I'm not sure that such cases would occur often enough to make the extra housekeeping worthwhile - so currently we may check records more than once.

pow2clk · 2025-01-30T22:50:00Z

tools/clang/lib/CodeGen/CGHLSLMS.cpp

+
+  // Collect any bases.
+  SmallVector<const CXXRecordDecl *, 4> Bases;
+  for (const CXXBaseSpecifier &Base : RD->bases()) {


If I understand this right, parent classes will be checked for every descendent and every struct that has a common ancestor. Maybe that's hard to avoid?

Yes, as per my comment above I don't expect such cases to occur often, so I believe the extra housekeeping probably isn't worthwhile.

tools/clang/lib/CodeGen/CGHLSLMS.cpp

tex3d · 2025-02-01T01:22:53Z

tools/clang/lib/CodeGen/CGHLSLMS.cpp

+        return true;
+    }
+    // Otherwise check this field for the SV_DispatchGrid semantic annotation
+    for (const hlsl::UnusualAnnotation *it : Field->getUnusualAnnotations()) {


According to the way semantics normally work, you could do this:

struct MyVec3 { uint3 value; }; struct Record { MyVec3 DispatchGrid : SV_DispatchGrid; };

Other code here won't handle this though.

Adopting code suggested in review - there can only be one concrete base so the sort step isn't required. Co-authored-by: Tex Riddell <[email protected]>

tcorringham requested review from tex3d and llvm-beanz September 24, 2024 15:31

tcorringham self-assigned this Sep 24, 2024

tcorringham requested a review from a team as a code owner September 24, 2024 15:31

damyanp assigned tex3d and unassigned tcorringham Sep 26, 2024

llvm-beanz reviewed Dec 2, 2024

View reviewed changes

Address review comments

4c74ce1

Remove the check for a virtual base class from the code in FindDispatchGridSemantic() as virtual classes can't appear in HLSL code.

llvm-beanz reviewed Dec 11, 2024

View reviewed changes

Add template test cases for nested SV_DispatchGrid

a2df420

Added test cases to cover nested SV_DispatchGrid used in records using templates.

damyanp requested a review from llvm-beanz January 8, 2025 18:43

damyanp requested review from pow2clk and bob80905 January 30, 2025 19:29

pow2clk reviewed Jan 30, 2025

View reviewed changes

damyanp added this to the Release 1.8.2502 milestone Jan 31, 2025

damyanp assigned llvm-beanz and pow2clk Jan 31, 2025

tex3d reviewed Feb 1, 2025

View reviewed changes

damyanp modified the milestones: Release 1.8.2502, Next+1 Release Feb 5, 2025

Update tools/clang/lib/CodeGen/CGHLSLMS.cpp

c45f825

Adopting code suggested in review - there can only be one concrete base so the sort step isn't required. Co-authored-by: Tex Riddell <[email protected]>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support SV_DispatchGrid semantic in a nested record #6931

Support SV_DispatchGrid semantic in a nested record #6931

tcorringham commented Sep 24, 2024

llvm-beanz left a comment

llvm-beanz Dec 2, 2024

tcorringham Dec 3, 2024

llvm-beanz Dec 3, 2024

tcorringham Dec 12, 2024

tcorringham commented Dec 3, 2024

llvm-beanz Dec 11, 2024

tcorringham Dec 12, 2024

damyanp commented Jan 30, 2025

pow2clk left a comment

pow2clk Jan 30, 2025

tex3d Feb 1, 2025

tcorringham Feb 6, 2025

pow2clk Jan 30, 2025

tcorringham Feb 6, 2025

pow2clk Jan 30, 2025

tcorringham Feb 6, 2025

pow2clk Jan 30, 2025

tcorringham Feb 6, 2025 •

edited

Loading

pow2clk Jan 30, 2025

tcorringham Feb 6, 2025

tex3d Feb 1, 2025

Support SV_DispatchGrid semantic in a nested record #6931

Are you sure you want to change the base?

Support SV_DispatchGrid semantic in a nested record #6931

Conversation

tcorringham commented Sep 24, 2024

llvm-beanz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tcorringham commented Dec 3, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

damyanp commented Jan 30, 2025

pow2clk left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tcorringham Feb 6, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tcorringham Feb 6, 2025 •

edited

Loading