SPV: Some group invocation operations should be applied to scalars.

SPIRV/GlslangToSpv.cpp

@@ -149,11 +149,14 @@ protected:
     spv::Id createBinaryOperation(glslang::TOperator op, spv::Decoration precision, spv::Decoration noContraction, spv::Id typeId, spv::Id left, spv::Id right, glslang::TBasicType typeProxy, bool reduceComparison = true);
     spv::Id createBinaryMatrixOperation(spv::Op, spv::Decoration precision, spv::Decoration noContraction, spv::Id typeId, spv::Id left, spv::Id right);
     spv::Id createUnaryOperation(glslang::TOperator op, spv::Decoration precision, spv::Decoration noContraction, spv::Id typeId, spv::Id operand,glslang::TBasicType typeProxy);
-    spv::Id createUnaryMatrixOperation(spv::Op, spv::Decoration precision, spv::Decoration noContraction, spv::Id typeId, spv::Id operand,glslang::TBasicType typeProxy);
+    spv::Id createUnaryMatrixOperation(spv::Op op, spv::Decoration precision, spv::Decoration noContraction, spv::Id typeId, spv::Id operand,glslang::TBasicType typeProxy);
     spv::Id createConversion(glslang::TOperator op, spv::Decoration precision, spv::Decoration noContraction, spv::Id destTypeId, spv::Id operand, glslang::TBasicType typeProxy);
     spv::Id makeSmearedConstant(spv::Id constant, int vectorSize);
     spv::Id createAtomicOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
-    spv::Id createInvocationsOperation(glslang::TOperator, spv::Id typeId, spv::Id operand, glslang::TBasicType typeProxy);
+    spv::Id createInvocationsOperation(glslang::TOperator op, spv::Id typeId, spv::Id operand, glslang::TBasicType typeProxy);
+#ifdef AMD_EXTENSIONS
+    spv::Id CreateInvocationsVectorOperation(spv::Op op, spv::Id typeId, spv::Id operand);
+#endif
     spv::Id createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
     spv::Id createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId);
     spv::Id getSymbolId(const glslang::TIntermSymbol* node);
@@ -3940,7 +3943,10 @@ spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op
                 spvOp = spv::OpGroupIAdd;
         }
 
-        return builder.createOp(spvOp, typeId, operands);
+        if (builder.isVectorType(typeId))
+            return CreateInvocationsVectorOperation(spvOp, typeId, operand);
+        else
+            return builder.createOp(spvOp, typeId, operands);
     }
     case glslang::EOpMinInvocationsNonUniform:
     case glslang::EOpMaxInvocationsNonUniform:
@@ -3974,7 +3980,10 @@ spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op
                 spvOp = spv::OpGroupIAddNonUniformAMD;
         }
 
-        return builder.createOp(spvOp, typeId, operands);
+        if (builder.isVectorType(typeId))
+            return CreateInvocationsVectorOperation(spvOp, typeId, operand);
+        else
+            return builder.createOp(spvOp, typeId, operands);
     }
 #endif
     default:
@@ -3983,6 +3992,48 @@ spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op
     }
 }
 
+#ifdef AMD_EXTENSIONS
+// Create group invocation operations on a vector
+spv::Id TGlslangToSpvTraverser::CreateInvocationsVectorOperation(spv::Op op, spv::Id typeId, spv::Id operand)
+{
+    assert(op == spv::OpGroupFMin || op == spv::OpGroupUMin || op == spv::OpGroupSMin ||
+           op == spv::OpGroupFMax || op == spv::OpGroupUMax || op == spv::OpGroupSMax ||
+           op == spv::OpGroupFAdd || op == spv::OpGroupIAdd ||
+           op == spv::OpGroupFMinNonUniformAMD || op == spv::OpGroupUMinNonUniformAMD || op == spv::OpGroupSMinNonUniformAMD ||
+           op == spv::OpGroupFMaxNonUniformAMD || op == spv::OpGroupUMaxNonUniformAMD || op == spv::OpGroupSMaxNonUniformAMD ||
+           op == spv::OpGroupFAddNonUniformAMD || op == spv::OpGroupIAddNonUniformAMD);
+
+    // Handle group invocation operations scalar by scalar.
+    // The result type is the same type as the original type.
+    // The algorithm is to:
+    //   - break the vector into scalars
+    //   - apply the operation to each scalar
+    //   - make a vector out the scalar results
+
+    // get the types sorted out
+    int numComponents = builder.getNumComponents(operand);
+    spv::Id scalarType = builder.getScalarTypeId(builder.getTypeId(operand));
+    std::vector<spv::Id> results;
+
+    // do each scalar op
+    for (int comp = 0; comp < numComponents; ++comp) {
+        std::vector<unsigned int> indexes;
+        indexes.push_back(comp);
+        spv::Id scalar = builder.createCompositeExtract(operand, scalarType, indexes);
+
+        std::vector<spv::Id> operands;
+        operands.push_back(builder.makeUintConstant(spv::ScopeSubgroup));
+        operands.push_back(spv::GroupOperationReduce);
+        operands.push_back(scalar);
+
+        results.push_back(builder.createOp(op, scalarType, operands));
+    }
+
+    // put the pieces together
+    return builder.createCompositeConstruct(typeId, results);
+}
+#endif
+
 spv::Id TGlslangToSpvTraverser::createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy)
 {
     bool isUnsigned = typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64;