Add `-fvk-use-dx-position-w` and fix ExecutionMode ordering for geometry shaders. (#3731)

* Add `-fvk-use-dx-position-w`.

* Fix ordering of OutputVertices and output primitive type decoration in spirv.

* Fix.

* fix

* Fix.

* Move test around.

15 files changed, 122 insertions, 29 deletions

diff --git a/docs/user-guide/08-compiling.md b/docs/user-guide/08-compiling.md
index fb9d8de6c..ba772d620 100644
--- a/docs/user-guide/08-compiling.md
+++ b/docs/user-guide/08-compiling.md
@@ -660,6 +660,7 @@ meanings of their `CompilerOptionValue` encodings.
 | VulkanBindShift | Specifies the `-fvk-bind-shift` option. `intValue0` (higher 8 bits): kind, `intValue0` (lower bits): set; `intValue1`: shift. |
 | VulkanBindGlobals | Specifies the `-fvk-bind-globals` option. `intValue0`: index, `intValue`: set. |
 | VulkanInvertY | Specifies the `-fvk-invert-y` option. `intValue0` specifies a bool value for the setting. |
+| VulkanUseDxPositionW | Specifies the `-fvk-use-dx-position-w` option. `intValue0` specifies a bool value for the setting. |
 | VulkanUseEntryPointName | When set, will keep the original name of entrypoints as they are defined in the source instead of renaming them to `main`. `intValue0` specifies a bool value for the setting. |
 | VulkanUseGLLayout | When set, will use std430 layout instead of D3D buffer layout for raw buffer load/stores. `intValue0` specifies a bool value for the setting. |
 | VulkanEmitReflection | Specifies the `-fspv-reflect` option. When set will include additional reflection instructions in the output SPIRV. `intValue0` specifies a bool value for the setting. |
diff --git a/slang.h b/slang.h
index fc242dbf1..28835f713 100644
--- a/slang.h
+++ b/slang.h
@@ -874,6 +874,7 @@ extern "C"
             VulkanBindShift,            // intValue0 (higher 8 bits): kind; intValue0(lower bits): set; intValue1: shift
             VulkanBindGlobals,          // intValue0: index; intValue1: set
             VulkanInvertY,              // bool
+            VulkanUseDxPositionW,       // bool
             VulkanUseEntryPointName,    // bool
             VulkanUseGLLayout,          // bool
             VulkanEmitReflection,       // bool
diff --git a/source/slang/slang-compiler-options.cpp b/source/slang/slang-compiler-options.cpp
index 100a5719f..974fbaaaa 100644
--- a/source/slang/slang-compiler-options.cpp
+++ b/source/slang/slang-compiler-options.cpp
@@ -113,6 +113,7 @@ namespace Slang
                 case CompilerOptionName::MatrixLayoutRow:
                 case CompilerOptionName::MatrixLayoutColumn:
                 case CompilerOptionName::VulkanInvertY:
+                case CompilerOptionName::VulkanUseDxPositionW:
                 case CompilerOptionName::VulkanUseEntryPointName:
                 case CompilerOptionName::VulkanUseGLLayout:
                 case CompilerOptionName::VulkanEmitReflection:
diff --git a/source/slang/slang-emit-spirv.cpp b/source/slang/slang-emit-spirv.cpp
index 2f318c464..8063975e2 100644
--- a/source/slang/slang-emit-spirv.cpp
+++ b/source/slang/slang-emit-spirv.cpp
@@ -3074,16 +3074,9 @@ struct SPIRVEmitContext
             }
             break;
         case kIROp_MaxVertexCountDecoration:
-            {
-                auto section = getSection(SpvLogicalSectionID::ExecutionModes);
-                auto maxVertexCount = cast<IRMaxVertexCountDecoration>(decoration);
-                emitOpExecutionModeOutputVertices(
-                    section,
-                    decoration,
-                    dstID,
-                    SpvLiteralInteger::from32(int32_t(getIntVal(maxVertexCount->getCount())))
-                );
-            }
+            // Don't do anything here, instead wait until we see OutputTopologyDecoration
+            // and emit them together to ensure MaxVertexCount always appears before OutputTopology,
+            // which seemed to be required by SPIRV.
             break;
         case kIROp_InstanceDecoration:
             {
@@ -3094,22 +3087,49 @@ struct SPIRVEmitContext
             }
             break;
         case kIROp_TriangleInputPrimitiveTypeDecoration:
-            emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), decoration, dstID, SpvExecutionModeTriangles);
-            break;
         case kIROp_LineInputPrimitiveTypeDecoration:
-            emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), decoration, dstID, SpvExecutionModeInputLines);
-            break;
         case kIROp_LineAdjInputPrimitiveTypeDecoration:
-            emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), decoration, dstID, SpvExecutionModeInputLinesAdjacency);
-            break;
         case kIROp_PointInputPrimitiveTypeDecoration:
-            emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), decoration, dstID, SpvExecutionModeInputPoints);
-            break;
         case kIROp_TriangleAdjInputPrimitiveTypeDecoration:
-            emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), decoration, dstID, SpvExecutionModeInputTrianglesAdjacency);
+            // Defer this until we see kIROp_StreamOutputTypeDecoration because the driver wants to see
+            // them before the output.
             break;
         case kIROp_StreamOutputTypeDecoration:
             {
+                for (auto inputDecor : decoration->getParent()->getDecorations())
+                {
+                    switch (inputDecor->getOp())
+                    {
+                    case kIROp_TriangleInputPrimitiveTypeDecoration:
+                        emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), inputDecor, dstID, SpvExecutionModeTriangles);
+                        break;
+                    case kIROp_LineInputPrimitiveTypeDecoration:
+                        emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), inputDecor, dstID, SpvExecutionModeInputLines);
+                        break;
+                    case kIROp_LineAdjInputPrimitiveTypeDecoration:
+                        emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), inputDecor, dstID, SpvExecutionModeInputLinesAdjacency);
+                        break;
+                    case kIROp_PointInputPrimitiveTypeDecoration:
+                        emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), inputDecor, dstID, SpvExecutionModeInputPoints);
+                        break;
+                    case kIROp_TriangleAdjInputPrimitiveTypeDecoration:
+                        emitOpExecutionMode(getSection(SpvLogicalSectionID::ExecutionModes), inputDecor, dstID, SpvExecutionModeInputTrianglesAdjacency);
+                        break;
+                    }
+                }
+                // SPIRV requires MaxVertexCount decoration to appear before OutputTopologyDecoration,
+                // so we emit them here.
+                if (auto maxVertexCount = decoration->getParent()->findDecoration<IRMaxVertexCountDecoration>())
+                {
+                    auto section = getSection(SpvLogicalSectionID::ExecutionModes);
+                    emitOpExecutionModeOutputVertices(
+                        section,
+                        maxVertexCount,
+                        dstID,
+                        SpvLiteralInteger::from32(int32_t(getIntVal(maxVertexCount->getCount())))
+                    );
+                }
+
                 auto decor = as<IRStreamOutputTypeDecoration>(decoration);
                 IRType* type = decor->getStreamType();
 
diff --git a/source/slang/slang-emit.cpp b/source/slang/slang-emit.cpp
index 5ec6fa62a..0ec9b0c1a 100644
--- a/source/slang/slang-emit.cpp
+++ b/source/slang/slang-emit.cpp
@@ -934,6 +934,8 @@ Result linkAndOptimizeIR(
         legalizeUniformBufferLoad(irModule);
         if (targetProgram->getOptionSet().getBoolOption(CompilerOptionName::VulkanInvertY))
             invertYOfPositionOutput(irModule);
+        if (targetProgram->getOptionSet().getBoolOption(CompilerOptionName::VulkanUseDxPositionW))
+            rcpWOfPositionInput(irModule);
     }
 
     // Lower sizeof/alignof
diff --git a/source/slang/slang-hlsl-to-vulkan-layout-options.cpp b/source/slang/slang-hlsl-to-vulkan-layout-options.cpp
index 0e6f70294..4246b7b39 100644
--- a/source/slang/slang-hlsl-to-vulkan-layout-options.cpp
+++ b/source/slang/slang-hlsl-to-vulkan-layout-options.cpp
@@ -53,7 +53,6 @@ void HLSLToVulkanLayoutOptions::loadFromOptionSet(CompilerOptionSet& optionSet)
             m_globalsBinding.set = (*bindGlobals)[0].intValue2;
         }
     }
-    m_invertY = optionSet.getBoolOption(CompilerOptionName::VulkanInvertY);
     m_useGLLayout = optionSet.getBoolOption(CompilerOptionName::VulkanUseGLLayout);
     m_useOriginalEntryPointName = optionSet.getBoolOption(CompilerOptionName::VulkanUseEntryPointName);
 
@@ -123,7 +122,7 @@ Index HLSLToVulkanLayoutOptions::getShift(Kind kind, Index set) const
 
 bool HLSLToVulkanLayoutOptions::hasState() const
 {
-    return canInferBindings() || hasGlobalsBinding() || shouldInvertY() || getUseOriginalEntryPointName()
+    return canInferBindings() || hasGlobalsBinding() || getUseOriginalEntryPointName()
         || shouldUseGLLayout() || shouldEmitSPIRVReflectionInfo();
 }
 
diff --git a/source/slang/slang-hlsl-to-vulkan-layout-options.h b/source/slang/slang-hlsl-to-vulkan-layout-options.h
index 2b8ec973c..1b7daf330 100644
--- a/source/slang/slang-hlsl-to-vulkan-layout-options.h
+++ b/source/slang/slang-hlsl-to-vulkan-layout-options.h
@@ -119,9 +119,6 @@ public:
         /// True if can infer a binding for a kind
     bool canInferBindingForKind(Kind kind) const { return (m_kindShiftEnabledFlags & getKindFlag(kind)) != 0; }
 
-        /// True if the compiler should invert the Y coordinate of any SV_Position output.
-    bool shouldInvertY() const { return m_invertY; }
-
     bool shouldUseGLLayout() const { return m_useGLLayout; }
 
     bool shouldEmitSPIRVReflectionInfo() const { return m_emitSPIRVReflectionInfo; }
@@ -152,8 +149,6 @@ public:
         /// Get the globals binding
     const Binding& getGlobalsBinding() const { return m_globalsBinding; }
 
-    void setInvertY(bool value) { m_invertY = value; }
-
     void setUseOriginalEntryPointName(bool value) { m_useOriginalEntryPointName = value; }
 
     void setUseGLLayout(bool value) { m_useGLLayout = value; }
@@ -186,9 +181,6 @@ protected:
         /// Maps a key to the amount of shift
     Dictionary<Key, Index> m_shifts;
 
-        /// Whether to invert the Y coordinate of SV_Position output.
-    bool m_invertY = false;
-
         /// If set, will use the original entry point name in the generated SPIRV instead of "main".
     bool m_useOriginalEntryPointName = false;
 
diff --git a/source/slang/slang-ir-glsl-legalize.cpp b/source/slang/slang-ir-glsl-legalize.cpp
index d2d09f45a..e40dc33ff 100644
--- a/source/slang/slang-ir-glsl-legalize.cpp
+++ b/source/slang/slang-ir-glsl-legalize.cpp
@@ -335,6 +335,7 @@ enum GLSLSystemValueKind
 {
     General,
     PositionOutput,
+    PositionInput,
 };
 
 struct GLSLSystemValueInfo
@@ -569,6 +570,7 @@ GLSLSystemValueInfo* getGLSLSystemValueInfo(
             && kind == LayoutResourceKind::VaryingInput )
         {
             name = "gl_FragCoord";
+            systemValueKind = GLSLSystemValueKind::PositionInput;
         }
         else if( stage == Stage::Geometry
             && kind == LayoutResourceKind::VaryingInput )
@@ -1008,6 +1010,9 @@ void createVarLayoutForLegalizedGlobalParam(
         case GLSLSystemValueKind::PositionOutput:
             builder->addGLPositionOutputDecoration(globalParam);
             break;
+        case GLSLSystemValueKind::PositionInput:
+            builder->addGLPositionInputDecoration(globalParam);
+            break;
         default:
             break;
         }
diff --git a/source/slang/slang-ir-inst-defs.h b/source/slang/slang-ir-inst-defs.h
index cd79f05a6..26e0ebd69 100644
--- a/source/slang/slang-ir-inst-defs.h
+++ b/source/slang/slang-ir-inst-defs.h
@@ -859,6 +859,9 @@ INST(HighLevelDeclDecoration,               highLevelDecl,          1, 0)
     INST(GLSLPrimitivesRateDecoration, perprimitive, 0, 0)
         // Marks an inst that represents the gl_Position output.
     INST(GLPositionOutputDecoration, PositionOutput, 0, 0)
+        // Marks an inst that represents the gl_Position input.
+    INST(GLPositionInputDecoration, PositionInput, 0, 0)
+
 
     /* StageAccessDecoration */
         INST(StageReadAccessDecoration, stageReadAccess, 0, 0)
diff --git a/source/slang/slang-ir-insts.h b/source/slang/slang-ir-insts.h
index 6fbccab5c..0a7f8c33b 100644
--- a/source/slang/slang-ir-insts.h
+++ b/source/slang/slang-ir-insts.h
@@ -1366,6 +1366,11 @@ struct IRGLPositionOutputDecoration : public IRDecoration
     IR_LEAF_ISA(GLPositionOutputDecoration)
 };
 
+struct IRGLPositionInputDecoration : public IRDecoration
+{
+    IR_LEAF_ISA(GLPositionInputDecoration)
+};
+
 struct IRMeshOutputRef : public IRInst
 {
     enum { kOp = kIROp_MeshOutputRef };
@@ -4297,6 +4302,11 @@ public:
         addDecoration(value, kIROp_GLPositionOutputDecoration);
     }
 
+    void addGLPositionInputDecoration(IRInst* value)
+    {
+        addDecoration(value, kIROp_GLPositionInputDecoration);
+    }
+
     void addInterpolationModeDecoration(IRInst* value, IRInterpolationMode mode)
     {
         addDecoration(value, kIROp_InterpolationModeDecoration, getIntValue(getIntType(), IRIntegerValue(mode)));
diff --git a/source/slang/slang-ir-vk-invert-y.cpp b/source/slang/slang-ir-vk-invert-y.cpp
index 5c21fee95..5e2dfcf2e 100644
--- a/source/slang/slang-ir-vk-invert-y.cpp
+++ b/source/slang/slang-ir-vk-invert-y.cpp
@@ -56,4 +56,36 @@ void invertYOfPositionOutput(IRModule* module)
     }
 }
 
+
+static IRInst* _invertWOfVector(IRBuilder& builder, IRInst* originalVector)
+{
+    auto vectorType = as<IRVectorType>(originalVector->getDataType());
+    SLANG_ASSERT(vectorType);
+    UInt elementIndexW = 3;
+    auto originalW = builder.emitSwizzle(vectorType->getElementType(), originalVector, 1, &elementIndexW);
+    auto rcpW = builder.emitDiv(originalW->getDataType(), builder.getFloatValue(originalW->getDataType(), 1.0), originalW);
+    auto newVal = builder.emitSwizzleSet(originalVector->getDataType(), originalVector, rcpW, 1, &elementIndexW);
+    return newVal;
+}
+
+// Find inputs of SV_Position and rcp the w coordinates of it right after the read.
+void rcpWOfPositionInput(IRModule* module)
+{
+    for (auto globalInst : module->getGlobalInsts())
+    {
+        if (globalInst->findDecoration<IRGLPositionInputDecoration>())
+        {
+            // Find all loads and replace them with reciprocals.
+            IRBuilder builder(module);
+            traverseUses(globalInst, [&](IRUse* use)
+                {
+                    // Get the inverted vector.
+                    builder.setInsertBefore(use->getUser());
+                    auto invertedVal = _invertWOfVector(builder, globalInst);
+                    // Replace original uses with the invertex vector.
+                    builder.replaceOperand(use, invertedVal);
+                });
+        }
+    }
+}
 }
diff --git a/source/slang/slang-ir-vk-invert-y.h b/source/slang/slang-ir-vk-invert-y.h
index b6b1e4a4c..4fc007169 100644
--- a/source/slang/slang-ir-vk-invert-y.h
+++ b/source/slang/slang-ir-vk-invert-y.h
@@ -5,6 +5,7 @@ namespace Slang
 {
     struct IRModule;
     void invertYOfPositionOutput(IRModule* module);
+    void rcpWOfPositionInput(IRModule* module);
 }
 
 #endif
diff --git a/source/slang/slang-options.cpp b/source/slang/slang-options.cpp
index 42eb66517..3d0b22edd 100644
--- a/source/slang/slang-options.cpp
+++ b/source/slang/slang-options.cpp
@@ -399,6 +399,7 @@ void initCommandOptions(CommandOptions& options)
         { OptionKind::VulkanBindGlobals, "-fvk-bind-globals", "-fvk-bind-globals <N> <descriptor-set>",
         "Places the $Globals cbuffer at descriptor set <descriptor-set> and binding <N>."},
         { OptionKind::VulkanInvertY, "-fvk-invert-y", nullptr, "Negates (additively inverts) SV_Position.y before writing to stage output."},
+        { OptionKind::VulkanUseDxPositionW, "-fvk-use-dx-position-w", nullptr, "Reciprocates (multiplicatively inverts) SV_Position.w after reading from stage input. For use in fragment shaders only."},
         { OptionKind::VulkanUseEntryPointName, "-fvk-use-entrypoint-name", nullptr, "Uses the entrypoint name from the source instead of 'main' in the spirv output."},
         { OptionKind::VulkanUseGLLayout, "-fvk-use-gl-layout", nullptr, "Use std430 layout instead of D3D buffer layout for raw buffer load/stores."},
         { OptionKind::VulkanEmitReflection, "-fspv-reflect", nullptr, "Include reflection decorations in the resulting SPIRV for shader parameters."},
@@ -1681,6 +1682,7 @@ SlangResult OptionsParser::_parse(
             case OptionKind::ValidateIr:
             case OptionKind::DumpIr:
             case OptionKind::VulkanInvertY:
+            case OptionKind::VulkanUseDxPositionW:
             case OptionKind::VulkanUseEntryPointName:
             case OptionKind::VulkanUseGLLayout:
             case OptionKind::VulkanEmitReflection:
diff --git a/tests/spirv/geometry-decoration-ordering.slang b/tests/spirv/geometry-decoration-ordering.slang
new file mode 100644
index 000000000..be6eee878
--- /dev/null
+++ b/tests/spirv/geometry-decoration-ordering.slang
@@ -0,0 +1,17 @@
+//TEST:SIMPLE(filecheck=CHECK):-target spirv -emit-spirv-directly -entry main -stage geometry
+
+// Make sure InputPoints always appears before OutputVertices and OutputTriangleStrip,
+// otherwise, the shader won't work correctly.
+
+// CHECK: OpExecutionMode {{.*}} InputPoints
+// CHECK: OpExecutionMode {{.*}} OutputVertices 4
+// CHECK: OpExecutionMode {{.*}} OutputTriangleStrip
+
+[maxvertexcount(4)]
+void main( point float4 input[1], inout TriangleStream<float4> OutputStream )
+{
+    OutputStream.Append(float4(0));
+    OutputStream.Append(float4(1));
+    OutputStream.Append(float4(2));
+    OutputStream.Append(float4(3));
+}
\ No newline at end of file
diff --git a/tests/spirv/rcp-w-fragment.slang b/tests/spirv/rcp-w-fragment.slang
new file mode 100644
index 000000000..1cd1b8ee8
--- /dev/null
+++ b/tests/spirv/rcp-w-fragment.slang
@@ -0,0 +1,7 @@
+//TEST:SIMPLE(filecheck=CHECK): -target spirv -entry main -stage fragment -fvk-use-dx-position-w -emit-spirv-directly
+
+// CHECK: OpFDiv
+float4 main(float4 pos : SV_Position) : SV_Target
+{
+    return pos;
+}
\ No newline at end of file