Merge pull request #1523 from sparmarNV/fix-SPV_NV_mesh_shader

StandAlone/StandAlone.cpp

@@ -1257,6 +1257,8 @@ int C_DECL main(int argc, char* argv[])
 //   .rchit = ray closest hit
 //   .rmiss = ray miss
 //   .rcall = ray callable
+//   .mesh  = mesh
+//   .task  = task
 //   Additionally, the file names may end in .<stage>.glsl and .<stage>.hlsl
 //   where <stage> is one of the stages listed above.
 //
@@ -1395,7 +1397,7 @@ void usage()
            "    .rahit   for a ray any hit shader\n"
            "    .rchit   for a ray closest hit shader\n"
            "    .rmiss   for a ray miss shader\n"
-           "    .rcall   for a ray callable shader"
+           "    .rcall   for a ray callable shader\n"
 #endif
            "    .glsl   for .vert.glsl, .tesc.glsl, ..., .comp.glsl compound suffixes\n"
            "    .hlsl   for .vert.hlsl, .tesc.hlsl, ..., .comp.hlsl compound suffixes\n"

Test/baseResults/spv.320.meshShaderUserDefined.mesh.out

-spv.meshShaderUserDefined.mesh
+spv.320.meshShaderUserDefined.mesh
 // Module Version 10000
 // Generated by (magic number): 80007
 // Id's are bound by 140

Test/spv.meshShaderRedeclBuiltins.mesh

+#version 460
+
+#define MAX_VER  81
+#define MAX_PRIM 32
+
+#define BARRIER() \
+    memoryBarrierShared(); \
+    barrier();
+
+#extension GL_NV_mesh_shader : enable
+
+layout(local_size_x = 32) in;
+
+layout(max_vertices=MAX_VER) out;
+layout(max_primitives=MAX_PRIM) out;
+layout(triangles) out;
+
+// test use of redeclared single-view builtins in mesh shaders:
+
+out gl_MeshPerVertexNV {
+    vec4 gl_Position;
+    float gl_PointSize;
+    float gl_ClipDistance[4];
+    float gl_CullDistance[4];
+} gl_MeshVerticesNV[];
+
+perprimitiveNV out gl_MeshPerPrimitiveNV {
+    int gl_PrimitiveID;
+    int gl_Layer;
+    int gl_ViewportIndex;
+    int gl_ViewportMask[];
+} gl_MeshPrimitivesNV[];
+
+void main()
+{
+    uint iid = gl_LocalInvocationID.x;
+    uint gid = gl_WorkGroupID.x;
+
+    gl_MeshVerticesNV[iid].gl_Position = vec4(1.0);
+    gl_MeshVerticesNV[iid].gl_PointSize = 2.0;
+    gl_MeshVerticesNV[iid].gl_ClipDistance[3] = 3.0;
+    gl_MeshVerticesNV[iid].gl_CullDistance[2] = 4.0;
+
+    BARRIER();
+
+    gl_MeshVerticesNV[iid+1].gl_Position = gl_MeshVerticesNV[iid].gl_Position;
+    gl_MeshVerticesNV[iid+1].gl_PointSize = gl_MeshVerticesNV[iid].gl_PointSize;
+    gl_MeshVerticesNV[iid+1].gl_ClipDistance[3] = gl_MeshVerticesNV[iid].gl_ClipDistance[3];
+    gl_MeshVerticesNV[iid+1].gl_CullDistance[2] = gl_MeshVerticesNV[iid].gl_CullDistance[2];
+
+    BARRIER();
+
+    gl_MeshPrimitivesNV[iid].gl_PrimitiveID = 6;
+    gl_MeshPrimitivesNV[iid].gl_Layer = 7;
+    gl_MeshPrimitivesNV[iid].gl_ViewportIndex = 8;
+    gl_MeshPrimitivesNV[iid].gl_ViewportMask[0] = 9;
+
+    BARRIER();
+
+    gl_MeshPrimitivesNV[iid+1].gl_PrimitiveID = gl_MeshPrimitivesNV[iid].gl_PrimitiveID;
+    gl_MeshPrimitivesNV[iid+1].gl_Layer = gl_MeshPrimitivesNV[iid].gl_Layer;
+    gl_MeshPrimitivesNV[iid+1].gl_ViewportIndex = gl_MeshPrimitivesNV[iid].gl_ViewportIndex;
+    gl_MeshPrimitivesNV[iid+1].gl_ViewportMask[0] = gl_MeshPrimitivesNV[iid].gl_ViewportMask[0];
+
+    BARRIER();
+}

Test/spv.meshShaderRedeclPerViewBuiltins.mesh

+#version 450
+
+#define MAX_VER  81
+#define MAX_PRIM 32
+#define MAX_VIEWS gl_MaxMeshViewCountNV
+
+#define BARRIER() \
+    memoryBarrierShared(); \
+    barrier();
+
+#extension GL_NV_mesh_shader : enable
+
+layout(local_size_x = 32) in;
+
+layout(max_vertices=MAX_VER) out;
+layout(max_primitives=MAX_PRIM) out;
+layout(triangles) out;
+
+// test use of redeclared per-view builtin attributes
+
+out gl_MeshPerVertexNV {
+    perviewNV vec4 gl_PositionPerViewNV[MAX_VIEWS];         // explicitly sized view dim
+    perviewNV float gl_ClipDistancePerViewNV[MAX_VIEWS][4]; // explicitly sized view dim
+    perviewNV float gl_CullDistancePerViewNV[MAX_VIEWS][4]; // explicitly sized view dim
+} gl_MeshVerticesNV[];
+
+perprimitiveNV out gl_MeshPerPrimitiveNV {
+    perviewNV int gl_LayerPerViewNV[];                      // implicitly sized view dim
+    perviewNV int gl_ViewportMaskPerViewNV[][1];            // implicitly sized view dim
+} gl_MeshPrimitivesNV[];
+
+void main()
+{
+    uint iid = gl_LocalInvocationID.x;
+    uint viewID = gl_MeshViewIndicesNV[gl_MeshViewCountNV%MAX_VIEWS];
+
+    gl_MeshVerticesNV[iid].gl_PositionPerViewNV[viewID]          = vec4(1.0, 2.0, 3.0, 4.0);
+    gl_MeshVerticesNV[iid].gl_ClipDistancePerViewNV[viewID][2]   = 5.0;
+    gl_MeshVerticesNV[iid].gl_CullDistancePerViewNV[viewID][3]   = 6.0;
+    gl_MeshPrimitivesNV[iid].gl_LayerPerViewNV[viewID]           = 7;
+    gl_MeshPrimitivesNV[iid].gl_ViewportMaskPerViewNV[viewID][0] = 8;
+
+    BARRIER();
+
+    gl_MeshVerticesNV[iid+1].gl_PositionPerViewNV[viewID]          = gl_MeshVerticesNV[iid].gl_PositionPerViewNV[viewID];
+    gl_MeshVerticesNV[iid+1].gl_ClipDistancePerViewNV[viewID][2]   = gl_MeshVerticesNV[iid].gl_ClipDistancePerViewNV[viewID][2];
+    gl_MeshVerticesNV[iid+1].gl_CullDistancePerViewNV[viewID][3]   = gl_MeshVerticesNV[iid].gl_CullDistancePerViewNV[viewID][3];
+    gl_MeshPrimitivesNV[iid+1].gl_LayerPerViewNV[viewID]           = gl_MeshPrimitivesNV[iid].gl_LayerPerViewNV[viewID];
+    gl_MeshPrimitivesNV[iid+1].gl_ViewportMaskPerViewNV[viewID][0] = gl_MeshPrimitivesNV[iid].gl_ViewportMaskPerViewNV[viewID][0];
+
+    BARRIER();
+}
+

glslang/Include/Types.h

@@ -1644,6 +1644,11 @@ public:
     {
         if (isUnsizedArray() && !(skipNonvariablyIndexed || isArrayVariablyIndexed()))
             changeOuterArraySize(getImplicitArraySize());
+#ifdef NV_EXTENSIONS
+        // For multi-dim per-view arrays, set unsized inner dimension size to 1
+        if (qualifier.isPerView() && arraySizes && arraySizes->isInnerUnsized())
+            arraySizes->clearInnerUnsized();
+#endif
         if (isStruct() && structure->size() > 0) {
             int lastMember = (int)structure->size() - 1;
             for (int i = 0; i < lastMember; ++i)

glslang/MachineIndependent/Initialize.cpp

@@ -5296,8 +5296,8 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
                 "float gl_ClipDistance[];"
                 "float gl_CullDistance[];"
                 "perviewNV vec4 gl_PositionPerViewNV[];"
-                "perviewNV float gl_ClipDistancePerViewNV[][8];"
-                "perviewNV float gl_CullDistancePerViewNV[][8];"
+                "perviewNV float gl_ClipDistancePerViewNV[][];"
+                "perviewNV float gl_CullDistancePerViewNV[][];"
             "} gl_MeshVerticesNV[];"
         );
 
@@ -5309,7 +5309,7 @@ void TBuiltIns::initialize(int version, EProfile profile, const SpvVersion& spvV
                 "int gl_ViewportIndex;"
                 "int gl_ViewportMask[];"
                 "perviewNV int gl_LayerPerViewNV[];"
-                "perviewNV int gl_ViewportMaskPerViewNV[][1];"
+                "perviewNV int gl_ViewportMaskPerViewNV[][];"
             "} gl_MeshPrimitivesNV[];"
         );
 

glslang/MachineIndependent/ParseHelper.cpp

@@ -400,9 +400,21 @@ TIntermTyped* TParseContext::handleBracketDereference(const TSourceLoc& loc, TIn
         handleIoResizeArrayAccess(loc, base);
 
     if (index->getQualifier().isFrontEndConstant()) {
-        if (base->getType().isUnsizedArray())
+        if (base->getType().isUnsizedArray()) {
             base->getWritableType().updateImplicitArraySize(indexValue + 1);
-        else
+#ifdef NV_EXTENSIONS
+            // For 2D per-view builtin arrays, update the inner dimension size in parent type
+            if (base->getQualifier().isPerView() && base->getQualifier().builtIn != EbvNone) {
+                TIntermBinary* binaryNode = base->getAsBinaryNode();
+                if (binaryNode) {
+                    TType& leftType = binaryNode->getLeft()->getWritableType();
+                    TArraySizes& arraySizes = *leftType.getArraySizes();
+                    assert(arraySizes.getNumDims() == 2);
+                    arraySizes.setDimSize(1, std::max(arraySizes.getDimSize(1), indexValue + 1));
+                }
+            }
+#endif
+        } else
             checkIndex(loc, base->getType(), indexValue);
         result = intermediate.addIndex(EOpIndexDirect, base, index, loc);
     } else {
@@ -813,6 +825,8 @@ TIntermTyped* TParseContext::handleDotDereference(const TSourceLoc& loc, TInterm
                 TIntermTyped* index = intermediate.addConstantUnion(member, loc);
                 result = intermediate.addIndex(EOpIndexDirectStruct, base, index, loc);
                 result->setType(*(*fields)[member].type);
+                if ((*fields)[member].type->getQualifier().isIo())
+                    intermediate.addIoAccessed(field);
             }
         } else
             error(loc, "no such field in structure", field.c_str(), "");
@@ -4059,10 +4073,31 @@ void TParseContext::redeclareBuiltinBlock(const TSourceLoc& loc, TTypeList& newT
                 error(memberLoc, "cannot redeclare block member with a different type", member->type->getFieldName().c_str(), "");
             if (oldType.isArray() != newType.isArray())
                 error(memberLoc, "cannot change arrayness of redeclared block member", member->type->getFieldName().c_str(), "");
-            else if (! oldType.sameArrayness(newType) && oldType.isSizedArray())
+            else if (! oldType.getQualifier().isPerView() && ! oldType.sameArrayness(newType) && oldType.isSizedArray())
                 error(memberLoc, "cannot change array size of redeclared block member", member->type->getFieldName().c_str(), "");
-            else if (newType.isArray())
+            else if (! oldType.getQualifier().isPerView() && newType.isArray())
                 arrayLimitCheck(loc, member->type->getFieldName(), newType.getOuterArraySize());
+#ifdef NV_EXTENSIONS
+            if (oldType.getQualifier().isPerView() && ! newType.getQualifier().isPerView())
+                error(memberLoc, "missing perviewNV qualifier to redeclared block member", member->type->getFieldName().c_str(), "");
+            else if (! oldType.getQualifier().isPerView() && newType.getQualifier().isPerView())
+                error(memberLoc, "cannot add perviewNV qualifier to redeclared block member", member->type->getFieldName().c_str(), "");
+            else if (newType.getQualifier().isPerView()) {
+                if (oldType.getArraySizes()->getNumDims() != newType.getArraySizes()->getNumDims())
+                    error(memberLoc, "cannot change arrayness of redeclared block member", member->type->getFieldName().c_str(), "");
+                else if (! newType.isUnsizedArray() && newType.getOuterArraySize() != resources.maxMeshViewCountNV)
+                    error(loc, "mesh view output array size must be gl_MaxMeshViewCountNV or implicitly sized", "[]", "");
+                else if (newType.getArraySizes()->getNumDims() == 2) {
+                    int innerDimSize = newType.getArraySizes()->getDimSize(1);
+                    arrayLimitCheck(memberLoc, member->type->getFieldName(), innerDimSize);
+                    oldType.getArraySizes()->setDimSize(1, innerDimSize);
+                }
+            }
+            if (oldType.getQualifier().isPerPrimitive() && ! newType.getQualifier().isPerPrimitive())
+                error(memberLoc, "missing perprimitiveNV qualifier to redeclared block member", member->type->getFieldName().c_str(), "");
+            else if (! oldType.getQualifier().isPerPrimitive() && newType.getQualifier().isPerPrimitive())
+                error(memberLoc, "cannot add perprimitiveNV qualifier to redeclared block member", member->type->getFieldName().c_str(), "");
+#endif
             if (newType.getQualifier().isMemory())
                 error(memberLoc, "cannot add memory qualifier to redeclared block member", member->type->getFieldName().c_str(), "");
             if (newType.getQualifier().hasNonXfbLayout())
@@ -4421,6 +4456,12 @@ void TParseContext::arrayLimitCheck(const TSourceLoc& loc, const TString& identi
         limitCheck(loc, size, "gl_MaxClipDistances", "gl_ClipDistance array size");
     else if (identifier.compare("gl_CullDistance") == 0)
         limitCheck(loc, size, "gl_MaxCullDistances", "gl_CullDistance array size");
+#ifdef NV_EXTENSIONS
+    else if (identifier.compare("gl_ClipDistancePerViewNV") == 0)
+        limitCheck(loc, size, "gl_MaxClipDistances", "gl_ClipDistancePerViewNV array size");
+    else if (identifier.compare("gl_CullDistancePerViewNV") == 0)
+        limitCheck(loc, size, "gl_MaxCullDistances", "gl_CullDistancePerViewNV array size");
+#endif
 }
 
 // See if the provided value is less than or equal to the symbol indicated by limit,

glslang/MachineIndependent/linkValidate.cpp

@@ -715,6 +715,17 @@ void TIntermediate::finalCheck(TInfoSink& infoSink, bool keepUncalled)
             error(infoSink, "Only one shaderRecordNVX buffer block is allowed per stage");
         break;
     case EShLangMeshNV:
+        // NV_mesh_shader doesn't allow use of both single-view and per-view builtins.
+        if (inIoAccessed("gl_Position") && inIoAccessed("gl_PositionPerViewNV"))
+            error(infoSink, "Can only use one of gl_Position or gl_PositionPerViewNV");
+        if (inIoAccessed("gl_ClipDistance") && inIoAccessed("gl_ClipDistancePerViewNV"))
+            error(infoSink, "Can only use one of gl_ClipDistance or gl_ClipDistancePerViewNV");
+        if (inIoAccessed("gl_CullDistance") && inIoAccessed("gl_CullDistancePerViewNV"))
+            error(infoSink, "Can only use one of gl_CullDistance or gl_CullDistancePerViewNV");
+        if (inIoAccessed("gl_Layer") && inIoAccessed("gl_LayerPerViewNV"))
+            error(infoSink, "Can only use one of gl_Layer or gl_LayerPerViewNV");
+        if (inIoAccessed("gl_ViewportMask") && inIoAccessed("gl_ViewportMaskPerViewNV"))
+            error(infoSink, "Can only use one of gl_ViewportMask or gl_ViewportMaskPerViewNV");
         if (outputPrimitive == ElgNone)
             error(infoSink, "At least one shader must specify an output layout primitive");
         if (vertices == TQualifier::layoutNotSet)

gtests/Spv.FromFile.cpp

@@ -524,6 +524,9 @@ INSTANTIATE_TEST_CASE_P(
     "spv.meshShaderPerViewUserDefined.mesh",
     "spv.meshShaderSharedMem.mesh",
     "spv.meshShaderTaskMem.mesh",
+    "spv.320.meshShaderUserDefined.mesh",
+    "spv.meshShaderRedeclBuiltins.mesh",
+    "spv.meshShaderRedeclPerViewBuiltins.mesh",
     "spv.meshTaskShader.task",
 })),
 FileNameAsCustomTestSuffix