New nonuniform analysis (KhronosGroup#2457)

This implements a new nonunifom analysis suggested by @jbolz. This change
generates nonUniform decorations that were previously missing and avoids
generation of incorrect decorations. Most notably, it now generates
decorations for nonuniform functions and out params. It avoids generating
decorations for lvalues which themselves are not nonuniform.

Author: greg-lunarg

SPIRV/GlslangToSpv.cpp

@@ -149,6 +149,7 @@ class TGlslangToSpvTraverser : public glslang::TIntermTraverser {
     spv::Decoration TranslateInterpolationDecoration(const glslang::TQualifier& qualifier);
     spv::Decoration TranslateAuxiliaryStorageDecoration(const glslang::TQualifier& qualifier);
     spv::Decoration TranslateNonUniformDecoration(const glslang::TQualifier& qualifier);
+    spv::Decoration TranslateNonUniformDecoration(const spv::Builder::AccessChain::CoherentFlags& coherentFlags);
     spv::Builder::AccessChain::CoherentFlags TranslateCoherent(const glslang::TType& type);
     spv::MemoryAccessMask TranslateMemoryAccess(const spv::Builder::AccessChain::CoherentFlags &coherentFlags);
     spv::ImageOperandsMask TranslateImageOperands(const spv::Builder::AccessChain::CoherentFlags &coherentFlags);
@@ -539,6 +540,20 @@ spv::Decoration TGlslangToSpvTraverser::TranslateNonUniformDecoration(const glsl
         return spv::DecorationMax;
 }
 
+// If lvalue flags contains nonUniform, return SPIR-V NonUniform decoration.
+spv::Decoration TGlslangToSpvTraverser::TranslateNonUniformDecoration(
+    const spv::Builder::AccessChain::CoherentFlags& coherentFlags)
+{
+#ifndef GLSLANG_WEB
+    if (coherentFlags.isNonUniform()) {
+        builder.addIncorporatedExtension("SPV_EXT_descriptor_indexing", spv::Spv_1_5);
+        builder.addCapability(spv::CapabilityShaderNonUniformEXT);
+        return spv::DecorationNonUniformEXT;
+    } else
+#endif
+        return spv::DecorationMax;
+}
+
 spv::MemoryAccessMask TGlslangToSpvTraverser::TranslateMemoryAccess(
     const spv::Builder::AccessChain::CoherentFlags &coherentFlags)
 {
@@ -614,6 +629,7 @@ spv::Builder::AccessChain::CoherentFlags TGlslangToSpvTraverser::TranslateCohere
                        flags.volatil;
     flags.isImage = type.getBasicType() == glslang::EbtSampler;
 #endif
+    flags.nonUniform = type.getQualifier().nonUniform;
     return flags;
 }
 
@@ -1376,6 +1392,8 @@ void InheritQualifiers(glslang::TQualifier& child, const glslang::TQualifier& pa
     if (parent.writeonly)
         child.writeonly = true;
 #endif
+    if (parent.nonUniform)
+        child.nonUniform = true;
 }
 
 bool HasNonLayoutQualifiers(const glslang::TType& type, const glslang::TQualifier& qualifier)
@@ -1881,9 +1899,11 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T
                 spv::Id leftRValue = accessChainLoad(node->getLeft()->getType());
 
                 // do the operation
+                spv::Builder::AccessChain::CoherentFlags coherentFlags = TranslateCoherent(node->getLeft()->getType());
+                coherentFlags |= TranslateCoherent(node->getRight()->getType());
                 OpDecorations decorations = { TranslatePrecisionDecoration(node->getOperationPrecision()),
                                               TranslateNoContractionDecoration(node->getType().getQualifier()),
-                                              TranslateNonUniformDecoration(node->getType().getQualifier()) };
+                                              TranslateNonUniformDecoration(coherentFlags) };
                 rValue = createBinaryOperation(node->getOp(), decorations,
                                                convertGlslangToSpvType(node->getType()), leftRValue, rValue,
                                                node->getType().getBasicType());
@@ -1914,13 +1934,16 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T
             if (! node->getLeft()->getType().isArray() &&
                 node->getLeft()->getType().isVector() &&
                 node->getOp() == glslang::EOpIndexDirect) {
+                // Swizzle is uniform so propagate uniform into access chain
+                spv::Builder::AccessChain::CoherentFlags coherentFlags = TranslateCoherent(node->getLeft()->getType());
+                coherentFlags.nonUniform = 0;
                 // This is essentially a hard-coded vector swizzle of size 1,
                 // so short circuit the access-chain stuff with a swizzle.
                 std::vector<unsigned> swizzle;
                 swizzle.push_back(glslangIndex);
                 int dummySize;
                 builder.accessChainPushSwizzle(swizzle, convertGlslangToSpvType(node->getLeft()->getType()),
-                                               TranslateCoherent(node->getLeft()->getType()),
+                                               coherentFlags,
                                                glslangIntermediate->getBaseAlignmentScalar(
                                                    node->getLeft()->getType(), dummySize));
             } else {
@@ -1951,9 +1974,14 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T
                     }
                 }
 
+                // Struct reference propagates uniform lvalue
+                spv::Builder::AccessChain::CoherentFlags coherentFlags =
+                        TranslateCoherent(node->getLeft()->getType());
+                coherentFlags.nonUniform = 0;
+
                 // normal case for indexing array or structure or block
                 builder.accessChainPush(builder.makeIntConstant(spvIndex),
-                    TranslateCoherent(node->getLeft()->getType()),
+                        coherentFlags,
                         node->getLeft()->getType().getBufferReferenceAlignment());
 
                 // Add capabilities here for accessing PointSize and clip/cull distance.
@@ -1987,15 +2015,20 @@ bool TGlslangToSpvTraverser::visitBinary(glslang::TVisit /* visit */, glslang::T
             // restore the saved access chain
             builder.setAccessChain(partial);
 
+            // Only if index is nonUniform should we propagate nonUniform into access chain
+            spv::Builder::AccessChain::CoherentFlags index_flags = TranslateCoherent(node->getRight()->getType());
+            spv::Builder::AccessChain::CoherentFlags coherent_flags = TranslateCoherent(node->getLeft()->getType());
+            coherent_flags.nonUniform = index_flags.nonUniform;
+
             if (! node->getLeft()->getType().isArray() && node->getLeft()->getType().isVector()) {
                 int dummySize;
-                builder.accessChainPushComponent(index, convertGlslangToSpvType(node->getLeft()->getType()),
-                                                TranslateCoherent(node->getLeft()->getType()),
+                builder.accessChainPushComponent(
+                    index, convertGlslangToSpvType(node->getLeft()->getType()), coherent_flags,
                                                 glslangIntermediate->getBaseAlignmentScalar(node->getLeft()->getType(),
                                                 dummySize));
             } else
-                builder.accessChainPush(index, TranslateCoherent(node->getLeft()->getType()),
-                    node->getLeft()->getType().getBufferReferenceAlignment());
+                builder.accessChainPush(index, coherent_flags,
+                                        node->getLeft()->getType().getBufferReferenceAlignment());
         }
         return false;
     case glslang::EOpVectorSwizzle:
@@ -2119,7 +2152,7 @@ spv::Id TGlslangToSpvTraverser::translateForcedType(spv::Id object)
             // handle 32-bit v.xy* -> 64-bit
             builder.clearAccessChain();
             builder.setAccessChainLValue(object);
-            object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, objectTypeId);
+            object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, spv::DecorationMax, objectTypeId);
             std::vector<spv::Id> components;
             components.push_back(builder.createCompositeExtract(object, builder.getContainedTypeId(objectTypeId), 0));
             components.push_back(builder.createCompositeExtract(object, builder.getContainedTypeId(objectTypeId), 1));
@@ -2135,7 +2168,7 @@ spv::Id TGlslangToSpvTraverser::translateForcedType(spv::Id object)
             // and we insert a transpose after loading the original non-transposed builtins
             builder.clearAccessChain();
             builder.setAccessChainLValue(object);
-            object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, objectTypeId);
+            object = builder.accessChainLoad(spv::NoPrecision, spv::DecorationMax, spv::DecorationMax, objectTypeId);
             return builder.createUnaryOp(spv::OpTranspose, desiredTypeId, object);
 
     } else  {
@@ -2322,7 +2355,7 @@ bool TGlslangToSpvTraverser::visitUnary(glslang::TVisit /* visit */, glslang::TI
             // The result of operation is always stored, but conditionally the
             // consumed result.  The consumed result is always an r-value.
             builder.accessChainStore(result,
-                                     TranslateNonUniformDecoration(node->getOperand()->getType().getQualifier()));
+                                     TranslateNonUniformDecoration(builder.getAccessChain().coherentFlags));
             builder.clearAccessChain();
             if (node->getOp() == glslang::EOpPreIncrement ||
                 node->getOp() == glslang::EOpPreDecrement)
@@ -2398,7 +2431,6 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt
     spv::Id invertedType = spv::NoType;                     // to use to override the natural type of the node
     std::vector<spv::Builder::AccessChain> complexLvalues;  // for holding swizzling l-values too complex for
                                                             // SPIR-V, for an out parameter
-    std::vector<glslang::TQualifier> complexLValueQualifiers;
     std::vector<spv::Id> temporaryLvalues;                  // temporaries to pass, as proxies for complexLValues
 
     auto resultType = [&invertedType, &node, this](){ return invertedType != spv::NoType ?
@@ -3020,7 +3052,6 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt
                 // receive the result, and must later swizzle that into the original
                 // l-value.
                 complexLvalues.push_back(builder.getAccessChain());
-                complexLValueQualifiers.push_back(glslangOperands[arg]->getAsTyped()->getType().getQualifier());
                 temporaryLvalues.push_back(builder.createVariable(
                     spv::NoPrecision, spv::StorageClassFunction,
                     builder.accessChainGetInferredType(), "swizzleTemp"));
@@ -3125,7 +3156,8 @@ bool TGlslangToSpvTraverser::visitAggregate(glslang::TVisit visit, glslang::TInt
 
         for (unsigned int i = 0; i < temporaryLvalues.size(); ++i) {
             builder.setAccessChain(complexLvalues[i]);
-            builder.accessChainStore(builder.createLoad(temporaryLvalues[i], spv::NoPrecision), TranslateNonUniformDecoration(complexLValueQualifiers[i]));
+            builder.accessChainStore(builder.createLoad(temporaryLvalues[i], spv::NoPrecision),
+                TranslateNonUniformDecoration(complexLvalues[i].coherentFlags));
         }
     }
 
@@ -4135,6 +4167,7 @@ spv::Id TGlslangToSpvTraverser::accessChainLoad(const glslang::TType& type)
     alignment |= type.getBufferReferenceAlignment();
 
     spv::Id loadedId = builder.accessChainLoad(TranslatePrecisionDecoration(type),
+        TranslateNonUniformDecoration(builder.getAccessChain().coherentFlags),
         TranslateNonUniformDecoration(type.getQualifier()),
         nominalTypeId,
         spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) & ~spv::MemoryAccessMakePointerAvailableKHRMask),
@@ -4202,7 +4235,7 @@ void TGlslangToSpvTraverser::accessChainStore(const glslang::TType& type, spv::I
     unsigned int alignment = builder.getAccessChain().alignment;
     alignment |= type.getBufferReferenceAlignment();
 
-    builder.accessChainStore(rvalue, TranslateNonUniformDecoration(type.getQualifier()),
+    builder.accessChainStore(rvalue, TranslateNonUniformDecoration(builder.getAccessChain().coherentFlags),
                              spv::MemoryAccessMask(TranslateMemoryAccess(coherentFlags) &
                                 ~spv::MemoryAccessMakePointerVisibleKHRMask),
                              TranslateMemoryScope(coherentFlags), alignment);
@@ -4734,8 +4767,10 @@ void TGlslangToSpvTraverser::translateArguments(const glslang::TIntermAggregate&
         }
 
         if (lvalue) {
-            arguments.push_back(builder.accessChainGetLValue());
+            spv::Id lvalue_id = builder.accessChainGetLValue();
+            arguments.push_back(lvalue_id);
             lvalueCoherentFlags = builder.getAccessChain().coherentFlags;
+            builder.addDecoration(lvalue_id, TranslateNonUniformDecoration(lvalueCoherentFlags));
             lvalueCoherentFlags |= TranslateCoherent(glslangArguments[i]->getAsTyped()->getType());
         } else
 #endif
@@ -5415,6 +5450,7 @@ spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAgg
     // 3. Make the call.
     spv::Id result = builder.createFunctionCall(function, spvArgs);
     builder.setPrecision(result, TranslatePrecisionDecoration(node->getType()));
+    builder.addDecoration(result, TranslateNonUniformDecoration(node->getType().getQualifier()));
 
     // 4. Copy back out an "out" arguments.
     lValueCount = 0;
@@ -5424,6 +5460,7 @@ spv::Id TGlslangToSpvTraverser::handleUserFunctionCall(const glslang::TIntermAgg
         else if (writableParam(qualifiers[a])) {
             if (qualifiers[a] == glslang::EvqOut || qualifiers[a] == glslang::EvqInOut) {
                 spv::Id copy = builder.createLoad(spvArgs[a], spv::NoPrecision);
+                builder.addDecoration(copy, TranslateNonUniformDecoration(argTypes[a]->getQualifier()));
                 builder.setAccessChain(lValues[lValueCount]);
                 multiTypeStore(*argTypes[a], copy);
             }
@@ -8228,9 +8265,6 @@ spv::Id TGlslangToSpvTraverser::getSymbolId(const glslang::TIntermSymbol* symbol
             builder.addDecoration(id, memory[i]);
     }
 
-    // nonuniform
-    builder.addDecoration(id, TranslateNonUniformDecoration(symbol->getType().getQualifier()));
-
     if (builtIn == spv::BuiltInSampleMask) {
           spv::Decoration decoration;
           // GL_NV_sample_mask_override_coverage extension

SPIRV/SpvBuilder.cpp

@@ -2798,8 +2798,9 @@ void Builder::accessChainStore(Id rvalue, Decoration nonUniform, spv::MemoryAcce
 }
 
 // Comments in header
-Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resultType,
-    spv::MemoryAccessMask memoryAccess, spv::Scope scope, unsigned int alignment)
+Id Builder::accessChainLoad(Decoration precision, Decoration l_nonUniform,
+    Decoration r_nonUniform, Id resultType, spv::MemoryAccessMask memoryAccess,
+    spv::Scope scope, unsigned int alignment)
 {
     Id id;
 
@@ -2863,9 +2864,9 @@ Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resu
         // Buffer accesses need the access chain decorated, and this is where
         // loaded image types get decorated. TODO: This should maybe move to
         // createImageTextureFunctionCall.
-        addDecoration(id, nonUniform);
+        addDecoration(id, l_nonUniform);
         id = createLoad(id, precision, memoryAccess, scope, alignment);
-        addDecoration(id, nonUniform);
+        addDecoration(id, r_nonUniform);
     }
 
     // Done, unless there are swizzles to do
@@ -2886,7 +2887,7 @@ Id Builder::accessChainLoad(Decoration precision, Decoration nonUniform, Id resu
     if (accessChain.component != NoResult)
         id = setPrecision(createVectorExtractDynamic(id, resultType, accessChain.component), precision);
 
-    addDecoration(id, nonUniform);
+    addDecoration(id, r_nonUniform);
     return id;
 }
 

SPIRV/SpvBuilder.h

@@ -625,6 +625,7 @@ class Builder {
             CoherentFlags operator |=(const CoherentFlags &other) { return *this; }
 #else
             bool isVolatile() const { return volatil; }
+            bool isNonUniform() const { return nonUniform; }
             bool anyCoherent() const {
                 return coherent || devicecoherent || queuefamilycoherent || workgroupcoherent ||
                     subgroupcoherent || shadercallcoherent;
@@ -639,6 +640,7 @@ class Builder {
             unsigned nonprivate : 1;
             unsigned volatil : 1;
             unsigned isImage : 1;
+            unsigned nonUniform : 1;
 
             void clear() {
                 coherent = 0;
@@ -650,6 +652,7 @@ class Builder {
                 nonprivate = 0;
                 volatil = 0;
                 isImage = 0;
+                nonUniform = 0;
             }
 
             CoherentFlags operator |=(const CoherentFlags &other) {
@@ -662,6 +665,7 @@ class Builder {
                 nonprivate |= other.nonprivate;
                 volatil |= other.volatil;
                 isImage |= other.isImage;
+                nonUniform |= other.nonUniform;
                 return *this;
             }
 #endif
@@ -727,7 +731,7 @@ class Builder {
         spv::Scope scope = spv::ScopeMax, unsigned int alignment = 0);
 
     // use accessChain and swizzle to load an r-value
-    Id accessChainLoad(Decoration precision, Decoration nonUniform, Id ResultType,
+    Id accessChainLoad(Decoration precision, Decoration l_nonUniform, Decoration r_nonUniform, Id ResultType,
         spv::MemoryAccessMask memoryAccess = spv::MemoryAccessMaskNone, spv::Scope scope = spv::ScopeMax,
             unsigned int alignment = 0);