Merge pull request #2659 from alan-baker/subgroup-uniform-control-flow

Implement GL_EXT_subgroup_uniform_control_flow.

SPIRV/GLSL.ext.KHR.h

@@ -51,5 +51,6 @@ static const char* const E_SPV_KHR_ray_query                    = "SPV_KHR_ray_q
 static const char* const E_SPV_KHR_fragment_shading_rate        = "SPV_KHR_fragment_shading_rate";
 static const char* const E_SPV_KHR_terminate_invocation         = "SPV_KHR_terminate_invocation";
 static const char* const E_SPV_KHR_workgroup_memory_explicit_layout = "SPV_KHR_workgroup_memory_explicit_layout";
+static const char* const E_SPV_KHR_subgroup_uniform_control_flow = "SPV_KHR_subgroup_uniform_control_flow";
 
 #endif  // #ifndef GLSLextKHR_H

SPIRV/GlslangToSpv.cpp

@@ -1526,6 +1526,13 @@ TGlslangToSpvTraverser::TGlslangToSpvTraverser(unsigned int spvVersion,
         builder.addCapability(spv::CapabilityRayTraversalPrimitiveCullingKHR);
     }
 
+#ifndef GLSLANG_WEB
+    if (glslangIntermediate->getSubgroupUniformControlFlow()) {
+        builder.addExtension(spv::E_SPV_KHR_subgroup_uniform_control_flow);
+        builder.addExecutionMode(shaderEntry, spv::ExecutionModeSubgroupUniformControlFlowKHR);
+    }
+#endif
+
     unsigned int mode;
     switch (glslangIntermediate->getStage()) {
     case EShLangVertex:

SPIRV/doc.cpp

@@ -188,6 +188,7 @@ const char* ExecutionModeString(int mode)
     case ExecutionModeRoundingModeRTE:          return "RoundingModeRTE";
     case ExecutionModeRoundingModeRTZ:          return "RoundingModeRTZ";
     case ExecutionModeStencilRefReplacingEXT:   return "StencilRefReplacingEXT";
+    case ExecutionModeSubgroupUniformControlFlowKHR: return "SubgroupUniformControlFlow";
 
     case ExecutionModeOutputLinesNV:            return "OutputLinesNV";
     case ExecutionModeOutputPrimitivesNV:       return "OutputPrimitivesNV";

SPIRV/spirv.hpp

@@ -150,6 +150,7 @@ enum ExecutionMode {
     ExecutionModeSubgroupsPerWorkgroupId = 37,
     ExecutionModeLocalSizeId = 38,
     ExecutionModeLocalSizeHintId = 39,
+    ExecutionModeSubgroupUniformControlFlowKHR = 4421,
     ExecutionModePostDepthCoverage = 4446,
     ExecutionModeDenormPreserve = 4459,
     ExecutionModeDenormFlushToZero = 4460,

Test/baseResults/spv.subgroupUniformControlFlow.vert.out

@@ -0,0 +1,23 @@
+spv.subgroupUniformControlFlow.vert
+WARNING: 0:7: '' : attribute with arguments not recognized, skipping 
+
+Validation failed
+// Module Version 10000
+// Generated by (magic number): 8000a
+// Id's are bound by 6
+
+                              Capability Shader
+                              Extension  "SPV_KHR_subgroup_uniform_control_flow"
+               1:             ExtInstImport  "GLSL.std.450"
+                              MemoryModel Logical GLSL450
+                              EntryPoint Vertex 4  "main"
+                              ExecutionMode 4 SubgroupUniformControlFlow
+                              Source GLSL 460
+                              SourceExtension  "GL_EXT_subgroup_uniform_control_flow"
+                              Name 4  "main"
+               2:             TypeVoid
+               3:             TypeFunction 2
+         4(main):           2 Function None 3
+               5:             Label
+                              Return
+                              FunctionEnd

Test/spv.subgroupUniformControlFlow.vert

@@ -0,0 +1,11 @@
+#version 460
+
+#ifdef GL_EXT_subgroup_uniform_control_flow
+
+#extension GL_EXT_subgroup_uniform_control_flow : enable
+
+[[random(4)]] void main() [[subgroup_uniform_control_flow]]
+{
+}
+
+#endif

glslang/MachineIndependent/ParseHelper.h

@@ -470,6 +470,8 @@ public:
     void handleSwitchAttributes(const TAttributes& attributes, TIntermNode*);
     // Determine loop control from attributes
     void handleLoopAttributes(const TAttributes& attributes, TIntermNode*);
+    // Function attributes
+    void handleFunctionAttributes(const TSourceLoc&, const TAttributes&, TFunction*);
 #endif
 
     void checkAndResizeMeshViewDim(const TSourceLoc&, TType&, bool isBlockMember);

glslang/MachineIndependent/Versions.cpp

@@ -251,6 +251,7 @@ void TParseVersions::initializeExtensionBehavior()
 
     extensionBehavior[E_GL_EXT_shader_16bit_storage]                    = EBhDisable;
     extensionBehavior[E_GL_EXT_shader_8bit_storage]                     = EBhDisable;
+    extensionBehavior[E_GL_EXT_subgroup_uniform_control_flow]           = EBhDisable;
 
     // #line and #include
     extensionBehavior[E_GL_GOOGLE_cpp_style_line_directive]          = EBhDisable;
@@ -415,6 +416,7 @@ void TParseVersions::getPreamble(std::string& preamble)
             }
             if (version >= 310) {
                 preamble += "#define GL_EXT_null_initializer 1\n";
+                preamble += "#define GL_EXT_subgroup_uniform_control_flow 1\n";
             }
 
     } else { // !isEsProfile()
@@ -546,6 +548,7 @@ void TParseVersions::getPreamble(std::string& preamble)
         }
         if (version >= 140) {
             preamble += "#define GL_EXT_null_initializer 1\n";
+            preamble += "#define GL_EXT_subgroup_uniform_control_flow 1\n";
         }
 #endif // GLSLANG_WEB
     }

glslang/MachineIndependent/Versions.h

@@ -204,6 +204,7 @@ const char* const E_GL_EXT_fragment_shading_rate            = "GL_EXT_fragment_s
 const char* const E_GL_EXT_shader_image_int64               = "GL_EXT_shader_image_int64";
 const char* const E_GL_EXT_null_initializer                 = "GL_EXT_null_initializer";
 const char* const E_GL_EXT_shared_memory_block              = "GL_EXT_shared_memory_block";
+const char* const E_GL_EXT_subgroup_uniform_control_flow    = "GL_EXT_subgroup_uniform_control_flow";
 
 // Arrays of extensions for the above viewportEXTs duplications
 

glslang/MachineIndependent/attribute.cpp

@@ -123,6 +123,8 @@ TAttributeType TParseContext::attributeFromName(const TString& name) const
         return EatPeelCount;
     else if (name == "partial_count")
         return EatPartialCount;
+    else if (name == "subgroup_uniform_control_flow")
+        return EatSubgroupUniformControlFlow;
     else
         return EatNone;
 }
@@ -341,6 +343,29 @@ void TParseContext::handleLoopAttributes(const TAttributes& attributes, TIntermN
     }
 }
 
+
+//
+// Function attributes
+//
+void TParseContext::handleFunctionAttributes(const TSourceLoc& loc, const TAttributes& attributes, TFunction* function)
+{
+    for (auto it = attributes.begin(); it != attributes.end(); ++it) {
+        if (it->size() > 0) {
+            warn(loc, "attribute with arguments not recognized, skipping", "", "");
+            continue;
+        }
+
+        switch (it->name) {
+        case EatSubgroupUniformControlFlow:
+            intermediate.setSubgroupUniformControlFlow();
+            break;
+        default:
+            warn(loc, "attribute does not apply to a function", "", "");
+            break;
+        }
+    }
+}
+
 } // end namespace glslang
 
 #endif // GLSLANG_WEB

glslang/MachineIndependent/attribute.h

@@ -118,7 +118,8 @@ namespace glslang {
         EatFormatR8ui,
         EatFormatUnknown,
         EatNonWritable,
-        EatNonReadable
+        EatNonReadable,
+        EatSubgroupUniformControlFlow,
     };
 
     class TIntermAggregate;

glslang/MachineIndependent/glslang.m4

@@ -944,6 +944,25 @@ function_prototype
         $$.function = $1;
         $$.loc = $2.loc;
     }
+    | function_declarator RIGHT_PAREN attribute {
+        $$.function = $1;
+        $$.loc = $2.loc;
+        parseContext.requireExtensions($2.loc, 1, &E_GL_EXT_subgroup_uniform_control_flow, "attribute");
+        parseContext.handleFunctionAttributes($2.loc, *$3, $$.function);
+    }
+    | attribute function_declarator RIGHT_PAREN {
+        $$.function = $2;
+        $$.loc = $3.loc;
+        parseContext.requireExtensions($3.loc, 1, &E_GL_EXT_subgroup_uniform_control_flow, "attribute");
+        parseContext.handleFunctionAttributes($3.loc, *$1, $$.function);
+    }
+    | attribute function_declarator RIGHT_PAREN attribute {
+        $$.function = $2;
+        $$.loc = $3.loc;
+        parseContext.requireExtensions($3.loc, 1, &E_GL_EXT_subgroup_uniform_control_flow, "attribute");
+        parseContext.handleFunctionAttributes($3.loc, *$1, $$.function);
+        parseContext.handleFunctionAttributes($3.loc, *$4, $$.function);
+    }
     ;
 
 function_declarator
@@ -3713,6 +3732,7 @@ selection_statement
     }
 GLSLANG_WEB_EXCLUDE_ON
     | attribute selection_statement_nonattributed {
+        parseContext.requireExtensions($2->getLoc(), 1, &E_GL_EXT_control_flow_attributes, "attribute");
         parseContext.handleSelectionAttributes(*$1, $2);
         $$ = $2;
     }
@@ -3760,6 +3780,7 @@ switch_statement
     }
 GLSLANG_WEB_EXCLUDE_ON
     | attribute switch_statement_nonattributed {
+        parseContext.requireExtensions($2->getLoc(), 1, &E_GL_EXT_control_flow_attributes, "attribute");
         parseContext.handleSwitchAttributes(*$1, $2);
         $$ = $2;
     }
@@ -3824,6 +3845,7 @@ iteration_statement
     }
 GLSLANG_WEB_EXCLUDE_ON
     | attribute iteration_statement_nonattributed {
+        parseContext.requireExtensions($2->getLoc(), 1, &E_GL_EXT_control_flow_attributes, "attribute");
         parseContext.handleLoopAttributes(*$1, $2);
         $$ = $2;
     }
@@ -4027,7 +4049,6 @@ GLSLANG_WEB_EXCLUDE_ON
 attribute
     : LEFT_BRACKET LEFT_BRACKET attribute_list RIGHT_BRACKET RIGHT_BRACKET {
         $$ = $3;
-        parseContext.requireExtensions($1.loc, 1, &E_GL_EXT_control_flow_attributes, "attribute");
     }
 
 attribute_list

glslang/MachineIndependent/glslang.y

@@ -944,6 +944,25 @@ function_prototype
         $$.function = $1;
         $$.loc = $2.loc;
     }
+    | function_declarator RIGHT_PAREN attribute {
+        $$.function = $1;
+        $$.loc = $2.loc;
+        parseContext.requireExtensions($2.loc, 1, &E_GL_EXT_subgroup_uniform_control_flow, "attribute");
+        parseContext.handleFunctionAttributes($2.loc, *$3, $$.function);
+    }
+    | attribute function_declarator RIGHT_PAREN {
+        $$.function = $2;
+        $$.loc = $3.loc;
+        parseContext.requireExtensions($3.loc, 1, &E_GL_EXT_subgroup_uniform_control_flow, "attribute");
+        parseContext.handleFunctionAttributes($3.loc, *$1, $$.function);
+    }
+    | attribute function_declarator RIGHT_PAREN attribute {
+        $$.function = $2;
+        $$.loc = $3.loc;
+        parseContext.requireExtensions($3.loc, 1, &E_GL_EXT_subgroup_uniform_control_flow, "attribute");
+        parseContext.handleFunctionAttributes($3.loc, *$1, $$.function);
+        parseContext.handleFunctionAttributes($3.loc, *$4, $$.function);
+    }
     ;
 
 function_declarator
@@ -3713,6 +3732,7 @@ selection_statement
     }
 
     | attribute selection_statement_nonattributed {
+        parseContext.requireExtensions($2->getLoc(), 1, &E_GL_EXT_control_flow_attributes, "attribute");
         parseContext.handleSelectionAttributes(*$1, $2);
         $$ = $2;
     }
@@ -3760,6 +3780,7 @@ switch_statement
     }
 
     | attribute switch_statement_nonattributed {
+        parseContext.requireExtensions($2->getLoc(), 1, &E_GL_EXT_control_flow_attributes, "attribute");
         parseContext.handleSwitchAttributes(*$1, $2);
         $$ = $2;
     }
@@ -3824,6 +3845,7 @@ iteration_statement
     }
 
     | attribute iteration_statement_nonattributed {
+        parseContext.requireExtensions($2->getLoc(), 1, &E_GL_EXT_control_flow_attributes, "attribute");
         parseContext.handleLoopAttributes(*$1, $2);
         $$ = $2;
     }
@@ -4027,7 +4049,6 @@ function_definition
 attribute
     : LEFT_BRACKET LEFT_BRACKET attribute_list RIGHT_BRACKET RIGHT_BRACKET {
         $$ = $3;
-        parseContext.requireExtensions($1.loc, 1, &E_GL_EXT_control_flow_attributes, "attribute");
     }
 
 attribute_list

glslang/MachineIndependent/glslang_tab.cpp.h

@@ -1,8 +1,8 @@
-/* A Bison parser, made by GNU Bison 3.7.4.  */
+/* A Bison parser, made by GNU Bison 3.7.5.  */
 
 /* Bison interface for Yacc-like parsers in C
 
-   Copyright (C) 1984, 1989-1990, 2000-2015, 2018-2020 Free Software Foundation,
+   Copyright (C) 1984, 1989-1990, 2000-2015, 2018-2021 Free Software Foundation,
    Inc.
 
    This program is free software: you can redistribute it and/or modify

glslang/MachineIndependent/intermOut.cpp

@@ -1487,6 +1487,9 @@ void TIntermediate::output(TInfoSink& infoSink, bool tree)
     if (xfbMode)
         infoSink.debug << "in xfb mode\n";
 
+    if (getSubgroupUniformControlFlow())
+        infoSink.debug << "subgroup_uniform_control_flow\n";
+
     switch (language) {
     case EShLangVertex:
         break;

glslang/MachineIndependent/localintermediate.h

@@ -328,6 +328,7 @@ public:
         textureSamplerTransformMode(EShTexSampTransKeep),
         needToLegalize(false),
         binaryDoubleOutput(false),
+        subgroupUniformControlFlow(false),
         usePhysicalStorageBuffer(false),
         uniformLocationBase(0)
 #endif
@@ -864,6 +865,9 @@ public:
 
     void setBinaryDoubleOutput() { binaryDoubleOutput = true; }
     bool getBinaryDoubleOutput() { return binaryDoubleOutput; }
+
+    void setSubgroupUniformControlFlow() { subgroupUniformControlFlow = true; }
+    bool getSubgroupUniformControlFlow() const { return subgroupUniformControlFlow; }
 #endif // GLSLANG_WEB
 
     void addBlockStorageOverride(const char* nameStr, TBlockStorageClass backing)
@@ -1115,6 +1119,7 @@ protected:
 
     bool needToLegalize;
     bool binaryDoubleOutput;
+    bool subgroupUniformControlFlow;
     bool usePhysicalStorageBuffer;
 
     std::unordered_map<std::string, int> uniformLocationOverrides;

gtests/Spv.FromFile.cpp

@@ -443,6 +443,7 @@ INSTANTIATE_TEST_SUITE_P(
         "spv.specConstant.int8.comp",
         "spv.storageBuffer.vert",
         "spv.terminate.frag",
+        "spv.subgroupUniformControlFlow.vert",
         "spv.precise.tese",
         "spv.precise.tesc",
         "spv.viewportindex.tese",