[dxbc] Use ballot to determine early-discard condition in fragment shaders

src/dxbc/dxbc_compiler.cpp

@@ -3898,16 +3898,20 @@ namespace dxvk {
       m_module.opStore(m_ps.killState, killState);
 
       if (m_moduleInfo.options.useSubgroupOpsForEarlyDiscard) {
-        uint32_t killSubgroup = m_module.opGroupNonUniformLogicalAnd(
+        uint32_t ballot = m_module.opGroupNonUniformBallot(
-          m_module.defBoolType(),
+          getVectorTypeId({ DxbcScalarType::Uint32, 4 }),
           m_module.constu32(spv::ScopeSubgroup),
-          m_moduleInfo.options.useSubgroupOpsClustered
+          killState);
-            ? spv::GroupOperationClusteredReduce
+        
-            : spv::GroupOperationReduce,
+        uint32_t invocationMask = m_module.opLoad(
-          killState,
+          getVectorTypeId({ DxbcScalarType::Uint32, 4 }),
-          m_moduleInfo.options.useSubgroupOpsClustered
+          m_ps.invocationMask);
-            ? m_module.constu32(4)
+        
-            : 0);
+        uint32_t killSubgroup = m_module.opAll(
+          m_module.defBoolType(),
+          m_module.opIEqual(
+            m_module.defVectorType(m_module.defBoolType(), 4),
+            ballot, invocationMask));
         
         DxbcConditional cond;
         cond.labelIf  = m_module.allocateId();
@@ -6297,6 +6301,17 @@ namespace dxvk {
       spv::BuiltInCullDistance,
       spv::StorageClassInput);
     
+    // Main function of the pixel shader
+    m_ps.functionId = m_module.allocateId();
+    m_module.setDebugName(m_ps.functionId, "ps_main");
+    
+    this->emitFunctionBegin(
+      m_ps.functionId,
+      m_module.defVoidType(),
+      m_module.defFunctionType(
+        m_module.defVoidType(), 0, nullptr));
+    this->emitFunctionLabel();
+
     // We may have to defer kill operations to the end of
     // the shader in order to keep derivatives correct.
     if (m_analysis->usesKill && m_analysis->usesDerivatives) {
@@ -6308,23 +6323,22 @@ namespace dxvk {
 
       if (m_moduleInfo.options.useSubgroupOpsForEarlyDiscard) {
         m_module.enableCapability(spv::CapabilityGroupNonUniform);
-        m_module.enableCapability(spv::CapabilityGroupNonUniformArithmetic);
+        m_module.enableCapability(spv::CapabilityGroupNonUniformBallot);
 
-        if (m_moduleInfo.options.useSubgroupOpsClustered)
+        DxbcRegisterInfo invocationMask;
-          m_module.enableCapability(spv::CapabilityGroupNonUniformClustered);
+        invocationMask.type = { DxbcScalarType::Uint32, 4, 0 };
+        invocationMask.sclass = spv::StorageClassFunction;
+
+        m_ps.invocationMask = emitNewVariable(invocationMask);
+        m_module.setDebugName(m_ps.invocationMask, "fInvocationMask");
+        
+        m_module.opStore(m_ps.invocationMask,
+          m_module.opGroupNonUniformBallot(
+            getVectorTypeId({ DxbcScalarType::Uint32, 4 }),
+            m_module.constu32(spv::ScopeSubgroup),
+            m_module.constBool(true)));
       }
     }
-    
-    // Main function of the pixel shader
-    m_ps.functionId = m_module.allocateId();
-    m_module.setDebugName(m_ps.functionId, "ps_main");
-    
-    this->emitFunctionBegin(
-      m_ps.functionId,
-      m_module.defVoidType(),
-      m_module.defFunctionType(
-        m_module.defVoidType(), 0, nullptr));
-    this->emitFunctionLabel();
   }
   
   

src/dxbc/dxbc_compiler.h

@@ -178,6 +178,7 @@ namespace dxvk {
     uint32_t builtinLayer         = 0;
     uint32_t builtinViewportId    = 0;
     
+    uint32_t invocationMask       = 0;
     uint32_t killState            = 0;
   };
   

src/dxbc/dxbc_options.cpp

@@ -18,9 +18,7 @@ namespace dxvk {
     useSubgroupOpsForEarlyDiscard
       = (devInfo.coreSubgroup.subgroupSize >= 4)
      && (devInfo.coreSubgroup.supportedStages     & VK_SHADER_STAGE_FRAGMENT_BIT)
-     && (devInfo.coreSubgroup.supportedOperations & VK_SUBGROUP_FEATURE_ARITHMETIC_BIT);
+     && (devInfo.coreSubgroup.supportedOperations & VK_SUBGROUP_FEATURE_BALLOT_BIT);
-    useSubgroupOpsClustered = useSubgroupOpsForEarlyDiscard
-     && (devInfo.coreSubgroup.supportedOperations & VK_SUBGROUP_FEATURE_CLUSTERED_BIT);
     
     zeroInitWorkgroupMemory = options.zeroInitWorkgroupMemory;
     
@@ -29,10 +27,8 @@ namespace dxvk {
     auto vendor = DxvkGpuVendor(devInfo.core.properties.vendorID);
 
     if (vendor == DxvkGpuVendor::Amd
-     || vendor == DxvkGpuVendor::Nvidia) {
+     || vendor == DxvkGpuVendor::Nvidia)
       useSubgroupOpsForEarlyDiscard = false;
-      useSubgroupOpsClustered       = false;
-    }
   }
   
 }

src/dxbc/dxbc_options.h

@@ -17,9 +17,6 @@ namespace dxvk {
     /// shader invocations if derivatives remain valid.
     bool useSubgroupOpsForEarlyDiscard = false;
 
-    /// Use clustered subgroup operations
-    bool useSubgroupOpsClustered = false;
-
     /// Clear thread-group shared memory to zero
     bool zeroInitWorkgroupMemory = false;
   };

tests/dxbc/test_dxbc_compiler.cpp

@@ -43,7 +43,6 @@ int WINAPI WinMain(HINSTANCE hInstance,
     
     DxbcModuleInfo moduleInfo;
     moduleInfo.options.useSubgroupOpsForEarlyDiscard = true;
-    moduleInfo.options.useSubgroupOpsClustered       = true;
     moduleInfo.xfb = nullptr;
 
     Rc<DxvkShader> shader = module.compile(moduleInfo, ifileName);