Determine threadscope for DXBC and DXIL shaders

Author: baldurk

renderdoc/driver/shaders/dxbc/dxbc_bytecode.h

@@ -1126,6 +1126,7 @@ class Program
   uint32_t GetMajorVersion() const { return m_Major; }
   uint32_t GetMinorVersion() const { return m_Minor; }
   bool IsShaderModel51() const { return m_Major == 5 && m_Minor == 1; }
+  DXBC::ThreadScope GetThreadScope() const { return m_Threadscope; }
   D3D_PRIMITIVE_TOPOLOGY GetOutputTopology();
   const rdcstr &GetDisassembly()
   {
@@ -1163,6 +1164,7 @@ class Program
 
   DXBC::ShaderType m_Type = DXBC::ShaderType::Max;
   uint32_t m_Major = 0, m_Minor = 0;
+  DXBC::ThreadScope m_Threadscope = DXBC::ThreadScope::Thread;
 
   rdcarray<uint32_t> m_ProgramWords;
 

renderdoc/driver/shaders/dxbc/dxbc_bytecode_ops.cpp

@@ -782,6 +782,10 @@ void Program::DecodeProgram()
         else
           m_LateDeclarations.back().push_back(decl);
       }
+
+      if(decl.declaration == OPCODE_DCL_THREAD_GROUP_SHARED_MEMORY_RAW ||
+         decl.declaration == OPCODE_DCL_THREAD_GROUP_SHARED_MEMORY_STRUCTURED)
+        m_Threadscope |= DXBC::ThreadScope::Workgroup;
     }
     else
     {
@@ -791,6 +795,9 @@ void Program::DecodeProgram()
       if(op.operation == OPCODE_HS_CONTROL_POINT_PHASE || op.operation == OPCODE_HS_FORK_PHASE ||
          op.operation == OPCODE_HS_JOIN_PHASE)
         m_LateDeclarations.push_back({});
+
+      if(decl.declaration == OPCODE_SYNC)
+        m_Threadscope |= DXBC::ThreadScope::Workgroup;
     }
   }
 

renderdoc/driver/shaders/dxbc/dxbc_common.h

@@ -85,6 +85,16 @@ enum class ShaderType : uint8_t
 
 ShaderStage GetShaderStage(ShaderType type);
 
+enum class ThreadScope : uint32_t
+{
+  Thread = 0,
+  Quad = 0x1,
+  Subgroup = 0x2,
+  Workgroup = 0x4,
+};
+
+BITMASK_OPERATORS(ThreadScope);
+
 /////////////////////////////////////////////////////////////////////////
 // the below classes basically mimics the existing reflection interface.
 //

renderdoc/driver/shaders/dxbc/dxbc_container.cpp

@@ -661,7 +661,7 @@ const rdcstr &DXBCContainer::GetDisassembly(bool dxcStyle)
         globalFlagsString += commentString + "       Resource descriptor heap indexing\n";
       if(m_GlobalFlags & GlobalShaderFlags::SamplerDescriptorHeapIndexing)
         globalFlagsString += commentString + "       Sampler descriptor heap indexing\n";
-      if(m_GlobalFlags & GlobalShaderFlags::Reserved)
+      if(m_GlobalFlags & GlobalShaderFlags::WaveMatrix)
         globalFlagsString += commentString + "       Wave Matrix\n";
       if(m_GlobalFlags & GlobalShaderFlags::AtomicInt64OnHeapResource)
         globalFlagsString += commentString + "       64-bit Atomics on Heap Resources\n";
@@ -697,6 +697,7 @@ const rdcstr &DXBCContainer::GetDisassembly(bool dxcStyle)
         m_Disassembly += "// Vendor shader extensions in use\n";
 
       m_Disassembly += m_DXBCByteCode->GetDisassembly();
+      m_Threadscope = m_DXBCByteCode->GetThreadScope();
     }
     else if(m_DXILByteCode)
     {
@@ -720,7 +721,14 @@ const rdcstr &DXBCContainer::GetDisassembly(bool dxcStyle)
 #endif
 
       m_Disassembly += m_DXILByteCode->GetDisassembly(dxcStyle, m_Reflection);
+      m_Threadscope = m_DXILByteCode->GetThreadScope();
     }
+
+    if(m_Type == DXBC::ShaderType::Pixel)
+      m_Threadscope |= ThreadScope::Quad;
+
+    if(m_GlobalFlags & GlobalShaderFlags::WaveOps)
+      m_Threadscope |= ThreadScope::Subgroup;
   }
 
   return m_Disassembly;

renderdoc/driver/shaders/dxbc/dxbc_container.h

@@ -138,7 +138,7 @@ enum class GlobalShaderFlags : int64_t
   DerivativesInMeshAndAmpShaders = 0x1000000,
   ResourceDescriptorHeapIndexing = 0x2000000,
   SamplerDescriptorHeapIndexing = 0x4000000,
-  Reserved = 0x8000000,
+  WaveMatrix = 0x8000000,
   AtomicInt64OnHeapResource = 0x10000000,
   AdvancedTextureOps = 0x20000000,
   WriteableMSAATextures = 0x40000000,
@@ -207,6 +207,11 @@ class DXBCContainer
   const IDebugInfo *GetDebugInfo() const { return m_DebugInfo; }
   const Reflection *GetReflection() const { return m_Reflection; }
   D3D_PRIMITIVE_TOPOLOGY GetOutputTopology();
+  ThreadScope GetThreadScope()
+  {
+    GetDisassembly(false);
+    return m_Threadscope;
+  }
 
   CBufferVariableType GetRayPayload(const ShaderEntryPoint &entry)
   {
@@ -306,6 +311,7 @@ class DXBCContainer
 
   rdcflatmap<ShaderEntryPoint, rdcpair<CBufferVariableType, CBufferVariableType>> m_RayPayloads;
 
+  ThreadScope m_Threadscope = ThreadScope::Thread;
   ShaderStatistics m_ShaderStats;
   DXBCBytecode::Program *m_DXBCByteCode = NULL;
   DXIL::Program *m_DXILByteCode = NULL;

renderdoc/driver/shaders/dxil/dxil_bytecode.h

@@ -1617,6 +1617,8 @@ class Program : public DXBC::IDebugInfo
   void FetchEntryPoint();
   DXBC::Reflection *BuildReflection();
 
+  DXBC::ThreadScope GetThreadScope() const { return m_Threadscope; }
+
   rdcstr GetDefaultCommandLine() const { return "-T " + m_Profile; }
 
   rdcstr GetDebugStatus();
@@ -1717,6 +1719,7 @@ class Program : public DXBC::IDebugInfo
 
   rdcstr m_CompilerSig, m_EntryPoint, m_Profile;
   ShaderCompileFlags m_CompileFlags;
+  DXBC::ThreadScope m_Threadscope = DXBC::ThreadScope::Thread;
 
   const Type *m_CurParseType = NULL;
 

renderdoc/driver/shaders/dxil/dxil_disassemble.cpp

@@ -1158,6 +1158,23 @@ void Program::Parse(const DXBC::Reflection *reflection)
   m_SsaAliases.clear();
   ParseReferences(reflection);
 
+  if(m_Type == DXBC::ShaderType::Compute || m_Type == DXBC::ShaderType::Amplification ||
+     m_Type == DXBC::ShaderType::Mesh)
+  {
+    for(GlobalVar *g : m_GlobalVars)
+    {
+      RDCASSERT(g->type->type == Type::Pointer);
+      if(g->type->type == Type::Pointer && g->type->addrSpace == Type::PointerAddrSpace::GroupShared)
+        m_Threadscope |= DXBC::ThreadScope::Workgroup;
+    }
+
+    for(Function *f : m_Functions)
+    {
+      if(f->name == "dx.op.barrier")
+        m_Threadscope |= DXBC::ThreadScope::Workgroup;
+    }
+  }
+
   m_Parsed = true;
 }
 

renderdoc/driver/shaders/spirv/spirv_processor.h

@@ -539,6 +539,7 @@ struct Section
 
 enum class ThreadScope : uint32_t
 {
+  Thread = 0,
   Quad = 0x1,
   Subgroup = 0x2,
   Workgroup = 0x4,
@@ -637,7 +638,7 @@ class Processor
 
   LogicalSection m_Sections[Section::Count];
 
-  ThreadScope m_ThreadScope;
+  ThreadScope m_ThreadScope = ThreadScope::Thread;
 
 private:
   struct DeferredMemberDecoration