Handle difference between MS and CS/AS and add a simple test

Author: Jack Elliott

include/dxc/DxilContainer/DxilPipelineStateValidation.h

@@ -176,7 +176,7 @@ struct PSVRuntimeInfo3 : public PSVRuntimeInfo2 {
 };
 
 struct PSVRuntimeInfo4 : public PSVRuntimeInfo3 {
-  uint32_t GroupSharedMemoryLimit;
+  uint32_t GroupSharedLimit;
 };
 
 enum class PSVResourceType {

lib/DxilContainer/DxilPipelineStateValidation.cpp

@@ -313,7 +313,7 @@ void hlsl::SetShaderProps(PSVRuntimeInfo4 *pInfo4, const DxilModule &DM) {
     case ShaderModel::Kind::Compute:
     case ShaderModel::Kind::Mesh:
     case ShaderModel::Kind::Amplification:
-      pInfo4->GroupSharedMemoryLimit = DM.GetGroupSharedLimit();
+      pInfo4->GroupSharedLimit = DM.GetGroupSharedLimit();
         break;
     default:
         break;
@@ -825,8 +825,8 @@ void hlsl::PrintPSVRuntimeInfo(llvm::raw_ostream &OS, PSVRuntimeInfo0 *pInfo0,
          << pInfo2->NumThreadsY << "," << pInfo2->NumThreadsZ << ")\n";
     }
     if (pInfo4) {
-      OS << Comment << " GroupSharedMemoryLimit="
-         << pInfo4->GroupSharedMemoryLimit << "\n";
+      OS << Comment << " GroupSharedLimit="
+         << pInfo4->GroupSharedLimit << "\n";
     }
     break;
   case PSVShaderKind::Amplification:
@@ -836,8 +836,8 @@ void hlsl::PrintPSVRuntimeInfo(llvm::raw_ostream &OS, PSVRuntimeInfo0 *pInfo0,
          << pInfo2->NumThreadsY << "," << pInfo2->NumThreadsZ << ")\n";
     }
     if (pInfo4) {
-      OS << Comment << " GroupSharedMemoryLimit="
-         << pInfo4->GroupSharedMemoryLimit << "\n";
+      OS << Comment << " GroupSharedLimit="
+         << pInfo4->GroupSharedLimit << "\n";
     }
     break;
   case PSVShaderKind::Mesh:
@@ -866,8 +866,8 @@ void hlsl::PrintPSVRuntimeInfo(llvm::raw_ostream &OS, PSVRuntimeInfo0 *pInfo0,
          << pInfo2->NumThreadsY << "," << pInfo2->NumThreadsZ << ")\n";
     }
     if (pInfo4) {
-      OS << Comment << " GroupSharedMemoryLimit="
-         << pInfo4->GroupSharedMemoryLimit << "\n";
+      OS << Comment << " GroupSharedLimit="
+         << pInfo4->GroupSharedLimit << "\n";
     }
     break;
   case PSVShaderKind::Library:

tools/clang/lib/CodeGen/CGHLSLMS.cpp

@@ -1647,8 +1647,6 @@ void CGMSHLSLRuntime::AddHLSLFunctionInfo(Function *F, const FunctionDecl *FD) {
   }
 
   if (const HLSLGroupSharedLimitAttr *Attr = FD->getAttr<HLSLGroupSharedLimitAttr>()) {
-    funcProps->groupSharedLimitBytes = Attr->getLimit();
-
     if (isEntry && !SM->IsCS() && !SM->IsMS() && !SM->IsAS()) {
       unsigned DiagID = Diags.getCustomDiagID(
           DiagnosticsEngine::Error,
@@ -1657,8 +1655,24 @@ void CGMSHLSLRuntime::AddHLSLFunctionInfo(Function *F, const FunctionDecl *FD) {
       return;
     }
 
+    // Only valid for SM6.10+
+    if (!SM->IsSM610Plus()) {
+      unsigned DiagID = Diags.getCustomDiagID(
+          DiagnosticsEngine::Error,
+          "attribute GroupSharedLimit only valid for Shader Model 6.10 and above.");
+      Diags.Report(Attr->getLocation(), DiagID);
+      return;
+    }
+
+    funcProps->groupSharedLimitBytes = Attr->getLimit();
   } else {
-    funcProps->groupSharedLimitBytes = DXIL::kMaxTGSMSize;
+    if (SM->IsMS()) { // Fallback to default limits
+        funcProps->groupSharedLimitBytes = DXIL::kMaxMSSMSize; // 28k For MS
+      } else if (SM->IsAS() || SM->IsCS()) {
+        funcProps->groupSharedLimitBytes = DXIL::kMaxTGSMSize; // 32k For AS/CS
+      } else {
+        funcProps->groupSharedLimitBytes = 0;
+      }
   }
 
   // Hull shader.

tools/clang/test/HLSLFileCheck/hlsl/entry/attributes/GroupSharedLimit.hlsl

@@ -0,0 +1,85 @@
+// RUN: %dxc -E MainPass -T cs_6_10 %s | FileCheck %s
+
+#define NUM_BYTES_OF_SHARED_MEM (32*1024)
+#define NUM_DWORDS_SHARED_MEM (NUM_BYTES_OF_SHARED_MEM / 4)
+#define THREAD_GROUP_SIZE_X 1024
+
+groupshared uint g_testBufferPASS[NUM_DWORDS_SHARED_MEM];
+
+RWStructuredBuffer <uint> g_output : register(u0);
+
+// CHECK: @MainPass
+
+[GroupSharedLimit(NUM_BYTES_OF_SHARED_MEM)]
+[numthreads(THREAD_GROUP_SIZE_X, 1, 1)]
+void MainPass( uint3 DTid : SV_DispatchThreadID )
+{
+    uint iterations = NUM_DWORDS_SHARED_MEM / THREAD_GROUP_SIZE_X;
+    
+    for (uint i = 0; i < iterations; i++)
+    {
+        uint index = DTid.x + i * THREAD_GROUP_SIZE_X;
+        g_testBufferPASS[index] = index;
+    }
+    
+    // synchronize all threads in the group
+    GroupMemoryBarrierWithGroupSync();
+
+    // write the shared data to the output buffer
+    for (uint i = 0; i < iterations; i++)
+    {
+        uint index = DTid.x + i * THREAD_GROUP_SIZE_X;
+        g_output[index] = g_testBufferPASS[index];
+    }
+}
+
+// RUN: not %dxc -E MainFail -T cs_6_10 %s 2>&1 | FileCheck %s --check-prefix=CHECK-ERROR
+// CHECK-ERROR: Total Thread Group Shared Memory storage is 32772, exceeded 32768.
+
+groupshared uint g_testBufferFAIL[NUM_DWORDS_SHARED_MEM + 1];
+
+[GroupSharedLimit(NUM_BYTES_OF_SHARED_MEM)]
+[numthreads(THREAD_GROUP_SIZE_X, 1, 1)]
+void MainFail( uint3 DTid : SV_DispatchThreadID )
+{
+    uint iterations = NUM_DWORDS_SHARED_MEM / THREAD_GROUP_SIZE_X;
+    
+    for (uint i = 0; i < iterations; i++)
+    {
+        uint index = DTid.x + i * THREAD_GROUP_SIZE_X;
+        g_testBufferFAIL[index] = index;
+    }
+    
+    // synchronize all threads in the group
+    GroupMemoryBarrierWithGroupSync();
+
+    // write the shared data to the output buffer
+    for (uint i = 0; i < iterations; i++)
+    {
+        uint index = DTid.x + i * THREAD_GROUP_SIZE_X;
+        g_output[index] = g_testBufferFAIL[index];
+    }
+}  
+
+// RUN: not %dxc -E MainFail2 -T cs_6_10 %s 2>&1 | FileCheck %s --check-prefix=CHECK-ERROR
+[numthreads(THREAD_GROUP_SIZE_X, 1, 1)]
+void MainFail2( uint3 DTid : SV_DispatchThreadID )
+{
+    uint iterations = NUM_DWORDS_SHARED_MEM / THREAD_GROUP_SIZE_X;
+    
+    for (uint i = 0; i < iterations; i++)
+    {
+        uint index = DTid.x + i * THREAD_GROUP_SIZE_X;
+        g_testBufferFAIL[index] = index;
+    }
+    
+    // synchronize all threads in the group
+    GroupMemoryBarrierWithGroupSync();
+
+    // write the shared data to the output buffer
+    for (uint i = 0; i < iterations; i++)
+    {
+        uint index = DTid.x + i * THREAD_GROUP_SIZE_X;
+        g_output[index] = g_testBufferFAIL[index];
+    }
+}