P4 to Git Change 1310350 by lmoriche@lmoriche_opencl_dev on 2016/09/02 12:29:12

SWDEV-101853 - Use the PointeeAlignment metadata to align the dynamic LDS allocations at dispatch.

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/amdgpu_metadata.cpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/amdgpu_metadata.hpp#3 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rockernel.cpp#11 edit

rocclr/runtime/device/rocm/amdgpu_metadata.cpp

@@ -124,7 +124,7 @@ namespace code {
   namespace KernelArg {
     using namespace AMDGPU::RuntimeMD::KernelArg;
     Metadata::Metadata()
-    : size(0), align(0),
+    : size(0), align(0), pointeeAlign(0),
       isConst(false), isRestrict(false), isVolatile(false), isPipe(false)
     {}
 
@@ -177,6 +177,7 @@ namespace code {
       case KeyArgName: return Read(in, name);
       case KeyArgTypeKind: return Read(in, typeKind);
       case KeyArgValueType: return Read(in, valueType);
+      case KeyArgPointeeAlign: return Read(in, pointeeAlign);
       case KeyArgAddrQual: return Read(in, addrQual);
       case KeyArgAccQual: return Read(in, accQual);
       case KeyArgIsConst: isConst = true; return true;
@@ -261,6 +262,7 @@ namespace code {
       case KeyArgName:
       case KeyArgTypeKind:
       case KeyArgValueType:
+      case KeyArgPointeeAlign:
       case KeyArgAddrQual:
       case KeyArgAccQual:
       case KeyArgIsConst:
@@ -424,6 +426,7 @@ namespace code {
         case KeyArgName:
         case KeyArgTypeKind:
         case KeyArgValueType:
+        case KeyArgPointeeAlign:
         case KeyArgAddrQual:
         case KeyArgAccQual:
         case KeyArgIsConst:

rocclr/runtime/device/rocm/amdgpu_metadata.hpp

@@ -61,6 +61,7 @@ namespace code {
     private:
       uint32_t size;
       uint32_t align;
+      uint32_t pointeeAlign;
       std::string typeName;
       std::string name;
       AMDGPU::RuntimeMD::KernelArg::TypeKind typeKind;
@@ -73,6 +74,7 @@ namespace code {
       Metadata();
       uint32_t Size() const { return size; }
       uint32_t Align() const { return align; }
+      uint32_t PointeeAlign() const { return pointeeAlign; }
       const std::string& TypeName() const { return typeName; }
       const std::string& Name() const { return name; }
       AMDGPU::RuntimeMD::KernelArg::TypeKind TypeKind() const { return typeKind; }

rocclr/runtime/device/rocm/rockernel.cpp

@@ -56,9 +56,17 @@ GetHSAILArgType(const aclArgData* argInfo)
 inline static size_t
 GetHSAILArgAlignment(const amd::hsa::code::KernelArg::Metadata* lcArg)
 {
-    if (lcArg->TypeKind() == AMDGPU::RuntimeMD::KernelArg::Pointer)
+    if (lcArg->TypeKind() == AMDGPU::RuntimeMD::KernelArg::Pointer) {
+        if (lcArg->AddrQual() == AMDGPU::RuntimeMD::KernelArg::Local) {
+            uint32_t align = lcArg->PointeeAlign();
+            if (align == 0) {
+                LogWarning("Missing dynamic_shared_pointer alignment");
+                align = 128; /* worst case alignment */;
+            }
+            return align;
+        }
         return lcArg->Align();
-
+    }
     return 1;
 }
 #endif // defined(WITH_LIGHTNING_COMPILER)