SWDEV-477324 - Graph Capture memcpy D2D

Change-Id: Ifaa4d78854c03b3150233142df187c9bbf731cab


[ROCm/clr commit: e98179d924]

projects/clr/hipamd/src/hip_graph.cpp

@@ -1202,8 +1202,16 @@ hipError_t hipGraphExecMemcpyNodeSetParams1D(hipGraphExec_t hGraphExec, hipGraph
   if (oldkind != kind) {
     HIP_RETURN(hipErrorInvalidValue);
   }
-  HIP_RETURN(reinterpret_cast<hip::GraphMemcpyNode1D*>(clonedNode)->SetParams(dst, src,
+  hipError_t status = reinterpret_cast<hip::GraphMemcpyNode1D*>(clonedNode)->SetParams(dst, src,
-                                                                              count, kind));
+                                                                              count, kind);
+  if (status != hipSuccess) {
+    HIP_RETURN(status);
+  }
+  if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
+    status = reinterpret_cast<hip::GraphExec*>(hGraphExec)
+                 ->UpdateAQLPacket(reinterpret_cast<hip::GraphKernelNode*>(clonedNode));
+  }
+  HIP_RETURN(status);
 }
 
 hipError_t hipGraphAddMemsetNode(hipGraphNode_t* pGraphNode, hipGraph_t graph,
@@ -1612,7 +1620,15 @@ hipError_t hipGraphExecMemcpyNodeSetParams(hipGraphExec_t hGraphExec, hipGraphNo
   if (oldkind != newkind) {
     HIP_RETURN(hipErrorInvalidValue);
   }
-  HIP_RETURN(reinterpret_cast<hip::GraphMemcpyNode*>(clonedNode)->SetParams(pNodeParams));
+  hipError_t status = reinterpret_cast<hip::GraphMemcpyNode*>(clonedNode)->SetParams(pNodeParams);
+  if (status != hipSuccess) {
+    HIP_RETURN(status);
+  }
+  if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
+    status = reinterpret_cast<hip::GraphExec*>(hGraphExec)
+                 ->UpdateAQLPacket(reinterpret_cast<hip::GraphKernelNode*>(clonedNode));
+  }
+  HIP_RETURN(status);
 }
 
 hipError_t hipGraphMemsetNodeGetParams(hipGraphNode_t node, hipMemsetParams* pNodeParams) {
@@ -2181,8 +2197,16 @@ hipError_t hipGraphExecMemcpyNodeSetParamsFromSymbol(hipGraphExec_t hGraphExec,
     HIP_RETURN(hipErrorInvalidValue);
   }
   constexpr bool kCheckDeviceIsSame = true;
-  HIP_RETURN(reinterpret_cast<hip::GraphMemcpyNodeFromSymbol*>(clonedNode)
+  hipError_t status = reinterpret_cast<hip::GraphMemcpyNodeFromSymbol*>(clonedNode)
-                 ->SetParams(dst, symbol, count, offset, kind, kCheckDeviceIsSame));
+               ->SetParams(dst, symbol, count, offset, kind, kCheckDeviceIsSame);
+  if (status != hipSuccess) {
+    HIP_RETURN(status);
+  }
+  if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
+    status = reinterpret_cast<hip::GraphExec*>(hGraphExec)
+                 ->UpdateAQLPacket(reinterpret_cast<hip::GraphKernelNode*>(clonedNode));
+  }
+  HIP_RETURN(status);
 }
 
 hipError_t hipGraphAddMemcpyNodeToSymbol(hipGraphNode_t* pGraphNode, hipGraph_t graph,
@@ -2255,8 +2279,16 @@ hipError_t hipGraphExecMemcpyNodeSetParamsToSymbol(hipGraphExec_t hGraphExec, hi
     HIP_RETURN(hipErrorInvalidValue);
   }
   constexpr bool kCheckDeviceIsSame = true;
-  HIP_RETURN(reinterpret_cast<hip::GraphMemcpyNodeToSymbol*>(clonedNode)
+  hipError_t status =  reinterpret_cast<hip::GraphMemcpyNodeToSymbol*>(clonedNode)
-                 ->SetParams(symbol, src, count, offset, kind, kCheckDeviceIsSame));
+                 ->SetParams(symbol, src, count, offset, kind, kCheckDeviceIsSame);
+  if (status != hipSuccess) {
+    HIP_RETURN(status);
+  }
+  if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
+    status = reinterpret_cast<hip::GraphExec*>(hGraphExec)
+                 ->UpdateAQLPacket(reinterpret_cast<hip::GraphKernelNode*>(clonedNode));
+  }
+  HIP_RETURN(status);
 }
 
 hipError_t hipGraphAddEventRecordNode(hipGraphNode_t* pGraphNode, hipGraph_t graph,

projects/clr/hipamd/src/hip_graph_internal.hpp

@@ -212,7 +212,7 @@ struct GraphNode : public hipGraphNodeDOTAttribute {
   std::vector<uint8_t *> gpuPackets_; //!< GPU Packet to enqueue during graph launch
   std::string capturedKernelName_;
   size_t alignedKernArgSize_ = 256;       //!< Aligned size required for kernel args
-  size_t kernargSegmentByteSize_ = 256;   //!< Kernel arg segment byte size
+  size_t kernargSegmentByteSize_ = 512;   //!< Kernel arg segment byte size
   size_t kernargSegmentAlignment_ = 256;  //!< Kernel arg segment alignment
 
  public:
@@ -434,13 +434,11 @@ struct GraphNode : public hipGraphNodeDOTAttribute {
   unsigned int GetEnabled() const { return isEnabled_; }
   void SetEnabled(unsigned int isEnabled) { isEnabled_ = isEnabled; }
   // Returns true if capture is enabled for the current node.
-  bool GraphCaptureEnabled() {
+  virtual bool GraphCaptureEnabled() {
     bool isGraphCapture = false;
     if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
       switch (GetType()) {
         case hipGraphNodeTypeKernel:
-          isGraphCapture = true;
-          break;
         case hipGraphNodeTypeMemset:
           isGraphCapture = true;
           break;
@@ -1525,6 +1523,19 @@ class GraphMemcpyNode : public GraphNode {
       return shape_;
     }
   }
+  virtual bool GraphCaptureEnabled() override {
+    bool isGraphCapture = false;
+    if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
+      switch (copyParams_.kind) {
+        case hipMemcpyDeviceToDevice:
+          isGraphCapture = true;
+          break;
+        default:
+          break;
+      }
+    }
+    return isGraphCapture;
+  }
 };
 
 class GraphMemcpyNode1D : public GraphMemcpyNode {
@@ -1704,6 +1715,19 @@ class GraphMemcpyNode1D : public GraphMemcpyNode {
       return shape_;
     }
   }
+  virtual bool GraphCaptureEnabled() override {
+    bool isGraphCapture = false;
+    if (DEBUG_CLR_GRAPH_PACKET_CAPTURE) {
+      switch (kind_) {
+        case hipMemcpyDeviceToDevice:
+          isGraphCapture = true;
+          break;
+        default:
+          break;
+      }
+    }
+    return isGraphCapture;
+  }
 };
 
 class GraphMemcpyNodeFromSymbol : public GraphMemcpyNode1D {