From 2bd18360498b66a87d0edb9d3e1f4097fff02ae2 Mon Sep 17 00:00:00 2001
From: Alex Xie <AlexBin.Xie@amd.com>
Date: Fri, 26 Mar 2021 03:31:59 -0400
Subject: [PATCH] SWDEV-276304 - [Navi][OpenCL][Linux]AMF Converter generate
 corrupted Frames

The copy image workaround could be recursively used by ROCclr blit kernel.
Avoid such situation by using stack variable.

Change-Id: Iadaa8cad9216220194760dd461a9533bb236aea0
---
 rocclr/device/rocm/rocvirtual.cpp | 20 +++++++++++---------
 rocclr/device/rocm/rocvirtual.hpp |  7 +++----
 2 files changed, 14 insertions(+), 13 deletions(-)

diff --git a/rocclr/device/rocm/rocvirtual.cpp b/rocclr/device/rocm/rocvirtual.cpp
index 06e3566917..00aa67fafd 100644
--- a/rocclr/device/rocm/rocvirtual.cpp
+++ b/rocclr/device/rocm/rocvirtual.cpp
@@ -433,7 +433,8 @@ void VirtualGPU::HwQueueTracker::ResetCurrentSignal() {
 
 // ================================================================================================
 bool VirtualGPU::processMemObjects(const amd::Kernel& kernel, const_address params,
-  size_t& ldsAddress, bool cooperativeGroups) {
+  size_t& ldsAddress, bool cooperativeGroups, bool& imageBufferWrtBack,
+  std::vector<device::Memory*>& wrtBackImageBuffer) {
   Kernel& hsaKernel = const_cast<Kernel&>(static_cast<const Kernel&>(*(kernel.getDeviceKernel(dev()))));
   const amd::KernelSignature& signature = kernel.signature();
   const amd::KernelParameters& kernelParams = kernel.parameters();
@@ -614,8 +615,8 @@ bool VirtualGPU::processMemObjects(const amd::Kernel& kernel, const_address para
 
               // If it's not a read only resource, then runtime has to write back
               if (!desc.info_.readOnly_) {
-                wrtBackImageBuffer_.push_back(mem->getDeviceMemory(dev()));
-                imageBufferWrtBack_ = true;
+                wrtBackImageBuffer.push_back(mem->getDeviceMemory(dev()));
+                imageBufferWrtBack = true;
               }
             }
           }
@@ -2451,10 +2452,14 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
   device::Kernel* devKernel = const_cast<device::Kernel*>(kernel.getDeviceKernel(dev()));
   Kernel& gpuKernel = static_cast<Kernel&>(*devKernel);
   size_t ldsUsage = gpuKernel.WorkgroupGroupSegmentByteSize();
+  bool imageBufferWrtBack = false; // Image buffer write back is required
+  std::vector<device::Memory*> wrtBackImageBuffer; // Array of images for write back
+
   setLastCommandSDMA(false);
   // Check memory dependency and SVM objects
   bool coopGroups = (vcmd != nullptr) ? vcmd->cooperativeGroups() : false;
-  if (!processMemObjects(kernel, parameters, ldsUsage, coopGroups)) {
+  if (!processMemObjects(kernel, parameters, ldsUsage, coopGroups,
+                         imageBufferWrtBack, wrtBackImageBuffer)) {
     LogError("Wrong memory objects!");
     return false;
   }
@@ -2698,12 +2703,10 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
   }
 
   // Check if image buffer write back is required
-  if (imageBufferWrtBack_) {
-    // Avoid recursive write back
-    imageBufferWrtBack_ = false;
+  if (imageBufferWrtBack) {
     // Make sure the original kernel execution is done
     releaseGpuMemoryFence();
-    for (const auto imageBuffer : wrtBackImageBuffer_) {
+    for (const auto imageBuffer : wrtBackImageBuffer) {
       Memory* buffer = dev().getGpuMemory(imageBuffer->owner()->parent());
       amd::Image* image = imageBuffer->owner()->asImage();
       Image* devImage = static_cast<Image*>(dev().getGpuMemory(imageBuffer->owner()));
@@ -2714,7 +2717,6 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
                                                 offs, image->getRegion(), true,
                                                 image->getRowPitch(), image->getSlicePitch());
     }
-    wrtBackImageBuffer_.clear();
   }
   return true;
 }
diff --git a/rocclr/device/rocm/rocvirtual.hpp b/rocclr/device/rocm/rocvirtual.hpp
index 964bba5338..d8b2a98a8b 100644
--- a/rocclr/device/rocm/rocvirtual.hpp
+++ b/rocclr/device/rocm/rocvirtual.hpp
@@ -305,7 +305,9 @@ class VirtualGPU : public device::VirtualDevice {
   bool processMemObjects(const amd::Kernel& kernel,  //!< AMD kernel object for execution
                          const_address params,       //!< Pointer to the param's store
                          size_t& ldsAddress,         //!< LDS usage
-                         bool cooperativeGroups      //!< Dispatch with cooperative groups
+                         bool cooperativeGroups,     //!< Dispatch with cooperative groups
+                         bool& imageBufferWrtBack,   //!< Image buffer write back is required
+                         std::vector<device::Memory*>& wrtBackImageBuffer //!< images for write back
                          );
 
   //! Adds a stage write buffer into a list
@@ -398,7 +400,6 @@ class VirtualGPU : public device::VirtualDevice {
   union {
     struct {
       uint32_t hasPendingDispatch_ : 1; //!< A kernel dispatch is outstanding
-      uint32_t imageBufferWrtBack_ : 1; //!< Image buffer write back is required
       uint32_t profiling_          : 1; //!< Profiling is enabled
       uint32_t cooperative_        : 1; //!< Cooperative launch is enabled
       uint32_t addSystemScope_     : 1; //!< Insert a system scope to the next aql
@@ -407,8 +408,6 @@ class VirtualGPU : public device::VirtualDevice {
     uint32_t  state_;
   };
 
-  std::vector<device::Memory*> wrtBackImageBuffer_;  //!< Array of images for write back
-
   Timestamp* timestamp_;
   hsa_agent_t gpu_device_;  //!< Physical device
   hsa_queue_t* gpu_queue_;  //!< Queue associated with a gpu