P4 to Git Change 1987582 by cpaquot@cpaquot-ocl-lc-lnx on 2019/08/22 13:20:28

SWDEV-185742 - [HIP] Add a svm path to avoid analyzing the mem objs
	and just memcpy the kernel args

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/hip/hip_module.cpp#36 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#82 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#25 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.cpp#96 edit
... //depot/stg/opencl/drivers/opencl/runtime/platform/command.hpp#94 edit

rocclr/runtime/platform/command.cpp

@@ -230,12 +230,13 @@ const Context& Command::context() const { return queue_->context(); }
 
 NDRangeKernelCommand::NDRangeKernelCommand(HostQueue& queue, const EventWaitList& eventWaitList,
                                            Kernel& kernel, const NDRangeContainer& sizes,
-                                           uint32_t sharedMemBytes, uint32_t extraParam)
+                                           uint32_t sharedMemBytes, uint32_t extraParam, bool svm)
     : Command(queue, CL_COMMAND_NDRANGE_KERNEL, eventWaitList, AMD_SERIALIZE_KERNEL)
     , kernel_(kernel)
     , sizes_(sizes)
     , sharedMemBytes_(sharedMemBytes)
-    , extraParam_(extraParam) {
+    , extraParam_(extraParam)
+    , svm_(svm) {
   auto& device = queue.device();
   auto devKernel = const_cast<device::Kernel*>(kernel.getDeviceKernel(device));
   profilingInfo_.setCallback(devKernel->getProfilingCallback(

rocclr/runtime/platform/command.hpp

@@ -764,6 +764,7 @@ class NDRangeKernelCommand : public Command {
   address parameters_;
   uint32_t sharedMemBytes_;
   uint32_t extraParam_;
+  bool svm_;
 
  public:
   enum {
@@ -774,7 +775,7 @@ class NDRangeKernelCommand : public Command {
   //! Construct an ExecuteKernel command
   NDRangeKernelCommand(HostQueue& queue, const EventWaitList& eventWaitList, Kernel& kernel,
                        const NDRangeContainer& sizes, uint32_t sharedMemBytes = 0,
-                       uint32_t extraParam = 0);
+                       uint32_t extraParam = 0, bool svm = false);
 
   virtual void submit(device::VirtualDevice& device) { device.submitKernel(*this); }
 
@@ -805,6 +806,9 @@ class NDRangeKernelCommand : public Command {
   void setLocalWorkSize(const NDRange& local) { sizes_.local() = local; }
 
   cl_int captureAndValidate();
+
+  //! Return if NDRange is svm or not
+  bool isSvm() const { return svm_; }
 };
 
 class NativeFnCommand : public Command {