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


[ROCm/clr commit: b02f98556a]

projects/clr/rocclr/runtime/device/rocm/rocvirtual.cpp

@@ -576,7 +576,7 @@ VirtualGPU::VirtualGPU(Device& device)
   kernarg_pool_base_ = nullptr;
   kernarg_pool_size_ = 0;
   kernarg_pool_cur_offset_ = 0;
-  aqlHeader_ = kDispatchPacketHeaderNoSync;
+  aqlHeader_ = kDispatchPacketHeader;
   barrier_signal_.handle = 0;
 
   // Note: Virtual GPU device creation must be a thread safe operation
@@ -1981,13 +1981,13 @@ bool VirtualGPU::createVirtualQueue(uint deviceQueueSize)
 }
 
 bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const amd::Kernel& kernel,
-  const_address parameters, void* eventHandle, uint32_t sharedMemBytes, bool cooperativeGroups) {
+  const_address parameters, void* eventHandle, uint32_t sharedMemBytes, bool cooperativeGroups, bool isSvm) {
   device::Kernel* devKernel = const_cast<device::Kernel*>(kernel.getDeviceKernel(dev()));
   Kernel& gpuKernel = static_cast<Kernel&>(*devKernel);
   size_t ldsUsage = gpuKernel.WorkgroupGroupSegmentByteSize();
 
   // Check memory dependency and SVM objects
-  if (!processMemObjects(kernel, parameters, ldsUsage, cooperativeGroups)) {
+  if (!isSvm && !processMemObjects(kernel, parameters, ldsUsage, cooperativeGroups)) {
     LogError("Wrong memory objects!");
     return false;
   }
@@ -2113,7 +2113,6 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
             AmdAqlWrap* wrap = reinterpret_cast<AmdAqlWrap*>(reinterpret_cast<uint64_t>(schedulerParam_->getHostMem()) + sizeof(SchedulerParam));
             memset(wrap, 0, sizeof(AmdAqlWrap));
             wrap->state = AQL_WRAP_DONE;
-
             spVA = reinterpret_cast<uint64_t>(schedulerMem->getDeviceMemory()) + sizeof(SchedulerParam);
           }
           WriteAqlArgAt(const_cast<address>(parameters), &spVA, it.size_, it.offset_);
@@ -2225,7 +2224,7 @@ void VirtualGPU::submitKernel(amd::NDRangeKernelCommand& vcmd) {
 
     // Submit kernel to HW
     if (!queue->submitKernelInternal(vcmd.sizes(), vcmd.kernel(), vcmd.parameters(),
-      static_cast<void*>(as_cl(&vcmd.event())), vcmd.sharedMemBytes(), vcmd.cooperativeGroups())) {
+      static_cast<void*>(as_cl(&vcmd.event())), vcmd.sharedMemBytes(), vcmd.cooperativeGroups()), vcmd.isSvm()) {
       LogError("AQL dispatch failed!");
       vcmd.setStatus(CL_INVALID_OPERATION);
     }

projects/clr/rocclr/runtime/device/rocm/rocvirtual.hpp

@@ -168,7 +168,8 @@ class VirtualGPU : public device::VirtualDevice {
                             const_address parameters,            //!< Parameters for the kernel
                             void* event_handle,  //!< Handle to OCL event for debugging
                             uint32_t sharedMemBytes = 0, //!< Shared memory size
-                            bool cooperativeGroups = false //!< TRUE if cooperative groups mode is required
+                            bool cooperativeGroups = false, //!< TRUE if cooperative groups mode is required
+                            bool isSvm = false //!< TRUE if all params are SVM
                             );
   void submitNativeFn(amd::NativeFnCommand& cmd);
   void submitMarker(amd::Marker& cmd);

projects/clr/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(

projects/clr/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 {