P4 to Git Change 1999915 by cpaquot@cpaquot-ocl-lc-lnx on 2019/09/17 19:43:34

SWDEV-201128 - [HIP] test_snli_cuda failure
	Default to sync packet
	Make sure GPU_NUM_MEM_DEPENDENCY is 0 for HIP
	No sync packet is only used when there are mem dependency check

Affected files ...

... //depot/stg/opencl/drivers/opencl/api/hip/hip_context.cpp#22 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.cpp#86 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/rocm/rocvirtual.hpp#28 edit


[ROCm/clr commit: 78cf6e473e]

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

@@ -195,7 +195,7 @@ bool VirtualGPU::processMemObjects(const amd::Kernel& kernel, const_address para
   const amd::KernelSignature& signature = kernel.signature();
   const amd::KernelParameters& kernelParams = kernel.parameters();
 
-  if (!cooperativeGroups) {
+  if (!cooperativeGroups && memoryDependency().maxMemObjectsInQueue() != 0) {
     // AQL packets
     setAqlHeader(kDispatchPacketHeaderNoSync);
   }
@@ -575,7 +575,7 @@ 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
@@ -2052,7 +2052,7 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
       return false;
     }
 
-    LogPrintfInfo("!\tShaderName : %s\n", gpuKernel.name().c_str());
+    LogPrintfInfo("[%zx]!\tShaderName : %s\n", std::this_thread::get_id(), gpuKernel.name().c_str());
 
     // Check if runtime has to setup hidden arguments
     for (uint32_t i = signature.numParameters(); i < signature.numParametersAll(); ++i) {

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

@@ -131,6 +131,9 @@ class VirtualGPU : public device::VirtualDevice {
     //! Clear memory dependency
     void clear(bool all = true);
 
+    //! Max number of mem objects in the queue
+    size_t maxMemObjectsInQueue() const { return maxMemObjectsInQueue_; }
+
    private:
     struct MemoryState {
       uint64_t start_;  //! Busy memory start address