P4 to Git Change 1114420 by wchau@wchau_WINDOWS7_OCL on 2015/01/22 16:45:46

ECR #399840 - OpenCL Runtime HW Debug support development - set aclBinary & event in the pre-dispatch callback function

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudebugger.hpp#4 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpudebugmanager.cpp#4 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.cpp#347 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/gpu/gpuvirtual.hpp#125 edit

rocclr/runtime/device/gpu/gpudebugger.hpp

@@ -18,6 +18,7 @@
 #include "sc-hsa/Interface/SCHSAInterface.h"
 #include "device/device.hpp"
 #include "device/hwdebug.hpp"
+#include "acl.h"
 
 static const int NumberReserveVgprs = 4;
 
@@ -93,6 +94,8 @@ struct DebugToolInfo
     amd::Memory*   trapHandler_;       //! Trap handler address
     amd::Memory*   trapBuffer_;        //! Trap buffer address
     bool     sqPerfcounterEnable_;     //! whether SQ perf counters are enabled
+    aclBinary*     aclBinary_;         //! pointer of the kernel ACL binary
+    amd::Event*    event_;             //! pointer of the kernel event in the enqueue command
 };
 
 /*!  \brief Message used by the KFD wave control for CI

rocclr/runtime/device/gpu/gpudebugmanager.cpp

@@ -85,6 +85,9 @@ GpuDebugManager::executePreDispatchCallBack(void*  aqlPacket,
 //      // for invalidate cache (BuildEndOfKernelNotifyCommands)
 //        aqlPacket->release_fence_scope = 2;
 
+        aclBinary_ = reinterpret_cast<void*>(info->aclBinary_);
+        oclEventHandle_ = reinterpret_cast<void*>(as_cl(info->event_));
+
         cl_device_id clDeviceId = as_cl(device_);
         preDispatchCallBackFunc_(clDeviceId,
                                  oclEventHandle_,

rocclr/runtime/device/gpu/gpuvirtual.cpp

@@ -1627,7 +1627,8 @@ VirtualGPU::submitKernel(amd::NDRangeKernelCommand& vcmd)
     profilingBegin(vcmd);
 
     // Submit kernel to HW
-    if (!submitKernelInternal(vcmd.sizes(), vcmd.kernel(), vcmd.parameters(), false)) {
+    if (!submitKernelInternal(vcmd.sizes(), vcmd.kernel(), vcmd.parameters(), false,
+                              &vcmd.event())) {
         vcmd.setStatus(CL_INVALID_OPERATION);
     }
 
@@ -1639,7 +1640,8 @@ VirtualGPU::submitKernelInternalHSA(
     const amd::NDRangeContainer& sizes,
     const amd::Kernel&  kernel,
     const_address parameters,
-    bool    nativeMem)
+    bool    nativeMem,
+    amd::Event* enqueueEvent)
 {
     uint64_t    vmParentWrap = 0;
     uint64_t    vmDefQueue = 0;
@@ -1766,7 +1768,7 @@ VirtualGPU::submitKernelInternalHSA(
     HwDbgKernelInfo *pKernelInfo = NULL;
 
     if (useHwDebug_) {
-        buildKernelInfo(hsaKernel, aqlPkt, kernelInfo);
+        buildKernelInfo(hsaKernel, aqlPkt, kernelInfo, enqueueEvent);
         pKernelInfo = &kernelInfo;
     }
 
@@ -1982,7 +1984,8 @@ VirtualGPU::submitKernelInternal(
     const amd::NDRangeContainer& sizes,
     const amd::Kernel&  kernel,
     const_address parameters,
-    bool    nativeMem)
+    bool    nativeMem,
+    amd::Event* enqueueEvent)
 {
     bool            result = true;
     uint            i;
@@ -1999,7 +2002,7 @@ VirtualGPU::submitKernelInternal(
     Kernel& gpuKernelOpt = static_cast<gpu::Kernel&>(*devKernel);
 
     if (gpuKernelOpt.hsa()) {
-        return submitKernelInternalHSA(sizes, kernel, parameters, nativeMem);
+        return submitKernelInternalHSA(sizes, kernel, parameters, nativeMem, enqueueEvent);
     }
     else if (state_.hsailKernel_) {
         // Reload GSL state to HW, so runtime could run AMDIL kernel
@@ -3458,7 +3461,8 @@ VirtualGPU::flushCuCaches(HwDbgGpuCacheMask cache_mask)
 void
 VirtualGPU::buildKernelInfo(const HSAILKernel& hsaKernel,
                             hsa_kernel_dispatch_packet_t* aqlPkt,
-                            HwDbgKernelInfo& kernelInfo)
+                            HwDbgKernelInfo& kernelInfo,
+                            amd::Event* enqueueEvent)
 {
     amd::HwDebugManager * dbgManager = dev().hwDebugMgr();
     assert (dbgManager && "No HW Debug Manager!");
@@ -3517,6 +3521,8 @@ VirtualGPU::buildKernelInfo(const HSAILKernel& hsaKernel,
         dbgSetting.scratchAddress_ = kernelInfo.scratchBufAddr;
         dbgSetting.scratchSize_ = kernelInfo.scratchBufferSizeInBytes;
         dbgSetting.globalAddress_ = kernelInfo.heapBufAddr;
+        dbgSetting.aclBinary_ = hsaKernel.prog().binaryElf();
+        dbgSetting.event_ = enqueueEvent;
 
         // Call the predispatch callback function & set the trap info
         AqlCodeInfo  aqlCodeInfo;

rocclr/runtime/device/gpu/gpuvirtual.hpp

@@ -223,13 +223,15 @@ public:
         const amd::NDRangeContainer& sizes, //!< Workload sizes
         const amd::Kernel&  kernel,         //!< Kernel for execution
         const_address parameters,           //!< Parameters for the kernel
-        bool    nativeMem = true            //!< Native memory objects
+        bool     nativeMem = true,          //!< Native memory objects
+        amd::Event* enqueueEvent = NULL     //!< Event provided in the enqueue kernel command
         );
     bool submitKernelInternalHSA(
         const amd::NDRangeContainer& sizes, //!< Workload sizes
         const amd::Kernel&  kernel,         //!< Kernel for execution
         const_address parameters,           //!< Parameters for the kernel
-        bool    nativeMem = true            //!< Native memory objects
+        bool    nativeMem = true,           //!< Native memory objects
+        amd::Event* enqueueEvent = NULL     //!< Event provided in the enqueue kernel command
         );
     void submitNativeFn(amd::NativeFnCommand& vcmd);
     void submitFillMemory(amd::FillMemoryCommand& vcmd);
@@ -505,7 +507,8 @@ private:
     void buildKernelInfo(
         const HSAILKernel& hsaKernel,       //!< hsa kernel
         hsa_kernel_dispatch_packet_t* aqlPkt,   //!< aql packet for dispatch
-        HwDbgKernelInfo& kernelInfo         //!< kernel info for the dispatch
+        HwDbgKernelInfo& kernelInfo,        //!< kernel info for the dispatch
+        amd::Event* enqueueEvent            //!< Event provided in the enqueue kernel command
         );
 
     void assignTrapHandler(