P4 to Git Change 1368703 by todli@todli-win-opencl-kv1 on 2017/02/02 18:12:31

SWDEV-96241 - Support SDI on PAL (runtime changes)

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palmemory.cpp#10 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#23 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#16 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#42 edit


[ROCm/clr commit: 1eda218988]

projects/clr/rocclr/runtime/device/pal/palmemory.cpp

@@ -168,7 +168,6 @@ Memory::create(
     if (result) {
         switch (memoryType()) {
         case Resource::Pinned:
-        case Resource::ExternalPhysical:
             // Marks memory object for direct GPU access to the host memory
             flags_ |= HostMemoryDirectAccess;
             break;

projects/clr/rocclr/runtime/device/pal/palresource.cpp

@@ -399,9 +399,9 @@ Resource::memTypeToHeap(Pal::GpuMemoryCreateInfo* createInfo)
         createInfo->heaps[0] = Pal::GpuHeapGartCacheable;
         desc_.cardMemory_ = false;
         break;
-    case Shader:
-    case BusAddressable:
     case ExternalPhysical:
+        desc_.cardMemory_ = false;
+    case Shader:
         // Fall through to process the memory allocation ...
     case Local:
         createInfo->heapCount = 2;
@@ -1075,6 +1075,18 @@ Resource::create(MemoryType memType, CreateParams* params)
     createInfo.alignment = MaxGpuAlignment;
     createInfo.vaRange = Pal::VaRange::Default;
     createInfo.priority  = Pal::GpuMemPriority::Normal;
+
+    if (memoryType() == ExternalPhysical){
+        cl_bus_address_amd bus_address =
+            (reinterpret_cast<amd::Buffer*>(params->owner_))->busAddress();
+        createInfo.surfaceBusAddr = bus_address.surface_bus_address;
+        createInfo.markerBusAddr = bus_address.marker_bus_address;
+        createInfo.flags.sdiExternal = true;
+    }
+    else if (memoryType() == BusAddressable){
+        createInfo.flags.busAddressable = true;
+    }
+
     memTypeToHeap(&createInfo);
     // createInfo.priority;
     memRef_ = dev().resourceCache().findGpuMemory(&desc_, createInfo.size, createInfo.alignment);

projects/clr/rocclr/runtime/device/pal/palsettings.cpp

@@ -353,13 +353,12 @@ Settings::create(
     }
 #endif // !defined(WITH_LIGHTNING_COMPILER)
 
-//! @todo
+    if (palProp.gpuMemoryProperties.busAddressableMemSize > 0) {
-/*
-    if (calAttr.totalSDIHeap > 0) {
         //Enable bus addressable memory extension
         enableExtension(ClAMDBusAddressableMemory);
     }
-
+//! @todo
+/*
     if (calAttr.longIdleDetect) {
         // KMD is unable to detect if we map the visible memory for CPU access, so
         // accessing persistent staged buffer may fail if LongIdleDetct is enabled.

projects/clr/rocclr/runtime/device/pal/palvirtual.cpp

@@ -1264,6 +1264,7 @@ VirtualGPU::copyMemory(cl_command_type type
 
     // Check if HW can be used for memory copy
     switch (type) {
+    case CL_COMMAND_MAKE_BUFFERS_RESIDENT_AMD:
     case CL_COMMAND_SVM_MEMCPY:
     case CL_COMMAND_COPY_BUFFER: {
         amd::Coord3D    realSrcOrigin(srcOrigin[0]);
@@ -2731,28 +2732,28 @@ VirtualGPU::submitSignal(amd::SignalCommand & vcmd)
 {
     amd::ScopedLock lock(execution());
     profilingBegin(vcmd);
-    pal::Memory* gpuMemory = dev().getGpuMemory(&vcmd.memory());
+    pal::Memory* pGpuMemory = dev().getGpuMemory(&vcmd.memory());
-    Unimplemented();
+
-/*
+    GpuEvent    gpuEvent;
+    eventBegin(MainEngine);
+
+    uint32_t value = vcmd.markerValue();
+    uint32_t size = vcmd.memory().getSize();
+
+    addVmMemory(pGpuMemory);
+
     if (vcmd.type() == CL_COMMAND_WAIT_SIGNAL_AMD) {
-        uint64_t surfAddr = gpuMemory->iMem()->getPhysicalAddress(cs());
+        iCmd()->CmdWaitMemoryValue(*(pGpuMemory->iMem()), size, value, 0xFFFFFFFF, Pal::CompareFunc::GreaterEqual);
-        uint64_t markerAddr = gpuMemory->iMem()->getMarkerAddress(cs());
-        uint64_t markerOffset = markerAddr - surfAddr;
-        cs()->p2pMarkerOp(gpuMemory->iMem(), vcmd.markerValue(),
-            markerOffset, false);
     }
     else if (vcmd.type() == CL_COMMAND_WRITE_SIGNAL_AMD) {
-        GpuEvent    gpuEvent;
+        iCmd()->CmdUpdateMemory(*(pGpuMemory->iMem()), size, 4, &value);
-        eventBegin(MainEngine);
-        cs()->p2pMarkerOp(gpuMemory->iMem(), vcmd.markerValue(),  vcmd.markerOffset(), true);
-        //! @todo We don't need flush if an event is tracked.
-        cs()->Flush();
-        eventEnd(MainEngine, gpuEvent);
-        gpuMemory->setBusy(*this, gpuEvent);
-        // Update the global GPU event
-        setGpuEvent(gpuEvent);
     }
-*/
+
+    eventEnd(MainEngine, gpuEvent);
+    pGpuMemory->setBusy(*this, gpuEvent);
+    // Update the global GPU event
+    setGpuEvent(gpuEvent);
+
     profilingEnd(vcmd);
 }
 
@@ -2761,39 +2762,24 @@ VirtualGPU::submitMakeBuffersResident(amd::MakeBuffersResidentCommand & vcmd)
 {
     amd::ScopedLock lock(execution());
     profilingBegin(vcmd);
+
     std::vector<amd::Memory*> memObjects = vcmd.memObjects();
-    cl_uint numObjects = memObjects.size();
+    uint32_t numObjects = memObjects.size();
-    Pal::IGpuMemory** pGpuMemObjects = new Pal::IGpuMemory*[numObjects];
 
-    for(cl_uint i = 0; i < numObjects; ++i)
+    for (int i = 0; i < numObjects; i++)
     {
-        pal::Memory* gpuMemory = dev().getGpuMemory(memObjects[i]);
+        // dummy render into the SDI surfaces so that KMD will be able to provide the bus addresses
-        pGpuMemObjects[i] = gpuMemory->iMem();
+        uint dummy = 0;
-        gpuMemory->syncCacheFromHost(*this);
+        static_cast<const KernelBlitManager&>(dev().xferMgr())
-    }
+            .writeRawData(*(dev().getGpuMemory(memObjects[i])), sizeof(dummy), &dummy);
 
-    uint64_t* surfBusAddr = new uint64_t[numObjects];
+        pal::Memory* pGpuMemory = dev().getGpuMemory(memObjects[i]);
-    uint64_t* markerBusAddr = new uint64_t[numObjects];
+
-    Unimplemented();
+        pGpuMemory->syncCacheFromHost(*this);
-/*
+
-    gslErrorCode res = cs()->makeBuffersResident(numObjects, pGpuMemObjects,
+        vcmd.busAddress()[i].surface_bus_address = pGpuMemory->iMem()->Desc().surfaceBusAddr;
-        surfBusAddr, markerBusAddr);
+        vcmd.busAddress()[i].marker_bus_address = pGpuMemory->iMem()->Desc().markerBusAddr;
-    if(res != GSL_NO_ERROR) {
-        LogError("MakeBuffersResident failed");
-        vcmd.setStatus(CL_INVALID_OPERATION);
     }
-    else {
-        cl_bus_address_amd* busAddr = vcmd.busAddress();
-        for(cl_uint i = 0; i < numObjects; ++i)
-        {
-            busAddr[i].surface_bus_address = surfBusAddr[i];
-            busAddr[i].marker_bus_address = markerBusAddr[i];
-        }
-    }
-*/
-    delete[] pGpuMemObjects;
-    delete[] surfBusAddr;
-    delete[] markerBusAddr;
     profilingEnd(vcmd);
 }