P4 to Git Change 1527848 by gandryey@gera-w8 on 2018/03/15 17:11:43

SWDEV-79445 - OCL generic changes and code clean-up
	- Add suballocations support for local(invisible) memory. It should significantly improve memory footprint and TLB usage with 2MB pages
	- Implementation uses BuddyAllocator provided in PAL
	- The chunk allocation size is 64MB, min allocation 4KB and max 4MB. GPU_MAX_SUBALLOC_SIZE controls the max size in KB

Affected files ...

... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldefs.hpp#33 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.cpp#76 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/paldevice.hpp#24 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palprogram.cpp#56 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.cpp#51 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palresource.hpp#17 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.cpp#45 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palsettings.hpp#16 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.cpp#77 edit
... //depot/stg/opencl/drivers/opencl/runtime/device/pal/palvirtual.hpp#42 edit
... //depot/stg/opencl/drivers/opencl/runtime/utils/flags.hpp#285 edit

rocclr/runtime/device/pal/palvirtual.cpp

@@ -409,7 +409,7 @@ void VirtualGPU::MemoryDependency::validate(VirtualGPU& gpu, const Memory* memor
   }
 
   uint64_t curStart = memory->vmAddress();
-  uint64_t curEnd = curStart + memory->vmSize();
+  uint64_t curEnd = curStart + memory->size();
 
   // Loop through all memory objects in the queue and find dependency
   // @note don't include objects from the current kernel
@@ -1974,6 +1974,7 @@ void VirtualGPU::PostDeviceEnqueue(
     uint64_t vmParentWrap,
     GpuEvent* gpuEvent)
 {
+  uint32_t id  = gpuEvent->id;
   amd::DeviceQueue* defQueue = kernel.program().context().defDeviceQueue(dev());
 
   // Make sure exculsive access to the device queue
@@ -2055,6 +2056,9 @@ void VirtualGPU::PostDeviceEnqueue(
     iCmd()->CmdVirtualQueueHandshake(vmParentWrap + offsetof(AmdAqlWrap, state), AQL_WRAP_DONE,
       vmParentWrap + offsetof(AmdAqlWrap, child_counter),
       signalAddr, dev().settings().useDeviceQueue_);
+    if (id != gpuEvent->id) {
+        LogError("Something is wrong. ID mismatch!\n");
+    }
     eventEnd(MainEngine, *gpuEvent);
   }
 
@@ -2203,6 +2207,9 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
     if (profiling() || state_.profileEnabled_) {
       addBarrier();
     }
+    if (id != gpuEvent.id) {
+      LogError("Something is wrong. ID mismatch!\n");
+    }
     eventEnd(MainEngine, gpuEvent);
 
     // Execute scheduler for device enqueue
@@ -2210,9 +2217,6 @@ bool VirtualGPU::submitKernelInternal(const amd::NDRangeContainer& sizes, const
       PostDeviceEnqueue(kernel, hsaKernel, gpuDefQueue, vmDefQueue, vmParentWrap, &gpuEvent);
     }
 
-    if (id != gpuEvent.id) {
-      LogError("Something is wrong. ID mismatch!\n");
-    }
     // Update the global GPU event
     setGpuEvent(gpuEvent, needFlush);
 
@@ -2266,7 +2270,7 @@ void VirtualGPU::submitMarker(amd::Marker& vcmd) {
   }
 }
 
-void VirtualGPU::releaseMemory(GpuMemoryReference* mem, GpuEvent* event) {
+void VirtualGPU::releaseMemory(GpuMemoryReference* mem) {
   queues_[MainEngine]->removeCmdMemRef(mem);
   queues_[SdmaEngine]->removeCmdMemRef(mem);
 }