Add support for memory dumps in TBX mode with AUB dump

Related-To: NEO-4600

Change-Id: I4aa45c2745400d49e5bd92e0fc0aa76aa5df8678
Signed-off-by: Milczarek, Slawomir <slawomir.milczarek@intel.com>

opencl/source/command_stream/aub_command_stream_receiver_hw.h

@@ -39,7 +39,6 @@ class AUBCommandStreamReceiverHw : public CommandStreamReceiverSimulatedHw<GfxFa
     using CommandStreamReceiverSimulatedCommonHw<GfxFamily>::stream;
 
     bool flush(BatchBuffer &batchBuffer, ResidencyContainer &allocationsForResidency) override;
-    void makeNonResident(GraphicsAllocation &gfxAllocation) override;
 
     void processResidency(const ResidencyContainer &allocationsForResidency, uint32_t handleId) override;
 
@@ -67,7 +66,7 @@ class AUBCommandStreamReceiverHw : public CommandStreamReceiverSimulatedHw<GfxFa
 
     uint32_t getDumpHandle();
     MOCKABLE_VIRTUAL void addContextToken(uint32_t dumpHandle);
-    MOCKABLE_VIRTUAL void dumpAllocation(GraphicsAllocation &gfxAllocation);
+    void dumpAllocation(GraphicsAllocation &gfxAllocation) override;
 
     static CommandStreamReceiver *create(const std::string &fileName, bool standalone, ExecutionEnvironment &executionEnvironment, uint32_t rootDeviceIndex);
 

opencl/source/command_stream/aub_command_stream_receiver_hw_base.inl

@@ -758,15 +758,6 @@ void AUBCommandStreamReceiverHw<GfxFamily>::dumpAllocation(GraphicsAllocation &g
     AubAllocDump::dumpAllocation<GfxFamily>(dumpFormat, gfxAllocation, getAubStream(), getDumpHandle());
 }
 
-template <typename GfxFamily>
-void AUBCommandStreamReceiverHw<GfxFamily>::makeNonResident(GraphicsAllocation &gfxAllocation) {
-    if (gfxAllocation.isResident(this->osContext->getContextId())) {
-        dumpAllocation(gfxAllocation);
-        this->getEvictionAllocations().push_back(&gfxAllocation);
-        gfxAllocation.releaseResidencyInOsContext(this->osContext->getContextId());
-    }
-}
-
 template <typename GfxFamily>
 AubSubCaptureStatus AUBCommandStreamReceiverHw<GfxFamily>::checkAndActivateAubSubCapture(const MultiDispatchInfo &dispatchInfo) {
     auto status = subCaptureManager->checkAndActivateSubCapture(dispatchInfo);

opencl/source/command_stream/command_stream_receiver_simulated_common_hw.h

@@ -61,6 +61,10 @@ class CommandStreamReceiverSimulatedCommonHw : public CommandStreamReceiverHw<Gf
     virtual void setTbxWritable(bool writable, GraphicsAllocation &graphicsAllocation) = 0;
     virtual bool isTbxWritable(GraphicsAllocation &graphicsAllocation) const = 0;
 
+    virtual void dumpAllocation(GraphicsAllocation &gfxAllocation) = 0;
+
+    void makeNonResident(GraphicsAllocation &gfxAllocation) override;
+
     size_t getPreferredTagPoolSize() const override { return 1; }
 
     aub_stream::AubManager *aubManager = nullptr;

opencl/source/command_stream/command_stream_receiver_simulated_common_hw_base.inl

@@ -90,6 +90,15 @@ void CommandStreamReceiverSimulatedCommonHw<GfxFamily>::freeEngineInfo(AddressMa
     engineInfo.pRingBuffer = nullptr;
 }
 
+template <typename GfxFamily>
+void CommandStreamReceiverSimulatedCommonHw<GfxFamily>::makeNonResident(GraphicsAllocation &gfxAllocation) {
+    if (gfxAllocation.isResident(osContext->getContextId())) {
+        dumpAllocation(gfxAllocation);
+        this->getEvictionAllocations().push_back(&gfxAllocation);
+        gfxAllocation.releaseResidencyInOsContext(this->osContext->getContextId());
+    }
+}
+
 template <typename GfxFamily>
 uint32_t CommandStreamReceiverSimulatedCommonHw<GfxFamily>::getDeviceIndex() const {
     return osContext->getDeviceBitfield().any() ? static_cast<uint32_t>(Math::log2(static_cast<uint32_t>(osContext->getDeviceBitfield().to_ulong()))) : 0u;

opencl/source/command_stream/tbx_command_stream_receiver_hw.h

@@ -56,6 +56,8 @@ class TbxCommandStreamReceiverHw : public CommandStreamReceiverSimulatedHw<GfxFa
     MOCKABLE_VIRTUAL void submitBatchBuffer(uint64_t batchBufferGpuAddress, const void *batchBuffer, size_t batchBufferSize, uint32_t memoryBank, uint64_t entryBits, bool overrideRingHead);
     void pollForCompletion() override;
 
+    void dumpAllocation(GraphicsAllocation &gfxAllocation) override;
+
     static CommandStreamReceiver *create(const std::string &baseName, bool withAubDump, ExecutionEnvironment &executionEnvironment, uint32_t rootDeviceIndex);
 
     TbxCommandStreamReceiverHw(ExecutionEnvironment &executionEnvironment, uint32_t rootDeviceIndex);

opencl/source/command_stream/tbx_command_stream_receiver_hw.inl

@@ -11,6 +11,7 @@
 #include "shared/source/helpers/aligned_memory.h"
 #include "shared/source/helpers/constants.h"
 #include "shared/source/helpers/debug_helpers.h"
+#include "shared/source/helpers/engine_node_helper.h"
 #include "shared/source/helpers/hw_helper.h"
 #include "shared/source/helpers/ptr_math.h"
 #include "shared/source/memory_manager/graphics_allocation.h"
@@ -18,6 +19,8 @@
 
 #include "opencl/source/aub/aub_center.h"
 #include "opencl/source/aub/aub_helper.h"
+#include "opencl/source/aub_mem_dump/aub_alloc_dump.h"
+#include "opencl/source/aub_mem_dump/aub_alloc_dump.inl"
 #include "opencl/source/aub_mem_dump/page_table_entry_bits.h"
 #include "opencl/source/command_stream/aub_command_stream_receiver.h"
 #include "opencl/source/command_stream/command_stream_receiver_with_aub_dump.h"
@@ -528,4 +531,29 @@ AubSubCaptureStatus TbxCommandStreamReceiverHw<GfxFamily>::checkAndActivateAubSu
     }
     return status;
 }
+
+template <typename GfxFamily>
+void TbxCommandStreamReceiverHw<GfxFamily>::dumpAllocation(GraphicsAllocation &gfxAllocation) {
+    if (!hardwareContextController) {
+        return;
+    }
+
+    if (EngineHelpers::isBcs(this->osContext->getEngineType())) {
+        return;
+    }
+
+    if (DebugManager.flags.AUBDumpAllocsOnEnqueueReadOnly.get()) {
+        if (!gfxAllocation.isAllocDumpable()) {
+            return;
+        }
+        gfxAllocation.setAllocDumpable(false);
+    }
+
+    auto dumpFormat = AubAllocDump::getDumpFormat(gfxAllocation);
+    auto surfaceInfo = std::unique_ptr<aub_stream::SurfaceInfo>(AubAllocDump::getDumpSurfaceInfo<GfxFamily>(gfxAllocation, dumpFormat));
+    if (surfaceInfo) {
+        hardwareContextController->pollForCompletion();
+        hardwareContextController->dumpSurface(*surfaceInfo.get());
+    }
+}
 } // namespace NEO