Add flag to determine if csr can allocate in internal heaps

Signed-off-by: Mateusz Jablonski <[email protected]>

Author: JablonskiMateusz

opencl/test/unit_test/command_queue/command_queue_tests.cpp

@@ -517,12 +517,13 @@ TEST_P(CommandQueueIndirectHeapTest, WhenGettingIndirectHeapThenValidObjectIsRet
     EXPECT_NE(nullptr, &indirectHeap);
 }
 
-TEST_P(CommandQueueIndirectHeapTest, givenIndirectObjectHeapWhenItIsQueriedForInternalAllocationThenTrueIsReturned) {
+HWTEST_P(CommandQueueIndirectHeapTest, givenIndirectObjectHeapWhenItIsQueriedForInternalAllocationThenTrueIsReturned) {
     const cl_queue_properties props[3] = {CL_QUEUE_PROPERTIES, 0, 0};
     MockCommandQueue cmdQ(context.get(), pClDevice, props);
+    auto &commandStreamReceiver = pClDevice->getUltCommandStreamReceiver<FamilyType>();
 
     auto &indirectHeap = cmdQ.getIndirectHeap(this->GetParam(), 8192);
-    if (this->GetParam() == IndirectHeap::INDIRECT_OBJECT) {
+    if (this->GetParam() == IndirectHeap::INDIRECT_OBJECT && commandStreamReceiver.canUse4GbHeaps) {
         EXPECT_TRUE(indirectHeap.getGraphicsAllocation()->is32BitAllocation());
     } else {
         EXPECT_FALSE(indirectHeap.getGraphicsAllocation()->is32BitAllocation());
@@ -587,7 +588,7 @@ TEST_P(CommandQueueIndirectHeapTest, WhenGettingIndirectHeapThenSizeIsAlignedToC
     EXPECT_TRUE(isAligned<MemoryConstants::cacheLineSize>(indirectHeap.getAvailableSpace()));
 }
 
-TEST_P(CommandQueueIndirectHeapTest, givenCommandStreamReceiverWithReusableAllocationsWhenAskedForHeapAllocationThenAllocationFromReusablePoolIsReturned) {
+HWTEST_P(CommandQueueIndirectHeapTest, givenCommandStreamReceiverWithReusableAllocationsWhenAskedForHeapAllocationThenAllocationFromReusablePoolIsReturned) {
     const cl_queue_properties props[3] = {CL_QUEUE_PROPERTIES, 0, 0};
     MockCommandQueue cmdQ(context.get(), pClDevice, props);
 
@@ -597,9 +598,9 @@ TEST_P(CommandQueueIndirectHeapTest, givenCommandStreamReceiverWithReusableAlloc
 
     GraphicsAllocation *allocation = nullptr;
 
-    auto &commandStreamReceiver = cmdQ.getGpgpuCommandStreamReceiver();
+    auto &commandStreamReceiver = pClDevice->getUltCommandStreamReceiver<FamilyType>();
     auto allocationType = GraphicsAllocation::AllocationType::LINEAR_STREAM;
-    if (this->GetParam() == IndirectHeap::INDIRECT_OBJECT) {
+    if (this->GetParam() == IndirectHeap::INDIRECT_OBJECT && commandStreamReceiver.canUse4GbHeaps) {
         allocationType = GraphicsAllocation::AllocationType::INTERNAL_HEAP;
     }
     allocation = memoryManager->allocateGraphicsMemoryWithProperties({pDevice->getRootDeviceIndex(), allocationSize, allocationType, pDevice->getDeviceBitfield()});
@@ -736,13 +737,14 @@ TEST_P(CommandQueueIndirectHeapTest, GivenCommandQueueWithHeapWhenGraphicAllocat
     memoryManager->freeGraphicsMemory(allocation);
 }
 
-TEST_P(CommandQueueIndirectHeapTest, givenCommandQueueWhenGetIndirectHeapIsCalledThenIndirectHeapAllocationTypeShouldBeSetToInternalHeapForIohAndLinearStreamForOthers) {
+HWTEST_P(CommandQueueIndirectHeapTest, givenCommandQueueWhenGetIndirectHeapIsCalledThenIndirectHeapAllocationTypeShouldBeSetToInternalHeapForIohAndLinearStreamForOthers) {
     const cl_queue_properties props[3] = {CL_QUEUE_PROPERTIES, 0, 0};
     MockCommandQueue cmdQ(context.get(), pClDevice, props);
+    auto &commandStreamReceiver = pClDevice->getUltCommandStreamReceiver<FamilyType>();
 
     auto heapType = this->GetParam();
 
-    bool requireInternalHeap = IndirectHeap::INDIRECT_OBJECT == heapType;
+    bool requireInternalHeap = IndirectHeap::INDIRECT_OBJECT == heapType && commandStreamReceiver.canUse4GbHeaps;
     const auto &indirectHeap = cmdQ.getIndirectHeap(heapType, 100);
     auto indirectHeapAllocation = indirectHeap.getGraphicsAllocation();
     ASSERT_NE(nullptr, indirectHeapAllocation);

opencl/test/unit_test/helpers/hardware_commands_helper_tests.cpp

@@ -206,7 +206,7 @@ HWTEST_F(HardwareCommandsTest, givenSendCrossThreadDataWhenWhenAddPatchInfoComme
     EXPECT_EQ(PatchInfoAllocationType::IndirectObjectHeap, kernel->getPatchInfoDataList()[0].targetType);
 }
 
-HWTEST_F(HardwareCommandsTest, givenIndirectHeapNotAllocatedFromInternalPoolWhenSendCrossThreadDataIsCalledThenOffsetZeroIsReturned) {
+HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenIndirectHeapNotAllocatedFromInternalPoolWhenSendCrossThreadDataIsCalledThenOffsetZeroIsReturned) {
     auto nonInternalAllocation = pDevice->getMemoryManager()->allocateGraphicsMemoryWithProperties(MockAllocationProperties{pDevice->getRootDeviceIndex(), MemoryConstants::pageSize});
     IndirectHeap indirectHeap(nonInternalAllocation, false);
 
@@ -221,7 +221,7 @@ HWTEST_F(HardwareCommandsTest, givenIndirectHeapNotAllocatedFromInternalPoolWhen
     pDevice->getMemoryManager()->freeGraphicsMemory(nonInternalAllocation);
 }
 
-HWTEST_F(HardwareCommandsTest, givenIndirectHeapAllocatedFromInternalPoolWhenSendCrossThreadDataIsCalledThenHeapBaseOffsetIsReturned) {
+HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenIndirectHeapAllocatedFromInternalPoolWhenSendCrossThreadDataIsCalledThenHeapBaseOffsetIsReturned) {
     auto internalAllocation = pDevice->getMemoryManager()->allocateGraphicsMemoryWithProperties(MockAllocationProperties(pDevice->getRootDeviceIndex(), true, MemoryConstants::pageSize, GraphicsAllocation::AllocationType::INTERNAL_HEAP, pDevice->getDeviceBitfield()));
     IndirectHeap indirectHeap(internalAllocation, true);
     auto expectedOffset = internalAllocation->getGpuAddressToPatch();
@@ -238,7 +238,7 @@ HWTEST_F(HardwareCommandsTest, givenIndirectHeapAllocatedFromInternalPoolWhenSen
     pDevice->getMemoryManager()->freeGraphicsMemory(internalAllocation);
 }
 
-HWTEST_F(HardwareCommandsTest, givenSendCrossThreadDataWhenWhenAddPatchInfoCommentsForAUBDumpIsSetThenAddPatchInfoDataOffsetsAreMoved) {
+HWCMDTEST_F(IGFX_GEN8_CORE, HardwareCommandsTest, givenSendCrossThreadDataWhenWhenAddPatchInfoCommentsForAUBDumpIsSetThenAddPatchInfoDataOffsetsAreMoved) {
     DebugManagerStateRestore dbgRestore;
     DebugManager.flags.AddPatchInfoCommentsForAUBDump.set(true);
 

opencl/test/unit_test/libult/ult_command_stream_receiver.h

@@ -54,6 +54,7 @@ class UltCommandStreamReceiver : public CommandStreamReceiverHw<GfxFamily>, publ
     using BaseClass::staticWorkPartitioningEnabled;
     using BaseClass::wasSubmittedToSingleSubdevice;
     using BaseClass::CommandStreamReceiver::bindingTableBaseAddressRequired;
+    using BaseClass::CommandStreamReceiver::canUse4GbHeaps;
     using BaseClass::CommandStreamReceiver::checkForNewResources;
     using BaseClass::CommandStreamReceiver::checkImplicitFlushForGpuIdle;
     using BaseClass::CommandStreamReceiver::cleanupResources;

shared/source/command_stream/command_stream_receiver.cpp

@@ -390,7 +390,7 @@ void CommandStreamReceiver::allocateHeapMemory(IndirectHeap::Type heapType,
     if (IndirectHeap::SURFACE_STATE == heapType) {
         finalHeapSize = defaultSshSize;
     }
-    bool requireInternalHeap = IndirectHeap::INDIRECT_OBJECT == heapType ? true : false;
+    bool requireInternalHeap = IndirectHeap::INDIRECT_OBJECT == heapType ? canUse4GbHeaps : false;
 
     if (DebugManager.flags.AddPatchInfoCommentsForAUBDump.get()) {
         requireInternalHeap = false;

shared/source/command_stream/command_stream_receiver.h

@@ -184,6 +184,7 @@ class CommandStreamReceiver {
     bool peekTimestampPacketWriteEnabled() const { return timestampPacketWriteEnabled; }
 
     size_t defaultSshSize;
+    bool canUse4GbHeaps = true;
 
     AllocationsList &getTemporaryAllocations();
     AllocationsList &getAllocationsForReuse();

shared/source/command_stream/command_stream_receiver_hw.h

@@ -145,6 +145,7 @@ class CommandStreamReceiverHw : public CommandStreamReceiver {
     void addPipeControlCmd(LinearStream &commandStream, PipeControlArgs &args);
     void addPipeControlBeforeStateBaseAddress(LinearStream &commandStream);
     size_t getSshHeapSize();
+    bool are4GbHeapsAvailable() const;
 
     uint64_t getScratchPatchAddress();
     void createScratchSpaceController();

shared/source/command_stream/command_stream_receiver_hw_base.inl

@@ -52,6 +52,7 @@ CommandStreamReceiverHw<GfxFamily>::CommandStreamReceiverHw(ExecutionEnvironment
     resetKmdNotifyHelper(new KmdNotifyHelper(&peekHwInfo().capabilityTable.kmdNotifyProperties));
     flatBatchBufferHelper.reset(new FlatBatchBufferHelperHw<GfxFamily>(executionEnvironment));
     defaultSshSize = getSshHeapSize();
+    canUse4GbHeaps = are4GbHeapsAvailable();
 
     timestampPacketWriteEnabled = hwHelper.timestampPacketWriteSupported();
     if (DebugManager.flags.EnableTimestampPacket.get() != -1) {

shared/source/command_stream/command_stream_receiver_hw_bdw_plus.inl

@@ -1,5 +1,5 @@
 /*
- * Copyright (C) 2019-2020 Intel Corporation
+ * Copyright (C) 2019-2021 Intel Corporation
  *
  * SPDX-License-Identifier: MIT
  *
@@ -14,6 +14,9 @@ size_t CommandStreamReceiverHw<GfxFamily>::getSshHeapSize() {
     return defaultHeapSize;
 }
 
+template <typename GfxFamily>
+bool CommandStreamReceiverHw<GfxFamily>::are4GbHeapsAvailable() const { return true; }
+
 template <typename GfxFamily>
 inline void CommandStreamReceiverHw<GfxFamily>::programL3(LinearStream &csr, DispatchFlags &dispatchFlags, uint32_t &newL3Config) {
     typedef typename GfxFamily::PIPE_CONTROL PIPE_CONTROL;