[22/n] Internal 4GB allocator.

- Finalize Instruction Heap removal.

Change-Id: Idd7df94a228238a5157c3251180fc3c8d3a189df

Author: MichalMrozek

runtime/command_queue/command_queue.cpp

@@ -217,11 +217,6 @@ uint32_t CommandQueue::getTaskLevelFromWaitList(uint32_t taskLevel,
     return taskLevel;
 }
 
-size_t CommandQueue::getInstructionHeapReservedBlockSize() const {
-    return alignUp(device->getCommandStreamReceiver().getInstructionHeapCmdStreamReceiverReservedSize(),
-                   MemoryConstants::cacheLineSize);
-}
-
 IndirectHeap &CommandQueue::getIndirectHeap(IndirectHeap::Type heapType,
                                             size_t minRequiredSize) {
     DEBUG_BREAK_IF(static_cast<uint32_t>(heapType) >= ARRAY_COUNT(indirectHeap));
@@ -243,10 +238,6 @@ IndirectHeap &CommandQueue::getIndirectHeap(IndirectHeap::Type heapType,
     if (!heapMemory) {
         size_t reservedSize = 0;
         auto finalHeapSize = defaultHeapSize;
-        if (heapType == IndirectHeap::INSTRUCTION) {
-            finalHeapSize = optimalInstructionHeapSize;
-            reservedSize = getInstructionHeapReservedBlockSize();
-        }
 
         minRequiredSize += reservedSize;
 
@@ -274,11 +265,6 @@ IndirectHeap &CommandQueue::getIndirectHeap(IndirectHeap::Type heapType,
             heap = new IndirectHeap(heapMemory);
             heap->overrideMaxSize(finalHeapSize);
         }
-
-        if (heapType == IndirectHeap::INSTRUCTION) {
-            device->getCommandStreamReceiver().initializeInstructionHeapCmdStreamReceiverReservedBlock(*heap);
-            heap->align(MemoryConstants::cacheLineSize);
-        }
     }
 
     return *heap;

runtime/command_queue/command_queue.h

@@ -333,7 +333,6 @@ class CommandQueue : public BaseObject<_cl_command_queue> {
     Context *getContextPtr() { return context; }
 
     LinearStream &getCS(size_t minRequiredSize = 1024u);
-    size_t getInstructionHeapReservedBlockSize() const;
     IndirectHeap &getIndirectHeap(IndirectHeap::Type heapType,
                                   size_t minRequiredSize = 0u);
 

runtime/command_stream/command_stream_receiver.cpp

@@ -219,14 +219,6 @@ void CommandStreamReceiver::setRequiredScratchSize(uint32_t newRequiredScratchSi
     }
 }
 
-size_t CommandStreamReceiver::getInstructionHeapCmdStreamReceiverReservedSize() const {
-    return PreemptionHelper::getInstructionHeapSipKernelReservedSize(*memoryManager->device);
-}
-
-void CommandStreamReceiver::initializeInstructionHeapCmdStreamReceiverReservedBlock(LinearStream &ih) const {
-    return PreemptionHelper::initializeInstructionHeapSipKernelReservedBlock(ih, *memoryManager->device);
-}
-
 GraphicsAllocation *CommandStreamReceiver::allocateDebugSurface(size_t size) {
     UNRECOVERABLE_IF(debugSurface != nullptr);
     debugSurface = memoryManager->allocateGraphicsMemory(size);

runtime/command_stream/command_stream_receiver.h

@@ -116,12 +116,6 @@ class CommandStreamReceiver {
     virtual void waitForTaskCountWithKmdNotifyFallback(uint32_t taskCountToWait, FlushStamp flushStampToWait, bool useQuickKmdSleep) = 0;
     MOCKABLE_VIRTUAL bool waitForCompletionWithTimeout(bool enableTimeout, int64_t timeoutMicroseconds, uint32_t taskCountToWait);
 
-    // returns size of block that needs to be reserved at the beginning of each instruction heap for CommandStreamReceiver
-    MOCKABLE_VIRTUAL size_t getInstructionHeapCmdStreamReceiverReservedSize() const;
-
-    // allows CommandStreamReceiver to prepopulate reserved block in instruction heap
-    MOCKABLE_VIRTUAL void initializeInstructionHeapCmdStreamReceiverReservedBlock(LinearStream &ih) const;
-
     void setSamplerCacheFlushRequired(SamplerCacheFlushState value) { this->samplerCacheFlushRequired = value; }
 
     // Collect patch info data

runtime/command_stream/preemption.cpp

@@ -104,29 +104,6 @@ void PreemptionHelper::adjustDefaultPreemptionMode(RuntimeCapabilityTable &devic
     }
 }
 
-size_t PreemptionHelper::getInstructionHeapSipKernelReservedSize(Device &device) {
-    if (device.getPreemptionMode() != PreemptionMode::MidThread) {
-        return 0;
-    }
-
-    return BuiltIns::getInstance().getSipKernel(SipKernelType::Csr, device).getBinarySize();
-}
-
-void PreemptionHelper::initializeInstructionHeapSipKernelReservedBlock(LinearStream &ih, Device &device) {
-    if (device.getPreemptionMode() != PreemptionMode::MidThread) {
-        return;
-    }
-
-    const SipKernel &sip = BuiltIns::getInstance().getSipKernel(SipKernelType::Csr, device);
-    size_t sipSize = sip.getBinarySize();
-    UNRECOVERABLE_IF(sipSize > ih.getAvailableSpace());
-    UNRECOVERABLE_IF(0 != ih.getUsed());
-    void *blockForSip = ih.getSpace(sipSize);
-    UNRECOVERABLE_IF(nullptr == blockForSip);
-    auto err = memcpy_s(blockForSip, sipSize, sip.getBinary(), sipSize);
-    UNRECOVERABLE_IF(err != 0);
-}
-
 PreemptionMode PreemptionHelper::getDefaultPreemptionMode(const HardwareInfo &hwInfo) {
     return DebugManager.flags.ForcePreemptionMode.get() == -1
                ? hwInfo.capabilityTable.defaultPreemptionMode

runtime/command_stream/preemption.h

@@ -41,9 +41,6 @@ class PreemptionHelper {
     static bool allowMidThreadPreemption(Kernel *kernel, Device &device);
     static void adjustDefaultPreemptionMode(RuntimeCapabilityTable &deviceCapabilities, bool allowMidThread, bool allowThreadGroup, bool allowMidBatch);
 
-    static size_t getInstructionHeapSipKernelReservedSize(Device &device);
-    static void initializeInstructionHeapSipKernelReservedBlock(LinearStream &ih, Device &device);
-
     template <typename GfxFamily>
     static size_t getRequiredPreambleSize(const Device &device);
 

runtime/indirect_heap/indirect_heap.h

@@ -30,7 +30,6 @@ namespace OCLRT {
 class GraphicsAllocation;
 
 constexpr size_t defaultHeapSize = 64 * KB;
-constexpr size_t optimalInstructionHeapSize = 512 * KB;
 constexpr size_t maxSshSize = defaultHeapSize - MemoryConstants::pageSize;
 
 class IndirectHeap : public LinearStream {
@@ -41,7 +40,6 @@ class IndirectHeap : public LinearStream {
         DYNAMIC_STATE = 0,
         GENERAL_STATE,
         INDIRECT_OBJECT,
-        INSTRUCTION,
         SURFACE_STATE,
         NUM_TYPES
     };

unit_tests/command_queue/command_queue_tests.cpp

@@ -421,8 +421,6 @@ TEST_P(CommandQueueIndirectHeapTest, IndirectHeapContainsAtLeast64KB) {
     auto &indirectHeap = cmdQ.getIndirectHeap(this->GetParam(), sizeof(uint32_t));
     if (this->GetParam() == IndirectHeap::SURFACE_STATE) {
         EXPECT_EQ(64 * KB - MemoryConstants::pageSize, indirectHeap.getAvailableSpace());
-    } else if (this->GetParam() == IndirectHeap::INSTRUCTION) {
-        EXPECT_EQ(optimalInstructionHeapSize, indirectHeap.getAvailableSpace());
     } else {
         EXPECT_EQ(64 * KB, indirectHeap.getAvailableSpace());
     }
@@ -478,9 +476,6 @@ TEST_P(CommandQueueIndirectHeapTest, MemoryManagerWithReusableAllocationsWhenAsk
 
     auto memoryManager = pDevice->getMemoryManager();
     auto allocationSize = defaultHeapSize * 2;
-    if (this->GetParam() == IndirectHeap::INSTRUCTION) {
-        allocationSize = optimalInstructionHeapSize * 2;
-    }
 
     auto allocation = memoryManager->allocateGraphicsMemory(allocationSize);
     memoryManager->storeAllocation(std::unique_ptr<GraphicsAllocation>(allocation), REUSABLE_ALLOCATION);
@@ -627,7 +622,6 @@ INSTANTIATE_TEST_CASE_P(
         IndirectHeap::DYNAMIC_STATE,
         IndirectHeap::GENERAL_STATE,
         IndirectHeap::INDIRECT_OBJECT,
-        IndirectHeap::INSTRUCTION,
         IndirectHeap::SURFACE_STATE));
 
 typedef Test<DeviceFixture> CommandQueueCSTest;
@@ -733,40 +727,6 @@ HWTEST_F(WaitForQueueCompletionTests, whenFinishIsCalledThenCallWaitWithoutQuick
     EXPECT_FALSE(cmdQ->requestedUseQuickKmdSleep);
 }
 
-constexpr char sipPattern[] = {2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 39, 41};
-static_assert(false == isAligned<MemoryConstants::cacheLineSize>(sizeof(sipPattern)),
-              "Will be checking for automatic cacheline alignment, so pattern length must not be a multiple of cacheline");
-constexpr size_t alignedPatternSize = alignUp(sizeof(sipPattern), MemoryConstants::cacheLineSize);
-
-TEST(CommandQueueGetIndirectHeap, whenNewInstructionHeapIsBeingCreatedThenCommandStreamReceiverCanReserveAMemoryBlockAtItsBegining) {
-    auto mockDevice = std::unique_ptr<MockDevice>(MockDevice::create<MockDevice>(nullptr));
-    MockCommandStreamReceiver *csr = new MockCommandStreamReceiver;
-    mockDevice->resetCommandStreamReceiver(csr);
-
-    csr->instructionHeapReserveredData.assign(sipPattern, sipPattern + sizeof(sipPattern));
-    MockCommandQueue cmdQ{nullptr, mockDevice.get(), nullptr};
-    IndirectHeap &heap = cmdQ.getIndirectHeap(OCLRT::IndirectHeap::INSTRUCTION, 8192);
-    EXPECT_LE(8192U, heap.getAvailableSpace());
-    EXPECT_EQ(alignedPatternSize, heap.getUsed());
-
-    ASSERT_LE(sizeof(sipPattern), heap.getMaxAvailableSpace());
-    char *reservedBlock = reinterpret_cast<char *>(heap.getCpuBase());
-    auto dataFoundInReservedBlock = ArrayRef<char>(reservedBlock, sizeof(sipPattern));
-    auto expectedData = ArrayRef<char>(csr->instructionHeapReserveredData);
-    EXPECT_THAT(dataFoundInReservedBlock, testing::ContainerEq(expectedData));
-}
-
-TEST(CommandQueueGetIndirectHeap, whenCheckingForCsrInstructionHeapReservedBlockSizeThenCachelineAlignmentIsExpected) {
-    auto mockDevice = std::unique_ptr<MockDevice>(MockDevice::create<MockDevice>(nullptr));
-    MockCommandStreamReceiver *csr = new MockCommandStreamReceiver;
-    mockDevice->resetCommandStreamReceiver(csr);
-    csr->instructionHeapReserveredData.assign(sipPattern, sipPattern + sizeof(sipPattern));
-    MockCommandQueue cmdQ{nullptr, mockDevice.get(), nullptr};
-
-    EXPECT_GE(alignedPatternSize, csr->getInstructionHeapCmdStreamReceiverReservedSize());
-    EXPECT_EQ(alignedPatternSize, cmdQ.getInstructionHeapReservedBlockSize());
-}
-
 TEST(CommandQueue, givenEnqueueAcquireSharedObjectsWhenNoObjectsThenReturnSuccess) {
     MockContext context;
     CommandQueue cmdQ(&context, nullptr, 0);

unit_tests/command_stream/command_stream_receiver_hw_tests.cpp

@@ -2763,6 +2763,7 @@ HWTEST_F(CommandStreamReceiverFlushTaskTests, givenCsrInBatchingModeWhenSusbsequ
                        dispatchFlags);
 
     EXPECT_EQ(expectedUsed + additionalSize, mockCsr->peekTotalMemoryUsed());
+    mockCsr->flushBatchedSubmissions();
 }
 
 struct MockedMemoryManager : public OsAgnosticMemoryManager {

unit_tests/command_stream/command_stream_receiver_tests.cpp

@@ -260,36 +260,4 @@ TEST_F(CommandStreamReceiverTest, givenForced32BitAddressingWhenDebugSurfaceIsAl
 HWTEST_F(CommandStreamReceiverTest, givenDefaultCommandStreamReceiverThenDefaultDispatchingPolicyIsImmediateSubmission) {
     auto &csr = pDevice->getUltCommandStreamReceiver<FamilyType>();
     EXPECT_EQ(CommandStreamReceiver::DispatchMode::ImmediateDispatch, csr.dispatchMode);
-}
-
-TEST(CommandStreamReceiver, cmdStreamReceiverReservedBlockInInstructionHeapIsBasedOnPreemptionHelper) {
-    auto mockDevice = std::unique_ptr<MockDevice>(MockDevice::create<MockDevice>(nullptr));
-    mockDevice->setPreemptionMode(PreemptionMode::MidThread);
-    {
-        MockBuiltins mockBuiltins;
-        mockBuiltins.overrideGlobalBuiltins();
-        {
-            auto sipOverride = std::unique_ptr<OCLRT::SipKernel>(new OCLRT::SipKernel(OCLRT::SipKernelType::Csr, getSipProgramWithCustomBinary()));
-            mockBuiltins.overrideSipKernel(std::move(sipOverride));
-        }
-
-        size_t reservedSize = mockDevice->getCommandStreamReceiver().getInstructionHeapCmdStreamReceiverReservedSize();
-        size_t expectedSize = OCLRT::PreemptionHelper::getInstructionHeapSipKernelReservedSize(*mockDevice);
-        EXPECT_NE(0U, expectedSize);
-        EXPECT_EQ(expectedSize, reservedSize);
-        ASSERT_LE(expectedSize, reservedSize);
-
-        StackVec<char, 4096> cmdStreamIhBuffer;
-        cmdStreamIhBuffer.resize(reservedSize);
-        LinearStream cmdStreamReceiverInstrucionHeap{cmdStreamIhBuffer.begin(), cmdStreamIhBuffer.size()};
-        mockDevice->getCommandStreamReceiver().initializeInstructionHeapCmdStreamReceiverReservedBlock(cmdStreamReceiverInstrucionHeap);
-
-        StackVec<char, 4096> preemptionHelperIhBuffer;
-        preemptionHelperIhBuffer.resize(expectedSize);
-        LinearStream preemptionHelperInstrucionHeap{preemptionHelperIhBuffer.begin(), preemptionHelperIhBuffer.size()};
-        PreemptionHelper::initializeInstructionHeapSipKernelReservedBlock(preemptionHelperInstrucionHeap, *mockDevice);
-
-        cmdStreamIhBuffer.resize(expectedSize);
-        EXPECT_THAT(preemptionHelperIhBuffer, testing::ContainerEq(cmdStreamIhBuffer));
-    }
-}
+}