Add capability to use malloc for Heap32 base.

- shift page tables to lower bits

Change-Id: I54dcba72255215cf5be75ba425fc27727b0bfd98

Author: MichalMrozek

runtime/aub_mem_dump/aub_mem_dump.cpp

@@ -37,18 +37,18 @@ const uint64_t PageTableTraits<48>::physicalMemory = 0; // 1ull <<addressingBits
 
 const uint64_t PageTableTraits<48>::numPTEntries  = BIT(PageTableTraits<48>::addressingBits - PageTableTraits<48>::NUM_OFFSET_BITS);
 const uint64_t PageTableTraits<48>::sizePT        = BIT(PageTableTraits<48>::addressingBits - PageTableTraits<48>::NUM_OFFSET_BITS) * sizeof(uint64_t);
-const uint64_t PageTableTraits<48>::ptBaseAddress = BIT(34);
+const uint64_t PageTableTraits<48>::ptBaseAddress = BIT(32);
 
 const uint64_t PageTableTraits<48>::numPDEntries  = BIT(PageTableTraits<48>::addressingBits - PageTableTraits<48>::NUM_OFFSET_BITS - PageTableTraits<48>::NUM_PTE_BITS);
 const uint64_t PageTableTraits<48>::sizePD        = BIT(PageTableTraits<48>::addressingBits - PageTableTraits<48>::NUM_OFFSET_BITS - PageTableTraits<48>::NUM_PTE_BITS) * sizeof(uint64_t);
-const uint64_t PageTableTraits<48>::pdBaseAddress = BIT(33);
+const uint64_t PageTableTraits<48>::pdBaseAddress = BIT(31);
 
 const uint64_t PageTableTraits<48>::numPDPEntries  = BIT(PageTableTraits<48>::addressingBits - PageTableTraits<48>::NUM_OFFSET_BITS - PageTableTraits<48>::NUM_PTE_BITS - PageTableTraits<48>::NUM_PDE_BITS);
 const uint64_t PageTableTraits<48>::sizePDP        = BIT(PageTableTraits<48>::addressingBits - PageTableTraits<48>::NUM_OFFSET_BITS - PageTableTraits<48>::NUM_PTE_BITS - PageTableTraits<48>::NUM_PDE_BITS) * sizeof(uint64_t);
-const uint64_t PageTableTraits<48>::pdpBaseAddress = BIT(32);
+const uint64_t PageTableTraits<48>::pdpBaseAddress = BIT(30);
 const uint64_t PageTableTraits<48>::numPML4Entries  = BIT(NUM_PML4_BITS);
 const uint64_t PageTableTraits<48>::sizePML4        = BIT(NUM_PML4_BITS) * sizeof(uint64_t);
-const uint64_t PageTableTraits<48>::pml4BaseAddress = BIT(31);
+const uint64_t PageTableTraits<48>::pml4BaseAddress = BIT(29);
 // clang-format on
 
 void LrcaHelper::setRingTail(void *pLRCIn, uint32_t ringTail) const {

runtime/memory_manager/os_agnostic_memory_manager.cpp

@@ -19,6 +19,9 @@
 namespace OCLRT {
 
 OsAgnosticMemoryManager::~OsAgnosticMemoryManager() {
+    if (DebugManager.flags.UseMallocToObtainHeap32Base.get()) {
+        alignedFreeWrapper(reinterpret_cast<void *>(allocator32Bit->getBase()));
+    }
     applyCommonCleanup();
 }
 
@@ -92,8 +95,15 @@ GraphicsAllocation *OsAgnosticMemoryManager::allocate32BitGraphicsMemory(size_t
     void *ptrAlloc = nullptr;
     auto gpuAddress = allocator32Bit->allocate(allocationSize);
 
-    if (size < 0xfffff000) {
-        ptrAlloc = alignedMallocWrapper(allocationSize, MemoryConstants::allocationAlignment);
+    auto freeCpuPointer = true;
+
+    if (DebugManager.flags.UseMallocToObtainHeap32Base.get()) {
+        ptrAlloc = reinterpret_cast<void *>(gpuAddress);
+        freeCpuPointer = false;
+    } else {
+        if (size < 0xfffff000) {
+            ptrAlloc = alignedMallocWrapper(allocationSize, MemoryConstants::allocationAlignment);
+        }
     }
 
     MemoryAllocation *memoryAllocation = nullptr;
@@ -102,7 +112,7 @@ GraphicsAllocation *OsAgnosticMemoryManager::allocate32BitGraphicsMemory(size_t
         memoryAllocation->is32BitAllocation = true;
         memoryAllocation->gpuBaseAddress = GmmHelper::canonize(allocator32Bit->getBase());
         memoryAllocation->sizeToFree = allocationSize;
-        memoryAllocation->cpuPtrAllocated = true;
+        memoryAllocation->cpuPtrAllocated = freeCpuPointer;
     }
     counter++;
     return memoryAllocation;
@@ -242,6 +252,13 @@ Allocator32bit *OsAgnosticMemoryManager::create32BitAllocator(bool aubUsage) {
     if (is32bit) {
         heap32Base = 0x0;
     }
+
+    if (DebugManager.flags.UseMallocToObtainHeap32Base.get()) {
+        size_t allocationSize = 40 * 1024 * 1024u;
+        allocatorSize = static_cast<uint64_t>(allocationSize);
+        heap32Base = castToUint64(alignedMallocWrapper(allocationSize, MemoryConstants::allocationAlignment));
+    }
+
     return new Allocator32bit(heap32Base, allocatorSize);
 }
 } // namespace OCLRT

runtime/os_interface/DebugVariables_base.inl

@@ -24,6 +24,7 @@ DECLARE_DEBUG_VARIABLE(int32_t, SetCommandStreamReceiver, 0, "Set command stream
 DECLARE_DEBUG_VARIABLE(int32_t, TbxPort, 4321, "TCP-IP port of TBX server")
 DECLARE_DEBUG_VARIABLE(bool, FlattenBatchBufferForAUBDump, false, "Dump multi-level batch buffers to AUB as single, flat batch buffer")
 DECLARE_DEBUG_VARIABLE(bool, AddPatchInfoCommentsForAUBDump, false, "Dump comments containing allocations and patching information")
+DECLARE_DEBUG_VARIABLE(bool, UseMallocToObtainHeap32Base, false, "Instead of using dedicated ranges, use pointer from malloc as heap base.")
 
 /*DEBUG FLAGS*/
 DECLARE_DEBUG_VARIABLE(int32_t, SchedulerSimulationReturnInstance, 0, "prints execution model related debug information")

unit_tests/aub_tests/command_stream/aub_mem_dump_tests.cpp

@@ -26,10 +26,10 @@ std::string getAubFileName(const OCLRT::Device *pDevice, const std::string baseN
 }
 
 TEST(PageTableTraits, when48BitTraitsAreUsedThenPageTableAddressesAreCorrect) {
-    EXPECT_EQ(BIT(34), AubMemDump::PageTableTraits<48>::ptBaseAddress);
-    EXPECT_EQ(BIT(33), AubMemDump::PageTableTraits<48>::pdBaseAddress);
-    EXPECT_EQ(BIT(32), AubMemDump::PageTableTraits<48>::pdpBaseAddress);
-    EXPECT_EQ(BIT(31), AubMemDump::PageTableTraits<48>::pml4BaseAddress);
+    EXPECT_EQ(BIT(32), AubMemDump::PageTableTraits<48>::ptBaseAddress);
+    EXPECT_EQ(BIT(31), AubMemDump::PageTableTraits<48>::pdBaseAddress);
+    EXPECT_EQ(BIT(30), AubMemDump::PageTableTraits<48>::pdpBaseAddress);
+    EXPECT_EQ(BIT(29), AubMemDump::PageTableTraits<48>::pml4BaseAddress);
 }
 
 TEST(PageTableTraits, when32BitTraitsAreUsedThenPageTableAddressesAreCorrect) {

unit_tests/memory_manager/memory_manager_tests.cpp

@@ -1878,3 +1878,19 @@ TEST(ResidencyDataTest, givenResidencyDataWhenUpdateCompletionDataIsCalledThenIt
     EXPECT_EQ(lastFenceValue3, residency.lastFenceValues[1]);
     EXPECT_EQ(lastFenceValue3, residency.getFenceValueForContextId(osContext2.getContextId()));
 }
+
+TEST(Heap32AllocationTests, givenDebugModeWhenMallocIsUsedToCreateAllocationWhenAllocationIsCreatedThenItDoesntRequireCpuPointerCleanup) {
+    DebugManagerStateRestore restore;
+    DebugManager.flags.UseMallocToObtainHeap32Base.set(true);
+    ExecutionEnvironment executionEnvironment;
+    executionEnvironment.incRefInternal();
+    OsAgnosticMemoryManager memoryManager(true, true, executionEnvironment);
+    auto internalBase = memoryManager.allocator32Bit->getBase();
+    EXPECT_NE(0x40000000000ul, internalBase);
+    EXPECT_NE(0x80000000000ul, internalBase);
+    EXPECT_NE(0x0ul, internalBase);
+
+    auto allocation = static_cast<MemoryAllocation *>(memoryManager.allocate32BitGraphicsMemory(4096u, nullptr, AllocationOrigin::EXTERNAL_ALLOCATION));
+    EXPECT_FALSE(allocation->cpuPtrAllocated);
+    memoryManager.freeGraphicsMemory(allocation);
+}

unit_tests/test_files/igdrcl.config

@@ -88,4 +88,5 @@ ReturnRawGpuTimestamps = 0
 DoNotRegisterTrimCallback = false
 AddClGlSharing = 0
 EnablePassInlineData = false
-LimitAmountOfReturnedDevices = 0
+LimitAmountOfReturnedDevices = 0
+UseMallocToObtainHeap32Base = false