Refactor clSharedMemAllocINTEL

- Choose allocation path according to given device,
- Connect device in createUnifiedAllocationWithDeviceStorage
- Fix type in clGetMemAllocInfoINTEL

Change-Id: I9e743001b4c032a712c939c2917f16de0a61b100
Signed-off-by: Krzysztof Gibala <krzysztof.gibala@intel.com>

Author: kgibala

core/memory_manager/unified_memory_manager.cpp

@@ -238,6 +238,7 @@ void *SVMAllocsManager::createUnifiedAllocationWithDeviceStorage(uint32_t rootDe
     SvmAllocationData allocData;
     allocData.gpuAllocation = allocationGpu;
     allocData.cpuAllocation = allocationCpu;
+    allocData.device = unifiedMemoryProperties.device;
     allocData.size = size;
 
     this->SVMAllocs.insert(allocData);

runtime/api/api.cpp

@@ -3525,7 +3525,7 @@ void *clSharedMemAllocINTEL(
 
     ErrorCodeHelper err(errcodeRet, CL_SUCCESS);
 
-    auto retVal = validateObjects(WithCastToInternal(context, &neoContext), WithCastToInternal(device, &neoDevice));
+    auto retVal = validateObjects(WithCastToInternal(context, &neoContext));
 
     if (retVal != CL_SUCCESS) {
         err.set(retVal);
@@ -3547,8 +3547,13 @@ void *clSharedMemAllocINTEL(
     }
 
     unifiedMemoryProperties.device = device;
-    unifiedMemoryProperties.subdeviceBitfield = neoDevice->getDefaultEngine().osContext->getDeviceBitfield();
+    if (!device) {
+        return neoContext->getSVMAllocsManager()->createUnifiedMemoryAllocation(neoContext->getDevice(0)->getRootDeviceIndex(), size, unifiedMemoryProperties);
+    }
+
+    validateObjects(WithCastToInternal(device, &neoDevice));
 
+    unifiedMemoryProperties.subdeviceBitfield = neoDevice->getDefaultEngine().osContext->getDeviceBitfield();
     return neoContext->getSVMAllocsManager()->createSharedUnifiedMemoryAllocation(neoContext->getDevice(0)->getRootDeviceIndex(), size, unifiedMemoryProperties, neoContext->getSpecialQueue());
 }
 
@@ -3626,7 +3631,7 @@ cl_int clGetMemAllocInfoINTEL(
         return retVal;
     }
     case CL_MEM_ALLOC_FLAGS_INTEL: {
-        retVal = info.set<uint32_t>(unifiedMemoryAllocation->allocationFlagsProperty.allAllocFlags);
+        retVal = info.set<cl_mem_alloc_flags_intel>(unifiedMemoryAllocation->allocationFlagsProperty.allAllocFlags);
         return retVal;
     }
     case CL_MEM_ALLOC_DEVICE_INTEL: {

unit_tests/api/cl_unified_shared_memory_tests.inl

@@ -437,6 +437,24 @@ TEST(clUnifiedSharedMemoryTests, givenSharedAllocationWhenItIsQueriedForDeviceTh
     EXPECT_EQ(CL_SUCCESS, retVal);
 }
 
+TEST(clUnifiedSharedMemoryTests, givenSharedAllocationWithoutDeviceWhenItIsQueriedForDeviceThenNullIsReturned) {
+    MockContext mockContext;
+    cl_int retVal = CL_SUCCESS;
+    size_t paramValueSizeRet = 0;
+    auto unifiedMemorySharedAllocation = clSharedMemAllocINTEL(&mockContext, nullptr, nullptr, 4, 0, &retVal);
+
+    cl_device_id returnedDevice;
+
+    retVal = clGetMemAllocInfoINTEL(&mockContext, unifiedMemorySharedAllocation, CL_MEM_ALLOC_DEVICE_INTEL, sizeof(returnedDevice), &returnedDevice, &paramValueSizeRet);
+
+    EXPECT_EQ(CL_SUCCESS, retVal);
+    EXPECT_EQ(paramValueSizeRet, sizeof(returnedDevice));
+    EXPECT_EQ(returnedDevice, nullptr);
+
+    retVal = clMemFreeINTEL(&mockContext, unifiedMemorySharedAllocation);
+    EXPECT_EQ(CL_SUCCESS, retVal);
+}
+
 TEST(clUnifiedSharedMemoryTests, givenHostAllocationWhenItIsQueriedForDeviceThenProperDeviceIsReturned) {
     MockContext mockContext;
     cl_int retVal = CL_SUCCESS;

unit_tests/memory_manager/unified_memory_manager_tests.cpp

@@ -250,6 +250,7 @@ TEST_F(SVMMemoryAllocatorTest, whenSharedAllocationIsCreatedThenItIsStoredWithPr
     SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties;
     unifiedMemoryProperties.memoryType = InternalMemoryType::SHARED_UNIFIED_MEMORY;
     auto allocationSize = 4096u;
+
     auto ptr = svmManager->createSharedUnifiedMemoryAllocation(0, 4096u, unifiedMemoryProperties, &cmdQ);
     EXPECT_NE(nullptr, ptr);
     auto allocation = svmManager->getSVMAlloc(ptr);
@@ -267,11 +268,13 @@ TEST_F(SVMMemoryAllocatorTest, whenSharedAllocationIsCreatedThenItIsStoredWithPr
 
 TEST_F(SVMLocalMemoryAllocatorTest, whenSharedAllocationIsCreatedWithDebugFlagSetThenItIsStoredWithProperTypeInAllocationMapAndHasCpuAndGpuStorage) {
     MockCommandQueue cmdQ;
+    MockContext mockContext;
     DebugManagerStateRestore restore;
     DebugManager.flags.AllocateSharedAllocationsWithCpuAndGpuStorage.set(true);
 
     SVMAllocsManager::UnifiedMemoryProperties unifiedMemoryProperties;
     unifiedMemoryProperties.memoryType = InternalMemoryType::SHARED_UNIFIED_MEMORY;
+    unifiedMemoryProperties.device = mockContext.getDevice(0u);
     auto allocationSize = 4096u;
     auto ptr = svmManager->createSharedUnifiedMemoryAllocation(0, 4096u, unifiedMemoryProperties, &cmdQ);
     EXPECT_NE(nullptr, ptr);
@@ -280,6 +283,7 @@ TEST_F(SVMLocalMemoryAllocatorTest, whenSharedAllocationIsCreatedWithDebugFlagSe
     EXPECT_NE(nullptr, allocation->gpuAllocation);
     EXPECT_EQ(InternalMemoryType::SHARED_UNIFIED_MEMORY, allocation->memoryType);
     EXPECT_EQ(allocationSize, allocation->size);
+    EXPECT_EQ(mockContext.getDevice(0u), allocation->device);
 
     EXPECT_EQ(alignUp(allocationSize, 2u * MB), allocation->gpuAllocation->getUnderlyingBufferSize());
     EXPECT_EQ(alignUp(allocationSize, 2u * MB), allocation->cpuAllocation->getUnderlyingBufferSize());