Enable memory transfer between images

Unlock flow for multi device setup in:
- enqueueCopyImage

Update cleanAllGraphicsAllocations test

Related-To: NEO-4589
Signed-off-by: Krzysztof Gibala <[email protected]>

Author: kgibala

opencl/source/command_queue/enqueue_copy_image.h

@@ -32,6 +32,11 @@ cl_int CommandQueueHw<GfxFamily>::enqueueCopyImage(
     const cl_event *eventWaitList,
     cl_event *event) {
 
+    auto rootDeviceIndex = getDevice().getRootDeviceIndex();
+
+    srcImage->getMigrateableMultiGraphicsAllocation().ensureMemoryOnDevice(*getDevice().getMemoryManager(), rootDeviceIndex);
+    dstImage->getMigrateableMultiGraphicsAllocation().ensureMemoryOnDevice(*getDevice().getMemoryManager(), rootDeviceIndex);
+
     auto &builder = BuiltInDispatchBuilderOp::getBuiltinDispatchInfoBuilder(EBuiltInOps::CopyImageToImage3d,
                                                                             this->getClDevice());
     BuiltInOwnershipWrapper builtInLock(builder);

opencl/test/unit_test/mem_obj/buffer_tests.cpp

@@ -1809,14 +1809,20 @@ TEST_F(BufferTransferTests, givenBufferWhenTransferFromHostPtrCalledThenCopyRequ
 
 using MultiRootDeviceBufferTest = MultiRootDeviceFixture;
 
-TEST_F(MultiRootDeviceBufferTest, WhenCleanAllGraphicsAllocationsCalledThenGraphicsAllocationsAreProperlyRemoved) {
+TEST_F(MultiRootDeviceBufferTest, WhenCleanAllGraphicsAllocationsCalledThenGraphicsAllocationsAreProperlyRemovedAccordingToIsParentObjectFlag) {
     AllocationInfoType allocationInfo;
     allocationInfo.resize(3u);
 
     allocationInfo[1u] = {};
     allocationInfo[1u].memory = mockMemoryManager->allocateGraphicsMemoryWithProperties(MockAllocationProperties{1u, MemoryConstants::pageSize});
 
-    Buffer::cleanAllGraphicsAllocations(*context, *context->getMemoryManager(), allocationInfo, false);
+    bool isParentObject = true;
+    Buffer::cleanAllGraphicsAllocations(*context, *context->getMemoryManager(), allocationInfo, isParentObject);
+    EXPECT_EQ(mockMemoryManager->freeGraphicsMemoryCalled, 0u);
+
+    isParentObject = false;
+    Buffer::cleanAllGraphicsAllocations(*context, *context->getMemoryManager(), allocationInfo, isParentObject);
+    EXPECT_EQ(mockMemoryManager->freeGraphicsMemoryCalled, 1u);
 }
 
 TEST_F(MultiRootDeviceBufferTest, WhenBufferIsCreatedThenBufferGraphicsAllocationHasCorrectRootDeviceIndex) {

opencl/test/unit_test/mem_obj/image_tests.cpp

@@ -1764,3 +1764,54 @@ TEST_F(ImageMultiRootDeviceTests, WhenImageIsCreatedAndEnqueueReadImageCalledThe
 
     alignedFree(hostBuffer);
 }
+
+TEST_F(ImageMultiRootDeviceTests, WhenImageIsCreatedAndEnqueueCopyImageCalledThenImagesMultiGraphicsAllocationLastUsedRootDeviceIndexHasCorrectRootDeviceIndex) {
+    REQUIRE_IMAGES_OR_SKIP(defaultHwInfo);
+
+    cl_int retVal = 0;
+
+    size_t height = 4;
+    size_t width = 4;
+    size_t region[] = {width * height, 1, 1};
+    size_t orgin[] = {0, 0, 0};
+
+    cl_image_format format;
+    format.image_channel_order = CL_RGBA;
+    format.image_channel_data_type = CL_UNSIGNED_INT8;
+
+    cl_image_desc desc{};
+    desc.image_type = CL_MEM_OBJECT_IMAGE2D;
+    desc.image_width = width * sizeof(unsigned int);
+    desc.image_height = height * sizeof(unsigned int);
+
+    cl_mem_flags flags = CL_MEM_READ_WRITE;
+
+    auto surfaceFormat = Image::getSurfaceFormatFromTable(
+        flags, &format, context->getDevice(0)->getHardwareInfo().capabilityTable.supportsOcl21Features);
+
+    std::unique_ptr<Image> image1(Image::create(context.get(), MemoryPropertiesHelper::createMemoryProperties(flags, 0, 0, &context->getDevice(0)->getDevice()), flags, 0, surfaceFormat, &desc, nullptr, retVal));
+    std::unique_ptr<Image> image2(Image::create(context.get(), MemoryPropertiesHelper::createMemoryProperties(flags, 0, 0, &context->getDevice(0)->getDevice()), flags, 0, surfaceFormat, &desc, nullptr, retVal));
+
+    auto cmdQ1 = context->getSpecialQueue(1u);
+    cmdQ1->enqueueCopyImage(image1.get(), image2.get(), orgin, orgin, region, 0, nullptr, nullptr);
+    EXPECT_EQ(image1->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 1u);
+    EXPECT_EQ(image2->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 1u);
+
+    cmdQ1->enqueueCopyImage(image1.get(), image2.get(), orgin, orgin, region, 0, nullptr, nullptr);
+    EXPECT_EQ(image1->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 1u);
+    EXPECT_EQ(image2->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 1u);
+
+    auto cmdQ2 = context->getSpecialQueue(2u);
+    cmdQ2->enqueueCopyImage(image1.get(), image2.get(), orgin, orgin, region, 0, nullptr, nullptr);
+    EXPECT_EQ(image1->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 2u);
+    EXPECT_EQ(image2->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 2u);
+
+    cmdQ1->enqueueCopyImage(image1.get(), image2.get(), orgin, orgin, region, 0, nullptr, nullptr);
+    EXPECT_EQ(image1->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 1u);
+    EXPECT_EQ(image2->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 1u);
+
+    static_cast<MemoryAllocation *>(image1->getMigrateableMultiGraphicsAllocation().getGraphicsAllocation(2u))->overrideMemoryPool(MemoryPool::LocalMemory);
+    cmdQ2->enqueueCopyImage(image1.get(), image2.get(), orgin, orgin, region, 0, nullptr, nullptr);
+    EXPECT_EQ(image1->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 2u);
+    EXPECT_EQ(image2->getMultiGraphicsAllocation().getLastUsedRootDeviceIndex(), 2u);
+}