cuda - 使用 Vulkan VkImage 作为 CUDA cuArray

Question

使用 Vulkan VkImage 作为 CUDA cuArray 的正确方法是什么？

我一直在尝试遵循一些示例，但是我在调​​用 cuExternalMemoryGetMappedMipmappedArray() 时收到 CUDA_ERROR_INVALID_VALUE

以有序的方式提供信息。

我正在使用 CUDA 10.1

基础代码来自https://github.com/SaschaWillems/Vulkan ，特别是我正在使用 01 - Vulkan Gears演示，通过 saveScreenshot 方法丰富 09 - Capturing screenshots

我不会将快照图像保存到文件中，而是将快照图像作为 CUarray 发送到 CUDA。

我已启用以下实例和设备扩展:

        std::vector<const char*> instanceExtensions = {
                VK_EXT_DEBUG_REPORT_EXTENSION_NAME,
                VK_KHR_GET_PHYSICAL_DEVICE_PROPERTIES_2_EXTENSION_NAME,
                VK_KHR_EXTERNAL_MEMORY_CAPABILITIES_EXTENSION_NAME,
                VK_KHR_EXTERNAL_SEMAPHORE_CAPABILITIES_EXTENSION_NAME };

        std::vector<const char*> deviceExtensions = { VK_KHR_EXTERNAL_MEMORY_EXTENSION_NAME,
                VK_KHR_EXTERNAL_MEMORY_FD_EXTENSION_NAME,
                VK_KHR_EXTERNAL_SEMAPHORE_EXTENSION_NAME,
                VK_KHR_EXTERNAL_SEMAPHORE_FD_EXTENSION_NAME };

我有一个 VkImage，创建如下:

        // Create the linear tiled destination image to copy to and to read the memory from
        VkImageCreateInfo imageCreateCI(vks::initializers::imageCreateInfo());
        imageCreateCI.imageType = VK_IMAGE_TYPE_2D;
        // Note that vkCmdBlitImage (if supported) will also do format conversions if the swapchain color format would differ
        imageCreateCI.format = VK_FORMAT_R8G8B8A8_UNORM;
        imageCreateCI.extent.width = width;
        imageCreateCI.extent.height = height;
        imageCreateCI.extent.depth = 1;
        imageCreateCI.arrayLayers = 1;
        imageCreateCI.mipLevels = 1;
        imageCreateCI.initialLayout = VK_IMAGE_LAYOUT_UNDEFINED;
        imageCreateCI.samples = VK_SAMPLE_COUNT_1_BIT;
        imageCreateCI.tiling = VK_IMAGE_TILING_LINEAR;
        imageCreateCI.sharingMode = VK_SHARING_MODE_EXCLUSIVE;
        imageCreateCI.usage = VK_IMAGE_USAGE_TRANSFER_SRC_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;

        VkExternalMemoryImageCreateInfoKHR extImageCreateInfo = {};

        /*
         * Indicate that the memory backing this image will be exported in an
         * fd. In some implementations, this may affect the call to
         * GetImageMemoryRequirements() with this image.
         */
        extImageCreateInfo.sType = VK_STRUCTURE_TYPE_EXTERNAL_MEMORY_IMAGE_CREATE_INFO_KHR;
        extImageCreateInfo.handleTypes |= VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR;

        imageCreateCI.pNext = &extImageCreateInfo;

        // Create the image
        VkImage dstImage;
        VK_CHECK_RESULT(vkCreateImage(device, &imageCreateCI, nullptr, &dstImage));
        // Create memory to back up the image
        VkMemoryRequirements memRequirements;
        VkMemoryAllocateInfo memAllocInfo(vks::initializers::memoryAllocateInfo());
        VkDeviceMemory dstImageMemory;
        vkGetImageMemoryRequirements(device, dstImage, &memRequirements);
        memAllocInfo.allocationSize = memRequirements.size;
        // Memory must be host visible to copy from
        memAllocInfo.memoryTypeIndex = vulkanDevice->getMemoryType(memRequirements.memoryTypeBits, VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_HOST_COHERENT_BIT);

        VkExportMemoryAllocateInfoKHR exportInfo = {};
        exportInfo.sType = VK_STRUCTURE_TYPE_EXPORT_MEMORY_ALLOCATE_INFO_KHR;
        exportInfo.handleTypes = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR;

        memAllocInfo.pNext = &exportInfo;


        VK_CHECK_RESULT(vkAllocateMemory(device, &memAllocInfo, nullptr, &dstImageMemory));
        VK_CHECK_RESULT(vkBindImageMemory(device, dstImage, dstImageMemory, 0));

从那里我会:

获取 Vulkan 内存处理程序:

int CuEncoderImpl::getVulkanMemoryHandle(VkDevice device,
        VkDeviceMemory memory) {
    // Get handle to memory of the VkImage

    int fd = -1;
    VkMemoryGetFdInfoKHR fdInfo = { };
    fdInfo.sType = VK_STRUCTURE_TYPE_MEMORY_GET_FD_INFO_KHR;
    fdInfo.memory = memory;
    fdInfo.handleType = VK_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD_BIT_KHR;

    auto func = (PFN_vkGetMemoryFdKHR) vkGetDeviceProcAddr(device,
            "vkGetMemoryFdKHR");

    if (!func) {
        printf("Failed to locate function vkGetMemoryFdKHR\n");
        return -1;
    }

    VkResult r = func(device, &fdInfo, &fd);
    if (r != VK_SUCCESS) {
        printf("Failed executing vkGetMemoryFdKHR [%d]\n", r);
        return -1;
    }

    return fd;

}

导入内存:

    CUDA_EXTERNAL_MEMORY_HANDLE_DESC memDesc = { };
    memDesc.type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD;
    memDesc.handle.fd = getVulkanMemoryHandle(device, memory);
    memDesc.size = extent.width*extent.height*4;

    CUDA_DRVAPI_CALL(cuImportExternalMemory(&externalMem, &memDesc));

并映射内存:这是失败的步骤。

CUarray CuEncoderImpl::getCUDAArrayFromExternalMemory(const VkExtent3D &extent,const CUexternalMemory &m_extMem) {
    CUmipmappedArray m_mipmapArray;
    CUresult result = CUDA_SUCCESS;
    CUarray array;

    CUDA_ARRAY3D_DESCRIPTOR arrayDesc = { };
    arrayDesc.Width = extent.width;
    arrayDesc.Height = extent.height;
    arrayDesc.Depth = 0;
    arrayDesc.Format = CU_AD_FORMAT_UNSIGNED_INT32;
    arrayDesc.NumChannels = 4;
    arrayDesc.Flags = CUDA_ARRAY3D_SURFACE_LDST;

    CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC mipmapArrayDesc = { };
    mipmapArrayDesc.arrayDesc = arrayDesc;
    mipmapArrayDesc.numLevels = 1;
    mipmapArrayDesc.offset = 0;

    CUDA_DRVAPI_CALL(cuExternalMemoryGetMappedMipmappedArray(&m_mipmapArray, m_extMem, &mipmapArrayDesc));

    CUDA_DRVAPI_CALL(cuMipmappedArrayGetLevel(&array, m_mipmapArray, 0));
    return array;
}

我一直在尝试多种参数组合，但到目前为止都失败了。该错误指向无效参数，但我不知道如何找到问题所在。

唯一有效的方法是将 Vulkan 图像内存映射到主机缓冲区，然后将其复制到 CUDA 数组中......但我想这很昂贵，如果可能的话我想避免它。

Answer 1

郑重声明，我终于让它发挥作用了。

我必须对问题中列出的代码进行的一些注释和修改:

1. Vulkan-CUDA 互操作性被宣传为 CUDA 10 的一项功能，请参阅 CUDA 10 Features revealed
2. 要映射的图像的平铺必须为“VK_IMAGE_TILING_OPTIMAL”

        imageCreateCI.tiling = VK_IMAGE_TILING_OPTIMAL;

必须使用 VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT 分配该图像的内存

        memAllocInfo.memoryTypeIndex = vulkanDevice->getMemoryType(memRequirements.memoryTypeBits, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT);

导入内存时的内存描述符应使用内存需求中返回的内存大小(下面的size是创建图像的代码中的memRequirements.size) :

    CUDA_EXTERNAL_MEMORY_HANDLE_DESC memDesc = { };
    memDesc.type = CU_EXTERNAL_MEMORY_HANDLE_TYPE_OPAQUE_FD;
    memDesc.handle.fd = getVulkanMemoryHandle(device, memory);
    memDesc.size = size;

    CUDA_DRVAPI_CALL(cuImportExternalMemory(&externalMem, &memDesc));

最后，映射数组被描述为 CU_AD_FORMAT_UNSIGNED_INT8，具有四个 channel 和 CUDA_ARRAY3D_COLOR_ATTACHMENT

    CUDA_ARRAY3D_DESCRIPTOR arrayDesc = { };
    arrayDesc.Width = extent.width;
    arrayDesc.Height = extent.height;
    arrayDesc.Depth = 0;
    arrayDesc.Format = CU_AD_FORMAT_UNSIGNED_INT8;
    arrayDesc.NumChannels = 4;
    arrayDesc.Flags = CUDA_ARRAY3D_COLOR_ATTACHMENT;

    CUDA_EXTERNAL_MEMORY_MIPMAPPED_ARRAY_DESC mipmapArrayDesc = { };
    mipmapArrayDesc.arrayDesc = arrayDesc;
    mipmapArrayDesc.numLevels = 1;
    mipmapArrayDesc.offset = 0;

    CUDA_DRVAPI_CALL(cuExternalMemoryGetMappedMipmappedArray(&m_mipmapArray, m_extMem, &mipmapArrayDesc));

经过这些更改后，我能够让它正常工作。我的一些变化是我这边明显的错误(比如大小)，我在第 100 次仔细地重新阅读文档时发现了一些东西，其他的东西是对文档中的提示的猜测，最后是大量的试验和错误.