[FFmpeg-devel] [PATCH] Vulkan hwcontext and filters
Anton Khirnov
anton at khirnov.net
Tue Jan 21 14:08:05 EET 2020
Quoting Lynne (2020-01-10 22:05:21)
> commit d5f1bbc61fab452803443511b1241931169359b7
> Author: Lynne <dev at lynne.ee>
> Date: Wed Aug 28 21:58:10 2019 +0100
>
> lavu: add Vulkan hwcontext code
>
> This commit adds the necessary code to initialize and use a Vulkan device
> within the hwcontext libavutil framework.
> Currently direct mapping to VAAPI and DRM frames is functional, and
> transfers to CUDA and native frames are supported.
>
> Lets hope the future Vulkan video decode extension fits well within this
> framework.
>
> /*
> * This file is part of FFmpeg.
> *
> * FFmpeg is free software; you can redistribute it and/or
> * modify it under the terms of the GNU Lesser General Public
> * License as published by the Free Software Foundation; either
> * version 2.1 of the License, or (at your option) any later version.
> *
> * FFmpeg is distributed in the hope that it will be useful,
> * but WITHOUT ANY WARRANTY; without even the implied warranty of
> * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
> * Lesser General Public License for more details.
> *
> * You should have received a copy of the GNU Lesser General Public
> * License along with FFmpeg; if not, write to the Free Software
> * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
> */
>
> #ifndef AVUTIL_HWCONTEXT_VULKAN_H
> #define AVUTIL_HWCONTEXT_VULKAN_H
>
> #include <vulkan/vulkan.h>
>
> /**
> * @file
> * API-specific header for AV_HWDEVICE_TYPE_VULKAN.
> *
> * For user-allocated pools, AVHWFramesContext.pool must return AVBufferRefs
> * with the data pointer set to an AVVkFrame.
> */
>
> /**
> * Main Vulkan context, allocated as AVHWDeviceContext.hwctx.
> * All of these can be set before init to change what the context uses
> */
> typedef struct AVVulkanDeviceContext {
> /**
> * Custom memory allocator, else NULL
> */
> const VkAllocationCallbacks *alloc;
> /**
> * Instance
> */
> VkInstance inst;
> /**
> * Physical device
> */
> VkPhysicalDevice phys_dev;
> /**
> * Activated physical device
> */
> VkDevice act_dev;
Where is this 'activated' terminology from? From what I can see, the
spec calls this a 'logical' device. Would be nice to be consistent with
it.
> /**
> * Queue family index for graphics
> */
> int queue_family_index;
What is this for? I don't see it used anywhere.
> /**
> * Queue family index for transfer ops only. By default, the priority order
> * is dedicated transfer > dedicated compute > graphics.
> */
IIUC the second sentence talks about how the internal device
creation/derivation code works, but that isn't made very clear. Maybe
make it a
@note the av_hwdevice_create() will set those fields like this...
> int queue_family_tx_index;
> /**
> * Queue family index for compute ops. Will be equal to the graphics
> * one unless a dedicated transfer queue is found.
> */
> int queue_family_comp_index;
Same comment as above.
> /**
> * The UUID of the selected physical device.
> */
> uint8_t device_uuid[VK_UUID_SIZE];
Do we really need to export this here? It can be queried from phys_dev,
no?
> } AVVulkanDeviceContext;
>
> /**
> * Allocated as AVHWFramesContext.hwctx, used to set pool-specific options
> */
> typedef struct AVVulkanFramesContext {
> /**
> * Controls the tiling of output frames.
> */
> VkImageTiling tiling;
> /**
> * Defines extra usage of output frames. This is bitwise OR'd with the
> * standard usage flags (SAMPLED, STORAGE, TRANSFER_SRC and TRANSFER_DST).
> */
> VkImageUsageFlagBits usage;
> /**
> * Extension data for image creation. By default, if the extension is
> * available, this will be chained to a VkImageFormatListCreateInfoKHR.
> */
> void *create_pnext;
> /**
> * Extension data for memory allocation. Must have as many entries as
> * the number of planes of the sw_format.
> * This will be chained to VkExportMemoryAllocateInfo, which is used
> * to make all pool images exportable to other APIs.
> */
> void *alloc_pnext[AV_NUM_DATA_POINTERS];
> } AVVulkanFramesContext;
>
> /*
> * Frame structure, the VkFormat of the image will always match
> * the pool's sw_format.
> * All frames, imported or allocated, will be created with the
> * VK_IMAGE_CREATE_ALIAS_BIT flag set, so the memory may be aliased if needed.
> */
> typedef struct AVVkFrame {
> /**
> * Vulkan images to which the memory is bound to.
> */
> VkImage img[AV_NUM_DATA_POINTERS];
>
> /**
> * Same tiling must be used for all images.
> */
> VkImageTiling tiling;
>
> /**
> * Memory backing the images. Could be less than the amount of images
> * if importing from a DRM or VAAPI frame.
> */
> VkDeviceMemory mem[AV_NUM_DATA_POINTERS];
> size_t size[AV_NUM_DATA_POINTERS];
>
> /**
> * OR'd flags for all memory allocated
> */
> VkMemoryPropertyFlagBits flags;
>
> /**
> * Updated after every barrier
> */
> VkAccessFlagBits access[AV_NUM_DATA_POINTERS];
> VkImageLayout layout[AV_NUM_DATA_POINTERS];
>
> /**
> * Per-image semaphores. Must not be freed manually. Must be waited on
> * and signalled at every queue submission.
> */
> VkSemaphore sem[AV_NUM_DATA_POINTERS];
>
> /**
> * Internal data.
> */
> struct AVVkFrameInternal *internal;
> } AVVkFrame;
This is a pretty big and complicated struct and its size has to be
hardcoded in the ABI IIUC. Do we want a constructor for it?
--
Anton Khirnov
More information about the ffmpeg-devel
mailing list