RetroArch/libretro-common/include/libretro_vulkan.h
Hans-Kristian Arntzen 9f7d0c74d5
[Vulkan] Define and implement v2 of context negotiation interface (#14890)
* [Vulkan] Add v2 of context negotiation interface.

* [Vulkan] Add vkEnumerateInstanceVersion symbol.

* [Vulkan] Implement v2 context negotiation

* [libretro] Add GET_HW_CONTEXT_NEGOTIATION_INTERFACE_SUPPORT.

Works around issues in v1 interface where it was not possible to
query what frontend would do when faces with newer interface versions.

This env-call gives stronger guarantees how things have to work.

* [runloop] Implement GET_HW_CONTEXT_NEGOATION_INTERFACE_SUPPORT.

Fairly trivial. Just report the latest version.

* [Vulkan] Add stricted wording around get_application_info in v2.

* [Vulkan] Be more defensive about get_application_info() in v1.
2023-02-04 16:51:50 +01:00

495 lines
24 KiB
C

/* Copyright (C) 2010-2020 The RetroArch team
*
* ---------------------------------------------------------------------------------------------
* The following license statement only applies to this libretro API header (libretro_vulkan.h)
* ---------------------------------------------------------------------------------------------
*
* Permission is hereby granted, free of charge,
* to any person obtaining a copy of this software and associated documentation files (the "Software"),
* to deal in the Software without restriction, including without limitation the rights to
* use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software,
* and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED,
* INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
* WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
*/
#ifndef LIBRETRO_VULKAN_H__
#define LIBRETRO_VULKAN_H__
#include <libretro.h>
#include <vulkan/vulkan.h>
#define RETRO_HW_RENDER_INTERFACE_VULKAN_VERSION 5
#define RETRO_HW_RENDER_CONTEXT_NEGOTIATION_INTERFACE_VULKAN_VERSION 2
struct retro_vulkan_image
{
VkImageView image_view;
VkImageLayout image_layout;
VkImageViewCreateInfo create_info;
};
typedef void (*retro_vulkan_set_image_t)(void *handle,
const struct retro_vulkan_image *image,
uint32_t num_semaphores,
const VkSemaphore *semaphores,
uint32_t src_queue_family);
typedef uint32_t (*retro_vulkan_get_sync_index_t)(void *handle);
typedef uint32_t (*retro_vulkan_get_sync_index_mask_t)(void *handle);
typedef void (*retro_vulkan_set_command_buffers_t)(void *handle,
uint32_t num_cmd,
const VkCommandBuffer *cmd);
typedef void (*retro_vulkan_wait_sync_index_t)(void *handle);
typedef void (*retro_vulkan_lock_queue_t)(void *handle);
typedef void (*retro_vulkan_unlock_queue_t)(void *handle);
typedef void (*retro_vulkan_set_signal_semaphore_t)(void *handle, VkSemaphore semaphore);
typedef const VkApplicationInfo *(*retro_vulkan_get_application_info_t)(void);
struct retro_vulkan_context
{
VkPhysicalDevice gpu;
VkDevice device;
VkQueue queue;
uint32_t queue_family_index;
VkQueue presentation_queue;
uint32_t presentation_queue_family_index;
};
/* This is only used in v1 of the negotiation interface.
* It is deprecated since it cannot express PDF2 features or optional extensions. */
typedef bool (*retro_vulkan_create_device_t)(
struct retro_vulkan_context *context,
VkInstance instance,
VkPhysicalDevice gpu,
VkSurfaceKHR surface,
PFN_vkGetInstanceProcAddr get_instance_proc_addr,
const char **required_device_extensions,
unsigned num_required_device_extensions,
const char **required_device_layers,
unsigned num_required_device_layers,
const VkPhysicalDeviceFeatures *required_features);
typedef void (*retro_vulkan_destroy_device_t)(void);
/* v2 CONTEXT_NEGOTIATION_INTERFACE only. */
typedef VkInstance (*retro_vulkan_create_instance_wrapper_t)(
void *opaque, const VkInstanceCreateInfo *create_info);
/* v2 CONTEXT_NEGOTIATION_INTERFACE only. */
typedef VkInstance (*retro_vulkan_create_instance_t)(
PFN_vkGetInstanceProcAddr get_instance_proc_addr,
const VkApplicationInfo *app,
retro_vulkan_create_instance_wrapper_t create_instance_wrapper,
void *opaque);
/* v2 CONTEXT_NEGOTIATION_INTERFACE only. */
typedef VkDevice (*retro_vulkan_create_device_wrapper_t)(
VkPhysicalDevice gpu, void *opaque,
const VkDeviceCreateInfo *create_info);
/* v2 CONTEXT_NEGOTIATION_INTERFACE only. */
typedef bool (*retro_vulkan_create_device2_t)(
struct retro_vulkan_context *context,
VkInstance instance,
VkPhysicalDevice gpu,
VkSurfaceKHR surface,
PFN_vkGetInstanceProcAddr get_instance_proc_addr,
retro_vulkan_create_device_wrapper_t create_device_wrapper,
void *opaque);
/* Note on thread safety:
* The Vulkan API is heavily designed around multi-threading, and
* the libretro interface for it should also be threading friendly.
* A core should be able to build command buffers and submit
* command buffers to the GPU from any thread.
*/
struct retro_hw_render_context_negotiation_interface_vulkan
{
/* Must be set to RETRO_HW_RENDER_CONTEXT_NEGOTIATION_INTERFACE_VULKAN. */
enum retro_hw_render_context_negotiation_interface_type interface_type;
/* Usually set to RETRO_HW_RENDER_CONTEXT_NEGOTIATION_INTERFACE_VULKAN_VERSION,
* but can be lower depending on GET_HW_RENDER_CONTEXT_NEGOTIATION_INTERFACE_SUPPORT. */
unsigned interface_version;
/* If non-NULL, returns a VkApplicationInfo struct that the frontend can use instead of
* its "default" application info.
* VkApplicationInfo::apiVersion also controls the target core Vulkan version for instance level functionality.
* Lifetime of the returned pointer must remain until the retro_vulkan_context is initialized.
*
* NOTE: For optimal compatibility with e.g. Android which is very slow to update its loader,
* a core version of 1.1 should be requested. Features beyond that can be requested with extensions.
* Vulkan 1.0 is only appropriate for legacy cores, but is still supported.
* A frontend is free to bump the instance creation apiVersion as necessary if the frontend requires more advanced core features.
*
* v2: This function must not be NULL, and must not return NULL.
* v1: It was not clearly defined if this function could return NULL.
* Frontends should be defensive and provide a default VkApplicationInfo
* if this function returns NULL or if this function is NULL.
*/
retro_vulkan_get_application_info_t get_application_info;
/* If non-NULL, the libretro core will choose one or more physical devices,
* create one or more logical devices and create one or more queues.
* The core must prepare a designated PhysicalDevice, Device, Queue and queue family index
* which the frontend will use for its internal operation.
*
* If gpu is not VK_NULL_HANDLE, the physical device provided to the frontend must be this PhysicalDevice if the call succeeds.
* The core is still free to use other physical devices for other purposes that are private to the core.
*
* The frontend will request certain extensions and layers for a device which is created.
* The core must ensure that the queue and queue_family_index support GRAPHICS and COMPUTE.
*
* If surface is not VK_NULL_HANDLE, the core must consider presentation when creating the queues.
* If presentation to "surface" is supported on the queue, presentation_queue must be equal to queue.
* If not, a second queue must be provided in presentation_queue and presentation_queue_index.
* If surface is not VK_NULL_HANDLE, the instance from frontend will have been created with supported for
* VK_KHR_surface extension.
*
* The core is free to set its own queue priorities.
* Device provided to frontend is owned by the frontend, but any additional device resources must be freed by core
* in destroy_device callback.
*
* If this function returns true, a PhysicalDevice, Device and Queues are initialized.
* If false, none of the above have been initialized and the frontend will attempt
* to fallback to "default" device creation, as if this function was never called.
*/
retro_vulkan_create_device_t create_device;
/* If non-NULL, this callback is called similar to context_destroy for HW_RENDER_INTERFACE.
* However, it will be called even if context_reset was not called.
* This can happen if the context never succeeds in being created.
* destroy_device will always be called before the VkInstance
* of the frontend is destroyed if create_device was called successfully so that the core has a chance of
* tearing down its own device resources.
*
* Only auxillary resources should be freed here, i.e. resources which are not part of retro_vulkan_context.
* v2: Auxillary instance resources created during create_instance can also be freed here.
*/
retro_vulkan_destroy_device_t destroy_device;
/* v2 API: If interface_version is < 2, fields below must be ignored.
* If the frontend does not support interface version 2, the v1 entry points will be used instead. */
/* If non-NULL, this is called to create an instance, otherwise a VkInstance is created by the frontend.
* v1 interface bug: The only way to enable instance features is through core versions signalled in VkApplicationInfo.
* The frontend may request that certain extensions and layers
* are enabled on the VkInstance. Application may add additional features.
* If app is non-NULL, apiVersion controls the minimum core version required by the application.
* Return a VkInstance or VK_NULL_HANDLE. The VkInstance is owned by the frontend.
*
* Rather than call vkCreateInstance directly, a core must call the CreateInstance wrapper provided with:
* VkInstance instance = create_instance_wrapper(opaque, &create_info);
* If the core wishes to create a private instance for whatever reason (relying on shared memory for example),
* it may call vkCreateInstance directly. */
retro_vulkan_create_instance_t create_instance;
/* If non-NULL and frontend recognizes negotiation interface >= 2, create_device2 takes precedence over create_device.
* Similar to create_device, but is extended to better understand new core versions and PDF2 feature enablement.
* Requirements for create_device2 are the same as create_device unless a difference is mentioned.
*
* v2 consideration:
* If the chosen gpu by frontend cannot be supported, a core must return false.
*
* NOTE: "Cannot be supported" is intentionally vaguely defined.
* Refusing to run on an iGPU for a very intensive core with desktop GPU as a minimum spec may be in the gray area.
* Not supporting optional features is not a good reason to reject a physical device, however.
*
* On device creation feature with explicit gpu, a frontend should fall back create_device2 with gpu == VK_NULL_HANDLE and let core
* decide on a supported device if possible.
*
* A core must assume that the explicitly provided GPU is the only guaranteed attempt it has to create a device.
* A fallback may not be attempted if there are particular reasons why only a specific physical device can work,
* but these situations should be esoteric and rare in nature, e.g. a libretro frontend is implemented with external memory
* and only LUID matching would work.
* Cores and frontends should ensure "best effort" when negotiating like this and appropriate logging is encouraged.
*
* v1 note: In the v1 version of create_device, it was never expected that create_device would fail like this,
* and frontends are not expected to attempt fall backs.
*
* Rather than call vkCreateDevice directly, a core must call the CreateDevice wrapper provided with:
* VkDevice device = create_device_wrapper(gpu, opaque, &create_info);
* If the core wishes to create a private device for whatever reason (relying on shared memory for example),
* it may call vkCreateDevice directly.
*
* This allows the frontend to add additional extensions that it requires as well as adjust the PDF2 pNext as required.
* It is also possible adjust the queue create infos in case the frontend desires to allocate some private queues.
*
* The get_instance_proc_addr provided in create_device2 must be the same as create_instance.
*
* NOTE: The frontend must not disable features requested by application.
* NOTE: The frontend must not add any robustness features as some API behavior may change (VK_EXT_descriptor_buffer comes to mind).
* I.e. robustBufferAccess and the like. (nullDescriptor from robustness2 is allowed to be enabled).
*/
retro_vulkan_create_device2_t create_device2;
};
struct retro_hw_render_interface_vulkan
{
/* Must be set to RETRO_HW_RENDER_INTERFACE_VULKAN. */
enum retro_hw_render_interface_type interface_type;
/* Must be set to RETRO_HW_RENDER_INTERFACE_VULKAN_VERSION. */
unsigned interface_version;
/* Opaque handle to the Vulkan backend in the frontend
* which must be passed along to all function pointers
* in this interface.
*
* The rationale for including a handle here (which libretro v1
* doesn't currently do in general) is:
*
* - Vulkan cores should be able to be freely threaded without lots of fuzz.
* This would break frontends which currently rely on TLS
* to deal with multiple cores loaded at the same time.
* - Fixing this in general is TODO for an eventual libretro v2.
*/
void *handle;
/* The Vulkan instance the context is using. */
VkInstance instance;
/* The physical device used. */
VkPhysicalDevice gpu;
/* The logical device used. */
VkDevice device;
/* Allows a core to fetch all its needed symbols without having to link
* against the loader itself. */
PFN_vkGetDeviceProcAddr get_device_proc_addr;
PFN_vkGetInstanceProcAddr get_instance_proc_addr;
/* The queue the core must use to submit data.
* This queue and index must remain constant throughout the lifetime
* of the context.
*
* This queue will be the queue that supports graphics and compute
* if the device supports compute.
*/
VkQueue queue;
unsigned queue_index;
/* Before calling retro_video_refresh_t with RETRO_HW_FRAME_BUFFER_VALID,
* set which image to use for this frame.
*
* If num_semaphores is non-zero, the frontend will wait for the
* semaphores provided to be signaled before using the results further
* in the pipeline.
*
* Semaphores provided by a single call to set_image will only be
* waited for once (waiting for a semaphore resets it).
* E.g. set_image, video_refresh, and then another
* video_refresh without set_image,
* but same image will only wait for semaphores once.
*
* For this reason, ownership transfer will only occur if semaphores
* are waited on for a particular frame in the frontend.
*
* Using semaphores is optional for synchronization purposes,
* but if not using
* semaphores, an image memory barrier in vkCmdPipelineBarrier
* should be used in the graphics_queue.
* Example:
*
* vkCmdPipelineBarrier(cmd,
* srcStageMask = VK_PIPELINE_STAGE_ALL_GRAPHICS_BIT,
* dstStageMask = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
* image_memory_barrier = {
* srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
* dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
* });
*
* The use of pipeline barriers instead of semaphores is encouraged
* as it is simpler and more fine-grained. A layout transition
* must generally happen anyways which requires a
* pipeline barrier.
*
* The image passed to set_image must have imageUsage flags set to at least
* VK_IMAGE_USAGE_TRANSFER_SRC_BIT and VK_IMAGE_USAGE_SAMPLED_BIT.
* The core will naturally want to use flags such as
* VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT and/or
* VK_IMAGE_USAGE_TRANSFER_DST_BIT depending
* on how the final image is created.
*
* The image must also have been created with MUTABLE_FORMAT bit set if
* 8-bit formats are used, so that the frontend can reinterpret sRGB
* formats as it sees fit.
*
* Images passed to set_image should be created with TILING_OPTIMAL.
* The image layout should be transitioned to either
* VK_IMAGE_LAYOUT_GENERIC or VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL.
* The actual image layout used must be set in image_layout.
*
* The image must be a 2D texture which may or not be layered
* and/or mipmapped.
*
* The image must be suitable for linear sampling.
* While the image_view is typically the only field used,
* the frontend may want to reinterpret the texture as sRGB vs.
* non-sRGB for example so the VkImageViewCreateInfo used to
* create the image view must also be passed in.
*
* The data in the pointer to the image struct will not be copied
* as the pNext field in create_info cannot be reliably deep-copied.
* The image pointer passed to set_image must be valid until
* retro_video_refresh_t has returned.
*
* If frame duping is used when passing NULL to retro_video_refresh_t,
* the frontend is free to either use the latest image passed to
* set_image or reuse the older pointer passed to set_image the
* frame RETRO_HW_FRAME_BUFFER_VALID was last used.
*
* Essentially, the lifetime of the pointer passed to
* retro_video_refresh_t should be extended if frame duping is used
* so that the frontend can reuse the older pointer.
*
* The image itself however, must not be touched by the core until
* wait_sync_index has been completed later. The frontend may perform
* layout transitions on the image, so even read-only access is not defined.
* The exception to read-only rule is if GENERAL layout is used for the image.
* In this case, the frontend is not allowed to perform any layout transitions,
* so concurrent reads from core and frontend are allowed.
*
* If frame duping is used, or if set_command_buffers is used,
* the frontend will not wait for any semaphores.
*
* The src_queue_family is used to specify which queue family
* the image is currently owned by. If using multiple queue families
* (e.g. async compute), the frontend will need to acquire ownership of the
* image before rendering with it and release the image afterwards.
*
* If src_queue_family is equal to the queue family (queue_index),
* no ownership transfer will occur.
* Similarly, if src_queue_family is VK_QUEUE_FAMILY_IGNORED,
* no ownership transfer will occur.
*
* The frontend will always release ownership back to src_queue_family.
* Waiting for frontend to complete with wait_sync_index() ensures that
* the frontend has released ownership back to the application.
* Note that in Vulkan, transfering ownership is a two-part process.
*
* Example frame:
* - core releases ownership from src_queue_index to queue_index with VkImageMemoryBarrier.
* - core calls set_image with src_queue_index.
* - Frontend will acquire the image with src_queue_index -> queue_index as well, completing the ownership transfer.
* - Frontend renders the frame.
* - Frontend releases ownership with queue_index -> src_queue_index.
* - Next time image is used, core must acquire ownership from queue_index ...
*
* Since the frontend releases ownership, we cannot necessarily dupe the frame because
* the core needs to make the roundtrip of ownership transfer.
*/
retro_vulkan_set_image_t set_image;
/* Get the current sync index for this frame which is obtained in
* frontend by calling e.g. vkAcquireNextImageKHR before calling
* retro_run().
*
* This index will correspond to which swapchain buffer is currently
* the active one.
*
* Knowing this index is very useful for maintaining safe asynchronous CPU
* and GPU operation without stalling.
*
* The common pattern for synchronization is to receive fences when
* submitting command buffers to Vulkan (vkQueueSubmit) and add this fence
* to a list of fences for frame number get_sync_index().
*
* Next time we receive the same get_sync_index(), we can wait for the
* fences from before, which will usually return immediately as the
* frontend will generally also avoid letting the GPU run ahead too much.
*
* After the fence has signaled, we know that the GPU has completed all
* GPU work related to work submitted in the frame we last saw get_sync_index().
*
* This means we can safely reuse or free resources allocated in this frame.
*
* In theory, even if we wait for the fences correctly, it is not technically
* safe to write to the image we earlier passed to the frontend since we're
* not waiting for the frontend GPU jobs to complete.
*
* The frontend will guarantee that the appropriate pipeline barrier
* in graphics_queue has been used such that
* VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT cannot
* start until the frontend is done with the image.
*/
retro_vulkan_get_sync_index_t get_sync_index;
/* Returns a bitmask of how many swapchain images we currently have
* in the frontend.
*
* If bit #N is set in the return value, get_sync_index can return N.
* Knowing this value is useful for preallocating per-frame management
* structures ahead of time.
*
* While this value will typically remain constant throughout the
* applications lifecycle, it may for example change if the frontend
* suddently changes fullscreen state and/or latency.
*
* If this value ever changes, it is safe to assume that the device
* is completely idle and all synchronization objects can be deleted
* right away as desired.
*/
retro_vulkan_get_sync_index_mask_t get_sync_index_mask;
/* Instead of submitting the command buffer to the queue first, the core
* can pass along its command buffer to the frontend, and the frontend
* will submit the command buffer together with the frontends command buffers.
*
* This has the advantage that the overhead of vkQueueSubmit can be
* amortized into a single call. For this mode, semaphores in set_image
* will be ignored, so vkCmdPipelineBarrier must be used to synchronize
* the core and frontend.
*
* The command buffers in set_command_buffers are only executed once,
* even if frame duping is used.
*
* If frame duping is used, set_image should be used for the frames
* which should be duped instead.
*
* Command buffers passed to the frontend with set_command_buffers
* must not actually be submitted to the GPU until retro_video_refresh_t
* is called.
*
* The frontend must submit the command buffer before submitting any
* other command buffers provided by set_command_buffers. */
retro_vulkan_set_command_buffers_t set_command_buffers;
/* Waits on CPU for device activity for the current sync index to complete.
* This is useful since the core will not have a relevant fence to sync with
* when the frontend is submitting the command buffers. */
retro_vulkan_wait_sync_index_t wait_sync_index;
/* If the core submits command buffers itself to any of the queues provided
* in this interface, the core must lock and unlock the frontend from
* racing on the VkQueue.
*
* Queue submission can happen on any thread.
* Even if queue submission happens on the same thread as retro_run(),
* the lock/unlock functions must still be called.
*
* NOTE: Queue submissions are heavy-weight. */
retro_vulkan_lock_queue_t lock_queue;
retro_vulkan_unlock_queue_t unlock_queue;
/* Sets a semaphore which is signaled when the image in set_image can safely be reused.
* The semaphore is consumed next call to retro_video_refresh_t.
* The semaphore will be signalled even for duped frames.
* The semaphore will be signalled only once, so set_signal_semaphore should be called every frame.
* The semaphore may be VK_NULL_HANDLE, which disables semaphore signalling for next call to retro_video_refresh_t.
*
* This is mostly useful to support use cases where you're rendering to a single image that
* is recycled in a ping-pong fashion with the frontend to save memory (but potentially less throughput).
*/
retro_vulkan_set_signal_semaphore_t set_signal_semaphore;
};
#endif