src/appendices/VK_KHR_external_memory.adoc

// Copyright 2016-2024 The Khronos Group Inc.
//
// SPDX-License-Identifier: CC-BY-4.0

include::{generated}/meta/{refprefix}VK_KHR_external_memory.adoc[]

=== Other Extension Metadata

*Last Modified Date*::
    2016-10-20
*IP Status*::
    No known IP claims.
*Interactions and External Dependencies*::
  - Interacts with `apiext:VK_KHR_dedicated_allocation`.
  - Interacts with `apiext:VK_NV_dedicated_allocation`.
*Contributors*::
  - Faith Ekstrand, Intel
  - Ian Elliott, Google
  - Jesse Hall, Google
  - Tobias Hector, Imagination Technologies
  - James Jones, NVIDIA
  - Jeff Juliano, NVIDIA
  - Matthew Netsch, Qualcomm Technologies, Inc.
  - Daniel Rakos, AMD
  - Carsten Rohde, NVIDIA
  - Ray Smith, ARM
  - Lina Versace, Google

=== Description

An application may wish to reference device memory in multiple Vulkan
logical devices or instances, in multiple processes, and/or in multiple
APIs.
This extension enables an application to export non-Vulkan handles from
Vulkan memory objects such that the underlying resources can be referenced
outside the scope of the Vulkan logical device that created them.

=== Promotion to Vulkan 1.1

All functionality in this extension is included in core Vulkan 1.1, with the
KHR suffix omitted.
The original type, enum and command names are still available as aliases of
the core functionality.

include::{generated}/interfaces/VK_KHR_external_memory.adoc[]

=== Issues

1) How do applications correlate two physical devices across process or
Vulkan instance boundaries?

*RESOLVED*: New device ID fields have been introduced by
`apiext:VK_KHR_external_memory_capabilities`.
These fields, combined with the existing
slink:VkPhysicalDeviceProperties::pname:driverVersion field can be used to
identify compatible devices across processes, drivers, and APIs.
slink:VkPhysicalDeviceProperties::pname:pipelineCacheUUID is not sufficient
for this purpose because despite its description in the specification, it
need only identify a unique pipeline cache format in practice.
Multiple devices may be able to use the same pipeline cache data, and hence
it would be desirable for all of them to have the same pipeline cache UUID.
However, only the same concrete physical device can be used when sharing
memory, so an actual unique device ID was introduced.
Further, the pipeline cache UUID was specific to Vulkan, but correlation
with other, non-extensible APIs is required to enable interoperation with
those APIs.

2) If memory objects are shared between processes and APIs, is this
considered aliasing according to the rules outlined in the
<<resources-memory-aliasing,Memory Aliasing>> section?

*RESOLVED*: Yes.
Applications must take care to obey all restrictions imposed on aliased
resources when using memory across multiple Vulkan instances or other APIs.

3) Are new image layouts or metadata required to specify image layouts and
layout transitions compatible with non-Vulkan APIs, or with other instances
of the same Vulkan driver?

*RESOLVED*: Separate instances of the same Vulkan driver running on the same
GPU should have identical internal layout semantics, so applications have
the tools they need to ensure views of images are consistent between the two
instances.
Other APIs will fall into two categories: Those that are Vulkan- compatible,
and those that are Vulkan-incompatible.
Vulkan-incompatible APIs will require the image to be in the GENERAL layout
whenever they are accessing them.

Note this does not attempt to address cross-device transitions, nor
transitions to engines on the same device which are not visible within the
Vulkan API.
Both of these are beyond the scope of this extension.

4) Is a new barrier flag or operation of some type needed to prepare
external memory for handoff to another Vulkan instance or API and/or receive
it from another instance or API?

*RESOLVED*: Yes.
Some implementations need to perform additional cache management when
transitioning memory between address spaces and other APIs, instances, or
processes which may operate in a separate address space.
Options for defining this transition include:

  * A new structure that can be added to the pname:pNext list in
    slink:VkMemoryBarrier, slink:VkBufferMemoryBarrier, and
    slink:VkImageMemoryBarrier.
  * A new bit in tlink:VkAccessFlags that can be set to indicate an
    "`external`" access.
  * A new bit in tlink:VkDependencyFlags
  * A new special queue family that represents an "`external`" queue.

A new structure has the advantage that the type of external transition can
be described in as much detail as necessary.
However, there is not currently a known need for anything beyond
differentiating between external and internal accesses, so this is likely an
over-engineered solution.
The access flag bit has the advantage that it can be applied at buffer,
image, or global granularity, and semantically it maps pretty well to the
operation being described.
Additionally, the API already includes ename:VK_ACCESS_MEMORY_READ_BIT and
ename:VK_ACCESS_MEMORY_WRITE_BIT which appear to be intended for this
purpose.
However, there is no obvious pipeline stage that would correspond to an
external access, and therefore no clear way to use
ename:VK_ACCESS_MEMORY_READ_BIT or ename:VK_ACCESS_MEMORY_WRITE_BIT.
tlink:VkDependencyFlags and tlink:VkPipelineStageFlags operate at command
granularity rather than image or buffer granularity, which would make an
entire pipeline barrier an internal->external or external->internal barrier.
This may not be a problem in practice, but seems like the wrong scope.
Another downside of tlink:VkDependencyFlags is that it lacks inherent
directionality: there are no ptext:src and ptext:dst variants of it in the
barrier or dependency description semantics, so two bits might need to be
added to describe both internal->external and external->internal
transitions.
Transitioning a resource to a special queue family corresponds well with the
operation of transitioning to a separate Vulkan instance, in that both
operations ideally include scheduling a barrier on both sides of the
transition: Both the releasing and the acquiring queue or process.
Using a special queue family requires adding an additional reserved queue
family index.
Re-using ename:VK_QUEUE_FAMILY_IGNORED would have left it unclear how to
transition a concurrent usage resource from one process to another, since
the semantics would have likely been equivalent to the currently-ignored
transition of
ename:VK_QUEUE_FAMILY_IGNORED{nbsp}->{nbsp}ename:VK_QUEUE_FAMILY_IGNORED.
Fortunately, creating a new reserved queue family index is not invasive.

Based on the above analysis, the approach of transitioning to a special
"`external`" queue family was chosen.

5) Do internal driver memory arrangements and/or other internal driver image
properties need to be exported and imported when sharing images across
processes or APIs.

*RESOLVED*: Some vendors claim this is necessary on their implementations,
but it was determined that the security risks of allowing opaque metadata to
be passed from applications to the driver were too high.
Therefore, implementations which require metadata will need to associate it
with the objects represented by the external handles, and rely on the
dedicated allocation mechanism to associate the exported and imported memory
objects with a single image or buffer.

6) Most prior interoperation and cross-process sharing APIs have been based
on image-level sharing.
Should Vulkan sharing be based on memory-object sharing or image sharing?

*RESOLVED*: These extensions have assumed memory-level sharing is the
correct granularity.
Vulkan is a lower-level API than most prior APIs, and as such attempts to
closely align with to the underlying primitives of the hardware and
system-level drivers it abstracts.
In general, the resource that holds the backing store for both images and
buffers of various types is memory.
Images and buffers are merely metadata containing brief descriptions of the
layout of bits within that memory.

Because memory object-based sharing is aligned with the overall Vulkan API
design, it enables the full range of Vulkan capabilities with external
objects.
External memory can be used as backing for sparse images, for example,
whereas such usage would be awkward at best with a sharing mechanism based
on higher-level primitives such as images.
Further, aligning the mechanism with the API in this way provides some hope
of trivial compatibility with future API enhancements.
If new objects backed by memory objects are added to the API, they too can
be used across processes with minimal additions to the base external memory
APIs.

Earlier APIs implemented interop at a higher level, and this necessitated
entirely separate sharing APIs for images and buffers.
To co-exist and interoperate with those APIs, the Vulkan external sharing
mechanism must accommodate their model.
However, if it can be agreed that memory-based sharing is the more desirable
and forward-looking design, legacy interoperation constraints can be
considered another reason to favor memory-based sharing: while native and
legacy driver primitives that may be used to implement sharing may not be as
low-level as the API here suggests, raw memory is still the least common
denominator among the types.
Image-based sharing can be cleanly derived from a set of base memory- object
sharing APIs with minimal effort, whereas image-based sharing does not
generalize well to buffer or raw-memory sharing.
Therefore, following the general Vulkan design principle of minimalism, it
is better to expose interopability with image-based native and external
primitives via the memory sharing API, and place sufficient limits on their
usage to ensure they can be used only as backing for equivalent Vulkan
images.
This provides a consistent API for applications regardless of which platform
or external API they are targeting, which makes development of multi-API and
multi-platform applications simpler.

7) Should Vulkan define a common external handle type and provide Vulkan
functions to facilitate cross-process sharing of such handles rather than
relying on native handles to define the external objects?

*RESOLVED*: No.
Cross-process sharing of resources is best left to native platforms.
There are myriad security and extensibility issues with such a mechanism,
and attempting to re-solve all those issues within Vulkan does not align
with Vulkan's purpose as a graphics API.
If desired, such a mechanism could be built as a layer or helper library on
top of the opaque native handle defined in this family of extensions.

8) Must implementations provide additional guarantees about state implicitly
included in memory objects for those memory objects that may be exported?

*RESOLVED*: Implementations must ensure that sharing memory objects does not
transfer any information between the exporting and importing instances and
APIs other than that required to share the data contained in the memory
objects explicitly shared.
As specific examples, data from previously freed memory objects that used
the same underlying physical memory, and data from memory objects using
adjacent physical memory must not be visible to applications importing an
exported memory object.

9) Must implementations validate external handles the application provides
as inputs to memory import operations?

*RESOLVED*: Implementations must return an error to the application if the
provided memory handle cannot be used to complete the requested import
operation.
However, implementations need not validate handles are of the exact type
specified by the application.

=== Version History

  * Revision 1, 2016-10-20 (James Jones)
  ** Initial version