Registered Extension Number




Ratification Status


Extension and Version Dependencies


Extension Proposal

Other Extension Metadata

Last Modified Date


  • Shahbaz Youssefi, Google

  • Faith Ekstrand, Collabora

  • Hans-Kristian Arntzen, Valve

  • Piers Daniell, NVIDIA

  • Jan-Harald Fredriksen, Arm

  • James Fitzpatrick, Imagination

  • Daniel Story, Nintendo


This extension allows applications to copy data between host memory and images on the host processor, without staging the data through a GPU-accessible buffer. This removes the need to allocate and manage the buffer and its associated memory. On some architectures it may also eliminate an extra copy operation. This extension additionally allows applications to copy data between images on the host.

To support initializing a new image in preparation for a host copy, it is now possible to transition a new image to VK_IMAGE_LAYOUT_GENERAL or other host-copyable layouts via vkTransitionImageLayoutEXT. Additionally, it is possible to perform copies that preserve the swizzling layout of the image by using the VK_HOST_IMAGE_COPY_MEMCPY_EXT flag. In that case, the memory size needed for copies to or from a buffer can be retrieved by chaining VkSubresourceHostMemcpySizeEXT to pLayout in vkGetImageSubresourceLayout2EXT.

New Commands

New Structures

New Enums

New Bitmasks

New Enum Constants



  • Extending VkFormatFeatureFlagBits2:


  • Extending VkImageUsageFlagBits:


  • Extending VkStructureType:












1) When uploading data to an image, the data is usually loaded from disk. Why not have the application load the data directly into a VkDeviceMemory bound to a buffer (instead of host memory), and use vkCmdCopyBufferToImage? The same could be done when downloading data from an image.

RESOLVED: This may not always be possible. Complicated Vulkan applications such as game engines often have decoupled subsystems for streaming data and rendering. It may be unreasonable to require the streaming subsystem to coordinate with the rendering subsystem to allocate memory on its behalf, especially as Vulkan may not be the only API supported by the engine. In emulation layers, the image data is necessarily provided by the application in host memory, so an optimization as suggested is not possible. Most importantly, the device memory may not be mappable by an application, but still accessible to the driver.

2) Are optimalBufferCopyOffsetAlignment and optimalBufferCopyRowPitchAlignment applicable to host memory as well with the functions introduced by this extension? Or should there be new limits?

RESOLVED: No alignment requirements for the host memory pointer.

3) Should there be granularity requirements for image offsets and extents?

RESOLVED: No granularity requirements, i.e. a granularity of 1 pixel (for non-compressed formats) and 1 texel block (for compressed formats) is assumed.

4) How should the application deal with layout transitions before or after copying to or from images?

RESOLVED: An existing issue with linear images is that when emulating other APIs, it is impossible to know when to transition them as they are written to by the host and then used bindlessly. The copy operations in this extension are affected by the same limitation. A new command is thus introduced by this extension to address this problem by allowing the host to perform an image layout transition between a handful of layouts.

Version History

  • Revision 0, 2021-01-20 (Faith Ekstrand)

    • Initial idea and xml

  • Revision 1, 2023-04-26 (Shahbaz Youssefi)

    • Initial revision

See Also

Document Notes

For more information, see the Vulkan Specification

This page is a generated document. Fixes and changes should be made to the generator scripts, not directly.

Copyright 2014-2023 The Khronos Group Inc.

SPDX-License-Identifier: CC-BY-4.0