VkSharingMode - Buffer and image sharing modes

Description

• SHARING_MODE_EXCLUSIVE specifies that access to any range or image subresource of the object will be exclusive to a single queue family at a time.

• SHARING_MODE_CONCURRENT specifies that concurrent access to any range or image subresource of the object from multiple queue families is supported.

SHARING_MODE_CONCURRENT may result in lower performance access to the buffer or image than SHARING_MODE_EXCLUSIVE.

Ranges of buffers and image subresources of image objects created using SHARING_MODE_EXCLUSIVE must only be accessed by queues in the queue family that has ownership of the resource. Upon creation, such resources are not owned by any queue family; ownership is implicitly acquired upon first use within a queue. Once a resource using SHARING_MODE_EXCLUSIVE is owned by some queue family, unless the maintenance9 feature is enabled, the application must perform a queue family ownership transfer if it wishes to make the memory contents of a range or image subresource accessible to a different queue family. SHARING_MODE_EXCLUSIVE resources that are already owned by a queue family may be acquired by a different queue family without a queue family ownership transfer, but unless the maintenance9 feature is enabled, their contents become undefined.

If the maintenance9 feature is enabled, the contents of buffer resources, and of linear image resources (i.e., those created with tiling set to IMAGE_TILING_LINEAR) are always preserved when they are implicitly acquired by a different queue family on the same logical device (i.e., neither queue family is QUEUE_FAMILY_FOREIGN_EXT or QUEUE_FAMILY_EXTERNAL). This means that whenever the maintenance9 feature is enabled, explicit queue family ownership transfers of buffer and linear image resources between different queue families on the same logical device are optional.

Additionally, if the maintenance9 feature is enabled, the contents of some optimal image resources (i.e., those created with IMAGE_TILING_OPTIMAL) are always preserved when they are implicitly acquired by a different queue family on the same logical device (i.e., neither queue family is QUEUE_FAMILY_FOREIGN_EXT or QUEUE_FAMILY_EXTERNAL). This applies only to optimal images that are being implicitly acquired by a queue family whose index bit is set in the current queue family’s QueueFamilyOwnershipTransferPropertiesKHR::optimalImageTransferToQueueFamilies, and that were created without any of the following bits set in usage:

• IMAGE_USAGE_COLOR_ATTACHMENT_BIT
• IMAGE_USAGE_DEPTH_STENCIL_ATTACHMENT_BIT
• IMAGE_USAGE_TRANSIENT_ATTACHMENT_BIT
• IMAGE_USAGE_INPUT_ATTACHMENT_BIT
• IMAGE_USAGE_ATTACHMENT_FEEDBACK_LOOP_BIT_EXT
• IMAGE_USAGE_FRAGMENT_SHADING_RATE_ATTACHMENT_BIT_KHR

This means that whenever the maintenance9 feature is enabled, explicit queue family ownership transfers of such image resources between such combinations of queue families are optional. For all other optimal images and/or combinations of queue families, the application must still perform an explicit queue family ownership transfer if it wishes to make the memory contents of an optimal image subresource already owned by a queue family accessible to a different queue family.

Applications are allowed to perform explicit queue family ownership transfers in circumstances where they are not required, but there is no functional nor performance advantage in doing so. Performing explicit transfers in such cases remains supported for backward compatibility and is not recommended for new applications.

Before being used on the first queue, images still require a layout transition from these layouts:

• IMAGE_LAYOUT_UNDEFINED
• IMAGE_LAYOUT_PREINITIALIZED
• IMAGE_LAYOUT_ZERO_INITIALIZED_EXT

A queue family can take ownership of an image subresource, tensor subresource, or buffer range of a resource created with SHARING_MODE_EXCLUSIVE, without an ownership transfer, in the same way as for a resource that was just created; however, taking ownership in this way has the effect that the contents of the image subresource or buffer range are undefined.

Ranges of buffers, tensor subresources of tensor objects, and image subresources of image objects created using SHARING_MODE_CONCURRENT must only be accessed by queues from the queue families specified through the queueFamilyIndexCount and pQueueFamilyIndices members of the corresponding create info structures.

See Also

VK_VERSION_1_0, BufferCreateInfo, ImageCreateInfo, PhysicalDeviceImageDrmFormatModifierInfoEXT, SwapchainCreateInfoKHR, TensorCreateInfoARM