Skip to content

Latest commit

 

History

History
344 lines (264 loc) · 14.8 KB

VK_KHR_synchronization2.adoc

File metadata and controls

344 lines (264 loc) · 14.8 KB

VK_KHR_synchronization2

Note

Promoted to core in Vulkan 1.3

The VK_KHR_synchronization2 extension provides improvements to pipeline barriers, events, image layout transitions and queue submission. This document shows the difference between the original Vulkan synchronization operations and those provided by the extension. There are also examples of how to update application code to make use of the extension.

Rethinking Pipeline Stages and Access Flags

One main change with the extension is to have pipeline stages and access flags now specified together in memory barrier structures. This makes the connection between the two more obvious.

The only new type of structure needed is VkDependencyInfoKHR, which wraps all the barriers into a single location.

VK_KHR_synchronization2_stage_access

Adding barriers for setting events

Note that with the introduction of VkDependencyInfoKHR that vkCmdSetEvent2KHR, unlike vkCmdSetEvent, has the ability to add barriers. This was added to allow the VkEvent to be more useful. Because the implementation of a synchronization2 VkEvent is likely to be substantially different from a Vulkan 1.2 VkEvent, you must not mix extension and core api calls for a single VkEvent. For example, you must not call vkCmdSetEvent2KHR() and then vkCmdWaitEvents().

Reusing the same pipeline stage and access flag names

Due to running out of the 32 bits for VkAccessFlag the VkAccessFlags2KHR type was created with a 64-bit range. To prevent the same issue for VkPipelineStageFlags, the VkPipelineStageFlags2KHR type was also created with a 64-bit range.

64-bit enumeration types are not available in all C/C++ compilers, so the code for the new fields uses static const values instead of an enum. As a result of this, there are no equivalent types to VkPipelineStageFlagBits and VkAccessFlagBits. Some code, including Vulkan functions such as vkCmdWriteTimestamp(), used the Bits type to indicate that the caller could only pass in a single bit value, rather than a mask of multiple bits. These calls need to be converted to take the Flags type and enforce the “only 1-bit” limitation via Valid Usage or the appropriate coding convention for your own code, as was done for vkCmdWriteTimestamp2KHR().

The new flags include identical bits to the original synchronization flags, with the same base name and identical values. Old flags can be used directly in the new APIs, subject to any typecasting constraints of the coding environment. The following 2 examples show the naming differences:

  • VK_PIPELINE_STAGE_COMPUTE_SHADER_BIT to VK_PIPELINE_STAGE_2_COMPUTE_SHADER_BIT_KHR

  • VK_ACCESS_SHADER_READ_BIT to VK_ACCESS_2_SHADER_READ_BIT_KHR

VkSubpassDependency

Updating the use of the pipeline stages and access flags in VkSubpassDependency requires simply using VkSubpassDependency2 which can have a VkMemoryBarrier2KHR passed in the pNext

Example would be taking

// Without VK_KHR_synchronization2
VkSubpassDependency dependency = {
    .srcSubpass = 0,
    .dstSubpass = 1,
    .srcStageMask = VK_PIPELINE_STAGE_EARLY_FRAGMENT_TESTS_BIT |
                    VK_PIPELINE_STAGE_LATE_FRAGMENT_TESTS_BIT,
    .dstStageMask = VK_PIPELINE_STAGE_FRAGMENT_SHADER_BIT,
    .srcAccessMask = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,
    .dstAccessMask = VK_ACCESS_INPUT_ATTACHMENT_READ_BIT,
    .dependencyFlags = VK_DEPENDENCY_BY_REGION_BIT
};

and turning it into

// With VK_KHR_synchronization2
VkMemoryBarrier2KHR memoryBarrier = {
    .sType = VK_STRUCTURE_TYPE_MEMORY_BARRIER_2_KHR,
    .pNext = nullptr,
    .srcStageMask = VK_PIPELINE_STAGE_2_EARLY_FRAGMENT_TESTS_BIT_KHR |
                    VK_PIPELINE_STAGE_2_LATE_FRAGMENT_TESTS_BIT_KHR,
    .dstStageMask = VK_PIPELINE_STAGE_2_FRAGMENT_SHADER_BIT_KHR,
    .srcAccessMask = VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT_KHR,
    .dstAccessMask = VK_ACCESS_2_INPUT_ATTACHMENT_READ_BIT_KHR
}

// The 4 fields unset are ignored according to the spec
// When VkMemoryBarrier2KHR is passed into pNext
VkSubpassDependency2 dependency = {
    .sType = VK_STRUCTURE_TYPE_SUBPASS_DEPENDENCY_2,
    .pNext = &memoryBarrier,
    .srcSubpass = 0,
    .dstSubpass = 1,
    .dependencyFlags = VK_DEPENDENCY_BY_REGION_BIT
};

Splitting up pipeline stages and access masks

Some VkAccessFlags and VkPipelineStageFlags had values that were ambiguous to what it was targeting in hardware. The new VkAccessFlags2KHR and VkPipelineStageFlags2KHR break these up for some cases while leaving the old value for maintability.

Splitting up VK_PIPELINE_STAGE_VERTEX_INPUT_BIT

The VK_PIPELINE_STAGE_VERTEX_INPUT_BIT (now VK_PIPELINE_STAGE_2_VERTEX_INPUT_BIT_KHR) was split into 2 new stage flags which specify a dedicated stage for both the index input and the vertex input instead of having them combined into a single pipeline stage flag.

  • VK_PIPELINE_STAGE_2_INDEX_INPUT_BIT_KHR

  • VK_PIPELINE_STAGE_2_VERTEX_ATTRIBUTE_INPUT_BIT_KHR

Splitting up VK_PIPELINE_STAGE_ALL_TRANSFER_BIT

The VK_PIPELINE_STAGE_ALL_TRANSFER_BIT (now VK_PIPELINE_STAGE_2_ALL_TRANSFER_BIT_KHR) was split into 4 new stage flags which specify a dedicated stage for the various staging commands instead of having them combined into a single pipeline stage flag.

  • VK_PIPELINE_STAGE_2_COPY_BIT_KHR

  • VK_PIPELINE_STAGE_2_RESOLVE_BIT_KHR

  • VK_PIPELINE_STAGE_2_BLIT_BIT_KHR

  • VK_PIPELINE_STAGE_2_CLEAR_BIT_KHR

Splitting up VK_ACCESS_SHADER_READ_BIT

The VK_ACCESS_SHADER_READ_BIT (now VK_ACCESS_2_SHADER_READ_BIT_KHR) was split into 3 new access flags which specify a dedicated access for the various case instead of having them combined into a single access flag.

  • VK_ACCESS_2_UNIFORM_READ_BIT_KHR

  • VK_ACCESS_2_SHADER_SAMPLED_READ_BIT_KHR

  • VK_ACCESS_2_SHADER_STORAGE_READ_BIT_KHR

Combining shader stages for pre-rasterization

Besides splitting up flags, the VK_PIPELINE_STAGE_2_PRE_RASTERIZATION_SHADERS_BIT_KHR was added to combine shader stages that occurs before rasterization in a single, convenient flag.

VK_ACCESS_SHADER_WRITE_BIT alias

The VK_ACCESS_SHADER_WRITE_BIT (now VK_ACCESS_2_SHADER_WRITE_BIT_KHR) was given an alias of VK_ACCESS_2_SHADER_STORAGE_WRITE_BIT_KHR to better describe the scope of what resources in the shader are described by the access flag.

TOP_OF_PIPE and BOTTOM_OF_PIPE deprecation

The use of VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT and VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT are now deprecated and updating is simple as following the following 4 case with the new equivalents.

  • VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT in first synchronization scope

    // From
      .srcStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
    
    // To
      .srcStageMask = VK_PIPELINE_STAGE_2_NONE_KHR;
      .srcAccessMask = VK_ACCESS_2_NONE_KHR;
  • VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT in second synchronization scope

    // From
      .dstStageMask = VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT;
    
    // To
      .dstStageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT_KHR;
      .dstAccessMask = VK_ACCESS_2_NONE_KHR;
  • VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT in first synchronization scope

    // From
      .srcStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT;
    
    // To
      .srcStageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT_KHR;
      .srcAccessMask = VK_ACCESS_2_NONE_KHR;
  • VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT in second synchronization scope

    // From
      .dstStageMask = VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT;
    
    // To
      .dstStageMask = VK_PIPELINE_STAGE_2_NONE_KHR;
      .dstAccessMask = VK_ACCESS_2_NONE_KHR;

Making use of new image layouts

VK_KHR_synchronization2 adds 2 new image layouts VK_IMAGE_LAYOUT_ATTACHMENT_OPTIMAL_KHR and VK_IMAGE_LAYOUT_READ_ONLY_OPTIMAL_KHR to help with making layout transition easier.

The following uses the example of doing a draw thats writes to both a color attachment and depth/stencil attachment which then are both sampled in the next draw. Prior a developer needed to make sure they matched up the layouts and access mask correctly such as the following:

VkImageMemoryBarrier colorImageMemoryBarrier = {
  .srcAccessMask = VK_ACCESS_COLOR_ATTACHMENT_WRITE_BIT,
  .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
  .oldLayout = VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL,
  .newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
};

VkImageMemoryBarrier depthStencilImageMemoryBarrier = {
  .srcAccessMask = VK_ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT,,
  .dstAccessMask = VK_ACCESS_SHADER_READ_BIT,
  .oldLayout = VK_IMAGE_LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL,
  .newLayout = VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL
};

but with VK_KHR_synchronization2 this is made simple

VkImageMemoryBarrier colorImageMemoryBarrier = {
  .srcAccessMask = VK_ACCESS_2_COLOR_ATTACHMENT_WRITE_BIT_KHR,
  .dstAccessMask = VK_ACCESS_2_SHADER_READ_BIT_KHR,
  .oldLayout = VK_IMAGE_LAYOUT_ATTACHMENT_OPTIMAL_KHR, // new layout from VK_KHR_synchronization2
  .newLayout = VK_IMAGE_LAYOUT_READ_ONLY_OPTIMAL_KHR   // new layout from VK_KHR_synchronization2
};

VkImageMemoryBarrier depthStencilImageMemoryBarrier = {
  .srcAccessMask = VK_ACCESS_2_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT_KHR,,
  .dstAccessMask = VK_ACCESS_2_SHADER_READ_BIT_KHR,
  .oldLayout = VK_IMAGE_LAYOUT_ATTACHMENT_OPTIMAL_KHR, // new layout from VK_KHR_synchronization2
  .newLayout = VK_IMAGE_LAYOUT_READ_ONLY_OPTIMAL_KHR   // new layout from VK_KHR_synchronization2
};

In the new case VK_IMAGE_LAYOUT_ATTACHMENT_OPTIMAL_KHR works by contextually appling itself based on the image format used. So as long as colorImageMemoryBarrier is used on a color format, VK_IMAGE_LAYOUT_ATTACHMENT_OPTIMAL_KHR maps to VK_IMAGE_LAYOUT_COLOR_ATTACHMENT_OPTIMAL

Additionally, with VK_KHR_synchronization2, if oldLayout is equal to newLayout, no layout transition is performed and the image contents are preserved. The layout used does not even need to match the layout of an image, so the following barrier is valid:

VkImageMemoryBarrier depthStencilImageMemoryBarrier = {
  // other fields omitted
  .oldLayout = VK_IMAGE_LAYOUT_UNDEFINED,
  .newLayout = VK_IMAGE_LAYOUT_UNDEFINED,
};

New submission flow

VK_KHR_synchronization2 adds the vkQueueSubmit2KHR command which main goal is to clean up the syntax for the function to wrap command buffers and semaphores in extensible structures, which incorporate changes from Vulkan 1.1, VK_KHR_device_group, and VK_KHR_timeline_semaphore.

Taking the following example of a normal queue submission call

VkSemaphore waitSemaphore;
VkSemaphore signalSemaphore;
VkCommandBuffer commandBuffers[8];

// Possible pNext from VK_KHR_timeline_semaphore
VkTimelineSemaphoreSubmitInfo timelineSemaphoreSubmitInfo = {
    // ...
    .pNext = nullptr
};

// Possible pNext from VK_KHR_device_group
VkDeviceGroupSubmitInfo deviceGroupSubmitInfo = {
    // ...
    .pNext = &timelineSemaphoreSubmitInfo
};

// Possible pNext from Vulkan 1.1
VkProtectedSubmitInfo = protectedSubmitInfo {
    // ...
    .pNext = &deviceGroupSubmitInfo
};

VkSubmitInfo submitInfo = {
    .pNext = &protectedSubmitInfo, // Chains all 3 extensible structures
    .waitSemaphoreCount = 1,
    .pWaitSemaphores = &waitSemaphore,
    .pWaitDstStageMask = VK_PIPELINE_STAGE_COLOR_ATTACHMENT_OUTPUT_BIT
    .commandBufferCount = 8,
    .pCommandBuffers = commandBuffers,
    .signalSemaphoreCount = 1,
    .pSignalSemaphores = signalSemaphore
};

vkQueueSubmit(queue, 1, submitInfo, fence);

this can now be transformed to vkQueueSubmit2KHR as

// Uses same semaphore and command buffer handles
VkSemaphore waitSemaphore;
VkSemaphore signalSemaphore;
VkCommandBuffer commandBuffers[8];

VkSemaphoreSubmitInfoKHR waitSemaphoreSubmitInfo = {
    .semaphore = waitSemaphore,
    .value = 1, // replaces VkTimelineSemaphoreSubmitInfo
    .stageMask = VK_PIPELINE_STAGE_2_COLOR_ATTACHMENT_OUTPUT_BIT_KHR,
    .deviceIndex = 0, // replaces VkDeviceGroupSubmitInfo
};

// Note this is allowing a stage to set the signal operation
VkSemaphoreSubmitInfoKHR signalSemaphoreSubmitInfo = {
    .semaphore = signalSemaphore,
    .value = 2, // replaces VkTimelineSemaphoreSubmitInfo
    .stageMask = VK_PIPELINE_STAGE_2_VERTEX_SHADER_BIT_KHR, // when to signal semaphore
    .deviceIndex = 0, // replaces VkDeviceGroupSubmitInfo
};

// Need one for each VkCommandBuffer
VkCommandBufferSubmitInfoKHR = commandBufferSubmitInfos[8] {
    // ...
    {
        .commandBuffer = commandBuffers[i],
        .deviceMask = 0 // replaces VkDeviceGroupSubmitInfo
    },
};

VkSubmitInfo2KHR submitInfo = {
    .pNext = nullptr, // All 3 struct above are built into VkSubmitInfo2KHR
    .flags = VK_SUBMIT_PROTECTED_BIT_KHR, // also can be zero, replaces VkProtectedSubmitInfo
    .waitSemaphoreInfoCount = 1,
    .pWaitSemaphoreInfos = waitSemaphoreSubmitInfo,
    .commandBufferInfoCount = 8,
    .pCommandBufferInfos = commandBufferSubmitInfos,
    .signalSemaphoreInfoCount = 1,
    .pSignalSemaphoreInfos = signalSemaphoreSubmitInfo
}

vkQueueSubmit2KHR(queue, 1, submitInfo, fence);

The difference between the two examples code snippets above is that the vkQueueSubmit2KHR will signal VkSemaphore signalSemaphore when the vertex shader stage is complete compared to the vkQueueSubmit call which will wait until the end of the submission.

To emulate the same behavior of semaphore signaling from vkQueueSubmit in vkQueueSubmit2KHR the stageMask can be set to VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT

// Waits until everything is done
VkSemaphoreSubmitInfoKHR signalSemaphoreSubmitInfo = {
    // ...
    .stageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT,
    // ...
};

Emulation Layer

For devices that do not natively support this extension, there is a portable implementation in the Vulkan-Extensionlayer repository. This layer should work with any Vulkan device. For more information see the layer documentation and the Sync2Compat.Vulkan10 test case.

Note

The VK_KHR_synchronization2 specification lists VK_KHR_create_renderpass2 and VK_KHR_get_phyiscal_device_properties2 as requirements. As a result, using synchronization2 without these extensions may result in validation errors. The extension requirements are being reevaluated and validation will be adjusted once this is complete.