VK_NV_device_generated_commands(3) Manual Page

278

3

Not ratified

This extension allows the device to generate a number of critical graphics commands for command buffers.

When rendering a large number of objects, the device can be leveraged to implement a number of critical functions, like updating matrices, or implementing occlusion culling, frustum culling, front to back sorting, etc. Implementing those on the device does not require any special extension, since an application is free to define its own data structures, and just process them using shaders.

However, if the application desires to quickly kick off the rendering of the final stream of objects, then unextended Vulkan forces the application to read back the processed stream and issue graphics command from the host. For very large scenes, the synchronization overhead and cost to generate the command buffer can become the bottleneck. This extension allows an application to generate a device side stream of state changes and commands, and convert it efficiently into a command buffer without having to read it back to the host.

Furthermore, it allows incremental changes to such command buffers by manipulating only partial sections of a command stream — for example pipeline bindings. Unextended Vulkan requires re-creation of entire command buffers in such a scenario, or updates synchronized on the host.

The intended usage for this extension is for the application to:

For each draw in a sequence, the following can be specified:

While the GPU can be faster than a CPU to generate the commands, it will not happen asynchronously to the device, therefore the primary use case is generating “less” total work (occlusion culling, classification to use specialized shaders, etc.).

1) How to name this extension ?

VK_NV_device_generated_commands

As usual, one of the hardest issues ;)

Alternatives: VK_gpu_commands, VK_execute_commands, VK_device_commands, VK_device_execute_commands, VK_device_execute, VK_device_created_commands, VK_device_recorded_commands, VK_device_generated_commands VK_indirect_generated_commands

2) Should we use a serial stateful token stream or stateless sequence descriptions?

Similarly to VkPipeline, fixed layouts have the most likelihood to be cross-vendor adoptable. They also benefit from being processable in parallel. This is a different design choice compared to the serial command stream generated through GL_NV_command_list.

3) How to name a sequence description?

VkIndirectCommandsLayout as in the NVX extension predecessor.

Alternative: VkGeneratedCommandsLayout

4) Do we want to provide indirectCommands inputs with layout or at indirectCommands time?

Separate layout from data as Vulkan does. Provide full flexibility for indirectCommands.

5) Should the input be provided as SoA or AoS?

Both ways are desirable. AoS can provide portability to other APIs and easier to setup, while SoA allows to update individual inputs in a cache-efficient manner, when others remain static.

6) How do we make developers aware of the memory requirements of implementation-dependent data used for the generated commands?

Make the API explicit and introduce a preprocess VkBuffer. Developers have to allocate it using vkGetGeneratedCommandsMemoryRequirementsNV.

In the NVX version the requirements were hidden implicitly as part of the command buffer reservation process, however as the memory requirements can be substantial, we want to give developers the ability to budget the memory themselves. By lowering the maxSequencesCount the memory consumption can be reduced. Furthermore reuse of the memory is possible, for example for doing explicit preprocessing and execution in a ping-pong fashion.

The actual buffer size is implementation-dependent and may be zero, i.e. not always required.

When making use of Graphics Shader Groups, the programs should behave similar with regards to vertex inputs, clipping and culling outputs of the geometry stage, as well as sample shading behavior in fragment shaders, to reduce the amount of the worst-case memory approximation.

7) Should we allow additional per-sequence dynamic state changes?

Yes

Introduced a lightweight indirect state flag VkIndirectStateFlagBitsNV. So far only switching front face winding state is exposed. Especially in CAD/DCC mirrored transforms that require such changes are common, and similar flexibility is given in the ray tracing instance description.

The flag could be extended further, for example to switch between primitive-lists or -strips, or make other state modifications.

Furthermore, as new tokens can be added easily, future extension could add the ability to change any VkDynamicState.

8) How do we allow re-using already “generated” indirectCommands?

Expose a preprocessBuffer to reuse implementation-dependencyFlags data. Set the isPreprocessed to true in vkCmdExecuteGeneratedCommandsNV.

9) Under which conditions is vkCmdExecuteGeneratedCommandsNV legal?

It behaves like a regular draw call command.

10) Is vkCmdPreprocessGeneratedCommandsNV copying the input data or referencing it?

There are multiple implementations possible:

If the data is mandated to be copied, then it puts a penalty on implementation that could process the inputs directly in pipe. If the data is “referenced”, then it allows both types of implementation.

The inputs are “referenced”, and must not be modified after the call to vkCmdExecuteGeneratedCommandsNV has completed.

11) Which buffer usage flags are required for the buffers referenced by VkGeneratedCommandsInfoNV ?

Reuse existing VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT

12) In which pipeline stage does the device generated command expansion happen?

vkCmdPreprocessGeneratedCommandsNV is treated as if it occurs in a separate logical pipeline from either graphics or compute, and that pipeline only includes VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, a new stage VK_PIPELINE_STAGE_COMMAND_PREPROCESS_BIT_NV, and VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT. This new stage has two corresponding new access types, VK_ACCESS_COMMAND_PREPROCESS_READ_BIT_NV and VK_ACCESS_COMMAND_PREPROCESS_WRITE_BIT_NV, used to synchronize reading the buffer inputs and writing the preprocess memory output.

The generated output written in the preprocess buffer memory by vkCmdExecuteGeneratedCommandsNV is considered to be consumed by the VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT pipeline stage.

Thus, to synchronize from writing the input buffers to preprocessing via vkCmdPreprocessGeneratedCommandsNV, use:

To synchronize from vkCmdPreprocessGeneratedCommandsNV to executing the generated commands by vkCmdExecuteGeneratedCommandsNV, use:

When vkCmdExecuteGeneratedCommandsNV is used with a isPreprocessed of VK_FALSE, the generated commands are implicitly preprocessed, therefore one only needs to synchronize the inputs via:

13) What if most token data is “static”, but we frequently want to render a subsection?

Added “sequencesIndexBuffer”. This allows to easier sort and filter what should actually be executed.

14) What are the changes compared to the previous NVX extension?

15) When porting from other APIs, their indirect buffers may use different enums, for example for index buffer types. How to solve this?

Added “pIndexTypeValues” to map custom uint32_t values to corresponding VkIndexType.

16) Do we need more shader group state overrides?

The NVX version allowed all PSO states to be different, however as the goal is not to replace all state setup, but focus on highly-frequent state changes for drawing lots of objects, we reduced the amount of state overrides. Especially VkPipelineLayout as well as VkRenderPass configuration should be left static, the rest is still open for discussion.

The current focus is just to allow VertexInput changes as well as shaders, while all shader groups use the same shader stages.

Too much flexibility will increase the test coverage requirement as well. However, further extensions could allow more dynamic state as well.

17) Do we need more detailed physical device feature queries/enables?

An EXT version would need detailed implementor feedback to come up with a good set of features. Please contact us if you are interested, we are happy to make more features optional, or add further restrictions to reduce the minimum feature set of an EXT.

18) Is there an interaction with VK_KHR_pipeline_library planned?

Yes, a future version of this extension will detail the interaction, once VK_KHR_pipeline_library is no longer provisional.

Open-Source samples illustrating the usage of the extension can be found at the following location (may not yet exist at time of writing):

https://github.com/nvpro-samples/vk_device_generated_cmds

No cross-references are available

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 2014-2024 The Khronos Group Inc.

SPDX-License-Identifier: CC-BY-4.0