Vulkan-Docs/doc/specs/vulkan/appendices/VK_NVX_device_generated_commands.txt

include::meta/VK_NVX_device_generated_commands.txt[]

*Last Modified Date*::
    2017-07-25
*Contributors*::
  - Pierre Boudier, NVIDIA
  - Christoph Kubisch, NVIDIA
  - Mathias Schott, NVIDIA
  - Jeff Bolz, NVIDIA
  - Eric Werness, NVIDIA
  - Detlef Roettger, NVIDIA
  - Daniel Koch, NVIDIA
  - Chris Hebert, NVIDIA

This extension allows the device to generate a number of critical 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...
Implementing those on the device does not require any special extension,
since an application is free to define its own data structure, 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 on 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
scenario, or updates synchronized on the host.

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

  * create its objects as in unextended Vulkan
  * create a slink:VkObjectTableNVX, and register the various Vulkan objects
    that are needed to evaluate the input parameters.
  * create a slink:VkIndirectCommandsLayoutNVX, which lists the
    slink:VkIndirectCommandsTokenTypeNVX it wants to dynamically change as
    atomic command sequence.
    This step likely involves some internal device code compilation, since
    the intent is for the GPU to generate the command buffer in the
    pipeline.
  * fill the input buffers with the data for each of the inputs it needs.
    Each input is an array that will be filled with an index in the object
    table, instead of using CPU pointers.
  * set up a target secondary command buffer
  * reserve command buffer space via flink:vkCmdReserveSpaceForCommandsNVX
    in a target command buffer at the position you want the generated
    commands to be executed.
  * call flink:vkCmdProcessCommandsNVX to create the actual device commands
    for all sequences based on the array contents into a provided target
    command buffer.
  * execute the target command buffer like a regular secondary command
    buffer

For each draw/dispatch, the following can be specified:

  * a different pipeline state object
  * a number of descriptor sets, with dynamic offsets
  * a number of vertex buffer bindings, with an optional dynamic offset
  * a different index buffer, with an optional dynamic offset

Applications should: register a small number of objects, and use dynamic
offsets whenever possible.

While the GPU can be faster than a CPU to generate the commands, it may not
happen asynchronously, therefore the primary use-case is generating "`less`"
total work (occlusion culling, classification to use specialized
shaders...).

=== New Object Types

  * slink:VkObjectTableNVX
  * slink:VkIndirectCommandsLayoutNVX

=== New Flag Types

  * elink:VkIndirectCommandsLayoutUsageFlagsNVX
  * elink:VkObjectEntryUsageFlagsNVX

=== New Enum Constants

Extending elink:VkStructureType:

  ** ename:VK_STRUCTURE_TYPE_OBJECT_TABLE_CREATE_INFO_NVX
  ** ename:VK_STRUCTURE_TYPE_INDIRECT_COMMANDS_LAYOUT_CREATE_INFO_NVX
  ** ename:VK_STRUCTURE_TYPE_CMD_PROCESS_COMMANDS_INFO_NVX
  ** ename:VK_STRUCTURE_TYPE_CMD_RESERVE_SPACE_FOR_COMMANDS_INFO_NVX
  ** ename:VK_STRUCTURE_TYPE_DEVICE_GENERATED_COMMANDS_LIMITS_NVX
  ** ename:VK_STRUCTURE_TYPE_DEVICE_GENERATED_COMMANDS_FEATURES_NVX

Extending elink:VkPipelineStageFlagBits:

  ** ename:VK_PIPELINE_STAGE_COMMAND_PROCESS_BIT_NVX

Extending elink:VkAccessFlagBits:

  ** ename:VK_ACCESS_COMMAND_PROCESS_READ_BIT_NVX
  ** ename:VK_ACCESS_COMMAND_PROCESS_WRITE_BIT_NVX

=== New Enums

  * elink:VkIndirectCommandsLayoutUsageFlagBitsNVX
  * elink:VkIndirectCommandsTokenTypeNVX
  * elink:VkObjectEntryUsageFlagBitsNVX
  * elink:VkObjectEntryTypeNVX

=== New Structures

  * slink:VkDeviceGeneratedCommandsFeaturesNVX
  * slink:VkDeviceGeneratedCommandsLimitsNVX
  * slink:VkIndirectCommandsTokenNVX
  * slink:VkIndirectCommandsLayoutTokenNVX
  * slink:VkIndirectCommandsLayoutCreateInfoNVX
  * slink:VkCmdProcessCommandsInfoNVX
  * slink:VkCmdReserveSpaceForCommandsInfoNVX
  * slink:VkObjectTableCreateInfoNVX
  * slink:VkObjectTableEntryNVX
  * slink:VkObjectTablePipelineEntryNVX
  * slink:VkObjectTableDescriptorSetEntryNVX
  * slink:VkObjectTableVertexBufferEntryNVX
  * slink:VkObjectTableIndexBufferEntryNVX
  * slink:VkObjectTablePushConstantEntryNVX

=== New Functions

  * flink:vkCmdProcessCommandsNVX
  * flink:vkCmdReserveSpaceForCommandsNVX
  * flink:vkCreateIndirectCommandsLayoutNVX
  * flink:vkDestroyIndirectCommandsLayoutNVX
  * flink:vkCreateObjectTableNVX
  * flink:vkDestroyObjectTableNVX
  * flink:vkRegisterObjectsNVX
  * flink:vkUnregisterObjectsNVX
  * flink:vkGetPhysicalDeviceGeneratedCommandsPropertiesNVX

=== Issues

1) How to name this extension ?

*RESOLVED*: +VK_NVX_device_generated_commands+

As usual one of the hardest issues ;)

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

2) Should we use serial tokens or redundant sequence description?

Similar to slink:VkPipeline, signatures have the most likeliness to be
cross-vendor adoptable.
They also benefit from being processable in parallel.

3) How to name sequence description

ExecuteCommandSignature a bit long, just ExecuteSignature or actually more
Vulkan nomenclature IndirectCommandsLayout

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

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

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

It is desired by application to reuse the list of objects and render them
with some kind override.
This can be done by just selecting a different input for a push constant or
a descriptor set, if they are defined as independent arrays.
If the data was interleaved, this would not be as easily possible.

Allowing input divisors can also reduce the conservative command buffer
allocation.

6) How do we know the size of the GPU command buffer generated by
flink:vkCmdProcessCommandsNVX ?

pname:maxSequenceCount can give an upper estimate, even if the actual count
is sourced from the gpu buffer at (buffer, countOffset).
As such pname:maxSequenceCount must always be set correctly.

Developers are encouraged to make well use the IndirectCommandsLayout's
pname:pTokens->divisor, as they allow less conservative storage costs.
Especially pipeline changes on a per-draw basis can be costly memory wise.

7) How to deal with dynamic offsets in DescriptorSets?

Maybe additional token etext:VK_EXECUTE_DESCRIPTOR_SET_OFFSET_COMMAND_NVX
that works for a "`single dynamic buffer`" descriptor set and then use (32
bit tableEntry + 32bit offset)

added dynamicCount field, variable sized input

8) Should we allow updates to the object table, similar to DescriptorSet?

Desired yes, people may change "`material`" shaders and not want to recreate
the entire register table.
However the developer must ensure to not overwrite a registered objectindex
while it is still being used.

9) Should we allow dynamic state changes?

Seems a bit excessive for "`per-draw`" type of scenario, but GPU could
partition work itself with viewport/scissor...

10) How do we allow re-using already "`filled`" indirectCommands buffers?

just use a slink:VkCommandBuffer for the output, and it can be reused
easily.

11) How portable should such re-use be?

Same as secondary command buffer

12) Should sequenceOrdered be part of IndirectCommandsLayout or
slink:vkCmdProcessCommandsNVX?

Seems better for IndirectCommandsLayout, as that is when most heavy lifting
in terms of internal device code generation is done.

13) Under which conditions is flink:vkCmdProcessCommandsNVX legal?

Options:

  a) on the host command buffer like a regular draw call b)
     flink:vkCmdProcessCommandsNVX makes use slink:VkCommandBufferBeginInfo
     and serves as flink:vkBeginCommandBuffer / flink:vkEndCommandBuffer
     implicitly.
  c) The targetCommandbuffer must be inside the "`begin`" state already at
     the moment of being passed.
     This very likely suggests a new slink:VkCommandBufferUsageFlags
     etext:VK_COMMAND_BUFFER_USAGE_DEVICE_GENERATED_BIT.
  d) The targetCommandbuffer must reserve space via a new function.

used a & d.

14) What if different pipelines have different DescriptorSetLayouts at a
certain set unit that mismatches in "`token.dynamicCount`"?

Considered legal, as long as the maximum dynamic count of all used
DescriptorSetLayouts is provided.

15) Should we add "`strides`" to input arrays, so that "`Array of
Structures`" type setups can be support more easily?

Maybe provide a usage flag for packed tokens stream (all inputs from same
buffer, implicit stride).

No, given performance test was worse.

16) Should we allow re-using the target command buffer directly, without
need to reset command buffer?

YES: new api flink:vkCmdReserveSpaceForCommandsNVX.

17) Is flink:vkCmdProcessCommandsNVX copying the input data or referencing
it ?

There are multiple implementations possible:

  * one could have some emulation code that parse the inputs, and generates
    an output command buffer, therefore copying the inputs.
  * one could just reference the inputs, and have the processing done in
    pipe at execution time.

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 should not be modified after the call to
flink:vkCmdProcessCommandsNVX and until after the rendering of the target
command buffer is finished.

18) Why is this +NVX+ and not +NV+?

To allow early experimentation and feedback.
We expect that a version with a refined design as multi-vendor variant will
follow up.

19) Should we make the availability for each token type a device limit?

Only distinguish between graphics/compute for now, further splitting up may
lead to too much fractioning.

20) When can the objectTable be modified?

Similar to the other inputs for flink:vkCmdProcessCommandsNVX, only when all
device access via flink:vkCmdProcessCommandsNVX or execution of target
command buffer has completed can an object at a given objectIndex be
unregistered or re-registered again.

21) Which buffer usage flags are required for the buffers referenced by
flink:vkCmdProcessCommandsNVX

reuse existing ename:VK_BUFFER_USAGE_INDIRECT_BUFFER_BIT

  * slink:VkCmdProcessCommandsInfoNVX::pname:sequencesCountBuffer
  * slink:VkCmdProcessCommandsInfoNVX::pname:sequencesIndexBuffer
  * slink:VkIndirectCommandsTokenNVX::pname:buffer

22) In which pipeline stage do the device generated command expansion
happen?

flink:vkCmdProcessCommandsNVX is treated as if it occurs in a separate
logical pipeline from either graphics or compute, and that pipeline only
includes ename:VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, a new stage
ename:VK_PIPELINE_STAGE_COMMAND_PROCESS_BIT_NVX, and
ename:VK_PIPELINE_STAGE_BOTTOM_OF_PIPE_BIT.
This new stage has two corresponding new access types,
ename:VK_ACCESS_COMMAND_PROCESS_READ_BIT_NVX and
ename:VK_ACCESS_COMMAND_PROCESS_WRITE_BIT_NVX, used to synchronize reading
the buffer inputs and writing the command buffer memory output.
The output written in the target command buffer is considered to be consumed
by the DRAW_INDIRECT pipeline stage.

Thus, to synchronize from writing the input buffers to executing
flink:vkCmdProcessCommandsNVX, use:

  * pname:dstStageMask = ename:VK_PIPELINE_STAGE_COMMAND_PROCESS_BIT_NVX
  * pname:dstAccessMask = ename:VK_ACCESS_COMMAND_PROCESS_READ_BIT_NVX

To synchronize from executing flink:vkCmdProcessCommandsNVX to executing the
generated commands, use

  * pname:srcStageMask = ename:VK_PIPELINE_STAGE_COMMAND_PROCESS_BIT_NVX
  * pname:srcAccessMask = ename:VK_ACCESS_COMMAND_PROCESS_WRITE_BIT_NVX
  * pname:dstStageMask = ename:VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT
  * pname:dstAccessMask = ename:VK_ACCESS_INDIRECT_COMMAND_READ_BIT

When flink:vkCmdProcessCommandsNVX is used with a pname:targetCommandBuffer
of `NULL`, the generated commands are immediately executed and there is
implicit synchronization between generation and execution.

23) 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 processed.

=== Example Code

Open-Source samples illustrating the usage of the extension can be found at
the following locations:

https://github.com/nvpro-samples/gl_vk_threaded_cadscene/blob/master/doc/vulkan_nvxdevicegenerated.md

https://github.com/NVIDIAGameWorks/GraphicsSamples/tree/master/samples/vk10-kepler/BasicDeviceGeneratedCommandsVk

[source,c]
---------------------------------------------------

  // setup secondary command buffer
    vkBeginCommandBuffer(generatedCmdBuffer, &beginInfo);
    ... setup its state as usual

  // insert the reservation (there can only be one per command buffer)
  // where the generated calls should be filled into
    VkCmdReserveSpaceForCommandsInfoNVX reserveInfo = { VK_STRUCTURE_TYPE_CMD_RESERVE_SPACE_FOR_COMMANDS_INFO_NVX };
    reserveInfo.objectTable = objectTable;
    reserveInfo.indirectCommandsLayout = deviceGeneratedLayout;
    reserveInfo.maxSequencesCount = myCount;
    vkCmdReserveSpaceForCommandsNVX(generatedCmdBuffer, &reserveInfo);

    vkEndCommandBuffer(generatedCmdBuffer);

  // trigger the generation at some point in another primary command buffer
    VkCmdProcessCommandsInfoNVX processInfo = { VK_STRUCTURE_TYPE_CMD_PROCESS_COMMANDS_INFO_NVX };
    processInfo.objectTable = objectTable;
    processInfo.indirectCommandsLayout = deviceGeneratedLayout;
    processInfo.maxSequencesCount = myCount;
    // set the target of the generation (if null we would directly execute with mainCmd)
    processInfo.targetCommandBuffer = generatedCmdBuffer;
    // provide input data
    processInfo.indirectCommandsTokenCount = 3;
    processInfo.pIndirectCommandsTokens = myTokens;

  // If you modify the input buffer data referenced by VkCmdProcessCommandsInfoNVX,
  // ensure you have added the appropriate barriers prior generation process.
  // When regenerating the content of the same reserved space, ensure prior operations have completed

    VkMemoryBarrier memoryBarrier = { VK_STRUCTURE_TYPE_MEMORY_BARRIER };
    memoryBarrier.srcAccessMask = ...;
    memoryBarrier.dstAccessMask = VK_ACCESS_COMMAND_PROCESS_READ_BIT_NVX;

    vkCmdPipelineBarrier(mainCmd,
                         /*srcStageMask*/VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
                         /*dstStageMask*/VK_PIPELINE_STAGE_COMMAND_PROCESS_BIT_NVX,
                         /*dependencyFlags*/0,
                         /*memoryBarrierCount*/1,
                         /*pMemoryBarriers*/&memoryBarrier,
                         ...);

    vkCmdProcessCommandsNVX(mainCmd, &processInfo);
    ...
  // execute the secondary command buffer and ensure the processing that modifies command-buffer content
  // has completed

    memoryBarrier.srcAccessMask = VK_ACCESS_COMMAND_PROCESS_WRITE_BIT_NVX;
    memoryBarrier.dstAccessMask = VK_ACCESS_INDIRECT_COMMAND_READ_BIT;

    vkCmdPipelineBarrier(mainCmd,
                         /*srcStageMask*/VK_PIPELINE_STAGE_COMMAND_PROCESS_BIT_NVX,
                         /*dstStageMask*/VK_PIPELINE_STAGE_DRAW_INDIRECT_BIT,
                         /*dependencyFlags*/0,
                         /*memoryBarrierCount*/1,
                         /*pMemoryBarriers*/&memoryBarrier,
                         ...)
    vkCmdExecuteCommands(mainCmd, 1, &generatedCmdBuffer);

---------------------------------------------------

=== Version History

 * Revision 3, 2017-07-25 (Chris Hebert)
   - Correction to specification of dynamicCount for push_constant token in
     VkIndirectCommandsLayoutNVX.
     Stride was incorrectly computed as dynamicCount was not treated as byte
     size.
 * Revision 2, 2017-06-01 (Christoph Kubisch)
   - header compatibility break: add missing _TYPE to
     VkIndirectCommandsTokenTypeNVX and VkObjectEntryTypeNVX enums to follow
     Vulkan naming convention
   - behavior clarification: only allow a single work provoking token per
     sequence when creating a slink:VkIndirectCommandsLayoutNVX
 * Revision 1, 2016-10-31 (Christoph Kubisch)
   - Initial draft