Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Is there any plan for Tessellation support in MoltenVK? #501

Closed
spencerkohan opened this Issue Feb 18, 2019 · 3 comments

Comments

Projects
None yet
3 participants
@spencerkohan
Copy link

commented Feb 18, 2019

It is understandable that geometry shaders are not supported in MoltenVK as this feature is not implemented in Metal2, but Tessellation is available. Is this on the roadmap for future versions?

@cdavis5e

This comment has been minimized.

Copy link
Collaborator

commented Feb 18, 2019

I'm working on it right now. A change to support tessellation evaluation shaders was just accepted into SPIRV-Cross. Hopefully, I'll be able to get this in by next week's SDK release.

@billhollings

This comment has been minimized.

Copy link
Contributor

commented Feb 18, 2019

Duplicate of #56.

cdavis5e added a commit to cdavis5e/MoltenVK that referenced this issue Feb 21, 2019

Add support for tessellation.
At long last, tessellation comes to MoltenVK! With this change, clients
will now be able to specify tessellation shaders when creating
pipelines, and then draw tessellated patches with them.

Unfortunately, there seem to be a few gotchas with tessellation in
Metal. For one thing, tessellation pipelines in Metal are structured
very differently from Vulkan. There is no tessellation control or even
vertex stage. Instead, the tessellation evaluation shader takes the
place of the vertex function as a "post-tessellation vertex function."
The tessellation levels are supplied in a buffer to the tessellator,
which you are expected to populate. The most common way to do this is by
running a compute shader. MoltenVK thus runs the vertex shader and
tessellation control shader by themselves; a single `VkPipeline` object
then requires at least *three* `MTLPipelineState` objects.

But wait, there's more! The tessellation-control-as-compute stage uses
Metal's support for vertex-style stage input to a compute shader. But,
this support requires one to declare indexing *ahead of time*, when the
pipeline state is created. So a single `VkPipeline` object could have as
many as *five* `MTLPipelineState` objects.

Further, if there are more output than input control points for the
tessellation control stage, then later invocations may end up fetching
the wrong attributes! To get around this, this change uses index buffers
to ensure that all tessellation control shaders see the correct input.
Unfortunately, in the indexed draw case, this means that the incoming
index buffer needs to be munged.

Instancing is another pain point here. In Vulkan, as in OpenGL and
Direct3D, instancing is done in the vertex shader; but in Metal, it is
done at the tessellation evaluation stage. For this reason, only the
vertex stage of a tessellated draw supports instancing. Additional
memory is required to hold data for the extra vertices generated by
instancing. This also requires still more munging of index buffers for
indexed draws.

Indirect draws are even more painful. Because the number of vertices and
instances is unknown, storage for the maximum possible number of
vertices must be allocated. This change imposes a totally arbitrary
limit of 131072 vertices from a single draw, including all vertices
generated by instancing. On a Mac, this requires about 194-256 MB of
VRAM for all the temporary buffers.

There are some possible optimizations here. If we could prove that the
vertex shader's output doesn't depend on the instance ID, either
directly or through a per-instance attribute, then we could avoid
running the vertex and tess. control stages per instance, and take
advantage of Metal's support for tess. eval instancing. If we could
also prove that the vertex shader simply passes instance attributes
through (similarly with the tess. control shader), we could do this for
many more instanced draws as well. It should also be possible to cache
the output from the tess. control stage; if the draw comes up again, we
can then skip the vertex and tess. control stages entirely!

Fixes KhronosGroup#56 and KhronosGroup#501.

cdavis5e added a commit to cdavis5e/MoltenVK that referenced this issue Mar 11, 2019

Add support for tessellation.
At long last, tessellation comes to MoltenVK! With this change, clients
will now be able to specify tessellation shaders when creating
pipelines, and then draw tessellated patches with them.

Unfortunately, there seem to be a few gotchas with tessellation in
Metal. For one thing, tessellation pipelines in Metal are structured
very differently from Vulkan. There is no tessellation control or even
vertex stage. Instead, the tessellation evaluation shader takes the
place of the vertex function as a "post-tessellation vertex function."
The tessellation levels are supplied in a buffer to the tessellator,
which you are expected to populate. The most common way to do this is by
running a compute shader. MoltenVK thus runs the vertex shader and
tessellation control shader by themselves; a single `VkPipeline` object
then requires at least *three* `MTLPipelineState` objects.

But wait, there's more! The tessellation-control-as-compute stage uses
Metal's support for vertex-style stage input to a compute shader. But,
this support requires one to declare indexing *ahead of time*, when the
pipeline state is created. So a single `VkPipeline` object could have as
many as *five* `MTLPipelineState` objects.

Further, if there are more output than input control points for the
tessellation control stage, then later invocations may end up fetching
the wrong attributes! To get around this, this change uses index buffers
to ensure that all tessellation control shaders see the correct input.
Unfortunately, in the indexed draw case, this means that the incoming
index buffer needs to be munged.

Instancing is another pain point here. In Vulkan, as in OpenGL and
Direct3D, instancing is done in the vertex shader; but in Metal, it is
done at the tessellation evaluation stage. For this reason, only the
vertex stage of a tessellated draw supports instancing. Additional
memory is required to hold data for the extra vertices generated by
instancing. This also requires still more munging of index buffers for
indexed draws.

Indirect draws are even more painful. Because the number of vertices and
instances is unknown, storage for the maximum possible number of
vertices must be allocated. This change imposes a totally arbitrary
limit of 131072 vertices from a single draw, including all vertices
generated by instancing. On a Mac, this requires about 194-256 MB of
VRAM for all the temporary buffers.

There are some possible optimizations here. If we could prove that the
vertex shader's output doesn't depend on the instance ID, either
directly or through a per-instance attribute, then we could avoid
running the vertex and tess. control stages per instance, and take
advantage of Metal's support for tess. eval instancing. If we could
also prove that the vertex shader simply passes instance attributes
through (similarly with the tess. control shader), we could do this for
many more instanced draws as well. It should also be possible to cache
the output from the tess. control stage; if the draw comes up again, we
can then skip the vertex and tess. control stages entirely!

Fixes KhronosGroup#56 and KhronosGroup#501.
@billhollings

This comment has been minimized.

Copy link
Contributor

commented Mar 16, 2019

Fixed in PR #508.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
You can’t perform that action at this time.