VK_EXT_mesh_shader, version 2 #1125
Conversation
| #[spirv_std_macros::gpu_only] | ||
| #[doc(alias = "OpEmitMeshTasksEXT")] | ||
| #[inline] | ||
| pub unsafe fn emit_mesh_tasks_ext(group_count_x: u32, group_count_y: u32, group_count_z: u32) -> ! { |
There was a problem hiding this comment.
The documentation on OpEmitMeshTasksEXT mentions an optional payload pointer, which does not exist on this function. The compiletests contain examples on mesh and task shaders with payloads. But I'm unsure how rust-gpu handles payloads in general, and if this is the correct way to handle them or not.
There was a problem hiding this comment.
Resolved it by adding an explicit emit_mesh_tasks_ext_payload() that takes a payload param simiarly to raytracing.
Firestar99
force-pushed
the
actual_mesh_shaders
branch
3 times, most recently
from
7f51741
to
8af2717
Compare
Firestar99
force-pushed
the
actual_mesh_shaders
branch
from
8af2717
to
05e8c8d
Compare
|
I've successfully used mesh shaders generated by this branch in a small demo: I had a workgroup size of 1 with each invocation processing a single triangle and was able to render the "lantern" model from glTF-sample-assets. Proper meshlets to be implemented and tested soon. |
| pub fn main(#[spirv(push_constant)] push: &u32) { | ||
| let count = 20 / *push; | ||
| unsafe { | ||
| emit_mesh_tasks_ext(1, 2, 3); |
There was a problem hiding this comment.
This causes rust-gpu to emit invalid byte code and I do not know why. Specifically, any possibility of a panic, like the 20 / *push when push is 0, will cause a mispile in any task shader that calls emit_mesh_tasks_ext() at the end (which is required by spec). Interestingly, this does not happen to OpKill in fragment shaders, so likely some misconfiguration on my part?
normalized stderr:
error: error:0:0 - Selection must be structured
OpBranchConditional %24 %25 %26
|
= note: spirv-val failed
= note: module `$TEST_BUILD_DIR/arch/task_shader_mispile.default`
error: aborting due to previous error
!!! builds on top of PR #1118, which needs to be merged first
Adds support for
VK_EXT_mesh_shader, rebased from the abandoned PR #1019 by @BeastLe9enD. Note that I have split the PR in two parts: PR #1118 handles updating to a more recent rspirv version, whereas this PR only handles the additions needed forVK_EXT_mesh_shadersupport.(Sorry if the diff currently looks horrid, as it currently shows 2 PRs in one. As soon as the other PR is merged, I'll rebase this one and the diff should be way easier to review on github)