Viewport: Hair ribbon from geometry node have clipped ends #121020

New Issue

Clément Foucault · 2024-04-24T13:06:12+02:00

Clément Foucault commented

2024-04-24 13:06:12 +02:00

This is visible in the microfacet hair orientation render test case.

This was observed on both M1 and M2 GPUs.

This is visible in the `microfacet hair orientation` render test case. This was observed on both M1 and M2 GPUs. ![](https://projects.blender.org/attachments/1af19e41-ac77-4318-8225-06c6c1ddc382)

Capture d’écran 2024-04-24 à 13.06.46.png

61 KiB

Clément Foucault added this to the 4.2 LTS milestone 2024-04-24 13:06:12 +02:00

Clément Foucault added the

labels 2024-04-24 13:06:13 +02:00

Michael Parkin-White commented

2024-04-24 16:02:57 +02:00

Will have a look into this

Michael Parkin-White commented

2024-04-27 10:08:38 +02:00

Currently a little tricky to see where this is going wrong. The backend seems to be rendering what it is given, although the geometry is doubling up on certain strands, when i'm unsure if it should. The strand buffer data input from the texture appears correct, but not sure if it is going wrong prematurely.

Is this definitely still working 100% correctly on all other platforms? (Or is the golden image possibly out of date).
I can see that the curves should have more segments, but this does not seem to be taking any of the topology optimization or emulation paths in the Metal backend, which was my first thought.

Currently a little tricky to see where this is going wrong. The backend seems to be rendering what it is given, although the geometry is doubling up on certain strands, when i'm unsure if it should. The strand buffer data input from the texture appears correct, but not sure if it is going wrong prematurely. Is this definitely still working 100% correctly on all other platforms? (Or is the golden image possibly out of date). I can see that the curves should have more segments, but this does not seem to be taking any of the topology optimization or emulation paths in the Metal backend, which was my first thought.

Clément Foucault commented

2024-04-27 14:00:41 +02:00

Is this definitely still working 100% correctly on all other platforms? (Or is the golden image possibly out of date).

It definitely is. I can also take a look at it later.

> Is this definitely still working 100% correctly on all other platforms? (Or is the golden image possibly out of date). It definitely is. I can also take a look at it later.

Michael Parkin-White referenced this issue

2024-04-29 17:08:49 +02:00

Metal: Workbench curve drawing artifacts #121065

Michael Parkin-White commented

2024-04-29 17:14:11 +02:00

So this is happening on Metal, but I imagine the same issue would occur in OpenGL if use_strip_prims was also false in OpenGL.

source/blender/draw/intern/draw_cache_impl_curves.cc - static void calc_final_indices

bool use_strip_prims = (GPU_backend_get_type() != GPU_BACKEND_METAL);

it looks like the code is going wrong somewhere in this process, whether it is the triangle index buffer building in: source/blender/gpu/shaders/gpu_shader_index_2d_array_tris.glsl called by GPU_indexbuf_build_curves_on_device(..) or some other part of the logic, this is something i'm investigating.

From what I can see, using the triangle strip path looks visually correct in Metal. The Triangles path is non-essential, but certainly faster for rendering in the case of lots of small curves, due to better GPU scaling of non-restart-compatible primitive types.

So this is happening on Metal, but I imagine the same issue would occur in OpenGL if `use_strip_prims` was also false in OpenGL. `source/blender/draw/intern/draw_cache_impl_curves.cc - static void calc_final_indices` ``` bool use_strip_prims = (GPU_backend_get_type() != GPU_BACKEND_METAL); ``` it looks like the code is going wrong somewhere in this process, whether it is the triangle index buffer building in: `source/blender/gpu/shaders/gpu_shader_index_2d_array_tris.glsl` called by `GPU_indexbuf_build_curves_on_device(..)` or some other part of the logic, this is something i'm investigating. From what I can see, using the triangle strip path looks visually correct in Metal. The Triangles path is non-essential, but certainly faster for rendering in the case of lots of small curves, due to better GPU scaling of non-restart-compatible primitive types.

Jason Fielder referenced a pull request that will close this issue

2024-04-29 17:56:30 +02:00

Fix #121020: Fix curve index buffer generation with Triangle prim #121218

Michael Parkin-White commented

2024-04-29 17:58:02 +02:00

Looks like a small stride error during indexbuf building for triangles. PR has been submitted, but not sure if this fix would work universally for all cases. (As it may depend on if this code is used elsewhere and expected to behave in the current way).

Hans Goudey added

and removed

labels 2024-04-29 18:14:50 +02:00

Clément Foucault referenced this issue from a commit

2024-04-30 10:49:21 +02:00

Fix #121020: Fix curve index buffer generation with Triangle prim

Clément Foucault closed this issue

2024-04-30 10:49:22 +02:00