Cycles Light Tree: Add case for when an emitter is not selected #111292

Alaska · 2023-08-19T11:17:06+02:00

Alaska commented

2023-08-19 11:17:06 +02:00

Discovered during an investigation into #111277
in rare situations (E.G. When normals are NaN), an emitter
won't be selected as part of light_tree_cluster_select_emitter()
and as a result of that, an emitter_index of -1 is passed to
kernel_data_fetch(light_tree_emitters, emitter_index) resulting in
an "illegal address" error on some devices.

Discovered during an investigation into #111277 in rare situations (E.G. When normals are NaN), an emitter won't be selected as part of `light_tree_cluster_select_emitter()` and as a result of that, an `emitter_index` of `-1` is passed to `kernel_data_fetch(light_tree_emitters, emitter_index)` resulting in an "illegal address" error on some devices.

Alaska added the

Module

Render & Cycles

label 2023-08-19 11:17:06 +02:00

Alaska added 1 commit 2023-08-19 11:17:18 +02:00

8b7d210301 Cycles Light Tree: Gracefully handle when an emitter is not selected

Discovered during an investigation into #111277 in rare situations,
an emitter won't be selected as part of
`light_tree_cluster_select_emitter()` and as a result of that,
an `emitter_index` of `-1` is passed to
`kernel_data_fetch(light_tree_emitters, emitter_index)` resulting in
an "illegal address" error on some devices.

Alaska requested review from Brecht Van Lommel 2023-08-19 11:17:30 +02:00

Alaska requested review from Weizhen Huang 2023-08-19 11:17:40 +02:00

Alaska commented

2023-08-19 11:18:06 +02:00

This commit should also be backported to Blender 3.6 if it is accepted.

Alaska referenced this pull request

2023-08-19 11:18:08 +02:00

Vector displacement suspected to cause geometry artefact and CUDA error #111277

Alaska referenced this pull request

2023-08-21 13:30:32 +02:00

Fix #111277: NaN in Vector Displacement leading to render errors #111294

Weizhen Huang commented

2023-08-22 10:31:31 +02:00

Is this still needed after #111294 is merged?
I am expecting light_tree_cluster_select_emitter() to pick an emitter if there is one, unless all emitters in the cluster have zero importance[0]. If any input is invalid it suggests that something else is wrong.

Is this still needed after #111294 is merged? I am expecting `light_tree_cluster_select_emitter()` to pick an emitter if there is one, unless all emitters in the cluster have zero `importance[0]`. If any input is invalid it suggests that something else is wrong.

Alaska commented

2023-08-22 11:24:24 +02:00

I am expecting light_tree_cluster_select_emitter() to pick an emitter if there is one, unless all emitters in the cluster have zero importance[0]. If any input is invalid it suggests that something else is wrong.

You're right. An emitter should be selected under normal conditions. But under bad conditions (E.G. When the normals are NaN), an emitter won't be selected and this catch is useful as it stops an illegal address error on the GPU. (E.G. If #111294 wasn't applied, then having this catch would still allow people to render the scene shared in #111277 on a GPU, even though it has NaN normals).

Since #111294 has been merged into main (and is being back ported to 3.6), this pull request is no longer needed to fix any of the issues reported in #111277. But having this pull request merged could still be useful for situations in the future when the normals or some other value is NaN (due to something else being broken in Blender/Cycles) and a user is trying to render on the GPU, because this pull request will stop the Light Tree Sampling code from reaching a kernel_assert and reporting an "illegal address" error on GPUs in those scenes.

It comes down to how you want to handle things.

Decline this pull request because under normal conditions everything works. If someone tries rendering a broken mesh/shader in the future, an emitter may not be selected, GPUs will spit out errors and refuse to render, a user makes a report, and a Blender/Cycles developer fixes the underlying geometry/shader issue.

Or

Accept this pull request so GPUs will still render even when rendering with a broken mesh/shader, and when a user reports rendering artifacts caused by the geometry/shader, then a Blender/Cycles developers goes in and fixes the issue. (Although as a side effect of avoiding the illegal address error, users are less likely to report these sorts of bugs for scenes that look alright, but do have issues)

From the perspective of a Blender user with no knowledge of programming, I would personally like this pull request to be accepted. It means I can keep rendering with the GPU even when something else is broken (especially if the broken thing only takes up a small part of the scene, such as a single face on a chair leg having NaN normals in a render of the exterior of a building).

As a casual developer, I can see both accepting and declining as valid options based on different criteria and preferences. I'll leave the decision up to you.

> I am expecting `light_tree_cluster_select_emitter()` to pick an emitter if there is one, unless all emitters in the cluster have zero `importance[0]`. If any input is invalid it suggests that something else is wrong. You're right. An emitter should be selected under normal conditions. But under bad conditions (E.G. When the normals are NaN), an emitter won't be selected and this catch is useful as it stops an illegal address error on the GPU. (E.G. If #111294 wasn't applied, then having this catch would still allow people to render the scene shared in #111277 on a GPU, even though it has NaN normals). Since #111294 has been merged into main (and is being back ported to 3.6), this pull request is no longer needed to fix any of the issues reported in #111277. But having this pull request merged could still be useful for situations in the future when the normals or some other value is NaN (due to something else being broken in Blender/Cycles) and a user is trying to render on the GPU, because this pull request will stop the Light Tree Sampling code from reaching a kernel_assert and reporting an "illegal address" error on GPUs in those scenes. It comes down to how you want to handle things. - Decline this pull request because under normal conditions everything works. If someone tries rendering a broken mesh/shader in the future, an emitter may not be selected, GPUs will spit out errors and refuse to render, a user makes a report, and a Blender/Cycles developer fixes the underlying geometry/shader issue. Or - Accept this pull request so GPUs will still render even when rendering with a broken mesh/shader, and when a user reports rendering artifacts caused by the geometry/shader, then a Blender/Cycles developers goes in and fixes the issue. (Although as a side effect of avoiding the illegal address error, users are less likely to report these sorts of bugs for scenes that look alright, but do have issues) --- From the perspective of a Blender user with no knowledge of programming, I would personally like this pull request to be accepted. It means I can keep rendering with the GPU even when something else is broken (especially if the broken thing only takes up a small part of the scene, such as a single face on a chair leg having NaN normals in a render of the exterior of a building). As a casual developer, I can see both accepting and declining as valid options based on different criteria and preferences. I'll leave the decision up to you.

Alaska commented

2023-08-23 09:09:10 +02:00

Extra note because I just remembered it.

It's possible a user could still create a NaN normal with something like a custom OSL script.

Extra note because I just remembered it. It's possible a user could still create a NaN normal with something like a custom OSL script.

Weizhen Huang approved these changes 2023-08-23 11:52:10 +02:00

Weizhen Huang left a comment

You are right, if the normal is wrong it's still visible as an artefact, this patch is just to cover illegal address error, so I think it's fine. Thanks!

Weizhen Huang merged commit 206ab6437b into main

2023-08-23 11:54:28 +02:00

Weizhen Huang referenced this issue from a commit

2023-08-23 11:54:28 +02:00

Fix illegal address error in Cycles Light Tree when no emitter is selected

Weizhen Huang commented

2023-08-23 17:14:13 +02:00

Hi, turns out that the selected emitter can be -1 because all the importance[1](min_importance) are zero, which is allowed at the line you changed, so the tests failed.
The problem is that total_importance[0] == 0 failed to detect the case where all the entries are NaN, a proper fix would be to use !has_importance instead (c076202e23). I also moved the selected_emitter < 0 detection a bit earlier. min_importance seems fine because there is a fmaxf(..., 0.0f) when computing it which is NaN-aware.
Again, many thanks for spotting and investigating the issue!

Hi, turns out that the selected emitter can be -1 because all the `importance[1]`(`min_importance`) are zero, which is allowed at the line you changed, so the tests failed. The problem is that `total_importance[0] == 0` failed to detect the case where all the entries are NaN, a proper fix would be to use `!has_importance` instead (c076202e23767e7c1a26b60c402118f808c3f499). I also moved the `selected_emitter < 0` detection a bit earlier. `min_importance` seems fine because there is a `fmaxf(..., 0.0f)` when computing it which is NaN-aware. Again, many thanks for spotting and investigating the issue!

Alaska commented

2023-08-23 22:16:07 +02:00

Thanks for implementing a better fix.

Alaska deleted branch gracefully-handle-111277-issue

2023-08-23 22:16:21 +02:00

Brecht Van Lommel referenced this issue from a commit

2023-12-05 01:37:12 +01:00

Fix illegal address error in Cycles Light Tree when no emitter is selected

Sign in to join this conversation.

No reviewers

No Label

Download

What's New

Blender Studio

Manual

Developers Blog

Documentation

Benchmark

Blender Conference

Development Fund

One-time Donations

Cycles Light Tree: Add case for when an emitter is not selected #111292