The light tree itself is disabled on the AMD GPUs due to a compiler issue.
There are couple of places where this was not fully checked:
- The `light_sample` function in the kernel.
- The light threshold during synchronization
The former one is solved as easy as just adding an ifdef block.
The latter one is solved by delaying the threshold assignment for
later on.
Pull Request #105022
This reverts commit 19222627c6.
Something went wrong here, seems like this commit merged the main branch
into the release branch, which should never be done.
This reverts commit 68181c2560.
I merged 3.6 into 3.5 by mistake. Basically I had a PR against main,
then changed it in the last minute to be against 3.5 via the
web-interface unaware that I shouldn't do it without updating the
patch.
Original Pull Request: #104889
Note that the node group has its sockets names
translated, while the built-in nodes don't.
So we need to use data_ for the built-in nodes names,
and the sockets of the created node groups.
Pull Request #104889
The proper fix (bb9eb262d4) caused compilation problems with HIP, so we're
delaying it until 3.6.
To fix the original bug report (#104586), this is a quick workaround that'll
hopefully not upset the compiler.
Pull Request #104723
The internal compiler error appears to be gone. Unclear why it appeared in the
first place and why it's gone now. Just random kernel code changes causing it.
Pull Request #104719
- Rename roughness variables for more clarity - before, the SVM/OSL code would
set s and v to the linear roughness values, and the setup function would over-
write them with the distribution parameters. This actually caused a bug in the
albedo code, since it intended to use the linear roughness value, but ended up
getting the remapped value.
- Deduplicate the evaluation and sample functions. Most of their code is the
same, only the middle part is different.
- Changed albedo computation to return the sum of the intensities of the four
BSDF lobes. Previously, the code applied the inverse of the color->sigma
mapping from the paper - this returns the color specified in the node, but
for very dark hair (e.g. when using the Melanin controls) the result is
extremely low (e.g. 0.000001) despite the hair still reflecting a significant
amount of light (since the R lobe is independent of sigma). This causes issues
with the light component passes, so this change fixes#104586.
- There's quite a few computations at the start of the evaluation function that
are needed for sampling, evaluation and albedo computation, but only depend on
the view direction. Therefore, just precompute them - we still have space in
PrincipledHairExtra after all.
- Fix a tiny bug - the direction sampling code did not account for the R lobe
roughness modifier.
Pull Request #104669
As a side effect of this change, more resolution divisions are now available.
Before this patch the possible resolution divisions were all powers of two.
Now the possible resolution divisions are the multiples of pixel_size.
This increase in possible resolution divisions is the same idea proposed in https://archive.blender.org/developer/D13590.
In that patch there were concerns that this will increase the time between a user navigating
and seeing the 1:1 render. To my knowledge this is a non-issue and there should be
little to no increase in time between those two events.
Pull Request #104450
(Follow on from D17043)
On AMD Navi2 devices the MetalRT checkbox was not hooked up properly and had no effect. This patch fixes it.
Co-authored-by: Michael Jones <michael_p_jones@apple.com>
Pull Request #104520
The required version numbers for various devices was hardcoded in the
UI messages. The result was that every time one of these versions was
bumped, every language team had to update the message in question.
Instead, the version numbers can be extracted, and injected into the
error messages using string formatting so that translation updates
need happen less frequently.
Pull Request #104488
Host memory fallback in CUDA and HIP devices is almost identical.
We remove duplicated code and create a shared generic version that
other devices (oneAPI) will be able to use.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D17173
This patch optimises subsurface intersection queries on MetalRT. Currently intersect_local traverses from the scene root, retrospectively discarding all non-local hits. Using a lookup of bottom level acceleration structures, we can explicitly query only the relevant instance. On M1 Max, with MetalRT selected, this can give a render speedup of 15-20% for scenes like Monster which make heavy use of subsurface scattering.
Patch authored by Marco Giordano.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D17153
This patch adds two new kernels: SORT_BUCKET_PASS and SORT_WRITE_PASS. These replace PREFIX_SUM and SORTED_PATHS_ARRAY on supported devices (currently implemented on Metal, but will be trivial to enable on the other backends). The new kernels exploit sort partitioning (see D15331) by sorting each partition separately using local atomics. This can give an overall render speedup of 2-3% depending on architecture. As before, we fall back to the original non-partitioned sorting when the shader count is "too high".
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16909
This patch removes the option to select both AMD and Intel GPUs on system that have both. Currently both devices will be selected by default which results in crashes and other poorly understood behaviour. This patch adds precedence for using any discrete AMD GPU over an integrated Intel one. This can be overridden with CYCLES_METAL_FORCE_INTEL.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D17166
This patch fixes T103393 by undefining `__LIGHT_TREE__` on Metal/AMD as it has an unexpected & major impact on performance even when light trees are not in use.
Patch authored by Prakash Kamliya.
Reviewed By: brecht
Maniphest Tasks: T103393
Differential Revision: https://developer.blender.org/D17167
This is both a cleanup and a preparation for the Principled v2 changes.
Notable changes:
- Clearcoat weight is now folded into the closure weight, there's no reason
to track this separately.
- There's a general-purpose helper for computing a Closure's albedo, which is
currently used by the denoising albedo and diffuse/gloss/transmission color
passes.
- The d/g/t color passes didn't account for closure albedo before, this means
that e.g. metallic shaders with Principled v2 now have their color texture
included in the glossy color pass. Also fixes T104041 (sheen albedo).
- Instead of precomputing and storing the albedo during shader setup, compute
it when needed. This is technically redundant since we still need to compute
it on shader setup to adjust the sample weight, but the operation is cheap
enough that freeing up the storage seems worth it.
- Future changes (Principled v2) are easier to integrate since the Fresnel
handling isn't all over the place anymore.
- Fresnel handling in the Multiscattering GGX code is still ugly, but since
removing that entirely is the next step, putting effort into cleaning it up
doesn't seem worth it.
- Apart from the d/g/t color passes, no changes to render results are expected.
Differential Revision: https://developer.blender.org/D17101
Cycles ignores the size of spot lights, therefore the illuminated area doesn't match the gizmo. This patch resolves this discrepancy.
| Before (Cycles) | After (Cycles) | Eevee
|{F14200605}|{F14200595}|{F14200600}|
This is done by scaling the ray direction by the size of the cone. The implementation of `spot_light_attenuation()` in `spot.h` matches `spot_attenuation()` in `lights_lib.glsl`.
**Test file**:
{F14200728}
Differential Revision: https://developer.blender.org/D17129
The image manager used to handle OSL textures on the GPU by
default loads images after displacement is evaluated. This is a
problem when the displacement shader uses any textures, hence
why the geometry manager already makes the image manager
load any images used in the displacement shader graph early
(`GeometryManager::device_update_displacement_images`).
This only handled Cycles image nodes however, not OSL nodes, so
if any `texture` calls were made in OSL those would be missed and
therefore crash when accessed on the GPU. Unfortunately it is not
simple to determine which textures referenced by OSL are needed
for displacement, so the solution for now is to simply load all of
them early if true displacement is used.
This patch also fixes the result of the displacement shader not
being used properly in OptiX.
Maniphest Tasks: T104240
Differential Revision: https://developer.blender.org/D17162
Also minor changes in comments:
- Reference BLENDER_HISTORY_FILE instead of the literal file-name
(simplifies looking up usage).
- Use usernames in tags, as noted in code-style.
The `MultiDevice` implementation of `get_cpu_osl_memory` returns a
nullptr when there is no CPU device in the mix. As such access to that
crashed in `update_osl_globals`. But that only updates maps that are not
currently used on the GPU anyway, so can just skip that when the CPU
is not used for rendering.
Maniphest Tasks: T104216
Removing all OSL script nodes from the shader graph would cause that
graph to no longer report it using `KERNEL_FEATURE_SHADER_RAYTRACE`
via `ShaderManager::get_graph_kernel_features`, but the shader object
itself still would have the `has_surface_raytrace` field set.
This caused kernels to be reloaded without shader raytracing support, but
later the `DEVICE_KERNEL_INTEGRATOR_SHADE_SURFACE_RAYTRACE`
kernel would still be invoked since the shader continued to report it
requiring that through the `SD_HAS_RAYTRACE` flag set because of
`has_surface_raytrace`.
Fix that by ensuring `has_surface_raytrace` is reset on every shader update,
so that when all OSL script nodes are deleted it is set to false, and only
stays true when there are still OSL script nodes (or other nodes using it).
Maniphest Tasks: T104157
Differential Revision: https://developer.blender.org/D17140
Based on "Sampling the GGX Distribution of Visible Normals" by Eric Heitz
(https://jcgt.org/published/0007/04/01/).
Also, this removes the lambdaI computation from the Beckmann sampling code and
just recomputes it below. We already need to recompute for two other cases
(GGX and clearcoat), so this makes the code more consistent.
In terms of performance, I don't expect a notable impact since the earlier
computation also was non-trivial, and while it probably was slightly more
accurate, I'd argue that being consistent between evaluation and sampling is
more important than absolute numerical accuracy anyways.
Differential Revision: https://developer.blender.org/D17100
This gives closer results to what I've seen in papers and other renderers when
using the code to precompute albedo later (to replace MultiGGX).
It's usually a tiny difference, the only case where I've seen it matter is
in the `shader/node_group_float.blend` test - but that's a (single-scatter) GGX
closure with 0.9 roughness, so it's not too surprising. In any case, the new
result looks closer to Eevee, so that's good I guess.
Differential Revision: https://developer.blender.org/D17099
Speckles and missing lights were experienced in scenes with Nishita Sky
Texture and a Sun Size smaller than 1.5°, such as in Lone Monk and Attic
scenes.
Increasing the precision of cosf fixes it.
A noticeable (>5%) performance regression in oneAPI backend came with
a501a2dbff. Updating to latest graphics
compiler from driver 101.4032 fixes it.
I've tested it with current min-supported drivers and it runs well but
since compatibility of graphics compiler with older drivers isn't
guaranteed, I'm also bumping the min-supported driver versions.
If end-users consider latest drivers too fresh to switch to (version
isn't released as stable on Linux as of today but should be before
Blender 3.5 release), CYCLES_ONEAPI_ALL_DEVICES=1 env variable can be
used.
Intel Graphics Compiler on Linux will be updated in a later commit
so we can then close D16984.
Reviewed By: sergey, LazyDodo
