Precompiled Cycles kernels make up a considerable fraction of the total size of
Blender builds nowadays. As we add more features and support for more
architectures, this will only continue to increase.
However, since these kernels tend to be quite compressible, we can save a lot
of storage by storing them in compressed form and decompressing the required
kernel(s) during loading.
By using Zstandard compression with a high level, we can get decent compression
ratios (~5x for the current kernels) while keeping decompression time low
(about 30ms in the worse case in my tests). And since we already require zstd
for Blender, this doesn't introduce a new dependency.
While the main improvement is to the size of the extracted Blender installation
(which is reduced by ~400-500MB currently), this also shrinks the download on
Windows, since .zip's deflate compression is less effective. It doesn't help on
Linux since we're already using .tar.xz there, but the smaller installed size
is still a good thing.
See #123522 for initial discussion.
Pull Request: https://projects.blender.org/blender/blender/pulls/123557
* Always define root directories in LIBDIR even when not needed,
to silence some warnings.
* Only show warnings about not finding libs when oneAPI is enabled.
* Prefix message for context.
Light linking was never working correctly in volume segment with light
tree, because `sd->object` was not assigned, thus
`light_link_receiver_nee(kg, sd)` always returned `OBJECT_NONE`, causing
the light tree sample to fail. This problem was revealed by fdc2962beb
since now the same light is used for volume segment and volume.
Also ensure we don't sample position on the light if sampling from
volume segment is failed, by setting `emitter_id` to `EMITTER_NONE` in
such cases.
Pull Request: https://projects.blender.org/blender/blender/pulls/122999
This patch implements blue-noise dithered sampling as described by Nathan Vegdahl (https://psychopath.io/post/2022_07_24_owen_scrambling_based_dithered_blue_noise_sampling), which in turn is based on "Screen-Space Blue-Noise Diffusion of Monte Carlo Sampling Error via Hierarchical Ordering of Pixels"(https://repository.kaust.edu.sa/items/1269ae24-2596-400b-a839-e54486033a93).
The basic idea is simple: Instead of generating independent sequences for each pixel by scrambling them, we use a single sequence for the entire image, with each pixel getting one chunk of the samples. The ordering across pixels is determined by hierarchical scrambling of the pixel's position along a space-filling curve, which ends up being pretty much the same operation as already used for the underlying sequence.
This results in a more high-frequency noise distribution, which appears smoother despite not being less noisy overall.
The main limitation at the moment is that the improvement is only clear if the full sample amount is used per pixel, so interactive preview rendering and adaptive sampling will not receive the benefit. One exception to this is that when using the new "Automatic" setting, the first sample in interactive rendering will also be blue-noise-distributed.
The sampling mode option is now exposed in the UI, with the three options being Blue Noise (the new mode), Classic (the previous Tabulated Sobol method) and the new default, Automatic (blue noise, with the additional property of ensuring the first sample is also blue-noise-distributed in interactive rendering). When debug mode is enabled, additional options appear, such as Sobol-Burley.
Note that the scrambling distance option is not compatible with the blue-noise pattern.
Pull Request: https://projects.blender.org/blender/blender/pulls/118479
On a M3 MacBook Pro, this change increases the benchmark score by 8% (with classroom seeing a path-tracing speedup of 15%).
The integrator state is currently store using struct-of-arrays, with one array per field. Such fine grained separation can result in poor GPU cache utilisation in cases where multiple fields of the same parent struct are accessed together. This PR changes the layout of the `ray`, `isect`, `subsurface`, and `shadow_ray` structs so that the data is interleaved (per parent struct) instead of separate. To try and keep this change localised, I encapsulated the layout change by extending the integrator state access macros, however maybe we want to do this more explicitly? (e.g. by updating every bit of code that accesses these parts of the state). Feedback welcome.
Pull Request: https://projects.blender.org/blender/blender/pulls/122015
This fixes#69535 and #98930.
We use a equi-solid-angle sampling algorithm for rectangular area lights,
but it is not particularly robust for small area lights (either small
in general and/or small because it's being viewed from grazing angles).
The actual sampling part is fine since it just gets clamped into the
valid area anyways, and the difference isn't notable for small lights.
However, we also need to compute the solid angle to get the sampling PDF,
and that computation is quite sensitive to numerical issues for small
values.
Therefore, this commit adds a fallback path for small values, which instead
uses the classic equi-area sampling PDF term times the area-to-solid-angle
Jacobian term. This approximation assumes that all points on the light have
the same distance and angle to the sampling point, which is of course not
strictly the case, but it's close enough for small area lights and better
than failing altogether.
Pull Request: https://projects.blender.org/blender/blender/pulls/122323
Reformulates some terms in the equi-solid-angle rectangle sampling code to
handle small area lamps better, and allows for some rounding error in the
check whether the sampled position is inside the area light.
Pull Request: https://projects.blender.org/blender/blender/pulls/122323
In the original paper, the falloff inside `bcone.theta_e` is assumed to
be `pi/2`, which is too large for spot light and resulted in an
overestimation near the cone boundary.
To address this issue, attenuate the energy of a spot light using the
minimal possible angle formed by the light axis and the shading point
when traversing the light tree.
Ref: #122362
Pull Request: https://projects.blender.org/blender/blender/pulls/122667
Since the previous fix to properly support volumes and transparent objects
it became very easy to make it so the intersection loop takes all 1024
iterations to find intersections.
This change makes it so the number of intersection is limited by the max
number of volume/transparent bounces.
This should minimize possible performance impact of the previous fix.
Pull Request: https://projects.blender.org/blender/blender/pulls/122448
We can't do the optimization to shorten the ray when we might still need
to go through transparent surfaces or volumes to reach the light.
This issue was not light tree specific, however in the test file it was
more noticable because the light tree poorly handles some areas. This in
in turn causes MIS weights for forward path tracing to become higher,
which is where the error was.
Pull Request: https://projects.blender.org/blender/blender/pulls/122404
This enables scenes with all textures not fitting in GPU
memory to finally render. For scenes that are fitting,
no functional change or performance change is expected.
Pull Request: https://projects.blender.org/blender/blender/pulls/122385
One of the properties of Perlin noise is that it always evaluates to 0.0
when not normalized (or 0.5 when normalized) when the input consists of
only whole integers in all vector components.
Blender's Perlin noise implementation uses single precision floats with
a machine epsilon of 1.19e-07 meaning that for numbers that are greater
than 1/(1.19e-07) = 8.40e6 there mantissa doesn't have any bits left to
store a rational part of the number, effectively meaning that any number
greater than 8.40e6 is a whole integer as far as Blender is concerned.
Therefore when evaluating Perlin noise for any coordinates greater than
that it always results in 0.0 (or 0.5 when normalized).
This fix works as follows: If the original input number is larger than
1.0e6 it is offset by 0.5 after it underwent modulo, which always outputs
numbers in a [0.0, 1.0e5) range leaving the mantissa room for a rational
part. This way the quantization error still persists however the outputs
are random again instead of a constant 0.0.
Pull Request: https://projects.blender.org/blender/blender/pulls/122112
The refactor in 97d9bbbc97 changed the way q is computed in the spherical triangle sampling code. While the new approach is more efficient and saves a few operations, it introduces numerical precision issues for skinny/small (spherical) triangles.
Therefore, this change moves the computation of q back to the method from the paper, while keeping the more efficient solid angle computation.
Pull Request: https://projects.blender.org/blender/blender/pulls/119224
when ray exceeds `max_bounce`, we do not allocate any closure at
intersection. However, Ray Portal BSDF still added `SD_BSDF` flag,
resulting in undefined behavior in
`integrate_surface_bsdf_bssrdf_bounce()`.
This part of code was similar to Transparent BSDF, however, Transparent
closure was still allocated in this case.
To fix the undefined behavior, add `SD_BSDF` flag only when the Ray
Portal closure was allocated.
This PR contains optimisations and a general tidy-up of the MetalRT backend.
- Currently `scene_intersect` is used for both normal and (opaque) shadow rays, however the usage patterns are different enough to warrant specialisation. Shadow intersection tests (flagged with `PATH_RAY_SHADOW_OPAQUE`) only need a bool result, but need a larger "self" payload in order to exclude hits against target lights. By specialising we can minimise the payload size in each case (which is helps performance) and avoid some dynamic branching. This PR introduces a new `scene_intersect_shadow` function which is specialised in Metal, and currently redirects to `scene_intersect` in the other backends.
- Currently `scene_intersect_local` is implemented for worst-case payload requirements as demanded by `subsurface_disk` (where `max_hits` is 4). The random_walk case only demands 1 hit result which we can retrieve directly from the intersector object (rather than stashing it in the payload). By specialising, we significantly reduce the payload size for random_walk queries, which has a big impact on performance. Additionally, we only need to use a custom intersection function for the first ray test in a random walk (for self-primitive filtering), so this PR forces faster `opaque` intersection testing for all but the first random walk test.
- Currently `scene_intersect_volume` has a lot of redundant code to handle non-triangle primitives despite volumes only being enclosed by trimeshes. This PR removes this code.
Additionally, this PR tidies up the convoluted intersection function linking code, removes some redundant intersection handlers, and uses more consistent naming of intersection functions.
On a M3 MacBook Pro, these changes give 2-3% performance increase on typical scenes with opaque trimesh materials (e.g. barbershop, classroom junkshop), but can give over 15% performance increase for certain scenes using random walk SSS (e.g. monster).
Pull Request: https://projects.blender.org/blender/blender/pulls/121397
This is an implementation of thin film iridescence in the Principled BSDF based on "A Practical Extension to Microfacet Theory for the Modeling of Varying Iridescence".
There are still several open topics that are left for future work:
- Currently, the thin film only affects dielectric Fresnel, not metallic. Properly specifying thin films on metals requires a proper conductive Fresnel term with complex IOR inputs, any attempt of trying to hack it into the F82 model we currently use for the Principled BSDF is fundamentally flawed. In the future, we'll add a node for proper conductive Fresnel, including thin films.
- The F0/F90 control is not very elegantly implemented right now. It fundamentally works, but enabling thin film while using a Specular Tint causes a jump in appearance since the models integrate it differently. Then again, thin film interference is a physical effect, so of course a non-physical tweak doesn't play nicely with it.
- The white point handling is currently quite crude. In short: The code computes XYZ values of the reflectance spectrum, but we'd need the XYZ values of the product of the reflectance spectrum and the neutral illuminant of the working color space. Currently, this is addressed by just dividing by the XYZ values of the illuminant, but it would be better to do a proper chromatic adaptation transform or to use the proper reference curves for the working space instead of the XYZ curves from the paper.
Pull Request: https://projects.blender.org/blender/blender/pulls/118477
This PR fixes the (currently unused) scene-based selective feature compilation macros. These feature based macros haven't been used for a few years, and enabling them currently results in compilation errors.
The only functional change in this PR is in geom/primitive.h where undef-ing `__HAIR__` had exposed an inconsistency in how pointcloud attributes were being fetched. Using the more general `primitive_surface_attribute_float4` (instead of `curve_attribute_float4`) fixed a compilation error that occurred when rendering pointcloud unit test scenes with adaptive compilation enabled.
Pull Request: https://projects.blender.org/blender/blender/pulls/121216
Transport rays that enter to another location in the scene, with
specified ray position and normal. This may be used to render portals
for visual effects, and other production rendering tricks.
This acts much like a Transparent BSDF. Render passes are passed
through, and this is affected by light path max transparent bounces.
Pull Request: https://projects.blender.org/blender/blender/pulls/114386
Clamp some of the inputs of the Glossy BSDF, Glass BSDF, Sheen BSDF,
and Subsurface Scattering nodes to improve consistency between render
engines and to avoid unexpected results.
* Clamp roughness to 0..1
* Clamp subsurface radius to 0..inf
* Clamp colors to 0..inf
Pull Request: https://projects.blender.org/blender/blender/pulls/120390
This replaces the fixed Tangent input in BsdfNode::compile
with a custom input.
This is done because very few nodes actually use the tangent input
and it would be better to have this slot available for other inputs
on different nodes in the future.
Pull Request: https://projects.blender.org/blender/blender/pulls/119042
by ensuring `KernelLight.lightgroup` is properly assigned in
`device_update_light()`. The value is later retrieved via
`lamp_lightgroup(kg, lamp)`.
`light_manager->device_update()` is called after
`background->device_update()`, so the background light group should
already been assigned when `device_update_lights()` is called.
Pull Request: https://projects.blender.org/blender/blender/pulls/120714
use available `film_pass_pixel_render_buffer()` to access the pointer
to the render buffer.
For shadow state, a similar function `film_pass_pixel_render_buffer_shadow()`
is created, because `shadow_path` instead of `path` is needed.
it is difficult to keep in mind when MIS weight is needed, better to
handle this logic in the lower-level functions.
This reduces code duplication in many places.
This is a regression since fdc2962beb
The size of state can not be different between CPU and GPU.
This change replaces compile-time condition with a kernel feature
check, which solves the render regression on AMD Metal. It also
minimizes the state size on other GPUs when Light Tree is disabled.
Pull Request: https://projects.blender.org/blender/blender/pulls/120476
This is a regression in 4.1, caused by 36e603c430.
For unbiased MIS weight in light tree, we should use sd->N for
mis_origin_n, since sc->N is not available in NEE.
The change also makes it so we do not sample lights below sd->N even
when bump map correction is disabled. This diverges from the original
idea of giving full control to artists, but ensures the internal math
is happy.
Pull Request: https://projects.blender.org/blender/blender/pulls/120216
fdc2962beb indirectly introduced a change
in inlining (light_tree_pdf started getting inlined) that led to a 5-10%
drop in performance for most scenes.
Dropping the noinline keyword for oneAPI device recovers it.
It however brings another performance regression to MNEE and Raytrace
kernels, that we'll look into separately.
The Perlin noise algorithms suffer from precision issues when a coordinate
is greater than about 250000.
To fix this the Perlin noise texture is repeated every 100000 on each axis.
This causes discontinuities every 100000, however at such scales this
usually shouldn't be noticeable.
Pull Request: https://projects.blender.org/blender/blender/pulls/119884
