This issue only affects profiling mode (`CYCLES_METAL_PROFILING=1`). There's a modest limit to the number of concurrent counter sampling buffers per device, so instead of creating one per device queue, we create one per device that can be reused by successive device queues.
Authored by Emma Liu.
Pull Request: https://projects.blender.org/blender/blender/pulls/136248
This PR removes a bunch of dead code following #123551 (removal of AMD and Intel GPU support). It is safe to assume that UMA will be available, so a lot of codepaths that dealt with copying between CPU and GPU are now just clutter.
Pull Request: https://projects.blender.org/blender/blender/pulls/136146
This PR removes a bunch of dead code following #123551 (removal of AMD and Intel GPU support). It is safe to assume that UMA will be available, so a lot of codepaths that dealt with copying between CPU and GPU are now just clutter.
Pull Request: https://projects.blender.org/blender/blender/pulls/136117
Check was misc-const-correctness, combined with readability-isolate-declaration
as suggested by the docs.
Temporarily clang-format "QualifierAlignment: Left" was used to get consistency
with the prevailing order of keywords.
Pull Request: https://projects.blender.org/blender/blender/pulls/132361
* Use .empty() and .data()
* Use nullptr instead of 0
* No else after return
* Simple class member initialization
* Add override for virtual methods
* Include C++ instead of C headers
* Remove some unused includes
* Use default constructors
* Always use braces
* Consistent names in definition and declaration
* Change typedef to using
Pull Request: https://projects.blender.org/blender/blender/pulls/132361
Running Xcode memory graphs and the Instruments tools revealed
memory leaks caused, in the main, by over-retained objects.
This removes the unnecessary 'retains' and adds some asserts
to guard against over-retaining in the future.
There are a few memory leaks remaining involving PyUnicode_DecodeUTF8
but I am unable to identify the cause of these at this time.
Authored by Apple: James McCarthy
Pull Request: https://projects.blender.org/blender/blender/pulls/129117
This PR fixes live viewport stability issues on Mac when MetalRT is enabled.
There were two sources of instability:
1) `MTLAccelerationStructure` instances were not being correctly retained meaning that use-after-free crashes could occur following a geometry sync.
2) `MTLIntersectionFunctionTable` objects could be unsafely shared between multiple `MetalDeviceQueue` instances (in this case, `setBuffer` being the unsafe mutation)
The solution to 2 involves creating a new `MetalDispatchPipeline` type which is strictly used by only 1 `MetalDeviceQueue` instance.
Pull Request: https://projects.blender.org/blender/blender/pulls/124055
This is because with the addition of new features to Cycles, these GPUs
experienced significant performance regressions and bugs, all stemming
from bugs in the Metal GPU driver/compiler. The only reasonable way to
work around these issues was to disable parts of Cycles code on
these GPUs to avoid the driver/compiler bugs.
This resulted in increased development time maintaining these platforms
while being unable to deliver feature parity with other
GPU backends.
It has been decided that this development time is better spent
maintaining platforms that are still actively maintained by
hardware/software vendors, and so AMD and Intel GPU support will be
removed from the Metal backend for Cycles.
Pull Request: https://projects.blender.org/blender/blender/pulls/123551
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 PR replaces the existing CPU wall-clock based profiling mechanism with more precise GPU counter based timestamps. As before, it is enabled by setting the env var `CYCLES_METAL_PROFILING=1`. Original implementation by Morteza Mostajabodaveh.
Pull Request: https://projects.blender.org/blender/blender/pulls/121208
Along with the 4.1 libraries upgrade, we are bumping the clang-format
version from 8-12 to 17. This affects quite a few files.
If not already the case, you may consider pointing your IDE to the
clang-format binary bundled with the Blender precompiled libraries.
This PR adds `@autoreleasepool` blocks around functions that have been observed to create hidden temporary NSObjects, and eventually cause command buffer failures. A couple of allocations needed to be tweaked in order to maintain correct retain/release behaviour. This PR also fixes the command buffer error text to show more useful information.
This PR fixes T39823, the sole failing unit test when running with MetalRT. It does so by implementing and binding a missing intersection handler (`__anyhit__cycles_metalrt_volume_test_tri`) which is required for `scene_intersect_volume` (as used by `integrator_volume_stack_update_for_subsurface`) to work as intended. This scene exposed the error as it uses subsurface scattering on a sphere which is intersected by volume.
Pull Request: https://projects.blender.org/blender/blender/pulls/112876
This patch updates the experimental MetalRT code path to use new [curve primitives](https://developer.apple.com/videos/play/wwdc2023/10128/) which were recently added in macOS 14. This replaces the previous custom box intersection implementation, allowing the driver to better optimise curve acceleration structures for the GPU. On existing hardware, this can speed up MetalRT renders by up to 40% for scenes that use hair / curve primitives extensively.
The MetalRT option will only be available on macOS >= 14, and requires Xcode >= 15 to build (otherwise the option will be compiled out).
Authored by Marco Giordano, Michael Jones, and Jason Fielder
---
Before / after render times (M1 Max MacBook Pro, macOS 14 beta, MetalRT enabled):
```
Custom box intersection MetalRT curve primitives Speedup
fishy_cat 111.5 80.5 1.39
koro 114.4 86.7 1.32
sinosauropteryx 291.8 279.2 1.05
spring 142.3 142.2 1.00
victor 442.7 347.7 1.27
```
---
Pull Request: https://projects.blender.org/blender/blender/pulls/111795
This patch replaces `dispatchThreadgroups` with `dispatchThreads` which takes care of non-uniform threadgroup bounds. This allows us to remove the bounds guards in the integrator kernel entry points.
Pull Request: https://projects.blender.org/blender/blender/pulls/106217
Contributed by Yulia Kuznetcova at Apple.
NanoVDB is patched to give add address spaces required by Metal. We hope that
in the future Metal will support the generic address space.
For AMD and Intel this is currently not available since it causes a performance
regression also on scenes without volumes.
Pull Request #104837
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 adds occupancy tuning for the newly announced high-end M2 machines, giving 10-15% render speedup over a pre-tuned build.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D17037
This patch adds a new "Kernel Optimization Level" dropdown menu to control Metal kernel specialisation. Currently this defaults to "full" optimisation, on the assumption that the changes proposed in D16371 will address usability concerns around app responsiveness and shader cache housekeeping.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16514
All kernel specialisation is now performed in the background regardless of kernel type, meaning that the first render will be visible a few seconds sooner. The only exception is during benchmark warm up, in which case we wait for all kernels to be cached. When stopping a render, we call a new `cancel()` method on the device which causes any outstanding compilation work to be cancelled, and we destroy the device in a detached thread so that any stale queued compilations can be safely purged without blocking the UI for longer than necessary.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16371
This patch tunes the integrator state sizing for Metal (`num_concurrent_states` and `num_concurrent_busy_states`).
On all GPUs architecture, we adjust the busy:total states ratio to be 1:4 which gives better rendering performance than the previous 1:16 ratio (independent of total state count). This gives a small performance uplift (e.g. 2-3% on M1 Ultra).
Additionally for M2 architectures, we double the overall state size if there is available headroom. Inclusive of the first change, we can expect uplift of close to 10% in future, as this results in larger dispatch sizes and minimises work submission overheads. In order to make an accurate determination of available headroom, we defer the calculation of `num_concurrent_states` and `num_concurrent_busy_states` until the time of integrator state allocation (i.e. after all of the scene data has been allocated). We also refactor `alloc_integrator_soa` to calculate an *exact* single-state-size in a first pass, right before allocating the integrator SoA buffers in a second pass.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D16313
This was added for Metal, but also gives good results with CUDA and OptiX.
Also enable it for future Apple GPUs instead of only M1 and M2, since this has
been shown to help across multiple GPUs so the better bet seems to enable
rather than disable it.
Also moves some of the logic outside of the Metal device code, and always
enables the code in the kernel since other devices don't do dynamic compile.
Time per sample with OptiX + RTX A6000:
new old
barbershop_interior 0.0730s 0.0727s
bmw27 0.0047s 0.0053s
classroom 0.0428s 0.0464s
fishy_cat 0.0102s 0.0108s
junkshop 0.0366s 0.0395s
koro 0.0567s 0.0578s
monster 0.0206s 0.0223s
pabellon 0.0158s 0.0174s
sponza 0.0088s 0.0100s
spring 0.1267s 0.1280s
victor 0.0524s 0.0531s
wdas_cloud 0.0817s 0.0816s
Ref D15331, T87836
This patch partitions the active indices into chunks prior to sorting by material in order to tradeoff some material coherence for better locality. On Apple Silicon GPUs (particularly higher end M1-family GPUs), we observe overall render time speedups of up to 15%. The partitioning is implemented by repeating the range of `shader_sort_key` for each partition, and encoding a "locator" key which distributes the indices into sorted chunks.
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D15331
If there is an error we should stop rendering, instead of finishing with a
wrong render result or reporting a wrong benchmark time.
Ref T96519
Differential Revision: https://developer.blender.org/D15287
* Rename "texture" to "data array". This has not used textures for a long time,
there are just global memory arrays now. (On old CUDA GPUs there was a cache
for textures but not global memory, so we used to put all data in textures.)
* For CUDA and HIP, put globals in KernelParams struct like other devices.
* Drop __ prefix for data array names, no possibility for naming conflict now that
these are in a struct.
This patch adds a new mode of gpu capture (env var `CYCLES_DEBUG_METAL_CAPTURE_SAMPLES`) to capture a block of dispatches between "reset" calls. It also fixes member data naming inconsistencies and adds some missing OS version checks.
Screenshot showing .gputrace capture in Xcode 14.0 beta (using `CYCLES_DEBUG_METAL_CAPTURE_SAMPLES="1"` and `CYCLES_DEBUG_METAL_CAPTURE_LIMIT="10"`):
{F13155703}
Reviewed By: sergey, brecht
Differential Revision: https://developer.blender.org/D15179
Move MNEE to own kernel, separate from shader ray-tracing. This does introduce
the limitation that a shader can't use both MNEE and AO/bevel, but that seems
like the better trade-off for now.
We can experiment with bigger kernel organization changes later.
Differential Revision: https://developer.blender.org/D15070
This is a stripped down version of D14645 without the scene specialisation optimisations.
The two major changes in this patch are:
- Enables more aggressive inlining on Apple Silicon resulting in a 1.1x speedup and 10% reduction in spill, at the cost of longer pipeline build times
- Revival of shader binary archives through a new ShaderCache which is shared between MetalDevice instances using the same physical MTLDevice. This mitigates the extra compile times via explicit caching (rather than, as before, relying on the implicit system shader cache which can be purged without notice)
Reviewed By: brecht
Differential Revision: https://developer.blender.org/D14763
