Cycles: refactor to move part of KernelData definition to template headerarcpatch-D14645

To be used  for specialization on Metal in a following commit, turning these
members into compile time constants.

Cycles: keep track of SVM nodes used in kernels

To be used for specialization in Metal, to automatically leave out unused nodes
from the kernel.

Cycles: Apple Silicon optimizations (~20% uplift on M1 Max)

M1 Max samples/min over 30 seconds (macOS 13.0):
```
                    PSO_GENERIC  PSO_SPECIALIZED_INTERSECT  PSO_SPECIALIZED_SHADE
barbershop_interior       83.4	            89.5                   93.7
bmw27                   1486.1	          1671.0                 1825.8
classroom                175.2	           196.8                  206.3
fishy_cat                674.2	           704.3                  719.3
junkshop                 205.4	           212.0                  257.7
koro                     310.1	           336.1                  342.8
monster                  376.7	           418.6                  424.1
pabellon                 273.5	           325.4                  339.8
sponza                   830.6	           929.6                 1142.4
victor                    86.7              96.4                   96.3
wdas_cloud               111.8	           112.7                  183.1
```

Next steps:

[ ] ~~Include SHADER_EVAL kernels in the "must cache" list~~ //(limited benefit to specializing one off shade steps)//
[ ] Adapt / merge with dynamic kernel compilation caching patch (D14754)
[x] Separate specialization of intersection (fast building) and shading (slow building) kernels
[x] Rate-limiting and invalidation of kernel compilation requests
[ ] UI for enabling / disabling background compilation

---

With this patch, the Metal backend compiles & caches a second set of kernels which are optimized for scene content, enabled for Apple Silicon.

The optimized kernels result in faster render times, but are slower to compile. They are compiled in the background and swapped in when ready. The optimizations are:

- ~~Aggressive inlining. This is not scene-specific, but hasn't been enabled for the generic kernels because it inflates compile time quite a lot. It results in better register usage, reducing the spill that we're seeing in some kernels. Possible adjustments: 1) take the compile hit for generic kernels since they're only compiled once (and it helps in general), or 2) add a _second_ set of generic_kernels with aggressive inlining enabled.~~ //(enabled by D14923)//
  - ~8% uplift in isolation for 3 benchmarking scenes

- Substitution of KernelData constants. Select members of KernelData struct are replaced with macros that are #defined at the top of source. Only constants pertaining to the rendering algorithm is specialized, rather than constants which might affect artistic look.
  - ~13% uplift in isolation for 3 benchmarking scenes

- Removal of unused SVM nodes in `svm_eval_nodes`. In combination with the other optimizations, this results in a further drop in register usage by eliminating dead code that can't be identified by static analysis.

Code contributed by Jason Fielder, Morteza Mostajabodaveh and Michael Jones

Differential Revision: https://developer.blender.org/D15456

1 files changed, 2 insertions, 0 deletions

diff --git a/intern/cycles/scene/scene.cpp b/intern/cycles/scene/scene.cpp
index 1fcc3331337..5a33f4a6cd1 100644
--- a/intern/cycles/scene/scene.cpp
+++ b/intern/cycles/scene/scene.cpp
@@ -369,6 +369,8 @@ void Scene::device_update(Device *device_, Progress &progress)
     device->const_copy_to("data", &dscene.data, sizeof(dscene.data));
   }
 
+  device->optimize_for_scene(this);
+
   if (print_stats) {
     size_t mem_used = util_guarded_get_mem_used();
     size_t mem_peak = util_guarded_get_mem_peak();