Cycles: CPU batched execution for performance #92572

New Issue

Brecht Van Lommel · 2021-10-28T15:04:16+02:00

Brecht Van Lommel commented

2021-10-28 15:04:16 +02:00

With the new Cycles architecture, it will be easier to take advantage of SIMD instructions. This can be done in 3 steps.

Use Embree batched ray tracing
Use OSL batched shading
Use batched execution for rest of the kernel

This requires some changes for CPU threads to render multiple paths in a batch. This is relatively easy with the new architecture, which already works like a state machine.

Some technical points:

For full batched execution of the kernel the integrator state likely needs a SoA memory layout for good performance. However for Embree and OSL this is likely not needed yet.
Full batched execution of the kernel could be implemented in two ways. Either we could use a dedicated compiler like ISPC for it, or we could use a template library like Enoki.
The optimal size of batches is unclear. It can range from 4/8/16 to potentially millions, depending if it pays of to extract as much coherence as possible with sorting etc, or if a small batch using incidental coherence is better due to less overhead.

With the new Cycles architecture, it will be easier to take advantage of SIMD instructions. This can be done in 3 steps. - [ ] Use Embree batched ray tracing - [ ] Use OSL batched shading - [ ] Use batched execution for rest of the kernel This requires some changes for CPU threads to render multiple paths in a batch. This is relatively easy with the new architecture, which already works like a state machine. Some technical points: * For full batched execution of the kernel the integrator state likely needs a SoA memory layout for good performance. However for Embree and OSL this is likely not needed yet. * Full batched execution of the kernel could be implemented in two ways. Either we could use a dedicated compiler like ISPC for it, or we could use a template library like Enoki. * The optimal size of batches is unclear. It can range from 4/8/16 to potentially millions, depending if it pays of to extract as much coherence as possible with sorting etc, or if a small batch using incidental coherence is better due to less overhead.

Brecht Van Lommel commented

2021-10-28 15:04:16 +02:00

Changed status from 'Needs Triage' to: 'Confirmed'