E-Cycles speed-up for Cycles X - PMJ extension
AbandonedPublic
Actions

Authored by Mathieu Menuet (E-Cycles) on Sep 16 2021, 8:31 AM.

Details

Reviewers

William Leeson (leesonw)
Brecht Van Lommel (brecht)
Sergey Sharybin (sergey)

Summary

As the title says, it adds the E-Cycles speed-up for sobol to PMJ .
Currently, it uses the original code while newer Cycles X got a new PMJ code. So this patch reverts the latest changes in the mean time and is provided here as a base line for noise levels and performance to be able to compare performance.
To see the full potential, the GPU occupancy is set to optimal values with fewer updates and higer number of max paths. The overall performance boost offered by this patch is of about 1.45x faster rendering on the dark interior scene compared to current PMJ.

Depends on https://developer.blender.org/D12318

Diff Detail

Repository: rB Blender

Event Timeline

Mathieu Menuet (E-Cycles) requested review of this revision.Sep 16 2021, 8:31 AM

Mathieu Menuet (E-Cycles) created this revision.

Mathieu Menuet (E-Cycles) mentioned this in D12318: Distance Scrambling for for Cycles X - Sobol version.

Thanks for this diff. Unfortunately, this version of the PMJ sampler has correlation issues between dimensions and also sample clumping due to the xor scrambling. Also the use of uint2 to generate the pixel hash to xor with samples causes correlation as this hash is also xor'ed in other places effectively undoing the xor operation. I don't know if it is a good one to base your changes off for profiling due to these issues. Looking at the code it should be easy enough to put your change onto the new PMJ sampler. I left a few tips if you want I can submit a diff later today?

intern/cycles/kernel/kernel_jitter.h
138	Again using the code on the left instead of this you could use this to generate the rv value.
150–151	It should be possible to put this into the CPR code on lines 102-111 to the left as they are very similar. If fact we just need to multiple `dx*(1.0f/NUM_PMJ_SAMPLES)` by (kernel_data.integrator.scrambling_distance/(float)0xFFFFFFFF) should do the trick.
158–171	For the code on the left just multiple `1.0f/NUM_PMJ_DIVISIONS` by `(kernel_data.integrator.scrambling_distance/(float)0xFFFFFFFF)`.
intern/cycles/render/integrator.cpp
191	This effectively gets undone along the edge pixels causing correlation.