First the BiDir result. I take is as reference.
- Default body 2 BiDir 1200 samples 1h33m
Due to the bright floor the caustics are faint. One has to look good at them to notice the fine lines.
On the next picture you can see the effect of increasing the amount of photons. For this test series indirect and caustic have the same amount.
- Influence of Photon amount on Caustics
Naturally the higher the number the more time the render takes during session initialization.
With 20 million photons one gets an image that has a steady character although with every million photons more the difference between to renders keep even from 49 to 50 millions comparable to differences between less photons.
The next animated gifs show effect of other changed parameters.
- Caustics-Look-up-Max-count
As you can see this parameter changes the 'noise pattern' a bit but only randomly.
(With a count above 800 the buffer size exceeded available GPU RAM of 3'422'552'064 bytes what lead to a warning before the render started (V2.2 beta 1, build 0a863a0)).
- Caustics-Look-up-radius
The dots of the 'noise pattern' reduces notable by decreasing this radius.
- Indirect-Cache-Look-up-radius
The animation shows that there is happening something but it is not clear to me what exactly.
Next pictures show openCL without PGIC.
- openCL_Metropolis_0011
- openCL_Sobol_0001
Metropolis pattern on the body has obviously a lower resolution than Sobol pattern.
Last image has the parameters: 10 million photons, indirect cache radius 0.1, caustic cache radius 0.00001, max count 64.
2500 samples, 2m19s init, 8m24s render
- Glass body, openCL PGIC 2500 samples
Conclusion: In comparison with BiDir PGIC fails to render the fine caustics of this body.
EDIT: I found the reason for the missing caustics. The calculation of the dispersion is wrong for this object, it is too high.