Different illuminance results produced by Radiance and Accelerad

[Aymeric DE LA BACHELERIE]

Hi Nathaniel,

Thanks a lot for sharing this great tool.

I noticed big discrepancies between the values I obtain with normal Radiance calculation (within honeybee) and Accelerad.

I made several tests on accelerad parameters to try to make the results be closer but without success.

I would like to share some of these experiments with you and if it is possible have your feedback and a possible explanation.

Case 0: calculation with the original rtrace (Radiance 5.0) using honeybee as an interface:

rtrace -I -h -dp 256 -ds 0.25 -dt 0.25 -dc 0.5 -dr 1 -st 0.5 -lr 6 -lw 0.01 -ab 5 -ad 2048 -as 2048 -ar 300 -aa 0.1 -e error.log 8118513_2.7_7.2_1.0_1.7_1010_RAD.oct

(estimation time: 12.6 min)

Mean Daylight Factor: 3.93%

Case1: Accelerad calculation with parameters as default. (I copy/paste accelerad_rtrace and Accelerad\lib to the Radiance directory to simplify integration in Honeybee)

*** PID   804: rtrace -I -h -dp 256 -ds 0.25 -dt 0.25 -dc 0.5 -dr 1 -st 0.5 -lr 6 -lw 0.01 -ab 5 -ad 2048 -as 2048 -ar 300 -aa 0.1 -e error.log 8118513_2.7_7.2_1.0_1.7_1010Accelerad_RAD.oct

OptiX 3.9.1 found driver 8.0.0 and 1 GPU device:

Device 0: Quadro M4000 with 13 multiprocessors, 1024 threads per block, 772500 kHz, 8589934592 bytes global memory, 128 hardware textures, compute capability 5.2, timeout disabled, Tesla compute cluster driver disabled, cuda device 0.

Geometry build time: 276 milliseconds for 340 objects.

OptiX kernel time: 174 milliseconds (0 seconds).

Adaptive sampling: 5 milliseconds.

Retrieved 513 of 513 potential seeds at level 0.

Using all 513 seeds at level 0 (4096 needed for k-means).

...

OptiX kernel time: 8 milliseconds (0 seconds).

rtrace: ray tracing time: 55314 milliseconds (56 seconds).

Mean Daylight Factor : 3.07%

Case 2: Accelerad calculation with -aa 0.2 -ac 8192 -g 8192 -ag 256

*** PID   808: rtrace -I -h -ds 0.25 -st 0.5 -lr 6 -lw 0.01 -ab 5 -ad 2048 -as 2048 -ar 300 -aa 0.2 -ac 8192 -ag 256 -g 8192 -at 0.02 -e error.log 8118513_2.7_7.2_1.0_1.7_1010Accelerad_RAD.oct

OptiX 3.9.1 found driver 8.0.0 and 1 GPU device:

Device 0: Quadro M4000 with 13 multiprocessors, 1024 threads per block, 772500 kHz, 8589934592 bytes global memory, 128 hardware textures, compute capability 5.2, timeout disabled, Tesla compute cluster driver disabled, cuda device 0.

Geometry build time: 280 milliseconds for 340 objects.

OptiX kernel time: 356 milliseconds (0 seconds).

Adaptive sampling: 5 milliseconds.

Retrieved 513 of 513 potential seeds at level 0.

Using all 513 seeds at level 0 (8192 needed for k-means).

...

OptiX kernel time: 8 milliseconds (0 seconds).

rtrace: ray tracing time: 293806 milliseconds (294 seconds).

Mean Daylight Factor: 3.16%

Conclusion:

The mean daylight factor is still very low in the case of using accelerad even when doubling some input parameters.

The difference is about 20%.

What should I do to reduce the difference between Radiance and Accelerad?

Any help would be very much appreciated!

Best regards,

Aymeric

[Nathaniel Jones]

Hi Aymeric,

I'd have to see the complete error.log file contents to give a more accurate guess as to what is going on. When you don't use -aa 0, Accelerad can give lower illuminance values than Radiance if its irradiance cache doesn't give good scene coverage. Coverage can be improved by using a larger -ac, larger -aa, or smaller ar. Making -ac too large will result in a longer simulation without an improvement in quality.

In the case of your model, you do have complete coverage of the first bounce, since you only have 513 sensors but you have set -ac to 8192. That means the -ag parameter won't have any affect because it only controls infill on the first bounce. However, coverage at the four higher ambient bounces could still be a problem.

I noticed that your -lw value is very low. Usually, -lw should be about 1/ad, so for -ad 2048, you want -lw 0.0005. If you use -aa 0, then -lw should be smaller still. It's possible that this is creating a problem, but hard to tell for sure. Also, once you use a smaller -lw, you can probably also use a smaller -as, like -as 20 or even -as 0, which may make the Accelerad simulation faster.

Unless you had stack overflow errors, it won't help to make -g larger. The -g command sets the stack size hint, so making it larger just makes the simulations slower.

Nathaniel

[Aymeric DE LA BACHELERIE]

Hi Nathaniel,

Many thanks for your quick and detailed answer.

Here you can find attached the error.log files of case 1 and case 2.

If I understand correctly:

- I should lower -lw down to 0.0005 (instead of 0.01) for -ad 2048. When doing so, -as can be fixed to 0 (instead of 2048) without any change on precision but with an improvement on simulation time.

- g doesn't need to be so high in my case

- make aa larger than 0.2 with -ac set to 8192?

- ag doesn't help in my case

I will do some more tests following your orientations.

Is it possible that the geometry and materials are not taken into account in the same manner in radiance and accelerad?

In my case, I have only void plastic and void glass materials. When I reimport the rad file in Rhino I get 320 trimed/untrimmed planar surfaces. Why the "Geometry build" in the error.log indicates 340 objects? 

I'm looking forward to hearing from you.

Aymeric

[Aymeric DE LA BACHELERIE]

Nathaniel, I have noticed that with -aa set to 0, the calculation is very fast but the results are even worst.

I get a mean daylight factor 2.64% which is 33% lower than the Radiance calculation.

-aa 0 is more precise than -aa 0.1 when using Radiance, right?

**************

*** PID 16284: rtrace -I -h -ds 0.25 -st 0.5 -lr 6 -lw 0.0005 -ab 6 -ad 2048 -as 1024 -ar 300 -aa 0 -ac 8192 -g 4096 -at 0.02 -e error.log 8118513_2.7_7.2_1.0_1.7_1010Accelerad_RAD.oct

OptiX 3.9.1 found driver 8.0.0 and 1 GPU device:

Device 0: Quadro M4000 with 13 multiprocessors, 1024 threads per block, 772500 kHz, 8589934592 bytes global memory, 128 hardware textures, compute capability 5.2, timeout disabled, Tesla compute cluster driver disabled, cuda device 0.

Geometry build time: 197 milliseconds for 340 objects.

OptiX kernel time: 1856 milliseconds (2 seconds).

rtrace: ray tracing time: 2693 milliseconds (3 seconds).

[Nathaniel Jones]

Hi Aymeric,

Accelerad takes geometry and materials into account in the same manner as Radiance. Both plastic and glass are implemented and should work correctly. The number of objects reported is the number of entries in the octree file, which includes geometry as well as materials and functions.

Based on the error logs, there do not appear to be any errors or stack overflows, so poor coverage in the ambient cache seems like the most likely reason for the discrepancy between Radiance and Accelerad. Because the number of seeds at levels two and three is so much higher than -ac, it's quite possible that the coverage at those levels is poor. There are a few possible solutions for this. In addition to increasing -ac and -aa or decreasing -ar, you could also try decreasing -ad. It's hard to imagine that you need a value has high as 2048 for a simple daylight factor calculation. You could also reduce the area that needs coverage by removing geometry that's not relevant, such as the sides of the building that don't reflect light into the courtyard.

When you use -aa 0, you should compare Accelerad to a Radiance simulation that also uses -aa 0. That's because -aa 0 triggers a fundamentally different global illumination algorithm, so you wouldn't expect the same results. When you use -aa 0, you also need a much lower -lw value in both Accelerad and Radiance, even as low as -lw 1e-7 or -lw 1e-10. Otherwise, the simulation will reach the ray weight limit after just one ambient bounce, which is why the daylight factor is so low and the simulation runs so fast.

Nathaniel

[Aymeric DE LA BACHELERIE]

Hi Nathaniel,

Thanks a lot for your help.

I tried with -aa 0 -lw 0 and I needed to set -g to 8192 to avoid stack overflow issue but it finally gave me results comparable to the Radiance simulation in a short time: 30sec instead of 12 minutes.

I've decided to simulate with Max settings of radiance rendering options: http://radsite.lbl.gov/radiance/refer/Notes/rpict_options.html

Both (accelerad and radiance) results are very closed but accelerad is taking a really short amount of time.

Congratulations for this great idea and tool!

Regards,

Aymeric