Different illuminance results produced by Radiance and Accelerad in DIVA4.0

[colin]

Hi,

I am currently trying to include Accelerad into the DIVA workflow of daylight simulation and am now doing some tests.

Basically I am running DIVA4.0 in normal Radiance as well as in Accelerad, comparing the speed and illuminance results for different accuracy(Low, Medium, High).

It is really fascinating that Accelerad has a perfect speed in no matter what level of accuracy. 

It finishes the calculation within 2 minutes, while for normal simulation of DIVA it takes around 40 mins for Medium accuracy and couples of hours for High accuracy.

While comparing the results produced by Accelerad and normal Radiance, there is a certain quantities of differences. Here are two of my observations.

1. The closer an evaluation point to an object is, the larger the different results becomes.(up to 1000lux).

2. As I compare all results produced by Accelerad, the amount of illuminance results are Low>Medium>High, for which I suppose it should be opposite, as the higher level of accuracy, the more times the light bounce are calculated, which can be verified in the normal Radiance calculation.

I am simulating for an exterior environment, using DIVA4.0 in grasshopper. The Accelerad version is 0,5beta.

Does anyone know why this happens?Any suggestions will be preferred.

Many thanks,

colin

[Nathaniel Jones]

Hi Colin,

The low, medium, and high presets in DIVA aren't very relevant to Accelerad; they certainly don't correspond to the accuracy or simulation time you should expect from Accelerad. For a full explanation of why, see my presentation from the Radiance workshop last year. Here's a brief summary.

Normally in Radiance, you control accuracy with the -aa parameter, which places an upper limit on the error introduced by the ambient calculation. Smaller -aa values result in more frequent ambient calculation and therefore longer simulation times.

This is not the case in Accelerad. Smaller -aa values still reduce error around individual calculation locations, but the frequency of calculation is determined in advance by the -ac parameter (in order to allow parallel calculation). Larger -ac values will result in more frequent ambient calculation and therefore longer Accelerad simulations. For a constant -ac value, smaller -aa values can make Accelerad calculations less accurate due to poor ambient coverage; this is the opposite of what you would expect in Radiance. See the documentation for details on -ac and other Accelerad parameters.

Your two observations probably indicate that either -aa or -ac is set too low in Accelerad, leading to lower-than-expected illuminance values. If Accelerad is producing too-high results anywhere, check your command-line output for warnings. The most common warning is stack overflow, which can be solved by increasing the -g parameter. In general, you should check visualizations to make sure that your settings produce reasonable images before using the same settings to calculate illuminance. I also do not use the same settings for Accelerad that I would use for Radiance. It is more valid to either compare quality from equal-time simulations or time from equal-quality results.

Best,

Nathaniel

[colin]

Hi Nathaniel,

Thanks for your reply, and it solves most of my confusion.

Here I attach the parameters setup for each accuracy level in DIVA.

Low              -ab 2    -ad  512     -ar 256    -as 128    -aa 0.15    -lw 0.004
Medium        -ab 4    -ad 1024    -ar 256    -as 256    -aa 0.1      -lw 0.004
High              -ab 7    -ad 4096   -ar 512    -as 1024   -aa 0.1      -lw 0.001

And if that is the case, if I want to increase the accuracy of the result in Accelerad, I should try to have a higher value of -ac and lower value of -aa, am I right?

And what else parameters I should modify in Accelerad in order to achieve higher accuracy?
I will try to do some tests based on your suggestions and post the feedback.

Many thanks!

best.

colin
 

[Nathaniel Jones]

Hi Colin,

In general, you can improve accuracy will smaller -aa or with larger -ac. I suggest you choose -aa first and then choose an appropriate -ac value using accelerad_rpict renderings to verify that the results are reasonable. I use larger values of -aa for Accelerad simulations than I do for classic Radiance simulations. The choice of other variables to experiment with will depend on the specific model you are working with. For instance, high -ab and -lr will be more useful in indoor scenes, high -ar will be helpful for larger scenes, and high -ad or -as will be helpful in scenes with lots of ambient variation. On the Accelerad side, varying -ag, -an, -at, and -ax may help improve speed and placement of ambient values.

Nathaniel

[colin]

Hi Nathaniel,

Great!

Thanks for the quick reply and all the information.

Will give it a shot soon.

Thanks!

best,

colin

[dan weissman]

Colin, Nathaniel -

this is a very helpful line of discussion as i'm operating in the same boat. I wonder, Nathaniel, if you'd be up for providing a similar "Low/Med/High" set of baseline settings to use in replace of the standard settings found in DIVA? Obviously specific settings must be tweaked, but many users (myself included) are not quite as familiar with ALL the radiance settings as you are. 

Also, can accelerad run the daylight coefficients (rtrace_dc.exe) as well?

many thanks

dan

[Nathaniel Jones]

Hi Dan,

As I said before, the "Low/Med/High" presets in DIVA aren't very relevant to Accelerad. DIVA's presets are built around the idea that sometimes you want a faster simulation and are willing to sacrifice accuracy for speed. However, Accelerad is built with the philosophy that accurate simulations should be fast, so purposefully offering low-quality settings runs counter to this. Of course, it's still possible to get inaccurate results with bad parameter combinations. It might be totally reasonable to suggest presets based on other factors -- large vs small scene size, indoor vs outdoor, high contrast vs low contrast, etc. -- at some point in the future. Right now, my impression is that a lot of these settings come from rules of thumb or trial and error.

The current version of Accelerad doesn't support daylight coefficients.

Nathaniel

[jwil...@gmail.com]

Hi Nathaniel,

do you have an estimation when the daylight coefficients will be added to Accelerad, so that it can be used with DAYSIM - at least a rough one?

Thank you.

Jiri

[kim.eyc...@gmail.com]

I'm also making use of Daysim and would like to use Accelerad with rtrace_dc.  Any idea when it will become available?

[Nathaniel Jones]

For daylight coefficient and sDA calculation, I recommend using the three-phase method instead of DAYSIM. The three-phase method using rcontrib is already much faster than rtrace-dc and has been shown to provide similar accuracy. Accelerad 0.6 includes a GPU-accelerated version of rcontrib that can be used for three-phase and five-phase calculations.

Nathaniel