"clipka" <nomail@nomail> wrote:
> Can anyone come up with a forumla for this?

I guess I have come up with something at last - can someone please verify that
I'm doing a valid test here?

This is the idea:

I will employ a "failure-free reliability testing" scheme to check whether I
have achieved the desired sample coverage, using the following parameters:

- reliability: user-specified sample coverage to achieve during pretrace
- confidence level: hard-coded to, say, 95%
- accepted failures: zero
- number of tests: n = ln (1-confidence) / ln (reliability)

with a "test" being a "radiosity query", that may either be satisfied by a
sample re-use ("pass"), or require a new sample to be gathered ("fail").

To get the required number of tests for the current tile, I will pretrace at a
suitable resolution, maybe multiple times (using jitter) if necessary, until I
have performed at least n tests, of which I will examine only the last n (in
fact, since I'll do a failure-free scheme, I'll just have to count the number
of re-uses since the last gather); if I find that I did not gather any new
sample for the last n tests, I can be confident enough that I have achieved the
desired coverage already.

If I find that I haven't reached the desired coverage yet, I check how many
consecutive re-uses I encountered since the last gather, and use this number N
to estimate the current coverage as N/(N+1). I use this estimate to get a first
estimate how many more rays I need to shoot in the current tile to achieve the
desired coverage, and maybe increase the pretrace resolution accordingly.

Post a reply to this message