On Monday, 22 August 2016 at 17:13:10 UTC, jmh530 wrote:
On Monday, 22 August 2016 at 15:34:47 UTC, Seb wrote:
Hey all,
I am proud to publish a report of my GSoC work as two
extensive blog posts, which explain non-uniform random
sampling and the mir.random.flex package (part of Mir >
0.16-beta2):
http://blog.mir.dlang.io/random/2016/08/19/intro-to-random-sampling.html
http://blog.mir.dlang.io/random/2016/08/22/transformed-density-rejection-sampling.html
Thanks for the well-done blog posts, especially the first one.
I am glad to hear this!
Does your implementation make any use of CTFE?
If you refer to whether the intervals can be calculated at CT,
unfortunately it can't be used due to four main reasons:
- FP-math at CT (it's already hard to deal with at RT, see e.g.
my recent complaint [1]) - the problem is that the Flex algorithm
is very sensitive to numerical errors and thus an erroneous
change at the lowest end (e.g 10^-15) can lead to totally
different numbers with a seeded random engine
- std.container due to pointers (I doubt this can/will be fixed
in the near future)
- std.math due to inline assembly and other tricks (this can be
fixed and I will submit a couple of PRs soon)
- speed of the CTFE engine (see e.g. [2] for std.regex)
That being said CTFE is of course used to compute mixins,
constants and specialize functions. Moreover thanks to all
speed-ups described in the second blog, constructing the
intervals takes about 0.1ms, so for the majority of the users it
shouldn't even be noticeable and for the tiny minority it does,
they can still manually inline the intervals.
[1]
http://forum.dlang.org/post/[email protected]
[2]
http://forum.dlang.org/post/[email protected]
