RFC: Adds random number generation for BigInts and BigFloats #4845
Conversation
This means that ranges between numbers bigger than `typemax(Int)` work now, but range lengths are still limited to up to `typemax(Int)`.
|
Very nice. Don't let this question slow you down, but is there any way we could get GMP to sync up with the Julia RNG state? |
|
This is a great pull request. I'm not sure how one would sync up the RNGs since it seems like BigRNG probably needs a lot more state than dSFMT. |
|
That's probably true: hadn't thought about that at all. |
|
Well, technically we could make GMP use Julia's dSFMT instead of its own internal Mersenne Twister, but it'll require large amounts of glue code to make it work. |
|
Probably not worth it. |
|
@andrioni thanks for picking this up again. I had completely forgotten about this. |
|
Do we want to be able to reproduce the same sequence of all random numbers (big or small) from a single |
|
It would be nice if we could, but may not be easy. What happens when we mix random numbers from BigRNG and existing RNGs? |
|
Currently, both RNGs are completely independent. You have to reseed them separately if you want to, and they don't affect each other's state. |
|
I would tend to prefer that |
|
That should be easy enough to do. |
|
I would prefer seeding both with |
OS X doesn't have `sha1sum` available, so the fallback used to get the initial seed whenever /dev/urandom isn't available doesn't work.
|
I've just changed the (I have no idea whatsoever why travis failed, especially since it worked here with the same commit) |
|
BTW, this should also get a mention in NEWS.md as we have started it for 0.3. Is this good to merge otherwise? |
|
I think everything I need/wanted to do is done, so barring other suggestions/objections, it should be good to merge. |
It requires JuliaLang/julia#4845 to be merged, since it uses BigRNG.
Now a macro is used to check if BigRNG is available, this means the same codebase is compatible with both julia post- JuliaLang/julia#4845 and julia 0.2.
|
The clang build prints a ton of garbage and dies. Perhaps worth investigating before merging. |
|
Bump. @andrioni Can you please rebase and check if the tests are passing? This would be good to have. |
|
I was able to merge and get this running with all the tests passing. |
| convert(BigInt, first(r) + randu(ulen)) | ||
| end | ||
|
|
||
| function rand!(r::Range1{BigInt}, A::Array{BigInt}) |
There was a problem hiding this comment.
This is redundant with definitions in random.jl
|
Since this patch was merged, why am I getting the error: ? |
|
It was rolled back after some time, since it was bugging out on some configurations (including the Travis one, IIRC). |
|
@andrioni Do you recall what exacly went wrong? |
|
@mschauer not off the top of my head, but I could take a look. |
This pull request extends the work done by @jiahao in #3077, adding a default RNG for BigInts and BigFloats (since they share the gmp_randstate interface) and BigFloat support (including for
randn).I'll update it later with the documentation as soon as we settle on the interface.