First draft of filterArray #144
Conversation
|
@andrewthad, here's a draft. As you can see from my comment on |
Data/Primitive/Internal/Bit.hs
Outdated
| let s = ((n + wordSize - 1) `unsafeShiftR` 3) | ||
| mary <- newByteArray s | ||
| fillByteArray mary 0 s 0 | ||
| return (MBA mary) |
There was a problem hiding this comment.
Alternatively, we could refrain from filling the array here, and just use a writeBitArray function that takes the value to write. That's probably better, actually.
| -- consider going word by word through the bit array and | ||
| -- using countTrailingZeroes. We could even choose | ||
| -- a different strategy for each word depending on its | ||
| -- popCount. |
There was a problem hiding this comment.
Also: I wouldn't be surprised if unordered-containers had some code or ideas we could steal here.
There was a problem hiding this comment.
I agree that CTZ (or equivalently CLZ) is going to do best here. I'm suspect that changing strategies with popCount would not be helpful because I cannot think of a strategy that performs better when most of the elements are preserved. However, when all of them corresponding to a word of bits are preserved (meaning: the word is equal to maxBound :: Word), we could use the functions that copy a slice of the array instead. I guess we could also do this when popCount is really high instead of just when it's 64 (or 32 depending on platform), but we'd be doing several copies instead. I wonder where the breakpoint is for this being effective.
We don't really need this general at the moment, so let's specialize to make things easier for GHC.
|
I still need to look at this. Sorry, been focused on other things the last few days. I'll try to look this over better (and add some benchmarks) later this week. |
| else check (i + 1) count ba | ||
| | otherwise | ||
| = do | ||
| mary <- newArray count (die "filterArray" "invalid") |
There was a problem hiding this comment.
If count equals the size of the original array, we have a much better option available to us. Reuse the original array. This prevents use runArray though.
There was a problem hiding this comment.
This is why we want runArrays and the like.
No description provided.