SoylentNews is people
SoylentNews
from the abusing-bits-for-fun-and...profit? dept.
Submitted via IRC for Bytram
Abusing A CPU's Adders To Optimize Bit Counting
If you like nitpicking around C code and generated assembly language — and we’ll admit that we do — you shouldn’t miss [Scaramanga’s] analysis of what’s known as Kernighan’s trick. If you haven’t heard of the trick, it’s a pretty efficient way of counting bits.
Like the Wheatstone bridge and a lot of other things, the Kernighan trick is misnamed. Brian Kernighan made it famous, but it was actually first published in 1960 and again in 1964 before he wrote about it in 1988. The naive way to count bits would be to scan through each bit position noting how many one bits you encounter, but the problem is, that takes a loop for each bit. A 64-bit word, then, takes 64 loops no matter what it contains. You can do slightly better by removing each bit you find and stopping when the word goes to zero, but that still could take 64 cycles if the last bit you test is set.
Using the trick, you make the observation that X & (X-1) will always clear the least significant bit of a word. Try a few examples:
X X-1 X&(X-1) 0001 0000 0000 0010 0001 0000 0011 0010 0010 1010 1001 1000 1100 1011 1000 1111 1110 1110 You can probably see where this is going. By computing X&(X-1) you clear a bit on each loop iteration and you only have to go through the number of bits that are actually set.
[...] If you like this sort of thing, be sure to check out [Sean Anderson’s] extensive list of bit hacks. It shows several different ways to count bits and do other common and uncommon tasks with different tradeoffs. For example, you could dedicate a 256-entry lookup table and do the whole thing with one loop per byte with great speed but bad memory utilization. Always a trade-off.
(Score: 0) by Anonymous Coward on Thursday June 20 2019, @11:10AM (1 child)
Not that good, as it contains a very slow division operation at the end, which not only has a high latency but ties up an arithmetic execution unit for the whole time. But this algorithm doesn't use any memory and is more consistent in speed than the Kernighan algorithm.
The lookup table has a long wait if it misses the cache, but at least the CPU can do other things while it waits. It's probably the best if you need to count a lot of bits all at once, and you can expect it to be in L1 most of the time. The thing is that if performance has any chance of mattering here, you probably do need to count a very large number of bits (or a small number very often, same thing). You could also count four bits at a time with only a 16-byte table, which might be better if you are very short on space.
Overall, today's CPUs are very fast and it's unlikely that the choice of algorithm here would ever make a difference. The obvious algorithm and lookup tables are the only ones that are immediately understandable, and the lookup table is faster in situations where it matters, so I'd probably go with one of those almost always. And x86/ARM have instructions to do this specific thing anyway.
But maybe you are not running on a powerful x86 or ARM, but instead on an attiny where you have a hard real time requirement and you can't spend 256, or even 16, bytes on a lookup table. You don't have a division instruction either. Maybe Kernighan's algorithm is best here.
(Score: 1, Informative) by Anonymous Coward on Thursday June 20 2019, @01:22PM
No division is needed:
https://godbolt.org/z/4PRXJw [godbolt.org]