Feature #15166
2.5 times faster implementation than current gcd implmentation
Description
This is to be more explicit (and accurate) than https://bugs.ruby-lang.org/issues/15161
This is my modified gcd benchmarks code, originally presented by Daniel Lemire (see 15161).
https://gist.github.com/jzakiya/44eae4feeda8f6b048e19ff41a0c6566
Ruby's current implementation of Stein's gcd algorithm is only slightly faster than the
code posted on the wikepedia page, and over 2.5 times slower than the fastest implementation
in the benchmarks.
[jzakiya@localhost ~]$ ./gcdbenchmarks gcd between numbers in [1 and 2000] gcdwikipedia7fast32 : time = 99 gcdwikipedia4fast : time = 121 gcdFranke : time = 126 gcdwikipedia3fast : time = 134 gcdwikipedia2fastswap : time = 136 gcdwikipedia5fast : time = 139 gcdwikipedia7fast : time = 138 gcdwikipedia2fast : time = 136 gcdwikipedia6fastxchg : time = 144 gcdwikipedia2fastxchg : time = 156 gcd_iterative_mod : time = 210 gcd_recursive : time = 215 basicgcd : time = 211 rubygcd : time = 267 gcdwikipedia2 : time = 321 gcd between numbers in [1000000001 and 1000002000] gcdwikipedia7fast32 : time = 100 gcdwikipedia4fast : time = 121 gcdFranke : time = 126 gcdwikipedia3fast : time = 134 gcdwikipedia2fastswap : time = 136 gcdwikipedia5fast : time = 138 gcdwikipedia7fast : time = 138 gcdwikipedia2fast : time = 136 gcdwikipedia6fastxchg : time = 144 gcdwikipedia2fastxchg : time = 156 gcd_iterative_mod : time = 210 gcd_recursive : time = 215 basicgcd : time = 211 rubygcd : time = 269 gcdwikipedia2 : time = 323
This is Ruby's code per: https://github.com/ruby/ruby/blob/3abbaab1a7a97d18f481164c7dc48749b86d7f39/rational.c#L285-L307
which is basically the wikepedia implementation.
inline static long
i_gcd(long x, long y)
{
unsigned long u, v, t;
int shift;
if (x < 0)
x = -x;
if (y < 0)
y = -y;
if (x == 0)
return y;
if (y == 0)
return x;
u = (unsigned long)x;
v = (unsigned long)y;
for (shift = 0; ((u | v) & 1) == 0; ++shift) {
u >>= 1;
v >>= 1;
}
while ((u & 1) == 0)
u >>= 1;
do {
while ((v & 1) == 0)
v >>= 1;
if (u > v) {
t = v;
v = u;
u = t;
}
v = v - u;
} while (v != 0);
return (long)(u << shift);
}
This is the fastest implementation from the benchmarks. (I originally, wrongly, cited
the implementation in the article, which is 4|5th fastest in benchmarks, but
still almost 2x faster than the Ruby implementation.)
// based on wikipedia's article,
// fixed by D. Lemire, K. Willets
unsigned int gcdwikipedia7fast32(unsigned int u, unsigned int v)
{
int shift, uz, vz;
if ( u == 0) return v;
if ( v == 0) return u;
uz = __builtin_ctz(u);
vz = __builtin_ctz(v);
shift = uz > vz ? vz : uz;
u >>= uz;
do {
v >>= vz;
int diff = v;
diff -= u;
if ( diff == 0 ) break;
vz = __builtin_ctz(diff);
if ( v < u ) u = v;
v = abs(diff);
} while( 1 );
return u << shift;
}
The key to speeding up all the algorithms is using the __builtin_ctz(x) directive
to determine the number of trailing binary '0's.
Files
History
Updated by jzakiya (Jabari Zakiya) about 1 year ago
Updated by - Backport deleted (
2.3: UNKNOWN, 2.4: UNKNOWN, 2.5: UNKNOWN) - Tracker changed from Bug to Feature
Updated by mame (Yusuke Endoh) about 1 year ago
Updated by - Assignee set to watson1978 (Shizuo Fujita)
- Status changed from Open to Assigned
Thanks. Assigned to @watson1978.
It would be very helpful if you could provide us a patch and perform the benchmark with Ruby implementation, not a toy benchmark program. Note that __builtin_ctzl is not available on some compilers. You need to check if it is available or not.
Updated by jzakiya (Jabari Zakiya) about 1 year ago
Hi
I just submitted this issue feature request:
https://bugs.ruby-lang.org/issues/15172
to deal with the issue of using (or not) the __builtin_ctz compiler directive.
I implemented code that mimicked it that also greatly increases the ruby gcd performance.
I included the new code and benchmarks to the gist I previously linked.
https://gist.github.com/jzakiya/44eae4feeda8f6b048e19ff41a0c6566
[jzakiya@jabari-pc ~]$ ./gcd2 gcd between numbers in [1 and 2000] gcdwikipedia7fast32 : time = 73 gcdwikipedia4fast : time = 113 gcdFranke : time = 133 gcdwikipedia3fast : time = 139 gcdwikipedia2fastswap : time = 162 gcdwikipedia5fast : time = 140 gcdwikipedia7fast : time = 129 gcdwikipedia2fast : time = 161 gcdwikipedia6fastxchg : time = 145 gcdwikipedia2fastxchg : time = 168 gcd_iterative_mod : time = 230 gcd_recursive : time = 232 basicgcd : time = 234 rubygcd : time = 305 gcdwikipedia2 : time = 312 gcdwikipedia7fast32_a : time = 129 gcdwikipedia4fast_a : time = 149 rubygcd_a : time = 193 rubygcd_b : time = 169 gcd between numbers in [1000000001 and 1000002000] gcdwikipedia7fast32 : time = 76 gcdwikipedia4fast : time = 106 gcdFranke : time = 121 gcdwikipedia3fast : time = 127 gcdwikipedia2fastswap : time = 153 gcdwikipedia5fast : time = 126 gcdwikipedia7fast : time = 118 gcdwikipedia2fast : time = 148 gcdwikipedia6fastxchg : time = 134 gcdwikipedia2fastxchg : time = 154 gcd_iterative_mod : time = 215 gcd_recursive : time = 214 basicgcd : time = 220 rubygcd : time = 287 gcdwikipedia2 : time = 289 gcdwikipedia7fast32_a : time = 116 gcdwikipedia4fast_a : time = 142 rubygcd_a : time = 180 rubygcd_b : time = 155
Note using the __builtin_ctz mimicking code, instead of the directive itself,
still makes the gcdwikipedia7fast32_a the third fastest version, and obviously
the preferred implementation if not using __builtin_ctz.
I present this in asking you how you want me to proceed, because I don't really know
C code and how to do PRs to Ruby. If you can lay out a detailed process for me to do
that maybe I can assess what is in my capacity to do.
At minimum, the code for rubygcd_a could|can be incorporated into the codebase
without dealing right now with the __builtin_ctz directive issue.
Updated by mame (Yusuke Endoh) about 1 year ago
No, no. You can just use __builtin_ctzl when it is available. All you need is check if it is available or not, and keep the original code for the case where __builtin_ctzl is unavailable. Gcc and clang provide it, so it is actually available in almost all cases. Even if __builtin_ctzl is unavailable, it should still build and work, but the performance does not matter, I think.
Ruby is already using __builtin_ctz and __builtin_ctzll. See configure.ac and internal.h.
Updated by ahorek (Pavel Rosický) 11 months ago
Updated by your micro-benchmarks aren't always fair, because some algorithms don't handle all edge cases, different data types etc.
for jruby I choose a different algorithm that is slightly slower than the fastest gcdwikipedia7fast32 (~15%) but in my opinion more readable.
https://github.com/jruby/jruby/blob/1d0c3d643a6841f388e646678ee243bff571450c/core/src/main/java/org/jruby/util/Numeric.java#L512
here's the PR (gcdwikipedia7fast32 + minor changes) https://github.com/ruby/ruby/pull/2060
and some ruby numbers (benchmark https://github.com/ruby/ruby/pull/1596)
all variants tested on AMD FX 8300 8C and gcc version 8.1.0 (Ubuntu 8.1.0-5ubuntu1~14.04)
ruby 2.7.0dev (2018-12-28 trunk 66617) [x86_64-linux]
Time#subsec 2.969M (± 9.6%) i/s - 14.733M in 5.010950s
Time#- 5.716M (±11.4%) i/s - 28.103M in 5.000934s
Time#round 400.712k (±11.9%) i/s - 1.992M in 5.046665s
Time#to_f 6.422M (±10.5%) i/s - 31.613M in 4.999488s
Time#to_r 2.251M (±10.4%) i/s - 11.124M in 5.002516s
Rational#+ 5.377M (±10.1%) i/s - 26.577M in 5.001636s
Rational#- 5.542M (± 9.5%) i/s - 27.419M in 5.001546s
Rational#* 6.341M (± 9.5%) i/s - 31.390M in 5.002212s
gcd 6.922M (± 9.0%) i/s - 34.285M in 5.001389s
trunk + new gcd
Time#subsec 3.348M (± 8.9%) i/s - 16.592M in 4.999620s / 1.13
Time#- 5.840M (±11.6%) i/s - 28.728M in 5.000946s / 1.02
Time#round 468.770k (±12.5%) i/s - 2.319M in 5.028050s / 1.17
Time#to_f 6.713M (± 9.8%) i/s - 33.214M in 4.999639s / 1.05
Time#to_r 3.191M (± 7.9%) i/s - 15.884M in 5.010305s / 1.42
Rational#+ 5.893M (±10.6%) i/s - 29.082M in 4.999884s / 1.10
Rational#- 6.183M (±11.2%) i/s - 30.443M in 4.999746s / 1.12
Rational#* 7.069M (±10.5%) i/s - 34.922M in 5.001804s / 1.11
gcd 9.742M (±10.4%) i/s - 48.159M in 5.007085s / 1.40
trunk + new gcd without __builtin_ctz support
Time#subsec 2.699M (± 8.9%) i/s - 13.385M in 5.002527s / 0.89
Time#- 5.734M (±10.6%) i/s - 28.224M in 5.002541s / 1.00
Time#round 392.314k (±13.8%) i/s - 1.926M in 5.012040s / 0.98
Time#to_f 6.725M (±10.5%) i/s - 33.163M in 4.999346s / 1.04
Time#to_r 2.366M (± 9.1%) i/s - 11.705M in 5.004491s / 1.05
Rational#+ 5.429M (±10.1%) i/s - 26.851M in 5.006358s / 1.01
Rational#- 5.544M (± 9.8%) i/s - 27.430M in 5.002418s / 0.98
Rational#* 6.225M (±10.7%) i/s - 30.833M in 5.018386s / 0.98
gcd 7.001M (± 7.1%) i/s - 34.855M in 5.006972s / 1.01
alternative implementations
jruby 9.2.6.0-SNAPSHOT (2.5.3) 2018-12-27 e51a3e4 Java HotSpot(TM) 64-Bit Server VM 11.0.1+13-LTS on 11.0.1+13-LTS +jit [linux-x86_64]
Time#subsec 5.018M (± 6.3%) i/s - 24.866M in 4.979170s
Time#- 7.868M (± 5.6%) i/s - 39.066M in 4.985576s
Time#round 3.461M (± 8.1%) i/s - 17.138M in 4.998527s
Time#to_f 8.198M (± 5.2%) i/s - 40.775M in 4.990224s
Time#to_r 4.789M (± 6.9%) i/s - 23.777M in 4.992261s
Rational#+ 5.217M (± 6.3%) i/s - 25.944M in 4.995694s
Rational#- 5.701M (± 7.4%) i/s - 28.329M in 4.998743s
Rational#* 6.290M (± 6.7%) i/s - 31.283M in 4.997365s
gcd 7.376M (± 7.2%) i/s - 36.625M in 4.995073s
truffleruby 1.0.0-rc10, like ruby 2.4.4, GraalVM CE Native [x86_64-linux]
Time#subsec 3.541M (±67.8%) i/s - 13.706M in 4.986699s
Time#- 8.279M (± 9.4%) i/s - 38.671M in 4.984896s
Time#round 311.696k (±43.3%) i/s - 502.226k in 4.991276s
Time#to_f 16.719M (± 9.2%) i/s - 75.067M in 4.981367s
Time#to_r 1.386M (±21.2%) i/s - 5.045M in 4.993055s
Rational#+ 7.332M (±14.7%) i/s - 28.100M in 4.982371s
Rational#- 7.354M (±24.3%) i/s - 22.682M in 4.992218s
Rational#* 7.340M (±19.3%) i/s - 28.534M in 5.003816s
gcd 68.576M (± 4.7%) i/s - 326.812M in 4.908116s
as you can see Time#to_r and Integer#gcd is about 40% faster which is the best case scenario even when in your micro-benchmark it was 300% faster.
using the new algorithm without __builtin_ctz introduced some perf regressions, but they're within margin of error
I don't think this change will have some impact on real application's performance, all of these cases are just micro-benchmarks...
Updated by ahorek (Pavel Rosický) 7 months ago
I'm curious if anyone have an explanation why the pure C version is 300% faster, but the ruby version of the same algorithm is 40% faster? Is it a VM overhead or something else?