Re: [PATCH v1] lib/vsprintf.c: Even faster decimal conversion

[Jeff Epler]

On Wed, Mar 11, 2015 at 12:01:11AM +0100, Rasmus Villemoes wrote:
> Test and verification code on github
> <https://github.com/Villemoes/dec>.  It would be nice if someone could
> verify the code on architectures other than x86 - in particular, I
> believe it _should_ work on big-endian, but I have no way of testing
> that. Benchmark numbers would be nice too, of course.

Results from Odroid U3 (32-bit arm exynos 4412) with gcc 4.8.2 after some
trivial modifications (e.g., no -m32) (ubuntu 14.04ish, I think):

Distribution              Function         nsecs/conv   Conv/1 sec      
uniform([10, 2^64-1])     linux_put_dec          139.94          7050452
uniform([10, 2^64-1])     rv_put_dec             105.55          9278367
                          +/-                   -24.57%          +31.60%
3 + neg_binom(0.05)       linux_put_dec           69.35         14015175
3 + neg_binom(0.05)       rv_put_dec              53.00         18201548
                          +/-                   -23.58%          +29.87%
3 + neg_binom(0.10)       linux_put_dec           48.89         19798285
3 + neg_binom(0.10)       rv_put_dec              38.14         25056279
                          +/-                   -22.00%          +26.56%
3 + neg_binom(0.15)       linux_put_dec           40.69         23576136
3 + neg_binom(0.15)       rv_put_dec              31.60         30009858
                          +/-                   -22.33%          +27.29%
3 + neg_binom(0.20)       linux_put_dec           37.50         25633406
3 + neg_binom(0.20)       rv_put_dec              28.86         32689927
                          +/-                   -23.04%          +27.53%
3 + neg_binom(0.50)       linux_put_dec           28.93         33481454
3 + neg_binom(0.50)       rv_put_dec              20.11         45681278
                          +/-                   -30.47%          +36.44%

Raspberry pi 2 (32-bit ARM BCM2836) with debian jessie armhf (not
raspbian) using a copy of the gcc 4.8.2 binary:
Distribution              Function         nsecs/conv   Conv/1 sec      
uniform([10, 2^64-1])     linux_put_dec          301.66          3234768
uniform([10, 2^64-1])     rv_put_dec             224.33          4313939
                          +/-                   -25.63%          +33.36%
3 + neg_binom(0.05)       linux_put_dec          153.66          6221185
3 + neg_binom(0.05)       rv_put_dec             119.23          7938177
                          +/-                   -22.41%          +27.60%
3 + neg_binom(0.10)       linux_put_dec          117.47          8040480
3 + neg_binom(0.10)       rv_put_dec              91.78         10155002
                          +/-                   -21.87%          +26.30%
3 + neg_binom(0.15)       linux_put_dec          101.14          9254714
3 + neg_binom(0.15)       rv_put_dec              79.15         11632168
                          +/-                   -21.73%          +25.69%
3 + neg_binom(0.20)       linux_put_dec           91.49         10152753
3 + neg_binom(0.20)       rv_put_dec              71.55         12739461
                          +/-                   -21.80%          +25.48%
3 + neg_binom(0.50)       linux_put_dec           77.28         11790819
3 + neg_binom(0.50)       rv_put_dec              57.64         15372506
                          +/-                   -25.42%          +30.38%

and using 4.9.1:
Distribution              Function         nsecs/conv   Conv/1 sec      
uniform([10, 2^64-1])     linux_put_dec          324.76          3009495
uniform([10, 2^64-1])     rv_put_dec             250.85          3870099
                          +/-                   -22.76%          +28.60%
3 + neg_binom(0.05)       linux_put_dec          156.64          6090859
3 + neg_binom(0.05)       rv_put_dec             128.18          7364603
                          +/-                   -18.17%          +20.91%
3 + neg_binom(0.10)       linux_put_dec          120.54          7813490
3 + neg_binom(0.10)       rv_put_dec              99.73          9320708
                          +/-                   -17.26%          +19.29%
3 + neg_binom(0.15)       linux_put_dec          102.22          9116936
3 + neg_binom(0.15)       rv_put_dec              85.22         10778178
                          +/-                   -16.63%          +18.22%
3 + neg_binom(0.20)       linux_put_dec           94.56          9797719
3 + neg_binom(0.20)       rv_put_dec              77.60         11750535
                          +/-                   -17.94%          +19.93%
3 + neg_binom(0.50)       linux_put_dec           80.37         11379434
3 + neg_binom(0.50)       rv_put_dec              62.16         14339337
                          +/-                   -22.67%          +26.01%

Verify passed on the odroid, but I got bored and sent this mail before it finished on the rpi2.

Jeff
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/