Glibc 2.27 Is Being Released Soon With Numerous Performance Optimizations

[heliosh]

https://www.phoronix.com/scan.php?pa...lease-Imminent

I wonder if some of the lesser optimized projects will benefit from this. For some operations the performance gain is said to be in the range of 50%.

[science_man_88]

Originally Posted by heliosh

https://www.phoronix.com/scan.php?pa...lease-Imminent

I wonder if some of the lesser optimized projects will benefit from this. For some operations the performance gain is said to be in the range of 50%.

Probably depends, on if the 30% performance hit from patches for spectre and meltdown just barely get nullified 100/70= 1.42857142857... so 1.5/1.42857142857...... ~5% gained.

[heliosh]

30% maybe on I/O workload. On computing workload it's more like in the lower one digit percent range, if there's a difference at all.
https://www.phoronix.com/scan.php?pa...etpoline&num=1

[science_man_88]

Originally Posted by heliosh

30% maybe on I/O workload. On computing workload it's more like in the lower one digit percent range, if there's a difference at all.
https://www.phoronix.com/scan.php?pa...etpoline&num=1

A 5% slowdown multiplied by a 50% speedup means a low 42% speedup.

[GP2]

For projects that make binaries available, would there be any benefit to using the Intel Math Kernel Library?

[rogue]

If you follow the link, this is only a single function and one that I've never used and I've written a lot of code. The actual gains for everything else is not known.

[heliosh]

Some more informations on improved functions (links in the article in the original posting)

There is now optimized asin, atan2, exp, expf, log, pow, atan, sin, and tan functions in glibc Git for benefiting from x86-64 for fused multiply-add,

Yesterday there was sinf() tuning including an FMA-tuned version. Longtime Binutils/GCC developer H.J. Lu of Intel found that his fused multiply–add version of the sinf call is 54% faster on average, the minimum is a 25% improvement, and the maximum time was also a 54% improvement. Tests were done on Skylake hardware.

Just the latest done by Intel's H.J. Lu this week is a FMA'ed cosf() yielding improvements up to ~45% faster. That test was done with a Skylake processor. And by removing older cosine code was also another win.

[heliosh]

glibc 2.27 was released yesterday, but will take a while until it makes its way to the distributions.

• Optimized x86-64 asin, atan2, exp, expf, log, pow, atan, sin, cosf, sinf, sincosf and tan with FMA, contributed by Arjan van de Ven and H.J. Lu from Intel.
• Optimized x86-64 trunc and truncf for processors with SSE4.1.
• Optimized generic expf, exp2f, logf, log2f, powf, sinf, cosf and sincosf.

https://sourceware.org/ml/libc-annou.../msg00000.html