AMD Developers Looking At GNU C Library Platform Optimizations For Zen

[dispat0r]

I did some testing memcpy is slower with avx2 than with sse on threadripper:
https://forum.level1techs.com/t/defa...r-3-bad/154393

I need to do some more testing could be that only this particular code is affected.

[mlau]

AMD seems to have been on good terms with SuSE in the past, why not write them a cheque and leave this to the experts?

[mlau]

I did some testing memcpy is slower with avx2 than with sse on threadripper:
https://forum.level1techs.com/t/defa...r-3-bad/154393

I need to do some more testing could be that only this particular code is affected.

The function mentioned in the first comment (__memmove_avx_unaligned_erms) is an intel-specific optimization, no wonder naive sse is faster on amd.

[hajj_3]

speaking of AMD, it seems that some of their source code has been leaked by a hacker and is threatening to release more: https://torrentfreak.com/amd-uses-dm...e-leak-200325/

[dispat0r]

The function mentioned in the first comment (__memmove_avx_unaligned_erms) is an intel-specific optimization, no wonder naive sse is faster on amd.

Most of the avx code for Intel is faster than the sse alternative on AMD memmove seems to be the exception.
So we need an AMD specific version?

[mlau]

Most of the avx code for Intel is faster than the sse alternative on AMD memmove seems to be the exception.
So we need an AMD specific version?

Well, this is where AMD needs to step up: find out which method works best on what core and which sizes, and wire that up. I read somewhere that on AMD memmove in reverse direction works best, unlike intel.

[carewolf]

I did some testing memcpy is slower with avx2 than with sse on threadripper:
https://forum.level1techs.com/t/defa...r-3-bad/154393

I need to do some more testing could be that only this particular code is affected.

That is a bit off-topic.. memcpy and other internal routines are not the optimizations were are talking about. glibc already optimizes those much more fine grained and with different paths for all kinds of things. This news is about how glibc picks up optimizations made in other libraries when one is compiled with AVX2 and the other only with SSE2.

[Guest]

It's good that AMD is actually doing the work now. Hopefully this increases, and comes to match Intel's level of software contributions.

[wizard69]

If they did this earlier they would have even higher advantage in the server area now.

True but the money simply wasn't there to fund such development. It appear that much of the world is either ignoring or doesn't understand how significant the super computer wins are to AMD and AMD's support of Linux. These contracts should give AMD the opportunity to optimize many parts of the open source stack that makes up Linux. For us more general users it should be a big win! Imagine your computer getting faster for the next 3-5 years.