1

Topic: Re: new commands - for suckers!

Hello, , you wrote: > By the way, the scoring from usage SSE2 too does not impress. void tsp (point cities [], double tourx [], double toury [], int ncities) {int i; double *ClosePt=tourx; double CloseDist; double *p;//... for (i=1; i <ncities; i ++) {double ThisX = tourx [i-1]; double ThisY = toury [i-1]; CloseDist = DBL_MAX; for (p=&tourx [ncities]; wink {double ThisDist; asm ("movapd % [closeDist], %% xmm4\n" "shufpd $0, %% xmm4, %% xmm4\n" "movapd % [thisx], %% xmm6\n" Perhaps, were necessary simply correctly  CloseDist, ThisX etc.? SSE after all makes certain demands to data smoothing...

2

Re: Re: new commands - for suckers!

Hello, , you wrote: > the Next programmer decided to check up a scoring from usage of new commands AVX-2 and 256-digit registers. Outlined the simple and clear test and.... > received double deceleration in comparison with "normal" commands. By the way, the scoring from usage SSE2 too does not impress. > some years ago I too computing  translated on SSE2 and did not see ANY acceleration. And it was any more the test, and real calculations. > attempts to explain all "cache misses" look very faintly.  all these new possibilities. > here its results: > cycles ____________ instructions > 187,888,737 ___ 366,382,169 ___ a C original > 167,129,257 ___ 282,694,918 ___ SSE2 > 390,340,078 ___ 168,337,307 ___ AVX2 > the Text of the original message: > https://groups.google.com/forum/#!topic … 7wBKVUOmfI very authoritative test

3

Re: Re: new commands - for suckers!

Hello, , you wrote: > Hello, kov_serg, you wrote: _>> very authoritative test > Yes, the hell with him with the test. But I translated the real task and from SSE2 received nothing. More precisely, speed did not change more than on 1-2 % and I translated the real task and received a productivity gain in 10 % on new instructions. Can you them be not able simply prepare?

4

Re: Re: new commands - for suckers!

Hello, , you wrote: At me implementation of dissymetric ciphers on AVX produces in times the big productivity, than implementation on "standard" instructions and registers. At first, there there are nuances of their application, secondly, not each task gives a gain. Was specific on tasks  these commands give a gain considerable.