Stereo Tool

It could very well be that AVX2 has no benefits over AVX; I'm going to run a build of the different targets and compare the assembly output. AVX2 mainly adds integers to AVX, and I don't do much with integers in loops.

Edit: Ok, there are benefits. AVX2 has the multiply-add instruction (which performs a multiplication and an addition in a single cycle) - that is really useful and I see it appear in multiple loops, causing the code to be shorter (hence faster). I don't know how big the effect is though, and until my test system is stable I cannot measure it.

This version should be the fastest on all platforms - maybe with an exception for the older Intel's, but they might also benefit slightly. Unless my own measurements here are showing me the wrong things (which is possible).
http://www.stereotool.com/download/dsp_ ... ll_ipo.dll

This version requires SSE2, but if possible it will make use of SSE3, SSE4.1, SSE4.2, AVX, AVX1 and AVX2. The only fear that I have is that the selection of the correct code might cause some overhead. I don't see any overhead on 2 systems that I've tested it on so far though.

According to the red bar , again NEW/NEW is a bit lower than the new all_ipo

O, I read your previous post as 'the last 3 are identical' - is that not what you meant? What hardware are you using? And how big is the difference? And is the all_ipo equal to the fas_ etc. versions?

Yes , this is what i mean . all_ipo is as i can see equal to the previous "fas", "nei" ect...
Even again as i do it a lots of time Old/old , old/new and new/new ... the New/New was the "bit lower" .
How much ? ... well..... one step lower ( there are steps on that bar ) . I`m with 1024x768 resolution , full screen , so i look at the above letters of the word "Percentage" . My bar is in the beggining of letter "n" , but with New/New it is one step lower .
Hardware is K10 Athlon II X2 240e , 3Gb ram , Xp pro SP3 / Full updates , MSI K9A2CF

If i remember right , time results of another Athlon few pages ago shows again the shortest time with NEw/New

On my E8400 is slowest.
Fastest is is "1". And vs this one is 2 seconds faster in 04:30 track. 01:12 bs 01:14
was using web preset.

_________________
control point
control point2

1:14 = 74.0? So even slower than 3? Or did you round the result? (And is the measurement really that accurate??)

Used other preset.. will measure all again.

_________________
control point
control point2

I'm still having weird hiccups on my test system, but I did a measurement anyway and the differences are big enough to mask the effect of the hiccups.

File, 4:00 long. Heavy preset with composite clipper and stokkemask, calculate RF spectrum is also enabled.
CPU: i7 4770S. Memory is too slow (only 1 stick is used instead of 2, since I'm still busy testing what's causing the problems on this system).

Results:
OLD/OLD: 64 seconds
OLD/NEW: 58 seconds
NEW/NEW: 56 seconds
ALL_IPO: 53 seconds

Moving from OLD/OLD (equal to the current release) to ALL_IPO (new compiler, new library, multiple processing paths for different CPU's) gives a 20% improvement!

i can`t test this way .. i don`t know why. Any suggestion ?

Are you all testers disable the FM output ? or all outputs