Assembly - divps and CPU load?

[trogluddite]

Replacing all the modlulus, division and rounding with SSE2 stuff and rcpps, I nearly halved the CPU load.

Rccps is as heavy i dunno who told you the opposite it`s 1/x.

Stoopid me! I always have to learn stuff the hard way. I'd read this one on the forums somewhere - the low CPU looked great at first - till I had all my audio paths running at once - then the denorms started breeding like bunnies! Oh well, back to divps then! All running nicely now - and still saved about 30% of the CPU .

If you know the internal representation of the floating values (e.g. if they are IEEE 754) you can mask them directly with an integer

Masking the mantissa but not the exponent - that kind of thing? Time for some experiments.

Use a text editor that has code highlighting.

Thanks for the tip, I'll go see what those nice folks at Source Forge have in their box of goodies...

The Luddite

[infuzion]

Rccps is as heavy i dunno who told you the opposite it`s 1/x.

Stoopid me! I always have to learn stuff the hard way. I'd read this one on the forums somewhere - the low CPU looked great at first - till I had all my audio paths running at once - then the denorms started breeding like bunnies! Oh well, back to divps then! All running nicely now - and still saved about 30% of the CPU .

Division by whole numbers take about half as much cycles no modern CPUs, so as long as your dependency interpolation is set up well, you should notice little impact. Post/send the OSM if you'd like to see if there could be more savings.

[trogluddite]

Post/send the OSM if you'd like to see if there could be more savings.

You guys are just too nice!
When I've next got time on my big computer, I'll try and break it down into bite sized chunks - already getting loads of ideas from my forum searches, but this ASM stuff is really taxing my grey matter!

[trogluddite]

Use a text editor that has code highlighting

Thanks again; that's Top Tip of the month for sure!
Just installed Notepad++ from sourceforge; wow, what a difference that makes. Just rooted out about a dozen unused variable declarations from 750+ lines of assembly in a few minutes, and searching for where you've used a register is a breeze!
If any Notepad++ users are interested here's a little user language file for the S|M ASM primitive with just the recognised opcodes, and the stream bits included...

SM Assembly for NotepadPP.zip

(901 Bytes) Downloaded 99 times

...had to zip it, SM won't attach .xml files. Put in in your Doc's and Settings/username/Notepad++ folder, or cut and paste the relevant bits into the user language file if you have one already.

[Andrew J]

If any Notepad++ users are interested here's a little user language file for the S|M ASM primitive with just the recognised opcodes, and the stream bits included...

SM Assembly for NotepadPP.zip

...had to zip it, SM won't attach .xml files. Put in in your Doc's and Settings/username/Notepad++ folder, or cut and paste the relevant bits into the user language file if you have one already.

Nice - thanks for that!

[TomC]

from 750+ lines of assembly in a few minutes

If the 750 lines are one single code I would rethink your approach.

If any Notepad++

SM Assembly for NotepadPP.zip

Thanks a lot for that.

Tom

[aliasant]

Thanx Luddite !

This style looks much better then the default Assembly style but I had some difficulties placing the langauge file in the right place
For others using Win7 drop the file in this folder and restart the app:
username/appdata/roaming/notepad++

[trogluddite]

For others using Win7

Ah yes, sorry, should have pointed out that I'm still on XP!

If the 750 lines are one single code I would rethink your approach

Not so easy with this one as it's not really a standard synth/FX architecture, more like a self comtained app - all of the core features rely on a common audio buffer - once that's inside a code primitive (as opposed to using mem array components), there's no way to share the buffer between modules. Now that I have a quck way to edit the code, I'm quite happy working this way - I'm used to it from the Z80 ASM I used to write on my ZX Spectrum years (decades!) ago (used to have to convert the code to hex by hand in those days too!). Managed to lose nearly 150 lines during optimisation already though.

[MegaHurtz]

If you get it to run you save some movaps.

funny that:

movaps xmm1,diva;
movaps xmm2,divb;
divps xmm1,xmm2;
movaps out,xmm1;

Uses less cpu than the divider primitive

[rl]

I don't believe your claim

If you connect both div methods to an analyser, that's the code for the div primitive:

Code: Select all

***** Init *****
push ebp;
mov ebp,edi;
pop ebp
ret;
***** Sample Rate *****
push ebp;
mov ebp,edi;
movaps xmm0,dword ptr[ebp+4208]
divps xmm0,dword ptr[ebp+4224]
movaps dword ptr[ebp+4192],xmm0
pop ebp
ret;


your asm code:

Code: Select all

***** Init *****
push ebp;
mov ebp,edi;
pop ebp
ret;
***** Sample Rate *****
push ebp;
mov ebp,edi;
movaps xmm1,dword ptr[ebp+4208]
movaps xmm2,dword ptr[ebp+4224]
divps xmm1,xmm2
movaps dword ptr[ebp+4192],xmm1
pop ebp
ret;


Pretty much the same

[infuzion]

If you get it to run you save some movaps.
movaps xmm1,diva;
movaps xmm2,divb;
divps xmm1,xmm2;
movaps out,xmm1;
Uses less cpu than the divider primitive

I don't believe your claim

If you have only a divide by itself, either one is the same. If you have a bunch of other opcodes before/after that divps in the same Assembly primitive, epically if they're outside of the divide's dependence chain, then that divps will indeed be faster.

[stw]

If you connect both div methods to an analyser, that's the code for the div primitive:

Code: Select all

***** Init *****
push ebp;
mov ebp,edi;
pop ebp
ret;
***** Sample Rate *****
push ebp;
mov ebp,edi;
movaps xmm0,dword ptr[ebp+4208]
divps xmm0,dword ptr[ebp+4224]
movaps dword ptr[ebp+4192],xmm0
pop ebp
ret;


where do you get that code from?

[rl]

through the signal analyzer. SM's mighty general purpose debugging tool:

analyzer.png (30.68 KiB) Viewed 1157 times

[stw]

wohoo...
Didn't know that. This is really cool!
Thanks for the info!

[MegaHurtz]

Thats a nice feat indeed, havnt seen that before.
Also tells you the SM version uses the smaller register
Sometimes higher cycles with lower cpu spike equals lower percentage.