Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[mpd]

Jumping in late...

I was going to post some comments on this (I used to be a DSP engineer), but I lost track of time and have some real work to do, but the following may be of interest for those who want to understand how floating point math works:

http://docs.sun.com/source/806-3568/ncg_goldberg.html

I will try to add some additional information to the thread later in the day.

[Ivo]

Thanks friends for all the replies. But to say the truth, now I am even less sure than before:

Shall I set Samplitude to 24 bit or 32 bit floating ? What would be better for a practical recording and editing ?

[UnderTow]

Ivo wrote on Thu, 16 February 2006 17:08

Shall I set Samplitude to 24 bit or 32 bit floating ? What would be better for a practical recording and editing ?

My choice would be 32 bit float.

UnderTow

[Buzz]

So what about 64 bit fixed ???? any insite there

LAter
Buzz

[Jon Hodgson]

Buzz wrote on Tue, 28 February 2006 05:46

So what about 64 bit fixed ???? any insite there LAter Buzz

Properly implemented it should be great, better than 32 bit float. Improperly implemented you might see no improvement over 24 bit integer. The thing is to scale values so you use as many of the bits as possible to maintain accuracy, whilst not saturating (Floats basically take care of this bit for you automatically).

Strictly speaking it can also be better than 64 bit float, but quite frankly if you can tell the difference between 64 bit float processing and 64 bit integer processing (in a proper blind test), then I'd say there's a DSP programmer who needs their wrist slapping.

Of the two, 64 bit float is easier to get right.

[Ronny]

Jon Hodgson wrote on Tue, 28 February 2006 07:34

Buzz wrote on Tue, 28 February 2006 05:46 So what about 64 bit fixed ???? any insite there LAter Buzz Properly implemented it should be great, better than 32 bit float. Improperly implemented you might see no improvement over 24 bit integer. The thing is to scale values so you use as many of the bits as possible to maintain accuracy, whilst not saturating (Floats basically take care of this bit for you automatically). Strictly speaking it can also be better than 64 bit float, but quite frankly if you can tell the difference between 64 bit float processing and 64 bit integer processing (in a proper blind test), then I'd say there's a DSP programmer who needs their wrist slapping. Of the two, 64 bit float is easier to get right.

I have a couple of 64 bit processors. What about CPU taxation, Jon? Would the difference between 64 bit fixed int. and 32 floar be significant in terms of how much power is needed for processing at the higher resolution. For example it seems that it's going to 64 bit with the new AMD's and Windows talking of 64 bit for the future. Does this mean that we are going to have to have more RAM and GHz at the CPU level to operate at native 64 bit?  

[UnderTow]

Ronny wrote on Tue, 28 February 2006 16:49

I have a couple of 64 bit processors. What about CPU taxation, Jon? Would the difference between 64 bit fixed int. and 32 floar be significant in terms of how much power is needed for processing at the higher resolution. For example it seems that it's going to 64 bit with the new AMD's and Windows talking of 64 bit for the future. Does this mean that we are going to have to have more RAM and GHz at the CPU level to operate at native 64 bit?

In Sonar, when I switch the audio engine from 32 bit float to 64 bit float I see a small increase in CPU. Nothing close to double.

UnderTow

[Buzz]

I was just curious I use SAW Studios as my DAW software which  operates at 64 bit FIXED so ???? ,

Anyhow I like the program and it runs seamlessly on XP also it is programmed in assembly language so there is low overhead on the CPU not having to do all those DLL calls

Anyhow just curious what you guys were thinking

LAter
Buzz

[Patrik T]

Quote:

I was just curious I use SAW Studios as my DAW software which operates at 64 bit FIXED so ???? , Anyhow I like the program and it runs seamlessly on XP also it is programmed in assembly language so there is low overhead on the CPU not having to do all those DLL calls Anyhow just curious what you guys were thinking LAter Buzz

After having purchased the SAW Basic just recently, I have to say that mixing in the SAW environment feels more like a behaviour of a analogue console - soundwise that is.

I can best describe it as like the audio's integrity is kept better throughout different scaling operations than doing the same scenario in my floating point host.

Simply less muffled and unpleasantly un-phased.

I wish I knew why. Until then I'm just happy that I gave SAW a chance and will remix all present projects in that host instead of with a floating point mix engine.

I guess I prefer 64 fixed bits before 64 floating bits for the moment.

[James Perrett]

Ronny wrote on Tue, 28 February 2006 16:49

I have a couple of 64 bit processors. What about CPU taxation, Jon? Would the difference between 64 bit fixed int. and 32 floar be significant in terms of how much power is needed for processing at the higher resolution.

I'm not a DSP engineer but the conventional wisdom in computing circles is that 64 bit integer calculations would be faster than 32 bit float calculations. That's why Bob Lentini always used to use integers in his software (I don't know if he still does). Modern processors reduce the difference by using hardware floating point processing.

Cheers

James.

[TotalSonic]

jamesp wrote on Tue, 07 March 2006 13:26

Ronny wrote on Tue, 28 February 2006 16:49 I have a couple of 64 bit processors. What about CPU taxation, Jon? Would the difference between 64 bit fixed int. and 32 floar be significant in terms of how much power is needed for processing at the higher resolution. I'm not a DSP engineer but the conventional wisdom in computing circles is that 64 bit integer calculations would be faster than 32 bit float calculations. That's why Bob Lentini always used to use integers in his software (I don't know if he still does). Modern processors reduce the difference by using hardware floating point processing. Cheers James.

In the case of Bob Lentini's SAWStudio: yes, I can confirm via many internet conversations with Bob that it currently uses 64bit integer math for all multiples and divides (except for the built in eq's which use 64bit floating point math) which return an undithered full 32bit DWORD (instead of the 24bit + 8bit mantissa of a 32bit IEEE file) and pass this on to the next process in line - until final output, where the end user has the option of either truncating or dithering to 24bit.

SAWStudio achieves its incredible cpu efficiency due to having its engine hand coded in assembly language (the entire program install file is under 4megs!), and by having many of its calls bypass Windows' various shells and instead address the machine more directly.  While it is certainly not well known, and does not do things in an "orthodox" manner typical of most DAW apps, I find that it is an extremely elegant piece of coding, and by far my favorite DAW app to work in.

Best regards,
Steve Berson

[Patrik T]

TotalSonic wrote on Tue, 07 March 2006 15:32

In the case of Bob Lentini's SAWStudio: yes, I can confirm via many internet conversations with Bob that it currently uses 64bit integer math for all multiples and divides (except for the built in eq's which use 64bit floating point math) which return an undithered full 32bit DWORD (instead of the 24bit + 8bit mantissa of a 32bit IEEE file) and pass this on to the next process in line - until final output, where the end user has the option of either truncating or dithering to 24bit.

Steve, just a rapid question - Do you dither the output (post fader) with 24 bit dither when mixing in SAW (in the case of recording and exporting audio at 24 bits)?

[Buzz]

Patrick there are 4 selections for dither in mixdown @ 24 bit in SAW

LAter
Buzz

[trock]

thanks buzz and steve etc for jumping in here, as i was reading this i was hoping someone would ask about Bob's code and the 64 bit interger math

[TotalSonic]

Patrik T wrote on Tue, 07 March 2006 18:37

TotalSonic wrote on Tue, 07 March 2006 15:32 In the case of Bob Lentini's SAWStudio: yes, I can confirm via many internet conversations with Bob that it currently uses 64bit integer math for all multiples and divides (except for the built in eq's which use 64bit floating point math) which return an undithered full 32bit DWORD (instead of the 24bit + 8bit mantissa of a 32bit IEEE file) and pass this on to the next process in line - until final output, where the end user has the option of either truncating or dithering to 24bit. Steve, just a rapid question - Do you dither the output (post fader) with 24 bit dither when mixing in SAW (in the case of recording and exporting audio at 24 bits)?

Hi Patrik -
I simply haven't been able to hear any difference between just truncating the 32bit integer sum to 24bit output against dithering at the 24th bit when bouncing to disc in SAWStudio to 24bit - as any distortion that is possibly happening from requantizing from 32bit to 24bit is far below the noise floor - so I currently don't add dither to my 24bit in-the-box mixes.  

When going to 16bit I always add dither though - as truncation is very easy to hear when going to that level.

Best regards,
Steve Berson