Any DSP that supports SDM & DXD

[audiventory]

Personally, though, I want to pretend that DSD doesn't exist, aside from a weird lossy compression format for hi-res distribution, and converting it to 24/96 or 24/172-192 is a convenient way to "decompress" it and get on with the normal signal chain and my favorite playback and DSP software.

 

More convenient (even right ) way convert DSD to 64-bit float for DSP. It's more suitable format for avoiding accumulating of errors and overloading. Using integer formats for DSP is issue of performance and calculating resources. Especially it actually for hardware processing. Though, now "hardware" is software anyway. Even PLD chips that programmed in programming language

[rcohen]

More convenient (even right ) way convert DSD to 64-bit float for DSP. It's more suitable format for avoiding accumulating of errors and overloading. Using integer formats for DSP is issue of performance and calculating resources. Especially it actually for hardware processing. Though, now "hardware" is software anyway. Even PLD chips that programmed in programming language

That makes sense. It's what I was speculating that you are doing.

 

The only bad news is the high performance cost and lack of software and lack of hardware support (DACs, HDMI, etc.) of very high sample rate double-precision DSD. So, you have to filter/downsample/redither at some point.

 

I understand what you are saying, though, that if the CPU time is available anyway, doing a 1-bit redither after processing may be better in some respects than converting to (dithered) 24/176-192 prior to processing.

 

So, in what ways is it better?

[Miska]

For online (on fly, real-time) algorithms need optimize performance also.

 

I do optimize performance, but I don't trade quality for speed. If in the end your computer is too slow, you need to go shopping for faster one...

 

SSE, AVX and hand-written assembly are good way to get faster code.

[Miska]

I understand what you are saying, though, that if the CPU time is available anyway, doing a 1-bit redither after processing may be better in some respects than converting to (dithered) 24/176-192 prior to processing.

 

...remember that DAC chips do especially bad job in converting that PCM back to SDM. They just don't have the computation resources to do it properly.

 

Adding extra rate conversion down and up doesn't improve things...

 

But for those who wish, HQPlayer supports converting DSD down to 1/16th rate (2.8 MHz -> 176.4 kHz) and then processing convolution at that speed if you like. Output is then PCM at rate of your choice. Or just plain conversion to 96 kHz PCM output if you'd like to use Dirac on Mac.

[rcohen]

...remember that DAC chips do especially bad job in converting that PCM back to SDM. They just don't have the computation resources to do it properly.

Really? I've never looked deeply into this, but I assumed they just did an upsample + noise shaping. In fact, I was under the impression that modern DACs did multi-bit noise shaping, allowing for superior results compared to software that can only able to send 1-bit SDM signals to the DAC.

 

Do DACs attempt to upsample DSD data and/or convert to multi-bit? I would think that would be more compute intensive and lossy than upsampling PCM data.

[Miska]

Really? I've never looked deeply into this, but I assumed they just did an upsample + noise shaping.

 

Yeah, the devil is the "just" part.

 

Yes, they will do upsampling, but you need to look at how they do it. And then the noise shaping is delta-sigma modulation which is the same stuff as when you convert to DSD.

 

In fact, I was under the impression that modern DACs did multi-bit noise shaping, allowing for superior results compared to software that can only able to send 1-bit SDM signals to the DAC.

 

Typical 2.5-bit vs 1-bit noise shaping the difference is 9 dB in ultrasonic noise level. So pretty much nothing. With ESS and 6-bit the difference is 30 dB which is still very small. Much bigger is the damage made by use of sample-and-hold oversampling.

 

Do DACs attempt to upsample DSD data and/or convert to multi-bit? I would think that would be more compute intensive and lossy than upsampling PCM data.

 

For those that do, it is less. Because you get to bypass the PCM upsampling part and do just the delta-sigma modulation to another word-length and maybe somewhat higher rate.

So it is otherwise same path, but one large processing block dropped out (ESS as an example).

 

And then you have those DACs chips like TI's that just perform straight conversion from DSD to analog bypassing all the digital processing.

[LewinskiH01]

I come at this from a very different angle: I know squat about DSD or PCM being inherently/theoretically better. From my little experiments I'm wondering if certain DACs might not be designed in such a way they just sound better with DSD. That would add yet another variable.

 

You can see my system in my signature below. Out of curiosity I purchased an exaSound e22 to explore Upconverting to DSD. For reference, the Metrum + Audiophilleo were $2k about 18/24 months ago. The e22 is $3.5k and newer.

 

Using JRiver + AO (but keeping WS2012 in GUI mode) and playing PCM, the Metrum/AP combo sounded better. There is glare to the e22.

Upconverting to DSD using JRiver didn't help: the e22 in this mode wasn't sounding better than Metrum/AP playing PCM. Maybe it wasn't worse either, but it wasn't better and running hardware significantly cheaper.

Then I tried HQPlayer converting PCM to DSD128 and have to admit it sounds better to me than the Metrum/AP combo. Is it $1.5k better? Haven't arrived at that conclusion yet.

 

I still need to try HQPlayer playing PCM to the e22 and Metrum/AP and watch for differences. Maybe the difference in sound is coming from HQPlayer rather than the e22 ability to play DSD?

[YashN]

Any link for Thorsten's demo, bibo01?

 

Just tested native DSD256 on Win with Foobar2000 and it sounds very good. My PC computer is too slow for real-time upsampling IMO.

 

HQ Player's up-conversion sounds fabulous (only tested DoP for now).

 

However, the algorythm used by JRiver for PCM-to-DSD conversion is not that good and it measures poorly even compared to Foobar's AsioProxy, as demonstrated by AMR/IFI Thorsten Loesch.