HQ Player

[Yviena]

27 minutes ago, Miska said:

 

It is hard to say, but I think they have made some drastic changes to the implementation. I'll try to get some new DACs that have filter switching over the spring. Both Topping and ADI-2 still retain this feature?

 

I have just the EVGA NuAudio card, but it doesn't have option for switching the filter. But since volume control doesn't work in DSD mode, I assume it is in DSD Direct mode.

 

The newest topping one has DSD filter switching implemented, along with the DSD direct mode in DAC mode.

 

Newest revision adi2 with AKM 4493 still has the 50/150khz filters I believe.

 

The PCM filters of the AKM4499 though  aren't particularly good specs wise with the sharpest filters having only -95dbfs attenuation.

[Account Closed]

1 hour ago, Miska said:

 

Yes, likely it won't work with the official Rendu's OS. You could try with my NAA image if it works better (I have not tested myself).

 

It also depends on the particular USB hardware. It should work with Intel USB controllers though, with my HQPlayer OS and NAA images.

 

When you are using NAA, the OS where HQPlayer is running doesn't  matter at all from that perspective since output is to a network endpoint bypassing the OS audio stack entirely.

 

OK thanks. I thought that might be it.

[AudioXP]

On 2/15/2020 at 10:44 AM, Miska said:

 

It usually still leaves plenty of images. But for example at 1.5 MHz Holo starts to become clean. But I think better way is to use DSD instead.

 

Depends on the OS and backend, there's practically no technical limit as such. On Linux at the moment PCM limit is 3.072 MHz. On macOS there's no limit (to be exact the limit is a bit over 4 GHz). On Windows the limit is at the moment 1.536 MHz. If needed, it is easy to extend the rates on Linux and Windows.

 

Got a Raspi 4 to work as NAA doing DSD256 with Holo 2 using the latest NAA image. When switching to PCM 1.536 MHz, I got sound, but with a horrible distortion, like a high-frequency ringing. Made me scared to destroy my tweeters. On a NUC NAA, got exactly this working fine (and enjoyed also the SQ :-). Is this to be expected? Can something be tuned...?

[Miska]

1 hour ago, Yviena said:

The PCM filters of the AKM4499 though  aren't particularly good specs wise with the sharpest filters having only -95dbfs attenuation.

 

Yes, and ESS and Chord settle on -120 dBFS which is not so great either, about 20-bit accuracy if not counting the images and phase shifts which depend on the final implementation with analog filter.

 

[zouyu513]

Dear miska

I have a 9900ks, but I run it in linear power supply and fanless way. I found that when the Turbo frequency is 4.2g, the PCM of any sampling rate can be upsampled to DSD256 with poly sinc xtr + asdm7EC, but dsd128 cannot be smoothly upsampled to DSD256. Does the DSD128 upsampling need more calculation?

[Miska]

5 hours ago, zouyu513 said:

I have a 9900ks, but I run it in linear power supply and fanless way. I found that when the Turbo frequency is 4.2g, the PCM of any sampling rate can be upsampled to DSD256 with poly sinc xtr + asdm7EC, but dsd128 cannot be smoothly upsampled to DSD256. Does the DSD128 upsampling need more calculation?

 

It is not necessarily heavier, but the load pattern is different. You could try enabling empty matrix for the first two channels with gains of 0 dB. This changes the workload distribution, but doesn't have other implications in this case.

 

[Yviena]

@Miska i don't much know much about programming but is there any technical reason why you can't parallelize the EC modulators more, or is it just too much work?

 

Looking at current developments i don't see us ever going beyond the 5ghz barrier especially in high core count CPU's without new materials to make processors from, so DSD512 with EC modulators may take a decade or more to become feasible.

[Miska]

30 minutes ago, Yviena said:

@Miska i don't much know much about programming but is there any technical reason why you can't parallelize the EC modulators more, or is it just too much work?

 

The reason is more mathematical, some algorithms can be parallelized more, some less.

 

30 minutes ago, Yviena said:

Looking at current developments i don't see us ever going beyond the 5ghz barrier especially in high core count CPU's without new materials to make processors from, so DSD512 with EC modulators may take a decade or more to become feasible.

 

That may be the case unless I find way to make it faster or processors become more efficient in terms of how many clock cycles are used for certain set of instructions. It would be already better if all instructions would execute in a single clock cycle.

 

There is always also overhead associated with parallelization in cases where cores need to communicate frequently or access shared data.

 

[Yviena]

30 minutes ago, Miska said:

 

The reason is more mathematical, some algorithms can be parallelized more, some less.

 

 

That may be the case unless I find way to make it faster or processors become more efficient in terms of how many clock cycles are used for certain set of instructions. It would be already better if all instructions would execute in a single clock cycle.

 

There is always also overhead associated with parallelization in cases where cores need to communicate frequently or access shared data.

 

Hmm i wonder if a CPU with more cache would help, maybe something like HBM on-die, or the crystal lake intel CPU's that have a 128mb L4 cache.

[Miska]

4 minutes ago, Yviena said:

Hmm i wonder if a CPU with more cache would help, maybe something like HBM on-die, or the crystal lake intel CPU's that have a 128mb L4 cache.

 

Maybe it does, but it is likely only the first cache levels that matter most. Compared to filters, memory footprint of modulators is fairly small, although the data rates in question are high.

 

[zouyu513]

3 hours ago, Miska said:

 

It is not necessarily heavier, but the load pattern is different. You could try enabling empty matrix for the first two channels with gains of 0 dB. This changes the workload distribution, but doesn't have other implications in this case.

 

I enabled matrix pipline，but it did not work. Cpu load was not very high(totally22.68% max core47.43% ).

[Miska]

2 minutes ago, zouyu513 said:

I enabled matrix pipline，but it did not work. Cpu load was not very high(totally22.68% max core47.43% ).

 

Which way it didn't work? Are you on latest version of HQPlayer?

 

[zouyu513]

I mean the upsampling is still not smooth(Source: DSD128 and DSD256), but DSD64 can be upsampled to DSD256(7ec+xtr) very smoothly. My hqplayerd is HQPlayer 4 Embedded 4.14.0.

[3ggerhappy]

@Miska I am also trying PCM 1.536, but it wont play well, having sluggish distorted sound same as @bobflood. Is there any setting I need to trigger? Using embedded hqp(Trial) on a Spring 2. Or do you think its a hardware issue? Im using an HP prodesk 600 g2 Intel I5 8gb memory. It could play all filters at 256DSD except non2s. For PCM 768 plays fine on all filters except poly-sinc-long-lp it just crashes after a while and i need to restart the machine, weird because I could use this fine on DSD 256 after a few minutes of wait.

[Miska]

5 hours ago, 3ggerhappy said:

@Miska I am also trying PCM 1.536, but it wont play well, having sluggish distorted sound same as @bobflood. Is there any setting I need to trigger? Using embedded hqp(Trial) on a Spring 2. Or do you think its a hardware issue? Im using an HP prodesk 600 g2 Intel I5 8gb memory. It could play all filters at 256DSD except non2s. For PCM 768 plays fine on all filters except poly-sinc-long-lp it just crashes after a while and i need to restart the machine, weird because I could use this fine on DSD 256 after a few minutes of wait.

 

Have you tried with my HQPlayer OS or are you on something else?

 

If you try to go from 44.1k base to higher 48k base with poly-sinc-long you are up for a very long wait (can be hours). While going from 48k base to 44.1k base is usually fine with not as long wait. So I'm a bit reluctant to disable support for non-integer ratios in general. I have not seen crashes, but I've seen for example four hour wait times. In fact it would be interesting to compare initialization times of poly-sinc-long between Intel and AMD CPUs. In similar cases in the past, AMD CPUs have been notably faster.

 

[3ggerhappy]

2 hours ago, Miska said:

 

Have you tried with my HQPlayer OS or are you on something else?

 

If you try to go from 44.1k base to higher 48k base with poly-sinc-long you are up for a very long wait (can be hours). While going from 48k base to 44.1k base is usually fine with not as long wait. So I'm a bit reluctant to disable support for non-integer ratios in general. I have not seen crashes, but I've seen for example four hour wait times. In fact it would be interesting to compare initialization times of poly-sinc-long between Intel and AMD CPUs. In similar cases in the past, AMD CPUs have been notably faster.

 

 

 

Thanks lemme try your OS, currently it is setup in Ubuntu. Yeah for long-lp, I think I did not wait long enough I just though it crashed because there is no response for some time, did not know it could take hours.

[StreamFidelity]

5 hours ago, Miska said:

If you try to go from 44.1k base to higher 48k base with poly-sinc-long you are up for a very long wait (can be hours).

 

This filter is currently my favorite filter for PCM 192kHz. I haven't used a stopwatch, but in this constellation, it takes about 5 minutes to load. I have an Intel Core i9-9900K on the ASUS ROG Maximus XI Gene mainboard.

 

It is interesting to use ASUS RAMCACHE III. This tool loads frequently used programs into the fast RAM and makes the content available the next time it is started. Sometimes poly-sinc-long runs immediately, sometimes not. Unfortunately I can not see any regularities yet. It would be helpful if the HQPlayer permanently saved the loaded files (44.1, 48, 88.2, 96,...) in the cache (until the next program start).

[Account Closed]

15 hours ago, Miska said:

 

Have you tried with my HQPlayer OS or are you on something else?

 

If you try to go from 44.1k base to higher 48k base with poly-sinc-long you are up for a very long wait (can be hours). While going from 48k base to 44.1k base is usually fine with not as long wait. So I'm a bit reluctant to disable support for non-integer ratios in general. I have not seen crashes, but I've seen for example four hour wait times. In fact it would be interesting to compare initialization times of poly-sinc-long between Intel and AMD CPUs. In similar cases in the past, AMD CPUs have been notably faster.

 

I am going from 44.1k to 48k DSD using ps long mp 2s and ASDM7 EC and the wait is only a few seconds if that.

[Miska]

8 minutes ago, bobflood said:

I am going from 44.1k to 48k DSD using ps long mp 2s and ASDM7 EC and the wait is only a few seconds if that.

 

Yes, 2s is possibly not too bad. But the time increases pretty rapidly, so for example same with single stage variant may be way too much wait.

 

[Account Closed]

5 minutes ago, Miska said:

 

Yes, 2s is possibly not too bad. But the time increases pretty rapidly, so for example same with single stage variant may be way too much wait.

 

I am glad I did not try the single stage, I would have thought something was wrong. Single stage filters push my CPU too hard as I don't have a GPU. I have an i5 8400 6 core on Z370i mb with bios set to get 3.9 Ghz all all cores and it runs at 50% total load. I thing ps long mp is your best filter right now.

[JTS]

I have a quick question about the AKM DSD direct mode and HQ Player. When using DSD direct, my understanding is that DSD will still go through AKMs DSD modulators/noise shapers even when using DSD direct on something like an RME ADI-2. Is this correct? If so, how would this operate differently than say a Holo Spring in NOS mode? What are the practical implications for the using HQPlayer?

[Yviena]

11 minutes ago, JTS said:

I have a quick question about the AKM DSD direct mode and HQ Player. When using DSD direct, my understanding is that DSD will still go through AKMs DSD modulators/noise shapers even when using DSD direct on something like an RME ADI-2. Is this correct? If so, how would this operate differently than say a Holo Spring in NOS mode? What are the practical implications for the using HQPlayer?

No according to the schematic of AKM chips DSDD mode will bypass both volume, and internal modulator.

[JTS]

12 minutes ago, Yviena said:

No according to the schematic of AKM chips DSDD mode will bypass both volume, and internal modulator.

 

You inspired me to look at the block diagram and you're right - the block diagram does show DSDD bypassing the modulator, but it still goes through a filter stage (SCF). I am curious about the interaction between HQ Player and the AKM filter vs something like the Holo Spring's NOS DSD setting.

[Miska]

3 hours ago, JTS said:

You inspired me to look at the block diagram and you're right - the block diagram does show DSDD bypassing the modulator, but it still goes through a filter stage (SCF). I am curious about the interaction between HQ Player and the AKM filter vs something like the Holo Spring's NOS DSD setting.

 

Holo Spring DSD DAC section is also a filter. These mixed domain conversion filters shouldn't be mixed with oversampling filters. In both cases there's also a subsequent analog filter. TI chips also have similar implementation for DSD conversion (they are just always in "direct" mode). Cirrus Logic chips also have AKM-style Direct DSD mode to goes straight to the output SCF. AKM has two filter configurations available, TI chips have four, Holo and Cirrus have just one fixed configuration. Other such DACs are Denafrips and T+A. And my DSC1 design.

 

DSD or PCM, correct reconstruction requires proper analog filter.

 

Characteristics of these filters vary and the end result depends on sampling rate and modulator used.

 

[JTS]

2 hours ago, Miska said:

AKM-style Direct DSD mode to goes straight to the output SCF.

 

 

Thanks for your response. I am trying to wrap my head around how DSD Direct with the AKM works. I am thinking of picking up a D90 for another system. Apparently their DSDD also goes straight to the analog filter, as it does in the RME, when the unit is in DAC mode.