HQ Player

[Miska]

7 hours ago, LoryWiv said:

These 2 settings in my Singxer SU-1 produce equivalent CPU / GPU utilization. Is one sonically preferred to the other?

 

 

Bigger buffer it better. Less context switching and less risk for drop-outs.

 

Buffer size is 262144 DSD samples in this case, the additional latencies reported are accounting for hardware buffers at the DAC end etc.

 

 

[k6davis]

Deleted

[StreamFidelity]

1 hour ago, Miska said:

Bigger buffer it better. Less context switching and less risk for drop-outs.

 

I am always in favor of taking other opinions as an impetus for further optimizations. I tried higher Latencies with following result:

 

poly-sinc-ext2 + ASDM7EC + 96/24/2 -> DSD 256
Asio buffer at maximum with about 50ms = CPU usage is about 1% less. Less micro details.

 

poly-sinc-ext2 + ASDM7EC + 96/24/2 -> DSD 256
Asio buffer at minimum with about 1.5ms = CPU utilization is at 20% -21%. More micro details, somehow smoother.

 

I think it depends on the hardware. Pink Faun uses very good components. I have also tried with my AudioPC. 😊

 

With the least latency it sounds better in my system and for my ears. For others, the results can be different.

[StreamFidelity]

Why is it not possible to play the following combinations?

 

closed-form + ASDM7EC + 96/24/2 -> DSD 256
or
closed-form-fast + ASDM7EC + 96/24/2 -> DSD 256

Roon tries to transmit the data, the HQPlayer does exactly nothing. Same effect with 48/24/2.

 

poly-sinc-ext2 + ASDM7EC + 96/24/2 -> DSD 256

or

closed-form + ASDM7EC + 44.1/16/2 -> DSD 256

works fine

[fred_com]

Hey @Miska I was wondering - is it possible to somehow spread CPU utilization over all existing cores? Is there something you could do in code, or is it controlled by OS? Because right now it seems that only 3-4 cores are used on my 1950x Threadripper. 

 

 

As you can see, I've pushed my CPU to the edge with 4.1GHz OC, and I can get relatively drop-free play only for DSD5EC, sinc-M -> DSD256 (other settings include CUDA offload, Multicore checkbox checked or grayed out, buffer is default).

 

Fun fact: this is true for Win10, on Ubuntu none of EC modulators are working without dropouts (with the same settings). I guess Windows is better suited for multi-core operation.

[brother love]

Update: i7-6700k music server with PCM 96k & 192k music files employing poly-sinc-ext 2 48k x 256 DSD @ different buffer settings:

 

ASDM7EC: continued drop-outs regardless of buffer settings (incl. max 250 ms)

 

ASDM5EC: not buffer setting- dependent, as posted prior: plays & sounds great w/ buffer time set to default

[Paetrick]

I have the 1950X too and I could not even get DSD5EC -> DSD128 running without hiccups. Regardless of the oversampling filter. I also use convolution, but very basic and for stereo. Otherwise I use a GTX 1080 Ti and have CUDA offload enabled. Running Windows 10.

I also tried to change the multicore option from greyed to enabled, but the result is the same. Normally I run closed-form-16M -> AMSDM 512+ -> DSD512. I guess it could be an optimisation issue with the 1950X, since most people seem to get DSD7EC -> DSD256 with (lesser) Intel CPUs.

[Hammer]

11 minutes ago, Paetrick said:

I have the 1950X too and I could not even get DSD5EC -> DSD128 running without hiccups. Regardless of the oversampling filter. I also use convolution, but very basic and for stereo. Otherwise I use a GTX 1080 Ti and have CUDA offload enabled. Running Windows 10.

I also tried to change the multicore option from greyed to enabled, but the result is the same. Normally I run closed-form-16M -> AMSDM 512+ -> DSD512. I guess it could be an optimisation issue with the 1950X, since most people seem to get DSD7EC -> DSD256 with (lesser) Intel CPUs.

 

Thanks for sharing your experience...I too have a 1950X and have been trying to run the new EC options to no avail.

 

Does anyone know if it’s more effective to upgrade the CPU or video card?

[Ernie]

Hello all...I am a new HQPlayer user (latest version 3) and enjoying it very much. I have two question though.

1) I am using hqplayer through roon and upsampling from pcm to dsd128. All is fine when upsampling redbook but when the source for hqplayer is hires ( e.g. 96 or 192 files streaming from qobuz) my dac (topping d10) reads 384 pcm instead of 5.64 Mhz. However the flowcharts in both roon and hqplayer show that things are correctly being upsampled to dsd 128. Anyway to know which is correct and why the dac seems to not be seeing this conversion when starting from hires recordings?

 

2). I have been trying to get NAA working using the following configuration. Roon to hqplayer on an imac i7. Downloaded NAA for mac and ran the script on a macbook pro connected to the dac (topping d10). Tried selecting NAA from hqplayer on the iMac and no devices were listed. The machine running roon and hqplayer is ethernet connected to router, but the machine running NAA is wireless. What am I missing?

 

thanks for any help

[ppy]

35 minutes ago, brother love said:

Update: i7-6700k music server with PCM 96k & 192k music files employing poly-sinc-ext 2 48k x 256 DSD @ different buffer settings:

ASDM7EC: continued drop-outs regardless of buffer settings (incl. max 250 ms)

ASDM5EC: not buffer setting- dependent, as posted prior: plays & sounds great w/ buffer time set to default

I5-7600K 

PCM96 -> poly-sinc-ext2 ASDM7EC DSD256 to  NAA(BeagleBone)

It sounds stable and without artifacts.

 

[fred_com]

2 hours ago, Hammer said:

 

Thanks for sharing your experience...I too have a 1950X and have been trying to run the new EC options to no avail.

 

Does anyone know if it’s more effective to upgrade the CPU or video card?

 

My GTX 1050Ti is pretty happy with processing sinc-M or closed-form-16M and old modulators into DSD512 with something like 50% load. But poly-sinc-xtr-lp maxes it out, so I have to switch to CPU, and 1950x does not disappoint here, with only 25% load on all cores. 

 

Btw, in the latter case, it seems that all the cores are loaded with work, unlike the case with new modulators. 

[LoryWiv]

6 hours ago, StreamFidelity said:

 

I am always in favor of taking other opinions as an impetus for further optimizations. I tried higher Latencies with following result:

 

poly-sinc-ext2 + ASDM7EC + 96/24/2 -> DSD 256
Asio buffer at maximum with about 50ms = CPU usage is about 1% less. Less micro details.

 

poly-sinc-ext2 + ASDM7EC + 96/24/2 -> DSD 256
Asio buffer at minimum with about 1.5ms = CPU utilization is at 20% -21%. More micro details, somehow smoother.

 

I think it depends on the hardware. Pink Faun uses very good components. I have also tried with my AudioPC. 😊

 

With the least latency it sounds better in my system and for my ears. For others, the results can be different.

Thanks for sharing your observations, StreamFidelity. I was (perhaps mistakenly) thinking higher buffers primary impact was to decrease risk of dropouts / stutter, and am trying to understand how it may additionally impact sound quality. I do realize cpu utilization and other characteristics of the audio PC affect noise floor and thus indirectly SQ. Is there something more specific to the filter combinations above at differing latencies that you experience? Of course, ultimately the most important way to judge is your ears, just trying to understand if there is a technical aspect at play as a basis for my own experimenting. Thanks.

[StreamFidelity]

37 minutes ago, LoryWiv said:

Of course, ultimately the most important way to judge is your ears, just trying to understand if there is a technical aspect at play as a basis for my own experimenting.

 

Thank you - And yes you put it right: there are observations in "my" system.

 

At first I also had dropouts on poly-sinc-ext2 + ASDM7EC + 44.1 / 16/2 -> DSD 256. And I experimented unsuccessfully with buffer increase. The solution was to increase the CPU clock frequency with 4GHz. HQPlayer likes to be supplied with higher frequencies. 😉

 

The very low latencies in my system are probably due to my hardware. The Windows 10 Pro operating system is not famous for low latencies. But OK. I optimized it with Audiophil Optimizer 3.00 and the manual shutdown of unnecessary services. That's why I come to around 60 processes. Without optimization, there can be 160 processes.

 

But back to the hardware. I chose a memory with extremely low latencies with the Corsair Dominator Platinum - DDR4 - 16 GB. The lowest I could find. With the application ASUS Ramcache III the processing basically only takes place in RAM. In addition there are the JCAT USB Card FEMTO and JCAT NET Card FEMTO. I think that the clocks before the DAC clock have a decisive influence on the timing of the music.

 

With Linux, Euphony, embedded HQPlayer, etc. the operating system can be further optimized. I know Windows very well. Therefore, I am not ready for a change to an unknown operating system.

 

Sorry for Off Topic. Many insights and suggestions I have from this thread: A novel way to massively improve the SQ of computer audio streaming.

[LoryWiv]

14 minutes ago, StreamFidelity said:

Many insights and suggestions I have from this thread: A novel way to massively improve the SQ of computer audio streaming.

 

I am well beyond Windows novice but quite a bit below your level. Perhaps when I retire that thread will be my spirit guide for further pursuit of the holy grail. 🔅

[k6davis]

5 hours ago, Paetrick said:

I have the 1950X too and I could not even get DSD5EC -> DSD128 running without hiccups. Regardless of the oversampling filter. I also use convolution, but very basic and for stereo. Otherwise I use a GTX 1080 Ti and have CUDA offload enabled. Running Windows 10.

I also tried to change the multicore option from greyed to enabled, but the result is the same. Normally I run closed-form-16M -> AMSDM 512+ -> DSD512. I guess it could be an optimisation issue with the 1950X, since most people seem to get DSD7EC -> DSD256 with (lesser) Intel CPUs.

 

I have a Ryzen 2700x, no Cuda and no convolution. I bought because it has 8 cores/16 threads. I can upsample all formats to DSD128 with any EC modulator. But I can't play anything at EC DSD256.

 

I like AMD chips. They offer value and great performance for certain tasks and perhaps most importantly, they push Intel to be better. I was planning to wait for the forthcoming 16-core Ryzen 3950x, but I've changed my mind. AMD's strengths (generally higher numbers of cores that run a bit slower than Intel's) haven't resulted in high performance for HQP. In fact, HQP EC seems to favor the opposite formula - it uses fewer cores and likes them to run fast. Jussi says he has some optimizations in mind for the future that may change that, but for now, that's the way it is.

 

Others here using Intel chips have reported getting all formats to upsample to EC DSD256, even with an overclocked i7-7700K, which is only 4 cores and several years old!!

 

So, that much is known. What I really want to know is: can any combination of Intel & Cuda upsample RedBook to EC DSD512? Jussi has said that the GPU handles the filters and the CPU handles the modulators. With that division of labor, it seems possible. That would allow for a huge amount of music to be enjoyed at a level beyond even what Jussi has heard. 😮😮😮

[Miska]

8 hours ago, fred_com said:

Hey @Miska I was wondering - is it possible to somehow spread CPU utilization over all existing cores? Is there something you could do in code, or is it controlled by OS? Because right now it seems that only 3-4 cores are used on my 1950x Threadripper. 

 

 

As you can see, I've pushed my CPU to the edge with 4.1GHz OC, and I can get relatively drop-free play only for DSD5EC, sinc-M -> DSD256 (other settings include CUDA offload, Multicore checkbox checked or grayed out, buffer is default).

 

Fun fact: this is true for Win10, on Ubuntu none of EC modulators are working without dropouts (with the same settings). I guess Windows is better suited for multi-core operation.

 

To me it looks like it is using all 16 cores while not touching the 16 siblings (threads), which is intended.

 

I will continue to try to spread the load more. In your case filters are running on the GPU, so there's less work left for CPU. And there's limited amount of possibility to spread modulator loads, so it they primarily utilize two cores per channel. Rest of the work is scattered to all available remaining physical cores. This seems to be working well in your case.

 

If you have very latest Windows 10 (1903), it has some special AMD support optimizations that may be still missing from some Linux kernels. Which Linux distro/kernel are you using by the way? It may be worth comparing Ubuntu 18.04 LTS using either my kernel or Ubuntu's lowlatency kernel, and Fedora 30 using it's stock kernel. Also note that only Ubuntu build of HQPlayer supports CUDA and the different builds are different also in other ways. While Fedora distro in general is very recent "bleeding edge" unlike Ubuntu (and I also use just LTS kernels).

 

[Miska]

49 minutes ago, k6davis said:

In fact, HQP EC seems to favor the opposite formula - it uses fewer cores and likes them to run fast. Jussi says he has some optimizations in mind for the future that may change that, but for now, that's the way it is.

 

Also earlier modulators followed similar pattern but weren't as heavy, now I have new optimizations and the modulator load can be spread more than before. But with EC, computational complexity grew roughly 3x compared to earlier.

 

However, filters can spread the load to much more cores due to different kind of algorithms.

 

It depends on settings which ones dominate the load, filters or modulators.

 

Combining CPU and GPU allows best of both worlds at the same time, few high clock speed cores in CPU and many lower clock speed cores in GPU.

[k6davis]

5 minutes ago, Miska said:

Combining CPU and GPU allows best of both worlds at the same time, few high clock speed cores in CPU and many lower clock speed cores in GPU.

 

Thanks for your support and explanations.

 

I'm hoping someone is able try a powerful CPU and a powerful GPU with a DSD512 DAC and see what's possible. Something like a Core i9-9900K and an RTX 2080 for example.

 

People have managed all formats in EC DSD256. I'm wondering if it's possible to go beyond that with the hardware that's available today.

[Miska]

10 hours ago, StreamFidelity said:

Why is it not possible to play the following combinations?

 

closed-form + ASDM7EC + 96/24/2 -> DSD 256
or
closed-form-fast + ASDM7EC + 96/24/2 -> DSD 256

Roon tries to transmit the data, the HQPlayer does exactly nothing. Same effect with 48/24/2.

 

poly-sinc-ext2 + ASDM7EC + 96/24/2 -> DSD 256

or

closed-form + ASDM7EC + 44.1/16/2 -> DSD 256

works fine

 

If your DAC cannot do 48k-base DSD, the conversion factor limitation prohibits use of some filters from 48k-base PCM rates to 44.1k-base DSD rates. Possible factors are listed in the manual table.

 

[k6davis]

Jussi,

 

Also, is it possible to get a DSD7EC modulator? Or is there no benefit to something like that?

 

The "DSD" modulators sound warmer and more natural to my ears, but only the "ASDM" versions get all the love, with 7th order and optimization for 512+. 😄

[DancingSea]

I have a DirectStream Junior DAC which upsamples to 20X DSD on its own.  I greatly enjoy HQP and its upsampling, but have read various places about users not doing any upsampling in HQP and passing it on to the DAC.  For instance, Roon —> HQP (no upsample) —> DAC.

 

How do I try this out?  I don’t see a specific option in HQP to “turn off upsampling”.  I see I can select “none” in the filters and modulators, but that’s all I see.

 

How do I turn off upsampling in HQP Desktop (4)?

[Ernie]

10 hours ago, Ernie said:

Hello all...I am a new HQPlayer user (latest version 3) and enjoying it very much. I have two question though.

1) I am using hqplayer through roon and upsampling from pcm to dsd128. All is fine when upsampling redbook but when the source for hqplayer is hires ( e.g. 96 or 192 files streaming from qobuz) my dac (topping d10) reads 384 pcm instead of 5.64 Mhz. However the flowcharts in both roon and hqplayer show that things are correctly being upsampled to dsd 128. Anyway to know which is correct and why the dac seems to not be seeing this conversion when starting from hires recordings?

 

2). I have been trying to get NAA working using the following configuration. Roon to hqplayer on an imac i7. Downloaded NAA for mac and ran the script on a macbook pro connected to the dac (topping d10). Tried selecting NAA from hqplayer on the iMac and no devices were listed. The machine running roon and hqplayer is ethernet connected to router, but the machine running NAA is wireless. What am I missing?

 

thanks for any help

 

just to elaborate on my earlier post.  I seem to be getting the incorrect reading on my dac when inputting anything higher than 44.1 into hqplayer. However I do not see the same behavior when the upsampling is done by roon or audirvana (dac shows 5.64 Mhz as it should) . Any help would be greatly appreciated.

[k6davis]

35 minutes ago, Ernie said:

 

just to elaborate on my earlier post.  I seem to be getting the incorrect reading on my dac when inputting anything higher than 44.1 into hqplayer. However I do not see the same behavior when the upsampling is done by roon or audirvana (dac shows 5.64 Mhz as it should) . Any help would be greatly appreciated.

 

Welcome Ernie.

 

Should be simple to solve. Can you post a screenshot of your settings? You can paste in image right into the editor on this forum. 

 

I'm headed out for a while, but most of the experienced users or the developer @Miska should hopefully be able to spot the issue.

[Ernie]

Thanks very much. Here are my hqplayer settings. Let me know if anything else will help.

 

 

[Ernie]

One more clue, if it helps. Just for kicks I tried downsampling a hires track to 44.1 with roon and then having roon send that to hqplayer.  In that case my dac shows accurately that the dsd conversion has taken place...reads 5.64.