HQ Player

[Whitigir]

Holy Xich! I am using HQPlayer playing PCM into the best DAC (PCM) I have at the moment, the AGD R2R-7 with Pinkfaun 32/192Khz Windows 10-64bits , HDPlex-dedicated PC....the performances is mind blowing.  Using the Poly-Sinc-Xtr-Mp into 32/192 without Dithering as the dithering will degrade my sound regardless of options.  May I ask @Miskawhat is this filter exactly ? I found that it has the most cohesiveness with the most extensions, details, and some of which in the back ground of the room reverberate and echoes or the rippled energy to be crazily enjoyable.  Dare I say that I never observed these before until now.  If one think that HQPlayer is only good with DSD upsampling, they are dead wrong, the PCM here is so another level.  The level of which I have never heard from Jriver, nor Audirvana.

 

such a blessing Audiophile playback software.  Thank you sir @Miska, I have bought a full retail versions of your HQPlayer.  Truly an extraordinary software 

[arglebargle]

[Edit: posted in wrong thread]

[fred_com]

Hi @Miska   

First of all, thank you for the great software! 

 

My HQPlayer processing PC consists of i7-6700K and Nvidia GTX 1050 Ti, and I have a question about the CUDA offload implementation.

 

I was under the impression that CUDA offload feature will take some of the load from CPU in order to use some heavy filters. 

But it seems that it's not the case - when CUDA Offload is turned on, the GPU load rises to 90% and even poly-sinc-hb begin to stutter. While this feature is off, only poly-sinc-xtr will stutter (which is understandable).

 

I've used search, but could not find straight answer how the CUDA offload works - when used for filters only - does it split the load between GPU and CPU, or just shifts the load to GPU completely?

 

Thanks,

Fedor

[shadowlight]

Has anyone tried to see if HQPlayer and NAA can be made to run under Windows IoT?

[Hoang Anh]

1 hour ago, fred_com said:

Hi @Miska   

First of all, thank you for the great software! 

 

My HQPlayer processing PC consists of i7-6700K and Nvidia GTX 1050 Ti, and I have a question about the CUDA offload implementation.

 

I was under the impression that CUDA offload feature will take some of the load from CPU in order to use some heavy filters. 

But it seems that it's not the case - when CUDA Offload is turned on, the GPU load rises to 90% and even poly-sinc-hb begin to stutter. While this feature is off, only poly-sinc-xtr will stutter (which is understandable).

 

I've used search, but could not find straight answer how the CUDA offload works - when used for filters only - does it split the load between GPU and CPU, or just shifts the load to GPU completely?

 

Thanks,

Fedor

You should use better GPU. Since Hqplayer not smart enough to share work between CPU and GPU well, GPU should better than CPU too make sure it work well. Since you have i7, Try at least 1060 or 1070.

[Miska]

6 hours ago, Hoang Anh said:

You should use better GPU. Since Hqplayer not smart enough to share work between CPU and GPU well, GPU should better than CPU too make sure it work well. Since you have i7, Try at least 1060 or 1070.

 

What player is smarter on this job?

 

I'd say HQPlayer is quite smart, but it is not about smartness, but about efficiency. It is not efficient to try to "share" work between CPU and GPU. Since both have their own RAM and the link between the CPU and GPU is relatively slow, you cannot "share" the work at micro level efficiently. It is like having two big factories, connected by 100 km of small road.

 

But at the moment I wouldn't buy any 10-series GPU, but wait for the new ones instead.

 

[Miska]

7 hours ago, fred_com said:

I was under the impression that CUDA offload feature will take some of the load from CPU in order to use some heavy filters. 

But it seems that it's not the case - when CUDA Offload is turned on, the GPU load rises to 90% and even poly-sinc-hb begin to stutter. While this feature is off, only poly-sinc-xtr will stutter (which is understandable).

 

I've used search, but could not find straight answer how the CUDA offload works - when used for filters only - does it split the load between GPU and CPU, or just shifts the load to GPU completely?

 

It depends on your use case. You can offload either just convolution engine, or both filters and convolution engine. Modulators are always run on the CPU.

 

So if you upsample for example to DSD512, you can choose to run filters on the GPU and leave just modulators running on the CPU. So with offload enabled, some (not all) of the tasks are moved to GPU. This is generally how GPU offloading works on other applications too. If you do for example digital room correction with the convolution engine, you can move this to the GPU and leave rest on the CPU.

 

If your GPU is slower running the filters than CPU, then it will just make things worse. But it also depends on type of the task, generally GPU is better on doing convolution engine or closed-form filters than for example poly-sinc filters. It depends on algorithm.

 

GeForce 10-series is relatively weak for HQPlayer type of processing, let's hope the new RTX 20-series wasn't crippled the same way (to sell more expensive models). But as a rule of thumb, if you have 500€ CPU, don't get under 500€ GPU, or it won't help you much. If you have 150€ CPU, then 500€ GPU is likely to help you relatively more.

 

[fred_com]

Thank you for the answers, Miska and Hoang Anh!

 

That's what I wanted to hear

And I guess I'll go with the Threadripper instead of upgrading to RTX. Hopefully the prices will drop after the second generation will start selling. Although Threadripper system will be more expensive - there will be less coolers and less noise as a result, both mechanical and electrical(if using some passive cooling video card).

[giordy60]

Hi Miska
can you briefly explain how to use the new filters you put in the 3.23.0 release?
thank you

[Hoang Anh]

Maybe buying server mainboard with 2 or more CPU Xeon slot is a future proof investment. When you want run more intensive filter, just buying more CPU and install. Since buy it later wI'll cheaper than investment very strong at once.

[louisxiawei]

Closed-form-16M filter is a new beast. 

 

Just give you an idea how heavy-computing it is:

 

7980XE CPU running 4.0GHz cannot run it alone and causes stutter. With the help of Titanxp (ticked), it can run smoothly while CUDA offload (grey) still cause stutter. 6700K cannot even run it at DSD128.

 

Regarding the SQ: don't have too much time testing all kinds music genre now. Just tested a live vocal track for four times, very impressive stuff! (that's the reason I listened to the same track for 4 times). I believe the mid range is amazingly natural and analogy. 

 

Thank you for brining the new great stuff @Miska

[Miska]

13 minutes ago, louisxiawei said:

Closed-form-16M filter is a new beast. 

 

Just give you an idea how heavy-computing it is:

 

7980XE CPU running 4.0GHz cannot run it alone and causes stutter. With the help of Titanxp (ticked), it can run smoothly while CUDA offload (grey) still cause stutter. 6700K cannot even run it at DSD128.

 

Regarding the SQ: don't have too much time testing all kinds music genre now. Just tested a live vocal track for four times, very impressive stuff! (that's the reason I listened to the same track for 4 times). I believe the mid range is amazingly natural and analogy. 

 

Thank you for brining the new great stuff @Miska

 

On my i7-6950X + GTX 1080, when running on GPU to DSD512 the load on GPU is around 15% and CPU is around 10%. On the CPU it doesn't run because it doesn't parallelize enough to run on all cores. I think I'll need to do some more optimizations for the CPU version.

 

Meanwhile, for PCM the closed-form-M from RedBook to 705.6 kHz (16x) on i7-7700K alone is around 10% CPU load, so quite light. On the same machine CPU RedBook to DSD256 with closed-form-16M works too with about 40% load (so 80% of physical cores). I didn't try DSD512 on it, but based on the load figures it likely won't work.

 

Processing 16 million taps at 20+ MHz rate is getting quite heavy.

 

[Hoang Anh]

Hope miska optimize for new gen 2x00 nvidia soon. 

[jimdukey]

With the 3.23, on my iMac, the Closed form 16M stutters 44 and 48 x 128.

Poly sinc extr mp is great, as always.

 

[Whitigir]

Thanks for letting me know 3.23, just realized that I used 3.22. Upgrading now!

[giordy60]

Miska a version of the closed-form filter 16M -2 is expected? 

[Miska]

1 hour ago, giordy60 said:

Miska a version of the closed-form filter 16M -2 is expected? 

 

No, not planned...

 

 

[Miska]

2 hours ago, Hoang Anh said:

Hope miska optimize for new gen 2x00 nvidia soon. 

 

It is already supported, no need to do anything...

[Nikko1960]

I've been playing around today with 16M upsampling redbook to DSD128.  Works a treat on a late 2014 Mac Mini with 3Ghz Intel core i7 and no stuttering.  Activity monitor shows around 40% cpu utilisation.  By contrast, poly-since-xtr-mp runs at about 48% utilisation and no stuttering.

 

As for SQ, I still can't get anything to sound better to my ears than poly-sinc-short-mp.

 

Nikko

[Miska]

Just now, Nikko1960 said:

As for SQ, I still can't get anything to sound better to my ears than poly-sinc-short-mp.

 

That is what I use myself, except when listening classical, then I use poly-sinc instead.

 

This was mostly answer to recent discussions about very long filters, so I added new variants where it made most sense. With most filters, number of taps is determined dynamically. Now I turned it other way around, so with these two variants you get fixed length. For example current normal "closed-form" doing RedBook to DSD512 is like "closed-form-8M".

 

[juanitox]

1 hour ago, Miska said:

 

That is what I use myself, except when listening classical, then I use poly-sinc instead.

 

This was mostly answer to recent discussions about very long filters, so I added new variants where it made most sense. With most filters, number of taps is determined dynamically. Now I turned it other way around, so with these two variants you get fixed length. For example current normal "closed-form" doing RedBook to DSD512 is like "closed-form-8M".

 

 

can you stop the good sounding filter variants , now i am changing filters for every recording

polysinc Xtr /DSD256 2s for rock , elcectro

polysync /512 for jazz and classic

closedform /256  for women voices

 

My Itunes period was more simple ?

[brightonjel]

Is there some technical reason why the 16M tap closed form version isn't offered for PCM upsampling?  I see it as an option for SDM, but for PCM only the closed-form-M is offered as an option?

[Miska]

42 minutes ago, brightonjel said:

Is there some technical reason why the 16M tap closed form version isn't offered for PCM upsampling?  I see it as an option for SDM, but for PCM only the closed-form-M is offered as an option?

 

Because it would be so excessive long. 16 million taps is around one second at DSD256/DSD512. At 705.6 kHz it would be around 24 seconds...

 

[brightonjel]

Thanks!

 

[Whitigir]

So I see closed-form medium.  Does that mean 8M for PCM ?