HQ Player

[sbenyo]

I am using hqplayer with ryzen threadripper 1920x. It works fine most of the time with the most demanding filters (xtr-lp) up sampling  to DSD512.

 

From some reason I cannot use CUDA offloading properly. I have Nvidia gtx 1060 6GB which should be compatible.

When I turn it on music staggers. I have to use it on partial which indicates convolution only but when playing  a song it displays CUDA offload details which says: upsampler=disabled convolution=disabled. I also don't see any GPU usage in the task manager for hqplayer.

 

Does anyone experience the same thing? Does anyone know why I cannot offload to GPU?

 

[sbenyo]

Thanks for all recommendations about the GPU. I have the latest driver (from 23/7/2019).

I still don't see any GPU usage on hqplayer process. I look in task manager on the GPU columns. 🤨

 

Not sure what else to do or try. It is frustrating as I am sure it should work better if GPU could be used.

 

[sbenyo]

Other issue which is probably related, is the CPU usage. I just moved from 3.25 with xtr to hqplayer 4.1 and using poly-sinc-xtr-lp. As far as I know this should be the same filter. CPU usage with hqplayer 4 is now between 50-99%. With hqplayer 3 it was stable around 40%. I wonder why there should be such a difference...

 

[sbenyo]

6 hours ago, Miska said:

 

 

If you have enabled GPU only for convolution, are you doing any convolution with convolution engine or pipeline matrix?

 

Not sure about it. I am just using hqplayer with simple xtr-lp filter to convert to DSD512. I don't use anything else.

[sbenyo]

Anyone using RTX 2070 with HQPlayer to do DSD512 xtr?

 

I have AMD 1920x and RTX 2070 super which from some unkown reason cannot do DSD512 properly. It plays well but has very short freezes every few seconds. CPU is low (10-15%) GPU (60-72%)

 

If I use CPU only it's better but not perfect for 48K family (still some freezes). CPU alone take ~40%.

 

Does anyone experience anything similar? Is it possible this is still not enough for HQPlayer xtr?

 

[sbenyo]

I am using Muso but I also tried to load a local file but it's the same. I even tried playing a simple 44.1k file and it does not matter if I try to up sample to 44kx512 or 48x512 (always using adaptive rate).

 

It seems strange as my AMD 1920x seems much stronger and is not used much. GPU can get up to 80% at peaks.

During freezes it does not seem specifically at peak. I still don't see a reason it should not work. I can hardly believe it's the difference between 2080 and 2070 super. 

 

I suspect something else is causing this issue. Nothing else is running during play. I also tried to stop anti-virus and it did not help. I may have to start killing processes to see if something else might be causing this.

 

 

 

[sbenyo]

@Miska, I just tested the new build. With CUDA enabled it's the same. No matter if I use the environment setting or not.

 

If I disable the CUDA and use CPU only everything works without any issues now!!

CPU is completely stable at 33% for 48k and 30% for 44k. I am really excited to see it so I will just do more tests now. 😀

 

I got back to using Muso which automatically launches hqplayer. It also works great.

Just to know, when issues started, I disabled both virtualization in bios and hyper-threading in Windows. Should I return hyper-threading now? Will it work better with it in the new build?

 

I am very pleased. I still wish I could make it work with CUDA offload as well.

I am not sure why the RTX 2070 super cannot handle it. If CUDA offload could somehow use more of the CPU power and balance  the GPU usage, it could be much better but I am not sure it this is possible.

 

[sbenyo]

Just to know, I disabled any virtualization, hyper-threading, core boosting (which was on by default) and OC in the bios since I got my PC. I wanted to keep temp as low as possible. I am running on default which is 3.5Mhz and everything with 4.1.1.1 is still  completely stable so no extra boost is needed. @asdf1000, I take your insight on the processes count and I will check it. I have win10 pro + fidelizer pro set at maximum optimization. I use this PC for all my regular work including gaming so I can guess it is not optimized. I still have a doubt regarding the rtx 2070 super. It has about 400 less CUDA cores than the 2080. Other than this they are not that different. I did not try to OC the card so it may also help. I wish Hqplayer could balance between CPU and GPU better so I can take advantage of it.

 

BTW, is it better to use the bios default clock boost option or Win10 hyper-threading?

[sbenyo]

For anyone using hqplayer as main player and wants a decent library, I recommend using Muso.

It works great with hqplayer and has a nice interface. 

http://klarita.net/muso.html

[sbenyo]

Just tested the new poly-sinc-long filter in 4.20 upsampling to DSD512 and it sounds fantastic, taking less resources from my PC. I still have to listen more to decide if I like it better than xtr but for now I keep it. Nice!

[sbenyo]

I experience a new issue with latest 4.2.3. I use the new poly-sinc-long-lp which supposed to be less extreme then an the xtr.

It used to work well used CUDA overload. I try to use it again with multi processors off and cuda on and it gets stuck with high CPU and not GPU.

 

It worked just the day before but today it not. Something in the settings does not work well as previous version. Is there anything that I can do to make it works (using GPU) as expected?

 

For now I moved to ext2 which works well with both multi processors and CUDA on.

 

 

[sbenyo]

On 12/31/2019 at 2:04 PM, Miska said:

 

If you use poly-sinc-long for converting between families, it can take a long time to initialize. Several minutes or hours, depending on machine and settings. Once it has been initialized once, and there are no changes, subsequent starts are faster. GPU is not used for initialization.

 

Did you use the same source rate + output rate combination today and the day before? The conversion ratio is what makes a huge difference (factor of several hundred in initialization times).

 

@Miska, here are my settings. Used to work but something happened. CPU gets high and I have to kill it.

I am using same tracks that used to work. CPU only, gives the same results.

 

Surprisingly, changing to the exterme xtr (CPU only) just works fine.I have AMD 1920x with gtx 2070 super.

 

[sbenyo]

What is the verdict for "Sinc L" filter? I am using a 2070RTX Super and it does not work converting to DSD512.

Does anyone know if this works with 1080TI or any other card? Is it the video card memory issue? it has only 8GB.

 

I am also running AMD 1920x which is the lowest CPU that allows me to run full XTR without GPU. I am not able to use the GPU with XTR and I have to run it on CPU only (which work fine).

 

[sbenyo]

22 hours ago, Miska said:

 

Depending on GPU generation, it can take quite a bit of RAM. So likely you are running out of GPU RAM. I posted some RAM figures earlier, and cannot check right now again. But IIRC, it would take closer to 16 GB, and that it didn't run for that reason on my RTX 2080Ti. You could check how much it takes for DSD256 and then multiply that by 2.

 

 

You don't have a Nvidia GPU on it or for some other reason?

 

 

Other reason. I have 1920x and RTX 2070S GPU. From some reason I cannot use both CPU and GPU for XTR (sinc-xtr-lp) for DSD512

If I use CPU and GPU, CPU is around 15% and GPU gets as high as 70%+. There are gaps in sound. Every few seconds there is a short stop. If I just use CPU, it gets as high as 35-40% but without and lags/stops.

 

It's not clear why I cannot benefit better from both and where is the bottle neck with the GPU (memory, clock?)

[sbenyo]

I am testing HQP with T+A DAC8. I am able to run Sinc-LI with any of the EC modulators with DSD256, but DSD512 with none.

 

Both options sound amazing. I still find the DSD512 a bit better to me than DSD256 + EC. Does anyone have the same experience?

Do you know if EC with DSD512 is going to be a significant difference?

[sbenyo]

3 hours ago, Rovo said:

no significant difference, a small step up, but all the small steps add up to a significant difference

 

Thanks. I did feel step up using sinc-M and sinc-L filters (the ones I can use). I did not feel a big difference with EC modulators in DSD256. The DSD512 without EC, sounds a bit better to me. It may also be specific to T+A DAC. May be that DSD512 is the sweet spot for this DAC.