I’m trying to setup HQPlayer on a dedicated machine to squeeze as much as possible from it. But I think maybe cuda isn’t working as should? I have installed all the various cuda drivers but it seems HQPlayer uses very little memory from the GPU. While playing a track this is the output from the GPU using the command nvidia-smi
Looking at the process for hqplayerd its only using 145MiB? should this not be a lot higher?
Some other info I have found using the nvcc --version command
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Sun_Jul_28_19:07:16_PDT_2019
Cuda compilation tools, release 10.1, V10.1.243
My processor is a AMD Ryzen 3 2200G, 16gig ram. During playback the speeds of the cores seem to rise around 3.6Ghz (2 examples below)
cpu MHz : 3425.257
cpu MHz : 3421.047
cpu MHz : 3442.948
cpu MHz : 3471.218
cpu MHz : 3693.204
cpu MHz : 3693.213
cpu MHz : 2460.553
cpu MHz : 2798.226
And finally a screenshot of roon’s signal path:
So… is cuda working? what can I do to improve the performance?
Just for info I followed this guide for installing cuda.
Looks like it is working as it should, hqplayerd is shown as “C” (compute) process as opposed to “G” (graphics). 145 MB is quite a lot of memory already. And poly-sinc-short-mp-2s among the lightest to process of all the filters. If you try with something like poly-sinc-xtr-lp (non-2s), or poly-sinc-short-mp instead, you’ll likely get higher figures.
I’m a little curious though why do you have graphical desktop on a dedicated machine with HQPlayer Embedded?
That is also normal, or actually quite low, because typical figures can easily go to 4 - 5 GHz range with ASDM7EC. But since your output is only DSD128 you get along with lower figure.
This got installed during the graphics drivers installation. I was going to do a full reinstall once I got the proper guides to follow to get everything running smoothly.
The figures for this option seems to be very similar:
+-----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 891 G /usr/lib/xorg/Xorg 8MiB |
| 0 N/A N/A 1014 G /usr/bin/gnome-shell 2MiB |
| 0 N/A N/A 4684 C hqplayerd 117MiB |
+-----------------------------------------------------------------------------+
cpu MHz : 3266.789
cpu MHz : 3297.548
cpu MHz : 2853.765
cpu MHz : 2855.525
First, get started by installing minimal Ubuntu Server 20.04 LTS (with just SSH server, nothing else). Then I would recommend to just add the CUDA repository and pull in for example “cuda-drivers” meta-package. That will also pull in lot of unnecessary stuff, but will keep things up to date and doesn’t force graphical desktop. Alternatively, more minimal by installing “nvidia-dkms-470” or “nvidia-driver-470”, but you’d need to watch out for major version updates manually.
How’s your GPU load? Which GPU is it? Name doesn’t fit in the screenshot of nvidia-smi.
Gonna have to get the thinking cap on libhsakmt.so.1: cannot open shared object file:
This error has came up after a fresh install of ubuntu server 20.04 and following the guide for hqplayerd
All dependencies are in after “sudo apt install -f”, including the ROCm stuff?
AMD is changing their packaging stuff on each release, but they don’t always seem to automatically include their library path in ld.so’s search path, so manually doing this will help:
This works as long as they establish symbolic link (through alternatives) from the version specific directory to /opt/rocm, which they have been doing so far.
I think I may have got sorted. I have done a full fresh install. I still managed to run into a couple of hiccups but nothing too brain taxing.
I tried to copy all the commands and steps as I took them for future ref.
After the cuda-drivers install xorg was still installed. This command prevents the server from going to sleep systemctl mask sleep.target suspend.target hibernate.target hybrid-sleep.target
This process did still install xserver when installing the cuda drivers, I found a command afterwards which may fix this sudo apt-get install --no-install-recommends cuda-drivers
Also for some odd reason the soundcard wouldn’t show up on alsa? To fix this I ran these commands:
And a look at what nvidia-smi looks like with a track playing through roon → HQPlayer:
And my signal path looks like this:
All playing with no glitches.
But I am still not happy lol. I’m not completely convinced I am using the full potential of cuda?
I will be doing another fresh install using the command sudo apt-get install --no-install-recommends cuda-drivers to see if it installs without all the xorg guff. But will also want to absolutely happy i’m using the full potential of the GFX card.
Well I found a nice wee tool called nvtop which seems to show the usage of my GPU.
It would seem playing a 192/24khz file with the same filters shown a few posts back the gpu is only running at around 35%.
So… Is it possible to increase this usage. Eg. offload more cpu usage to cuda?
Swapped about with the filters and can bring the gpu up to 100% so its working perfect!
I have noticed that some filters seem to use more cpu than gpu so time for a bigger cpu now that I am certain I am getting the full potential from my gpu.
Thanks @jussi_laako
If you have latest HQPlayer, then the driver is too old. In the nvidia-smi it shows “CUDA Version: 11.8” while HQPlayer requires driver that supports latest CUDA which is 12.1. This means driver >= 530
Easiest way to handle this is to add the Nvidia’s CUDA repository according to instructions here.
Then just install package called “nvidia-headless-530” this should pull in drivers needed for headless server operation.
Note that they don’t have a generic headless package, so once HQPlayer is updated with new CUDA version again, then you need to manually look up the latest headless package and move over to a new version.
Nvidia seems to have a tight pairing between the CUDA stack and their driver. Generally applications built with older CUDA SDK work with newer drivers, but not vice versa.
If the driver is not new enough the CUDA stack doesn’t get initialized and it just reports that there are no GPUs available. In such case HQPlayer just falls back to the regular CPU code.
I now managed to get to 12.0 using unstable channels for NixOS, but I still got the same results.
±----------------------------------------------------------------------------+
| NVIDIA-SMI 525.89.02 Driver Version: 525.89.02 CUDA Version: 12.0 |
|-------------------------------±---------------------±---------------------+
| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|===============================+======================+======================|
| 0 NVIDIA RTX A4000 Off | 00000000:01:00.0 Off | Off |
| 41% 34C P8 12W / 140W | 4MiB / 16376MiB | 0% Default |
| | | N/A |
±------------------------------±---------------------±---------------------+
±----------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=============================================================================|
| 0 N/A N/A 2082 G …hell-43.3/bin/gnome-shell 3MiB |
±----------------------------------------------------------------------------+
This “sse” was a good pointer. Manually switched to non-sse one and everything is working! What does the “sse” stand for? I didn’t find anything with quick googling.
EDIT: I can now listen to gauss-xla on PCM192->DSD1024. I’m a super happy camper!
