In the end, simply, what is the best way to play DSD via Roon / HQP> Allo USBridge Sig (dop)> RME ADI-2 ?
For the rest I use HQP in PCM: poly-ins-ext2 / LNS15 / 768 adapative
If I try with my NUC i5, in DSD: it’s max DSD128 with mp-2s, ADSM7 … and I have blanks at times.
However, the CPU does not exceed 65% with 12GB of memory, auto multicore (one blank every 3min approximately).
So PCM in 705/768 is good.
Why would you send PCM to ADI-2 instead of DSD256 in DSD Direct mode?
I’ve been doing DSD256 output from NUC5i5RYH to various DACs for years using similar settings. NUC running first Debian Stretch and later Debian Buster and HQPlayer Embedded. So it should certainly work. All the -2s filters and ext2, with ASDM7 modulator to DSD256 output.
This for all PCM sources. I don’t remember what I’ve had for DSD sources, but if nothing else, at least Direct SDM will work.
Filter is more up to your preferences. At the moment I personally prefer poly-sinc-ext2.
For R2R DAC it would be best to have linearity measurement results for fine tuning. But noise shaped dithers like NS9, NS5 or LNS15 are good, depending on what is maximum sampling rate your R2R can take.
I just got my Chord Hugo Mscaler today. If haven’t done any side by side testing yet but I do find if much better than when I gave HQPlayer sinc-m, LND15 settings a few listens in the past. With HQPlayer and the Dave, the vocals were too bright but I do not find that the case with the mscaler.
At first I thought the mscaler also had an issue with brightness but then realized the video setting was on which I believe is worse quality than the Dave alone.
So HQPlayer is a great product for other dacs but it doesn’t replace the mscaler for Chord dacs.
I’m investigating about my cpu overload (i5 3740T - 12go + SSD)
The problem comes from the DSD + convolution pipeline.
I asked HAF for an acoustic correction (crosstalk convolution) which is then injected via 4 pipelines.
The result is great, wonderful and works well in flac> PCM.
Following the advice, I requested a convolution in maximum resolution : 4 wav files in 768k.
If I read a direct DSD (DirectSDM) without HAF convolution: CPU at 5%
A Flac> 768 + HAF convolution pipeline = 40% HQP (48% overall CPU)
A direct DSD + HAF pipeline: 85/95% overall CPU !! Unable to read
Tests done live on HQP or via Roon.
I will request some 384 wav files in pipelines to try.
Yes, but with a simple convolution (convolution setup), it’s possible.
Yes, it’s only two files (R + L)
So DSD + convolution from the pipeline … not possible except with a war machine ?? i9 ??
Another question or request … if I play a DSD file with HQP in PCM, it is not possible to pass DSD direct (with no ressampling) ?
Currently it is double sampling and CPU overload.
Have you decided to encapsulate HQP in an android application for my Shield 2019?
Maybe a joke but if it were possible, that would make the ultimate HQP server at $ 150 !
If we think about optimal quality and processing, it could have a “direct DSD” option even if we upsample 44> 705, isn’t it stupid to want to keep DSD natively ?
I did all of these listening with HQP 4.3.3 (win10)
HAF convolution is 4 files : HLL, HLR, HRL and HRR.
You can try to limit the length of the files to allow convolution in DSD to work.
Length in seconds = number of samples in IR / sampling rate of IR
If you go below 0,3s this should work, HAF files are typically 0,7s-1,2s long.
I can do this for you if you send me your HAF files.
Thanks alec_eiffel, this is the proposal made to me by Home Audio Fidelity with some short filters … but, it’s the same.
4 pipelines … 84% CPU HQP (92% all processes)
A convolution via the “convolution setup” menu does not have the same impact (HQP 4.3.3)
1.2 second filter at DSD64 rate is about 3 million points. 4x 3 million points is not very light to process. But probably works fine on a powerful GPU or powerful enough CPU.
If you email me your filters and matrix setup, I can test how it runs on my i9-9900KS + RTX2080Ti.
@jussi_laako, quick technical questions to optimise filter lengths…
Do you anyway round up filter length to the nearest power of 2 anyway due to Fast convolution DFT algorithm ? If yes, there is limited interest to reduce filter length from let’s say 50000 taps to 40000 taps as they will be rounded to 65536 and have the same processing cost.