HeadMania UpSampler v2.2.1— Reference-precision filters up to 16 million taps, higher native DSD, and a live performance readout
Support the ongoing development of this project on Patreon: https://www.patreon.com/c/HeadMania
Need help, want to share feedback, or chat with other listeners? Join the Telegram group: https://t.me/+lIFYDf6ZmMk4N2E0
Found a bug or have a feature request? Report it on the issue tracker: https://github.com/danghe/HeadManiaReleases/issues
v2.2.1
This release rolls up everything since 1.35.3 into one update: a new extreme-precision filter tier, instant startup for the big filters, higher native DSD — and now smoother high-rate DSD plus a live real-time performance readout. The precision work is a measured-spec capability, not a new in-band sound claim — judge it by ear on your own gear.
What's new
Reference Precision — a new extreme-precision filter tier. New filters that compute the reconstruction in 106-bit double-double precision at up to 16 million taps — the precision class of PGGB and HQPlayer's 128-bit modes, at and beyond the tap count of Chord's £25,000 Quartet. Available as Reference Linear, Reference Organik, Organik Halfband and Kinetic V3, at 4M / 8M / 16M taps, for file and network (UPnP/DLNA) playback. They're SIMD-accelerated — AVX2 (256-bit) on Intel/AMD, NEON on Apple Silicon — for roughly 2× the throughput of a scalar build on x86 (NEON's narrower 128-bit vectors give a smaller but real gain), which is what lets even the 16-million-tap variants run with real-time headroom to spare (a 16M filter plays comfortably above real time; 4M with room to burn). And they're honest about themselves: at a 24/32-bit transport the result is byte-identical to the standard path — and the SIMD path is bit-for-bit identical to the scalar one — a computed-precision capability, never an audible-quality claim.
Fast Start — the standard million-tap filters begin instantly. The Mega / Ultimate filter family starts playing in tens of milliseconds instead of buffering for several seconds. On by default; switch it off in Settings for a lower-CPU, bursty mode on older machines. The sound is bit-for-bit identical either way.
Higher, smoother native DSD. Native DSD output over ASIO up to DSD1024, with rate-tuned noise shaping for cleaner DSD256 / DSD512 (and DSD2048 on DDCs that generate it). New in this build: the DSD upsampling and the 1-bit modulator now run on separate CPU cores, so the highest DSD rates carry more real-time headroom — DSD1024 in particular is steadier. Identical audio either way; this is about keeping up cleanly, not sound. It engages automatically on machines with cores to spare. DSD is still labelled honestly, never as a fake PCM rate.
A live real-time performance readout. The Debug tab now shows, live, how much real-time headroom each stage has — the filter, the DSD upsampling, and the DSD modulator — and whether DSD is running split across two cores. No setup, no flags: you can see exactly how much margin you have on your own hardware, with any filter and rate.
Honest caveats
The Reference tier is file / server source only. It's too heavy for the real-time live-capture path. If you start live system-audio capture with a Reference preset selected, HeadMania substitutes a live-capable filter for the capture, shows that filter as active in the grid, and tells you — your Reference choice returns the moment you stop capture and play a file or stream from the server.
Reference precision is computed, transport-limited, and inaudible by construction. Every figure sits well below the best-DAC analog wall. It matches the computation precision and tap count of the most extreme upsamplers made, and is never presented as a sound upgrade.
The first time you pick a given Reference filter, it builds its filter once — a few seconds, longer for 16M — then caches it, so every later start is near-instant. On a file or network source the player buffers ahead, so this one-time build stays out of the way.
At the very highest DSD rates (DSD1024), the readout is your friend: a heavier dual-path filter leaves less headroom for the modulator than a single-path one, so if you push DSD1024 with the biggest filters, watch the readout and lean single-path for the steadiest result. The core-splitting auto-engages only where there are enough cores; on smaller machines DSD stays on the proven single-core path. Identical audio either way.
Windows release only. The macOS build is pending its own Clang gate.
Need help, want to share feedback, or chat with other listeners? Join the Telegram group: https://t.me/+lIFYDf6ZmMk4N2E0
Found a bug or have a feature request? Report it on the issue tracker: https://github.com/danghe/HeadManiaReleases/issues