Performance

This page reports how much memory a sampled organ needs, how long it takes to open, and what each of Masterpiece’s engine settings gains. Every figure is measured on a real sample set, and the method is at the end.

Summary

The baseline is the conventional approach: every sample decoded to 32-bit floating point and held in memory.

Test bed

The figures are the project’s own measurements, made with Masterpiece 0.3.3.

Techniques

The settings are on the Loading tab of Organ settings.

The Memory profile list sets several of these at once: Best quality - 24-bit, hold everything, Recommended - 16-bit, stream releases and Smallest - 16-bit mono, stream releases.

Memory

Release streaming is the largest single saving: it leaves 56% of the sample data on disk. The settings multiply, and each step is an exact ratio of the one before:

The exact ratios confirm that the accounting measures the sample data itself.

Memory while loading and playing

Each run opens the organ, then walks the registration from a single Principal 8′ to full organ in eight steps. At each step it plays the same 15-second passage: notes build up to sixteen held voices, then all are released together, so every voice is in its release tail at once.

Memory rises only while the organ loads. After that it stays level through every registration step and every chord, including sixteen simultaneous release tails streaming from disk.

Memory against registration

Every rank is loaded whether its stop is drawn or not, so full organ costs the same memory as a single stop. Peak process memory at each step:

Drawing a stop waits for no samples, and a registration change allocates nothing.

Load time

Holding less also opens faster: a streaming load reads no release tails, and they are more than half the data on disk.

Sample cache

The cache file is 2,439 MB and stands in for about 7 GB of source files.

The cache is keyed to a fingerprint of everything that changes the bytes in memory: the organ definition’s size and modification time, storage width, mono folding, load rate, preload head, loop selection, release streaming and the ranks requested. Any change is a cache miss that rebuilds the cache, so a stale cache is never read. This was verified on a five-rank load: a 16-bit cache is reused at 16-bit, and rebuilt at 24-bit and at 16-bit mono to exactly the expected sizes (291.8, 437.7 and 146.0 MB).

By default Masterpiece keeps a single cache file and replaces it when the organ or settings change, so the cache stays the size of one organ. The Sample cache list offers One cache, replaced as organs change, One cache per organ (uses more disk) and Off.

Sample-rate conversion

Memory falls in exact proportion to the load rate: a 96 kHz sample set played through a 48 kHz device takes half the memory. Verified on Friesach (five ranks) by converting from its native 48 kHz:

Pitch and level are unchanged at every rate. A converted attack joined to a converted streamed release tail matches a fully resident render to 0.0 dB across the whole release.

The Sample rate list offers As recorded, 48 kHz and 44.1 kHz.

Audio fidelity

Measured on the sample data directly, applying exactly what the loader applies, to 60 files chosen at random from the 12,148:

Method

• Held / left on disk. The sample library’s own accounting, logged at the end of each load: residentBytes() sums the sample buffers actually held, streamedBytesSaved() sums what streaming left on disk.
• Process memory. The process’s private bytes, sampled from outside it every 250 ms.
• Load time. From the start of the load to the organ being ready to play, as logged by the application.

Friesach is a freely published sample set by Piotr Grabowski and is not part of this repository. See ATTRIBUTION.md in the repository root.