NEWS
readNSx 0.2.0 (2026-11-05)
NEV data packets are now read in chunks, and the packet types of a whole
chunk are checked at once. Excluded events (by default spike) no longer cost
any parsing, and memory use no longer grows with the file size, so importing
a NEV file that stores the waveform of every spike is many times faster
- The waveform of each spike is parsed for a whole chunk at once and written to
waveforms.h5 while reading
- Fixed the spike waveform being cut to its first 4 (
NEV 2.2 and 2.3) or 6
(NEV 3.0) samples. The waveform data set in waveforms.h5 is now a
matrix with all samples of each spike, one column per spike (zero-padded when
electrodes store samples in different sizes); it used to be a vector
- Fixed an error when importing spikes from a recording that has no spikes
- Added
read_nev_header() to read the NEV headers, such as the channel
labels, without reading the data packets
- Importing
NSx files that store one data point per data block (recordings
with precision time protocol, PTP, time-stamps) no longer takes days: data
points are read and written to HDF5 in batches instead of once per data
block and channel. Partitions are found while scanning the data block
headers, so memory use no longer grows with the length of the recording
- In those files, clock jumps that cancel out within a frame of
readNSx.packets_per_frame data blocks, and backward jumps, no longer split
a continuous recording into several partitions. As in NPMK openNSx, a
forward jump of more than readNSx.max_tick_multiple sampling intervals
still starts a new partition. In other NSx files, each data block is one
partition: unchanged for files with a 30 kHz time-stamp clock, while files
with nanosecond time-stamps used to merge data blocks whose time-stamps lined
up. If such a file's time-stamps do not tick at the header's
time_resolution_timestamp (more than half of its data points would start a
new partition), the import stops with an error naming that field instead of
writing one partition per data point
NSx time-stamps are converted to seconds with the file's
time_resolution_timestamp instead of always as nanoseconds. For files with
a 30 kHz time-stamp clock, the relative_time of each partition used to be
33,333 times too small and now agrees with the NEV event times. A
resolution of 2000 or less is treated as 30000. The relative_time stored
with each channel keeps full precision; it used to be rounded to 4 decimals
(up to 1.5 samples at 30 kHz). Imports made by earlier versions keep the
old values until they are imported again
- A data block whose header declares more data points than the file holds is
read up to its last complete data point instead of failing. A warning names
the file when it ends early, has a corrupt data block header, or holds no
data points. A partition can now hold more than 2^31 - 1 data points
- Invalid values of the options
readNSx.chunk_size,
readNSx.packets_per_frame and readNSx.max_tick_multiple give an error
naming the option before anything is written; readNSx.chunk_size = 0 used
to loop forever
- Fixed results of the
C++ readers (NSx scanning, the legacy NSx reader,
and BCI2000 parameters) being exposed to garbage collection before they
were returned, which could give wrong values, hangs, or crashes
- Writing
HDF5 data to an existing file that is not HDF5 now gives an error
and leaves the file alone. The file used to be replaced with an empty HDF5
file without warning, even when only checking whether it could be written. An
empty file is still replaced, so importing again repairs one left behind by
an interrupted import
- Fixed the
WebAssembly build, which has failed since HDF5 support moved to
hdf5lib, leaving webR users on version 0.0.7, which cannot read HDF5
data there. WebAssembly builds now link the HDF5 library of the webR
build tools
readNSx 0.1.0 (2026-09-02)
HDF5 is now read and written directly by readNSx, linking against
hdf5lib for a static build of the HDF5 library. This removes the
dependency on hdf5r and the HDF5 system requirement; no system HDF5
installation is needed on any platform
- Removed the
hdf5r code path and the plain-R .ralt fallback store; there
is now a single storage implementation. The layout on disk is unchanged, so
files exported by earlier versions read back unchanged and files written now
stay readable by hdf5r, h5py, and command-line HDF5 tools
- Channel datasets are no longer capped at
2^31 samples: dimensions and
offsets are carried as doubles all the way into HDF5
- Reading a contiguous range from a channel now reads only that range instead of
the whole dataset
h5_valid(file, "w") no longer truncates the file it is asked about
- Added complex support, stored as the
H5T_COMPOUND {"r", "i"} of doubles that
numpy and h5py use for complex128. Complex datasets written elsewhere are
read back as complex whatever their member names, including the
{"Real", "Imaginary"} layout hdf5r uses and HDF5 2.0 native complex
- Strings are now written as
UTF-8 rather than being labeled ASCII
regardless of content, so non-ASCII text survives. For plain ASCII the
stored bytes are unchanged. Strings are read using whatever character set the
file declares, so files written by earlier versions are unaffected
NA_character_ is now stored as an HDF5 NULL and read back as NA.
Previously it became the literal text "NA", which no reader could tell apart
from a genuine "NA" string
NA in integer, double and logical vectors round-trips exactly, matching what
hdf5r stores bit for bit
readNSx 0.0.7 (2026-05-23)
import_nsp Supports streaming the data with lengths larger than the bound of 32-bit integers. This supports high-frequency long-time continuous recording, extending the 30,000 Hz recording duration from 10 minutes (100 channels) to Almost a half day recording. The individual channel size is still limited by the 32-bit integer, that is around 19 hours
- Fixed an
rchk issue reported by CRAN
- Replaced
c.integer64 with c
readNSx 0.0.6 (2025-10-24)
HDF5 falls back to alternatives in WASM to avoid compilation issues when running in web browser
readNSx 0.0.5 (2024-09-03)
- Fixed calling
R_MakeExternalPtr with Rf_mkChar, using Rf_install instead
- Fixed a potential imbalanced protect
readNSx 0.0.4 (2024-02-24)
- Fixed a
ASAN/UBSAN error in read_bci2000
readNSx 0.0.3 (2024-02-17)
- Added
read_bci2000 to read BCI2000 data
readNSx 0.0.2 (2023-06-21)
- Added support:
NEV/NSx 3.0 specification
- Corrected
NEV 3.0 file bug. In some cases, the acquisition software may incorrectly set "Time Resolution Time-stamp", resulting in wrong event time-stamps
readNSx 0.0.1 (2023-02-01)
- Added a
NEWS.md file to track changes to the package.