Precompute little-endian sample dtype once per stream

CHANGELOG.md

@@ -1,4 +1,7 @@
 ## [UNRELEASED] - YYYY-MM-DD
+### Changed
+- Precompute the little-endian sample dtype once per stream instead of per sample in `_read_chunk3`, speeding up loading of large numeric streams (~8% faster on a 553k-sample, 99-channel EEG recording) with byte-identical output ([#170](https://github.com/xdf-modules/pyxdf/pull/170) by [Stefan Appelhoff](https://github.com/sappelhoff))
+
 ### Fixed
 - Fix `playback_lsl.py` for cases where no metadata is available in the to-be-streamed XDF file ([#160] by [Stefan Appelhoff](https://github.com/sappelhoff))
 

src/pyxdf/pyxdf.pxd

@@ -7,7 +7,7 @@ cdef class StreamData:
     cdef readonly int nchns, samplebytes
     cdef readonly fmt
     cdef public double effective_srate
-    cdef public time_stamps, time_series, clock_times, clock_values, dtype, fmts
+    cdef public time_stamps, time_series, clock_times, clock_values, dtype, dtype_le, fmts
 
 cdef bytearray _read_varlen_int_buf
 

src/pyxdf/pyxdf.py

@@ -68,6 +68,9 @@ def __init__(self, xml):
         # pre-calc some parsing parameters for efficiency
         if self.fmt != "string":
             self.dtype = np.dtype(fmts[self.fmt])
+            # little-endian view of the dtype, precomputed once because it is used
+            # once per sample in _read_chunk3 (XDF sample data is little-endian)
+            self.dtype_le = self.dtype.newbyteorder("<")
             # number of bytes to read from stream to handle one sample
             self.samplebytes = self.nchns * self.dtype.itemsize
 
@@ -478,9 +481,7 @@ def _read_chunk3(f, s):
             # read the values
             f.readinto(raw)
             # no fromfile(), see https://github.com/numpy/numpy/issues/13319
-            values[k, :] = np.frombuffer(
-                raw, dtype=s.dtype.newbyteorder("<"), count=s.nchns
-            )
+            values[k, :] = np.frombuffer(raw, dtype=s.dtype_le, count=s.nchns)
     return nsamples, stamps, values