The reader inflated a whole deflated chunk (4 MB for a 1 Mpx int32 frame, 2 MB for its uint16 per-frame
mask) into a buffer and then converted it in a second pass; with 16 reading threads those buffers never
stay in the cache, so the conversion ran at memory speed. A deflated, unshuffled chunk is now inflated in
128 kB pieces (InflateInPieces, the zlib stream API - the stream's checksum is checked as before), and
each piece is converted into the image, or marked in the mask, before the next is inflated. Shuffled,
bitshuffle/LZ4, HDF5-read and single-panel (placed) frames are converted as a whole, as before.
Bit-identical: the FNV pixel hashes of 2000 frames on the five reader configurations of the previous
commit are unchanged, and a new test reads 300 x 300 frames (several pieces, a final partial one) with a
per-frame mask and a gain pixel by pixel. Reader alone, 16 threads, 3000 frames: 2.43 -> 1.73 s (the
previous commit's state -> this). rugnux on the 14,445-frame CXI run, against rc178 (A/B, two runs each):
geometry refinement sample 11.6 -> 5.1 s, image pass 27.7 -> 12.9 s, whole run 67.7 -> 46.2 s; the
geometry refinement, module placement and tilt log lines are identical.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015CeEGgpKdfkCTJb3YNMJcu
leonarski_fandClaude Opus 5.5
f5a990dc8d