MRIFiles: Faster ISMRMRD read via single bulk HDF5 call

[hakkelt]

Summary

• In the unfiltered case (no slice/repetition/contrast filter), replace the per-profile element reads with a single read(dset) call that loads the entire dataset at once.
• Remove the now-redundant separate header-only bulk read that preceded the loop.
• The filtered path is unchanged: it still bulk-reads headers to decide which profiles to keep, then fetches only matching profiles individually.

Why this helps

The ISMRMRD dataset is written with chunkshape = [1] (one HDF5 chunk per profile). The previous code called h["…/data"][m] inside the loop, issuing one HDF5 element read per profile — O(N) separate chunk lookups, each with its own dataspace selection and variable-length heap traversal. On a 186 MB file with 3952 profiles this was the dominant cost.

Benchmark (186 MB OCMR file, 3952 profiles, 3 runs each)

Path	Time (min)	Allocations	Speedup
Original — per-element reads	3.14 s	436 MiB	1×
New — single bulk read	1.26 s	395 MiB	2.5×
Filtered worst case — original	2.71 s	436 MiB	—
Filtered worst case — new	2.48 s	436 MiB	~1.1×

The speedup grows with profile count; larger cardiac/fMRI raw datasets (10k–200k profiles) will see a proportionally larger improvement.

Correctness

All 3952 profiles match the original read bit-for-bit (data and trajectory arrays) in both the unfiltered and filtered paths.

🤖 Generated with Claude Code