Thanks again!
The internal {T,ndims} specification does seem to be important. I’m not running BenchmarkTools yet, but that small change did appear to make my test case snappier.
The bottleneck I’m seeing seems to be related to the amount of file I/O I’m having to do, with lots of seeks and reads into an IOBuffer, followed by ProtoBuf.readproto(iob) calls, followed by unpacking samples from UInt32 arrays [using bit-shift .>>, bit-and .&, Base.signed.()] and converting to Complex. Our target system will have 10-Gig ethernet and I’ll probably be reading files from an NFS-mounted drive.
This is fun.