I’m well aware of what the major use for this library is. For my particular needs I don’t even care about Java computation performance, I do almost everything with OpenCL. I just wanted to share my findings for people that may see this library and think all their performance problems are magically gone, just because there’s no bounds checks with Unsafe or whatever. Since computation has to happen eventually, there may be a temptation to do it on mapped data, which may lead to the performance gotchas I described.
Indeed: but look at it from this perspective - using mapped objects is so much vastly more simple and less error-prone than directly peeking and poking into ByteBuffers that simply getting it to roughly the same speed is a massive win. If by some miracle it goes twice as fast or whatever - that’s just gravy!
Cas 