[cctalk] Re: Intel's i860, Cray-On-A-Chip

On Fri, 23 Sep 2022, emanuel stiebler via cctalk wrote: I can't help with your quest, however I wish you both luck and the energy to preserve your old grey cells when dealing with these beasts. Back in the day I did some work on the i860. Fascinating chip. Very capable of doing high speed vector math and multiply + add in a single cycle. As long as you were prepared to sacrifice the old grey cells to it's dual-instruction mode programming. I didn't do anything in C that I recall, all assembler (I was involved mostly in test/diagnostics and low-level accesses - we had them on a shared memory system with Transputers being relegated (by then) to nothing more than "smart" comms chips... (Late 80s) I think we used the Portland Group C and FORTRAN compilers, but even then we had large librarys hand-coded in assembler to achieve that fabled 3 instructions per cycle quote - you have to effectively pump the floating point pipeline by hand so you ran operations that assembled to a 64-bit word which was one 32-bit instructions for the integer unit (say a load from RAM into FPU pipeline) and one for the FPU (say, a multiply and add instruction). So some 8 cycles later you'd actually start to get results out as it took 8 cycles to actually do the multiply and add - however it was one cycle after that - good for vector work, but sub-optimal for a single number multiply. And lets hope you never, ever, have to take an interrupt during an FP operation. It's 100's of cycles to save/restore the pipeline. I think there was a unix/unix-like OS for them, but I imagine context switching was slow... Cheers, Gordon