Lambda the Ultimate

I think SIMT is going to be added to the traditional CPU, much like SIMD already has been. In any case SIMT already has a foothold in the GPGPU architectures which are SIMT.

SIMT is where a single instruction is executed by multiple threads. It differs from SIMD in that each execution unit has its own register set and control flow, but shares instruction fetch and decode, whereas in SIMD registers and control flow are shared. This needs special care to deal with conditional branches. Generally only if/then/else or switch style conditional branches are allowed and as all instructions are shared by all SIMT units, they have to ignore the instructions in the branch they are not taking. In the case of loops you can have early termination, but all units wait for the last one to exit the loop (effectively instructions still get fed to the units, but they are ignored). This reminds me somewhat of the conditional execution of any instruction in the original ARM instruction set - however I think this was dropped in ARM64 in favour of traditional branches because branch prediction results in faster execution in the (more common) single threaded case.

I am interested in languages and compilers designed to take advantage of SIMT architectures, and that may make efficient programming for such architectures easier for the programmer and compiler technologies that may make things more efficient for the processor to execute. I wanted to start this topic to see what has been done in this field, and collect links to relevant papers. Please post if you know of any, come across any, or have any ideas for how to handle SIMT.