Hi
I have some question about the instructions listed in CUDA Binary Utility page:
1- Some uniform datapath instructions are indeed integer instructions (LEA vs. ULEA). So, why they are separated? Does that mean profiling tools don’t count ULEA as integer instruction?
2- Some instructions from different categories looks similar. For example, B2R looks like a MOV instructions. Does profiling tools, e.g. nsight-compute rely only on the same table for categorization? I ask that because I want to manually process the opcode_histogram from Nvbit.
My suggestion is to ask profiling questions like this on one of the profiler forums (e.g. perhaps Nsight compute forum).
The uniform datapath instructions and mechanism are not well documented by NVIDIA, which is generally true for many things that manifest at the SASS level. However the uniform datapath is essentially an alternate computation path, that some newer GPUs can use in certain circumstances, to achieve additional throughput. The instructions for the uniform datapath are separated because they are separate instructions that must be compiled as such by the compiler, to indicate alternate datapath usage/scheduling. An ordinary instruction would not be processed by the uniform datapath mechanism.
Thanks for the clarification.