GCC's .cpp files are already compiled with -O2, and here are the NVCC flags I put in the Makefile:

NVCCFLAGS=-std=c++11 --ptxas-options="-v --opt-level 0" -Xcicc -O0 --compile --compiler-options -O2 -gencode=arch=compute_${COMPUTE_CAP},code=sm_${COMPUTE_CAP}

So the C++ files are compiled with -O2 but I completely disabled cicc and ptxas optimization. This resulted in several more CUDA functions being included in the Cubin (versus about 6 when the default optimization level 3 is used).

I want to see if t behaves properly under optimization levels 1 and 2 first. At 20 MKey/s it looks like there's still more work to do though I am aware that I can't push the optimization all the way up, so speed's never going to be as fast as it should be.

(Default block thread and points settings this time)