Honored Contributor
Honored ContributorI recommend instrumenting your kernel with macros to control things like SIMD lanes, number of compute units, loop unrolling, etc... and pass them in at compile time to implement the same type of optimizations that the -O3 flag did automatically. If you script your compile then you could do your own exploration turning various knobs to see what improves the performance in an automated way.
Even when -O3 was supported I found it was easier to do my own exploration since it's hard to tell which optimizations work well and which ones do not when the compiler is turning all the knobs for you.
