The -cl-fast-relaxed-math compiler option is the most general and powerful among other performance related options. Notice that the option affects the compilation of the entire OpenCL program, so it does not permit fine control of the resulting numeric accuracy. You may want to consider experimenting with native_* equivalents separately for each specific built-in instead, keeping track of the resulting accuracy. Please find more details on this approach in the “Considering native_ and half_ versions of Math Built-Ins” section.
Refer to the User Manual - OpenCL™ Code Builder for the list of compiler options for the specific optimizations.
Intel® CPU and Intel® Graphics devices support different sets of options.