OpenCL-Docs
OpenCL-Docs copied to clipboard
Math optimization defaults
The API registry says at 5.8.6.3. Optimization Options that
These options are not turned on by default since it can result in incorrect output...
Specifically referring to fmad optimization I found this is not entirely true since there also is 6.15.2.1. Floating-point macros and pragmas in the C registry.
I found that on many hardware FP_CONTRACT is on by default, which enables fmadby default even without the compiler option -cl-mad-enable beeing set. This may be clarified.
I've moved this issue from OpenCL-Registry to OpenCL-Docs. It should get more attention here.