You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Originally reported by: edanor (Bitbucket: edanor, GitHub: edanor)
Because we want to explicitly specialize only operations that can be expressed in specific instruction set, all unsupported operations should be left with resolution to scalar emulation.
Most of emulated functions are using insert/extract on a per-element basis. Performing insert/extract operations on vectors is slow for non-emulated vectors. Doing load/store on emulated data types shouldn't create any slow-downs due to compiler optimizations, and even if not, it will still reduce slow-down on target vector code.
This proposal is to re-write scalar emulation functions with LOAD/STORE instead of INSERT/EXTRACT.
Originally reported by: edanor (Bitbucket: edanor, GitHub: edanor)
Because we want to explicitly specialize only operations that can be expressed in specific instruction set, all unsupported operations should be left with resolution to scalar emulation.
Most of emulated functions are using insert/extract on a per-element basis. Performing insert/extract operations on vectors is slow for non-emulated vectors. Doing load/store on emulated data types shouldn't create any slow-downs due to compiler optimizations, and even if not, it will still reduce slow-down on target vector code.
This proposal is to re-write scalar emulation functions with LOAD/STORE instead of INSERT/EXTRACT.
The text was updated successfully, but these errors were encountered: