I think I read the figures 5% -> 15% on the intel linked site:-
https://software.intel.com/en-us/art...-applications/
(Or maybe it was on StackExchange).
My only experience coding for 64 bit did result in slower 64 bit, but may have been down to ME
Here (WaterMark):-
https://forum.doom9.org/showthread.p...40#post1826340
EDIT: Although it did make it more awkward in my 64bit version as I changed from using RGB blue channel of mask image for the mask (which I did not like) to the YUV luma channel, and in doing so made it necessary to track the mask in YUV luma and apply it to RGB32 R,G,B channels, and so having Y the right way up, and RGB upside down.