Commits on Nov 4, 2024

1. Fix incorrect outputs and improve performance of commonMemSetLargePat… …

   …tern
   
   Change the implementation of commonMemSetLargePattern to use the largest
   pattern word size supported by the backend into which the pattern can be
   divided. That is, use 4-byte words if the pattern size is a multiple of 4,
   2-byte words for even sizes and 1-byte words for odd sizes.
   
   Keep the idea of filling the entire destination region with the first word,
   and only start strided fill from the second, but implement it correctly.
   The previous implementation produced incorrect results for any pattern size
   which wasn't a multiple of 4. For HIP, the strided fill remains to be always
   in 1-byte increments because HIP API doesn't provide strided multi-byte
   memset functions like CUDA does. For CUDA, both the initial memset and the
   strided ones use the largest possible word size.
   
   Add a new optimisation skipping the strided fills completely if the pattern
   is equal to the first word repeated throughout. This is most commonly the
   case for a pattern of all zeros, but other cases are possible. This
   optimisation is implemented in both CUDA and HIP adapters.

   

   rafbiels committed Nov 4, 2024
   Configuration menu

   • View commit details
   • Copy full SHA for cc52821
   • Browse repository at this point

   Copy the full SHA

   cc52821 View commit details

   Browse the repository at this point in the history