Hi,
have seen AMD IL/HW guides and seems evergreen an up support expanded GDS operations for example atomic logic ops (or,and,etc..) and arithmetic (atomic add etc..) so I request improve upon atomic counter support from least effort to more:
(Note response in twitter from Graham Sellers (OGL AMD driver guy) from AMD seems working to expose that):
*at least expose more than 8 atomic counts: >=256 ideal case as these allows me to calculate very fast histograms using an array of 256 atomic counters atomiccounter[inputdata]
*expose as said existing GDS HW support for arithmetic and logic ops
*the general way expose a GDS adress space similar to __global or __local could be __fastglobal
Hope supports for this comes this year
