The visual profiler gives me several lines of the foll;owing form when I ask it to analyse a kernel for Uncoalesced Global Memory.
“Line: xxx Global Load L2 Transactions/Access = 96.0 [3686400 L2 transactions for 38400 total executions]”
What do the numbers mean? I have a hard time accounting for any of them. I know I can reduce the number of transactions by reducing the number of global memory reads, which makes it faster. But, is it also telling me something actionable about cache hit rates or non-coalescing of reads?