Does access constant memory is faster than global memory and how fast ? If the access latency is the same , why do we need to use the constant memory ? Thanks.
Constant Memory is faster because the constant memory cache is implemented onchip. Global memory resides offchip so it's slower. Note that the data should ideally fit inside the constant memory. If it doesn't and there's a cache miss, the fetch operation for the constant cache is expensive. To change the size of the constant cache you can use the --const-cache-bytes <N> where N is the number of bytes you want the cache to be
Therefore access constant memory is faster ( > 1x ) than global memory ? Or is there any detail number ? If I didn't use --const-cache-bytes <N> , the default max allocation is 1G (use helloworld to show the infomation) ? Thanks
Yes, constant memory access is definitely faster than global memory. How much faster, I'm not sure on the number. You can probably perform some tests on that.
