Question 01:
Does the use of data storage in constant memory provides any benefit over texture in the Pascal architecture if the data request frequence is very high among threads (every thread pick at least one data from a specific column)?
Question 02:
If every thread in a warp request in the very same instruction data from constant memory, but every every thread request it from a different address (no broadcasting), the performance will be worse than the texture memory?
Question 03:
Does the cache for constant memory and texture have the same performance? Do they both have similar dimensions? Can I manipulate the cache dimension for constant or texture on Pascal architecture?
Background:
There is a portion in a kernel that is sure to request some float data from a specific column of a 2D matrix in the texture memory, usually, every thread request the data stored in adjacent cell "below" the requested data from the previous thread.
The frequence of which two threads in the same warp request the very same address if bellow 25% for most warps, but can reach 50% in a less than 10% of the warps.
It is sure that this specific column fits cmem for 95% of the cases.