How to use local video card memory and is it faster than shared?

Question

GPGPU

There are identifiers for shared and constant memory.
Is it true that variables declared inside the kernel use local memory? And what is its speed compared to shared and constant?

Answer 1

Local and global memory is the slowest. Register is the fastest, variables declared inside the kernel are supposedly stored in it. shared memory is divided into a pool of threads (warp in the context of CUDA).
www.microway.com/hpc-tech-tips/gpu-memory-types-pe...
The hardest thing to do at first is to achieve transactions when copying memory from global to shared, so that the data is copied in blocks.