texture memory or shared memory? which is faster, and by what factor?

For read of threads in half warp (read address is consecutive, each thread reads one float), is it necesary to move data from texture memory to shared memory for performance? I’m not sure how fast shared memory compared to texture memory cache.