What's in this chapter?
In the last chapter, you studied how threads are executed on GPUs and learned how to optimize kernel performance by manipulating warp execution. However, kernel performance cannot be explained purely through an understanding of warp execution. Recall in Chapter 3, from the section titled “Checking Memory Options with
nvprof,” that setting the innermost dimension of a thread block to half the warp size caused memory load efficiency to drop substantially. This performance loss could not be explained by the scheduling of warps or the amount of exposed parallelism. The real cause of the performance loss was poor access patterns to global memory.
In this chapter, you are going to dissect kernel interaction with global memory to help understand how those interactions affect performance. This chapter will explain the CUDA memory model and, by analyzing different global memory access patterns, teach you how to use global memory efficiently from your kernel.