* Reading from and writing to global memory is generally very expensive. * It often involves copying data across an off-chip bus. * This means you generally want to avoid unnecessary accesses. * Memory access operations is done in chunks. * This means accessing data that is physically close together in memory is more efficient.







* Coalescing global memory access is particularly important when working in multiple dimensions. * This is because when doing so you have to convert from a position in 2d space to a linear memory space. * There are two ways to do this; generally referred to as row-major and column-major.




* Another area this is a factor is when composing data structures. * It's often instinctive to have struct representing a collection of data and then have an array of this - often referred to as Array of Structs (AoS). * But for data parallel architectures such as a GPU it's more efficient to have sequential elements of the same type stored contiguously in memory - often referred to as Struct of Arrays (SoA).












Code_Exercises/Coalesced_Global_Memory/source

Try inverting the dimensions when calculating the linear address in memory and measure the performance.