GatherScatter

Gatherscatter refers to a class of memory operations used in vector and parallel processing to handle non-contiguous data layouts. In practice, gather operations load elements from memory into a vector register using a vector of indices, while scatter operations write elements from a vector to memory at addresses specified by indices. Some architectures expose gather and scatter as separate instructions, while others provide variants that combine the two or offer masked and indirect addressing options. The term can also describe the broader pattern of indirect memory accesses that underpins these instructions.

Implementation and hardware support vary by architecture. Modern CPUs and GPUs commonly provide gather and scatter instructions as part of their vector or SIMD instruction sets, often with masking to enable selective processing and to handle out-of-bounds indices. The exact semantics—such as whether indices are used in bytes or elements, or how partial vectors are handled—depend on the ISA. Because gather/scatter relies on indirect addressing, performance is highly sensitive to memory layout, cache behavior, and memory bandwidth.

Common use cases include irregular data processing where elements are selected by index rather than by a regular stride. This includes sparse linear algebra (such as sparse matrix-vector multiplication), graph processing, data permutation and reordering, and certain stencil or stencil-like computations where data access patterns vary across a vector lane.

Performance considerations are central. Gather and scatter typically incur higher latency and lower throughput than contiguous loads and stores due to irregular memory access, cache misses, and potential memory bank conflicts. Effective use often requires careful data-layout design, masking strategies, and sometimes software prefetching to mitigate latency and improve cache utilization.