Glossary of rocPRIM terms

Glossary of rocPRIM terms#

This glossary is to help users understand the basic concepts or terminologies used in the rocPRIM library.

Warp#: A group of threads that runs using the single instruction, multiple thread (SIMT) model.
Wavefront#: The AMD-specific term for a warp. rocPRIM uses “warp”. However “wavefront” is also used in AMD documentation.
Hardware warp size#: The number of threads in a warp as defined by the hardware. On NVIDIA GPUs, the warp size is thirty-two (32) threads. On AMD GPUs, the warp size can be either thirty-two (32) or sixty-four (64) threads.
Logical warp size#: The number of threads in a warp as defined by the user. This can be equal to or less than the size of the hardware warp size.
Block#: A group of warps that run on the same streaming multiprocessor (SM). Threads in the block can be indexed using one dimension, {X}, two dimensions, {X, Y}, or three dimensions, {X, Y, Z}.
Stride#: The number of threads per block.
Tile#: The name for a block in C++ AMP. In rocPRIM the tile size is always the same as the block size. This is different than a tile in the NVIDIA CCCL library. In the NVIDIA CCCL library a tile refers to a chunk of data that is processed by a block. This size of the chunk of data can be different than the block size.
Grid#: A group of blocks that all run the same kernel call.
Warp ID#: The identifier of the warp within a block. A warp’s warp ID is guaranteed to be unique.
Thread ID#: The identifier of the thread within a block.
Lane ID#: The identifier of the thread within the warp.
Flat ID#: The flattened block or thread idex. The flat ID is a one-dimensional index created from two-dimensional or three-dimensional indices. For example the flat ID of a two-dimensional thread ID {X, Y} in a two-dimensional 128x4 block is Y*128*X.