Each of the Warp Selectors (light orange) selects a single warp from the set of available warps in the SMX core. Once a warp is selected, up to two independent instructions are selected from it by the fetch/decode units (dark orange) for execution. These 8 selected instructions are then mapped to the 8 blocks of execution units (yellow/red/blue).
In an ideal world the 8 instructions selected would be composed of 6 typical ALU operations (yellow blocks), 1 special ALU operation (red blocks), and 1 memory load/store (blue blocks). In reality this may not be the case and thus only a subset of these instructions may be executed resulting in lower than ideal occupancy.
A summary of an execution cycle:
Each of the Warp Selectors (light orange) selects a single warp from the set of available warps in the SMX core. Once a warp is selected, up to two independent instructions are selected from it by the fetch/decode units (dark orange) for execution. These 8 selected instructions are then mapped to the 8 blocks of execution units (yellow/red/blue). In an ideal world the 8 instructions selected would be composed of 6 typical ALU operations (yellow blocks), 1 special ALU operation (red blocks), and 1 memory load/store (blue blocks). In reality this may not be the case and thus only a subset of these instructions may be executed resulting in lower than ideal occupancy.
This comment was marked helpful 0 times.