Complex streamed instructions: introduction and initial evaluation

An architectural paradigm intended to improve the performance of streaming operations is introduced. The proposed complex streamed instructions perform setting, controlling and executing vector operations simultaneously. That is, each of the instructions has the capability of sectioning a vector in addition to performing complex memory accesses and SIMD execution. To provide an initial validation, a video compression application is considered. We identify 5 streamed operations and perform simulations using the SimpleScalar toolset. The experiments show the following: the cycle count diminishes by a factor of 3.1, the number of executed instructions reduces by a factor of 3.0 to 3.2, and the CPI stays about the same. These results show that substantial performance improvements can be expected when complex streamed instructions are employed.