GPU Assist using DSP Pre-processor

摘要

There is an ever increasing need for higher GPU performance to render sophisticated User Interface, latest high end 3D games and general purpose compute (GPGPU) applications. GPU SW programming models such as OpenGL have evolved over decades to cater to the unique mixed pipeline 3D GPU architectures. Due to the sticky nature of GPU SW programming model, leveraging other HW blocks to enhance graphics performance has been a most challenging task for SW architects. System designers have usually responded to the GFLOPS demand by increasing the GPU HW specifications. This paper proposes enhancing GPU performance by leveraging DSP transparently in background without impacting GPU software programming model. The proposed solution consists of multiple novel techniques namely ability to offload vertex shader to DSP, 3 stage pipelined execution and ability to re-use GPU internal pipeline. The proposed solution is prototyped in Jacinoto6 Platform from Texas Instruments. The default GPU spec performance is increased by up-to 41% by leveraging dual core C66x DSP in Jacinto6 Platform using proposed solution for different use-cases. The proposed solution is fully transparent to application software stack. In addition, the solution is directly applicable to any GPU + DSP architecture making it attractive approach for cost optimized solutions.