Efficient Mapping of Graphic Software on DSP

• Main Authors: Mody, Mihir, Jayaraj, Ajay, Hariyani, Hemant, Balagopalakrishnan, Anand, Jones, Jason, Narvaez, Erick, Govindarajan, Sriramakrishnan, Shankar, Prithvi, Garud, Hrushikesh
• Format: Conference Proceeding
• Language: English
• Subjects: Computational modeling; Compute Primitives; Conferences; DSP; GLOPS; GPu; Graphics processing units; Jacinto; LLVM; OpenGL; Programming; Rendering (computer graphics); Shader; Software; User interfaces; Vulkan
• ISSN: 2766-2101
• DOI: 10.1109/CONECCT52877.2021.9622613

High Performance GPU is critical to render sophisticated user interfaces and general-purpose compute for image and vision processing in the Automotive and Industrial applications. The paper provides an overview of methods that can be used for offloading of part of GPU software (namely shaders) on DSP keeping traditional GPU software programming model intact to improve overall performance in the system. The methods allow efficient and transparent mapping of GPU shaders to DSP without alterations to existing software model. The proposed solution is prototyped in Jacinoto6 Platform of Texas Instruments. Experiments indicate that multiple shader primitives can be efficiently mapped to DSP achieving performance similar to GPU, with exception of trigonometric primitives. This allows DSP to be used as co-processor for graphics rendering applications. Leveraging compute capabilities of dual core C66x DSP in J acinto6 Platform using proposed solution can improve overall GPU Shader performance by up-to 41 % for different shaders corresponding to different use-cases.