After implementing a performant prefix-sum in Vulkan, I extended the work to WebGPU by designing high-performance WGSL shaders inspired by my earlier OpenCL kernels. I refined these implementations through in-browser benchmarking, maximizing scan throughput while addressing the significant portability challenges inherent to WebGPU (github). More details are available in my honors thesis.

Developed a Vulkan prefix-sum GPU kernel which achieves up to 43% higher throughput than Nvidia's proprietary implementation of prefix-scan (CUB) on small inputs. Additionally, achieves performance within 1% of Nvidia CUB on RTX 4070 and 1.5% on AMD 7900 XT, relative to device throughput limits at the largest input sizes (github). More details can be found in my honors thesis.

Reimplemented memory management in our lab's open-source Vulkan simplification tool to use device-local buffers, enabling accurate GPU benchmarking. This facilitated efficient performance characterization of Vulkan GPU kernels and contributed to the development of the Vulkan prefix-scan (main PR).