badnikhil/xpose

Xpose is a C and C++ toolkit that lets Android phones, and desktop computers, run heavy GPU calculations with an experience similar to CUDA, the toolkit Nvidia graphics cards use for this kind of work. Normally, Android phones do not have anything like CUDA available, so getting a phone's graphics chip to do general purpose computing means writing a lot of low level code using Vulkan, the graphics standard most modern phones support, or using another heavyweight tool that still asks the programmer to manage a lot of technical setup. Xpose tries to remove that pain. With Xpose, you write your GPU program, called a kernel, using a C like programming language called OpenCL C. That kernel gets compiled ahead of time into a format called SPIR-V using an existing open source tool. Then Xpose's own runtime, a small piece of code that ships with your app, loads that compiled kernel, works out how to connect its inputs and outputs automatically, and lets you run it with a single line of code instead of hundreds of lines of manual setup. The project pays special attention to a real problem with phone hardware: unlike desktop graphics cards, many phone GPUs use a type of memory that needs to be manually kept in sync between the processor and the graphics chip. Xpose is built and tested on real phone hardware to handle this correctly, aiming for zero copy transfers when possible instead of unnecessary copying. The README demonstrates two example workloads running on a real phone: a small AI language model responding at real time speeds entirely on the phone's GPU, and a very large physics simulation running much faster on the GPU than on the phone's regular processor. Xpose is a small library, no separate Vulkan development kit installation is required, and it aims to work across Android, Linux, and desktop GPUs from major manufacturers, as well as running on the CPU as a fallback. It was built by a team of three during a hackathon hosted by Qualcomm in July 2026.