Code for the project course on tvm. The goal of the course is to get familiar with tvm first. Then try to implement a new optimization for nvidia gpu using it and do some benchmarks from time, power consumption, and memory footprint perspectives. For detailed information about what I have done, refer to the report.
- byoc: Bring Your Own Codegen study, where I successfully use it to offload computation to ncnn, a lightweight neural network inference engine designed for mobile devices. You can check the implementation here
- cmake: my cmake file backup for build tvm with various support
- codegen: codegen stuff
- cu_prog: cuda programs
- docker: dockerfile and scripts for setting up development environment
- ncnn: demo showing the usage of ncnn for later BYOC intergrating
- schedule: benchmark different strategies in tvm for optimizing mobilenet_v2 and yolov8
- relay: relay related learning materials
- howto: scripts from official docs
- tvmcon2023: scripts from tvm conference 2023