TIRx Harness¶
Give your agent the language, tools, and execution environment to develop kernels for graphics processing units (GPUs). Use the components directly or connect them in an agent loop.
TIRx Harness brings together tirx-lite for kernel authoring, tirx_tools for analysis, and kcoral for remote execution. Skills guide the agent in using them, while workload contracts define correctness and performance. Start with the component or task you need.
How the pieces fit¶
Your agent orchestrates the loop.
Skills guide its choices. tirx-lite produces a kernel; tirx_tools inspects its TIRx function. A benchmark runs the compiled candidate locally or through kcoral. Correctness results, timings, and profiler artifacts inform the next edit.
Choose a component¶
Kernel language & examples
Write kernels in tirx-lite, a domain-specific language over the TIRx intermediate representation. Explore complete implementations and launch examples.
Domain-specific compiler analysis
Check synchronization, memory races, and numerical behavior. Inspect compiler output and generated GPU instructions.
Remote GPU execution
Keep your agent on one machine and run GPU work on another. Use remote execution adapters to run kernels and retrieve diagnostic artifacts.
Choose your next step¶
What you want to do |
Start here |
|---|---|
Optimize a kernel in your own project |
|
Run a registered workload |
|
Define a new optimization task |
|
Publish a kernel produced by a run |
|
Reproduce and resolve a defect |
For hardware and kernel-programming background, see Modern GPU Programming for MLSys. For complete implementations, browse the Kernel Zoo.