Writing and Running Kernels¶
Document status
Owner: CL Team · Baseline: Legato 0.14.1+public, PyTorch 2.10.0+cpu, Bertha EVT0 · Last verified: 2026-10-19
Exercises for writing Legato kernels and compiling and running them in legato.session(...).
The examples load inputs, compute on cores, store outputs, and compare them with CPU reference values.
Prerequisites¶
- A Bertha LPU server with the SDK installed.
- An environment with Legato 0.14.1, PyTorch 2.10,
torch_ha, and the HW runtime. - Target: Bertha EVT0, 32 cores, BF16. Each example uses a fixed shape.
Save and run the Python files in the SDK environment. The first example uses core 0 only.
| Order | Example | Topics |
|---|---|---|
| 1 | Vector addition | Kernel definition, device memory, compilation, execution, and result checking |
| 2 | Matrix multiplication | MPU weight layout and matrix multiplication |
| 3 | Multi-core addition | Tile partitioning and per-core outputs |
Each page provides a standalone executable file. Changes to shape, dtype, or core count require revalidation of memory layout and results. See the Compiler documentation for the full API.