Fill-down kernel benchmark
Evaluates one formula filled down N rows three ways: the tree-walking interpreter, the compiled JS kernel (f64, bit-identical to the interpreter), and a WebGPU compute shader (f32). Everything runs in this tab.
| formula | method | rows | median ms | M rows/s | vs interpreter | notes |
|---|
Method
- Inputs A, B, C are random positive doubles (B never 0), so every row takes the fast path.
- Each method runs once to warm up, then runs times; the median is reported.
- The interpreter is capped at 2M rows (it's slow); its rows/s is measured on that slice.
- JS kernel results are checked against the interpreter row-by-row on the interpreted slice (expect 0 mismatches).
- GPU timing is end to end: f64→f32 conversion, upload, compute and readback. "compute" in notes is the dispatch alone.
- GPU error is the max relative difference vs the f64 kernel.