pub fn reference_eval(
bp: &TensorWasmKernelBlueprint,
inputs: &[u8],
) -> Result<Vec<u8>, ReferenceEvalError>Expand description
Interpret a blueprint over inputs (treated as a little-endian
f32 stream) and produce the corresponding little-endian f32 output
stream. This is the ground truth the proptest harness asserts
the GPU path against.
The interpreter mirrors the PTX emitter’s offset accounting in
[crate::ptx_emit::lower_body]: every LoadUnified { lanes }
reads lanes consecutive f32s from the input cursor; every
StoreUnified { lanes } writes lanes consecutive f32s to the
output cursor; arithmetic ops pop two (or three for FMA) values
from the value stack and push one result per lane.