precision-binary-search

Insert checkpoint tensors as kernel input parameters at key computation nodes, write intermediate results into them with pypto.assemble or slicing, then compare each checkpoint against the golden function's corresponding return value. Bisect the computation flow until the first mismatching checkpoint identifies the faulty op.

Declare checkpoint tensors as input parameters in the kernel signature with explicit shape and dtype, initialize them with torch.empty() in the test function, and write intermediate results in place using pypto.assemble or view-based slicing. The kernel should not return values; outputs are written via out parameters.

pypto.assemble requires source and destination tensors to have identical dtypes, otherwise compilation fails with a dtype error. Insert an explicit cast before assembling, and when comparing with golden, convert both sides to float32 for consistent comparison.

Create a large tensor outside the loop and use view or pypto.assemble inside the loop to store each iteration's intermediate result. For complex multi-level loops where kernel and golden implementations match, comparing only the in-loop temporary variables is acceptable.

PyPTO's final block can contain dirty data in non-aligned tail cases, causing false assertion failures. Compare the first block's data instead, or ensure shapes are aligned so tail blocks do not introduce invalid values.