Method log_results

torch/_inductor/select_algorithm.py:894–929

(name, input_nodes, timings, elapse)

    @staticmethod
    def log_results(name, input_nodes, timings, elapse):
        if not (config.max_autotune or config.max_autotune_gemm) or not PRINT_AUTOTUNE:
            return
        sizes = ", ".join(
            [
                "x".join(
                    map(
                        str,
                        V.graph.sizevars.size_hints(
                            n.get_size(), fallback=config.unbacked_symint_fallback
                        ),
                    )
                )
                for n in input_nodes
            ]
        )
        n = None if log.getEffectiveLevel() == logging.DEBUG else 10
        top_k = sorted(timings, key=timings.__getitem__)[:n]
        best = top_k[0]
        best_time = timings[best]
        sys.stderr.write(f"AUTOTUNE {name}({sizes})\n")
        for choice in top_k:
            result = timings[choice]
            if result:
                sys.stderr.write(
                    f"  {choice.name} {result:.4f} ms {best_time/result:.1%}\n"
                )
            else:
                sys.stderr.write(
                    f"  {choice.name} {result:.4f} ms <DIVIDED BY ZERO ERROR>\n"
                )

        autotune_type_str = (
            "SubProcess" if config.autotune_in_subproc else "SingleProcess"
        )
        sys.stderr.write(f"{autotune_type_str} AUTOTUNE takes {elapse:.4f} seconds\n")

    @staticmethod
    def benchmark_example_value(node):