gbso/gbso/optimize.py

from dataclasses import dataclass, replace
from math import log
from random import random
from typing import Callable, List, Optional, Tuple

from gbso.program.test_case import Output, TestCase, eq_on_testcase
from gbso.program.mutate import create_random_program, mutate_program
from gbso.program.program import Program

EPSILON = 0.00001

DEFAULT_ANNEALING_CONSTANT = 0.5
DEFAULT_SYNTHESIS_ITERS = 0
DEFAULT_OPTIMIZE_ITERS = 5_000_000
DEFAULT_NUM_CANDIDATES = 1

DEFAULT_PROB_OPCODE = 0.25
DEFAULT_PROB_OPERAND = 0.25
DEFAULT_PROB_SWAP = 0.25
DEFAULT_PROB_INSN = 0.25

DEFAULT_PROB_INSN_UNUSED = 0.1


def cost(
    orig_prgm: Program, test_cases: List[TestCase], outputs: List[Output], prgm: Program
) -> Tuple[float, bool]:
    # Since each instruction executes in 4*k cycles (for some k), this can have
    # the undesirable effect of performance improvements being weighted much
    # higher than correctness. This hurts convergence pretty badly, so we scale
    # by 1/4 to compensate.
    perf = (prgm.perf() - orig_prgm.perf()) / 4.0
    eq = 0

    for test_case in test_cases:
        eq += eq_on_testcase(orig_prgm, prgm, test_case, outputs)

    return perf + eq, eq == 0


def cost_noperf(
    orig_prgm: Program, test_cases: List[TestCase], outputs: List[Output], prgm: Program
) -> Tuple[float, bool]:
    eq = 0
    for test_case in test_cases:
        eq += eq_on_testcase(orig_prgm, prgm, test_case, outputs)
    return eq, eq == 0


@dataclass
class OptimizationParameters:
    max_size: int
    beta: float = DEFAULT_ANNEALING_CONSTANT
    synthesis_iters: int = DEFAULT_SYNTHESIS_ITERS
    optimize_iters: int = DEFAULT_OPTIMIZE_ITERS
    num_candidates: int = DEFAULT_NUM_CANDIDATES
    prob_opcode: float = DEFAULT_PROB_OPCODE
    prob_operand: float = DEFAULT_PROB_OPERAND
    prob_swap: float = DEFAULT_PROB_SWAP
    prob_insn: float = DEFAULT_PROB_INSN
    prob_insn_unused: float = DEFAULT_PROB_INSN_UNUSED
    cost_fn: Callable[
        [Program, List[TestCase], List[Output], Program], Tuple[float, bool]
    ] = cost


# Perform one round of optimization
def _optimize(
    target_prgm: Program,
    test_cases: List[TestCase],
    outputs: List[Output],
    params: OptimizationParameters,
    num_iters: int = DEFAULT_OPTIMIZE_ITERS,
    init_prgm: Optional[Program] = None,
) -> Program:
    padded_prgm = target_prgm.pad(params.max_size)
    if init_prgm is not None:
        padded_prgm = init_prgm.pad(params.max_size)

    last_prgm = padded_prgm
    last_cost, _last_eq = params.cost_fn(target_prgm, test_cases, outputs, last_prgm)

    best_prgm = target_prgm.pad(params.max_size)
    best_cost = 0.0

    num_candidates = 0

    for _ in range(num_iters):
        candidate_prgm = mutate_program(
            last_prgm,
            params.prob_opcode,
            params.prob_operand,
            params.prob_swap,
            params.prob_insn,
            params.prob_insn_unused,
        )
        candidate_cost, candidate_eq = params.cost_fn(
            target_prgm, test_cases, outputs, candidate_prgm
        )

        if candidate_cost < best_cost and candidate_eq:
            best_prgm = candidate_prgm
            best_cost = candidate_cost
            num_candidates += 1

        if candidate_cost < last_cost - log(random()) / params.beta:
            last_prgm = candidate_prgm
            last_cost = candidate_cost

    return best_prgm


def optimize(
    target_prgm: Program,
    test_cases: List[TestCase],
    outputs: List[Output],
    params: OptimizationParameters,
) -> Program:
    print("Synthesizing candidates...")
    candidates = [
        _optimize(
            target_prgm,
            test_cases,
            outputs,
            replace(params, cost_fn=cost_noperf),
            num_iters=params.synthesis_iters,
            init_prgm=create_random_program(params.max_size),
        )
        for _ in range(params.num_candidates)
    ]
    best_candidate = min(
        candidates, key=lambda p: cost(target_prgm, test_cases, outputs, p)[0]
    )
    print("Optimizing...")
    return _optimize(
        target_prgm,
        test_cases,
        outputs,
        params,
        num_iters=params.optimize_iters,
        init_prgm=best_candidate,
    )