qibotn/tools/benchmark_qredtea_svd_controls.py

#!/usr/bin/env python
"""Benchmark qredtea/qtealeaves SVD control modes.

This isolates the tensor split used by MPS updates: a rank-2 tensor is split
with singular values contracted either left or right, then reconstructed to
measure numerical error and timing.
"""

from __future__ import annotations

import argparse
import gc
import statistics
import time

import torch

import qmatchatea
from qredtea.torchapi import QteaTorchTensor


def _dtype(name: str):
    return {
        "complex64": torch.complex64,
        "complex128": torch.complex128,
        "float64": torch.float64,
        "float32": torch.float32,
    }[name]


def _random_matrix(shape, dtype, seed):
    gen = torch.Generator(device="cpu")
    gen.manual_seed(seed)
    if dtype.is_complex:
        real_dtype = torch.float32 if dtype == torch.complex64 else torch.float64
        real = torch.randn(shape, dtype=real_dtype, generator=gen)
        imag = torch.randn(shape, dtype=real_dtype, generator=gen)
        return torch.complex(real, imag).to(dtype)
    return torch.randn(shape, dtype=dtype, generator=gen)


def _sync():
    if torch.cuda.is_available():
        torch.cuda.synchronize()


def run_one(matrix, ctrl, max_bond, contract_singvals, repeats):
    conv = qmatchatea.QCConvergenceParameters(
        max_bond_dimension=max_bond,
        cut_ratio=0.0,
        svd_ctrl=ctrl,
    )
    qtensor = QteaTorchTensor.from_elem_array(matrix, dtype=matrix.dtype, device="cpu")

    times = []
    rel_error = None
    kept = None
    status = "ok"
    error = ""

    for i in range(repeats):
        gc.collect()
        _sync()
        t0 = time.perf_counter()
        try:
            left, right, singvals, _ = qtensor.split_svd(
                [0],
                [1],
                contract_singvals=contract_singvals,
                conv_params=conv,
            )
        except Exception as exc:  # noqa: BLE001 - benchmark should keep going
            status = "error"
            error = repr(exc)
            break
        _sync()
        times.append(time.perf_counter() - t0)

        if i == repeats - 1:
            left_matrix = left.elem.reshape(matrix.shape[0], -1)
            right_matrix = right.elem.reshape(-1, matrix.shape[1])
            recon = left_matrix @ right_matrix
            rel_error = (
                torch.linalg.vector_norm(matrix - recon)
                / torch.linalg.vector_norm(matrix)
            ).item()
            kept = int(singvals.numel())

    return {
        "ctrl": ctrl,
        "contract_singvals": contract_singvals,
        "status": status,
        "median_ms": float("nan") if not times else statistics.median(times) * 1000,
        "min_ms": float("nan") if not times else min(times) * 1000,
        "rel_error": rel_error,
        "kept": kept,
        "error": error,
    }


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--shapes", nargs="+", default=("256x1024", "1024x256", "512x512"))
    parser.add_argument("--max-bond", type=int, default=128)
    parser.add_argument("--dtype", choices=("complex64", "complex128", "float32", "float64"), default="complex128")
    parser.add_argument("--threads", type=int, default=8)
    parser.add_argument("--repeats", type=int, default=3)
    parser.add_argument(
        "--controls",
        nargs="+",
        default=("A", "D", "V", "R", "E", "E!", "X", "X!"),
    )
    args = parser.parse_args()

    torch.set_num_threads(args.threads)
    dtype = _dtype(args.dtype)

    print(
        "svd_benchmark "
        f"dtype={args.dtype} threads={torch.get_num_threads()} "
        f"max_bond={args.max_bond} repeats={args.repeats}",
        flush=True,
    )
    print(
        "columns shape contract ctrl status median_ms min_ms kept rel_error error",
        flush=True,
    )

    for shape_text in args.shapes:
        m_text, n_text = shape_text.lower().split("x", 1)
        shape = (int(m_text), int(n_text))
        matrix = _random_matrix(shape, dtype, seed=sum(shape))
        for contract_singvals in ("L", "R"):
            for ctrl in args.controls:
                result = run_one(
                    matrix,
                    ctrl=ctrl,
                    max_bond=args.max_bond,
                    contract_singvals=contract_singvals,
                    repeats=args.repeats,
                )
                print(
                    f"row shape={shape_text} "
                    f"contract={contract_singvals} "
                    f"ctrl={ctrl} "
                    f"status={result['status']} "
                    f"median_ms={result['median_ms']:.3f} "
                    f"min_ms={result['min_ms']:.3f} "
                    f"kept={result['kept']} "
                    f"rel_error={result['rel_error']} "
                    f"error={result['error']}",
                    flush=True,
                )


if __name__ == "__main__":
    main()