pcmt/bitset.py

from __future__ import annotations

import time
from dataclasses import dataclass
from typing import Iterable

import numpy as np

from .schema import Window


@dataclass(frozen=True)
class EngineInfo:
    requested: str
    actual: str
    torch_available: bool
    cuda_available: bool
    device_name: str | None = None


def resolve_engine(requested: str = "auto") -> EngineInfo:
    requested = requested.lower()
    try:
        import torch

        cuda = bool(torch.cuda.is_available())
        name = torch.cuda.get_device_name(0) if cuda else None
        if requested in {"cuda", "gpu"} and cuda:
            return EngineInfo(requested=requested, actual="cuda", torch_available=True, cuda_available=True, device_name=name)
        if requested == "auto" and cuda:
            return EngineInfo(requested=requested, actual="cuda", torch_available=True, cuda_available=True, device_name=name)
        if requested in {"cuda", "gpu"} and not cuda:
            return EngineInfo(requested=requested, actual="cuda-requested-fell-back", torch_available=True, cuda_available=False)
        return EngineInfo(requested=requested, actual="cpu", torch_available=True, cuda_available=cuda, device_name=name)
    except Exception:
        if requested in {"cuda", "gpu"}:
            return EngineInfo(requested=requested, actual="cuda-requested-fell-back", torch_available=False, cuda_available=False)
        return EngineInfo(requested=requested, actual="cpu", torch_available=False, cuda_available=False)


def atom_matrix(windows: list[Window], atoms: Iterable[str], engine: EngineInfo):
    atom_list = list(atoms)
    matrix = np.zeros((len(windows), len(atom_list)), dtype=bool)
    for row, window in enumerate(windows):
        present = window.atom_set()
        for col, atom in enumerate(atom_list):
            matrix[row, col] = atom in present
    if engine.actual == "cuda":
        import torch

        return torch.as_tensor(matrix, device="cuda"), atom_list
    return matrix, atom_list


def within_radius(mask, radius: int):
    if radius <= 0:
        return mask
    if hasattr(mask, "device") and str(mask.device).startswith("cuda"):
        import torch

        out = mask.clone()
        for shift in range(1, radius + 1):
            out[shift:] |= mask[:-shift]
            out[:-shift] |= mask[shift:]
        return out
    out = mask.copy()
    for shift in range(1, radius + 1):
        out[shift:] |= mask[:-shift]
        out[:-shift] |= mask[shift:]
    return out


def to_numpy_bool(mask) -> np.ndarray:
    if hasattr(mask, "detach"):
        return mask.detach().cpu().numpy().astype(bool)
    return np.asarray(mask, dtype=bool)


def stress_check(windows: list[Window], engine: EngineInfo, repeats: int = 1) -> dict[str, float | int | str]:
    atoms = sorted({atom for window in windows for atom in window.atom_set()})
    start = time.time()
    matrix, _ = atom_matrix(windows, atoms, engine)
    checksum = 0
    if engine.actual == "cuda":
        import torch

        work = matrix
        for _ in range(max(1, repeats)):
            work = torch.logical_xor(within_radius(work, 1), torch.roll(work, shifts=1, dims=0))
            checksum += int(work.sum().detach().cpu().item())
        torch.cuda.synchronize()
    else:
        work = matrix
        for _ in range(max(1, repeats)):
            work = np.logical_xor(within_radius(work, 1), np.roll(work, shift=1, axis=0))
            checksum += int(work.sum())
    elapsed = time.time() - start
    return {
        "engine": engine.actual,
        "atom_count": len(atoms),
        "window_count": len(windows),
        "repeats": repeats,
        "checksum": checksum,
        "elapsed_seconds": elapsed,
    }