blocksolver 0.8.3__cp310-cp310-macosx_10_9_universal2.whl

blocksolver/blqmr.py

@@ -0,0 +1,1319 @@
+"""
+BLQMR - Block Quasi-Minimal-Residual sparse linear solver.
+
+This module provides a unified interface that uses the Fortran extension
+when available, falling back to a pure-Python implementation otherwise.
+"""
+
+import numpy as np
+from scipy import sparse
+from scipy.sparse.linalg import splu, spilu
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+import warnings
+
+__all__ = [
+    "blqmr_solve",
+    "blqmr_solve_multi",
+    "blqmr_scipy",
+    "blqmr",
+    "BLQMRResult",
+    "BLQMR_EXT",
+    "qqr",
+    "BLQMRWorkspace",
+    "SparsePreconditioner",
+    "DensePreconditioner",
+    "make_preconditioner",
+]
+
+# =============================================================================
+# Backend Detection
+# =============================================================================
+
+BLQMR_EXT = False
+_blqmr = None
+
+try:
+    from blocksolver import _blqmr
+
+    BLQMR_EXT = True
+except ImportError:
+    try:
+        import _blqmr
+
+        BLQMR_EXT = True
+    except ImportError:
+        pass
+
+# Optional Numba acceleration
+try:
+    from numba import njit
+
+    HAS_NUMBA = True
+except (ImportError, Exception) as e:
+    HAS_NUMBA = False
+
+    def njit(*args, **kwargs):
+        def decorator(func):
+            return func
+
+        return decorator if not args or callable(args[0]) else decorator
+
+
+# =============================================================================
+# Result Container
+# =============================================================================
+
+
+@dataclass
+class BLQMRResult:
+    """Result container for BLQMR solver."""
+
+    x: np.ndarray
+    flag: int
+    iter: int
+    relres: float
+    resv: Optional[np.ndarray] = None
+
+    @property
+    def converged(self) -> bool:
+        return self.flag == 0
+
+    def __repr__(self) -> str:
+        status = "converged" if self.converged else f"flag={self.flag}"
+        backend = "fortran" if BLQMR_EXT else "python"
+        return f"BLQMRResult({status}, iter={self.iter}, relres={self.relres:.2e}, backend={backend})"
+
+
+# =============================================================================
+# Quasi-QR Decomposition
+# =============================================================================
+
+
+@njit(cache=True)
+def _qqr_kernel_complex(Q, R, n, m):
+    """Numba-accelerated quasi-QR kernel for complex arrays."""
+    for j in range(m):
+        r_jj_sq = 0.0j
+        for i in range(n):
+            r_jj_sq += Q[i, j] * Q[i, j]
+        r_jj = np.sqrt(r_jj_sq)
+        R[j, j] = r_jj
+        if abs(r_jj) > 1e-14:
+            inv_r_jj = 1.0 / r_jj
+            for i in range(n):
+                Q[i, j] *= inv_r_jj
+            for k in range(j + 1, m):
+                dot = 0.0j
+                for i in range(n):
+                    dot += Q[i, j] * Q[i, k]
+                R[j, k] = dot
+                for i in range(n):
+                    Q[i, k] -= Q[i, j] * dot
+
+
+@njit(cache=True)
+def _qqr_kernel_real(Q, R, n, m):
+    """Numba-accelerated quasi-QR kernel for real arrays."""
+    for j in range(m):
+        r_jj_sq = 0.0
+        for i in range(n):
+            r_jj_sq += Q[i, j] * Q[i, j]
+        r_jj = np.sqrt(r_jj_sq)
+        R[j, j] = r_jj
+        if abs(r_jj) > 1e-14:
+            inv_r_jj = 1.0 / r_jj
+            for i in range(n):
+                Q[i, j] *= inv_r_jj
+            for k in range(j + 1, m):
+                dot = 0.0
+                for i in range(n):
+                    dot += Q[i, j] * Q[i, k]
+                R[j, k] = dot
+                for i in range(n):
+                    Q[i, k] -= Q[i, j] * dot
+
+
+@njit(cache=True)
+def _qqr_kernel_complex(Q, R, n, m):
+    """Numba-accelerated quasi-QR kernel for complex arrays."""
+    for j in range(m):
+        # Quasi inner product: sum(q*q) WITHOUT conjugation
+        r_jj_sq = 0.0j
+        for i in range(n):
+            r_jj_sq += Q[i, j] * Q[i, j]  # No conjugation!
+        r_jj = np.sqrt(r_jj_sq)
+        R[j, j] = r_jj
+        if abs(r_jj) > 1e-14:
+            inv_r_jj = 1.0 / r_jj
+            for i in range(n):
+                Q[i, j] *= inv_r_jj
+            for k in range(j + 1, m):
+                # Quasi inner product: sum(q_j * q_k) WITHOUT conjugation
+                dot = 0.0j
+                for i in range(n):
+                    dot += Q[i, j] * Q[i, k]  # No conjugation!
+                R[j, k] = dot
+                for i in range(n):
+                    Q[i, k] -= Q[i, j] * dot
+
+
+@njit(cache=True)
+def _qqr_kernel_real(Q, R, n, m):
+    """Numba-accelerated quasi-QR kernel for real arrays."""
+    for j in range(m):
+        r_jj_sq = 0.0
+        for i in range(n):
+            r_jj_sq += Q[i, j] * Q[i, j]
+        r_jj = np.sqrt(r_jj_sq)
+        R[j, j] = r_jj
+        if abs(r_jj) > 1e-14:
+            inv_r_jj = 1.0 / r_jj
+            for i in range(n):
+                Q[i, j] *= inv_r_jj
+            for k in range(j + 1, m):
+                dot = 0.0
+                for i in range(n):
+                    dot += Q[i, j] * Q[i, k]
+                R[j, k] = dot
+                for i in range(n):
+                    Q[i, k] -= Q[i, j] * dot
+
+
+def qqr(
+    A: np.ndarray, tol: float = 0, use_numba: bool = True
+) -> Tuple[np.ndarray, np.ndarray]:
+    """
+    Quasi-QR decomposition using modified Gram-Schmidt with quasi inner product.
+
+    For complex symmetric systems, uses <x,y>_Q = sum(x_k * y_k) without conjugation.
+
+    Parameters
+    ----------
+    A : ndarray
+        Input matrix (n x m)
+    tol : float
+        Tolerance (unused, for API compatibility)
+    use_numba : bool
+        If True and Numba available, use JIT-compiled kernel
+
+    Returns
+    -------
+    Q : ndarray
+        Quasi-orthonormal columns (n x m)
+    R : ndarray
+        Upper triangular matrix (m x m)
+    """
+    n, m = A.shape
+    is_complex = np.iscomplexobj(A)
+    dtype = np.complex128 if is_complex else np.float64
+
+    Q = np.ascontiguousarray(A, dtype=dtype)
+    R = np.zeros((m, m), dtype=dtype)
+
+    if use_numba and HAS_NUMBA:
+        if is_complex:
+            _qqr_kernel_complex(Q, R, n, m)
+        else:
+            _qqr_kernel_real(Q, R, n, m)
+    else:
+        for j in range(m):
+            qj = Q[:, j]
+            # CRITICAL FIX: Use sum(qj * qj) NOT np.dot(qj, qj)
+            # np.dot conjugates the first argument for complex arrays!
+            # Fortran: R(k,k)=dsqrt(sum(Q(:,k)*Q(:,k))) - no conjugation
+            r_jj_sq = np.sum(qj * qj)  # Quasi inner product - NO conjugation
+            r_jj = np.sqrt(r_jj_sq)
+            R[j, j] = r_jj
+            if np.abs(r_jj) > 1e-14:
+                Q[:, j] *= 1.0 / r_jj
+                if j < m - 1:
+                    # CRITICAL FIX: Quasi inner product for off-diagonal
+                    # Fortran: R(k,j)=sum(Q(:,k)*Q(:,j)) - no conjugation
+                    for k in range(j + 1, m):
+                        R[j, k] = np.sum(Q[:, j] * Q[:, k])  # NO conjugation
+                        Q[:, k] -= R[j, k] * Q[:, j]
+
+    return Q, R
+
+
+# =============================================================================
+# Preconditioner Classes
+# =============================================================================
+
+
+class _ILUPreconditioner:
+    """Wrapper for ILU preconditioner to work with blqmr."""
+
+    def __init__(self, ilu_factor):
+        self.ilu = ilu_factor
+        self.shape = (ilu_factor.shape[0], ilu_factor.shape[1])
+        self.dtype = ilu_factor.L.dtype
+
+    def solve(self, b):
+        # Convert to real if needed for real ILU
+        b_solve = b.real if np.isrealobj(self.ilu.L.data) and np.iscomplexobj(b) else b
+        if b_solve.ndim == 1:
+            return self.ilu.solve(b_solve)
+        else:
+            x = np.zeros_like(b_solve)
+            for i in range(b_solve.shape[1]):
+                x[:, i] = self.ilu.solve(b_solve[:, i])
+            return x
+
+
+class SparsePreconditioner:
+    """Efficient sparse preconditioner using LU factorization."""
+
+    __slots__ = ("lu1", "lu2", "is_two_part", "is_ilu1", "is_ilu2")
+
+    def __init__(self, M1, M2=None):
+        self.is_two_part = M2 is not None
+        self.is_ilu1 = isinstance(M1, (_ILUPreconditioner, _LUPreconditioner))
+        self.is_ilu2 = (
+            isinstance(M2, (_ILUPreconditioner, _LUPreconditioner))
+            if M2 is not None
+            else False
+        )
+
+        if M1 is not None:
+            if self.is_ilu1:
+                self.lu1 = M1
+            else:
+                M1_csc = sparse.csc_matrix(M1) if not sparse.isspmatrix_csc(M1) else M1
+                self.lu1 = splu(M1_csc)
+        else:
+            self.lu1 = None
+
+        if M2 is not None:
+            if self.is_ilu2:
+                self.lu2 = M2
+            else:
+                M2_csc = sparse.csc_matrix(M2) if not sparse.isspmatrix_csc(M2) else M2
+                self.lu2 = splu(M2_csc)
+        else:
+            self.lu2 = None
+
+    def solve(self, b: np.ndarray, out: Optional[np.ndarray] = None) -> np.ndarray:
+        if self.lu1 is None:
+            return b
+        if out is None:
+            out = np.empty_like(b)
+
+        # Handle dtype conversion for ILU with real data
+        if self.is_ilu1:
+            result = self.lu1.solve(b)
+            if out.dtype != result.dtype:
+                out = np.asarray(out, dtype=result.dtype)
+            out[:] = result
+        else:
+            if b.ndim == 1:
+                out[:] = self.lu1.solve(b)
+            else:
+                for i in range(b.shape[1]):
+                    out[:, i] = self.lu1.solve(b[:, i])
+
+        if self.is_two_part:
+            if self.is_ilu2:
+                out[:] = self.lu2.solve(out)
+            else:
+                if b.ndim == 1:
+                    out[:] = self.lu2.solve(out)
+                else:
+                    for i in range(b.shape[1]):
+                        out[:, i] = self.lu2.solve(out[:, i])
+        return out
+
+
+class DensePreconditioner:
+    """Efficient dense preconditioner using LU factorization."""
+
+    __slots__ = ("lu1", "piv1", "lu2", "piv2", "is_two_part")
+
+    def __init__(self, M1: Optional[np.ndarray], M2: Optional[np.ndarray] = None):
+        from scipy.linalg import lu_factor
+
+        self.is_two_part = M2 is not None
+        if M1 is not None:
+            self.lu1, self.piv1 = lu_factor(M1)
+        else:
+            self.lu1 = self.piv1 = None
+        if M2 is not None:
+            self.lu2, self.piv2 = lu_factor(M2)
+        else:
+            self.lu2 = self.piv2 = None
+
+    def solve(self, b: np.ndarray, out: Optional[np.ndarray] = None) -> np.ndarray:
+        from scipy.linalg import lu_solve
+
+        if self.lu1 is None:
+            return b
+        result = lu_solve((self.lu1, self.piv1), b)
+        if self.is_two_part:
+            result = lu_solve((self.lu2, self.piv2), result)
+        if out is not None:
+            out[:] = result
+            return out
+        return result
+
+
+# =============================================================================
+# BL-QMR Workspace
+# =============================================================================
+
+
+class BLQMRWorkspace:
+    """Pre-allocated workspace for BL-QMR iterations."""
+
+    __slots__ = (
+        "v",
+        "vt",
+        "beta",
+        "alpha",
+        "omega",
+        "theta",
+        "Qa",
+        "Qb",
+        "Qc",
+        "Qd",
+        "zeta",
+        "zetat",
+        "eta",
+        "tau",
+        "taot",
+        "p",
+        "stacked",
+        "QQ_full",
+        "tmp0",
+        "tmp1",
+        "tmp2",
+        "Av",
+        "precond_tmp",
+        "n",
+        "m",
+        "dtype",
+    )
+
+    def __init__(self, n: int, m: int, dtype=np.float64):
+        self.n, self.m = n, m
+        self.dtype = dtype
+        self.v = np.zeros((n, m, 3), dtype=dtype)
+        self.vt = np.zeros((n, m), dtype=dtype)
+        self.beta = np.zeros((m, m, 3), dtype=dtype)
+        self.alpha = np.zeros((m, m), dtype=dtype)
+        self.omega = np.zeros((m, m, 3), dtype=dtype)
+        self.theta = np.zeros((m, m), dtype=dtype)
+        self.Qa = np.zeros((m, m, 3), dtype=dtype)
+        self.Qb = np.zeros((m, m, 3), dtype=dtype)
+        self.Qc = np.zeros((m, m, 3), dtype=dtype)
+        self.Qd = np.zeros((m, m, 3), dtype=dtype)
+        self.zeta = np.zeros((m, m), dtype=dtype)
+        self.zetat = np.zeros((m, m), dtype=dtype)
+        self.eta = np.zeros((m, m), dtype=dtype)
+        self.tau = np.zeros((m, m), dtype=dtype)
+        self.taot = np.zeros((m, m), dtype=dtype)
+        self.p = np.zeros((n, m, 3), dtype=dtype)
+        self.stacked = np.zeros((2 * m, m), dtype=dtype)
+        self.QQ_full = np.zeros((2 * m, 2 * m), dtype=dtype)
+        self.tmp0 = np.zeros((m, m), dtype=dtype)
+        self.tmp1 = np.zeros((m, m), dtype=dtype)
+        self.tmp2 = np.zeros((m, m), dtype=dtype)
+        self.Av = np.zeros((n, m), dtype=dtype)
+        self.precond_tmp = np.zeros((n, m), dtype=dtype)
+
+    def reset(self):
+        self.v.fill(0)
+        self.beta.fill(0)
+        self.omega.fill(0)
+        self.Qa.fill(0)
+        self.Qb.fill(0)
+        self.Qc.fill(0)
+        self.Qd.fill(0)
+        self.p.fill(0)
+        self.taot.fill(0)
+
+
+# =============================================================================
+# Preconditioner Factory
+# =============================================================================
+
+
+def make_preconditioner(A: sparse.spmatrix, precond_type: str = "diag", **kwargs):
+    """
+    Create a preconditioner for iterative solvers.
+
+    Parameters
+    ----------
+    A : sparse matrix
+        System matrix
+    precond_type : str
+        'diag' or 'jacobi': Diagonal (Jacobi) preconditioner
+        'ilu' or 'ilu0': Incomplete LU with minimal fill
+        'ilut': Incomplete LU with threshold (better quality)
+        'lu': Full LU factorization (exact, use as reference)
+        'ssor': Symmetric SOR
+    **kwargs : dict
+        Additional parameters for ILU:
+        - drop_tol: Drop tolerance (default: 1e-4 for ilut, 0 for ilu0)
+        - fill_factor: Fill factor (default: 10 for ilut, 1 for ilu0)
+
+    Returns
+    -------
+    M : preconditioner object
+        Preconditioner (use as M1 in blqmr)
+    """
+    if precond_type in ("diag", "jacobi"):
+        diag = A.diagonal().copy()
+        diag[np.abs(diag) < 1e-14] = 1.0
+        return sparse.diags(
+            1.0 / diag, format="csr"
+        )  # Return inverse for preconditioning!
+
+    elif precond_type == "ilu0":
+        # ILU(0) - no fill-in, fast but may be poor quality
+        try:
+            ilu = spilu(A.tocsc(), drop_tol=0, fill_factor=1)
+            return _ILUPreconditioner(ilu)
+        except Exception as e:
+            warnings.warn(f"ILU(0) factorization failed: {e}, falling back to diagonal")
+            return make_preconditioner(A, "diag")
+
+    elif precond_type in ("ilu", "ilut"):
+        # ILUT - ILU with threshold, better quality (similar to UMFPACK)
+        drop_tol = kwargs.get("drop_tol", 1e-4)
+        fill_factor = kwargs.get("fill_factor", 10)
+        try:
+            ilu = spilu(A.tocsc(), drop_tol=drop_tol, fill_factor=fill_factor)
+            return _ILUPreconditioner(ilu)
+        except Exception as e:
+            warnings.warn(f"ILUT factorization failed: {e}, trying ILU(0)")
+            try:
+                ilu = spilu(A.tocsc(), drop_tol=0, fill_factor=1)
+                return _ILUPreconditioner(ilu)
+            except Exception as e2:
+                warnings.warn(f"ILU(0) also failed: {e2}, falling back to diagonal")
+                return make_preconditioner(A, "diag")
+
+    elif precond_type == "lu":
+        # Full LU - exact factorization (for reference/debugging)
+        try:
+            lu = splu(A.tocsc())
+            return _LUPreconditioner(lu)
+        except Exception as e:
+            warnings.warn(f"LU factorization failed: {e}, falling back to ILUT")
+            return make_preconditioner(A, "ilut")
+
+    elif precond_type == "ssor":
+        omega = kwargs.get("omega", 1.0)
+        D = sparse.diags(A.diagonal(), format="csr")
+        L = sparse.tril(A, k=-1, format="csr")
+        return (D + omega * L).tocsr()
+
+    else:
+        raise ValueError(f"Unknown preconditioner type: {precond_type}")
+
+
+class _LUPreconditioner:
+    """Wrapper for full LU preconditioner."""
+
+    def __init__(self, lu_factor):
+        self.lu = lu_factor
+        self.shape = (lu_factor.shape[0], lu_factor.shape[1])
+        self.dtype = np.float64  # Assume real for now
+
+    def solve(self, b):
+        if b.ndim == 1:
+            return self.lu.solve(b)
+        else:
+            x = np.zeros_like(b)
+            for i in range(b.shape[1]):
+                x[:, i] = self.lu.solve(b[:, i])
+            return x
+
+
+# =============================================================================
+# Pure-Python Block QMR Solver
+# =============================================================================
+
+
+def _blqmr_python_impl(
+    A: Union[np.ndarray, sparse.spmatrix],
+    B: np.ndarray,
+    tol: float = 1e-6,
+    maxiter: Optional[int] = None,
+    M1=None,
+    M2=None,
+    x0: Optional[np.ndarray] = None,
+    residual: bool = False,
+    workspace: Optional[BLQMRWorkspace] = None,
+) -> Tuple[np.ndarray, int, float, int, np.ndarray]:
+    """Native Python Block QMR implementation (internal)."""
+    if B.ndim == 1:
+        B = B.reshape(-1, 1)
+
+    n, m = B.shape
+    is_complex_input = np.iscomplexobj(A) or np.iscomplexobj(B)
+    dtype = np.complex128 if is_complex_input else np.float64
+
+    if maxiter is None:
+        maxiter = min(n, 100)
+
+    if (
+        workspace is None
+        or workspace.n != n
+        or workspace.m != m
+        or workspace.dtype != dtype
+    ):
+        ws = BLQMRWorkspace(n, m, dtype)
+    else:
+        ws = workspace
+        ws.reset()
+
+    # Setup preconditioner
+    if M1 is not None:
+        if isinstance(M1, (_ILUPreconditioner, _LUPreconditioner)):
+            precond = SparsePreconditioner(M1, M2)
+        elif sparse.issparse(M1):
+            precond = SparsePreconditioner(M1, M2)
+        elif hasattr(M1, "solve"):
+            # Custom preconditioner with .solve() method
+            precond = M1  # Use directly
+        else:
+            precond = DensePreconditioner(M1, M2)
+    else:
+        precond = None
+
+    if x0 is None:
+        x = np.zeros((n, m), dtype=dtype)
+    else:
+        x = np.asarray(x0, dtype=dtype).reshape(n, m).copy()
+
+    # Initialize indices: Fortran t3=mod(0,3)+1=1 -> Python t3=0
+    t3 = 0
+    t3n = 2
+    t3p = 1
+
+    # Initialize Q matrices (identity)
+    ws.Qa[:, :, :] = 0
+    ws.Qb[:, :, :] = 0
+    ws.Qc[:, :, :] = 0
+    ws.Qd[:, :, :] = 0
+    ws.Qa[:, :, t3] = np.eye(m, dtype=dtype)
+    ws.Qd[:, :, t3n] = np.eye(m, dtype=dtype)
+    ws.Qd[:, :, t3] = np.eye(m, dtype=dtype)
+
+    A_is_sparse = sparse.issparse(A)
+    if A_is_sparse:
+        ws.vt[:] = B - A @ x
+    else:
+        np.subtract(B, A @ x, out=ws.vt)
+
+    if precond is not None:
+        precond.solve(ws.vt, out=ws.vt)
+        if np.any(np.isnan(ws.vt)):
+            return x, 2, 1.0, 0, np.array([])
+
+    # QQR decomposition
+    Q, R = qqr(ws.vt)
+    ws.v[:, :, t3p] = Q
+    ws.beta[:, :, t3p] = R
+
+    # Compute omega - standard norm WITH conjugation (Hermitian norm)
+    # Fortran: omega(i,i,t3p)=sqrt(sum(conjg(v(:,i,t3p))*v(:,i,t3p)))
+    for i in range(m):
+        col = ws.v[:, i, t3p]
+        if is_complex_input:
+            ws.omega[i, i, t3p] = np.sqrt(np.sum(np.conj(col) * col).real)
+        else:
+            ws.omega[i, i, t3p] = np.sqrt(np.sum(col * col))
+
+    # taut = omega * beta
+    ws.taot[:] = ws.omega[:, :, t3p] @ ws.beta[:, :, t3p]
+
+    isquasires = not residual
+    if isquasires:
+        # Fortran: Qres0=maxval(sqrt(sum(abs(conjg(taut)*taut),1))) for complex
+        if is_complex_input:
+            Qres0 = np.max(np.sqrt(np.sum(np.abs(np.conj(ws.taot) * ws.taot), axis=0)))
+        else:
+            Qres0 = np.max(np.sqrt(np.sum(ws.taot * ws.taot, axis=0)))
+    else:
+        omegat = np.zeros((n, m), dtype=dtype)
+        for i in range(m):
+            if np.abs(ws.omega[i, i, t3p]) > 1e-14:
+                omegat[:, i] = ws.v[:, i, t3p] / ws.omega[i, i, t3p]
+        if is_complex_input:
+            Qres0 = np.max(np.sqrt(np.sum(np.abs(np.conj(ws.vt) * ws.vt), axis=0)))
+        else:
+            Qres0 = np.max(np.sqrt(np.sum(ws.vt * ws.vt, axis=0)))
+
+    if Qres0 < 1e-16:
+        result = x.real if not is_complex_input else x
+        return result, 0, 0.0, 0, np.array([0.0])
+
+    flag, resv, Qres1, relres, iter_count = 1, np.zeros(maxiter), -1.0, 1.0, 0
+
+    for k in range(1, maxiter + 1):
+        # Index cycling
+        t3 = k % 3
+        t3p = (k + 1) % 3
+        t3n = (k - 1) % 3
+        t3nn = (k - 2) % 3
+
+        # tmp = A * v(:,:,t3)
+        if A_is_sparse:
+            ws.Av[:] = A @ ws.v[:, :, t3]
+        else:
+            np.matmul(A, ws.v[:, :, t3], out=ws.Av)
+
+        # Apply preconditioner
+        if precond is not None:
+            precond.solve(ws.Av, out=ws.vt)
+            ws.vt[:] = ws.vt - ws.v[:, :, t3n] @ ws.beta[:, :, t3].T
+        else:
+            ws.vt[:] = ws.Av - ws.v[:, :, t3n] @ ws.beta[:, :, t3].T
+
+        # alpha = v^T * vt (transpose, not conjugate transpose)
+        ws.alpha[:] = ws.v[:, :, t3].T @ ws.vt
+        ws.vt[:] = ws.vt - ws.v[:, :, t3] @ ws.alpha
+
+        # QQR decomposition
+        Q, R = qqr(ws.vt)
+        ws.v[:, :, t3p] = Q
+        ws.beta[:, :, t3p] = R
+
+        # Compute omega (standard Hermitian norm)
+        for i in range(m):
+            col = ws.v[:, i, t3p]
+            if is_complex_input:
+                ws.omega[i, i, t3p] = np.sqrt(np.sum(np.conj(col) * col).real)
+            else:
+                ws.omega[i, i, t3p] = np.sqrt(np.sum(col * col))
+
+        # Compute intermediate matrices
+        ws.tmp0[:] = ws.omega[:, :, t3n] @ ws.beta[:, :, t3].T
+        ws.theta[:] = ws.Qb[:, :, t3nn] @ ws.tmp0
+        ws.tmp1[:] = ws.Qd[:, :, t3nn] @ ws.tmp0
+        ws.tmp2[:] = ws.omega[:, :, t3] @ ws.alpha
+        ws.eta[:] = ws.Qa[:, :, t3n] @ ws.tmp1 + ws.Qb[:, :, t3n] @ ws.tmp2
+        ws.zetat[:] = ws.Qc[:, :, t3n] @ ws.tmp1 + ws.Qd[:, :, t3n] @ ws.tmp2
+
+        # Build ZZ matrix and do standard QR
+        ws.stacked[:m, :] = ws.zetat
+        ws.stacked[m:, :] = ws.omega[:, :, t3p] @ ws.beta[:, :, t3p]
+
+        QQ, zeta_full = np.linalg.qr(ws.stacked, mode="complete")
+        ws.zeta[:] = zeta_full[:m, :]
+
+        if is_complex_input:
+            ws.QQ_full[:] = np.conj(QQ.T)
+        else:
+            ws.QQ_full[:] = QQ.T
+
+        ws.Qa[:, :, t3] = ws.QQ_full[:m, :m]
+        ws.Qb[:, :, t3] = ws.QQ_full[:m, m : 2 * m]
+        ws.Qc[:, :, t3] = ws.QQ_full[m : 2 * m, :m]
+        ws.Qd[:, :, t3] = ws.QQ_full[m : 2 * m, m : 2 * m]
+
+        # Invert zeta
+        try:
+            zeta_inv = np.linalg.inv(ws.zeta)
+        except np.linalg.LinAlgError:
+            zeta_inv = np.linalg.pinv(ws.zeta)
+
+        # Update p, tau, x, taut
+        ws.p[:, :, t3] = (
+            ws.v[:, :, t3] - ws.p[:, :, t3n] @ ws.eta - ws.p[:, :, t3nn] @ ws.theta
+        ) @ zeta_inv
+        ws.tau[:] = ws.Qa[:, :, t3] @ ws.taot
+        x[:] = x + ws.p[:, :, t3] @ ws.tau
+        ws.taot[:] = ws.Qc[:, :, t3] @ ws.taot
+
+        # Compute residual
+        if isquasires:
+            if is_complex_input:
+                Qres = np.max(
+                    np.sqrt(np.sum(np.abs(np.conj(ws.taot) * ws.taot), axis=0))
+                )
+            else:
+                Qres = np.max(np.sqrt(np.sum(ws.taot * ws.taot, axis=0)))
+        else:
+            tmp0_diag = np.zeros((m, m), dtype=dtype)
+            for i in range(m):
+                if np.abs(ws.omega[i, i, t3p]) > 1e-14:
+                    tmp0_diag[i, :] = ws.Qd[:, i, t3] / ws.omega[i, i, t3p]
+            if is_complex_input:
+                omegat = omegat @ np.conj(ws.Qc[:, :, t3].T) + ws.v[
+                    :, :, t3p
+                ] @ np.conj(tmp0_diag)
+                tmp_res = np.conj(omegat @ ws.taot)
+                Qres = np.max(
+                    np.sqrt(np.sum(np.abs(np.conj(tmp_res) * tmp_res), axis=0))
+                )
+            else:
+                omegat = omegat @ ws.Qc[:, :, t3].T + ws.v[:, :, t3p] @ tmp0_diag
+                tmp_res = omegat @ ws.taot
+                Qres = np.max(np.sqrt(np.sum(tmp_res * tmp_res, axis=0)))
+
+        resv[k - 1] = Qres
+
+        if k > 1 and abs(Qres - Qres1) < np.finfo(dtype).eps:
+            flag, iter_count = 3, k
+            break
+
+        Qres1, relres, iter_count = Qres, Qres / Qres0, k
+
+        if relres <= tol:
+            flag = 0
+            break
+
+    resv = resv[:iter_count]
+    result = x.real if not is_complex_input else x
+    return result, flag, relres, iter_count, resv
+
+
+# =============================================================================
+# High-Level Solver Interface
+# =============================================================================
+
+
+def blqmr_solve(
+    Ap: np.ndarray,
+    Ai: np.ndarray,
+    Ax: np.ndarray,
+    b: np.ndarray,
+    *,
+    x0: Optional[np.ndarray] = None,
+    tol: float = 1e-6,
+    maxiter: Optional[int] = None,
+    droptol: float = 0.001,
+    use_precond: bool = True,
+    zero_based: bool = True,
+) -> BLQMRResult:
+    """
+    Solve sparse linear system Ax = b using Block QMR algorithm.
+
+    Uses Fortran extension if available, otherwise falls back to pure Python.
+
+    Parameters
+    ----------
+    Ap : ndarray of int32
+        Column pointers for CSC format. Length n+1.
+    Ai : ndarray of int32
+        Row indices for CSC format. Length nnz.
+    Ax : ndarray of float64
+        Non-zero values. Length nnz.
+    b : ndarray of float64
+        Right-hand side vector. Length n.
+    x0 : ndarray, optional
+        Initial guess.
+    tol : float, default 1e-6
+        Convergence tolerance for relative residual.
+    maxiter : int, optional
+        Maximum iterations. Default is n.
+    droptol : float, default 0.001
+        Drop tolerance for ILU preconditioner (Fortran only).
+    use_precond : bool, default True
+        Whether to use ILU preconditioning.
+    zero_based : bool, default True
+        If True, Ap and Ai use 0-based indexing (Python/C convention).
+        If False, uses 1-based indexing (Fortran convention).
+
+    Returns
+    -------
+    BLQMRResult
+        Result object containing solution and convergence info.
+    """
+    n = len(Ap) - 1
+
+    if maxiter is None:
+        maxiter = n
+
+    if BLQMR_EXT:
+        return _blqmr_solve_fortran(
+            Ap,
+            Ai,
+            Ax,
+            b,
+            x0=x0,
+            tol=tol,
+            maxiter=maxiter,
+            droptol=droptol,
+            use_precond=use_precond,
+            zero_based=zero_based,
+        )
+    else:
+        return _blqmr_solve_native_csc(
+            Ap,
+            Ai,
+            Ax,
+            b,
+            x0=x0,
+            tol=tol,
+            maxiter=maxiter,
+            use_precond=use_precond,
+            zero_based=zero_based,
+        )
+
+
+def _blqmr_solve_fortran(
+    Ap, Ai, Ax, b, *, x0, tol, maxiter, droptol, use_precond, zero_based
+) -> BLQMRResult:
+    """Fortran backend for blqmr_solve."""
+    n = len(Ap) - 1
+    nnz = len(Ax)
+
+    Ap = np.asfortranarray(Ap, dtype=np.int32)
+    Ai = np.asfortranarray(Ai, dtype=np.int32)
+    Ax = np.asfortranarray(Ax, dtype=np.float64)
+    b = np.asfortranarray(b, dtype=np.float64)
+
+    if len(Ai) != nnz:
+        raise ValueError(f"Ai length ({len(Ai)}) must match Ax length ({nnz})")
+    if len(b) != n:
+        raise ValueError(f"b length ({len(b)}) must match matrix size ({n})")
+
+    if zero_based:
+        Ap = Ap + 1
+        Ai = Ai + 1
+
+    dopcond = 1 if use_precond else 0
+
+    x, flag, niter, relres = _blqmr.blqmr_solve_real(
+        n, nnz, Ap, Ai, Ax, b, maxiter, tol, droptol, dopcond
+    )
+
+    return BLQMRResult(
+        x=x.copy(), flag=int(flag), iter=int(niter), relres=float(relres)
+    )
+
+
+def _blqmr_solve_native_csc(
+    Ap, Ai, Ax, b, *, x0, tol, maxiter, use_precond, zero_based
+) -> BLQMRResult:
+    """Native Python backend for blqmr_solve with CSC input."""
+    n = len(Ap) - 1
+
+    if not zero_based:
+        Ap = Ap - 1
+        Ai = Ai - 1
+
+    A = sparse.csc_matrix((Ax, Ai, Ap), shape=(n, n))
+
+    M1 = None
+    if use_precond:
+        try:
+            M1 = make_preconditioner(A, "ilu")
+        except Exception:
+            try:
+                M1 = make_preconditioner(A, "diag")  # FIX: Changed A_sp to A
+            except Exception:
+                M1 = None  # Fall back to no preconditioning
+
+    x, flag, relres, niter, resv = _blqmr_python_impl(
+        A, b, tol=tol, maxiter=maxiter, M1=M1, x0=x0
+    )
+
+    if x.ndim > 1:
+        x = x.ravel()
+
+    return BLQMRResult(x=x, flag=flag, iter=niter, relres=relres, resv=resv)
+
+
+def blqmr_solve_multi(
+    Ap: np.ndarray,
+    Ai: np.ndarray,
+    Ax: np.ndarray,
+    B: np.ndarray,
+    *,
+    tol: float = 1e-6,
+    maxiter: Optional[int] = None,
+    droptol: float = 0.001,
+    use_precond: bool = True,
+    zero_based: bool = True,
+) -> BLQMRResult:
+    """
+    Solve sparse linear system AX = B with multiple right-hand sides.
+
+    Uses Fortran extension if available, otherwise falls back to pure Python.
+    """
+    n = len(Ap) - 1
+
+    if maxiter is None:
+        maxiter = n
+
+    if BLQMR_EXT:
+        return _blqmr_solve_multi_fortran(
+            Ap,
+            Ai,
+            Ax,
+            B,
+            tol=tol,
+            maxiter=maxiter,
+            droptol=droptol,
+            use_precond=use_precond,
+            zero_based=zero_based,
+        )
+    else:
+        return _blqmr_solve_multi_native(
+            Ap,
+            Ai,
+            Ax,
+            B,
+            tol=tol,
+            maxiter=maxiter,
+            use_precond=use_precond,
+            zero_based=zero_based,
+        )
+
+
+def _blqmr_solve_multi_fortran(
+    Ap, Ai, Ax, B, *, tol, maxiter, droptol, use_precond, zero_based
+) -> BLQMRResult:
+    """Fortran backend for blqmr_solve_multi."""
+    n = len(Ap) - 1
+    nnz = len(Ax)
+
+    Ap = np.asfortranarray(Ap, dtype=np.int32)
+    Ai = np.asfortranarray(Ai, dtype=np.int32)
+    Ax = np.asfortranarray(Ax, dtype=np.float64)
+    B = np.asfortranarray(B, dtype=np.float64)
+
+    if B.ndim == 1:
+        B = B.reshape(-1, 1, order="F")
+    nrhs = B.shape[1]
+
+    if zero_based:
+        Ap = Ap + 1
+        Ai = Ai + 1
+
+    dopcond = 1 if use_precond else 0
+
+    X, flag, niter, relres = _blqmr.blqmr_solve_real_multi(
+        n, nnz, nrhs, Ap, Ai, Ax, B, maxiter, tol, droptol, dopcond
+    )
+
+    return BLQMRResult(
+        x=X.copy(), flag=int(flag), iter=int(niter), relres=float(relres)
+    )
+
+
+def _blqmr_solve_multi_native(
+    Ap, Ai, Ax, B, *, tol, maxiter, use_precond, zero_based
+) -> BLQMRResult:
+    """Native Python backend for blqmr_solve_multi."""
+    n = len(Ap) - 1
+
+    if not zero_based:
+        Ap = Ap - 1
+        Ai = Ai - 1
+
+    A = sparse.csc_matrix((Ax, Ai, Ap), shape=(n, n))
+
+    M1 = None
+    if use_precond:
+        try:
+            M1 = make_preconditioner(A, "ilu")
+        except Exception:
+            try:
+                M1 = make_preconditioner(A, "diag")  # FIX: Changed A_sp to A
+            except Exception:
+                M1 = None  # Fall back to no preconditioning
+
+    if B.ndim == 1:
+        B = B.reshape(-1, 1)
+
+    x, flag, relres, niter, resv = _blqmr_python_impl(
+        A, B, tol=tol, maxiter=maxiter, M1=M1
+    )
+
+    return BLQMRResult(x=x, flag=flag, iter=niter, relres=relres, resv=resv)
+
+
+def blqmr_scipy(
+    A,
+    b: np.ndarray,
+    x0: Optional[np.ndarray] = None,
+    tol: float = 1e-6,
+    maxiter: Optional[int] = None,
+    M=None,
+    **kwargs,
+) -> Tuple[np.ndarray, int]:
+    """
+    SciPy-compatible interface for BLQMR solver.
+
+    Parameters
+    ----------
+    A : sparse matrix or ndarray
+        System matrix
+    b : ndarray
+        Right-hand side vector
+    x0 : ndarray, optional
+        Initial guess
+    tol : float
+        Convergence tolerance
+    maxiter : int, optional
+        Maximum iterations
+    M : preconditioner, optional
+        Preconditioner (used as M1 for Python backend)
+    **kwargs
+        Additional arguments passed to blqmr()
+
+    Returns
+    -------
+    x : ndarray
+        Solution vector
+    flag : int
+        Convergence flag (0 = converged)
+    """
+    result = blqmr(A, b, x0=x0, tol=tol, maxiter=maxiter, M1=M, **kwargs)
+    return result.x, result.flag
+
+
+def blqmr(
+    A: Union[np.ndarray, sparse.spmatrix],
+    B: np.ndarray,
+    tol: float = 1e-6,
+    maxiter: Optional[int] = None,
+    M1=None,
+    M2=None,
+    x0: Optional[np.ndarray] = None,
+    residual: bool = False,
+    workspace: Optional[BLQMRWorkspace] = None,
+    droptol: float = 0.001,
+    use_precond: bool = True,
+) -> BLQMRResult:
+    """
+    Block Quasi-Minimal-Residual (BL-QMR) solver - main interface.
+
+    Uses Fortran extension if available, otherwise falls back to pure Python.
+
+    Parameters
+    ----------
+    A : ndarray or sparse matrix
+        Symmetric n x n matrix (can be complex)
+    B : ndarray
+        Right-hand side vector/matrix (n,) or (n x m)
+    tol : float
+        Convergence tolerance (default: 1e-6)
+    maxiter : int, optional
+        Maximum iterations (default: n for Fortran, min(n, 20) for Python)
+    M1, M2 : preconditioner, optional
+        Preconditioner M = M1 @ M2 (Python backend only)
+    x0 : ndarray, optional
+        Initial guess
+    residual : bool
+        If True, use true residual for convergence (Python backend only)
+    workspace : BLQMRWorkspace, optional
+        Pre-allocated workspace (Python backend only)
+    droptol : float, default 0.001
+        Drop tolerance for ILU preconditioner (Fortran backend only)
+    use_precond : bool, default True
+        Whether to use ILU preconditioning (Fortran backend only)
+
+    Returns
+    -------
+    BLQMRResult
+        Result object containing:
+        - x: Solution array
+        - flag: 0 = converged, 1 = max iterations, 2 = preconditioner singular, 3 = stagnated
+        - iter: Number of iterations
+        - relres: Final relative residual
+        - resv: Residual history (Python backend only)
+    """
+    if BLQMR_EXT:
+        return _blqmr_fortran(
+            A,
+            B,
+            tol=tol,
+            maxiter=maxiter,
+            x0=x0,
+            droptol=droptol,
+            use_precond=use_precond,
+        )
+    else:
+        return _blqmr_native(
+            A,
+            B,
+            tol=tol,
+            maxiter=maxiter,
+            M1=M1,
+            M2=M2,
+            x0=x0,
+            residual=residual,
+            workspace=workspace,
+            use_precond=use_precond,
+        )
+
+
+def _blqmr_fortran(
+    A: Union[np.ndarray, sparse.spmatrix],
+    B: np.ndarray,
+    *,
+    tol: float,
+    maxiter: Optional[int],
+    x0: Optional[np.ndarray],
+    droptol: float,
+    use_precond: bool,
+) -> BLQMRResult:
+    """Fortran backend for blqmr()."""
+    A_csc = sparse.csc_matrix(A)
+
+    # CRITICAL: Sort indices for UMFPACK compatibility
+    if not A_csc.has_sorted_indices:
+        A_csc.sort_indices()
+
+    Ap = A_csc.indptr.astype(np.int32)
+    Ai = A_csc.indices.astype(np.int32)
+
+    n = A_csc.shape[0]
+    nnz = A_csc.nnz
+
+    if maxiter is None:
+        maxiter = n
+
+    # Convert to Fortran format (1-based indexing)
+    Ap_f = np.asfortranarray(Ap + 1, dtype=np.int32)
+    Ai_f = np.asfortranarray(Ai + 1, dtype=np.int32)
+
+    dopcond = 1 if use_precond else 0
+
+    # Check if complex
+    is_complex = np.iscomplexobj(A) or np.iscomplexobj(B)
+
+    if is_complex:
+        # Complex path
+        Ax_f = np.asfortranarray(A_csc.data, dtype=np.complex128)
+
+        if B.ndim == 1 or (B.ndim == 2 and B.shape[1] == 1):
+            # Single RHS
+            b_f = np.asfortranarray(B.ravel(), dtype=np.complex128)
+            x, flag, niter, relres = _blqmr.blqmr_solve_complex(
+                n, nnz, Ap_f, Ai_f, Ax_f, b_f, maxiter, tol, droptol, dopcond
+            )
+            return BLQMRResult(
+                x=x.copy(), flag=int(flag), iter=int(niter), relres=float(relres)
+            )
+        else:
+            # Multiple RHS - use block method
+            B_f = np.asfortranarray(B, dtype=np.complex128)
+            nrhs = B_f.shape[1]
+            X, flag, niter, relres = _blqmr.blqmr_solve_complex_multi(
+                n, nnz, nrhs, Ap_f, Ai_f, Ax_f, B_f, maxiter, tol, droptol, dopcond
+            )
+            return BLQMRResult(
+                x=X.copy(), flag=int(flag), iter=int(niter), relres=float(relres)
+            )
+    else:
+        # Real path
+        Ax_f = np.asfortranarray(A_csc.data, dtype=np.float64)
+
+        if B.ndim == 1 or (B.ndim == 2 and B.shape[1] == 1):
+            # Single RHS
+            b_f = np.asfortranarray(B.ravel(), dtype=np.float64)
+            x, flag, niter, relres = _blqmr.blqmr_solve_real(
+                n, nnz, Ap_f, Ai_f, Ax_f, b_f, maxiter, tol, droptol, dopcond
+            )
+            return BLQMRResult(
+                x=x.copy(), flag=int(flag), iter=int(niter), relres=float(relres)
+            )
+        else:
+            # Multiple RHS - use block method
+            B_f = np.asfortranarray(B, dtype=np.float64)
+            nrhs = B_f.shape[1]
+            X, flag, niter, relres = _blqmr.blqmr_solve_real_multi(
+                n, nnz, nrhs, Ap_f, Ai_f, Ax_f, B_f, maxiter, tol, droptol, dopcond
+            )
+            return BLQMRResult(
+                x=X.copy(), flag=int(flag), iter=int(niter), relres=float(relres)
+            )
+
+
+def _blqmr_native(
+    A: Union[np.ndarray, sparse.spmatrix],
+    B: np.ndarray,
+    *,
+    tol: float,
+    maxiter: Optional[int],
+    M1,
+    M2,
+    x0: Optional[np.ndarray],
+    residual: bool,
+    workspace: Optional[BLQMRWorkspace],
+    use_precond: bool,
+) -> BLQMRResult:
+    """Native Python backend for blqmr()."""
+    # Auto-create preconditioner if requested and not provided
+    if use_precond and M1 is None:
+        A_sp = sparse.csc_matrix(A) if not sparse.issparse(A) else A
+        try:
+            M1 = make_preconditioner(A_sp, "ilu")
+        except Exception:
+            try:
+                M1 = make_preconditioner(A_sp, "diag")
+            except Exception:
+                M1 = None  # Fall back to no preconditioning
+
+    x, flag, relres, niter, resv = _blqmr_python_impl(
+        A,
+        B,
+        tol=tol,
+        maxiter=maxiter,
+        M1=M1,
+        M2=M2,
+        x0=x0,
+        residual=residual,
+        workspace=workspace,
+    )
+
+    # Flatten x if single RHS
+    if x.ndim > 1 and x.shape[1] == 1:
+        x = x.ravel()
+
+    return BLQMRResult(x=x, flag=flag, iter=niter, relres=relres, resv=resv)
+
+
+# =============================================================================
+# Test Function
+# =============================================================================
+
+
+def _test():
+    """Quick test to verify installation."""
+    print("BLIT BLQMR Test")
+    print("=" * 40)
+    print(f"Fortran backend available: {BLQMR_EXT}")
+    print(f"Numba acceleration available: {HAS_NUMBA}")
+    print(f"Using backend: {'Fortran' if BLQMR_EXT else 'Pure Python'}")
+    print()
+
+    # Build test matrix from CSC components
+    n = 5
+    Ap = np.array([0, 2, 5, 9, 10, 12], dtype=np.int32)
+    Ai = np.array([0, 1, 0, 2, 4, 1, 2, 3, 4, 2, 1, 4], dtype=np.int32)
+    Ax = np.array(
+        [2.0, 3.0, 3.0, -1.0, 4.0, 4.0, -3.0, 1.0, 2.0, 2.0, 6.0, 1.0], dtype=np.float64
+    )
+    b = np.array([8.0, 45.0, -3.0, 3.0, 19.0], dtype=np.float64)
+
+    # Create sparse matrix
+    A = sparse.csc_matrix((Ax, Ai, Ap), shape=(n, n))
+
+    print(f"Matrix: {n}x{n}, nnz={len(Ax)}")
+    print(f"b: {b}")
+    print("\nCalling blqmr()...")
+
+    result = blqmr(A, b, tol=1e-8)
+
+    print(f"\n{result}")
+    print(f"Solution: {result.x}")
+
+    # Verify
+    res = np.linalg.norm(A @ result.x - b)
+    print(f"||Ax - b|| = {res:.2e}")
+
+    return result.converged
+
+
+if __name__ == "__main__":
+    _test()