da4ml 0.5.1.post1__cp311-cp311-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl

da4ml/trace/fixed_variable_array.py

@@ -0,0 +1,600 @@
+from collections.abc import Callable
+from inspect import signature
+from typing import TypeVar
+
+import numpy as np
+from numba.typed import List as NumbaList
+from numpy.typing import NDArray
+
+from ..cmvm.api import solve, solver_options_t
+from .fixed_variable import FixedVariable, FixedVariableInput, HWConfig, LookupTable, QInterval
+from .ops import _quantize, einsum, reduce
+
+T = TypeVar('T')
+
+
+def to_raw_arr(obj: T) -> T:
+    if isinstance(obj, tuple):
+        return tuple(to_raw_arr(x) for x in obj)  # type: ignore
+    elif isinstance(obj, list):
+        return [to_raw_arr(x) for x in obj]  # type: ignore
+    elif isinstance(obj, dict):
+        return {k: to_raw_arr(v) for k, v in obj.items()}  # type: ignore
+    if isinstance(obj, FixedVariableArray):
+        return obj._vars  # type: ignore
+    return obj
+
+
+def _max_of(a, b):
+    if isinstance(a, FixedVariable):
+        return a.max_of(b)
+    elif isinstance(b, FixedVariable):
+        return b.max_of(a)
+    else:
+        return max(a, b)
+
+
+def _min_of(a, b):
+    if isinstance(a, FixedVariable):
+        return a.min_of(b)
+    elif isinstance(b, FixedVariable):
+        return b.min_of(a)
+    else:
+        return min(a, b)
+
+
+def mmm(mat0: np.ndarray, mat1: np.ndarray):
+    shape = mat0.shape[:-1] + mat1.shape[1:]
+    mat0, mat1 = mat0.reshape((-1, mat0.shape[-1])), mat1.reshape((mat1.shape[0], -1))
+    _shape = (mat0.shape[0], mat1.shape[1])
+    _vars = np.empty(_shape, dtype=object)
+    for i in range(mat0.shape[0]):
+        for j in range(mat1.shape[1]):
+            vec0 = mat0[i]
+            vec1 = mat1[:, j]
+            _vars[i, j] = reduce(lambda x, y: x + y, vec0 * vec1)
+    return _vars.reshape(shape)
+
+
+def cmvm(cm: np.ndarray, v: 'FixedVariableArray', solver_options: solver_options_t) -> np.ndarray:
+    mask = offload_mask(cm, v)
+    if np.any(mask):
+        offload_cm = cm * mask.astype(cm.dtype)
+        cm = cm * (~mask).astype(cm.dtype)
+    else:
+        offload_cm = None
+    _qintervals = [QInterval(float(_v.low), float(_v.high), float(_v.step)) for _v in v._vars]
+    _latencies = [float(_v.latency) for _v in v._vars]
+    qintervals = NumbaList(_qintervals)  # type: ignore
+    latencies = NumbaList(_latencies)  # type: ignore
+    hwconf = v._vars.ravel()[0].hwconf
+    solver_options.setdefault('adder_size', hwconf.adder_size)
+    solver_options.setdefault('carry_size', hwconf.carry_size)
+    _mat = np.ascontiguousarray(cm.astype(np.float32))
+    sol = solve(_mat, qintervals=qintervals, latencies=latencies, **solver_options)
+    _r: np.ndarray = sol(v._vars)
+    if offload_cm is not None:
+        _r = _r + mmm(v._vars, offload_cm)
+    return _r
+
+
+def offload_mask(cm: NDArray, v: 'FixedVariableArray') -> NDArray[np.bool_]:
+    assert v.ndim == 1
+    assert cm.ndim == 2
+    assert cm.shape[0] == v.shape[0]
+    bits = np.sum(v.kif, axis=0)[:, None]
+    return (bits == 0) & (cm != 0)
+
+
+_unary_functions = (
+    np.sin,
+    np.cos,
+    np.tan,
+    np.exp,
+    np.log,
+    np.invert,
+    np.sqrt,
+    np.tanh,
+    np.sinh,
+    np.cosh,
+    np.arccos,
+    np.arcsin,
+    np.arctan,
+    np.arcsinh,
+    np.arccosh,
+    np.arctanh,
+    np.exp2,
+    np.expm1,
+    np.log2,
+    np.log10,
+    np.log1p,
+    np.cbrt,
+    np.reciprocal,
+)
+
+
+class FixedVariableArray:
+    """Symbolic array of FixedVariable for tracing operations. Supports numpy ufuncs and array functions."""
+
+    __array_priority__ = 100
+
+    def __array_function__(self, func, types, args, kwargs):
+        if func in (np.mean, np.sum, np.amax, np.amin, np.prod, np.max, np.min):
+            match func:
+                case np.mean:
+                    _x = reduce(lambda x, y: x + y, *args, **kwargs)
+                    return _x * (_x.size / self._vars.size)
+                case np.sum:
+                    return reduce(lambda x, y: x + y, *args, **kwargs)
+                case np.max | np.amax:
+                    return reduce(_max_of, *args, **kwargs)
+                case np.min | np.amin:
+                    return reduce(_min_of, *args, **kwargs)
+                case np.prod:
+                    return reduce(lambda x, y: x * y, *args, **kwargs)
+                case _:
+                    raise NotImplementedError(f'Unsupported function: {func}')
+
+        if func is np.clip:
+            assert len(args) == 3, 'Clip function requires exactly three arguments'
+            x, low, high = args
+            _x, low, high = np.broadcast_arrays(x, low, high)
+            x = FixedVariableArray(_x, self.solver_options)
+            x = np.amax(np.stack((x, low), axis=-1), axis=-1)  # type: ignore
+            return np.amin(np.stack((x, high), axis=-1), axis=-1)
+
+        if func is np.einsum:
+            # assert len(args) == 2
+            sig = signature(np.einsum)
+            bind = sig.bind(*args, **kwargs)
+            eq = args[0]
+            operands = bind.arguments['operands']
+            if isinstance(operands[0], str):
+                operands = operands[1:]
+            assert len(operands) == 2, 'Einsum on FixedVariableArray requires exactly two operands'
+            assert bind.arguments.get('out', None) is None, 'Output argument is not supported'
+            return einsum(eq, *operands)
+
+        if func is np.dot:
+            assert len(args) in (2, 3), 'Dot function requires exactly two or three arguments'
+
+            assert len(args) == 2
+            a, b = args
+            if not isinstance(a, FixedVariableArray):
+                a = np.array(a)
+            if not isinstance(b, FixedVariableArray):
+                b = np.array(b)
+            if a.shape[-1] == b.shape[0]:
+                return a @ b
+
+            assert a.size == 1 or b.size == 1, f'Error in dot product: {a.shape} @ {b.shape}'
+            return a * b
+
+        args, kwargs = to_raw_arr(args), to_raw_arr(kwargs)
+        return FixedVariableArray(
+            func(*args, **kwargs),
+            self.solver_options,
+        )
+
+    def __array_ufunc__(self, ufunc, method, *inputs, **kwargs):
+        assert method == '__call__', f'Only __call__ method is supported for ufuncs, got {method}'
+
+        match ufunc:
+            case np.add | np.subtract | np.multiply | np.true_divide | np.negative:
+                inputs = [to_raw_arr(x) for x in inputs]
+                return FixedVariableArray(ufunc(*inputs, **kwargs), self.solver_options)
+
+            case np.negative:
+                assert len(inputs) == 1
+                return FixedVariableArray(ufunc(to_raw_arr(inputs[0]), **kwargs), self.solver_options)
+
+            case np.maximum | np.minimum:
+                op = _max_of if ufunc is np.maximum else _min_of
+                a, b = np.broadcast_arrays(inputs[0], inputs[1])
+                shape = a.shape
+                a, b = a.ravel(), b.ravel()
+                r = np.empty(a.size, dtype=object)
+                for i in range(a.size):
+                    r[i] = op(a[i], b[i])
+                return FixedVariableArray(r.reshape(shape), self.solver_options)
+
+            case np.matmul:
+                assert len(inputs) == 2
+                assert isinstance(inputs[0], FixedVariableArray) or isinstance(inputs[1], FixedVariableArray)
+                if isinstance(inputs[0], FixedVariableArray):
+                    return inputs[0].matmul(inputs[1])
+                else:
+                    return inputs[1].rmatmul(inputs[0])
+
+            case np.power:
+                assert len(inputs) == 2
+                base, exp = inputs
+                return base**exp
+
+            case np.abs | np.absolute:
+                assert len(inputs) == 1
+                assert inputs[0] is self
+                arr = self._vars.ravel()
+                r = np.array([v.__abs__() for v in arr])
+                return FixedVariableArray(r.reshape(self.shape), self.solver_options)
+
+            case np.square:
+                assert len(inputs) == 1
+                assert inputs[0] is self
+                return self**2
+
+        if ufunc in _unary_functions:
+            assert len(inputs) == 1
+            assert inputs[0] is self
+            return self.apply(ufunc)
+
+        raise NotImplementedError(f'Unsupported ufunc: {ufunc}')
+
+    def __init__(
+        self,
+        vars: NDArray,
+        solver_options: solver_options_t | None = None,
+    ):
+        _vars = np.array(vars)
+        _vars_f = _vars.ravel()
+        hwconf = next(iter(v for v in _vars_f if isinstance(v, FixedVariable))).hwconf
+        for i, v in enumerate(_vars_f):
+            if not isinstance(v, FixedVariable):
+                _vars_f[i] = FixedVariable(float(v), float(v), 1.0, hwconf=hwconf)
+        self._vars = _vars
+        _solver_options = signature(solve).parameters
+        _solver_options = {k: v.default for k, v in _solver_options.items() if v.default is not v.empty}
+        if solver_options is not None:
+            _solver_options.update(solver_options)
+        _solver_options.pop('qintervals', None)
+        _solver_options.pop('latencies', None)
+        self.solver_options: solver_options_t = _solver_options  # type: ignore
+
+    @classmethod
+    def from_lhs(
+        cls,
+        low: NDArray[np.floating],
+        high: NDArray[np.floating],
+        step: NDArray[np.floating],
+        hwconf: HWConfig | tuple[int, int, int] = HWConfig(1, -1, -1),
+        latency: np.ndarray | float = 0.0,
+        solver_options: solver_options_t | None = None,
+    ):
+        low, high, step = np.array(low), np.array(high), np.array(step)
+        shape = low.shape
+        assert shape == high.shape == step.shape
+
+        low, high, step = low.ravel(), high.ravel(), step.ravel()
+        latency = np.full_like(low, latency) if isinstance(latency, (int, float)) else latency.ravel()
+
+        vars = []
+        for l, h, s, lat in zip(low, high, step, latency):
+            var = FixedVariable(
+                low=float(l),
+                high=float(h),
+                step=float(s),
+                hwconf=hwconf,
+                latency=float(
+                    lat,
+                ),
+            )
+            vars.append(var)
+        vars = np.array(vars).reshape(shape)
+        return cls(vars, solver_options)
+
+    __array_priority__ = 100
+
+    @classmethod
+    def from_kif(
+        cls,
+        k: NDArray[np.bool_ | np.integer],
+        i: NDArray[np.integer],
+        f: NDArray[np.integer],
+        hwconf: HWConfig | tuple[int, int, int] = HWConfig(1, -1, -1),
+        latency: NDArray[np.floating] | float = 0.0,
+        solver_options: solver_options_t | None = None,
+    ):
+        mask = k + i + f <= 0
+        k = np.where(mask, 0, k)
+        i = np.where(mask, 0, i)
+        f = np.where(mask, 0, f)
+        step = 2.0**-f
+        _high = 2.0**i
+        high, low = _high - step, -_high * k
+        return cls.from_lhs(low, high, step, hwconf, latency, solver_options)
+
+    def matmul(self, other) -> 'FixedVariableArray':
+        if self.collapsed:
+            self_mat = np.array([v.low for v in self._vars.ravel()], dtype=np.float64).reshape(self._vars.shape)
+            if isinstance(other, FixedVariableArray):
+                if not other.collapsed:
+                    return self_mat @ other  # type: ignore
+                other_mat = np.array([v.low for v in other._vars.ravel()], dtype=np.float64).reshape(other._vars.shape)
+            else:
+                other_mat = np.array(other, dtype=np.float64)
+
+            r = self_mat @ other_mat
+            return FixedVariableArray.from_lhs(
+                low=r,
+                high=r,
+                step=np.ones_like(r),
+                hwconf=self._vars.ravel()[0].hwconf,
+                solver_options=self.solver_options,
+            )
+
+        if isinstance(other, FixedVariableArray):
+            other = other._vars
+        if not isinstance(other, np.ndarray):
+            other = np.array(other)
+        if any(isinstance(x, FixedVariable) for x in other.ravel()):
+            mat0, mat1 = self._vars, other
+            _vars = mmm(mat0, mat1)
+            return FixedVariableArray(_vars, self.solver_options)
+
+        solver_options = (self.solver_options or {}).copy()
+        shape0, shape1 = self.shape, other.shape
+        assert shape0[-1] == shape1[0], f'Matrix shapes do not match: {shape0} @ {shape1}'
+        contract_len = shape1[0]
+        out_shape = shape0[:-1] + shape1[1:]
+        mat0, mat1 = self.reshape((-1, contract_len)), other.reshape((contract_len, -1))
+        r = []
+        for i in range(mat0.shape[0]):
+            vec = mat0[i]
+            _r = cmvm(mat1, vec, solver_options)
+            r.append(_r)
+        r = np.array(r).reshape(out_shape)
+        return FixedVariableArray(r, self.solver_options)
+
+    def __matmul__(self, other):
+        return self.matmul(other)
+
+    def rmatmul(self, other):
+        mat1 = np.moveaxis(other, -1, 0)
+        mat0 = np.moveaxis(self, 0, -1)  # type: ignore
+        ndim0, ndim1 = mat0.ndim, mat1.ndim
+        r = mat0 @ mat1
+
+        _axes = tuple(range(0, ndim0 + ndim1 - 2))
+        axes = _axes[ndim0 - 1 :] + _axes[: ndim0 - 1]
+        return r.transpose(axes)
+
+    def __rmatmul__(self, other):
+        return self.rmatmul(other)
+
+    def __getitem__(self, item):
+        vars = self._vars[item]
+        if isinstance(vars, np.ndarray):
+            return FixedVariableArray(vars, self.solver_options)
+        else:
+            return vars
+
+    def __len__(self):
+        return len(self._vars)
+
+    @property
+    def shape(self):
+        return self._vars.shape
+
+    def __add__(self, other):
+        if isinstance(other, FixedVariableArray):
+            return FixedVariableArray(self._vars + other._vars, self.solver_options)
+        return FixedVariableArray(self._vars + other, self.solver_options)
+
+    def __sub__(self, other):
+        if isinstance(other, FixedVariableArray):
+            return FixedVariableArray(self._vars - other._vars, self.solver_options)
+        return FixedVariableArray(self._vars - other, self.solver_options)
+
+    def __mul__(self, other):
+        if isinstance(other, FixedVariableArray):
+            return FixedVariableArray(self._vars * other._vars, self.solver_options)
+        return FixedVariableArray(self._vars * other, self.solver_options)
+
+    def __truediv__(self, other):
+        return FixedVariableArray(self._vars * (1 / other), self.solver_options)
+
+    def __radd__(self, other):
+        return self + other
+
+    def __neg__(self):
+        return FixedVariableArray(-self._vars, self.solver_options)
+
+    def __repr__(self):
+        shape = self._vars.shape
+        hwconf_str = str(self._vars.ravel()[0].hwconf)[8:]
+        max_lat = max(v.latency for v in self._vars.ravel())
+        return f'FixedVariableArray(shape={shape}, hwconf={hwconf_str}, latency={max_lat})'
+
+    def __pow__(self, power: int | float):
+        _power = int(power)
+        if _power == power and _power >= 0:
+            return FixedVariableArray(self._vars**_power, self.solver_options)
+        else:
+            return self.apply(lambda x: x**power)
+
+    def relu(
+        self,
+        i: NDArray[np.integer] | None = None,
+        f: NDArray[np.integer] | None = None,
+        round_mode: str = 'TRN',
+    ):
+        shape = self._vars.shape
+        i = np.broadcast_to(i, shape) if i is not None else np.full(shape, None)
+        f = np.broadcast_to(f, shape) if f is not None else np.full(shape, None)
+        ret = []
+        for v, i, f in zip(self._vars.ravel(), i.ravel(), f.ravel()):  # type: ignore
+            ret.append(v.relu(i=i, f=f, round_mode=round_mode))
+        return FixedVariableArray(np.array(ret).reshape(shape), self.solver_options)
+
+    def quantize(
+        self,
+        k: NDArray[np.integer] | np.integer | int | None = None,
+        i: NDArray[np.integer] | np.integer | int | None = None,
+        f: NDArray[np.integer] | np.integer | int | None = None,
+        overflow_mode: str = 'WRAP',
+        round_mode: str = 'TRN',
+    ):
+        shape = self._vars.shape
+        if any(x is None for x in (k, i, f)):
+            kif = self.kif
+        k = np.broadcast_to(k, shape) if k is not None else kif[0]  # type: ignore
+        i = np.broadcast_to(i, shape) if i is not None else kif[1]  # type: ignore
+        f = np.broadcast_to(f, shape) if f is not None else kif[2]  # type: ignore
+        ret = []
+        for v, k, i, f in zip(self._vars.ravel(), k.ravel(), i.ravel(), f.ravel()):  # type: ignore
+            ret.append(v.quantize(k=k, i=i, f=f, overflow_mode=overflow_mode, round_mode=round_mode))
+        return FixedVariableArray(np.array(ret).reshape(shape), self.solver_options)
+
+    def flatten(self):
+        return FixedVariableArray(self._vars.flatten(), self.solver_options)
+
+    def reshape(self, *shape):
+        return FixedVariableArray(self._vars.reshape(*shape), self.solver_options)
+
+    def transpose(self, axes=None):
+        return FixedVariableArray(self._vars.transpose(axes), self.solver_options)
+
+    def ravel(self):
+        return FixedVariableArray(self._vars.ravel(), self.solver_options)
+
+    @property
+    def dtype(self):
+        return self._vars.dtype
+
+    @property
+    def size(self):
+        return self._vars.size
+
+    @property
+    def ndim(self):
+        return self._vars.ndim
+
+    @property
+    def kif(self):
+        """[k, i, f] array"""
+        shape = self._vars.shape
+        kif = np.array([v.kif for v in self._vars.ravel()]).reshape(*shape, 3)
+        return np.moveaxis(kif, -1, 0)
+
+    @property
+    def lhs(self):
+        """[low, high, step] array"""
+        shape = self._vars.shape
+        lhs = np.array([(v.low, v.high, v.step) for v in self._vars.ravel()], dtype=np.float32).reshape(*shape, 3)
+        return np.moveaxis(lhs, -1, 0)
+
+    @property
+    def latency(self):
+        """Maximum latency among all elements."""
+        return np.array([v.latency for v in self._vars.ravel()]).reshape(self._vars.shape)
+
+    @property
+    def collapsed(self):
+        return all(v.low == v.high for v in self._vars.ravel())
+
+    def apply(self, fn: Callable[[NDArray], NDArray]) -> 'RetardedFixedVariableArray':
+        """Apply a unary operator to all elements, returning a RetardedFixedVariableArray."""
+        return RetardedFixedVariableArray(
+            self._vars,
+            self.solver_options,
+            operator=fn,
+        )
+
+    @property
+    def T(self):
+        return self.transpose()
+
+
+class FixedVariableArrayInput(FixedVariableArray):
+    """Similar to FixedVariableArray, but initializes all elements as FixedVariableInput - the precisions are unspecified when initialized, and the highest precision requested (i.e., quantized to) will be recorded for generation of the logic."""
+
+    def __init__(
+        self,
+        shape: tuple[int, ...] | int,
+        hwconf: HWConfig | tuple[int, int, int] = HWConfig(1, -1, -1),
+        solver_options: solver_options_t | None = None,
+        latency=0.0,
+    ):
+        _vars = np.empty(shape, dtype=object)
+        _vars_f = _vars.ravel()
+        for i in range(_vars.size):
+            _vars_f[i] = FixedVariableInput(latency, hwconf)
+        super().__init__(_vars, solver_options)
+
+
+def make_table(fn: Callable[[NDArray], NDArray], qint: QInterval) -> LookupTable:
+    low, high, step = qint
+    n = round(abs(high - low) / step) + 1
+    return LookupTable(fn(np.linspace(low, high, n)))
+
+
+class RetardedFixedVariableArray(FixedVariableArray):
+    """Ephemeral FixedVariableArray generated from operations of unspecified output precision.
+    This object translates to normal FixedVariableArray upon quantization.
+    Does not inherit the maximum precision like FixedVariableArrayInput.
+
+    This object can be used in two ways:
+    1. Quantization with specified precision, which converts to FixedVariableArray.
+    2. Apply an further unary operation, which returns another RetardedFixedVariableArray. (e.g., composite functions)
+    """
+
+    def __init__(self, vars: NDArray, solver_options: solver_options_t | None, operator: Callable[[NDArray], NDArray]):
+        self._operator = operator
+        super().__init__(vars, solver_options)
+
+    def __array_function__(self, ufunc, method, *inputs, **kwargs):
+        raise RuntimeError('RetardedFixedVariableArray only supports quantization or further unary operations.')
+
+    def apply(self, fn: Callable[[NDArray], NDArray]) -> 'RetardedFixedVariableArray':
+        return RetardedFixedVariableArray(
+            self._vars,
+            self.solver_options,
+            operator=lambda x: fn(self._operator(x)),
+        )
+
+    def quantize(
+        self,
+        k: NDArray[np.integer] | np.integer | int | None = None,
+        i: NDArray[np.integer] | np.integer | int | None = None,
+        f: NDArray[np.integer] | np.integer | int | None = None,
+        overflow_mode: str = 'WRAP',
+        round_mode: str = 'TRN',
+    ):
+        if any(x is None for x in (k, i, f)):
+            assert all(x is not None for x in (k, i, f)), 'Either all or none of k, i, f must be specified'
+            _k = _i = _f = [None] * self.size
+        else:
+            _k = np.broadcast_to(k, self.shape).ravel()  # type: ignore
+            _i = np.broadcast_to(i, self.shape).ravel()  # type: ignore
+            _f = np.broadcast_to(f, self.shape).ravel()  # type: ignore
+
+            op = lambda x: _quantize(self._operator(x), k, i, f, overflow_mode, round_mode)  # type: ignore
+
+        local_tables: dict[tuple[QInterval, tuple[int, int, int]] | QInterval, LookupTable] = {}
+        variables = []
+        for v, _kk, _ii, _ff in zip(self._vars.ravel(), _k, _i, _f):
+            v: FixedVariable
+            qint = v.qint if v._factor >= 0 else QInterval(v.qint.max, v.qint.min, v.qint.step)
+            if (_kk is None) or (_ii is None) or (_ff is None):
+                op = self._operator
+                _key = qint
+            else:
+                op = lambda x: _quantize(self._operator(x), _kk, _ii, _ff, overflow_mode, round_mode)  # type: ignore
+                _key = (qint, (int(_kk), int(_ii), int(_ff)))
+
+            if _key in local_tables:
+                table = local_tables[_key]
+            else:
+                table = make_table(op, qint)
+                local_tables[_key] = table
+            variables.append(v.lookup(table))
+
+        variables = np.array(variables).reshape(self._vars.shape)
+        return FixedVariableArray(variables, self.solver_options)
+
+    def __repr__(self):
+        return 'Retarded' + super().__repr__()
+
+    @property
+    def kif(self):
+        raise RuntimeError('RetardedFixedVariableArray does not have defined kif until quantized.')

da4ml/trace/ops/__init__.py

@@ -0,0 +1,13 @@
+from .einsum_utils import einsum
+from .quantization import _quantize, quantize, relu
+from .reduce_utils import reduce
+
+__all__ = [
+    'einsum',
+    'relu',
+    'quantization',
+    'reduce',
+    '_quantize',
+    'relu',
+    'quantize',
+]