5579681118

pending completion

Build # 5579681118

Build Type

push

github

Committed by

web-flow

Commit Message

Merge pull request #1281 from IntelPython/unary_out_overlap

Created a temporary copy in case of overlap for unary function

Run Details

2269 of 2783 branches covered (81.53%)

Branch coverage included in aggregate %.

15 of 16 new or added lines in 1 file covered. (93.75%)

34 existing lines in 1 file now uncovered.

8281 of 9898 relevant lines covered (83.66%)

5828.35 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

87.54

/dpctl/tensor/_elementwise_common.py

#                       Data Parallel Control (dpctl)
#
#  Copyright 2020-2023 Intel Corporation
#
#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.

import numbers

import numpy as np

import dpctl
import dpctl.memory as dpm
import dpctl.tensor as dpt
import dpctl.tensor._tensor_impl as ti
from dpctl.tensor._manipulation_functions import _broadcast_shape_impl
from dpctl.tensor._usmarray import _is_object_with_buffer_protocol as _is_buffer
from dpctl.utils import ExecutionPlacementError

from ._type_utils import (
    _empty_like_orderK,
    _empty_like_pair_orderK,
    _find_buf_dtype,
    _find_buf_dtype2,
    _find_inplace_dtype,
    _to_device_supported_dtype,
)


class UnaryElementwiseFunc:
    """
    Class that implements unary element-wise functions.
    """

    def __init__(self, name, result_type_resolver_fn, unary_dp_impl_fn, docs):
        self.__name__ = "UnaryElementwiseFunc"
        self.name_ = name
        self.result_type_resolver_fn_ = result_type_resolver_fn
        self.unary_fn_ = unary_dp_impl_fn
        self.__doc__ = docs

    def __call__(self, x, out=None, order="K"):
        if not isinstance(x, dpt.usm_ndarray):
            raise TypeError(f"Expected dpctl.tensor.usm_ndarray, got {type(x)}")

        if order not in ["C", "F", "K", "A"]:
            order = "K"
        buf_dt, res_dt = _find_buf_dtype(
            x.dtype, self.result_type_resolver_fn_, x.sycl_device
        )
        if res_dt is None:
            raise RuntimeError

        orig_out = out
        if out is not None:
            if not isinstance(out, dpt.usm_ndarray):
                raise TypeError(
                    f"output array must be of usm_ndarray type, got {type(out)}"
                )

            if out.shape != x.shape:
                raise TypeError(
                    "The shape of input and output arrays are inconsistent."
                    f"Expected output shape is {x.shape}, got {out.shape}"
                )

            if res_dt != out.dtype:
                raise TypeError(
                    f"Output array of type {res_dt} is needed,"
                    f" got {out.dtype}"
                )

            if (
                buf_dt is None
                and ti._array_overlap(x, out)
                and not ti._same_logical_tensors(x, out)
            ):
                # Allocate a temporary buffer to avoid memory overlapping.
                # Note if `buf_dt` is not None, a temporary copy of `x` will be
                # created, so the array overlap check isn't needed.
                out = dpt.empty_like(out)

            if (
                dpctl.utils.get_execution_queue((x.sycl_queue, out.sycl_queue))
                is None
            ):
                raise TypeError(
                    "Input and output allocation queues are not compatible"
                )

        exec_q = x.sycl_queue
        if buf_dt is None:
            if out is None:
                if order == "K":
                    out = _empty_like_orderK(x, res_dt)
                else:
                    if order == "A":
                        order = "F" if x.flags.f_contiguous else "C"
                    out = dpt.empty_like(x, dtype=res_dt, order=order)

            ht_unary_ev, unary_ev = self.unary_fn_(x, out, sycl_queue=exec_q)

            if not (orig_out is None or orig_out is out):
                # Copy the out data from temporary buffer to original memory
                ht_copy_ev, _ = ti._copy_usm_ndarray_into_usm_ndarray(
                    src=out, dst=orig_out, sycl_queue=exec_q, depends=[unary_ev]
                )
                ht_copy_ev.wait()
                out = orig_out

            ht_unary_ev.wait()
            return out

        if order == "K":
            buf = _empty_like_orderK(x, buf_dt)
        else:
            if order == "A":
                order = "F" if x.flags.f_contiguous else "C"
            buf = dpt.empty_like(x, dtype=buf_dt, order=order)

        ht_copy_ev, copy_ev = ti._copy_usm_ndarray_into_usm_ndarray(
            src=x, dst=buf, sycl_queue=exec_q
        )
        if out is None:
            if order == "K":
                out = _empty_like_orderK(buf, res_dt)
            else:
                out = dpt.empty_like(buf, dtype=res_dt, order=order)

        ht, _ = self.unary_fn_(buf, out, sycl_queue=exec_q, depends=[copy_ev])
        ht_copy_ev.wait()
        ht.wait()

        return out


def _get_queue_usm_type(o):
    """Return SYCL device where object `o` allocated memory, or None."""
    if isinstance(o, dpt.usm_ndarray):
        return o.sycl_queue, o.usm_type
    elif hasattr(o, "__sycl_usm_array_interface__"):
        try:
            m = dpm.as_usm_memory(o)
            return m.sycl_queue, m.get_usm_type()
        except Exception:
            return None, None
    return None, None


class WeakBooleanType:
    "Python type representing type of Python boolean objects"

    def __init__(self, o):
        self.o_ = o

    def get(self):
        return self.o_


class WeakIntegralType:
    "Python type representing type of Python integral objects"

    def __init__(self, o):
        self.o_ = o

    def get(self):
        return self.o_


class WeakFloatingType:
    """Python type representing type of Python floating point objects"""

    def __init__(self, o):
        self.o_ = o

    def get(self):
        return self.o_


class WeakComplexType:
    """Python type representing type of Python complex floating point objects"""

    def __init__(self, o):
        self.o_ = o

    def get(self):
        return self.o_


def _get_dtype(o, dev):
    if isinstance(o, dpt.usm_ndarray):
        return o.dtype
    if hasattr(o, "__sycl_usm_array_interface__"):
        return dpt.asarray(o).dtype
    if _is_buffer(o):
        host_dt = np.array(o).dtype
        dev_dt = _to_device_supported_dtype(host_dt, dev)
        return dev_dt
    if hasattr(o, "dtype"):
        dev_dt = _to_device_supported_dtype(o.dtype, dev)
        return dev_dt
    if isinstance(o, bool):
        return WeakBooleanType(o)
    if isinstance(o, int):
        return WeakIntegralType(o)
    if isinstance(o, float):
        return WeakFloatingType(o)
    if isinstance(o, complex):
        return WeakComplexType(o)
    return np.object_


def _validate_dtype(dt) -> bool:
    return isinstance(
        dt,
        (WeakBooleanType, WeakIntegralType, WeakFloatingType, WeakComplexType),
    ) or (
        isinstance(dt, dpt.dtype)
        and dt
        in [
            dpt.bool,
            dpt.int8,
            dpt.uint8,
            dpt.int16,
            dpt.uint16,
            dpt.int32,
            dpt.uint32,
            dpt.int64,
            dpt.uint64,
            dpt.float16,
            dpt.float32,
            dpt.float64,
            dpt.complex64,
            dpt.complex128,
        ]
    )


def _weak_type_num_kind(o):
    _map = {"?": 0, "i": 1, "f": 2, "c": 3}
    if isinstance(o, WeakBooleanType):
        return _map["?"]
    if isinstance(o, WeakIntegralType):
        return _map["i"]
    if isinstance(o, WeakFloatingType):
        return _map["f"]
    if isinstance(o, WeakComplexType):
        return _map["c"]
    raise TypeError(
        f"Unexpected type {o} while expecting "
        "`WeakBooleanType`, `WeakIntegralType`,"
        "`WeakFloatingType`, or `WeakComplexType`."
    )


def _strong_dtype_num_kind(o):
    _map = {"b": 0, "i": 1, "u": 1, "f": 2, "c": 3}
    if not isinstance(o, dpt.dtype):
        raise TypeError
    k = o.kind
    if k in _map:
        return _map[k]
    raise ValueError(f"Unrecognized kind {k} for dtype {o}")


def _resolve_weak_types(o1_dtype, o2_dtype, dev):
    "Resolves weak data type per NEP-0050"
    if isinstance(
        o1_dtype,
        (WeakBooleanType, WeakIntegralType, WeakFloatingType, WeakComplexType),
    ):
        if isinstance(
            o2_dtype,
            (
                WeakBooleanType,
                WeakIntegralType,
                WeakFloatingType,
                WeakComplexType,
            ),
        ):
            raise ValueError
        o1_kind_num = _weak_type_num_kind(o1_dtype)
        o2_kind_num = _strong_dtype_num_kind(o2_dtype)
        if o1_kind_num > o2_kind_num:
            if isinstance(o1_dtype, WeakBooleanType):
                return dpt.bool, o2_dtype
            if isinstance(o1_dtype, WeakIntegralType):
                return dpt.int64, o2_dtype
            if isinstance(o1_dtype, WeakComplexType):
                if o2_dtype is dpt.float16 or o2_dtype is dpt.float32:
                    return dpt.complex64, o2_dtype
                return (
                    _to_device_supported_dtype(dpt.complex128, dev),
                    o2_dtype,
                )
            return _to_device_supported_dtype(dpt.float64, dev), o2_dtype
        else:
            return o2_dtype, o2_dtype
    elif isinstance(
        o2_dtype,
        (WeakBooleanType, WeakIntegralType, WeakFloatingType, WeakComplexType),
    ):
        o1_kind_num = _strong_dtype_num_kind(o1_dtype)
        o2_kind_num = _weak_type_num_kind(o2_dtype)
        if o2_kind_num > o1_kind_num:
            if isinstance(o2_dtype, WeakBooleanType):
                return o1_dtype, dpt.bool
            if isinstance(o2_dtype, WeakIntegralType):
                return o1_dtype, dpt.int64
            if isinstance(o2_dtype, WeakComplexType):
                if o1_dtype is dpt.float16 or o1_dtype is dpt.float32:
                    return o1_dtype, dpt.complex64
                return o1_dtype, _to_device_supported_dtype(dpt.complex128, dev)
            return (
                o1_dtype,
                _to_device_supported_dtype(dpt.float64, dev),
            )
        else:
            return o1_dtype, o1_dtype
    else:
        return o1_dtype, o2_dtype


def _get_shape(o):
    if isinstance(o, dpt.usm_ndarray):
        return o.shape
    if _is_buffer(o):
        return memoryview(o).shape
    if isinstance(o, numbers.Number):
        return tuple()
    return getattr(o, "shape", tuple())


class BinaryElementwiseFunc:
    """
    Class that implements binary element-wise functions.
    """

    def __init__(
        self,
        name,
        result_type_resolver_fn,
        binary_dp_impl_fn,
        docs,
        binary_inplace_fn=None,
    ):
        self.__name__ = "BinaryElementwiseFunc"
        self.name_ = name
        self.result_type_resolver_fn_ = result_type_resolver_fn
        self.binary_fn_ = binary_dp_impl_fn
        self.binary_inplace_fn_ = binary_inplace_fn
        self.__doc__ = docs

    def __str__(self):
        return f"<BinaryElementwiseFunc '{self.name_}'>"

    def __repr__(self):
        return f"<BinaryElementwiseFunc '{self.name_}'>"

    def __call__(self, o1, o2, out=None, order="K"):
        # FIXME: replace with check against base array
        # when views can be identified
        if o1 is out:
            return self._inplace(o1, o2)
        elif o2 is out:
            return self._inplace(o2, o1)

        if order not in ["K", "C", "F", "A"]:
            order = "K"
        q1, o1_usm_type = _get_queue_usm_type(o1)
        q2, o2_usm_type = _get_queue_usm_type(o2)
        if q1 is None and q2 is None:
            raise ExecutionPlacementError(
                "Execution placement can not be unambiguously inferred "
                "from input arguments. "
                "One of the arguments must represent USM allocation and "
                "expose `__sycl_usm_array_interface__` property"
            )
        if q1 is None:
            exec_q = q2
            res_usm_type = o2_usm_type
        elif q2 is None:
            exec_q = q1
            res_usm_type = o1_usm_type
        else:
            exec_q = dpctl.utils.get_execution_queue((q1, q2))
            if exec_q is None:
                raise ExecutionPlacementError(
                    "Execution placement can not be unambiguously inferred "
                    "from input arguments."
                )
            res_usm_type = dpctl.utils.get_coerced_usm_type(
                (
                    o1_usm_type,
                    o2_usm_type,
                )
            )
        dpctl.utils.validate_usm_type(res_usm_type, allow_none=False)
        o1_shape = _get_shape(o1)
        o2_shape = _get_shape(o2)
        if not all(
            isinstance(s, (tuple, list))
            for s in (
                o1_shape,
                o2_shape,
            )
        ):
            raise TypeError(
                "Shape of arguments can not be inferred. "
                "Arguments are expected to be "
                "lists, tuples, or both"
            )
        try:
            res_shape = _broadcast_shape_impl(
                [
                    o1_shape,
                    o2_shape,
                ]
            )
        except ValueError:
            raise ValueError(
                "operands could not be broadcast together with shapes "
                f"{o1_shape} and {o2_shape}"
            )
        sycl_dev = exec_q.sycl_device
        o1_dtype = _get_dtype(o1, sycl_dev)
        o2_dtype = _get_dtype(o2, sycl_dev)
        if not all(_validate_dtype(o) for o in (o1_dtype, o2_dtype)):
            raise ValueError("Operands of unsupported types")

        o1_dtype, o2_dtype = _resolve_weak_types(o1_dtype, o2_dtype, sycl_dev)

        buf1_dt, buf2_dt, res_dt = _find_buf_dtype2(
            o1_dtype, o2_dtype, self.result_type_resolver_fn_, sycl_dev
        )

        if res_dt is None:
            raise TypeError(
                f"function '{self.name_}' does not support input types "
                f"({o1_dtype}, {o2_dtype}), "
                "and the inputs could not be safely coerced to any "
                "supported types according to the casting rule ''safe''."
            )

        if out is not None:
            if not isinstance(out, dpt.usm_ndarray):
                raise TypeError(
                    f"output array must be of usm_ndarray type, got {type(out)}"
                )

            if out.shape != res_shape:
                raise TypeError(
                    "The shape of input and output arrays are inconsistent."
                    f"Expected output shape is {o1_shape}, got {out.shape}"
                )

            if ti._array_overlap(o1, out) or ti._array_overlap(o2, out):
                raise TypeError("Input and output arrays have memory overlap")

            if (
                dpctl.utils.get_execution_queue(
                    (o1.sycl_queue, o2.sycl_queue, out.sycl_queue)
                )
                is None
            ):
                raise TypeError(
                    "Input and output allocation queues are not compatible"
                )

        if isinstance(o1, dpt.usm_ndarray):
            src1 = o1
        else:
            src1 = dpt.asarray(o1, dtype=o1_dtype, sycl_queue=exec_q)
        if isinstance(o2, dpt.usm_ndarray):
            src2 = o2
        else:
            src2 = dpt.asarray(o2, dtype=o2_dtype, sycl_queue=exec_q)

        if buf1_dt is None and buf2_dt is None:
            if out is None:
                if order == "K":
                    out = _empty_like_pair_orderK(
                        src1, src2, res_dt, res_shape, res_usm_type, exec_q
                    )
                else:
                    if order == "A":
                        order = (
                            "F"
                            if all(
                                arr.flags.f_contiguous
                                for arr in (
                                    src1,
                                    src2,
                                )
                            )
                            else "C"
                        )
                    out = dpt.empty(
                        res_shape,
                        dtype=res_dt,
                        usm_type=res_usm_type,
                        sycl_queue=exec_q,
                        order=order,
                    )
            else:
                if res_dt != out.dtype:
                    raise TypeError(
                        f"Output array of type {res_dt} is needed,"
                        f"got {out.dtype}"
                    )

            src1 = dpt.broadcast_to(src1, res_shape)
            src2 = dpt.broadcast_to(src2, res_shape)
            ht_, _ = self.binary_fn_(
                src1=src1, src2=src2, dst=out, sycl_queue=exec_q
            )
            ht_.wait()
            return out
        elif buf1_dt is None:
            if order == "K":
                buf2 = _empty_like_orderK(src2, buf2_dt)
            else:
                if order == "A":
                    order = "F" if src1.flags.f_contiguous else "C"
                buf2 = dpt.empty_like(src2, dtype=buf2_dt, order=order)
            ht_copy_ev, copy_ev = ti._copy_usm_ndarray_into_usm_ndarray(
                src=src2, dst=buf2, sycl_queue=exec_q
            )
            if out is None:
                if order == "K":
                    out = _empty_like_pair_orderK(
                        src1, buf2, res_dt, res_shape, res_usm_type, exec_q
                    )
                else:
                    out = dpt.empty(
                        res_shape,
                        dtype=res_dt,
                        usm_type=res_usm_type,
                        sycl_queue=exec_q,
                        order=order,
                    )
            else:
                if res_dt != out.dtype:
                    raise TypeError(
                        f"Output array of type {res_dt} is needed,"
                        f"got {out.dtype}"
                    )

            src1 = dpt.broadcast_to(src1, res_shape)
            buf2 = dpt.broadcast_to(buf2, res_shape)
            ht_, _ = self.binary_fn_(
                src1=src1,
                src2=buf2,
                dst=out,
                sycl_queue=exec_q,
                depends=[copy_ev],
            )
            ht_copy_ev.wait()
            ht_.wait()
            return out
        elif buf2_dt is None:
            if order == "K":
                buf1 = _empty_like_orderK(src1, buf1_dt)
            else:
                if order == "A":
                    order = "F" if src1.flags.f_contiguous else "C"
                buf1 = dpt.empty_like(src1, dtype=buf1_dt, order=order)
            ht_copy_ev, copy_ev = ti._copy_usm_ndarray_into_usm_ndarray(
                src=src1, dst=buf1, sycl_queue=exec_q
            )
            if out is None:
                if order == "K":
                    out = _empty_like_pair_orderK(
                        buf1, src2, res_dt, res_shape, res_usm_type, exec_q
                    )
                else:
                    out = dpt.empty(
                        res_shape,
                        dtype=res_dt,
                        usm_type=res_usm_type,
                        sycl_queue=exec_q,
                        order=order,
                    )
            else:
                if res_dt != out.dtype:
                    raise TypeError(
                        f"Output array of type {res_dt} is needed,"
                        f"got {out.dtype}"
                    )

            buf1 = dpt.broadcast_to(buf1, res_shape)
            src2 = dpt.broadcast_to(src2, res_shape)
            ht_, _ = self.binary_fn_(
                src1=buf1,
                src2=src2,
                dst=out,
                sycl_queue=exec_q,
                depends=[copy_ev],
            )
            ht_copy_ev.wait()
            ht_.wait()
            return out

        if order in ["K", "A"]:
            if src1.flags.f_contiguous and src2.flags.f_contiguous:
                order = "F"
            elif src1.flags.c_contiguous and src2.flags.c_contiguous:
                order = "C"
            else:
                order = "C" if order == "A" else "K"
        if order == "K":
            buf1 = _empty_like_orderK(src1, buf1_dt)
        else:
            buf1 = dpt.empty_like(src1, dtype=buf1_dt, order=order)
        ht_copy1_ev, copy1_ev = ti._copy_usm_ndarray_into_usm_ndarray(
            src=src1, dst=buf1, sycl_queue=exec_q
        )
        if order == "K":
            buf2 = _empty_like_orderK(src2, buf2_dt)
        else:
            buf2 = dpt.empty_like(src2, dtype=buf2_dt, order=order)
        ht_copy2_ev, copy2_ev = ti._copy_usm_ndarray_into_usm_ndarray(
            src=src2, dst=buf2, sycl_queue=exec_q
        )
        if out is None:
            if order == "K":
                out = _empty_like_pair_orderK(
                    buf1, buf2, res_dt, res_shape, res_usm_type, exec_q
                )
            else:
                out = dpt.empty(
                    res_shape,
                    dtype=res_dt,
                    usm_type=res_usm_type,
                    sycl_queue=exec_q,
                    order=order,
                )
        else:
            if res_dt != out.dtype:
                raise TypeError(
                    f"Output array of type {res_dt} is needed, got {out.dtype}"
                )

        buf1 = dpt.broadcast_to(buf1, res_shape)
        buf2 = dpt.broadcast_to(buf2, res_shape)
        ht_, _ = self.binary_fn_(
            src1=buf1,
            src2=buf2,
            dst=out,
            sycl_queue=exec_q,
            depends=[copy1_ev, copy2_ev],
        )
        dpctl.SyclEvent.wait_for([ht_copy1_ev, ht_copy2_ev, ht_])
        return out

    def _inplace(self, lhs, val):
        if self.binary_inplace_fn_ is None:
            raise ValueError(
                f"In-place operation not supported for ufunc '{self.name_}'"
            )
        if not isinstance(lhs, dpt.usm_ndarray):
            raise TypeError(
                f"Expected dpctl.tensor.usm_ndarray, got {type(lhs)}"
            )
        q1, lhs_usm_type = _get_queue_usm_type(lhs)
        q2, val_usm_type = _get_queue_usm_type(val)
        if q2 is None:
            exec_q = q1
            usm_type = lhs_usm_type
        else:
            exec_q = dpctl.utils.get_execution_queue((q1, q2))
            if exec_q is None:
                raise ExecutionPlacementError(
                    "Execution placement can not be unambiguously inferred "
                    "from input arguments."
                )
            usm_type = dpctl.utils.get_coerced_usm_type(
                (
                    lhs_usm_type,
                    val_usm_type,
                )
            )
        dpctl.utils.validate_usm_type(usm_type, allow_none=False)
        lhs_shape = _get_shape(lhs)
        val_shape = _get_shape(val)
        if not all(
            isinstance(s, (tuple, list))
            for s in (
                lhs_shape,
                val_shape,
            )
        ):
            raise TypeError(
                "Shape of arguments can not be inferred. "
                "Arguments are expected to be "
                "lists, tuples, or both"
            )
        try:
            res_shape = _broadcast_shape_impl(
                [
                    lhs_shape,
                    val_shape,
                ]
            )
        except ValueError:
            raise ValueError(
                "operands could not be broadcast together with shapes "
                f"{lhs_shape} and {val_shape}"
            )
        if res_shape != lhs_shape:
            raise ValueError(
                f"output shape {lhs_shape} does not match "
                f"broadcast shape {res_shape}"
            )
        sycl_dev = exec_q.sycl_device
        lhs_dtype = lhs.dtype
        val_dtype = _get_dtype(val, sycl_dev)
        if not _validate_dtype(val_dtype):
            raise ValueError("Input operand of unsupported type")

        lhs_dtype, val_dtype = _resolve_weak_types(
            lhs_dtype, val_dtype, sycl_dev
        )

        buf_dt = _find_inplace_dtype(
            lhs_dtype, val_dtype, self.result_type_resolver_fn_, sycl_dev
        )

        if buf_dt is None:
            raise TypeError(
                f"In-place '{self.name_}' does not support input types "
                f"({lhs_dtype}, {val_dtype}), "
                "and the inputs could not be safely coerced to any "
                "supported types according to the casting rule ''safe''."
            )

        if isinstance(val, dpt.usm_ndarray):
            rhs = val
            overlap = ti._array_overlap(lhs, rhs)
        else:
            rhs = dpt.asarray(val, dtype=val_dtype, sycl_queue=exec_q)
            overlap = False

        if buf_dt == val_dtype and overlap is False:
            rhs = dpt.broadcast_to(rhs, res_shape)
            ht_, _ = self.binary_inplace_fn_(
                lhs=lhs, rhs=rhs, sycl_queue=exec_q
            )
            ht_.wait()

        else:
            buf = dpt.empty_like(rhs, dtype=buf_dt)
            ht_copy_ev, copy_ev = ti._copy_usm_ndarray_into_usm_ndarray(
                src=rhs, dst=buf, sycl_queue=exec_q
            )

            buf = dpt.broadcast_to(buf, res_shape)
            ht_, _ = self.binary_inplace_fn_(
                lhs=lhs,
                rhs=buf,
                sycl_queue=exec_q,
                depends=[copy_ev],
            )
            ht_copy_ev.wait()
            ht_.wait()

        return lhs

IntelPython / dpctl / 5579681118

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous