mindspore 2.2.14__cp39-cp39-win_amd64.whl → 2.4.0__cp39-cp39-win_amd64.whl

mindspore/hal/memory.py

@@ -0,0 +1,326 @@
+# Copyright 2024 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+
+"""Hardware memory interfaces."""
+from mindspore._c_expression import _memory_stats, _reset_max_mem_reserved, _reset_max_mem_allocated
+from mindspore import log as logger
+from .device import _check_inputs_validation, is_initialized
+
+
+@_check_inputs_validation
+def memory_stats(device_target=None):
+    """
+    Returns status information queried from the memory pool.
+
+    Note:
+        - If `device_target` is not specified, get the device capability of the current backend set by context.
+        - For the `CPU` backend, a dictionary with empty data is always returned.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Returns:
+        dict, the queried memory information.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.hal.memory_stats())
+        {'total_reserved_memory': 1073741824, 'total_allocated_memory': 1024, 'total_idle_memory': 1073740800,
+        'total_eager_free_memory': 0, 'max_reserved_memory': 1073741824, 'max_allocated_memory': 1536,
+        'common_mem_pool_stats': {'block_unit_size': 1073741824, 'block_counts': 1, 'blocks_info':
+        {<capsule object NULL at 0x7f7e8c27b030>: {'block_stream_id': 0, 'block_memory_size': 1073741824}}},
+        'persistent_mem_pool_stats': {'block_unit_size': 1073741824, 'block_counts': 0, 'blocks_info': {}}}
+    """
+    if not is_initialized(device_target):
+        logger.warning(f"Backend {device_target} is not initialized yet. Return empty dict.")
+        return {}
+    return _memory_stats(device_target)
+
+
+@_check_inputs_validation
+def memory_reserved(device_target=None):
+    """
+    Returns the total amount of memory currently managed by the memory pool.
+
+    Note:
+        - If `device_target` is not specified, get the device capability of the current backend set by context.
+        - For the `CPU` backend, 0 is always returned.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Returns:
+        int, in Byte.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.hal.memory_reserved())
+        1073741824
+    """
+    return _memory_stats(device_target).get("total_reserved_memory", 0)
+
+
+@_check_inputs_validation
+def max_memory_reserved(device_target=None):
+    """
+    Returns the peak value of the total memory managed by the memory pool since the process was started.
+
+    Note:
+        - If `device_target` is not specified, get the device capability of the current backend set by context.
+        - For the `CPU` backend, 0 is always returned.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Returns:
+        int, in Byte.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.hal.max_memory_reserved())
+        1073741824
+    """
+    return _memory_stats(device_target).get("max_reserved_memory", 0)
+
+
+@_check_inputs_validation
+def empty_cache():
+    """
+    Release all memory fragments in the memory pool, so that memory arrangement
+    will be optimized.
+
+    Note:
+        Currently, the MindSpore memory pool does not have the function of releasing memory fragments.
+        This interface is reserved but implemented as an empty method and prompted in log mode.
+    """
+    logger.warning(f"The empty_cache operation is currently not supported.")
+
+
+@_check_inputs_validation
+def reset_peak_memory_stats(device_target=None):
+    """
+    Reset the "peak" stats tracked by memory manager.
+
+    Note:
+        If `device_target` is not specified, get the device capability of the current backend set by context.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.hal.max_memory_reserved())
+        1073741824
+        >>> print(ms.hal.max_memory_allocated())
+        1536
+        >>> ms.hal.reset_peak_memory_stats()
+        >>> print(ms.hal.max_memory_reserved())
+        0
+        >>> print(ms.hal.max_memory_allocated())
+        0
+    """
+    _reset_max_mem_reserved(device_target)
+    _reset_max_mem_allocated(device_target)
+
+
+@_check_inputs_validation
+def memory_summary(device_target=None):
+    """
+    Returns readable memory pool status information.
+
+    Note:
+        If `device_target` is not specified, get the device capability of the current backend set by context.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Returns:
+        str, readable memory pool status information in tabular form.
+    """
+    stats = _memory_stats(device_target)
+
+    def _format_size(sz, pref_sz):
+        prefixes = ["B  ", "KB", "MB", "GB", "TB", "PB"]
+        prefix = prefixes[0]
+        for new_prefix in prefixes[1:]:
+            if pref_sz < 768 * 1024:
+                break
+            prefix = new_prefix
+            sz //= 1024
+            pref_sz /= 1024
+        return f"{sz:6d} {prefix}"
+
+    metrics_to_display = [
+        ("total_reserved_memory", "Reserved memory", _format_size),
+        ("total_allocatd_memory", "Allocated memory", _format_size),
+        ("total_idle_memory", "Idle memory", _format_size),
+        ("total_eager_free_memory", "Eager free memory", _format_size),
+        ("max_reserved_memory", "Max reserved memory", _format_size),
+        ("max_allocated_memory", "Max allocated memory", _format_size),
+    ]
+
+    lines = []
+    lines.append("=" * 45)
+    lines.append(" {:^43} ".format('Memory summary'))
+    lines.append("=" * 45)
+    lines.append(" {:<20} | {:<20} ".format('Metric', 'Data'))
+
+    for metric_key, metric_name, formatter in metrics_to_display:
+        lines.append("-" * 45)
+        data = stats[metric_key]
+        lines.append(" {:<20} | {:<20} ".format(metric_name, formatter(data, data)))
+
+    lines.append("=" * 45)
+
+    return "|" + "|\n|".join(lines) + "|\n"
+
+
+@_check_inputs_validation
+def memory_allocated(device_target=None):
+    """
+    Returns the actual memory size currently occupied by Tensor.
+
+    Note:
+        - If `device_target` is not specified, get the device capability of the current backend set by context.
+        - For the `CPU` backend, 0 is always returned.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Returns:
+        int, in Byte.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.hal.memory_allocated())
+        1024
+    """
+    return _memory_stats(device_target).get("total_allocatd_memory", 0)
+
+
+@_check_inputs_validation
+def max_memory_allocated(device_target=None):
+    """
+    Returns the peak memory size of the memory pool actually occupied by Tensor since the process was started.
+
+    Note:
+        - If `device_target` is not specified, get the device capability of the current backend set by context.
+        - For the `CPU` backend, 0 is always returned.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Returns:
+        int, in Byte.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.hal.max_memory_allocated())
+        1536
+    """
+    return _memory_stats(device_target).get("max_allocated_memory", 0)
+
+
+@_check_inputs_validation
+def reset_max_memory_reserved(device_target=None):
+    """
+    Reset the peak memory size managed by the memory pool.
+
+    Note:
+        If `device_target` is not specified, get the device capability of the current backend set by context.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.hal.max_memory_reserved())
+        1073741824
+        >>> ms.hal.reset_max_memory_reserved()
+        >>> print(ms.hal.max_memory_reserved())
+        0
+    """
+    _reset_max_mem_reserved(device_target)
+
+
+@_check_inputs_validation
+def reset_max_memory_allocated(device_target=None):
+    """
+    Reset the peak memory size of the memory pool actually occupied by Tensor.
+
+    Note:
+        If `device_target` is not specified, get the device capability of the current backend set by context.
+
+    Args:
+        device_target (str, optional): The device name of backend, should be one of "CPU", "GPU" and "Ascend".
+            Default value: ``None``.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1, 2]), ms.float32)
+        >>> b = Tensor(np.ones([1, 2]), ms.float32)
+        >>> c = ops.add(a, b).asnumpy()
+        >>> print(ms.hal.max_memory_allocated())
+        1536
+        >>> ms.hal.reset_max_memory_allocated()
+        >>> print(ms.hal.max_memory_allocated())
+        0
+    """
+    _reset_max_mem_allocated(device_target)

mindspore/hal/stream.py

@@ -0,0 +1,357 @@
+# Copyright 2023 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+"""Hal stream class"""
+from mindspore._c_expression import Stream as Stream_
+from mindspore._c_expression import set_cur_stream as set_cur_stream_
+from mindspore._c_expression import synchronize as synchronize_
+from mindspore._c_expression import current_stream as current_stream_
+from mindspore._c_expression import default_stream as default_stream_
+from mindspore._c_expression import communication_stream as communication_stream_
+
+from mindspore import _checkparam as Validator
+from .event import Event
+
+
+class Stream(Stream_):
+    r"""
+    Wrapper around a device stream.
+
+    A device stream is a linear sequence of execution that belongs to a specific device,
+    independent from other streams.
+
+    For a quick start of using Stream, please refer to `Illustration of stream management <https://www.mindspore.cn/docs/en/master/api_python/samples/hal/stream_manager.html>`_ .
+
+    Args:
+        priority (int, optional): priority of the stream, lower numbers represent higher priorities.
+            By default, streams have priority ``0``.
+        kwargs (dict): keyword arguments.
+    """
+    def __init__(self, priority=0, **kwargs):
+        self.init_finished = False
+        if 'stream' in kwargs and kwargs['stream'] is not None:
+            super().__init__(kwargs['stream'])
+
+        Validator.check_is_int(priority, 'priority', "Stream")
+        if 'stream_id' in kwargs:
+            Validator.check_is_int(kwargs['stream_id'], 'stream_id', "Stream")
+            super().__init__(priority, kwargs['stream_id'])
+        else:
+            super().__init__(priority)
+        self.init_finished = True
+
+    def record_event(self, event=None):
+        r"""
+        Records an event.
+
+        Args:
+            event (Event, optional): event to record. If not given, a new one
+                will be allocated.
+
+        Returns:
+            Event, recorded event. If this argument is ``None``, a new one will be allocated. Default is ``None``.
+
+        Raises:
+            TypeError: If 'event' is neither a :class:`mindspore.hal.Event` nor a ``None``.
+
+        Examples:
+            >>> import mindspore as ms
+            >>> import numpy as np
+            >>> from mindspore import Tensor, ops
+            >>> a = Tensor(np.ones([3, 3]), ms.float32)
+            >>> b = Tensor(np.ones([3, 3]), ms.float32)
+            >>> s1 = ms.hal.Stream()
+            >>> with ms.hal.StreamCtx(s1):
+            ...     c = a + b
+            ...     event = s1.record_event()
+            ...     d = a * b
+            >>> cur_stream = ms.hal.current_stream()
+            >>> cur_stream.wait_event(event)
+            >>> e = c + 3
+            >>> print(e)
+            [[5. 5. 5.]
+             [5. 5. 5.]
+             [5. 5. 5.]]
+        """
+        if event is None:
+            event = Event()
+        if not isinstance(event, Event):
+            raise TypeError(f"For 'record_event', the argument 'event' should be Event,"
+                            f" but got {type(event)}.")
+        event.record(self)
+        return event
+
+    def wait_event(self, event):
+        r"""
+        Makes all future work submitted to the stream wait for an event.
+
+        Args:
+            event (Event): an event to wait for.
+
+        Raises:
+            TypeError: If 'event' is not a :class:`mindspore.hal.Event`.
+
+        Examples:
+            >>> import mindspore as ms
+            >>> import numpy as np
+            >>> from mindspore import Tensor, ops
+            >>> a = Tensor(np.ones([3, 3]), ms.float32)
+            >>> b = Tensor(np.ones([3, 3]), ms.float32)
+            >>> s1 = ms.hal.Stream()
+            >>> with ms.hal.StreamCtx(s1):
+            ...     c = a + b
+            ...     event = s1.record_event()
+            ...     d = a * b
+            >>> cur_stream = ms.hal.current_stream()
+            >>> cur_stream.wait_event(event)
+            >>> e = c + 3
+            >>> print(e)
+            [[5. 5. 5.]
+             [5. 5. 5.]
+             [5. 5. 5.]]
+        """
+        if not isinstance(event, Event):
+            raise TypeError(f"For 'wait_event', the argument 'event' should be Event,"
+                            f" but got {type(event)}.")
+        event.wait(self)
+
+    def wait_stream(self, stream):
+        r"""
+        Synchronizes with another stream.
+
+        All future work submitted to this stream will wait until all kernels
+        submitted to a given stream at the time of call complete.
+
+        Args:
+            stream (Stream): a stream to synchronize.
+
+        Raises:
+            TypeError: If 'stream' is not a :class:`mindspore.hal.Stream`.
+
+        Examples:
+            >>> import mindspore as ms
+            >>> import numpy as np
+            >>> from mindspore import Tensor, ops
+            >>> s1 = ms.hal.Stream()
+            >>> s2 = ms.hal.Stream()
+            >>> a = Tensor(np.ones([1, 2]), ms.float32)
+            >>> b = Tensor(np.ones([2, 2]), ms.float32)
+            >>> with ms.hal.StreamCtx(s1):
+            ...     c = ops.matmul(a, b)
+            >>> with ms.hal.StreamCtx(s2):
+            ...     s2.wait_stream(s1)
+            ...     d = ops.matmul(c, b)
+            >>> ms.hal.synchronize()
+            >>> print(d)
+            [[4. 4.]]
+        """
+        if not isinstance(stream, Stream):
+            raise TypeError(f"For 'wait_stream', the argument 'stream' should be Stream,"
+                            f" but got {type(stream)}.")
+        self.wait_event(stream.record_event())
+
+    def synchronize(self):
+        r"""
+        Wait for all the kernels in this stream to complete.
+
+        Examples:
+            >>> import mindspore as ms
+            >>> import numpy as np
+            >>> from mindspore import Tensor, ops
+            >>> a = Tensor(np.ones([1024, 2048]), ms.float32)
+            >>> b = Tensor(np.ones([2048, 4096]), ms.float32)
+            >>> s1 = ms.hal.Stream()
+            >>> with ms.hal.StreamCtx(s1):
+            ...     c = ops.matmul(a, b)
+            >>> s1.synchronize()
+            >>> assert s1.query()
+        """
+        # pylint: disable=useless-super-delegation
+        super().synchronize()
+
+    def query(self):
+        r"""
+        Checks if all the work submitted has been completed.
+
+        Returns:
+            A boolean indicating if all kernels in this stream are completed.
+
+        Examples:
+            >>> import mindspore as ms
+            >>> import numpy as np
+            >>> from mindspore import Tensor, ops
+            >>> a = Tensor(np.ones([1024, 2048]), ms.float32)
+            >>> b = Tensor(np.ones([2048, 4096]), ms.float32)
+            >>> s1 = ms.hal.Stream()
+            >>> with ms.hal.StreamCtx(s1):
+            ...     c = ops.matmul(a, b)
+            >>> s1.synchronize()
+            >>> assert s1.query()
+        """
+        # pylint: disable=useless-super-delegation
+        return super().query()
+
+    def __eq__(self, other):
+        if not isinstance(other, Stream):
+            raise TypeError(f"For '__eq__', the argument 'other' should be Stream,"
+                            f" but got {type(other)}.")
+        return super().__eq__(other)
+
+    def __hash__(self):
+        return hash((self.id, self.device_id))
+
+    def __repr__(self):
+        if self.init_finished:
+            return super().__repr__()
+        return ''
+
+
+def synchronize():
+    r"""
+    Synchronize all streams on current device.(Each MindSpore process only occupies one device)
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1024, 2048]), ms.float32)
+        >>> b = Tensor(np.ones([2048, 4096]), ms.float32)
+        >>> s1 = ms.hal.Stream()
+        >>> with ms.hal.StreamCtx(s1):
+        ...     c = ops.matmul(a, b)
+        >>> ms.hal.synchronize()
+        >>> assert s1.query()
+    """
+    synchronize_()
+
+
+def set_cur_stream(stream):
+    r"""
+    Sets the current stream.This is a wrapper API to set the stream.
+    Usage of this function is discouraged in favor of the ``stream`` context manager.
+
+    Args:
+        stream (Stream): selected stream. This function is a no-op
+            if this argument is ``None``.
+
+    Raises:
+        TypeError: If 'stream' is neither a :class:`mindspore.hal.Stream` nor a ``None``.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> cur_stream = ms.hal.current_stream()
+        >>> assert cur_stream == ms.hal.default_stream()
+        >>> s1 = ms.hal.Stream()
+        >>> ms.hal.set_cur_stream(s1)
+        >>> assert ms.hal.current_stream() == s1
+        >>> ms.hal.set_cur_stream(ms.hal.default_stream())
+    """
+    if stream is None:
+        return
+    if not isinstance(stream, Stream):
+        raise TypeError(f"For 'set_cur_stream', the argument 'stream' should be Stream,"
+                        f" but got {type(stream)}.")
+    set_cur_stream_(stream)
+
+
+def current_stream():
+    r"""
+    Return current stream used on this device.
+
+    Returns:
+        stream (Stream), current stream.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> cur_stream = ms.hal.current_stream()
+        >>> assert cur_stream == ms.hal.default_stream()
+    """
+    return Stream(stream=current_stream_())
+
+
+def default_stream():
+    r"""
+    Return default stream on this device.
+
+    Returns:
+        stream (Stream), default stream.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> cur_stream = ms.hal.current_stream()
+        >>> assert cur_stream == ms.hal.default_stream()
+    """
+    return Stream(stream=default_stream_())
+
+
+def communication_stream():
+    r"""
+    Return default stream on this device.
+
+    Returns:
+        stream (Stream), default stream.
+
+    TODO
+    Examples:
+        >>> import mindspore as ms
+        >>> cur_stream = ms.hal.current_stream()
+        >>> assert cur_stream == ms.hal.default_stream()
+    """
+    return Stream(stream=communication_stream_())
+
+
+class StreamCtx():
+    r"""
+    Context-manager that selects a given stream.
+
+    All kernels queued within its context will be enqueued on a selected
+    stream.
+
+    Args:
+        ctx_stream (Stream): selected stream. This manager is a no-op if it's ``None``.
+
+    Raises:
+        TypeError: If 'stream' is neither a :class:`mindspore.hal.Stream` nor a ``None``.
+
+    Examples:
+        >>> import mindspore as ms
+        >>> import numpy as np
+        >>> from mindspore import Tensor, ops
+        >>> a = Tensor(np.ones([1024, 2048]), ms.float32)
+        >>> b = Tensor(np.ones([2048, 4096]), ms.float32)
+        >>> s1 = ms.hal.Stream()
+        >>> with ms.hal.StreamCtx(s1):
+        ...     c = ops.matmul(a, b)
+        >>> ms.hal.synchronize()
+        >>> assert s1.query()
+    """
+    def __init__(self, ctx_stream):
+        if ctx_stream is not None and not isinstance(ctx_stream, Stream):
+            raise TypeError(f"For 'StreamCtx', the argument 'ctx_stream' should be Stream,"
+                            f" but got {type(ctx_stream)}.")
+        self.stream = ctx_stream
+        self.prev_stream = None
+
+    def __enter__(self):
+        if self.stream is None:
+            return
+        self.prev_stream = current_stream()
+        set_cur_stream(self.stream)
+        return
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if self.stream is None:
+            return
+        set_cur_stream(self.prev_stream)
+        return

mindspore/include/api/data_type.h

@@ -38,8 +38,8 @@ enum class DataType : int {
   kNumberTypeFloat16 = 42,
   kNumberTypeFloat32 = 43,
   kNumberTypeFloat64 = 44,
-  kNumberTypeBFloat16 = 46,
-  kNumberTypeEnd = 53,
+  kNumberTypeBFloat16 = 45,
+  kNumberTypeEnd = 52,
   // add new enum here
   kInvalidType = INT32_MAX,
 };