mindstudio-probe 8.3.2__py3-none-any.whl → 26.0.0a1__py3-none-any.whl

msprobe/msaccucmp/dump_parse/proto_dump_data.py

@@ -0,0 +1,308 @@
+# coding=utf-8
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
+import json
+from enum import Enum
+import base64
+from typing import List, Any, Dict
+
+
+class EnumEncoder(json.JSONEncoder):
+    def default(self, o):
+        if isinstance(o, Enum):
+            return o.value
+        if isinstance(o, bytes):
+            return base64.b64encode(o).decode("ascii")
+        return super().default(o)
+
+
+class RepeatedField(list):
+    """模仿 protobuf repeated 字段，支持 add()"""
+    def __init__(self, cls_type):
+        super().__init__()
+        self._cls_type = cls_type
+
+    def add(self, **kwargs):
+        obj = self._cls_type(**kwargs)
+        self.append(obj)
+        return obj
+
+
+class Shape:
+    def __init__(self, dim=None):
+        self._dim = []
+        if dim:
+            self.extend(dim)
+
+    @property
+    def dim(self):
+        return self._dim
+
+    def __setattr__(self, name, value):
+        if name == "dim":
+            object.__setattr__(self, "_dim", [int(v) for v in value])
+        else:
+            object.__setattr__(self, name, value)
+
+    def append(self, value):
+        self._dim.append(int(value))
+
+    def extend(self, values):
+        self._dim.extend(int(v) for v in values)
+
+    def Clear(self):
+        self._dim.clear()
+
+    def to_dict(self):
+        return {"dim": self._dim}
+
+    def __repr__(self):
+        return f"Shape(dim={self._dim})"
+
+
+class DimRange:
+    def __init__(self, dim_start=0, dim_end=0):
+        self.dim_start = dim_start
+        self.dim_end = dim_end
+
+    def to_dict(self):
+        return {"dim_start": self.dim_start, "dim_end": self.dim_end}
+
+
+class OriginalOp:
+    def __init__(self, name="", output_index=0, data_type=0, data_format=0):
+        self.name = name
+        self.output_index = output_index
+        self.data_type = data_type
+        self.format = data_format
+
+    def to_dict(self):
+        return {
+            "name": self.name,
+            "output_index": self.output_index,
+            "data_type": self.data_type,
+            "format": self.format,
+        }
+
+
+class OpOutput():
+    def __init__(self):
+        self.data_type = 0
+        self.format = 0
+        self.shape = Shape()
+        self.original_op = OriginalOp()
+        self.data = b""
+        self.size = 0
+        self.original_shape = Shape()
+        self.sub_format = 0
+        self.address = 0
+        self.dim_range: List[DimRange] = []
+        self.offset = 0
+
+    def to_dict(self):
+        return {
+            "data_type": self.data_type,
+            "format": self.format,
+            "shape": self.shape.to_dict(),
+            "original_op": self.original_op.to_dict(),
+            "data": self.data,
+            "size": self.size,
+            "original_shape": self.original_shape.to_dict(),
+            "sub_format": self.sub_format,
+            "address": self.address,
+            "dim_range": [d.to_dict() for d in self.dim_range],
+            "offset": self.offset,
+        }
+
+
+class OpInput():
+    def __init__(self):
+        self.data_type = 0
+        self.format = 0
+        self.shape = Shape()
+        self.data = b""
+        self.size = 0
+        self.original_shape = Shape()
+        self.sub_format = 0
+        self.address = 0
+        self.offset = 0
+
+    def to_dict(self):
+        return {
+            "data_type": self.data_type,
+            "format": self.format,
+            "shape": self.shape.to_dict(),
+            "data": self.data,
+            "size": self.size,
+            "original_shape": self.original_shape.to_dict(),
+            "sub_format": self.sub_format,
+            "address": self.address,
+            "offset": self.offset,
+        }
+
+
+class OpBuffer:
+    def __init__(self):
+        self.buffer_type = 0
+        self.data = b""
+        self.size = 0
+
+    def to_dict(self):
+        return {"buffer_type": self.buffer_type, "data": self.data, "size": self.size}
+
+
+class OpAttr:
+    def __init__(self, name="", value=""):
+        self.name = name
+        self.value = value
+
+    def to_dict(self):
+        return {"name": self.name, "value": self.value}
+
+
+class Workspace:
+    def __init__(self, data_type=0):
+        self.type = data_type
+        self.data = b""
+        self.size = 0
+
+    def to_dict(self):
+        return {"type": self.type, "data": self.data, "size": self.size}
+
+
+class DumpData:
+    def __init__(self):
+        self.version = ""
+        self.dump_time = 0
+        self.output = RepeatedField(OpOutput)
+        self.input = RepeatedField(OpInput)
+        self.buffer = RepeatedField(OpBuffer)
+        self.op_name = ""
+        self.attr = RepeatedField(OpAttr)
+        self.space = RepeatedField(Workspace)
+    
+    @classmethod
+    def from_dict(cls, data: Dict):
+        obj = cls()
+        fill_dump_data(obj, data)
+        return obj
+
+    def to_dict(self):
+        return {
+            "version": self.version,
+            "dump_time": self.dump_time,
+            "op_name": self.op_name,
+            "output": [data.to_dict() for data in self.output],
+            "input": [data.to_dict() for data in self.input],
+            "buffer": [data.to_dict() for data in self.buffer],
+            "attr": [data.to_dict() for data in self.attr],
+            "space": [data.to_dict() for data in self.space],
+        }
+
+    def SerializeToString(self) -> bytes:
+        """模拟 protobuf 的序列化，返回 JSON bytes"""
+        return json.dumps(self.to_dict(), cls=EnumEncoder).encode("utf-8")
+
+    def ParseFromString(self, data: bytes):
+        """
+        模拟 protobuf 的 ParseFromString
+        :param data: bytes (JSON 格式)
+        :return: self
+        """
+        obj_dict = json.loads(data.decode("utf-8"))
+        # 用 from_dict 填充当前实例
+        fill_dump_data(self, obj_dict)
+        return self
+
+
+TYPE_MAP = {
+    "input": OpInput,
+    "output": OpOutput,
+    "buffer": OpBuffer,
+    "attr": OpAttr,
+    "space": Workspace,
+}
+
+
+def _handle_repeated_field(attr: Any, key: str, value_list: list) -> None:
+    """处理 repeated 字段：当 TYPE_MAP 有对应类时用 add() 并递归填充，否则 append 原始元素。"""
+    cls_type = TYPE_MAP.get(key)
+    if cls_type:
+        for item in value_list:
+            # 保持原行为：使用 add() 创建子对象
+            sub_obj = attr.add()  
+            # 只有 dict 才递归填充，否则跳过（与原行为一致）
+            if isinstance(item, dict):
+                fill_dump_data(sub_obj, item)
+    else:
+        # 保持原行为：逐项 append（而不是 extend），以保留原先每项可能的非 list 语义
+        for item in value_list:
+            attr.append(item)
+
+
+def _handle_shape(attr: "Shape", value: Any) -> None:
+    """填充 Shape 类型：支持 dict {'dim': [...]} 或直接 list."""
+    if isinstance(value, dict):
+        dims = value.get("dim")
+        if dims is None:
+            return
+        # 保证全部转换为 int
+        attr.dim = [int(v) for v in dims]
+    elif isinstance(value, list):
+        # 如果直接给了 list，也允许赋值（向后兼容）
+        attr.dim = [int(v) for v in value]
+
+
+def _handle_original_op(attr: "OriginalOp", value: Any) -> None:
+    """按字段赋值 OriginalOp（仅处理 dict）。"""
+    if not isinstance(value, dict):
+        return
+    for sub_key, sub_val in value.items():
+        if hasattr(attr, sub_key):
+            setattr(attr, sub_key, sub_val)
+
+
+def fill_dump_data(obj: Any, data: Dict) -> Any:
+    """
+    将字典 data 递归填充到 obj（DumpData 或子消息）上。
+    """
+    for key, value in data.items():
+        if not hasattr(obj, key):
+            # 字典中含有类没有的字段，跳过
+            continue
+
+        attr = getattr(obj, key)
+
+        # 1) repeated field (RepeatedField)
+        if isinstance(attr, RepeatedField) and isinstance(value, list):
+            _handle_repeated_field(attr, key, value)
+            continue
+
+        # 2) Shape 嵌套
+        if isinstance(attr, Shape):
+            _handle_shape(attr, value)
+            continue
+
+        # 3) OriginalOp 嵌套
+        if isinstance(attr, OriginalOp):
+            _handle_original_op(attr, value)
+            continue
+
+        # 4) 其它基本类型（bytes / int / str / list 等），直接赋值
+        setattr(obj, key, value)
+
+    return obj

msprobe/msaccucmp/dump_parser.py

@@ -0,0 +1,90 @@
+# coding=utf-8
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
+"""
+Function:
+This class mainly involves the main function.
+"""
+
+import argparse
+import sys
+import time
+
+from msprobe.msaccucmp.cmp_utils import log, file_utils
+from msprobe.msaccucmp.cmp_utils.utils import safe_path_string
+from msprobe.msaccucmp.cmp_utils.constant.compare_error import CompareError
+from msprobe.msaccucmp.dump_parse.dump_data_parser import DumpDataParser
+
+
+def _save_log_parser(save_log_parser: argparse.ArgumentParser) -> None:
+    save_log_parser.add_argument(
+        '-d', '--dump_file', dest='dump_path', default='', type=safe_path_string,
+        help='<Required> the dump file path, supports one AICPU custom operator dump file.',
+        required=True)
+    save_log_parser.add_argument('-out', '--output', dest='output_path', default='', type=safe_path_string,
+                                 help='<Optional> the output path')
+
+
+def _do_cmd() -> int:
+    parser = argparse.ArgumentParser()
+    subparsers = parser.add_subparsers(help='commands')
+    save_log_parser = subparsers.add_parser('save_log', help='Save AICPU custom operator log.')
+
+    _save_log_parser(save_log_parser)
+
+    args = parser.parse_args(sys.argv[1:])
+    if len(sys.argv) < 2:
+        parser.print_help()
+        raise CompareError(CompareError.MSACCUCMP_INVALID_PARAM_ERROR)
+
+    if sys.argv[1] == 'save_log':
+        args.dump_version = 2
+        args.output_file_type = None
+        ret = _do_save_log(args)
+        return ret
+    else:
+        return 0
+
+
+def _do_save_log(args: argparse.Namespace) -> int:
+    ret = DumpDataParser(args).parse_log_data()
+    return ret
+
+
+def main() -> None:
+    """
+    parse argument and run command
+    :return:
+    """
+    start = time.time()
+    with file_utils.UmaskWrapper():
+        try:
+            ret = _do_cmd()
+        except CompareError as err:
+            ret = err.code
+        except Exception as base_err:
+            log.print_error_log(f'Basic error running {sys.argv[0]}: {base_err}')
+            sys.exit(1)
+    end = time.time()
+    if ret != 0:
+        log.print_error_log("Failed to parse dump log.")
+    log.print_info_log(
+        'The command was completed and took %d seconds.' % (end - start))
+
+
+if __name__ == '__main__':
+    main()

msprobe/msaccucmp/format_manager/__init__.py

@@ -0,0 +1,16 @@
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+

msprobe/msaccucmp/format_manager/builtin_format_convert/convert_FRACTAL_NZ_to_NCHW.py

@@ -0,0 +1,53 @@
+# coding=utf-8
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
+"""
+Function:
+convert format from FRACTAL_NZ to NCHW.
+"""
+from functools import reduce
+
+import numpy as np
+
+
+def convert(shape_from: list, shape_to: list, array: any) -> any:
+    """
+    Convert the data format from FRACTAL_NZ to NCHW
+    :param shape_from: the shape before convert
+    :param shape_to: the shape after convert
+    :param array: the one-dimensional array
+    :return: the data array of NCHW shape
+    """
+    if len(shape_to) == 1:
+        axis_h, axis_n, axis_c = 1, 1, shape_to[0]
+    elif len(shape_to) == 2:
+        axis_h, axis_n, axis_c = 1, shape_to[0], shape_to[1]
+    else:
+        axis_h, axis_n, axis_c = reduce(lambda x, y: x * y, shape_to[:-2]), shape_to[-2], shape_to[-1]
+    axis_c0 = shape_from[-1]
+    axis_ni = shape_from[-2]
+    axis_no = shape_from[-3]
+    axis_c1 = shape_from[-4]
+    c_pad = None if axis_c1 * axis_c0 == axis_c else axis_c - axis_c1 * axis_c0
+    n_pad = None if axis_no * axis_ni == axis_n else axis_n - axis_no * axis_ni
+    tmp_input_tensor = array.reshape(axis_h, axis_c1, axis_no, axis_ni, axis_c0)
+    tmp_input_tensor = np.transpose(tmp_input_tensor, axes=(0, 2, 3, 1, 4))
+    tmp_input_tensor = tmp_input_tensor.reshape((axis_h, axis_no * axis_ni, axis_c1 * axis_c0))
+    data_y = tmp_input_tensor[:, :n_pad, :c_pad]
+    if len(shape_to) <= 2:
+        data_y = data_y.reshape([data_y.shape[1], data_y.shape[2]])
+    return data_y

msprobe/msaccucmp/format_manager/builtin_format_convert/convert_FRACTAL_NZ_to_ND.py

@@ -0,0 +1,52 @@
+# coding=utf-8
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
+"""
+Function:
+convert format from FRACTAL_NZ to ND.
+"""
+from functools import reduce
+import numpy as np
+
+
+def convert(shape_from: list, shape_to: list, array: any) -> any:
+    """
+    Convert the data format from FRACTAL_NZ to ND
+    :param shape_from: the shape before convert
+    :param shape_to: the shape after convert
+    :param array: the one-dimensional array
+    :return: the data array of ND shape
+    """
+    if len(shape_to) == 1:
+        axis_h, axis_n, axis_c = 1, 1, shape_to[0]
+    elif len(shape_to) == 2:
+        axis_h, axis_n, axis_c = 1, shape_to[0], shape_to[1]
+    else:
+        axis_h, axis_n, axis_c = reduce(lambda x, y: x * y, shape_to[:-2]), shape_to[-2], shape_to[-1]
+    axis_c1 = shape_from[-4]
+    axis_no = shape_from[-3]
+    axis_ni = shape_from[-2]
+    axis_c0 = shape_from[-1]
+    c_pad = None if axis_c1 * axis_c0 == axis_c else axis_c - axis_c1 * axis_c0
+    n_pad = None if axis_no * axis_ni == axis_n else axis_n - axis_no * axis_ni
+    tmp_input_tensor = array.reshape(axis_h, axis_c1, axis_no, axis_ni, axis_c0)
+    tmp_input_tensor = np.transpose(tmp_input_tensor, axes=(0, 2, 3, 1, 4))
+    tmp_input_tensor = tmp_input_tensor.reshape((axis_h, axis_no * axis_ni, axis_c1 * axis_c0))
+    data_y = tmp_input_tensor[:, :n_pad, :c_pad]
+    if len(shape_to) <= 2:
+        data_y = data_y.reshape([data_y.shape[1], data_y.shape[2]])
+    return data_y

msprobe/msaccucmp/format_manager/builtin_format_convert/convert_FRACTAL_NZ_to_NHWC.py

@@ -0,0 +1,53 @@
+# coding=utf-8
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
+"""
+Function:
+convert format from FRACTAL_NZ to NHWC.
+"""
+
+from functools import reduce
+import numpy as np
+
+
+def convert(shape_from: list, shape_to: list, array: any) -> any:
+    """
+    Convert the data format from FRACTAL_NZ to NHWC
+    :param shape_from: the shape before convert
+    :param shape_to: the shape after convert
+    :param array: the one-dimensional array
+    :return: the data array of NHWC shape
+    """
+    if len(shape_to) == 1:
+        axis_h, axis_n, axis_c = 1, 1, shape_to[0]
+    elif len(shape_to) == 2:
+        axis_h, axis_n, axis_c = 1, shape_to[0], shape_to[1]
+    else:
+        axis_h, axis_n, axis_c = reduce(lambda x, y: x * y, shape_to[:-2]), shape_to[-2], shape_to[-1]
+    axis_ni = shape_from[-2]
+    axis_no = shape_from[-3]
+    axis_c0 = shape_from[-1]
+    axis_c1 = shape_from[-4]
+    n_pad = None if axis_no * axis_ni == axis_n else axis_n - axis_no * axis_ni
+    c_pad = None if axis_c1 * axis_c0 == axis_c else axis_c - axis_c1 * axis_c0
+    tmp_input_tensor = array.reshape(axis_h, axis_c1, axis_no, axis_ni, axis_c0)
+    tmp_input_tensor = np.transpose(tmp_input_tensor, axes=(0, 2, 3, 1, 4))
+    tmp_input_tensor = tmp_input_tensor.reshape((axis_h, axis_no * axis_ni, axis_c1 * axis_c0))
+    data_y = tmp_input_tensor[:, :n_pad, :c_pad]
+    if len(shape_to) <= 2:
+        data_y = data_y.reshape([data_y.shape[1], data_y.shape[2]])
+    return data_y

msprobe/msaccucmp/format_manager/builtin_format_convert/convert_FRACTAL_Z_to_HWCN.py

@@ -0,0 +1,47 @@
+# coding=utf-8
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
+"""
+Function:
+convert format from FRACTAL_Z to HWCN.
+"""
+import numpy as np
+
+
+def convert(shape_from: list, shape_to: list, array: any) -> any:
+    """
+    Convert the data format from FRACTAL_Z to HWCN
+    :param shape_from: the shape before convert
+    :param shape_to: the shape after convert
+    :param array: the one-dimensional array
+    :return: the data array of HWCN shape
+    """
+    axis_c = shape_to[2]
+    axis_n = shape_to[3]
+    axis_no = shape_from[1]
+    axis_ni = shape_from[2]
+    axis_h = shape_to[0]
+    axis_w = shape_to[1]
+    axis_c1 = shape_from[0] // (axis_h * axis_w)
+    axis_c0 = shape_from[3]
+    c_pad = None if axis_c1 * axis_c0 == axis_c else axis_c - axis_c1 * axis_c0
+    n_pad = None if axis_no * axis_ni == axis_n else axis_n - axis_no * axis_ni
+    tmp_input_tensor = array.reshape(axis_c1, axis_h, axis_w, axis_no, axis_ni, axis_c0)
+    # transpose the shape from (c1,h,w,no,ni,c0) to (h,w,c1,c0,no,ni)
+    tmp_input_tensor = np.transpose(tmp_input_tensor, (1, 2, 0, 5, 3, 4))
+    tmp_input_tensor = tmp_input_tensor.reshape((axis_h, axis_w, axis_c1 * axis_c0, axis_no * axis_ni))
+    return tmp_input_tensor[:, :, :c_pad, :n_pad]

msprobe/msaccucmp/format_manager/builtin_format_convert/convert_FRACTAL_Z_to_NCHW.py

@@ -0,0 +1,47 @@
+# coding=utf-8
+# -------------------------------------------------------------------------
+#  This file is part of the MindStudio project.
+# Copyright (c) 2025 Huawei Technologies Co.,Ltd.
+#
+# MindStudio is licensed under Mulan PSL v2.
+# You can use this software according to the terms and conditions of the Mulan PSL v2.
+# You may obtain a copy of Mulan PSL v2 at:
+#
+#          http://license.coscl.org.cn/MulanPSL2
+#
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND,
+# EITHER EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT,
+# MERCHANTABILITY OR FIT FOR A PARTICULAR PURPOSE.
+# See the Mulan PSL v2 for more details.
+# -------------------------------------------------------------------------
+
+"""
+Function:
+convert format from FRACTAL_Z to NCHW.
+"""
+import numpy as np
+
+
+def convert(shape_from: list, shape_to: list, array: any) -> any:
+    """
+    Convert the data format from FRACTAL_Z to NCHW
+    :param shape_from: the shape before convert
+    :param shape_to: the shape after convert
+    :param array: the one-dimensional array
+    :return: the data array of NCHW shape
+    """
+    axis_c = shape_to[1]
+    axis_n = shape_to[0]
+    axis_no = shape_from[1]
+    axis_ni = shape_from[2]
+    axis_h = shape_to[2]
+    axis_w = shape_to[3]
+    axis_c1 = shape_from[0] // (axis_h * axis_w)
+    axis_c0 = shape_from[3]
+    c_pad = None if axis_c1 * axis_c0 == axis_c else axis_c - axis_c1 * axis_c0
+    n_pad = None if axis_no * axis_ni == axis_n else axis_n - axis_no * axis_ni
+    tmp_input_tensor = array.reshape(axis_c1, axis_h, axis_w, axis_no, axis_ni, axis_c0)
+    # transpose the shape from (c1,h,w,no,ni,c0) to (no,ni,c1,c0,h,w)
+    tmp_input_tensor = np.transpose(tmp_input_tensor, (3, 4, 0, 5, 1, 2))
+    tmp_input_tensor = tmp_input_tensor.reshape((axis_no * axis_ni, axis_c1 * axis_c0, axis_h, axis_w))
+    return tmp_input_tensor[:n_pad, :c_pad, :, :, ]