Star1369
1460
代码
代码
Issues119
Pull Requests292
Actions
项目讨论
Wiki
分析
Star1369
1460
已合并
Math仓算子新增Golden文件，方便算子质量看护和精度测试。 #1782
ExerrCise创建于 3月20日
Math仓算子新增Golden文件，方便算子质量看护和精度测试。 #1782
已合并
ExerrCise创建于 3月20日
从moonw0lf/ops-mathdl_ttk_to_math合入到cann/ops-mathmaster
共 15 个文件变更+1004-0
Aconversion/clip_by_value/tests/assets/golden.py+44-0
@@ -0,0 +1,44 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "clip_by_value": "clip_by_value_golden"
18+ }
19+}
20+ 
21+ 
22+def clip_by_value_golden(x, clip_value_min, clip_value_max, **kwargs):
23+ '''
24+ Golden function for clip_by_value.
25+ All the parameters (names and order) follow @clip_by_value_def.cpp without outputs.
26+ All the input Tensors are numpy.ndarray.
27+ 
28+ Args:
29+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
30+ full_soc_version, short_soc_version, testcase_name
31+ 
32+ Returns:
33+ Output tensor
34+ '''
35+ if "bfloat16" in str(x.dtype):
36+ x = x.astype("float32")
37+ clip_value_min = clip_value_min.astype("float32")
38+ clip_value_max = clip_value_max.astype("float32")
39+ min_ = np.minimum(x, clip_value_max)
40+ res = np.maximum(min_, clip_value_min)
41+ if "bfloat16" in str(x.dtype):
42+ return res.astype(x.dtype, copy=False)
43+ return res
44+ 
Aconversion/clip_by_value_v2/tests/assets/golden.py+43-0
@@ -0,0 +1,43 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "clip_by_value_v2": "clip_by_value_v2_golden"
18+ }
19+}
20+ 
21+ 
22+def clip_by_value_v2_golden(x, clip_value_min, clip_value_max, **kwargs):
23+ '''
24+ Golden function for clip_by_value_v2.
25+ All the parameters (names and order) follow @clip_by_value_v2_def.cpp without outputs.
26+ All the input Tensors are numpy.ndarray.
27+ 
28+ Args:
29+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
30+ full_soc_version, short_soc_version, testcase_name
31+ 
32+ Returns:
33+ Output tensor
34+ '''
35+ if "bfloat16" in str(x.dtype):
36+ x = x.astype("float32")
37+ clip_value_min = clip_value_min.astype("float32")
38+ clip_value_max = clip_value_max.astype("float32")
39+ max_ = np.maximum(x, clip_value_min)
40+ res = np.minimum(max_, clip_value_max)
41+ if "bfloat16" in str(x.dtype):
42+ return res.astype(x.dtype, copy=False)
43+ return res
Aconversion/concat/tests/assets/golden.py+83-0
@@ -0,0 +1,83 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "concat": "concat_golden"
18+ }
19+}
20+ 
21+ 
22+def update_axis_for_hw_inner_format(ori_shape, axis, input_format, ori_format, reduce_mode=False):
23+ if input_format in ("NDC1HWC0", "NC1HWC0"):
24+ ori_shape_len = len(ori_shape) if -2 not in ori_shape else len(ori_format)
25+ axis = axis % ori_shape_len
26+ offset_6hd = 1 if input_format == "NDC1HWC0" else 0
27+ format_c_axis = 1 + offset_6hd if not reduce_mode else [1 + offset_6hd, 4 + offset_6hd]
28+ format_axis_map = {
29+ "N": 0,
30+ "C": format_c_axis,
31+ "H": 2 + offset_6hd,
32+ "W": 3 + offset_6hd,
33+ "D": 1
34+ }
35+ concat_dim_name = ori_format[axis]
36+ axis = format_axis_map[concat_dim_name]
37+ 
38+ if input_format in ("FRACTAL_NZ",):
39+ axis = axis % len(ori_shape)
40+ if axis == len(ori_shape) - 1:
41+ axis = len(ori_shape) - 2 if not reduce_mode else [len(ori_shape) - 2, len(ori_shape) + 1]
42+ elif axis == len(ori_shape) - 2:
43+ axis = len(ori_shape) - 1 if not reduce_mode else [len(ori_shape) - 1, len(ori_shape) + 0]
44+ 
45+ if input_format in ("FRACTAL_Z", "FRACTAL_Z_3D"):
46+ axis = axis % len(ori_shape)
47+ offset_3d = 1 if input_format == "FRACTAL_Z_3D" else 0
48+ format_c_axis = 0 + offset_3d if not reduce_mode else [0 + offset_3d, 5 + offset_3d]
49+ format_n_axis = 3 + offset_3d if not reduce_mode else [3 + offset_3d, 4 + offset_3d]
50+ format_axis_map = {
51+ "N": format_n_axis,
52+ "C": format_c_axis,
53+ "H": 1 + offset_3d,
54+ "W": 2 + offset_3d,
55+ "D": 0
56+ }
57+ concat_dim_name = ori_format[axis]
58+ axis = format_axis_map[concat_dim_name]
59+ 
60+ return axis
61+ 
62+ 
63+def concat_golden(concat_dim, x, *, N=1, **kwargs):
64+ '''
65+ Golden function for concat.
66+ All the parameters (names and order) follow @concat_def.cpp without outputs.
67+ All the input Tensors are numpy.ndarray.
68+ 
69+ Args:
70+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
71+ full_soc_version, short_soc_version, testcase_name
72+ 
73+ Returns:
74+ Output tensor
75+ '''
76+ x_arrays = list(x)
77+ 
78+ ori_shape = kwargs.get('input_ori_shapes', [x[0].shape])[0]
79+ input_formats = kwargs.get('input_formats', ['ND'])
80+ input_ori_formats = kwargs.get('input_ori_formats', ['ND'])
81+
82+ concat_dim = update_axis_for_hw_inner_format(ori_shape, concat_dim, input_formats[0], input_ori_formats[0])
83+ return numpy.concatenate(x_arrays, axis=concat_dim)
Aconversion/concat_d/tests/assets/golden.py+83-0
@@ -0,0 +1,83 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "concat_d": "concat_d_golden"
18+ }
19+}
20+ 
21+ 
22+def update_axis_for_hw_inner_format(ori_shape, axis, input_format, ori_format, reduce_mode=False):
23+ if input_format in ("NDC1HWC0", "NC1HWC0"):
24+ ori_shape_len = len(ori_shape) if -2 not in ori_shape else len(ori_format)
25+ axis = axis % ori_shape_len
26+ offset_6hd = 1 if input_format == "NDC1HWC0" else 0
27+ format_c_axis = 1 + offset_6hd if not reduce_mode else [1 + offset_6hd, 4 + offset_6hd]
28+ format_axis_map = {
29+ "N": 0,
30+ "C": format_c_axis,
31+ "H": 2 + offset_6hd,
32+ "W": 3 + offset_6hd,
33+ "D": 1
34+ }
35+ concat_dim_name = ori_format[axis]
36+ axis = format_axis_map[concat_dim_name]
37+ 
38+ if input_format in ("FRACTAL_NZ",):
39+ axis = axis % len(ori_shape)
40+ if axis == len(ori_shape) - 1:
41+ axis = len(ori_shape) - 2 if not reduce_mode else [len(ori_shape) - 2, len(ori_shape) + 1]
42+ elif axis == len(ori_shape) - 2:
43+ axis = len(ori_shape) - 1 if not reduce_mode else [len(ori_shape) - 1, len(ori_shape) + 0]
44+ 
45+ if input_format in ("FRACTAL_Z", "FRACTAL_Z_3D"):
46+ axis = axis % len(ori_shape)
47+ offset_3d = 1 if input_format == "FRACTAL_Z_3D" else 0
48+ format_c_axis = 0 + offset_3d if not reduce_mode else [0 + offset_3d, 5 + offset_3d]
49+ format_n_axis = 3 + offset_3d if not reduce_mode else [3 + offset_3d, 4 + offset_3d]
50+ format_axis_map = {
51+ "N": format_n_axis,
52+ "C": format_c_axis,
53+ "H": 1 + offset_3d,
54+ "W": 2 + offset_3d,
55+ "D": 0
56+ }
57+ concat_dim_name = ori_format[axis]
58+ axis = format_axis_map[concat_dim_name]
59+ 
60+ return axis
61+ 
62+ 
63+def concat_d_golden(x, *, concat_dim, N=1, **kwargs):
64+ '''
65+ Golden function for concat.
66+ All the parameters (names and order) follow @concat_d_def.cpp without outputs.
67+ All the input Tensors are numpy.ndarray.
D
Ddengguojie3月24日

@concat_d_def.cpp

likedislike
68+ 
69+ Args:
70+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
71+ full_soc_version, short_soc_version, testcase_name
72+ 
73+ Returns:
74+ Output tensor
75+ '''
76+ x_arrays = list(x)
77+ 
78+ ori_shape = kwargs.get('input_ori_shapes', [x[0].shape])[0]
79+ input_formats = kwargs.get('input_formats', ['ND'])
80+ input_ori_formats = kwargs.get('input_ori_formats', ['ND'])
81+
82+ concat_dim = update_axis_for_hw_inner_format(ori_shape, concat_dim, input_formats[0], input_ori_formats[0])
83+ return numpy.concatenate(x_arrays, axis=concat_dim)
Aconversion/concat_v2/tests/assets/golden.py+83-0
@@ -0,0 +1,83 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "concat_v2": "concat_v2_golden"
18+ }
19+}
20+ 
21+ 
22+def update_axis_for_hw_inner_format(ori_shape, axis, input_format, ori_format, reduce_mode=False):
23+ if input_format in ("NDC1HWC0", "NC1HWC0"):
24+ ori_shape_len = len(ori_shape) if -2 not in ori_shape else len(ori_format)
25+ axis = axis % ori_shape_len
26+ offset_6hd = 1 if input_format == "NDC1HWC0" else 0
27+ format_c_axis = 1 + offset_6hd if not reduce_mode else [1 + offset_6hd, 4 + offset_6hd]
28+ format_axis_map = {
29+ "N": 0,
30+ "C": format_c_axis,
31+ "H": 2 + offset_6hd,
32+ "W": 3 + offset_6hd,
33+ "D": 1
34+ }
35+ concat_dim_name = ori_format[axis]
36+ axis = format_axis_map[concat_dim_name]
37+ 
38+ if input_format in ("FRACTAL_NZ",):
39+ axis = axis % len(ori_shape)
40+ if axis == len(ori_shape) - 1:
41+ axis = len(ori_shape) - 2 if not reduce_mode else [len(ori_shape) - 2, len(ori_shape) + 1]
42+ elif axis == len(ori_shape) - 2:
43+ axis = len(ori_shape) - 1 if not reduce_mode else [len(ori_shape) - 1, len(ori_shape) + 0]
44+ 
45+ if input_format in ("FRACTAL_Z", "FRACTAL_Z_3D"):
46+ axis = axis % len(ori_shape)
47+ offset_3d = 1 if input_format == "FRACTAL_Z_3D" else 0
48+ format_c_axis = 0 + offset_3d if not reduce_mode else [0 + offset_3d, 5 + offset_3d]
49+ format_n_axis = 3 + offset_3d if not reduce_mode else [3 + offset_3d, 4 + offset_3d]
50+ format_axis_map = {
51+ "N": format_n_axis,
52+ "C": format_c_axis,
53+ "H": 1 + offset_3d,
54+ "W": 2 + offset_3d,
55+ "D": 0
56+ }
57+ concat_dim_name = ori_format[axis]
58+ axis = format_axis_map[concat_dim_name]
59+ 
60+ return axis
61+ 
62+ 
63+def concat_v2_golden(x, concat_dim, *, N=1, **kwargs):
64+ '''
65+ Golden function for concat.
66+ All the parameters (names and order) follow @concat_def.cpp without outputs.
67+ All the input Tensors are numpy.ndarray.
68+ 
69+ Args:
70+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
71+ full_soc_version, short_soc_version, testcase_name
72+ 
73+ Returns:
74+ Output tensor
75+ '''
76+ x_arrays = list(x)
77+ 
78+ ori_shape = kwargs.get('input_ori_shapes', [x[0].shape])[0]
79+ input_formats = kwargs.get('input_formats', ['ND'])
80+ input_ori_formats = kwargs.get('input_ori_formats', ['ND'])
81+
82+ concat_dim = update_axis_for_hw_inner_format(ori_shape, concat_dim, input_formats[0], input_ori_formats[0])
83+ return numpy.concatenate(x_arrays, axis=concat_dim)
Aconversion/depth_to_space/tests/ut/assets/golden.py+57-0
@@ -0,0 +1,57 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "depth_to_space": "depth_to_space_golden"
18+ }
19+}
20+ 
21+ 
22+def depth_to_space_golden(x, *, block_size, mode, data_format, **kwargs):
23+ '''
24+ Golden function for depth_to_space.
25+ All the parameters (names and order) follow @depth_to_space_def.cpp without outputs.
26+ All the input Tensors are numpy.ndarray.
27+ 
28+ Args:
29+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
30+ full_soc_version, short_soc_version, testcase_name
31+ 
32+ Returns:
33+ Output tensor
34+ '''
35+ shapes = x.shape
36+ 
37+ if data_format == "NCHW":
38+ n, c, h, w = shapes
39+ output_shapes = [n, c // (block_size ** 2), h * block_size, w * block_size]
40+ if mode == "CRD":
41+ input_shapes = [n, c // (block_size ** 2), block_size, block_size, h, w]
42+ perm = [0, 1, 4, 2, 5, 3]
43+ else: # mode == "DCR"
44+ input_shapes = [n, block_size, block_size, c // (block_size ** 2), h, w]
45+ perm = [0, 3, 4, 1, 5, 2]
46+ else: # data_format == "NHWC":
47+ n, h, w, c = shapes
48+ output_shapes = [n, h * block_size, w * block_size, c // (block_size ** 2)]
49+ if mode == "CRD":
50+ input_shapes = [n, h, w, c // (block_size ** 2), block_size, block_size]
51+ perm = [0, 1, 4, 2, 5, 3]
52+ else: # mode == "DCR"
53+ input_shapes = [n, h, w, block_size, block_size, c // (block_size ** 2)]
54+ perm = [0, 1, 3, 2, 4, 5]
55+ tmp = x.reshape(input_shapes)
56+ tmp = numpy.transpose(tmp, perm)
57+ return tmp.reshape(output_shapes)
Aconversion/diag_v2/tests/assets/golden.py+36-0
@@ -0,0 +1,36 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "diag_v2": "diag_v2_golden"
18+ }
19+}
20+ 
21+ 
22+def diag_v2_golden(x, *, diagonal, **kwargs):
23+ '''
24+ Golden function for diag_v2.
25+ All the parameters (names and order) follow @diag_v2_def.cpp without outputs.
26+ All the input Tensors are numpy.ndarray.
27+ 
28+ Args:
29+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
30+ full_soc_version, short_soc_version, testcase_name
31+ 
32+ Returns:
33+ Output tensor
34+ '''
35+
36+ return np.diag(x, k=diagonal)
Amath/bias_add/tests/assets/golden.py+48-0
@@ -0,0 +1,48 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "bias_add": "bias_add_golden"
18+ }
19+}
20+ 
21+ 
22+def bias_add_golden(x, bias, *, data_format=None, **kwargs):
23+ '''
24+ Golden function for bias_add.
25+ All the parameters (names and order) follow @bias_add_def.cpp without outputs.
26+ All the input Tensors are numpy.ndarray.
27+ 
28+ Args:
29+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
30+ full_soc_version, short_soc_version, testcase_name
31+ 
32+ Returns:
33+ Output tensor
34+ '''
35+ import tensorflow as tf
36+ 
37+ tf.compat.v1.disable_eager_execution()
38+ x_input = tf.compat.v1.placeholder(shape=x.shape, dtype=x.dtype)
39+ bias_input = tf.compat.v1.placeholder(shape=bias.shape, dtype=bias.dtype)
40+ 
41+ out = tf.nn.bias_add(x_input, bias_input, data_format=data_format, name="biasadd")
42+ feed_dict = {x_input: x, bias_input: bias}
43+ init_op = tf.compat.v1.global_variables_initializer()
44+ 
45+ with tf.compat.v1.Session() as sess:
46+ sess.run(init_op)
47+ res = sess.run(out, feed_dict=feed_dict)
48+ return res.astype(kwargs['output_dtypes'][0])
Amath/bias_add_grad/tests/assets/golden.py+85-0
@@ -0,0 +1,85 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "bias_add_grad": "bias_add_grad_golden"
18+ }
19+}
20+ 
21+ 
22+def _infer_axes(input_data_format, data_format, shape):
23+ g_shape_list = []
24+ if input_data_format == 'FRACTAL_NZ':
25+ if data_format == "NCHW":
26+ if len(shape) == 4:
27+ for i in range(-1 * len(shape), 0):
28+ if i not in (-1, -4):
29+ g_shape_list += [i + len(shape)]
30+ elif len(shape) == 5:
31+ for i in range(-1 * len(shape), 0):
32+ if i not in (-2, -3):
33+ g_shape_list += [i + len(shape)]
34+ else:
35+ g_shape_list.append(0)
36+ for i in range(2, len(shape)):
37+ g_shape_list = g_shape_list + [i]
38+ else:
39+ if len(shape) < 4:
40+ raise RuntimeError("cce_bias_add_grad_nz_2_nhwc only support shape larger than 4D")
41+ for i in range(-1 * len(shape), 0):
42+ if i not in (-1, -4):
43+ g_shape_list += [i + len(shape)]
44+ elif input_data_format in ("FRACTAL_Z", "FRACTAL_Z_3D", "NC1HWC0", "NDC1HWC0"):
45+ if input_data_format == "FRACTAL_Z":
46+ g_shape_list = [1, 2, 3, 4]
47+ elif input_data_format == "FRACTAL_Z_3D":
48+ g_shape_list = [0, 2, 3, 4, 5]
49+ elif input_data_format == "NC1HWC0":
50+ g_shape_list = [0, 2, 3]
51+ elif input_data_format == "NDC1HWC0":
52+ g_shape_list = [0, 1, 3, 4]
53+ else:
54+ if data_format == "NCHW":
55+ g_shape_list = [0]
56+ for i in range(2, len(shape)):
57+ g_shape_list += [i]
58+ else:
59+ if len(shape) < 2:
60+ raise RuntimeError("cce_bias_add_grad only support shape larger than 2D")
61+ g_shape_list = [x for x in range(len(shape) - 1)]
62+ return g_shape_list
63+ 
64+ 
65+def __eliminate_duplicate_axes(axis, x):
66+ axis = tuple(set([_ax if _ax >= 0 else len(x.shape) + _ax for _ax in axis]))
67+ return axis
68+ 
69+ 
70+def bias_add_grad_golden(x, *, data_format, **kwargs):
71+ '''
72+ Golden function for bias_add_grad.
73+ All the parameters (names and order) follow @bias_add_grad_def.cpp without outputs.
74+ All the input Tensors are numpy.ndarray.
75+ 
76+ Args:
77+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
78+ full_soc_version, short_soc_version, testcase_name
79+ 
80+ Returns:
81+ Output tensor
82+ '''
83+ actual_formats = kwargs.get('input_formats', ['ND'])
84+ axis = __eliminate_duplicate_axes(_infer_axes(actual_formats[0], data_format, x.shape), x)
85+ return np.sum(x, axis=axis, dtype="float64")
Amath/bincount/tests/assets/golden.py+74-0
@@ -0,0 +1,74 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "bincount": "bincount_golden"
18+ }
19+}
20+ 
21+ 
22+def numpy_to_torch_tensor(np_array):
23+ import torch
24+ if np_array is None:
25+ return None
26+ np_dtype = np_array.dtype.name
27+ if "bfloat16" in np_dtype:
28+ np_int16 = np_array.view(dtype=np.int16)
29+ t_int16 = torch.from_numpy(np_int16)
30+ return t_int16.view(torch.bfloat16)
31+ else:
32+ return torch.from_numpy(np_array)
33+ 
34+ 
35+def torch_to_numpy_tensor(torch_tensor):
36+ import torch
37+ if torch_tensor is None:
38+ return None
39+ if not isinstance(torch_tensor, torch.Tensor):
40+ raise RuntimeError(f"Only support torch.Tensor. But got {type(torch_tensor)}")
41+ torch_dtype = torch_tensor.dtype
42+ if torch_dtype == torch.bfloat16:
43+ t_int16 = torch_tensor.view(torch.int16)
44+ np_int16 = t_int16.numpy()
45+ return np_int16.view(dtype=np.bfloat16)
46+ else:
47+ return torch_tensor.numpy()
48+ 
49+ 
50+def bincount_golden(array, size, weight, **kwargs):
51+ '''
52+ Golden function for bincount.
53+ All the parameters (names and order) follow @bincount_def.cpp without outputs.
54+ All the input Tensors are numpy.ndarray.
55+ 
56+ Args:
57+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
58+ full_soc_version, short_soc_version, testcase_name
59+ 
60+ Returns:
61+ Output tensor
62+ '''
63+ import torch
64+ 
65+ array_tensor = numpy_to_torch_tensor(array)
66+ weight_tensor = numpy_to_torch_tensor(weight)
67+ 
68+ if weight_tensor.numel() == 0:
69+ res = torch.bincount(array_tensor).to(dtype=torch.int32)
70+ else:
71+ size = int(size) if size is not None else 0
72+ res = torch.bincount(array_tensor, weights=weight_tensor, minlength=size)
73+ 
74+ return torch_to_numpy_tensor(res)
Amath/cumsum/tests/assets/golden.py+48-0
@@ -0,0 +1,48 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "cumsum": "cumsum_golden"
18+ }
19+}
20+ 
21+def cumsum_golden(x, axis, *, exclusive, reverse, **kwargs):
22+ '''
23+ Golden function for cumsum.
24+ All the parameters (names and order) follow @cumsum_def.cpp without outputs.
25+ All the input Tensors are numpy.ndarray.
26+ 
27+ Args:
28+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
29+ full_soc_version, short_soc_version, testcase_name
30+ 
31+ Returns:
32+ Output tensor
33+ '''
34+ import tensorflow.compat.v1 as tf
35+ tf.disable_eager_execution()
36+ 
37+ x_dtype = x.dtype
38+ if x_dtype.name == "bfloat16" or x_dtype.name == "float16":
39+ x = x.astype("float32")
40+ 
41+ p0 = tf.constant(x)
42+ out = tf.cumsum(x=p0, axis=axis, reverse=reverse, exclusive=exclusive)
43+ 
44+ with tf.Session() as sess:
45+ res = sess.run(out)
46+ if x_dtype.name == "bfloat16" or x_dtype.name == "float16":
47+ res = res.astype(x_dtype, copy=False)
48+ return res
Amath/diag_part/tests/assets/golden.py+47-0
@@ -0,0 +1,47 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "diag_part": "diag_part_golden"
18+ }
19+}
20+ 
21+ 
22+def diag_part_golden(x, **kwargs):
23+ '''
24+ Golden function for diag_part.
25+ All the parameters (names and order) follow @diag_part_def.cpp without outputs.
26+ All the input Tensors are numpy.ndarray.
27+ 
28+ Args:
29+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
30+ full_soc_version, short_soc_version, testcase_name
31+ 
32+ Returns:
33+ Output tensor
34+ '''
35+ import tensorflow.compat.v1 as tf
36+ tf.disable_v2_behavior()
37+ 
38+ dtype = x.dtype
39+ if "bfloat16" in str(dtype):
40+ x = x.view("float16")
41+ x_holder = tf.placeholder(x.dtype, shape=x.shape)
42+ res = tf.diag_part(x_holder)
43+ with tf.Session() as session:
44+ res = session.run(res, feed_dict={x_holder: x})
45+ if "bfloat16" in str(dtype):
46+ res = res.view(dtype)
47+ return res
Amath/div/tests/assets/golden.py+124-0
@@ -0,0 +1,124 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "div": "div_golden"
18+ }
19+}
20+ 
21+ 
22+def broadcast_to_maxshape(shapes: list):
23+ def _max(_shape):
24+ no_one_shape = [s for s in _shape if s != 1]
25+ if len(no_one_shape) == 0:
26+ max_value = 1
27+ else:
28+ max_value = no_one_shape[0]
29+ return max_value
30+ max_dim_length = max(len(list(shape)) for shape in shapes)
31+ input_shapes = []
32+ for shape in shapes:
33+ input_shapes.append([1 for _ in range(max_dim_length - len(shape))] + list(shape))
34+ input_shapes = list(map(list, zip(*input_shapes)))
35+ max_shape = [_max(shape) for shape in input_shapes]
36+ input_shapes = list(map(list, zip(*input_shapes)))
37+ return (*input_shapes, max_shape)
38+ 
39+ 
40+def div_golden(x1, x2, **kwargs):
41+ '''
42+ Golden function for div.
43+ All the parameters (names and order) follow @div_def.cpp without outputs.
44+ All the input Tensors are numpy.ndarray.
45+ 
46+ Args:
47+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
48+ full_soc_version, short_soc_version, testcase_name
49+ 
50+ Returns:
51+ Output tensor
52+ '''
53+ ori_dtype = x1.dtype
54+ input_dtypes = kwargs.get('input_dtypes', [ori_dtype.name, ori_dtype.name])
55+
56+ if "float16" in str(ori_dtype): #include bfloat16 & float16
57+ x1, x2 = x1.astype("float32"), x2.astype("float32")
58+ 
59+ if "complex32" in input_dtypes:
60+ import torch
61+ 
62+ def complex32_div(reala, imaga, realb, imagb):
63+ abs_b1 = numpy.abs(realb)
64+ abs_b2 = numpy.abs(imagb)
65+ if abs_b1 >= abs_b2:
66+ if abs_b1 == 0 and abs_b2 == 0:
67+ real_ = reala / abs_b1
68+ imag_ = imaga / abs_b2
69+ else:
70+ temp1 = imagb / realb
71+ temp2 = realb + imagb * temp1
72+ tensor_one = numpy.array(1.0, dtype=numpy.float16)
73+ cm = tensor_one / temp2
74+ real_ = (reala + imaga * temp1) * cm
75+ imag_ = (imaga - reala * temp1) * cm
76+ else:
77+ temp1 = realb / imagb
78+ temp2 = imagb + realb * temp1
79+ tensor_one = numpy.array(1.0, dtype=numpy.float16)
80+ cm = tensor_one / temp2
81+ real_ = (imaga + reala * temp1) * cm
82+ imag_ = (imaga * temp1 - reala) * cm
83+ 
84+ return real_, imag_
85+ 
86+ x1, x2 = numpy.broadcast_arrays(x1, x2)
87+ xreal, ximag = numpy.split(x1, 2, axis=-1)
88+ yreal, yimag = numpy.split(x2, 2, axis=-1)
89+ ori_shape = xreal.shape
90+ xreal = xreal.reshape(-1)
91+ ximag = ximag.reshape(-1)
92+ yreal = yreal.reshape(-1)
93+ yimag = yimag.reshape(-1)
94+ input_xr = torch.from_numpy(xreal)
95+ input_xi = torch.from_numpy(ximag)
96+ input_yr = torch.from_numpy(yreal)
97+ input_yi = torch.from_numpy(yimag)
98+ zreal = torch.zeros_like(input_xr)
99+ zimag = torch.zeros_like(input_xr)
100+ for i in range(len(input_xr)):
101+ zreal[i], zimag[i] = complex32_div(input_xr[i],input_xi[i], input_yr[i], input_yi[i])
102+ 
103+ zreal = zreal.numpy()
104+ zimag = zimag.numpy()
105+ zreal = zreal.reshape(ori_shape)
106+ zimag = zimag.reshape(ori_shape)
107+ res = numpy.concatenate((zreal, zimag), axis=-1)
108+ return res
109+ else:
110+ import tensorflow as tf
111+ tf.compat.v1.disable_eager_execution()
112+ _, _, shape_max = broadcast_to_maxshape([x1.shape, x2.shape])
113+ x1 = numpy.broadcast_to(x1, shape_max)
114+ x2 = numpy.broadcast_to(x2, shape_max)
115+ x = tf.compat.v1.placeholder(shape=x1.shape, dtype=x1.dtype)
116+ y = tf.compat.v1.placeholder(shape=x2.shape, dtype=x2.dtype)
117+ out = tf.compat.v1.div(x, y)
118+ feed_dict = {x: x1, y: x2}
119+ init_op = tf.compat.v1.global_variables_initializer()
120+ 
121+ with tf.compat.v1.Session() as sess:
122+ sess.run(init_op)
123+ res = sess.run(out, feed_dict=feed_dict)
124+ return res.astype(ori_dtype, copy=False)
Amath/div_no_nan/tests/ut/op_host/assets/golden.py+67-0
@@ -0,0 +1,67 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "div_no_nan": "div_no_nan_golden"
18+ }
19+}
20+ 
21+ 
22+def div_no_nan_golden(x1, x2, **kwargs):
23+ '''
24+ Golden function for div_no_nan.
25+ All the parameters (names and order) follow @div_no_nan_def.cpp without outputs.
26+ All the input Tensors are numpy.ndarray.
27+ 
28+ Args:
29+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
30+ full_soc_version, short_soc_version, testcase_name
31+ 
32+ Returns:
33+ Output tensor
34+ '''
35+ import tensorflow as tf
36+ tf.compat.v1.disable_eager_execution()
37+ 
38+ x1_placeholder = tf.compat.v1.placeholder(shape=x1.shape, dtype=x1.dtype)
39+ x2_placeholder = tf.compat.v1.placeholder(shape=x2.shape, dtype=x2.dtype)
40+ 
41+ if x1.dtype == np.int8 or x1.dtype == np.int8:
42+ x1_fp16 = tf.cast(x1_placeholder, tf.float16)
43+ x2_fp16 = tf.cast(x2_placeholder, tf.float16)
44+ out = tf.compat.v1.div_no_nan(x1_fp16, x2_fp16, name="divnonan")
45+ out = tf.cast(out, tf.int8)
46+ elif x1.dtype == np.uint8 or x1.dtype == np.uint8:
47+ x1_fp16 = tf.cast(x1_placeholder, tf.float16)
48+ x2_fp16 = tf.cast(x2_placeholder, tf.float16)
49+ out = tf.compat.v1.div_no_nan(x1_fp16, x2_fp16, name="divnonan")
50+ out = tf.cast(out, tf.uint8)
51+ elif x1.dtype == np.int32 or x1.dtype == np.int32:
52+ x1_fp32 = tf.cast(x1_placeholder, tf.float32)
53+ x2_fp32 = tf.cast(x2_placeholder, tf.float32)
54+ out = tf.compat.v1.div_no_nan(x1_fp32, x2_fp32, name="divnonan")
55+ out = tf.cast(out, tf.int32)
56+ else:
57+ out = tf.compat.v1.div_no_nan(x1_placeholder, x2_placeholder, name="divnonan")
58+ 
59+ feed_dict = {x1_placeholder: x1, x2_placeholder: x2}
60+ init_op = tf.compat.v1.global_variables_initializer()
61+ 
62+ with tf.compat.v1.Session() as sess:
63+ sess.run(init_op)
64+ res = sess.run(out, feed_dict=feed_dict)
65+ 
66+ output_dtypes = kwargs.get('output_dtypes', [None])
67+ return res.astype(output_dtypes[0]) if output_dtypes[0] else res
Arandom/drop_out_do_mask/tests/assets/golden.py+82-0
@@ -0,0 +1,82 @@
1+#!/usr/bin/env python3
2+# -*- coding: UTF-8 -*-
3+# ----------------------------------------------------------------------------
4+# Copyright (c) 2026 Huawei Technologies Co., Ltd.
5+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
6+# CANN Open Software License Agreement Version 2.0 (the "License").
7+# Please refer to the License for details. You may not use this file except in compliance with the License.
8+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
9+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
10+# See LICENSE in the root of the software repository for the full text of the License.
11+# ----------------------------------------------------------------------------
12+ 
13+import numpy as np
14+ 
15+__golden__ = {
16+ "kernel": {
17+ "drop_out_do_mask": "drop_out_do_mask_golden"
18+ }
19+}
20+ 
21+ 
22+def revert_bit(n):
23+ result = 0
24+ for i in range(8):
25+ result <<= 1
26+ result |= n & 1
27+ n >>= 1
28+ return result
29+ 
30+ 
31+def revert_array_bit(arr):
32+ res = []
33+ for item in arr.flatten():
34+ res.append(revert_bit(item))
35+ return np.array(res, dtype=np.uint8).reshape(arr.shape)
36+ 
37+ 
38+def drop_out_do_mask_golden(x, mask, keep_prob, **kwargs):
39+ '''
40+ Golden function for drop_out_do_mask.
41+ All the parameters (names and order) follow @drop_out_do_mask_def.cpp without outputs.
42+ All the input Tensors are numpy.ndarray.
43+ 
44+ Args:
45+ **kwargs: {input,output}_{dtypes,ori_shapes,formats,ori_formats},
46+ full_soc_version, short_soc_version, testcase_name
47+ 
48+ Returns:
49+ Output tensor
50+ '''
51+ short_soc_version = kwargs.get('short_soc_version', '')
52+ dtype = str(x.dtype)
53+ 
54+ if short_soc_version in ("Ascend950",):
55+ if dtype in ("bfloat16", "float16"):
56+ x = x.astype("float32")
57+ keep_prob = keep_prob.astype("float32")
58+ if keep_prob.flat[0] == 1.0:
59+ x = x.astype(dtype)
60+ return x
61+ elif keep_prob.flat[0] == 0.0:
62+ y_out = np.zeros(x.shape, dtype=dtype)
63+ return y_out
64+ else:
65+ if str(x.dtype) == "bfloat16":
66+ x = x.astype("float32")
67+ keep_prob = keep_prob.astype("float32")
68+ 
69+ shape_x = x.shape
70+ x_scale = x * (1.0 / keep_prob)
71+ mask = revert_array_bit(mask)
72+ mask_dtype = np.unpackbits(mask, axis=-1).astype(x.dtype)
73+ 
74+ size_x = 1
75+ x_scale = x_scale.flatten()
76+ for i in shape_x:
77+ size_x = size_x * i
78+ expect = x_scale
79+ mask_dtype = mask_dtype[:size_x]
80+ expect[mask_dtype == 0] = 0
81+ output = expect.reshape(shape_x)
82+ return output.astype(dtype)