Paddle/test/onednn/test_elementwise_mul_bf16_onednn_op.py at 43f1b17474b009491781e2a5c741b862f5885312 · co63oc/Paddle · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
#  Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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 unittest

import numpy as np
from op_test import OpTest, convert_float_to_uint16

from paddle import enable_static
from paddle.base import core


@unittest.skipIf(
    not core.supports_bfloat16(), "place does not support BF16 evaluation"
)
class TestElementwiseMulBf16OneDNNOp(OpTest):
    def setUp(self):
        self.op_type = "elementwise_mul"
        self.use_onednn = True
        self.onednn_data_type = "bfloat16"
        self.axis = -1

        self.generate_data()
        self.x_bf16 = convert_float_to_uint16(self.x)
        self.y_bf16 = convert_float_to_uint16(self.y)
        self.inputs = {'X': self.x_bf16, 'Y': self.y_bf16}
        self.attrs = {'axis': self.axis, 'use_onednn': self.use_onednn}
        self.outputs = {'Out': convert_float_to_uint16(self.out)}

    def generate_data(self):
        self.x = np.random.random(
            100,
        ).astype(np.float32)
        self.y = np.random.random(
            100,
        ).astype(np.float32)
        self.out = np.multiply(self.x, self.y)

    def test_check_output(self):
        self.check_output_with_place(core.CPUPlace(), check_pir_onednn=True)

    def test_check_grad_normal(self):
        self.check_grad_with_place(
            core.CPUPlace(),
            ["X", "Y"],
            "Out",
            check_dygraph=False,
            user_defined_grads=[
                np.multiply(self.x, self.y),
                np.multiply(self.x, self.x),
            ],
            user_defined_grad_outputs=[self.x_bf16],
            check_pir_onednn=True,
        )

    def test_check_grad_ignore_x(self):
        self.check_grad_with_place(
            core.CPUPlace(),
            ["Y"],
            "Out",
            check_dygraph=False,
            user_defined_grads=[np.multiply(self.y, self.x)],
            user_defined_grad_outputs=[self.y_bf16],
            check_pir_onednn=True,
        )

    def test_check_grad_ignore_y(self):
        self.check_grad_with_place(
            core.CPUPlace(),
            ["X"],
            "Out",
            check_dygraph=False,
            user_defined_grads=[np.multiply(self.x, self.y)],
            user_defined_grad_outputs=[self.x_bf16],
            check_pir_onednn=True,
        )


class TestElementwiseMulBroadcastingBf16OneDNNOp(
    TestElementwiseMulBf16OneDNNOp
):
    def generate_data(self):
        self.x = np.random.uniform(1, 2, [1, 2, 3, 100]).astype(np.float32)
        self.y = np.random.uniform(1, 2, [100]).astype(np.float32)
        self.out = np.multiply(self.x, self.y)

    # Compute partial sums along all axes but last one
    def compute_reduced_gradients(self, out_grads):
        part_sum = np.add.reduceat(out_grads, [0], axis=0)
        part_sum = np.add.reduceat(part_sum, [0], axis=1)
        part_sum = np.add.reduceat(part_sum, [0], axis=2)
        return part_sum.flatten()

    # TODO(jczaja): elementwise_mul bf16 grad got some potential
    # accuracy problems that need to be explained
    def test_check_grad_normal(self):
        pass
        # self.check_grad_with_place(
        #    core.CPUPlace(), ["X", "Y"],
        #    "Out",
        #    check_dy_graph=False,
        #    user_defined_grads=[
        #        np.multiply(self.x, self.y),
        #        self.compute_reduced_gradients(np.multiply(self.x, self.x))
        #    ],
        #    user_defined_grad_outputs=[self.x_bf16])

    def test_check_grad_ignore_x(self):
        pass
        # self.check_grad_with_place(
        #    core.CPUPlace(), ["Y"],
        #    "Out",
        #    check_dy_graph=False,
        #    user_defined_grads=[
        #        self.compute_reduced_gradients(np.multiply(self.x, self.x))
        #    ],
        #    user_defined_grad_outputs=[self.x_bf16])


if __name__ == '__main__':
    enable_static()
    unittest.main()