• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# Copyright 2020 Huawei Technologies Co., Ltd
2#
3# Licensed under the Apache License, Version 2.0 (the "License");
4# you may not use this file except in compliance with the License.
5# You may obtain a copy of the License at
6#
7# http://www.apache.org/licenses/LICENSE-2.0
8#
9# Unless required by applicable law or agreed to in writing, software
10# distributed under the License is distributed on an "AS IS" BASIS,
11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12# See the License for the specific language governing permissions and
13# limitations under the License.
14
15import numpy as np
16
17import mindspore as ms
18from mindspore import context, Tensor, Parameter
19from mindspore.common.api import _cell_graph_executor
20from mindspore.nn import Cell, TrainOneStepCell, Momentum
21from mindspore.ops import operations as P
22
23
24class Net(Cell):
25    def __init__(self, mul_weight, batch_matmul_weight, transpose_b=False, strategy1=None, strategy2=None):
26        super().__init__()
27        self.mul = P.Mul().shard(strategy1)
28        self.batch_matmul = P.BatchMatMul(transpose_b=transpose_b).shard(strategy2)
29        self.mul_weight = Parameter(mul_weight, "w1")
30        self.batch_matmul_weight = Parameter(batch_matmul_weight, "w2")
31
32    def construct(self, x, b):
33        out = self.mul(x, self.mul_weight)
34        out = self.batch_matmul(out, self.batch_matmul_weight)
35        return out
36
37
38_x = Tensor(np.ones([128, 64, 32]), dtype=ms.float32)
39_w1 = Tensor(np.ones([128, 64, 32]), dtype=ms.float32)
40_w2 = Tensor(np.ones([128, 32, 32]), dtype=ms.float32)
41_b = Tensor(np.ones([128, 64, 16]), dtype=ms.float32)
42
43
44def compile_net(net):
45    optimizer = Momentum(net.trainable_params(), learning_rate=0.1, momentum=0.9)
46    train_net = TrainOneStepCell(net, optimizer)
47    train_net.set_auto_parallel()
48    train_net.set_train()
49    _cell_graph_executor.compile(train_net, _x, _b)
50    context.reset_auto_parallel_context()
51
52
53def test_batch_matmul_data_parallel():
54    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
55    strategy1 = ((16, 1, 1), (16, 1, 1))
56    strategy2 = ((16, 1, 1), (16, 1, 1))
57    net = Net(_w1, _w2, False, strategy1, strategy2)
58    compile_net(net)
59
60
61def test_batch_matmul_model_parallel():
62    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
63    strategy1 = ((1, 1, 1), (1, 1, 1))
64    strategy2 = ((1, 1, 1), (1, 1, 16))
65    net = Net(_w1, _w2, False, strategy1, strategy2)
66    compile_net(net)
67
68
69def test_batch_matmul_hybrid_parallel():
70    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
71    strategy1 = ((2, 2, 2), (2, 2, 2))
72    strategy2 = ((2, 2, 2), (2, 2, 2))
73    net = Net(_w1, _w2, False, strategy1, strategy2)
74    compile_net(net)
75
76
77def test_batch_matmul_auto_parallel():
78    context.set_auto_parallel_context(parallel_mode="auto_parallel", device_num=16, global_rank=0)
79    net = Net(_w1, _w2, False)
80    compile_net(net)
81
82
83def test_batch_matmul_repeat_calc():
84    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
85    strategy1 = ((2, 2, 4), (2, 2, 4))
86    strategy2 = ((1, 2, 2), (1, 2, 2))
87    net = Net(_w1, _w2, False, strategy1, strategy2)
88    compile_net(net)
89
90
91def test_batch_matmul_transpose_b():
92    context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
93    strategy1 = ((2, 2, 4), (2, 2, 4))
94    strategy2 = ((1, 2, 2), (1, 2, 2))
95    net = Net(_w1, _w2, True, strategy1, strategy2)
96    compile_net(net)
97