1# Copyright 2019-2021 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.context as context 18import mindspore.nn as nn 19from mindspore import Tensor 20from mindspore.common.initializer import initializer 21from mindspore.common.parameter import Parameter 22from mindspore.communication.management import init, NCCL_WORLD_COMM_GROUP, get_rank, get_group_size 23from mindspore.ops import operations as P 24from mindspore.ops.operations import _inner_ops as inner 25 26context.set_context(mode=context.GRAPH_MODE, device_target='GPU') 27 28init() 29rank = get_rank() 30size = get_group_size() 31x = np.ones([3, 1, 3, 3]).astype(np.float32) * 0.01 * (rank + 1) 32y = np.ones([3, 4, 6, 3]).astype(np.float32) * 0.01 * (rank + 1) 33 34class Net(nn.Cell): 35 def __init__(self): 36 super(Net, self).__init__() 37 self.x1 = Parameter(initializer(Tensor(x), x.shape), name='x1') 38 self.x2 = Parameter(initializer(Tensor(x), x.shape), name='x2') 39 self.x3 = Parameter(initializer(Tensor(x), x.shape), name='x3') 40 41 self.op0 = "sum" 42 self.op1 = "sum" 43 self.op2 = "sum" 44 45 self.all_reduce1 = P.AllReduce(self.op0, group=NCCL_WORLD_COMM_GROUP) 46 self.all_reduce2 = P.AllReduce(self.op1, group=NCCL_WORLD_COMM_GROUP) 47 self.all_reduce3 = P.AllReduce(self.op2, group=NCCL_WORLD_COMM_GROUP) 48 49 def construct(self): 50 return (self.all_reduce1(self.x1), 51 self.all_reduce2(self.x2), 52 self.all_reduce3(self.x3)) 53 54 55def test_AllReduce(): 56 all_reduce = Net() 57 output = all_reduce() 58 59 expect0 = np.ones([3, 1, 3, 3]).astype(np.float32) * 0 60 for i in range(size): 61 part = np.ones([3, 1, 3, 3]).astype(np.float32) * 0.01 * (i + 1) 62 expect0 += part 63 diff0 = output[0].asnumpy() - expect0 64 error0 = np.ones(shape=expect0.shape) * 1.0e-5 65 assert np.all(diff0 < error0) 66 assert output[0].shape == expect0.shape 67 68 expect1 = expect0 69 diff1 = output[1].asnumpy() - expect1 70 error1 = np.ones(shape=expect1.shape) * 1.0e-5 71 assert np.all(diff1 < error1) 72 assert output[1].shape == expect1.shape 73 74 expect2 = expect1 75 diff2 = output[2].asnumpy() - expect2 76 error2 = np.ones(shape=expect2.shape) * 1.0e-5 77 assert np.all(diff2 < error2) 78 assert output[2].shape == expect2.shape 79 80 81class Net2(nn.Cell): 82 def __init__(self): 83 super(Net2, self).__init__() 84 self.x1 = Parameter(initializer(Tensor(x), x.shape), name='x1') 85 86 self.op0 = "sum" 87 self.op1 = "sum" 88 self.op2 = "sum" 89 90 self.all_reduce1 = P.AllReduce(self.op0, group=NCCL_WORLD_COMM_GROUP) 91 self.all_reduce2 = P.AllReduce(self.op1, group=NCCL_WORLD_COMM_GROUP) 92 self.all_reduce3 = P.AllReduce(self.op2, group=NCCL_WORLD_COMM_GROUP) 93 94 def construct(self): 95 x_ = self.all_reduce1(self.x1) 96 y_ = self.all_reduce2(x_) 97 z_ = self.all_reduce3(y_) 98 return (x_, y_, z_) 99 100 101def test_AllReduce2(): 102 all_reduce = Net2() 103 output = all_reduce() 104 105 expect0 = np.ones([3, 1, 3, 3]).astype(np.float32) * 0 106 for i in range(size): 107 part = np.ones([3, 1, 3, 3]).astype(np.float32) * 0.01 * (i + 1) 108 expect0 += part 109 diff0 = abs(output[0].asnumpy() - expect0) 110 error0 = np.ones(shape=expect0.shape) * 1.0e-5 111 assert np.all(diff0 < error0) 112 assert output[0].shape == expect0.shape 113 114 expect1 = expect0 * size 115 diff1 = abs(output[1].asnumpy() - expect1) 116 error1 = np.ones(shape=expect1.shape) * 1.0e-5 117 assert np.all(diff1 < error1) 118 assert output[1].shape == expect1.shape 119 120 expect2 = expect1 * size 121 diff2 = abs(output[2].asnumpy() - expect2) 122 error2 = np.ones(shape=expect2.shape) * 1.0e-5 123 assert np.all(diff2 < error2) 124 assert output[2].shape == expect2.shape 125 126 127class DynamicAllReduceNet(nn.Cell): 128 def __init__(self): 129 super(DynamicAllReduceNet, self).__init__() 130 self.op = "sum" 131 self.all_reduce = P.AllReduce(self.op, group=NCCL_WORLD_COMM_GROUP) 132 self.d = inner.GpuConvertToDynamicShape() 133 134 def construct(self, input_x): 135 out = self.d(input_x) 136 out = self.all_reduce(out) 137 return out 138 139 140def test_all_reduce_dynamic(): 141 context.set_context(mode=context.GRAPH_MODE, device_target='GPU') 142 input1 = Tensor(x) 143 input2 = Tensor(y) 144 net = DynamicAllReduceNet() 145 146 output1 = net(input1) 147 expect1 = np.ones([3, 1, 3, 3]).astype(np.float32) * 0 148 for i in range(size): 149 part = np.ones([3, 1, 3, 3]).astype(np.float32) * 0.01 * (i + 1) 150 expect1 += part 151 diff1 = abs(output1.asnumpy() - expect1) 152 error1 = np.ones(shape=expect1.shape) * 1.0e-5 153 assert np.all(diff1 < error1) 154 assert output1.shape == expect1.shape 155 156 output2 = net(input2) 157 expect2 = np.ones([3, 4, 6, 3]).astype(np.float32) * 0 158 for i in range(size): 159 part = np.ones([3, 4, 6, 3]).astype(np.float32) * 0.01 * (i + 1) 160 expect2 += part 161 diff2 = abs(output2.asnumpy() - expect2) 162 error2 = np.ones(shape=expect2.shape) * 1.0e-5 163 assert np.all(diff2 < error2) 164 assert output2.shape == expect2.shape 165