1# Copyright 2019 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 18import mindspore.nn as nn 19from mindspore import Tensor 20from mindspore import context 21from mindspore.common.api import _cell_graph_executor 22from mindspore.ops import composite as C 23from mindspore.ops import operations as P 24from tests.ut.python.ops.test_math_ops import VirtualLoss 25 26 27grad_all = C.GradOperation(get_all=True) 28 29 30class NetWithLoss(nn.Cell): 31 def __init__(self, network): 32 super(NetWithLoss, self).__init__() 33 self.loss = VirtualLoss() 34 self.network = network 35 36 def construct(self, x, y, b): 37 predict = self.network(x, y, b) 38 return self.loss(predict) 39 40 41class GradWrap(nn.Cell): 42 def __init__(self, network): 43 super(GradWrap, self).__init__() 44 self.network = network 45 46 def construct(self, x, y, b): 47 return grad_all(self.network)(x, y, b) 48 49 50def compile_net(net, x, y, b): 51 net.set_auto_parallel() 52 net.set_train() 53 _cell_graph_executor.compile(net, x, y, b) 54 55 56def test_matmul_tanh(): 57 class Net(nn.Cell): 58 def __init__(self, strategy1, strategy2, strategy3): 59 super().__init__() 60 self.matmul1 = P.MatMul().shard(strategy1) 61 self.matmul2 = P.MatMul().shard(strategy2) 62 self.tanh = P.Tanh().shard(strategy3) 63 64 def construct(self, x, y, b): 65 out = self.tanh(self.matmul1(x, y)) 66 out = self.matmul2(out, b) 67 return out 68 69 strategy1 = ((16, 1), (1, 1)) 70 strategy2 = ((1, 1), (1, 16)) 71 strategy3 = ((4, 4),) 72 net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3))) 73 context.set_auto_parallel_context(parallel_mode="semi_auto_parallel") 74 context.set_auto_parallel_context(device_num=16, global_rank=0) 75 76 x = Tensor(np.ones([128, 32]), dtype=ms.float32) 77 y = Tensor(np.ones([32, 64]), dtype=ms.float32) 78 b = Tensor(np.ones([64, 64]), dtype=ms.float32) 79 compile_net(net, x, y, b) 80 81 82def test_matmul_activation(): 83 class Net(nn.Cell): 84 def __init__(self, strategy1, strategy2, strategy3): 85 super().__init__() 86 self.matmul1 = P.MatMul().shard(strategy1) 87 self.matmul2 = P.MatMul().shard(strategy2) 88 self.activation = P.ReLU().shard(strategy3) 89 90 def construct(self, x, y, b): 91 out = self.activation(self.matmul1(x, y)) 92 out = self.matmul2(out, b) 93 return out 94 95 strategy1 = ((16, 1), (1, 1)) 96 strategy2 = ((1, 1), (1, 16)) 97 strategy3 = ((4, 4),) 98 net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3))) 99 context.set_auto_parallel_context(parallel_mode="semi_auto_parallel") 100 context.set_auto_parallel_context(device_num=16, global_rank=0) 101 102 x = Tensor(np.ones([128, 32]), dtype=ms.float32) 103 y = Tensor(np.ones([32, 64]), dtype=ms.float32) 104 b = Tensor(np.ones([64, 64]), dtype=ms.float32) 105 compile_net(net, x, y, b) 106 107 108def test_matmul_softmax(): 109 class Net(nn.Cell): 110 def __init__(self, strategy1, strategy2, strategy3): 111 super().__init__() 112 self.matmul1 = P.MatMul().shard(strategy1) 113 self.matmul2 = P.MatMul().shard(strategy2) 114 self.softmax = P.Softmax().shard(strategy3) 115 116 def construct(self, x, y, b): 117 out = self.softmax(self.matmul1(x, y)) 118 out = self.matmul2(out, b) 119 return out 120 121 strategy1 = ((16, 1), (1, 1)) 122 strategy2 = ((1, 1), (1, 16)) 123 strategy3 = ((16, 1),) 124 net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3))) 125 context.set_auto_parallel_context(parallel_mode="semi_auto_parallel") 126 context.set_auto_parallel_context(device_num=16, global_rank=0) 127 128 x = Tensor(np.ones([128, 32]), dtype=ms.float32) 129 y = Tensor(np.ones([32, 64]), dtype=ms.float32) 130 b = Tensor(np.ones([64, 64]), dtype=ms.float32) 131 compile_net(net, x, y, b) 132 133 134def test_matmul_logsoftmax(): 135 class Net(nn.Cell): 136 def __init__(self, strategy1, strategy2, strategy3): 137 super().__init__() 138 self.matmul1 = P.MatMul().shard(strategy1) 139 self.matmul2 = P.MatMul().shard(strategy2) 140 self.logsoftmax = P.LogSoftmax().shard(strategy3) 141 142 def construct(self, x, y, b): 143 out = self.logsoftmax(self.matmul1(x, y)) 144 out = self.matmul2(out, b) 145 return out 146 147 strategy1 = ((4, 2), (2, 2)) 148 strategy2 = ((2, 4), (4, 2)) 149 strategy3 = ((16, 1),) 150 net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3))) 151 context.set_auto_parallel_context(parallel_mode="semi_auto_parallel") 152 context.set_auto_parallel_context(device_num=16, global_rank=0) 153 154 x = Tensor(np.ones([128, 32]), dtype=ms.float32) 155 y = Tensor(np.ones([32, 64]), dtype=ms.float32) 156 b = Tensor(np.ones([64, 64]), dtype=ms.float32) 157 compile_net(net, x, y, b) 158 159 160def test_activations(): 161 class Net(nn.Cell): 162 def __init__(self, strategy1, strategy2, strategy3): 163 super().__init__() 164 self.matmul1 = P.MatMul().shard(strategy1) 165 self.matmul2 = P.MatMul().shard(strategy2) 166 self.gelu = P.GeLU().shard(strategy3) 167 self.tanh = P.Tanh().shard(strategy3) 168 self.softmax = P.Softmax().shard(strategy3) 169 self.logsoftmax = P.LogSoftmax().shard(strategy3) 170 171 def construct(self, x, y, b): 172 out = self.gelu(self.tanh(self.matmul1(x, y))) 173 out = self.logsoftmax(self.softmax(self.matmul2(out, b))) 174 return out 175 176 strategy1 = ((1, 2), (2, 2)) 177 strategy2 = ((2, 2), (2, 1)) 178 strategy3 = ((4, 1),) 179 net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3))) 180 context.set_auto_parallel_context(parallel_mode="semi_auto_parallel") 181 context.set_auto_parallel_context(device_num=4, global_rank=0) 182 183 x = Tensor(np.ones([128, 32]), dtype=ms.float32) 184 y = Tensor(np.ones([32, 64]), dtype=ms.float32) 185 b = Tensor(np.ones([64, 64]), dtype=ms.float32) 186 compile_net(net, x, y, b) 187 188 189def test_activations_repeated_calculation(): 190 class Net(nn.Cell): 191 def __init__(self, strategy1, strategy2, strategy3, strategy4, strategy5, strategy6): 192 super().__init__() 193 self.matmul1 = P.MatMul().shard(strategy1) 194 self.matmul2 = P.MatMul().shard(strategy2) 195 self.gelu = P.GeLU().shard(strategy3) 196 self.tanh = P.Tanh().shard(strategy4) 197 self.softmax = P.Softmax().shard(strategy5) 198 self.logsoftmax = P.LogSoftmax().shard(strategy6) 199 200 def construct(self, x, y, b): 201 out = self.gelu(self.tanh(self.matmul1(x, y))) 202 out = self.logsoftmax(self.softmax(self.matmul2(out, b))) 203 return out 204 205 strategy1 = ((2, 4), (4, 8)) 206 strategy2 = ((2, 2), (2, 1)) 207 strategy3 = ((2, 1),) 208 strategy4 = ((2, 2),) 209 strategy5 = ((4, 1),) 210 strategy6 = ((8, 1),) 211 net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3, strategy4, strategy5, strategy6))) 212 context.set_auto_parallel_context(parallel_mode="semi_auto_parallel") 213 context.set_auto_parallel_context(device_num=64, global_rank=0) 214 215 x = Tensor(np.ones([128, 32]), dtype=ms.float32) 216 y = Tensor(np.ones([32, 64]), dtype=ms.float32) 217 b = Tensor(np.ones([64, 64]), dtype=ms.float32) 218 compile_net(net, x, y, b) 219 220 221def test_activations_axis_tuple(): 222 class Net(nn.Cell): 223 def __init__(self, strategy1, strategy2, strategy3, strategy4, strategy5, strategy6): 224 super().__init__() 225 self.matmul1 = P.MatMul().shard(strategy1) 226 self.matmul2 = P.MatMul().shard(strategy2) 227 self.gelu = P.GeLU().shard(strategy3) 228 self.tanh = P.Tanh().shard(strategy4) 229 self.softmax = P.Softmax(axis=(0, 1)).shard(strategy5) 230 self.logsoftmax = P.LogSoftmax().shard(strategy6) 231 232 def construct(self, x, y, b): 233 out = self.gelu(self.tanh(self.matmul1(x, y))) 234 out = self.logsoftmax(self.softmax(self.matmul2(out, b))) 235 return out 236 237 strategy1 = ((2, 4), (4, 8)) 238 strategy2 = ((2, 2), (2, 1)) 239 strategy3 = ((2, 1),) 240 strategy4 = ((2, 2),) 241 strategy5 = ((1, 1),) 242 strategy6 = ((8, 1),) 243 net = GradWrap(NetWithLoss(Net(strategy1, strategy2, strategy3, strategy4, strategy5, strategy6))) 244 context.set_auto_parallel_context(parallel_mode="semi_auto_parallel") 245 context.set_auto_parallel_context(device_num=64, global_rank=0) 246 247 x = Tensor(np.ones([128, 32]), dtype=ms.float32) 248 y = Tensor(np.ones([32, 64]), dtype=ms.float32) 249 b = Tensor(np.ones([64, 64]), dtype=ms.float32) 250 compile_net(net, x, y, b) 251