!1023 add_gatherv2_distributed_op
Merge pull request !1023 from lichen/add_gatherv2_distributed_oppull/1023/MERGE
mindspore-ci-bot
6 years ago
committed by
Gitee
commit
dd2062bf8d
File diff suppressed because it is too large
Load Diff
@ -0,0 +1,67 @@
|
||||
/**
|
||||
* Copyright 2020 Huawei Technologies Co., Ltd
|
||||
*
|
||||
* 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.
|
||||
*/
|
||||
|
||||
#ifndef MINDSPORE_CCSRC_PARALLEL_OPS_INFO_GATHER_V2_P_INFO_H_
|
||||
#define MINDSPORE_CCSRC_PARALLEL_OPS_INFO_GATHER_V2_P_INFO_H_
|
||||
|
||||
#include <memory>
|
||||
#include <string>
|
||||
#include <unordered_map>
|
||||
#include <vector>
|
||||
|
||||
#include "ir/value.h"
|
||||
#include "parallel/auto_parallel/operator_costmodel.h"
|
||||
#include "parallel/ops_info/operator_info.h"
|
||||
#include "parallel/strategy.h"
|
||||
|
||||
namespace mindspore {
|
||||
namespace parallel {
|
||||
class GatherV2PInfo : public OperatorInfo {
|
||||
public:
|
||||
GatherV2PInfo(const std::string &name, const Shapes &inputs_shape, const Shapes &outputs_shape,
|
||||
const PrimitiveAttrs &attrs)
|
||||
: OperatorInfo(name, inputs_shape, outputs_shape, attrs, std::make_shared<GatherV2PCost>()) {}
|
||||
~GatherV2PInfo() override = default;
|
||||
Status Init(const StrategyPtr &strategy) override;
|
||||
Status InitForCostModel(const StrategyPtr &strategy) override;
|
||||
|
||||
Status GenerateStrategies(int32_t stage_id) override;
|
||||
Status SetCostUnderStrategy(const StrategyPtr &strategy) override;
|
||||
std::shared_ptr<std::vector<std::vector<int32_t>>> GenerateBatchStrategies() override;
|
||||
|
||||
protected:
|
||||
Status CheckStrategy(const StrategyPtr &strategy) override;
|
||||
Status InferMirrorOps() override { return SUCCESS; }
|
||||
Status InferForwardCommunication() override { return SUCCESS; }
|
||||
Status InferTensorInfo() override;
|
||||
Status InferDevMatrixShape() override;
|
||||
Status InferTensorMap() override;
|
||||
Status GetAttrs() override;
|
||||
|
||||
private:
|
||||
Status ComputeReplaceGraph(const CNodePtr &cnode);
|
||||
Status InferBias();
|
||||
Status InferGroup();
|
||||
|
||||
int32_t axis_;
|
||||
int32_t bias_;
|
||||
int32_t slice_size_;
|
||||
Shape out_dev_matrix_shape_;
|
||||
Group group_;
|
||||
};
|
||||
} // namespace parallel
|
||||
} // namespace mindspore
|
||||
#endif // MINDSPORE_CCSRC_PARALLEL_OPS_INFO_GATHER_V2_P_INFO_H_
|
||||
@ -0,0 +1,173 @@
|
||||
# Copyright 2019 Huawei Technologies Co., Ltd
|
||||
#
|
||||
# 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 numpy as np
|
||||
import mindspore as ms
|
||||
from mindspore import Tensor
|
||||
from mindspore import context
|
||||
import mindspore.nn as nn
|
||||
from mindspore.ops import operations as P
|
||||
from mindspore.ops import composite as C
|
||||
from mindspore.common import dtype as mstype
|
||||
from mindspore.common.api import _executor
|
||||
from tests.ut.python.ops.test_math_ops import VirtualLoss
|
||||
|
||||
|
||||
class NetWithLoss(nn.Cell):
|
||||
def __init__(self, network):
|
||||
super(NetWithLoss, self).__init__()
|
||||
self.loss = VirtualLoss()
|
||||
self.network = network
|
||||
|
||||
def construct(self, x, y):
|
||||
predict = self.network(x, y)
|
||||
return self.loss(predict)
|
||||
|
||||
|
||||
class GradWrap(nn.Cell):
|
||||
def __init__(self, network):
|
||||
super(GradWrap, self).__init__()
|
||||
self.network = network
|
||||
|
||||
def construct(self, x, y):
|
||||
return C.grad_all(self.network)(x, y)
|
||||
|
||||
|
||||
class Net(nn.Cell):
|
||||
def __init__(self, axis=0, strategy1=None, strategy2=None, shape=[64, 64]):
|
||||
super().__init__()
|
||||
self.gatherv2 = P.GatherV2().set_strategy(strategy1)
|
||||
self.mul = P.Mul().set_strategy(strategy2)
|
||||
self.index = Tensor(np.ones(shape), dtype=ms.int32)
|
||||
self.axis = axis
|
||||
|
||||
def construct(self, x, y):
|
||||
out = self.gatherv2(x, self.index, self.axis)
|
||||
out = self.mul(out, y)
|
||||
return out
|
||||
|
||||
|
||||
def test_gatherv2_semi_auto0():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy1 = ((1, 8), )
|
||||
strategy2 = ((4, 2, 1), (4, 2, 1))
|
||||
net = GradWrap(NetWithLoss(Net(0, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 32]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_auto1():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy1 = ((8, 1), )
|
||||
strategy2 = ((4, 2, 1), (4, 2, 1))
|
||||
net = GradWrap(NetWithLoss(Net(0, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 32]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_auto2():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy1 = ((2, 4), )
|
||||
strategy2 = ((4, 2, 1), (4, 2, 1))
|
||||
net = GradWrap(NetWithLoss(Net(0, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 32]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_auto3():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy1 = ((1, 8), )
|
||||
strategy2 = ((4, 2, 1), (4, 2, 1))
|
||||
net = GradWrap(NetWithLoss(Net(1, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 64]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_auto4():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy1 = ((8, 1), )
|
||||
strategy2 = ((4, 2, 1), (4, 2, 1))
|
||||
net = GradWrap(NetWithLoss(Net(1, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 64]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_auto5():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy1 = ((2, 4), )
|
||||
strategy2 = ((4, 2, 1), (4, 2, 1))
|
||||
net = GradWrap(NetWithLoss(Net(1, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 64]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_auto6():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy2 = ((4, 2, 1), (4, 2, 1))
|
||||
net = GradWrap(NetWithLoss(Net(0, None, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 32]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_auto7():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy2 = ((4, 2, 1), (4, 2, 1))
|
||||
net = GradWrap(NetWithLoss(Net(1, None, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 64]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_semi_auto8():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="semi_auto_parallel")
|
||||
strategy1 = ((8, ), )
|
||||
strategy2 = ((4, 2), (4, 2))
|
||||
net = GradWrap(NetWithLoss(Net(0, strategy1, strategy2)))
|
||||
net.set_auto_parallel()
|
||||
|
||||
x = Tensor(np.ones([64]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_auto0():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="auto_parallel")
|
||||
net = GradWrap(NetWithLoss(Net(0)))
|
||||
net.set_auto_parallel()
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 32]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
def test_gatherv2_auto1():
|
||||
context.set_auto_parallel_context(device_num=8, global_rank=0, parallel_mode="auto_parallel")
|
||||
net = GradWrap(NetWithLoss(Net(1)))
|
||||
net.set_auto_parallel()
|
||||
x = Tensor(np.ones([64, 32]), dtype=ms.float32)
|
||||
y = Tensor(np.ones([64, 64, 64]), dtype=ms.float32)
|
||||
_executor.compile(net, x, y)
|
||||
|
||||
Loading…
Reference in new issue