Skip to content

Commit a20f8e2

Browse files
jacquesqiaojacquesqiao
committed
Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into dist-table-support-optimizer-regular
2 parents d9dc81a + 1001f8e commit a20f8e2

File tree

6 files changed

+35
-63
lines changed

6 files changed

+35
-63
lines changed

‎paddle/fluid/operators/adagrad_op.cc

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -119,8 +119,8 @@ struct SparseAdagradFunctor<platform::CPUDeviceContext, T> {
119119
auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
120120

121121
// 2. m += g_m * g_m
122-
math::scatter::Mul<platform::CPUDeviceContext, T> sqare_func;
123-
auto grad_square = sqare_func(context, grad_merge, grad_merge);
122+
auto grad_square =
123+
SquareSelectedRows<platform::CPUDeviceContext, T>(context, grad_merge);
124124

125125
math::SelectedRowsAddToTensor<platform::CPUDeviceContext, T> functor;
126126
functor(context, grad_square, moment);

‎paddle/fluid/operators/adagrad_op.cu

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -84,8 +84,8 @@ struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
8484
auto* grad_merge_data = grad_merge.mutable_value()->template data<T>();
8585
framework::Vector<int64_t> merge_rows(grad_merge.rows());
8686
// 2. m += g_m * g_m
87-
math::scatter::Mul<platform::CUDADeviceContext, T> sqare_func;
88-
auto grad_square = sqare_func(context, grad_merge, grad_merge);
87+
auto grad_square =
88+
SquareSelectedRows<platform::CUDADeviceContext, T>(context, grad_merge);
8989

9090
math::SelectedRowsAddToTensor<platform::CUDADeviceContext, T> functor;
9191
functor(context, grad_square, moment);

‎paddle/fluid/operators/adagrad_op.h

Lines changed: 14 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -28,6 +28,20 @@ struct SparseAdagradFunctor {
2828
framework::Tensor *moment, framework::Tensor *param);
2929
};
3030

31+
template <typename DeviceContext, typename T>
32+
framework::SelectedRows SquareSelectedRows(
33+
const DeviceContext &context, const framework::SelectedRows &input) {
34+
framework::SelectedRows out;
35+
out.set_rows(input.rows());
36+
out.set_height(input.height());
37+
out.mutable_value()->mutable_data<T>(input.value().dims(),
38+
context.GetPlace());
39+
auto e_out = framework::EigenVector<T>::Flatten(*(out.mutable_value()));
40+
auto e_in = framework::EigenVector<T>::Flatten(input.value());
41+
e_out.device(*context.eigen_device()) = e_in.square();
42+
return out;
43+
}
44+
3145
template <typename DeviceContext, typename T>
3246
class AdagradOpKernel : public framework::OpKernel<T> {
3347
public:

‎paddle/fluid/operators/math/jit_kernel_blas.cc

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -168,9 +168,11 @@ class VAddKernelImpl : public VAddKernel<T> {
168168
#endif
169169
this->Compute = VAddRefer<T>;
170170
}
171+
#ifdef PADDLE_WITH_XBYAK
171172

172173
private:
173174
std::unique_ptr<gen::VVVJitCode> jitcode_{nullptr};
175+
#endif
174176
};
175177

176178
#ifdef PADDLE_WITH_XBYAK
@@ -210,9 +212,11 @@ class VAddReluKernelImpl : public VAddReluKernel<T> {
210212
#endif
211213
this->Compute = VAddReluRefer<T>;
212214
}
215+
#ifdef PADDLE_WITH_XBYAK
213216

214217
private:
215218
std::unique_ptr<gen::VVVJitCode> jitcode_{nullptr};
219+
#endif
216220
};
217221

218222
#ifdef PADDLE_WITH_XBYAK

‎paddle/fluid/operators/math/selected_rows_functor.h

Lines changed: 0 additions & 51 deletions
Original file line numberDiff line numberDiff line change
@@ -88,57 +88,6 @@ struct MergeAdd {
8888
framework::SelectedRows* output);
8989
};
9090

91-
template <typename DeviceContext, typename T>
92-
struct Add {
93-
framework::SelectedRows operator()(const DeviceContext& context,
94-
const framework::SelectedRows& input1,
95-
const framework::SelectedRows& input2) {
96-
framework::SelectedRows out;
97-
out.set_rows(input1.rows());
98-
out.set_height(input1.height());
99-
out.mutable_value()->mutable_data<T>(input1.value().dims(),
100-
context.GetPlace());
101-
auto e_out = framework::EigenVector<T>::Flatten(*(out.mutable_value()));
102-
auto e_in1 = framework::EigenVector<T>::Flatten(input1.value());
103-
auto e_in2 = framework::EigenVector<T>::Flatten(input2.value());
104-
e_out.device(*context.eigen_device()) = e_in1 + e_in2;
105-
return out;
106-
}
107-
};
108-
109-
template <typename DeviceContext, typename T>
110-
struct Mul {
111-
// multiply two SelectedRows
112-
framework::SelectedRows operator()(const DeviceContext& context,
113-
const framework::SelectedRows& input1,
114-
const framework::SelectedRows& input2) {
115-
framework::SelectedRows out;
116-
out.set_rows(input1.rows());
117-
out.set_height(input1.height());
118-
out.mutable_value()->mutable_data<T>(input1.value().dims(),
119-
context.GetPlace());
120-
auto e_out = framework::EigenVector<T>::Flatten(*(out.mutable_value()));
121-
auto e_in1 = framework::EigenVector<T>::Flatten(input1.value());
122-
auto e_in2 = framework::EigenVector<T>::Flatten(input2.value());
123-
e_out.device(*context.eigen_device()) = e_in1 * e_in2;
124-
return out;
125-
}
126-
// multiply scalar to SelectedRows
127-
framework::SelectedRows operator()(const DeviceContext& context,
128-
const framework::SelectedRows& input1,
129-
const T input2) {
130-
framework::SelectedRows out;
131-
out.set_rows(input1.rows());
132-
out.set_height(input1.height());
133-
out.mutable_value()->mutable_data<T>(input1.value().dims(),
134-
context.GetPlace());
135-
auto e_out = framework::EigenVector<T>::Flatten(*(out.mutable_value()));
136-
auto e_in1 = framework::EigenVector<T>::Flatten(input1.value());
137-
e_out.device(*context.eigen_device()) = input2 * e_in1;
138-
return out;
139-
}
140-
};
141-
14291
enum class ScatterOps { ASSIGN, ADD, SUB, SUBBY, MUL, DIV, DIVBY };
14392

14493
// out = seleted_rows_in / tensor

‎python/paddle/fluid/tests/unittests/test_parallel_executor_crf.py

Lines changed: 13 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -16,6 +16,7 @@
1616

1717
import paddle.dataset.conll05 as conll05
1818
import paddle.fluid as fluid
19+
import paddle.fluid.core as core
1920
import unittest
2021
import paddle
2122
import numpy as np
@@ -177,32 +178,36 @@ def check_network_convergence(self,
177178
def test_update_sparse_parameter_all_reduce(self):
178179
build_strategy = fluid.BuildStrategy()
179180
build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
180-
self.check_network_convergence(
181-
is_sparse=True, build_strategy=build_strategy, use_cuda=True)
181+
if core.is_compiled_with_cuda():
182+
self.check_network_convergence(
183+
is_sparse=True, build_strategy=build_strategy, use_cuda=True)
182184
self.check_network_convergence(
183185
is_sparse=True, build_strategy=build_strategy, use_cuda=False)
184186

185187
def test_update_dense_parameter_all_reduce(self):
186188
build_strategy = fluid.BuildStrategy()
187189
build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
188-
self.check_network_convergence(
189-
is_sparse=False, build_strategy=build_strategy, use_cuda=True)
190+
if core.is_compiled_with_cuda():
191+
self.check_network_convergence(
192+
is_sparse=False, build_strategy=build_strategy, use_cuda=True)
190193
self.check_network_convergence(
191194
is_sparse=False, build_strategy=build_strategy, use_cuda=False)
192195

193196
def test_update_sparse_parameter_reduce(self):
194197
build_strategy = fluid.BuildStrategy()
195198
build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
196-
self.check_network_convergence(
197-
is_sparse=True, build_strategy=build_strategy, use_cuda=True)
199+
if core.is_compiled_with_cuda():
200+
self.check_network_convergence(
201+
is_sparse=True, build_strategy=build_strategy, use_cuda=True)
198202
self.check_network_convergence(
199203
is_sparse=True, build_strategy=build_strategy, use_cuda=False)
200204

201205
def test_update_dense_parameter_reduce(self):
202206
build_strategy = fluid.BuildStrategy()
203207
build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
204-
self.check_network_convergence(
205-
is_sparse=False, build_strategy=build_strategy, use_cuda=True)
208+
if core.is_compiled_with_cuda():
209+
self.check_network_convergence(
210+
is_sparse=False, build_strategy=build_strategy, use_cuda=True)
206211
self.check_network_convergence(
207212
is_sparse=False, build_strategy=build_strategy, use_cuda=False)
208213

0 commit comments

Comments
 (0)