提交 07cf302e 编写于 作者: Y Yang Yang

first commit

上级 7905e367
...@@ -64,6 +64,12 @@ static void SplitTensorAndMoveTensorToScopes( ...@@ -64,6 +64,12 @@ static void SplitTensorAndMoveTensorToScopes(
} }
} }
void WaitOnPlace(const platform::Place place) {
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
auto &dev_ctx = *pool.Get(place);
dev_ctx.Wait();
}
void WaitOnPlaces(const std::vector<platform::Place> places) { void WaitOnPlaces(const std::vector<platform::Place> places) {
platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance(); platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
...@@ -214,6 +220,7 @@ class ParallelDoGradOp : public framework::OperatorBase { ...@@ -214,6 +220,7 @@ class ParallelDoGradOp : public framework::OperatorBase {
auto &tensor_to_merge = sub_scopes[i]->FindVar(s)->Get<LoDTensor>(); auto &tensor_to_merge = sub_scopes[i]->FindVar(s)->Get<LoDTensor>();
if (!(places[i] == places[0])) { if (!(places[i] == places[0])) {
framework::Copy(tensor_to_merge, places[0], tmp); framework::Copy(tensor_to_merge, places[0], tmp);
WaitOnPlace(places[0]);
} else { } else {
tmp->ShareDataWith(tensor_to_merge); tmp->ShareDataWith(tensor_to_merge);
} }
...@@ -222,12 +229,13 @@ class ParallelDoGradOp : public framework::OperatorBase { ...@@ -222,12 +229,13 @@ class ParallelDoGradOp : public framework::OperatorBase {
"sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}}, "sum", {{"X", {s, tmp_name}}}, {{"Out", {s}}},
framework::AttributeMap{}); framework::AttributeMap{});
sum_op->Run(*sub_scopes[0], places[0]); sum_op->Run(*sub_scopes[0], places[0]);
WaitOnPlaces(places); WaitOnPlace(places[0]);
} }
VLOG(3) << result; VLOG(3) << result;
framework::Copy(result, place, scope.FindVar(s)->GetMutable<LoDTensor>()); framework::Copy(result, place, scope.FindVar(s)->GetMutable<LoDTensor>());
} }
WaitOnPlaces(places);
} }
}; };
......
...@@ -15,9 +15,6 @@ import unittest ...@@ -15,9 +15,6 @@ import unittest
import paddle.v2.fluid as fluid import paddle.v2.fluid as fluid
import numpy import numpy
import sys
# TODO(dzhwinter): get places op check need to be enhanced.
sys.exit(0)
class BaseParallelForTest(unittest.TestCase): class BaseParallelForTest(unittest.TestCase):
...@@ -165,13 +162,13 @@ class ParallelOpTest(BaseParallelForTest): ...@@ -165,13 +162,13 @@ class ParallelOpTest(BaseParallelForTest):
feed={ feed={
'img': numpy.random.random(size=(51, 784)).astype('float32') 'img': numpy.random.random(size=(51, 784)).astype('float32')
}, },
fetch='fc1.w@GRAD') fetch=['fc1.w@GRAD'])
def test_fc_with_tiny_data(self): def test_fc_with_tiny_data(self):
self.run_test( self.run_test(
callback=ParallelOpTest.__network__, callback=ParallelOpTest.__network__,
feed={'img': numpy.random.random(size=(1, 784)).astype('float32')}, feed={'img': numpy.random.random(size=(1, 784)).astype('float32')},
fetch='fc1.w@GRAD') fetch=['fc1.w@GRAD'])
if __name__ == '__main__': if __name__ == '__main__':
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册