Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into py_calc_memory

revert-12646-feature/jit/xbyak
chenweihang 7 years ago
commit 999d097bbb

@ -1,16 +1,16 @@
## Motivation
There is a ```gap``` between the ```Program``` defined by
user and the ```Executable``` that can be scheduled
There is a `gap` between the `Program` defined by
user and the `Executable` that can be scheduled
efficiently on heterogeneous hardware, either locally
or distributedly.
Usually, the ```gap``` is bridged by
Usually, the `gap` is bridged by
* A serious transformations with defined order.
* These transformations usually involve
```insert, delete, clustering, split, dependency analysis```.
`insert, delete, clustering, split, dependency analysis`.
* Has a simple way to verify and debug each transformation.
@ -38,44 +38,44 @@ design below.
#### Node
```Node``` represents an operation that performs some computation or
`Node` represents an operation that performs some computation or
a variable that is input or output of operation.
```Node```s are connected to other ```Node```s via inputs and outputs.
`Node`s are connected to other `Node`s via inputs and outputs.
Other properties (maybe device placement information) can be added
to ```Node``` in the future if it's a
common requirement of many other ```Pass```es. Otherwise, it should live
in a ```Node``` wrapper class that is private to some ```Pass``` or be
a local member of a ```Pass```.
to `Node` in the future if it's a
common requirement of many other `Pass`es. Otherwise, it should live
in a `Node` wrapper class that is private to some `Pass` or be
a local member of a `Pass`.
#### Graph
```Graph``` contains a list of ```Node```s, which are connected to
`Graph` contains a list of `Node`s, which are connected to
each other via inputs and outputs.
TODO: Better definitions for the graph.
```Graph``` can also contain ```Attribute```s. ```Attribute```s
can be ``any`` thing. For example, it can be a list of "wraper"
nodes. The ```wrapper``` nodes compose ```Node```s and provide
helper method for execution or transformation. ```Attribute```
`Graph` can also contain `Attribute`s. `Attribute`s
can be `any` thing. For example, it can be a list of "wraper"
nodes. The `wrapper` nodes compose `Node`s and provide
helper method for execution or transformation. `Attribute`
can also contain other things that describe some properties of
the ```Graph``` or ```Graph``` nodes. ```Attribute``` can be passed
across ```Pass```. However, it should be used with care.
the `Graph` or `Graph` nodes. `Attribute` can be passed
across `Pass`. However, it should be used with care.
#### Pass
```Pass``` represents a transformation of ```Graph```. Its input
is a ```Graph``` and its output is also a ```Graph```. For example,
a ```Pass``` can simply print out the ```Graph```. A ```Pass```
can also fuse some ```Graph```'s ```Node```s.
`Pass` represents a transformation of `Graph`. Its input
is a `Graph` and its output is also a `Graph`. For example,
a `Pass` can simply print out the `Graph`. A `Pass`
can also fuse some `Graph`'s `Node`s.
#### Optimize
```Optimize``` contains a series of ```Pass``` with defined order.
```Optimize``` transforms a ```Graph``` that only contains raw
modeling logic to a ```Graph``` that can be run efficiently while
`Optimize` contains a series of `Pass` with defined order.
`Optimize` transforms a `Graph` that only contains raw
modeling logic to a `Graph` that can be run efficiently while
maintaining the original modeling logic.

@ -35,8 +35,7 @@ paddle.fluid.program_guard ArgSpec(args=[], varargs='args', keywords='kwds', def
paddle.fluid.get_var ArgSpec(args=['name', 'program'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.Executor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.as_lodtensor ArgSpec(args=['self', 'data'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.begin_pass ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.end_pass ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False))
paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.scope_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
@ -200,31 +199,23 @@ paddle.fluid.layers.argsort ArgSpec(args=['input', 'axis', 'name'], varargs=None
paddle.fluid.layers.ones ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.zeros ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.reverse ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.split_lod_tensor ArgSpec(args=['input', 'mask', 'level'], varargs=None, keywords=None, defaults=(0,))
paddle.fluid.layers.merge_lod_tensor ArgSpec(args=['in_true', 'in_false', 'x', 'mask', 'level'], varargs=None, keywords=None, defaults=(0,))
paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.While.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.While.complete ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.Switch.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.Switch.case ArgSpec(args=['self', 'condition'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.Switch.default ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.lod_rank_table ArgSpec(args=['x', 'level'], varargs=None, keywords=None, defaults=(0,))
paddle.fluid.layers.max_sequence_len ArgSpec(args=['rank_table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.lod_tensor_to_array ArgSpec(args=['x', 'table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.array_to_lod_tensor ArgSpec(args=['x', 'table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.increment ArgSpec(args=['x', 'value', 'in_place'], varargs=None, keywords=None, defaults=(1.0, True))
paddle.fluid.layers.array_write ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.create_array ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.less_than ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords='ignored', defaults=(None, None))
paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords='ignored', defaults=(None,))
paddle.fluid.layers.array_read ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.shrink_memory ArgSpec(args=['x', 'i', 'table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.array_length ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.IfElse.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.IfElse.false_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.IfElse.input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.IfElse.output ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None)
paddle.fluid.layers.IfElse.parent_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.IfElse.true_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.DynamicRNN.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
@ -233,9 +224,6 @@ paddle.fluid.layers.DynamicRNN.output ArgSpec(args=['self'], varargs='outputs',
paddle.fluid.layers.DynamicRNN.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.update_memory ArgSpec(args=['self', 'ex_mem', 'new_mem'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.ConditionalBlock.__init__ ArgSpec(args=['self', 'inputs', 'is_scalar_condition', 'name'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.layers.ConditionalBlock.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.ConditionalBlock.complete ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.StaticRNN.complete_op ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.memory ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1))

@ -22,7 +22,12 @@ endif()
cc_test(eigen_test SRCS eigen_test.cc DEPS tensor)
nv_test(mixed_vector_test SRCS mixed_vector_test.cu DEPS place memory device_context tensor)
if(WITH_GPU)
nv_test(mixed_vector_test SRCS mixed_vector_test.cc mixed_vector_test.cu DEPS place memory device_context tensor)
else()
cc_test(mixed_vector_test SRCS mixed_vector_test.cc DEPS place memory device_context tensor)
endif()
cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto recordio)
cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)

@ -1,11 +1,11 @@
cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto)
cc_library(var_handle SRCS var_handle.cc DEPS place framework_proto node)
cc_library(op_handle_base SRCS op_handle_base.cc DEPS var_handle device_context lod_tensor)
cc_library(scale_loss_grad_op_handle SRCS scale_loss_grad_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(fetch_op_handle SRCS fetch_op_handle.cc DEPS op_handle_base scope lod_tensor ddim memory)
cc_library(computation_op_handle SRCS computation_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(rpc_op_handle SRCS rpc_op_handle.cc DEPS framework_proto scope place operator op_registry)
cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS graph)
cc_library(ssa_graph_builder SRCS ssa_graph_builder.cc DEPS graph graph_helper)
cc_library(ssa_graph_printer SRCS ssa_graph_printer.cc DEPS ssa_graph_builder)
cc_library(ssa_graph_checker SRCS ssa_graph_checker.cc DEPS ssa_graph_builder)

@ -46,11 +46,13 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
const std::vector<Scope *> &local_scopes,
const BuildStrategy &strategy);
#endif
std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph) const override;
std::unique_ptr<ir::Graph> Apply(
std::unique_ptr<ir::Graph> graph) const override;
int GetVarDeviceID(const std::string &varname) const override;
private:
void CreateOpHandleIOs(Graph *result, ir::Node *node, size_t device_id) const;
void CreateOpHandleIOs(ir::Graph *result, ir::Node *node,
size_t device_id) const;
private:
std::string loss_var_name_;
@ -64,8 +66,8 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
bool IsScaleLossOp(ir::Node *node) const;
void CreateRPCOp(Graph *result, ir::Node *node) const;
void CreateDistTrainOp(Graph *result, ir::Node *node) const;
void CreateRPCOp(ir::Graph *result, ir::Node *node) const;
void CreateDistTrainOp(ir::Graph *result, ir::Node *node) const;
/**
* Is this operator as the end-point operator before/after send operator.
@ -74,21 +76,22 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
const std::vector<std::string> &recv_vars) const;
std::vector<std::string> FindDistTrainSendVars(
const std::vector<std::unique_ptr<ir::Node>> &nodes) const;
const std::vector<ir::Node *> &nodes) const;
std::vector<std::string> FindDistTrainRecvVars(
const std::vector<std::unique_ptr<ir::Node>> &nodes) const;
const std::vector<ir::Node *> &nodes) const;
void ConnectOp(Graph *result, OpHandleBase *op,
void ConnectOp(ir::Graph *result, OpHandleBase *op,
const std::string &prev_op_name) const;
void CreateComputationalOps(Graph *result, ir::Node *node,
void CreateComputationalOps(ir::Graph *result, ir::Node *node,
size_t num_places) const;
void CreateScaleLossGradOp(Graph *result) const;
VarHandle *CreateReduceOp(Graph *result, const std::string &og,
void CreateScaleLossGradOp(ir::Graph *result) const;
VarHandle *CreateReduceOp(ir::Graph *result, const std::string &og,
int dst_dev_id) const;
void CreateComputationalOp(Graph *result, ir::Node *node, int dev_id) const;
void CreateComputationalOp(ir::Graph *result, ir::Node *node,
int dev_id) const;
bool IsParameterGradientOnce(
const std::string &og,
@ -96,12 +99,12 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
int GetOpDeviceID(ir::Node *node) const;
void InsertAllReduceOp(Graph *result, const std::string &og) const;
void InsertAllReduceOp(ir::Graph *result, const std::string &og) const;
void InsertDataBalanceOp(Graph *result,
void InsertDataBalanceOp(ir::Graph *result,
const std::vector<std::string> &datas) const;
void CreateBroadcastOp(Graph *result, const std::string &p_name,
void CreateBroadcastOp(ir::Graph *result, const std::string &p_name,
size_t src_dev_id) const;
bool IsSparseGradient(const std::string &og) const;

@ -13,6 +13,7 @@
// limitations under the License.
#include "paddle/fluid/framework/details/rpc_op_handle.h"
#include "paddle/fluid/framework/ir/graph.h"
namespace paddle {
namespace framework {
@ -33,7 +34,7 @@ void RPCOpHandle::RunImpl() {
for (auto *in : inputs_) {
auto &p = static_cast<VarHandle *>(in)->place_;
// FIXME(Yancey1989): need a better solution instead of use DebugString()
if (in->DebugString() == "dummy") { // HACK
if (ir::IsControlDepVar(*in->Node())) { // HACK
continue;
}
if (in->GeneratedOp()) {

@ -17,7 +17,7 @@
namespace paddle {
namespace framework {
namespace details {
void SSAGraphBuilder::PolishGraphToSupportDataHazards(Graph *graph) {
void SSAGraphBuilder::PolishGraphToSupportDataHazards(ir::Graph *graph) {
for (auto &var_map : graph->Get<GraphVars>("vars")) {
for (auto &name_pair : var_map) {
if (name_pair.second.size() <= 1) {
@ -36,9 +36,18 @@ void SSAGraphBuilder::PolishGraphToSupportDataHazards(Graph *graph) {
// Read Write is the same op.
continue;
}
bool has_dep = false;
for (auto *r_out : read_op->Outputs()) {
for (auto *w_in : write_op->Inputs()) {
if (r_out->Node() == w_in->Node()) {
has_dep = true;
break;
}
}
}
if (has_dep) continue;
auto *dep_var = new DummyVarHandle(
graph->CreateEmptyNode("dummy", ir::Node::Type::kVariable));
auto *dep_var = new DummyVarHandle(graph->CreateControlDepVar());
read_op->AddOutput(dep_var);
write_op->AddInput(dep_var);
graph->Get<GraphDepVars>("dep_vars").emplace(dep_var);
@ -49,7 +58,7 @@ void SSAGraphBuilder::PolishGraphToSupportDataHazards(Graph *graph) {
}
VarHandle *SSAGraphBuilder::CreateOrGetLatestVarHandle(
Graph *graph, ir::Node *node, const platform::Place &place,
ir::Graph *graph, ir::Node *node, const platform::Place &place,
size_t place_offset) {
auto &var_holders = graph->Get<GraphVars>("vars")[place_offset];
auto &var_holder = var_holders[node->Name()];
@ -70,7 +79,7 @@ VarHandle *SSAGraphBuilder::CreateOrGetLatestVarHandle(
return var;
}
void SSAGraphBuilder::CreateOpOutput(Graph *graph, OpHandleBase *op_handle,
void SSAGraphBuilder::CreateOpOutput(ir::Graph *graph, OpHandleBase *op_handle,
ir::Node *new_node,
const platform::Place &place,
size_t place_offset) {
@ -82,13 +91,12 @@ void SSAGraphBuilder::CreateOpOutput(Graph *graph, OpHandleBase *op_handle,
op_handle->AddOutput(var);
}
void SSAGraphBuilder::AddOutputToLeafOps(Graph *graph) {
void SSAGraphBuilder::AddOutputToLeafOps(ir::Graph *graph) {
for (auto &op : graph->Get<GraphOps>("ops")) {
if (!op->Outputs().empty()) {
continue;
}
auto *dummy_leaf = new DummyVarHandle(
graph->CreateEmptyNode("dummy", ir::Node::Type::kVariable));
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>("dep_vars").emplace(dummy_leaf);
op->AddOutput(dummy_leaf);
}

@ -57,26 +57,23 @@ class SSAGraphBuilder : public ir::Pass {
DISABLE_COPY_AND_ASSIGN(SSAGraphBuilder);
protected:
/**
* We only handle write after read(WAR), since it should not have a write
* after write in program. If there are write after write operators, we need
* prune them.
*
* https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR)
/*
Dependency graph has been constructed. However, there are still data
hazards need to be handled.
*/
static void PolishGraphToSupportDataHazards(Graph *graph);
static void PolishGraphToSupportDataHazards(ir::Graph *graph);
static VarHandle *CreateOrGetLatestVarHandle(Graph *graph, ir::Node *node,
static VarHandle *CreateOrGetLatestVarHandle(ir::Graph *graph, ir::Node *node,
const platform::Place &place,
size_t place_offset);
// Add an output variable (each_var_name, place, place_offset) to op_handle,
// which belongs to graph
static void CreateOpOutput(Graph *graph, OpHandleBase *op_handle,
static void CreateOpOutput(ir::Graph *graph, OpHandleBase *op_handle,
ir::Node *new_node, const platform::Place &place,
size_t place_offset);
static void AddOutputToLeafOps(Graph *graph);
static void AddOutputToLeafOps(ir::Graph *graph);
};
} // namespace details
} // namespace framework

@ -20,7 +20,7 @@ namespace paddle {
namespace framework {
namespace details {
bool SSAGraghBuilderWithChecker::IsValidGraph(const Graph *graph) const {
bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const {
std::unordered_map<OpHandleBase *, size_t> pending_ops;
std::unordered_set<VarHandleBase *> pending_vars;
std::unordered_set<VarHandleBase *> ready_vars;

@ -28,7 +28,8 @@ class SSAGraghBuilderWithChecker : public SSAGraphBuilder {
std::unique_ptr<SSAGraphBuilder>&& builder)
: builder_(std::move(builder)) {}
std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph) const override {
std::unique_ptr<ir::Graph> Apply(
std::unique_ptr<ir::Graph> graph) const override {
auto new_graph = builder_->Apply(std::move(graph));
PADDLE_ENFORCE(IsValidGraph(new_graph.get()));
return new_graph;
@ -38,7 +39,7 @@ class SSAGraghBuilderWithChecker : public SSAGraphBuilder {
return builder_->GetVarDeviceID(var_name);
}
bool IsValidGraph(const Graph* graph) const;
bool IsValidGraph(const ir::Graph* graph) const;
private:
std::unique_ptr<SSAGraphBuilder> builder_;

@ -21,7 +21,7 @@ namespace framework {
namespace details {
template <typename Callback>
static inline void IterAllVar(const Graph &graph, Callback callback) {
static inline void IterAllVar(const ir::Graph &graph, Callback callback) {
for (auto &each : graph.Get<GraphVars>("vars")) {
for (auto &pair1 : each) {
for (auto &pair2 : pair1.second) {
@ -35,7 +35,7 @@ static inline void IterAllVar(const Graph &graph, Callback callback) {
}
}
void GraphvizSSAGraphPrinter::Print(const Graph &graph,
void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph,
std::ostream &sout) const {
size_t var_id = 0;
std::unordered_map<const VarHandleBase *, size_t> vars;

@ -25,12 +25,12 @@ namespace details {
class SSAGraphPrinter {
public:
virtual ~SSAGraphPrinter() {}
virtual void Print(const Graph& graph, std::ostream& sout) const = 0;
virtual void Print(const ir::Graph& graph, std::ostream& sout) const = 0;
};
class GraphvizSSAGraphPrinter : public SSAGraphPrinter {
public:
void Print(const Graph& graph, std::ostream& sout) const override;
void Print(const ir::Graph& graph, std::ostream& sout) const override;
};
class SSAGraghBuilderWithPrinter : public SSAGraphBuilder {
@ -50,7 +50,8 @@ class SSAGraghBuilderWithPrinter : public SSAGraphBuilder {
stream_ptr_(std::move(sout)),
stream_ref_(*stream_ptr_) {}
std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph) const override {
std::unique_ptr<ir::Graph> Apply(
std::unique_ptr<ir::Graph> graph) const override {
auto new_graph = builder_->Apply(std::move(graph));
printer_->Print(*new_graph, stream_ref_);
return new_graph;

@ -21,7 +21,8 @@ namespace framework {
namespace details {
ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places, std::unique_ptr<Graph> &&graph)
const std::vector<platform::Place> &places,
std::unique_ptr<ir::Graph> &&graph)
: graph_(std::move(graph)),
pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_)
: nullptr),

@ -40,7 +40,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy,
const std::vector<Scope *> &local_scopes,
const std::vector<platform::Place> &places,
std::unique_ptr<Graph> &&graph);
std::unique_ptr<ir::Graph> &&graph);
// Run a SSAGraph by a thread pool
// Use topological sort algorithm
@ -53,7 +53,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
details::OpHandleBase *op);
private:
std::unique_ptr<Graph> graph_;
std::unique_ptr<ir::Graph> graph_;
std::unique_ptr<::ThreadPool> pool_;
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;

@ -26,7 +26,7 @@ std::string VarHandle::DebugString() const {
return ss.str();
}
std::string DummyVarHandle::DebugString() const { return "dummy"; }
std::string DummyVarHandle::DebugString() const { return node_->Name(); }
} // namespace details
} // namespace framework
} // namespace paddle

@ -45,19 +45,13 @@ ExecutorPrepareContext::~ExecutorPrepareContext() {
Executor::Executor(const platform::Place& place) : place_(place) {}
void Executor::Close() {
#ifdef PADDLE_WITH_DISTRIBUTE
void Executor::BeginPass() {
::paddle::operators::distributed::RPCClient::GetInstance<
::paddle::operators::distributed::GRPCClient>()
->SendBeginPass();
}
void Executor::EndPass() {
::paddle::operators::distributed::RPCClient::GetInstance<
::paddle::operators::distributed::GRPCClient>()
->SendEndPass();
}
->SendComplete();
#endif
}
void InitializeVariable(Variable* var, proto::VarType::Type var_type) {
if (var_type == proto::VarType::LOD_TENSOR) {

@ -44,17 +44,11 @@ class Executor {
explicit Executor(const platform::Place& place);
#ifdef PADDLE_WITH_DISTRIBUTE
/*
* Sending signal to pserver to mark current pass started.
* Close this Executor.
* Calling this method will send complete messages to all pserver instances.
*/
void BeginPass();
/*
* Sending signal to pserver to mark current pass finished.
*/
void EndPass();
#endif
void Close();
/* @Brief
* Runtime evaluation of the given ProgramDesc under certain Scope

@ -1,5 +1,6 @@
cc_library(node SRCS node.cc DEPS proto_desc)
cc_library(graph SRCS graph.cc DEPS node)
cc_library(graph_helper SRCS graph_helper.cc DEPS graph)
cc_library(pass SRCS pass.cc DEPS graph node)
cc_test(graph_test SRCS graph_test.cc DEPS graph proto_desc op_registry)
cc_test(graph_test SRCS graph_test.cc DEPS graph op_registry)
cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph_helper op_registry)

@ -12,14 +12,18 @@ 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. */
#include <algorithm>
#include <unordered_set>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/op_proto_maker.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/var_desc.h"
namespace paddle {
namespace framework {
namespace ir {
// NOTE(paddle-dev): This graph contains circle.
Graph::Graph(const ProgramDesc &program) : program_(program) {
VLOG(3) << "block in program:" << program_.Size();
std::unordered_map<std::string, VarDesc *> all_vars;
@ -27,40 +31,87 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
all_vars.emplace(var->Name(), var);
}
std::map<std::string, ir::Node *> var_nodes;
std::map<std::string, std::vector<ir::Node *>> var_nodes;
for (auto *op : program.Block(0).AllOps()) {
ir::Node *node = CreateOpNode(op);
// For input args, reuse the same var name if it was created before.
// Otherwise, create a new one.
for (auto &each_var_name : op->InputArgumentNames()) {
ir::Node *var = nullptr;
if (var_nodes.find(each_var_name) != var_nodes.end()) {
var = var_nodes.at(each_var_name);
var = var_nodes.at(each_var_name).back();
} else if (all_vars.count(each_var_name) != 0) {
var = CreateVarNode(all_vars.at(each_var_name));
var_nodes[each_var_name] = var;
var_nodes[each_var_name].push_back(var);
} else {
// TODO(paddle-dev): Seems some assumption doesn't hold?
VLOG(3) << op->Type()
<< " input var not in all_var list: " << each_var_name;
// Operation input var can be optional (dispensable). Which means
// the operation doesn't really need the var at runtime. In this
// case, the no-existed var is ready at the beginning.
var = CreateEmptyNode(each_var_name, ir::Node::Type::kVariable);
var_nodes[each_var_name] = var;
var_nodes[each_var_name].push_back(var);
}
node->inputs.push_back(var);
var->outputs.push_back(node);
}
// For output args, always create a new var.
for (auto &each_var_name : op->OutputArgumentNames()) {
ir::Node *var = nullptr;
if (var_nodes.find(each_var_name) != var_nodes.end()) {
var = var_nodes.at(each_var_name);
} else {
var = CreateVarNode(all_vars.at(each_var_name));
var_nodes[each_var_name] = var;
}
ir::Node *var = CreateVarNode(all_vars.at(each_var_name));
var_nodes[each_var_name].push_back(var);
node->outputs.push_back(var);
var->inputs.push_back(node);
}
}
/**
* We only handle write after read(WAR), since it should not have a write
* after write in program. If there are write after write operators, we need
* prune them.
*
* https://en.wikipedia.org/wiki/Hazard_(computer_architecture)#Write_after_read_(WAR)
*/
for (auto &var : var_nodes) {
auto &versions = var.second;
if (versions.size() <= 1) continue;
auto it_new = versions.rbegin();
auto it_old = versions.rbegin();
++it_old;
for (; it_old != versions.rend(); it_new = it_old, ++it_old) {
ir::Node *write_op =
(*it_new)->inputs.empty() ? nullptr : (*it_new)->inputs[0];
const auto &read_ops = (*it_old)->outputs;
for (auto *read_op : read_ops) {
// Manually add a dependency var from read_op to write_op;
if (read_op == write_op) {
// Read Write is the same op.
continue;
}
// 2 ops might have been connected via other vars.
bool has_dep = false;
for (ir::Node *r_out : read_op->outputs) {
for (ir::Node *w_in : write_op->inputs) {
if (r_out == w_in) {
has_dep = true;
break;
}
}
}
if (has_dep) continue;
ir::Node *dep_var = CreateControlDepVar();
read_op->outputs.push_back(dep_var);
dep_var->inputs.push_back(read_op);
write_op->inputs.push_back(dep_var);
dep_var->outputs.push_back(write_op);
}
}
}
}
bool IsControlDepVar(const ir::Node &var) {
return var.Name().find(ir::Node::kControlDepVarName) != std::string::npos;
}
} // namespace ir
} // namespace framework
} // namespace paddle

@ -26,6 +26,7 @@ limitations under the License. */
namespace paddle {
namespace framework {
namespace ir {
class Graph {
public:
@ -54,29 +55,70 @@ class Graph {
};
}
const std::unordered_set<ir::Node *> &Nodes() const { return node_set_; }
// Create a normal variable with non-null VarDesc.
ir::Node *CreateVarNode(VarDesc *var_desc) {
nodes.emplace_back(new ir::Node(var_desc));
return nodes.back().get();
return AddNode(new ir::Node(var_desc));
}
// Create a normal runnable operator with OpDesc.
ir::Node *CreateOpNode(OpDesc *op_desc) {
nodes.emplace_back(new ir::Node(op_desc));
return nodes.back().get();
return AddNode(new ir::Node(op_desc));
}
// Create a control dependency var that connects 2 operations. The
// var doesn't hold any data. Other than that, it's no different from
// other var, considering dependency analysis.
ir::Node *CreateControlDepVar() {
// TODO(panyx0718): control var name should be really unique.
const std::string name = string::Sprintf(
"%s@%llu", ir::Node::kControlDepVarName, node_set_.size());
return AddNode(new ir::Node(name, ir::Node::Type::kVariable));
}
// A more free style way of creating a graph node. Mostly use for test
// or "copy" from another node. Avoid using it if possible.
ir::Node *CreateEmptyNode(const std::string &name, ir::Node::Type type) {
nodes.emplace_back(new ir::Node(name, type));
return nodes.back().get();
return AddNode(new ir::Node(name, type));
}
std::vector<std::unique_ptr<ir::Node>> nodes;
// Clear all node information of the graph and return the ownership of the
// nodes.
std::vector<std::unique_ptr<ir::Node>> ReleaseNodes() {
std::vector<std::unique_ptr<ir::Node>> ret;
for (auto &n : nodes_) {
ret.emplace_back(n.second.release());
}
nodes_.clear();
node_set_.clear();
return ret;
}
private:
// This method takes ownership of `node`.
ir::Node *AddNode(ir::Node *node) {
PADDLE_ENFORCE(node_set_.find(node) == node_set_.end());
nodes_[node].reset(node);
node_set_.insert(node);
return node;
}
void RemoveNode(ir::Node *node) {
PADDLE_ENFORCE(node_set_.find(node) != node_set_.end());
node_set_.erase(node);
nodes_.erase(node);
}
// NOTE: program_ shouldn't be exposed to user.
const ProgramDesc &program_;
std::map<std::string, boost::any> attrs_;
std::map<std::string, std::function<void(void)>> attr_dels_;
std::map<ir::Node *, std::unique_ptr<ir::Node>> nodes_;
std::unordered_set<ir::Node *> node_set_;
};
bool IsControlDepVar(const ir::Node &var);
} // namespace ir
} // namespace framework
} // namespace paddle

@ -0,0 +1,118 @@
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
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. */
#include <algorithm>
#include <unordered_set>
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {
namespace ir {
namespace {
void SortHelper(
const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
ir::Node *node, std::unordered_set<ir::Node *> *visited,
std::vector<ir::Node *> *ret) {
visited->insert(node);
for (auto adj : adj_list.at(node)) {
if (visited->find(adj) == visited->end()) {
SortHelper(adj_list, adj, visited, ret);
}
}
VLOG(3) << "topology sort insert: " << node->Name()
<< reinterpret_cast<void *>(node) << " input " << node->inputs.size();
ret->push_back(node);
}
bool HasCircleHelper(
ir::Node *node,
const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list,
std::unordered_set<ir::Node *> *visited,
std::unordered_set<ir::Node *> *in_trace) {
if (visited->find(node) == visited->end()) {
visited->insert(node);
in_trace->insert(node);
for (ir::Node *in : adj_list.at(node)) {
if (visited->find(in) == visited->end() &&
HasCircleHelper(in, adj_list, visited, in_trace)) {
return true;
} else if (in_trace->find(in) != in_trace->end()) {
return true;
}
}
}
in_trace->erase(node);
return false;
}
bool HasCircleInternal(
const std::map<ir::Node *, std::unordered_set<ir::Node *>> &adj_list) {
std::unordered_set<ir::Node *> visited;
std::unordered_set<ir::Node *> in_trace;
for (auto &adj : adj_list) {
if (HasCircleHelper(adj.first, adj_list, &visited, &in_trace)) {
return true;
}
}
return false;
}
} // namespace
bool HasCircle(const Graph &graph) {
return HasCircleInternal(BuildOperationAdjList(graph));
}
std::vector<ir::Node *> TopologySortOperations(const Graph &graph) {
std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list =
BuildOperationAdjList(graph);
PADDLE_ENFORCE(!HasCircleInternal(adj_list));
std::unordered_set<ir::Node *> visited;
std::vector<ir::Node *> ret;
for (auto adj : adj_list) {
if (visited.find(adj.first) == visited.end()) {
SortHelper(adj_list, adj.first, &visited, &ret);
}
}
return ret;
}
std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
const Graph &graph) {
std::map<ir::Node *, std::unordered_set<ir::Node *>> adj_list;
for (auto &n : graph.Nodes()) {
if (n->NodeType() != ir::Node::Type::kOperation) continue;
if (adj_list.find(n) == adj_list.end()) {
adj_list[n] = std::unordered_set<ir::Node *>();
}
for (auto &var : n->inputs) {
for (auto &adj_n : var->inputs) {
PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
adj_list[n].insert(adj_n);
VLOG(3) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
<< " -> " << n->Name() << reinterpret_cast<void *>(n)
<< " via " << var->Name() << reinterpret_cast<void *>(var);
}
}
}
return adj_list;
}
} // namespace ir
} // namespace framework
} // namespace paddle

@ -0,0 +1,40 @@
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
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. */
#pragma once
#include <map>
#include <memory>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/node.h"
namespace paddle {
namespace framework {
namespace ir {
// Test if the graph contains circle.
bool HasCircle(const Graph &graph);
// Topology Sort the operations in the graph from inputs to outputs.
// `graph` cannot contain circle.
std::vector<ir::Node *> TopologySortOperations(const Graph &graph);
// Build an adjacency list of operations for the `graph`.
std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
const Graph &graph);
} // namespace ir
} // namespace framework
} // namespace paddle

@ -0,0 +1,125 @@
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
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. */
#include "paddle/fluid/framework/ir/graph.h"
#include <string>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/program_desc.h"
namespace paddle {
namespace framework {
namespace ir {
void BuildCircleGraph(Graph* g) {
ir::Node* o1 = g->CreateEmptyNode("op1", Node::Type::kOperation);
ir::Node* v1 = g->CreateEmptyNode("var1", Node::Type::kVariable);
o1->outputs.push_back(v1);
o1->inputs.push_back(v1);
v1->inputs.push_back(o1);
v1->outputs.push_back(o1);
}
void BuildCircleGraph2(Graph* g) {
ir::Node* o1 = g->CreateEmptyNode("op1", Node::Type::kOperation);
ir::Node* o2 = g->CreateEmptyNode("op2", Node::Type::kOperation);
ir::Node* v1 = g->CreateEmptyNode("var1", Node::Type::kVariable);
ir::Node* v2 = g->CreateEmptyNode("var2", Node::Type::kVariable);
o1->outputs.push_back(v1);
o2->inputs.push_back(v1);
v1->inputs.push_back(o1);
v1->outputs.push_back(o2);
o2->outputs.push_back(v2);
o1->inputs.push_back(v2);
v2->inputs.push_back(o2);
v2->outputs.push_back(o1);
}
void BuildNoCircleGraph(Graph* g) {
ir::Node* o1 = g->CreateEmptyNode("op1", Node::Type::kOperation);
ir::Node* o2 = g->CreateEmptyNode("op2", Node::Type::kOperation);
ir::Node* o3 = g->CreateEmptyNode("op3", Node::Type::kOperation);
ir::Node* o4 = g->CreateEmptyNode("op4", Node::Type::kOperation);
ir::Node* o5 = g->CreateEmptyNode("op5", Node::Type::kOperation);
ir::Node* v1 = g->CreateEmptyNode("var1", Node::Type::kVariable);
ir::Node* v2 = g->CreateEmptyNode("var2", Node::Type::kVariable);
ir::Node* v3 = g->CreateEmptyNode("var3", Node::Type::kVariable);
ir::Node* v4 = g->CreateEmptyNode("var4", Node::Type::kVariable);
// o1->v1->o2
o1->outputs.push_back(v1);
o2->inputs.push_back(v1);
v1->inputs.push_back(o1);
v1->outputs.push_back(o2);
// o2->v2->o3
// o2->v2->o4
o2->outputs.push_back(v2);
o3->inputs.push_back(v2);
o4->inputs.push_back(v2);
v2->inputs.push_back(o2);
v2->outputs.push_back(o3);
v2->outputs.push_back(o4);
// o2->v3->o5
o2->outputs.push_back(v3);
o5->inputs.push_back(v3);
v3->inputs.push_back(o2);
v3->outputs.push_back(o5);
// o3-v4->o5
o3->outputs.push_back(v4);
o5->inputs.push_back(v4);
v4->inputs.push_back(o3);
v4->outputs.push_back(o5);
}
TEST(GraphHelperTest, Basic) {
ProgramDesc prog;
Graph g(prog);
BuildCircleGraph(&g);
ASSERT_TRUE(HasCircle(g));
Graph g2(prog);
BuildCircleGraph2(&g2);
ASSERT_TRUE(HasCircle(g2));
auto adj_list = BuildOperationAdjList(g2);
for (auto& adj : adj_list) {
auto& adj_set = adj.second;
if (adj.first->Name() == "op1") {
ASSERT_EQ((*adj_set.begin())->Name(), "op2");
} else if (adj.first->Name() == "op2") {
ASSERT_EQ((*adj_set.begin())->Name(), "op1");
} else {
ASSERT_TRUE(false);
}
}
Graph g3(prog);
BuildNoCircleGraph(&g3);
ASSERT_FALSE(HasCircle(g3));
auto sorted = TopologySortOperations(g3);
std::map<std::string, size_t> node_map;
for (size_t i = 0; i < sorted.size(); ++i) {
node_map[sorted[i]->Name()] = i;
}
ASSERT_EQ(node_map.at("op1"), 0);
ASSERT_EQ(node_map.at("op2"), 1);
ASSERT_TRUE(node_map.at("op3") < node_map.at("op5"));
}
} // namespace ir
} // namespace framework
} // namespace paddle

@ -76,6 +76,7 @@ TEST(GraphTest, Basic) {
op->SetType("sum");
op->SetInput("X", {"test_a", "test_b", "test_c"});
op->SetOutput("Out", {"test_out"});
op->SetAttr("op_role", 1);
prog.MutableBlock(0)->Var("test_a")->SetType(proto::VarType::SELECTED_ROWS);
prog.MutableBlock(0)->Var("test_b")->SetType(proto::VarType::SELECTED_ROWS);
@ -92,21 +93,22 @@ TEST(GraphTest, Basic) {
ASSERT_EQ(proto::VarType::LOD_TENSOR,
prog.MutableBlock(0)->Var("test_out")->GetType());
std::unique_ptr<Graph> g(new Graph(prog));
ASSERT_EQ(g->nodes[0]->Name(), "sum");
ASSERT_EQ(g->nodes[0]->inputs[0]->Name(), "test_a");
ASSERT_EQ(g->nodes[0]->inputs[1]->Name(), "test_b");
ASSERT_EQ(g->nodes[0]->inputs[2]->Name(), "test_c");
ASSERT_EQ(g->nodes[0]->outputs[0]->Name(), "test_out");
ASSERT_EQ(g->nodes[1]->Name(), "test_a");
ASSERT_EQ(g->nodes[1]->outputs[0]->Name(), "sum");
ASSERT_EQ(g->nodes[2]->Name(), "test_b");
ASSERT_EQ(g->nodes[2]->outputs[0]->Name(), "sum");
ASSERT_EQ(g->nodes[3]->Name(), "test_c");
ASSERT_EQ(g->nodes[3]->outputs[0]->Name(), "sum");
ASSERT_EQ(g->nodes[4]->Name(), "test_out");
ASSERT_EQ(g->nodes[4]->inputs[0]->Name(), "sum");
ASSERT_EQ(g->nodes.size(), 5);
std::unique_ptr<ir::Graph> g(new ir::Graph(prog));
std::vector<ir::Node *> nodes(g->Nodes().begin(), g->Nodes().end());
for (ir::Node *n : nodes) {
if (n->Name() == "sum") {
ASSERT_EQ(n->inputs.size(), 3);
ASSERT_EQ(n->outputs.size(), 1);
} else if (n->Name() == "test_a" || n->Name() == "test_b" ||
n->Name() == "test_c") {
ASSERT_EQ(n->inputs.size(), 0);
ASSERT_EQ(n->outputs.size(), 1);
} else if (n->Name() == "test_out") {
ASSERT_EQ(n->inputs.size(), 1);
ASSERT_EQ(n->outputs.size(), 0);
}
}
ASSERT_EQ(nodes.size(), 5);
}
} // namespace framework
} // namespace paddle

Some files were not shown because too many files have changed in this diff Show More

Loading…
Cancel
Save