Skip to content

P
PaddleDetection

PaddlePaddle / PaddleDetection
大约 1 年 前同步成功

通知 695
Star 11112
Fork 2696
P
PaddleDetection
提交 a5c96af3 编写于 作者: N nhzlx
浏览文件
操作

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into add_tensorrt_conv2d_converter

上级 f05c7fb8 baff71d5
隐藏空白更改
内联 并排
Showing with 1849 addition and 569 deletion
doc/fluid/design/ir/draft.md
浏览文件 @ a5c96af3
......@@ -64,6 +64,41 @@ 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.
```cpp
class Graph {
public:
explicit Graph(const ProgramDesc &program);
bool Has(const std::string &attr_name) const;
template <typename AttrType>
AttrType &Get(const std::string &attr_name) const;
template <typename AttrType>
void Set(const std::string &attr_name, AttrType *attr);
const std::unordered_set<ir::Node *> &Nodes() const;
// Create a normal variable with non-null VarDesc.
ir::Node *CreateVarNode(VarDesc *var_desc);
// Create a normal runnable operator with OpDesc.
ir::Node *CreateOpNode(OpDesc *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();
// 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);
// Clear all node information of the graph and return the ownership of the
// nodes.
std::vector<std::unique_ptr<ir::Node>> ReleaseNodes();
};
```
#### Pass
`Pass` represents a transformation of `Graph`. Its input
......@@ -71,6 +106,54 @@ 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.
```cpp
class Pass {
public:
std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph) const {
// Some correctness check.
auto new_graph = ApplyImpl(std::move(graph));
// Some correctness check.
return new_graph;
}
// Get a reference to the attributed previously set.
template <typename AttrType>
AttrType &Get(const std::string &attr_name) const;
// Set a pointer to the attribute. Pass takes ownership of the attribute.
template <typename AttrType>
void Set(const std::string &attr_name, AttrType *attr) ;
// Set a pointer to the attribute. Pass doesn't take ownership. Caller
// should delete the attribute.
template <typename AttrType>
void SetNotOwned(const std::string &attr_name, AttrType *attr);
protected:
virtual std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const = 0;
};
// In my_pass.cc
class MyPass : public Pass {
protected:
std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const override {
// do something.
return graph;
}
}
REGISTER_PASS(my_pass, MyPass)
.RequirePassAttr("places")
.RequireGraphAttr("dep_vars");
// To use the pass.
auto my_pass = ir::PassRegistry::Instance().Get("my_pass");
graph = my_pass->Apply(std::move(graph));
// Note: to force link my_pass.cc, in the code:
USE_PASS(my_pass);
```
#### Optimize
`Optimize` contains a series of `Pass` with defined order.
......@@ -86,4 +169,17 @@ maintaining the original modeling logic.
* Graph is transformed from raw model logic to a
form that is efficient to execute.
Program->ProgramToGraph->Graph->Pass1->Graph->Pass2->Graph->Pass3->Graph->Executor
```
// Program->ProgramToGraph->Graph->Pass1->Graph->Pass2->Graph->Pass3->Graph->Executor
auto graph = Graph(program);
graph = PassRegistry::Instance().Get("op_fuse_pass").Apply(std::move(grah));
// For more complex Pass, Optimize Process can provide Pass attributes.
auto mem_opt_pass = PassRegistry::Instance().Get("memory_optimization_pass");
mem_opt_pass.SetNotOwned<int>("optimize_level", 1);
mem_opt_pass->Apply(std::move(graph));
graph = PassRegistry::Instance().Get("multi_device_pass").Apply(std::move(grah));
graph = PassRegistry::Instance().Get("multi_device_check_pass").Apply(std::move(grah));
Executor exe;
exe.Run(graph);
```
paddle/fluid/API.spec
浏览文件 @ a5c96af3
......@@ -170,6 +170,7 @@ paddle.fluid.layers.mean_iou ArgSpec(args=['input', 'label', 'num_classes'], var
paddle.fluid.layers.relu ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.log ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_recordio_file ArgSpec(args=['filename', 'shapes', 'lod_levels', 'dtypes', 'pass_num', 'for_parallel'], varargs=None, keywords=None, defaults=(1, True))
paddle.fluid.layers.open_files ArgSpec(args=['filenames', 'shapes', 'lod_levels', 'dtypes', 'thread_num', 'buffer_size', 'pass_num', 'is_test'], varargs=None, keywords=None, defaults=(None, None, 1, None))
......@@ -201,7 +202,6 @@ paddle.fluid.layers.zeros ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=
paddle.fluid.layers.reverse ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=None)
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)
......@@ -225,17 +225,14 @@ paddle.fluid.layers.DynamicRNN.static_input ArgSpec(args=['self', 'x'], varargs=
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.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))
paddle.fluid.layers.StaticRNN.output ArgSpec(args=['self'], varargs='outputs', keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.parent_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.step ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.step_output ArgSpec(args=['self', 'o'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.update_memory ArgSpec(args=['self', 'mem', 'var'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.reorder_lod_tensor_by_rank ArgSpec(args=['x', 'rank_table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.ParallelDo.__init__ ArgSpec(args=['self', 'places', 'use_nccl', 'name'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.layers.ParallelDo.complete_op ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.ParallelDo.do ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.ParallelDo.get_parameters ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.ParallelDo.parent_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
......
paddle/fluid/framework/CMakeLists.txt
浏览文件 @ a5c96af3
......@@ -99,7 +99,7 @@ else()
endif()
cc_library(parallel_executor SRCS parallel_executor.cc DEPS ssa_graph_builder_factory threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph)
cc_library(parallel_executor SRCS parallel_executor.cc DEPS threaded_ssa_graph_executor scope_buffered_ssa_graph_executor graph graph_viz_pass multi_devices_graph_builder ssa_graph_printer ssa_graph_checker)
cc_library(prune SRCS prune.cc DEPS framework_proto)
cc_test(prune_test SRCS prune_test.cc DEPS op_info prune recurrent_op device_context)
......
paddle/fluid/framework/details/CMakeLists.txt
浏览文件 @ a5c96af3
......@@ -31,9 +31,6 @@ cc_library(fuse_vars_op_handle SRCS fuse_vars_op_handle.cc DEPS op_handle_base s
cc_library(multi_devices_graph_builder SRCS multi_devices_graph_builder.cc DEPS ssa_graph_builder computation_op_handle
scale_loss_grad_op_handle rpc_op_handle all_reduce_op_handle reduce_op_handle broadcast_op_handle data_balance_op_handle)
cc_library(ssa_graph_builder_factory SRCS ssa_graph_builder_factory.cc DEPS multi_devices_graph_builder ssa_graph_printer ssa_graph_checker)
cc_library(ssa_graph_executor SRCS ssa_graph_executor.cc DEPS graph framework_proto)
cc_library(threaded_ssa_graph_executor SRCS threaded_ssa_graph_executor.cc DEPS fetch_op_handle ssa_graph_executor scope
simple_threadpool device_context)
......
paddle/fluid/framework/details/ssa_graph_builder_factory.h → paddle/fluid/framework/details/exception_holder.h
浏览文件 @ a5c96af3
......@@ -13,57 +13,69 @@
// limitations under the License.
#pragma once
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/details/build_strategy.h"
#include "paddle/fluid/framework/details/ssa_graph_builder.h"
#include "paddle/fluid/platform/place.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/nccl_helper.h"
#endif
#include "paddle/fluid/platform/enforce.h"
namespace paddle {
namespace framework {
class Scope;
namespace details {
class SSAGraphBuilderFactory {
class ExceptionHolder {
public:
SSAGraphBuilderFactory(const std::vector<platform::Place>& places,
const std::string& loss_var_name,
const std::unordered_set<std::string>& param_names,
const std::vector<Scope*>& local_scopes,
const BuildStrategy& strategy)
: places_(places),
loss_var_name_(loss_var_name),
param_names_(param_names),
local_scopes_(local_scopes),
strategy_(strategy) {
#ifdef PADDLE_WITH_CUDA
nccl_ctxs_ = nullptr;
#endif
void Catch(const platform::EnforceNotMet& exp) {
std::lock_guard<std::mutex> lock(mu_);
exception_.reset(new platform::EnforceNotMet(exp));
type_ = kEnforceNotMet;
}
#ifdef PADDLE_WITH_CUDA
void SetNCCLContextMap(platform::NCCLContextMap* nccl_ctxs) {
nccl_ctxs_ = nccl_ctxs;
void Catch(const platform::EOFException& exp) {
std::lock_guard<std::mutex> lock(mu_);
// EOFException will not cover up existing EnforceNotMet.
if (exception_.get() == nullptr) {
exception_.reset(new platform::EOFException(exp));
type_ = kEOF;
}
}
#endif
std::unique_ptr<SSAGraphBuilder> Create();
bool ExceptionCatched() const {
std::lock_guard<std::mutex> lock(mu_);
return exception_.get() != nullptr;
}
void Throw() {
std::lock_guard<std::mutex> lock(mu_);
switch (type_) {
case kNone:
break;
case kEnforceNotMet: {
auto e = *static_cast<platform::EnforceNotMet*>(exception_.get());
throw e;
break;
}
case kEOF: {
auto e = *static_cast<platform::EOFException*>(exception_.get());
throw e;
break;
}
default:
LOG(FATAL) << "Unknown exception.";
}
exception_.reset();
type_ = kNone;
}
void Clear() {
std::lock_guard<std::mutex> lock(mu_);
exception_.reset();
type_ = kNone;
}
private:
std::vector<platform::Place> places_;
std::string loss_var_name_;
std::unordered_set<std::string> param_names_;
std::vector<Scope*> local_scopes_;
BuildStrategy strategy_;
enum ExceptionType { kNone, kEnforceNotMet, kEOF };
ExceptionType type_{kNone};
#ifdef PADDLE_WITH_CUDA
platform::NCCLContextMap* nccl_ctxs_;
#endif
std::unique_ptr<std::exception> exception_;
mutable std::mutex mu_;
};
} // namespace details
......
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @ a5c96af3
......@@ -34,30 +34,22 @@ namespace paddle {
namespace framework {
namespace details {
static const char kLossVarName[] = "loss_var_name";
static const char kPlaces[] = "places";
static const char kParams[] = "params";
static const char kLocalScopes[] = "local_scopes";
static const char kStrategy[] = "strategy";
void MultiDevSSAGraphBuilder::Init() const {
loss_var_name_ = Get<const std::string>(kLossVarName);
places_ = Get<const std::vector<platform::Place>>(kPlaces);
local_scopes_ = Get<const std::vector<Scope *>>(kLocalScopes);
strategy_ = Get<const BuildStrategy>(kStrategy);
#ifdef PADDLE_WITH_CUDA
MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::unordered_set<std::string> &params,
const std::vector<Scope *> &local_scopes,
platform::NCCLContextMap *nccl_ctxs, const BuildStrategy &strategy)
: loss_var_name_(loss_var_name),
places_(places),
local_scopes_(local_scopes),
nccl_ctxs_(nccl_ctxs),
strategy_(strategy) {
#else
MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::unordered_set<std::string> &params,
const std::vector<Scope *> &local_scopes, const BuildStrategy &strategy)
: loss_var_name_(loss_var_name),
places_(places),
local_scopes_(local_scopes),
strategy_(strategy) {
nccl_ctxs_ = &Get<platform::NCCLContextMap>("nccl_ctxs");
#endif
for (auto &p : params) {
for (auto &p : Get<const std::unordered_set<std::string>>(kParams)) {
grad_names_.insert(GradVarName(p));
}
balance_vars_.resize(places_.size(), 0);
......@@ -72,7 +64,7 @@ void MultiDevSSAGraphBuilder::CreateOpHandleIOs(ir::Graph *result,
ir::Node *node,
size_t place_id) const {
auto p = places_[place_id];
auto *op_handle = result->Get<GraphOps>("ops").back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
op_handle->SetDeviceContext(p,
platform::DeviceContextPool::Instance().Get(p));
......@@ -239,8 +231,9 @@ std::vector<ir::Node *> SortOpsAndDelayOptimizeOp(const ir::Graph &graph) {
return sorted_ret;
}
std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::Apply(
std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
Init();
// Give the topology sort order and rebuild the graph structure.
std::vector<ir::Node *> sorted_ops = SortOpsAndDelayOptimizeOp(*graph);
auto nodes = graph->ReleaseNodes();
......@@ -254,9 +247,10 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::Apply(
std::unordered_set<std::string> og_has_been_broadcast;
// We cannot invoke resize. It is a bug of GCC 4.8
result.Set("vars", new GraphVars(places_.size()));
result.Set("dep_vars", new GraphDepVars);
result.Set("ops", new GraphOps);
result.Set(kGraphVars, new GraphVars(places_.size()));
result.Set(kGraphDepVars, new GraphDepVars);
result.Set(kGraphOps, new GraphOps);
result.Set(kShardedVarDevice, new ShardedVarDevice);
// find send/recv vars so that we can place the distributed training
// related op in the place 0
......@@ -289,11 +283,12 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::Apply(
// the block.
is_forwarding = false;
} else {
int op_dev_id = GetOpDeviceID(node);
int op_dev_id = GetOpDeviceID(result, node);
if (op_dev_id != -1) { // This op only runs on one specific device.
CreateComputationalOp(&result, node, op_dev_id);
for (ir::Node *n : node->outputs) {
var_name_on_devices_.emplace(n->Name(), op_dev_id);
graph->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(n->Name(), op_dev_id);
}
} else {
// This op runs on all devices, and its output may have parameter's
......@@ -330,7 +325,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::Apply(
case BuildStrategy::ReduceStrategy::kReduce:
cur_device_id = GetAppropriateDeviceID({g_name});
CreateReduceOp(&result, g_name, cur_device_id);
var_name_on_devices_.emplace(g_name, cur_device_id);
graph->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(g_name, cur_device_id);
bcast_var_name_set[cur_device_id].emplace(p_name);
break;
case BuildStrategy::ReduceStrategy::kAllReduce:
......@@ -416,16 +412,16 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
result->CreateEmptyNode("broadcast", ir::Node::Type::kOperation),
local_scopes_, places_);
#endif
result->Get<GraphOps>("ops").emplace_back(op_handle);
result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
auto *in =
result->Get<GraphVars>("vars").at(src_dev_id).at(p_name).back().get();
result->Get<GraphVars>(kGraphVars).at(src_dev_id).at(p_name).back().get();
op_handle->AddInput(in);
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
auto &vars = result->Get<GraphVars>("vars").at(i).at(p_name);
auto &vars = result->Get<GraphVars>(kGraphVars).at(i).at(p_name);
auto *out_var = new VarHandle(
result->CreateEmptyNode(p_name, ir::Node::Type::kVariable), vars.size(),
i, p_name, p);
......@@ -437,7 +433,7 @@ void MultiDevSSAGraphBuilder::CreateBroadcastOp(ir::Graph *result,
void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
ir::Node *node,
int dev_id) const {
result->Get<GraphOps>("ops").emplace_back(
result->Get<GraphOps>(kGraphOps).emplace_back(
new ComputationOpHandle(result->CreateOpNode(node->Op()),
local_scopes_[dev_id], places_[dev_id]));
CreateOpHandleIOs(result, node, dev_id);
......@@ -446,20 +442,20 @@ void MultiDevSSAGraphBuilder::CreateComputationalOp(ir::Graph *result,
void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
const std::string &og) const {
#ifdef PADDLE_WITH_CUDA
result->Get<GraphOps>("ops").emplace_back(new AllReduceOpHandle(
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_));
#else
result->Get<GraphOps>("ops").emplace_back(new AllReduceOpHandle(
result->Get<GraphOps>(kGraphOps).emplace_back(new AllReduceOpHandle(
result->CreateEmptyNode("allreduce", ir::Node::Type::kOperation),
local_scopes_, places_));
#endif
auto *op_handle = result->Get<GraphOps>("ops").back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
auto &vars = result->Get<GraphVars>("vars")[i][og];
auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
PADDLE_ENFORCE(!vars.empty());
auto &prev_grad = vars.back();
op_handle->AddInput(prev_grad.get());
......@@ -475,20 +471,20 @@ void MultiDevSSAGraphBuilder::InsertAllReduceOp(ir::Graph *result,
void MultiDevSSAGraphBuilder::InsertDataBalanceOp(
ir::Graph *result, const std::vector<std::string> &datas) const {
#ifdef PADDLE_WITH_CUDA
result->Get<GraphOps>("ops").emplace_back(new DataBalanceOpHandle(
result->Get<GraphOps>(kGraphOps).emplace_back(new DataBalanceOpHandle(
result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_));
#else
result->Get<GraphOps>("ops").emplace_back(new DataBalanceOpHandle(
result->Get<GraphOps>(kGraphOps).emplace_back(new DataBalanceOpHandle(
result->CreateEmptyNode("data_balance", ir::Node::Type::kOperation),
local_scopes_, places_));
#endif
auto *op_handle = result->Get<GraphOps>("ops").back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
for (const std::string &d_name : datas) {
auto &vars = result->Get<GraphVars>("vars")[i][d_name];
auto &vars = result->Get<GraphVars>(kGraphVars)[i][d_name];
PADDLE_ENFORCE(!vars.empty());
op_handle->AddInput(vars.back().get());
auto var = new VarHandle(
......@@ -512,7 +508,8 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
return is_pg_once;
}
int MultiDevSSAGraphBuilder::GetOpDeviceID(ir::Node *node) const {
int MultiDevSSAGraphBuilder::GetOpDeviceID(const ir::Graph &graph,
ir::Node *node) const {
if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) {
return -1;
}
......@@ -525,15 +522,17 @@ int MultiDevSSAGraphBuilder::GetOpDeviceID(ir::Node *node) const {
node->Op()->GetAttr(OpProtoAndCheckerMaker::OpRoleVarAttrName()));
PADDLE_ENFORCE_EQ(param_grad.size(), 2U);
int dev_id = GetVarDeviceID(param_grad[1]);
int dev_id = GetVarDeviceID(graph, param_grad[1]);
PADDLE_ENFORCE_NE(dev_id, -1, "dev_id should not be -1.[%s, %s, %s]",
node->Op()->Type(), param_grad[0], param_grad[1]);
return dev_id;
}
int MultiDevSSAGraphBuilder::GetVarDeviceID(const std::string &varname) const {
auto got = var_name_on_devices_.find(varname);
return got == var_name_on_devices_.end() ? -1 : got->second;
int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
const std::string &varname) const {
auto &sharded_var_device = graph.Get<ShardedVarDevice>(kShardedVarDevice);
auto got = sharded_var_device.find(varname);
return got == sharded_var_device.end() ? -1 : got->second;
}
void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(ir::Graph *result) const {
......@@ -551,7 +550,7 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(ir::Graph *result) const {
result->CreateEmptyNode("scale_loss_grad", ir::Node::Type::kOperation),
local_scopes_.size(), local_scopes_[i], places_[i],
communication_dev_ctx);
result->Get<GraphOps>("ops").emplace_back(op_handle);
result->Get<GraphOps>(kGraphOps).emplace_back(op_handle);
// FIXME: Currently ScaleLossGradOp only use device_count as scale
// factor. So it does not depend on any other operators.
......@@ -572,7 +571,7 @@ void MultiDevSSAGraphBuilder::CreateComputationalOps(ir::Graph *result,
for (size_t scope_idx = 0; scope_idx < num_places; ++scope_idx) {
auto p = places_[scope_idx];
auto s = local_scopes_[scope_idx];
result->Get<GraphOps>("ops").emplace_back(
result->Get<GraphOps>(kGraphOps).emplace_back(
new ComputationOpHandle(result->CreateOpNode(node->Op()), s, p));
CreateOpHandleIOs(result, node, scope_idx);
}
......@@ -582,25 +581,25 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
const std::string &og,
int dst_dev_id) const {
#ifdef PADDLE_WITH_CUDA
result->Get<GraphOps>("ops").emplace_back(new ReduceOpHandle(
result->Get<GraphOps>(kGraphOps).emplace_back(new ReduceOpHandle(
result->CreateEmptyNode("reduce", ir::Node::Type::kOperation),
local_scopes_, places_, nccl_ctxs_));
#else
result->Get<GraphOps>("ops").emplace_back(new ReduceOpHandle(
result->Get<GraphOps>(kGraphOps).emplace_back(new ReduceOpHandle(
result->CreateEmptyNode("reduce", ir::Node::Type::kOperation),
local_scopes_, places_));
#endif
auto *op_handle = result->Get<GraphOps>("ops").back().get();
auto *op_handle = result->Get<GraphOps>(kGraphOps).back().get();
for (size_t i = 0; i < places_.size(); ++i) {
auto &p = places_[i];
SetCommunicationContext(op_handle, p);
auto &vars = result->Get<GraphVars>("vars")[i][og];
auto &vars = result->Get<GraphVars>(kGraphVars)[i][og];
PADDLE_ENFORCE(!vars.empty());
auto &prev_grad = vars.back();
op_handle->AddInput(prev_grad.get());
}
auto &vars = result->Get<GraphVars>("vars")[dst_dev_id][og];
auto &vars = result->Get<GraphVars>(kGraphVars)[dst_dev_id][og];
auto var =
new VarHandle(result->CreateEmptyNode(og, ir::Node::Type::kVariable),
vars.size(), dst_dev_id, og, places_[dst_dev_id]);
......@@ -613,11 +612,11 @@ VarHandle *MultiDevSSAGraphBuilder::CreateReduceOp(ir::Graph *result,
// on it.
void MultiDevSSAGraphBuilder::ConnectOp(ir::Graph *result, OpHandleBase *op,
const std::string &prev_op_name) const {
for (auto &prev_op : result->Get<GraphOps>("ops")) {
for (auto &prev_op : result->Get<GraphOps>(kGraphOps)) {
if (prev_op->Name() == prev_op_name) {
auto *dep_var = new DummyVarHandle(result->CreateControlDepVar());
prev_op->AddOutput(dep_var);
result->Get<GraphDepVars>("dep_vars").emplace(dep_var);
result->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
op->AddInput(dep_var);
}
}
......@@ -638,20 +637,23 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
if (node->Op()->Type() == "split_byref" ||
node->Op()->Type() == "split_selected_rows") {
// TODO(paddle-dev): getting the first var is not safe.
op_dev_id = GetVarDeviceID(input_var_names[0]);
op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
if (strategy_.reduce_ == BuildStrategy::ReduceStrategy::kAllReduce) {
op_dev_id = GetAppropriateDeviceID(input_var_names);
for (auto &varname : input_var_names) {
var_name_on_devices_.emplace(varname, op_dev_id);
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
}
}
for (auto &varname : output_var_names) {
var_name_on_devices_.emplace(varname, op_dev_id);
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
}
} else if (node->Op()->Type() == "concat") {
op_dev_id = GetVarDeviceID(input_var_names[0]);
op_dev_id = GetVarDeviceID(*result, input_var_names[0]);
for (auto &varname : output_var_names) {
var_name_on_devices_.emplace(varname, op_dev_id);
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
}
} else {
PADDLE_ENFORCE(
......@@ -665,7 +667,7 @@ void MultiDevSSAGraphBuilder::CreateDistTrainOp(ir::Graph *result,
CreateComputationalOp(result, node, op_dev_id);
if (node->Op()->Type() == "concat") {
ConnectOp(result, result->Get<GraphOps>("ops").back().get(),
ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(),
"fetch_barrier");
}
}
......@@ -676,7 +678,7 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
int op_dev_id = -1;
if (node->Op()->Type() == "send") {
// TODO(paddle-dev): getting the first var is not safe.
op_dev_id = GetVarDeviceID(node->inputs[0]->Name());
op_dev_id = GetVarDeviceID(*result, node->inputs[0]->Name());
PADDLE_ENFORCE(!ir::IsControlDepVar(*node->inputs[0]),
"This hack no longer holds, please fix.");
// the variable name which contains .block means it was splited by
......@@ -691,7 +693,8 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
}
op_dev_id = GetAppropriateDeviceID(input_var_names);
for (auto &varname : input_var_names) {
var_name_on_devices_.emplace(varname, op_dev_id);
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
}
}
} else if (node->Op()->Type() == "recv") {
......@@ -701,7 +704,8 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
}
op_dev_id = GetAppropriateDeviceID(output_var_names);
for (auto &varname : output_var_names) {
var_name_on_devices_.emplace(varname, op_dev_id);
result->Get<ShardedVarDevice>(kShardedVarDevice)
.emplace(varname, op_dev_id);
}
} else {
// send_barrier and fetch_barrier op can be scheduled on device 0
......@@ -711,18 +715,18 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
PADDLE_ENFORCE(op_dev_id != -1, "can not find the right place for rpc op: %s",
node->Op()->Type());
result->Get<GraphOps>("ops").emplace_back(new RPCOpHandle(
result->Get<GraphOps>(kGraphOps).emplace_back(new RPCOpHandle(
result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
node->Op()->Type(), places_[op_dev_id]));
// TODO(panyx0718): This might not be needed anymore.
if (node->Op()->Type() == "send_barrier") {
ConnectOp(result, result->Get<GraphOps>("ops").back().get(), "send");
ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(), "send");
} else if (node->Op()->Type() == "recv") {
ConnectOp(result, result->Get<GraphOps>("ops").back().get(),
ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(),
"send_barrier");
} else if (node->Op()->Type() == "fetch_barrier") {
ConnectOp(result, result->Get<GraphOps>("ops").back().get(), "recv");
ConnectOp(result, result->Get<GraphOps>(kGraphOps).back().get(), "recv");
} else if (node->Op()->Type() == "send") {
// do nothing
} else {
......@@ -744,3 +748,11 @@ bool MultiDevSSAGraphBuilder::IsScaleLossOp(ir::Node *node) const {
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(multi_device_pass,
paddle::framework::details::MultiDevSSAGraphBuilder)
.RequirePassAttr(paddle::framework::details::kLossVarName)
.RequirePassAttr(paddle::framework::details::kPlaces)
.RequirePassAttr(paddle::framework::details::kParams)
.RequirePassAttr(paddle::framework::details::kLocalScopes)
.RequirePassAttr(paddle::framework::details::kStrategy);
paddle/fluid/framework/details/multi_devices_graph_builder.h
浏览文件 @ a5c96af3
......@@ -31,39 +31,27 @@ class Scope;
namespace details {
class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
public:
#ifdef PADDLE_WITH_CUDA
MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::unordered_set<std::string> &params,
const std::vector<Scope *> &local_scopes,
platform::NCCLContextMap *nccl_ctxs,
const BuildStrategy &strategy);
#else
MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::unordered_set<std::string> &params,
const std::vector<Scope *> &local_scopes,
const BuildStrategy &strategy);
#endif
std::unique_ptr<ir::Graph> Apply(
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
int GetVarDeviceID(const std::string &varname) const override;
private:
void CreateOpHandleIOs(ir::Graph *result, ir::Node *node,
size_t device_id) const;
void Init() const;
private:
std::string loss_var_name_;
const std::vector<platform::Place> &places_;
const std::vector<Scope *> &local_scopes_;
std::unordered_set<std::string> grad_names_;
mutable std::string loss_var_name_;
mutable std::vector<platform::Place> places_;
mutable std::vector<Scope *> local_scopes_;
mutable std::unordered_set<std::string> grad_names_;
#ifdef PADDLE_WITH_CUDA
platform::NCCLContextMap *nccl_ctxs_;
mutable platform::NCCLContextMap *nccl_ctxs_;
#endif
int GetVarDeviceID(const ir::Graph &graph, const std::string &varname) const;
bool IsScaleLossOp(ir::Node *node) const;
void CreateRPCOp(ir::Graph *result, ir::Node *node) const;
......@@ -97,7 +85,7 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
const std::string &og,
std::unordered_set<std::string> *og_has_been_broadcast) const;
int GetOpDeviceID(ir::Node *node) const;
int GetOpDeviceID(const ir::Graph &graph, ir::Node *node) const;
void InsertAllReduceOp(ir::Graph *result, const std::string &og) const;
......@@ -113,9 +101,8 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
const std::vector<std::string> &var_names) const;
private:
BuildStrategy strategy_;
mutable BuildStrategy strategy_;
mutable std::unordered_map<std::string, VarDesc *> all_vars_;
mutable std::unordered_map<std::string, int> var_name_on_devices_;
mutable std::vector<int64_t> balance_vars_;
void SetCommunicationContext(OpHandleBase *op_handle,
......
paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h
浏览文件 @ a5c96af3
......@@ -40,6 +40,11 @@ class ScopeBufferedSSAGraphExecutor : public SSAGraphExecutor {
ExecutionStrategy strategy, std::vector<Scope*> local_scopes,
std::vector<VariableInfo> var_infos, std::vector<platform::Place> places,
std::unique_ptr<SSAGraphExecutor>&& underlying_executor);
const ir::Graph& Graph() const override {
return underlying_executor_->Graph();
}
FeedFetchList Run(const std::vector<std::string>& fetch_tensors) override;
private:
......
paddle/fluid/framework/details/ssa_graph_builder.cc
浏览文件 @ a5c96af3
......@@ -18,7 +18,7 @@ namespace paddle {
namespace framework {
namespace details {
void SSAGraphBuilder::PolishGraphToSupportDataHazards(ir::Graph *graph) {
for (auto &var_map : graph->Get<GraphVars>("vars")) {
for (auto &var_map : graph->Get<GraphVars>(kGraphVars)) {
for (auto &name_pair : var_map) {
if (name_pair.second.size() <= 1) {
continue;
......@@ -50,7 +50,7 @@ void SSAGraphBuilder::PolishGraphToSupportDataHazards(ir::Graph *graph) {
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);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dep_var);
}
}
}
......@@ -60,7 +60,7 @@ void SSAGraphBuilder::PolishGraphToSupportDataHazards(ir::Graph *graph) {
VarHandle *SSAGraphBuilder::CreateOrGetLatestVarHandle(
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_holders = graph->Get<GraphVars>(kGraphVars)[place_offset];
auto &var_holder = var_holders[node->Name()];
VarHandle *var = nullptr;
if (var_holder.empty()) {
......@@ -83,7 +83,8 @@ void SSAGraphBuilder::CreateOpOutput(ir::Graph *graph, OpHandleBase *op_handle,
ir::Node *new_node,
const platform::Place &place,
size_t place_offset) {
auto &vars = graph->Get<GraphVars>("vars")[place_offset][new_node->Name()];
auto &vars =
graph->Get<GraphVars>(kGraphVars)[place_offset][new_node->Name()];
size_t version = vars.size();
auto var =
new VarHandle(new_node, version, place_offset, new_node->Name(), place);
......@@ -92,12 +93,12 @@ void SSAGraphBuilder::CreateOpOutput(ir::Graph *graph, OpHandleBase *op_handle,
}
void SSAGraphBuilder::AddOutputToLeafOps(ir::Graph *graph) {
for (auto &op : graph->Get<GraphOps>("ops")) {
for (auto &op : graph->Get<GraphOps>(kGraphOps)) {
if (!op->Outputs().empty()) {
continue;
}
auto *dummy_leaf = new DummyVarHandle(graph->CreateControlDepVar());
graph->Get<GraphDepVars>("dep_vars").emplace(dummy_leaf);
graph->Get<GraphDepVars>(kGraphDepVars).emplace(dummy_leaf);
op->AddOutput(dummy_leaf);
}
}
......
paddle/fluid/framework/details/ssa_graph_builder.h
浏览文件 @ a5c96af3
......@@ -39,21 +39,25 @@ namespace details {
typedef std::vector<
std::unordered_map<std::string, std::vector<std::unique_ptr<VarHandle>>>>
GraphVars;
const char kGraphVars[] = "vars";
// aux variables to represent dependency. Useful to resolve data hazard.
typedef std::unordered_set<std::unique_ptr<VarHandleBase>> GraphDepVars;
const char kGraphDepVars[] = "dep_vars";
// all operators. NOTE that even we use a vector here, the operators is
// unordered.
typedef std::vector<std::unique_ptr<OpHandleBase>> GraphOps;
const char kGraphOps[] = "ops";
typedef std::unordered_map<std::string, int> ShardedVarDevice;
const char kShardedVarDevice[] = "sharded_var_device";
class SSAGraphBuilder : public ir::Pass {
public:
SSAGraphBuilder() {}
virtual ~SSAGraphBuilder() {}
virtual int GetVarDeviceID(const std::string &var_name) const = 0;
DISABLE_COPY_AND_ASSIGN(SSAGraphBuilder);
protected:
......
paddle/fluid/framework/details/ssa_graph_checker.cc
浏览文件 @ a5c96af3
......@@ -33,7 +33,7 @@ bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const {
}
};
for (auto &var_map : graph->Get<GraphVars>("vars")) {
for (auto &var_map : graph->Get<GraphVars>(kGraphVars)) {
for (auto &name_pair : var_map) {
for (auto &version_pair : name_pair.second) {
insert_pending_var(version_pair.get());
......@@ -41,11 +41,11 @@ bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const {
}
}
for (auto &var : graph->Get<GraphDepVars>("dep_vars")) {
for (auto &var : graph->Get<GraphDepVars>(kGraphDepVars)) {
insert_pending_var(var.get());
}
for (auto &op : graph->Get<GraphOps>("ops")) {
for (auto &op : graph->Get<GraphOps>(kGraphOps)) {
if (op->Inputs().empty()) {
ready_ops.insert(op.get());
} else {
......@@ -85,3 +85,10 @@ bool SSAGraghBuilderWithChecker::IsValidGraph(const ir::Graph *graph) const {
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(multi_device_check_pass,
paddle::framework::details::SSAGraghBuilderWithChecker)
.RequireGraphAttr(paddle::framework::details::kGraphVars)
.RequireGraphAttr(paddle::framework::details::kGraphDepVars)
.RequireGraphAttr(paddle::framework::details::kGraphOps)
.RequireGraphAttr(paddle::framework::details::kShardedVarDevice);
paddle/fluid/framework/details/ssa_graph_checker.h
浏览文件 @ a5c96af3
......@@ -23,26 +23,14 @@ namespace framework {
namespace details {
class SSAGraghBuilderWithChecker : public SSAGraphBuilder {
public:
explicit SSAGraghBuilderWithChecker(
std::unique_ptr<SSAGraphBuilder>&& builder)
: builder_(std::move(builder)) {}
std::unique_ptr<ir::Graph> Apply(
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
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;
}
int GetVarDeviceID(const std::string& var_name) const override {
return builder_->GetVarDeviceID(var_name);
PADDLE_ENFORCE(IsValidGraph(graph.get()));
return graph;
}
bool IsValidGraph(const ir::Graph* graph) const;
private:
std::unique_ptr<SSAGraphBuilder> builder_;
};
} // namespace details
......
paddle/fluid/framework/details/ssa_graph_executor.h
浏览文件 @ a5c96af3
......@@ -32,7 +32,9 @@ class SSAGraphExecutor {
virtual ~SSAGraphExecutor();
virtual FeedFetchList Run(const std::vector<std::string> &fetch_tensors) = 0;
virtual const ir::Graph& Graph() const = 0;
virtual FeedFetchList Run(const std::vector<std::string>& fetch_tensors) = 0;
};
} // namespace details
} // namespace framework
......
paddle/fluid/framework/details/ssa_graph_printer.cc
浏览文件 @ a5c96af3
......@@ -22,7 +22,7 @@ namespace details {
template <typename Callback>
static inline void IterAllVar(const ir::Graph &graph, Callback callback) {
for (auto &each : graph.Get<GraphVars>("vars")) {
for (auto &each : graph.Get<GraphVars>(kGraphVars)) {
for (auto &pair1 : each) {
for (auto &pair2 : pair1.second) {
callback(*pair2);
......@@ -30,7 +30,7 @@ static inline void IterAllVar(const ir::Graph &graph, Callback callback) {
}
}
for (auto &var : graph.Get<GraphDepVars>("dep_vars")) {
for (auto &var : graph.Get<GraphDepVars>(kGraphDepVars)) {
callback(*var);
}
}
......@@ -61,7 +61,7 @@ void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph,
});
size_t op_id = 0;
for (auto &op : graph.Get<GraphOps>("ops")) {
for (auto &op : graph.Get<GraphOps>(kGraphOps)) {
std::string op_name = "op_" + std::to_string(op_id++);
sout << op_name << " [label=\"" << op->Name() << "\", shape=rect]"
<< std::endl;
......@@ -81,3 +81,6 @@ void GraphvizSSAGraphPrinter::Print(const ir::Graph &graph,
} // namespace details
} // namespace framework
} // namespace paddle
REGISTER_PASS(multi_device_print_pass,
paddle::framework::details::SSAGraghBuilderWithPrinter);
paddle/fluid/framework/details/ssa_graph_printer.h
浏览文件 @ a5c96af3
......@@ -14,7 +14,9 @@
#pragma once
#include <fstream>
#include <iosfwd>
#include <ostream>
#include <string>
#include "paddle/fluid/framework/details/ssa_graph_builder.h"
......@@ -34,38 +36,15 @@ class GraphvizSSAGraphPrinter : public SSAGraphPrinter {
};
class SSAGraghBuilderWithPrinter : public SSAGraphBuilder {
public:
SSAGraghBuilderWithPrinter(std::ostream& sout,
std::unique_ptr<SSAGraphPrinter>&& printer,
std::unique_ptr<SSAGraphBuilder>&& builder)
: printer_(std::move(printer)),
builder_(std::move(builder)),
stream_ref_(sout) {}
SSAGraghBuilderWithPrinter(std::unique_ptr<std::ostream>&& sout,
std::unique_ptr<SSAGraphPrinter>&& printer,
std::unique_ptr<SSAGraphBuilder>&& builder)
: printer_(std::move(printer)),
builder_(std::move(builder)),
stream_ptr_(std::move(sout)),
stream_ref_(*stream_ptr_) {}
std::unique_ptr<ir::Graph> Apply(
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
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;
std::unique_ptr<std::ostream> fout(
new std::ofstream(Get<const std::string>("debug_graphviz_path")));
PADDLE_ENFORCE(fout->good());
Get<GraphvizSSAGraphPrinter>("graph_printer").Print(*graph, *fout);
return graph;
}
int GetVarDeviceID(const std::string& var_name) const override {
return builder_->GetVarDeviceID(var_name);
}
private:
std::unique_ptr<SSAGraphPrinter> printer_;
std::unique_ptr<SSAGraphBuilder> builder_;
std::unique_ptr<std::ostream> stream_ptr_;
std::ostream& stream_ref_;
};
} // namespace details
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @ a5c96af3
......@@ -45,18 +45,18 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
std::unordered_set<OpHandleBase *> delayed_ops;
// Transform SSAGraph to pending_ops & pending_vars
for (auto &var_map : graph_->Get<details::GraphVars>("vars")) {
for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
for (auto &name_pair : var_map) {
for (auto &version_pair : name_pair.second) {
InsertPendingVar(&pending_vars, &ready_vars, version_pair.get());
}
}
}
for (auto &var : graph_->Get<details::GraphDepVars>("dep_vars")) {
for (auto &var : graph_->Get<details::GraphDepVars>(details::kGraphDepVars)) {
InsertPendingVar(&pending_vars, &ready_vars, var.get());
}
for (auto &op : graph_->Get<details::GraphOps>("ops")) {
for (auto &op : graph_->Get<details::GraphOps>(details::kGraphOps)) {
if (op->Inputs().empty()) { // Special case, Op has no input.
ready_ops.insert(op.get());
} else {
......@@ -83,7 +83,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
// Clean run context
run_op_futures_.clear();
exception_.reset();
exception_holder_.Clear();
// Step 3. Execution
while (!pending_vars.empty()) {
......@@ -103,23 +103,11 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
auto cur_ready_vars = ready_vars.PopAll(1, &timeout);
if (timeout) {
std::unique_lock<std::mutex> l(exception_mu_);
if (exception_) {
l.unlock();
if (exception_holder_.ExceptionCatched()) {
for (auto &run_op_future : run_op_futures_) {
run_op_future.wait();
}
l.lock();
std::exception *exp = exception_.get();
if (dynamic_cast<platform::EOFException *>(exp)) {
auto e = *static_cast<platform::EOFException *>(exp);
throw e;
} else if (dynamic_cast<platform::EnforceNotMet *>(exp)) {
auto e = *static_cast<platform::EnforceNotMet *>(exp);
throw e;
} else {
LOG(FATAL) << "Unknown exception.";
}
exception_holder_.Throw();
} else {
continue;
}
......@@ -162,7 +150,7 @@ void ThreadedSSAGraphExecutor::InsertFetchOps(
std::unordered_map<std::string, std::vector<VarHandleBase *>> fetched_vars;
for (auto &fetch_var_name : fetch_tensors) {
for (auto &var_map : graph_->Get<details::GraphVars>("vars")) {
for (auto &var_map : graph_->Get<details::GraphVars>(details::kGraphVars)) {
auto it = var_map.find(fetch_var_name);
if (it != var_map.end()) {
fetched_vars[fetch_var_name].push_back(it->second.rbegin()->get());
......@@ -229,14 +217,9 @@ void ThreadedSSAGraphExecutor::RunOp(
ready_var_q->Extend(op->Outputs());
VLOG(10) << op << " " << op->Name() << "Signal posted";
} catch (platform::EOFException ex) {
std::lock_guard<std::mutex> l(exception_mu_);
// EOFException will not cover up existing EnforceNotMet.
if (exception_.get() == nullptr) {
exception_.reset(new platform::EOFException(ex));
}
exception_holder_.Catch(ex);
} catch (platform::EnforceNotMet ex) {
std::lock_guard<std::mutex> l(exception_mu_);
exception_.reset(new platform::EnforceNotMet(ex));
exception_holder_.Catch(ex);
} catch (...) {
LOG(FATAL) << "Unknown exception catched";
}
......
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
浏览文件 @ a5c96af3
......@@ -24,6 +24,7 @@
#include <functional>
#include "ThreadPool.h" // ThreadPool in thrird party
#include "paddle/fluid/framework/blocking_queue.h"
#include "paddle/fluid/framework/details/exception_holder.h"
#include "paddle/fluid/framework/details/execution_strategy.h"
#include "paddle/fluid/framework/details/fetch_op_handle.h"
#include "paddle/fluid/framework/details/ssa_graph_executor.h"
......@@ -42,6 +43,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
const std::vector<platform::Place> &places,
std::unique_ptr<ir::Graph> &&graph);
const ir::Graph &Graph() const override { return *graph_; }
// Run a SSAGraph by a thread pool
// Use topological sort algorithm
FeedFetchList Run(const std::vector<std::string> &fetch_tensors) override;
......@@ -58,8 +60,7 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
std::vector<Scope *> local_scopes_;
std::vector<platform::Place> places_;
platform::DeviceContextPool fetch_ctxs_;
std::mutex exception_mu_;
std::unique_ptr<std::exception> exception_;
ExceptionHolder exception_holder_;
std::atomic<int> running_ops_;
void InsertPendingOp(std::unordered_map<OpHandleBase *, size_t> *pending_ops,
......
paddle/fluid/framework/ir/CMakeLists.txt
浏览文件 @ a5c96af3
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 op_registry)
cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph_helper op_registry)
cc_library(pass SRCS pass.cc DEPS graph node graph_helper)
cc_library(graph_viz_pass SRCS graph_viz_pass.cc DEPS graph pass graph_helper)
cc_test(pass_test SRCS pass_test.cc DEPS graph pass graph_helper)
cc_test(graph_test SRCS graph_test.cc DEPS graph graph_helper op_registry)
cc_test(graph_helper_test SRCS graph_helper_test.cc DEPS graph graph_helper op_registry)
paddle/fluid/framework/ir/graph.h
浏览文件 @ a5c96af3
......@@ -40,14 +40,21 @@ class Graph {
attr_dels_.clear();
}
bool Has(const std::string &attr_name) const {
return attrs_.find(attr_name) != attrs_.end();
}
template <typename AttrType>
AttrType &Get(const std::string &attr_name) const {
PADDLE_ENFORCE(Has(attr_name), "%s attr not registered for graph.",
attr_name);
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
}
template <typename AttrType>
void Set(const std::string &attr_name, AttrType *attr) {
PADDLE_ENFORCE(attrs_.count(attr_name) == 0);
PADDLE_ENFORCE(attrs_.count(attr_name) == 0, "%s already set in the graph",
attr_name);
attrs_[attr_name] = attr;
attr_dels_[attr_name] = [attr, attr_name]() {
VLOG(3) << "deleting " << attr_name;
......
paddle/fluid/framework/ir/graph_viz_pass.cc 0 → 100644
浏览文件 @ a5c96af3
/* 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_viz_pass.h"
namespace paddle {
namespace framework {
namespace ir {
static const char kGraphVizPath[] = "graph_viz_path";
std::unique_ptr<ir::Graph> GraphVizPass::ApplyImpl(
std::unique_ptr<ir::Graph> graph) const {
const std::string graph_viz_path = Get<std::string>(kGraphVizPath);
std::unique_ptr<std::ostream> fout(new std::ofstream(graph_viz_path));
PADDLE_ENFORCE(fout->good());
std::ostream& sout = *fout;
size_t var_id = 0;
std::unordered_map<const ir::Node*, size_t> vars;
sout << "digraph G {\n";
for (const ir::Node* n : graph->Nodes()) {
if (n->NodeType() != ir::Node::Type::kVariable) continue;
size_t cur_var_id = var_id++;
vars[n] = cur_var_id;
sout << "var_" << cur_var_id << " [label=\"" << n->Name() << "\"]"
<< std::endl;
}
size_t op_id = 0;
for (const ir::Node* n : graph->Nodes()) {
if (n->NodeType() != ir::Node::Type::kOperation) continue;
std::string op_name = "op_" + std::to_string(op_id++);
sout << op_name << " [label=\"" << n->Name() << "\", shape=rect]"
<< std::endl;
for (auto in : n->inputs) {
std::string var_name = "var_" + std::to_string(vars[in]);
sout << var_name << " -> " << op_name << std::endl;
}
for (auto out : n->outputs) {
std::string var_name = "var_" + std::to_string(vars[out]);
sout << op_name << " -> " << var_name << std::endl;
}
}
sout << "}\n";
return graph;
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(graph_viz_pass, paddle::framework::ir::GraphVizPass)
.RequirePassAttr(paddle::framework::ir::kGraphVizPath);
paddle/fluid/framework/ir/graph_viz_pass.h 0 → 100644
浏览文件 @ a5c96af3
/* 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 <fstream>
#include <map>
#include <memory>
#include <string>
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/pass.h"
namespace paddle {
namespace framework {
namespace ir {
class GraphVizPass : public Pass {
protected:
std::unique_ptr<ir::Graph> ApplyImpl(
std::unique_ptr<ir::Graph> graph) const override;
};
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/pass.cc
浏览文件 @ a5c96af3
......@@ -13,7 +13,34 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
namespace paddle {
namespace framework {} // namespace framework
namespace framework {
namespace ir {
std::unique_ptr<Graph> Pass::Apply(std::unique_ptr<Graph> graph) const {
PADDLE_ENFORCE(!applied_, "Pass can only Apply() once.");
PADDLE_ENFORCE(graph.get(), "graph passed to Pass::Apply() cannot be empty.");
for (const std::string& attr : required_pass_attrs_) {
PADDLE_ENFORCE(attrs_.find(attr) != attrs_.end(),
"Required pass atrribute %s not set.", attr);
}
for (const std::string& attr : required_graph_attrs_) {
PADDLE_ENFORCE(graph->Has(attr), "Required graph atrribute %s not set.",
attr);
}
auto applied_graph = ApplyImpl(std::move(graph));
// TODO(panyx0718): Add more verifications.
PADDLE_ENFORCE(!HasCircle(*applied_graph),
"Illegal Pass. Generated graph shouldn't has cycle.");
applied_ = true;
return applied_graph;
}
PassRegistry& PassRegistry::Instance() {
static PassRegistry g_pass_info_map;
return g_pass_info_map;
}
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/pass.h
浏览文件 @ a5c96af3
......@@ -14,21 +14,187 @@ limitations under the License. */
#pragma once
#include <functional>
#include <map>
#include <string>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/node.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/platform/variant.h"
namespace paddle {
namespace framework {
namespace ir {
template <typename PassType>
struct PassRegistrar;
class Pass {
public:
Pass() = default;
virtual ~Pass() {}
virtual ~Pass() {
for (auto &attr : attrs_) {
if (attr_dels_.find(attr.first) != attr_dels_.end()) {
attr_dels_[attr.first]();
}
}
attrs_.clear();
attr_dels_.clear();
}
std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph) const;
// Get a reference to the attributed previously set.
template <typename AttrType>
AttrType &Get(const std::string &attr_name) const {
PADDLE_ENFORCE(attrs_.find(attr_name) != attrs_.end(),
"%s attr not registered for pass.", attr_name);
return *boost::any_cast<AttrType *>(attrs_.at(attr_name));
}
// Set a pointer to the attribute. Pass takes ownership of the attribute.
template <typename AttrType>
void Set(const std::string &attr_name, AttrType *attr) {
PADDLE_ENFORCE(attrs_.count(attr_name) == 0, "%s already set in the pass",
attr_name);
attrs_[attr_name] = attr;
attr_dels_[attr_name] = [attr, attr_name]() {
VLOG(3) << "deleting " << attr_name;
delete attr;
};
}
// Set a pointer to the attribute. Pass doesn't take ownership. Caller
// should delete the attribute.
template <typename AttrType>
void SetNotOwned(const std::string &attr_name, AttrType *attr) {
PADDLE_ENFORCE(attrs_.count(attr_name) == 0);
attrs_[attr_name] = attr;
}
protected:
virtual std::unique_ptr<Graph> ApplyImpl(
std::unique_ptr<Graph> graph) const = 0;
private:
template <typename PassType>
friend struct PassRegistrar;
void RegisterRequiredPassAttrs(const std::unordered_set<std::string> &attrs) {
required_pass_attrs_.insert(attrs.begin(), attrs.end());
}
void RegisterRequiredGraphAttrs(
const std::unordered_set<std::string> &attrs) {
required_graph_attrs_.insert(attrs.begin(), attrs.end());
}
mutable bool applied_{false};
std::unordered_set<std::string> required_pass_attrs_;
std::unordered_set<std::string> required_graph_attrs_;
std::map<std::string, boost::any> attrs_;
std::map<std::string, std::function<void(void)>> attr_dels_;
};
using PassCreator = std::function<std::unique_ptr<Pass>()>;
class Registrar {
public:
// In our design, various kinds of passes,
// have their corresponding registry and registrar. The action of
// registration is in the constructor of a global registrar variable, which
// are not used in the code that calls package framework, and would
// be removed from the generated binary file by the linker. To avoid such
// removal, we add Touch to all registrar classes and make USE_PASS macros to
// call this method. So, as long as the callee code calls USE_PASS, the global
// registrar variable won't be removed by the linker.
void Touch() {}
};
virtual std::unique_ptr<Graph> Apply(std::unique_ptr<Graph> graph) const = 0;
class PassRegistry {
public:
static PassRegistry &Instance();
bool Has(const std::string &pass_type) const {
return map_.find(pass_type) != map_.end();
}
void Insert(const std::string &pass_type, const PassCreator &pass_creator) {
PADDLE_ENFORCE(!Has(pass_type), "Pass %s has been registered", pass_type);
map_.insert({pass_type, pass_creator});
}
std::unique_ptr<Pass> Get(const std::string &pass_type) const {
PADDLE_ENFORCE(Has(pass_type), "Pass %s has not been registered",
pass_type);
return map_.at(pass_type)();
}
private:
PassRegistry() = default;
std::unordered_map<std::string, PassCreator> map_;
DISABLE_COPY_AND_ASSIGN(PassRegistry);
};
template <typename PassType>
struct PassRegistrar : public Registrar {
explicit PassRegistrar(const char *pass_type) {
PADDLE_ENFORCE(!PassRegistry::Instance().Has(pass_type),
"'%s' is registered more than once.", pass_type);
PassRegistry::Instance().Insert(
pass_type, [this]() -> std::unique_ptr<Pass> {
std::unique_ptr<Pass> pass(new PassType());
pass->RegisterRequiredPassAttrs(this->required_pass_attrs_);
pass->RegisterRequiredGraphAttrs(this->required_graph_attrs_);
return pass;
});
}
PassRegistrar<PassType> &RequirePassAttr(const std::string &attr) {
required_pass_attrs_.insert(attr);
return *this;
}
PassRegistrar<PassType> &RequireGraphAttr(const std::string &attr) {
required_graph_attrs_.insert(attr);
return *this;
}
private:
std::unordered_set<std::string> required_pass_attrs_;
std::unordered_set<std::string> required_graph_attrs_;
};
#define STATIC_ASSERT_PASS_GLOBAL_NAMESPACE(uniq_name, msg) \
struct __test_global_namespace_##uniq_name##__ {}; \
static_assert(std::is_same<::__test_global_namespace_##uniq_name##__, \
__test_global_namespace_##uniq_name##__>::value, \
msg)
// Register a new pass that can be applied on the IR.
#define REGISTER_PASS(pass_type, pass_class) \
STATIC_ASSERT_PASS_GLOBAL_NAMESPACE( \
__reg_pass__##pass_type, \
"REGISTER_PASS must be called in global namespace"); \
static ::paddle::framework::ir::PassRegistrar<pass_class> \
__pass_registrar_##pass_type##__(#pass_type); \
int TouchPassRegistrar_##pass_type() { \
__pass_registrar_##pass_type##__.Touch(); \
return 0; \
} \
static ::paddle::framework::ir::PassRegistrar<pass_class> \
&__pass_tmp_registrar_##pass_type##__ __attribute__((unused)) = \
__pass_registrar_##pass_type##__
#define USE_PASS(pass_type) \
STATIC_ASSERT_PASS_GLOBAL_NAMESPACE( \
__use_pass_itself_##pass_type, \
"USE_PASS must be called in global namespace"); \
extern int TouchPassRegistrar_##pass_type(); \
static int use_pass_itself_##pass_type##_ __attribute__((unused)) = \
TouchPassRegistrar_##pass_type()
} // namespace ir
} // namespace framework
} // namespace paddle
paddle/fluid/framework/ir/pass_test.cc 0 → 100644
浏览文件 @ a5c96af3
/* 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/pass.h"
#include <string>
#include "gtest/gtest.h"
#include "paddle/fluid/framework/ir/graph.h"
namespace paddle {
namespace framework {
namespace ir {
void BuildCircleGraph(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);
}
class TestPass : public Pass {
protected:
std::unique_ptr<Graph> ApplyImpl(std::unique_ptr<Graph> graph) const {
graph->Set<int>("copy_test_pass_attr", new int);
graph->Set<int>("copy_test_graph_attr", new int);
int test_pass_attr = this->Get<int>("test_pass_attr");
graph->Get<int>("copy_test_pass_attr") = test_pass_attr + 1;
int test_graph_attr = graph->Get<int>("test_graph_attr");
graph->Get<int>("copy_test_graph_attr") = test_graph_attr + 1;
return graph;
}
};
TEST(PassTest, TestPassAttrCheck) {
ProgramDesc prog;
auto pass = PassRegistry::Instance().Get("test_pass");
std::unique_ptr<Graph> graph(new Graph(prog));
std::string exception;
try {
graph = pass->Apply(std::move(graph));
} catch (paddle::platform::EnforceNotMet e) {
exception = std::string(e.what());
}
ASSERT_TRUE(exception.find("test_pass_attr not set") != exception.npos);
int val = 1;
graph.reset(new Graph(prog));
pass->SetNotOwned<int>("test_pass_attr", &val);
try {
graph = pass->Apply(std::move(graph));
} catch (paddle::platform::EnforceNotMet e) {
exception = std::string(e.what());
}
ASSERT_TRUE(exception.find("test_graph_attr not set") != exception.npos);
graph.reset(new Graph(prog));
graph->Set<int>("test_graph_attr", new int);
graph->Get<int>("test_graph_attr") = 1;
graph = pass->Apply(std::move(graph));
ASSERT_EQ(graph->Get<int>("copy_test_pass_attr"), 2);
ASSERT_EQ(graph->Get<int>("copy_test_graph_attr"), 2);
try {
graph = pass->Apply(std::move(graph));
} catch (paddle::platform::EnforceNotMet e) {
exception = std::string(e.what());
}
ASSERT_TRUE(exception.find("Pass can only Apply() once") != exception.npos);
pass = PassRegistry::Instance().Get("test_pass");
pass->SetNotOwned<int>("test_pass_attr", &val);
graph.reset(new Graph(prog));
BuildCircleGraph(graph.get());
graph->Set<int>("test_graph_attr", new int);
graph->Get<int>("test_graph_attr") = 2;
try {
auto tmp = pass->Apply(std::move(graph));
} catch (paddle::platform::EnforceNotMet e) {
exception = std::string(e.what());
}
ASSERT_TRUE(exception.find("shouldn't has cycle") != exception.npos);
}
} // namespace ir
} // namespace framework
} // namespace paddle
REGISTER_PASS(test_pass, paddle::framework::ir::TestPass)
.RequirePassAttr("test_pass_attr")
.RequireGraphAttr("test_graph_attr");
paddle/fluid/framework/parallel_executor.cc
浏览文件 @ a5c96af3
......@@ -19,19 +19,80 @@ limitations under the License. */
#include <vector>
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_viz_pass.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/nccl_helper.h"
#endif
#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
#include "paddle/fluid/framework/details/ssa_graph_builder_factory.h"
#include "paddle/fluid/framework/details/ssa_graph_checker.h"
#include "paddle/fluid/framework/details/ssa_graph_printer.h"
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
#include "paddle/fluid/platform/profiler.h"
namespace paddle {
namespace framework {
std::unique_ptr<ir::Graph> ApplyParallelExecutorPass(
const ProgramDesc &main_program, const std::vector<platform::Place> &places,
const std::string &loss_var_name,
const std::unordered_set<std::string> &param_names,
const std::vector<Scope *> &local_scopes, const bool use_cuda,
#ifdef PADDLE_WITH_CUDA
const BuildStrategy &strategy, platform::NCCLContextMap *nccl_ctxs) {
#else
const BuildStrategy &strategy) {
#endif
// Convert the program to graph.
std::unique_ptr<ir::Graph> graph(new ir::Graph(main_program));
// Apply a graph viz pass to record a graph.
if (!strategy.debug_graphviz_path_.empty()) {
auto viz_pass = ir::PassRegistry::Instance().Get("graph_viz_pass");
const std::string graph_path = string::Sprintf(
"%s%s", strategy.debug_graphviz_path_.c_str(), "_original_graph");
viz_pass->Set<std::string>("graph_viz_path", new std::string(graph_path));
graph = viz_pass->Apply(std::move(graph));
}
// Convert graph to run on multi-devices.
auto multi_device_pass =
ir::PassRegistry::Instance().Get("multi_device_pass");
multi_device_pass->SetNotOwned<const std::vector<platform::Place>>("places",
&places);
multi_device_pass->SetNotOwned<const std::string>("loss_var_name",
&loss_var_name);
multi_device_pass->SetNotOwned<const std::unordered_set<std::string>>(
"params", &param_names);
multi_device_pass->SetNotOwned<const std::vector<Scope *>>("local_scopes",
&local_scopes);
multi_device_pass->SetNotOwned<const BuildStrategy>("strategy", &strategy);
#ifdef PADDLE_WITH_CUDA
platform::NCCLContextMap *nctx = use_cuda ? nccl_ctxs : nullptr;
multi_device_pass->SetNotOwned<platform::NCCLContextMap>("nccl_ctxs", nctx);
#endif
graph = multi_device_pass->Apply(std::move(graph));
// Apply a graph print pass to record a graph with device info.
if (!strategy.debug_graphviz_path_.empty()) {
auto multi_device_print_pass =
ir::PassRegistry::Instance().Get("multi_device_print_pass");
multi_device_print_pass->SetNotOwned<const std::string>(
"debug_graphviz_path", &strategy.debug_graphviz_path_);
multi_device_print_pass->Set<details::GraphvizSSAGraphPrinter>(
"graph_printer", new details::GraphvizSSAGraphPrinter);
graph = multi_device_print_pass->Apply(std::move(graph));
}
// Verify that the graph is correct for multi-device executor.
auto multi_device_check_pass =
ir::PassRegistry::Instance().Get("multi_device_check_pass");
graph = multi_device_check_pass->Apply(std::move(graph));
return graph;
}
class ParallelExecutorPrivate {
public:
explicit ParallelExecutorPrivate(const std::vector<platform::Place> &places)
......@@ -119,21 +180,19 @@ ParallelExecutor::ParallelExecutor(
var_infos.back().persistable_ = var->Persistable();
}
// Step 3. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
details::SSAGraphBuilderFactory builder_factory(
member_->places_, loss_var_name, params, member_->local_scopes_,
build_strategy);
if (member_->use_cuda_) {
// Step 3. Convert main_program to SSA form and dependency graph. Also, insert
// ncclOp
#ifdef PADDLE_WITH_CUDA
builder_factory.SetNCCLContextMap(member_->nccl_ctxs_.get());
std::unique_ptr<ir::Graph> graph = ApplyParallelExecutorPass(
main_program, member_->places_, loss_var_name, params,
member_->local_scopes_, member_->use_cuda_, build_strategy,
member_->nccl_ctxs_.get());
#else
PADDLE_THROW("Not compiled with CUDA.");
std::unique_ptr<ir::Graph> graph = ApplyParallelExecutorPass(
main_program, member_->places_, loss_var_name, params,
member_->local_scopes_, member_->use_cuda_, build_strategy);
#endif
}
builder_ = builder_factory.Create();
std::unique_ptr<ir::Graph> graph(new ir::Graph(main_program));
graph = builder_->Apply(std::move(graph));
member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
exec_strategy, member_->local_scopes_, places, std::move(graph)));
member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
......@@ -146,11 +205,18 @@ void ParallelExecutor::BCastParamsToDevices(
// the initializing bcast, all vars would be bcast from device(0),
// otherwise
// bcast from the specified device.
bool initializing = builder_.get() == nullptr ? true : false;
bool initializing = member_->executor_ ? false : true;
for (auto &var : vars) {
int var_dev_id =
builder_.get() == nullptr ? -1 : builder_->GetVarDeviceID(var);
int var_dev_id = -1;
if (member_->executor_) {
auto &sharded_var_device =
member_->executor_->Graph().Get<details::ShardedVarDevice>(
details::kShardedVarDevice);
if (sharded_var_device.find(var) != sharded_var_device.end()) {
var_dev_id = sharded_var_device.at(var);
}
}
if (!initializing && var_dev_id == -1) continue;
framework::Variable *main_var = nullptr;
......@@ -286,3 +352,8 @@ ParallelExecutor::~ParallelExecutor() {
} // namespace framework
} // namespace paddle
USE_PASS(graph_viz_pass);
USE_PASS(multi_device_pass);
USE_PASS(multi_device_check_pass);
USE_PASS(multi_device_print_pass);
paddle/fluid/framework/parallel_executor.h
浏览文件 @ a5c96af3
......@@ -70,7 +70,6 @@ class ParallelExecutor {
private:
ParallelExecutorPrivate *member_;
std::unique_ptr<details::SSAGraphBuilder> builder_;
};
} // namespace framework
......
paddle/fluid/inference/analysis/CMakeLists.txt
浏览文件 @ a5c96af3
......@@ -6,9 +6,11 @@ cc_library(analysis SRCS pass_manager.cc dot.cc node.cc data_flow_graph.cc graph
tensorrt_subgraph_node_mark_pass.cc
analyzer.cc
helper.cc
model_store_pass.cc
DEPS framework_proto proto_desc)
cc_test(test_node SRCS node_tester.cc DEPS analysis)
cc_test(test_dot SRCS dot_tester.cc DEPS analysis)
cc_binary(inference_analyzer SRCS analyzer_main.cc DEPS analysis)
set(PYTHON_TESTS_DIR ${PADDLE_BINARY_DIR}/python/paddle/fluid/tests)
......@@ -40,3 +42,4 @@ inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_
inference_analysis_test(test_pass_manager SRCS pass_manager_tester.cc)
inference_analysis_test(test_tensorrt_subgraph_node_mark_pass SRCS tensorrt_subgraph_node_mark_pass_tester.cc)
inference_analysis_test(test_analyzer SRCS analyzer_tester.cc)
inference_analysis_test(test_model_store_pass SRCS model_store_pass_tester.cc)
paddle/fluid/inference/analysis/analyzer.cc
浏览文件 @ a5c96af3
......@@ -17,6 +17,7 @@
#include "paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass.h"
#include "paddle/fluid/inference/analysis/dfg_graphviz_draw_pass.h"
#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
#include "paddle/fluid/inference/analysis/model_store_pass.h"
#include "paddle/fluid/inference/analysis/pass_manager.h"
#include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h"
#include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h"
......@@ -29,6 +30,9 @@ DEFINE_bool(inference_analysis_enable_tensorrt_subgraph_engine, false,
DEFINE_string(inference_analysis_graphviz_log_root, "./",
"Graphviz debuger for data flow graphs.");
DEFINE_string(inference_analysis_output_storage_path, "",
"optimized model output path");
namespace inference {
namespace analysis {
......@@ -47,6 +51,9 @@ class DfgPassManagerImpl final : public DfgPassManager {
AddPass("tensorrt-subgraph", new TensorRTSubGraphPass(trt_teller));
}
AddPass("data-flow-graph-to-fluid", new DataFlowGraphToFluidPass);
if (!FLAGS_inference_analysis_output_storage_path.empty()) {
AddPass("model-store-pass", new ModelStorePass);
}
}
std::string repr() const override { return "dfg-pass-manager"; }
......
paddle/fluid/inference/analysis/analyzer.h
浏览文件 @ a5c96af3
......@@ -16,28 +16,23 @@ limitations under the License. */
/*
* This file contains Analyzer, an class that exposed as a library that analyze
* and optimize
* Fluid ProgramDesc for inference. Similar to LLVM, it has multiple flags to
* control whether
* an process is applied on the program.
* and optimize Fluid ProgramDesc for inference. Similar to LLVM, it has
* multiple flags to
* control whether an process is applied on the program.
*
* The processes are called Passes in analysis, the Passes are placed in a
* pipeline, the first
* Pass is the FluidToDataFlowGraphPass which transforms a Fluid ProgramDesc to
* a data flow
* graph, the last Pass is DataFlowGraphToFluidPass which transforms a data flow
* graph to a
* Fluid ProgramDesc. The passes in the middle of the pipeline can be any Passes
* which take a
* node or data flow graph as input.
* pipeline, the first Pass is the FluidToDataFlowGraphPass which transforms a
* Fluid ProgramDesc to
* a data flow graph, the last Pass is DataFlowGraphToFluidPass which transforms
* a data flow graph to a Fluid ProgramDesc. The passes in the middle of the
* pipeline can be any Passes
* which take a node or data flow graph as input.
*
* The Analyzer can be used in two methods, the first is a executable file which
* can be used to
* pre-process the inference model and can be controlled by passing difference
* command flags;
* can be used to pre-process the inference model and can be controlled by
* passing difference command flags;
* the other way is to compose inside the inference API as a runtime pre-process
* phase in the
* inference service.
* phase in the inference service.
*/
#include <gflags/gflags.h>
......@@ -50,6 +45,7 @@ namespace paddle {
// flag if not available.
DECLARE_bool(inference_analysis_enable_tensorrt_subgraph_engine);
DECLARE_string(inference_analysis_graphviz_log_root);
DECLARE_string(inference_analysis_output_storage_path);
namespace inference {
namespace analysis {
......
paddle/fluid/inference/analysis/analyzer_main.cc 0 → 100644
浏览文件 @ a5c96af3
// 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.
/*
* This file implements analysizer -- an executation help to analyze and
* optimize trained model.
*/
#include "paddle/fluid/inference/analysis/analyzer.h"
#include <gflags/gflags.h>
#include <glog/logging.h>
int main(int argc, char** argv) {
google::ParseCommandLineFlags(&argc, &argv, true);
using paddle::inference::analysis::Analyzer;
using paddle::inference::analysis::Argument;
Argument argument;
Analyzer analyzer;
analyzer.Run(&argument);
return 0;
}
paddle/fluid/inference/analysis/analyzer_tester.cc
浏览文件 @ a5c96af3
......@@ -20,14 +20,18 @@ namespace paddle {
namespace inference {
namespace analysis {
TEST_F(DFG_Tester, analysis_without_tensorrt) {
TEST(Analyzer, analysis_without_tensorrt) {
FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = false;
Argument argument;
argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
Analyzer analyser;
analyser.Run(&argument);
}
TEST_F(DFG_Tester, analysis_with_tensorrt) {
TEST(Analyzer, analysis_with_tensorrt) {
FLAGS_inference_analysis_enable_tensorrt_subgraph_engine = true;
Argument argument;
argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
Analyzer analyser;
analyser.Run(&argument);
}
......
paddle/fluid/inference/analysis/argument.h
浏览文件 @ a5c96af3
......@@ -36,6 +36,16 @@ namespace analysis {
* All the fields should be registered here for clearness.
*/
struct Argument {
Argument() = default;
explicit Argument(const std::string& fluid_model_dir)
: fluid_model_dir(new std::string(fluid_model_dir)) {}
// The directory of the trained model.
std::unique_ptr<std::string> fluid_model_dir;
// The path of `__model__` and `param`, this is used when the file name of
// model and param is changed.
std::unique_ptr<std::string> fluid_model_program_path;
std::unique_ptr<std::string> fluid_model_param_path;
// The graph that process by the Passes or PassManagers.
std::unique_ptr<DataFlowGraph> main_dfg;
......@@ -44,6 +54,9 @@ struct Argument {
// The processed program desc.
std::unique_ptr<framework::proto::ProgramDesc> transformed_program_desc;
// The output storage path of ModelStorePass.
std::unique_ptr<std::string> model_output_store_path;
};
#define UNLIKELY(condition) __builtin_expect(static_cast<bool>(condition), 0)
......
paddle/fluid/inference/analysis/data_flow_graph.h
浏览文件 @ a5c96af3
......@@ -36,6 +36,8 @@ namespace analysis {
/*
* DataFlowGraph - A container of Value and Function Nodes.
*
* This is the base graph for any other type of graphs, such as SSA or CFG.
*/
struct DataFlowGraph {
NodeMap nodes;
......
paddle/fluid/inference/analysis/data_flow_graph_tester.cc
浏览文件 @ a5c96af3
......@@ -20,7 +20,7 @@ namespace inference {
namespace analysis {
TEST(DataFlowGraph, BFS) {
auto desc = LoadProgramDesc();
auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__");
auto dfg = ProgramDescToDFG(desc);
dfg.Build();
......@@ -44,7 +44,7 @@ TEST(DataFlowGraph, BFS) {
}
TEST(DataFlowGraph, DFS) {
auto desc = LoadProgramDesc();
auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__");
auto dfg = ProgramDescToDFG(desc);
dfg.Build();
GraphTraits<DataFlowGraph> trait(&dfg);
......
paddle/fluid/inference/analysis/data_flow_graph_to_fluid_pass_tester.cc
浏览文件 @ a5c96af3
......@@ -26,21 +26,21 @@ namespace paddle {
namespace inference {
namespace analysis {
TEST_F(DFG_Tester, Test) {
DataFlowGraph graph;
TEST(DataFlowGraph, Test) {
Argument argument(FLAGS_inference_model_dir);
FluidToDataFlowGraphPass pass0;
DataFlowGraphToFluidPass pass1;
ASSERT_TRUE(pass0.Initialize(&argument));
ASSERT_TRUE(pass1.Initialize(&argument));
pass0.Run(&graph);
pass1.Run(&graph);
pass0.Run(argument.main_dfg.get());
pass1.Run(argument.main_dfg.get());
pass0.Finalize();
pass1.Finalize();
LOG(INFO) << graph.nodes.size();
LOG(INFO) << argument.main_dfg->nodes.size();
}
}; // namespace analysis
......
paddle/fluid/inference/analysis/dfg_graphviz_draw_pass_tester.cc
浏览文件 @ a5c96af3
......@@ -23,12 +23,18 @@ namespace paddle {
namespace inference {
namespace analysis {
TEST_F(DFG_Tester, dfg_graphviz_draw_pass_tester) {
auto dfg = ProgramDescToDFG(*argument.origin_program_desc);
TEST(DFG_GraphvizDrawPass, dfg_graphviz_draw_pass_tester) {
Argument argument(FLAGS_inference_model_dir);
FluidToDataFlowGraphPass pass0;
ASSERT_TRUE(pass0.Initialize(&argument));
pass0.Run(argument.main_dfg.get());
// auto dfg = ProgramDescToDFG(*argument.origin_program_desc);
DFG_GraphvizDrawPass::Config config("./", "test");
DFG_GraphvizDrawPass pass(config);
pass.Initialize(&argument);
pass.Run(&dfg);
pass.Run(argument.main_dfg.get());
// test content
std::ifstream file("./0-graph_test.dot");
......
paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.cc
浏览文件 @ a5c96af3
......@@ -12,6 +12,7 @@ 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 <glog/logging.h>
#include <string>
#include <vector>
......@@ -25,8 +26,20 @@ namespace analysis {
bool FluidToDataFlowGraphPass::Initialize(Argument *argument) {
ANALYSIS_ARGUMENT_CHECK_FIELD(argument);
ANALYSIS_ARGUMENT_CHECK_FIELD(argument->origin_program_desc);
PADDLE_ENFORCE(argument);
if (argument->origin_program_desc) {
LOG(WARNING) << "argument's origin_program_desc is already set, might "
"duplicate called";
}
if (!argument->fluid_model_program_path) {
ANALYSIS_ARGUMENT_CHECK_FIELD(argument->fluid_model_dir);
argument->fluid_model_program_path.reset(
new std::string(*argument->fluid_model_dir + "/__model__"));
}
ANALYSIS_ARGUMENT_CHECK_FIELD(argument->fluid_model_program_path);
auto program = LoadProgramDesc(*argument->fluid_model_program_path);
argument->origin_program_desc.reset(
new framework::proto::ProgramDesc(program));
if (!argument->main_dfg) {
argument->main_dfg.reset(new DataFlowGraph);
}
......@@ -40,6 +53,8 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) {
PADDLE_ENFORCE(graph);
PADDLE_ENFORCE(desc_);
// insert vars
// The `var2id` keeps a map from a variable's name to its Node-id, the Node-id
// will keep updating to its latest alias during the graph-building.
std::unordered_map<std::string, size_t> var2id;
auto &main_block = desc_->blocks(framework::kRootBlockIndex);
for (int i = 0; i < main_block.vars_size(); i++) {
......@@ -51,6 +66,15 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) {
var2id[var.name()] = v->id();
}
// The variables in a SSA can only write once, so if a variable is written
// multiple times(quite common in our ProgramDesc design), multiple alias
// Nodes of this variable will be created, and each will just write once.
// An set that keep all the names of the variables(the original, not alias)
// that have been written(as outputs). Once an Op's output variable hit the
// set, it should create a new alias and update the global alias for this
// variable. And that make a Data Flow Graph a SSA.
std::unordered_set<Node *> unique_written_vars;
for (int i = 0; i < main_block.ops_size(); i++) {
const auto &op = main_block.ops(i);
auto *o = graph->nodes.Create(Node::Type::kFunction);
......@@ -62,33 +86,33 @@ void FluidToDataFlowGraphPass::Run(DataFlowGraph *graph) {
o->SetPbMsg(op.SerializeAsString());
// set inputs and outputs
std::unordered_set<Node *> inlinks;
for (int j = 0; j < op.inputs_size(); j++) {
auto &in_var = op.inputs(j);
for (int k = 0; k < in_var.arguments_size(); k++) {
auto *in = graph->nodes.GetMutable(var2id.at(in_var.arguments(k)));
in->outlinks.push_back(o);
o->inlinks.push_back(in);
inlinks.insert(in);
}
}
for (int j = 0; j < op.outputs_size(); j++) {
auto &out_var = op.outputs(j);
for (int k = 0; k < out_var.arguments_size(); k++) {
auto *out = graph->nodes.GetMutable(var2id[out_var.arguments(k)]);
if (inlinks.count(out)) {
if (unique_written_vars.count(out)) {
// Loop found, for example, a = op(a), use SSA, change to a1 = op(a).
auto *out_alias = graph->nodes.Create(Node::Type::kValue);
out_alias->SetName(out->name());
out_alias->SetPbDesc(out->pb_desc());
out_alias->SetPbMsg(out->pb_msg());
var2id[out_alias->name()] = out_alias->id(); // update a -> a0
var2id[out_alias->name()] =
out_alias->id(); // update variable's alias Node
LOG(INFO) << "loop found in graph, create SSA alias node ["
<< out_alias->repr() << "] for [" << out->repr() << "]";
out = out_alias;
}
out->inlinks.push_back(o);
o->outlinks.push_back(out);
unique_written_vars.insert(out);
}
}
}
......
paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h
浏览文件 @ a5c96af3
......@@ -30,7 +30,7 @@ namespace inference {
namespace analysis {
/*
* Transform a FluidDesc to a data flow graph.
* Transform a FluidDesc to a SSA.
*/
class FluidToDataFlowGraphPass final : public DataFlowGraphPass {
public:
......
paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass_tester.cc
浏览文件 @ a5c96af3
......@@ -21,8 +21,9 @@ namespace paddle {
namespace inference {
namespace analysis {
TEST_F(DFG_Tester, Init) {
TEST(FluidToDataFlowGraphPass, Test) {
FluidToDataFlowGraphPass pass;
Argument argument(FLAGS_inference_model_dir);
pass.Initialize(&argument);
pass.Run(argument.main_dfg.get());
// Analysis is sensitive to ProgramDesc, careful to change the original model.
......
paddle/fluid/inference/analysis/helper.h
浏览文件 @ a5c96af3
......@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include <cstdio>
#include <fstream>
#include <string>
#include <typeindex>
#include <unordered_map>
......@@ -136,6 +137,20 @@ static void ExecShellCommand(const std::string &cmd, std::string *message) {
}
}
static framework::proto::ProgramDesc LoadProgramDesc(
const std::string &model_path) {
std::ifstream fin(model_path, std::ios::in | std::ios::binary);
PADDLE_ENFORCE(fin.is_open(), "Cannot open file %s", model_path);
fin.seekg(0, std::ios::end);
std::string buffer(fin.tellg(), ' ');
fin.seekg(0, std::ios::beg);
fin.read(&buffer[0], buffer.size());
fin.close();
framework::proto::ProgramDesc program_desc;
program_desc.ParseFromString(buffer);
return program_desc;
}
} // namespace analysis
} // namespace inference
} // namespace paddle
......
paddle/fluid/inference/analysis/model_store_pass.cc 0 → 100644
浏览文件 @ a5c96af3
// 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/inference/analysis/model_store_pass.h"
#include <stdio.h>
#include <stdlib.h>
#include "paddle/fluid/inference/analysis/analyzer.h"
#include "paddle/fluid/inference/analysis/argument.h"
namespace paddle {
namespace inference {
namespace analysis {
void ModelStorePass::Run(DataFlowGraph *x) {
if (!argument_->fluid_model_param_path) {
PADDLE_ENFORCE_NOT_NULL(argument_->fluid_model_dir);
argument_->fluid_model_param_path.reset(
new std::string(*argument_->fluid_model_dir + "param"));
}
PADDLE_ENFORCE_NOT_NULL(argument_->model_output_store_path);
// Directly copy param file to destination.
std::stringstream ss;
// NOTE these commands only works on linux.
ss << "mkdir -p " << *argument_->model_output_store_path;
LOG(INFO) << "run command: " << ss.str();
PADDLE_ENFORCE_EQ(system(ss.str().c_str()), 0);
ss.str("");
ss << "cp " << *argument_->fluid_model_dir << "/*"
<< " " << *argument_->model_output_store_path;
LOG(INFO) << "run command: " << ss.str();
PADDLE_ENFORCE_EQ(system(ss.str().c_str()), 0);
// Store program
PADDLE_ENFORCE_NOT_NULL(argument_->transformed_program_desc,
"program desc is not transformed, should call "
"DataFlowGraphToFluidPass first.");
const std::string program_output_path =
*argument_->model_output_store_path + "/__model__";
std::ofstream file(program_output_path, std::ios::binary);
PADDLE_ENFORCE(file.is_open(), "failed to open %s to write.",
program_output_path);
const std::string serialized_message =
argument_->transformed_program_desc->SerializeAsString();
file.write(serialized_message.c_str(), serialized_message.size());
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/framework/details/ssa_graph_builder_factory.cc → paddle/fluid/inference/analysis/model_store_pass.h
浏览文件 @ a5c96af3
......@@ -12,39 +12,40 @@
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/framework/details/ssa_graph_builder_factory.h"
#include <fstream>
#include "paddle/fluid/framework/details/multi_devices_graph_builder.h"
#include "paddle/fluid/framework/details/ssa_graph_checker.h"
#include "paddle/fluid/framework/details/ssa_graph_printer.h"
/*
* This file defines ModelStorePass, which store the runtime DFG to a Paddle
* model in the disk, and that model can be reloaded for prediction.
*/
#include "paddle/fluid/inference/analysis/pass.h"
namespace paddle {
namespace framework {
namespace details {
std::unique_ptr<SSAGraphBuilder> SSAGraphBuilderFactory::Create() {
std::unique_ptr<SSAGraphBuilder> res(
#ifdef PADDLE_WITH_CUDA
new MultiDevSSAGraphBuilder(places_, loss_var_name_, param_names_,
local_scopes_, nccl_ctxs_, strategy_)
#else
new MultiDevSSAGraphBuilder(places_, loss_var_name_, param_names_,
local_scopes_, strategy_)
#endif
); // NOLINT
if (!strategy_.debug_graphviz_path_.empty()) {
std::unique_ptr<std::ostream> fout(
new std::ofstream(strategy_.debug_graphviz_path_));
PADDLE_ENFORCE(fout->good());
std::unique_ptr<GraphvizSSAGraphPrinter> graphviz_printer(
new GraphvizSSAGraphPrinter());
res.reset(new SSAGraghBuilderWithPrinter(
std::move(fout), std::move(graphviz_printer), std::move(res)));
namespace inference {
namespace analysis {
class ModelStorePass : public DataFlowGraphPass {
public:
bool Initialize(Argument* argument) override {
if (!argument) {
LOG(ERROR) << "invalid argument";
return false;
}
argument_ = argument;
return true;
}
void Run(DataFlowGraph* x) override;
std::string repr() const override { return "DFG-store-pass"; }
std::string description() const override {
return R"DD(This file defines ModelStorePass, which store the runtime DFG to a Paddle
model in the disk, and that model can be reloaded for prediction again.)DD";
}
res.reset(new SSAGraghBuilderWithChecker(std::move(res)));
return res;
}
} // namespace details
} // namespace framework
private:
Argument* argument_{nullptr};
};
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/model_store_pass_tester.cc 0 → 100644
浏览文件 @ a5c96af3
// 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/inference/analysis/model_store_pass.h"
#include <gflags/gflags.h>
#include <gtest/gtest.h>
#include "paddle/fluid/inference/analysis/analyzer.h"
namespace paddle {
namespace inference {
namespace analysis {
DEFINE_string(inference_model_dir, "", "Model path");
TEST(DFG_StorePass, test) {
Analyzer analyzer;
Argument argument(FLAGS_inference_model_dir);
argument.model_output_store_path.reset(
new std::string("./_dfg_store_pass_tmp"));
// disable storage in alalyzer
FLAGS_inference_analysis_output_storage_path = "";
analyzer.Run(&argument);
ModelStorePass pass;
pass.Initialize(&argument);
pass.Run(argument.main_dfg.get());
}
} // namespace analysis
} // namespace inference
} // namespace paddle
paddle/fluid/inference/analysis/pass.h
浏览文件 @ a5c96af3
......@@ -50,6 +50,7 @@ class Pass {
// Create a debugger Pass that draw the DFG by graphviz toolkit.
virtual Pass *CreateGraphvizDebugerPass() const { return nullptr; }
virtual void Run() { LOG(FATAL) << "not valid"; }
// Run on a single Node.
virtual void Run(Node *x) { LOG(FATAL) << "not valid"; }
// Run on a single Function.
......
paddle/fluid/inference/analysis/pass_manager_tester.cc
浏览文件 @ a5c96af3
......@@ -56,7 +56,7 @@ class TestNodePass final : public NodePass {
std::string description() const override { return "some doc"; }
};
TEST_F(DFG_Tester, DFG_pass_manager) {
TEST(PassManager, DFG_pass_manager) {
TestDfgPassManager manager;
DFG_GraphvizDrawPass::Config config("./", "dfg.dot");
......@@ -64,12 +64,15 @@ TEST_F(DFG_Tester, DFG_pass_manager) {
manager.Register("graphviz", new DFG_GraphvizDrawPass(config));
manager.Register("dfg-to-fluid", new DataFlowGraphToFluidPass);
Argument argument(FLAGS_inference_model_dir);
ASSERT_TRUE(&argument);
ASSERT_TRUE(manager.Initialize(&argument));
manager.RunAll();
}
TEST_F(DFG_Tester, Node_pass_manager) {
TEST(PassManager, Node_pass_manager) {
Argument argument(FLAGS_inference_model_dir);
// Pre-process: initialize the DFG with the ProgramDesc first.
FluidToDataFlowGraphPass pass0;
pass0.Initialize(&argument);
......
paddle/fluid/inference/analysis/subgraph_splitter_tester.cc
浏览文件 @ a5c96af3
......@@ -31,8 +31,8 @@ SubGraphSplitter::NodeInsideSubgraphTeller teller = [](const Node* node) {
return false;
};
TEST_F(DFG_Tester, Split) {
auto desc = LoadProgramDesc();
TEST(SubGraphSplitter, Split) {
auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__");
auto dfg = ProgramDescToDFG(desc);
LOG(INFO) << "spliter\n" << dfg.DotString();
......@@ -63,8 +63,8 @@ TEST_F(DFG_Tester, Split) {
ASSERT_EQ(subgraphs.back().size(), 6UL);
}
TEST_F(DFG_Tester, Fuse) {
auto desc = LoadProgramDesc();
TEST(SubGraphSplitter, Fuse) {
auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__");
auto dfg = ProgramDescToDFG(desc);
size_t count0 = dfg.nodes.size();
......
paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass_tester.cc
浏览文件 @ a5c96af3
......@@ -22,11 +22,11 @@ namespace paddle {
namespace inference {
namespace analysis {
TEST_F(DFG_Tester, tensorrt_subgraph_node_mark_pass) {
TEST(TensorRTSubgraphNodeMarkPass, test) {
// init
FluidToDataFlowGraphPass pass;
Argument argument(FLAGS_inference_model_dir);
ASSERT_TRUE(pass.Initialize(&argument));
argument.main_dfg.reset(new DataFlowGraph);
pass.Run(argument.main_dfg.get());
TensorRTSubgraphNodeMarkPass::teller_t teller = [](const Node* node) {
......@@ -41,7 +41,7 @@ TEST_F(DFG_Tester, tensorrt_subgraph_node_mark_pass) {
for (auto& node : argument.main_dfg->nodes.nodes()) {
counter += node->attr(ATTR_supported_by_tensorrt).Bool();
}
ASSERT_EQ(counter, 2);
LOG(INFO) << counter << " nodes marked";
}
......
paddle/fluid/inference/analysis/tensorrt_subgraph_pass_tester.cc
浏览文件 @ a5c96af3
......@@ -25,7 +25,7 @@ namespace analysis {
DEFINE_string(dot_dir, "./", "");
TEST_F(DFG_Tester, tensorrt_single_pass) {
TEST(TensorRTSubGraphPass, main) {
std::unordered_set<std::string> teller_set(
{"elementwise_add", "mul", "sigmoid"});
SubGraphSplitter::NodeInsideSubgraphTeller teller = [&](const Node* node) {
......@@ -35,7 +35,8 @@ TEST_F(DFG_Tester, tensorrt_single_pass) {
return false;
};
LOG(INFO) << "init";
Argument argument(FLAGS_inference_model_dir);
DFG_GraphvizDrawPass::Config config{FLAGS_dot_dir, "origin"};
DFG_GraphvizDrawPass::Config config1{FLAGS_dot_dir, "fusion"};
......@@ -44,13 +45,11 @@ TEST_F(DFG_Tester, tensorrt_single_pass) {
FluidToDataFlowGraphPass pass0;
TensorRTSubGraphPass trt_pass(std::move(teller));
LOG(INFO) << "Initialize";
dfg_pass.Initialize(&argument);
dfg_pass1.Initialize(&argument);
pass0.Initialize(&argument);
trt_pass.Initialize(&argument);
LOG(INFO) << "Run";
argument.main_dfg.reset(new DataFlowGraph);
pass0.Run(argument.main_dfg.get());
dfg_pass.Run(argument.main_dfg.get());
......
paddle/fluid/inference/analysis/ut_helper.h
浏览文件 @ a5c96af3
......@@ -20,7 +20,7 @@ limitations under the License. */
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/inference/analysis/data_flow_graph.h"
#include "paddle/fluid/inference/analysis/fluid_to_data_flow_graph_pass.h"
#include "paddle/fluid/inference/analysis/ut_helper.h"
#include "paddle/fluid/inference/analysis/helper.h"
namespace paddle {
namespace inference {
......@@ -32,27 +32,12 @@ namespace analysis {
DEFINE_string(inference_model_dir, "", "inference test model dir");
static framework::proto::ProgramDesc LoadProgramDesc(
const std::string& model_dir = FLAGS_inference_model_dir) {
std::string msg;
std::string net_file = FLAGS_inference_model_dir + "/__model__";
std::ifstream fin(net_file, std::ios::in | std::ios::binary);
PADDLE_ENFORCE(static_cast<bool>(fin), "Cannot open file %s", net_file);
fin.seekg(0, std::ios::end);
msg.resize(fin.tellg());
fin.seekg(0, std::ios::beg);
fin.read(&(msg.at(0)), msg.size());
fin.close();
framework::proto::ProgramDesc program_desc;
program_desc.ParseFromString(msg);
return program_desc;
}
static DataFlowGraph ProgramDescToDFG(
const framework::proto::ProgramDesc& desc) {
DataFlowGraph graph;
FluidToDataFlowGraphPass pass;
Argument argument;
argument.fluid_model_dir.reset(new std::string(FLAGS_inference_model_dir));
argument.origin_program_desc.reset(new framework::proto::ProgramDesc(desc));
pass.Initialize(&argument);
pass.Run(&graph);
......@@ -63,7 +48,7 @@ static DataFlowGraph ProgramDescToDFG(
class DFG_Tester : public ::testing::Test {
protected:
void SetUp() override {
auto desc = LoadProgramDesc(FLAGS_inference_model_dir);
auto desc = LoadProgramDesc(FLAGS_inference_model_dir + "/__model__");
argument.origin_program_desc.reset(new framework::proto::ProgramDesc(desc));
}
......
paddle/fluid/inference/api/api_anakin_engine_tester.cc
浏览文件 @ a5c96af3
......@@ -37,19 +37,21 @@ TEST(inference, anakin) {
float data[1 * 3 * 224 * 224] = {1.0f};
PaddleTensor tensor{.name = "input_0",
.shape = std::vector<int>({1, 3, 224, 224}),
.data = PaddleBuf(data, sizeof(data)),
.dtype = PaddleDType::FLOAT32};
PaddleTensor tensor;
tensor.name = "input_0";
tensor.shape = std::vector<int>({1, 3, 224, 224});
tensor.data = PaddleBuf(data, sizeof(data));
tensor.dtype = PaddleDType::FLOAT32;
// For simplicity, we set all the slots with the same data.
std::vector<PaddleTensor> paddle_tensor_feeds;
paddle_tensor_feeds.emplace_back(std::move(tensor));
PaddleTensor tensor_out{.name = "prob_out",
.shape = std::vector<int>({1000, 1}),
.data = PaddleBuf(),
.dtype = PaddleDType::FLOAT32};
PaddleTensor tensor_out;
tensor_out.name = "prob_out";
tensor_out.shape = std::vector<int>({1000, 1});
tensor_out.data = PaddleBuf();
tensor_out.dtype = PaddleDType::FLOAT32;
std::vector<PaddleTensor> outputs;
outputs.emplace_back(std::move(tensor_out));
......
paddle/fluid/inference/api/api_impl.cc
浏览文件 @ a5c96af3
......@@ -183,6 +183,13 @@ bool NativePaddlePredictor::SetFeed(const std::vector<PaddleTensor> &inputs,
// TODO(panyx0718): Init LoDTensor from existing memcpy to save a copy.
std::memcpy(static_cast<void *>(input_ptr), inputs[i].data.data(),
inputs[i].data.length());
// TODO(Superjomn) Low performance, need optimization for heavy LoD copy.
framework::LoD lod;
for (auto &level : inputs[i].lod) {
lod.emplace_back(level);
}
input.set_lod(lod);
feeds->push_back(input);
}
return true;
......@@ -248,6 +255,10 @@ bool NativePaddlePredictor::GetFetch(
buffer.Resize(sizeof(float) * data.size());
}
std::memcpy(buffer.data(), data.data(), buffer.length());
// copy LoD
for (const auto &level : fetchs[i].lod()) {
outputs->at(i).lod.emplace_back(level);
}
outputs->at(i).dtype = PaddleDType::FLOAT32;
// TODO(panyx0718): support other types? fill tensor name? avoid a copy.
}
......
paddle/fluid/inference/api/api_tensorrt_subgraph_engine.cc
浏览文件 @ a5c96af3
......@@ -90,6 +90,18 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor {
void OptimizeInferenceProgram() {
// Analyze inference_program
Argument argument;
if (!config_.model_dir.empty()) {
argument.fluid_model_dir.reset(new std::string(config_.model_dir));
} else {
PADDLE_ENFORCE(
!config_.param_file.empty(),
"Either model_dir or (param_file, prog_file) should be set.");
PADDLE_ENFORCE(!config_.prog_file.empty());
argument.fluid_model_program_path.reset(
new std::string(config_.prog_file));
argument.fluid_model_param_path.reset(
new std::string(config_.param_file));
}
argument.origin_program_desc.reset(
new ProgramDesc(*inference_program_->Proto()));
Singleton<Analyzer>::Global().Run(&argument);
......
paddle/fluid/inference/api/api_tensorrt_subgraph_engine_tester.cc
浏览文件 @ a5c96af3
......@@ -49,11 +49,10 @@ void CompareTensorRTWithFluid(bool enable_tensorrt) {
std::vector<int64_t> data(20);
for (int i = 0; i < 20; i++) data[i] = i;
PaddleTensor tensor{
.name = "",
.shape = std::vector<int>({10, 1}),
.data = PaddleBuf(data.data(), data.size() * sizeof(int64_t)),
.dtype = PaddleDType::INT64};
PaddleTensor tensor;
tensor.shape = std::vector<int>({10, 1});
tensor.data = PaddleBuf(data.data(), data.size() * sizeof(int64_t));
tensor.dtype = PaddleDType::INT64;
// For simplicity, we set all the slots with the same data.
std::vector<PaddleTensor> slots(4, tensor);
......
paddle/fluid/inference/api/demo_ci/simple_on_word2vec.cc
浏览文件 @ a5c96af3
......@@ -47,10 +47,10 @@ void Main(bool use_gpu) {
//# 2. Prepare input.
int64_t data[4] = {1, 2, 3, 4};
PaddleTensor tensor{.name = "",
.shape = std::vector<int>({4, 1}),
.data = PaddleBuf(data, sizeof(data)),
.dtype = PaddleDType::INT64};
PaddleTensor tensor;
tensor.shape = std::vector<int>({4, 1});
tensor.data = PaddleBuf(data, sizeof(data));
tensor.dtype = PaddleDType::INT64;
// For simplicity, we set all the slots with the same data.
std::vector<PaddleTensor> slots(4, tensor);
......@@ -94,10 +94,11 @@ void MainThreads(int num_threads, bool use_gpu) {
for (int batch_id = 0; batch_id < num_batches; ++batch_id) {
// 2. Dummy Input Data
int64_t data[4] = {1, 2, 3, 4};
PaddleTensor tensor{.name = "",
.shape = std::vector<int>({4, 1}),
.data = PaddleBuf(data, sizeof(data)),
.dtype = PaddleDType::INT64};
PaddleTensor tensor;
tensor.shape = std::vector<int>({4, 1});
tensor.data = PaddleBuf(data, sizeof(data));
tensor.dtype = PaddleDType::INT64;
std::vector<PaddleTensor> inputs(4, tensor);
std::vector<PaddleTensor> outputs;
// 3. Run
......
paddle/fluid/inference/api/demo_ci/vis_demo.cc
浏览文件 @ a5c96af3
......@@ -123,11 +123,11 @@ void Main(bool use_gpu) {
file.close();
// Inference.
PaddleTensor input{
.name = "xx",
.shape = record.shape,
.data = PaddleBuf(record.data.data(), record.data.size() * sizeof(float)),
.dtype = PaddleDType::FLOAT32};
PaddleTensor input;
input.shape = record.shape;
input.data =
PaddleBuf(record.data.data(), record.data.size() * sizeof(float));
input.dtype = PaddleDType::FLOAT32;
VLOG(3) << "run executor";
std::vector<PaddleTensor> output;
......
paddle/fluid/inference/api/paddle_inference_api.h
浏览文件 @ a5c96af3
......@@ -67,9 +67,9 @@ struct PaddleTensor {
PaddleTensor() = default;
std::string name; // variable name.
std::vector<int> shape;
// TODO(Superjomn) for LoD support, add a vector<vector<int>> field if needed.
PaddleBuf data; // blob of data.
PaddleDType dtype;
std::vector<std::vector<uint64_t>> lod; // lod data
};
enum class PaddleEngineKind {
......
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @ a5c96af3
......@@ -19,12 +19,17 @@ limitations under the License. */
#include <thread> // NOLINT
#include <vector>
#include "gflags/gflags.h"
#include "paddle/fluid/operators/detail/macros.h"
#include "paddle/fluid/operators/distributed/request_handler_impl.h"
#include "paddle/fluid/operators/listen_and_serv_op.h"
#include "paddle/fluid/platform/profiler.h"
DEFINE_int32(listen_and_serv_profile_period, 0,
"the period of listen_and_serv to do profile");
namespace paddle {
namespace operators {
......@@ -122,7 +127,18 @@ void ListenAndServOp::RunSyncLoop(
std::shared_ptr<framework::ExecutorPrepareContext>(nullptr));
rpc_service_->ResetBarrierCounter();
int32_t profile_step = 0;
while (true) {
PADDLE_ENFORCE_LE(profile_step, FLAGS_listen_and_serv_profile_period,
"profile_step should not be larger then "
"FLAGS_listen_and_serv_profile_period");
if (FLAGS_listen_and_serv_profile_period > 0) {
if (profile_step == 0) {
auto pf_state = paddle::platform::ProfilerState::kCPU;
paddle::platform::EnableProfiler(pf_state);
}
}
// Get from multiple trainers, we don't care about the order in which
// the gradients arrives, just add suffix 0~n and merge the gradient.
rpc_service_->SetCond(distributed::kRequestSend);
......@@ -164,6 +180,15 @@ void ListenAndServOp::RunSyncLoop(
// reset received sparse vars to avoid reuse it in the next mini-batch
dynamic_cast<distributed::RequestSendHandler *>(request_send_handler_.get())
->ResetSparseVarRecorder();
if (FLAGS_listen_and_serv_profile_period > 0) {
if (profile_step == FLAGS_listen_and_serv_profile_period) {
paddle::platform::DisableProfiler(
paddle::platform::EventSortingKey::kTotal, "/dev/null");
profile_step = 0;
} else {
profile_step++;
}
}
} // while(true)
}
......
paddle/fluid/operators/math/im2col.cc
浏览文件 @ a5c96af3
......@@ -14,6 +14,7 @@ limitations under the License. */
#include "paddle/fluid/operators/math/im2col.h"
#include <vector>
#include "paddle/fluid/operators/math/im2col_cfo_cpu.h"
namespace paddle {
namespace operators {
......@@ -35,61 +36,18 @@ class Im2ColFunctor<paddle::operators::math::ColFormat::kCFO,
PADDLE_ENFORCE(im.dims().size() == 3);
PADDLE_ENFORCE(col->dims().size() == 5);
int im_channels = im.dims()[0];
int im_height = im.dims()[1];
int im_width = im.dims()[2];
int filter_height = col->dims()[1];
int filter_width = col->dims()[2];
int output_height = col->dims()[3];
int output_width = col->dims()[4];
int channels_col = im_channels * filter_height * filter_width;
const T* im_data = im.data<T>();
T* col_data = col->data<T>();
// TODO(TJ): change me to template
// further optimaze:
// 1. padding != 1
// 2. could also support stride_h != 1
if (stride[0] == 1 && stride[1] == 1 && dilation[0] == 1 &&
dilation[1] == 1 && padding[0] == 0 && padding[1] == 0) {
int col_matrix_width = output_width * output_height;
size_t copy_size = sizeof(T) * output_width;
for (int oh = 0; oh < output_height; ++oh) {
const T* im_data_start = im_data + oh * im_width;
T* dst_data = col_data + oh * output_width;
for (int ic = 0; ic < im_channels; ++ic) {
const T* src_data = im_data_start + ic * im_height * im_width;
for (int kh = 0; kh < filter_height; ++kh) {
for (int kw = 0; kw < filter_width; ++kw) {
std::memcpy(dst_data, src_data + kw, copy_size);
dst_data = dst_data + col_matrix_width;
}
src_data = src_data + im_width;
}
}
}
return;
}
for (int c = 0; c < channels_col; ++c) {
int w_offset = c % filter_width;
int h_offset = (c / filter_width) % filter_height;
int c_im = c / (filter_width * filter_height);
for (int h = 0; h < output_height; ++h) {
int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0];
for (int w = 0; w < output_width; ++w) {
int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1];
int col_idx = (c * output_height + h) * output_width + w;
int im_idx = (im_row_idx + c_im * im_height) * im_width + im_col_idx;
col_data[col_idx] = (im_row_idx < 0 || im_row_idx >= im_height ||
im_col_idx < 0 || im_col_idx >= im_width)
? static_cast<T>(0)
: im_data[im_idx];
}
dilation[1] == 1) {
if (padding[0] == 0 && padding[1] == 0) {
im2col_sh1sw1dh1dw1ph0pw0<T>(im, col);
return;
} else if (padding[0] == 1 && padding[1] == 1) {
im2col_sh1sw1dh1dw1ph1pw1<T>(im, col);
return;
}
// TODO(TJ): complete padding >=2
}
im2col_common<T>(im, dilation, stride, padding, col);
}
};
......
paddle/fluid/operators/math/im2col_cfo_cpu.h 0 → 100644
浏览文件 @ a5c96af3
/* Copyright (c) 2016 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 <vector>
#include "paddle/fluid/framework/tensor.h"
namespace paddle {
namespace operators {
namespace math {
/**
* The most common im2col algorithm.
* Support dilation, stride and padding.
*/
template <typename T>
inline void im2col_common(const framework::Tensor& im,
const std::vector<int>& dilation,
const std::vector<int>& stride,
const std::vector<int>& padding,
framework::Tensor* col) {
int im_channels = im.dims()[0];
int im_height = im.dims()[1];
int im_width = im.dims()[2];
int filter_height = col->dims()[1];
int filter_width = col->dims()[2];
int output_height = col->dims()[3];
int output_width = col->dims()[4];
int channels_col = im_channels * filter_height * filter_width;
const T* im_data = im.data<T>();
T* col_data = col->data<T>();
for (int c = 0; c < channels_col; ++c) {
int w_offset = c % filter_width;
int h_offset = (c / filter_width) % filter_height;
int c_im = c / (filter_width * filter_height);
for (int h = 0; h < output_height; ++h) {
int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0];
for (int w = 0; w < output_width; ++w) {
int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1];
int col_idx = (c * output_height + h) * output_width + w;
int im_idx = (im_row_idx + c_im * im_height) * im_width + im_col_idx;
col_data[col_idx] = (im_row_idx < 0 || im_row_idx >= im_height ||
im_col_idx < 0 || im_col_idx >= im_width)
? static_cast<T>(0)
: im_data[im_idx];
}
}
}
}
/**
* im2col algorithm with strides == 1, dilations == 1, paddings == 0
*/
template <typename T>
inline void im2col_sh1sw1dh1dw1ph0pw0(const framework::Tensor& im,
framework::Tensor* col) {
int im_channels = im.dims()[0];
int im_height = im.dims()[1];
int im_width = im.dims()[2];
int filter_height = col->dims()[1];
int filter_width = col->dims()[2];
int output_height = col->dims()[3];
int output_width = col->dims()[4];
const T* im_data = im.data<T>();
T* col_data = col->data<T>();
int col_matrix_width = output_width * output_height;
int im_size = im_height * im_width;
size_t copy_size = sizeof(T) * output_width;
const T* im_data_oh = im_data;
T* dst_data_oh = col_data;
for (int oh = 0; oh < output_height; ++oh) {
const T* src_data_ic = im_data_oh;
T* dst_data = dst_data_oh;
for (int ic = 0; ic < im_channels; ++ic) {
const T* src_data = src_data_ic;
for (int kh = 0; kh < filter_height; ++kh) {
for (int kw = 0; kw < filter_width; ++kw) {
std::memcpy(dst_data, src_data + kw, copy_size);
dst_data = dst_data + col_matrix_width;
}
src_data = src_data + im_width;
}
src_data_ic = src_data_ic + im_size;
}
im_data_oh = im_data_oh + im_width;
dst_data_oh = dst_data_oh + output_width;
}
}
/**
* im2col algorithm with strides == 1, dilations == 1, paddings == 1
* and filter_width == 1 have a special implementation
*/
template <typename T>
inline void im2col_sh1sw1dh1dw1ph1pw1(const framework::Tensor& im,
framework::Tensor* col) {
int im_channels = im.dims()[0];
int im_height = im.dims()[1];
int im_width = im.dims()[2];
int filter_height = col->dims()[1];
int filter_width = col->dims()[2];
int output_height = col->dims()[3];
int output_width = col->dims()[4];
constexpr int plh = 1;
constexpr int prh = 1;
constexpr int plw = 1;
constexpr int prw = 1;
const T* im_data = im.data<T>();
T* col_data = col->data<T>();
int im_size = im_height * im_width;
int col_matrix_width = output_width * output_height;
int col_block_fh = filter_width * col_matrix_width; // fw*oh*ow
int col_block_ic = filter_height * col_block_fh; // fh*fw*oh*ow
// fill height padding
{
size_t copy_size = sizeof(T) * output_width;
T* col_start_l = col_data;
T* col_start_r = col_data + (filter_height - 1) * col_block_fh +
col_matrix_width - output_width;
for (int ic = 0; ic < im_channels; ++ic) {
T* dst_data_l = col_start_l;
T* dst_data_r = col_start_r;
for (int kw = 0; kw < filter_width; ++kw) {
std::memset(dst_data_l, 0, copy_size);
std::memset(dst_data_r, 0, copy_size);
dst_data_l = dst_data_l + col_matrix_width;
dst_data_r = dst_data_r + col_matrix_width;
}
col_start_l = col_start_l + col_block_ic;
col_start_r = col_start_r + col_block_ic;
}
}
auto pad = static_cast<T>(0);
if (filter_width == 1) {
// fill width padding
T* dst_data_ic = col_data;
for (int ic = 0; ic < im_channels; ++ic) {
T* dst_data_kh = dst_data_ic;
for (int kh = 0; kh < filter_height; ++kh) {
T* dst_data = dst_data_kh;
for (int oh = 0; oh < output_height; ++oh) {
*dst_data = pad;
dst_data = dst_data + output_width - 1;
*dst_data = pad;
++dst_data;
}
dst_data_kh = dst_data_kh + col_block_fh;
}
dst_data_ic = dst_data_ic + col_block_ic;
}
// fill core
size_t copy_size = sizeof(T) * (output_width - plw - prw);
for (int oh = 0; oh < output_height; ++oh) {
const T* im_data_start =
im_data + (oh - plh > 0 ? oh - plh : 0) * im_width;
T* dst_data = col_data + oh * output_width;
for (int ic = 0; ic < im_channels; ++ic) {
const T* src_data = im_data_start + ic * im_size;
for (int kh = 0; kh < filter_height; ++kh) {
if ((oh < plh && kh < plh) || (oh > (output_height - prh - 1) &&
kh > (filter_height - prh - 1))) {
dst_data = dst_data + col_matrix_width;
continue;
}
std::memcpy(dst_data + plw, src_data, copy_size);
dst_data = dst_data + col_matrix_width;
src_data = src_data + im_width;
}
}
}
return;
}
// filter_width != 1
// fill width padding
T* dst_data_ic = col_data;
for (int ic = 0; ic < im_channels; ++ic) {
T* dst_data_kh = dst_data_ic;
for (int kh = 0; kh < filter_height; ++kh) {
for (T* dst_data :
{dst_data_kh, dst_data_kh + (filter_width - prw) * col_matrix_width +
output_width - 1}) {
// TODO(TJ): from plh, saving repeated assignment
for (int oh = 0; oh < output_height; ++oh) {
*dst_data = pad;
dst_data = dst_data + output_width;
}
}
dst_data_kh = dst_data_kh + col_block_fh;
}
dst_data_ic = dst_data_ic + col_block_ic;
}
// TODO(TJ): use array like: size_t copy_size[kw]={sizeof(T) *
// (output_width-1)}
// length of copy_size is equal kw.
for (int oh = 0; oh < output_height; ++oh) {
const T* im_data_start = im_data + (oh - plh > 0 ? oh - plh : 0) * im_width;
T* dst_data = col_data + oh * output_width;
for (int ic = 0; ic < im_channels; ++ic) {
const T* src_data = im_data_start + ic * im_size;
for (int kh = 0; kh < filter_height; ++kh) {
if ((oh < plh && kh < plh) || (oh > (output_height - prh - 1) &&
kh > (filter_height - prh - 1))) {
dst_data = dst_data + filter_width * col_matrix_width;
continue;
}
// TODO(TJ): reuse plw-kw outside this for
// try to unify
for (int kw = 0; kw < plw; ++kw) {
std::memcpy(dst_data + (plw - kw), src_data,
sizeof(T) * (output_width - (plw - kw)));
dst_data = dst_data + col_matrix_width;
}
for (int kw = plw; kw < filter_width - prw; ++kw) {
std::memcpy(dst_data, src_data + (kw - plw),
sizeof(T) * output_width);
dst_data = dst_data + col_matrix_width;
}
int i = 1;
for (int kw = filter_width - prw; kw < filter_width; ++kw, ++i) {
std::memcpy(dst_data, src_data + (kw - plw),
sizeof(T) * (output_width - i));
dst_data = dst_data + col_matrix_width;
}
src_data = src_data + im_width;
}
}
}
}
} // namespace math
} // namespace operators
} // namespace paddle
paddle/fluid/operators/math/im2col_test.cc
浏览文件 @ a5c96af3
......@@ -14,7 +14,9 @@ limitations under the License. */
#include "paddle/fluid/operators/math/im2col.h"
#include <gtest/gtest.h>
#include <sys/time.h>
#include <vector>
#include "paddle/fluid/operators/math/im2col_cfo_cpu.h"
template <typename DeviceContext, typename Place>
void testIm2col() {
......@@ -160,82 +162,111 @@ void testIm2col() {
delete context;
}
void testIm2colCPU(int ic, int ih, int iw, int fh, int fw, int ph, int pw) {
paddle::framework::Tensor input;
paddle::framework::Tensor output;
paddle::framework::Tensor ref_output;
std::vector<int> padding({ph, pw});
std::vector<int> stride({1, 1}); // stride_y, stride_x
std::vector<int> dilation({1, 1}); // dilation_y, dilation_x
int output_height = (ih - fh + padding[0] * 2) / stride[0] + 1;
int output_width = (iw - fw + padding[1] * 2) / stride[1] + 1;
float* input_ptr =
input.mutable_data<float>({ic, ih, iw}, paddle::platform::CPUPlace());
for (int i = 0; i < input.numel(); ++i) {
input_ptr[i] = static_cast<float>(i + 1);
}
paddle::platform::CPUPlace place;
paddle::platform::CPUDeviceContext context(place);
output.mutable_data<float>({ic, fh, fw, output_height, output_width}, place);
ref_output.mutable_data<float>({ic, fh, fw, output_height, output_width},
place);
paddle::operators::math::Im2ColFunctor<
paddle::operators::math::ColFormat::kCFO,
paddle::platform::CPUDeviceContext, float>
im2col;
im2col(context, input, dilation, stride, padding, &output);
auto ref_im2col = [&](
const paddle::framework::Tensor& im, const std::vector<int>& dilation,
const std::vector<int>& stride, const std::vector<int>& padding,
paddle::framework::Tensor* col) {
int im_channels = im.dims()[0];
int im_height = im.dims()[1];
int im_width = im.dims()[2];
int filter_height = col->dims()[1];
int filter_width = col->dims()[2];
int output_height = col->dims()[3];
int output_width = col->dims()[4];
int channels_col = im_channels * filter_height * filter_width;
const float* im_data = im.data<float>();
float* col_data = col->data<float>();
for (int c = 0; c < channels_col; ++c) {
int w_offset = c % filter_width;
int h_offset = (c / filter_width) % filter_height;
int c_im = c / (filter_width * filter_height);
for (int h = 0; h < output_height; ++h) {
int im_row_idx = h * stride[0] - padding[0] + h_offset * dilation[0];
for (int w = 0; w < output_width; ++w) {
int im_col_idx = w * stride[1] - padding[1] + w_offset * dilation[1];
int col_idx = (c * output_height + h) * output_width + w;
int im_idx = (im_row_idx + c_im * im_height) * im_width + im_col_idx;
col_data[col_idx] = (im_row_idx < 0 || im_row_idx >= im_height ||
im_col_idx < 0 || im_col_idx >= im_width)
? 0.f
: im_data[im_idx];
}
}
}
};
ref_im2col(input, dilation, stride, padding, &ref_output);
float* out_cfo_ptr = output.data<float>();
float* out_ref_ptr = ref_output.data<float>();
for (int i = 0; i < output.numel(); ++i) {
EXPECT_EQ(out_cfo_ptr[i], out_ref_ptr[i]);
}
}
TEST(math, im2col) {
testIm2col<paddle::platform::CPUDeviceContext, paddle::platform::CPUPlace>();
testIm2colCPU(/*ic*/ 3, /*ih*/ 5, /*iw*/ 5, /*fh*/ 3, /*fw*/ 2, /*ph*/ 0,
/*pw*/ 0);
testIm2colCPU(/*ic*/ 2, /*ih*/ 5, /*iw*/ 4, /*fh*/ 3, /*fw*/ 3, /*ph*/ 1,
/*pw*/ 1);
#ifdef PADDLE_WITH_CUDA
testIm2col<paddle::platform::CUDADeviceContext,
paddle::platform::CUDAPlace>();
#endif
}
#define PREPARE_IM2COL_CPU \
paddle::platform::CPUPlace place; \
paddle::platform::CPUDeviceContext context(place); \
paddle::framework::Tensor input; \
paddle::framework::Tensor out; \
paddle::framework::Tensor ref; \
std::vector<int> padding({ph, pw}); \
std::vector<int> stride({1, 1}); \
std::vector<int> dilation({1, 1}); \
float* input_ptr = input.mutable_data<float>({ic, ih, iw}, place); \
for (int i = 0; i < input.numel(); ++i) { \
input_ptr[i] = static_cast<float>(i + 1); \
} \
int output_height = (ih - fh + padding[0] * 2) / stride[0] + 1; \
int output_width = (iw - fw + padding[1] * 2) / stride[1] + 1; \
out.mutable_data<float>({ic, fh, fw, output_height, output_width}, place); \
ref.mutable_data<float>({ic, fh, fw, output_height, output_width}, place); \
paddle::operators::math::Im2ColFunctor< \
paddle::operators::math::ColFormat::kCFO, \
paddle::platform::CPUDeviceContext, float> \
im2col
void testIm2colCPU(int ic, int ih, int iw, int fh, int fw, int ph, int pw) {
PREPARE_IM2COL_CPU;
im2col(context, input, dilation, stride, padding, &out);
paddle::operators::math::im2col_common<float>(input, dilation, stride,
padding, &ref);
float* ref_data = ref.data<float>();
float* out_data = out.data<float>();
for (int i = 0; i < out.numel(); ++i) {
EXPECT_EQ(out_data[i], ref_data[i]);
}
}
void benchIm2col(int ic, int ih, int iw, int fh, int fw, int ph, int pw) {
PREPARE_IM2COL_CPU;
constexpr int repeat = 100;
auto GetCurrentMs = []() -> double {
struct timeval time;
gettimeofday(&time, NULL);
return 1e+3 * time.tv_sec + 1e-3 * time.tv_usec;
};
auto t1 = GetCurrentMs();
for (int i = 0; i < repeat; ++i) {
im2col(context, input, dilation, stride, padding, &out);
}
auto t2 = GetCurrentMs();
for (int i = 0; i < repeat; ++i) {
paddle::operators::math::im2col_common<float>(input, dilation, stride,
padding, &ref);
}
auto t3 = GetCurrentMs();
LOG(INFO) << "before: " << (t3 - t2) / repeat
<< ",after: " << (t2 - t1) / repeat
<< ",boost: " << ((t3 - t2) / (t2 - t1) - 1) * 100 << "%";
}
TEST(math, im2col_cputest) {
// padding_h == padding_w
for (int p = 0; p < 4; ++p) {
// width == height
testIm2colCPU(/*ic*/ 2, /*ih*/ 5, /*iw*/ 5, /*fh*/ 4, /*fw*/ 4, /*ph*/ p,
/*pw*/ p);
testIm2colCPU(/*ic*/ 2, /*ih*/ 4, /*iw*/ 4, /*fh*/ 3, /*fw*/ 3, /*ph*/ p,
/*pw*/ p);
testIm2colCPU(/*ic*/ 2, /*ih*/ 4, /*iw*/ 4, /*fh*/ 2, /*fw*/ 2, /*ph*/ p,
/*pw*/ p);
// height != width
testIm2colCPU(/*ic*/ 2, /*ih*/ 5, /*iw*/ 4, /*fh*/ 2, /*fw*/ 3, /*ph*/ p,
/*pw*/ p);
testIm2colCPU(/*ic*/ 2, /*ih*/ 5, /*iw*/ 4, /*fh*/ 1, /*fw*/ 3, /*ph*/ p,
/*pw*/ p);
testIm2colCPU(/*ic*/ 2, /*ih*/ 4, /*iw*/ 5, /*fh*/ 3, /*fw*/ 1, /*ph*/ p,
/*pw*/ p);
// filter == 1
testIm2colCPU(/*ic*/ 3, /*ih*/ 4, /*iw*/ 4, /*fh*/ 1, /*fw*/ 1, /*ph*/ p,
/*pw*/ p);
testIm2colCPU(/*ic*/ 3, /*ih*/ 3, /*iw*/ 4, /*fh*/ 1, /*fw*/ 1, /*ph*/ p,
/*pw*/ p);
}
// padding_h != padding_w
testIm2colCPU(/*ic*/ 2, /*ih*/ 4, /*iw*/ 4, /*fh*/ 2, /*fw*/ 3, /*ph*/ 1,
/*pw*/ 2);
// benchmark
for (int p : {0, 1}) {
for (int k : {1, 3, 5}) {
LOG(INFO) << "padding == " << p << ", filter == " << k;
benchIm2col(/*ic*/ 3, /*ih*/ 224, /*iw*/ 224, /*fh*/ k, /*fw*/ k,
/*ph*/ p, /*pw*/ p);
}
}
}
paddle/fluid/operators/reshape_op.cc
浏览文件 @ a5c96af3
......@@ -127,12 +127,6 @@ class ReshapeOpMaker : public framework::OpProtoAndCheckerMaker {
AddOutput("Out", "(Tensor). The output tensor of reshape operator.");
AddAttr<std::vector<int>>(
"shape", "(std::vector<int>) Target shape of reshape operator.");
AddAttr<bool>("inplace",
"(default: false) Change the source tensor's shape without "
"memory copy. When Attr(inplace) is set true, the output "
"tensor shares memory with Input(X), otherwise, a new output "
"tensor is created, and its data are copied from Input(x).")
.SetDefault(false);
AddComment(R"DOC(
Reshape Operator.
......@@ -233,16 +227,9 @@ class ReshapeKernel {
"sequence_reshape op.");
}
bool inplace = ctx.Attr<bool>("inplace");
out->mutable_data(ctx.GetPlace(), in->type());
framework::TensorCopySync(*in, ctx.GetPlace(), out);
out->Resize(out_dims);
if (!inplace) {
out->mutable_data(ctx.GetPlace(), in->type());
framework::TensorCopySync(*in, ctx.GetPlace(), out);
out->Resize(out_dims);
} else {
out->ShareDataWith(*in);
out->Resize(out_dims);
}
}
};
......@@ -251,19 +238,11 @@ class ReshapeGradKernel {
void operator()(const framework::ExecutionContext &ctx) const {
auto *d_out = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
auto *d_x = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
auto in_dims = d_x->dims();
d_x->mutable_data(ctx.GetPlace(), d_out->type());
bool inplace = ctx.Attr<bool>("inplace");
auto in_dims = d_x->dims();
if (!inplace) {
framework::TensorCopy(*d_out, ctx.GetPlace(), ctx.device_context(), d_x);
ctx.device_context().Wait();
d_x->Resize(in_dims);
} else {
d_x->ShareDataWith(*d_out);
d_x->Resize(in_dims);
}
framework::TensorCopySync(*d_out, ctx.GetPlace(), d_x);
d_x->Resize(in_dims);
}
};
......
paddle/fluid/operators/split_ids_op.h
浏览文件 @ a5c96af3
......@@ -14,6 +14,7 @@ limitations under the License. */
#pragma once
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
......@@ -67,10 +68,15 @@ class SplitIdsOpKernel : public framework::OpKernel<T> {
const auto &ids_rows = ids_selected_rows->rows();
auto outs = ctx.MultiOutput<framework::SelectedRows>("Out");
const size_t shard_num = outs.size();
for (auto &out : outs) {
out->mutable_rows()->clear();
}
// get rows for outputs
for (auto &id : ids_rows) {
size_t shard_id = static_cast<size_t>(id) % shard_num;
outs[shard_id]->mutable_rows()->push_back(id);
std::unordered_map<int64_t, size_t> id_to_index;
for (size_t i = 0; i < ids_rows.size(); ++i) {
id_to_index[ids_rows[i]] = i;
size_t shard_id = static_cast<size_t>(ids_rows[i]) % shard_num;
outs[shard_id]->mutable_rows()->push_back(ids_rows[i]);
}
int64_t row_width = ids_dims[1];
......@@ -80,7 +86,8 @@ class SplitIdsOpKernel : public framework::OpKernel<T> {
{static_cast<int64_t>(out->rows().size()), row_width});
T *output = out->mutable_value()->mutable_data<T>(ddim, place);
for (int64_t i = 0; i < ddim[0]; ++i) {
memcpy(output + i * row_width, ids + out->rows()[i] * row_width,
memcpy(output + i * row_width,
ids + id_to_index[out->rows()[i]] * row_width,
row_width * sizeof(T));
}
}
......
paddle/fluid/platform/cuda_helper_test.cu
浏览文件 @ a5c96af3
......@@ -13,7 +13,6 @@
// limitations under the License.
#include <gtest/gtest.h>
#include <bitset>
#include <iostream>
#include <random>
......@@ -25,13 +24,13 @@
using paddle::platform::PADDLE_CUDA_NUM_THREADS;
using paddle::platform::float16;
#define CUDA_ATOMIC_KERNEL(op, T) \
__global__ void op##Kernel(const T* data_a, T* data_b, size_t num) { \
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num; \
i += blockDim.x * gridDim.x) { \
paddle::platform::CudaAtomic##op(&data_b[i], data_a[i]); \
} \
template <typename T>
__global__ void AddKernel(const T* data_a, T* data_b, size_t num) {
for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < num;
i += blockDim.x * gridDim.x) {
paddle::platform::CudaAtomicAdd(&data_b[i], data_a[i]);
}
}
template <typename T>
struct AddFunctor {
......@@ -39,80 +38,116 @@ struct AddFunctor {
};
template <typename T>
struct SubFunctor {
T operator()(const T& a, const T& b) { return a - b; }
};
// NOTE(dzhwinter): the float16 add has small underflow/overflow
// so we use EXPECT_NEAR to check the result.
#define ARITHMETIC_KERNEL_LAUNCH(op, T) \
void Test##T##op(size_t num) { \
T *in1, *in2, *out; \
T *d_in1, *d_in2; \
size_t size = sizeof(T) * num; \
cudaMalloc(reinterpret_cast<void**>(&d_in1), size); \
cudaMalloc(reinterpret_cast<void**>(&d_in2), size); \
in1 = reinterpret_cast<T*>(malloc(size)); \
in2 = reinterpret_cast<T*>(malloc(size)); \
out = reinterpret_cast<T*>(malloc(size)); \
std::minstd_rand engine; \
std::uniform_real_distribution<double> dist(0.0, 1.0); \
for (size_t i = 0; i < num; ++i) { \
in1[i] = static_cast<T>(dist(engine)); \
in2[i] = static_cast<T>(dist(engine)); \
} \
cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice); \
cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice); \
op##Kernel<<<1, PADDLE_CUDA_NUM_THREADS>>>(d_in1, d_in2, num); \
cudaDeviceSynchronize(); \
cudaMemcpy(out, d_in2, size, cudaMemcpyDeviceToHost); \
cudaDeviceSynchronize(); \
for (size_t i = 0; i < num; ++i) { \
EXPECT_NEAR(static_cast<float>(out[i]), \
static_cast<float>(op##Functor<T>()(in1[i], in2[i])), \
0.001); \
} \
free(in1); \
free(in2); \
free(out); \
cudaFree(d_in1); \
cudaFree(d_in2); \
void TestCase(size_t num) {
T *in1, *in2, *out;
T *d_in1, *d_in2;
size_t size = sizeof(T) * num;
cudaMalloc(reinterpret_cast<void**>(&d_in1), size);
cudaMalloc(reinterpret_cast<void**>(&d_in2), size);
in1 = reinterpret_cast<T*>(malloc(size));
in2 = reinterpret_cast<T*>(malloc(size));
out = reinterpret_cast<T*>(malloc(size));
std::minstd_rand engine;
std::uniform_real_distribution<double> dist(0.0, 1.0);
for (size_t i = 0; i < num; ++i) {
in1[i] = static_cast<T>(dist(engine));
in2[i] = static_cast<T>(dist(engine));
}
CUDA_ATOMIC_KERNEL(Add, float);
CUDA_ATOMIC_KERNEL(Add, double);
CUDA_ATOMIC_KERNEL(Add, float16);
ARITHMETIC_KERNEL_LAUNCH(Add, float);
ARITHMETIC_KERNEL_LAUNCH(Add, double);
ARITHMETIC_KERNEL_LAUNCH(Add, float16);
namespace paddle {
namespace platform {
USE_CUDA_ATOMIC(Sub, int);
};
};
CUDA_ATOMIC_KERNEL(Sub, int);
ARITHMETIC_KERNEL_LAUNCH(Sub, int);
cudaMemcpy(d_in1, in1, size, cudaMemcpyHostToDevice);
cudaMemcpy(d_in2, in2, size, cudaMemcpyHostToDevice);
AddKernel<T><<<1, PADDLE_CUDA_NUM_THREADS>>>(d_in1, d_in2, num);
cudaDeviceSynchronize();
cudaMemcpy(out, d_in2, size, cudaMemcpyDeviceToHost);
cudaDeviceSynchronize();
for (size_t i = 0; i < num; ++i) {
// NOTE(dzhwinter): the float16 add has small underflow/overflow
// so we use EXPECT_NEAR to check the result.
EXPECT_NEAR(static_cast<float>(out[i]),
static_cast<float>(AddFunctor<T>()(in1[i], in2[i])), 0.001);
}
free(in1);
free(in2);
free(out);
cudaFree(d_in1);
cudaFree(d_in2);
}
// cuda primitives
TEST(CudaAtomic, Add) {
TestfloatAdd(static_cast<size_t>(10));
TestfloatAdd(static_cast<size_t>(1024 * 1024));
TestdoubleAdd(static_cast<size_t>(10));
TestdoubleAdd(static_cast<size_t>(1024 * 1024));
}
TestCase<float>(static_cast<size_t>(10));
TestCase<float>(static_cast<size_t>(1024 * 1024));
TEST(CudaAtomic, Sub) {
TestintSub(static_cast<size_t>(10));
TestintSub(static_cast<size_t>(1024 * 1024));
TestCase<double>(static_cast<size_t>(10));
TestCase<double>(static_cast<size_t>(1024 * 1024));
}
TEST(CudaAtomic, float16) {
using paddle::platform::float16;
Testfloat16Add(static_cast<size_t>(1));
Testfloat16Add(static_cast<size_t>(2));
Testfloat16Add(static_cast<size_t>(3));
TestCase<float16>(static_cast<size_t>(1));
TestCase<float16>(static_cast<size_t>(2));
TestCase<float16>(static_cast<size_t>(3));
TestCase<float16>(static_cast<size_t>(10));
TestCase<float16>(static_cast<size_t>(1024 * 1024));
}
// unalignment of uint8
void TestUnalign(size_t num, const int shift_bit) {
PADDLE_ENFORCE(num % 2 == 0, "must be a multiple of 2");
float16 *in1, *in2, *out;
float16 *d_in1, *d_in2;
size_t size = sizeof(uint8_t) * (num + shift_bit);
size_t array_size = sizeof(float16) * (num / 2);
cudaMalloc(reinterpret_cast<void**>(&d_in1), size);
cudaMalloc(reinterpret_cast<void**>(&d_in2), size);
in1 = reinterpret_cast<float16*>(malloc(size));
in2 = reinterpret_cast<float16*>(malloc(size));
out = reinterpret_cast<float16*>(malloc(size));
// right shift 1, mimic the unalignment of address
float16* r_in1 =
reinterpret_cast<float16*>(reinterpret_cast<uint8_t*>(in1) + shift_bit);
float16* r_in2 =
reinterpret_cast<float16*>(reinterpret_cast<uint8_t*>(in2) + shift_bit);
std::minstd_rand engine;
std::uniform_real_distribution<double> dist(0.0, 1.0);
for (size_t i = 0; i < num / 2; ++i) {
r_in1[i] = static_cast<float16>(dist(engine));
r_in2[i] = static_cast<float16>(dist(engine));
}
cudaMemcpy(d_in1, r_in1, array_size, cudaMemcpyHostToDevice);
cudaMemcpy(d_in2, r_in2, array_size, cudaMemcpyHostToDevice);
AddKernel<float16><<<1, PADDLE_CUDA_NUM_THREADS>>>(d_in1, d_in2, num / 2);
cudaDeviceSynchronize();
cudaMemcpy(out, d_in2, array_size, cudaMemcpyDeviceToHost);
cudaDeviceSynchronize();
for (size_t i = 0; i < num / 2; ++i) {
// NOTE(dzhwinter): the float16 add has small underflow/overflow
// so we use EXPECT_NEAR to check the result.
EXPECT_NEAR(static_cast<float>(out[i]),
static_cast<float>(AddFunctor<float16>()(r_in1[i], r_in2[i])),
0.001);
}
free(in1);
free(in2);
free(out);
cudaFree(d_in1);
cudaFree(d_in2);
}
TEST(CudaAtomic, float16Unalign) {
// same with float16 testcase
TestUnalign(static_cast<size_t>(2), /*shift_bit*/ 2);
TestUnalign(static_cast<size_t>(1024), /*shift_bit*/ 2);
TestUnalign(static_cast<size_t>(1024 * 1024), /*shift_bit*/ 2);
// shift the address.
TestUnalign(static_cast<size_t>(2), /*shift_bit*/ 1);
TestUnalign(static_cast<size_t>(1024), /*shift_bit*/ 1);
TestUnalign(static_cast<size_t>(1024 * 1024), /*shift_bit*/ 1);
Testfloat16Add(static_cast<size_t>(10));
Testfloat16Add(static_cast<size_t>(1024 * 1024));
TestUnalign(static_cast<size_t>(2), /*shift_bit*/ 3);
TestUnalign(static_cast<size_t>(1024), /*shift_bit*/ 3);
TestUnalign(static_cast<size_t>(1024 * 1024), /*shift_bit*/ 3);
}
paddle/fluid/platform/cuda_primitives.h
浏览文件 @ a5c96af3
......@@ -79,41 +79,41 @@ CUDA_ATOMIC_WRAPPER(Add, double) {
// convert the value into float and do the add arithmetic.
// then store the result into a uint32.
inline __device__ uint32_t add_to_low_half(uint32_t val, float x) {
inline static __device__ uint32_t add_to_low_half(uint32_t val, float x) {
float16 low_half;
// the float16 in lower 16bits
low_half.x = static_cast<uint16_t>(val & 0xffffu);
low_half.x = static_cast<uint16_t>(val & 0xFFFFu);
low_half = static_cast<float16>(static_cast<float>(low_half) + x);
return (val & 0xffff0000u) | low_half.x;
return (val & 0xFFFF0000u) | low_half.x;
}
inline __device__ uint32_t add_to_high_half(uint32_t val, float x) {
inline static __device__ uint32_t add_to_high_half(uint32_t val, float x) {
float16 high_half;
// the float16 in higher 16bits
high_half.x = static_cast<uint16_t>(val >> 16);
high_half = static_cast<float16>(static_cast<float>(high_half) + x);
return (val & 0xffffu) | (static_cast<uint32_t>(high_half.x) << 16);
return (val & 0xFFFFu) | (static_cast<uint32_t>(high_half.x) << 16);
}
CUDA_ATOMIC_WRAPPER(Add, float16) {
// concrete packed float16 value may exsits in lower or higher 16bits
// of the 32bits address.
uint32_t *address_as_ui =
reinterpret_cast<uint32_t *>(reinterpret_cast<char *>(address) -
(reinterpret_cast<size_t>(address) & 2));
uint32_t *address_as_ui = reinterpret_cast<uint32_t *>(
reinterpret_cast<char *>(address) -
(reinterpret_cast<uintptr_t>(address) & 0x02));
float val_f = static_cast<float>(val);
uint32_t old = *address_as_ui;
uint32_t sum;
uint32_t newval;
uint32_t assumed;
if (((size_t)address & 2) == 0) {
if (((uintptr_t)address & 0x02) == 0) {
// the float16 value stay at lower 16 bits of the address.
do {
assumed = old;
old = atomicCAS(address_as_ui, assumed, add_to_low_half(assumed, val_f));
} while (old != assumed);
float16 ret;
ret.x = old & 0xffffu;
ret.x = old & 0xFFFFu;
return ret;
} else {
// the float16 value stay at higher 16 bits of the address.
......
paddle/scripts/paddle_build.sh
浏览文件 @ a5c96af3
......@@ -534,7 +534,7 @@ EOF
make -j `nproc` inference_lib_dist
cd ${PADDLE_ROOT}/build
cp -r fluid_install_dir fluid
tar -cf fluid.tgz fluid
tar -czf fluid.tgz fluid
fi
}
......
python/paddle/fluid/__init__.py
浏览文件 @ a5c96af3
......@@ -127,6 +127,7 @@ def __bootstrap__():
]
if core.is_compiled_with_dist():
read_env_flags.append('rpc_deadline')
read_env_flags.append('listen_and_serv_profile_period')
if core.is_compiled_with_cuda():
read_env_flags += [
......
python/paddle/fluid/layers/control_flow.py
浏览文件 @ a5c96af3
......@@ -21,6 +21,7 @@ from ..layer_helper import LayerHelper, unique_name
from ..initializer import force_init_on_cpu
from ops import logical_and, logical_not, logical_or
import numpy
import warnings
__all__ = [
'While',
......@@ -280,6 +281,9 @@ class ParallelDo(object):
"""
def __init__(self, places, use_nccl=False, name=None):
warnings.warn(
"API ParallelDo is deprecated since 0.15.0. Please use ParallelExecutor instead.",
Warning)
self.helper = LayerHelper("parallel_do", name=name)
self.inputs = []
self.places = places
......@@ -338,7 +342,7 @@ class ParallelDo(object):
return [parent_block.var(name) for name in params]
def complete_op(self):
def _complete_op(self):
main_program = self.helper.main_program
current_block = main_program.current_block()
parent_block = self.parent_block()
......@@ -394,7 +398,7 @@ class BlockGuardWithCompletion(BlockGuard):
if exc_type is not None:
return False
self.rnn.status = StaticRNN.AFTER_RNN_BLOCK
self.rnn.complete_op()
self.rnn._complete_op()
return super(BlockGuardWithCompletion, self).__exit__(exc_type, exc_val,
exc_tb)
......@@ -470,7 +474,7 @@ class StaticRNN(object):
if shape is None or batch_ref is None:
raise ValueError(
"if init is None, memory at least need shape and batch_ref")
parent_block = self.parent_block()
parent_block = self._parent_block()
var_name = unique_name.generate("@".join(
[self.helper.name, "memory_boot"]))
boot_var = parent_block.create_var(
......@@ -527,7 +531,7 @@ class StaticRNN(object):
outputs={'Out': tmp_o},
attrs={'dtype': o.dtype})
out_var = self.parent_block().create_var(
out_var = self._parent_block().create_var(
name=tmp_o.name,
shape=[self.seq_len] + list(tmp_o.shape),
dtype=tmp_o.dtype)
......@@ -543,7 +547,7 @@ class StaticRNN(object):
raise TypeError("update memory should take variables")
self.memories[mem.name].mem = var
def parent_block(self):
def _parent_block(self):
prog = self.helper.main_program
parent_idx = prog.current_block().parent_idx
assert parent_idx >= 0
......@@ -560,10 +564,10 @@ class StaticRNN(object):
else:
return self.outputs
def complete_op(self):
def _complete_op(self):
main_program = self.helper.main_program
rnn_block = main_program.current_block()
parent_block = self.parent_block()
parent_block = self._parent_block()
local_inputs = set()
......@@ -643,7 +647,7 @@ class WhileGuard(BlockGuard):
if exc_type is not None:
return False
self.while_op.status = While.AFTER_WHILE_BLOCK
self.while_op.complete()
self.while_op._complete()
return super(WhileGuard, self).__exit__(exc_type, exc_val, exc_tb)
......@@ -690,7 +694,7 @@ class While(object):
def block(self):
return WhileGuard(self)
def complete(self):
def _complete(self):
main_program = self.helper.main_program
while_block = main_program.current_block()
parent_block = main_program.block(main_program.current_block()
......
python/paddle/fluid/layers/nn.py
浏览文件 @ a5c96af3
......@@ -4473,15 +4473,14 @@ def reshape(x, shape, actual_shape=None, act=None, inplace=True, name=None):
"except one unknown dimension.")
helper = LayerHelper("reshape", **locals())
reshaped = helper.create_tmp_variable(dtype=x.dtype)
out = helper.create_tmp_variable(dtype=x.dtype)
helper.append_op(
type="reshape",
inputs=inputs,
attrs={"shape": shape,
"inplace": inplace},
outputs={"Out": reshaped})
attrs={"shape": shape},
outputs={"Out": out})
return helper.append_activation(reshaped)
return helper.append_activation(out)
def lod_reset(x, y=None, target_lod=None):
......
python/paddle/fluid/tests/unittests/test_memory_optimization_transpiler.py
浏览文件 @ a5c96af3
......@@ -43,5 +43,29 @@ class TestControlFlowGraph(unittest.TestCase):
print(str(result_program))
class TestMemoryTranspiler2(unittest.TestCase):
def setUp(self):
program = Program()
with program_guard(program, startup_program=Program()):
x = layers.data(name='x', shape=[13], dtype='float32')
fc = layers.fc(input=x, size=10, act=None)
reshape = layers.reshape(x=fc, shape=[-1, 2, 5])
fc = layers.reshape(x=reshape, shape=[-1, 5, 2])
y_predict = layers.fc(input=fc, size=1, act=None)
y = layers.data(name='y', shape=[1], dtype='float32')
cost = layers.square_error_cost(input=y_predict, label=y)
avg_cost = layers.mean(cost)
opt = optimizer.SGD(learning_rate=0.001)
opt.minimize(avg_cost)
self.program = program
def test_inplace_ops(self):
print("before optimization")
print(str(self.program))
result_program = memory_optimize(self.program)
print("after optimization")
print(str(result_program))
if __name__ == "__main__":
unittest.main()
python/paddle/fluid/tests/unittests/test_reshape_op.py
浏览文件 @ a5c96af3
......@@ -25,7 +25,7 @@ class TestReshapeOp(OpTest):
self.op_type = "reshape"
self.inputs = {"X": np.random.random(ori_shape).astype("float32")}
self.attrs = {"shape": new_shape, "inplace": False}
self.attrs = {"shape": new_shape}
self.outputs = {"Out": self.inputs["X"].reshape(new_shape)}
def test_check_output(self):
......@@ -42,7 +42,7 @@ class TestReshapeOpDimInfer1(OpTest):
self.op_type = "reshape"
self.inputs = {"X": np.random.random(ori_shape).astype("float32")}
self.attrs = {"shape": new_shape, "inplace": False}
self.attrs = {"shape": new_shape}
self.outputs = {"Out": self.inputs["X"].reshape(self.attrs["shape"])}
def test_check_output(self):
......@@ -60,7 +60,7 @@ class TestReshapeOpDimInfer2(OpTest):
self.op_type = "reshape"
self.inputs = {"X": np.random.random(ori_shape).astype("float32")}
self.attrs = {"shape": new_shape, "inplace": False}
self.attrs = {"shape": new_shape}
self.outputs = {"Out": self.inputs["X"].reshape(infered_shape)}
def test_check_output(self):
......
python/paddle/fluid/tests/unittests/test_split_ids_op.py
浏览文件 @ a5c96af3
......@@ -15,6 +15,8 @@
import unittest
import numpy as np
from op_test import OpTest
import paddle.fluid.core as core
from paddle.fluid.op import Operator
class TestSplitIdsOp(OpTest):
......@@ -31,5 +33,55 @@ class TestSplitIdsOp(OpTest):
self.check_output()
class TestSpliteIds(unittest.TestCase):
def get_places(self):
places = [core.CPUPlace()]
return places
def test_check_output(self):
for place in self.get_places():
self.check_with_place(place)
def check_with_place(self, place):
scope = core.Scope()
rows = [0, 5, 7, 4, 9]
height = 20
row_numel = 2
# initialize input variable X
x = scope.var('X').get_selected_rows()
x.set_rows(rows)
x.set_height(height)
np_array = np.ones((len(rows), row_numel)).astype("float32")
for i in range(len(rows)):
for j in range(row_numel):
np_array[i, j] = rows[i] + j
x_tensor = x.get_tensor()
x_tensor.set(np_array, place)
outs_name = ["out%d" % i for i in xrange(3)]
outs = [
scope.var(var_name).get_selected_rows() for var_name in outs_name
]
# expected output selected rows
expected_out_rows = [[0, 9], [7, 4], [5]]
op = Operator("split_ids", Ids="X", Out=outs_name)
for _ in range(3):
op.run(scope, place)
for i in range(len(outs)):
expected_rows = expected_out_rows[i]
self.assertEqual(outs[i].rows(), expected_rows)
for j in range(len(expected_rows)):
row = expected_rows[j]
self.assertAlmostEqual(
float(row), np.array(outs[i].get_tensor())[j, 0])
self.assertAlmostEqual(
float(row + 1), np.array(outs[i].get_tensor())[j, 1])
if __name__ == '__main__':
unittest.main()
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @ a5c96af3
......@@ -495,6 +495,7 @@ class DistributeTranspiler(object):
pserver_index = self.pserver_endpoints.index(endpoint)
table_opt_block = self._create_table_optimize_block(
pserver_index, pserver_program, pre_block_idx, grad_to_block_id)
optimize_blocks.append(table_opt_block)
prefetch_var_name_to_block_id = self._create_prefetch_block(
pserver_index, pserver_program, table_opt_block)
checkpoint_block_id = self._create_checkpoint_save_block(
......
tools/manylinux1/Dockerfile.x64
浏览文件 @ a5c96af3
......@@ -13,7 +13,7 @@ ENV PATH /opt/rh/devtoolset-2/root/usr/bin:$PATH
ENV LD_LIBRARY_PATH /opt/rh/devtoolset-2/root/usr/lib64:/opt/rh/devtoolset-2/root/usr/lib:/usr/local/lib64:/usr/local/lib:${LD_LIBRARY_PATH}
ENV PKG_CONFIG_PATH=/usr/local/lib/pkgconfig
RUN yum install -y sqlite-devel zlib-devel openssl-devel pcre-devel vim tk-devel tkinter libtool xz
RUN yum install -y sqlite-devel zlib-devel openssl-devel pcre-devel vim tk-devel tkinter libtool xz graphviz
COPY build_scripts /build_scripts
RUN bash build_scripts/build.sh && \
bash build_scripts/install_nccl2.sh && rm -r build_scripts
......
Markdown is supported
0% 或 .
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 或