Tao Luo
6 years ago
commit
923b18877e
@ -0,0 +1,54 @@
|
||||
/* 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/op_kernel_type.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace framework {
|
||||
|
||||
size_t OpKernelType::Hash::operator()(const OpKernelType& key) const {
|
||||
int cur_loc = 0;
|
||||
|
||||
int place = key.place_.which();
|
||||
cur_loc += OpKernelType::kPlaceBits;
|
||||
|
||||
int data_type = static_cast<int>(key.data_type_) << cur_loc;
|
||||
cur_loc += OpKernelType::kPrimaryDTypeBits;
|
||||
|
||||
int data_layout = static_cast<int>(key.data_layout_) << cur_loc;
|
||||
cur_loc += OpKernelType::kLayoutBits;
|
||||
|
||||
int library_type = static_cast<int>(key.library_type_) << cur_loc;
|
||||
cur_loc += OpKernelType::kLibBits;
|
||||
|
||||
int customized_value = key.customized_type_value_;
|
||||
PADDLE_ENFORCE(customized_value < (1 << OpKernelType::kCustomizeBits));
|
||||
customized_value = customized_value << cur_loc;
|
||||
cur_loc += OpKernelType::kCustomizeBits;
|
||||
PADDLE_ENFORCE(cur_loc < 64);
|
||||
|
||||
std::hash<int> hasher;
|
||||
return hasher(place + data_type + data_layout + library_type +
|
||||
customized_value);
|
||||
}
|
||||
|
||||
bool OpKernelType::operator==(const OpKernelType& o) const {
|
||||
return platform::places_are_same_class(place_, o.place_) &&
|
||||
data_type_ == o.data_type_ && data_layout_ == o.data_layout_ &&
|
||||
library_type_ == o.library_type_ &&
|
||||
customized_type_value_ == o.customized_type_value_;
|
||||
}
|
||||
|
||||
} // namespace framework
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,3 @@
|
||||
cc_library(layer SRCS layer.cc DEPS proto_desc operator)
|
||||
cc_library(tracer SRCS tracer.cc DEPS proto_desc)
|
||||
cc_library(engine SRCS engine.cc)
|
||||
@ -0,0 +1,53 @@
|
||||
// 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/imperative/engine.h"
|
||||
|
||||
#include <mutex> // NOLINT
|
||||
#include <vector>
|
||||
|
||||
#include "glog/logging.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace imperative {
|
||||
|
||||
static std::once_flag init_engine;
|
||||
static Engine* engine;
|
||||
|
||||
class DummyEngine : public Engine {
|
||||
public:
|
||||
void Enqueue(Runnable* runnable) override {
|
||||
queued_runnables_.push_back(runnable);
|
||||
}
|
||||
|
||||
size_t Size() const override { return queued_runnables_.size(); }
|
||||
|
||||
void Sync() override {
|
||||
for (Runnable* l : queued_runnables_) {
|
||||
LOG(INFO) << "running " << reinterpret_cast<void*>(l);
|
||||
}
|
||||
queued_runnables_.clear();
|
||||
}
|
||||
|
||||
private:
|
||||
std::vector<Runnable*> queued_runnables_;
|
||||
};
|
||||
|
||||
Engine* GetEngine() {
|
||||
std::call_once(init_engine, []() { engine = new DummyEngine(); });
|
||||
return engine;
|
||||
}
|
||||
|
||||
} // namespace imperative
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,39 @@
|
||||
// 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 <cstddef>
|
||||
#include <cstdint>
|
||||
|
||||
namespace paddle {
|
||||
namespace imperative {
|
||||
|
||||
struct Runnable {};
|
||||
|
||||
class Engine {
|
||||
public:
|
||||
virtual ~Engine() {}
|
||||
|
||||
virtual void Enqueue(Runnable* runnable) = 0;
|
||||
|
||||
virtual size_t Size() const = 0;
|
||||
|
||||
virtual void Sync() = 0;
|
||||
};
|
||||
|
||||
Engine* GetEngine();
|
||||
|
||||
} // namespace imperative
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,221 @@
|
||||
// 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/imperative/layer.h"
|
||||
#include <deque>
|
||||
#include <limits>
|
||||
#include <map>
|
||||
#include <random>
|
||||
#include <utility>
|
||||
|
||||
#include "paddle/fluid/framework/lod_tensor.h"
|
||||
#include "paddle/fluid/framework/op_registry.h"
|
||||
#include "paddle/fluid/string/printf.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace imperative {
|
||||
|
||||
using framework::Variable;
|
||||
|
||||
void AddTo(Variable* src, Variable* dst) {
|
||||
framework::LoDTensor* dst_tensor = dst->GetMutable<framework::LoDTensor>();
|
||||
framework::LoDTensor* src_tensor = src->GetMutable<framework::LoDTensor>();
|
||||
PADDLE_ENFORCE(dst_tensor->numel() == src_tensor->numel(), "%lld vs %lld",
|
||||
dst_tensor->numel(), src_tensor->numel());
|
||||
float* dst_data = dst_tensor->mutable_data<float>(platform::CPUPlace());
|
||||
const float* src_data = src_tensor->data<float>();
|
||||
for (size_t i = 0; i < src_tensor->numel(); ++i) {
|
||||
dst_data[i] += src_data[i];
|
||||
}
|
||||
}
|
||||
|
||||
class Autograd {
|
||||
public:
|
||||
explicit Autograd(framework::Scope* scope) : scope_(scope) {}
|
||||
|
||||
void RunBackward(VarBase* var) {
|
||||
PADDLE_ENFORCE(var->pre_op_->op_desc_);
|
||||
// TODO(panyx0718): Only create for vars that "require_grad"
|
||||
(*var->pre_op_->output_vars_)[var->pre_op_out_idx_]->grads_ = var->grads_;
|
||||
|
||||
std::deque<OpBase*> ready;
|
||||
ready.push_back(var->pre_op_);
|
||||
|
||||
std::map<OpBase*, int> dep_counts = ComputeDepCounts(var->pre_op_);
|
||||
|
||||
while (!ready.empty()) {
|
||||
OpBase* ready_op = ready.front();
|
||||
ready.pop_front();
|
||||
std::vector<Variable*> input_grads = ready_op->ApplyGrad(scope_);
|
||||
|
||||
for (size_t i = 0; i < input_grads.size(); ++i) {
|
||||
if (!input_grads[i]) continue;
|
||||
OpBase* pre_op = ready_op->pre_ops_->at(i);
|
||||
if (!pre_op) continue;
|
||||
|
||||
dep_counts[pre_op] -= 1;
|
||||
PADDLE_ENFORCE(dep_counts[pre_op] >= 0);
|
||||
bool pre_op_ready = dep_counts[pre_op] == 0;
|
||||
if (pre_op_ready) {
|
||||
ready.push_back(pre_op);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
private:
|
||||
std::map<OpBase*, int> ComputeDepCounts(OpBase* op) {
|
||||
std::map<OpBase*, int> ret;
|
||||
|
||||
std::deque<OpBase*> queue;
|
||||
queue.push_back(op);
|
||||
std::unordered_set<OpBase*> visited;
|
||||
visited.insert(op);
|
||||
while (!queue.empty()) {
|
||||
OpBase* candidate = queue.front();
|
||||
queue.pop_front();
|
||||
for (OpBase* pre_op : *(candidate->pre_ops_)) {
|
||||
if (!pre_op) continue;
|
||||
if (visited.find(pre_op) == visited.end()) {
|
||||
visited.insert(pre_op);
|
||||
queue.push_back(pre_op);
|
||||
}
|
||||
ret[pre_op] += 1;
|
||||
}
|
||||
}
|
||||
|
||||
return ret;
|
||||
}
|
||||
|
||||
framework::Scope* scope_;
|
||||
};
|
||||
|
||||
framework::Variable* CreateVariable(const std::string& name,
|
||||
const framework::DDim& dim, float val,
|
||||
framework::Scope* scope,
|
||||
bool random_name = true) {
|
||||
std::string varname = name;
|
||||
if (random_name) {
|
||||
std::mt19937 rng;
|
||||
rng.seed(std::random_device()());
|
||||
std::uniform_int_distribution<std::mt19937::result_type> dist6(
|
||||
1, std::numeric_limits<int>::max());
|
||||
int id = dist6(rng);
|
||||
varname = string::Sprintf("%s@%d", varname, id);
|
||||
}
|
||||
|
||||
VLOG(3) << "creating var " << varname;
|
||||
framework::Variable* var = scope->Var(varname);
|
||||
framework::LoDTensor* tensor = var->GetMutable<framework::LoDTensor>();
|
||||
|
||||
float* data = tensor->mutable_data<float>(dim, platform::CPUPlace());
|
||||
std::fill(data, data + tensor->numel(), val);
|
||||
return var;
|
||||
}
|
||||
|
||||
framework::LoDTensor& VarBase::Grad() {
|
||||
VLOG(3) << "get var grad " << var_desc_->Name();
|
||||
return *grads_->GetMutable<framework::LoDTensor>();
|
||||
}
|
||||
|
||||
void VarBase::ApplyGrad(framework::Scope* scope, Variable* grad) {
|
||||
VLOG(3) << "apply var grad " << var_desc_->Name() << " "
|
||||
<< grad->Get<framework::LoDTensor>().data<float>()[0];
|
||||
if (!grads_) {
|
||||
grads_ =
|
||||
CreateVariable(string::Sprintf("%s@IGrad", var_desc_->Name()),
|
||||
var_->Get<framework::LoDTensor>().dims(), 0.0, scope);
|
||||
}
|
||||
AddTo(grad, grads_);
|
||||
VLOG(3) << "grad_ after apply var grad " << var_desc_->Name() << " "
|
||||
<< grads_->Get<framework::LoDTensor>().data<float>()[0];
|
||||
}
|
||||
|
||||
std::vector<Variable*> OpBase::ApplyGrad(framework::Scope* scope) {
|
||||
VLOG(3) << "op grad " << grad_op_desc_->Type();
|
||||
|
||||
for (const std::string& grad_invar : grad_op_desc_->InputArgumentNames()) {
|
||||
if (grad_to_var_->find(grad_invar) == grad_to_var_->end()) {
|
||||
// grad op inputs can be forward inputs, so not in grad_to_var.
|
||||
continue;
|
||||
}
|
||||
VLOG(3) << "op grad in var " << grad_invar;
|
||||
block_->FindRecursiveOrCreateVar(grad_invar);
|
||||
framework::Variable* var = scope->Var(grad_invar);
|
||||
const std::string& invar = grad_to_var_->at(grad_invar);
|
||||
for (VarBase* varbase : *output_vars_) {
|
||||
// Use the accumulated grads_ by sharing the input with grads_.
|
||||
if (varbase->var_desc_->Name() == invar) {
|
||||
var->GetMutable<framework::LoDTensor>()->ShareDataWith(
|
||||
varbase->grads_->Get<framework::LoDTensor>());
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
|
||||
VLOG(3) << "grad outvar " << outvar;
|
||||
block_->FindRecursiveOrCreateVar(outvar);
|
||||
framework::Variable* var = scope->Var(outvar);
|
||||
if (!var->IsInitialized()) {
|
||||
framework::VarDesc* var_desc = block_->FindVar(outvar);
|
||||
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
|
||||
var->GetMutable<framework::LoDTensor>();
|
||||
} else {
|
||||
LOG(ERROR) << "tracer doesn't support yet";
|
||||
}
|
||||
}
|
||||
}
|
||||
grad_op_desc_->InferShape(*block_);
|
||||
grad_op_desc_->InferVarType(block_);
|
||||
std::unique_ptr<framework::OperatorBase> opbase =
|
||||
framework::OpRegistry::CreateOp(*grad_op_desc_);
|
||||
|
||||
opbase->Run(*scope, platform::CPUPlace());
|
||||
|
||||
// `ret` matches exactly with `input_vars_` of forward op.
|
||||
std::vector<Variable*> ret;
|
||||
for (size_t i = 0; i < input_vars_->size(); ++i) {
|
||||
bool found = false;
|
||||
for (const std::string& outvar : grad_op_desc_->OutputArgumentNames()) {
|
||||
Variable* var = scope->FindVar(outvar);
|
||||
VarBase* origin_var = (*input_vars_)[i];
|
||||
std::string orig_var = grad_to_var_->at(outvar);
|
||||
PADDLE_ENFORCE(origin_var->var_desc_->Name() == orig_var);
|
||||
VLOG(3) << "apply grad " << outvar << " with origin " << orig_var;
|
||||
origin_var->ApplyGrad(scope, var);
|
||||
found = true;
|
||||
ret.push_back(var);
|
||||
// TODO(panyx0718): There might be another outvar with the same name.
|
||||
// In that case, it doesn't matter the first one or the second one is
|
||||
// used.
|
||||
break;
|
||||
}
|
||||
if (!found) {
|
||||
ret.push_back(nullptr);
|
||||
}
|
||||
}
|
||||
return ret;
|
||||
}
|
||||
|
||||
void VarBase::RunBackward(framework::Scope* scope) {
|
||||
grads_ = CreateVariable(framework::GradVarName(var_desc_->Name()),
|
||||
var_->Get<framework::LoDTensor>().dims(), 1.0, scope,
|
||||
false);
|
||||
if (!pre_op_) return;
|
||||
Autograd(scope).RunBackward(this);
|
||||
}
|
||||
|
||||
} // namespace imperative
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,102 @@
|
||||
// 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 <string>
|
||||
#include <vector>
|
||||
#include "paddle/fluid/framework/op_desc.h"
|
||||
#include "paddle/fluid/framework/operator.h"
|
||||
#include "paddle/fluid/framework/scope.h"
|
||||
#include "paddle/fluid/framework/var_desc.h"
|
||||
#include "paddle/fluid/platform/enforce.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace imperative {
|
||||
|
||||
class OpBase;
|
||||
|
||||
class VarBase {
|
||||
public:
|
||||
VarBase()
|
||||
: pre_op_(nullptr),
|
||||
pre_op_out_idx_(-1),
|
||||
var_desc_(nullptr),
|
||||
var_(nullptr),
|
||||
grads_(nullptr) {}
|
||||
|
||||
virtual ~VarBase() {}
|
||||
|
||||
void ApplyGrad(framework::Scope* scope, framework::Variable* grad);
|
||||
|
||||
void RunBackward(framework::Scope* scope);
|
||||
|
||||
framework::LoDTensor& Grad();
|
||||
|
||||
OpBase* pre_op_;
|
||||
int pre_op_out_idx_;
|
||||
|
||||
framework::VarDesc* var_desc_;
|
||||
framework::Variable* var_;
|
||||
framework::Variable* grads_;
|
||||
};
|
||||
|
||||
class OpBase {
|
||||
public:
|
||||
OpBase()
|
||||
: input_vars_(new std::vector<VarBase*>()),
|
||||
output_vars_(new std::vector<VarBase*>()),
|
||||
pre_ops_(new std::vector<OpBase*>()),
|
||||
pre_ops_out_idx_(new std::vector<int>()),
|
||||
op_desc_(nullptr),
|
||||
grad_op_desc_(nullptr) {}
|
||||
|
||||
virtual ~OpBase() {
|
||||
delete input_vars_;
|
||||
delete output_vars_;
|
||||
|
||||
delete pre_ops_;
|
||||
delete pre_ops_out_idx_;
|
||||
|
||||
if (grad_op_desc_) delete grad_op_desc_;
|
||||
if (grad_to_var_) delete grad_to_var_;
|
||||
}
|
||||
|
||||
std::vector<framework::Variable*> ApplyGrad(framework::Scope* scope);
|
||||
|
||||
std::vector<VarBase*>* input_vars_;
|
||||
std::vector<VarBase*>* output_vars_;
|
||||
std::vector<OpBase*>* pre_ops_;
|
||||
std::vector<int>* pre_ops_out_idx_;
|
||||
framework::OpDesc* op_desc_;
|
||||
|
||||
framework::OpDesc* grad_op_desc_;
|
||||
std::unordered_map<std::string, std::string>* grad_to_var_;
|
||||
framework::BlockDesc* block_;
|
||||
};
|
||||
|
||||
class Layer {
|
||||
public:
|
||||
virtual ~Layer() {}
|
||||
|
||||
virtual std::vector<VarBase> Forward(const std::vector<VarBase>& inputs) {
|
||||
std::vector<VarBase> vars;
|
||||
return vars;
|
||||
}
|
||||
|
||||
virtual void Backward() { LOG(ERROR) << "To support customize"; }
|
||||
};
|
||||
|
||||
} // namespace imperative
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,19 @@
|
||||
// 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/imperative/tracer.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace imperative {} // namespace imperative
|
||||
} // namespace paddle
|
||||
@ -0,0 +1,128 @@
|
||||
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
|
||||
//
|
||||
// Licensed under the Apache License, Version 2.0 (the "License");
|
||||
// you may not use this file except in compliance with the License.
|
||||
// You may obtain a copy of the License at
|
||||
//
|
||||
// http://www.apache.org/licenses/LICENSE-2.0
|
||||
//
|
||||
// Unless required by applicable law or agreed to in writing, software
|
||||
// distributed under the License is distributed on an "AS IS" BASIS,
|
||||
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
// See the License for the specific language governing permissions and
|
||||
// limitations under the License.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <map>
|
||||
#include <string>
|
||||
#include <vector>
|
||||
|
||||
#include "paddle/fluid/framework/op_desc.h"
|
||||
#include "paddle/fluid/framework/op_registry.h"
|
||||
#include "paddle/fluid/framework/scope.h"
|
||||
#include "paddle/fluid/imperative/engine.h"
|
||||
#include "paddle/fluid/imperative/layer.h"
|
||||
|
||||
namespace paddle {
|
||||
namespace imperative {
|
||||
|
||||
void CreateGradOp(const framework::OpDesc& op_desc,
|
||||
const std::unordered_set<std::string>& no_grad_set,
|
||||
const std::vector<framework::BlockDesc*>& grad_sub_block,
|
||||
framework::OpDesc** grad_op_desc,
|
||||
std::unordered_map<std::string, std::string>* grad_to_var) {
|
||||
std::vector<std::unique_ptr<framework::OpDesc>> grad_op_descs =
|
||||
framework::OpInfoMap::Instance()
|
||||
.Get(op_desc.Type())
|
||||
.GradOpMaker()(op_desc, no_grad_set, grad_to_var, grad_sub_block);
|
||||
PADDLE_ENFORCE(grad_op_descs.size() == 1, "Only support 1 grad op now.");
|
||||
// TODO(panyx0718): Leak?
|
||||
*grad_op_desc = grad_op_descs[0].release();
|
||||
}
|
||||
|
||||
class Tracer {
|
||||
public:
|
||||
explicit Tracer(framework::BlockDesc* root_block) : root_block_(root_block) {
|
||||
root_scope_ = new framework::Scope();
|
||||
scopes_[root_block_] = root_scope_;
|
||||
}
|
||||
|
||||
virtual ~Tracer() { delete root_scope_; }
|
||||
|
||||
void Trace(OpBase* op, const std::vector<VarBase*>& inputs,
|
||||
const std::vector<VarBase*>& outputs,
|
||||
framework::BlockDesc* block) {
|
||||
framework::Scope* scope = GetScope(block);
|
||||
framework::OpDesc* op_desc = op->op_desc_;
|
||||
VLOG(3) << "tracer tracing " << op_desc->Type();
|
||||
op_desc->InferShape(*block);
|
||||
op_desc->InferVarType(block);
|
||||
std::unique_ptr<framework::OperatorBase> op_base =
|
||||
framework::OpRegistry::CreateOp(*op_desc);
|
||||
|
||||
*op->input_vars_ = inputs;
|
||||
for (VarBase* input : inputs) {
|
||||
const std::string vname = input->var_desc_->Name();
|
||||
framework::Variable* var = scope->Var(vname);
|
||||
input->var_ = var;
|
||||
if (!var->IsInitialized()) {
|
||||
framework::VarDesc* var_desc = block->FindVar(vname);
|
||||
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
|
||||
var->GetMutable<framework::LoDTensor>();
|
||||
} else {
|
||||
LOG(ERROR) << "tracer doesn't support yet";
|
||||
}
|
||||
}
|
||||
if (input->pre_op_) {
|
||||
op->pre_ops_->push_back(input->pre_op_);
|
||||
op->pre_ops_out_idx_->push_back(input->pre_op_out_idx_);
|
||||
} else {
|
||||
op->pre_ops_->push_back(nullptr);
|
||||
}
|
||||
}
|
||||
|
||||
*op->output_vars_ = outputs;
|
||||
for (size_t i = 0; i < outputs.size(); ++i) {
|
||||
const std::string vname = outputs[i]->var_desc_->Name();
|
||||
framework::Variable* var = scope->Var(vname);
|
||||
if (!var->IsInitialized()) {
|
||||
framework::VarDesc* var_desc = block->FindVar(vname);
|
||||
if (var_desc->GetType() == framework::proto::VarType::LOD_TENSOR) {
|
||||
var->GetMutable<framework::LoDTensor>();
|
||||
} else {
|
||||
LOG(ERROR) << "tracer doesn't support yet";
|
||||
}
|
||||
}
|
||||
outputs[i]->var_ = var;
|
||||
outputs[i]->pre_op_ = op;
|
||||
outputs[i]->pre_op_out_idx_ = i;
|
||||
}
|
||||
op_base->Run(*scope, platform::CPUPlace());
|
||||
framework::OpDesc* grad_op_desc;
|
||||
auto grad_to_var = new std::unordered_map<std::string, std::string>();
|
||||
CreateGradOp(*op_desc, {}, {block}, &grad_op_desc, grad_to_var);
|
||||
op->grad_op_desc_ = grad_op_desc;
|
||||
op->grad_to_var_ = grad_to_var;
|
||||
op->block_ = block;
|
||||
}
|
||||
|
||||
framework::Scope* GetScope(framework::BlockDesc* block) {
|
||||
if (scopes_.find(block) != scopes_.end()) {
|
||||
return scopes_.at(block);
|
||||
}
|
||||
framework::BlockDesc* parent_block = block->ParentBlock();
|
||||
PADDLE_ENFORCE(scopes_.find(parent_block) != scopes_.end());
|
||||
framework::Scope* scope = &scopes_[parent_block]->NewScope();
|
||||
scopes_[block] = scope;
|
||||
return scope;
|
||||
}
|
||||
|
||||
private:
|
||||
std::map<framework::BlockDesc*, framework::Scope*> scopes_;
|
||||
framework::BlockDesc* root_block_;
|
||||
framework::Scope* root_scope_;
|
||||
};
|
||||
|
||||
} // namespace imperative
|
||||
} // namespace paddle
|
||||
@ -1,4 +1,4 @@
|
||||
nv_library(tensorrt_plugin
|
||||
SRCS trt_plugin.cc split_op_plugin.cu elementwise_op_plugin.cu prelu_op_plugin.cu
|
||||
avg_pool_op_plugin.cu
|
||||
DEPS enforce tensorrt_engine)
|
||||
DEPS enforce tensorrt_engine prelu)
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Loading…
Reference in new issue