m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge branch 'dis_ckpt_fix' of github.com:seiriosPlus/Paddle into dis_ckpt_fix

Browse Source
revert-12864-feature/process_lod_grad
tangwei12 7 years ago
parent 74c5476f68 04b1df2ac8
commit 018e2f3a6c
53 changed files with 513 additions and 633 deletions
Show Stats Download Patch File Download Diff File

21
paddle/fluid/API.spec
Unescape View File

@ -1,8 +1,3 @@
paddle.fluid.Variable.__init__ ArgSpec(args=['self', 'block', 'type', 'name', 'shape', 'dtype', 'lod_level', 'capacity', 'persistable', 'error_clip', 'stop_gradient', 'is_data'], varargs=None, keywords='kwargs', defaults=(VarType.LOD_TENSOR, None, None, None, None, None, None, None, False, False))
paddle.fluid.Variable.astype ArgSpec(args=['self', 'dtype'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Variable.set_desc ArgSpec(args=['self', 'input'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Variable.set_error_clip ArgSpec(args=['self', 'error_clip'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Variable.to_string ArgSpec(args=['self', 'throw_on_error', 'with_details'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.Program.__init__ ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Program.block ArgSpec(args=['self', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Program.clone ArgSpec(args=['self', 'for_test'], varargs=None, keywords=None, defaults=(False,))
@ -33,8 +28,6 @@ paddle.fluid.Operator.set_attr ArgSpec(args=['self', 'name', 'val'], varargs=Non
paddle.fluid.Operator.to_string ArgSpec(args=['self', 'throw_on_error'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Parameter.__init__ ArgSpec(args=['self', 'block', 'shape', 'dtype'], varargs=None, keywords='kwargs', defaults=None)
paddle.fluid.Parameter.astype ArgSpec(args=['self', 'dtype'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Parameter.set_desc ArgSpec(args=['self', 'input'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Parameter.set_error_clip ArgSpec(args=['self', 'error_clip'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Parameter.to_string ArgSpec(args=['self', 'throw_on_error', 'with_details'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.default_startup_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.default_main_program ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
@ -42,8 +35,7 @@ paddle.fluid.program_guard ArgSpec(args=[], varargs='args', keywords='kwds', def
paddle.fluid.get_var ArgSpec(args=['name', 'program'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.Executor.__init__ ArgSpec(args=['self', 'place'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.as_lodtensor ArgSpec(args=['self', 'data'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.begin_pass ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.end_pass ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.close ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.Executor.run ArgSpec(args=['self', 'program', 'feed', 'fetch_list', 'feed_var_name', 'fetch_var_name', 'scope', 'return_numpy', 'use_program_cache'], varargs=None, keywords=None, defaults=(None, None, None, 'feed', 'fetch', None, True, False))
paddle.fluid.global_scope ArgSpec(args=[], varargs=None, keywords=None, defaults=None)
paddle.fluid.scope_guard ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
@ -207,31 +199,23 @@ paddle.fluid.layers.argsort ArgSpec(args=['input', 'axis', 'name'], varargs=None
paddle.fluid.layers.ones ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.zeros ArgSpec(args=['shape', 'dtype', 'force_cpu'], varargs=None, keywords=None, defaults=(False,))
paddle.fluid.layers.reverse ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.split_lod_tensor ArgSpec(args=['input', 'mask', 'level'], varargs=None, keywords=None, defaults=(0,))
paddle.fluid.layers.merge_lod_tensor ArgSpec(args=['in_true', 'in_false', 'x', 'mask', 'level'], varargs=None, keywords=None, defaults=(0,))
paddle.fluid.layers.While.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.While.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.While.complete ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.Switch.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.Switch.case ArgSpec(args=['self', 'condition'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.Switch.default ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.lod_rank_table ArgSpec(args=['x', 'level'], varargs=None, keywords=None, defaults=(0,))
paddle.fluid.layers.max_sequence_len ArgSpec(args=['rank_table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.lod_tensor_to_array ArgSpec(args=['x', 'table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.array_to_lod_tensor ArgSpec(args=['x', 'table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.increment ArgSpec(args=['x', 'value', 'in_place'], varargs=None, keywords=None, defaults=(1.0, True))
paddle.fluid.layers.array_write ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.create_array ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.less_than ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords='ignored', defaults=(None, None))
paddle.fluid.layers.equal ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords='ignored', defaults=(None,))
paddle.fluid.layers.array_read ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.shrink_memory ArgSpec(args=['x', 'i', 'table'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.array_length ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.IfElse.__init__ ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.IfElse.false_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.IfElse.input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.IfElse.output ArgSpec(args=['self'], varargs='outs', keywords=None, defaults=None)
paddle.fluid.layers.IfElse.parent_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.IfElse.true_block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.DynamicRNN.block ArgSpec(args=[], varargs='args', keywords='kwds', defaults=None)
@ -240,9 +224,6 @@ paddle.fluid.layers.DynamicRNN.output ArgSpec(args=['self'], varargs='outputs',
paddle.fluid.layers.DynamicRNN.static_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.step_input ArgSpec(args=['self', 'x'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.DynamicRNN.update_memory ArgSpec(args=['self', 'ex_mem', 'new_mem'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.ConditionalBlock.__init__ ArgSpec(args=['self', 'inputs', 'is_scalar_condition', 'name'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.layers.ConditionalBlock.block ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.ConditionalBlock.complete ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.__init__ ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.StaticRNN.complete_op ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.StaticRNN.memory ArgSpec(args=['self', 'init', 'shape', 'batch_ref', 'init_value', 'init_batch_dim_idx', 'ref_batch_dim_idx'], varargs=None, keywords=None, defaults=(None, None, None, 0.0, 0, 1))

12
paddle/fluid/framework/executor.cc
Unescape View File

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

12
paddle/fluid/framework/executor.h
Unescape View File

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

1
paddle/fluid/inference/api/api_impl.cc
Unescape View File

@ -137,6 +137,7 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
executor_->RunPreparedContext(
ctx_.get(), sub_scope_ != nullptr ? sub_scope_ : scope_.get(),
&feed_targets, &fetch_targets,
false, /* don't create local scope each time*/
false /* don't create variable eatch time */);
VLOG(4) << "Finish prepared context";
if (!GetFetch(fetchs, output_data)) {

14
paddle/fluid/inference/tensorrt/convert/fc_op.cc
Unescape View File

@ -32,11 +32,11 @@ void Reorder2(nvinfer1::DimsHW shape, const T* idata, nvinfer1::DimsHW istrides,
for (int h = 0; h < shape.h(); ++h) {
for (int w = 0; w < shape.w(); ++w) {
odata[h * ostrides.h() + w * ostrides.w()] =
idata[h * ostrides.h() + w * ostrides.w()];
idata[h * istrides.h() + w * istrides.w()];
}
}
}
// indata c * k
// Reorder the data layout from CK to KC.
void ReorderCKtoKC(TensorRTEngine::Weight& iweights,
TensorRTEngine::Weight* oweights) {
@ -79,9 +79,8 @@ class FcOpConverter : public OpConverter {
framework::LoDTensor tmp;
tmp.Resize(Y_t->dims());
memcpy(tmp.mutable_data<float>(platform::CPUPlace()), Y_t->data<float>(),
Y_t->dims()[0] * Y_t->dims()[1]);
memcpy(tmp.mutable_data<float>(platform::CPUPlace()), weight_data,
Y_t->dims()[0] * Y_t->dims()[1] * sizeof(float));
TensorRTEngine::Weight weight{nvinfer1::DataType::kFLOAT,
static_cast<void*>(weight_data),
Y_t->memory_size() / sizeof(float)};
@ -93,7 +92,7 @@ class FcOpConverter : public OpConverter {
// The data layout of TRT FC layer's weight is different from fluid's FC,
// need to reorder the elements.
ReorderCKtoKC(tmp_weight, &weight);
ReorderCKtoKC(weight, &tmp_weight);
// Currently, the framework can only handle one fluid op -> one TRT layer,
// but fc fuses `mul` and `bias` (2 fluid ops), so here is a trick, just
@ -103,7 +102,7 @@ class FcOpConverter : public OpConverter {
auto* layer = TRT_ENGINE_ADD_LAYER(engine_, FullyConnected,
*const_cast<nvinfer1::ITensor*>(X),
n_output, weight.get(), bias.get());
n_output, tmp_weight.get(), bias.get());
auto output_name = op_desc.Output("Out").front();
engine_->SetITensor(output_name, layer->getOutput(0));
@ -118,4 +117,3 @@ class FcOpConverter : public OpConverter {
} // namespace paddle
REGISTER_TRT_OP_CONVERTER(fc, FcOpConverter);
USE_OP(mul);

2
paddle/fluid/inference/tensorrt/convert/test_activation_op.cc
Unescape View File

@ -37,7 +37,7 @@ TEST(ReluOpConverter, main) {
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(10);
validator.Execute(1);
}
} // namespace tensorrt

13
paddle/fluid/inference/tensorrt/convert/test_fc_op.cc
Unescape View File

@ -23,11 +23,11 @@ namespace tensorrt {
TEST(fc_op, test) {
std::unordered_set<std::string> parameters({"mul-Y"});
framework::Scope scope;
TRTConvertValidation validator(20, parameters, scope, 1000);
validator.DeclInputVar("mul-X", nvinfer1::Dims4(8, 3, 1, 1));
validator.DeclParamVar("mul-Y", nvinfer1::Dims2(3, 2));
validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(8, 2));
TRTConvertValidation validator(10, parameters, scope, 1000);
validator.DeclInputVar("mul-X", nvinfer1::Dims4(1, 10, 1, 1));
validator.DeclParamVar("mul-Y", nvinfer1::Dims2(10, 2));
// validator.DeclParamVar("mul-Y", nvinfer1::Dims2(8, 2));
validator.DeclOutputVar("mul-Out", nvinfer1::Dims2(1, 2));
// Prepare Op description
framework::OpDesc desc;
@ -38,9 +38,10 @@ TEST(fc_op, test) {
validator.SetOp(*desc.Proto());
validator.Execute(10);
validator.Execute(1);
}
} // namespace tensorrt
} // namespace inference
} // namespace paddle
USE_OP(mul);

2
paddle/fluid/inference/tensorrt/convert/test_mul_op.cc
Unescape View File

@ -39,7 +39,7 @@ TEST(MulOpConverter, main) {
validator.SetOp(*desc.Proto());
LOG(INFO) << "execute";
validator.Execute(10);
validator.Execute(1);
}
} // namespace tensorrt

10
paddle/fluid/inference/tensorrt/convert/ut_helper.h
Unescape View File

@ -39,7 +39,7 @@ namespace tensorrt {
float random(float low, float high) {
static std::random_device rd;
static std::mt19937 mt(rd());
std::uniform_real_distribution<double> dist(1.0, 10.0);
std::uniform_real_distribution<double> dist(low, high);
return dist(mt);
}
@ -49,6 +49,7 @@ void RandomizeTensor(framework::LoDTensor* tensor, const platform::Place& place,
size_t num_elements = analysis::AccuDims(dims, dims.size());
PADDLE_ENFORCE_GT(num_elements, 0);
auto* data = tensor->mutable_data<float>(place);
for (size_t i = 0; i < num_elements; i++) {
*(data + i) = random(0., 1.);
}
@ -68,7 +69,7 @@ class TRTConvertValidation {
int workspace_size = 1 << 10)
: parameters_(parameters), scope_(scope) {
// create engine.
engine_.reset(new TensorRTEngine(10, 1 << 10, &stream_));
engine_.reset(new TensorRTEngine(batch_size, workspace_size, &stream_));
engine_->InitNetwork();
PADDLE_ENFORCE_EQ(cudaStreamCreate(&stream_), 0);
@ -138,12 +139,11 @@ class TRTConvertValidation {
cudaStreamSynchronize(*engine_->stream());
ASSERT_FALSE(op_desc_->OutputArgumentNames().empty());
const size_t output_space_size = 200;
const size_t output_space_size = 2000;
for (const auto& output : op_desc_->OutputArgumentNames()) {
std::vector<float> fluid_out;
std::vector<float> trt_out(output_space_size);
engine_->GetOutputInCPU(output, &trt_out[0],
output_space_size * sizeof(float));
engine_->GetOutputInCPU(output, &trt_out[0], output_space_size);
cudaStreamSynchronize(*engine_->stream());
auto* var = scope_.FindVar(output);

62
paddle/fluid/inference/tensorrt/engine.cc
Unescape View File

@ -1,7 +1,7 @@
/* 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.
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
@ -26,6 +26,8 @@ namespace paddle {
namespace inference {
namespace tensorrt {
int TensorRTEngine::runtime_batch_ = 1;
void TensorRTEngine::Build(const DescType &paddle_model) {
PADDLE_ENFORCE(false, "not implemented");
}
@ -42,6 +44,7 @@ void TensorRTEngine::Execute(int batch_size) {
PADDLE_ENFORCE_NOT_NULL(stream_);
infer_context_->enqueue(batch_size, buffers.data(), *stream_, nullptr);
cudaStreamSynchronize(*stream_);
SetRuntimeBatch(batch_size);
}
TensorRTEngine::~TensorRTEngine() {
@ -80,17 +83,17 @@ void TensorRTEngine::FreezeNetwork() {
auto dims = infer_engine_->getBindingDimensions(slot_offset);
item.second = kDataTypeSize[static_cast<int>(
infer_engine_->getBindingDataType(slot_offset))] *
analysis::AccuDims(dims.d, dims.nbDims);
analysis::AccuDims(dims.d, dims.nbDims) * max_batch_;
PADDLE_ENFORCE_GT(item.second, 0);
}
auto &buf = buffer(item.first);
buf.max_size = item.second * max_batch_;
CHECK(buf.buffer == nullptr); // buffer should be allocated only once.
PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, buf.max_size));
PADDLE_ENFORCE_LE(buf.max_size, 1 << 30); // 10G
// buf.size will changed in the runtime.
PADDLE_ENFORCE_EQ(0, cudaMalloc(&buf.buffer, item.second * max_batch_));
buf.size = 0;
PADDLE_ENFORCE_LE(buf.max_size, 1 << 30); // 10G
buf.device = DeviceType::GPU;
}
}
@ -105,7 +108,7 @@ nvinfer1::ITensor *TensorRTEngine::DeclareInput(const std::string &name,
auto *input = infer_network_->addInput(name.c_str(), dtype, dims);
PADDLE_ENFORCE(input, "infer network add input %s failed", name);
buffer_sizes_[name] = kDataTypeSize[static_cast<int>(dtype)] *
analysis::AccuDims(dims.d, dims.nbDims);
analysis::AccuDims(dims.d, dims.nbDims) * max_batch_;
PADDLE_ENFORCE(input->isNetworkInput());
TensorRTEngine::SetITensor(name, input);
return input;
@ -149,35 +152,42 @@ void *TensorRTEngine::GetOutputInGPU(const std::string &name) {
void TensorRTEngine::GetOutputInGPU(const std::string &name, void *dst,
size_t max_size) {
// determine data size
auto *output = TensorRTEngine::GetITensor(name);
nvinfer1::Dims dims = output->getDimensions();
auto dim_size = analysis::AccuDims(dims.d, dims.nbDims);
size_t dst_size = dim_size * runtime_batch_ *
kDataTypeSize[static_cast<int>(output->getType())];
auto it = buffer_sizes_.find(name);
PADDLE_ENFORCE(it != buffer_sizes_.end());
PADDLE_ENFORCE_GT(it->second, 0);
PADDLE_ENFORCE_GE(max_size, it->second);
PADDLE_ENFORCE_LE(dst_size, it->second);
PADDLE_ENFORCE_GE(max_size, dst_size);
auto &buf = buffer(name);
PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before");
PADDLE_ENFORCE_EQ(cudaMemcpyAsync(dst, buf.buffer, it->second,
PADDLE_ENFORCE_EQ(cudaMemcpyAsync(dst, buf.buffer, dst_size,
cudaMemcpyDeviceToDevice, *stream_),
0);
}
void TensorRTEngine::GetOutputInCPU(const std::string &name, void *dst,
size_t max_size) {
VLOG(4) << "get output in cpu";
auto &buf = buffer(name);
// Update needed buffer size.
auto slot_offset = infer_engine_->getBindingIndex(name.c_str());
auto dims = infer_engine_->getBindingDimensions(slot_offset);
buf.size = kDataTypeSize[static_cast<int>(
infer_engine_->getBindingDataType(slot_offset))] *
analysis::AccuDims(dims.d, dims.nbDims);
PADDLE_ENFORCE_LE(buf.size, buf.max_size);
// determine data size
auto *output = TensorRTEngine::GetITensor(name);
nvinfer1::Dims dims = output->getDimensions();
auto dim_size = analysis::AccuDims(dims.d, dims.nbDims);
size_t dst_size = dim_size * runtime_batch_ *
kDataTypeSize[static_cast<int>(output->getType())];
auto it = buffer_sizes_.find(name);
PADDLE_ENFORCE(it != buffer_sizes_.end());
PADDLE_ENFORCE_GT(it->second, 0);
PADDLE_ENFORCE_LE(dst_size, it->second);
PADDLE_ENFORCE_GE(max_size, dst_size);
auto &buf = buffer(name);
PADDLE_ENFORCE_NOT_NULL(buf.buffer, "buffer should be allocated before");
// DEBUG
memset(dst, 0, buf.size);
PADDLE_ENFORCE_EQ(
0, cudaMemcpy(dst, buf.buffer, buf.size, cudaMemcpyDeviceToHost));
PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(dst, buf.buffer, dst_size,
cudaMemcpyDeviceToHost, *stream_));
}
Buffer &TensorRTEngine::buffer(const std::string &name) {
@ -225,6 +235,12 @@ nvinfer1::ITensor *TensorRTEngine::GetITensor(const std::string &name) {
return itensor_map_[name];
}
void TensorRTEngine::SetRuntimeBatch(size_t batch_size) {
runtime_batch_ = batch_size;
}
int TensorRTEngine::GetRuntimeBatch() { return runtime_batch_; }
} // namespace tensorrt
} // namespace inference
} // namespace paddle

4
paddle/fluid/inference/tensorrt/engine.h
Unescape View File

@ -117,10 +117,14 @@ class TensorRTEngine : public EngineBase {
nvinfer1::ICudaEngine* engine() { return infer_engine_.get(); }
nvinfer1::INetworkDefinition* network() { return infer_network_.get(); }
void SetRuntimeBatch(size_t batch_size);
int GetRuntimeBatch();
private:
// the max batch size
int max_batch_;
// the runtime batch size
static int runtime_batch_;
// the max memory size the engine uses
int max_workspace_;

40
paddle/fluid/inference/tensorrt/test_engine.cc
Unescape View File

@ -28,7 +28,7 @@ class TensorRTEngineTest : public ::testing::Test {
protected:
void SetUp() override {
ASSERT_EQ(0, cudaStreamCreate(&stream_));
engine_ = new TensorRTEngine(1, 1 << 10, &stream_);
engine_ = new TensorRTEngine(10, 1 << 10, &stream_);
engine_->InitNetwork();
}
@ -71,7 +71,7 @@ TEST_F(TensorRTEngineTest, add_layer) {
LOG(INFO) << "to get output";
float y_cpu;
engine_->GetOutputInCPU("y", &y_cpu, sizeof(float));
engine_->GetOutputInCPU("y", &y_cpu, 1 * sizeof(float));
LOG(INFO) << "to checkout output";
ASSERT_EQ(y_cpu, x_v * 2 + 3);
@ -103,15 +103,49 @@ TEST_F(TensorRTEngineTest, add_layer_multi_dim) {
LOG(INFO) << "to get output";
float y_cpu[2] = {-1., -1.};
auto dims = engine_->GetITensor("y")->getDimensions();
ASSERT_EQ(dims.nbDims, 3);
ASSERT_EQ(dims.d[0], 2);
ASSERT_EQ(dims.d[1], 1);
engine_->GetOutputInCPU("y", &y_cpu[0], sizeof(float) * 2);
engine_->GetOutputInCPU("y", &y_cpu[0], 2 * sizeof(float));
ASSERT_EQ(y_cpu[0], 4.5);
ASSERT_EQ(y_cpu[1], 14.5);
}
TEST_F(TensorRTEngineTest, test_conv2d_temp) {
// Weight in CPU memory.
float raw_weight[9] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
float raw_bias[1] = {0};
TensorRTEngine::Weight weight(nvinfer1::DataType::kFLOAT, raw_weight, 9);
TensorRTEngine::Weight bias(nvinfer1::DataType::kFLOAT, raw_bias, 1);
auto* x = engine_->DeclareInput("x", nvinfer1::DataType::kFLOAT,
nvinfer1::Dims3{1, 3, 3});
auto* conv_layer =
TRT_ENGINE_ADD_LAYER(engine_, Convolution, *x, 1, nvinfer1::DimsHW{3, 3},
weight.get(), bias.get());
PADDLE_ENFORCE(conv_layer != nullptr);
conv_layer->setStride(nvinfer1::DimsHW{1, 1});
conv_layer->setPadding(nvinfer1::DimsHW{1, 1});
engine_->DeclareOutput(conv_layer, 0, "y");
engine_->FreezeNetwork();
ASSERT_EQ(engine_->engine()->getNbBindings(), 2);
float x_v[18] = {1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0,
1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0};
engine_->SetInputFromCPU("x", reinterpret_cast<void*>(&x_v),
18 * sizeof(float));
engine_->Execute(2);
LOG(INFO) << "to get output";
float* y_cpu = new float[18];
engine_->GetOutputInCPU("y", &y_cpu[0], 18 * sizeof(float));
ASSERT_EQ(y_cpu[0], 4.0);
ASSERT_EQ(y_cpu[1], 6.0);
}
} // namespace tensorrt
} // namespace inference
} // namespace paddle

10
paddle/fluid/inference/tests/test_helper.h
Unescape View File

@ -210,13 +210,14 @@ void TestInference(const std::string& dirname,
// Ignore the profiling results of the first run
std::unique_ptr<paddle::framework::ExecutorPrepareContext> ctx;
bool CreateLocalScope = CreateVars;
if (PrepareContext) {
ctx = executor.Prepare(*inference_program, 0);
executor.RunPreparedContext(ctx.get(), scope, &feed_targets,
&fetch_targets, true, CreateVars);
&fetch_targets, CreateLocalScope, CreateVars);
} else {
executor.Run(*inference_program, scope, &feed_targets, &fetch_targets,
true, CreateVars);
CreateLocalScope, CreateVars);
}
// Enable the profiler
@ -232,10 +233,11 @@ void TestInference(const std::string& dirname,
// Note: if you change the inference_program, you need to call
// executor.Prepare() again to get a new ExecutorPrepareContext.
executor.RunPreparedContext(ctx.get(), scope, &feed_targets,
&fetch_targets, CreateVars);
&fetch_targets, CreateLocalScope,
CreateVars);
} else {
executor.Run(*inference_program, scope, &feed_targets, &fetch_targets,
CreateVars);
CreateLocalScope, CreateVars);
}
}

2
paddle/fluid/operators/distributed/CMakeLists.txt
Unescape View File

@ -18,7 +18,7 @@ if(WITH_GRPC)
set_source_files_properties(grpc_serde_test.cc rpc_server_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
cc_test(grpc_serde_test SRCS grpc_serde_test.cc
DEPS grpc++_unsecure grpc_unsecure gpr cares zlib protobuf sendrecvop_grpc scope profiler math_function SERIAL)
cc_test(grpc_server_test SRCS rpc_server_test.cc
cc_test(rpc_server_test SRCS rpc_server_test.cc
DEPS sendrecvop_grpc grpc++_unsecure grpc_unsecure gpr cares zlib protobuf executor proto_desc lookup_table_op SERIAL)
return()
endif()

39
paddle/fluid/operators/distributed/grpc_client.cc
Unescape View File

@ -36,20 +36,16 @@ void GRPCClient::InitEventLoop() {
client_thread_.reset(new std::thread(std::bind(&GRPCClient::Proceed, this)));
}
void GRPCClient::SendBeginPass() {
for (auto& it : channels_) {
VLOG(3) << "send begin pass to: " << it.first;
this->AsyncSendBeginPass(it.first);
}
this->Wait();
}
void GRPCClient::SendEndPass() {
for (auto& it : channels_) {
VLOG(3) << "send end pass to " << it.first;
this->AsyncSendEndPass(it.first);
void GRPCClient::SendComplete() {
std::unique_lock<std::mutex> lk(completed_mutex_);
if (!completed_) {
for (auto& it : channels_) {
VLOG(3) << "send complete message to " << it.first;
this->AsyncSendComplete(it.first);
}
PADDLE_ENFORCE(this->Wait(), "internal grpc error");
completed_ = true;
}
this->Wait();
}
GRPCClient::~GRPCClient() {
@ -239,32 +235,19 @@ void GRPCClient::AsyncSendFetchBarrier(const std::string& ep,
req_count_++;
}
void GRPCClient::AsyncSendBeginPass(const std::string& ep, int64_t time_out) {
void GRPCClient::AsyncSendComplete(const std::string& ep, int64_t time_out) {
const auto ch = GetChannel(ep);
BatchBarrierProcessor* s = new BatchBarrierProcessor(ch);
s->Prepare(time_out);
sendrecv::VariableMessage req;
req.set_varname(BEGIN_PASS_MESSAGE);
req.set_varname(COMPLETE_MESSAGE);
auto rpc = s->stub_->AsyncSendVariable(s->context_.get(), req, &cq_);
rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
req_count_++;
}
void GRPCClient::AsyncSendEndPass(const std::string& ep, int64_t time_out) {
const auto ch = GetChannel(ep);
FetchBarrierProcessor* s = new FetchBarrierProcessor(ch);
s->Prepare(time_out);
sendrecv::VariableMessage req;
req.set_varname(END_PASS_MESSAGE);
auto rpc = s->stub_->AsyncGetVariable(s->context_.get(), req, &cq_);
rpc->Finish(&s->reply_, &s->status_, reinterpret_cast<void*>(s));
req_count_++;
}
void GRPCClient::AsyncCheckpointNotify(const std::string& ep,
const std::string& dir,
int64_t time_out) {

17
paddle/fluid/operators/distributed/grpc_client.h
Unescape View File

@ -174,7 +174,7 @@ class CheckpointNotifyProcessor : public BaseProcessor {
class GRPCClient : public RPCClient {
public:
GRPCClient() : ok_(true) {}
GRPCClient() : ok_(true), completed_(false) {}
virtual ~GRPCClient();
bool AsyncSendVar(const std::string& ep, const platform::DeviceContext& ctx,
@ -201,17 +201,12 @@ class GRPCClient : public RPCClient {
void AsyncCheckpointNotify(const std::string& ep, const std::string& dir,
int64_t time_out = FLAGS_rpc_deadline) override;
void AsyncSendBeginPass(const std::string& ep,
int64_t time_out = FLAGS_rpc_deadline) override;
void AsyncSendEndPass(const std::string& ep,
int64_t time_out = FLAGS_rpc_deadline) override;
void AsyncSendComplete(const std::string& ep,
int64_t time_out = FLAGS_rpc_deadline) override;
bool Wait() override;
void SendBeginPass() override;
void SendEndPass() override;
void SendComplete() override;
protected:
void InitImpl() override;
@ -238,6 +233,10 @@ class GRPCClient : public RPCClient {
// mutex for GetChannel thread safety
std::mutex chan_mutex_;
DISABLE_COPY_AND_ASSIGN(GRPCClient);
// mutex for sending complete message only once
std::mutex completed_mutex_;
bool completed_;
};
} // namespace distributed

2
paddle/fluid/operators/distributed/request_handler.h
Unescape View File

@ -43,8 +43,6 @@ constexpr char kRequestPassBarrier[] = "RequestPassBarrier";
#define BATCH_BARRIER_MESSAGE "BATCH_BARRIER@RECV"
#define FETCH_BARRIER_MESSAGE "FETCH_BARRIER@RECV"
#define COMPLETE_MESSAGE "COMPLETE@RECV"
#define BEGIN_PASS_MESSAGE "BEGIN_PASS@RECV"
#define END_PASS_MESSAGE "END_PASS@RECV"
#define CHECKPOINT_SAVE_MESSAGE "SAVE@CHECKPOINTNOTIFY"
#define CHECKPOINT_LOAD_MESSAGE "LOAD@CHECKPOINTNOTIFY"

11
paddle/fluid/operators/distributed/request_handler_impl.cc
Unescape View File

@ -55,10 +55,9 @@ bool RequestSendHandler::Handle(const std::string& varname,
if (varname == BATCH_BARRIER_MESSAGE) {
VLOG(3) << "sync: recv BATCH_BARRIER_MESSAGE";
rpc_server_->IncreaseBatchBarrier(kRequestSend);
} else if (varname == BEGIN_PASS_MESSAGE) {
VLOG(3) << "sync: recv begin pass message";
rpc_server_->WaitCond(kRequestSend);
rpc_server_->BeginPass();
} else if (varname == COMPLETE_MESSAGE) {
VLOG(3) << "sync: recv complete message";
rpc_server_->Complete();
} else {
VLOG(3) << "sync: received var_name: " << varname;
rpc_server_->WaitCond(kRequestSend);
@ -94,14 +93,12 @@ bool RequestGetHandler::Handle(const std::string& varname,
if (varname == FETCH_BARRIER_MESSAGE) {
VLOG(3) << "sync: recv fetch barrier message";
rpc_server_->IncreaseBatchBarrier(kRequestGet);
} else if (varname == END_PASS_MESSAGE) {
rpc_server_->EndPass();
} else {
rpc_server_->WaitCond(kRequestGet);
*outvar = scope_->FindVar(varname);
}
} else {
if (varname != FETCH_BARRIER_MESSAGE && varname != END_PASS_MESSAGE) {
if (varname != FETCH_BARRIER_MESSAGE && varname != COMPLETE_MESSAGE) {
*outvar = scope_->FindVar(varname);
}
}

14
paddle/fluid/operators/distributed/rpc_client.h
Unescape View File

@ -60,17 +60,13 @@ class RPCClient {
const std::string& dir,
int64_t time_out = FLAGS_rpc_deadline) = 0;
virtual void AsyncSendBeginPass(const std::string& ep,
int64_t time_out = FLAGS_rpc_deadline) = 0;
virtual void AsyncSendComplete(const std::string& ep,
int64_t time_out = FLAGS_rpc_deadline) = 0;
virtual void AsyncSendEndPass(const std::string& ep,
int64_t time_out = FLAGS_rpc_deadline) = 0;
// BeginePass/EndPass tells all the pserver that start/end a pass, so that
// the pserver can increase/reduce it's barrier count, and continue to train
// Complete tells all the pserver instances that finishe the training,
// the pserver can reduce it's barrier count, and continue to train
// with other trainers.
virtual void SendBeginPass() = 0;
virtual void SendEndPass() = 0;
virtual void SendComplete() = 0;
virtual bool Wait() = 0;

18
paddle/fluid/operators/distributed/rpc_server.cc
Unescape View File

@ -64,18 +64,7 @@ void RPCServer::IncreaseBatchBarrier(const std::string rpc_name) {
}
}
void RPCServer::BeginPass() {
VLOG(4) << "RPCServer begin increase pass barrier";
{
std::unique_lock<std::mutex> lock(mutex_);
client_num_++;
VLOG(4) << "increase client_num to: " << client_num_;
}
barrier_cond_.notify_all();
}
void RPCServer::EndPass() {
VLOG(4) << "RPCServer begin increase pass barrier";
void RPCServer::Complete() {
{
std::unique_lock<std::mutex> lock(mutex_);
client_num_--;
@ -87,6 +76,11 @@ void RPCServer::EndPass() {
barrier_cond_.notify_all();
}
int RPCServer::GetClientNum() {
std::unique_lock<std::mutex> lock(mutex_);
return client_num_;
}
void RPCServer::ResetBarrierCounter() {
VLOG(3) << "RPCServer ResetBarrierCounter ";
std::unique_lock<std::mutex> lock(mutex_);

5
paddle/fluid/operators/distributed/rpc_server.h
Unescape View File

@ -44,7 +44,7 @@ class RPCServer {
int GetSelectedPort() const { return selected_port_; }
int GetClientNum() const;
int GetClientNum();
void SavePort() const;
@ -64,8 +64,7 @@ class RPCServer {
void WaitCond(const std::string& rpc_name);
void IncreaseBatchBarrier(const std::string rpc_name);
void BeginPass();
void EndPass();
void Complete();
void ResetBarrierCounter();

29
paddle/fluid/operators/distributed/rpc_server_test.cc
Unescape View File

@ -91,7 +91,7 @@ void InitTensorsOnServer(framework::Scope* scope, platform::CPUPlace* place,
}
}
void StartServer() {
void StartServer(const std::string& rpc_name) {
framework::ProgramDesc program;
framework::Scope scope;
platform::CPUPlace place;
@ -107,14 +107,14 @@ void StartServer() {
std::shared_ptr<framework::ExecutorPrepareContext>>
prefetch_var_name_to_prepared;
prefetch_var_name_to_prepared[in_var_name] = prepared[0];
g_req_handler->SetProgram(&program);
g_req_handler->SetPrefetchPreparedCtx(&prefetch_var_name_to_prepared);
g_req_handler->SetDevCtx(&ctx);
g_req_handler->SetScope(&scope);
g_req_handler->SetExecutor(&exe);
g_rpc_service->RegisterRPC(distributed::kRequestPrefetch,
g_req_handler.get());
g_rpc_service->RegisterRPC(rpc_name, g_req_handler.get());
g_req_handler->SetRPCServer(g_rpc_service.get());
std::thread server_thread(
@ -129,7 +129,7 @@ TEST(PREFETCH, CPU) {
distributed::RPCClient* client =
distributed::RPCClient::GetInstance<RPCCLIENT_T>();
std::thread server_thread(StartServer);
std::thread server_thread(StartServer, distributed::kRequestPrefetch);
g_rpc_service->WaitServerReady();
int port = g_rpc_service->GetSelectedPort();
@ -162,3 +162,24 @@ TEST(PREFETCH, CPU) {
g_rpc_service.reset(nullptr);
g_req_handler.reset(nullptr);
}
TEST(COMPLETE, CPU) {
g_req_handler.reset(new distributed::RequestSendHandler(true));
g_rpc_service.reset(new RPCSERVER_T("127.0.0.1:0", 2));
distributed::RPCClient* client =
distributed::RPCClient::GetInstance<RPCCLIENT_T>();
PADDLE_ENFORCE(client != nullptr);
std::thread server_thread(StartServer, distributed::kRequestSend);
g_rpc_service->WaitServerReady();
int port = g_rpc_service->GetSelectedPort();
std::string ep = paddle::string::Sprintf("127.0.0.1:%d", port);
client->AsyncSendComplete(ep);
client->Wait();
EXPECT_EQ(g_rpc_service->GetClientNum(), 1);
g_rpc_service->ShutDown();
server_thread.join();
g_rpc_service.reset(nullptr);
g_req_handler.reset(nullptr);
}

17
paddle/fluid/operators/reader/lod_tensor_blocking_queue.h
Unescape View File

@ -38,12 +38,10 @@ class LoDTensorBlockingQueue {
public:
bool Push(const std::vector<framework::LoDTensor>& lod_tensor_vec) {
CheckDims(lod_tensor_vec);
return queue_.Send(lod_tensor_vec);
}
bool Push(std::vector<framework::LoDTensor>&& lod_tensor_vec) {
CheckDims(lod_tensor_vec);
return queue_.Send(std::move(lod_tensor_vec));
}
@ -65,21 +63,6 @@ class LoDTensorBlockingQueue {
inline bool IsClosed() const { return queue_.IsClosed(); }
private:
void CheckDims(
const std::vector<framework::LoDTensor>& lod_tensor_vec) const {
PADDLE_ENFORCE(dims_.size() == lod_tensor_vec.size(),
"Expect input size is %d but found %s", dims_.size(),
lod_tensor_vec.size());
for (size_t i = 0; i < dims_.size(); ++i) {
const auto& in_dims = framework::slice_ddim(
lod_tensor_vec[i].dims(), 1, lod_tensor_vec[i].dims().size());
const auto& expect_dims =
framework::slice_ddim(dims_[i], 1, dims_[i].size());
PADDLE_ENFORCE(in_dims == expect_dims,
"Dims of the %d-th input tensor do not match", i);
}
}
BlockingQueue<std::vector<framework::LoDTensor>> queue_;
std::vector<framework::DDim> dims_;
};

2
paddle/fluid/operators/reshape_op.cc
Unescape View File

@ -216,7 +216,7 @@ class ReshapeKernel {
if (shape_tensor) {
auto *shape_data = shape_tensor->data<int>();
framework::Tensor cpu_shape_tensor;
if (platform::is_gpu_place(ctx.GetPlace())) {
if (platform::is_gpu_place(shape_tensor->place())) {
TensorCopySync(*shape_tensor, platform::CPUPlace(), &cpu_shape_tensor);
shape_data = cpu_shape_tensor.data<int>();
}

7
paddle/fluid/operators/tensorrt_engine_op.cc
Unescape View File

@ -55,13 +55,14 @@ nvinfer1::Dims Vec2TRT_Dims(const std::vector<int64_t> &shape) {
"TensorRT' tensor input requires at least 2 dimensions");
PADDLE_ENFORCE_LE(shape.size(), 4UL,
"TensorRT' tensor input requires at most 4 dimensions");
switch (shape.size()) {
case 2:
return nvinfer1::Dims2(shape[0], shape[1]);
return nvinfer1::Dims2(1, shape[1]);
case 3:
return nvinfer1::Dims3(shape[0], shape[1], shape[2]);
return nvinfer1::Dims3(1, shape[1], shape[2]);
case 4:
return nvinfer1::Dims4(shape[0], shape[1], shape[2], shape[3]);
return nvinfer1::Dims4(1, shape[1], shape[2], shape[3]);
default:
return nvinfer1::Dims();
}

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

Write Preview
Loading…
Cancel
Save