提交 31f587e7 编写于 作者: M minqiyang

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

......@@ -2,6 +2,11 @@ if (NOT WITH_ANAKIN)
return()
endif()
option(ANAKIN_ENABLE_OP_TIMER "Get more detailed information with Anakin op time" OFF)
if(ANAKIN_ENABLE_OP_TIMER)
add_definitions(-DPADDLE_ANAKIN_ENABLE_OP_TIMER)
endif()
INCLUDE(ExternalProject)
set(ANAKIN_SOURCE_DIR ${THIRD_PARTY_PATH}/anakin)
# the anakin install dir is only default one now
......@@ -11,23 +16,34 @@ set(ANAKIN_LIBRARY ${ANAKIN_INSTALL_DIR})
set(ANAKIN_SHARED_LIB ${ANAKIN_LIBRARY}/libanakin.so)
set(ANAKIN_SABER_LIB ${ANAKIN_LIBRARY}/libanakin_saber_common.so)
# TODO(luotao): ANAKIN_MODLE_URL will move to demo ci later.
set(ANAKIN_MODLE_URL "http://paddle-inference-dist.bj.bcebos.com/mobilenet_v2.anakin.bin")
# TODO(luotao): ANAKIN_MODLE_URL etc will move to demo ci later.
set(INFERENCE_URL "http://paddle-inference-dist.bj.bcebos.com")
set(ANAKIN_MODLE_URL "${INFERENCE_URL}/mobilenet_v2.anakin.bin")
set(ANAKIN_RNN_MODLE_URL "${INFERENCE_URL}/anakin_test%2Fditu_rnn.anakin2.model.bin")
set(ANAKIN_RNN_DATA_URL "${INFERENCE_URL}/anakin_test%2Fditu_rnn_data.txt")
execute_process(COMMAND bash -c "mkdir -p ${ANAKIN_SOURCE_DIR}")
execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_MODLE_URL}")
execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_MODLE_URL} -N")
execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_RNN_MODLE_URL} -N")
execute_process(COMMAND bash -c "cd ${ANAKIN_SOURCE_DIR}; wget -q --no-check-certificate ${ANAKIN_RNN_DATA_URL} -N")
include_directories(${ANAKIN_INCLUDE})
include_directories(${ANAKIN_INCLUDE}/saber/)
include_directories(${ANAKIN_INCLUDE}/saber/core/)
include_directories(${ANAKIN_INCLUDE}/saber/funcs/impl/x86/)
include_directories(${ANAKIN_INCLUDE}/saber/funcs/impl/cuda/base/cuda_c/)
set(ANAKIN_COMPILE_EXTRA_FLAGS
-Wno-error=unused-but-set-variable -Wno-unused-but-set-variable
-Wno-error=unused-variable -Wno-unused-variable
-Wno-error=format-extra-args -Wno-format-extra-args
-Wno-error=comment -Wno-comment
-Wno-error=format -Wno-format
-Wno-error=comment -Wno-comment
-Wno-error=format -Wno-format
-Wno-error=maybe-uninitialized -Wno-maybe-uninitialized
-Wno-error=switch -Wno-switch
-Wno-error=return-type -Wno-return-type
-Wno-error=non-virtual-dtor -Wno-non-virtual-dtor
-Wno-error=ignored-qualifiers
-Wno-ignored-qualifiers
-Wno-sign-compare
-Wno-reorder
-Wno-error=cpp)
......@@ -38,7 +54,7 @@ ExternalProject_Add(
DEPENDS ${MKLML_PROJECT}
# Anakin codes error on Intel(R) Xeon(R) Gold 5117 CPU, temporary do not compile avx512 related code.
GIT_REPOSITORY "https://github.com/luotao1/Anakin"
GIT_TAG "bcf17aabe7921ceb7bce591244b4f9dce7dba5c8"
GIT_TAG "211d1fc5d813d70c0c14072f9083cf25f40940ea"
PREFIX ${ANAKIN_SOURCE_DIR}
UPDATE_COMMAND ""
CMAKE_ARGS -DUSE_GPU_PLACE=YES
......@@ -48,6 +64,7 @@ ExternalProject_Add(
-DMKLML_ROOT=${THIRD_PARTY_PATH}/install/mklml
-DCUDNN_ROOT=${CUDNN_ROOT}
-DCUDNN_INCLUDE_DIR=${CUDNN_INCLUDE_DIR}
-DENABLE_OP_TIMER=${ANAKIN_ENABLE_OP_TIMER}
${EXTERNAL_OPTIONAL_ARGS}
CMAKE_CACHE_ARGS -DCMAKE_INSTALL_PREFIX:PATH=${ANAKIN_INSTALL_DIR}
)
......
......@@ -119,10 +119,29 @@ $$Out = scale*X$$
这个例子有`AddAttr<AttrType>("scale", "...").SetDefault(1.0);` : 增加`scale`系数,作为参数属性,并且设置默认值为1.0。
### 定义GradProtoMaker类
每个Op的必须有一个对应的GraProtoMaker,若未定制对应前向Op的GradProtoMaker,fluid提供了DefaultGradProtoMaker,默认注册会使用全部输入输出,包括Input, Output, Output@Grad等,使用不需要的变量的会造成显存浪费。
下面示例定义了ScaleOp的GradProtoMaker。
```cpp
class ScaleGradMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
std::unique_ptr<framework::OpDesc> Apply() const override {
auto *grad_op = new framework::OpDesc();
grad_op->SetType("scale");
grad_op->SetInput("X", OutputGrad("Out"));
grad_op->SetOutput("Out", InputGrad("X"));
grad_op->SetAttr("scale", GetAttr("scale"));
return std::unique_ptr<framework::OpDesc>(grad_op);
}
};
```
### 定义Operator类
下面的点实现了MulOp的定义:
下面实现了MulOp的定义:
```cpp
class MulOp : public framework::OperatorWithKernel {
......@@ -334,3 +353,83 @@ ctest -R test_mul_op
- 注册Op时的类型名,需要和该Op的名字一样。即不允许在`A_op.cc`里面,注册`REGISTER_OPERATOR(B, ...)`等,这将会导致单元测试出错。
- 如果Op没有实现CUDA Kernel,请不要创建空的`*_op.cu`,这将会导致单元测试出错。
- 如果多个Op依赖一些共用的函数,可以创建非`*_op.*`格式的文件来存放,如`gather.h`文件。
### PADDLE_ENFORCE使用注意
实现Op时检查数据的合法性需要使用PADDLE_ENFORCE以及PADDLE_ENFORCE_EQ等宏定义,基本格式如下:
```
PADDLE_ENFORCE(表达式, 错误提示信息)
PADDLE_ENFORCE_EQ(比较对象A, 比较对象B, 错误提示信息)
```
如果表达式为真,或者比较对象A=B,则检查通过,否则会终止程序运行,向用户反馈相应的错误提示信息。
为了确保提示友好易懂,开发者需要注意其使用方法。
#### 总体原则
任何使用了PADDLE_ENFORCE与PADDLE_ENFORCE_**检查的地方,必须有详略得当的备注解释!**错误提示信息**不能为空!
#### 提示信息书写标准
1. [required] 哪里错了?为什么错了?
- 例如:`ValueError: Mismatched label shape`
2. [optional] 期望的输入是什么样的?实际的输入是怎样的?
- 例如:`Expected labels dimension=1. Received 4.`
3. [optional] 能否给出修改意见?
- 例如:`Suggested Fix:If your classifier expects one-hot encoding label,check your n_classes argument to the estimatorand/or the shape of your label.Otherwise, check the shape of your label.`
如果并非必要或者简洁的描述即可表达清楚以上要点,根据情况书写亦可。
##### FAQ 典型问题
1. 无报错信息或报错信息过于简单,不能给用户提供有效的提示!
问题示例1 :未写提示信息
```
PADDLE_ENFORCE(ctx->HasInput("X"), "");
```
问题示例2 :提示信息过于简单
```
PADDLE_ENFORCE(i != nullptr, "I must be set"); // I是什么?
```
2. 在报错信息中使用开发人员定义的变量缩写,不易理解!
问题示例:
```
PADDLE_ENFORCE(forward_pd != nullptr,
"Fail to find eltwise_fwd_pd in device context"); //eltwise_fwd_pd用户可能看不懂
```
3. OP内部调用非法接口:Op内部如果出现Output = ShareDataWith(Input)
问题示例:
```cpp
auto *out = ctx.Output<framework::LoDTensor>("Out");
auto *in = ctx.Input<framework::LoDTensor>("X");
out->ShareDataWith(*in);
```
Op内部如果出现Output = ShareDataWith(Input),相当于operator图的中有一条隐藏边,连接了Input和Output,这条边无法在图分析中表达,引发基于图优化的错误。
4. OP实现的性能实践
调用了eigen的broadcast, chop等操作,性能会比手写cuda kernel差几倍以上。此时cpu的实现可以复用eigen,gpu实现可以实现cuda kernel.
#### OP InferShape检查提示信息特别说明
- 检查输入输出变量,请统一遵循以下格式
`Input(变量名) of OP名 operator should not be null.`
正确示例:
```
PADDLE_ENFORCE(ctx->HasInput("Input"),
"Input(Input) of LSTMP operator should not be null.");
```
- 反向Op的输入输出检查,要写明反向Op的名字
正确示例:
```
PADDLE_ENFORCE(ctx->HasInput("X"),
"Input(X) of LoDResetGrad opreator should not be null.");
```
......@@ -104,7 +104,7 @@ std::map<ir::Node *, std::unordered_set<ir::Node *>> BuildOperationAdjList(
for (auto &adj_n : var->inputs) {
PADDLE_ENFORCE(adj_n->NodeType() == ir::Node::Type::kOperation);
adj_list[n].insert(adj_n);
VLOG(3) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
VLOG(4) << "adj " << adj_n->Name() << reinterpret_cast<void *>(adj_n)
<< " -> " << n->Name() << reinterpret_cast<void *>(n)
<< " via " << var->Name() << reinterpret_cast<void *>(var);
}
......
......@@ -22,7 +22,7 @@ function (inference_analysis_test TARGET)
if(WITH_TESTING)
set(options "")
set(oneValueArgs "")
set(multiValueArgs SRCS)
set(multiValueArgs SRCS EXTRA_DEPS)
cmake_parse_arguments(analysis_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
set(mem_opt "")
......@@ -31,22 +31,43 @@ function (inference_analysis_test TARGET)
endif()
cc_test(${TARGET}
SRCS "${analysis_test_SRCS}"
DEPS analysis graph fc_fuse_pass graph_viz_pass infer_clean_graph_pass graph_pattern_detecter pass
DEPS analysis graph fc_fuse_pass graph_viz_pass infer_clean_graph_pass graph_pattern_detecter pass ${analysis_test_EXTRA_DEPS}
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model ${mem_opt})
set_tests_properties(${TARGET} PROPERTIES DEPENDS test_word2vec)
endif(WITH_TESTING)
endfunction(inference_analysis_test)
cc_test(test_analyzer SRCS analyzer_tester.cc DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis
# ir
fc_fuse_pass
graph_viz_pass
infer_clean_graph_pass
graph_pattern_detecter
pass
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model)
#set_tests_properties(test_analyzer PROPERTIES DEPENDS test_word2vec)
#inference_api_test(test_analyzer SRC analyzer_tester.cc ARGS test_word2vec)
set(DITU_RNN_MODEL_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fmodel.tar.gz")
set(DITU_RNN_DATA_URL "http://paddle-inference-dist.bj.bcebos.com/ditu_rnn_fluid%2Fdata.txt.tar.gz")
set(DITU_INSTALL_DIR "${THIRD_PARTY_PATH}/install/ditu_rnn" CACHE PATH "Ditu RNN model and data root." FORCE)
set(DITU_RNN_MODEL ${DITU_INSTALL_DIR}/model)
set(DITU_RNN_DATA ${DITU_INSTALL_DIR}/data.txt)
function (inference_download_and_uncompress target url gz_filename)
message(STATUS "Download inference test stuff ${gz_filename} from ${url}")
execute_process(COMMAND bash -c "mkdir -p ${DITU_INSTALL_DIR}")
execute_process(COMMAND bash -c "cd ${DITU_INSTALL_DIR} && wget -q ${url}")
execute_process(COMMAND bash -c "cd ${DITU_INSTALL_DIR} && tar xzf ${gz_filename}")
message(STATUS "finish downloading ${gz_filename}")
endfunction(inference_download_and_uncompress)
if (NOT EXISTS ${DITU_INSTALL_DIR})
inference_download_and_uncompress(ditu_rnn_model ${DITU_RNN_MODEL_URL} "ditu_rnn_fluid%2Fmodel.tar.gz")
inference_download_and_uncompress(ditu_rnn_data ${DITU_RNN_DATA_URL} "ditu_rnn_fluid%2Fdata.txt.tar.gz")
endif()
inference_analysis_test(test_analyzer SRCS analyzer_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis
# ir
fc_fuse_pass
graph_viz_pass
infer_clean_graph_pass
graph_pattern_detecter
infer_clean_graph_pass
pass
ARGS --inference_model_dir=${PYTHON_TESTS_DIR}/book/word2vec.inference.model
--infer_ditu_rnn_model=${DITU_INSTALL_DIR}/model
--infer_ditu_rnn_data=${DITU_INSTALL_DIR}/data.txt)
inference_analysis_test(test_data_flow_graph SRCS data_flow_graph_tester.cc)
inference_analysis_test(test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc)
......
......@@ -23,8 +23,6 @@
#include "paddle/fluid/inference/analysis/tensorrt_subgraph_node_mark_pass.h"
#include "paddle/fluid/inference/analysis/tensorrt_subgraph_pass.h"
namespace paddle {
DEFINE_bool(IA_enable_tensorrt_subgraph_engine, false,
"Enable subgraph to TensorRT engine for acceleration");
......@@ -35,6 +33,7 @@ DEFINE_string(IA_graphviz_log_root, "./",
DEFINE_string(IA_output_storage_path, "", "optimized model output path");
namespace paddle {
namespace inference {
namespace analysis {
......@@ -72,7 +71,8 @@ class DfgPassManagerImpl final : public DfgPassManager {
if (FLAGS_IA_enable_tensorrt_subgraph_engine) {
auto trt_teller = [&](const Node* node) {
std::unordered_set<std::string> teller_set(
{"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax"});
{"elementwise_add", "mul", "conv2d", "pool2d", "relu", "softmax",
"depthwise_conv2d", "batch_norm"});
if (!node->IsFunction()) return false;
const auto* func = static_cast<const Function*>(node);
......
......@@ -39,8 +39,6 @@ limitations under the License. */
#include "paddle/fluid/inference/analysis/pass.h"
#include "paddle/fluid/inference/analysis/pass_manager.h"
namespace paddle {
// TODO(Superjomn) add a definition flag like PADDLE_WITH_TENSORRT and hide this
// flag if not available.
DECLARE_bool(IA_enable_tensorrt_subgraph_engine);
......@@ -48,6 +46,7 @@ DECLARE_string(IA_graphviz_log_root);
DECLARE_string(IA_output_storage_path);
DECLARE_bool(IA_enable_ir);
namespace paddle {
namespace inference {
namespace analysis {
......
......@@ -13,11 +13,17 @@
// limitations under the License.
#include "paddle/fluid/inference/analysis/analyzer.h"
#include <google/protobuf/text_format.h>
#include <gtest/gtest.h>
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/inference/analysis/ut_helper.h"
#include "paddle/fluid/inference/api/helper.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
DEFINE_string(infer_ditu_rnn_model, "", "model path for ditu RNN");
DEFINE_string(infer_ditu_rnn_data, "", "data path for ditu RNN");
namespace paddle {
namespace inference {
namespace analysis {
......@@ -38,7 +44,7 @@ TEST(Analyzer, analysis_with_tensorrt) {
analyser.Run(&argument);
}
void TestWord2vecPrediction(const std::string& model_path) {
void TestWord2vecPrediction(const std::string &model_path) {
NativeConfig config;
config.model_dir = model_path;
config.use_gpu = false;
......@@ -69,12 +75,245 @@ void TestWord2vecPrediction(const std::string& model_path) {
// The outputs' buffers are in CPU memory.
for (size_t i = 0; i < std::min(5UL, num_elements); i++) {
LOG(INFO) << "data: "
<< static_cast<float*>(outputs.front().data.data())[i];
PADDLE_ENFORCE(static_cast<float*>(outputs.front().data.data())[i],
<< static_cast<float *>(outputs.front().data.data())[i];
PADDLE_ENFORCE(static_cast<float *>(outputs.front().data.data())[i],
result[i]);
}
}
namespace {
struct DataRecord {
std::vector<std::vector<std::vector<float>>> link_step_data_all;
std::vector<std::vector<float>> week_data_all, minute_data_all;
std::vector<size_t> lod1, lod2, lod3;
std::vector<std::vector<float>> rnn_link_data, rnn_week_datas,
rnn_minute_datas;
size_t batch_iter{0};
size_t batch_size{1};
DataRecord() = default;
DataRecord(const std::string &path, int batch_size = 1)
: batch_size(batch_size) {
Load(path);
}
DataRecord NextBatch() {
DataRecord data;
size_t batch_end = batch_iter + batch_size;
// NOTE skip the final batch, if no enough data is provided.
if (batch_end <= link_step_data_all.size()) {
data.link_step_data_all.assign(link_step_data_all.begin() + batch_iter,
link_step_data_all.begin() + batch_end);
data.week_data_all.assign(week_data_all.begin() + batch_iter,
week_data_all.begin() + batch_end);
data.minute_data_all.assign(minute_data_all.begin() + batch_iter,
minute_data_all.begin() + batch_end);
// Prepare LoDs
data.lod1.push_back(0);
data.lod2.push_back(0);
data.lod3.push_back(0);
CHECK(!data.link_step_data_all.empty()) << "empty";
CHECK(!data.week_data_all.empty());
CHECK(!data.minute_data_all.empty());
CHECK_EQ(data.link_step_data_all.size(), data.week_data_all.size());
CHECK_EQ(data.minute_data_all.size(), data.link_step_data_all.size());
for (size_t j = 0; j < data.link_step_data_all.size(); j++) {
for (const auto &d : data.link_step_data_all[j]) {
data.rnn_link_data.push_back(d);
}
data.rnn_week_datas.push_back(data.week_data_all[j]);
data.rnn_minute_datas.push_back(data.minute_data_all[j]);
// calculate lod
data.lod1.push_back(data.lod1.back() +
data.link_step_data_all[j].size());
data.lod3.push_back(data.lod3.back() + 1);
for (size_t i = 1; i < data.link_step_data_all[j].size() + 1; i++) {
data.lod2.push_back(data.lod2.back() +
data.link_step_data_all[j].size());
}
}
}
batch_iter += batch_size;
return data;
}
void Load(const std::string &path) {
std::ifstream file(path);
std::string line;
int num_lines = 0;
while (std::getline(file, line)) {
num_lines++;
std::vector<std::string> data;
split(line, ':', &data);
std::vector<std::vector<float>> link_step_data;
std::vector<std::string> link_datas;
split(data[0], '|', &link_datas);
for (auto &step_data : link_datas) {
std::vector<float> tmp;
split_to_float(step_data, ',', &tmp);
link_step_data.push_back(tmp);
}
// load week data
std::vector<float> week_data;
split_to_float(data[2], ',', &week_data);
// load minute data
std::vector<float> minute_data;
split_to_float(data[1], ',', &minute_data);
link_step_data_all.push_back(std::move(link_step_data));
week_data_all.push_back(std::move(week_data));
minute_data_all.push_back(std::move(minute_data));
}
}
};
void PrepareInputs(std::vector<PaddleTensor> *input_slots, DataRecord *data,
int batch_size) {
// DataRecord data(FLAGS_datapath, batch_size);
PaddleTensor lod_attention_tensor, init_zero_tensor, lod_tensor_tensor,
week_tensor, minute_tensor;
lod_attention_tensor.name = "data_lod_attention";
init_zero_tensor.name = "cell_init";
lod_tensor_tensor.name = "data";
week_tensor.name = "week";
minute_tensor.name = "minute";
auto one_batch = data->NextBatch();
// clang-format off
std::vector<int> rnn_link_data_shape
({static_cast<int>(one_batch.rnn_link_data.size()), static_cast<int>(one_batch.rnn_link_data.front().size())});
lod_attention_tensor.shape.assign({1, 2});
lod_attention_tensor.lod.assign({one_batch.lod1, one_batch.lod2});
init_zero_tensor.shape.assign({batch_size, 15});
init_zero_tensor.lod.assign({one_batch.lod3});
lod_tensor_tensor.shape = rnn_link_data_shape;
lod_tensor_tensor.lod.assign({one_batch.lod1});
week_tensor.shape.assign({(int) one_batch.rnn_week_datas.size(), (int) one_batch.rnn_week_datas.front().size()});
week_tensor.lod.assign({one_batch.lod3});
minute_tensor.shape.assign({(int) one_batch.rnn_minute_datas.size(),
(int) one_batch.rnn_minute_datas.front().size()});
minute_tensor.lod.assign({one_batch.lod3});
// assign data
TensorAssignData(&lod_attention_tensor, std::vector<std::vector<float>>({{0, 0}}));
std::vector<float> tmp_zeros(batch_size * 15, 0.);
TensorAssignData(&init_zero_tensor, {tmp_zeros});
TensorAssignData(&lod_tensor_tensor, one_batch.rnn_link_data);
TensorAssignData(&week_tensor, one_batch.rnn_week_datas);
TensorAssignData(&minute_tensor, one_batch.rnn_minute_datas);
// clang-format on
// Set inputs.
auto init_zero_tensor1 = init_zero_tensor;
init_zero_tensor1.name = "hidden_init";
input_slots->assign({week_tensor, init_zero_tensor, minute_tensor,
init_zero_tensor1, lod_attention_tensor,
lod_tensor_tensor});
for (auto &tensor : *input_slots) {
tensor.dtype = PaddleDType::FLOAT32;
}
}
std::string DescribeTensor(const PaddleTensor &tensor) {
std::stringstream os;
os << "Tensor [" << tensor.name << "]\n";
os << " - type: ";
switch (tensor.dtype) {
case PaddleDType::FLOAT32:
os << "float32";
break;
case PaddleDType::INT64:
os << "int64";
break;
default:
os << "unset";
}
os << '\n';
os << " - shape: " << to_string(tensor.shape) << '\n';
os << " - lod: ";
for (auto &l : tensor.lod) {
os << to_string(l) << "; ";
}
os << "\n";
os << " - data: ";
// clang-format off
int dim = std::accumulate(tensor.shape.begin(),
tensor.shape.end(),
1,
[](int a, int b) { return a * b; }); // clang-format on
for (size_t i = 0; i < dim; i++) {
os << static_cast<float *>(tensor.data.data())[i] << " ";
}
os << '\n';
return os.str();
}
} // namespace
const float ditu_rnn_target_data[] = {
104.711, 11.2431, 1.35422, 0, 0, 0, 0, 0,
27.7039, 1.41486, 7.09526, 0, 0, 0, 0, 0,
7.6481, 6.5324, 56.383, 2.88018, 8.92918, 132.007, 4.27429, 2.02934,
14.1727, 10.7461, 25.0616, 16.0197, 14.4163, 16.9199, 6.75517, 0,
80.0249, 4.77739, 0, 0, 0, 0, 0, 0,
47.5643, 2.67029, 8.76252, 0, 0, 0, 0, 0,
51.8822, 4.4411, 0, 0, 0, 0, 0, 0,
10.7286, 12.0595, 10.6672, 0, 0, 0, 0, 0,
93.5771, 3.84641, 0, 0, 0, 0, 0, 0,
169.426, 0, 0, 0, 0, 0, 0, 0};
// Test with a really complicate model.
void TestDituRNNPrediction(const std::string &model_path,
const std::string &data_path, int batch_size,
bool use_analysis, bool activate_ir,
int num_times = 1) {
FLAGS_IA_enable_ir = activate_ir;
FLAGS_IA_enable_tensorrt_subgraph_engine = false;
FLAGS_IA_output_storage_path = "./analysis.out";
std::string model_out;
if (use_analysis) {
Argument argument(model_path);
argument.model_output_store_path.reset(new std::string("./analysis.out"));
Analyzer analyzer;
analyzer.Run(&argument);
// Should get the transformed model stored to ./analysis.out
model_out = "./analysis.out";
ASSERT_TRUE(PathExists(model_out));
} else {
model_out = FLAGS_infer_ditu_rnn_model;
}
NativeConfig config;
config.prog_file = model_out + "/__model__";
config.param_file = model_out + "/param";
config.use_gpu = false;
config.device = 0;
config.specify_input_name = true;
auto predictor =
CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
std::vector<PaddleTensor> input_slots;
DataRecord data(data_path, batch_size);
// Prepare inputs.
PrepareInputs(&input_slots, &data, batch_size);
std::vector<PaddleTensor> outputs;
Timer timer;
timer.tic();
for (int i = 0; i < num_times; i++) {
predictor->Run(input_slots, &outputs);
}
LOG(INFO) << "time/batch: " << timer.toc() / num_times;
for (auto &out : outputs) {
size_t size = std::accumulate(out.shape.begin(), out.shape.end(), 1,
[](int a, int b) { return a * b; });
float *data = static_cast<float *>(out.data.data());
for (int i = 0;
i < std::min(sizeof(ditu_rnn_target_data) / sizeof(float), size);
i++) {
EXPECT_NEAR(data[i], ditu_rnn_target_data[i], 1e-3);
}
}
}
// Turn on the IR pass supportion, run a real inference and check the result.
TEST(Analyzer, SupportIRPass) {
FLAGS_IA_enable_ir = true;
......@@ -94,6 +333,27 @@ TEST(Analyzer, SupportIRPass) {
TestWord2vecPrediction("./analysis.out");
}
// Directly infer with the original model.
TEST(Analyzer, DituRNN_without_analysis) {
TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
10, false, false);
}
// Inference with the original model with the analysis turned on, the analysis
// module will transform the program to a data flow graph.
TEST(Analyzer, DituRNN_with_analysis) {
LOG(INFO) << "ditu rnn with analysis";
TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
10, true, false, 1);
}
// Inference with analysis and IR. The IR module will fuse some large kernels.
TEST(Analyzer, DituRNN_with_analysis_with_IR) {
LOG(INFO) << "ditu rnn with analysis and IR fuse";
TestDituRNNPrediction(FLAGS_infer_ditu_rnn_model, FLAGS_infer_ditu_rnn_data,
10, true, true, 1);
}
} // namespace analysis
} // namespace inference
} // namespace paddle
......
......@@ -97,6 +97,7 @@ void DataFlowGraph::Build(const framework::proto::ProgramDesc &prog) {
auto *in = nodes.GetMutable(var2id.at(in_var.arguments(k)));
in->outlinks.push_back(o);
o->inlinks.push_back(in);
unique_written_vars.insert(in);
}
}
for (int j = 0; j < op.outputs_size(); j++) {
......@@ -117,7 +118,6 @@ void DataFlowGraph::Build(const framework::proto::ProgramDesc &prog) {
}
out->inlinks.push_back(o);
o->outlinks.push_back(out);
unique_written_vars.insert(out);
}
}
}
......
......@@ -23,9 +23,6 @@
namespace paddle {
namespace inference {
DEFINE_int32(tensorrt_max_batchsize, 1, "TensorRT maximum batch size");
DEFINE_int32(tensorrt_workspace_size, 2048, "TensorRT workspace size");
namespace analysis {
using framework::proto::ProgramDesc;
......@@ -199,8 +196,6 @@ void CreateTrtEngineOp(Node *node, const DataFlowGraph &graph,
// Set attrs
SetAttr(desc.Proto(), "subgraph", block->SerializeAsString());
SetAttr(desc.Proto(), "engine_uniq_key", "trt-" + std::to_string(counter++));
SetAttr(desc.Proto(), "max_batch", FLAGS_tensorrt_max_batchsize);
SetAttr(desc.Proto(), "max_workspace", FLAGS_tensorrt_workspace_size);
SetAttr(desc.Proto(), "parameters", ExtractParameters(graph.nodes.nodes()));
SetAttr(desc.Proto(), "output_name_mapping", output_mapping);
node->SetPbMsg(desc.Proto()->SerializeAsString());
......
......@@ -27,9 +27,6 @@
namespace paddle {
namespace inference {
DECLARE_int32(tensorrt_max_batchsize);
DECLARE_int32(tensorrt_workspace_size);
namespace analysis {
class DataFlowGraphToFluidPass final : public DataFlowGraphPass {
public:
......
......@@ -18,7 +18,10 @@ if(APPLE)
endif(APPLE)
set(inference_deps paddle_inference_api paddle_fluid_api)
set(inference_deps paddle_inference_api paddle_fluid_api analysis pass ir_pass_manager
graph_viz_pass fc_fuse_pass
infer_clean_graph_pass
)
if(WITH_GPU AND TENSORRT_FOUND)
set(inference_deps ${inference_deps} paddle_inference_tensorrt_subgraph_engine)
......@@ -62,7 +65,7 @@ endif()
if (WITH_ANAKIN AND WITH_GPU) # only needed in CI
# compile the libinference_anakin_api.a and anakin.so.
cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber)
cc_library(inference_anakin_api SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber mklml)
cc_library(inference_anakin_api_shared SHARED SRCS api.cc api_anakin_engine.cc DEPS anakin_shared anakin_saber)
function(anakin_target target_name)
target_compile_options(${target_name} BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
......@@ -70,9 +73,12 @@ if (WITH_ANAKIN AND WITH_GPU) # only needed in CI
anakin_target(inference_anakin_api)
anakin_target(inference_anakin_api_shared)
if (WITH_TESTING)
cc_test(inference_anakin_test SRCS api_anakin_engine_tester.cc
cc_test(api_anakin_engine_tester SRCS api_anakin_engine_tester.cc
ARGS --model=${ANAKIN_SOURCE_DIR}/mobilenet_v2.anakin.bin
DEPS inference_anakin_api dynload_cuda SERIAL)
target_compile_options(inference_anakin_test BEFORE PUBLIC ${ANAKIN_COMPILE_EXTRA_FLAGS})
DEPS inference_anakin_api_shared dynload_cuda SERIAL)
cc_test(api_anakin_engine_rnn_tester SRCS api_anakin_engine_rnn_tester.cc
ARGS --model=${ANAKIN_SOURCE_DIR}/anakin_test%2Fditu_rnn.anakin2.model.bin
--datapath=${ANAKIN_SOURCE_DIR}/anakin_test%2Fditu_rnn_data.txt
DEPS inference_anakin_api_shared dynload_cuda SERIAL)
endif(WITH_TESTING)
endif()
/* 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.
......
......@@ -13,9 +13,22 @@
// limitations under the License.
#include "paddle/fluid/inference/api/api_anakin_engine.h"
#ifdef PADDLE_WITH_CUDA
#include <cuda.h>
#endif
#include <mkl_service.h>
#include <omp.h>
#include <map>
#include <string>
#include <utility>
#include <vector>
#include "framework/core/net/net.h"
#include "framework/operators/ops.h"
#include "saber/funcs/timer.h"
namespace paddle {
template <typename Target>
......@@ -23,16 +36,24 @@ PaddleInferenceAnakinPredictor<Target>::PaddleInferenceAnakinPredictor(
const AnakinConfig &config) {
CHECK(Init(config));
}
template <>
PaddleInferenceAnakinPredictor<anakin::X86>::PaddleInferenceAnakinPredictor(
const AnakinConfig &config) {
omp_set_dynamic(0);
omp_set_num_threads(1);
mkl_set_num_threads(1);
CHECK(Init(config));
}
template <typename Target>
bool PaddleInferenceAnakinPredictor<Target>::Init(const AnakinConfig &config) {
if (!(graph_.load(config.model_file))) {
LOG(FATAL) << "fail to load graph from " << config.model_file;
VLOG(3) << "fail to load graph from " << config.model_file;
return false;
}
auto inputs = graph_.get_ins();
for (auto &input_str : inputs) {
graph_.ResetBatchSize(input_str, config.max_batch_size);
max_batch_size_ = config.max_batch_size;
}
// optimization for graph
if (!(graph_.Optimize())) {
......@@ -52,15 +73,15 @@ bool PaddleInferenceAnakinPredictor<Target>::Run(
std::vector<PaddleTensor> *output_data, int batch_size) {
for (const auto &input : inputs) {
if (input.dtype != PaddleDType::FLOAT32) {
LOG(ERROR) << "Only support float type inputs. " << input.name
<< "'s type is not float";
VLOG(3) << "Only support float type inputs. " << input.name
<< "'s type is not float";
return false;
}
auto d_tensor_in_p = executor_p_->get_in(input.name);
auto net_shape = d_tensor_in_p->valid_shape();
auto net_shape = d_tensor_in_p->shape();
if (net_shape.size() != input.shape.size()) {
LOG(ERROR) << " input " << input.name
<< "'s shape size should be equal to that of net";
VLOG(3) << " input " << input.name
<< "'s shape size should be equal to that of net";
return false;
}
int sum = 1;
......@@ -79,21 +100,45 @@ bool PaddleInferenceAnakinPredictor<Target>::Run(
}
d_tensor_in_p->reshape(tmp_shape);
if (input.lod.size() > 0) {
if (input.lod.size() > 1) {
VLOG(3) << " input lod first dim should <=1, but you set "
<< input.lod.size();
return false;
}
std::vector<int> offset(input.lod[0].begin(), input.lod[0].end());
d_tensor_in_p->set_seq_offset(offset);
VLOG(3) << "offset.size(): " << offset.size();
for (int i = 0; i < offset.size(); i++) {
VLOG(3) << offset[i];
}
}
float *d_data_p = d_tensor_in_p->mutable_data();
if (cudaMemcpy(d_data_p, static_cast<float *>(input.data.data()),
d_tensor_in_p->valid_size() * sizeof(float),
cudaMemcpyHostToDevice) != 0) {
LOG(ERROR) << "copy data from CPU to GPU error";
return false;
#ifdef PADDLE_WITH_CUDA
if (std::is_same<anakin::NV, Target>::value) {
if (cudaMemcpy(d_data_p, static_cast<float *>(input.data.data()),
d_tensor_in_p->valid_size() * sizeof(float),
cudaMemcpyHostToDevice) != 0) {
VLOG(3) << "copy data from CPU to GPU error";
return false;
}
}
#endif
if (std::is_same<anakin::X86, Target>::value) {
memcpy(d_data_p, static_cast<float *>(input.data.data()),
d_tensor_in_p->valid_size() * sizeof(float));
}
cudaStreamSynchronize(NULL);
}
#ifdef PADDLE_WITH_CUDA
cudaDeviceSynchronize();
executor_p_->prediction();
cudaDeviceSynchronize();
#endif
if (output_data->empty()) {
LOG(ERROR) << "At least one output should be set with tensors' names.";
VLOG(3) << "At least one output should be set with tensors' names.";
return false;
}
for (auto &output : *output_data) {
......@@ -102,14 +147,22 @@ bool PaddleInferenceAnakinPredictor<Target>::Run(
if (output.data.length() < tensor->valid_size() * sizeof(float)) {
output.data.Resize(tensor->valid_size() * sizeof(float));
}
// Copy data from GPU -> CPU
if (cudaMemcpy(output.data.data(), tensor->mutable_data(),
tensor->valid_size() * sizeof(float),
cudaMemcpyDeviceToHost) != 0) {
LOG(ERROR) << "copy data from GPU to CPU error";
return false;
#if PADDLE_WITH_CUDA
if (std::is_same<anakin::NV, Target>::value) {
// Copy data from GPU -> CPU
if (cudaMemcpy(output.data.data(), tensor->mutable_data(),
tensor->valid_size() * sizeof(float),
cudaMemcpyDeviceToHost) != 0) {
VLOG(3) << "copy data from GPU to CPU error";
return false;
}
}
#endif
if (std::is_same<anakin::X86, Target>::value) {
memcpy(output.data.data(), tensor->mutable_data(),
tensor->valid_size() * sizeof(float));
}
cudaStreamSynchronize(NULL);
}
return true;
}
......@@ -132,7 +185,7 @@ PaddleInferenceAnakinPredictor<Target>::Clone() {
auto anakin_predictor_p =
dynamic_cast<PaddleInferenceAnakinPredictor<Target> *>(cls.get());
if (!anakin_predictor_p) {
LOG(ERROR) << "fail to call Init";
VLOG(3) << "fail to call Init";
return nullptr;
}
anakin_predictor_p->get_executer().init(graph_);
......@@ -162,6 +215,44 @@ std::unique_ptr<PaddlePredictor> CreatePaddlePredictor<
VLOG(3) << "Anakin Predictor create on unknown platform.";
return nullptr;
}
};
}
#ifdef PADDLE_ANAKIN_ENABLE_OP_TIMER
template <typename Target>
using executor_t =
anakin::Net<Target, anakin::saber::AK_FLOAT, anakin::Precision::FP32>;
template <typename Target>
void DisplayOpTimer(executor_t<Target> *net_executor, int epoch) {
std::vector<float> op_time = net_executor->get_op_time();
auto exec_funcs = net_executor->get_exec_funcs();
auto op_param = net_executor->get_op_param();
for (int i = 0; i < op_time.size(); i++) {
LOG(INFO) << "name: " << exec_funcs[i].name
<< " op_type: " << exec_funcs[i].op_name
<< " op_param: " << op_param[i] << " time " << op_time[i] / epoch;
}
std::map<std::string, float> op_map;
for (int i = 0; i < op_time.size(); i++) {
auto it = op_map.find(op_param[i]);
if (it != op_map.end())
op_map[op_param[i]] += op_time[i];
else
op_map.insert(std::pair<std::string, float>(op_param[i], op_time[i]));
}
for (auto it = op_map.begin(); it != op_map.end(); ++it) {
LOG(INFO) << it->first << " " << (it->second) / epoch << " ms";
}
}
#endif
template <typename Target>
PaddleInferenceAnakinPredictor<Target>::~PaddleInferenceAnakinPredictor() {
#ifdef PADDLE_ANAKIN_ENABLE_OP_TIMER
DisplayOpTimer<Target>(executor_p_, max_batch_size_);
#endif
delete executor_p_;
executor_p_ = nullptr;
}
} // namespace paddle
......@@ -47,10 +47,7 @@ class PaddleInferenceAnakinPredictor : public PaddlePredictor {
anakin::Net<Target, anakin::saber::AK_FLOAT, anakin::Precision::FP32>&
get_executer();
~PaddleInferenceAnakinPredictor() override {
delete executor_p_;
executor_p_ = nullptr;
};
~PaddleInferenceAnakinPredictor() override;
private:
bool Init(const AnakinConfig& config);
......@@ -60,6 +57,7 @@ class PaddleInferenceAnakinPredictor : public PaddlePredictor {
anakin::Net<Target, anakin::saber::AK_FLOAT, anakin::Precision::FP32>*
executor_p_{nullptr};
AnakinConfig config_;
int max_batch_size_{0};
};
} // namespace paddle
/* 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 <gflags/gflags.h>
#include <sys/time.h>
#include <time.h>
#include <algorithm>
#include <fstream>
#include <iostream>
#include <thread> // NOLINT
#include <vector>
#include "framework/core/net/net.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
DEFINE_string(model, "", "Directory of the inference model.");
DEFINE_string(datapath, "", "Path of the dataset.");
DEFINE_int32(batch_size, 1, "batch size.");
DEFINE_int32(repeat, 1, "Running the inference program repeat times.");
// Timer for timer
class Timer {
public:
double start;
double startu;
void tic() {
struct timeval tp;
gettimeofday(&tp, NULL);
start = tp.tv_sec;
startu = tp.tv_usec;
}
double toc() {
struct timeval tp;
gettimeofday(&tp, NULL);
double used_time_ms =
(tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0;
return used_time_ms;
}
};
std::vector<std::string> string_split(std::string in_str,
std::string delimiter) {
std::vector<std::string> seq;
int found = in_str.find(delimiter);
int pre_found = -1;
while (found != std::string::npos) {
if (pre_found == -1) {
seq.push_back(in_str.substr(0, found));
} else {
seq.push_back(in_str.substr(pre_found + delimiter.length(),
found - delimiter.length() - pre_found));
}
pre_found = found;
found = in_str.find(delimiter, pre_found + delimiter.length());
}
seq.push_back(
in_str.substr(pre_found + 1, in_str.length() - (pre_found + 1)));
return seq;
}
std::vector<std::string> string_split(
std::string in_str, std::vector<std::string>& delimiter) { // NOLINT
std::vector<std::string> in;
std::vector<std::string> out;
out.push_back(in_str);
for (auto del : delimiter) {
in = out;
out.clear();
for (auto s : in) {
auto out_s = string_split(s, del);
for (auto o : out_s) {
out.push_back(o);
}
}
}
return out;
}
class Data {
public:
Data(std::string file_name, int batch_size)
: _batch_size(batch_size), _total_length(0) {
_file.open(file_name);
_file.seekg(_file.end);
_total_length = _file.tellg();
_file.seekg(_file.beg);
}
void get_batch_data(std::vector<std::vector<float>>& fea, // NOLINT
std::vector<std::vector<float>>& week_fea, // NOLINT
std::vector<std::vector<float>>& time_fea, // NOLINT
std::vector<long unsigned int>& seq_offset); // NOLINT
private:
std::fstream _file;
int _total_length;
int _batch_size;
};
void Data::get_batch_data(
std::vector<std::vector<float>>& fea, // NOLINT
std::vector<std::vector<float>>& week_fea, // NOLINT
std::vector<std::vector<float>>& time_fea, // NOLINT
std::vector<long unsigned int>& seq_offset) { // NOLINT
int seq_num = 0;
long unsigned int cum = 0; // NOLINT
char buf[10000];
seq_offset.clear();
seq_offset.push_back(0);
fea.clear();
week_fea.clear();
time_fea.clear();
while (_file.getline(buf, 10000)) {
std::string s = buf;
std::vector<std::string> deli_vec = {":"};
std::vector<std::string> data_vec = string_split(s, deli_vec);
std::vector<std::string> seq;
seq = string_split(data_vec[0], {"|"});
for (auto link : seq) {
std::vector<std::string> data = string_split(link, ",");
std::vector<float> vec;
for (int i = 0; i < data.size(); i++) {
vec.push_back(atof(data[i].c_str()));
}
fea.push_back(vec);
}
std::vector<std::string> week_data;
std::vector<std::string> time_data;
week_data = string_split(data_vec[2], ",");
std::vector<float> vec_w;
for (int i = 0; i < week_data.size(); i++) {
vec_w.push_back(atof(week_data[i].c_str()));
}
week_fea.push_back(vec_w);
time_data = string_split(data_vec[1], ",");
std::vector<float> vec_t;
for (int i = 0; i < time_data.size(); i++) {
vec_t.push_back(atof(time_data[i].c_str()));
}
time_fea.push_back(vec_t);
cum += seq.size();
seq_offset.push_back(cum);
seq_num++;
if (seq_num >= _batch_size) {
break;
}
}
}
namespace paddle {
AnakinConfig GetConfig() {
AnakinConfig config;
// using AnakinConfig::X86 if you need to use cpu to do inference
config.target_type = AnakinConfig::X86;
config.model_file = FLAGS_model;
config.device = 0;
config.max_batch_size = 1000; // the max number of token
return config;
}
void set_tensor(std::string name, std::vector<int> shape,
std::vector<PaddleTensor>& vec) { // NOLINT
int sum = 1;
std::for_each(shape.begin(), shape.end(), [&](int n) { sum *= n; });
float* data = new float[sum];
PaddleTensor tensor;
tensor.name = name;
tensor.shape = shape;
tensor.data = PaddleBuf(data, sum);
tensor.dtype = PaddleDType::FLOAT32;
vec.push_back(tensor);
}
void single_test() {
AnakinConfig config = GetConfig();
auto predictor =
CreatePaddlePredictor<AnakinConfig, PaddleEngineKind::kAnakin>(config);
int max_batch_size = 1000;
std::string feature_file = FLAGS_datapath;
Data map_data(feature_file, FLAGS_batch_size);
std::vector<std::vector<float>> fea;
std::vector<std::vector<float>> week_fea;
std::vector<std::vector<float>> time_fea;
std::vector<long unsigned int> seq_offset; // NOLINT
paddle::PaddleTensor tensor_0, tensor_1, tensor_2;
tensor_0.name = "input_0";
tensor_1.name = "input_4";
tensor_2.name = "input_5";
PaddleTensor tensor_out;
tensor_out.name = "final_output.tmp_1_gout";
tensor_out.shape = std::vector<int>({});
tensor_out.data = PaddleBuf();
tensor_out.dtype = PaddleDType::FLOAT32;
std::vector<PaddleTensor> inputs;
std::vector<PaddleTensor> outputs(1, tensor_out);
int data_0_dim = 38;
int data_1_dim = 10;
int data_2_dim = 10;
float data_0[max_batch_size * data_0_dim]; // NOLINT
float data_1[max_batch_size * data_1_dim]; // NOLINT
float data_2[max_batch_size * data_2_dim]; // NOLINT
int count = 0;
while (true) {
if (count++ > 0) break; // only run the first batch in ci.
seq_offset.clear();
map_data.get_batch_data(fea, week_fea, time_fea, seq_offset);
if (seq_offset.size() <= 1) {
LOG(FATAL) << "seq_offset.size() <= 1, exit.";
break;
}
std::vector<std::vector<long unsigned int>> seq_offset_vec; // NOLINT
seq_offset_vec.push_back(seq_offset);
tensor_0.lod = seq_offset_vec;
int p_shape_0[] = {(int)fea.size(), 1, 1, data_0_dim}; // NOLINT
int p_shape_1[] = {(int)week_fea.size(), data_1_dim, 1, 1}; // NOLINT
int p_shape_2[] = {(int)time_fea.size(), data_2_dim, 1, 1}; // NOLINT
std::vector<int> shape_0(p_shape_0, p_shape_0 + 4);
std::vector<int> shape_1(p_shape_1, p_shape_1 + 4);
std::vector<int> shape_2(p_shape_2, p_shape_2 + 4);
tensor_0.shape = shape_0;
tensor_1.shape = shape_1;
tensor_2.shape = shape_2;
for (int i = 0; i < fea.size(); i++) {
memcpy(data_0 + i * data_0_dim, &fea[i][0], sizeof(float) * data_0_dim);
}
for (int i = 0; i < week_fea.size(); i++) {
memcpy(data_1 + i * data_1_dim, &week_fea[i][0],
sizeof(float) * data_1_dim);
}
for (int i = 0; i < time_fea.size(); i++) {
memcpy(data_2 + i * data_2_dim, &time_fea[i][0],
sizeof(float) * data_2_dim);
}
tensor_0.data =
paddle::PaddleBuf(data_0, fea.size() * sizeof(float) * data_0_dim);
tensor_1.data =
paddle::PaddleBuf(data_1, week_fea.size() * sizeof(float) * data_1_dim);
tensor_2.data =
paddle::PaddleBuf(data_2, time_fea.size() * sizeof(float) * data_2_dim);
tensor_0.dtype = paddle::PaddleDType::FLOAT32;
tensor_1.dtype = paddle::PaddleDType::FLOAT32;
tensor_2.dtype = paddle::PaddleDType::FLOAT32;
inputs.clear();
inputs.push_back(tensor_1);
inputs.push_back(tensor_2);
inputs.push_back(tensor_0);
Timer timer;
timer.tic();
for (int i = 0; i < FLAGS_repeat; i++) predictor->Run(inputs, &outputs);
LOG(INFO) << "batch_size = " << FLAGS_batch_size
<< ", repeat = " << FLAGS_repeat
<< ", sequence_length = " << seq_offset[seq_offset.size() - 1]
<< ", latency: " << timer.toc() / FLAGS_repeat << "ms";
float* data_o = static_cast<float*>(outputs[0].data.data());
VLOG(3) << "outputs[0].data.length() = " << outputs[0].data.length();
for (size_t j = 0; j < outputs[0].data.length(); ++j) {
VLOG(3) << "output[" << j << "]: " << data_o[j];
}
}
}
} // namespace paddle
int main(int argc, char** argv) {
google::ParseCommandLineFlags(&argc, &argv, true);
logger::init(argv[0]);
paddle::single_test();
/* multi-threads
std::vector<std::thread> threads;
int num = 1;
for (int i = 0; i < num; i++) {
LOG(INFO) << " thread id : " << i;
threads.emplace_back(paddle::single_test);
}
for (int i = 0; i < num; i++) {
threads[i].join();
}
threads.clear();
*/
return 0;
}
......@@ -137,8 +137,11 @@ bool NativePaddlePredictor::Run(const std::vector<PaddleTensor> &inputs,
return false;
}
for (size_t i = 0; i < feed_target_names_.size(); ++i) {
VLOG(4) << "setting " << i << "-th target";
feed_targets[feed_target_names_[i]] = &feeds[i];
if (config_.specify_input_name) {
feed_targets[inputs[i].name] = &feeds[i];
} else {
feed_targets[feed_target_names_[i]] = &feeds[i];
}
}
// get fetch variable
std::map<std::string, framework::LoDTensor *> fetch_targets;
......
......@@ -15,6 +15,7 @@
#include "paddle/fluid/inference/analysis/analyzer.h"
#include "paddle/fluid/inference/api/api_impl.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
#include "paddle/fluid/inference/utils/singleton.h"
#include "paddle/fluid/operators/tensorrt_engine_op.h"
......@@ -32,7 +33,8 @@ class TensorRTSubgraphPredictor : public NativePaddlePredictor {
bool Init(const std::shared_ptr<framework::Scope>& parent_scope) {
VLOG(3) << "Predictor::init()";
FLAGS_tensorrt_max_batch_size = config_.max_batch_size;
FLAGS_tensorrt_workspace_size = config_.workspace_size;
if (config_.use_gpu) {
place_ = paddle::platform::CUDAPlace(config_.device);
} else {
......@@ -150,3 +152,12 @@ CreatePaddlePredictor<TensorRTConfig, PaddleEngineKind::kAutoMixedTensorRT>(
}
} // namespace paddle
USE_TRT_CONVERTER(elementwise_add_weight);
USE_TRT_CONVERTER(mul);
USE_TRT_CONVERTER(conv2d);
USE_TRT_CONVERTER(relu);
USE_TRT_CONVERTER(fc);
USE_TRT_CONVERTER(pool2d);
USE_TRT_CONVERTER(softmax);
USE_TRT_CONVERTER(batch_norm);
// 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 <sys/time.h>
#include <algorithm>
#include <sstream>
#include <string>
#include <vector>
#include "paddle/fluid/inference/api/paddle_inference_api.h"
namespace paddle {
namespace inference {
// Timer for timer
class Timer {
public:
double start;
double startu;
void tic() {
struct timeval tp;
gettimeofday(&tp, NULL);
start = tp.tv_sec;
startu = tp.tv_usec;
}
double toc() {
struct timeval tp;
gettimeofday(&tp, NULL);
double used_time_ms =
(tp.tv_sec - start) * 1000.0 + (tp.tv_usec - startu) / 1000.0;
return used_time_ms;
}
};
void split(const std::string &str, char sep, std::vector<std::string> *pieces) {
pieces->clear();
if (str.empty()) {
return;
}
size_t pos = 0;
size_t next = str.find(sep, pos);
while (next != std::string::npos) {
pieces->push_back(str.substr(pos, next - pos));
pos = next + 1;
next = str.find(sep, pos);
}
if (!str.substr(pos).empty()) {
pieces->push_back(str.substr(pos));
}
}
void split_to_float(const std::string &str, char sep, std::vector<float> *fs) {
std::vector<std::string> pieces;
split(str, sep, &pieces);
std::transform(pieces.begin(), pieces.end(), std::back_inserter(*fs),
[](const std::string &v) { return std::stof(v); });
}
template <typename T>
std::string to_string(const std::vector<T> &vec) {
std::stringstream ss;
for (const auto &c : vec) {
ss << c << " ";
}
return ss.str();
}
template <>
std::string to_string<std::vector<float>>(
const std::vector<std::vector<float>> &vec) {
std::stringstream ss;
for (const auto &piece : vec) {
ss << to_string(piece) << "\n";
}
return ss.str();
}
template <>
std::string to_string<std::vector<std::vector<float>>>(
const std::vector<std::vector<std::vector<float>>> &vec) {
std::stringstream ss;
for (const auto &line : vec) {
for (const auto &rcd : line) {
ss << to_string(rcd) << ";\t";
}
ss << '\n';
}
return ss.str();
}
// clang-format off
void TensorAssignData(PaddleTensor *tensor, const std::vector<std::vector<float>> &data) {
// Assign buffer
int dim = std::accumulate(tensor->shape.begin(), tensor->shape.end(), 1, [](int a, int b) { return a * b; });
tensor->data.Resize(sizeof(float) * dim);
int c = 0;
for (const auto &f : data) {
for (float v : f) { static_cast<float *>(tensor->data.data())[c++] = v; }
}
}
} // namespace inference
} // namespace paddle
......@@ -45,7 +45,7 @@ class PaddleBuf {
PaddleBuf(void* data, size_t length)
: data_(data), length_(length), memory_owned_{false} {}
// Own memory.
explicit PaddleBuf(size_t length)
PaddleBuf(size_t length)
: data_(new char[length]), length_(length), memory_owned_(true) {}
// Resize to `length` bytes.
void Resize(size_t length);
......@@ -70,7 +70,7 @@ struct PaddleTensor {
std::vector<int> shape;
PaddleBuf data; // blob of data.
PaddleDType dtype;
std::vector<std::vector<uint64_t>> lod; // lod data
std::vector<std::vector<size_t>> lod; // Tensor+LoD equals LoDTensor
};
enum class PaddleEngineKind {
......@@ -120,6 +120,8 @@ struct NativeConfig : public PaddlePredictor::Config {
bool use_gpu{false};
int device{0};
float fraction_of_gpu_memory{-1.f}; // Negative to notify initialization.
// Specify the variable's name of each input.
bool specify_input_name{false};
std::string prog_file;
std::string param_file;
......@@ -137,6 +139,14 @@ struct AnakinConfig : public PaddlePredictor::Config {
struct TensorRTConfig : public NativeConfig {
// Determine whether a subgraph will be executed by TRT.
int min_subgraph_size{1};
// While TensorRT allows an engine optimized for a given max batch size
// to run at any smaller size, the performance for those smaller
// sizes may not be as well-optimized. Therefore, Max batch is best
// equivalent to the runtime batch size.
int max_batch_size{1};
// For workspace_size, refer it from here:
// https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#troubleshooting
int workspace_size{1 << 30};
};
// A factory to help create different predictors.
......
......@@ -135,7 +135,7 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput("Variance",
"The global variance (for training) "
"or estimated Variance (for testing)");
AddOutput("Y", "result after normalization").Reuse("X");
AddOutput("Y", "result after normalization");
AddOutput("MeanOut",
"Share memory with Mean. "
"Store the global mean when training")
......
......@@ -54,9 +54,9 @@ class MulOp : public framework::OperatorWithKernel {
auto x_mat_dims = framework::flatten_to_2d(x_dims, x_num_col_dims);
auto y_mat_dims = framework::flatten_to_2d(y_dims, y_num_col_dims);
PADDLE_ENFORCE_EQ(
x_mat_dims[1], y_mat_dims[0],
"First matrix's width must be equal with second matrix's height.");
PADDLE_ENFORCE_EQ(x_mat_dims[1], y_mat_dims[0],
"First matrix's width must be equal with second matrix's "
"height. %s, %s");
std::vector<int64_t> output_dims;
output_dims.reserve(
static_cast<size_t>(x_num_col_dims + y_dims.size() - y_num_col_dims));
......
......@@ -22,6 +22,8 @@
namespace paddle {
DEFINE_int32(tensorrt_engine_batch_size, 1, "the batch_size of TensorRT");
DEFINE_int32(tensorrt_max_batch_size, 1, "TensorRT maximum batch size");
DEFINE_int32(tensorrt_workspace_size, 16 << 20, "TensorRT workspace size");
namespace operators {
......@@ -32,8 +34,6 @@ class TensorRTEngineOpMaker : public framework::OpProtoAndCheckerMaker {
AddOutput("Ys", "A list of outputs").AsDuplicable();
AddAttr<std::string>("subgraph", "the subgraph.");
AddAttr<std::string>("engine_uniq_key", "unique key for the TRT engine.");
AddAttr<int>("max_batch", "the maximum batch size.");
AddAttr<int>("max_workspace", "the maximum batch size.");
AddComment("TensorRT engine operator.");
}
};
......
......@@ -28,6 +28,8 @@
namespace paddle {
DECLARE_int32(tensorrt_engine_batch_size);
DECLARE_int32(tensorrt_max_batch_size);
DECLARE_int32(tensorrt_workspace_size);
namespace operators {
......@@ -54,8 +56,10 @@ 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");
PADDLE_ENFORCE_EQ(shape.size(), 4UL);
return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]);
PADDLE_ENFORCE(shape.size() == 4UL || shape.size() == 2UL);
if (shape.size() == 4UL)
return nvinfer1::DimsCHW(shape[1], shape[2], shape[3]);
return nvinfer1::DimsCHW(shape[1], 1, 1);
}
} // namespace
......@@ -95,7 +99,7 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
auto input_names = context.op().Inputs("Xs");
PADDLE_ENFORCE(!input_names.empty(), "should pass more than one inputs");
PADDLE_ENFORCE_LE(FLAGS_tensorrt_engine_batch_size,
context.Attr<int>("max_batch"));
FLAGS_tensorrt_max_batch_size);
std::vector<std::string> output_maps =
context.Attr<std::vector<std::string>>("output_name_mapping");
......@@ -132,7 +136,12 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
nvinfer1::ITensor* trt_t = engine->GetITensor(output_maps[output_index]);
auto dims = trt_t->getDimensions();
// Use the output ITensor's dims to reshape the Fluid Tensor.
std::vector<int> ddim(dims.d, dims.d + dims.nbDims);
// The ITensor doesn't contain the batch size dim.
std::vector<int> ddim;
ddim.push_back(FLAGS_tensorrt_engine_batch_size);
for (int i = 0; i < dims.nbDims; i++) {
ddim.push_back(dims.d[i]);
}
auto* fluid_v = context.scope().FindVar(y);
PADDLE_ENFORCE_NOT_NULL(fluid_v, "no output variable called %s", y);
......@@ -168,8 +177,8 @@ class TensorRTEngineKernel : public framework::OpKernel<T> {
// Get the ProgramDesc and pass to convert.
framework::proto::BlockDesc block_desc;
block_desc.ParseFromString(context.Attr<std::string>("subgraph"));
int max_batch = context.Attr<int>("max_batch");
auto max_workspace = context.Attr<int>("max_workspace");
int max_batch = FLAGS_tensorrt_max_batch_size;
auto max_workspace = FLAGS_tensorrt_workspace_size;
auto params = context.Attr<std::vector<std::string>>("parameters");
std::unordered_set<std::string> parameters;
for (const auto& param : params) {
......
......@@ -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 "paddle/fluid/operators/tensorrt_engine_op.h"
#include <gtest/gtest.h>
#include "paddle/fluid/framework/block_desc.h"
#include "paddle/fluid/framework/lod_tensor.h"
......@@ -57,6 +58,8 @@ void AddTensorToBlockDesc(framework::proto::BlockDesc* block,
using inference::analysis::SetAttr;
TEST(TensorRTEngineOp, manual) {
FLAGS_tensorrt_engine_batch_size = 2;
FLAGS_tensorrt_max_batch_size = 2;
framework::ProgramDesc program;
auto* block_ = program.Proto()->add_blocks();
block_->set_idx(0);
......@@ -98,8 +101,6 @@ TEST(TensorRTEngineOp, manual) {
engine_op_desc.SetOutput("Ys", std::vector<std::string>({"z0"}));
SetAttr<std::string>(engine_op_desc.Proto(), "subgraph",
block_->SerializeAsString());
SetAttr<int>(engine_op_desc.Proto(), "max_batch", 100);
SetAttr<int>(engine_op_desc.Proto(), "max_workspace", 1 << 10);
SetAttr<std::string>(engine_op_desc.Proto(), "engine_uniq_key", "a_engine");
SetAttr<std::vector<std::string>>(engine_op_desc.Proto(), "parameters",
std::vector<std::string>({}));
......@@ -128,6 +129,8 @@ TEST(TensorRTEngineOp, manual) {
}
void Execute(int batch_size, int input_dim, int output_dim, int nlayers = 1) {
FLAGS_tensorrt_engine_batch_size = batch_size;
FLAGS_tensorrt_max_batch_size = batch_size;
framework::ProgramDesc program;
framework::Scope scope;
platform::CUDAPlace place;
......
......@@ -57,12 +57,12 @@ class WhileOp : public framework::OperatorBase {
PADDLE_ENFORCE(platform::is_cpu_place(cond.place()),
"Condition of while op must in CPU memory.");
auto ctx = executor.Prepare(*program, block->ID());
while (cond.data<bool>()[0]) {
auto &current_scope = scope.NewScope();
step_scopes->push_back(&current_scope);
executor.Run(*program, &current_scope, block->ID(),
false /*create_local_scope*/);
executor.RunPreparedContext(ctx.get(), &current_scope, false);
}
}
};
......@@ -109,6 +109,7 @@ class WhileGradOp : public framework::OperatorBase {
framework::Executor executor(dev_place);
auto *block = Attr<framework::BlockDesc *>(kStepBlock);
auto *program = block->Program();
auto ctx = executor.Prepare(*program, block->ID());
auto *step_scopes =
scope.FindVar(Input(kStepScopes))->GetMutable<StepScopeVar>();
......@@ -161,8 +162,7 @@ class WhileGradOp : public framework::OperatorBase {
}
}
}
executor.Run(*program, *cur_scope_iter, block->ID(), false);
executor.RunPreparedContext(ctx.get(), *cur_scope_iter, false);
auto &pg_names = Outputs(kXGRAD);
auto &p_names = Inputs(kX);
......
......@@ -406,6 +406,9 @@ def load_vars(executor,
attrs={'file_path': os.path.join(dirname, filename)})
executor.run(load_prog)
if main_program is None:
main_program = default_main_program()
# load slice vars on pserver, if have it.
_load_slice_up_vars(executor, dirname,
main_program._slice_vars_and_attrs)
......
......@@ -27,6 +27,7 @@ from . import utils
import random
from .. import unique_name
from functools import reduce
import warnings
__all__ = [
'fc',
......@@ -2046,7 +2047,7 @@ def batch_norm(input,
param_attr(ParamAttr): The parameter attribute for Parameter `scale`.
bias_attr(ParamAttr): The parameter attribute for Parameter `bias`.
data_layout(string, default NCHW): NCHW|NHWC
in_place(bool, Default False): Make the input and output of batch norm reuse memory.
in_place(bool, Default False): This argument is deprecated since 0.15.0.
use_mkldnn(bool, Default false): ${use_mkldnn_comment}
name(string, Default None): A name for this layer(optional). If set None, the layer
will be named automatically.
......@@ -2068,6 +2069,10 @@ def batch_norm(input,
helper = LayerHelper('batch_norm', **locals())
dtype = helper.input_dtype()
if in_place:
raise warnings.warn("The argument in_place is deprecated since 0.15.0, "
"please do not set it True.")
input_shape = input.shape
if data_layout == 'NCHW':
channel_num = input_shape[1]
......@@ -2117,7 +2122,7 @@ def batch_norm(input,
saved_mean = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
saved_variance = helper.create_tmp_variable(dtype=dtype, stop_gradient=True)
batch_norm_out = input if in_place else helper.create_tmp_variable(dtype)
batch_norm_out = helper.create_tmp_variable(dtype)
helper.append_op(
type="batch_norm",
......
......@@ -229,7 +229,7 @@ def img_conv_group(input,
use_mkldnn=use_mkldnn)
if conv_with_batchnorm[i]:
tmp = layers.batch_norm(input=tmp, act=conv_act, in_place=True)
tmp = layers.batch_norm(input=tmp, act=conv_act)
drop_rate = conv_batchnorm_drop_rate[i]
if abs(drop_rate) > 1e-5:
tmp = layers.dropout(x=tmp, dropout_prob=drop_rate)
......
......@@ -256,7 +256,10 @@ def main(net_type, use_cuda, is_local=True):
save_dirname = "image_classification_" + net_type + ".inference.model"
train(net_type, use_cuda, save_dirname, is_local)
infer(use_cuda, save_dirname)
# There is bug in fluid.InferenceTranspiler for VGG.
if net_type == "resnet":
infer(use_cuda, save_dirname)
class TestImageClassification(unittest.TestCase):
......
......@@ -56,8 +56,8 @@ def get_numeric_gradient(place,
def get_output():
sum = []
op.run(scope, place)
for output_name in output_names:
op.run(scope, place)
sum.append(
np.array(scope.find_var(output_name).get_tensor()).mean())
return np.array(sum).mean()
......
......@@ -25,9 +25,6 @@ import paddle.fluid.core as core
class TestProfiler(unittest.TestCase):
def net_profiler(self, state, profile_path='/tmp/profile'):
enable_if_gpu = state == 'GPU' or state == "All"
if enable_if_gpu and not core.is_compiled_with_cuda():
return
startup_program = fluid.Program()
main_program = fluid.Program()
......@@ -81,8 +78,6 @@ class TestProfiler(unittest.TestCase):
pass_acc_calculator.add(value=acc, weight=b_size)
pass_acc = pass_acc_calculator.eval()
@unittest.skipIf(not core.is_compiled_with_cuda(),
"profiler is enabled only with GPU")
def test_cpu_profiler(self):
self.net_profiler('CPU')
......
......@@ -153,9 +153,6 @@ def append_input_output(block, op_proto, np_list, is_input, dtype):
def append_loss_ops(block, output_names):
mean_inputs = list(map(block.var, output_names))
# for item in mean_inputs:
# print(item)
# print("Item", item.dtype)
if len(mean_inputs) == 1:
loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1])
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册