tiny modify

fea/anakin-support-x86
nhzlx 7 years ago
commit 5fcdd81da7

@ -46,6 +46,7 @@
| tianbingsz | Tian-Bing Xu |
| tpatejko | Tomasz Patejko |
| typhoonzero | Yi Wu |
| velconia | Qi-Yang Min |
| wanghaoshuang | Hao-Shuang Wang |
| wangyang59 | Yang Wang |
| wangzhen-nlp | Zhen Wang |

@ -85,8 +85,7 @@ def dist_transpile(trainer_id, args):
trainer_id,
pservers=pserver_endpoints,
trainers=trainers,
sync_mode=not args.async_mode,
slice_var_up=not args.no_split_var)
sync_mode=not args.async_mode)
if training_role == "PSERVER":
pserver_program = t.get_pserver_program(current_endpoint)
pserver_startup_program = t.get_startup_program(current_endpoint,

@ -50,7 +50,7 @@ ExternalProject_Add(
UPDATE_COMMAND ""
CONFIGURE_COMMAND ""
BUILD_IN_SOURCE 1
PATCH_COMMAND git apply ${PADDLE_SOURCE_DIR}/patches/grpc/fix_too_early_destory.patch
PATCH_COMMAND cp ${PADDLE_SOURCE_DIR}/patches/grpc/grpc_library.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/grpc_library.h && cp ${PADDLE_SOURCE_DIR}/patches/grpc/completion_queue.h ${GRPC_SOURCES_DIR}/src/extern_grpc/include/grpcpp/impl/codegen/completion_queue.h
# NOTE(yuyang18):
# Disable -Werror, otherwise the compile will fail in MacOS.
# It seems that we cannot configure that by make command.

@ -263,7 +263,7 @@ function(cc_test TARGET_NAME)
COMMAND ${TARGET_NAME} ${cc_test_ARGS}
WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
if (${cc_test_SERIAL})
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
endif()
endif()
@ -328,7 +328,7 @@ function(nv_test TARGET_NAME)
add_dependencies(${TARGET_NAME} ${nv_test_DEPS} paddle_gtest_main lod_tensor memory gtest gflags glog)
add_test(${TARGET_NAME} ${TARGET_NAME})
if (nv_test_SERIAL)
set_property(TEST ${TARGET_NAME} PROPERTY SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
endif()
endif()

@ -148,18 +148,11 @@ if (WITH_ANAKIN AND WITH_GPU)
list(APPEND inference_deps anakin_inference_lib)
endif()
copy(inference_api_lib DEPS paddle_inference_api paddle_inference_api_shared
SRCS ${src_dir}/${module}/paddle_inference_api.h
${src_dir}/${module}/demo_ci
${PADDLE_BINARY_DIR}/paddle/fluid/inference/api/libpaddle_inference_api*
DSTS ${dst_dir}/inference ${dst_dir}/inference ${dst_dir}/inference
)
list(APPEND inference_deps inference_api_lib)
set(module "inference")
copy(inference_lib DEPS ${inference_deps}
SRCS ${src_dir}/${module}/*.h ${PADDLE_BINARY_DIR}/paddle/fluid/inference/libpaddle_fluid.*
DSTS ${dst_dir}/${module} ${dst_dir}/${module}
${src_dir}/${module}/api/paddle_inference_api.h ${src_dir}/${module}/api/demo_ci
DSTS ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module} ${dst_dir}/${module}
)
set(module "platform")

@ -8,9 +8,9 @@ cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
nv_test(dim_test SRCS dim_test.cu DEPS ddim)
cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context)
if(WITH_GPU)
nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type)
nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type device_context)
else()
cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type)
cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type device_context)
endif()
cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)
@ -110,7 +110,7 @@ cc_test(selected_rows_test SRCS selected_rows_test.cc DEPS selected_rows)
cc_test(op_kernel_type_test SRCS op_kernel_type_test.cc DEPS place device_context framework_proto)
cc_test(cow_ptr_tests SRCS details/cow_ptr_test.cc)
# cc_test(channel_test SRCS channel_test.cc)
cc_test(tuple_test SRCS tuple_test.cc )

@ -88,9 +88,8 @@ class BlockDesc {
OpDesc *InsertOp(size_t index);
/*
* Remove Op and its input/output variables.
* Note that for either input or output variable, if it is also an input or
* output variable of other ops, we should remain it.
* Only remove op itself,
* do nothing to its input and output variables
*/
void RemoveOp(size_t s, size_t e);

@ -259,7 +259,7 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::Apply(
result.Set("ops", new GraphOps);
// find send/recv vars so that we can place the distributed training
// realted op in the place 0
// related op in the place 0
auto send_vars = FindDistTrainSendVars(sorted_ops);
auto recv_vars = FindDistTrainRecvVars(sorted_ops);
@ -715,6 +715,7 @@ void MultiDevSSAGraphBuilder::CreateRPCOp(ir::Graph *result,
result->CreateOpNode(node->Op()), *node->Op(), local_scopes_[op_dev_id],
node->Op()->Type(), places_[op_dev_id]));
// TODO(panyx0718): This might not be needed anymore.
if (node->Op()->Type() == "send_barrier") {
ConnectOp(result, result->Get<GraphOps>("ops").back().get(), "send");
} else if (node->Op()->Type() == "recv") {

@ -24,6 +24,68 @@ namespace paddle {
namespace framework {
namespace ir {
std::vector<std::string> FindDistTrainSendVars(
const std::vector<ir::Node *> &nodes) {
std::vector<std::string> send_vars;
// since parameters are all in block 0,
// it's enough to only scan send ops in block 0
for (auto &node : nodes) {
auto op_vars = node->Op()->InputArgumentNames();
send_vars.reserve(send_vars.size() +
std::distance(op_vars.begin(), op_vars.end()));
send_vars.insert(send_vars.end(), op_vars.begin(), op_vars.end());
}
return send_vars;
}
std::vector<std::string> FindDistTrainRecvVars(
const std::vector<ir::Node *> &nodes) {
std::vector<std::string> recv_vars;
for (auto &node : nodes) {
auto op_vars = node->Op()->OutputArgumentNames();
recv_vars.reserve(recv_vars.size() +
std::distance(op_vars.begin(), op_vars.end()));
recv_vars.insert(recv_vars.end(), op_vars.begin(), op_vars.end());
}
return recv_vars;
}
bool IsDistTrainOp(ir::Node *node, const std::vector<std::string> &send_vars,
const std::vector<std::string> &recv_vars) {
if (send_vars.size() == 0 || recv_vars.size() == 0) {
return false;
}
/**
* Check any of opvars contains `.block` and in sendvars
*/
auto checker = [](const std::vector<std::string> &opvars,
const std::vector<std::string> &rpc_vars) -> bool {
for (auto &var : opvars) {
// a variable name with the suffix `.block` means it's a splited
// variable by (DistributeTranspiler)
// [python/paddle/fluid/transpiler/distribute_transpiler.py]
if (var.find(".block") != std::string::npos &&
std::find(rpc_vars.begin(), rpc_vars.end(), var) != rpc_vars.end()) {
return true;
}
}
return false;
};
std::vector<std::string> input_var_names;
std::vector<std::string> output_var_names;
for (ir::Node *input : node->inputs) {
input_var_names.push_back(input->Name());
}
for (ir::Node *output : node->outputs) {
output_var_names.push_back(output->Name());
}
return checker(output_var_names, send_vars) ||
checker(input_var_names, recv_vars);
}
Graph::Graph(const ProgramDesc &program) : program_(program) {
VLOG(3) << "block in program:" << program_.Size();
std::unordered_map<std::string, VarDesc *> all_vars;
@ -61,6 +123,64 @@ Graph::Graph(const ProgramDesc &program) : program_(program) {
var->inputs.push_back(node);
}
}
std::vector<ir::Node *> send_ops;
ir::Node *send_bar = nullptr;
std::vector<ir::Node *> recv_ops;
ir::Node *fetch_bar = nullptr;
for (ir::Node *node : Nodes()) {
if (node->Name() == "send") {
send_ops.push_back(node);
} else if (node->Name() == "send_barrier") {
PADDLE_ENFORCE(!send_bar, "only has one send barrier");
send_bar = node;
} else if (node->Name() == "recv") {
recv_ops.push_back(node);
} else if (node->Name() == "fetch_barrier") {
PADDLE_ENFORCE(!fetch_bar, "only has one fetch barrier");
fetch_bar = node;
}
}
if (send_bar) {
for (ir::Node *send : send_ops) {
ir::Node *dep_var = CreateControlDepVar();
send->outputs.push_back(dep_var);
dep_var->inputs.push_back(send);
send_bar->inputs.push_back(dep_var);
dep_var->outputs.push_back(send_bar);
}
for (ir::Node *recv : recv_ops) {
ir::Node *dep_var = CreateControlDepVar();
recv->inputs.push_back(dep_var);
dep_var->outputs.push_back(recv);
send_bar->outputs.push_back(dep_var);
dep_var->inputs.push_back(send_bar);
}
}
if (fetch_bar) {
for (ir::Node *recv : recv_ops) {
ir::Node *dep_var = CreateControlDepVar();
recv->outputs.push_back(dep_var);
dep_var->inputs.push_back(recv);
fetch_bar->inputs.push_back(dep_var);
dep_var->outputs.push_back(fetch_bar);
}
}
std::vector<std::string> send_vars = FindDistTrainSendVars(send_ops);
std::vector<std::string> recv_vars = FindDistTrainRecvVars(recv_ops);
for (ir::Node *node : Nodes()) {
if (IsDistTrainOp(node, send_vars, recv_vars)) {
if (fetch_bar && node->Name() == "concat") {
ir::Node *dep_var = CreateControlDepVar();
fetch_bar->outputs.push_back(dep_var);
dep_var->inputs.push_back(fetch_bar);
node->inputs.push_back(dep_var);
dep_var->outputs.push_back(node);
}
}
}
/**
* We only handle write after read(WAR), since it should not have a write
* after write in program. If there are write after write operators, we need

@ -679,6 +679,8 @@ void OperatorWithKernel::RunImpl(const Scope& scope,
if (var == nullptr) continue;
if (var->IsType<framework::LoDTensor>()) {
CheckTensorNANOrInf(vname, var->Get<framework::LoDTensor>());
} else if (var->IsType<framework::SelectedRows>()) {
CheckTensorNANOrInf(vname, var->Get<framework::SelectedRows>().value());
}
}
}

@ -14,8 +14,15 @@ cc_library(paddle_fluid_api
get_property(fluid_modules GLOBAL PROPERTY FLUID_MODULES)
# paddle_fluid_origin exclude inference api interface
cc_library(paddle_fluid_origin DEPS ${fluid_modules} paddle_fluid_api)
if(NOT APPLE)
add_subdirectory(api)
endif()
# Create static library
cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api)
cc_library(paddle_fluid DEPS ${fluid_modules} paddle_fluid_api paddle_inference_api)
if(NOT APPLE)
# TODO(liuyiqu: Temporarily disable the link flag because it is not support on Mac.
set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.sym")
@ -24,7 +31,7 @@ endif()
# Create shared library
cc_library(paddle_fluid_shared SHARED
SRCS io.cc
SRCS io.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api.cc ${CMAKE_CURRENT_SOURCE_DIR}/api/api_impl.cc
DEPS ${fluid_modules} paddle_fluid_api)
set_target_properties(paddle_fluid_shared PROPERTIES OUTPUT_NAME paddle_fluid)
@ -32,12 +39,21 @@ if(NOT APPLE)
# TODO(liuyiqun): Temporarily disable the link flag because it is not support on Mac.
set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/paddle_fluid.map")
set_target_properties(paddle_fluid_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
# check symbol hidden
FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake
"execute_process(COMMAND bash -c \"${CMAKE_CURRENT_SOURCE_DIR}/check_symbol.sh"
" ${CMAKE_CURRENT_BINARY_DIR}/libpaddle_fluid.so\" RESULT_VARIABLE symbol_res)\n"
"if(NOT \"\${symbol_res}\" STREQUAL \"0\")\n"
" message(FATAL_ERROR \"Check symbol failed.\")\n"
"endif()\n")
add_custom_command(
OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol"
COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake"
DEPENDS paddle_fluid_shared)
add_custom_target(check_symbol ALL DEPENDS "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol")
endif()
if(WITH_TESTING)
# both tests/book and analysis depends the models that generated by python/paddle/fluid/tests/book
add_subdirectory(tests/book)
endif()
if(NOT APPLE)
add_subdirectory(api)
endif()

@ -42,35 +42,8 @@ function(inference_api_test TARGET_NAME)
endif(WITH_TESTING)
endfunction(inference_api_test)
cc_library(paddle_inference_api
SRCS api.cc api_impl.cc
DEPS ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
if(NOT APPLE)
set(LINK_FLAGS "-Wl,--retain-symbols-file ${CMAKE_CURRENT_SOURCE_DIR}/api.sym")
set_target_properties(paddle_inference_api PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
endif()
# Here the shared library doesn't depend on other fluid libraries, or double free will occur.
cc_library(paddle_inference_api_shared SHARED
SRCS api.cc api_impl.cc)
add_dependencies(paddle_inference_api_shared ${FLUID_CORE_MODULES} ${GLOB_OP_LIB})
set_target_properties(paddle_inference_api_shared PROPERTIES OUTPUT_NAME paddle_inference_api)
cc_library(paddle_inference_api SRCS api.cc api_impl.cc DEPS lod_tensor)
if(NOT APPLE)
set(LINK_FLAGS "-Wl,--version-script ${CMAKE_CURRENT_SOURCE_DIR}/api.map")
set_target_properties(paddle_inference_api_shared PROPERTIES LINK_FLAGS "${LINK_FLAGS}")
FILE(WRITE ${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake
"execute_process(COMMAND bash -c \"${CMAKE_CURRENT_SOURCE_DIR}/check_symbol.sh"
" ${CMAKE_CURRENT_BINARY_DIR}/libpaddle_inference_api.so\" RESULT_VARIABLE symbol_res)\n"
"if(NOT \"\${symbol_res}\" STREQUAL \"0\")\n"
" message(FATAL_ERROR \"Check symbol failed.\")\n"
"endif()\n")
add_custom_command(
OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol"
COMMAND ${CMAKE_COMMAND} -P "${CMAKE_CURRENT_BINARY_DIR}/check_symbol.cmake"
DEPENDS paddle_inference_api_shared)
add_custom_target(check_symbol ALL DEPENDS "${CMAKE_CURRENT_BINARY_DIR}/.check_symbol")
endif()
cc_test(test_paddle_inference_api
SRCS api_tester.cc

@ -1,6 +0,0 @@
{
global:
*paddle*;
local:
*;
};

@ -55,11 +55,9 @@ endif()
# Note: libpaddle_inference_api.so/a must put before libpaddle_fluid.so/a
if(WITH_STATIC_LIB)
set(DEPS
${PADDLE_LIB}/paddle/fluid/inference/libpaddle_inference_api.a
${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.a)
else()
set(DEPS
${PADDLE_LIB}/paddle/fluid/inference/libpaddle_inference_api.so
${PADDLE_LIB}/paddle/fluid/inference/libpaddle_fluid.so)
endif()
set(EXTERNAL_LIB "-lrt -ldl -lpthread")

@ -0,0 +1,4 @@
set -x
cd `dirname $0`
rm -rf build/ data/
set +x

@ -3,8 +3,8 @@
lib=$1
if [ $# -ne 1 ]; then echo "No input library"; exit -1 ; fi
num_paddle_syms=$(nm -D --defined-only ${lib} | grep paddle | wc -l)
num_google_syms=$(nm -D --defined-only ${lib} | grep google | wc -l)
num_paddle_syms=$(nm -D ${lib} | grep paddle | wc -l)
num_google_syms=$(nm -D ${lib} | grep google | grep -v paddle | grep T | wc -l)
if [ $num_paddle_syms -le 0 ]; then echo "Have no paddle symbols"; exit -1 ; fi
if [ $num_google_syms -ge 1 ]; then echo "Have some google symbols"; exit -1 ; fi

@ -1,7 +1,7 @@
# Add TRT tests
nv_library(tensorrt_converter
SRCS mul_op.cc conv2d_op.cc fc_op.cc pool2d_op.cc elementwise_op.cc
DEPS tensorrt_engine mul_op)
DEPS tensorrt_engine operator scope framework_proto op_registry)
nv_test(test_op_converter SRCS test_op_converter.cc DEPS
${FLUID_CORE_MODULES} tensorrt_engine tensorrt_converter)

@ -109,7 +109,7 @@ class ElementwiseTensorOpConverter : public OpConverter {
nvinfer1::Dims dims_x = X->getDimensions();
nvinfer1::Dims dims_y = Y->getDimensions();
// only support the C * H * W input format
// The two input tensor should have the same dims
PADDLE_ENFORCE(dims_x.nbDims >= 3);
if (dims_x.nbDims == dims_y.nbDims) {
for (int i = 0; i < dims_x.nbDims; i++) {

@ -49,5 +49,4 @@ class MulOpConverter : public OpConverter {
} // namespace inference
} // namespace paddle
USE_OP(mul);
REGISTER_TRT_OP_CONVERTER(mul, MulOpConverter);

@ -47,7 +47,7 @@ TEST(elementwise_op, add_weight_test) {
TEST(elementwise_op, add_tensor_test) {
std::unordered_set<std::string> parameters;
framework::Scope scope;
TRTConvertValidation validator(1, parameters, scope, 1 << 15);
TRTConvertValidation validator(2, parameters, scope, 1 << 15);
validator.DeclInputVar("elementwise_add-X", nvinfer1::DimsCHW(10, 3, 3));
validator.DeclInputVar("elementwise_add-Y", nvinfer1::Dims3(10, 3, 3));
// validator.DeclParamVar("mul-Y", nvinfer1::Dims2(8, 2));
@ -60,8 +60,7 @@ TEST(elementwise_op, add_tensor_test) {
desc.SetInput("Y", {"elementwise_add-Y"});
desc.SetOutput("Out", {"elementwise_add-Out"});
int axis = 1;
desc.SetAttr("axis", axis);
// the defalut axis of elementwise op is -1
validator.SetOp(*desc.Proto());

@ -17,7 +17,7 @@ function(inference_test TARGET_NAME)
string(REGEX REPLACE "^_$" "" arg "${arg}")
cc_test(test_inference_${TARGET_NAME}${arg}
SRCS test_inference_${TARGET_NAME}.cc
DEPS paddle_fluid
DEPS paddle_fluid_origin
ARGS --dirname=${PYTHON_TESTS_DIR}/book/${TARGET_NAME}${arg}.inference.model)
set_tests_properties(test_inference_${TARGET_NAME}${arg}
PROPERTIES DEPENDS test_${TARGET_NAME})
@ -43,6 +43,6 @@ inference_test(word2vec)
# TODO(TJ): clean me up
cc_test(test_inference_nlp
SRCS test_inference_nlp.cc
DEPS paddle_fluid
DEPS paddle_fluid_origin
ARGS
--model_path=${PADDLE_BINARY_DIR}/python/paddle/fluid/tests/book/recognize_digits_mlp.inference.model)

@ -20,9 +20,6 @@ limitations under the License. */
#include "gtest/gtest.h"
#include "paddle/fluid/inference/tests/test_helper.h"
#include "paddle/fluid/platform/cpu_helper.h"
#ifdef PADDLE_WITH_MKLML
#include <omp.h>
#endif
DEFINE_string(model_path, "", "Directory of the inference model.");
DEFINE_string(data_file, "", "File of input index data.");
@ -30,6 +27,7 @@ DEFINE_int32(repeat, 100, "Running the inference program repeat times");
DEFINE_bool(prepare_vars, true, "Prepare variables before executor");
DEFINE_int32(num_threads, 1, "Number of threads should be used");
DECLARE_bool(use_mkldnn);
DECLARE_int32(paddle_num_threads);
inline double GetCurrentMs() {
struct timeval time;
@ -160,12 +158,7 @@ TEST(inference, nlp) {
std::unique_ptr<paddle::framework::Scope> scope(
new paddle::framework::Scope());
#ifdef PADDLE_WITH_MKLML
// only use 1 thread number per std::thread
omp_set_dynamic(0);
omp_set_num_threads(1);
paddle::platform::SetNumThreads(1);
#endif
paddle::platform::SetNumThreads(FLAGS_paddle_num_threads);
double start_ms = 0, stop_ms = 0;
if (FLAGS_num_threads > 1) {

@ -15,6 +15,10 @@ limitations under the License. */
#include "paddle/fluid/memory/detail/buddy_allocator.h"
#include "glog/logging.h"
DEFINE_bool(free_idle_memory, false,
"If it is true, Paddle will try to free idle memory trunks during "
"running time.");
namespace paddle {
namespace memory {
namespace detail {
@ -152,13 +156,14 @@ void BuddyAllocator::Free(void* p) {
pool_.insert(
IndexSizeAddress(block->index(cache_), block->total_size(cache_), block));
// Clean up if existing too much free memory
// Prefer freeing fallback allocation first
CleanIdleFallBackAlloc();
if (FLAGS_free_idle_memory) {
// Clean up if existing too much free memory
// Prefer freeing fallback allocation first
CleanIdleFallBackAlloc();
// Free normal allocation
CleanIdleNormalAlloc();
// Free normal allocation
CleanIdleNormalAlloc();
}
}
size_t BuddyAllocator::Used() { return total_used_; }

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

Loading…
Cancel
Save