Skip to content

N
ngraph
Commit a903a001 authored by fenglei.tian's avatar fenglei.tian
Browse files
Options

Merge remote-tracking branch 'origin/master' into tfl/gpu_add_nvrtc

parents ee220ffb ce3670b5
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 688 additions and 269 deletions
.gitignore
View file @ a903a001
......@@ -48,6 +48,7 @@ output/
*.mpg
*.cpio
*.wav
*.backup
doc/source/generated
.cache/
nervana_aeon.egg-info/
......
cmake/external_argon_transformer.cmake
View file @ a903a001
......@@ -30,20 +30,24 @@ if (NGRAPH_ARGON_ENABLE)
endif()
# Repository
set(ARGON_TRANSFORMER_CMAKE_GIT_REPOSITORY git@github.com:NervanaSystems/argon-transformer.git)
if (DEFINED CUSTOM_ARGON_TRANSFORMER_GIT_REPOSITORY)
set(ARGON_TRANSFORMER_GIT_REPOSITORY ${CUSTOM_ARGON_TRANSFORMER_GIT_REPOSITORY})
else()
set(ARGON_TRANSFORMER_GIT_REPOSITORY git@github.com:NervanaSystems/argon-transformer.git)
endif()
# Set argon_transformer tag
# Notes:
# - Before we have ngraph CI job for argon transformer, ngraph master might not be
# compatible with argon transformer. To ensure compatibility, checkout the ngraph commit point
# where the following `ARGON_TRANSFORMER_CMAKE_GIT_TAG` is set and build ngraph with argon using this
# where the following `ARGON_TRANSFORMER_GIT_TAG` is set and build ngraph with argon using this
# commit.
# - After we have ngraph CI job for argon transformer, ngraph master will be compatible with
# argon transformer guaranteed by CI.
set(ARGON_TRANSFORMER_CMAKE_GIT_TAG cpp-master)
set(ARGON_TRANSFORMER_GIT_TAG cpp-master)
# Determines where argon-transformer will be located
set(ARGON_TRANSFORMER_CMAKE_PREFIX ${CMAKE_CURRENT_BINARY_DIR}/argon_transformer)
set(ARGON_TRANSFORMER_PREFIX ${CMAKE_CURRENT_BINARY_DIR}/argon_transformer)
# Print
message(STATUS "NGRAPH_INCLUDE_PATH: ${NGRAPH_INCLUDE_PATH}")
......@@ -56,27 +60,35 @@ if (NGRAPH_ARGON_ENABLE)
ExternalProject_Add(
ext_argon_transformer
SOURCE_DIR ${CUSTOM_ARGON_TRANSFORMER_DIR}
PREFIX ${ARGON_TRANSFORMER_CMAKE_PREFIX}
PREFIX ${ARGON_TRANSFORMER_PREFIX}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-DNGRAPH_INSTALL_PREFIX=${ARGON_TRANSFORMER_CMAKE_PREFIX}
CMAKE_ARGS
-DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DNGRAPH_INSTALL_PREFIX=${ARGON_TRANSFORMER_PREFIX}
-DPREBUILD_ARGON_API_PATH=${NGRAPH_PREBUILD_ARGON_API_PATH}
-DEXTERNAL_NGRAPH_INCLUDE_DIR=${NGRAPH_INCLUDE_PATH}
-DINSTALLED_HEADERS_PATH=${CMAKE_INSTALL_PREFIX}/include
-DMKLDNN_INCLUDE_DIR=${MKLDNN_INCLUDE_DIR}
BUILD_ALWAYS 1
)
else()
ExternalProject_Add(
ext_argon_transformer
GIT_REPOSITORY ${ARGON_TRANSFORMER_CMAKE_GIT_REPOSITORY}
GIT_TAG ${ARGON_TRANSFORMER_CMAKE_GIT_TAG}
PREFIX ${ARGON_TRANSFORMER_CMAKE_PREFIX}
GIT_REPOSITORY ${ARGON_TRANSFORMER_GIT_REPOSITORY}
GIT_TAG ${ARGON_TRANSFORMER_GIT_TAG}
PREFIX ${ARGON_TRANSFORMER_PREFIX}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-DNGRAPH_INSTALL_PREFIX=${ARGON_TRANSFORMER_CMAKE_PREFIX}
CMAKE_ARGS
-DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DNGRAPH_INSTALL_PREFIX=${ARGON_TRANSFORMER_PREFIX}
-DPREBUILD_ARGON_API_PATH=${NGRAPH_PREBUILD_ARGON_API_PATH}
-DEXTERNAL_NGRAPH_INCLUDE_DIR=${NGRAPH_INCLUDE_PATH}
-DINSTALLED_HEADERS_PATH=${CMAKE_INSTALL_PREFIX}/include
-DMKLDNN_INCLUDE_DIR=${MKLDNN_INCLUDE_DIR}
BUILD_ALWAYS 1
)
endif()
......@@ -85,29 +97,37 @@ if (NGRAPH_ARGON_ENABLE)
ExternalProject_Add(
ext_argon_transformer
SOURCE_DIR ${CUSTOM_ARGON_TRANSFORMER_DIR}
PREFIX ${ARGON_TRANSFORMER_CMAKE_PREFIX}
PREFIX ${ARGON_TRANSFORMER_PREFIX}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-DNGRAPH_INSTALL_PREFIX=${ARGON_TRANSFORMER_CMAKE_PREFIX}
CMAKE_ARGS
-DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DNGRAPH_INSTALL_PREFIX=${ARGON_TRANSFORMER_PREFIX}
-DPREBUILD_ARGON_API_PATH=${NGRAPH_PREBUILD_ARGON_API_PATH}
-DEXTERNAL_NGRAPH_INCLUDE_DIR=${NGRAPH_INCLUDE_PATH}
-DINSTALLED_HEADERS_PATH=${CMAKE_INSTALL_PREFIX}/include
BUILD_BYPRODUCTS ${ARGON_TRANSFORMER_CMAKE_PREFIX}
-DMKLDNN_INCLUDE_DIR=${MKLDNN_INCLUDE_DIR}
BUILD_BYPRODUCTS ${ARGON_TRANSFORMER_PREFIX}
BUILD_ALWAYS 1
)
else()
ExternalProject_Add(
ext_argon_transformer
GIT_REPOSITORY ${ARGON_TRANSFORMER_CMAKE_GIT_REPOSITORY}
GIT_TAG ${ARGON_TRANSFORMER_CMAKE_GIT_TAG}
PREFIX ${ARGON_TRANSFORMER_CMAKE_PREFIX}
GIT_REPOSITORY ${ARGON_TRANSFORMER_GIT_REPOSITORY}
GIT_TAG ${ARGON_TRANSFORMER_GIT_TAG}
PREFIX ${ARGON_TRANSFORMER_PREFIX}
UPDATE_COMMAND ""
CMAKE_ARGS -DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-DNGRAPH_INSTALL_PREFIX=${ARGON_TRANSFORMER_CMAKE_PREFIX}
CMAKE_ARGS
-DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
-DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
-DNGRAPH_INSTALL_PREFIX=${ARGON_TRANSFORMER_PREFIX}
-DPREBUILD_ARGON_API_PATH=${NGRAPH_PREBUILD_ARGON_API_PATH}
-DEXTERNAL_NGRAPH_INCLUDE_DIR=${NGRAPH_INCLUDE_PATH}
-DINSTALLED_HEADERS_PATH=${CMAKE_INSTALL_PREFIX}/include
BUILD_BYPRODUCTS ${ARGON_TRANSFORMER_CMAKE_PREFIX}
-DMKLDNN_INCLUDE_DIR=${MKLDNN_INCLUDE_DIR}
BUILD_BYPRODUCTS ${ARGON_TRANSFORMER_PREFIX}
BUILD_ALWAYS 1
)
endif()
......@@ -115,8 +135,8 @@ if (NGRAPH_ARGON_ENABLE)
ExternalProject_Get_Property(ext_argon_transformer source_dir)
set(ARGON_TRANSFORMER_SOURCE_DIR ${source_dir} PARENT_SCOPE)
set(ARGON_TRANSFORMER_INCLUDE_DIR ${ARGON_TRANSFORMER_CMAKE_PREFIX}/include PARENT_SCOPE)
set(ARGON_TRANSFORMER_LIB_DIR ${ARGON_TRANSFORMER_CMAKE_PREFIX}/lib PARENT_SCOPE)
set(ARGON_TRANSFORMER_INCLUDE_DIR ${ARGON_TRANSFORMER_PREFIX}/include PARENT_SCOPE)
set(ARGON_TRANSFORMER_LIB_DIR ${ARGON_TRANSFORMER_PREFIX}/lib PARENT_SCOPE)
set(ARGON_API_INCLUDE_DIR ${NGRAPH_PREBUILD_ARGON_API_PATH}/include PARENT_SCOPE)
set(ARGON_API_LIB_DIR ${NGRAPH_PREBUILD_ARGON_API_PATH}/lib) # Used by find_library below
set(ARGON_API_LIB_DIR ${NGRAPH_PREBUILD_ARGON_API_PATH}/lib PARENT_SCOPE)
......
cmake/external_mkldnn.cmake
View file @ a903a001
......@@ -81,4 +81,8 @@ if(NGRAPH_CPU_ENABLE)
set(MKLDNN_INCLUDE_DIR "${EXTERNAL_PROJECTS_ROOT}/mkldnn/include" PARENT_SCOPE)
set(MKLDNN_LIB_DIR "${EXTERNAL_PROJECTS_ROOT}/mkldnn/lib" PARENT_SCOPE)
# Other .cmake files in current scope (e.g. Argon Transformer) needs this path as well
set(MKLDNN_INCLUDE_DIR "${EXTERNAL_PROJECTS_ROOT}/mkldnn/include")
set(MKLDNN_LIB_DIR "${EXTERNAL_PROJECTS_ROOT}/mkldnn/lib")
endif()
doc/sphinx/ngraph.doxyfile_backup deleted 100644 → 0
View file @ ee220ffb
This diff is collapsed.
doc/sphinx/ngraph_theme/static/css/theme.css.backup deleted 100644 → 0
View file @ ee220ffb
This diff is collapsed.
maint/apply-code-format.sh
View file @ a903a001
......@@ -42,21 +42,23 @@ bash_lib_status "Verified that '${CLANG_FORMAT_PROG}' has version '${REQUIRED_CL
pushd "${THIS_SCRIPT_DIR}/.."
declare ARGON_SRC_DIR="build/third-party/argon_transformer/src/ext_argon_transformer/src"
declare ARGON_TEST_DIR="build/third-party/argon_transformer/src/ext_argon_transformer/test"
declare ROOT_SUBDIR
for ROOT_SUBDIR in src test; do
for ROOT_SUBDIR in src test ${ARGON_SRC_DIR} ${ARGON_TEST_DIR}; do
if ! [[ -d "${ROOT_SUBDIR}" ]]; then
bash_lib_die "In directory '$(pwd)', no subdirectory named '${ROOT_SUBDIR}' was found."
fi
bash_lib_status "About to format C/C++ code in directory tree '$(pwd)/${ROOT_SUBDIR}' ..."
bash_lib_status "In directory '$(pwd)', no subdirectory named '${ROOT_SUBDIR}' was found."
else
bash_lib_status "About to format C/C++ code in directory tree '$(pwd)/${ROOT_SUBDIR}' ..."
# Note that we restrict to "-type f" to exclude symlinks. Emacs sometimes
# creates dangling symlinks with .cpp/.hpp suffixes as a sort of locking
# mechanism, and this confuses clang-format.
find "${ROOT_SUBDIR}" -type f -and \( -name '*.cpp' -or -name '*.hpp' \) | xargs "${CLANG_FORMAT_PROG}" -i -style=file
# Note that we restrict to "-type f" to exclude symlinks. Emacs sometimes
# creates dangling symlinks with .cpp/.hpp suffixes as a sort of locking
# mechanism, and this confuses clang-format.
find "${ROOT_SUBDIR}" -type f -and \( -name '*.cpp' -or -name '*.hpp' \) | xargs "${CLANG_FORMAT_PROG}" -i -style=file
bash_lib_status "Done."
bash_lib_status "Done."
fi
done
popd
maint/check-code-format.sh
View file @ a903a001
......@@ -45,23 +45,26 @@ declare NUM_FILES_CHECKED=0
pushd "${THIS_SCRIPT_DIR}/.."
declare ARGON_SRC_DIR="build/third-party/argon_transformer/src/ext_argon_transformer/src"
declare ARGON_TEST_DIR="build/third-party/argon_transformer/src/ext_argon_transformer/test"
declare ROOT_SUBDIR
for ROOT_SUBDIR in src test; do
for ROOT_SUBDIR in src test ${ARGON_SRC_DIR} ${ARGON_TEST_DIR}; do
if ! [[ -d "${ROOT_SUBDIR}" ]]; then
bash_lib_die "In directory '$(pwd)', no subdirectory named '${ROOT_SUBDIR}' was found."
bash_lib_status "In directory '$(pwd)', no subdirectory named '${ROOT_SUBDIR}' was found."
else
bash_lib_status "About to format C/C++ code in directory tree '$(pwd)/${ROOT_SUBDIR}' ..."
declare SRC_FILE
# Note that we restrict to "-type f" to exclude symlinks. Emacs sometimes
# creates dangling symlinks with .cpp/.hpp suffixes as a sort of locking
# mechanism, and this confuses clang-format.
for SRC_FILE in $(find "${ROOT_SUBDIR}" -type f -and \( -name '*.cpp' -or -name '*.hpp' \) ); do
if "${CLANG_FORMAT_PROG}" -style=file -output-replacements-xml "${SRC_FILE}" | grep -c "<replacement " >/dev/null; then
FAILED_FILES+=( "${SRC_FILE}" )
fi
NUM_FILES_CHECKED=$((NUM_FILES_CHECKED+1))
done
fi
bash_lib_status "About to format C/C++ code in directory tree '$(pwd)/${ROOT_SUBDIR}' ..."
declare SRC_FILE
# Note that we restrict to "-type f" to exclude symlinks. Emacs sometimes
# creates dangling symlinks with .cpp/.hpp suffixes as a sort of locking
# mechanism, and this confuses clang-format.
for SRC_FILE in $(find "${ROOT_SUBDIR}" -type f -and \( -name '*.cpp' -or -name '*.hpp' \) ); do
if "${CLANG_FORMAT_PROG}" -style=file -output-replacements-xml "${SRC_FILE}" | grep -c "<replacement " >/dev/null; then
FAILED_FILES+=( "${SRC_FILE}" )
fi
NUM_FILES_CHECKED=$((NUM_FILES_CHECKED+1))
done
done
popd
......@@ -76,4 +79,3 @@ else
done
exit 1
fi
src/ngraph/CMakeLists.txt
View file @ a903a001
......@@ -38,7 +38,7 @@ set (SRC
ops/avg_pool.cpp
ops/batch_norm.cpp
ops/broadcast.cpp
ops/concatenate.cpp
ops/concat.cpp
ops/constant.cpp
ops/convert.cpp
ops/convolution.cpp
......@@ -78,12 +78,13 @@ set (SRC
ops/tan.cpp
ops/tanh.cpp
ops/util/arithmetic_reduction.cpp
ops/util/binary_elementwise_arithmetic.cpp
ops/util/binary_elementwise_comparison.cpp
ops/util/binary_elementwise.cpp
ops/util/requires_tensor_view_args.cpp
ops/util/unary_elementwise_arithmetic.cpp
ops/util/binary_elementwise_arithmetic.cpp
ops/util/binary_elementwise_comparison.cpp
ops/util/binary_elementwise.cpp
ops/util/requires_tensor_view_args.cpp
ops/util/unary_elementwise_arithmetic.cpp
ops/util/unary_elementwise.cpp
pass/assign_placement.cpp
pass/dump_sorted.cpp
pass/graph_rewrite.cpp
pass/inliner.cpp
......@@ -112,6 +113,7 @@ set (SRC
types/type.cpp
util.cpp
graph_util.cpp
placement.cpp
)
message(STATUS ${CMAKE_CURRENT_SOURCE_DIR}/ops)
......
src/ngraph/codegen/compiler.cpp
View file @ a903a001
......@@ -411,24 +411,21 @@ void codegen::StaticCompiler::configure_search_path()
void codegen::StaticCompiler::load_headers_from_resource()
{
const string builtin_root = "/$builtin";
HeaderSearchOptions& hso = m_compiler->getInvocation().getHeaderSearchOpts();
PreprocessorOptions& preprocessor_options = m_compiler->getInvocation().getPreprocessorOpts();
std::set<std::string> header_search_paths;
for (const HeaderInfo& hi : header_info)
for (const string& search_path : builtin_search_paths)
{
string search_path = hi.search_path;
string absolute_path = file_util::path_join(search_path, hi.header_path);
string builtin = "/$builtin" + absolute_path;
string builtin = builtin_root + search_path;
hso.AddPath(builtin, clang::frontend::System, false, false);
}
for (const pair<string, string>& header_info : builtin_headers)
{
string absolute_path = header_info.first;
string builtin = builtin_root + absolute_path;
std::unique_ptr<llvm::MemoryBuffer> mb(
llvm::MemoryBuffer::getMemBuffer(hi.header_data, builtin));
llvm::MemoryBuffer::getMemBuffer(header_info.second, builtin));
preprocessor_options.addRemappedFile(builtin, mb.release());
if (!contains(header_search_paths, search_path))
{
string builtin = "/$builtin" + search_path;
hso.AddPath(builtin, clang::frontend::System, false, false);
header_search_paths.insert(search_path);
}
}
}
......
src/ngraph/graph_util.cpp
View file @ a903a001
This diff is collapsed.
src/ngraph/graph_util.hpp
View file @ a903a001
......@@ -30,11 +30,24 @@
#include <unordered_set>
#include <vector>
#include "ngraph/placement.hpp"
namespace ngraph
{
class Node;
class Function;
namespace descriptor
{
class Input;
class Output;
}
namespace op
{
class Parameter;
}
void traverse_nodes(const std::shared_ptr<const Function> p,
std::function<void(std::shared_ptr<Node>)> f);
void traverse_nodes(const Function* p, std::function<void(std::shared_ptr<Node>)> f);
......@@ -60,7 +73,7 @@ namespace ngraph
class NodeMap
{
public:
// map original node to replcacement node
// map original node to replacement node
// throws ngraph_error if key already exists
void add(std::shared_ptr<ngraph::Node> orig, std::shared_ptr<ngraph::Node> replacement);
......@@ -100,4 +113,18 @@ namespace ngraph
// NodeMap output (by reference) fully maps input and cloned function ops
std::shared_ptr<ngraph::Function> clone_function(std::shared_ptr<ngraph::Function> func,
NodeMap& node_map);
// Assert that nodes in the function is colocated and return that placement
Placement get_colocated_function_placement(std::shared_ptr<Function> func);
// Split function to function(s) with unique placement
std::vector<std::shared_ptr<Function>> split_function_by_placement(
std::shared_ptr<Function> f,
std::unordered_map<std::shared_ptr<op::Parameter>, std::shared_ptr<Node>>&
map_parameter_to_source_node);
// Insert parameter node between src_node and dst_node by splitting the graph
void insert_parameter_split_between(std::shared_ptr<Node> src_node,
std::shared_ptr<Node> dst_node,
std::shared_ptr<op::Parameter> p_node);
}
src/ngraph/ngraph.hpp
View file @ a903a001
......@@ -73,7 +73,7 @@
#include "ngraph/ops/avg_pool.hpp"
#include "ngraph/ops/broadcast.hpp"
#include "ngraph/ops/ceiling.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/ops/constant.hpp"
#include "ngraph/ops/convert.hpp"
#include "ngraph/ops/convolution.hpp"
......
src/ngraph/node.cpp
View file @ a903a001
......@@ -23,6 +23,7 @@
#include "ngraph/descriptor/layout/tensor_view_layout.hpp"
#include "ngraph/descriptor/primary_tensor_view.hpp"
#include "ngraph/ops/parameter.hpp"
#include "ngraph/placement.hpp"
using namespace std;
using namespace ngraph;
......@@ -144,6 +145,16 @@ void Node::set_name(const string& name)
}
}
Placement Node::get_placement() const
{
return m_placement;
}
void Node::set_placement(Placement placement)
{
m_placement = placement;
}
std::shared_ptr<Node> Node::get_input_op(size_t index)
{
for (auto arg : m_arguments)
......@@ -304,3 +315,27 @@ bool Node::has_same_type(std::shared_ptr<const Node> node) const
}
return true;
}
descriptor::Input* Node::get_input_from(const shared_ptr<Node>& src)
{
for (size_t i = 0; i < this->get_input_size(); ++i)
{
if (this->get_input_op(i) == src)
{
return &(this->get_inputs().at(i));
}
}
throw ngraph_error("Error: src is not one of self's input Node");
}
descriptor::Output* Node::get_output_to(const shared_ptr<Node>& dst)
{
for (size_t i = 0; i < dst->get_input_size(); ++i)
{
if (dst->get_input_op(i).get() == this)
{
return &(dst->get_inputs().at(i).get_output());
}
}
throw ngraph_error("Error: dst is not one of self's output Node");
}
src/ngraph/node.hpp
View file @ a903a001
......@@ -32,6 +32,7 @@
#include "ngraph/descriptor/input.hpp"
#include "ngraph/descriptor/output.hpp"
#include "ngraph/descriptor/tensor.hpp"
#include "ngraph/placement.hpp"
#include "ngraph/types/type.hpp"
namespace ngraph
......@@ -39,6 +40,10 @@ namespace ngraph
void replace_node_users_arguments(std::shared_ptr<Node> target,
std::shared_ptr<Node> replacement);
void insert_parameter_split_between(std::shared_ptr<Node> src_node,
std::shared_ptr<Node> dst_node,
std::shared_ptr<op::Parameter> p_node);
/// Nodes are the backbone of the graph of Value dataflow. Every node has
/// zero or more nodes as arguments and one value, which is either a tensor
/// view or a (possibly empty) tuple of values.
......@@ -49,6 +54,9 @@ namespace ngraph
friend class descriptor::Input;
friend void replace_node_users_arguments(std::shared_ptr<Node> target,
std::shared_ptr<Node> replacement);
friend void insert_parameter_split_between(std::shared_ptr<Node> src_node,
std::shared_ptr<Node> dst_node,
std::shared_ptr<op::Parameter> p_node);
protected:
Node(const std::string& node_type, const Nodes& arguments);
......@@ -165,9 +173,21 @@ namespace ngraph
virtual std::vector<std::shared_ptr<Function>> get_functions() const;
// True if this and node have one output with same element type and shape
/// True if this and node have one output with same element type and shape
bool has_same_type(std::shared_ptr<const Node> node) const;
/// Get device placement
Placement get_placement() const;
/// Set device placement
void set_placement(Placement placement);
/// Get input descriptor that is connected to src
descriptor::Input* get_input_from(const std::shared_ptr<Node>& src);
/// Get ouput descriptor that outputs to dst
descriptor::Output* get_output_to(const std::shared_ptr<Node>& dst);
protected:
void add_output(const element::Type& element_type, const Shape& shape);
......@@ -180,9 +200,11 @@ namespace ngraph
std::deque<descriptor::Output> m_outputs;
bool m_is_output;
std::unordered_map<Node*, autodiff::Adjoints> m_adjoint_map;
Placement m_placement = Placement::DEFAULT;
private:
Nodes m_arguments;
//m_arguments still needs to be kept in sync with i/o since get_input_ops
//is pretty ubiquitous and might be called after the original graph was modified.
//get_input_ops uses m_arguments to check if a node view reconstruction from i/o
......
src/ngraph/ops/concat.cpp
View file @ a903a001
......@@ -17,7 +17,7 @@
#include <cassert>
#include <memory>
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/ops/slice.hpp"
using namespace std;
......
src/ngraph/ops/concatenate.cpp deleted 100644 → 0
View file @ ee220ffb
/*******************************************************************************
* Copyright 2017-2018 Intel Corporation
*
* 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 <cassert>
#include <memory>
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/slice.hpp"
using namespace std;
using namespace ngraph;
op::Concat::Concat(const Nodes& args, size_t concatenation_axis)
: RequiresTensorViewArgs("Concat", args)
, m_concatenation_axis(concatenation_axis)
{
if (m_inputs.size() < 1)
{
throw ngraph_error("At least one argument required");
}
auto& input_0 = get_inputs().at(0);
auto input_0_shape = input_0.get_shape();
if (m_concatenation_axis >= input_0_shape.size())
{
throw ngraph_error("Concatenation axis is out of bounds");
}
size_t concatenation_axis_length = input_0_shape.at(m_concatenation_axis);
auto& input_0_element_type = input_0.get_element_type();
for (auto i = 1; i < get_inputs().size(); i++)
{
auto& input_i = get_inputs().at(i);
auto input_i_shape = input_i.get_shape();
if (input_i_shape.size() != input_0_shape.size())
{
throw ngraph_error("Arguments to concat do not have same rank");
}
if (input_i.get_element_type() != input_0_element_type)
{
throw ngraph_error("Argument element types do not match");
}
for (auto j = 0; j < input_i_shape.size(); j++)
{
if (j != m_concatenation_axis && input_0_shape.at(j) != input_i_shape.at(j))
{
throw ngraph_error(
"Arguments to concat do not have same dimension on a non-concatenation axis");
}
else if (j == m_concatenation_axis)
{
concatenation_axis_length += input_i_shape.at(j);
}
}
}
vector<size_t> concatenated_shape = input_0_shape;
concatenated_shape.at(m_concatenation_axis) = concatenation_axis_length;
set_value_type_checked(make_shared<TensorViewType>(input_0_element_type, concatenated_shape));
}
void op::Concat::generate_adjoints(autodiff::Adjoints& adjoints, const std::shared_ptr<Node>& delta)
{
auto concat_result_shape = get_outputs().at(0).get_shape();
Coordinate arg_delta_slice_lower = Coordinate(concat_result_shape.size(), 0);
Coordinate arg_delta_slice_upper = concat_result_shape;
Coordinate arg_delta_slice_strides = Coordinate(concat_result_shape.size(), 1);
size_t pos = 0;
for (auto arg : get_input_ops())
{
auto arg_shape = arg->get_shape();
auto slice_width = arg_shape[m_concatenation_axis];
size_t next_pos = pos + slice_width;
arg_delta_slice_lower[m_concatenation_axis] = pos;
arg_delta_slice_upper[m_concatenation_axis] = next_pos;
adjoints.add_delta(
arg,
make_shared<op::Slice>(
delta, arg_delta_slice_lower, arg_delta_slice_upper, arg_delta_slice_strides));
pos = next_pos;
}
}
src/ngraph/pass/assign_placement.cpp 0 → 100644
View file @ a903a001
/*******************************************************************************
* Copyright 2017-2018 Intel Corporation
*
* 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 "ngraph/pass/assign_placement.hpp"
#include "ngraph/node.hpp"
#include "ngraph/placement.hpp"
using namespace std;
using namespace ngraph;
ngraph::pass::AssignPlacement::AssignPlacement(
std::function<Placement(std::shared_ptr<Node>)> placement_policy)
: m_placement_policy(placement_policy)
{
}
bool ngraph::pass::AssignPlacement::run_on_call_graph(const std::list<std::shared_ptr<Node>>& nodes)
{
for (const std::shared_ptr<Node>& node : nodes)
{
run_on_node(node);
}
return false;
}
bool ngraph::pass::AssignPlacement::run_on_node(shared_ptr<Node> node)
{
node->set_placement(m_placement_policy(node));
return false;
}
src/ngraph/ops/concatenate.hpp → src/ngraph/pass/assign_placement.hpp
View file @ a903a001
......@@ -16,36 +16,26 @@
#pragma once
#include <memory>
#include <exception>
#include <sstream>
#include "ngraph/ops/util/requires_tensor_view_args.hpp"
#include "ngraph/pass/pass.hpp"
#include "ngraph/placement.hpp"
namespace ngraph
{
namespace op
namespace pass
{
/// \brief Concatenation operation.
class Concat : public util::RequiresTensorViewArgs
class AssignPlacement : public CallGraphPass
{
public:
/// \brief Constructs a concatenation operation.
///
/// \param args The nodes producing the input tensors.
/// \param concatenation_axis The axis along which to concatenate the input tensors.
Concat(const Nodes& args, size_t concatenation_axis);
// TODO: make policy a class
AssignPlacement(std::function<Placement(std::shared_ptr<Node>)> placement_policy);
virtual bool run_on_call_graph(const std::list<std::shared_ptr<Node>>& nodes) override;
virtual std::shared_ptr<Node> copy_with_new_args(
const std::vector<std::shared_ptr<Node>>& new_args) const override
{
return std::make_shared<Concat>(new_args, m_concatenation_axis);
}
/// \return The concatenation axis.
size_t get_concatenation_axis() const { return m_concatenation_axis; }
protected:
virtual void generate_adjoints(autodiff::Adjoints& adjoints,
const std::shared_ptr<Node>& delta) override;
const size_t m_concatenation_axis;
private:
bool run_on_node(std::shared_ptr<Node> node);
std::function<Placement(std::shared_ptr<Node>)> m_placement_policy;
};
}
}
src/ngraph/pass/manager.cpp
View file @ a903a001
......@@ -30,6 +30,12 @@ using namespace std;
using namespace ngraph;
ngraph::pass::Manager::Manager()
: m_to_set_is_output(true)
{
}
ngraph::pass::Manager::Manager(bool to_set_is_output)
: m_to_set_is_output(to_set_is_output)
{
}
......@@ -50,11 +56,14 @@ void ngraph::pass::Manager::run_passes(shared_ptr<Function> func)
set<shared_ptr<Function>> tfs(begin(fs), end(fs));
get_state().set_functions(tfs);
for (shared_ptr<Function> f : get_state().get_functions())
if (m_to_set_is_output)
{
for (size_t i = 0; i < f->get_output_size(); ++i)
for (shared_ptr<Function> f : get_state().get_functions())
{
f->get_output_op(i)->set_is_output();
for (size_t i = 0; i < f->get_output_size(); ++i)
{
f->get_output_op(i)->set_is_output();
}
}
}
......
src/ngraph/pass/manager.hpp
View file @ a903a001
......@@ -36,6 +36,7 @@ class ngraph::pass::Manager
{
public:
Manager();
Manager(bool to_set_is_output);
~Manager();
void initialize_default_passes();
......@@ -56,4 +57,5 @@ public:
private:
std::vector<std::shared_ptr<PassBase>> m_pass_list;
ManagerState m_state;
bool m_to_set_is_output;
};
src/ngraph/placement.cpp 0 → 100644
View file @ a903a001
/*******************************************************************************
* Copyright 2017-2018 Intel Corporation
*
* 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 "ngraph/placement.hpp"
std::string ngraph::placement_to_string(Placement placement)
{
switch (placement)
{
case Placement::DEFAULT: return "DEFAULT";
case Placement::INTERPRETER: return "INTERPRETER";
case Placement::CPU: return "CPU";
case Placement::GPU: return "GPU";
case Placement::ARGON: return "ARGON";
}
}
src/ngraph/placement.hpp 0 → 100644
View file @ a903a001
/*******************************************************************************
* Copyright 2017-2018 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/
#pragma once
#include <string>
namespace ngraph
{
enum class Placement
{
DEFAULT,
INTERPRETER,
CPU,
GPU,
ARGON,
};
std::string placement_to_string(Placement placement);
}
src/ngraph/runtime/cpu/cpu_emitter.cpp
View file @ a903a001
This diff is collapsed.
src/ngraph/runtime/cpu/cpu_external_function.cpp
View file @ a903a001
......@@ -43,7 +43,7 @@
#include "ngraph/ops/batch_norm.hpp"
#include "ngraph/ops/broadcast.hpp"
#include "ngraph/ops/ceiling.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/ops/constant.hpp"
#include "ngraph/ops/convert.hpp"
#include "ngraph/ops/convolution.hpp"
......
src/ngraph/runtime/cpu/cpu_tensor_view.cpp
View file @ a903a001
......@@ -107,8 +107,9 @@ void runtime::cpu::CPUTensorView::read(void* target, size_t tensor_offset, size_
auto tvl = this->get_tensor_view_layout();
auto cpu_tvl = dynamic_cast<runtime::cpu::LayoutDescriptor*>(tvl.get());
if (cpu_tvl && cpu_tvl->get_mkldnn_format() != memory::format::format_undef &&
cpu_tvl->get_mkldnn_format() !=
runtime::cpu::mkldnn_utils::CreateNativeDataFormat(*cpu_tvl))
!runtime::cpu::mkldnn_utils::compare_mkldnn_formats(
cpu_tvl->get_mkldnn_format(),
runtime::cpu::mkldnn_utils::CreateNativeDataFormat(*cpu_tvl)))
{
auto tensor_shape = this->get_shape();
auto input_format = cpu_tvl->get_mkldnn_format();
......
src/ngraph/runtime/cpu/mkldnn_utils.cpp
View file @ a903a001
......@@ -25,12 +25,15 @@
#include "ngraph/ops/convolution.hpp"
#include "ngraph/ops/max_pool.hpp"
#include "ngraph/ops/relu.hpp"
#include "ngraph/runtime/cpu/cpu_layout_descriptor.hpp"
#include "ngraph/runtime/cpu/cpu_op_annotations.hpp"
#include "ngraph/types/element_type.hpp"
#include "mkldnn_utils.hpp"
using namespace mkldnn;
using namespace ngraph;
using namespace std;
#define TI(x) std::type_index(typeid(x))
......@@ -94,6 +97,7 @@ static const std::map<memory::format, const std::string> s_mkldnn_format_string_
{memory::format::oIhw16i, "memory::format::oIhw16i"},
{memory::format::OIhw8i8o, "memory::format::OIhw8i8o"},
{memory::format::OIhw16i16o, "memory::format::OIhw16i16o"},
{memory::format::IOhw16o16i, "memory::format::IOhw16o16i"},
{memory::format::OIhw8o8i, "memory::format::OIhw8o8i"},
{memory::format::OIhw16o16i, "memory::format::OIhw16o16i"},
{memory::format::Oihw8o, "memory::format::Oihw8o"},
......@@ -148,3 +152,38 @@ const std::string& runtime::cpu::mkldnn_utils::get_mkldnn_format_string(memory::
std::to_string(fmt));
return it->second;
}
mkldnn::memory::format runtime::cpu::mkldnn_utils::get_input_mkldnn_format(const Node* node,
int index)
{
auto tvl = node->get_inputs()[index].get_output().get_tensor_view()->get_tensor_view_layout();
return dynamic_cast<runtime::cpu::LayoutDescriptor&>(*tvl).get_mkldnn_format();
}
mkldnn::memory::format runtime::cpu::mkldnn_utils::get_output_mkldnn_format(const Node* node,
int index)
{
auto tvl = node->get_output_tensor_view(0)->get_tensor_view_layout();
return dynamic_cast<runtime::cpu::LayoutDescriptor&>(*tvl).get_mkldnn_format();
}
bool runtime::cpu::mkldnn_utils::use_mkldnn_kernel(const ngraph::Node* node)
{
auto op_annotations = static_cast<const ngraph::op::Op*>(node)->get_op_annotations();
return (op_annotations &&
static_pointer_cast<ngraph::runtime::cpu::CPUOpAnnotations>(op_annotations)
->is_mkldnn_op());
}
bool runtime::cpu::mkldnn_utils::compare_mkldnn_formats(mkldnn::memory::format fmt1,
mkldnn::memory::format fmt2)
{
set<mkldnn::memory::format> similar_4d_formats{mkldnn::memory::format::nchw,
mkldnn::memory::format::oihw};
if ((fmt1 == fmt2) || (similar_4d_formats.find(fmt1) != similar_4d_formats.end() &&
similar_4d_formats.find(fmt2) != similar_4d_formats.end()))
{
return true;
}
return false;
}
src/ngraph/runtime/cpu/mkldnn_utils.hpp
View file @ a903a001
......@@ -38,6 +38,12 @@ namespace ngraph
const std::string& get_mkldnn_data_type_string(const ngraph::element::Type& type);
mkldnn::memory::data_type get_mkldnn_data_type(const ngraph::element::Type& type);
const std::string& get_mkldnn_format_string(mkldnn::memory::format fmt);
mkldnn::memory::format get_input_mkldnn_format(const Node* node, int index);
mkldnn::memory::format get_output_mkldnn_format(const Node* node, int index);
bool use_mkldnn_kernel(const ngraph::Node* node);
bool compare_mkldnn_formats(mkldnn::memory::format fmt1,
mkldnn::memory::format fmt2);
}
}
}
......
src/ngraph/runtime/cpu/pass/cpu_assignment.cpp
View file @ a903a001
......@@ -25,7 +25,9 @@
#include <mkldnn.hpp>
#include "ngraph/descriptor/output.hpp"
#include "ngraph/ops/avg_pool.hpp"
#include "ngraph/ops/convolution.hpp"
#include "ngraph/ops/relu.hpp"
#include "ngraph/runtime/cpu/cpu_op_annotations.hpp"
#include "ngraph/runtime/cpu/mkldnn_utils.hpp"
......@@ -66,6 +68,134 @@ namespace ngraph
convolution->set_op_annotations(op_annotations);
}
}
template <>
void CPUAssignment::ASSIGN_DECL(ngraph::op::ConvolutionBackpropData)
{
auto convolution = static_cast<op::ConvolutionBackpropData*>(node);
auto arg0_shape = node->get_input_shape(0);
auto arg1_shape = node->get_input_shape(1);
auto result_shape = node->get_output_shape(0);
auto arg0_rank = arg0_shape.size();
auto arg1_rank = arg1_shape.size();
bool data_dilated = false;
for (size_t s : convolution->get_data_dilation_strides_forward())
{
data_dilated = data_dilated || (s != 1);
}
if (!data_dilated && arg0_rank == 4 && arg1_rank == 4 &&
node->get_input_element_type(0) == element::f32)
{
auto op_annotations =
std::make_shared<ngraph::runtime::cpu::CPUOpAnnotations>();
op_annotations->set_mkldnn_op(true);
convolution->set_op_annotations(op_annotations);
}
}
template <>
void CPUAssignment::ASSIGN_DECL(ngraph::op::ConvolutionBackpropFilters)
{
auto convolution = static_cast<op::ConvolutionBackpropFilters*>(node);
auto arg0_shape = node->get_input_shape(0);
auto arg1_shape = node->get_input_shape(1);
auto result_shape = node->get_output_shape(0);
auto arg0_rank = arg0_shape.size();
auto arg1_rank = arg1_shape.size();
bool data_dilated = false;
for (size_t s : convolution->get_data_dilation_strides_forward())
{
data_dilated = data_dilated || (s != 1);
}
if (!data_dilated && arg0_rank == 4 && arg1_rank == 4 &&
node->get_input_element_type(0) == element::f32)
{
auto op_annotations =
std::make_shared<ngraph::runtime::cpu::CPUOpAnnotations>();
op_annotations->set_mkldnn_op(true);
convolution->set_op_annotations(op_annotations);
}
}
template <>
void CPUAssignment::ASSIGN_DECL(ngraph::op::AvgPool)
{
auto avg_pool = static_cast<op::AvgPool*>(node);
auto arg0_shape = node->get_input_shape(0);
auto arg0_rank = arg0_shape.size();
auto result_shape = node->get_output_shape(0);
if (arg0_rank == 4 && avg_pool->get_window_shape().size() == 2 &&
node->get_input_element_type(0) == element::f32)
{
auto op_annotations =
std::make_shared<ngraph::runtime::cpu::CPUOpAnnotations>();
op_annotations->set_mkldnn_op(true);
avg_pool->set_op_annotations(op_annotations);
}
}
template <>
void CPUAssignment::ASSIGN_DECL(ngraph::op::AvgPoolBackprop)
{
auto avg_pool = static_cast<op::AvgPoolBackprop*>(node);
auto arg0_shape = node->get_input_shape(0);
auto arg0_rank = arg0_shape.size();
auto result_shape = node->get_output_shape(0);
if (arg0_rank == 4 && avg_pool->get_window_shape().size() == 2 &&
node->get_input_element_type(0) == element::f32)
{
auto op_annotations =
std::make_shared<ngraph::runtime::cpu::CPUOpAnnotations>();
op_annotations->set_mkldnn_op(true);
avg_pool->set_op_annotations(op_annotations);
}
}
template <>
void CPUAssignment::ASSIGN_DECL(ngraph::op::Relu)
{
auto avg_pool = static_cast<op::Relu*>(node);
auto arg0_shape = node->get_input_shape(0);
auto arg0_rank = arg0_shape.size();
auto result_shape = node->get_output_shape(0);
if (arg0_rank == 4 && node->get_input_element_type(0) == element::f32)
{
auto op_annotations =
std::make_shared<ngraph::runtime::cpu::CPUOpAnnotations>();
op_annotations->set_mkldnn_op(true);
avg_pool->set_op_annotations(op_annotations);
}
}
template <>
void CPUAssignment::ASSIGN_DECL(ngraph::op::ReluBackprop)
{
auto avg_pool = static_cast<op::ReluBackprop*>(node);
auto arg0_shape = node->get_input_shape(0);
auto arg0_rank = arg0_shape.size();
auto result_shape = node->get_output_shape(0);
if (arg0_rank == 4 && node->get_input_element_type(0) == element::f32)
{
auto op_annotations =
std::make_shared<ngraph::runtime::cpu::CPUOpAnnotations>();
op_annotations->set_mkldnn_op(true);
avg_pool->set_op_annotations(op_annotations);
}
}
}
}
}
......@@ -76,6 +206,16 @@ namespace ngraph
static const runtime::cpu::pass::AssignOpMap s_dispatcher{
{TI(ngraph::op::Convolution),
&runtime::cpu::pass::CPUAssignment::assign<ngraph::op::Convolution>},
{TI(ngraph::op::ConvolutionBackpropData),
&runtime::cpu::pass::CPUAssignment::assign<ngraph::op::ConvolutionBackpropData>},
{TI(ngraph::op::ConvolutionBackpropFilters),
&runtime::cpu::pass::CPUAssignment::assign<ngraph::op::ConvolutionBackpropFilters>},
{TI(ngraph::op::AvgPool), &runtime::cpu::pass::CPUAssignment::assign<ngraph::op::AvgPool>},
{TI(ngraph::op::AvgPoolBackprop),
&runtime::cpu::pass::CPUAssignment::assign<ngraph::op::AvgPoolBackprop>},
{TI(ngraph::op::Relu), &runtime::cpu::pass::CPUAssignment::assign<ngraph::op::Relu>},
{TI(ngraph::op::ReluBackprop),
&runtime::cpu::pass::CPUAssignment::assign<ngraph::op::ReluBackprop>},
};
bool runtime::cpu::pass::CPUAssignment::run_on_call_graph(
......
src/ngraph/runtime/cpu/pass/cpu_layout.cpp
View file @ a903a001
This diff is collapsed.
src/ngraph/runtime/cpu/pass/cpu_layout.hpp
View file @ a903a001
......@@ -53,6 +53,13 @@ namespace ngraph
private:
std::shared_ptr<CPU_ExternalFunction> m_external_function;
static std::shared_ptr<Node> insert_input_conversions(
CPU_ExternalFunction* external_function,
std::shared_ptr<Node>& node,
const std::vector<mkldnn::memory::format>& required_formats);
static void set_output_layouts(
std::shared_ptr<Node>& node,
const std::vector<mkldnn::memory::format>& output_formats);
static void set_default_layouts(CPU_ExternalFunction* external_function,
std::shared_ptr<Node> node);
};
......
src/ngraph/runtime/gpu/gpu_emitter.cpp
View file @ a903a001
......@@ -29,7 +29,7 @@
#include "ngraph/node.hpp"
#include "ngraph/ops/broadcast.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/ops/constant.hpp"
#include "ngraph/ops/convolution.hpp"
#include "ngraph/ops/dot.hpp"
......
src/ngraph/runtime/gpu/gpu_external_function.cpp
View file @ a903a001
......@@ -45,7 +45,7 @@
#include "ngraph/ops/atan.hpp"
#include "ngraph/ops/broadcast.hpp"
#include "ngraph/ops/ceiling.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/ops/constant.hpp"
#include "ngraph/ops/convert.hpp"
#include "ngraph/ops/convolution.hpp"
......
src/ngraph/runtime/interpreter/int_call_frame.hpp
View file @ a903a001
......@@ -25,7 +25,7 @@
#include "ngraph/node.hpp"
#include "ngraph/ops/avg_pool.hpp"
#include "ngraph/ops/broadcast.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/ops/constant.hpp"
#include "ngraph/ops/convolution.hpp"
#include "ngraph/ops/dot.hpp"
......
src/ngraph/runtime/interpreter/int_external_function.cpp
View file @ a903a001
......@@ -34,7 +34,7 @@
#include "ngraph/ops/asin.hpp"
#include "ngraph/ops/atan.hpp"
#include "ngraph/ops/broadcast.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/ops/constant.hpp"
#include "ngraph/ops/convert.hpp"
#include "ngraph/ops/cos.hpp"
......
src/ngraph/serializer.cpp
View file @ a903a001
......@@ -25,7 +25,7 @@
#include "ngraph/ops/batch_norm.hpp"
#include "ngraph/ops/broadcast.hpp"
#include "ngraph/ops/ceiling.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/ops/constant.hpp"
#include "ngraph/ops/convert.hpp"
#include "ngraph/ops/convolution.hpp"
......@@ -372,7 +372,7 @@ static shared_ptr<ngraph::Function>
auto padding_below = node_js.at("padding_below").get<vector<size_t>>();
auto padding_above = node_js.at("padding_above").get<vector<size_t>>();
auto include_padding_in_avg_computation =
node_js.at("include_padding_in_avg_computation").get<bool>();
get_or_default<bool>(node_js, "include_padding_in_avg_computation", false);
node = make_shared<op::AvgPoolBackprop>(forward_arg_shape,
args[0],
window_shape,
......@@ -796,6 +796,11 @@ static shared_ptr<ngraph::Function>
throw runtime_error(ss.str());
}
node_map[node_name] = node;
// Typically, it could be unsafe to change the name of a node since it may break nameing
// uniqueness. However, it could sometimes be helpful to use the original name from
// the serialization for debugging.
// node->set_name(node_name);
}
std::vector<std::shared_ptr<Node>> result;
......
src/ngraph/serializer.hpp
View file @ a903a001
......@@ -28,4 +28,19 @@ namespace ngraph
std::string serialize(std::shared_ptr<ngraph::Function>, size_t indent = 0);
std::shared_ptr<ngraph::Function> deserialize(std::istream&);
std::shared_ptr<ngraph::Function> deserialize(const std::string&);
template <typename T>
T get_or_default(nlohmann::json& j, const std::string& key, const T& default_value)
{
T rc;
try
{
rc = j.at(key).get<T>();
}
catch (...)
{
rc = default_value;
}
return rc;
}
}
src/resource/main.cpp
View file @ a903a001
......@@ -100,8 +100,8 @@ int main(int argc, char** argv)
include_paths.push_back({CLANG_BUILTIN_HEADERS_PATH, {}, true});
include_paths.push_back({"/usr/include/x86_64-linux-gnu", {"asm", "sys", "bits", "gnu"}});
include_paths.push_back({"/usr/include", {"asm", "sys", "bits", "gnu"}});
include_paths.push_back({"/usr/include", {"linux", "asm-generic"}});
include_paths.push_back(
{"/usr/include", {"asm", "sys", "bits", "gnu", "linux", "asm-generic"}});
include_paths.push_back({cpp0, {"bits"}});
include_paths.push_back({"/usr/include/c++/4.8.2/x86_64-redhat-linux", {"bits"}});
include_paths.push_back({cpp1, {"bits", "ext", "debug", "backward"}});
......@@ -168,63 +168,41 @@ int main(int argc, char** argv)
if (update_needed)
{
size_t total_size = 0;
size_t total_count = 0;
const string prefix = "pReFiX";
ofstream out(output_path);
out << "#pragma clang diagnostic ignored \"-Weverything\"\n";
out << "#include <vector>\n";
out << "namespace ngraph\n";
out << "{\n";
out << " static const uint8_t header_resources[] =\n";
out << " const std::vector<std::string> builtin_search_paths =\n";
out << " {\n";
vector<pair<size_t, size_t>> offset_size_list;
size_t offset = 0;
size_t total_size = 0;
size_t total_count = 0;
for (const ResourceInfo& path : include_paths)
{
for (const string& header_file : path.files)
out << " \"" << path.search_path << "\",\n";
}
out << " };\n";
out << " const std::vector<std::pair<std::string, std::string>> builtin_headers =\n";
out << " {\n";
for (const ResourceInfo& path : include_paths)
{
for (const string& header_path : path.files)
{
string header_data = read_file_to_string(header_file);
string base_path = header_file.substr(path.search_path.size() + 1);
header_data = rewrite_header(header_data, base_path);
string header_data = read_file_to_string(header_path);
string relative_path = header_path.substr(path.search_path.size() + 1);
header_data = rewrite_header(header_data, relative_path);
// header_data = uncomment(header_data);
total_size += header_data.size();
total_count++;
// data layout is triplet of strings containing:
// 1) search path
// 2) header path within search path
// 3) header data
// all strings are null terminated and the length includes the null
// The + 1 below is to account for the null terminator
dump(out, path.search_path.c_str(), path.search_path.size() + 1);
offset_size_list.push_back({offset, path.search_path.size() + 1});
offset += path.search_path.size() + 1;
dump(out, header_file.c_str(), header_file.size() + 1);
offset_size_list.push_back({offset, header_file.size() + 1});
offset += header_file.size() + 1;
dump(out, header_data.c_str(), header_data.size() + 1);
offset_size_list.push_back({offset, header_data.size() + 1});
offset += header_data.size() + 1;
out << " {";
out << "\"" << header_path << "\",\nR\"" << prefix << "(" << header_data << ")"
<< prefix << "\"},\n";
}
}
out << " };\n";
out << " struct HeaderInfo\n";
out << " {\n";
out << " const char* search_path;\n";
out << " const char* header_path;\n";
out << " const char* header_data;\n";
out << " };\n";
out << " std::vector<HeaderInfo> header_info\n";
out << " {\n";
for (size_t i = 0; i < offset_size_list.size();)
{
out << " {(char*)(&header_resources[" << offset_size_list[i++].first;
out << "]), (char*)(&header_resources[" << offset_size_list[i++].first;
out << "]), (char*)(&header_resources[" << offset_size_list[i++].first << "])},\n";
}
out << " };\n";
out << "}\n";
cout.imbue(locale(""));
cout << "Total size " << total_size << " in " << total_count << " files\n";
......
test/CMakeLists.txt
View file @ a903a001
......@@ -39,6 +39,7 @@ set (SRC
input_output_assign.cpp
main.cpp
op.cpp
graph_partition.cpp
pass_liveness.cpp
pass_manager.cpp
pass_memory_layout.cpp
......@@ -98,7 +99,7 @@ if(NGRAPH_ARGON_ENABLE)
set(BACKEND_NAMES ${BACKEND_NAMES} "ARGON")
set(SRC ${SRC} argon_fusion.cpp)
# enable additional Argon backend test
# enable additional Argon backend test
set(ARGON_ADDITIONAL_BACKEND_TESTS ${ARGON_TRANSFORMER_SOURCE_DIR}/test/test_argon_backend)
message(STATUS "ARGON_ADDITIONAL_BACKEND_TESTS path: ${ARGON_ADDITIONAL_BACKEND_TESTS}")
set(ADDITIONAL_ARGON_TEST
......@@ -107,7 +108,7 @@ if(NGRAPH_ARGON_ENABLE)
${ARGON_ADDITIONAL_BACKEND_TESTS}/test_broadcast.cpp
${ARGON_ADDITIONAL_BACKEND_TESTS}/test_conv.cpp
${ARGON_ADDITIONAL_BACKEND_TESTS}/test_binary_ew.cpp)
# ensures ADDITIONAL_ARGON_TEST are a dependency on argon transformer
add_custom_command(OUTPUT ${ADDITIONAL_ARGON_TEST} DEPENDS ext_argon_transformer COMMAND "")
set(SRC ${SRC} ${ADDITIONAL_ARGON_TEST})
......@@ -193,4 +194,3 @@ add_custom_target(check
style-check
unit-test-check
)
test/autodiff.in.cpp
View file @ a903a001
......@@ -497,6 +497,8 @@ TEST(${BACKEND_NAME}, backwards_broadcast1)
TEST(${BACKEND_NAME}, backwards_concat_vector)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -522,6 +524,8 @@ TEST(${BACKEND_NAME}, backwards_concat_vector)
TEST(${BACKEND_NAME}, backwards_concat_axis_0)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -547,6 +551,8 @@ TEST(${BACKEND_NAME}, backwards_concat_axis_0)
TEST(${BACKEND_NAME}, backwards_concat_axis_1)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -572,6 +578,8 @@ TEST(${BACKEND_NAME}, backwards_concat_axis_1)
TEST(${BACKEND_NAME}, backwards_ceiling)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -632,6 +640,8 @@ TEST(${BACKEND_NAME}, backwards_cos)
TEST(${BACKEND_NAME}, backwards_cosh)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -654,6 +664,8 @@ TEST(${BACKEND_NAME}, backwards_cosh)
TEST(${BACKEND_NAME}, backwards_divide)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -851,6 +863,8 @@ TEST(${BACKEND_NAME}, backwards_exp)
TEST(${BACKEND_NAME}, backwards_floor)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -1000,6 +1014,8 @@ TEST(${BACKEND_NAME}, backwards_parameter)
TEST(${BACKEND_NAME}, backwards_power)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -1068,6 +1084,8 @@ TEST(${BACKEND_NAME}, backwards_relu)
TEST(${BACKEND_NAME}, backwards_replace_slice)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -1113,6 +1131,8 @@ TEST(${BACKEND_NAME}, backwards_reshape)
TEST(${BACKEND_NAME}, backwards_select)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -1147,6 +1167,8 @@ TEST(${BACKEND_NAME}, backwards_select)
TEST(${BACKEND_NAME}, backwards_select_nested)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -1181,6 +1203,8 @@ TEST(${BACKEND_NAME}, backwards_select_nested)
TEST(${BACKEND_NAME}, backwards_sign)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -1235,6 +1259,8 @@ TEST(${BACKEND_NAME}, backwards_sin)
TEST(${BACKEND_NAME}, backwards_sinh)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -1257,9 +1283,10 @@ TEST(${BACKEND_NAME}, backwards_sinh)
TEST(${BACKEND_NAME}, backwards_slice)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
test::Uniform<float> rng(-10.0f, 10.0f);
Shape shape{5, 5};
auto make_graph = [shape]() {
......@@ -1394,6 +1421,8 @@ TEST(${BACKEND_NAME}, backwards_sum_m2v_1)
TEST(${BACKEND_NAME}, backwards_tan)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......@@ -1429,6 +1458,8 @@ TEST(${BACKEND_NAME}, backwards_tan)
TEST(${BACKEND_NAME}, backwards_tanh)
{
SKIP_TEST_FOR("GPU", "${BACKEND_NAME}");
SKIP_TEST_FOR("ARGON", "${BACKEND_NAME}");
auto manager = runtime::Manager::get("${BACKEND_NAME}");
auto backend = manager->allocate_backend();
......
test/backend_performance.cpp
View file @ a903a001
......@@ -24,7 +24,7 @@
#include "ngraph/codegen/execution_engine.hpp"
#include "ngraph/file_util.hpp"
#include "ngraph/log.hpp"
#include "ngraph/ops/concatenate.hpp"
#include "ngraph/ops/concat.hpp"
#include "ngraph/runtime/backend.hpp"
#include "ngraph/runtime/call_frame.hpp"
#include "ngraph/runtime/cpu/cpu_call_frame.hpp"
......
test/backend_test.in.cpp
View file @ a903a001
This diff is collapsed.
test/build_graph.cpp
View file @ a903a001
......@@ -16,7 +16,9 @@
#include "gtest/gtest.h"
#include "ngraph/file_util.hpp"
#include "ngraph/ngraph.hpp"
#include "ngraph/serializer.hpp"
#include "util/test_tools.hpp"
#include <memory>
......
test/convolution_test.in.cpp
View file @ a903a001
This diff is collapsed.
test/graph_partition.cpp 0 → 100644
View file @ a903a001
This diff is collapsed.
test/ref_generators/generate_convolution_ref.py
View file @ a903a001
......@@ -213,11 +213,11 @@ def emit_test(t,f):
TEST (${BACKEND_NAME}, %s)
{
Shape shape_a{%s};
auto A = make_shared<op::Parameter>(element::f32, shape_a);
Shape shape_b{%s};
auto B = make_shared<op::Parameter>(element::f32, shape_b);
Shape shape_r{%s};
auto make_graph = [A, B] {
auto make_graph = [shape_a, shape_b] {
auto A = make_shared<op::Parameter>(element::f32, shape_a);
auto B = make_shared<op::Parameter>(element::f32, shape_b);
return make_shared<Function>(make_shared<op::Convolution>(A, B,
Strides{%s}, // move_strides
Strides{%s}, // filter_dilation
......
test/serialize.cpp
View file @ a903a001
......@@ -106,6 +106,18 @@ TEST(serialize, existing_models)
}
}
TEST(serialize, default_value)
{
json j = {{"test1", 1}, {"test2", 2}};
int x1 = j.at("test1").get<int>();
EXPECT_EQ(x1, 1);
int x2 = get_or_default<int>(j, "test2", 0);
EXPECT_EQ(x2, 2);
int x3 = get_or_default<int>(j, "test3", 3);
EXPECT_EQ(x3, 3);
}
TEST(benchmark, serialize)
{
stopwatch timer;
......
test/util/all_close.hpp
View file @ a903a001
......@@ -69,7 +69,9 @@ namespace ngraph
}
if (a->get_shape() != b->get_shape())
{
return false;
}
return all_close(read_vector<T>(a), read_vector<T>(b), rtol, atol);
}
......
test/util/autodiff/numeric_compare.hpp
View file @ a903a001
......@@ -14,9 +14,13 @@
* limitations under the License.
*******************************************************************************/
#include "ngraph/log.hpp"
#include "ngraph/runtime/manager.hpp"
#include "ngraph/types/element_type.hpp"
#include "util/all_close.hpp"
#include "util/autodiff/backprop_derivative.hpp"
#include "util/autodiff/numeric_derivative.hpp"
#include "util/test_tools.hpp"
template <typename T>
bool autodiff_numeric_compare(const std::shared_ptr<ngraph::runtime::Manager>& manager,
......@@ -27,15 +31,48 @@ bool autodiff_numeric_compare(const std::shared_ptr<ngraph::runtime::Manager>& m
T atol)
{
T delta = static_cast<T>(0.001);
// Use INTERPRETER to compute numerical derivatives
auto interpreter_manager = ngraph::runtime::Manager::get("INTERPRETER");
auto interpreter_backend = interpreter_manager->allocate_backend();
auto f = make_graph();
std::vector<std::shared_ptr<ngraph::runtime::TensorView>> interpreter_args;
for (auto arg : args)
{
auto interpreter_arg = interpreter_backend->make_primary_tensor_view(
arg->get_tensor().get_element_type(), arg->get_shape());
// TODO: copy_data should not require T. Quick fix here for bool used in `Select`
if (arg->get_tensor().get_element_type() == ngraph::element::boolean)
{
copy_data(interpreter_arg, read_vector<char>(arg));
}
else
{
copy_data(interpreter_arg, read_vector<T>(arg));
}
interpreter_args.push_back(interpreter_arg);
}
auto results_num = ngraph::autodiff::numeric_derivative<T>(
manager, backend, f, args, delta, f->get_parameters());
interpreter_manager, interpreter_backend, f, interpreter_args, delta, f->get_parameters());
// Use the backend being tested to compute symbolic derivatives
auto g = make_graph();
auto results_sym =
ngraph::autodiff::backprop_derivative<T>(manager, backend, g, args, g->get_parameters());
return ngraph::test::all_close(results_num, results_sym, rtol, atol);
// Cast to HostTensorView for comparision
std::vector<std::shared_ptr<ngraph::runtime::TensorView>> interpreter_results_sym;
for (auto result : results_sym)
{
auto interpreter_result = interpreter_backend->make_primary_tensor_view(
ngraph::element::from<T>(), result->get_shape());
copy_data(interpreter_result, read_vector<T>(result));
interpreter_results_sym.push_back(interpreter_result);
}
return ngraph::test::all_close(results_num, interpreter_results_sym, rtol, atol);
}
template <typename T>
......@@ -48,6 +85,7 @@ bool autodiff_numeric_compare_selective(
T atol,
const std::vector<bool>& indep_param_mask)
{
// Use INTERPRETER to compute numerical derivatives
std::vector<std::shared_ptr<ngraph::op::Parameter>> f_indep_params;
auto f = make_graph();
......@@ -62,9 +100,30 @@ bool autodiff_numeric_compare_selective(
i++;
}
auto results_num =
ngraph::autodiff::numeric_derivative<T>(manager, backend, f, args, .001f, f_indep_params);
auto interpreter_manager = ngraph::runtime::Manager::get("INTERPRETER");
auto interpreter_backend = interpreter_manager->allocate_backend();
std::vector<std::shared_ptr<ngraph::runtime::TensorView>> interpreter_args;
for (auto arg : args)
{
auto interpreter_arg = interpreter_backend->make_primary_tensor_view(
arg->get_tensor().get_element_type(), arg->get_shape());
// TODO: copy_data should not require T. Quick fix here for bool used in `Select`
if (arg->get_tensor().get_element_type() == ngraph::element::boolean)
{
copy_data(interpreter_arg, read_vector<char>(arg));
}
else
{
copy_data(interpreter_arg, read_vector<T>(arg));
}
interpreter_args.push_back(interpreter_arg);
}
auto results_num = ngraph::autodiff::numeric_derivative<T>(
interpreter_manager, interpreter_backend, f, interpreter_args, .001f, f_indep_params);
// Use the backend being tested to compute symbolic derivatives
std::vector<std::shared_ptr<ngraph::op::Parameter>> g_indep_params;
auto g = make_graph();
......@@ -82,5 +141,15 @@ bool autodiff_numeric_compare_selective(
auto results_sym =
ngraph::autodiff::backprop_derivative<T>(manager, backend, g, args, g_indep_params);
return ngraph::test::all_close(results_num, results_sym, rtol, atol);
// Cast to HostTensorView for comparision
std::vector<std::shared_ptr<ngraph::runtime::TensorView>> interpreter_results_sym;
for (auto result : results_sym)
{
auto interpreter_result = interpreter_backend->make_primary_tensor_view(
ngraph::element::from<T>(), result->get_shape());
copy_data(interpreter_result, read_vector<T>(result));
interpreter_results_sym.push_back(interpreter_result);
}
return ngraph::test::all_close(results_num, interpreter_results_sym, rtol, atol);
}
test/util/test_tools.hpp
View file @ a903a001
This diff is collapsed.
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment