Skip to content

T
tic
Commit 65246a71 by tqchen Committed by Tianqi Chen
Options

change project to NNVM

parent cd49ed0e
Showing with 129 additions and 98 deletions
nnvm/README.md
# mxnngraph
Prototype of graph optimizer and construction API for next generation core engine in MXNet and beyond
# NNVM: Build deep learning system by parts
## Goal
- Construct graph easily
- Pluggable inference, optimization algorithms
- Can be used as input graph execution in various settings, specifically be able to support mxnet's symbolic execution API
NNVM is not a deep learning library. It is a modular, lightweight library to
help build deep learning libraries efficiently.
## What is it
While most deep learning systems offer end to end solutions,
it is interesting to ask if we can actually assemble a deep learning system by parts.
The goal is to enable hackers can customize optimizations, target platforms and set of operators they care about.
We believe that the modular system is an interesting direction.
The hope is that effective parts can be assembled together just like you assemble your own desktops.
So the customized deep learning solution can be minimax, minimum in terms of dependencies,
while maxiziming the users' need.
NNVM offers one such part, it provides a generic to do generic
computation graph optimization such as memory reduction, device allocation,
operator fusion while being agnostic to the operator
interface defintion and how operators are executed.
NNVM is inspired by LLVM, aiming to be an intermediate representation library
for neural nets and computation graphs in general.
## Deep learning system by parts
This is one way to divide the deep learning system into common parts.
Each can be isolated to a modular part.
- Computation graph definition, manipulation.
- Computation graph intermediate optimization.
- Computation graph execution.
- Operator kernel libraries.
- Imperative task scheduling and parallel task coordination.
We hope that there will be more modular parts in the future,
so system building can be fun and rewarding.
## Links
[MXNet](https://github.com/dmlc/mxnet) will be using NNVM as its intermediate
representation layer for symbolic graphs.
nnvm/include/nngraph/base.h → nnvm/include/nnvm/base.h
/*!
* Copyright (c) 2016 by Contributors
* \file base.h
* \brief Configuation of nngraph as well as basic data structure.
* \brief Configuation of nnvm as well as basic data structure.
*/
#ifndef NNGRAPH_BASE_H_
#define NNGRAPH_BASE_H_
#ifndef NNVM_BASE_H_
#define NNVM_BASE_H_
#include <dmlc/base.h>
#include <dmlc/any.h>
......@@ -12,7 +12,7 @@
#include <dmlc/registry.h>
#include <dmlc/array_view.h>
namespace nngraph {
namespace nnvm {
/*! \brief any type */
using any = dmlc::any;
......@@ -47,6 +47,6 @@ inline const T& get(const any& src) {
return dmlc::get<T>(src);
}
} // namespace nngraph
} // namespace nnvm
#endif // NNGRAPH_BASE_H_
#endif // NNVM_BASE_H_
nnvm/include/nngraph/graph.h → nnvm/include/nnvm/graph.h
/*!
* Copyright (c) 2016 by Contributors
* \file graph.h
* \brief Configuation of nngraph as well as basic data structure.
* \brief Configuation of nnvm as well as basic data structure.
*/
#ifndef NNGRAPH_GRAPH_H_
#define NNGRAPH_GRAPH_H_
#ifndef NNVM_GRAPH_H_
#define NNVM_GRAPH_H_
#include <vector>
#include <string>
......@@ -15,7 +15,7 @@
#include "./base.h"
#include "./node.h"
namespace nngraph {
namespace nnvm {
/*!
* \brief Symbolic computation graph.
......@@ -98,6 +98,6 @@ inline void Graph::DFSVisit(FVisit fvisit) const {
});
}
} // namespace nngraph
} // namespace nnvm
#endif // NNGRAPH_GRAPH_H_
#endif // NNVM_GRAPH_H_
nnvm/include/nngraph/graph_attr_types.h → nnvm/include/nnvm/graph_attr_types.h
......@@ -3,14 +3,14 @@
* \file graph_attr_types.h
* \brief Data structures that can appear in graph attributes.
*/
#ifndef NNGRAPH_GRAPH_ATTR_TYPES_H_
#define NNGRAPH_GRAPH_ATTR_TYPES_H_
#ifndef NNVM_GRAPH_ATTR_TYPES_H_
#define NNVM_GRAPH_ATTR_TYPES_H_
#include <vector>
#include <unordered_map>
#include "./graph.h"
namespace nngraph {
namespace nnvm {
/*!
* \brief Auxililary data structure to index a graph.
......@@ -39,7 +39,7 @@ struct IndexedGraph {
/*! \brief Node data structure in IndexedGraph */
struct Node {
/*! \brief pointer to the source node */
const nngraph::Node* source;
const nnvm::Node* source;
/*! \brief inputs to the node */
array_view<NodeEntry> inputs;
/*! \brief control flow dependencies to the node */
......@@ -68,7 +68,7 @@ struct IndexedGraph {
* \param e The entry to query for index.
* \return the unique index.
*/
inline uint32_t entry_id(const nngraph::NodeEntry& e) const {
inline uint32_t entry_id(const nnvm::NodeEntry& e) const {
return entry_rptr_[node_id(e.node.get())] + e.index;
}
/*!
......@@ -76,7 +76,7 @@ struct IndexedGraph {
* \param node The Node to query for index.
* \return the node index.
*/
inline uint32_t node_id(const nngraph::Node* node) const {
inline uint32_t node_id(const nnvm::Node* node) const {
return node2index_.at(node);
}
/*!
......@@ -92,7 +92,7 @@ struct IndexedGraph {
* \param node The pointer to the Node structure
* \return const reference to the corresponding IndexedGraph::Node
*/
inline const Node& operator[](const nngraph::Node* node) const {
inline const Node& operator[](const nnvm::Node* node) const {
return nodes_[node_id(node)];
}
/*! \return list of argument nodes */
......@@ -113,7 +113,7 @@ struct IndexedGraph {
// index to argument nodes
std::vector<uint32_t> arg_nodes_;
// mapping from node to index.
std::unordered_map<const nngraph::Node*, uint32_t> node2index_;
std::unordered_map<const nnvm::Node*, uint32_t> node2index_;
// CSR pointer of node entries
std::vector<size_t> entry_rptr_;
// space to store input entries of each
......@@ -122,6 +122,6 @@ struct IndexedGraph {
std::vector<uint32_t> control_deps_;
};
} // namespace nngraph
} // namespace nnvm
#endif // NNGRAPH_GRAPH_ATTR_TYPES_H_
#endif // NNVM_GRAPH_ATTR_TYPES_H_
nnvm/include/nngraph/node.h → nnvm/include/nnvm/node.h
/*!
* Copyright (c) 2016 by Contributors
* \file base.h
* \file node.h
* \brief Graph node data structure.
*/
#ifndef NNGRAPH_NODE_H_
#define NNGRAPH_NODE_H_
#ifndef NNVM_NODE_H_
#define NNVM_NODE_H_
#include <memory>
#include <string>
......@@ -13,7 +13,7 @@
#include "./base.h"
#include "./op.h"
namespace nngraph {
namespace nnvm {
// Forward declare node.
class Node;
......@@ -93,6 +93,6 @@ inline uint32_t Node::num_outputs() const {
}
}
} // namespace nngraph
} // namespace nnvm
#endif // NNGRAPH_NODE_H_
#endif // NNVM_NODE_H_
nnvm/include/nngraph/op.h → nnvm/include/nnvm/op.h
......@@ -3,8 +3,8 @@
* \file op.h
* \brief Operator information structor.
*/
#ifndef NNGRAPH_OP_H_
#define NNGRAPH_OP_H_
#ifndef NNVM_OP_H_
#define NNVM_OP_H_
#include <string>
#include <vector>
......@@ -13,7 +13,7 @@
#include <functional>
#include "./base.h"
namespace nngraph {
namespace nnvm {
// forward declarations
class Node;
......@@ -38,18 +38,18 @@ static const int kVarg = -1;
* // registeration of oeprators
* // NOTE that the attr function can register any
* // additional attributes to the operator
* NNGRAPH_REGISTER_OP(add)
* NNVM_REGISTER_OP(add)
* .describe("add two inputs together")
* .set_num_inputs(2)
* .attr<OpKernel>("gpu_kernel", AddKernel);
*
* NNGRAPH_REGISTER_OP(sub)
* NNVM_REGISTER_OP(sub)
* .describe("substract one tensor from another")
* .set_num_inputs(2);
*
* // Can call regster multiple times in different files
* // to register different part of information
* NNGRAPH_REGISTER_OP(sub)
* NNVM_REGISTER_OP(sub)
* .attr<OpKernel>("gpu_kernel", SubKernel);
*
* // get operators from registry.
......@@ -124,7 +124,7 @@ class Op {
* const std::vector<TShape>& ishapes) {
* // we can use the parsed version of param
* // without repeatively parsing the parameter
* const SumParam& param = nngraph::get<SumParam>(attrs.parsed);
* const SumParam& param = nnvm::get<SumParam>(attrs.parsed);
* }
* \endcode
*/
......@@ -234,36 +234,36 @@ class OpMap {
};
// internal macros to make
#define NNGRAPH_STR_CONCAT_(__x, __y) __x##__y
#define NNGRAPH_STR_CONCAT(__x, __y) NNGRAPH_STR_CONCAT_(__x, __y)
#define NNGRAPH_REGISTER_VAR_DEF(OpName) \
static ::nngraph::Op & __make_ ## NNGraphOp ## _ ## OpName
#define NNVM_STR_CONCAT_(__x, __y) __x##__y
#define NNVM_STR_CONCAT(__x, __y) NNVM_STR_CONCAT_(__x, __y)
#define NNVM_REGISTER_VAR_DEF(OpName) \
static ::nnvm::Op & __make_ ## NnvmOp ## _ ## OpName
/*!
* \def NNGRAPH_REGISTER_OP
* \def NNVM_REGISTER_OP
* \brief Register
* This macro must be used under namespace dmlc, and only used once in cc file.
* \param OpName The name of registry
*
* \code
*
* NNGRAPH_REGISTER_OP(add)
* NNVM_REGISTER_OP(add)
* .describe("add two inputs together")
* .set_num_inputs(2)
* .attr<OpKernel>("gpu_kernel", AddKernel);
*
* \endcode
*/
#define NNGRAPH_REGISTER_OP(OpName) \
NNGRAPH_STR_CONCAT(NNGRAPH_REGISTER_VAR_DEF(OpName), __COUNTER__) = \
::dmlc::Registry<::nngraph::Op>::Get()->__REGISTER_OR_GET__(#OpName)
#define NNVM_REGISTER_OP(OpName) \
NNVM_STR_CONCAT(NNVM_REGISTER_VAR_DEF(OpName), __COUNTER__) = \
::dmlc::Registry<::nnvm::Op>::Get()->__REGISTER_OR_GET__(#OpName)
// implementations of template functions after this.
// member function of Op
template<typename ValueType>
inline const OpMap<ValueType>& Op::GetAttr(const std::string& key) {
const any& ref = GetAttrMap(key);
return nngraph::get<OpMap<ValueType> >(ref);
return nnvm::get<OpMap<ValueType> >(ref);
}
template<typename ValueType>
......@@ -282,7 +282,7 @@ inline Op& Op::attr( // NOLINT(*)
<< " previously " << pmap->type().name()
<< " current " << typeid(OpMap<ValueType>).name();
std::vector<std::pair<ValueType, int> >& vec =
nngraph::get<OpMap<ValueType> >(*pmap).data_;
nnvm::get<OpMap<ValueType> >(*pmap).data_;
// resize the value type.
vec.resize(index_ + 1,
std::make_pair(ValueType(), 0));
......@@ -338,6 +338,6 @@ inline const ValueType& OpMap<ValueType>::operator[](const Op* op) const {
return data_[idx].first;
}
} // namespace nngraph
} // namespace nnvm
#endif // NNGRAPH_OP_H_
#endif // NNVM_OP_H_
nnvm/include/nngraph/pass.h → nnvm/include/nnvm/pass.h
......@@ -3,15 +3,15 @@
* \file pass.h
* \brief Pass that can be applied to a graph.
*/
#ifndef NNGRAPH_PASS_H_
#define NNGRAPH_PASS_H_
#ifndef NNVM_PASS_H_
#define NNVM_PASS_H_
#include <vector>
#include <functional>
#include "./base.h"
#include "./graph.h"
namespace nngraph {
namespace nnvm {
/*!
* \brief A PassFunction is a basic "Operator on Graph"
......@@ -90,12 +90,12 @@ struct PassFunctionReg
};
/*!
* \def NNGRAPH_REGISTER_PASS
* \def NNVM_REGISTER_PASS
* \brief Macro to register pass fuctions.
*
* \code
* // example of registering a shape inference pass
* NNGRAPH_REGISTER_PASS(InferShape)
* NNVM_REGISTER_PASS(InferShape)
* .describe("Shape Inference function, generate graph attributes")
* .provide_graph_attr("data_shape")
* .depend_graph_attr("indexed_graph")
......@@ -105,9 +105,9 @@ struct PassFunctionReg
* });
* \endcode
*/
#define NNGRAPH_REGISTER_PASS(name) \
DMLC_REGISTRY_REGISTER(::nngraph::PassFunctionReg, PassFunctionReg, name)
#define NNVM_REGISTER_PASS(name) \
DMLC_REGISTRY_REGISTER(::nnvm::PassFunctionReg, PassFunctionReg, name)
} // namespace nngraph
} // namespace nnvm
#endif // NNGRAPH_PASS_H_
#endif // NNVM_PASS_H_
nnvm/include/nngraph/tuple.h → nnvm/include/nnvm/tuple.h
......@@ -3,15 +3,15 @@
* \file tuple.h
* \brief Data structure Tuple and TShape to store dynamic sized shapes.
*/
#ifndef NNGRAPH_TUPLE_H_
#define NNGRAPH_TUPLE_H_
#ifndef NNVM_TUPLE_H_
#define NNVM_TUPLE_H_
#include <vector>
#include <type_traits>
#include <algorithm>
#include <iostream>
namespace nngraph {
namespace nnvm {
/*! \brief data type to store array index */
typedef uint32_t index_t;
......@@ -329,6 +329,6 @@ class TShape : public Tuple<index_t> {
}
};
} // namespace nngraph
} // namespace nnvm
#endif // NNGRAPH_TUPLE_H_
#endif // NNVM_TUPLE_H_
nnvm/src/core/graph_attr_types.cc
......@@ -3,10 +3,10 @@
* \file graph_attr_types.cc
* \brief Graph node data structure.
*/
#include <nngraph/graph_attr_types.h>
#include <nnvm/graph_attr_types.h>
#include <limits>
namespace nngraph {
namespace nnvm {
// implement constructor from graph
IndexedGraph::IndexedGraph(const Graph &g) {
......@@ -14,7 +14,7 @@ IndexedGraph::IndexedGraph(const Graph &g) {
std::vector<size_t> inputs_rptr{0}, control_rptr{0};
g.DFSVisit([this, &inputs_rptr, &control_rptr]
(const std::shared_ptr<nngraph::Node>& n) {
(const std::shared_ptr<nnvm::Node>& n) {
CHECK_LT(nodes_.size(), std::numeric_limits<uint32_t>::max());
uint32_t nid = static_cast<uint32_t>(nodes_.size());
// nodes_
......@@ -59,4 +59,4 @@ IndexedGraph::IndexedGraph(const Graph &g) {
}
}
} // namespace nngraph
} // namespace nnvm
nnvm/src/core/node.cc
......@@ -3,9 +3,9 @@
* \file node.cc
* \brief Graph node data structure.
*/
#include <nngraph/node.h>
#include <nnvm/node.h>
namespace nngraph {
namespace nnvm {
Node::~Node() {
if (inputs.size() != 0) {
......@@ -42,4 +42,4 @@ std::shared_ptr<Node> Node::Create() {
return std::make_shared<Node>();
}
} // namespace nngraph
} // namespace nnvm
nnvm/src/core/op.cc
......@@ -3,18 +3,18 @@
* \file op.cc
* \brief Support for operator registry.
*/
#include <nngraph/base.h>
#include <nngraph/op.h>
#include <nnvm/base.h>
#include <nnvm/op.h>
#include <atomic>
#include <mutex>
namespace dmlc {
// enable registry
DMLC_REGISTRY_ENABLE(nngraph::Op);
DMLC_REGISTRY_ENABLE(nnvm::Op);
} // namespace dmlc
namespace nngraph {
namespace nnvm {
// single manager of operator information.
struct OpManager {
......@@ -66,4 +66,4 @@ void Op::UpdateAttrMap(const std::string& key,
updater(&value);
}
} // namespace nngraph
} // namespace nnvm
nnvm/src/core/pass.cc
......@@ -3,15 +3,15 @@
* \file pass.cc
* \brief Support for pass registry.
*/
#include <nngraph/pass.h>
#include <nnvm/pass.h>
#include <algorithm>
namespace dmlc {
// enable registry
DMLC_REGISTRY_ENABLE(nngraph::PassFunctionReg);
DMLC_REGISTRY_ENABLE(nnvm::PassFunctionReg);
} // namespace dmlc
namespace nngraph {
namespace nnvm {
const PassFunctionReg* FindPassDep(const std::string&attr_name) {
for (auto* r : dmlc::Registry<PassFunctionReg>::List()) {
......@@ -54,4 +54,4 @@ Graph ApplyPass(const Graph& src,
return g;
}
} // namespace nngraph
} // namespace nnvm
nnvm/src/example/operator.cc
// Copyright (c) 2016 by Contributors
// This is an example on how we can register operator information to NNGRAPH
// This is an example on how we can register operator information to NNVM
#include <nngraph/op.h>
#include <nnvm/op.h>
#include <utility>
NNGRAPH_REGISTER_OP(add)
NNVM_REGISTER_OP(add)
.describe("add two data together")
.set_num_inputs(2)
.attr("inplace_pair", std::make_pair(0, 0));
NNGRAPH_REGISTER_OP(add)
NNVM_REGISTER_OP(add)
.attr<std::string>("nick_name", "plus");
nnvm/src/test_main.cc
// Copyright (c) 2016 by Contributors
#include <nngraph/op.h>
#include <nngraph/graph.h>
#include <nngraph/tuple.h>
#include <nngraph/graph_attr_types.h>
#include <nnvm/op.h>
#include <nnvm/graph.h>
#include <nnvm/tuple.h>
#include <nnvm/graph_attr_types.h>
#include <string>
void test_op() {
using namespace nngraph;
using namespace nnvm;
auto add = Op::Get("add");
auto nick = Op::GetAttr<std::string>("nick_name");
LOG(INFO) << "nick=" << nick[add];
}
void test_tuple() {
using nngraph::Tuple;
using nngraph::TShape;
using nnvm::Tuple;
using nnvm::TShape;
Tuple<int> x{1, 2, 3};
Tuple<int> y{1, 2, 3, 5, 6};
x = std::move(y);
......@@ -27,7 +27,7 @@ void test_tuple() {
std::istringstream is(os.str());
is >> y;
CHECK_EQ(x, y);
Tuple<nngraph::index_t> ss{1, 2, 3};
Tuple<nnvm::index_t> ss{1, 2, 3};
TShape s = ss;
s = std::move(ss);
CHECK((s == TShape{1, 2, 3}));
......@@ -35,8 +35,8 @@ void test_tuple() {
void test_graph() {
nngraph::Graph g;
g.DFSVisit([](const std::shared_ptr<const nngraph::Node>& n){
nnvm::Graph g;
g.DFSVisit([](const std::shared_ptr<const nnvm::Node>& n){
});
}
int main() {
......
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