Skip to content

N
ngraph
Unverified Commit 621df65a authored by Ashok Emani's avatar Ashok Emani Committed by GitHub
Browse files
Options

Merge branch 'master' into examples

parents 7d160dad 607bcbc4
Show whitespace changes
Inline Side-by-side
Showing with 104 additions and 180 deletions
src/ngraph/codegen/compiler.cpp
View file @ 621df65a
......@@ -208,19 +208,8 @@ void codegen::StaticCompiler::initialize()
}
// Enable various target features
// Most of these are for Eigen
auto& TO = m_compiler->getInvocation().getTargetOpts();
TO.CPU = sys::getHostCPUName();
TO.FeaturesAsWritten.emplace_back("+sse");
TO.FeaturesAsWritten.emplace_back("+sse2");
TO.FeaturesAsWritten.emplace_back("+sse3");
TO.FeaturesAsWritten.emplace_back("+ssse3");
TO.FeaturesAsWritten.emplace_back("+sse4.1");
TO.FeaturesAsWritten.emplace_back("+sse4.2");
TO.FeaturesAsWritten.emplace_back("+avx");
TO.FeaturesAsWritten.emplace_back("+avx2");
TO.FeaturesAsWritten.emplace_back("+fma");
}
codegen::StaticCompiler::~StaticCompiler()
......
src/ngraph/node.cpp
View file @ 621df65a
......@@ -144,42 +144,6 @@ void Node::set_name(const string& name)
}
}
void Node::assert_argument_list_equivalency(const Nodes& b)
{
bool arguments_equal = true;
if (this->m_arguments.size() == b.size())
{
for (size_t i = 0; i < this->m_arguments.size(); i++)
{
arguments_equal = arguments_equal && this->m_arguments.at(i) == b.at(i);
}
}
else
{
arguments_equal = false;
}
if (!arguments_equal)
{
std::cout << "node = " << this->get_name() << std::endl;
std::cout << "m_arguments" << std::endl;
for (auto arg : this->m_arguments)
{
std::cout << "arg = " << arg->get_name() << std::endl;
}
std::cout << "results" << std::endl;
for (auto arg : b)
{
std::cout << "arg = " << arg->get_name() << std::endl;
}
}
if (!arguments_equal)
{
throw "Arguments aren't equal";
}
}
std::shared_ptr<Node> Node::get_input_op(size_t index)
{
for (auto arg : m_arguments)
......@@ -201,7 +165,10 @@ Nodes Node::get_input_ops() //const
result.push_back(i.get_output().get_node());
}
}
assert_argument_list_equivalency(result);
if (m_arguments != result)
{
throw ngraph_error("Arguments aren't equal: different values");
}
return result;
}
......
src/ngraph/node.hpp
View file @ 621df65a
......@@ -170,8 +170,6 @@ namespace ngraph
protected:
void add_output(const element::Type& element_type, const Shape& shape);
void assert_argument_list_equivalency(const Nodes& b);
bool test_identical(const Node&) const;
std::string m_node_type;
std::multiset<Node*> m_users;
......
src/ngraph/runtime/cpu/cpu_emitter.cpp
View file @ 621df65a
This source diff could not be displayed because it is too large. You can view the blob instead.
src/ngraph/runtime/cpu/cpu_emitter.hpp
View file @ 621df65a
......@@ -24,12 +24,12 @@
#include "ngraph/runtime/cpu/cpu_external_function.hpp"
#include "ngraph/runtime/cpu/cpu_tensor_view_wrapper.hpp"
#define EMITTER_DECL(E) \
E(ngraph::runtime::cpu::CPU_ExternalFunction* external_function, \
codegen::CodeWriter& writer, \
#define EMITTER_DECL(op_name) \
emit<op_name>(CPU_ExternalFunction * external_function, \
codegen::CodeWriter & writer, \
const ngraph::Node* node, \
const std::vector<ngraph::runtime::cpu::TensorViewWrapper>& args, \
const std::vector<ngraph::runtime::cpu::TensorViewWrapper>& out)
const std::vector<TensorViewWrapper>& args, \
const std::vector<TensorViewWrapper>& out)
namespace ngraph
{
......@@ -40,72 +40,25 @@ namespace ngraph
class CPU_Emitter
{
public:
static void EMITTER_DECL(EmitNop);
static void EMITTER_DECL(EmitAdd);
#ifdef NGRAPH_DISTRIBUTED
static void EMITTER_DECL(EmitAllReduce);
#endif
static void EMITTER_DECL(EmitDot);
static void EMITTER_DECL(EmitMultiply);
static void EMITTER_DECL(EmitGetOutputElement);
static void EMITTER_DECL(EmitXLAGetTupleElement);
static void EMITTER_DECL(EmitTuple);
static void EMITTER_DECL(EmitAbs);
static void EMITTER_DECL(EmitConcat);
static void EMITTER_DECL(EmitDivide);
static void EMITTER_DECL(EmitEqual);
static void EMITTER_DECL(EmitGreater);
static void EMITTER_DECL(EmitGreaterEq);
static void EMITTER_DECL(EmitLess);
static void EMITTER_DECL(EmitLessEq);
static void EMITTER_DECL(EmitLog);
static void EMITTER_DECL(EmitMaximum);
static void EMITTER_DECL(EmitMinimum);
static void EMITTER_DECL(EmitNegative);
static void EMITTER_DECL(EmitNotEqual);
static void EMITTER_DECL(EmitSelect);
static void EMITTER_DECL(EmitSubtract);
static void EMITTER_DECL(EmitBroadcast);
static void EMITTER_DECL(EmitMatmulBias);
static void EMITTER_DECL(EmitConvert);
static void EMITTER_DECL(EmitConstant);
static void EMITTER_DECL(EmitReshape);
static void EMITTER_DECL(EmitFunctionCall);
static void EMITTER_DECL(EmitReduce);
static void EMITTER_DECL(EmitSign);
static void EMITTER_DECL(EmitSlice);
static void EMITTER_DECL(EmitSum);
static void EMITTER_DECL(EmitExp);
static void EMITTER_DECL(EmitSin);
static void EMITTER_DECL(EmitSinh);
static void EMITTER_DECL(EmitCos);
static void EMITTER_DECL(EmitCosh);
static void EMITTER_DECL(EmitTan);
static void EMITTER_DECL(EmitTanh);
static void EMITTER_DECL(EmitAsin);
static void EMITTER_DECL(EmitAcos);
static void EMITTER_DECL(EmitAtan);
static void EMITTER_DECL(EmitPower);
static void EMITTER_DECL(EmitReplaceSlice);
static void EMITTER_DECL(EmitOneHot);
static void EMITTER_DECL(EmitFloor);
static void EMITTER_DECL(EmitCeiling);
static void EMITTER_DECL(EmitSqrt);
static void EMITTER_DECL(EmitConvolution);
static void EMITTER_DECL(EmitConvolutionBackpropFilters);
static void EMITTER_DECL(EmitConvolutionBackpropData);
static void EMITTER_DECL(EmitNot);
static void EMITTER_DECL(EmitMaxPool);
static void EMITTER_DECL(EmitReverse);
static void EMITTER_DECL(EmitReduceWindow);
static void EMITTER_DECL(EmitSelectAndScatter);
static void EMITTER_DECL(EmitAvgPool);
static void EMITTER_DECL(EmitAvgPoolBackprop);
static void EMITTER_DECL(EmitPad);
static void EMITTER_DECL(EmitBatchNorm);
static void EMITTER_DECL(EmitMaxPoolBackprop);
template <typename OP>
static void emit(CPU_ExternalFunction* external_function,
codegen::CodeWriter& writer,
const ngraph::Node* node,
const std::vector<TensorViewWrapper>& args,
const std::vector<TensorViewWrapper>& out)
{
throw std::runtime_error("Unimplemented op in CPU emitter");
}
static void nop(CPU_ExternalFunction* external_function,
codegen::CodeWriter& writer,
const ngraph::Node* node,
const std::vector<TensorViewWrapper>& args,
const std::vector<TensorViewWrapper>& out)
{
}
static void EmitMKLDNNPreamble(codegen::CodeWriter& writer);
static void emit_mkldnn_preamble(codegen::CodeWriter& writer);
private:
static std::string emit_vector(const TensorViewWrapper&,
......
src/ngraph/runtime/cpu/cpu_external_function.cpp
View file @ 621df65a
......@@ -36,6 +36,7 @@
#include "ngraph/ops/abs.hpp"
#include "ngraph/ops/acos.hpp"
#include "ngraph/ops/add.hpp"
#include "ngraph/ops/allreduce.hpp"
#include "ngraph/ops/asin.hpp"
#include "ngraph/ops/atan.hpp"
#include "ngraph/ops/avg_pool.hpp"
......@@ -54,6 +55,7 @@
#include "ngraph/ops/exp.hpp"
#include "ngraph/ops/floor.hpp"
#include "ngraph/ops/function_call.hpp"
#include "ngraph/ops/get_output_element.hpp"
#include "ngraph/ops/greater.hpp"
#include "ngraph/ops/greater_eq.hpp"
#include "ngraph/ops/less.hpp"
......@@ -67,10 +69,13 @@
#include "ngraph/ops/not.hpp"
#include "ngraph/ops/not_equal.hpp"
#include "ngraph/ops/one_hot.hpp"
#include "ngraph/ops/op.hpp"
#include "ngraph/ops/pad.hpp"
#include "ngraph/ops/parameter.hpp"
#include "ngraph/ops/power.hpp"
#include "ngraph/ops/reduce.hpp"
#include "ngraph/ops/reduce_window.hpp"
#include "ngraph/ops/remainder.hpp"
#include "ngraph/ops/replace_slice.hpp"
#include "ngraph/ops/reshape.hpp"
#include "ngraph/ops/reverse.hpp"
......@@ -155,69 +160,69 @@ static StaticInitializers s_static_initializers;
#define TI(x) type_index(typeid(x))
static const runtime::cpu::OpMap dispatcher{
{TI(ngraph::op::Add), &runtime::cpu::CPU_Emitter::EmitAdd},
{TI(ngraph::op::Add), &runtime::cpu::CPU_Emitter::emit<op::Add>},
#ifdef NGRAPH_DISTRIBUTED
{TI(ngraph::op::AllReduce), &runtime::cpu::CPU_Emitter::EmitAllReduce},
{TI(ngraph::op::AllReduce), &runtime::cpu::CPU_Emitter::emit<op::AllReduce>},
#endif
{TI(ngraph::op::MatmulBias), &runtime::cpu::CPU_Emitter::EmitMatmulBias},
{TI(ngraph::op::Dot), &runtime::cpu::CPU_Emitter::EmitDot},
{TI(ngraph::op::Multiply), &runtime::cpu::CPU_Emitter::EmitMultiply},
{TI(ngraph::op::Parameter), &runtime::cpu::CPU_Emitter::EmitNop},
{TI(ngraph::op::Abs), &runtime::cpu::CPU_Emitter::EmitAbs},
{TI(ngraph::op::Concat), &runtime::cpu::CPU_Emitter::EmitConcat},
{TI(ngraph::op::Divide), &runtime::cpu::CPU_Emitter::EmitDivide},
{TI(ngraph::op::Equal), &runtime::cpu::CPU_Emitter::EmitEqual},
{TI(ngraph::op::Greater), &runtime::cpu::CPU_Emitter::EmitGreater},
{TI(ngraph::op::GreaterEq), &runtime::cpu::CPU_Emitter::EmitGreaterEq},
{TI(ngraph::op::Less), &runtime::cpu::CPU_Emitter::EmitLess},
{TI(ngraph::op::LessEq), &runtime::cpu::CPU_Emitter::EmitLessEq},
{TI(ngraph::op::Log), &runtime::cpu::CPU_Emitter::EmitLog},
{TI(ngraph::op::Maximum), &runtime::cpu::CPU_Emitter::EmitMaximum},
{TI(ngraph::op::Minimum), &runtime::cpu::CPU_Emitter::EmitMinimum},
{TI(ngraph::op::Negative), &runtime::cpu::CPU_Emitter::EmitNegative},
{TI(ngraph::op::NotEqual), &runtime::cpu::CPU_Emitter::EmitNotEqual},
{TI(ngraph::op::Power), &runtime::cpu::CPU_Emitter::EmitPower},
{TI(ngraph::op::Select), &runtime::cpu::CPU_Emitter::EmitSelect},
{TI(ngraph::op::Subtract), &runtime::cpu::CPU_Emitter::EmitSubtract},
{TI(ngraph::op::Broadcast), &runtime::cpu::CPU_Emitter::EmitBroadcast},
{TI(ngraph::op::Convert), &runtime::cpu::CPU_Emitter::EmitConvert},
{TI(ngraph::op::Constant), &runtime::cpu::CPU_Emitter::EmitConstant},
{TI(ngraph::op::Reshape), &runtime::cpu::CPU_Emitter::EmitReshape},
{TI(ngraph::op::FunctionCall), &runtime::cpu::CPU_Emitter::EmitFunctionCall},
{TI(ngraph::op::Reduce), &runtime::cpu::CPU_Emitter::EmitReduce},
{TI(ngraph::op::Sign), &runtime::cpu::CPU_Emitter::EmitSign},
{TI(ngraph::op::Slice), &runtime::cpu::CPU_Emitter::EmitSlice},
{TI(ngraph::op::Sum), &runtime::cpu::CPU_Emitter::EmitSum},
{TI(ngraph::op::Exp), &runtime::cpu::CPU_Emitter::EmitExp},
{TI(ngraph::op::Sin), &runtime::cpu::CPU_Emitter::EmitSin},
{TI(ngraph::op::Sinh), &runtime::cpu::CPU_Emitter::EmitSinh},
{TI(ngraph::op::Cos), &runtime::cpu::CPU_Emitter::EmitCos},
{TI(ngraph::op::Cosh), &runtime::cpu::CPU_Emitter::EmitCosh},
{TI(ngraph::op::Tan), &runtime::cpu::CPU_Emitter::EmitTan},
{TI(ngraph::op::Tanh), &runtime::cpu::CPU_Emitter::EmitTanh},
{TI(ngraph::op::Asin), &runtime::cpu::CPU_Emitter::EmitAsin},
{TI(ngraph::op::Acos), &runtime::cpu::CPU_Emitter::EmitAcos},
{TI(ngraph::op::Atan), &runtime::cpu::CPU_Emitter::EmitAtan},
{TI(ngraph::op::ReplaceSlice), &runtime::cpu::CPU_Emitter::EmitReplaceSlice},
{TI(ngraph::op::OneHot), &runtime::cpu::CPU_Emitter::EmitOneHot},
{TI(ngraph::op::Floor), &runtime::cpu::CPU_Emitter::EmitFloor},
{TI(ngraph::op::Ceiling), &runtime::cpu::CPU_Emitter::EmitCeiling},
{TI(ngraph::op::Sqrt), &runtime::cpu::CPU_Emitter::EmitSqrt},
{TI(ngraph::op::Convolution), &runtime::cpu::CPU_Emitter::EmitConvolution},
{TI(ngraph::op::MatmulBias), &runtime::cpu::CPU_Emitter::emit<op::MatmulBias>},
{TI(ngraph::op::Dot), &runtime::cpu::CPU_Emitter::emit<op::Dot>},
{TI(ngraph::op::Multiply), &runtime::cpu::CPU_Emitter::emit<op::Multiply>},
{TI(ngraph::op::Parameter), &runtime::cpu::CPU_Emitter::nop},
{TI(ngraph::op::Abs), &runtime::cpu::CPU_Emitter::emit<op::Abs>},
{TI(ngraph::op::Concat), &runtime::cpu::CPU_Emitter::emit<op::Concat>},
{TI(ngraph::op::Divide), &runtime::cpu::CPU_Emitter::emit<op::Divide>},
{TI(ngraph::op::Equal), &runtime::cpu::CPU_Emitter::emit<op::Equal>},
{TI(ngraph::op::Greater), &runtime::cpu::CPU_Emitter::emit<op::Greater>},
{TI(ngraph::op::GreaterEq), &runtime::cpu::CPU_Emitter::emit<op::GreaterEq>},
{TI(ngraph::op::Less), &runtime::cpu::CPU_Emitter::emit<op::Less>},
{TI(ngraph::op::LessEq), &runtime::cpu::CPU_Emitter::emit<op::LessEq>},
{TI(ngraph::op::Log), &runtime::cpu::CPU_Emitter::emit<op::Log>},
{TI(ngraph::op::Maximum), &runtime::cpu::CPU_Emitter::emit<op::Maximum>},
{TI(ngraph::op::Minimum), &runtime::cpu::CPU_Emitter::emit<op::Minimum>},
{TI(ngraph::op::Negative), &runtime::cpu::CPU_Emitter::emit<op::Negative>},
{TI(ngraph::op::NotEqual), &runtime::cpu::CPU_Emitter::emit<op::NotEqual>},
{TI(ngraph::op::Power), &runtime::cpu::CPU_Emitter::emit<op::Power>},
{TI(ngraph::op::Select), &runtime::cpu::CPU_Emitter::emit<op::Select>},
{TI(ngraph::op::Subtract), &runtime::cpu::CPU_Emitter::emit<op::Subtract>},
{TI(ngraph::op::Broadcast), &runtime::cpu::CPU_Emitter::emit<op::Broadcast>},
{TI(ngraph::op::Convert), &runtime::cpu::CPU_Emitter::emit<op::Convert>},
{TI(ngraph::op::Constant), &runtime::cpu::CPU_Emitter::emit<op::Constant>},
{TI(ngraph::op::Reshape), &runtime::cpu::CPU_Emitter::emit<op::Reshape>},
{TI(ngraph::op::FunctionCall), &runtime::cpu::CPU_Emitter::emit<op::FunctionCall>},
{TI(ngraph::op::Reduce), &runtime::cpu::CPU_Emitter::emit<op::Reduce>},
{TI(ngraph::op::Sign), &runtime::cpu::CPU_Emitter::emit<op::Sign>},
{TI(ngraph::op::Slice), &runtime::cpu::CPU_Emitter::emit<op::Slice>},
{TI(ngraph::op::Sum), &runtime::cpu::CPU_Emitter::emit<op::Sum>},
{TI(ngraph::op::Exp), &runtime::cpu::CPU_Emitter::emit<op::Exp>},
{TI(ngraph::op::Sin), &runtime::cpu::CPU_Emitter::emit<op::Sin>},
{TI(ngraph::op::Sinh), &runtime::cpu::CPU_Emitter::emit<op::Sinh>},
{TI(ngraph::op::Cos), &runtime::cpu::CPU_Emitter::emit<op::Cos>},
{TI(ngraph::op::Cosh), &runtime::cpu::CPU_Emitter::emit<op::Cosh>},
{TI(ngraph::op::Tan), &runtime::cpu::CPU_Emitter::emit<op::Tan>},
{TI(ngraph::op::Tanh), &runtime::cpu::CPU_Emitter::emit<op::Tanh>},
{TI(ngraph::op::Asin), &runtime::cpu::CPU_Emitter::emit<op::Asin>},
{TI(ngraph::op::Acos), &runtime::cpu::CPU_Emitter::emit<op::Acos>},
{TI(ngraph::op::Atan), &runtime::cpu::CPU_Emitter::emit<op::Atan>},
{TI(ngraph::op::ReplaceSlice), &runtime::cpu::CPU_Emitter::emit<op::ReplaceSlice>},
{TI(ngraph::op::OneHot), &runtime::cpu::CPU_Emitter::emit<op::OneHot>},
{TI(ngraph::op::Floor), &runtime::cpu::CPU_Emitter::emit<op::Floor>},
{TI(ngraph::op::Ceiling), &runtime::cpu::CPU_Emitter::emit<op::Ceiling>},
{TI(ngraph::op::Sqrt), &runtime::cpu::CPU_Emitter::emit<op::Sqrt>},
{TI(ngraph::op::Convolution), &runtime::cpu::CPU_Emitter::emit<op::Convolution>},
{TI(ngraph::op::ConvolutionBackpropFilters),
&runtime::cpu::CPU_Emitter::EmitConvolutionBackpropFilters},
&runtime::cpu::CPU_Emitter::emit<op::ConvolutionBackpropFilters>},
{TI(ngraph::op::ConvolutionBackpropData),
&runtime::cpu::CPU_Emitter::EmitConvolutionBackpropData},
{TI(ngraph::op::Not), &runtime::cpu::CPU_Emitter::EmitNot},
{TI(ngraph::op::MaxPool), &runtime::cpu::CPU_Emitter::EmitMaxPool},
{TI(ngraph::op::Reverse), &runtime::cpu::CPU_Emitter::EmitReverse},
{TI(ngraph::op::ReduceWindow), &runtime::cpu::CPU_Emitter::EmitReduceWindow},
{TI(ngraph::op::SelectAndScatter), &runtime::cpu::CPU_Emitter::EmitSelectAndScatter},
{TI(ngraph::op::AvgPool), &runtime::cpu::CPU_Emitter::EmitAvgPool},
{TI(ngraph::op::AvgPoolBackprop), &runtime::cpu::CPU_Emitter::EmitAvgPoolBackprop},
{TI(ngraph::op::Pad), &runtime::cpu::CPU_Emitter::EmitPad},
{TI(ngraph::op::BatchNorm), &runtime::cpu::CPU_Emitter::EmitBatchNorm},
{TI(ngraph::op::MaxPoolBackprop), &runtime::cpu::CPU_Emitter::EmitMaxPoolBackprop},
&runtime::cpu::CPU_Emitter::emit<op::ConvolutionBackpropData>},
{TI(ngraph::op::Not), &runtime::cpu::CPU_Emitter::emit<op::Not>},
{TI(ngraph::op::MaxPool), &runtime::cpu::CPU_Emitter::emit<op::MaxPool>},
{TI(ngraph::op::Reverse), &runtime::cpu::CPU_Emitter::emit<op::Reverse>},
{TI(ngraph::op::ReduceWindow), &runtime::cpu::CPU_Emitter::emit<op::ReduceWindow>},
{TI(ngraph::op::SelectAndScatter), &runtime::cpu::CPU_Emitter::emit<op::SelectAndScatter>},
{TI(ngraph::op::AvgPool), &runtime::cpu::CPU_Emitter::emit<op::AvgPool>},
{TI(ngraph::op::AvgPoolBackprop), &runtime::cpu::CPU_Emitter::emit<op::AvgPoolBackprop>},
{TI(ngraph::op::Pad), &runtime::cpu::CPU_Emitter::emit<op::Pad>},
{TI(ngraph::op::BatchNorm), &runtime::cpu::CPU_Emitter::emit<op::BatchNorm>},
{TI(ngraph::op::MaxPoolBackprop), &runtime::cpu::CPU_Emitter::emit<op::MaxPoolBackprop>},
};
runtime::cpu::CPU_ExternalFunction::CPU_ExternalFunction(
......@@ -308,7 +313,7 @@ using namespace ngraph::runtime;
if (include_mkldnn_headers)
{
runtime::cpu::CPU_Emitter::EmitMKLDNNPreamble(writer);
runtime::cpu::CPU_Emitter::emit_mkldnn_preamble(writer);
}
string pch_header_source = writer.get_code();
......
test/backend_performance.cpp
View file @ 621df65a
......@@ -74,6 +74,18 @@ TEST(benchmark, mxnet_seq2seq_backward)
run_benchmark(json_path, "CPU", 10);
}
TEST(benchmark, mxnet_sockeye_seq2seq_forward)
{
const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/Sockeye_Seq2Seq_forward.json");
run_benchmark(json_path, "CPU", 10);
}
TEST(benchmark, mxnet_sockeye_seq2seq_backward)
{
const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/Sockeye_Seq2Seq_backward.json");
run_benchmark(json_path, "CPU", 10);
}
//
// Benchmarks a graph that concatenates six 32x1x200 arrays along the middle axis.
//
......
test/models/mxnet/Sockeye_Seq2Seq_backward.json 0 → 100644
View file @ 621df65a
This source diff could not be displayed because it is too large. You can view the blob instead.
test/models/mxnet/Sockeye_Seq2Seq_forward.json 0 → 100644
View file @ 621df65a
This source diff could not be displayed because it is too large. You can view the blob instead.
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