Skip to content

N
ngraph
Commit 46e0dea7 authored by Jayaram Bobba's avatar Jayaram Bobba
Browse files
Options

Enable optimal layouts on MKLDNN convolution backprop ops

parent d0f8dff2
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 173 additions and 53 deletions
src/ngraph/runtime/cpu/cpu_emitter.cpp
View file @ 46e0dea7
...@@ -2001,11 +2001,7 @@ namespace ngraph ...@@ -2001,11 +2001,7 @@ namespace ngraph
auto arg1_shape = args[1].get_shape(); auto arg1_shape = args[1].get_shape();
auto result_shape = out[0].get_shape(); auto result_shape = out[0].get_shape();
auto op_annotations = if (runtime::cpu::mkldnn_utils::use_mkldnn_kernel(node))
static_cast<const ngraph::op::Op*>(node)->get_op_annotations();
if (op_annotations &&
static_pointer_cast<ngraph::runtime::cpu::CPUOpAnnotations>(op_annotations)
->is_mkldnn_op())
{ {
// For dilation, MKLDNN wants to know how many elements to insert between, not how far // For dilation, MKLDNN wants to know how many elements to insert between, not how far
// apart to space the elements like nGraph. So we have to subtract 1 from each pos. // apart to space the elements like nGraph. So we have to subtract 1 from each pos.
...@@ -2014,22 +2010,13 @@ namespace ngraph ...@@ -2014,22 +2010,13 @@ namespace ngraph
{ {
window_dilation_strides_adjusted.push_back(s - 1); window_dilation_strides_adjusted.push_back(s - 1);
} }
auto input_tvl = node->get_inputs()[0]
.get_output() auto input_format =
.get_tensor_view() runtime::cpu::mkldnn_utils::get_input_mkldnn_format(node, 0);
->get_tensor_view_layout();
auto weights_tvl = node->get_inputs()[1]
.get_output()
.get_tensor_view()
->get_tensor_view_layout();
auto output_tvl = node->get_output_tensor_view(0)->get_tensor_view_layout();
auto input_format = dynamic_cast<runtime::cpu::LayoutDescriptor&>(*input_tvl)
.get_mkldnn_format();
auto weights_format = auto weights_format =
dynamic_cast<runtime::cpu::LayoutDescriptor&>(*weights_tvl) runtime::cpu::mkldnn_utils::get_input_mkldnn_format(node, 1);
.get_mkldnn_format(); auto output_format =
auto output_format = dynamic_cast<runtime::cpu::LayoutDescriptor&>(*output_tvl) runtime::cpu::mkldnn_utils::get_output_mkldnn_format(node, 0);
.get_mkldnn_format();
auto& mkldnn_emitter = external_function->get_mkldnn_emitter(); auto& mkldnn_emitter = external_function->get_mkldnn_emitter();
auto input_data_desc = auto input_data_desc =
...@@ -2091,17 +2078,8 @@ namespace ngraph ...@@ -2091,17 +2078,8 @@ namespace ngraph
auto arg0_shape = args[0].get_shape(); auto arg0_shape = args[0].get_shape();
auto arg1_shape = args[1].get_shape(); auto arg1_shape = args[1].get_shape();
auto result_shape = out[0].get_shape(); auto result_shape = out[0].get_shape();
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 && if (runtime::cpu::mkldnn_utils::use_mkldnn_kernel(node))
args[0].get_element_type() == element::f32)
{ {
const string& elem_type = const string& elem_type =
runtime::cpu::mkldnn_utils::get_mkldnn_data_type_string( runtime::cpu::mkldnn_utils::get_mkldnn_data_type_string(
...@@ -2112,12 +2090,19 @@ namespace ngraph ...@@ -2112,12 +2090,19 @@ namespace ngraph
{ {
window_dilation_strides_adjusted.push_back(s - 1); window_dilation_strides_adjusted.push_back(s - 1);
} }
auto data_format = runtime::cpu::mkldnn_utils::get_input_mkldnn_format(node, 0);
auto delta_format =
runtime::cpu::mkldnn_utils::get_input_mkldnn_format(node, 1);
auto result_format =
runtime::cpu::mkldnn_utils::get_output_mkldnn_format(node, 0);
auto emit_memory_desc = [&writer](const std::string& var, auto emit_memory_desc = [&writer](const std::string& var,
const std::string& shape, const std::string& shape,
const std::string& type, const std::string& type,
const std::string& layout) { const std::string& layout) {
writer << "memory::desc " << var << " = memory::desc({" << shape << "}, " writer << "memory::desc " << var << " = memory::desc({" << shape << "}, "
<< type << ", memory::format::" << layout << ");\n"; << type << ", " << layout << ");\n";
}; };
auto emit_memory = [&writer]( auto emit_memory = [&writer](
...@@ -2135,9 +2120,21 @@ namespace ngraph ...@@ -2135,9 +2120,21 @@ namespace ngraph
writer << "try\n"; writer << "try\n";
writer.block_begin(); writer.block_begin();
writer << "engine cpu_engine = engine(engine::cpu, 0);\n"; writer << "engine cpu_engine = engine(engine::cpu, 0);\n";
emit_memory_desc("data_desc", join(arg0_shape), elem_type, "nchw"); emit_memory_desc(
emit_memory_desc("delta_desc", join(arg1_shape), elem_type, "nchw"); "data_desc",
emit_memory_desc("result_desc", join(result_shape), elem_type, "oihw"); join(arg0_shape),
elem_type,
runtime::cpu::mkldnn_utils::get_mkldnn_format_string(data_format));
emit_memory_desc(
"delta_desc",
join(arg1_shape),
elem_type,
runtime::cpu::mkldnn_utils::get_mkldnn_format_string(delta_format));
emit_memory_desc(
"result_desc",
join(result_shape),
elem_type,
runtime::cpu::mkldnn_utils::get_mkldnn_format_string(result_format));
emit_memory("data", "data_desc", args[0].get_name()); emit_memory("data", "data_desc", args[0].get_name());
emit_memory("delta", "delta_desc", args[1].get_name()); emit_memory("delta", "delta_desc", args[1].get_name());
emit_memory("result", "result_desc", out[0].get_name()); emit_memory("result", "result_desc", out[0].get_name());
...@@ -2202,17 +2199,8 @@ namespace ngraph ...@@ -2202,17 +2199,8 @@ namespace ngraph
auto arg0_shape = args[0].get_shape(); auto arg0_shape = args[0].get_shape();
auto arg1_shape = args[1].get_shape(); auto arg1_shape = args[1].get_shape();
auto result_shape = out[0].get_shape(); auto result_shape = out[0].get_shape();
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 && if (runtime::cpu::mkldnn_utils::use_mkldnn_kernel(node))
args[0].get_element_type() == element::f32)
{ {
const string& elem_type = const string& elem_type =
runtime::cpu::mkldnn_utils::get_mkldnn_data_type_string( runtime::cpu::mkldnn_utils::get_mkldnn_data_type_string(
...@@ -2224,12 +2212,19 @@ namespace ngraph ...@@ -2224,12 +2212,19 @@ namespace ngraph
window_dilation_strides_adjusted.push_back(s - 1); window_dilation_strides_adjusted.push_back(s - 1);
} }
auto weight_format =
runtime::cpu::mkldnn_utils::get_input_mkldnn_format(node, 0);
auto delta_format =
runtime::cpu::mkldnn_utils::get_input_mkldnn_format(node, 1);
auto result_format =
runtime::cpu::mkldnn_utils::get_output_mkldnn_format(node, 0);
auto emit_memory_desc = [&writer](const std::string& var, auto emit_memory_desc = [&writer](const std::string& var,
const std::string& shape, const std::string& shape,
const std::string& type, const std::string& type,
const std::string& layout) { const std::string& layout) {
writer << "memory::desc " << var << " = memory::desc({" << shape << "}, " writer << "memory::desc " << var << " = memory::desc({" << shape << "}, "
<< type << ", memory::format::" << layout << ");\n"; << type << ", " << layout << ");\n";
}; };
auto emit_memory = [&writer]( auto emit_memory = [&writer](
...@@ -2247,9 +2242,21 @@ namespace ngraph ...@@ -2247,9 +2242,21 @@ namespace ngraph
writer << "try\n"; writer << "try\n";
writer.block_begin(); writer.block_begin();
writer << "engine cpu_engine = engine(engine::cpu, 0);\n"; writer << "engine cpu_engine = engine(engine::cpu, 0);\n";
emit_memory_desc("weight_desc", join(arg0_shape), elem_type, "oihw"); emit_memory_desc(
emit_memory_desc("delta_desc", join(arg1_shape), elem_type, "nchw"); "weight_desc",
emit_memory_desc("result_desc", join(result_shape), elem_type, "nchw"); join(arg0_shape),
elem_type,
runtime::cpu::mkldnn_utils::get_mkldnn_format_string(weight_format));
emit_memory_desc(
"delta_desc",
join(arg1_shape),
elem_type,
runtime::cpu::mkldnn_utils::get_mkldnn_format_string(delta_format));
emit_memory_desc(
"result_desc",
join(result_shape),
elem_type,
runtime::cpu::mkldnn_utils::get_mkldnn_format_string(result_format));
emit_memory("weight", "weight_desc", args[0].get_name()); emit_memory("weight", "weight_desc", args[0].get_name());
emit_memory("delta", "delta_desc", args[1].get_name()); emit_memory("delta", "delta_desc", args[1].get_name());
emit_memory("result", "result_desc", out[0].get_name()); emit_memory("result", "result_desc", out[0].get_name());
......
src/ngraph/runtime/cpu/cpu_tensor_view.cpp
View file @ 46e0dea7
...@@ -107,8 +107,9 @@ void runtime::cpu::CPUTensorView::read(void* target, size_t tensor_offset, size_ ...@@ -107,8 +107,9 @@ void runtime::cpu::CPUTensorView::read(void* target, size_t tensor_offset, size_
auto tvl = this->get_tensor_view_layout(); auto tvl = this->get_tensor_view_layout();
auto cpu_tvl = dynamic_cast<runtime::cpu::LayoutDescriptor*>(tvl.get()); auto cpu_tvl = dynamic_cast<runtime::cpu::LayoutDescriptor*>(tvl.get());
if (cpu_tvl && cpu_tvl->get_mkldnn_format() != memory::format::format_undef && if (cpu_tvl && cpu_tvl->get_mkldnn_format() != memory::format::format_undef &&
cpu_tvl->get_mkldnn_format() != !runtime::cpu::mkldnn_utils::compare_mkldnn_formats(
runtime::cpu::mkldnn_utils::CreateNativeDataFormat(*cpu_tvl)) cpu_tvl->get_mkldnn_format(),
runtime::cpu::mkldnn_utils::CreateNativeDataFormat(*cpu_tvl)))
{ {
auto tensor_shape = this->get_shape(); auto tensor_shape = this->get_shape();
auto input_format = cpu_tvl->get_mkldnn_format(); auto input_format = cpu_tvl->get_mkldnn_format();
......
src/ngraph/runtime/cpu/mkldnn_utils.cpp
View file @ 46e0dea7
...@@ -19,18 +19,21 @@ ...@@ -19,18 +19,21 @@
#include <typeinfo> #include <typeinfo>
#include <unordered_set> #include <unordered_set>
#include "ngraph/types/element_type.hpp"
#include "ngraph/node.hpp" #include "ngraph/node.hpp"
#include "ngraph/ops/avg_pool.hpp" #include "ngraph/ops/avg_pool.hpp"
#include "ngraph/ops/batch_norm.hpp" #include "ngraph/ops/batch_norm.hpp"
#include "ngraph/ops/convolution.hpp" #include "ngraph/ops/convolution.hpp"
#include "ngraph/ops/max_pool.hpp" #include "ngraph/ops/max_pool.hpp"
#include "ngraph/ops/relu.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" #include "mkldnn_utils.hpp"
using namespace mkldnn; using namespace mkldnn;
using namespace ngraph; using namespace ngraph;
using namespace std;
#define TI(x) std::type_index(typeid(x)) #define TI(x) std::type_index(typeid(x))
...@@ -120,7 +123,8 @@ mkldnn::memory::format runtime::cpu::mkldnn_utils::CreateNativeDataFormat( ...@@ -120,7 +123,8 @@ mkldnn::memory::format runtime::cpu::mkldnn_utils::CreateNativeDataFormat(
} }
} }
const std::string& runtime::cpu::mkldnn_utils::get_mkldnn_data_type_string(const ngraph::element::Type& type) const std::string&
runtime::cpu::mkldnn_utils::get_mkldnn_data_type_string(const ngraph::element::Type& type)
{ {
auto it = s_mkldnn_data_type_string_map.find(type); auto it = s_mkldnn_data_type_string_map.find(type);
if (it == s_mkldnn_data_type_string_map.end() || it->second.empty()) if (it == s_mkldnn_data_type_string_map.end() || it->second.empty())
...@@ -128,7 +132,8 @@ const std::string& runtime::cpu::mkldnn_utils::get_mkldnn_data_type_string(const ...@@ -128,7 +132,8 @@ const std::string& runtime::cpu::mkldnn_utils::get_mkldnn_data_type_string(const
return it->second; return it->second;
} }
mkldnn::memory::data_type runtime::cpu::mkldnn_utils::get_mkldnn_data_type(const ngraph::element::Type& type) mkldnn::memory::data_type
runtime::cpu::mkldnn_utils::get_mkldnn_data_type(const ngraph::element::Type& type)
{ {
auto it = s_mkldnn_data_type_map.find(type); auto it = s_mkldnn_data_type_map.find(type);
if (it == s_mkldnn_data_type_map.end() || it->second == memory::data_type::data_undef) if (it == s_mkldnn_data_type_map.end() || it->second == memory::data_type::data_undef)
...@@ -146,3 +151,38 @@ const std::string& runtime::cpu::mkldnn_utils::get_mkldnn_format_string(memory:: ...@@ -146,3 +151,38 @@ const std::string& runtime::cpu::mkldnn_utils::get_mkldnn_format_string(memory::
std::to_string(fmt)); std::to_string(fmt));
return it->second; 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;
}
\ No newline at end of file
src/ngraph/runtime/cpu/mkldnn_utils.hpp
View file @ 46e0dea7
...@@ -38,6 +38,12 @@ namespace ngraph ...@@ -38,6 +38,12 @@ namespace ngraph
const std::string& get_mkldnn_data_type_string(const ngraph::element::Type& type); 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); mkldnn::memory::data_type get_mkldnn_data_type(const ngraph::element::Type& type);
const std::string& get_mkldnn_format_string(mkldnn::memory::format fmt); 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 @ 46e0dea7
...@@ -66,6 +66,60 @@ namespace ngraph ...@@ -66,6 +66,60 @@ namespace ngraph
convolution->set_op_annotations(op_annotations); 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);
}
}
} }
} }
} }
...@@ -76,6 +130,10 @@ namespace ngraph ...@@ -76,6 +130,10 @@ namespace ngraph
static const runtime::cpu::pass::AssignOpMap s_dispatcher{ static const runtime::cpu::pass::AssignOpMap s_dispatcher{
{TI(ngraph::op::Convolution), {TI(ngraph::op::Convolution),
&runtime::cpu::pass::CPUAssignment::assign<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>},
}; };
bool runtime::cpu::pass::CPUAssignment::run_on_call_graph( bool runtime::cpu::pass::CPUAssignment::run_on_call_graph(
......
src/ngraph/runtime/cpu/pass/cpu_layout.cpp
View file @ 46e0dea7
This diff is collapsed.
src/ngraph/runtime/cpu/pass/cpu_layout.hpp
View file @ 46e0dea7
...@@ -53,6 +53,13 @@ namespace ngraph ...@@ -53,6 +53,13 @@ namespace ngraph
private: private:
std::shared_ptr<CPU_ExternalFunction> m_external_function; 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, static void set_default_layouts(CPU_ExternalFunction* external_function,
std::shared_ptr<Node> node); std::shared_ptr<Node> node);
}; };
......
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