Skip to content

N
ngraph
Commit b025ac0e authored by Amy Zhuang's avatar Amy Zhuang Committed by Michał Karzyński
Browse files
Options

Do not change the offset of the slice when processing in place (#2063)

parent 599fb3e7
Show whitespace changes
Inline Side-by-side
Showing with 58 additions and 45 deletions
src/ngraph/runtime/cpu/cpu_emitter.cpp
View file @ b025ac0e
......@@ -1048,15 +1048,14 @@ namespace ngraph
{
writer << "if (" << args[i].get_name() << " < " << out[0].get_name()
<< " || " << args[i].get_name() << " >= " << out[0].get_name()
<< " + " << out[0].get_size() * out[0].get_element_type().size()
<< ")\n";
<< " + " << out[0].get_size() << ")\n";
writer.block_begin();
writer << "memcpy(" << out[0].get_name() << " + " << offset << ", "
<< args[i].get_name() << ", "
<< args[i].get_size() * out[0].get_element_type().size()
<< ");\n";
writer.block_end();
offset += args[i].get_size() * out[0].get_element_type().size();
offset += args[i].get_size();
}
return;
}
......@@ -2025,11 +2024,10 @@ namespace ngraph
}
writer << "if (" << out[0].get_name() << " < " << args[0].get_name()
<< " || " << out[0].get_name() << " >= " << args[0].get_name()
<< " + " << args[0].get_size() * out[0].get_element_type().size()
<< ")\n";
<< " + " << args[0].get_size() << ")\n";
writer.block_begin();
writer << "memcpy(" << out[0].get_name() << ", " << args[0].get_name()
<< " + " << start * out[0].get_element_type().size() << ", "
<< " + " << start << ", "
<< out[0].get_size() * out[0].get_element_type().size() << ");\n";
writer.block_end();
return;
......
src/ngraph/runtime/cpu/cpu_external_function.cpp
View file @ b025ac0e
......@@ -668,9 +668,6 @@ using namespace ngraph::runtime;
}
}
// In place slice optimization
process_in_place_slice(ordered_ops);
// In place concatenation optimization
process_in_place_concat(ordered_ops);
......@@ -693,19 +690,6 @@ using namespace ngraph::runtime;
writer << "size_t pool_base_ptr = (size_t) ctx->memory_buffers["
<< m_memory_buffer_sizes.size() - 1 << "]->get_ptr();\n";
writer << "\n";
// Add temporaries to the variable name map
for (shared_ptr<Node> node : ordered_ops)
{
for (descriptor::Tensor* tensor : node->liveness_new_list)
{
stringstream ss;
ss << "((" << tensor->get_element_type().c_type_string()
<< "*)(pool_base_ptr + " << tensor->get_pool_offset() << "))";
m_variable_name_map[tensor->get_name()] = ss.str();
m_tensor_roles[tensor->get_name()] = CPUTensorRole::INTERMEDIATE;
}
}
}
writer << "bool* t_en = (bool*)" << current_function->get_name() << "_t_en;\n";
......@@ -749,6 +733,28 @@ using namespace ngraph::runtime;
}
}
// In place slice optimization
process_in_place_slice(ordered_ops);
if (temporaries_used)
{
// Add temporaries to the variable name map
for (shared_ptr<Node> node : ordered_ops)
{
for (descriptor::Tensor* tensor : node->liveness_new_list)
{
stringstream ss;
ss << "((" << tensor->get_element_type().c_type_string()
<< "*)(pool_base_ptr + " << tensor->get_pool_offset() << "))";
m_variable_name_map[tensor->get_name()] = ss.str();
if (m_tensor_roles.find(tensor->get_name()) == m_tensor_roles.end())
{
m_tensor_roles[tensor->get_name()] = CPUTensorRole::INTERMEDIATE;
}
}
}
}
// Add outputs to the variable name map
for (size_t i = 0; i < current_function->get_output_size(); ++i)
{
......@@ -1342,7 +1348,7 @@ void runtime::cpu::CPU_ExternalFunction::propagate_in_place_concat(
auto old_offset = input_tensor->get_pool_offset();
input_tensor->set_pool_offset(offset);
NGRAPH_DEBUG
<< "cpu_external_function, propagate: change offset, old offset is "
<< "cpu_external_function, propagate: concat, change offset, old offset is "
<< old_offset << ", new offset is " << offset << std::endl;
offset += input_tensor->size();
if (auto arg_concat = std::dynamic_pointer_cast<ngraph::op::Concat>(arg))
......@@ -1368,11 +1374,17 @@ void runtime::cpu::CPU_ExternalFunction::process_in_place_slice(
auto in_place_oi_pairs = op_annotations->get_in_place_oi_pairs();
if (in_place_oi_pairs.size() > 0)
{
auto arg = slice->get_argument(0);
auto input = slice->get_input_from(arg);
auto input = &slice->get_inputs().at(0);
auto arg = input->get_output().get_node();
auto index = input->get_output().get_index();
auto input_node = std::dynamic_pointer_cast<ngraph::op::Op>(arg);
auto input_tensor = &input_node->get_output_tensor(index);
if (m_tensor_roles[input_tensor->get_name()] == CPUTensorRole::INPUT)
{
NGRAPH_DEBUG << "cpu_external_function: function input pointer passed to "
"slice, do not change offset.";
continue;
}
auto offset = input_tensor->get_pool_offset();
auto lower_bounds = slice->get_lower_bounds();
auto start = 0, accumulated = 1;
......@@ -1387,7 +1399,7 @@ void runtime::cpu::CPU_ExternalFunction::process_in_place_slice(
auto old_offset = output_tensor->get_pool_offset();
output_tensor->set_pool_offset(offset);
NGRAPH_DEBUG << "cpu_external_function: change offset, old offset is "
NGRAPH_DEBUG << "cpu_external_function: slice, change offset, old offset is "
<< old_offset << ", new offset is " << offset << std::endl;
}
}
......@@ -1459,28 +1471,9 @@ void runtime::cpu::CPU_ExternalFunction::build()
// Build executor
// In place slice optimization
process_in_place_slice(m_function->get_ordered_ops());
// In place concatenation optimization
process_in_place_concat(m_function->get_ordered_ops());
// Intermediates
if (m_function->get_temporary_pool_size())
{
m_memory_buffer_sizes.push_back(m_function->get_temporary_pool_size());
for (auto& node : m_function->get_ordered_ops())
{
for (auto tensor : node->liveness_new_list)
{
intermediates_offsets.emplace_back(tensor_data[tensor->get_name()],
tensor->get_pool_offset());
m_tensor_roles[tensor->get_name()] = CPUTensorRole::INTERMEDIATE;
}
}
}
// Constants
for (auto& node : m_function->get_ordered_ops())
{
......@@ -1509,6 +1502,28 @@ void runtime::cpu::CPU_ExternalFunction::build()
}
}
// In place slice optimization
process_in_place_slice(m_function->get_ordered_ops());
// Intermediates
if (m_function->get_temporary_pool_size())
{
m_memory_buffer_sizes.push_back(m_function->get_temporary_pool_size());
for (auto& node : m_function->get_ordered_ops())
{
for (auto tensor : node->liveness_new_list)
{
intermediates_offsets.emplace_back(tensor_data[tensor->get_name()],
tensor->get_pool_offset());
if (m_tensor_roles.find(tensor->get_name()) == m_tensor_roles.end())
{
m_tensor_roles[tensor->get_name()] = CPUTensorRole::INTERMEDIATE;
}
}
}
}
// Outputs
for (size_t i = 0; i < m_function->get_output_size(); ++i)
{
......
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