IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Update debug config verbose option to show the name of tensors allocated (#7407)

Browse Source 
Signed-off-by: Andrew Kwangwoong Park <andrew.kwangwoong.park@intel.com>
This commit is contained in:
Andrew Kwangwoong Park
2021-09-17 21:12:53 +03:00
committed by GitHub
parent 7772aeb3bb
commit 277a23b8e4
11 changed files with 74 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
inference-engine/src/cldnn_engine/cldnn_infer_request.cpp
+9
View File
@@ -12,6 +12,7 @@
#include "cldnn_remote_context.h"
#include "cldnn_executable_network.h"
#include "cldnn_itt.h"
#include "cldnn/runtime/debug_configuration.hpp"
#include <ie_algorithm.hpp>
#include <debug.h>
@@ -622,6 +623,10 @@ void CLDNNInferRequest::allocate_inputs() {
IE_THROW() << "Input layout for " << name << " is not found";
}
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << name << ": input blob]" << std::endl;
}
if (desc.getPrecision() == Precision::I16 || desc.getPrecision() == Precision::U16) {
TensorDesc desc_fp32 = desc;
desc_fp32.setPrecision(Precision::FP32);
@@ -673,6 +678,10 @@ void CLDNNInferRequest::allocate_outputs() {
const cldnn::layout output_layout = m_graph->GetNetwork()->get_output_memory(outputID)->get_layout();
const TensorDesc& desc = no.second->getTensorDesc();
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << no.first << ": output blob]" << std::endl;
}
auto blobPtr = create_device_blob(desc, output_layout);
_deviceOutputs[no.first] = blobPtr;
_outputs[no.first] = blobPtr;
inference-engine/src/cldnn_engine/ops/constant.cpp
+5
View File
@@ -18,6 +18,7 @@
#include "ngraph/op/util/op_types.hpp"
#include "cldnn/primitives/data.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
namespace CLDNNPlugin {
@@ -169,6 +170,10 @@ void CreateConstantOp(Program& p, const std::shared_ptr<ngraph::op::v0::Constant
if (bufIter != p.blobMemCache.end()) {
constPrimID = bufIter->second;
} else {
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << initialconstPrimID << ": constant]" << std::endl;
}
cldnn::memory::ptr mem = p.GetEngine().allocate_memory(constLayout, false);
auto& stream = p.GetEngine().get_program_stream();
cldnn::mem_lock<char> lock{mem, stream};
inference-engine/src/cldnn_engine/ops/ctc_greedy_decoder.cpp
+5
View File
@@ -11,6 +11,7 @@
#include "cldnn/primitives/ctc_greedy_decoder.hpp"
#include "cldnn/primitives/reorder.hpp"
#include "cldnn/primitives/mutable_data.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
#include "transformations/utils/utils.hpp"
@@ -73,6 +74,10 @@ void CreateCommonCTCGreedyDecoderOp(Program& p, const std::shared_ptr<ngraph::No
DefaultFormatForDims(op->get_output_shape(1).size()),
CldnnTensorFromIEDims(op->get_output_shape(1)));
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << layer_type_name_ID(op) << ": mutable data]" << std::endl;
}
shared_memory.emplace_back(p.GetEngine().allocate_memory(mutableLayout));
cldnn::primitive_id ctc_gd_mutable_id_w = layer_type_name_ID(op) + "_md_write";
inference-engine/src/cldnn_engine/ops/non_max_suppression.cpp
+8
View File
@@ -12,6 +12,7 @@
#include "cldnn/primitives/reorder.hpp"
#include "cldnn/primitives/mutable_data.hpp"
#include "cldnn/primitives/non_max_suppression.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
namespace CLDNNPlugin {
@@ -66,6 +67,7 @@ void CreateNonMaxSuppressionIEInternalOp(Program& p, const std::shared_ptr<ngrap
std::size_t num_output = op->get_output_size();
std::vector<cldnn::memory::ptr> shared_memory;
GPU_DEBUG_GET_INSTANCE(debug_config);
switch (num_output) {
case 3: {
auto mutable_precision_second = op->get_output_element_type(2);
@@ -77,6 +79,9 @@ void CreateNonMaxSuppressionIEInternalOp(Program& p, const std::shared_ptr<ngrap
DefaultFormatForDims(op->get_output_shape(2).size()),
CldnnTensorFromIEDims(op->get_output_shape(2)));
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << layer_type_name_ID(op) << ": mutable data]" << std::endl;
}
shared_memory.emplace_back(p.GetEngine().allocate_memory(mutableLayoutSecond));
cldnn::primitive_id non_max_supression_mutable_id_w_second = layer_type_name_ID(op) + "_md_write_second";
@@ -95,6 +100,9 @@ void CreateNonMaxSuppressionIEInternalOp(Program& p, const std::shared_ptr<ngrap
cldnn::format::bfyx,
cldnn::tensor(static_cast<int32_t>(outputIndices), 3, 1, 1));
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << layer_type_name_ID(op) << ": mutable data]" << std::endl;
}
shared_memory.emplace_back(p.GetEngine().allocate_memory(mutableLayoutFirst));
cldnn::primitive_id non_max_supression_mutable_id_w_first = layer_type_name_ID(op) + "_md_write_first";
inference-engine/src/cldnn_engine/ops/proposal.cpp
+5
View File
@@ -9,6 +9,7 @@
#include "cldnn/primitives/proposal.hpp"
#include "cldnn/primitives/mutable_data.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
namespace CLDNNPlugin {
@@ -62,6 +63,10 @@ void CreateProposalOp(Program& p, const std::shared_ptr<ngraph::op::v0::Proposal
DefaultFormatForDims(op->get_output_shape(1).size()),
CldnnTensorFromIEDims(op->get_output_shape(1)));
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << layer_type_name_ID(op) << ": mutable data]" << std::endl;
}
auto shared_memory = p.GetEngine().allocate_memory(mutableLayout);
cldnn::primitive_id proposal_mutable_id_w = layer_type_name_ID(op) + "_md_write";
inference-engine/src/cldnn_engine/ops/topk.cpp
+5
View File
@@ -9,6 +9,7 @@
#include "cldnn/primitives/arg_max_min.hpp"
#include "cldnn/primitives/mutable_data.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
namespace CLDNNPlugin {
@@ -71,6 +72,10 @@ void CreateTopKOp(Program& p, const std::shared_ptr<ngraph::op::v1::TopK>& op) {
DefaultFormatForDims(op->get_output_shape(1).size()),
CldnnTensorFromIEDims(op->get_output_shape(1)));
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << layer_type_name_ID(op) << ": mutable data]" << std::endl;
}
auto shared_memory = p.GetEngine().allocate_memory(mutableLayout);
cldnn::primitive_id argmax_mutable_id_w = layer_type_name_ID(op) + "_md_write";
inference-engine/thirdparty/clDNN/api/cldnn/runtime/debug_configuration.hpp
Vendored
+1 -1
View File
@@ -12,7 +12,7 @@
#define GPU_DEBUG_IF(cond) if (0)
#endif
#define GPU_DEBUG_COUT std::cout << debug_configuration::prefix
#define GPU_DEBUG_COUT std::cout << cldnn::debug_configuration::prefix
// Macro below is inserted to avoid unused variable warning when GPU_DEBUG_CONFIG is OFF
#define GPU_DEBUG_GET_INSTANCE(name) auto name = cldnn::debug_configuration::get_instance(); (void)(name);
inference-engine/thirdparty/clDNN/runtime/memory_pool.cpp
Vendored
+9
View File
@@ -9,6 +9,7 @@
#include "cldnn/runtime/memory.hpp"
#include "cldnn/runtime/engine.hpp"
#include "cldnn/runtime/memory_pool.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
#include <list>
#include <string>
@@ -136,6 +137,10 @@ memory::ptr memory_pool::get_from_non_padded_pool(const layout& layout,
++it;
}
}
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << id << ": output]" << std::endl;
}
// didn't find anything for you? create new resource
auto mem = alloc_memory(layout, type);
{
@@ -174,6 +179,10 @@ memory::ptr memory_pool::get_from_padded_pool(const layout& layout,
memory_record({{id, network_id}}, mem, network_id, type));
return mem;
}
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << id << ": output]" << std::endl;
}
auto mem = alloc_memory(layout, type);
std::list<memory_record> list = {memory_record({{id, network_id}}, mem, network_id, type)};
_padded_pool.emplace(layout, std::move(list));
inference-engine/thirdparty/clDNN/src/impls/ocl/primitive_base.hpp
Vendored
+9
View File
@@ -9,6 +9,7 @@
#include "primitive_inst.h"
#include "cldnn/graph/program.hpp"
#include "cldnn/runtime/error_handler.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
#include "kernel_selector_helper.h"
#include "cldnn/graph/network.hpp"
#include "register.hpp"
@@ -46,6 +47,10 @@ struct typed_primitive_impl_ocl : public typed_primitive_impl<PType> {
_kernels.emplace_back(other._kernels[k]->clone());
}
for (auto& mem : other._intermediates_memory) {
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << _kernel_data.params->layerID << ": internal buf]" << std::endl;
}
auto& engine = _outer.get_program().get_engine();
auto new_mem = engine.allocate_memory(mem->get_layout(), mem->get_allocation_type());
_intermediates_memory.push_back(new_mem);
@@ -75,6 +80,10 @@ struct typed_primitive_impl_ocl : public typed_primitive_impl<PType> {
{1, 1, 1, (tensor::value_type)(size / bpp)}};
auto& eimpl = arg.get_program().get_engine();
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << _kernel_data.params->layerID << ": internal buf]" << std::endl;
}
_intermediates_memory.push_back(eimpl.allocate_memory(expected_layout));
}
}
inference-engine/thirdparty/clDNN/src/primitive_inst.cpp
Vendored
+14 -1
View File
@@ -15,6 +15,7 @@
#include "cldnn/runtime/memory.hpp"
#include "cldnn/runtime/error_handler.hpp"
#include "cldnn/runtime/debug_configuration.hpp"
#include "json_object.h"
#include <string>
#include <stack>
@@ -174,8 +175,11 @@ memory::ptr primitive_inst::allocate_output() {
allocation_type alloc_type = use_lockable_memory ?
engine.get_lockable_preffered_memory_allocation_type(layout.format.is_image_2d())
: allocation_type::usm_device;
GPU_DEBUG_GET_INSTANCE(debug_config);
if (!_network.is_internal() && (_node.can_be_optimized() || _node.is_type<generic_layer>())) {
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << _node.id() << ": output]" << std::endl;
}
return _network.get_memory_from_pool(layout,
_node.id(),
_node.get_memory_dependencies(),
@@ -183,12 +187,21 @@ memory::ptr primitive_inst::allocate_output() {
false);
} else if (_network.is_internal() && _node.is_output() && _node.is_type<generic_layer>() &&
engine.supports_allocation(allocation_type::usm_device)) {
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << _node.id() << ": output]" << std::endl;
}
return engine.allocate_memory(layout, allocation_type::usm_device, false);
} else if (_network.is_internal() && !_node.is_output() && _node.is_type<input_layout>()) {
// Skip memory reset for input_layout primitives, since data will be copied from cldnn::data primitive
// or just reuse primitive's memory
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << _node.id() << ": constant]" << std::endl;
}
return engine.allocate_memory(layout, alloc_type, false);
} else if (_network.is_internal() || (!_node.can_share_buffer()) || _node.can_be_optimized() || _node.is_output()) {
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << _node.id() << ": output]" << std::endl;
}
return engine.allocate_memory(layout, alloc_type);
} else {
return _network.get_memory_from_pool(layout,
inference-engine/thirdparty/clDNN/src/program.cpp
Vendored
+4
View File
@@ -602,6 +602,10 @@ void program::transfer_memory_to_device() {
if (alloc_type == allocation_type::usm_host || alloc_type == allocation_type::usm_shared) {
GPU_DEBUG_GET_INSTANCE(debug_config);
GPU_DEBUG_IF(debug_config->verbose >= 2) {
GPU_DEBUG_COUT << "[" << data_node.id() << ": constant]" << std::endl;
}
// Allocate and transfer memory
auto device_mem = mem.get_engine()->allocate_memory(data_node_layout, allocation_type::usm_device, false);
device_mem->copy_from(get_stream(), mem);