experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Add more TraceMe to track function runtime overhead.

Browse Source 
PiperOrigin-RevId: 254091355
This commit is contained in:
Ruoxin Sang 2019-06-19 16:09:32 -07:00 committed by TensorFlower Gardener
parent 3877e8ee47
commit c2890a06ac
3 changed files with 40 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
tensorflow/core/common_runtime/eager/BUILD
View File

@ -245,6 +245,7 @@ cc_library(
":kernel_and_device",
":tensor_handle",
"@com_google_absl//absl/strings",
"//tensorflow/core/profiler/lib:traceme",
] + select({
"//tensorflow:android": [
"//tensorflow/core:android_tensorflow_lib_lite",

20
tensorflow/core/common_runtime/eager/execute.cc
View File

@ -23,6 +23,7 @@ limitations under the License.
// clang-format on
#include "absl/strings/match.h"
#include "absl/strings/str_cat.h"
#include "tensorflow/core/common_runtime/device.h"
#include "tensorflow/core/common_runtime/device_set.h"
#include "tensorflow/core/common_runtime/eager/context.h"
@ -36,6 +37,7 @@ limitations under the License.
#include "tensorflow/core/framework/node_def_util.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/profiler/lib/traceme.h"
#if !defined(IS_MOBILE_PLATFORM)
#include "tensorflow/core/distributed_runtime/eager/eager_client.h"
#include "tensorflow/core/distributed_runtime/eager/remote_execute_node.h"
@ -293,6 +295,8 @@ Status AddInputDevicesToCacheKey(const EagerContext* ctx,
const EagerOperation* op,
std::vector<Device*>* input_dev_ptrs,
Fprint128* cache_key) {
profiler::TraceMe activity("AddInputDevicesToCacheKey",
profiler::TraceMeLevel::kVerbose);
input_dev_ptrs->reserve(op->Inputs().size());
Device* cpu_device = ctx->HostCPU();
for (TensorHandle* tensor_handle : op->Inputs()) {
@ -347,6 +351,8 @@ Status AddInputTensorShapesToCacheKey(
const EagerContext* ctx, const EagerOperation* op,
std::unordered_map<int, TensorShape>* input_tensor_shapes,
Fprint128* cache_key) {
profiler::TraceMe activity("AddInputTensorShapesToCacheKey",
profiler::TraceMeLevel::kVerbose);
for (int i = 0; i < op->Inputs().size(); i++) {
TensorHandle* tensor_handle = op->Inputs()[i];
@ -382,6 +388,8 @@ Status AddInputResourceDtypesAndShapesToCacheKey(
std::unordered_map<int, std::pair<DataType, TensorShape>>*
input_resource_dtypes_shapes,
Fprint128* cache_key) {
profiler::TraceMe activity("AddInputResourceDtypesAndShapesToCacheKey",
profiler::TraceMeLevel::kVerbose);
for (int i = 0; i < op->Inputs().size(); i++) {
TensorHandle* tensor_handle = op->Inputs()[i];
@ -475,6 +483,9 @@ Status ShouldCompileWithXLA(const EagerOperation* op, const Device* device,
Status EagerLocalExecute(EagerOperation* op,
gtl::InlinedVector<TensorHandle*, 2>* retvals,
int* num_retvals) {
profiler::TraceMe activity(
[&] { return absl::StrCat("EagerLocalExecute: ", op->Name()); },
profiler::TraceMeLevel::kInfo);
const string unspecified_device_name("<unspecified>");
EagerContext* ctx = op->EagerContext();
auto status = ctx->GetStatus();
@ -505,6 +516,9 @@ Status EagerLocalExecute(EagerOperation* op,
// Once that is the case, we will be able to write a thin wrapper layer over
// the EagerService that behaves similar to the current
// ClusterFunctionLibraryRuntime/DistributedFunctionLibraryRuntime.
{
profiler::TraceMe activity("EagerCopyToDevice",
profiler::TraceMeLevel::kInfo);
for (int i = 0; i < op->Inputs().size(); i++) {
TensorHandle* input = op->Inputs()[i];
if (input->IsRemote()) {
@ -517,6 +531,7 @@ Status EagerLocalExecute(EagerOperation* op,
handle->Unref();
}
}
}
TF_RETURN_IF_ERROR(
AddInputDevicesToCacheKey(ctx, op, &input_dev_ptrs, &cache_key));
TF_RETURN_IF_ERROR(AddInputTensorShapesToCacheKey(
@ -1038,6 +1053,9 @@ Status MaybeUpdateOpDevice(EagerOperation* op) {
Status EagerExecute(EagerOperation* op,
gtl::InlinedVector<TensorHandle*, 2>* retvals,
int* num_retvals) {
profiler::TraceMe activity(
[&] { return absl::StrCat("EagerExecute: ", op->Name()); },
profiler::TraceMeLevel::kInfo);
TF_RETURN_IF_ERROR(MaybeUpdateOpDevice(op));
bool op_is_local = op->EagerContext()->IsLocal(op->Device());
@ -1063,6 +1081,8 @@ Status EagerKernelExecute(EagerContext* ctx,
StepStats* maybe_step_stats,
GraphCollector* graph_collector,
TensorHandle** retvals, int num_retvals) {
profiler::TraceMe activity("EagerKernelExecute",
profiler::TraceMeLevel::kInfo);
std::vector<Tensor> outputs(1);
// If there are multiple references to a TensorHandle in 'op_inputs' we must

4
tensorflow/core/distributed_runtime/eager/eager_service_impl.cc
View File

@ -180,6 +180,9 @@ Status EagerServiceImpl::ExecuteOp(const Operation& operation,
TF_RETURN_IF_ERROR(op->SetDevice(operation.device().c_str()));
{
profiler::TraceMe activity("EagerService:RemoteTensorHandleInternal",
profiler::TraceMeLevel::kVerbose);
for (const auto& remote_handle : operation.inputs()) {
tensorflow::TensorHandle* handle;
TF_RETURN_IF_ERROR(server_context->GetTensorHandle(
@ -187,6 +190,7 @@ Status EagerServiceImpl::ExecuteOp(const Operation& operation,
op->AddInput(handle);
}
}
for (const auto& attr : operation.attrs()) {
op->MutableAttrs()->Set(attr.first, attr.second);