Skip to content

[LTC] Replace XLA_TIMED with TORCH_LAZY_TIMED #4250

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Nov 29, 2022
+29 −16
Merged

[LTC] Replace XLA_TIMED with TORCH_LAZY_TIMED #4250

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
55a2dbe
[LTC] Replace XLA_TIMED with TORCH_LAZY_TIMED
alanwaketan Nov 28, 2022
4e962cf
Fix linters
alanwaketan Nov 28, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
18 changes: 14 additions & 4 deletions test/test_metrics.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,15 +134,25 @@ def test_metrics_report(self):
self.assertIn("InputOutputAliasCount", metric_names)
self.assertNotEqual(met.metric_data("InputOutputAliasCount"), None)

met.clear_metrics()
self.assertNotIn("InputOutputAliasCount", met.metric_names())
self.assertEqual(met.metric_data("InputOutputAliasCount"), None)

# timed metrics
self.assertIn("TensorToData", report)
self.assertIn("UnwrapXlaData", report)
self.assertIn("WrapXlaData", report)
self.assertIn("DeviceLockWait", report)
self.assertIn("TensorToData", metric_names)
self.assertNotEqual(met.metric_data("TensorToData"), None)
self.assertIn("UnwrapXlaData", metric_names)
self.assertNotEqual(met.metric_data("UnwrapXlaData"), None)
self.assertIn("WrapXlaData", metric_names)
self.assertNotEqual(met.metric_data("WrapXlaData"), None)
self.assertIn("DeviceLockWait", metric_names)
self.assertNotEqual(met.metric_data("DeviceLockWait"), None)

met.clear_metrics()
self.assertNotIn("InputOutputAliasCount", met.metric_names())
self.assertEqual(met.metric_data("InputOutputAliasCount"), None)
self.assertNotIn("TensorToData", met.metric_names())
self.assertEqual(met.metric_data("TensorToData"), None)

# repeat the same computation and expect to see the CachedCompile counter
t3 = t1 * 2
Expand Down
3 changes: 3 additions & 0 deletions third_party/xla_client/metrics.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -246,6 +246,9 @@ class TimedSection {
int64_t start_;
};

// XLA_TIMED should only be used within xla_client. Please use
// TORCH_LAZY_TIMED in pytorch/xla. For more information, see
// NOTE: [TORCH_LAZY_COUNTER v.s. XLA_COUNTER].
#define XLA_TIMED(name) \
static xla::metrics::Metric* timed_metric = \
new xla::metrics::Metric(name, xla::metrics::MetricFnTime); \
Expand Down
4 changes: 2 additions & 2 deletions torch_xla/csrc/init_python_bindings.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1667,7 +1667,7 @@ void InitXlaModuleBindings(py::module m) {
std::vector<torch::lazy::BackendDataPtr> parameters_data;
torch::lazy::BackendDevice device = torch_xla::GetCurrentDevice();
{
XLA_TIMED("RunCachedGraphInputData");
TORCH_LAZY_TIMED("RunCachedGraphInputData");
// setup the parameters_data
int idx = 0;
for (auto& ivalue : graph_inputs) {
Expand All @@ -1688,7 +1688,7 @@ void InitXlaModuleBindings(py::module m) {
cachedComputation->computation, parameters_data, device);
std::vector<at::Tensor> retlist;
{
XLA_TIMED("RunCachedGraphOutputData");
TORCH_LAZY_TIMED("RunCachedGraphOutputData");
// Convert result back to at::tensor
int i = 0;
for (auto& data : results) {
Expand Down
4 changes: 2 additions & 2 deletions torch_xla/csrc/tensor.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,7 +151,7 @@ xla::util::ExceptionCleanup LockDevice(
TF_VLOG(4) << "Waiting on device barrier for device " << device << " ...";
std::shared_ptr<DeviceLocker> locker;
{
XLA_TIMED("DeviceLockWait");
TORCH_LAZY_TIMED("DeviceLockWait");
locker = DeviceLockerArena::Get()->GetLocker(device);
locker->Lock();
}
Expand Down Expand Up @@ -806,7 +806,7 @@ torch::lazy::Value XLATensor::GetIrValueForTensor(
data = GetDeviceData(tensor, device);
read_only = true;
} else {
XLA_TIMED("IrValueTensorToXlaData");
TORCH_LAZY_TIMED("IrValueTensorToXlaData");
data = TensorToXlaData(tensor, device);
}
return CreateTensorNode(std::move(data), read_only);
Expand Down
16 changes: 8 additions & 8 deletions torch_xla/csrc/tensor_util.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -37,7 +37,7 @@ struct DataAsync {

void TransferToServerAsync(std::shared_ptr<DataAsync> async,
const std::vector<std::string>& devices) {
XLA_TIMED("TransferToServerAsync");
TORCH_LAZY_TIMED("TransferToServerAsync");

std::vector<xla::ComputationClient::DataPtr> async_xla_datas =
xla::ComputationClient::Get()->CreateAsyncDatas(async->source_tensors);
Expand Down Expand Up @@ -652,7 +652,7 @@ void PopulateTensorBuffer(const at::Tensor& tensor,
torch::lazy::BackendDataPtr TensorToXlaData(
const at::Tensor& tensor, const xla::Shape& shape,
const torch::lazy::BackendDevice& device) {
XLA_TIMED("TensorToData");
TORCH_LAZY_TIMED("TensorToData");
if (device.type() == (int8_t)XlaDeviceType::SPMD) {
// When SPMD is enabled, we want to delay the data transfer for XLA
// tensors until the data is sharded. So, we skip the data transfer
Expand Down Expand Up @@ -763,13 +763,13 @@ at::Tensor XlaLiteralToTensorHelper(const xla::Literal& literal,

xla::ComputationClient::DataPtr UnwrapXlaData(
const torch::lazy::BackendDataPtr& data) {
XLA_TIMED("UnwrapXlaData");
TORCH_LAZY_TIMED("UnwrapXlaData");
return dynamic_cast<XLAData*>(data.get())->xla_data();
}

std::vector<xla::ComputationClient::DataPtr> UnwrapXlaData(
absl::Span<const torch::lazy::BackendDataPtr> datas) {
XLA_TIMED("UnwrapXlaData");
TORCH_LAZY_TIMED("UnwrapXlaData");
std::vector<xla::ComputationClient::DataPtr> xla_datas;
xla_datas.reserve(datas.size());
for (const auto& data : datas) {
Expand All @@ -780,13 +780,13 @@ std::vector<xla::ComputationClient::DataPtr> UnwrapXlaData(

torch::lazy::BackendDataPtr WrapXlaData(
const xla::ComputationClient::DataPtr& xla_data) {
XLA_TIMED("WrapXlaData");
TORCH_LAZY_TIMED("WrapXlaData");
return std::make_shared<XLAData>(xla_data);
}

std::vector<torch::lazy::BackendDataPtr> WrapXlaData(
absl::Span<const xla::ComputationClient::DataPtr> xla_datas) {
XLA_TIMED("WrapXlaData");
TORCH_LAZY_TIMED("WrapXlaData");
std::vector<torch::lazy::BackendDataPtr> datas;
datas.reserve(xla_datas.size());
for (const auto& xla_data : xla_datas) {
Expand Down Expand Up @@ -868,7 +868,7 @@ torch::lazy::BackendDataPtr TensorToXlaData(
std::vector<torch::lazy::BackendDataPtr> CreateTensorsData(
const std::vector<at::Tensor>& tensors,
const std::vector<std::string>& devices, bool transfer_async) {
XLA_TIMED("TensorToData");
TORCH_LAZY_TIMED("TensorToData");
XLA_CHECK_EQ(tensors.size(), devices.size());
if (transfer_async) {
std::shared_ptr<DataAsync> async = std::make_shared<DataAsync>();
Expand Down Expand Up @@ -918,7 +918,7 @@ std::vector<torch::lazy::BackendDataPtr> CreateTensorsData(
const std::vector<at::Tensor>& tensors,
const std::vector<XLATensor::ShardingSpecPtr>& shardings,
const std::vector<std::string>& devices) {
XLA_TIMED("TensorToData");
TORCH_LAZY_TIMED("TensorToData");
XLA_CHECK_EQ(tensors.size(), shardings.size());
XLA_CHECK_EQ(tensors.size(), devices.size());

Expand Down