Skip to content
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

Use upstream XLA concurrency utilities #5799

Merged
merged 7 commits into from
Nov 15, 2023
Merged

Use upstream XLA concurrency utilities #5799

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
bf4ff6f
Use TSL threadpool
will-cromar Nov 8, 2023
59de529
remove multiwait
will-cromar Nov 9, 2023
48e6333
fix test build
will-cromar Nov 9, 2023
b688515
Move threadpool namespace
will-cromar Nov 14, 2023
8198cb4
formatting
will-cromar Nov 14, 2023
72c1361
fix test build
will-cromar Nov 14, 2023
0712756
Use BlockingCounter
will-cromar Nov 15, 2023
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
4 changes: 2 additions & 2 deletions test/cpp/BUILD
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -75,12 +75,12 @@ ptxla_cc_test(
srcs = ["test_replication.cpp"],
deps = [
":cpp_test_util",
":thread_pool",
":torch_xla_test",
"//torch_xla/csrc/runtime:runtime",
"//torch_xla/csrc/runtime:debug_macros",
"//torch_xla/csrc/runtime:multi_wait",
"//torch_xla/csrc/runtime:thread_pool",
"//torch_xla/csrc:tensor",
"@com_google_absl//absl/synchronization",
"@com_google_googletest//:gtest_main",
"@xla//xla:shape_util",
"@xla//xla/client:xla_builder",
Expand Down
10 changes: 5 additions & 5 deletions test/cpp/test_replication.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,15 +3,15 @@

#include <iostream>

#include "absl/synchronization/blocking_counter.h"
#include "test/cpp/cpp_test_util.h"
#include "test/cpp/torch_xla_test.h"
#include "torch_xla/csrc/aten_xla_bridge.h"
#include "torch_xla/csrc/helpers.h"
#include "torch_xla/csrc/runtime/debug_macros.h"
#include "torch_xla/csrc/runtime/multi_wait.h"
#include "torch_xla/csrc/runtime/runtime.h"
#include "torch_xla/csrc/runtime/thread_pool.h"
#include "torch_xla/csrc/tensor_util.h"
#include "torch_xla/csrc/thread_pool.h"
#include "torch_xla/csrc/torch_util.h"
#include "xla/client/xla_builder.h"
#include "xla/shape_util.h"
Expand Down Expand Up @@ -57,7 +57,7 @@ void TestSingleReplication(

std::vector<std::vector<torch_xla::runtime::ComputationClient::DataPtr>>
results(device_strings.size());
torch_xla::runtime::util::MultiWait mwait(device_strings.size());
absl::BlockingCounter mwait(device_strings.size());
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

let's rename it to bc or something, mwait would be confusing if the type is not MultiWait.

torch_xla::runtime::ComputationClient::ExecuteComputationOptions exec_options;
for (size_t i = 0; i < device_strings.size(); ++i) {
auto executor = [&, i]() {
Expand All @@ -68,9 +68,9 @@ void TestSingleReplication(
torch_xla::runtime::ComputationClient::Data>(
tensors_data[i])},
device_strings[i], exec_options);
mwait.DecrementCount();
};
torch_xla::runtime::env::ScheduleIoClosure(
mwait.Completer(std::move(executor)));
torch_xla::thread::Schedule(std::move(executor));
}
mwait.Wait();

Expand Down
13 changes: 11 additions & 2 deletions torch_xla/csrc/BUILD
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -269,15 +269,14 @@ ptxla_cc_library(
"//torch_xla/csrc/runtime:metrics",
"//torch_xla/csrc/runtime:metrics_analysis",
"//torch_xla/csrc/runtime:metrics_reader",
"//torch_xla/csrc/runtime:multi_wait",
"//torch_xla/csrc/runtime:profiler",
"//torch_xla/csrc/runtime:sys_util",
"//torch_xla/csrc/runtime:thread_pool",
"//torch_xla/csrc/runtime:util",
"//torch_xla/csrc/runtime:xla_coordinator",
"//torch_xla/csrc/runtime:xla_util",
"@com_google_absl//absl/container:flat_hash_map",
"@com_google_absl//absl/strings",
"@com_google_absl//absl/synchronization",
"@com_google_absl//absl/types:variant",
"@tsl//tsl/profiler/lib:traceme",
"@tsl//tsl/profiler/lib:traceme_encode",
Expand Down Expand Up @@ -320,6 +319,16 @@ cc_library(
],
)

cc_library(
name = "thread_pool",
srcs = ["thread_pool.cc"],
hdrs = ["thread_pool.h"],
deps = [
"//torch_xla/csrc/runtime:sys_util",
"@tsl//tsl/platform:env"
],
)

ptxla_cc_library(
name = "unwrap_data",
srcs = ["unwrap_data.cpp"],
Expand Down
3 changes: 1 addition & 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 @@ -20,6 +20,7 @@

#include "absl/container/flat_hash_map.h"
#include "absl/strings/str_cat.h"
#include "absl/synchronization/blocking_counter.h"
#include "absl/types/variant.h"
#include "pybind11/attr.h"
#include "pybind11/cast.h"
Expand All @@ -43,11 +44,9 @@
#include "torch_xla/csrc/runtime/metrics.h"
#include "torch_xla/csrc/runtime/metrics_analysis.h"
#include "torch_xla/csrc/runtime/metrics_reader.h"
#include "torch_xla/csrc/runtime/multi_wait.h"
#include "torch_xla/csrc/runtime/profiler.h"
#include "torch_xla/csrc/runtime/runtime.h"
#include "torch_xla/csrc/runtime/sys_util.h"
#include "torch_xla/csrc/runtime/thread_pool.h"
#include "torch_xla/csrc/runtime/util.h"
#include "torch_xla/csrc/runtime/xla_coordinator.h"
#include "torch_xla/csrc/runtime/xla_util.h"
Expand Down
23 changes: 2 additions & 21 deletions torch_xla/csrc/runtime/BUILD
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,13 +82,12 @@ cc_library(
":computation_client",
":debug_macros",
":env_vars",
":multi_wait",
":profiler",
":stablehlo_helper",
":tensor_source",
":tf_logging",
":thread_pool",
":xla_coordinator",
"//torch_xla/csrc:thread_pool",
"@xla//xla:literal",
"@xla//xla:shape_util",
"@xla//xla/client:xla_computation",
Expand All @@ -102,6 +101,7 @@ cc_library(
"@tsl//tsl/profiler/lib:traceme",
"@tsl//tsl/platform/cloud:gcs_file_system",
"@com_google_absl//absl/strings",
"@com_google_absl//absl/synchronization",
"@com_google_absl//absl/types:span",
],
)
Expand Down Expand Up @@ -187,15 +187,6 @@ cc_library(
],
)

cc_library(
name = "multi_wait",
srcs = ["multi_wait.cc"],
hdrs = ["multi_wait.h"],
deps = [
"@xla//xla:types",
],
)

# Profiler silently fails unless we link these backends
cc_library(
name = "profiler_backends",
Expand Down Expand Up @@ -279,16 +270,6 @@ cc_library(
],
)

cc_library(
name = "thread_pool",
srcs = ["thread_pool.cc"],
hdrs = ["thread_pool.h"],
deps = [
":metrics",
":tf_logging",
],
)

cc_library(
name = "tensor_source",
hdrs = ["tensor_source.h"],
Expand Down
73 changes: 0 additions & 73 deletions torch_xla/csrc/runtime/multi_wait.cc
View file
Open in desktop

This file was deleted.

60 changes: 0 additions & 60 deletions torch_xla/csrc/runtime/multi_wait.h
View file
Open in desktop

This file was deleted.

31 changes: 14 additions & 17 deletions torch_xla/csrc/runtime/pjrt_computation_client.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,18 +5,18 @@
#include <vector>

#include "absl/strings/ascii.h"
#include "absl/synchronization/blocking_counter.h"
#include "absl/types/span.h"
#include "pjrt_computation_client.h"
#include "torch_xla/csrc/runtime/computation_client.h"
#include "torch_xla/csrc/runtime/debug_macros.h"
#include "torch_xla/csrc/runtime/env_vars.h"
#include "torch_xla/csrc/runtime/multi_wait.h"
#include "torch_xla/csrc/runtime/profiler.h"
#include "torch_xla/csrc/runtime/stablehlo_helper.h"
#include "torch_xla/csrc/runtime/tensor_source.h"
#include "torch_xla/csrc/runtime/tf_logging.h"
#include "torch_xla/csrc/runtime/thread_pool.h"
#include "torch_xla/csrc/runtime/xla_coordinator.h"
#include "torch_xla/csrc/thread_pool.h"
#include "tsl/profiler/lib/traceme.h"
#include "xla/client/xla_builder.h"
#include "xla/client/xla_computation.h"
Expand Down Expand Up @@ -619,9 +619,9 @@ PjRtComputationClient::ExecuteComputation(
}
CreateDataHandlesCounter()->AddValue(datas.size());

auto mwait = std::make_shared<util::MultiWait>(1);
auto lockfn = [&, this, device, returned_future = std::move(*returned_future),
timed]() mutable {
thread::Schedule(std::move([&, this, device,
returned_future = std::move(*returned_future),
timed]() mutable {
TF_VLOG(5) << "ExecuteComputation acquiring PJRT device lock for "
<< device;
// Grab the shared lock and block the `WaitDeviceOps` until buffer is
Expand All @@ -642,9 +642,7 @@ PjRtComputationClient::ExecuteComputation(
timed.reset();
TF_VLOG(3) << "ExecuteComputation returned_future->OnReady finished";
});
};

env::ScheduleIoClosure(util::MultiWait::Completer(mwait, std::move(lockfn)));
}));

TF_VLOG(1) << "Returning " << datas.size() << " results";
return datas;
Expand All @@ -668,7 +666,7 @@ PjRtComputationClient::ExecuteReplicated(
XLA_CHECK(devices.size() == arguments.size())
<< "ExecuteReplicated over " << devices.size() << " devices, but "
<< arguments.size() << " arguments devices.";
auto mwait_argument = std::make_shared<util::MultiWait>(devices.size());
absl::BlockingCounter mwait(devices.size());
std::vector<std::vector<xla::PjRtBuffer*>> argument_handles(devices.size());
{
tsl::profiler::TraceMe activity(
Expand All @@ -689,11 +687,11 @@ PjRtComputationClient::ExecuteReplicated(
buffers.push_back(pjrt_data->buffer.get());
}
argument_handles[i] = std::move(buffers);
mwait.DecrementCount();
};
env::ScheduleIoClosure(util::MultiWait::Completer(
mwait_argument, std::move(buffer_converter)));
thread::Schedule(std::move(buffer_converter));
}
mwait_argument->Wait();
mwait.Wait();
}

xla::ExecuteOptions execute_options;
Expand Down Expand Up @@ -748,9 +746,9 @@ PjRtComputationClient::ExecuteReplicated(
}
}

auto mwait = std::make_shared<util::MultiWait>(1);
auto lockfn = [&, this, returned_futures = std::move(*returned_futures),
timed]() mutable {
thread::Schedule(std::move([&, this,
returned_futures = std::move(*returned_futures),
timed]() mutable {
// Grab the shared lock and block the `WaitDeviceOps` until buffer is
// ready. Since this is the SPMD code path. There is no points to grab
// devices lock for every individual device.
Expand All @@ -771,8 +769,7 @@ PjRtComputationClient::ExecuteReplicated(
timed.reset();
TF_VLOG(3) << "ExecuteReplicated returned_future->OnReady finished";
});
};
env::ScheduleIoClosure(util::MultiWait::Completer(mwait, std::move(lockfn)));
}));

TF_VLOG(1) << "Returning " << data_handles.size() << " sets of results "
<< "with dimensions [" << absl::StrJoin(dims, ",") << "].";
Expand Down
Loading
Loading