wehub-resource-sync
120 lines
4.6 KiB
C++
120 lines
4.6 KiB
C++
/* Copyright 2022 The TensorFlow Authors. All Rights Reserved.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
==============================================================================*/
|
|
|
|
#include "tensorflow/compiler/jit/xla_compile_util.h"
|
|
|
|
#include <memory>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
#include "absl/status/status.h"
|
|
#include "tensorflow/compiler/jit/flags.h"
|
|
#include "tensorflow/core/framework/op_kernel.h"
|
|
#include "tensorflow/core/framework/resource_mgr.h"
|
|
#include "tensorflow/core/framework/types.h"
|
|
#include "tensorflow/core/graph/algorithm.h"
|
|
#include "tensorflow/core/graph/node_builder.h"
|
|
#include "tensorflow/core/tfrt/common/global_state.h"
|
|
#include "tensorflow/core/util/determinism.h"
|
|
|
|
namespace tensorflow {
|
|
namespace {
|
|
constexpr const char* kPjRtDeviceCompilerResourceName = "pjrt_device_compiler";
|
|
constexpr const char* kPjRtDeviceCompilationProfilerResourceName =
|
|
"pjrt_device_compilation_profiler";
|
|
} // namespace
|
|
|
|
absl::StatusOr<std::unique_ptr<Graph>> CreateSingleOpGraph(
|
|
const NodeDef& node_def, absl::Span<const XlaArgument> args,
|
|
absl::Span<const DataType> result_types) {
|
|
// TODO(b/74182462): We implement this by creating a new dummy Graph including
|
|
// _Arg nodes, and let CompileGraph walk it. This could be optimized.
|
|
std::unique_ptr<Graph> graph(new Graph(OpRegistry::Global()));
|
|
|
|
// First create the actual node we care about computing.
|
|
TF_ASSIGN_OR_RETURN(Node * main_node, graph->AddNode(node_def));
|
|
|
|
// Create dummy _Arg nodes. Link these to `node` and also via a control
|
|
// dependency edge to the _SOURCE node.
|
|
for (int64_t i = 0, end = args.size(); i < end; ++i) {
|
|
Node* node;
|
|
std::string arg_name = absl::StrCat("_arg", i);
|
|
absl::Status status =
|
|
NodeBuilder(arg_name, FunctionLibraryDefinition::kArgOp)
|
|
.ControlInput(graph->source_node())
|
|
.Attr("T", args[i].kind == XlaArgument::kResource ? DT_RESOURCE
|
|
: args[i].type)
|
|
.Attr("index", i)
|
|
.Finalize(graph.get(), &node);
|
|
TF_RETURN_IF_ERROR(status);
|
|
graph->AddEdge(node, 0, main_node, i);
|
|
}
|
|
|
|
// Similarly with return values, create dummy _Retval nodes fed by `node`.
|
|
for (int64_t i = 0, end = result_types.size(); i < end; ++i) {
|
|
Node* node;
|
|
std::string retval_name = absl::StrCat("_retval", i);
|
|
absl::Status status =
|
|
NodeBuilder(retval_name, FunctionLibraryDefinition::kRetOp)
|
|
.Input(main_node, i)
|
|
.Attr("T", result_types[i])
|
|
.Attr("index", i)
|
|
.Finalize(graph.get(), &node);
|
|
TF_RETURN_IF_ERROR(status);
|
|
}
|
|
FixupSourceAndSinkEdges(graph.get());
|
|
return graph;
|
|
}
|
|
|
|
bool UsePjRtForSingleDeviceCompilation(const DeviceType& device_type) {
|
|
const auto& rollout_config = GetXlaOpsCommonFlags()->tf_xla_use_device_api;
|
|
return rollout_config.IsEnabledInXlaLaunchForDevice(device_type) ||
|
|
rollout_config.IsEnabledInXlaCompileOnDemandForDevice(device_type) ||
|
|
rollout_config.IsEnabledInXlaCompileAndRunForDevice(device_type);
|
|
}
|
|
|
|
std::string GetPjRtDeviceCompilerResourceName(const DeviceType& device_type) {
|
|
return absl::StrCat(kPjRtDeviceCompilerResourceName, "_",
|
|
device_type.type_string());
|
|
}
|
|
|
|
std::string GetPjRtDeviceCompilationProfilerResourceName(
|
|
const DeviceType& device_type) {
|
|
return absl::StrCat(kPjRtDeviceCompilationProfilerResourceName, "_",
|
|
device_type.type_string());
|
|
}
|
|
|
|
absl::StatusOr<ResourceMgr*> GetResourceMgrForDeviceCompiler(
|
|
const OpKernelContext& ctx, const DeviceType& device_type) {
|
|
// We store information about the JIT-compiled XLA computation in the
|
|
// ResourceMgr. The DeviceCompiler (which contains the DeviceCompilationCache)
|
|
// is stored in the tfrt_global ResourceMgr for TPU and the Device ResourceMgr
|
|
// for CPU/GPU. This is to make sure the DeviceCompiler's lifecycle is
|
|
// maintained appropriately.
|
|
ResourceMgr* rm = nullptr;
|
|
if (device_type == DEVICE_TPU) {
|
|
rm = tfrt_global::GetTFGlobalResourceMgr();
|
|
} else {
|
|
rm = ctx.resource_manager();
|
|
}
|
|
|
|
if (!rm) {
|
|
return absl::InternalError("No resource manager found.");
|
|
}
|
|
return rm;
|
|
}
|
|
|
|
} // namespace tensorflow
|