• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
2 
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6 
7     http://www.apache.org/licenses/LICENSE-2.0
8 
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15 
16 #include "tensorflow/compiler/xla/service/executable.h"
17 
18 #include "absl/memory/memory.h"
19 #include "absl/strings/str_format.h"
20 #include "tensorflow/compiler/xla/debug_options_flags.h"
21 #include "tensorflow/compiler/xla/service/dump.h"
22 #include "tensorflow/compiler/xla/service/hlo_graph_dumper.h"
23 #include "tensorflow/compiler/xla/service/maybe_owning_device_memory.h"
24 #include "tensorflow/compiler/xla/status.h"
25 #include "tensorflow/compiler/xla/status_macros.h"
26 #include "tensorflow/core/lib/core/status.h"
27 #include "tensorflow/core/lib/hash/hash.h"
28 #include "tensorflow/core/lib/io/path.h"
29 #include "tensorflow/core/lib/strings/proto_serialization.h"
30 #include "tensorflow/core/platform/env.h"
31 #include "tensorflow/core/platform/errors.h"
32 #include "tensorflow/stream_executor/device_description.h"
33 
34 namespace xla {
35 
~ExecutionInput()36 ExecutionInput::~ExecutionInput() {
37   for (auto& index : unowned_indices_) {
38     auto buffer = buffers_.mutable_element(index)->Release();
39     if (buffer) {
40       buffer->Release();
41     }
42   }
43 }
44 
SetDynamicShape(Shape dynamic_shape)45 Status ExecutionInput::SetDynamicShape(Shape dynamic_shape) {
46   const Shape& input_shape = shape();
47   if (!ShapeUtil::DynamicShapeIsCompatible(input_shape, dynamic_shape)) {
48     return tensorflow::errors::InvalidArgument(
49         "Cannot set dynamic shape: ", input_shape.DebugString(), " vs. ",
50         dynamic_shape.DebugString());
51   }
52   dynamic_shape_ = absl::make_unique<Shape>(std::move(dynamic_shape));
53   return Status::OK();
54 }
55 
SetUnownedBuffer(const ShapeIndex & index,MaybeOwningDeviceMemory buffer)56 void ExecutionInput::SetUnownedBuffer(const ShapeIndex& index,
57                                       MaybeOwningDeviceMemory buffer) {
58   *buffers_.mutable_element(index) = std::move(buffer);
59   unowned_indices_.insert(index);
60 }
61 
ToShapedBuffer(se::DeviceMemoryAllocator * allocator,int device_ordinal) const62 StatusOr<ShapedBuffer> ExecutionInput::ToShapedBuffer(
63     se::DeviceMemoryAllocator* allocator, int device_ordinal) const {
64   const Shape& input_shape = shape();
65   ShapedBuffer shaped_buffer(input_shape, device_ordinal);
66   for (const auto& index_buffer : Buffers()) {
67     const tensorflow::se::OwningDeviceMemory* mem =
68         index_buffer.second.AsOwningDeviceMemory();
69     if (mem != nullptr && (mem->allocator() != allocator ||
70                            mem->device_ordinal() != device_ordinal)) {
71       return tensorflow::errors::InvalidArgument(
72           "Device buffer at index ", index_buffer.first.ToString(),
73           " has mismatching allocator/device");
74     }
75     shaped_buffer.set_buffer(index_buffer.second.AsDeviceMemoryBase(),
76                              index_buffer.first);
77   }
78   return std::move(shaped_buffer);
79 }
80 
ExecuteOnStream(const ServiceExecutableRunOptions * run_options,absl::Span<const ShapedBuffer * const> arguments,HloExecutionProfile * hlo_execution_profile)81 StatusOr<ScopedShapedBuffer> Executable::ExecuteOnStream(
82     const ServiceExecutableRunOptions* run_options,
83     absl::Span<const ShapedBuffer* const> arguments,
84     HloExecutionProfile* hlo_execution_profile) {
85   StatusOr<ScopedShapedBuffer> result =
86       ExecuteAsyncOnStream(run_options, arguments, hlo_execution_profile);
87   Status blocking_status = run_options->stream()->BlockHostUntilDone();
88   TF_RETURN_IF_ERROR(result.status());
89   TF_RETURN_IF_ERROR(blocking_status);
90   return result;
91 }
92 
MakeMaybeOwningDeviceMemoryTree(const ShapedBuffer & shaped_buffer)93 static ExecutionInput MakeMaybeOwningDeviceMemoryTree(
94     const ShapedBuffer& shaped_buffer) {
95   ExecutionInput result(shaped_buffer.on_device_shape());
96   shaped_buffer.buffers().ForEachElement(
97       [&](const ShapeIndex& index, const se::DeviceMemoryBase& mem) {
98         result.SetBuffer(index, MaybeOwningDeviceMemory(mem));
99       });
100   return result;
101 }
102 
ExecuteAsyncOnStream(const ServiceExecutableRunOptions * run_options,absl::Span<const ShapedBuffer * const> arguments,HloExecutionProfile * hlo_execution_profile)103 StatusOr<ScopedShapedBuffer> Executable::ExecuteAsyncOnStream(
104     const ServiceExecutableRunOptions* run_options,
105     absl::Span<const ShapedBuffer* const> arguments,
106     HloExecutionProfile* hlo_execution_profile) {
107   std::vector<ExecutionInput> args;
108   args.reserve(arguments.size());
109   for (const ShapedBuffer* arg : arguments) {
110     args.emplace_back(MakeMaybeOwningDeviceMemoryTree(*arg));
111   }
112   TF_ASSIGN_OR_RETURN(ExecutionOutput out,
113                       ExecuteAsyncOnStream(run_options, std::move(args),
114                                            hlo_execution_profile));
115   return out.ConsumeResult();
116 }
117 
ExecuteOnStream(const ServiceExecutableRunOptions * run_options,std::vector<ExecutionInput> arguments,HloExecutionProfile * hlo_execution_profile)118 StatusOr<ExecutionOutput> Executable::ExecuteOnStream(
119     const ServiceExecutableRunOptions* run_options,
120     std::vector<ExecutionInput> arguments,
121     HloExecutionProfile* hlo_execution_profile) {
122   StatusOr<ExecutionOutput> result = ExecuteAsyncOnStream(
123       run_options, std::move(arguments), hlo_execution_profile);
124   Status blocking_status = run_options->stream()->BlockHostUntilDone();
125   TF_RETURN_IF_ERROR(result.status());
126   TF_RETURN_IF_ERROR(blocking_status);
127   return result;
128 }
129 
ExecuteOnStreams(absl::Span<const ServiceExecutableRunOptions> run_options,absl::Span<const absl::Span<const ShapedBuffer * const>> arguments)130 StatusOr<std::vector<ScopedShapedBuffer>> Executable::ExecuteOnStreams(
131     absl::Span<const ServiceExecutableRunOptions> run_options,
132     absl::Span<const absl::Span<const ShapedBuffer* const>> arguments) {
133   TF_RET_CHECK(run_options.size() == arguments.size());
134 
135   std::vector<ScopedShapedBuffer> return_values;
136   return_values.reserve(run_options.size());
137 
138   if (run_options.size() == 1) {
139     TF_ASSIGN_OR_RETURN(auto rv,
140                         ExecuteOnStream(&run_options[0], arguments[0],
141                                         /*hlo_execution_profile=*/nullptr));
142     return_values.push_back(std::move(rv));
143     return std::move(return_values);
144   }
145 
146   for (size_t i = 0; i < run_options.size(); ++i) {
147     // We cannot BlockHostUntilDone() on the already-launched executions in case
148     // of error, since if the executions communicate, the initially launched
149     // executions may never complete if not all executions are running.
150     TF_ASSIGN_OR_RETURN(
151         auto rv, ExecuteAsyncOnStream(&run_options[i], arguments[i],
152                                       /*hlo_execution_profile=*/nullptr));
153     return_values.push_back(std::move(rv));
154   }
155   for (const auto& options : run_options) {
156     TF_RET_CHECK(options.stream() != nullptr);
157     TF_RETURN_IF_ERROR(options.stream()->BlockHostUntilDone());
158   }
159   return std::move(return_values);
160 }
161 
ExecuteOnStreamWrapper(const ServiceExecutableRunOptions * run_options,absl::Span<const ShapedBuffer * const> arguments)162 StatusOr<ScopedShapedBuffer> Executable::ExecuteOnStreamWrapper(
163     const ServiceExecutableRunOptions* run_options,
164     absl::Span<const ShapedBuffer* const> arguments) {
165   StatusOr<ScopedShapedBuffer> result =
166       ExecuteAsyncOnStreamWrapper(run_options, arguments);
167   Status block_status = run_options->stream()->BlockHostUntilDone();
168   TF_RETURN_IF_ERROR(result.status());
169   TF_RETURN_IF_ERROR(block_status);
170   return result;
171 }
172 
ExecuteOnStreamWrapper(const ServiceExecutableRunOptions * run_options,std::vector<ExecutionInput> arguments)173 StatusOr<ExecutionOutput> Executable::ExecuteOnStreamWrapper(
174     const ServiceExecutableRunOptions* run_options,
175     std::vector<ExecutionInput> arguments) {
176   StatusOr<ExecutionOutput> result =
177       ExecuteAsyncOnStreamWrapper(run_options, std::move(arguments));
178   Status block_status = run_options->stream()->BlockHostUntilDone();
179   TF_RETURN_IF_ERROR(result.status());
180   TF_RETURN_IF_ERROR(block_status);
181   return result;
182 }
183 
184 struct ExecuteAsyncOnStreamWrapperState {
185   ExecutionProfile* profile;
186   std::shared_ptr<se::Timer> timer;
187   std::shared_ptr<HloExecutionProfile> profile_ptr;
188 };
189 
ExecuteWrapperBeforeExecution(const Executable & executable,const ServiceExecutableRunOptions * run_options)190 static ExecuteAsyncOnStreamWrapperState ExecuteWrapperBeforeExecution(
191     const Executable& executable,
192     const ServiceExecutableRunOptions* run_options) {
193   ExecuteAsyncOnStreamWrapperState state;
194   se::Stream* stream = run_options->stream();
195   state.profile = run_options->run_options().execution_profile();
196   if (state.profile != nullptr) {
197     state.timer = std::make_shared<se::Timer>(stream->parent());
198     stream->InitTimer(state.timer.get()).ThenStartTimer(state.timer.get());
199   }
200 
201   VLOG(1) << "enqueueing executable on stream...";
202   // If the profiling flag isn't enabled, we pass nullptr as the profile to
203   // indicate profiling is not requested.
204   state.profile_ptr =
205       executable.module_config().debug_options().xla_hlo_profile() &&
206               executable.hlo_profiling_enabled()
207           ? std::make_shared<HloExecutionProfile>(
208                 &executable.hlo_profile_printer_data(),
209                 &executable.hlo_profile_index_map())
210           : nullptr;
211   return state;
212 }
213 
ExecuteWrapperAfterExecution(Executable * executable,const ExecuteAsyncOnStreamWrapperState & state,Status return_status,se::Stream * stream)214 Status ExecuteWrapperAfterExecution(
215     Executable* executable, const ExecuteAsyncOnStreamWrapperState& state,
216     Status return_status, se::Stream* stream) {
217   if (!return_status.ok()) {
218     if (state.profile != nullptr) {
219       // Ensure the ThenStartTimer call has completed before we destroy timer.
220       // We already have a failure status to return, so just log this if it
221       // fails.
222       Status status = stream->BlockHostUntilDone();
223       if (!status.ok()) {
224         LOG(ERROR) << "Failed to BlockHostUntilDone: " << status;
225       }
226     }
227     return return_status;
228   }
229 
230   if (state.profile != nullptr) {
231     VLOG(1) << "enqueueing 'stop timer' and profiling callback...";
232     stream->ThenStopTimer(state.timer.get());
233 
234     // We block instead of using an async callback because reading the timer
235     // value may call back into the driver on GPU, which is not allowed.
236     TF_RETURN_IF_ERROR(stream->BlockHostUntilDone());
237 
238     const int64 executable_size_in_bytes =
239         executable->SizeOfGeneratedCodeInBytes();
240     // Merge in run-time profile information from execution_profile.
241 
242     // Overall execution time (in nanoseconds) from the executor timer.
243     state.profile->set_compute_and_transfer_time_ns(state.timer->Nanoseconds());
244 
245     // TODO(b/28447609): The value in compute_and_transfer_time_ns is actually
246     // the compute time without the transfer time, so this way we get the
247     // correct compute time. We should instead have the correct value for
248     // compute_and_transfer_time and set compute_time to the compute time.
249     if (state.profile->compute_time_ns() == 0) {
250       state.profile->set_compute_time_ns(
251           state.profile->compute_and_transfer_time_ns());
252     }
253 
254     if (executable_size_in_bytes != 0) {
255       state.profile->set_executable_size_in_bytes(executable_size_in_bytes);
256     }
257   }
258 
259   if (executable->module_config().debug_options().xla_hlo_profile() &&
260       state.profile_ptr != nullptr) {
261     DumpToFileInDir(executable->module(), /*file_prefix=*/"",
262                     /*file_suffix=*/"hlo_execution_profile_data",
263                     state.profile_ptr->ToProto().SerializeAsString());
264   }
265 
266   if (state.profile_ptr != nullptr) {
267     const se::DeviceDescription* device_description =
268         &stream->parent()->GetDeviceDescription();
269     std::shared_ptr<HloExecutionProfile> profile = state.profile_ptr;
270     stream->ThenDoHostCallback([profile, device_description]() {
271       XLA_LOG_LINES(tensorflow::INFO,
272                     profile->ToString(device_description->clock_rate_ghz()));
273     });
274   }
275 
276   return return_status;
277 }
278 
ExecuteAsyncOnStreamWrapper(const ServiceExecutableRunOptions * run_options,absl::Span<const ShapedBuffer * const> arguments)279 StatusOr<ScopedShapedBuffer> Executable::ExecuteAsyncOnStreamWrapper(
280     const ServiceExecutableRunOptions* run_options,
281     absl::Span<const ShapedBuffer* const> arguments) {
282   auto state = ExecuteWrapperBeforeExecution(*this, run_options);
283   StatusOr<ScopedShapedBuffer> return_value =
284       ExecuteAsyncOnStream(run_options, arguments, state.profile_ptr.get());
285   TF_RETURN_IF_ERROR(ExecuteWrapperAfterExecution(
286       this, state, return_value.status(), run_options->stream()));
287   return return_value;
288 }
289 
ExecuteAsyncOnStreamWrapper(const ServiceExecutableRunOptions * run_options,std::vector<ExecutionInput> arguments)290 StatusOr<ExecutionOutput> Executable::ExecuteAsyncOnStreamWrapper(
291     const ServiceExecutableRunOptions* run_options,
292     std::vector<ExecutionInput> arguments) {
293   auto state = ExecuteWrapperBeforeExecution(*this, run_options);
294   StatusOr<ExecutionOutput> return_value = ExecuteAsyncOnStream(
295       run_options, std::move(arguments), state.profile_ptr.get());
296   TF_RETURN_IF_ERROR(ExecuteWrapperAfterExecution(
297       this, state, return_value.status(), run_options->stream()));
298   return return_value;
299 }
300 
SizeOfGeneratedCodeInBytes() const301 int64 Executable::SizeOfGeneratedCodeInBytes() const { return -1; }
302 
MarkToBeReleasedArguments(absl::Span<ExecutionInput> arguments,ExecutionOutput & result)303 void Executable::MarkToBeReleasedArguments(absl::Span<ExecutionInput> arguments,
304                                            ExecutionOutput& result) {
305   for (ExecutionInput& argument : arguments) {
306     for (auto& index_buffer : *argument.MutableBuffers()) {
307       if (absl::optional<se::OwningDeviceMemory> maybe_owning_buffer =
308               index_buffer.second.Release()) {
309         result.AddToBeReleased(std::move(*maybe_owning_buffer));
310       }
311     }
312   }
313 }
314 
315 }  // namespace xla
316