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/interpreter/executable.h"
17
18 #include <algorithm>
19 #include <cstring>
20 #include <string>
21 #include <utility>
22 #include <vector>
23
24 #include "absl/memory/memory.h"
25 #include "tensorflow/compiler/xla/literal.h"
26 #include "tensorflow/compiler/xla/service/hlo_computation.h"
27 #include "tensorflow/compiler/xla/service/hlo_instruction.h"
28 #include "tensorflow/compiler/xla/service/interpreter/executor.h"
29 #include "tensorflow/compiler/xla/service/transfer_manager.h"
30 #include "tensorflow/compiler/xla/shape_util.h"
31 #include "tensorflow/compiler/xla/status_macros.h"
32 #include "tensorflow/core/lib/core/errors.h"
33 #include "tensorflow/core/platform/env.h"
34 #include "tensorflow/core/platform/stream_executor_no_cuda.h"
35
36 namespace xla {
37 namespace interpreter {
38
InterpreterExecutable(std::unique_ptr<HloModule> hlo_module,std::unique_ptr<HloEvaluator> evaluator)39 InterpreterExecutable::InterpreterExecutable(
40 std::unique_ptr<HloModule> hlo_module,
41 std::unique_ptr<HloEvaluator> evaluator)
42 : Executable(std::move(hlo_module), /*hlo_profile_printer=*/nullptr,
43 /*hlo_profile_index_map=*/nullptr),
44 evaluator_(std::move(evaluator)) {}
45
~InterpreterExecutable()46 InterpreterExecutable::~InterpreterExecutable() {}
47
ExecuteOnStream(const ServiceExecutableRunOptions * run_options,absl::Span<const ShapedBuffer * const> arguments,HloExecutionProfile * hlo_execution_profile)48 StatusOr<ScopedShapedBuffer> InterpreterExecutable::ExecuteOnStream(
49 const ServiceExecutableRunOptions* run_options,
50 absl::Span<const ShapedBuffer* const> arguments,
51 HloExecutionProfile* hlo_execution_profile) {
52 se::Stream* stream = run_options->stream();
53 se::StreamExecutor* executor = stream->parent();
54 const se::Platform* platform = executor->platform();
55
56 VLOG(1) << "Execute " << module().name();
57 if (VLOG_IS_ON(2)) {
58 for (const auto& a : arguments) {
59 VLOG(2) << "-- argument " << *a;
60 }
61 }
62
63 uint64 start_micros = tensorflow::Env::Default()->NowMicros();
64
65 const HloComputation* computation = module().entry_computation();
66 if (computation->num_parameters() != arguments.size()) {
67 return tensorflow::errors::Internal(
68 "Mismatch between argument count and graph parameter count.");
69 }
70
71 // Check that the args have the right shape.
72 for (int64 i = 0; i < computation->num_parameters(); ++i) {
73 const auto& expected_shape = computation->parameter_instruction(i)->shape();
74 const auto& actual_shape = arguments[i]->on_device_shape();
75 if (!ShapeUtil::Equal(expected_shape, actual_shape)) {
76 return InvalidArgument(
77 "Shape mismatch on parameter %d. Expected %s, but was %s.", i,
78 ShapeUtil::HumanString(expected_shape),
79 ShapeUtil::HumanString(actual_shape));
80 }
81 }
82
83 TF_ASSIGN_OR_RETURN(TransferManager * transfer_manager,
84 TransferManager::GetForPlatform(platform));
85
86 // Transform the ShapedBuffer arguments into literals which the evaluator
87 // consumes.
88 std::vector<Literal> arg_literals;
89 for (int64 p = 0; p < computation->num_parameters(); ++p) {
90 TF_ASSIGN_OR_RETURN(Literal arg_literal,
91 transfer_manager->TransferLiteralFromDevice(
92 run_options->stream(), *arguments[p]));
93 arg_literals.push_back(std::move(arg_literal));
94 }
95
96 // Execute the graph using the HloEvaluator.
97 Literal result_literal;
98 {
99 tensorflow::mutex_lock lock(evaluator_lock_);
100 evaluator_->ResetVisitStates();
101 TF_ASSIGN_OR_RETURN(result_literal,
102 evaluator_->Evaluate(*computation, arg_literals));
103 }
104
105 // Transform the result literal back into a ShapedBuffer.
106 TF_ASSIGN_OR_RETURN(ScopedShapedBuffer result,
107 transfer_manager->AllocateScopedShapedBuffer(
108 result_literal.shape(), run_options->allocator(),
109 executor->device_ordinal()));
110 TF_RETURN_IF_ERROR(transfer_manager->TransferLiteralToDevice(
111 run_options->stream(), result_literal, result));
112
113 uint64 end_micros = tensorflow::Env::Default()->NowMicros();
114
115 {
116 tensorflow::mutex_lock lock(mutex_);
117 const double nanoseconds = (end_micros - start_micros) * 1000.0;
118 execution_profile_.set_compute_time_ns(std::max(nanoseconds, 1.0));
119 }
120
121 return std::move(result);
122 }
123
ExecuteAsyncOnStream(const ServiceExecutableRunOptions * run_options,absl::Span<const ShapedBuffer * const> arguments)124 StatusOr<ScopedShapedBuffer> InterpreterExecutable::ExecuteAsyncOnStream(
125 const ServiceExecutableRunOptions* run_options,
126 absl::Span<const ShapedBuffer* const> arguments) {
127 return tensorflow::errors::Unimplemented(
128 "ExecuteAsyncOnStream is not yet supported on Interpreter.");
129 }
130
ShapeSizeBytes(const Shape & shape)131 /*static*/ int64 InterpreterExecutable::ShapeSizeBytes(const Shape& shape) {
132 if (shape.IsOpaque()) {
133 return sizeof(void*);
134 }
135 return ShapeUtil::ByteSizeOf(shape, sizeof(void*));
136 }
137
138 } // namespace interpreter
139 } // namespace xla
140