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/compile_only_service.h"
17
18 #include <string>
19 #include <utility>
20 #include <vector>
21
22 #include "absl/strings/str_cat.h"
23 #include "tensorflow/compiler/xla/debug_options_flags.h"
24 #include "tensorflow/compiler/xla/service/backend.h"
25 #include "tensorflow/compiler/xla/service/computation_layout.h"
26 #include "tensorflow/compiler/xla/service/dump.h"
27 #include "tensorflow/compiler/xla/service/platform_util.h"
28 #include "tensorflow/compiler/xla/status_macros.h"
29 #include "tensorflow/compiler/xla/types.h"
30 #include "tensorflow/compiler/xla/util.h"
31 #include "tensorflow/core/lib/gtl/cleanup.h"
32 #include "tensorflow/core/lib/io/path.h"
33 #include "tensorflow/core/platform/host_info.h"
34 #include "tensorflow/core/platform/logging.h"
35 #include "tensorflow/core/platform/stream_executor_no_cuda.h"
36
37 namespace xla {
38
39 /* static */ StatusOr<std::unique_ptr<CompileOnlyService>>
NewService(se::Platform * platform)40 CompileOnlyService::NewService(se::Platform* platform) {
41 ServiceOptions default_options;
42 default_options.set_platform(platform);
43 return NewService(default_options);
44 }
45
46 /* static */ StatusOr<std::unique_ptr<CompileOnlyService>>
NewService(const ServiceOptions & options)47 CompileOnlyService::NewService(const ServiceOptions& options) {
48 se::Platform* platform = options.platform();
49 if (platform == nullptr) {
50 TF_ASSIGN_OR_RETURN(platform, PlatformUtil::GetDefaultPlatform());
51 }
52
53 TF_ASSIGN_OR_RETURN(auto compiler, Compiler::GetForPlatform(platform));
54
55 std::unique_ptr<CompileOnlyService> service(
56 new CompileOnlyService(options, compiler));
57 return std::move(service);
58 }
59
CompileOnlyService(const ServiceOptions & options,Compiler * compiler)60 CompileOnlyService::CompileOnlyService(const ServiceOptions& options,
61 Compiler* compiler)
62 : Service(options, /*execute_backend=*/nullptr), compiler_(compiler) {}
63
64 StatusOr<std::vector<std::unique_ptr<AotCompilationResult>>>
CompileAheadOfTime(const absl::Span<const AotXlaComputationInstance> computations,const AotCompilationOptions & options,std::unique_ptr<AotCompilationMetadata> * metadata)65 CompileOnlyService::CompileAheadOfTime(
66 const absl::Span<const AotXlaComputationInstance> computations,
67 const AotCompilationOptions& options,
68 std::unique_ptr<AotCompilationMetadata>* metadata) {
69 std::vector<std::unique_ptr<HloModule>> hlo_modules;
70
71 const DebugOptions& debug_options = options.debug_options();
72 ExecutionOptions execution_options;
73 *execution_options.mutable_debug_options() = debug_options;
74 // Capture replica_count, num_cores, and device_assignment in ExecutionOptions
75 // to later save in a proto dump.
76 if (options.replica_count() > 0) {
77 execution_options.set_num_replicas(options.replica_count());
78 if (options.has_static_device_assignment()) {
79 CHECK_EQ(options.replica_count(),
80 options.static_device_assignment().replica_count());
81 }
82 }
83 if (options.num_cores() > 0) {
84 execution_options.set_num_partitions(options.num_cores());
85 if (options.has_static_device_assignment()) {
86 CHECK_EQ(options.num_cores(),
87 options.static_device_assignment().computation_count());
88 }
89 }
90 if (options.has_static_device_assignment()) {
91 TF_RETURN_IF_ERROR(options.static_device_assignment().Serialize(
92 execution_options.mutable_device_assignment()));
93 }
94 execution_options.set_use_spmd_partitioning(options.use_spmd_partitioning());
95 execution_options.set_deduplicate_hlo(options.deduplicate_hlo());
96 execution_options.set_broadcast_replicated_parameters_via_collectives(
97 options.broadcast_replicated_params());
98 for (const AotXlaComputationInstance& instance : computations) {
99 TF_RET_CHECK(instance.computation.has_host_program_shape());
100 *execution_options.mutable_shape_with_output_layout() =
101 instance.result_layout->ToProto();
102
103 TF_ASSIGN_OR_RETURN(
104 std::unique_ptr<HloModuleConfig> module_config,
105 CreateModuleConfig(
106 ProgramShape(instance.computation.host_program_shape()),
107 instance.argument_layouts, &execution_options, &options));
108
109 TF_ASSIGN_OR_RETURN(
110 std::unique_ptr<HloModule> hlo_module,
111 HloModule::CreateFromProto(instance.computation, *module_config));
112 DumpHloModuleIfEnabled(*hlo_module, "before_optimizations");
113 hlo_modules.push_back(std::move(hlo_module));
114 }
115
116 execution_options.clear_shape_with_output_layout();
117 DumpExecutionOptions(execution_options, debug_options);
118
119 return compiler_->CompileAheadOfTime(
120 absl::make_unique<HloModuleGroup>(hlo_modules[0]->name(),
121 absl::MakeSpan(hlo_modules)),
122 options, metadata);
123 }
124
125 } // namespace xla
126