1 /* Copyright 2019 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/compilation_stats.h"
17
18 #include <iostream>
19 #include <memory>
20 #include <string>
21
22 #include "absl/container/flat_hash_map.h"
23 #include "absl/strings/str_format.h"
24 #include "tensorflow/compiler/xla/types.h"
25 #include "tensorflow/core/platform/env.h"
26
27 namespace xla {
28
29 class NoopStats : public CompilationStats {
30 public:
31 NoopStats() = default;
32
StartPass(absl::string_view pass_name)33 void StartPass(absl::string_view pass_name) override {}
34
EndPass(absl::string_view pass_name)35 void EndPass(absl::string_view pass_name) override {}
36
CompilationReport()37 void CompilationReport() override {}
38
GetPassesSize()39 int GetPassesSize() override { return 0; }
40 };
41
42 class Stats : public CompilationStats {
43 public:
44 Stats() = default;
45
46 void StartPass(absl::string_view pass_name) override;
47
48 void EndPass(absl::string_view pass_name) override;
49
50 void CompilationReport() override;
51
52 int GetPassesSize() override;
53
54 private:
55 struct PassInfo {
PassInfoxla::Stats::PassInfo56 PassInfo(absl::string_view name, double duration)
57 : name(name), duration_ms(duration) {}
58
59 std::string name;
60 int num_runs = 1;
61 double duration_ms;
62 };
63
64 // Info about the passes that have been run so far.
65 std::vector<PassInfo> passes_;
66 // Used to avoid nested calls to StartPass.
67 bool pass_running_ = false;
68 std::string current_pass_;
69 // The start time of the currently running pass.
70 uint64_t start_micros_;
71 };
72
73 /* static */
MakeNoopStats()74 std::unique_ptr<CompilationStats> CompilationStats::MakeNoopStats() {
75 return std::make_unique<NoopStats>();
76 }
77
78 /* static */
MakeStats()79 std::unique_ptr<CompilationStats> CompilationStats::MakeStats() {
80 return std::make_unique<Stats>();
81 }
82
StartPass(absl::string_view pass_name)83 void Stats::StartPass(absl::string_view pass_name) {
84 CHECK(!pass_running_) << "Can't start " << pass_name << " while running "
85 << current_pass_;
86 pass_running_ = true;
87 current_pass_ = std::string(pass_name);
88 start_micros_ = tensorflow::Env::Default()->NowMicros();
89 }
90
EndPass(absl::string_view pass_name)91 void Stats::EndPass(absl::string_view pass_name) {
92 CHECK(pass_running_);
93 CHECK_EQ(current_pass_, std::string(pass_name));
94 pass_running_ = false;
95 uint64_t end_micros = tensorflow::Env::Default()->NowMicros();
96 double duration_ms = (end_micros - start_micros_) / 1000.0;
97 passes_.push_back(PassInfo(current_pass_, duration_ms));
98 }
99
CompilationReport()100 void Stats::CompilationReport() {
101 CHECK(!pass_running_) << "EndPass never called for " << current_pass_;
102 absl::flat_hash_map<std::string, PassInfo> summary;
103 double total_duration = 0;
104
105 for (auto& pass_run : passes_) {
106 auto pass_name = pass_run.name;
107 total_duration += pass_run.duration_ms;
108 auto it = summary.find(pass_name);
109 if (it == summary.end()) {
110 summary.insert(std::make_pair(pass_name, pass_run));
111 } else {
112 ++summary.at(pass_name).num_runs;
113 summary.at(pass_name).duration_ms += pass_run.duration_ms;
114 }
115 }
116
117 std::vector<PassInfo> sorted_summary;
118 sorted_summary.reserve(summary.size());
119 for (auto& it : summary) {
120 sorted_summary.push_back(it.second);
121 }
122 absl::c_sort(sorted_summary, [](const PassInfo& a, const PassInfo& b) {
123 // Sort passes that take the longest first, break ties using pass names.
124 return std::make_pair(b.duration_ms, a.name) <
125 std::make_pair(a.duration_ms, b.name);
126 });
127 LOG(INFO) << "Total runtime (ms) of HLO passes: " << total_duration;
128 LOG(INFO) << "Pass name, num runs, time (ms)";
129 for (auto& pass_info : sorted_summary) {
130 LOG(INFO) << pass_info.name << ", " << pass_info.num_runs << ", "
131 << pass_info.duration_ms;
132 }
133 }
134
GetPassesSize()135 int Stats::GetPassesSize() { return passes_.size(); }
136
137 } // namespace xla
138