]> git.sesse.net Git - nageru/blob - nageru/print_latency.cpp
Make number of cards flexible at runtime.
[nageru] / nageru / print_latency.cpp
1 #include "print_latency.h"
2
3 #include "flags.h"
4 #include "shared/metrics.h"
5 #include "mixer.h"
6
7 #include <stdio.h>
8 #include <algorithm>
9 #include <chrono>
10 #include <string>
11
12 using namespace std;
13 using namespace std::chrono;
14
15 ReceivedTimestamps find_received_timestamp(const vector<RefCountedFrame> &input_frames)
16 {
17         assert(input_frames.size() == MAX_VIDEO_CARDS * FRAME_HISTORY_LENGTH);
18
19         ReceivedTimestamps ts;
20         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
21                 for (unsigned frame_index = 0; frame_index < FRAME_HISTORY_LENGTH; ++frame_index) {
22                         const RefCountedFrame &input_frame = input_frames[card_index * FRAME_HISTORY_LENGTH + frame_index];
23                         if (input_frame == nullptr ||
24                             (frame_index > 0 && input_frame.get() == input_frames[card_index * FRAME_HISTORY_LENGTH + frame_index - 1].get())) {
25                                 ts.ts.push_back(steady_clock::time_point::min());
26                         } else {
27                                 ts.ts.push_back(input_frame->received_timestamp);
28                         }
29                 }
30         }
31         return ts;
32 }
33
34 void LatencyHistogram::init(const string &measuring_point)
35 {
36         summaries.resize(MAX_VIDEO_CARDS * FRAME_HISTORY_LENGTH * 2);
37         for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
38                 char card_index_str[64];
39                 snprintf(card_index_str, sizeof(card_index_str), "%u", card_index);
40                 summaries[card_index].resize(FRAME_HISTORY_LENGTH);
41                 for (unsigned frame_index = 0; frame_index < FRAME_HISTORY_LENGTH; ++frame_index) {
42                         char frame_index_str[64];
43                         snprintf(frame_index_str, sizeof(frame_index_str), "%u", frame_index);
44
45                         vector<double> quantiles{0.01, 0.1, 0.25, 0.5, 0.75, 0.9, 0.99};
46                         summaries[card_index][frame_index].reset(new Summary[3]);
47                         summaries[card_index][frame_index][0].init(quantiles, 60.0);
48                         summaries[card_index][frame_index][1].init(quantiles, 60.0);
49                         summaries[card_index][frame_index][2].init(quantiles, 60.0);
50                         global_metrics.add("latency_seconds",
51                                 {{ "measuring_point", measuring_point },
52                                  { "card", card_index_str },
53                                  { "frame_age", frame_index_str },
54                                  { "frame_type", "i/p" }},
55                                  &summaries[card_index][frame_index][0],
56                                 Metrics::PRINT_WHEN_NONEMPTY);
57                         global_metrics.add("latency_seconds",
58                                 {{ "measuring_point", measuring_point },
59                                  { "card", card_index_str },
60                                  { "frame_age", frame_index_str },
61                                  { "frame_type", "b" }},
62                                  &summaries[card_index][frame_index][1],
63                                 Metrics::PRINT_WHEN_NONEMPTY);
64                         global_metrics.add("latency_seconds",
65                                 {{ "measuring_point", measuring_point },
66                                  { "card", card_index_str },
67                                  { "frame_age", frame_index_str },
68                                  { "frame_type", "total" }},
69                                  &summaries[card_index][frame_index][2],
70                                 Metrics::PRINT_WHEN_NONEMPTY);
71                 }
72         }
73 }
74
75 void print_latency(const char *header, const ReceivedTimestamps &received_ts, bool is_b_frame, int *frameno, LatencyHistogram *histogram)
76 {
77         if (received_ts.ts.empty())
78                 return;
79
80         const steady_clock::time_point now = steady_clock::now();
81
82         if (global_mixer == nullptr) {
83                 // Kaeru.
84                 assert(received_ts.ts.size() == 1);
85                 steady_clock::time_point ts = received_ts.ts[0];
86                 if (ts != steady_clock::time_point::min()) {
87                         duration<double> latency = now - ts;
88                         histogram->summaries[0][0][is_b_frame].count_event(latency.count());
89                         histogram->summaries[0][0][2].count_event(latency.count());
90                 }
91         } else {
92                 assert(received_ts.ts.size() == MAX_VIDEO_CARDS * FRAME_HISTORY_LENGTH);
93                 for (unsigned card_index = 0; card_index < MAX_VIDEO_CARDS; ++card_index) {
94                         for (unsigned frame_index = 0; frame_index < FRAME_HISTORY_LENGTH; ++frame_index) {
95                                 steady_clock::time_point ts = received_ts.ts[card_index * FRAME_HISTORY_LENGTH + frame_index];
96                                 if (ts == steady_clock::time_point::min()) {
97                                         continue;
98                                 }
99                                 duration<double> latency = now - ts;
100                                 histogram->summaries[card_index][frame_index][is_b_frame].count_event(latency.count());
101                                 histogram->summaries[card_index][frame_index][2].count_event(latency.count());
102                         }
103                 }
104         }
105
106         // 101 is chosen so that it's prime, which is unlikely to get the same frame type every time.
107         if (global_flags.print_video_latency && (++*frameno % 101) == 0) {
108                 // Find min and max timestamp of all input frames that have a timestamp.
109                 steady_clock::time_point min_ts = steady_clock::time_point::max(), max_ts = steady_clock::time_point::min();
110                 for (const auto &ts : received_ts.ts) {
111                         if (ts > steady_clock::time_point::min()) {
112                                 min_ts = min(min_ts, ts);
113                                 max_ts = max(max_ts, ts);
114                         }
115                 }
116                 duration<double> lowest_latency = now - max_ts;
117                 duration<double> highest_latency = now - min_ts;
118
119                 printf("%-60s %4.0f ms (lowest-latency input), %4.0f ms (highest-latency input)",
120                         header, 1e3 * lowest_latency.count(), 1e3 * highest_latency.count());
121
122                 if (is_b_frame) {
123                         printf("  [on B-frame; potential extra latency]\n");
124                 } else {
125                         printf("\n");
126                 }
127         }
128 }