]> git.sesse.net Git - pkanalytics/blob - video_widget.cpp
Count interceptions towards (successful) catches.
[pkanalytics] / video_widget.cpp
1 #define GL_GLEXT_PROTOTYPES
2
3 #include "video_widget.h"
4
5 #include <assert.h>
6 #include <pthread.h>
7 #include <stdint.h>
8 #include <stdio.h>
9 #include <stdlib.h>
10 #include <string.h>
11 #include <sys/stat.h>
12 #include <unistd.h>
13
14 extern "C" {
15 #include <libavcodec/avcodec.h>
16 #include <libavformat/avformat.h>
17 #include <libavutil/avutil.h>
18 #include <libavutil/error.h>
19 #include <libavutil/frame.h>
20 #include <libavutil/imgutils.h>
21 #include <libavutil/mem.h>
22 #include <libavutil/pixfmt.h>
23 #include <libavutil/opt.h>
24 #include <libswscale/swscale.h>
25 }
26
27 #include <chrono>
28 #include <cstdint>
29 #include <utility>
30 #include <vector>
31 #include <unordered_set>
32
33 #include "post_to_main_thread.h"
34
35 #include <QOpenGLFunctions>
36 #include <QWheelEvent>
37 #include <QMouseEvent>
38 #include <QMouseEvent>
39 #include <QHBoxLayout>
40
41 #define BUFFER_OFFSET(i) ((char *)nullptr + (i))
42
43 using namespace std;
44 using namespace std::chrono;
45
46 namespace {
47
48 bool is_full_range(const AVPixFmtDescriptor *desc)
49 {
50         // This is horrible, but there's no better way that I know of.
51         return (strchr(desc->name, 'j') != nullptr);
52 }
53
54 AVPixelFormat decide_dst_format(AVPixelFormat src_format)
55 {
56         // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
57         // and be done with it. It's too strange to spend a lot of time on.
58         // (Let's hope there's no alpha.)
59         const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
60         if (src_desc == nullptr ||
61             src_desc->nb_components != 3 ||
62             (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
63                 return AV_PIX_FMT_YUV444P;
64         }
65
66         // The best for us would be Cb and Cr together if possible,
67         // but FFmpeg doesn't support that except in the special case of
68         // NV12, so we need to go to planar even for the case of NV12.
69         // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
70         // that matches in color range. (This will also include the case of
71         // the source format already being acceptable.)
72         bool src_full_range = is_full_range(src_desc);
73         const char *best_format = "yuv444p";
74         unsigned best_score = numeric_limits<unsigned>::max();
75         for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
76              desc;
77              desc = av_pix_fmt_desc_next(desc)) {
78                 // Find planar Y'CbCr formats only.
79                 if (desc->nb_components != 3) continue;
80                 if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
81                 if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
82                 if (desc->comp[0].plane != 0 ||
83                     desc->comp[1].plane != 1 ||
84                     desc->comp[2].plane != 2) continue;
85
86                 // 8-bit formats only.
87                 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
88                 if (desc->comp[0].depth != 8) continue;
89
90                 // Same or better chroma resolution only.
91                 int chroma_w_diff = src_desc->log2_chroma_w - desc->log2_chroma_w;
92                 int chroma_h_diff = src_desc->log2_chroma_h - desc->log2_chroma_h;
93                 if (chroma_w_diff < 0 || chroma_h_diff < 0)
94                         continue;
95
96                 // Matching full/limited range only.
97                 if (is_full_range(desc) != src_full_range)
98                         continue;
99
100                 // Pick something with as little excess chroma resolution as possible.
101                 unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
102                 if (score < best_score) {
103                         best_score = score;
104                         best_format = desc->name;
105                 }
106         }
107         return av_get_pix_fmt(best_format);
108 }
109
110 }  // namespace
111
112 bool VideoWidget::process_queued_commands(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, int video_stream_index, bool *seeked)
113 {
114         // Process any queued commands from other threads.
115         vector<QueuedCommand> commands;
116         {
117                 lock_guard<mutex> lock(queue_mu);
118                 swap(commands, command_queue);
119         }
120
121         for (const QueuedCommand &cmd : commands) {
122                 switch (cmd.command) {
123                 case QueuedCommand::PAUSE:
124                         paused = true;
125                         break;
126                 case QueuedCommand::RESUME:
127                         paused = false;
128                         pts_origin = last_pts;
129                         start = next_frame_start = steady_clock::now();
130                         break;
131                 case QueuedCommand::SEEK:
132                 case QueuedCommand::SEEK_ABSOLUTE:
133                         // Dealt with below.
134                         break;
135                 }
136         }
137
138         // Combine all seeks into one big one. (There are edge cases where this is probably
139         // subtly wrong, but we'll live with it.)
140         int64_t base_pts = last_pts;
141         int64_t relative_seek_ms = 0;
142         int64_t relative_seek_frames = 0;
143         for (const QueuedCommand &cmd : commands) {
144                 if (cmd.command == QueuedCommand::SEEK) {
145                         relative_seek_ms += cmd.relative_seek_ms;
146                         relative_seek_frames += cmd.relative_seek_frames;
147                 } else if (cmd.command == QueuedCommand::SEEK_ABSOLUTE) {
148                         base_pts = av_rescale_q(cmd.seek_ms, AVRational{ 1, 1000 }, video_timebase);
149                         relative_seek_ms = 0;
150                         relative_seek_frames = 0;
151                 }
152         }
153         int64_t relative_seek_pts = av_rescale_q(relative_seek_ms, AVRational{ 1, 1000 }, video_timebase);
154         if (relative_seek_ms != 0 && relative_seek_pts == 0) {
155                 // Just to be sure rounding errors don't move us into nothingness.
156                 relative_seek_pts = (relative_seek_ms > 0) ? 1 : -1;
157         }
158         int64_t goal_pts = base_pts + relative_seek_pts;
159         if (goal_pts != last_pts || relative_seek_frames < 0) {
160                 avcodec_flush_buffers(video_codec_ctx);
161                 queued_frames.clear();
162
163                 // Seek to the last keyframe before this point.
164                 int64_t seek_pts = goal_pts;
165                 if (relative_seek_frames < 0) {
166                         // If we're frame-skipping backwards, add 100 ms of slop for each frame
167                         // so we're fairly certain we are able to see the ones we want.
168                         seek_pts -= av_rescale_q(-relative_seek_frames, AVRational{ 1, 10 }, video_timebase);
169                 }
170                 av_seek_frame(format_ctx, video_stream_index, seek_pts, AVSEEK_FLAG_BACKWARD);
171
172                 // Decode frames until EOF, or until we see something past our seek point.
173                 std::deque<AVFrameWithDeleter> queue;
174                 for ( ;; ) {
175                         bool error = false;
176                         AVFrameWithDeleter frame = decode_frame(format_ctx, video_codec_ctx,
177                                 pathname, video_stream_index, &error);
178                         if (frame == nullptr || error) {
179                                 break;
180                         }
181
182                         int64_t frame_pts = frame->pts;
183                         if (relative_seek_frames < 0) {
184                                 // Buffer this frame; don't display it unless we know it's the Nth-latest.
185                                 queue.push_back(std::move(frame));
186                                 if (queue.size() > uint64_t(-relative_seek_frames) + 1) {
187                                         queue.pop_front();
188                                 }
189                         }
190                         if (frame_pts >= goal_pts) {
191                                 if (relative_seek_frames > 0) {
192                                         --relative_seek_frames;
193                                 } else {
194                                         if (relative_seek_frames < 0) {
195                                                 // Hope we have the right amount.
196                                                 // The rest will remain in the queue for when we play forward again.
197                                                 frame = std::move(queue.front());
198                                                 queue.pop_front();
199                                                 queued_frames = std::move(queue);
200                                         }
201                                         video_window->set_current_frame(make_video_frame(frame.get()));
202                                         update();
203                                         store_pts(frame->pts);
204                                         break;
205                                 }
206                         }
207                 }
208
209                 // NOTE: We keep pause status as-is.
210
211                 pts_origin = last_pts;
212                 start = next_frame_start = last_frame = steady_clock::now();
213                 if (seeked) {
214                         *seeked = true;
215                 }
216         } else if (relative_seek_frames > 0) {
217                 // The base PTS is fine, we only need to skip a few frames forwards.
218                 while (relative_seek_frames > 1) {
219                         // Eat a frame (ignore errors).
220                         bool error;
221                         decode_frame(format_ctx, video_codec_ctx, pathname, video_stream_index, &error);
222                         --relative_seek_frames;
223                 }
224
225                 // Display the last one.
226                 bool error;
227                 AVFrameWithDeleter frame = decode_frame(format_ctx, video_codec_ctx,
228                         pathname, video_stream_index, &error);
229                 if (frame == nullptr || error) {
230                         return true;
231                 }
232                 video_window->set_current_frame(make_video_frame(frame.get()));
233                 update();
234                 store_pts(frame->pts);
235         }
236         return false;
237 }
238
239 VideoWidget::VideoWidget(QWidget *parent)
240         : QWidget(parent),
241           video_window(new VideoWindow(this)) {
242         setLayout(new QHBoxLayout);
243         layout()->addWidget(QWidget::createWindowContainer(video_window));
244
245         // ...
246         connect(video_window, &VideoWindow::mouse_wheel, this, &VideoWidget::wheelEvent);
247         connect(video_window, &VideoWindow::mouse_pressed, this, &VideoWidget::mousePressEvent);
248         connect(video_window, &VideoWindow::mouse_released, this, &VideoWidget::mouseReleaseEvent);
249         connect(video_window, &VideoWindow::mouse_moved, this, &VideoWidget::mouseMoveEvent);
250 }
251
252 GLuint compile_shader(const string &shader_src, GLenum type)
253 {
254         GLuint obj = glCreateShader(type);
255         const GLchar* source[] = { shader_src.data() };
256         const GLint length[] = { (GLint)shader_src.size() };
257         glShaderSource(obj, 1, source, length);
258         glCompileShader(obj);
259
260         GLchar info_log[4096];
261         GLsizei log_length = sizeof(info_log) - 1;
262         glGetShaderInfoLog(obj, log_length, &log_length, info_log);
263         info_log[log_length] = 0;
264         if (strlen(info_log) > 0) {
265                 fprintf(stderr, "Shader compile log: %s\n", info_log);
266         }
267
268         GLint status;
269         glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
270         if (status == GL_FALSE) {
271                 // Add some line numbers to easier identify compile errors.
272                 string src_with_lines = "/*   1 */ ";
273                 size_t lineno = 1;
274                 for (char ch : shader_src) {
275                         src_with_lines.push_back(ch);
276                         if (ch == '\n') {
277                                 char buf[32];
278                                 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
279                                 src_with_lines += buf;
280                         }
281                 }
282
283                 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
284                 exit(1);
285         }
286
287         return obj;
288 }
289
290 void VideoWindow::initializeGL()
291 {
292         glDisable(GL_BLEND);
293         glDisable(GL_DEPTH_TEST);
294         glDepthMask(GL_FALSE);
295         glCreateTextures(GL_TEXTURE_2D, 3, tex);
296
297         ycbcr_vertex_shader = compile_shader(R"(
298 #version 460 core
299
300 layout(location = 0) in vec2 position;
301 layout(location = 1) in vec2 texcoord;
302 out vec2 tc;
303
304 void main()
305 {
306         // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is:
307         //
308         //   2.000  0.000  0.000 -1.000
309         //   0.000  2.000  0.000 -1.000
310         //   0.000  0.000 -2.000 -1.000
311         //   0.000  0.000  0.000  1.000
312         gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0);
313         tc = texcoord;
314         tc.y = 1.0f - tc.y;
315 }
316 )", GL_VERTEX_SHADER);
317         ycbcr_fragment_shader = compile_shader(R"(
318 #version 460 core
319
320 layout(location = 0) uniform sampler2D tex_y;
321 layout(location = 1) uniform sampler2D tex_cb;
322 layout(location = 2) uniform sampler2D tex_cr;
323 layout(location = 3) uniform vec2 cbcr_offset;
324
325 in vec2 tc;
326 out vec4 FragColor;
327
328 // Computed statically by Movit, for limited-range BT.709.
329 // (We don't check whether the input could be BT.601 or BT.2020 currently, or full-range)
330 const mat3 inv_ycbcr_matrix = mat3(
331         1.16438f, 1.16438f, 1.16438f,
332         0.0f, -0.21325f, 2.11240f,
333         1.79274f, -0.53291f, 0.0f
334 );
335
336 void main()
337 {
338         if (tc.x < 0.0 || tc.x > 1.0 || tc.y < 0.0 || tc.y > 1.0) {
339                 FragColor.rgba = vec4(0.0f, 0.0f, 0.0f, 1.0f);
340                 return;
341         }
342
343         vec3 ycbcr;
344         ycbcr.r = texture(tex_y, tc).r;
345         ycbcr.g = texture(tex_cb, tc + cbcr_offset).r;
346         ycbcr.b = texture(tex_cr, tc + cbcr_offset).r;
347         ycbcr -= vec3(16.0f / 255.0f, 128.0f / 255.0f, 128.0f / 255.0f);
348         FragColor.rgb = inv_ycbcr_matrix * ycbcr;
349         FragColor.a = 1.0f;
350 }
351 )", GL_FRAGMENT_SHADER);
352         ycbcr_program = glCreateProgram();
353         glAttachShader(ycbcr_program, ycbcr_vertex_shader);
354         glAttachShader(ycbcr_program, ycbcr_fragment_shader);
355         glLinkProgram(ycbcr_program);
356
357         GLint success;
358         glGetProgramiv(ycbcr_program, GL_LINK_STATUS, &success);
359         if (success == GL_FALSE) {
360                 GLchar error_log[1024] = {0};
361                 glGetProgramInfoLog(ycbcr_program, 1024, nullptr, error_log);
362                 fprintf(stderr, "Error linking program: %s\n", error_log);
363                 exit(1);
364         }
365
366         glCreateSamplers(1, &bilinear_sampler);
367         glSamplerParameteri(bilinear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
368         glSamplerParameteri(bilinear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
369         glSamplerParameteri(bilinear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
370         glSamplerParameteri(bilinear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
371 }
372
373 void VideoWindow::resizeGL(int w, int h)
374 {
375         glViewport(0, 0, w, h);
376         display_aspect = double(w) / h;
377 }
378
379 int num_levels(GLuint width, GLuint height)
380 {
381         int levels = 1;
382         while (width > 1 || height > 1) {
383                 width = max(width / 2, 1u);
384                 height = max(height / 2, 1u);
385                 ++levels;
386         }
387         return levels;
388 }
389
390 void VideoWindow::paintGL()
391 {
392         std::shared_ptr<VideoWidget::Frame> frame;
393         {
394                 lock_guard lock(current_frame_mu);
395                 frame = current_frame;
396         }
397         if (frame == nullptr) {
398                 glClear(GL_COLOR_BUFFER_BIT);
399                 return;
400         }
401
402         glUseProgram(ycbcr_program);
403         if (frame->width != last_width || frame->height != last_height) {
404                 glTextureStorage2D(tex[0], num_levels(frame->width, frame->height), GL_R8, frame->width, frame->height);
405         }
406         if (frame->chroma_width != last_chroma_width || frame->chroma_height != last_chroma_height) {
407                 for (GLuint num : { tex[1], tex[2] }) {
408                         glTextureStorage2D(num, num_levels(frame->chroma_width, frame->chroma_height), GL_R8, frame->chroma_width, frame->chroma_height);
409                 }
410         }
411
412         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, frame->pbo);
413
414         if (frame->need_flush_len > 0) {
415                 glFlushMappedNamedBufferRange(frame->pbo, 0, frame->need_flush_len);
416                 frame->need_flush_len = 0;
417         }
418
419         glPixelStorei(GL_UNPACK_ROW_LENGTH, 0);
420         glTextureSubImage2D(tex[0], 0, 0, 0, frame->width, frame->height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(0));
421         glGenerateTextureMipmap(tex[0]);
422
423         glTextureSubImage2D(tex[1], 0, 0, 0, frame->chroma_width, frame->chroma_height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(frame->width * frame->height));
424         glGenerateTextureMipmap(tex[1]);
425
426         glTextureSubImage2D(tex[2], 0, 0, 0, frame->chroma_width, frame->chroma_height, GL_RED, GL_UNSIGNED_BYTE, BUFFER_OFFSET(frame->width * frame->height + frame->chroma_width * frame->chroma_height));
427         glGenerateTextureMipmap(tex[2]);
428
429         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
430
431         glBindTextureUnit(0, tex[0]);
432         glBindTextureUnit(1, tex[1]);
433         glBindTextureUnit(2, tex[2]);
434         glBindSampler(0, bilinear_sampler);
435         glBindSampler(1, bilinear_sampler);
436         glBindSampler(2, bilinear_sampler);
437         glProgramUniform1i(ycbcr_program, 0, 0);
438         glProgramUniform1i(ycbcr_program, 1, 1);
439         glProgramUniform1i(ycbcr_program, 2, 2);
440         glProgramUniform2f(ycbcr_program, 3, cbcr_offset[0], -cbcr_offset[1]);
441
442         float tx1 = 0.0f;
443         float tx2 = 1.0f;
444         float ty1 = 0.0f;
445         float ty2 = 1.0f;
446
447         double video_aspect = double(frame->width) / frame->height;
448         if (display_aspect > video_aspect) {
449                 double extra_width = frame->height * display_aspect - frame->width;
450                 tx1 = -0.5 * extra_width / frame->width;
451                 tx2 = 1.0 + 0.5 * extra_width / frame->width;
452         } else if (display_aspect < video_aspect) {
453                 double extra_height = frame->width / display_aspect - frame->height;
454                 ty1 = -0.5 * extra_height / frame->height;
455                 ty2 = 1.0 + 0.5 * extra_height / frame->height;
456         }
457
458         glBegin(GL_QUADS);
459
460         // (0,0)
461         glVertexAttrib2f(1, tx1, ty1);
462         glVertex2f(zoom_matrix[2 * 3 + 0], zoom_matrix[2 * 3 + 1]);
463
464         // (0,1)
465         glVertexAttrib2f(1, tx1, ty2);
466         glVertex2f(zoom_matrix[1 * 3 + 0] + zoom_matrix[2 * 3 + 0], zoom_matrix[1 * 3 + 1] + zoom_matrix[2 * 3 + 1]);
467
468         // (1,1)
469         glVertexAttrib2f(1, tx2, ty2);
470         glVertex2f(zoom_matrix[0 * 3 + 0] + zoom_matrix[1 * 3 + 0] + zoom_matrix[2 * 3 + 0],
471                    zoom_matrix[1 * 3 + 0] + zoom_matrix[1 * 3 + 1] + zoom_matrix[2 * 3 + 1]);
472
473         // (1,0)
474         glVertexAttrib2f(1, tx2, ty1);
475         glVertex2f(zoom_matrix[0 * 3 + 0] + zoom_matrix[2 * 3 + 0],
476                    zoom_matrix[1 * 3 + 0] + zoom_matrix[2 * 3 + 1]);
477
478         glEnd();
479 }
480
481 void VideoWindow::set_current_frame(shared_ptr<VideoWidget::Frame> new_frame)
482 {
483         {
484                 lock_guard lock(current_frame_mu);
485                 current_frame = std::move(new_frame);
486         }
487         update();
488 }
489
490 void matmul3x3(const double a[9], const double b[9], double res[9])
491 {
492         for (int i = 0; i < 3; ++i) {
493                 for (int j = 0; j < 3; ++j) {
494                         double sum = 0.0;
495                         for (int k = 0; k < 3; ++k) {
496                                 sum += a[i * 3 + k] * b[k * 3 + j];
497                         }
498                         res[i * 3 + j] = sum;
499                 }
500         }
501 }
502
503 void VideoWidget::wheelEvent(QWheelEvent *event)
504 {
505         int delta = event->angleDelta().y();
506         if (delta == 0) {
507                 return;
508         }
509         double x = event->position().x() / width();
510         double y = 1.0 - event->position().y() / height();
511         double zoom = delta > 0 ? pow(1.005, delta) : pow(1/1.005, -delta);
512
513         const double inv_translation_matrix[9] = {
514                 1.0, 0.0, 0.0,
515                 0.0, 1.0, 0.0,
516                 -x, -y, 1.0
517         };
518         const double scale_matrix[9] = {
519                 zoom, 0.0, 0.0,
520                 0.0, zoom, 0.0,
521                 0.0, 0.0, 1.0
522         };
523         const double translation_matrix[9] = {
524                 1.0, 0.0, 0.0,
525                 0.0, 1.0, 0.0,
526                 x, y, 1.0
527         };
528         double tmp1[9], tmp2[9];
529         matmul3x3(zoom_matrix, inv_translation_matrix, tmp1);
530         matmul3x3(tmp1, scale_matrix, tmp2);
531         matmul3x3(tmp2, translation_matrix, zoom_matrix);
532
533         fixup_zoom_matrix();
534         video_window->set_zoom_matrix(zoom_matrix);
535         update();
536 }
537
538 void VideoWidget::mousePressEvent(QMouseEvent *e)
539 {
540         if (e->button() == Qt::BackButton) {
541                 emit mouse_back_clicked();
542         } else if (e->button() == Qt::ForwardButton) {
543                 emit mouse_forward_clicked();
544         } else if (e->button() == Qt::LeftButton) {
545                 dragging = true;
546                 last_drag_x = e->position().x();
547                 last_drag_y = e->position().y();
548         }
549 }
550
551 void VideoWidget::mouseReleaseEvent(QMouseEvent *e)
552 {
553         if (e->button() == Qt::LeftButton) {
554                 dragging = false;
555         }
556 }
557
558 void VideoWidget::mouseMoveEvent(QMouseEvent *e)
559 {
560         if (!dragging) {
561                 return;
562         }
563         float dx = (e->position().x() - last_drag_x) / width();
564         float dy = (e->position().y() - last_drag_y) / height();
565
566         //zoom_matrix[6] += dx * zoom_matrix[0];
567         //zoom_matrix[7] += dy * zoom_matrix[4];
568         zoom_matrix[6] += dx;
569         zoom_matrix[7] -= dy;
570         fixup_zoom_matrix();
571         video_window->set_zoom_matrix(zoom_matrix);
572
573         last_drag_x = e->position().x();
574         last_drag_y = e->position().y();
575
576         update();
577 }
578
579 // Normalize the matrix so that we never get skew or similar,
580 // and also never can zoom or pan too far out.
581 void VideoWidget::fixup_zoom_matrix()
582 {
583         // Correct for any numerical errors (we know the matrix must be orthogonal
584         // and have zero rotation).
585         zoom_matrix[4] = zoom_matrix[0];
586         zoom_matrix[1] = zoom_matrix[2] = zoom_matrix[3] = zoom_matrix[5] = 0.0;
587         zoom_matrix[8] = 1.0;
588
589         // We can't zoom further out than 1:1. (Perhaps it would be nice to
590         // reuse the last zoom-in point to do this, but the center will have to do
591         // for now.)
592         if (zoom_matrix[0] < 1.0) {
593                 const double zoom = 1.0 / zoom_matrix[0];
594                 const double inv_translation_matrix[9] = {
595                         1.0, 0.0, 0.0,
596                         0.0, 1.0, 0.0,
597                         -0.5, -0.5, 1.0
598                 };
599                 const double scale_matrix[9] = {
600                         zoom, 0.0, 0.0,
601                         0.0, zoom, 0.0,
602                         0.0, 0.0, 1.0
603                 };
604                 const double translation_matrix[9] = {
605                         1.0, 0.0, 0.0,
606                         0.0, 1.0, 0.0,
607                         0.5, 0.5, 1.0
608                 };
609                 double tmp1[9], tmp2[9];
610                 matmul3x3(zoom_matrix, inv_translation_matrix, tmp1);
611                 matmul3x3(tmp1, scale_matrix, tmp2);
612                 matmul3x3(tmp2, translation_matrix, zoom_matrix);
613         }
614
615         // Looking at the points we'll draw with glVertex2f(), make sure none of them are
616         // inside the square (which would generally mean we've panned ourselves out-of-bounds).
617         // We simply adjust the translation, which is possible because we fixed scaling above.
618         zoom_matrix[6] = min(zoom_matrix[6], 0.0);  // Left side (x=0).
619         zoom_matrix[7] = min(zoom_matrix[7], 0.0);  // Bottom side (y=0).
620         zoom_matrix[6] = std::max(zoom_matrix[6], 1.0 - zoom_matrix[0]);  // Right side (x=1).
621         zoom_matrix[7] = std::max(zoom_matrix[7], 1.0 - zoom_matrix[4]);  // Top side (y=1).
622 }
623
624 bool VideoWidget::open(const string &filename)
625 {
626         stop();
627         internal_rewind();
628         pathname = filename;
629         play();
630
631         while (running == STARTING) {
632                 // Poor man's condition variable...
633                 usleep(10000);
634                 sched_yield();
635         }
636         return (running != VIDEO_FILE_ERROR);   
637 }
638
639 void VideoWidget::play()
640 {
641         if (running != NOT_RUNNING && running != VIDEO_FILE_ERROR) {
642                 std::lock_guard<std::mutex> lock(queue_mu);
643                 command_queue.push_back(QueuedCommand { QueuedCommand::RESUME });
644                 producer_thread_should_quit.wakeup();
645                 return;
646         }
647         running = STARTING;
648         producer_thread_should_quit.unquit();
649         if (producer_thread.joinable()) {
650                 producer_thread.join();
651         }
652         producer_thread = std::thread(&VideoWidget::producer_thread_func, this);
653 }
654
655 void VideoWidget::pause()
656 {
657         if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
658                 return;
659         }
660         std::lock_guard<std::mutex> lock(queue_mu);
661         command_queue.push_back(QueuedCommand { QueuedCommand::PAUSE });
662         producer_thread_should_quit.wakeup();
663 }
664
665 void VideoWidget::seek(int64_t relative_seek_ms)
666 {
667         if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
668                 return;
669         }
670         std::lock_guard<std::mutex> lock(queue_mu);
671         command_queue.push_back(QueuedCommand { QueuedCommand::SEEK, relative_seek_ms, 0, 0 });
672         producer_thread_should_quit.wakeup();
673 }
674
675 void VideoWidget::seek_frames(int64_t relative_seek_frames)
676 {
677         if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
678                 return;
679         }
680         std::lock_guard<std::mutex> lock(queue_mu);
681         command_queue.push_back(QueuedCommand { QueuedCommand::SEEK, 0, relative_seek_frames, 0 });
682         producer_thread_should_quit.wakeup();
683 }
684
685 void VideoWidget::seek_absolute(int64_t position_ms)
686 {
687         if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
688                 return;
689         }
690         std::lock_guard<std::mutex> lock(queue_mu);
691         command_queue.push_back(QueuedCommand { QueuedCommand::SEEK_ABSOLUTE, 0, 0, position_ms });
692         producer_thread_should_quit.wakeup();
693 }
694
695 void VideoWidget::stop()
696 {
697         if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
698                 return;
699         }
700         producer_thread_should_quit.quit();
701         producer_thread.join();
702 }
703
704 void VideoWidget::producer_thread_func()
705 {
706         if (!producer_thread_should_quit.should_quit()) {
707                 if (!play_video(pathname)) {
708                         running = VIDEO_FILE_ERROR;
709                 } else {
710                         running = NOT_RUNNING;
711                 }
712         }
713 }
714
715 void VideoWidget::internal_rewind()
716 {
717         pts_origin = last_pts = 0;
718         last_position = 0;
719         start = next_frame_start = steady_clock::now();
720 }
721
722 template<AVHWDeviceType type>
723 AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
724 {
725         bool found_config_of_right_type = false;
726         for (int i = 0;; ++i) {  // Termination condition inside loop.
727                 const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
728                 if (config == nullptr) {  // End of list.
729                         break;
730                 }
731                 if (!(config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) ||
732                     config->device_type != type) {
733                         // Not interesting for us.
734                         continue;
735                 }
736
737                 // We have a config of the right type, but does it actually support
738                 // the pixel format we want? (Seemingly, FFmpeg's way of signaling errors
739                 // is to just replace the pixel format with a software-decoded one,
740                 // such as yuv420p.)
741                 found_config_of_right_type = true;
742                 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
743                         if (config->pix_fmt == *fmt_ptr) {
744                                 fprintf(stderr, "Initialized '%s' hardware decoding for codec '%s'.\n",
745                                         av_hwdevice_get_type_name(type), ctx->codec->name);
746                                 if (ctx->profile == FF_PROFILE_H264_BASELINE) {
747                                         fprintf(stderr, "WARNING: Stream claims to be H.264 Baseline, which is generally poorly supported in hardware decoders.\n");
748                                         fprintf(stderr, "         Consider encoding it as Constrained Baseline, Main or High instead.\n");
749                                         fprintf(stderr, "         Decoding might fail and fall back to software.\n");
750                                 }
751                                 return config->pix_fmt;
752                         }
753                 }
754                 fprintf(stderr, "Decoder '%s' supports only these pixel formats:", ctx->codec->name);
755                 unordered_set<AVPixelFormat> seen;
756                 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
757                         if (!seen.count(*fmt_ptr)) {
758                                 fprintf(stderr, " %s", av_get_pix_fmt_name(*fmt_ptr));
759                                 seen.insert(*fmt_ptr);
760                         }
761                 }
762                 fprintf(stderr, " (wanted %s for hardware acceleration)\n", av_get_pix_fmt_name(config->pix_fmt));
763
764         }
765
766         if (!found_config_of_right_type) {
767                 fprintf(stderr, "Decoder '%s' does not support device type '%s'.\n", ctx->codec->name, av_hwdevice_get_type_name(type));
768         }
769
770         // We found no VA-API formats, so take the first software format.
771         for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
772                 if ((av_pix_fmt_desc_get(*fmt_ptr)->flags & AV_PIX_FMT_FLAG_HWACCEL) == 0) {
773                         fprintf(stderr, "Falling back to software format %s.\n", av_get_pix_fmt_name(*fmt_ptr));
774                         return *fmt_ptr;
775                 }
776         }
777
778         // Fallback: Just return anything. (Should never really happen.)
779         return fmt[0];
780 }
781
782 AVFrameWithDeleter VideoWidget::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx,
783         const std::string &pathname, int video_stream_index,
784         bool *error)
785 {
786         *error = false;
787
788         if (!queued_frames.empty()) {
789                 AVFrameWithDeleter frame = std::move(queued_frames.front());
790                 queued_frames.pop_front();
791                 return frame;
792         }
793
794         // Read packets until we have a frame or there are none left.
795         bool frame_finished = false;
796         AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
797         bool eof = false;
798         do {
799                 AVPacket *pkt = av_packet_alloc();
800                 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
801                         pkt, av_packet_unref);
802                 pkt->data = nullptr;
803                 pkt->size = 0;
804                 if (av_read_frame(format_ctx, pkt) == 0) {
805                         if (pkt->stream_index == video_stream_index) {
806                                 if (avcodec_send_packet(video_codec_ctx, pkt) < 0) {
807                                         fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
808                                         *error = true;
809                                         return AVFrameWithDeleter(nullptr);
810                                 }
811                         }
812                 } else {
813                         eof = true;  // Or error, but ignore that for the time being.
814                 }
815
816                 // Decode video, if we have a frame.
817                 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
818                 if (err == 0) {
819                         frame_finished = true;
820                         break;
821                 } else if (err != AVERROR(EAGAIN)) {
822                         fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
823                         *error = true;
824                         return AVFrameWithDeleter(nullptr);
825                 }
826         } while (!eof);
827
828         if (frame_finished)
829                 return video_avframe;
830         else
831                 return AVFrameWithDeleter(nullptr);
832 }
833
834 int find_stream_index(AVFormatContext *ctx, AVMediaType media_type)
835 {
836         for (unsigned i = 0; i < ctx->nb_streams; ++i) {
837                 if (ctx->streams[i]->codecpar->codec_type == media_type) {
838                         return i;
839                 }
840         }
841         return -1;
842 }
843
844 steady_clock::time_point compute_frame_start(int64_t frame_pts, int64_t pts_origin, const AVRational &video_timebase, const steady_clock::time_point &origin, double rate)
845 {
846         const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
847         return origin + duration_cast<steady_clock::duration>(pts / rate);
848 }
849
850 bool VideoWidget::play_video(const string &pathname)
851 {
852         queued_frames.clear();
853         AVFormatContextWithCloser format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
854                 /*options=*/nullptr);
855         if (format_ctx == nullptr) {
856                 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
857                 return false;
858         }
859
860         if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
861                 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
862                 return false;
863         }
864
865         int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
866         if (video_stream_index == -1) {
867                 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
868                 return false;
869         }
870
871         // Open video decoder.
872         const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
873         const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
874
875         video_timebase = format_ctx->streams[video_stream_index]->time_base;
876         AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
877         if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
878                 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
879                 return false;
880         }
881         if (video_codec == nullptr) {
882                 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
883                 return false;
884         }
885
886         // Seemingly, it's not too easy to make something that just initializes
887         // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
888         // without enumerating through several different types.
889         // VA-API and VDPAU will do for now. We prioritize VDPAU for the
890         // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
891         // cards that seems to work, but just hangs when trying to transfer the frame.
892         //
893         // Note that we don't actually check codec support beforehand,
894         // so if you have a low-end VDPAU device but a high-end VA-API device,
895         // you lose out on the extra codec support from the latter.
896         AVBufferRef *hw_device_ctx = nullptr;
897         if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
898                 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
899                 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
900         } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
901                 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
902                 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
903         } else {
904                 fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
905         }
906
907         if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
908                 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
909                 return false;
910         }
911         unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
912                 video_codec_ctx.get(), avcodec_close);
913
914         internal_rewind();
915
916         running = RUNNING;
917
918         // Main loop.
919         int consecutive_errors = 0;
920         double rate = 1.0;
921         while (!producer_thread_should_quit.should_quit()) {
922                 if (process_queued_commands(format_ctx.get(), video_codec_ctx.get(), video_stream_index, /*seeked=*/nullptr)) {
923                         return true;
924                 }
925                 if (paused) {
926                         producer_thread_should_quit.sleep_for(hours(1));
927                         continue;
928                 }
929
930                 bool error;
931                 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(),
932                         pathname, video_stream_index, &error);
933                 if (error) {
934                         if (++consecutive_errors >= 100) {
935                                 fprintf(stderr, "More than 100 consecutive video frames, aborting playback.\n");
936                                 return false;
937                         } else {
938                                 continue;
939                         }
940                 } else {
941                         consecutive_errors = 0;
942                 }
943                 if (frame == nullptr) {
944                         // EOF.
945                         return false;
946                 }
947
948                 // Sleep until it's time to present this frame.
949                 for ( ;; ) {
950                         if (last_pts == 0 && pts_origin == 0) {
951                                 pts_origin = frame->pts;
952                         }
953                         steady_clock::time_point now = steady_clock::now();
954                         next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
955
956                         if (duration<double>(now - next_frame_start).count() >= 0.1) {
957                                 // If we don't have enough CPU to keep up, or if we have a live stream
958                                 // where the initial origin was somehow wrong, we could be behind indefinitely.
959                                 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
960                                         pathname.c_str(),
961                                         1e3 * duration<double>(now - next_frame_start).count());
962                                 pts_origin = frame->pts;
963                                 start = next_frame_start = now;
964                         }
965                         bool finished_wakeup;
966                         finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
967                         if (finished_wakeup) {
968                                 video_window->set_current_frame(make_video_frame(frame.get()));
969                                 last_frame = steady_clock::now();
970                                 update();
971                                 break;
972                         } else {
973                                 if (producer_thread_should_quit.should_quit()) break;
974
975                                 bool seeked = false;
976                                 if (process_queued_commands(format_ctx.get(), video_codec_ctx.get(), video_stream_index, &seeked)) {
977                                         return true;
978                                 }
979
980                                 if (paused) {
981                                         // Just paused, so present the frame immediately and then go into deep sleep.
982                                         video_window->set_current_frame(make_video_frame(frame.get()));
983                                         last_frame = steady_clock::now();
984                                         update();
985                                         break;
986                                 }
987
988                                 // If we just seeked, drop this frame on the floor and be done.
989                                 if (seeked) {
990                                         break;
991                                 }
992                         }
993                 }
994                 store_pts(frame->pts);
995         }
996         return true;
997 }
998
999 void VideoWidget::store_pts(int64_t pts)
1000 {
1001         last_pts = pts;
1002         last_position = lrint(pts * double(video_timebase.num) / double(video_timebase.den) * 1000);
1003         post_to_main_thread([this, last_position{last_position.load()}] {
1004                 emit position_changed(last_position);
1005         });
1006 }
1007
1008 // Taken from Movit (see the comment there for explanation)
1009 float compute_chroma_offset(float pos, unsigned subsampling_factor, unsigned resolution)
1010 {
1011         float local_chroma_pos = (0.5 + pos * (subsampling_factor - 1)) / subsampling_factor;
1012         if (fabs(local_chroma_pos - 0.5) < 1e-10) {
1013                 // x + (-0) can be optimized away freely, as opposed to x + 0.
1014                 return -0.0;
1015         } else {
1016                 return (0.5 - local_chroma_pos) / resolution;
1017         }
1018 }
1019
1020 shared_ptr<VideoWidget::Frame> VideoWidget::alloc_frame(unsigned width, unsigned height, unsigned chroma_width, unsigned chroma_height)
1021 {
1022         lock_guard lock(freelist_mu);
1023         for (auto it = frame_freelist.begin(); it != frame_freelist.end(); ++it) {
1024                 if ((*it)->width == width &&
1025                     (*it)->height == height &&
1026                     (*it)->chroma_width == chroma_width &&
1027                     (*it)->chroma_height == chroma_height) {
1028                         Frame *frame = *it;
1029                         frame_freelist.erase(it);
1030                         return shared_ptr<Frame>{frame, free_frame};
1031                 }
1032         }
1033
1034         Frame *frame = new Frame;
1035         frame->owner = this;
1036         frame->width = width;
1037         frame->height = height;
1038         frame->chroma_width = chroma_width;
1039         frame->chroma_height = chroma_height;
1040
1041         size_t len = frame->width * frame->height + 2 * frame->chroma_width * frame->chroma_height;
1042
1043         while (!video_window->isValid()) {
1044                 usleep(100000);
1045         }
1046
1047         // Augh :-)
1048         mutex mu;
1049         condition_variable done_cv;
1050         bool done = false;
1051
1052         post_to_main_thread([this, &frame, len, &done, &mu, &done_cv]{
1053                 video_window->makeCurrent();
1054                 glCreateBuffers(1, &frame->pbo);
1055                 glNamedBufferStorage(frame->pbo, len, nullptr, GL_MAP_WRITE_BIT | GL_MAP_PERSISTENT_BIT);
1056                 frame->data = (uint8_t *)glMapNamedBufferRange(frame->pbo, 0, len, GL_MAP_WRITE_BIT | GL_MAP_FLUSH_EXPLICIT_BIT | GL_MAP_PERSISTENT_BIT);
1057                 video_window->doneCurrent();
1058
1059                 lock_guard lock(mu);
1060                 done = true;
1061                 done_cv.notify_all();
1062         });
1063         {
1064                 unique_lock lock(mu);
1065                 done_cv.wait(lock, [&done]{ return done; });
1066         }
1067
1068         return shared_ptr<Frame>{frame, free_frame};
1069 }
1070
1071 void VideoWidget::free_frame(VideoWidget::Frame *frame)
1072 {
1073         VideoWidget *self = frame->owner;
1074         lock_guard lock(self->freelist_mu);
1075         if (self->frame_freelist.size() >= 16) {
1076                 GLuint pbo = frame->pbo;
1077                 post_to_main_thread([self, pbo]{
1078                         self->video_window->makeCurrent();
1079                         glUnmapNamedBuffer(pbo);
1080                         glDeleteBuffers(1, &pbo);
1081                         self->video_window->doneCurrent();
1082                 });
1083                 delete self->frame_freelist.front();
1084                 self->frame_freelist.pop_front();
1085         }
1086         self->frame_freelist.push_back(frame);
1087 }
1088
1089 shared_ptr<VideoWidget::Frame> VideoWidget::make_video_frame(const AVFrame *frame)
1090 {
1091         AVFrameWithDeleter sw_frame;
1092
1093         if (frame->format == AV_PIX_FMT_VAAPI ||
1094             frame->format == AV_PIX_FMT_VDPAU) {
1095                 // Get the frame down to the CPU. (TODO: See if we can keep it
1096                 // on the GPU all the way, since it will be going up again later.
1097                 // However, this only works if the OpenGL GPU is the same one.)
1098                 sw_frame = av_frame_alloc_unique();
1099                 int err = av_hwframe_transfer_data(sw_frame.get(), frame, 0);
1100                 if (err != 0) {
1101                         fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
1102                 } else {
1103                         sw_frame->pts = frame->pts;
1104                         sw_frame->pkt_duration = frame->pkt_duration;
1105                         frame = sw_frame.get();
1106                 }
1107         }
1108
1109         if (sws_ctx == nullptr ||
1110             sws_last_width != frame->width ||
1111             sws_last_height != frame->height ||
1112             sws_last_src_format != frame->format) {
1113                 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format));
1114                 sws_ctx.reset(
1115                         sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1116                                 frame->width, frame->height, sws_dst_format,
1117                                 SWS_BICUBIC, nullptr, nullptr, nullptr));
1118                 sws_last_width = frame->width;
1119                 sws_last_height = frame->height;
1120                 sws_last_src_format = frame->format;
1121         }
1122         if (sws_ctx == nullptr) {
1123                 fprintf(stderr, "Could not create scaler context\n");
1124                 abort();
1125         }
1126
1127         uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1128         int linesizes[4] = { 0, 0, 0, 0 };
1129         const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1130
1131         shared_ptr<Frame> video_frame = alloc_frame(
1132                 frame->width,
1133                 frame->height,
1134                 AV_CEIL_RSHIFT(int(frame->width), desc->log2_chroma_w),
1135                 AV_CEIL_RSHIFT(int(frame->height), desc->log2_chroma_h));
1136
1137         // We always assume left chroma placement for now.
1138         video_window->set_cbcr_offset(
1139                 compute_chroma_offset(0.0f, 1 << desc->log2_chroma_w, video_frame->chroma_width),
1140                 compute_chroma_offset(0.5f, 1 << desc->log2_chroma_h, video_frame->chroma_height)
1141         );
1142
1143         pic_data[0] = video_frame->data;
1144         linesizes[0] = frame->width;
1145
1146         pic_data[1] = pic_data[0] + frame->width * frame->height;
1147         linesizes[1] = video_frame->chroma_width;
1148
1149         pic_data[2] = pic_data[1] + video_frame->chroma_width * video_frame->chroma_height;
1150         linesizes[2] = video_frame->chroma_width;
1151
1152         sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);
1153
1154         video_frame->need_flush_len = video_frame->width * video_frame->height + 2 * video_frame->chroma_width * video_frame->chroma_height;
1155
1156         return video_frame;
1157 }
1158