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