1 #define GL_GLEXT_PROTOTYPES
3 #include "video_widget.h"
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>
31 #include <unordered_set>
33 #include <QOpenGLFunctions>
34 #include <QWheelEvent>
35 #include <QMouseEvent>
38 using namespace std::chrono;
42 bool is_full_range(const AVPixFmtDescriptor *desc)
44 // This is horrible, but there's no better way that I know of.
45 return (strchr(desc->name, 'j') != nullptr);
48 AVPixelFormat decide_dst_format(AVPixelFormat src_format)
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;
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);
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;
80 // 8-bit formats only.
81 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
82 if (desc->comp[0].depth != 8) continue;
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)
90 // Matching full/limited range only.
91 if (is_full_range(desc) != src_full_range)
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) {
98 best_format = desc->name;
101 return av_get_pix_fmt(best_format);
106 bool VideoWidget::process_queued_commands(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, int video_stream_index, bool *seeked)
108 // Process any queued commands from other threads.
109 vector<QueuedCommand> commands;
111 lock_guard<mutex> lock(queue_mu);
112 swap(commands, command_queue);
115 for (const QueuedCommand &cmd : commands) {
116 switch (cmd.command) {
117 case QueuedCommand::PAUSE:
120 case QueuedCommand::RESUME:
122 pts_origin = last_pts;
123 start = next_frame_start = steady_clock::now();
125 case QueuedCommand::SEEK:
126 case QueuedCommand::SEEK_ABSOLUTE:
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 = av_rescale_q(cmd.seek_ms, AVRational{ 1, 1000 }, video_timebase);
143 relative_seek_ms = 0;
144 relative_seek_frames = 0;
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;
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();
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);
164 av_seek_frame(format_ctx, video_stream_index, seek_pts, AVSEEK_FLAG_BACKWARD);
166 // Decode frames until EOF, or until we see something past our seek point.
167 std::deque<AVFrameWithDeleter> queue;
170 AVFrameWithDeleter frame = decode_frame(format_ctx, video_codec_ctx,
171 pathname, video_stream_index, &error);
172 if (frame == nullptr || error) {
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) {
184 if (frame_pts >= goal_pts) {
185 if (relative_seek_frames > 0) {
186 --relative_seek_frames;
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());
193 queued_frames = std::move(queue);
195 current_frame.reset(new Frame(make_video_frame(frame.get())));
197 store_pts(frame->pts);
203 // NOTE: We keep pause status as-is.
205 pts_origin = last_pts;
206 start = next_frame_start = last_frame = steady_clock::now();
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).
215 decode_frame(format_ctx, video_codec_ctx, pathname, video_stream_index, &error);
216 --relative_seek_frames;
219 // Display the last one.
221 AVFrameWithDeleter frame = decode_frame(format_ctx, video_codec_ctx,
222 pathname, video_stream_index, &error);
223 if (frame == nullptr || error) {
226 current_frame.reset(new Frame(make_video_frame(frame.get())));
228 store_pts(frame->pts);
233 VideoWidget::VideoWidget(QWidget *parent)
234 : QOpenGLWidget(parent) {}
236 GLuint compile_shader(const string &shader_src, GLenum type)
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);
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);
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 */ ";
258 for (char ch : shader_src) {
259 src_with_lines.push_back(ch);
262 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
263 src_with_lines += buf;
267 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
274 void VideoWidget::initializeGL()
277 glDisable(GL_DEPTH_TEST);
278 glDepthMask(GL_FALSE);
279 glCreateTextures(GL_TEXTURE_2D, 3, tex);
281 ycbcr_vertex_shader = compile_shader(R"(
284 layout(location = 0) in vec2 position;
285 layout(location = 1) in vec2 texcoord;
290 // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is:
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);
300 )", GL_VERTEX_SHADER);
301 ycbcr_fragment_shader = compile_shader(R"(
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;
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
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);
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;
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);
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);
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);
357 void VideoWidget::resizeGL(int w, int h)
359 glViewport(0, 0, w, h);
360 display_aspect = double(w) / h;
363 int num_levels(GLuint width, GLuint height)
366 while (width > 1 || height > 1) {
367 width = max(width / 2, 1u);
368 height = max(height / 2, 1u);
374 void VideoWidget::paintGL()
376 std::shared_ptr<Frame> frame = current_frame;
377 if (frame == nullptr) {
378 glClear(GL_COLOR_BUFFER_BIT);
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);
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);
392 glTextureSubImage2D(tex[0], 0, 0, 0, frame->width, frame->height, GL_RED, GL_UNSIGNED_BYTE, frame->data.get());
393 glGenerateTextureMipmap(tex[0]);
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]);
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]);
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]);
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;
431 glVertexAttrib2f(1, tx1, ty1);
432 glVertex2f(zoom_matrix[2 * 3 + 0], zoom_matrix[2 * 3 + 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]);
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]);
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]);
451 void matmul3x3(const double a[9], const double b[9], double res[9])
453 for (int i = 0; i < 3; ++i) {
454 for (int j = 0; j < 3; ++j) {
456 for (int k = 0; k < 3; ++k) {
457 sum += a[i * 3 + k] * b[k * 3 + j];
459 res[i * 3 + j] = sum;
464 void VideoWidget::wheelEvent(QWheelEvent *event)
466 int delta = event->angleDelta().y();
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);
474 const double inv_translation_matrix[9] = {
479 const double scale_matrix[9] = {
484 const double translation_matrix[9] = {
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);
498 void VideoWidget::mousePressEvent(QMouseEvent *e)
500 if (e->button() == Qt::BackButton) {
501 emit mouse_back_clicked();
502 } else if (e->button() == Qt::ForwardButton) {
503 emit mouse_forward_clicked();
504 } else if (e->button() == Qt::LeftButton) {
506 last_drag_x = e->position().x();
507 last_drag_y = e->position().y();
511 void VideoWidget::mouseReleaseEvent(QMouseEvent *e)
513 if (e->button() == Qt::LeftButton) {
518 void VideoWidget::mouseMoveEvent(QMouseEvent *e)
523 float dx = (e->position().x() - last_drag_x) / width();
524 float dy = (e->position().y() - last_drag_y) / height();
526 //zoom_matrix[6] += dx * zoom_matrix[0];
527 //zoom_matrix[7] += dy * zoom_matrix[4];
528 zoom_matrix[6] += dx;
529 zoom_matrix[7] -= dy;
532 last_drag_x = e->position().x();
533 last_drag_y = e->position().y();
538 // Normalize the matrix so that we never get skew or similar,
539 // and also never can zoom or pan too far out.
540 void VideoWidget::fixup_zoom_matrix()
542 // Correct for any numerical errors (we know the matrix must be orthogonal
543 // and have zero rotation).
544 zoom_matrix[4] = zoom_matrix[0];
545 zoom_matrix[1] = zoom_matrix[2] = zoom_matrix[3] = zoom_matrix[5] = 0.0;
546 zoom_matrix[8] = 1.0;
548 // We can't zoom further out than 1:1. (Perhaps it would be nice to
549 // reuse the last zoom-in point to do this, but the center will have to do
551 if (zoom_matrix[0] < 1.0) {
552 const double zoom = 1.0 / zoom_matrix[0];
553 const double inv_translation_matrix[9] = {
558 const double scale_matrix[9] = {
563 const double translation_matrix[9] = {
568 double tmp1[9], tmp2[9];
569 matmul3x3(zoom_matrix, inv_translation_matrix, tmp1);
570 matmul3x3(tmp1, scale_matrix, tmp2);
571 matmul3x3(tmp2, translation_matrix, zoom_matrix);
574 // Looking at the points we'll draw with glVertex2f(), make sure none of them are
575 // inside the square (which would generally mean we've panned ourselves out-of-bounds).
576 // We simply adjust the translation, which is possible because we fixed scaling above.
577 zoom_matrix[6] = min(zoom_matrix[6], 0.0); // Left side (x=0).
578 zoom_matrix[7] = min(zoom_matrix[7], 0.0); // Bottom side (y=0).
579 zoom_matrix[6] = std::max(zoom_matrix[6], 1.0 - zoom_matrix[0]); // Right side (x=1).
580 zoom_matrix[7] = std::max(zoom_matrix[7], 1.0 - zoom_matrix[4]); // Top side (y=1).
583 void VideoWidget::open(const string &filename)
591 void VideoWidget::play()
594 std::lock_guard<std::mutex> lock(queue_mu);
595 command_queue.push_back(QueuedCommand { QueuedCommand::RESUME });
596 producer_thread_should_quit.wakeup();
600 producer_thread_should_quit.unquit();
601 producer_thread = std::thread(&VideoWidget::producer_thread_func, this);
604 void VideoWidget::pause()
609 std::lock_guard<std::mutex> lock(queue_mu);
610 command_queue.push_back(QueuedCommand { QueuedCommand::PAUSE });
611 producer_thread_should_quit.wakeup();
614 void VideoWidget::seek(int64_t relative_seek_ms)
619 std::lock_guard<std::mutex> lock(queue_mu);
620 command_queue.push_back(QueuedCommand { QueuedCommand::SEEK, relative_seek_ms, 0, 0 });
621 producer_thread_should_quit.wakeup();
624 void VideoWidget::seek_frames(int64_t relative_seek_frames)
629 std::lock_guard<std::mutex> lock(queue_mu);
630 command_queue.push_back(QueuedCommand { QueuedCommand::SEEK, 0, relative_seek_frames, 0 });
631 producer_thread_should_quit.wakeup();
634 void VideoWidget::seek_absolute(int64_t position_ms)
639 std::lock_guard<std::mutex> lock(queue_mu);
640 command_queue.push_back(QueuedCommand { QueuedCommand::SEEK_ABSOLUTE, 0, 0, position_ms });
641 producer_thread_should_quit.wakeup();
644 void VideoWidget::stop()
650 producer_thread_should_quit.quit();
651 producer_thread.join();
654 void VideoWidget::producer_thread_func()
656 if (!producer_thread_should_quit.should_quit()) {
657 if (!play_video(pathname)) {
658 // TODO: Send the error back to the UI somehow.
663 void VideoWidget::internal_rewind()
665 pts_origin = last_pts = 0;
667 start = next_frame_start = steady_clock::now();
670 template<AVHWDeviceType type>
671 AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
673 bool found_config_of_right_type = false;
674 for (int i = 0;; ++i) { // Termination condition inside loop.
675 const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
676 if (config == nullptr) { // End of list.
679 if (!(config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) ||
680 config->device_type != type) {
681 // Not interesting for us.
685 // We have a config of the right type, but does it actually support
686 // the pixel format we want? (Seemingly, FFmpeg's way of signaling errors
687 // is to just replace the pixel format with a software-decoded one,
689 found_config_of_right_type = true;
690 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
691 if (config->pix_fmt == *fmt_ptr) {
692 fprintf(stderr, "Initialized '%s' hardware decoding for codec '%s'.\n",
693 av_hwdevice_get_type_name(type), ctx->codec->name);
694 if (ctx->profile == FF_PROFILE_H264_BASELINE) {
695 fprintf(stderr, "WARNING: Stream claims to be H.264 Baseline, which is generally poorly supported in hardware decoders.\n");
696 fprintf(stderr, " Consider encoding it as Constrained Baseline, Main or High instead.\n");
697 fprintf(stderr, " Decoding might fail and fall back to software.\n");
699 return config->pix_fmt;
702 fprintf(stderr, "Decoder '%s' supports only these pixel formats:", ctx->codec->name);
703 unordered_set<AVPixelFormat> seen;
704 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
705 if (!seen.count(*fmt_ptr)) {
706 fprintf(stderr, " %s", av_get_pix_fmt_name(*fmt_ptr));
707 seen.insert(*fmt_ptr);
710 fprintf(stderr, " (wanted %s for hardware acceleration)\n", av_get_pix_fmt_name(config->pix_fmt));
714 if (!found_config_of_right_type) {
715 fprintf(stderr, "Decoder '%s' does not support device type '%s'.\n", ctx->codec->name, av_hwdevice_get_type_name(type));
718 // We found no VA-API formats, so take the first software format.
719 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
720 if ((av_pix_fmt_desc_get(*fmt_ptr)->flags & AV_PIX_FMT_FLAG_HWACCEL) == 0) {
721 fprintf(stderr, "Falling back to software format %s.\n", av_get_pix_fmt_name(*fmt_ptr));
726 // Fallback: Just return anything. (Should never really happen.)
730 AVFrameWithDeleter VideoWidget::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx,
731 const std::string &pathname, int video_stream_index,
736 if (!queued_frames.empty()) {
737 AVFrameWithDeleter frame = std::move(queued_frames.front());
738 queued_frames.pop_front();
742 // Read packets until we have a frame or there are none left.
743 bool frame_finished = false;
744 AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
747 AVPacket *pkt = av_packet_alloc();
748 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
749 pkt, av_packet_unref);
752 if (av_read_frame(format_ctx, pkt) == 0) {
753 if (pkt->stream_index == video_stream_index) {
754 if (avcodec_send_packet(video_codec_ctx, pkt) < 0) {
755 fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
757 return AVFrameWithDeleter(nullptr);
761 eof = true; // Or error, but ignore that for the time being.
764 // Decode video, if we have a frame.
765 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
767 frame_finished = true;
769 } else if (err != AVERROR(EAGAIN)) {
770 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
772 return AVFrameWithDeleter(nullptr);
777 return video_avframe;
779 return AVFrameWithDeleter(nullptr);
782 int find_stream_index(AVFormatContext *ctx, AVMediaType media_type)
784 for (unsigned i = 0; i < ctx->nb_streams; ++i) {
785 if (ctx->streams[i]->codecpar->codec_type == media_type) {
792 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)
794 const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
795 return origin + duration_cast<steady_clock::duration>(pts / rate);
798 bool VideoWidget::play_video(const string &pathname)
800 queued_frames.clear();
801 AVFormatContextWithCloser format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
802 /*options=*/nullptr);
803 if (format_ctx == nullptr) {
804 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
808 if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
809 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
813 int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
814 if (video_stream_index == -1) {
815 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
819 // Open video decoder.
820 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
821 const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
823 video_timebase = format_ctx->streams[video_stream_index]->time_base;
824 AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
825 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
826 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
829 if (video_codec == nullptr) {
830 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
834 // Seemingly, it's not too easy to make something that just initializes
835 // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
836 // without enumerating through several different types.
837 // VA-API and VDPAU will do for now. We prioritize VDPAU for the
838 // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
839 // cards that seems to work, but just hangs when trying to transfer the frame.
841 // Note that we don't actually check codec support beforehand,
842 // so if you have a low-end VDPAU device but a high-end VA-API device,
843 // you lose out on the extra codec support from the latter.
844 AVBufferRef *hw_device_ctx = nullptr;
845 if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
846 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
847 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
848 } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
849 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
850 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
852 fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
855 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
856 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
859 unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
860 video_codec_ctx.get(), avcodec_close);
865 int consecutive_errors = 0;
867 while (!producer_thread_should_quit.should_quit()) {
868 if (process_queued_commands(format_ctx.get(), video_codec_ctx.get(), video_stream_index, /*seeked=*/nullptr)) {
872 producer_thread_should_quit.sleep_for(hours(1));
877 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(),
878 pathname, video_stream_index, &error);
880 if (++consecutive_errors >= 100) {
881 fprintf(stderr, "More than 100 consecutive video frames, aborting playback.\n");
887 consecutive_errors = 0;
889 if (frame == nullptr) {
894 // Sleep until it's time to present this frame.
896 if (last_pts == 0 && pts_origin == 0) {
897 pts_origin = frame->pts;
899 steady_clock::time_point now = steady_clock::now();
900 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
902 if (duration<double>(now - next_frame_start).count() >= 0.1) {
903 // If we don't have enough CPU to keep up, or if we have a live stream
904 // where the initial origin was somehow wrong, we could be behind indefinitely.
905 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
907 1e3 * duration<double>(now - next_frame_start).count());
908 pts_origin = frame->pts;
909 start = next_frame_start = now;
911 bool finished_wakeup;
912 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
913 if (finished_wakeup) {
914 current_frame.reset(new Frame(make_video_frame(frame.get())));
915 last_frame = steady_clock::now();
919 if (producer_thread_should_quit.should_quit()) break;
922 if (process_queued_commands(format_ctx.get(), video_codec_ctx.get(), video_stream_index, &seeked)) {
927 // Just paused, so present the frame immediately and then go into deep sleep.
928 current_frame.reset(new Frame(make_video_frame(frame.get())));
929 last_frame = steady_clock::now();
934 // If we just seeked, drop this frame on the floor and be done.
940 store_pts(frame->pts);
945 void VideoWidget::store_pts(int64_t pts)
948 last_position = lrint(pts * double(video_timebase.num) / double(video_timebase.den) * 1000);
949 emit position_changed(last_position);
952 // Taken from Movit (see the comment there for explanation)
953 float compute_chroma_offset(float pos, unsigned subsampling_factor, unsigned resolution)
955 float local_chroma_pos = (0.5 + pos * (subsampling_factor - 1)) / subsampling_factor;
956 if (fabs(local_chroma_pos - 0.5) < 1e-10) {
957 // x + (-0) can be optimized away freely, as opposed to x + 0.
960 return (0.5 - local_chroma_pos) / resolution;
964 VideoWidget::Frame VideoWidget::make_video_frame(const AVFrame *frame)
967 AVFrameWithDeleter sw_frame;
969 if (frame->format == AV_PIX_FMT_VAAPI ||
970 frame->format == AV_PIX_FMT_VDPAU) {
971 // Get the frame down to the CPU. (TODO: See if we can keep it
972 // on the GPU all the way, since it will be going up again later.
973 // However, this only works if the OpenGL GPU is the same one.)
974 sw_frame = av_frame_alloc_unique();
975 int err = av_hwframe_transfer_data(sw_frame.get(), frame, 0);
977 fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
979 sw_frame->pts = frame->pts;
980 sw_frame->pkt_duration = frame->pkt_duration;
981 frame = sw_frame.get();
985 if (sws_ctx == nullptr ||
986 sws_last_width != frame->width ||
987 sws_last_height != frame->height ||
988 sws_last_src_format != frame->format) {
989 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format));
991 sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
992 frame->width, frame->height, sws_dst_format,
993 SWS_BICUBIC, nullptr, nullptr, nullptr));
994 sws_last_width = frame->width;
995 sws_last_height = frame->height;
996 sws_last_src_format = frame->format;
998 if (sws_ctx == nullptr) {
999 fprintf(stderr, "Could not create scaler context\n");
1003 uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1004 int linesizes[4] = { 0, 0, 0, 0 };
1005 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1007 video_frame.width = frame->width;
1008 video_frame.height = frame->height;
1009 video_frame.chroma_width = AV_CEIL_RSHIFT(int(frame->width), desc->log2_chroma_w);
1010 video_frame.chroma_height = AV_CEIL_RSHIFT(int(frame->height), desc->log2_chroma_h);
1012 // We always assume left chroma placement for now.
1013 cbcr_offset[0] = compute_chroma_offset(0.0f, 1 << desc->log2_chroma_w, video_frame.chroma_width);
1014 cbcr_offset[1] = compute_chroma_offset(0.5f, 1 << desc->log2_chroma_h, video_frame.chroma_height);
1016 size_t len = frame->width * frame->height + 2 * video_frame.chroma_width * video_frame.chroma_height;
1017 video_frame.data.reset(new uint8_t[len]);
1019 pic_data[0] = video_frame.data.get();
1020 linesizes[0] = frame->width;
1022 pic_data[1] = pic_data[0] + frame->width * frame->height;
1023 linesizes[1] = video_frame.chroma_width;
1025 pic_data[2] = pic_data[1] + video_frame.chroma_width * video_frame.chroma_height;
1026 linesizes[2] = video_frame.chroma_width;
1028 sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);