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 "post_to_main_thread.h"
35 #include <QOpenGLFunctions>
36 #include <QWheelEvent>
37 #include <QMouseEvent>
40 using namespace std::chrono;
44 bool is_full_range(const AVPixFmtDescriptor *desc)
46 // This is horrible, but there's no better way that I know of.
47 return (strchr(desc->name, 'j') != nullptr);
50 AVPixelFormat decide_dst_format(AVPixelFormat src_format)
52 // If this is a non-Y'CbCr format, just convert to 4:4:4 Y'CbCr
53 // and be done with it. It's too strange to spend a lot of time on.
54 // (Let's hope there's no alpha.)
55 const AVPixFmtDescriptor *src_desc = av_pix_fmt_desc_get(src_format);
56 if (src_desc == nullptr ||
57 src_desc->nb_components != 3 ||
58 (src_desc->flags & AV_PIX_FMT_FLAG_RGB)) {
59 return AV_PIX_FMT_YUV444P;
62 // The best for us would be Cb and Cr together if possible,
63 // but FFmpeg doesn't support that except in the special case of
64 // NV12, so we need to go to planar even for the case of NV12.
65 // Thus, look for the closest (but no worse) 8-bit planar Y'CbCr format
66 // that matches in color range. (This will also include the case of
67 // the source format already being acceptable.)
68 bool src_full_range = is_full_range(src_desc);
69 const char *best_format = "yuv444p";
70 unsigned best_score = numeric_limits<unsigned>::max();
71 for (const AVPixFmtDescriptor *desc = av_pix_fmt_desc_next(nullptr);
73 desc = av_pix_fmt_desc_next(desc)) {
74 // Find planar Y'CbCr formats only.
75 if (desc->nb_components != 3) continue;
76 if (desc->flags & AV_PIX_FMT_FLAG_RGB) continue;
77 if (!(desc->flags & AV_PIX_FMT_FLAG_PLANAR)) continue;
78 if (desc->comp[0].plane != 0 ||
79 desc->comp[1].plane != 1 ||
80 desc->comp[2].plane != 2) continue;
82 // 8-bit formats only.
83 if (desc->flags & AV_PIX_FMT_FLAG_BE) continue;
84 if (desc->comp[0].depth != 8) continue;
86 // Same or better chroma resolution only.
87 int chroma_w_diff = src_desc->log2_chroma_w - desc->log2_chroma_w;
88 int chroma_h_diff = src_desc->log2_chroma_h - desc->log2_chroma_h;
89 if (chroma_w_diff < 0 || chroma_h_diff < 0)
92 // Matching full/limited range only.
93 if (is_full_range(desc) != src_full_range)
96 // Pick something with as little excess chroma resolution as possible.
97 unsigned score = (1 << (chroma_w_diff)) << chroma_h_diff;
98 if (score < best_score) {
100 best_format = desc->name;
103 return av_get_pix_fmt(best_format);
108 bool VideoWidget::process_queued_commands(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx, int video_stream_index, bool *seeked)
110 // Process any queued commands from other threads.
111 vector<QueuedCommand> commands;
113 lock_guard<mutex> lock(queue_mu);
114 swap(commands, command_queue);
117 for (const QueuedCommand &cmd : commands) {
118 switch (cmd.command) {
119 case QueuedCommand::PAUSE:
122 case QueuedCommand::RESUME:
124 pts_origin = last_pts;
125 start = next_frame_start = steady_clock::now();
127 case QueuedCommand::SEEK:
128 case QueuedCommand::SEEK_ABSOLUTE:
134 // Combine all seeks into one big one. (There are edge cases where this is probably
135 // subtly wrong, but we'll live with it.)
136 int64_t base_pts = last_pts;
137 int64_t relative_seek_ms = 0;
138 int64_t relative_seek_frames = 0;
139 for (const QueuedCommand &cmd : commands) {
140 if (cmd.command == QueuedCommand::SEEK) {
141 relative_seek_ms += cmd.relative_seek_ms;
142 relative_seek_frames += cmd.relative_seek_frames;
143 } else if (cmd.command == QueuedCommand::SEEK_ABSOLUTE) {
144 base_pts = av_rescale_q(cmd.seek_ms, AVRational{ 1, 1000 }, video_timebase);
145 relative_seek_ms = 0;
146 relative_seek_frames = 0;
149 int64_t relative_seek_pts = av_rescale_q(relative_seek_ms, AVRational{ 1, 1000 }, video_timebase);
150 if (relative_seek_ms != 0 && relative_seek_pts == 0) {
151 // Just to be sure rounding errors don't move us into nothingness.
152 relative_seek_pts = (relative_seek_ms > 0) ? 1 : -1;
154 int64_t goal_pts = base_pts + relative_seek_pts;
155 if (goal_pts != last_pts || relative_seek_frames < 0) {
156 avcodec_flush_buffers(video_codec_ctx);
157 queued_frames.clear();
159 // Seek to the last keyframe before this point.
160 int64_t seek_pts = goal_pts;
161 if (relative_seek_frames < 0) {
162 // If we're frame-skipping backwards, add 100 ms of slop for each frame
163 // so we're fairly certain we are able to see the ones we want.
164 seek_pts -= av_rescale_q(-relative_seek_frames, AVRational{ 1, 10 }, video_timebase);
166 av_seek_frame(format_ctx, video_stream_index, seek_pts, AVSEEK_FLAG_BACKWARD);
168 // Decode frames until EOF, or until we see something past our seek point.
169 std::deque<AVFrameWithDeleter> queue;
172 AVFrameWithDeleter frame = decode_frame(format_ctx, video_codec_ctx,
173 pathname, video_stream_index, &error);
174 if (frame == nullptr || error) {
178 int64_t frame_pts = frame->pts;
179 if (relative_seek_frames < 0) {
180 // Buffer this frame; don't display it unless we know it's the Nth-latest.
181 queue.push_back(std::move(frame));
182 if (queue.size() > uint64_t(-relative_seek_frames) + 1) {
186 if (frame_pts >= goal_pts) {
187 if (relative_seek_frames > 0) {
188 --relative_seek_frames;
190 if (relative_seek_frames < 0) {
191 // Hope we have the right amount.
192 // The rest will remain in the queue for when we play forward again.
193 frame = std::move(queue.front());
195 queued_frames = std::move(queue);
197 Frame *new_frame = new Frame(make_video_frame(frame.get()));
199 lock_guard lock(current_frame_mu);
200 current_frame.reset(new_frame);
203 store_pts(frame->pts);
209 // NOTE: We keep pause status as-is.
211 pts_origin = last_pts;
212 start = next_frame_start = last_frame = steady_clock::now();
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).
221 decode_frame(format_ctx, video_codec_ctx, pathname, video_stream_index, &error);
222 --relative_seek_frames;
225 // Display the last one.
227 AVFrameWithDeleter frame = decode_frame(format_ctx, video_codec_ctx,
228 pathname, video_stream_index, &error);
229 if (frame == nullptr || error) {
232 Frame *new_frame = new Frame(make_video_frame(frame.get()));
234 lock_guard lock(current_frame_mu);
235 current_frame.reset(new_frame);
238 store_pts(frame->pts);
243 VideoWidget::VideoWidget(QWidget *parent)
244 : QOpenGLWidget(parent) {}
246 GLuint compile_shader(const string &shader_src, GLenum type)
248 GLuint obj = glCreateShader(type);
249 const GLchar* source[] = { shader_src.data() };
250 const GLint length[] = { (GLint)shader_src.size() };
251 glShaderSource(obj, 1, source, length);
252 glCompileShader(obj);
254 GLchar info_log[4096];
255 GLsizei log_length = sizeof(info_log) - 1;
256 glGetShaderInfoLog(obj, log_length, &log_length, info_log);
257 info_log[log_length] = 0;
258 if (strlen(info_log) > 0) {
259 fprintf(stderr, "Shader compile log: %s\n", info_log);
263 glGetShaderiv(obj, GL_COMPILE_STATUS, &status);
264 if (status == GL_FALSE) {
265 // Add some line numbers to easier identify compile errors.
266 string src_with_lines = "/* 1 */ ";
268 for (char ch : shader_src) {
269 src_with_lines.push_back(ch);
272 snprintf(buf, sizeof(buf), "/* %3zu */ ", ++lineno);
273 src_with_lines += buf;
277 fprintf(stderr, "Failed to compile shader:\n%s\n", src_with_lines.c_str());
284 void VideoWidget::initializeGL()
287 glDisable(GL_DEPTH_TEST);
288 glDepthMask(GL_FALSE);
289 glCreateTextures(GL_TEXTURE_2D, 3, tex);
291 ycbcr_vertex_shader = compile_shader(R"(
294 layout(location = 0) in vec2 position;
295 layout(location = 1) in vec2 texcoord;
300 // The result of glOrtho(0.0, 1.0, 0.0, 1.0, 0.0, 1.0) is:
302 // 2.000 0.000 0.000 -1.000
303 // 0.000 2.000 0.000 -1.000
304 // 0.000 0.000 -2.000 -1.000
305 // 0.000 0.000 0.000 1.000
306 gl_Position = vec4(2.0 * position.x - 1.0, 2.0 * position.y - 1.0, -1.0, 1.0);
310 )", GL_VERTEX_SHADER);
311 ycbcr_fragment_shader = compile_shader(R"(
314 layout(location = 0) uniform sampler2D tex_y;
315 layout(location = 1) uniform sampler2D tex_cb;
316 layout(location = 2) uniform sampler2D tex_cr;
317 layout(location = 3) uniform vec2 cbcr_offset;
322 // Computed statically by Movit, for limited-range BT.709.
323 // (We don't check whether the input could be BT.601 or BT.2020 currently, or full-range)
324 const mat3 inv_ycbcr_matrix = mat3(
325 1.16438f, 1.16438f, 1.16438f,
326 0.0f, -0.21325f, 2.11240f,
327 1.79274f, -0.53291f, 0.0f
332 if (tc.x < 0.0 || tc.x > 1.0 || tc.y < 0.0 || tc.y > 1.0) {
333 FragColor.rgba = vec4(0.0f, 0.0f, 0.0f, 1.0f);
338 ycbcr.r = texture(tex_y, tc).r;
339 ycbcr.g = texture(tex_cb, tc + cbcr_offset).r;
340 ycbcr.b = texture(tex_cr, tc + cbcr_offset).r;
341 ycbcr -= vec3(16.0f / 255.0f, 128.0f / 255.0f, 128.0f / 255.0f);
342 FragColor.rgb = inv_ycbcr_matrix * ycbcr;
345 )", GL_FRAGMENT_SHADER);
346 ycbcr_program = glCreateProgram();
347 glAttachShader(ycbcr_program, ycbcr_vertex_shader);
348 glAttachShader(ycbcr_program, ycbcr_fragment_shader);
349 glLinkProgram(ycbcr_program);
352 glGetProgramiv(ycbcr_program, GL_LINK_STATUS, &success);
353 if (success == GL_FALSE) {
354 GLchar error_log[1024] = {0};
355 glGetProgramInfoLog(ycbcr_program, 1024, nullptr, error_log);
356 fprintf(stderr, "Error linking program: %s\n", error_log);
360 glCreateSamplers(1, &bilinear_sampler);
361 glSamplerParameteri(bilinear_sampler, GL_TEXTURE_MIN_FILTER, GL_LINEAR_MIPMAP_NEAREST);
362 glSamplerParameteri(bilinear_sampler, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
363 glSamplerParameteri(bilinear_sampler, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
364 glSamplerParameteri(bilinear_sampler, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
367 void VideoWidget::resizeGL(int w, int h)
369 glViewport(0, 0, w, h);
370 display_aspect = double(w) / h;
373 int num_levels(GLuint width, GLuint height)
376 while (width > 1 || height > 1) {
377 width = max(width / 2, 1u);
378 height = max(height / 2, 1u);
384 void VideoWidget::paintGL()
386 std::shared_ptr<Frame> frame;
388 lock_guard lock(current_frame_mu);
389 frame = current_frame;
391 if (frame == nullptr) {
392 glClear(GL_COLOR_BUFFER_BIT);
396 glUseProgram(ycbcr_program);
397 if (frame->width != last_width || frame->height != last_height) {
398 glTextureStorage2D(tex[0], num_levels(frame->width, frame->height), GL_R8, frame->width, frame->height);
400 if (frame->chroma_width != last_chroma_width || frame->chroma_height != last_chroma_height) {
401 for (GLuint num : { tex[1], tex[2] }) {
402 glTextureStorage2D(num, num_levels(frame->chroma_width, frame->chroma_height), GL_R8, frame->chroma_width, frame->chroma_height);
406 glTextureSubImage2D(tex[0], 0, 0, 0, frame->width, frame->height, GL_RED, GL_UNSIGNED_BYTE, frame->data.get());
407 glGenerateTextureMipmap(tex[0]);
409 glTextureSubImage2D(tex[1], 0, 0, 0, frame->chroma_width, frame->chroma_height, GL_RED, GL_UNSIGNED_BYTE, frame->data.get() + frame->width * frame->height);
410 glGenerateTextureMipmap(tex[1]);
412 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);
413 glGenerateTextureMipmap(tex[2]);
415 glBindTextureUnit(0, tex[0]);
416 glBindTextureUnit(1, tex[1]);
417 glBindTextureUnit(2, tex[2]);
418 glBindSampler(0, bilinear_sampler);
419 glBindSampler(1, bilinear_sampler);
420 glBindSampler(2, bilinear_sampler);
421 glProgramUniform1i(ycbcr_program, 0, 0);
422 glProgramUniform1i(ycbcr_program, 1, 1);
423 glProgramUniform1i(ycbcr_program, 2, 2);
424 glProgramUniform2f(ycbcr_program, 3, cbcr_offset[0], -cbcr_offset[1]);
431 double video_aspect = double(frame->width) / frame->height;
432 if (display_aspect > video_aspect) {
433 double extra_width = frame->height * display_aspect - frame->width;
434 tx1 = -0.5 * extra_width / frame->width;
435 tx2 = 1.0 + 0.5 * extra_width / frame->width;
436 } else if (display_aspect < video_aspect) {
437 double extra_height = frame->width / display_aspect - frame->height;
438 ty1 = -0.5 * extra_height / frame->height;
439 ty2 = 1.0 + 0.5 * extra_height / frame->height;
445 glVertexAttrib2f(1, tx1, ty1);
446 glVertex2f(zoom_matrix[2 * 3 + 0], zoom_matrix[2 * 3 + 1]);
449 glVertexAttrib2f(1, tx1, ty2);
450 glVertex2f(zoom_matrix[1 * 3 + 0] + zoom_matrix[2 * 3 + 0], zoom_matrix[1 * 3 + 1] + zoom_matrix[2 * 3 + 1]);
453 glVertexAttrib2f(1, tx2, ty2);
454 glVertex2f(zoom_matrix[0 * 3 + 0] + zoom_matrix[1 * 3 + 0] + zoom_matrix[2 * 3 + 0],
455 zoom_matrix[1 * 3 + 0] + zoom_matrix[1 * 3 + 1] + zoom_matrix[2 * 3 + 1]);
458 glVertexAttrib2f(1, tx2, ty1);
459 glVertex2f(zoom_matrix[0 * 3 + 0] + zoom_matrix[2 * 3 + 0],
460 zoom_matrix[1 * 3 + 0] + zoom_matrix[2 * 3 + 1]);
465 void matmul3x3(const double a[9], const double b[9], double res[9])
467 for (int i = 0; i < 3; ++i) {
468 for (int j = 0; j < 3; ++j) {
470 for (int k = 0; k < 3; ++k) {
471 sum += a[i * 3 + k] * b[k * 3 + j];
473 res[i * 3 + j] = sum;
478 void VideoWidget::wheelEvent(QWheelEvent *event)
480 int delta = event->angleDelta().y();
484 double x = event->position().x() / width();
485 double y = 1.0 - event->position().y() / height();
486 double zoom = delta > 0 ? pow(1.005, delta) : pow(1/1.005, -delta);
488 const double inv_translation_matrix[9] = {
493 const double scale_matrix[9] = {
498 const double translation_matrix[9] = {
503 double tmp1[9], tmp2[9];
504 matmul3x3(zoom_matrix, inv_translation_matrix, tmp1);
505 matmul3x3(tmp1, scale_matrix, tmp2);
506 matmul3x3(tmp2, translation_matrix, zoom_matrix);
512 void VideoWidget::mousePressEvent(QMouseEvent *e)
514 if (e->button() == Qt::BackButton) {
515 emit mouse_back_clicked();
516 } else if (e->button() == Qt::ForwardButton) {
517 emit mouse_forward_clicked();
518 } else if (e->button() == Qt::LeftButton) {
520 last_drag_x = e->position().x();
521 last_drag_y = e->position().y();
525 void VideoWidget::mouseReleaseEvent(QMouseEvent *e)
527 if (e->button() == Qt::LeftButton) {
532 void VideoWidget::mouseMoveEvent(QMouseEvent *e)
537 float dx = (e->position().x() - last_drag_x) / width();
538 float dy = (e->position().y() - last_drag_y) / height();
540 //zoom_matrix[6] += dx * zoom_matrix[0];
541 //zoom_matrix[7] += dy * zoom_matrix[4];
542 zoom_matrix[6] += dx;
543 zoom_matrix[7] -= dy;
546 last_drag_x = e->position().x();
547 last_drag_y = e->position().y();
552 // Normalize the matrix so that we never get skew or similar,
553 // and also never can zoom or pan too far out.
554 void VideoWidget::fixup_zoom_matrix()
556 // Correct for any numerical errors (we know the matrix must be orthogonal
557 // and have zero rotation).
558 zoom_matrix[4] = zoom_matrix[0];
559 zoom_matrix[1] = zoom_matrix[2] = zoom_matrix[3] = zoom_matrix[5] = 0.0;
560 zoom_matrix[8] = 1.0;
562 // We can't zoom further out than 1:1. (Perhaps it would be nice to
563 // reuse the last zoom-in point to do this, but the center will have to do
565 if (zoom_matrix[0] < 1.0) {
566 const double zoom = 1.0 / zoom_matrix[0];
567 const double inv_translation_matrix[9] = {
572 const double scale_matrix[9] = {
577 const double translation_matrix[9] = {
582 double tmp1[9], tmp2[9];
583 matmul3x3(zoom_matrix, inv_translation_matrix, tmp1);
584 matmul3x3(tmp1, scale_matrix, tmp2);
585 matmul3x3(tmp2, translation_matrix, zoom_matrix);
588 // Looking at the points we'll draw with glVertex2f(), make sure none of them are
589 // inside the square (which would generally mean we've panned ourselves out-of-bounds).
590 // We simply adjust the translation, which is possible because we fixed scaling above.
591 zoom_matrix[6] = min(zoom_matrix[6], 0.0); // Left side (x=0).
592 zoom_matrix[7] = min(zoom_matrix[7], 0.0); // Bottom side (y=0).
593 zoom_matrix[6] = std::max(zoom_matrix[6], 1.0 - zoom_matrix[0]); // Right side (x=1).
594 zoom_matrix[7] = std::max(zoom_matrix[7], 1.0 - zoom_matrix[4]); // Top side (y=1).
597 bool VideoWidget::open(const string &filename)
604 while (running == STARTING) {
605 // Poor man's condition variable...
609 return (running != VIDEO_FILE_ERROR);
612 void VideoWidget::play()
614 if (running != NOT_RUNNING && running != VIDEO_FILE_ERROR) {
615 std::lock_guard<std::mutex> lock(queue_mu);
616 command_queue.push_back(QueuedCommand { QueuedCommand::RESUME });
617 producer_thread_should_quit.wakeup();
621 producer_thread_should_quit.unquit();
622 if (producer_thread.joinable()) {
623 producer_thread.join();
625 producer_thread = std::thread(&VideoWidget::producer_thread_func, this);
628 void VideoWidget::pause()
630 if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
633 std::lock_guard<std::mutex> lock(queue_mu);
634 command_queue.push_back(QueuedCommand { QueuedCommand::PAUSE });
635 producer_thread_should_quit.wakeup();
638 void VideoWidget::seek(int64_t relative_seek_ms)
640 if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
643 std::lock_guard<std::mutex> lock(queue_mu);
644 command_queue.push_back(QueuedCommand { QueuedCommand::SEEK, relative_seek_ms, 0, 0 });
645 producer_thread_should_quit.wakeup();
648 void VideoWidget::seek_frames(int64_t relative_seek_frames)
650 if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
653 std::lock_guard<std::mutex> lock(queue_mu);
654 command_queue.push_back(QueuedCommand { QueuedCommand::SEEK, 0, relative_seek_frames, 0 });
655 producer_thread_should_quit.wakeup();
658 void VideoWidget::seek_absolute(int64_t position_ms)
660 if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
663 std::lock_guard<std::mutex> lock(queue_mu);
664 command_queue.push_back(QueuedCommand { QueuedCommand::SEEK_ABSOLUTE, 0, 0, position_ms });
665 producer_thread_should_quit.wakeup();
668 void VideoWidget::stop()
670 if (running == NOT_RUNNING || running == VIDEO_FILE_ERROR) {
673 producer_thread_should_quit.quit();
674 producer_thread.join();
677 void VideoWidget::producer_thread_func()
679 if (!producer_thread_should_quit.should_quit()) {
680 if (!play_video(pathname)) {
681 running = VIDEO_FILE_ERROR;
683 running = NOT_RUNNING;
688 void VideoWidget::internal_rewind()
690 pts_origin = last_pts = 0;
692 start = next_frame_start = steady_clock::now();
695 template<AVHWDeviceType type>
696 AVPixelFormat get_hw_format(AVCodecContext *ctx, const AVPixelFormat *fmt)
698 bool found_config_of_right_type = false;
699 for (int i = 0;; ++i) { // Termination condition inside loop.
700 const AVCodecHWConfig *config = avcodec_get_hw_config(ctx->codec, i);
701 if (config == nullptr) { // End of list.
704 if (!(config->methods & AV_CODEC_HW_CONFIG_METHOD_HW_DEVICE_CTX) ||
705 config->device_type != type) {
706 // Not interesting for us.
710 // We have a config of the right type, but does it actually support
711 // the pixel format we want? (Seemingly, FFmpeg's way of signaling errors
712 // is to just replace the pixel format with a software-decoded one,
714 found_config_of_right_type = true;
715 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
716 if (config->pix_fmt == *fmt_ptr) {
717 fprintf(stderr, "Initialized '%s' hardware decoding for codec '%s'.\n",
718 av_hwdevice_get_type_name(type), ctx->codec->name);
719 if (ctx->profile == FF_PROFILE_H264_BASELINE) {
720 fprintf(stderr, "WARNING: Stream claims to be H.264 Baseline, which is generally poorly supported in hardware decoders.\n");
721 fprintf(stderr, " Consider encoding it as Constrained Baseline, Main or High instead.\n");
722 fprintf(stderr, " Decoding might fail and fall back to software.\n");
724 return config->pix_fmt;
727 fprintf(stderr, "Decoder '%s' supports only these pixel formats:", ctx->codec->name);
728 unordered_set<AVPixelFormat> seen;
729 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
730 if (!seen.count(*fmt_ptr)) {
731 fprintf(stderr, " %s", av_get_pix_fmt_name(*fmt_ptr));
732 seen.insert(*fmt_ptr);
735 fprintf(stderr, " (wanted %s for hardware acceleration)\n", av_get_pix_fmt_name(config->pix_fmt));
739 if (!found_config_of_right_type) {
740 fprintf(stderr, "Decoder '%s' does not support device type '%s'.\n", ctx->codec->name, av_hwdevice_get_type_name(type));
743 // We found no VA-API formats, so take the first software format.
744 for (const AVPixelFormat *fmt_ptr = fmt; *fmt_ptr != -1; ++fmt_ptr) {
745 if ((av_pix_fmt_desc_get(*fmt_ptr)->flags & AV_PIX_FMT_FLAG_HWACCEL) == 0) {
746 fprintf(stderr, "Falling back to software format %s.\n", av_get_pix_fmt_name(*fmt_ptr));
751 // Fallback: Just return anything. (Should never really happen.)
755 AVFrameWithDeleter VideoWidget::decode_frame(AVFormatContext *format_ctx, AVCodecContext *video_codec_ctx,
756 const std::string &pathname, int video_stream_index,
761 if (!queued_frames.empty()) {
762 AVFrameWithDeleter frame = std::move(queued_frames.front());
763 queued_frames.pop_front();
767 // Read packets until we have a frame or there are none left.
768 bool frame_finished = false;
769 AVFrameWithDeleter video_avframe = av_frame_alloc_unique();
772 AVPacket *pkt = av_packet_alloc();
773 unique_ptr<AVPacket, decltype(av_packet_unref)*> pkt_cleanup(
774 pkt, av_packet_unref);
777 if (av_read_frame(format_ctx, pkt) == 0) {
778 if (pkt->stream_index == video_stream_index) {
779 if (avcodec_send_packet(video_codec_ctx, pkt) < 0) {
780 fprintf(stderr, "%s: Cannot send packet to video codec.\n", pathname.c_str());
782 return AVFrameWithDeleter(nullptr);
786 eof = true; // Or error, but ignore that for the time being.
789 // Decode video, if we have a frame.
790 int err = avcodec_receive_frame(video_codec_ctx, video_avframe.get());
792 frame_finished = true;
794 } else if (err != AVERROR(EAGAIN)) {
795 fprintf(stderr, "%s: Cannot receive frame from video codec.\n", pathname.c_str());
797 return AVFrameWithDeleter(nullptr);
802 return video_avframe;
804 return AVFrameWithDeleter(nullptr);
807 int find_stream_index(AVFormatContext *ctx, AVMediaType media_type)
809 for (unsigned i = 0; i < ctx->nb_streams; ++i) {
810 if (ctx->streams[i]->codecpar->codec_type == media_type) {
817 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)
819 const duration<double> pts((frame_pts - pts_origin) * double(video_timebase.num) / double(video_timebase.den));
820 return origin + duration_cast<steady_clock::duration>(pts / rate);
823 bool VideoWidget::play_video(const string &pathname)
825 queued_frames.clear();
826 AVFormatContextWithCloser format_ctx = avformat_open_input_unique(pathname.c_str(), /*fmt=*/nullptr,
827 /*options=*/nullptr);
828 if (format_ctx == nullptr) {
829 fprintf(stderr, "%s: Error opening file\n", pathname.c_str());
833 if (avformat_find_stream_info(format_ctx.get(), nullptr) < 0) {
834 fprintf(stderr, "%s: Error finding stream info\n", pathname.c_str());
838 int video_stream_index = find_stream_index(format_ctx.get(), AVMEDIA_TYPE_VIDEO);
839 if (video_stream_index == -1) {
840 fprintf(stderr, "%s: No video stream found\n", pathname.c_str());
844 // Open video decoder.
845 const AVCodecParameters *video_codecpar = format_ctx->streams[video_stream_index]->codecpar;
846 const AVCodec *video_codec = avcodec_find_decoder(video_codecpar->codec_id);
848 video_timebase = format_ctx->streams[video_stream_index]->time_base;
849 AVCodecContextWithDeleter video_codec_ctx = avcodec_alloc_context3_unique(nullptr);
850 if (avcodec_parameters_to_context(video_codec_ctx.get(), video_codecpar) < 0) {
851 fprintf(stderr, "%s: Cannot fill video codec parameters\n", pathname.c_str());
854 if (video_codec == nullptr) {
855 fprintf(stderr, "%s: Cannot find video decoder\n", pathname.c_str());
859 // Seemingly, it's not too easy to make something that just initializes
860 // “whatever goes”, so we don't get CUDA or VULKAN or whatever here
861 // without enumerating through several different types.
862 // VA-API and VDPAU will do for now. We prioritize VDPAU for the
863 // simple reason that there's a VA-API-via-VDPAU emulation for NVidia
864 // cards that seems to work, but just hangs when trying to transfer the frame.
866 // Note that we don't actually check codec support beforehand,
867 // so if you have a low-end VDPAU device but a high-end VA-API device,
868 // you lose out on the extra codec support from the latter.
869 AVBufferRef *hw_device_ctx = nullptr;
870 if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VDPAU, nullptr, nullptr, 0) >= 0) {
871 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
872 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VDPAU>;
873 } else if (av_hwdevice_ctx_create(&hw_device_ctx, AV_HWDEVICE_TYPE_VAAPI, nullptr, nullptr, 0) >= 0) {
874 video_codec_ctx->hw_device_ctx = av_buffer_ref(hw_device_ctx);
875 video_codec_ctx->get_format = get_hw_format<AV_HWDEVICE_TYPE_VAAPI>;
877 fprintf(stderr, "Failed to initialize VA-API or VDPAU for FFmpeg acceleration. Decoding video in software.\n");
880 if (avcodec_open2(video_codec_ctx.get(), video_codec, nullptr) < 0) {
881 fprintf(stderr, "%s: Cannot open video decoder\n", pathname.c_str());
884 unique_ptr<AVCodecContext, decltype(avcodec_close)*> video_codec_ctx_cleanup(
885 video_codec_ctx.get(), avcodec_close);
892 int consecutive_errors = 0;
894 while (!producer_thread_should_quit.should_quit()) {
895 if (process_queued_commands(format_ctx.get(), video_codec_ctx.get(), video_stream_index, /*seeked=*/nullptr)) {
899 producer_thread_should_quit.sleep_for(hours(1));
904 AVFrameWithDeleter frame = decode_frame(format_ctx.get(), video_codec_ctx.get(),
905 pathname, video_stream_index, &error);
907 if (++consecutive_errors >= 100) {
908 fprintf(stderr, "More than 100 consecutive video frames, aborting playback.\n");
914 consecutive_errors = 0;
916 if (frame == nullptr) {
921 // Sleep until it's time to present this frame.
923 if (last_pts == 0 && pts_origin == 0) {
924 pts_origin = frame->pts;
926 steady_clock::time_point now = steady_clock::now();
927 next_frame_start = compute_frame_start(frame->pts, pts_origin, video_timebase, start, rate);
929 if (duration<double>(now - next_frame_start).count() >= 0.1) {
930 // If we don't have enough CPU to keep up, or if we have a live stream
931 // where the initial origin was somehow wrong, we could be behind indefinitely.
932 fprintf(stderr, "%s: Playback %.0f ms behind, resetting time scale\n",
934 1e3 * duration<double>(now - next_frame_start).count());
935 pts_origin = frame->pts;
936 start = next_frame_start = now;
938 bool finished_wakeup;
939 finished_wakeup = producer_thread_should_quit.sleep_until(next_frame_start);
940 if (finished_wakeup) {
941 Frame *new_frame = new Frame(make_video_frame(frame.get()));
943 lock_guard lock(current_frame_mu);
944 current_frame.reset(new_frame);
946 last_frame = steady_clock::now();
950 if (producer_thread_should_quit.should_quit()) break;
953 if (process_queued_commands(format_ctx.get(), video_codec_ctx.get(), video_stream_index, &seeked)) {
958 // Just paused, so present the frame immediately and then go into deep sleep.
959 Frame *new_frame = new Frame(make_video_frame(frame.get()));
961 lock_guard lock(current_frame_mu);
962 current_frame.reset(new_frame);
964 last_frame = steady_clock::now();
969 // If we just seeked, drop this frame on the floor and be done.
975 store_pts(frame->pts);
980 void VideoWidget::store_pts(int64_t pts)
983 last_position = lrint(pts * double(video_timebase.num) / double(video_timebase.den) * 1000);
984 post_to_main_thread([this, last_position{last_position.load()}] {
985 emit position_changed(last_position);
989 // Taken from Movit (see the comment there for explanation)
990 float compute_chroma_offset(float pos, unsigned subsampling_factor, unsigned resolution)
992 float local_chroma_pos = (0.5 + pos * (subsampling_factor - 1)) / subsampling_factor;
993 if (fabs(local_chroma_pos - 0.5) < 1e-10) {
994 // x + (-0) can be optimized away freely, as opposed to x + 0.
997 return (0.5 - local_chroma_pos) / resolution;
1001 VideoWidget::Frame VideoWidget::make_video_frame(const AVFrame *frame)
1004 AVFrameWithDeleter sw_frame;
1006 if (frame->format == AV_PIX_FMT_VAAPI ||
1007 frame->format == AV_PIX_FMT_VDPAU) {
1008 // Get the frame down to the CPU. (TODO: See if we can keep it
1009 // on the GPU all the way, since it will be going up again later.
1010 // However, this only works if the OpenGL GPU is the same one.)
1011 sw_frame = av_frame_alloc_unique();
1012 int err = av_hwframe_transfer_data(sw_frame.get(), frame, 0);
1014 fprintf(stderr, "%s: Cannot transfer hardware video frame to software.\n", pathname.c_str());
1016 sw_frame->pts = frame->pts;
1017 sw_frame->pkt_duration = frame->pkt_duration;
1018 frame = sw_frame.get();
1022 if (sws_ctx == nullptr ||
1023 sws_last_width != frame->width ||
1024 sws_last_height != frame->height ||
1025 sws_last_src_format != frame->format) {
1026 sws_dst_format = decide_dst_format(AVPixelFormat(frame->format));
1028 sws_getContext(frame->width, frame->height, AVPixelFormat(frame->format),
1029 frame->width, frame->height, sws_dst_format,
1030 SWS_BICUBIC, nullptr, nullptr, nullptr));
1031 sws_last_width = frame->width;
1032 sws_last_height = frame->height;
1033 sws_last_src_format = frame->format;
1035 if (sws_ctx == nullptr) {
1036 fprintf(stderr, "Could not create scaler context\n");
1040 uint8_t *pic_data[4] = { nullptr, nullptr, nullptr, nullptr };
1041 int linesizes[4] = { 0, 0, 0, 0 };
1042 const AVPixFmtDescriptor *desc = av_pix_fmt_desc_get(sws_dst_format);
1044 video_frame.width = frame->width;
1045 video_frame.height = frame->height;
1046 video_frame.chroma_width = AV_CEIL_RSHIFT(int(frame->width), desc->log2_chroma_w);
1047 video_frame.chroma_height = AV_CEIL_RSHIFT(int(frame->height), desc->log2_chroma_h);
1049 // We always assume left chroma placement for now.
1050 cbcr_offset[0] = compute_chroma_offset(0.0f, 1 << desc->log2_chroma_w, video_frame.chroma_width);
1051 cbcr_offset[1] = compute_chroma_offset(0.5f, 1 << desc->log2_chroma_h, video_frame.chroma_height);
1053 size_t len = frame->width * frame->height + 2 * video_frame.chroma_width * video_frame.chroma_height;
1054 video_frame.data.reset(new uint8_t[len]);
1056 pic_data[0] = video_frame.data.get();
1057 linesizes[0] = frame->width;
1059 pic_data[1] = pic_data[0] + frame->width * frame->height;
1060 linesizes[1] = video_frame.chroma_width;
1062 pic_data[2] = pic_data[1] + video_frame.chroma_width * video_frame.chroma_height;
1063 linesizes[2] = video_frame.chroma_width;
1065 sws_scale(sws_ctx.get(), frame->data, frame->linesize, 0, frame->height, pic_data, linesizes);