#include <assert.h>
#include <epoxy/egl.h>
#include <libavcodec/avcodec.h>
-#include <libavformat/avio.h>
-#include <libavutil/mathematics.h>
+#include <libavutil/channel_layout.h>
+#include <libavutil/frame.h>
#include <libavutil/rational.h>
+#include <libavutil/samplefmt.h>
#include <libdrm/drm_fourcc.h>
-#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <va/va_enc_h264.h>
#include <va/va_x11.h>
#include <condition_variable>
+#include <cstdint>
+#include <memory>
#include <mutex>
#include <queue>
#include <string>
#include <thread>
+#include <utility>
#include "context.h"
+#include "defs.h"
+#include "httpd.h"
#include "timebase.h"
class QOpenGLContext;
static int frame_height = 144;
static int frame_width_mbaligned;
static int frame_height_mbaligned;
-static int frame_rate = 60;
static unsigned int frame_bitrate = 0;
static unsigned int frame_slices = 1;
static double frame_size = 0;
//static int initial_qp = 28;
static int minimal_qp = 0;
static int intra_period = 30;
-static int intra_idr_period = 60;
+static int intra_idr_period = MAX_FPS; // About a second; more at lower frame rates. Not ideal.
static int ip_period = 3;
static int rc_mode = -1;
static int rc_default_modes[] = {
using namespace std;
+// Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly,
+// but if we don't delete it here, we get leaks. The GStreamer implementation
+// does the same.
+static void render_picture_and_delete(VADisplay dpy, VAContextID context, VABufferID *buffers, int num_buffers)
+{
+ VAStatus va_status = vaRenderPicture(dpy, context, buffers, num_buffers);
+ CHECK_VASTATUS(va_status, "vaRenderPicture");
+
+ for (int i = 0; i < num_buffers; ++i) {
+ va_status = vaDestroyBuffer(va_dpy, buffers[i]);
+ CHECK_VASTATUS(va_status, "vaDestroyBuffer");
+ }
+}
+
static unsigned int
va_swap32(unsigned int val)
{
bitstream_put_ui(bs, 1, 1); /* vui_parameters_present_flag */
bitstream_put_ui(bs, 0, 1); /* aspect_ratio_info_present_flag */
bitstream_put_ui(bs, 0, 1); /* overscan_info_present_flag */
- bitstream_put_ui(bs, 0, 1); /* video_signal_type_present_flag */
+ bitstream_put_ui(bs, 1, 1); /* video_signal_type_present_flag */
+ {
+ bitstream_put_ui(bs, 5, 3); /* video_format (5 = Unspecified) */
+ bitstream_put_ui(bs, 0, 1); /* video_full_range_flag */
+ bitstream_put_ui(bs, 1, 1); /* colour_description_present_flag */
+ {
+ bitstream_put_ui(bs, 1, 8); /* colour_primaries (1 = BT.709) */
+ bitstream_put_ui(bs, 2, 8); /* transfer_characteristics (2 = unspecified, since we use sRGB) */
+ bitstream_put_ui(bs, 6, 8); /* matrix_coefficients (6 = BT.601/SMPTE 170M) */
+ }
+ }
bitstream_put_ui(bs, 0, 1); /* chroma_loc_info_present_flag */
bitstream_put_ui(bs, 1, 1); /* timing_info_present_flag */
{
{IDR(PBB)(PBB)}.
*/
-/*
- * Return displaying order with specified periods and encoding order
- * displaying_order: displaying order
- * frame_type: frame type
- */
+// General pts/dts strategy:
+//
+// Getting pts and dts right with variable frame rate (VFR) and B-frames can be a
+// bit tricky. We assume first of all that the frame rate never goes _above_
+// MAX_FPS, which gives us a frame period N. The decoder can always decode
+// in at least this speed, as long at dts <= pts (the frame is not attempted
+// presented before it is decoded). Furthermore, we never have longer chains of
+// B-frames than a fixed constant C. (In a B-frame chain, we say that the base
+// I/P-frame has order O=0, the B-frame depending on it directly has order O=1,
+// etc. The last frame in the chain, which no B-frames depend on, is the “tip”
+// frame, with an order O <= C.)
+//
+// Many strategies are possible, but we establish these rules:
+//
+// - Tip frames have dts = pts - (C-O)*N.
+// - Non-tip frames have dts = dts_last + N.
+//
+// An example, with C=2 and N=10 and the data flow showed with arrows:
+//
+// I B P B B P
+// pts: 30 40 50 60 70 80
+// ↓ ↓ ↓
+// dts: 10 30 20 60 50←40
+// | | ↑ ↑
+// `--|--' |
+// `----------'
+//
+// To show that this works fine also with irregular spacings, let's say that
+// the third frame is delayed a bit (something earlier was dropped). Now the
+// situation looks like this:
+//
+// I B P B B P
+// pts: 30 40 80 90 100 110
+// ↓ ↓ ↓
+// dts: 10 30 20 90 50←40
+// | | ↑ ↑
+// `--|--' |
+// `----------'
+//
+// The resetting on every tip frame makes sure dts never ends up lagging a lot
+// behind pts, and the subtraction of (C-O)*N makes sure pts <= dts.
+//
+// In the output of this function, if <dts_lag> is >= 0, it means to reset the
+// dts from the current pts minus <dts_lag>, while if it's -1, the frame is not
+// a tip frame and should be given a dts based on the previous one.
#define FRAME_P 0
#define FRAME_B 1
#define FRAME_I 2
unsigned long long encoding_order, int intra_period,
int intra_idr_period, int ip_period,
unsigned long long *displaying_order,
- int *frame_type)
+ int *frame_type, int *pts_lag)
{
int encoding_order_gop = 0;
+ *pts_lag = 0;
+
if (intra_period == 1) { /* all are I/IDR frames */
*displaying_order = encoding_order;
if (intra_idr_period == 0)
if (intra_period == 0)
intra_idr_period = 0;
- /* new sequence like
- * IDR PPPPP IPPPPP
- * IDR (PBB)(PBB)(IBB)(PBB)
- */
- encoding_order_gop = (intra_idr_period == 0)? encoding_order:
- (encoding_order % (intra_idr_period + ((ip_period == 1)?0:1)));
+ if (ip_period == 1) {
+ // No B-frames, sequence is like IDR PPPPP IPPPPP.
+ encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % intra_idr_period);
+ *displaying_order = encoding_order;
+
+ if (encoding_order_gop == 0) { /* the first frame */
+ *frame_type = FRAME_IDR;
+ } else if (intra_period != 0 && /* have I frames */
+ encoding_order_gop >= 2 &&
+ (encoding_order_gop % intra_period == 0)) {
+ *frame_type = FRAME_I;
+ } else {
+ *frame_type = FRAME_P;
+ }
+ return;
+ }
+
+ // We have B-frames. Sequence is like IDR (PBB)(PBB)(IBB)(PBB).
+ encoding_order_gop = (intra_idr_period == 0) ? encoding_order : (encoding_order % (intra_idr_period + 1));
+ *pts_lag = -1; // Most frames are not tip frames.
if (encoding_order_gop == 0) { /* the first frame */
*frame_type = FRAME_IDR;
*displaying_order = encoding_order;
+ // IDR frames are a special case; I honestly can't find the logic behind
+ // why this is the right thing, but it seems to line up nicely in practice :-)
+ *pts_lag = TIMEBASE / MAX_FPS;
} else if (((encoding_order_gop - 1) % ip_period) != 0) { /* B frames */
- *frame_type = FRAME_B;
+ *frame_type = FRAME_B;
*displaying_order = encoding_order - 1;
- } else if ((intra_period != 0) && /* have I frames */
- (encoding_order_gop >= 2) &&
- ((ip_period == 1 && encoding_order_gop % intra_period == 0) || /* for IDR PPPPP IPPPP */
- /* for IDR (PBB)(PBB)(IBB) */
- (ip_period >= 2 && ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0))) {
- *frame_type = FRAME_I;
- *displaying_order = encoding_order + ip_period - 1;
+ if ((encoding_order_gop % ip_period) == 0) {
+ *pts_lag = 0; // Last B-frame.
+ }
+ } else if (intra_period != 0 && /* have I frames */
+ encoding_order_gop >= 2 &&
+ ((encoding_order_gop - 1) / ip_period % (intra_period / ip_period)) == 0) {
+ *frame_type = FRAME_I;
+ *displaying_order = encoding_order + ip_period - 1;
} else {
- *frame_type = FRAME_P;
- *displaying_order = encoding_order + ip_period - 1;
+ *frame_type = FRAME_P;
+ *displaying_order = encoding_order + ip_period - 1;
}
}
}
if (frame_bitrate == 0)
- frame_bitrate = frame_width * frame_height * 12 * frame_rate / 50;
+ frame_bitrate = frame_width * frame_height * 12 * MAX_FPS / 50;
if (coded_fn == NULL) {
struct stat buf;
render_id[0] = seq_param_buf;
render_id[1] = rc_param_buf;
- va_status = vaRenderPicture(va_dpy, context_id, &render_id[0], 2);
- CHECK_VASTATUS(va_status, "vaRenderPicture");;
+ render_picture_and_delete(va_dpy, context_id, &render_id[0], 2);
if (misc_priv_type != 0) {
va_status = vaCreateBuffer(va_dpy, context_id,
misc_param_tmp->data[0] = misc_priv_value;
vaUnmapBuffer(va_dpy, misc_param_tmpbuf);
- va_status = vaRenderPicture(va_dpy, context_id, &misc_param_tmpbuf, 1);
+ render_picture_and_delete(va_dpy, context_id, &misc_param_tmpbuf, 1);
}
return 0;
va_status = vaCreateBuffer(va_dpy, context_id, VAEncPictureParameterBufferType,
sizeof(pic_param), 1, &pic_param, &pic_param_buf);
- CHECK_VASTATUS(va_status, "vaCreateBuffer");;
+ CHECK_VASTATUS(va_status, "vaCreateBuffer");
- va_status = vaRenderPicture(va_dpy, context_id, &pic_param_buf, 1);
- CHECK_VASTATUS(va_status, "vaRenderPicture");
+ render_picture_and_delete(va_dpy, context_id, &pic_param_buf, 1);
return 0;
}
render_id[0] = packedseq_para_bufid;
render_id[1] = packedseq_data_bufid;
- va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
- CHECK_VASTATUS(va_status, "vaRenderPicture");
+ render_picture_and_delete(va_dpy, context_id, render_id, 2);
free(packedseq_buffer);
render_id[0] = packedpic_para_bufid;
render_id[1] = packedpic_data_bufid;
- va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
- CHECK_VASTATUS(va_status, "vaRenderPicture");
+ render_picture_and_delete(va_dpy, context_id, render_id, 2);
free(packedpic_buffer);
render_id[0] = packedslice_para_bufid;
render_id[1] = packedslice_data_bufid;
- va_status = vaRenderPicture(va_dpy, context_id, render_id, 2);
- CHECK_VASTATUS(va_status, "vaRenderPicture");
+ render_picture_and_delete(va_dpy, context_id, render_id, 2);
free(packedslice_buffer);
}
va_status = vaCreateBuffer(va_dpy, context_id, VAEncSliceParameterBufferType,
sizeof(slice_param), 1, &slice_param, &slice_param_buf);
- CHECK_VASTATUS(va_status, "vaCreateBuffer");;
+ CHECK_VASTATUS(va_status, "vaCreateBuffer");
+
+ render_picture_and_delete(va_dpy, context_id, &slice_param_buf, 1);
- va_status = vaRenderPicture(va_dpy, context_id, &slice_param_buf, 1);
- CHECK_VASTATUS(va_status, "vaRenderPicture");
-
return 0;
}
-int H264Encoder::save_codeddata(storage_task task)
+void H264Encoder::save_codeddata(storage_task task)
{
VACodedBufferSegment *buf_list = NULL;
VAStatus va_status;
string data;
+ const int64_t global_delay = (ip_period - 1) * (TIMEBASE / MAX_FPS); // So we never get negative dts.
+
va_status = vaMapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf, (void **)(&buf_list));
CHECK_VASTATUS(va_status, "vaMapBuffer");
while (buf_list != NULL) {
}
vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf);
- const int64_t pts_dts_delay = (ip_period - 1) * (TIMEBASE / frame_rate);
- const int64_t av_delay = TIMEBASE / 30; // Corresponds to the fixed delay in resampler.h. TODO: Make less hard-coded.
{
- int64_t pts, dts;
- {
- unique_lock<mutex> lock(frame_queue_mutex);
- assert(timestamps.count(task.display_order));
- assert(timestamps.count(task.encode_order));
- pts = timestamps[task.display_order];
- dts = timestamps[task.encode_order];
- }
// Add video.
AVPacket pkt;
memset(&pkt, 0, sizeof(pkt));
pkt.buf = nullptr;
- pkt.pts = av_rescale_q(pts + av_delay + pts_dts_delay, AVRational{1, TIMEBASE}, avstream_video->time_base);
- pkt.dts = av_rescale_q(dts + av_delay, AVRational{1, TIMEBASE}, avstream_video->time_base);
pkt.data = reinterpret_cast<uint8_t *>(&data[0]);
pkt.size = data.size();
pkt.stream_index = 0;
pkt.flags = 0;
}
//pkt.duration = 1;
- av_interleaved_write_frame(avctx, &pkt);
+ httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
}
// Encode and add all audio frames up to and including the pts of this video frame.
- // (They can never be queued to us after the video frame they belong to, only before.)
for ( ;; ) {
- int display_order;
- int64_t pts;
+ int64_t audio_pts;
std::vector<float> audio;
{
unique_lock<mutex> lock(frame_queue_mutex);
- if (pending_audio_frames.empty()) break;
+ frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || !pending_audio_frames.empty(); });
+ if (copy_thread_should_quit) return;
auto it = pending_audio_frames.begin();
- if (it->first > int(task.display_order)) break;
- display_order = it->first;
+ if (it->first > task.pts) break;
+ audio_pts = it->first;
audio = move(it->second);
pending_audio_frames.erase(it);
-
- auto pts_it = timestamps.find(display_order);
- assert(pts_it != timestamps.end());
- pts = pts_it->second;
}
+
AVFrame *frame = avcodec_alloc_frame();
frame->nb_samples = audio.size() / 2;
- frame->format = AV_SAMPLE_FMT_FLT;
+ frame->format = AV_SAMPLE_FMT_S32;
frame->channel_layout = AV_CH_LAYOUT_STEREO;
- unique_ptr<float[]> planar_samples(new float[audio.size()]);
- avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_FLTP, (const uint8_t*)planar_samples.get(), audio.size() * sizeof(float), 0);
- for (int i = 0; i < frame->nb_samples; ++i) {
- planar_samples[i] = audio[i * 2 + 0];
- planar_samples[i + frame->nb_samples] = audio[i * 2 + 1];
+ unique_ptr<int32_t[]> int_samples(new int32_t[audio.size()]);
+ int ret = avcodec_fill_audio_frame(frame, 2, AV_SAMPLE_FMT_S32, (const uint8_t*)int_samples.get(), audio.size() * sizeof(int32_t), 1);
+ if (ret < 0) {
+ fprintf(stderr, "avcodec_fill_audio_frame() failed with %d\n", ret);
+ exit(1);
+ }
+ for (int i = 0; i < frame->nb_samples * 2; ++i) {
+ if (audio[i] >= 1.0f) {
+ int_samples[i] = 2147483647;
+ } else if (audio[i] <= -1.0f) {
+ int_samples[i] = -2147483647;
+ } else {
+ int_samples[i] = lrintf(audio[i] * 2147483647.0f);
+ }
}
AVPacket pkt;
pkt.data = nullptr;
pkt.size = 0;
int got_output;
- avcodec_encode_audio2(avstream_audio->codec, &pkt, frame, &got_output);
+ avcodec_encode_audio2(context_audio, &pkt, frame, &got_output);
if (got_output) {
- pkt.pts = av_rescale_q(pts + pts_dts_delay, AVRational{1, TIMEBASE}, avstream_audio->time_base);
- pkt.dts = pkt.pts;
pkt.stream_index = 1;
- av_interleaved_write_frame(avctx, &pkt);
+ httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
}
// TODO: Delayed frames.
avcodec_free_frame(&frame);
- }
- {
- unique_lock<mutex> lock(frame_queue_mutex);
- timestamps.erase(task.encode_order - (ip_period - 1));
+ av_free_packet(&pkt);
+ if (audio_pts == task.pts) break;
}
#if 0
printf("%08lld", encode_order);
printf("(%06d bytes coded)", coded_size);
#endif
-
- return 0;
}
if (rc_mode != -1)
printf("INPUT: RateControl : %s\n", rc_to_string(rc_mode));
printf("INPUT: Resolution : %dx%dframes\n", frame_width, frame_height);
- printf("INPUT: FrameRate : %d\n", frame_rate);
printf("INPUT: Bitrate : %d\n", frame_bitrate);
printf("INPUT: Slieces : %d\n", frame_slices);
printf("INPUT: IntraPeriod : %d\n", intra_period);
return 0;
}
-
-//H264Encoder::H264Encoder(SDL_Window *window, SDL_GLContext context, int width, int height, const char *output_filename)
-H264Encoder::H264Encoder(QSurface *surface, int width, int height, const char *output_filename)
- : current_storage_frame(0), surface(surface)
- //: width(width), height(height), current_encoding_frame(0)
+H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
+ : current_storage_frame(0), surface(surface), httpd(httpd)
{
- av_register_all();
- avctx = avformat_alloc_context();
- avctx->oformat = av_guess_format(NULL, output_filename, NULL);
- strcpy(avctx->filename, output_filename);
- if (avio_open2(&avctx->pb, output_filename, AVIO_FLAG_WRITE, &avctx->interrupt_callback, NULL) < 0) {
- fprintf(stderr, "%s: avio_open2() failed\n", output_filename);
- exit(1);
- }
- AVCodec *codec_video = avcodec_find_encoder(AV_CODEC_ID_H264);
- avstream_video = avformat_new_stream(avctx, codec_video);
- if (avstream_video == nullptr) {
- fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
- exit(1);
- }
- avstream_video->time_base = AVRational{1, TIMEBASE};
- avstream_video->codec->width = width;
- avstream_video->codec->height = height;
- avstream_video->codec->time_base = AVRational{1, TIMEBASE};
- avstream_video->codec->ticks_per_frame = 1; // or 2?
-
- AVCodec *codec_audio = avcodec_find_encoder(AV_CODEC_ID_MP3);
- avstream_audio = avformat_new_stream(avctx, codec_audio);
- if (avstream_audio == nullptr) {
- fprintf(stderr, "%s: avformat_new_stream() failed\n", output_filename);
- exit(1);
- }
- avstream_audio->time_base = AVRational{1, TIMEBASE};
- avstream_audio->codec->bit_rate = 256000;
- avstream_audio->codec->sample_rate = 48000;
- avstream_audio->codec->sample_fmt = AV_SAMPLE_FMT_FLTP;
- avstream_audio->codec->channels = 2;
- avstream_audio->codec->channel_layout = AV_CH_LAYOUT_STEREO;
- avstream_audio->codec->time_base = AVRational{1, TIMEBASE};
-
- /* open it */
- if (avcodec_open2(avstream_audio->codec, codec_audio, NULL) < 0) {
+ AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC);
+ context_audio = avcodec_alloc_context3(codec_audio);
+ context_audio->bit_rate = AUDIO_OUTPUT_BIT_RATE;
+ context_audio->sample_rate = OUTPUT_FREQUENCY;
+ context_audio->sample_fmt = AUDIO_OUTPUT_SAMPLE_FMT;
+ context_audio->channels = 2;
+ context_audio->channel_layout = AV_CH_LAYOUT_STEREO;
+ context_audio->time_base = AVRational{1, TIMEBASE};
+ if (avcodec_open2(context_audio, codec_audio, NULL) < 0) {
fprintf(stderr, "Could not open codec\n");
exit(1);
}
- if (avformat_write_header(avctx, NULL) < 0) {
- fprintf(stderr, "%s: avformat_write_header() failed\n", output_filename);
- exit(1);
- }
-
frame_width = width;
frame_height = height;
frame_width_mbaligned = (frame_width + 15) & (~15);
copy_thread = std::thread([this]{
//SDL_GL_MakeCurrent(window, context);
- QOpenGLContext *context = create_context();
+ QOpenGLContext *context = create_context(this->surface);
eglBindAPI(EGL_OPENGL_API);
if (!make_current(context, this->surface)) {
printf("display=%p surface=%p context=%p curr=%p err=%d\n", eglGetCurrentDisplay(), this->surface, context, eglGetCurrentContext(),
{
unique_lock<mutex> lock(frame_queue_mutex);
copy_thread_should_quit = true;
- frame_queue_nonempty.notify_one();
+ frame_queue_nonempty.notify_all();
}
storage_thread.join();
copy_thread.join();
release_encode();
deinit_va();
-
- av_write_trailer(avctx);
- avformat_free_context(avctx);
}
bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
return true;
}
-void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, std::vector<float> audio, const std::vector<RefCountedFrame> &input_frames)
+void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
+{
+ {
+ unique_lock<mutex> lock(frame_queue_mutex);
+ pending_audio_frames[pts] = move(audio);
+ }
+ frame_queue_nonempty.notify_all();
+}
+
+void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
{
{
unique_lock<mutex> lock(frame_queue_mutex);
- pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames };
- pending_audio_frames[current_storage_frame] = move(audio);
- timestamps[current_storage_frame] = pts;
+ pending_video_frames[current_storage_frame] = PendingFrame{ fence, input_frames, pts };
++current_storage_frame;
}
- frame_queue_nonempty.notify_one();
+ frame_queue_nonempty.notify_all();
}
void H264Encoder::copy_thread_func()
{
+ int64_t last_dts = -1;
for ( ;; ) {
PendingFrame frame;
+ int pts_lag;
encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
- ¤t_frame_display, ¤t_frame_type);
+ ¤t_frame_display, ¤t_frame_type, &pts_lag);
if (current_frame_type == FRAME_IDR) {
numShortTerm = 0;
current_frame_num = 0;
va_status = vaEndPicture(va_dpy, context_id);
CHECK_VASTATUS(va_status, "vaEndPicture");
+ // Determine the pts and dts of this frame.
+ int64_t pts = frame.pts;
+ int64_t dts;
+ if (pts_lag == -1) {
+ assert(last_dts != -1);
+ dts = last_dts + (TIMEBASE / MAX_FPS);
+ } else {
+ dts = pts - pts_lag;
+ }
+ last_dts = dts;
+
// so now the data is done encoding (well, async job kicked off)...
// we send that to the storage thread
storage_task tmp;
tmp.display_order = current_frame_display;
- tmp.encode_order = current_frame_encoding;
tmp.frame_type = current_frame_type;
+ tmp.pts = pts;
+ tmp.dts = dts;
storage_task_enqueue(move(tmp));
update_ReferenceFrames();