Reduce latency by 1 frame

To packetize the H.264 raw stream, av_parser_parse2() (called by
av_read_frame()) knows that it has received a full frame only after it
has received some data for the next frame. As a consequence, the client
always waited until the next frame before sending the current frame to
the decoder!

On the device side, we know packets boundaries. To reduce latency,
make the device always transmit the "frame meta" to packetize the stream
manually (it was already implemented to send PTS, but only enabled on
recording).

On the client side, replace av_read_frame() by manual packetizing and
parsing.

<https://stackoverflow.com/questions/50682518/replacing-av-read-frame-to-reduce-delay>
<https://trac.ffmpeg.org/ticket/3354>

app/src/recorder.c

@@ -166,11 +166,21 @@ recorder_rescale_packet(struct recorder *recorder, AVPacket *packet) {
 bool
 recorder_write(struct recorder *recorder, AVPacket *packet) {
     if (!recorder->header_written) {
+        if (packet->pts != AV_NOPTS_VALUE) {
+            LOGE("The first packet is not a config packet");
+            return false;
+        }
         bool ok = recorder_write_header(recorder, packet);
         if (!ok) {
             return false;
         }
         recorder->header_written = true;
+        return true;
+    }
+
+    if (packet->pts == AV_NOPTS_VALUE) {
+        // ignore config packets
+        return true;
     }
 
     recorder_rescale_packet(recorder, packet);

app/src/scrcpy.c

@@ -277,7 +277,6 @@ scrcpy(const struct scrcpy_options *options) {
         .local_port = options->port,
         .max_size = options->max_size,
         .bit_rate = options->bit_rate,
-        .send_frame_meta = record,
         .control = options->control,
     };
     if (!server_start(&server, options->serial, &params)) {

app/src/server.c

@@ -130,7 +130,7 @@ execute_server(struct server *server, const struct server_params *params) {
         bit_rate_string,
         server->tunnel_forward ? "true" : "false",
         params->crop ? params->crop : "-",
-        params->send_frame_meta ? "true" : "false",
+        "true", // always send frame meta (packet boundaries + timestamp)
         params->control ? "true" : "false",
     };
     return adb_execute(server->serial, cmd, sizeof(cmd) / sizeof(cmd[0]));

app/src/server.h

@@ -34,7 +34,6 @@ struct server_params {
     uint16_t local_port;
     uint16_t max_size;
     uint32_t bit_rate;
-    bool send_frame_meta;
     bool control;
 };
 

app/src/stream.c

@@ -22,54 +22,8 @@
 #define HEADER_SIZE 12
 #define NO_PTS UINT64_C(-1)
 
-static struct frame_meta *
-frame_meta_new(uint64_t pts) {
-    struct frame_meta *meta = SDL_malloc(sizeof(*meta));
-    if (!meta) {
-        return meta;
-    }
-    meta->pts = pts;
-    meta->next = NULL;
-    return meta;
-}
-
-static void
-frame_meta_delete(struct frame_meta *frame_meta) {
-    SDL_free(frame_meta);
-}
-
 static bool
-receiver_state_push_meta(struct receiver_state *state, uint64_t pts) {
+stream_recv_packet(struct stream *stream, AVPacket *packet) {
-    struct frame_meta *frame_meta = frame_meta_new(pts);
-    if (!frame_meta) {
-        return false;
-    }
-
-    // append to the list
-    // (iterate to find the last item, in practice the list should be tiny)
-    struct frame_meta **p = &state->frame_meta_queue;
-    while (*p) {
-        p = &(*p)->next;
-    }
-    *p = frame_meta;
-    return true;
-}
-
-static uint64_t
-receiver_state_take_meta(struct receiver_state *state) {
-    struct frame_meta *frame_meta = state->frame_meta_queue; // first item
-    SDL_assert(frame_meta); // must not be empty
-    uint64_t pts = frame_meta->pts;
-    state->frame_meta_queue = frame_meta->next; // remove the item
-    frame_meta_delete(frame_meta);
-    return pts;
-}
-
-static int
-read_packet_with_meta(void *opaque, uint8_t *buf, int buf_size) {
-    struct stream *stream = opaque;
-    struct receiver_state *state = &stream->receiver_state;
-
     // The video stream contains raw packets, without time information. When we
     // record, we retrieve the timestamps separately, from a "meta" header
     // added by the server before each raw packet.
@@ -82,60 +36,30 @@ read_packet_with_meta(void *opaque, uint8_t *buf, int buf_size) {
     //
     // It is followed by <packet_size> bytes containing the packet/frame.
 
-    if (!state->remaining) {
-#define HEADER_SIZE 12
     uint8_t header[HEADER_SIZE];
     ssize_t r = net_recv_all(stream->socket, header, HEADER_SIZE);
-        if (r == -1) {
+    if (r < HEADER_SIZE) {
-            return AVERROR(errno);
+        return false;
     }
-        if (r == 0) {
-            return AVERROR_EOF;
-        }
-        // no partial read (net_recv_all())
-        SDL_assert_release(r == HEADER_SIZE);
 
     uint64_t pts = buffer_read64be(header);
-        state->remaining = buffer_read32be(&header[8]);
+    uint32_t len = buffer_read32be(&header[8]);
+    SDL_assert(len);
 
-        if (pts != NO_PTS && !receiver_state_push_meta(state, pts)) {
+    if (av_new_packet(packet, len)) {
-            LOGE("Could not store PTS for recording");
+        LOGE("Could not allocate packet");
-            // we could not save the PTS, the recording would be broken
+        return false;
-            return AVERROR(ENOMEM);
-        }
     }
 
-    SDL_assert(state->remaining);
+    r = net_recv_all(stream->socket, packet->data, len);
-
+    if (r < len) {
-    if (buf_size > state->remaining) {
+        av_packet_unref(packet);
-        buf_size = state->remaining;
+        return false;
     }
 
-    ssize_t r = net_recv(stream->socket, buf, buf_size);
+    packet->pts = pts != NO_PTS ? pts : AV_NOPTS_VALUE;
-    if (r == -1) {
-        return errno ? AVERROR(errno) : AVERROR_EOF;
-    }
-    if (r == 0) {
-        return AVERROR_EOF;
-    }
 
-    SDL_assert(state->remaining >= r);
+    return true;
-    state->remaining -= r;
-
-    return r;
-}
-
-static int
-read_raw_packet(void *opaque, uint8_t *buf, int buf_size) {
-    struct stream *stream = opaque;
-    ssize_t r = net_recv(stream->socket, buf, buf_size);
-    if (r == -1) {
-        return errno ? AVERROR(errno) : AVERROR_EOF;
-    }
-    if (r == 0) {
-        return AVERROR_EOF;
-    }
-    return r;
 }
 
 static void
@@ -145,55 +69,136 @@ notify_stopped(void) {
     SDL_PushEvent(&stop_event);
 }
 
+static bool
+process_config_packet(struct stream *stream, AVPacket *packet) {
+    if (stream->recorder && !recorder_write(stream->recorder, packet)) {
+        LOGE("Could not send config packet to recorder");
+        return false;
+    }
+    return true;
+}
+
+static bool
+process_frame(struct stream *stream, AVPacket *packet) {
+    if (stream->decoder && !decoder_push(stream->decoder, packet)) {
+        return false;
+    }
+
+    if (stream->recorder) {
+        packet->dts = packet->pts;
+
+        if (!recorder_write(stream->recorder, packet)) {
+            LOGE("Could not write frame to output file");
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static bool
+stream_parse(struct stream *stream, AVPacket *packet) {
+    uint8_t *in_data = packet->data;
+    int in_len = packet->size;
+    uint8_t *out_data = NULL;
+    int out_len = 0;
+    int r = av_parser_parse2(stream->parser, stream->codec_ctx,
+                             &out_data, &out_len, in_data, in_len,
+                             AV_NOPTS_VALUE, AV_NOPTS_VALUE, -1);
+
+    // PARSER_FLAG_COMPLETE_FRAMES is set
+    SDL_assert(r == in_len);
+    SDL_assert(out_len == in_len);
+
+    if (stream->parser->key_frame == 1) {
+        packet->flags |= AV_PKT_FLAG_KEY;
+    }
+
+    bool ok = process_frame(stream, packet);
+    if (!ok) {
+        LOGE("Could not process frame");
+        return false;
+    }
+
+    return true;
+}
+
+static bool
+stream_push_packet(struct stream *stream, AVPacket *packet) {
+    bool is_config = packet->pts == AV_NOPTS_VALUE;
+
+    // A config packet must not be decoded immetiately (it contains no
+    // frame); instead, it must be concatenated with the future data packet.
+    if (stream->has_pending || is_config) {
+        size_t offset;
+        if (stream->has_pending) {
+            offset = stream->pending.size;
+            if (av_grow_packet(&stream->pending, packet->size)) {
+                LOGE("Could not grow packet");
+                return false;
+            }
+        } else {
+            offset = 0;
+            if (av_new_packet(&stream->pending, packet->size)) {
+                LOGE("Could not create packet");
+                return false;
+            }
+            stream->has_pending = true;
+        }
+
+        memcpy(stream->pending.data + offset, packet->data, packet->size);
+
+        if (!is_config) {
+            // prepare the concat packet to send to the decoder
+            stream->pending.pts = packet->pts;
+            stream->pending.dts = packet->dts;
+            stream->pending.flags = packet->flags;
+            packet = &stream->pending;
+        }
+    }
+
+    if (is_config) {
+        // config packet
+        bool ok = process_config_packet(stream, packet);
+        if (!ok) {
+            return false;
+        }
+    } else {
+        // data packet
+        bool ok = stream_parse(stream, packet);
+
+        if (stream->has_pending) {
+            // the pending packet must be discarded (consumed or error)
+            stream->has_pending = false;
+            av_packet_unref(&stream->pending);
+        }
+
+        if (!ok) {
+            return false;
+        }
+    }
+    return true;
+}
+
 static int
 run_stream(void *data) {
     struct stream *stream = data;
 
-    AVFormatContext *format_ctx = avformat_alloc_context();
-    if (!format_ctx) {
-        LOGC("Could not allocate format context");
-        goto end;
-    }
-
-    unsigned char *buffer = av_malloc(BUFSIZE);
-    if (!buffer) {
-        LOGC("Could not allocate buffer");
-        goto finally_free_format_ctx;
-    }
-
-    // initialize the receiver state
-    stream->receiver_state.frame_meta_queue = NULL;
-    stream->receiver_state.remaining = 0;
-
-    // if recording is enabled, a "header" is sent between raw packets
-    int (*read_packet)(void *, uint8_t *, int) =
-            stream->recorder ? read_packet_with_meta : read_raw_packet;
-    AVIOContext *avio_ctx = avio_alloc_context(buffer, BUFSIZE, 0, stream,
-                                               read_packet, NULL, NULL);
-    if (!avio_ctx) {
-        LOGC("Could not allocate avio context");
-        // avformat_open_input takes ownership of 'buffer'
-        // so only free the buffer before avformat_open_input()
-        av_free(buffer);
-        goto finally_free_format_ctx;
-    }
-
-    format_ctx->pb = avio_ctx;
-
-    if (avformat_open_input(&format_ctx, NULL, NULL, NULL) < 0) {
-        LOGE("Could not open video stream");
-        goto finally_free_avio_ctx;
-    }
-
     AVCodec *codec = avcodec_find_decoder(AV_CODEC_ID_H264);
     if (!codec) {
         LOGE("H.264 decoder not found");
         goto end;
     }
 
+    stream->codec_ctx = avcodec_alloc_context3(codec);
+    if (!stream->codec_ctx) {
+        LOGC("Could not allocate codec context");
+        goto end;
+    }
+
     if (stream->decoder && !decoder_open(stream->decoder, codec)) {
         LOGE("Could not open decoder");
-        goto finally_close_input;
+        goto finally_free_codec_ctx;
     }
 
     if (stream->recorder && !recorder_open(stream->recorder, codec)) {
@@ -201,50 +206,40 @@ run_stream(void *data) {
         goto finally_close_decoder;
     }
 
+    stream->parser = av_parser_init(AV_CODEC_ID_H264);
+    if (!stream->parser) {
+        LOGE("Could not initialize parser");
+        goto finally_close_recorder;
+    }
+
+    // We must only pass complete frames to av_parser_parse2()!
+    // It's more complicated, but this allows to reduce the latency by 1 frame!
+    stream->parser->flags |= PARSER_FLAG_COMPLETE_FRAMES;
+
+    for (;;) {
         AVPacket packet;
-    av_init_packet(&packet);
+        bool ok = stream_recv_packet(stream, &packet);
-    packet.data = NULL;
+        if (!ok) {
-    packet.size = 0;
+            // end of stream
-
+            break;
-    while (!av_read_frame(format_ctx, &packet)) {
-        if (SDL_AtomicGet(&stream->stopped)) {
-            // if the stream is stopped, the socket had been shutdown, so the
-            // last packet is probably corrupted (but not detected as such by
-            // FFmpeg) and will not be decoded correctly
-            av_packet_unref(&packet);
-            goto quit;
-        }
-        if (stream->decoder && !decoder_push(stream->decoder, &packet)) {
-            av_packet_unref(&packet);
-            goto quit;
         }
 
-        if (stream->recorder) {
+        ok = stream_push_packet(stream, &packet);
-            // we retrieve the PTS in order they were received, so they will
-            // be assigned to the correct frame
-            uint64_t pts = receiver_state_take_meta(&stream->receiver_state);
-            packet.pts = pts;
-            packet.dts = pts;
-
-            // no need to rescale with av_packet_rescale_ts(), the timestamps
-            // are in microseconds both in input and output
-            if (!recorder_write(stream->recorder, &packet)) {
-                LOGE("Could not write frame to output file");
         av_packet_unref(&packet);
-                goto quit;
+        if (!ok) {
-            }
+            // cannot process packet (error already logged)
-        }
-
-        av_packet_unref(&packet);
-
-        if (avio_ctx->eof_reached) {
             break;
         }
     }
 
     LOGD("End of frames");
 
-quit:
+    if (stream->has_pending) {
+        av_packet_unref(&stream->pending);
+    }
+
+    av_parser_close(stream->parser);
+finally_close_recorder:
     if (stream->recorder) {
         recorder_close(stream->recorder);
     }
@@ -252,13 +247,8 @@ finally_close_decoder:
     if (stream->decoder) {
         decoder_close(stream->decoder);
     }
-finally_close_input:
+finally_free_codec_ctx:
-    avformat_close_input(&format_ctx);
+    avcodec_free_context(&stream->codec_ctx);
-finally_free_avio_ctx:
-    av_free(avio_ctx->buffer);
-    av_free(avio_ctx);
-finally_free_format_ctx:
-    avformat_free_context(format_ctx);
 end:
     notify_stopped();
     return 0;
@@ -270,7 +260,7 @@ stream_init(struct stream *stream, socket_t socket,
     stream->socket = socket;
     stream->decoder = decoder,
     stream->recorder = recorder;
-    SDL_AtomicSet(&stream->stopped, 0);
+    stream->has_pending = false;
 }
 
 bool
@@ -287,7 +277,6 @@ stream_start(struct stream *stream) {
 
 void
 stream_stop(struct stream *stream) {
-    SDL_AtomicSet(&stream->stopped, 1);
     if (stream->decoder) {
         decoder_interrupt(stream->decoder);
     }

app/src/stream.h

@@ -3,6 +3,7 @@
 
 #include <stdbool.h>
 #include <stdint.h>
+#include <libavformat/avformat.h>
 #include <SDL2/SDL_atomic.h>
 #include <SDL2/SDL_thread.h>
 
@@ -10,23 +11,18 @@
 
 struct video_buffer;
 
-struct frame_meta {
-    uint64_t pts;
-    struct frame_meta *next;
-};
-
 struct stream {
     socket_t socket;
     struct video_buffer *video_buffer;
     SDL_Thread *thread;
-    SDL_atomic_t stopped;
     struct decoder *decoder;
     struct recorder *recorder;
-    struct receiver_state {
+    AVCodecContext *codec_ctx;
-        // meta (in order) for frames not consumed yet
+    AVCodecParserContext *parser;
-        struct frame_meta *frame_meta_queue;
+    // successive packets may need to be concatenated, until a non-config
-        size_t remaining; // remaining bytes to receive for the current frame
+    // packet is available
-    } receiver_state;
+    bool has_pending;
+    AVPacket pending;
 };
 
 void