Add support for uncompressed video instead of H.264 (while still storing H.264 to...

Note that microhttpd and VLC is really inefficient for large data amounts, so
this is not actually a CPU win right now.

diff --git a/flags.cpp b/flags.cpp
index 21e71e99208765ee1122e6900fa3464f98b3839a..c06be70e4aed1b47cd28d46aedc0dccd0a18ad7c 100644 (file)
--- a/flags.cpp
+++ b/flags.cpp
@@ -11,6 +11,7 @@ void parse_flags(int argc, char * const argv[])
        static const option long_options[] = {
                { "num-cards", required_argument, 0, 'c' },
                { "va-display", required_argument, 0, 1000 },
+               { "http-uncompressed-video", no_argument, 0, 1001 },
                { 0, 0, 0, 0 }
        };
        for ( ;; ) {
@@ -27,6 +28,9 @@ void parse_flags(int argc, char * const argv[])
                case 1000:
                        global_flags.va_display = optarg;
                        break;
+               case 1001:
+                       global_flags.uncompressed_video_to_http = true;
+                       break;
                default:
                        fprintf(stderr, "Unknown option '%s'\n", argv[option_index]);
                        exit(1);

diff --git a/flags.h b/flags.h
index 52a9a50dff680c0a71f2dd208e86620396dd7342..2a65c2490889b81a1fab89f4ec7d75aeb61f3a3e 100644 (file)
--- a/flags.h
+++ b/flags.h
@@ -6,6 +6,7 @@
 struct Flags {
        int num_cards = 2;
        std::string va_display;
+       bool uncompressed_video_to_http = false;
 };
 extern Flags global_flags;
 

diff --git a/h264encode.cpp b/h264encode.cpp
index 876e956ec36cfcf7f0abb38c59ad3dc9772933b2..fbeeba707949284bb0ba982cdd86d46a4b0c01b6 100644 (file)
--- a/h264encode.cpp
+++ b/h264encode.cpp
@@ -38,6 +38,7 @@ extern "C" {
 
 #include "context.h"
 #include "defs.h"
+#include "flags.h"
 #include "httpd.h"
 #include "timebase.h"
 
@@ -113,6 +114,82 @@ typedef struct __bitstream bitstream;
 
 using namespace std;
 
+// H.264 video comes out in encoding order (e.g. with two B-frames:
+// 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
+// come in the right order. Since we do everything, including waiting
+// for the frames to come out of OpenGL, in encoding order, we need
+// a reordering buffer for uncompressed frames so that they come out
+// correctly. We go the super-lazy way of not making it understand
+// anything about the true order (which introduces some extra latency,
+// though); we know that for N B-frames we need at most (N-1) frames
+// in the reorder buffer, and can just sort on that.
+//
+// The class also deals with keeping a freelist as needed.
+class FrameReorderer {
+public:
+       FrameReorderer(unsigned queue_length, int width, int height);
+
+       // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
+       // Does _not_ take ownership of data; a copy is taken if needed.
+       // The returned pointer is valid until the next call to reorder_frame, or destruction.
+       // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
+       pair<int64_t, const uint8_t *> reorder_frame(int64_t pts, const uint8_t *data);
+
+       // The same as reorder_frame, but without inserting anything. Used to empty the queue.
+       pair<int64_t, const uint8_t *> get_first_frame();
+
+       bool empty() const { return frames.empty(); }
+
+private:
+       unsigned queue_length;
+       int width, height;
+
+       priority_queue<pair<int64_t, uint8_t *>> frames;
+       stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
+
+       // Owns all the pointers. Normally, freelist and frames could do this themselves,
+       // except priority_queue doesn't work well with movable-only types.
+       vector<unique_ptr<uint8_t[]>> owner;
+};
+
+FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height)
+    : queue_length(queue_length), width(width), height(height)
+{
+       for (unsigned i = 0; i < queue_length; ++i) {
+               owner.emplace_back(new uint8_t[width * height * 2]);
+               freelist.push(owner.back().get());
+       }
+}
+
+pair<int64_t, const uint8_t *> FrameReorderer::reorder_frame(int64_t pts, const uint8_t *data)
+{
+       if (queue_length == 0) {
+               return make_pair(pts, data);
+       }
+
+       assert(!freelist.empty());
+       uint8_t *storage = freelist.top();
+       freelist.pop();
+       memcpy(storage, data, width * height * 2);
+       frames.emplace(-pts, storage);  // Invert pts to get smallest first.
+
+       if (frames.size() >= queue_length) {
+               return get_first_frame();
+       } else {
+               return make_pair(-1, nullptr);
+       }
+}
+
+pair<int64_t, const uint8_t *> FrameReorderer::get_first_frame()
+{
+       assert(!frames.empty());
+       pair<int64_t, uint8_t *> storage = frames.top();
+       frames.pop();
+       int64_t pts = storage.first;
+       freelist.push(storage.second);
+       return make_pair(-pts, storage.second);  // Re-invert pts (see reorder_frame()).
+}
+
 class H264EncoderImpl {
 public:
        H264EncoderImpl(QSurface *surface, const string &va_display, int width, int height, HTTPD *httpd);
@@ -137,6 +214,7 @@ private:
 
        void encode_thread_func();
        void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
+       void add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data);
        void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
                          int frame_type, int64_t pts, int64_t dts);
        void storage_task_thread();
@@ -188,6 +266,7 @@ private:
        AVCodecContext *context_audio;
        AVFrame *audio_frame = nullptr;
        HTTPD *httpd;
+       unique_ptr<FrameReorderer> reorderer;
 
        Display *x11_display = nullptr;
 
@@ -856,6 +935,10 @@ VADisplay H264EncoderImpl::va_open_display(const string &va_display)
                        return NULL;
                }
                use_zerocopy = true;
+               if (global_flags.uncompressed_video_to_http) {
+                       fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n");
+                       use_zerocopy = false;
+               }
                return vaGetDisplay(x11_display);
        } else if (va_display[0] != '/') {
                x11_display = XOpenDisplay(va_display.c_str());
@@ -864,6 +947,10 @@ VADisplay H264EncoderImpl::va_open_display(const string &va_display)
                        return NULL;
                }
                use_zerocopy = true;
+               if (global_flags.uncompressed_video_to_http) {
+                       fprintf(stderr, "Disabling zerocopy H.264 encoding due to --uncompressed_video_to_http.\n");
+                       use_zerocopy = false;
+               }
                return vaGetDisplay(x11_display);
        } else {
                drm_fd = open(va_display.c_str(), O_RDWR);
@@ -1524,7 +1611,8 @@ void H264EncoderImpl::save_codeddata(storage_task task)
             pkt.flags = 0;
         }
         //pkt.duration = 1;
-        httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay);
+        httpd->add_packet(pkt, task.pts + global_delay, task.dts + global_delay,
+               global_flags.uncompressed_video_to_http ? HTTPD::DESTINATION_FILE_ONLY : HTTPD::DESTINATION_FILE_AND_HTTP);
     }
     // Encode and add all audio frames up to and including the pts of this video frame.
     for ( ;; ) {
@@ -1569,7 +1657,7 @@ void H264EncoderImpl::save_codeddata(storage_task task)
         avcodec_encode_audio2(context_audio, &pkt, audio_frame, &got_output);
         if (got_output) {
             pkt.stream_index = 1;
-            httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
+            httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay, HTTPD::DESTINATION_FILE_AND_HTTP);
         }
         // TODO: Delayed frames.
         av_frame_unref(audio_frame);
@@ -1672,6 +1760,10 @@ H264EncoderImpl::H264EncoderImpl(QSurface *surface, const string &va_display, in
 
        //print_input();
 
+       if (global_flags.uncompressed_video_to_http) {
+               reorderer.reset(new FrameReorderer(ip_period - 1, frame_width, frame_height));
+       }
+
        init_va(va_display);
        setup_encode();
 
@@ -1917,6 +2009,26 @@ void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int g
                encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts);
                last_dts = dts;
        }
+
+       if (global_flags.uncompressed_video_to_http) {
+               // Add frames left in reorderer.
+               while (!reorderer->empty()) {
+                       pair<int64_t, const uint8_t *> output_frame = reorderer->get_first_frame();
+                       add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
+               }
+       }
+}
+
+void H264EncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, const uint8_t *data)
+{
+       AVPacket pkt;
+       memset(&pkt, 0, sizeof(pkt));
+       pkt.buf = nullptr;
+       pkt.data = const_cast<uint8_t *>(data);
+       pkt.size = frame_width * frame_height * 2;
+       pkt.stream_index = 0;
+       pkt.flags = AV_PKT_FLAG_KEY;
+       httpd->add_packet(pkt, pts, pts, HTTPD::DESTINATION_HTTP_ONLY);
 }
 
 namespace {
@@ -1970,6 +2082,15 @@ void H264EncoderImpl::encode_frame(H264EncoderImpl::PendingFrame frame, int enco
 
                va_status = vaUnmapBuffer(va_dpy, surf->surface_image.buf);
                CHECK_VASTATUS(va_status, "vaUnmapBuffer");
+
+               if (global_flags.uncompressed_video_to_http) {
+                       // Add uncompressed video. (Note that pts == dts here.)
+                       const int64_t global_delay = int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);  // Needs to match audio.
+                       pair<int64_t, const uint8_t *> output_frame = reorderer->reorder_frame(pts + global_delay, reinterpret_cast<uint8_t *>(surf->y_ptr));
+                       if (output_frame.second != nullptr) {
+                               add_packet_for_uncompressed_frame(output_frame.first, output_frame.second);
+                       }
+               }
        }
 
        va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);

diff --git a/httpd.cpp b/httpd.cpp
index 401509d265f5404a63119691e345d35f7f954d86..d9bc022bc9029844936b9a0a65490571a9ef7482 100644 (file)
--- a/httpd.cpp
+++ b/httpd.cpp
@@ -19,6 +19,7 @@ extern "C" {
 #include "httpd.h"
 
 #include "defs.h"
+#include "flags.h"
 #include "timebase.h"
 
 struct MHD_Connection;
@@ -41,13 +42,15 @@ void HTTPD::start(int port)
                         MHD_OPTION_END);
 }
 
-void HTTPD::add_packet(const AVPacket &pkt, int64_t pts, int64_t dts)
+void HTTPD::add_packet(const AVPacket &pkt, int64_t pts, int64_t dts, PacketDestination destination)
 {
        unique_lock<mutex> lock(streams_mutex);
-       for (Stream *stream : streams) {
-               stream->add_packet(pkt, pts, dts);
+       if (destination != DESTINATION_FILE_ONLY) {
+               for (Stream *stream : streams) {
+                       stream->add_packet(pkt, pts, dts);
+               }
        }
-       if (file_mux) {
+       if (file_mux && destination != DESTINATION_HTTP_ONLY) {
                file_mux->add_packet(pkt, pts, dts);
        }
 }
@@ -67,7 +70,7 @@ void HTTPD::open_output_file(const string &filename)
                exit(1);
        }
 
-       file_mux.reset(new Mux(avctx, width, height));
+       file_mux.reset(new Mux(avctx, width, height, Mux::CODEC_NV12));
 }
 
 void HTTPD::close_output_file()
@@ -135,10 +138,10 @@ void HTTPD::request_completed(struct MHD_Connection *connection, void **con_cls,
        }
 }
 
-HTTPD::Mux::Mux(AVFormatContext *avctx, int width, int height)
+HTTPD::Mux::Mux(AVFormatContext *avctx, int width, int height, Codec codec)
        : avctx(avctx)
 {
-       AVCodec *codec_video = avcodec_find_encoder(AV_CODEC_ID_H264);
+       AVCodec *codec_video = avcodec_find_encoder((codec == CODEC_H264) ? AV_CODEC_ID_H264 : AV_CODEC_ID_RAWVIDEO);
        avstream_video = avformat_new_stream(avctx, codec_video);
        if (avstream_video == nullptr) {
                fprintf(stderr, "avformat_new_stream() failed\n");
@@ -146,7 +149,13 @@ HTTPD::Mux::Mux(AVFormatContext *avctx, int width, int height)
        }
        avstream_video->time_base = AVRational{1, TIMEBASE};
        avstream_video->codec->codec_type = AVMEDIA_TYPE_VIDEO;
-       avstream_video->codec->codec_id = AV_CODEC_ID_H264;
+       if (codec == CODEC_H264) {
+               avstream_video->codec->codec_id = AV_CODEC_ID_H264;
+       } else {
+               assert(codec == CODEC_NV12);
+               avstream_video->codec->codec_id = AV_CODEC_ID_RAWVIDEO;
+               avstream_video->codec->codec_tag = avcodec_pix_fmt_to_codec_tag(AV_PIX_FMT_NV12);
+       }
        avstream_video->codec->width = width;
        avstream_video->codec->height = height;
        avstream_video->codec->time_base = AVRational{1, TIMEBASE};
@@ -236,9 +245,17 @@ HTTPD::Stream::Stream(AVOutputFormat *oformat, int width, int height)
        avctx->oformat = oformat;
        uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
        avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, &HTTPD::Stream::write_packet_thunk, nullptr);
+
+       Mux::Codec codec;
+       if (global_flags.uncompressed_video_to_http) {
+               codec = Mux::CODEC_NV12;
+       } else {
+               codec = Mux::CODEC_H264;
+       }
+
        avctx->flags = AVFMT_FLAG_CUSTOM_IO;
 
-       mux.reset(new Mux(avctx, width, height));
+       mux.reset(new Mux(avctx, width, height, codec));
 }
 
 ssize_t HTTPD::Stream::reader_callback_thunk(void *cls, uint64_t pos, char *buf, size_t max)

diff --git a/httpd.h b/httpd.h
index 88aeecc8263dffc2f4bc961db424324a9744b794..f5bf55f411574ab20dafc47bb9635728019a45fc 100644 (file)
--- a/httpd.h
+++ b/httpd.h
@@ -29,9 +29,15 @@ extern "C" {
 
 class HTTPD {
 public:
+       enum PacketDestination {
+               DESTINATION_FILE_ONLY,
+               DESTINATION_HTTP_ONLY,
+               DESTINATION_FILE_AND_HTTP
+       };
+
        HTTPD(int width, int height);
        void start(int port);
-       void add_packet(const AVPacket &pkt, int64_t pts, int64_t dts);
+       void add_packet(const AVPacket &pkt, int64_t pts, int64_t dts, PacketDestination destination);
 
        // You can only have one going at the same time.
        void open_output_file(const std::string &filename);
@@ -56,7 +62,12 @@ private:
 
        class Mux {
        public:
-               Mux(AVFormatContext *avctx, int width, int height);  // Takes ownership of avctx.
+               enum Codec {
+                       CODEC_H264,
+                       CODEC_NV12,  // Uncompressed 4:2:0.
+               };
+
+               Mux(AVFormatContext *avctx, int width, int height, Codec codec);  // Takes ownership of avctx.
                ~Mux();
                void add_packet(const AVPacket &pkt, int64_t pts, int64_t dts);