]> git.sesse.net Git - nageru/commitdiff
Add functionality for MJPEG export.
authorSteinar H. Gunderson <sgunderson@bigfoot.com>
Sun, 14 Oct 2018 22:14:22 +0000 (00:14 +0200)
committerSteinar H. Gunderson <sgunderson@bigfoot.com>
Sun, 11 Nov 2018 12:40:11 +0000 (13:40 +0100)
This simply exports all incoming frames under /multicam.mp4,
as one big 4:2:2 MJPEG multi-video stream. The intended use is
for multi-camera recording, or for instant replay.

15 files changed:
bmusb
decklink_capture.cpp
httpd.cpp
httpd.h
kaeru.cpp
meson.build
mixer.cpp
mixer.h
mjpeg_encoder.cpp [new file with mode: 0644]
mjpeg_encoder.h [new file with mode: 0644]
pbo_frame_allocator.cpp
quicksync_encoder.cpp
quicksync_encoder_impl.h
va_display_with_cleanup.h [new file with mode: 0644]
video_encoder.cpp

diff --git a/bmusb b/bmusb
index b447be096623b949a2117c71af03bb96d329d988..e9aa80b8e6a09cf2c8b3458a1c395b6f1b6a58cc 160000 (submodule)
--- a/bmusb
+++ b/bmusb
@@ -1 +1 @@
-Subproject commit b447be096623b949a2117c71af03bb96d329d988
+Subproject commit e9aa80b8e6a09cf2c8b3458a1c395b6f1b6a58cc
index 881e1817b1b7262f295584ac24e7f6d80fd1b65e..0f48e3e77f48ffe18c74175675060b52ea02f50b 100644 (file)
@@ -265,6 +265,9 @@ HRESULT STDMETHODCALLTYPE DeckLinkCapture::VideoInputFrameArrived(
                        } else {
                                memcpy(current_video_frame.data, frame_bytes, num_bytes);
                        }
+                       if (current_video_frame.data_copy != nullptr) {
+                               memcpy(current_video_frame.data_copy, frame_bytes, num_bytes);
+                       }
                        current_video_frame.len += num_bytes;
 
                        video_format.width = width;
index f6441768bbdde1df3d31e2b42a2cb13430019376..8d33e75a4f6b5bde8e6f8b79aa23ffa007e8351f 100644 (file)
--- a/httpd.cpp
+++ b/httpd.cpp
@@ -58,11 +58,13 @@ void HTTPD::stop()
        }
 }
 
-void HTTPD::add_data(const char *buf, size_t size, bool keyframe, int64_t time, AVRational timebase)
+void HTTPD::add_data(StreamType stream_type, const char *buf, size_t size, bool keyframe, int64_t time, AVRational timebase)
 {
        unique_lock<mutex> lock(streams_mutex);
        for (Stream *stream : streams) {
-               stream->add_data(buf, size, keyframe ? Stream::DATA_TYPE_KEYFRAME : Stream::DATA_TYPE_OTHER, time, timebase);
+               if (stream->get_stream_type() == stream_type) {
+                       stream->add_data(buf, size, keyframe ? Stream::DATA_TYPE_KEYFRAME : Stream::DATA_TYPE_OTHER, time, timebase);
+               }
        }
 }
 
@@ -87,6 +89,12 @@ int HTTPD::answer_to_connection(MHD_Connection *connection,
        } else {
                framing = HTTPD::Stream::FRAMING_RAW;
        }
+       HTTPD::StreamType stream_type;
+       if (strcmp(url, "/multicam.mp4") == 0) {
+               stream_type = HTTPD::StreamType::MULTICAM_STREAM;
+       } else {
+               stream_type = HTTPD::StreamType::MAIN_STREAM;
+       }
 
        if (strcmp(url, "/metrics") == 0) {
                string contents = global_metrics.serialize();
@@ -121,8 +129,8 @@ int HTTPD::answer_to_connection(MHD_Connection *connection,
                return ret;
        }
 
-       HTTPD::Stream *stream = new HTTPD::Stream(this, framing);
-       stream->add_data(header.data(), header.size(), Stream::DATA_TYPE_HEADER, AV_NOPTS_VALUE, AVRational{ 1, 0 });
+       HTTPD::Stream *stream = new HTTPD::Stream(this, framing, stream_type);
+       stream->add_data(header[stream_type].data(), header[stream_type].size(), Stream::DATA_TYPE_HEADER, AV_NOPTS_VALUE, AVRational{ 1, 0 });
        {
                unique_lock<mutex> lock(streams_mutex);
                streams.insert(stream);
diff --git a/httpd.h b/httpd.h
index 57c649b61158c6f73b498aeca6556b97bc52f179..1ff5c51108facf03ca1b0d8fc2f076a63938e595 100644 (file)
--- a/httpd.h
+++ b/httpd.h
@@ -31,9 +31,15 @@ public:
        HTTPD();
        ~HTTPD();
 
+       enum StreamType {
+               MAIN_STREAM,
+               MULTICAM_STREAM,
+               NUM_STREAM_TYPES
+       };
+
        // Should be called before start().
-       void set_header(const std::string &data) {
-               header = data;
+       void set_header(StreamType stream_type, const std::string &data) {
+               header[stream_type] = data;
        }
 
        // Should be called before start() (due to threading issues).
@@ -47,7 +53,7 @@ public:
 
        void start(int port);
        void stop();
-       void add_data(const char *buf, size_t size, bool keyframe, int64_t time, AVRational timebase);
+       void add_data(StreamType stream_type, const char *buf, size_t size, bool keyframe, int64_t time, AVRational timebase);
        int64_t get_num_connected_clients() const {
                return metric_num_connected_clients.load();
        }
@@ -72,7 +78,8 @@ private:
                        FRAMING_RAW,
                        FRAMING_METACUBE
                };
-               Stream(HTTPD *parent, Framing framing) : parent(parent), framing(framing) {}
+               Stream(HTTPD *parent, Framing framing, StreamType stream_type)
+                       : parent(parent), framing(framing), stream_type(stream_type) {}
 
                static ssize_t reader_callback_thunk(void *cls, uint64_t pos, char *buf, size_t max);
                ssize_t reader_callback(uint64_t pos, char *buf, size_t max);
@@ -85,6 +92,7 @@ private:
                void add_data(const char *buf, size_t size, DataType data_type, int64_t time, AVRational timebase);
                void stop();
                HTTPD *get_parent() const { return parent; }
+               StreamType get_stream_type() const { return stream_type; }
 
        private:
                HTTPD *parent;
@@ -96,6 +104,7 @@ private:
                std::deque<std::string> buffered_data;  // Protected by <buffer_mutex>.
                size_t used_of_buffered_data = 0;  // How many bytes of the first element of <buffered_data> that is already used. Protected by <mutex>.
                size_t seen_keyframe = false;
+               StreamType stream_type;
        };
 
        MHD_Daemon *mhd = nullptr;
@@ -106,7 +115,7 @@ private:
                CORSPolicy cors_policy;
        };
        std::unordered_map<std::string, Endpoint> endpoints;
-       std::string header;
+       std::string header[NUM_STREAM_TYPES];
 
        // Metrics.
        std::atomic<int64_t> metric_num_connected_clients{0};
index 10f1e9394b2a5c0ddab4e42a226b0ff7af4ccc73..b3a9bb3e43f92e09501198cd54c190d3a5579538 100644 (file)
--- a/kaeru.cpp
+++ b/kaeru.cpp
@@ -31,6 +31,8 @@ BasicStats *global_basic_stats = nullptr;
 QuittableSleeper should_quit;
 MuxMetrics stream_mux_metrics;
 
+namespace {
+
 int write_packet(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
 {
        static bool seen_sync_markers = false;
@@ -47,13 +49,15 @@ int write_packet(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType ty
 
        if (type == AVIO_DATA_MARKER_HEADER) {
                stream_mux_header.append((char *)buf, buf_size);
-               httpd->set_header(stream_mux_header);
+               httpd->set_header(HTTPD::MAIN_STREAM, stream_mux_header);
        } else {
-               httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
+               httpd->add_data(HTTPD::MAIN_STREAM, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
        }
        return buf_size;
 }
 
+}  // namespace
+
 unique_ptr<Mux> create_mux(HTTPD *httpd, AVOutputFormat *oformat, X264Encoder *x264_encoder, AudioEncoder *audio_encoder)
 {
        AVFormatContext *avctx = avformat_alloc_context();
index cc161f2dae4e4a12576960e7c89d398289564fa7..d35dcb5661a7c0b3a66d880a87003f79de3d131c 100644 (file)
@@ -190,7 +190,7 @@ nageru_link_with += audio
 # Mixer objects.
 srcs += ['chroma_subsampler.cpp', 'v210_converter.cpp', 'mixer.cpp', 'pbo_frame_allocator.cpp',
        'context.cpp', 'theme.cpp', 'image_input.cpp', 'alsa_output.cpp',
-       'disk_space_estimator.cpp', 'timecode_renderer.cpp', 'tweaked_inputs.cpp']
+       'disk_space_estimator.cpp', 'timecode_renderer.cpp', 'tweaked_inputs.cpp', 'mjpeg_encoder.cpp']
 
 # Streaming and encoding objects (largely the set that is shared between Nageru and Kaeru).
 stream_srcs = ['quicksync_encoder.cpp', 'x264_encoder.cpp', 'x264_dynamic.cpp', 'x264_speed_control.cpp', 'video_encoder.cpp',
index deaa8e73079710b7f13b8dffd6a9f084f0ae9b4c..294040f1bfd6fc6dde71a793af91c536e2c9f12e 100644 (file)
--- a/mixer.cpp
+++ b/mixer.cpp
 #include "flags.h"
 #include "input_mapping.h"
 #include "metrics.h"
+#include "mjpeg_encoder.h"
 #include "pbo_frame_allocator.h"
 #include "ref_counted_gl_sync.h"
 #include "resampling_queue.h"
 #include "timebase.h"
 #include "timecode_renderer.h"
 #include "v210_converter.h"
+#include "va_display_with_cleanup.h"
 #include "video_encoder.h"
 
 #undef Status
@@ -356,6 +358,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
        display_chain->finalize();
 
        video_encoder.reset(new VideoEncoder(resource_pool.get(), h264_encoder_surface, global_flags.va_display, global_flags.width, global_flags.height, &httpd, global_disk_space_estimator));
+       mjpeg_encoder.reset(new MJPEGEncoder(&httpd, global_flags.va_display));
 
        // Must be instantiated after VideoEncoder has initialized global_flags.use_zerocopy.
        theme.reset(new Theme(global_flags.theme_filename, global_flags.theme_dirs, resource_pool.get(), num_cards));
@@ -499,6 +502,7 @@ Mixer::Mixer(const QSurfaceFormat &format, unsigned num_cards)
 
 Mixer::~Mixer()
 {
+       mjpeg_encoder->stop();
        httpd.stop();
        BMUSBCapture::stop_bm_thread();
 
@@ -954,6 +958,9 @@ void Mixer::bm_frame(unsigned card_index, uint16_t timecode,
                        new_frame.upload_func = upload_func;
                        new_frame.dropped_frames = dropped_frames;
                        new_frame.received_timestamp = video_frame.received_timestamp;  // Ignore the audio timestamp.
+                       new_frame.video_format = video_format;
+                       new_frame.y_offset = y_offset;
+                       new_frame.cbcr_offset = cbcr_offset;
                        card->new_frames.push_back(move(new_frame));
                        card->jitter_history.frame_arrived(video_frame.received_timestamp, frame_length, dropped_frames);
                        card->may_have_dropped_last_frame = false;
@@ -1063,6 +1070,14 @@ void Mixer::thread_func()
                                new_frame->upload_func();
                                new_frame->upload_func = nullptr;
                        }
+
+                       // There are situations where we could possibly want to
+                       // include FFmpeg inputs (CEF inputs are unlikely),
+                       // but they're not necessarily in 4:2:2 Y'CbCr, so it would
+                       // require more functionality the the JPEG encoder.
+                       if (card_index < num_cards) {
+                               mjpeg_encoder->upload_frame(pts_int, card_index, new_frame->frame, new_frame->video_format, new_frame->y_offset, new_frame->cbcr_offset);
+                       }
                }
 
                int64_t frame_duration = output_frame_info.frame_duration;
diff --git a/mixer.h b/mixer.h
index 32a1ea476f21547137e680fb13675c80b6b42835..84313e5130b0816b2a997ee9265d637781cb3aed 100644 (file)
--- a/mixer.h
+++ b/mixer.h
@@ -42,6 +42,7 @@
 class ALSAOutput;
 class ChromaSubsampler;
 class DeckLinkOutput;
+class MJPEGEncoder;
 class QSurface;
 class QSurfaceFormat;
 class TimecodeRenderer;
@@ -475,6 +476,7 @@ private:
        std::unique_ptr<ChromaSubsampler> chroma_subsampler;
        std::unique_ptr<v210Converter> v210_converter;
        std::unique_ptr<VideoEncoder> video_encoder;
+       std::unique_ptr<MJPEGEncoder> mjpeg_encoder;
 
        std::unique_ptr<TimecodeRenderer> timecode_renderer;
        std::atomic<bool> display_timecode_in_stream{false};
@@ -530,6 +532,12 @@ private:
                        std::function<void()> upload_func;  // Needs to be called to actually upload the texture to OpenGL.
                        unsigned dropped_frames = 0;  // Number of dropped frames before this one.
                        std::chrono::steady_clock::time_point received_timestamp = std::chrono::steady_clock::time_point::min();
+
+                       // Used for MJPEG encoding. (upload_func packs everything it needs
+                       // into the functor, but would otherwise also use these.)
+                       // width=0 or height=0 means a broken frame, ie., do not upload.
+                       bmusb::VideoFormat video_format;
+                       size_t y_offset, cbcr_offset;
                };
                std::deque<NewFrame> new_frames;
                std::condition_variable new_frames_changed;  // Set whenever new_frames is changed.
diff --git a/mjpeg_encoder.cpp b/mjpeg_encoder.cpp
new file mode 100644 (file)
index 0000000..740b059
--- /dev/null
@@ -0,0 +1,677 @@
+#include "mjpeg_encoder.h"
+
+#include <jpeglib.h>
+#include <unistd.h>
+#if __SSE2__
+#include <immintrin.h>
+#endif
+#include <list>
+
+extern "C" {
+#include <libavformat/avformat.h>
+}
+
+#include "defs.h"
+#include "ffmpeg_raii.h"
+#include "flags.h"
+#include "httpd.h"
+#include "memcpy_interleaved.h"
+#include "pbo_frame_allocator.h"
+#include "timebase.h"
+#include "va_display_with_cleanup.h"
+
+#include <va/va.h>
+#include <va/va_drm.h>
+#include <va/va_x11.h>
+
+using namespace bmusb;
+using namespace std;
+
+extern void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height);
+
+#define CHECK_VASTATUS(va_status, func)                                 \
+    if (va_status != VA_STATUS_SUCCESS) {                               \
+        fprintf(stderr, "%s:%d (%s) failed with %d\n", __func__, __LINE__, func, va_status); \
+        exit(1);                                                        \
+    }
+
+// From libjpeg (although it's of course identical between implementations).
+static const int jpeg_natural_order[DCTSIZE2] = {
+        0,  1,  8, 16,  9,  2,  3, 10,
+       17, 24, 32, 25, 18, 11,  4,  5,
+       12, 19, 26, 33, 40, 48, 41, 34,
+       27, 20, 13,  6,  7, 14, 21, 28,
+       35, 42, 49, 56, 57, 50, 43, 36,
+       29, 22, 15, 23, 30, 37, 44, 51,
+       58, 59, 52, 45, 38, 31, 39, 46,
+       53, 60, 61, 54, 47, 55, 62, 63,
+};
+
+struct VectorDestinationManager {
+       jpeg_destination_mgr pub;
+       std::vector<uint8_t> dest;
+
+       VectorDestinationManager()
+       {
+               pub.init_destination = init_destination_thunk;
+               pub.empty_output_buffer = empty_output_buffer_thunk;
+               pub.term_destination = term_destination_thunk;
+       }
+
+       static void init_destination_thunk(j_compress_ptr ptr)
+       {
+               ((VectorDestinationManager *)(ptr->dest))->init_destination();
+       }
+
+       inline void init_destination()
+       {
+               make_room(0);
+       }
+
+       static boolean empty_output_buffer_thunk(j_compress_ptr ptr)
+       {
+               return ((VectorDestinationManager *)(ptr->dest))->empty_output_buffer();
+       }
+
+       inline bool empty_output_buffer()
+       {
+               make_room(dest.size());  // Should ignore pub.free_in_buffer!
+               return true;
+       }
+
+       inline void make_room(size_t bytes_used)
+       {
+               dest.resize(bytes_used + 4096);
+               dest.resize(dest.capacity());
+               pub.next_output_byte = dest.data() + bytes_used;
+               pub.free_in_buffer = dest.size() - bytes_used;
+       }
+
+       static void term_destination_thunk(j_compress_ptr ptr)
+       {
+               ((VectorDestinationManager *)(ptr->dest))->term_destination();
+       }
+
+       inline void term_destination()
+       {
+               dest.resize(dest.size() - pub.free_in_buffer);
+       }
+};
+static_assert(std::is_standard_layout<VectorDestinationManager>::value, "");
+
+int MJPEGEncoder::write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
+{
+       MJPEGEncoder *engine = (MJPEGEncoder *)opaque;
+       return engine->write_packet2(buf, buf_size, type, time);
+}
+
+int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time)
+{
+       if (type == AVIO_DATA_MARKER_HEADER) {
+               mux_header.append((char *)buf, buf_size);
+               httpd->set_header(HTTPD::MULTICAM_STREAM, mux_header);
+       } else {
+               httpd->add_data(HTTPD::MULTICAM_STREAM, (char *)buf, buf_size, /*keyframe=*/true, AV_NOPTS_VALUE, AVRational{ AV_TIME_BASE, 1 });
+       }
+       return buf_size;
+}
+
+MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display)
+       : httpd(httpd)
+{
+       encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this);
+
+       // Set up the mux. We don't use the Mux wrapper, because it's geared towards
+       // a situation with only one video stream (and possibly one audio stream)
+       // with known width/height, and we don't need the extra functionality it provides.
+       avctx.reset(avformat_alloc_context());
+       avctx->oformat = av_guess_format("mp4", nullptr, nullptr);
+
+       uint8_t *buf = (uint8_t *)av_malloc(MUX_BUFFER_SIZE);
+       avctx->pb = avio_alloc_context(buf, MUX_BUFFER_SIZE, 1, this, nullptr, nullptr, nullptr);
+       avctx->pb->write_data_type = &MJPEGEncoder::write_packet2_thunk;
+       avctx->flags = AVFMT_FLAG_CUSTOM_IO;
+
+       for (int card_idx = 0; card_idx < global_flags.num_cards; ++card_idx) {
+               AVStream *stream = avformat_new_stream(avctx.get(), nullptr);
+               if (stream == nullptr) {
+                       fprintf(stderr, "avformat_new_stream() failed\n");
+                       exit(1);
+               }
+               stream->time_base = AVRational{ 1, TIMEBASE };
+               stream->codecpar->codec_type = AVMEDIA_TYPE_VIDEO;
+               stream->codecpar->codec_id = AV_CODEC_ID_MJPEG;
+
+               // Used for aspect ratio only. Can change without notice (the mux won't care).
+               stream->codecpar->width = global_flags.width;
+               stream->codecpar->height = global_flags.height;
+
+               // TODO: We could perhaps use the interpretation for each card here
+               // (or at least the command-line flags) instead of the defaults,
+               // but what would we do when they change?
+               stream->codecpar->color_primaries = AVCOL_PRI_BT709;
+               stream->codecpar->color_trc = AVCOL_TRC_IEC61966_2_1;
+               stream->codecpar->color_space = AVCOL_SPC_BT709;
+               stream->codecpar->color_range = AVCOL_RANGE_MPEG;
+               stream->codecpar->chroma_location = AVCHROMA_LOC_LEFT;
+               stream->codecpar->field_order = AV_FIELD_PROGRESSIVE;
+       }
+
+       AVDictionary *options = NULL;
+       vector<pair<string, string>> opts = MUX_OPTS;
+       for (pair<string, string> opt : opts) {
+               av_dict_set(&options, opt.first.c_str(), opt.second.c_str(), 0);
+       }
+       if (avformat_write_header(avctx.get(), &options) < 0) {
+               fprintf(stderr, "avformat_write_header() failed\n");
+               exit(1);
+       }
+
+       // Initialize VA-API.
+       string error;
+       va_dpy = try_open_va(va_display, &error, &config_id);
+       if (va_dpy == nullptr) {
+               fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str());
+       }
+
+       running = true;
+}
+
+void MJPEGEncoder::stop()
+{
+       if (!running) {
+               return;
+       }
+       running = false;
+       should_quit = true;
+       any_frames_to_be_encoded.notify_all();
+       encoder_thread.join();
+}
+
+unique_ptr<VADisplayWithCleanup> MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id)
+{
+       unique_ptr<VADisplayWithCleanup> va_dpy = va_open_display(va_display);
+       if (va_dpy == nullptr) {
+               if (error) *error = "Opening VA display failed";
+               return nullptr;
+       }
+       int major_ver, minor_ver;
+       VAStatus va_status = vaInitialize(va_dpy->va_dpy, &major_ver, &minor_ver);
+       if (va_status != VA_STATUS_SUCCESS) {
+               char buf[256];
+               snprintf(buf, sizeof(buf), "vaInitialize() failed with status %d\n", va_status);
+               if (error != nullptr) *error = buf;
+               return nullptr;
+       }
+
+       VAConfigAttrib attr = { VAConfigAttribRTFormat, VA_RT_FORMAT_YUV422 };
+       va_status = vaCreateConfig(va_dpy->va_dpy, VAProfileJPEGBaseline, VAEntrypointEncPicture,
+               &attr, 1, config_id);
+       if (va_status == VA_STATUS_ERROR_UNSUPPORTED_ENTRYPOINT) {
+               if (error != nullptr) *error = "No hardware support";
+               return nullptr;
+       } else if (va_status != VA_STATUS_SUCCESS) {
+               char buf[256];
+               snprintf(buf, sizeof(buf), "vaCreateConfig() failed with status %d\n", va_status);
+               if (error != nullptr) *error = buf;
+               return nullptr;
+       }
+
+       int num_formats = vaMaxNumImageFormats(va_dpy->va_dpy);
+       assert(num_formats > 0);
+
+       unique_ptr<VAImageFormat[]> formats(new VAImageFormat[num_formats]);
+       va_status = vaQueryImageFormats(va_dpy->va_dpy, formats.get(), &num_formats);
+       if (va_status != VA_STATUS_SUCCESS) {
+               char buf[256];
+               snprintf(buf, sizeof(buf), "vaQueryImageFormats() failed with status %d\n", va_status);
+               if (error != nullptr) *error = buf;
+               return nullptr;
+       }
+
+       return va_dpy;
+}
+
+void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFrame frame, const bmusb::VideoFormat &video_format, size_t y_offset, size_t cbcr_offset)
+{
+       PBOFrameAllocator::Userdata *userdata = (PBOFrameAllocator::Userdata *)frame->userdata;
+       if (video_format.width == 0 || video_format.height == 0) {
+               return;
+       }
+       if (video_format.interlaced) {
+               fprintf(stderr, "Card %u: Ignoring JPEG encoding for interlaced frame\n", card_index);
+               return;
+       }
+       if (userdata->pixel_format != PixelFormat_8BitYCbCr ||
+           !frame->interleaved) {
+               fprintf(stderr, "Card %u: Ignoring JPEG encoding for unsupported pixel format\n", card_index);
+               return;
+       }
+       if (video_format.width > 4096 || video_format.height > 4096) {
+               fprintf(stderr, "Card %u: Ignoring JPEG encoding for oversized frame\n", card_index);
+               return;
+       }
+
+       lock_guard<mutex> lock(mu);
+       frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset });
+       any_frames_to_be_encoded.notify_all();
+}
+
+void MJPEGEncoder::encoder_thread_func()
+{
+       pthread_setname_np(pthread_self(), "MJPEG_Encode");
+       posix_memalign((void **)&tmp_y, 4096, 4096 * 8);
+       posix_memalign((void **)&tmp_cbcr, 4096, 4096 * 8);
+       posix_memalign((void **)&tmp_cb, 4096, 4096 * 8);
+       posix_memalign((void **)&tmp_cr, 4096, 4096 * 8);
+
+       unique_lock<mutex> lock(mu);
+       for (;;) {
+               any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; });
+               if (should_quit) return;
+               QueuedFrame qf = move(frames_to_be_encoded.front());
+               frames_to_be_encoded.pop();
+
+               vector<uint8_t> jpeg = encode_jpeg(qf);
+
+               AVPacket pkt;
+               memset(&pkt, 0, sizeof(pkt));
+               pkt.buf = nullptr;
+               pkt.data = &jpeg[0];
+               pkt.size = jpeg.size();
+               pkt.stream_index = qf.card_index;
+               pkt.flags = AV_PKT_FLAG_KEY;
+               pkt.pts = pkt.dts = qf.pts;
+
+               if (av_write_frame(avctx.get(), &pkt) < 0) {
+                       fprintf(stderr, "av_write_frame() failed\n");
+                       exit(1);
+               }
+       }
+}
+
+class VABufferDestroyer {
+public:
+       VABufferDestroyer(VADisplay dpy, VABufferID buf)
+               : dpy(dpy), buf(buf) {}
+
+       ~VABufferDestroyer() {
+               VAStatus va_status = vaDestroyBuffer(dpy, buf);
+               CHECK_VASTATUS(va_status, "vaDestroyBuffer");
+       }
+
+private:
+       VADisplay dpy;
+       VABufferID buf;
+};
+
+MJPEGEncoder::VAResources MJPEGEncoder::get_va_resources(unsigned width, unsigned height)
+{
+       {
+               lock_guard<mutex> lock(va_resources_mutex);
+               for (auto it = va_resources_freelist.begin(); it != va_resources_freelist.end(); ++it) {
+                       if (it->width == width && it->height == height) {
+                               VAResources ret = *it;
+                               va_resources_freelist.erase(it);
+                               return ret;
+                       }
+               }
+       }
+
+       VAResources ret;
+
+       ret.width = width;
+       ret.height = height;
+
+       VASurfaceAttrib attrib;
+       attrib.flags = VA_SURFACE_ATTRIB_SETTABLE;
+       attrib.type = VASurfaceAttribPixelFormat;
+       attrib.value.type = VAGenericValueTypeInteger;
+       attrib.value.value.i = VA_FOURCC_UYVY;
+
+       VAStatus va_status = vaCreateSurfaces(va_dpy->va_dpy, VA_RT_FORMAT_YUV422,
+               width, height,
+               &ret.surface, 1, &attrib, 1);
+       CHECK_VASTATUS(va_status, "vaCreateSurfaces");
+
+       va_status = vaCreateContext(va_dpy->va_dpy, config_id, width, height, 0, &ret.surface, 1, &ret.context);
+       CHECK_VASTATUS(va_status, "vaCreateContext");
+
+       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncCodedBufferType, width * height * 3 + 8192, 1, nullptr, &ret.data_buffer);
+       CHECK_VASTATUS(va_status, "vaCreateBuffer");
+
+       return ret;
+}
+
+void MJPEGEncoder::release_va_resources(MJPEGEncoder::VAResources resources)
+{
+       lock_guard<mutex> lock(va_resources_mutex);
+       if (va_resources_freelist.size() > 10) {
+               auto it = va_resources_freelist.end();
+               --it;
+
+               VAStatus va_status = vaDestroyBuffer(va_dpy->va_dpy, it->data_buffer);
+               CHECK_VASTATUS(va_status, "vaDestroyBuffer");
+
+               va_status = vaDestroyContext(va_dpy->va_dpy, it->context);
+               CHECK_VASTATUS(va_status, "vaDestroyContext");
+
+               va_status = vaDestroySurfaces(va_dpy->va_dpy, &it->surface, 1);
+               CHECK_VASTATUS(va_status, "vaDestroySurfaces");
+
+               va_resources_freelist.erase(it);
+       }
+
+       va_resources_freelist.push_front(resources);
+}
+
+void MJPEGEncoder::init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo)
+{
+       jpeg_error_mgr jerr;
+       cinfo->err = jpeg_std_error(&jerr);
+       jpeg_create_compress(cinfo);
+
+       cinfo->dest = (jpeg_destination_mgr *)dest;
+
+       cinfo->input_components = 3;
+       jpeg_set_defaults(cinfo);
+       jpeg_set_quality(cinfo, quality, /*force_baseline=*/false);
+
+       cinfo->image_width = width;
+       cinfo->image_height = height;
+       cinfo->raw_data_in = true;
+       jpeg_set_colorspace(cinfo, JCS_YCbCr);
+       cinfo->comp_info[0].h_samp_factor = 2;
+       cinfo->comp_info[0].v_samp_factor = 1;
+       cinfo->comp_info[1].h_samp_factor = 1;
+       cinfo->comp_info[1].v_samp_factor = 1;
+       cinfo->comp_info[2].h_samp_factor = 1;
+       cinfo->comp_info[2].v_samp_factor = 1;
+       cinfo->CCIR601_sampling = true;  // Seems to be mostly ignored by libjpeg, though.
+       jpeg_start_compress(cinfo, true);
+}
+
+vector<uint8_t> MJPEGEncoder::get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo)
+{
+       VectorDestinationManager dest;
+       init_jpeg_422(width, height, &dest, cinfo);
+
+       // Make a dummy black image; there's seemingly no other easy way of
+       // making libjpeg outputting all of its headers.
+       JSAMPROW yptr[8], cbptr[8], crptr[8];
+       JSAMPARRAY data[3] = { yptr, cbptr, crptr };
+       memset(tmp_y, 0, 4096);
+       memset(tmp_cb, 0, 4096);
+       memset(tmp_cr, 0, 4096);
+       for (unsigned yy = 0; yy < 8; ++yy) {
+               yptr[yy] = tmp_y;
+               cbptr[yy] = tmp_cb;
+               crptr[yy] = tmp_cr;
+       }
+       for (unsigned y = 0; y < height; y += 8) {
+               jpeg_write_raw_data(cinfo, data, /*num_lines=*/8);
+       }
+       jpeg_finish_compress(cinfo);
+
+       // We're only interested in the header, not the data after it.
+       dest.term_destination();
+       for (size_t i = 0; i < dest.dest.size() - 1; ++i) {
+               if (dest.dest[i] == 0xff && dest.dest[i + 1] == 0xda) {  // Start of scan (SOS).
+                       unsigned len = dest.dest[i + 2] * 256 + dest.dest[i + 3];
+                       dest.dest.resize(i + len + 2);
+                       break;
+               }
+       }
+
+       return dest.dest;
+}
+
+MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, unsigned height)
+{
+       pair<unsigned, unsigned> key(width, height);
+       if (va_data_for_resolution.count(key)) {
+               return va_data_for_resolution[key];
+       }
+
+       // Use libjpeg to generate a header and set sane defaults for e.g.
+       // quantization tables. Then do the actual encode with VA-API.
+       jpeg_compress_struct cinfo;
+       vector<uint8_t> jpeg_header = get_jpeg_header(width, height, &cinfo);
+
+       // Picture parameters.
+       VAEncPictureParameterBufferJPEG pic_param;
+       memset(&pic_param, 0, sizeof(pic_param));
+       pic_param.reconstructed_picture = VA_INVALID_ID;
+       pic_param.picture_width = cinfo.image_width;
+       pic_param.picture_height = cinfo.image_height;
+       for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
+               const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
+               pic_param.component_id[component_idx] = comp->component_id;
+               pic_param.quantiser_table_selector[component_idx] = comp->quant_tbl_no;
+       }
+       pic_param.num_components = cinfo.num_components;
+       pic_param.num_scan = 1;
+       pic_param.sample_bit_depth = 8;
+       pic_param.coded_buf = VA_INVALID_ID;  // To be filled out by caller.
+       pic_param.pic_flags.bits.huffman = 1;
+       pic_param.quality = 50;  // Don't scale the given quantization matrices. (See gen8_mfc_jpeg_fqm_state)
+
+       // Quantization matrices.
+       VAQMatrixBufferJPEG q;
+       memset(&q, 0, sizeof(q));
+
+       q.load_lum_quantiser_matrix = true;
+       q.load_chroma_quantiser_matrix = true;
+       for (int quant_tbl_idx = 0; quant_tbl_idx < min(4, NUM_QUANT_TBLS); ++quant_tbl_idx) {
+               const JQUANT_TBL *qtbl = cinfo.quant_tbl_ptrs[quant_tbl_idx];
+               assert((qtbl == nullptr) == (quant_tbl_idx >= 2));
+               if (qtbl == nullptr) continue;
+
+               uint8_t *qmatrix = (quant_tbl_idx == 0) ? q.lum_quantiser_matrix : q.chroma_quantiser_matrix;
+               for (int i = 0; i < 64; ++i) {
+                       if (qtbl->quantval[i] > 255) {
+                               fprintf(stderr, "Baseline JPEG only!\n");
+                               abort();
+                       }
+                       qmatrix[i] = qtbl->quantval[jpeg_natural_order[i]];
+               }
+       }
+
+       // Huffman tables (arithmetic is not supported).
+       VAHuffmanTableBufferJPEGBaseline huff;
+       memset(&huff, 0, sizeof(huff));
+
+       for (int huff_tbl_idx = 0; huff_tbl_idx < min(2, NUM_HUFF_TBLS); ++huff_tbl_idx) {
+               const JHUFF_TBL *ac_hufftbl = cinfo.ac_huff_tbl_ptrs[huff_tbl_idx];
+               const JHUFF_TBL *dc_hufftbl = cinfo.dc_huff_tbl_ptrs[huff_tbl_idx];
+               if (ac_hufftbl == nullptr) {
+                       assert(dc_hufftbl == nullptr);
+                       huff.load_huffman_table[huff_tbl_idx] = 0;
+               } else {
+                       assert(dc_hufftbl != nullptr);
+                       huff.load_huffman_table[huff_tbl_idx] = 1;
+
+                       for (int i = 0; i < 16; ++i) {
+                               huff.huffman_table[huff_tbl_idx].num_dc_codes[i] = dc_hufftbl->bits[i + 1];
+                       }
+                       for (int i = 0; i < 12; ++i) {
+                               huff.huffman_table[huff_tbl_idx].dc_values[i] = dc_hufftbl->huffval[i];
+                       }
+                       for (int i = 0; i < 16; ++i) {
+                               huff.huffman_table[huff_tbl_idx].num_ac_codes[i] = ac_hufftbl->bits[i + 1];
+                       }
+                       for (int i = 0; i < 162; ++i) {
+                               huff.huffman_table[huff_tbl_idx].ac_values[i] = ac_hufftbl->huffval[i];
+                       }
+               }
+       }
+
+       // Slice parameters (metadata about the slice).
+       VAEncSliceParameterBufferJPEG parms;
+       memset(&parms, 0, sizeof(parms));
+       for (int component_idx = 0; component_idx < cinfo.num_components; ++component_idx) {
+               const jpeg_component_info *comp = &cinfo.comp_info[component_idx];
+               parms.components[component_idx].component_selector = comp->component_id;
+               parms.components[component_idx].dc_table_selector = comp->dc_tbl_no;
+               parms.components[component_idx].ac_table_selector = comp->ac_tbl_no;
+               if (parms.components[component_idx].dc_table_selector > 1 ||
+                   parms.components[component_idx].ac_table_selector > 1) {
+                       fprintf(stderr, "Uses too many Huffman tables\n");
+                       abort();
+               }
+       }
+       parms.num_components = cinfo.num_components;
+       parms.restart_interval = cinfo.restart_interval;
+
+       jpeg_destroy_compress(&cinfo);
+
+       VAData ret;
+       ret.jpeg_header = move(jpeg_header);
+       ret.pic_param = pic_param;
+       ret.q = q;
+       ret.huff = huff;
+       ret.parms = parms;
+       va_data_for_resolution[key] = ret;
+       return ret;
+}
+
+vector<uint8_t> MJPEGEncoder::encode_jpeg(const QueuedFrame &qf)
+{
+       if (va_dpy != nullptr) {
+               return encode_jpeg_va(qf);
+       } else {
+               return encode_jpeg_libjpeg(qf);
+       }
+}
+
+vector<uint8_t> MJPEGEncoder::encode_jpeg_va(const QueuedFrame &qf)
+{
+       unsigned width = qf.video_format.width;
+       unsigned height = qf.video_format.height;
+
+       VAResources resources = get_va_resources(width, height);
+       ReleaseVAResources release(this, resources);
+
+       VAData va_data = get_va_data_for_resolution(width, height);
+       va_data.pic_param.coded_buf = resources.data_buffer;
+
+       VABufferID pic_param_buffer;
+       VAStatus va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPictureParameterBufferType, sizeof(va_data.pic_param), 1, &va_data.pic_param, &pic_param_buffer);
+       CHECK_VASTATUS(va_status, "vaCreateBuffer");
+       VABufferDestroyer destroy_pic_param(va_dpy->va_dpy, pic_param_buffer);
+
+       VABufferID q_buffer;
+       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAQMatrixBufferType, sizeof(va_data.q), 1, &va_data.q, &q_buffer);
+       CHECK_VASTATUS(va_status, "vaCreateBuffer");
+       VABufferDestroyer destroy_iq(va_dpy->va_dpy, q_buffer);
+
+       VABufferID huff_buffer;
+       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAHuffmanTableBufferType, sizeof(va_data.huff), 1, &va_data.huff, &huff_buffer);
+       CHECK_VASTATUS(va_status, "vaCreateBuffer");
+       VABufferDestroyer destroy_huff(va_dpy->va_dpy, huff_buffer);
+
+       VABufferID slice_param_buffer;
+       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncSliceParameterBufferType, sizeof(va_data.parms), 1, &va_data.parms, &slice_param_buffer);
+       CHECK_VASTATUS(va_status, "vaCreateBuffer");
+       VABufferDestroyer destroy_slice_param(va_dpy->va_dpy, slice_param_buffer);
+
+       VAImage image;
+       va_status = vaDeriveImage(va_dpy->va_dpy, resources.surface, &image);
+       CHECK_VASTATUS(va_status, "vaDeriveImage");
+
+       // Upload the pixel data.
+       uint8_t *surface_p = nullptr;
+       vaMapBuffer(va_dpy->va_dpy, image.buf, (void **)&surface_p);
+
+       size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
+       size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
+
+       {
+               const uint8_t *src = qf.frame->data_copy + field_start;
+               uint8_t *dst = (unsigned char *)surface_p + image.offsets[0];
+               memcpy_with_pitch(dst, src, qf.video_format.width * 2, image.pitches[0], qf.video_format.height);
+       }
+
+       va_status = vaUnmapBuffer(va_dpy->va_dpy, image.buf);
+       CHECK_VASTATUS(va_status, "vaUnmapBuffer");
+       va_status = vaDestroyImage(va_dpy->va_dpy, image.image_id);
+       CHECK_VASTATUS(va_status, "vaDestroyImage");
+
+       // Finally, stick in the JPEG header.
+       VAEncPackedHeaderParameterBuffer header_parm;
+       header_parm.type = VAEncPackedHeaderRawData;
+       header_parm.bit_length = 8 * va_data.jpeg_header.size();
+
+       VABufferID header_parm_buffer;
+       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderParameterBufferType, sizeof(header_parm), 1, &header_parm, &header_parm_buffer);
+       CHECK_VASTATUS(va_status, "vaCreateBuffer");
+       VABufferDestroyer destroy_header(va_dpy->va_dpy, header_parm_buffer);
+
+       VABufferID header_data_buffer;
+       va_status = vaCreateBuffer(va_dpy->va_dpy, config_id, VAEncPackedHeaderDataBufferType, va_data.jpeg_header.size(), 1, va_data.jpeg_header.data(), &header_data_buffer);
+       CHECK_VASTATUS(va_status, "vaCreateBuffer");
+       VABufferDestroyer destroy_header_data(va_dpy->va_dpy, header_data_buffer);
+
+       va_status = vaBeginPicture(va_dpy->va_dpy, resources.context, resources.surface);
+       CHECK_VASTATUS(va_status, "vaBeginPicture");
+       va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &pic_param_buffer, 1);
+       CHECK_VASTATUS(va_status, "vaRenderPicture(pic_param)");
+       va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &q_buffer, 1);
+       CHECK_VASTATUS(va_status, "vaRenderPicture(q)");
+       va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &huff_buffer, 1);
+       CHECK_VASTATUS(va_status, "vaRenderPicture(huff)");
+       va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &slice_param_buffer, 1);
+       CHECK_VASTATUS(va_status, "vaRenderPicture(slice_param)");
+       va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_parm_buffer, 1);
+       CHECK_VASTATUS(va_status, "vaRenderPicture(header_parm)");
+       va_status = vaRenderPicture(va_dpy->va_dpy, resources.context, &header_data_buffer, 1);
+       CHECK_VASTATUS(va_status, "vaRenderPicture(header_data)");
+       va_status = vaEndPicture(va_dpy->va_dpy, resources.context);
+       CHECK_VASTATUS(va_status, "vaEndPicture");
+
+       va_status = vaSyncSurface(va_dpy->va_dpy, resources.surface);
+       CHECK_VASTATUS(va_status, "vaSyncSurface");
+
+       VACodedBufferSegment *segment;
+       va_status = vaMapBuffer(va_dpy->va_dpy, resources.data_buffer, (void **)&segment);
+       CHECK_VASTATUS(va_status, "vaMapBuffer");
+
+       const char *coded_buf = reinterpret_cast<char *>(segment->buf);
+       vector<uint8_t> jpeg(coded_buf, coded_buf + segment->size); 
+
+       va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.data_buffer);
+       CHECK_VASTATUS(va_status, "vaUnmapBuffer");
+
+       return jpeg;
+}
+
+vector<uint8_t> MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf)
+{
+       unsigned width = qf.video_format.width;
+       unsigned height = qf.video_format.height;
+
+       VectorDestinationManager dest;
+       jpeg_compress_struct cinfo;
+       init_jpeg_422(width, height, &dest, &cinfo);
+
+       size_t field_start_line = qf.video_format.extra_lines_top;  // No interlacing support.
+       size_t field_start = qf.cbcr_offset * 2 + qf.video_format.width * field_start_line * 2;
+
+       JSAMPROW yptr[8], cbptr[8], crptr[8];
+       JSAMPARRAY data[3] = { yptr, cbptr, crptr };
+       for (unsigned y = 0; y < qf.video_format.height; y += 8) {
+               const uint8_t *src = qf.frame->data_copy + field_start + y * qf.video_format.width * 2;
+
+               memcpy_interleaved(tmp_y, tmp_cbcr, src, qf.video_format.width * 8 * 2);
+               memcpy_interleaved(tmp_cb, tmp_cr, tmp_cbcr, qf.video_format.width * 8);
+               for (unsigned yy = 0; yy < 8; ++yy) {
+                       yptr[yy] = tmp_y + yy * width;
+                       cbptr[yy] = tmp_cb + yy * width / 2;
+                       crptr[yy] = tmp_cr + yy * width / 2;
+               }
+               jpeg_write_raw_data(&cinfo, data, /*num_lines=*/8);
+       }
+       jpeg_finish_compress(&cinfo);
+
+       return dest.dest;
+}
diff --git a/mjpeg_encoder.h b/mjpeg_encoder.h
new file mode 100644 (file)
index 0000000..ab8632a
--- /dev/null
@@ -0,0 +1,121 @@
+#ifndef _MJPEG_ENCODER_H
+#define _MJPEG_ENCODER_H 1
+
+#include "ffmpeg_raii.h"
+#include "ref_counted_frame.h"
+
+extern "C" {
+
+#include <libavformat/avio.h>
+
+}  // extern "C"
+
+#include <atomic>
+#include <bmusb/bmusb.h>
+#include <condition_variable>
+#include <list>
+#include <mutex>
+#include <queue>
+#include <stdint.h>
+#include <string>
+#include <thread>
+
+#include <va/va.h>
+
+class HTTPD;
+struct jpeg_compress_struct;
+struct VADisplayWithCleanup;
+struct VectorDestinationManager;
+
+class MJPEGEncoder {
+public:
+       MJPEGEncoder(HTTPD *httpd, const std::string &va_display);
+       void stop();
+       void upload_frame(int64_t pts, unsigned card_index, RefCountedFrame frame, const bmusb::VideoFormat &video_format, size_t y_offset, size_t cbcr_offset);
+
+private:
+       static constexpr int quality = 90;
+
+       struct QueuedFrame {
+               int64_t pts;
+               unsigned card_index;
+               RefCountedFrame frame;
+               bmusb::VideoFormat video_format;
+               size_t y_offset, cbcr_offset;
+       };
+
+       void encoder_thread_func();
+       std::vector<uint8_t> encode_jpeg(const QueuedFrame &qf);
+       std::vector<uint8_t> encode_jpeg_va(const QueuedFrame &qf);
+       std::vector<uint8_t> encode_jpeg_libjpeg(const QueuedFrame &qf);
+       void init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo);
+       std::vector<uint8_t> get_jpeg_header(unsigned width, unsigned height, jpeg_compress_struct *cinfo);
+
+       static int write_packet2_thunk(void *opaque, uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time);
+       int write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType type, int64_t time);
+
+       std::thread encoder_thread;
+
+       std::mutex mu;
+       std::queue<QueuedFrame> frames_to_be_encoded;  // Under mu.
+       std::condition_variable any_frames_to_be_encoded;
+
+       std::queue<QueuedFrame> frames_encoding;  // Under mu.
+       std::condition_variable any_frames_encoding;
+
+       AVFormatContextWithCloser avctx;
+       HTTPD *httpd;
+       std::string mux_header;
+       std::atomic<bool> should_quit{false};
+       bool running = false;
+
+       std::unique_ptr<VADisplayWithCleanup> va_dpy;
+       VAConfigID config_id;
+
+       struct VAData {
+               std::vector<uint8_t> jpeg_header;
+               VAEncPictureParameterBufferJPEG pic_param;
+               VAQMatrixBufferJPEG q;
+               VAHuffmanTableBufferJPEGBaseline huff;
+               VAEncSliceParameterBufferJPEG parms;
+       };
+       std::map<std::pair<unsigned, unsigned>, VAData> va_data_for_resolution;
+       VAData get_va_data_for_resolution(unsigned width, unsigned height);
+
+       struct VAResources {
+               unsigned width, height;
+               VASurfaceID surface;
+               VAContextID context;
+               VABufferID data_buffer;
+       };
+       std::list<VAResources> va_resources_freelist;
+       std::mutex va_resources_mutex;
+       VAResources get_va_resources(unsigned width, unsigned height);
+       void release_va_resources(VAResources resources);
+
+       // RAII wrapper to release VAResources on return (even on error).
+       class ReleaseVAResources {
+       public:
+               ReleaseVAResources(MJPEGEncoder *mjpeg, const VAResources &resources)
+                       : mjpeg(mjpeg), resources(resources) {}
+               ~ReleaseVAResources()
+               {
+                       if (!committed) {
+                               mjpeg->release_va_resources(resources);
+                       }
+               }
+
+               void commit() { committed = true; }
+
+       private:
+               MJPEGEncoder * const mjpeg;
+               const VAResources &resources;
+               bool committed = false;
+       };
+
+       static std::unique_ptr<VADisplayWithCleanup> try_open_va(const std::string &va_display, std::string *error, VAConfigID *config_id);
+
+       uint8_t *tmp_y, *tmp_cbcr, *tmp_cb, *tmp_cr;  // Private to the encoder thread.
+};
+
+#endif  // !defined(_MJPEG_ENCODER_H)
index ea17f1259fde4076320c9ebdb37da6c714a6c5f3..c8687025e2f2f1d5dc0d193a07035b01e272f970 100644 (file)
@@ -52,6 +52,7 @@ void PBOFrameAllocator::init_frame(size_t frame_idx, size_t frame_size, GLuint w
        Frame frame;
        frame.data = (uint8_t *)glMapBufferRange(buffer, 0, frame_size, permissions | map_bits | GL_MAP_PERSISTENT_BIT);
        frame.data2 = frame.data + frame_size / 2;
+       frame.data_copy = new uint8_t[frame_size];
        check_error();
        frame.size = frame_size;
        frame.userdata = &userdata[frame_idx];
@@ -214,6 +215,8 @@ PBOFrameAllocator::~PBOFrameAllocator()
 
 void PBOFrameAllocator::destroy_frame(Frame *frame)
 {
+       delete[] frame->data_copy;
+
        GLuint pbo = ((Userdata *)frame->userdata)->pbo;
        glBindBuffer(buffer, pbo);
        check_error();
index 5200ce61b4911ab4e18692da0bfc17b483ae548d..bca7ffd91983e73ed819d7889c9925cd58acae2f 100644 (file)
@@ -1939,8 +1939,6 @@ void QuickSyncEncoderImpl::add_packet_for_uncompressed_frame(int64_t pts, int64_
        stream_mux->add_packet(pkt, pts, pts);
 }
 
-namespace {
-
 void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_t dst_pitch, size_t height)
 {
        if (src_width == dst_pitch) {
@@ -1954,8 +1952,6 @@ void memcpy_with_pitch(uint8_t *dst, const uint8_t *src, size_t src_width, size_
        }
 }
 
-}  // namespace
-
 void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration)
 {
        // Wait for the GPU to be done with the frame.
index 7645b99b6b34edf5e8afd83b91a6a9e7f981ba10..0317b6af0393723bb1ca38ac0100947d8c800c85 100644 (file)
@@ -20,6 +20,7 @@
 #include "timebase.h"
 #include "print_latency.h"
 #include "ref_counted_gl_sync.h"
+#include "va_display_with_cleanup.h"
 
 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
@@ -39,16 +40,6 @@ class DiskSpaceEstimator;
 class QSurface;
 class X264Encoder;
 
-struct VADisplayWithCleanup {
-       ~VADisplayWithCleanup();
-
-       VADisplay va_dpy;
-       Display *x11_display = nullptr;
-       bool can_use_zerocopy = true;
-       int drm_fd = -1;
-};
-std::unique_ptr<VADisplayWithCleanup> va_open_display(const std::string &va_display);  // Can return nullptr on failure.
-
 class QuickSyncEncoderImpl {
 public:
        QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const std::string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator);
diff --git a/va_display_with_cleanup.h b/va_display_with_cleanup.h
new file mode 100644 (file)
index 0000000..3f9b1c5
--- /dev/null
@@ -0,0 +1,19 @@
+#ifndef _VA_DISPLAY_WITH_CLEANUP
+#define _VA_DISPLAY_WITH_CLEANUP 1
+
+#include <va/va.h>
+#include <X11/Xlib.h>
+
+#include <memory>
+
+struct VADisplayWithCleanup {
+       ~VADisplayWithCleanup();
+
+       VADisplay va_dpy;
+       Display *x11_display = nullptr;
+       bool can_use_zerocopy = true;
+       int drm_fd = -1;
+};
+std::unique_ptr<VADisplayWithCleanup> va_open_display(const std::string &va_display);  // Can return nullptr on failure.
+
+#endif  // !defined(_VA_DISPLAY_WITH_CLEANUP)
index 6344b8c6a04a89e0fe6e9e312999a2ad449ab93f..2b8fcd5edec9e0a9ea3a79aed79d39d02494375d 100644 (file)
@@ -216,9 +216,9 @@ int VideoEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType t
 
        if (type == AVIO_DATA_MARKER_HEADER) {
                stream_mux_header.append((char *)buf, buf_size);
-               httpd->set_header(stream_mux_header);
+               httpd->set_header(HTTPD::MAIN_STREAM, stream_mux_header);
        } else {
-               httpd->add_data((char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
+               httpd->add_data(HTTPD::MAIN_STREAM, (char *)buf, buf_size, type == AVIO_DATA_MARKER_SYNC_POINT, time, AVRational{ AV_TIME_BASE, 1 });
        }
        return buf_size;
 }