From c913b0254382c2a1bd7410d720e45dfc63223c9c Mon Sep 17 00:00:00 2001 From: "Steinar H. Gunderson" Date: Sun, 11 Nov 2018 17:58:24 +0100 Subject: [PATCH] Make VA-API encoded MJPEGs be encoded asynchronously. This also frees up the mixer thread, which was stuck locking on the mutex whenever something was under encoding. --- nageru/mjpeg_encoder.cpp | 120 +++++++++++++++++++++++++-------------- nageru/mjpeg_encoder.h | 93 +++++++++++++++++++----------- 2 files changed, 139 insertions(+), 74 deletions(-) diff --git a/nageru/mjpeg_encoder.cpp b/nageru/mjpeg_encoder.cpp index d82107f..9ddb9ff 100644 --- a/nageru/mjpeg_encoder.cpp +++ b/nageru/mjpeg_encoder.cpp @@ -119,8 +119,6 @@ int MJPEGEncoder::write_packet2(uint8_t *buf, int buf_size, AVIODataMarkerType t 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. @@ -174,6 +172,11 @@ MJPEGEncoder::MJPEGEncoder(HTTPD *httpd, const string &va_display) fprintf(stderr, "Could not initialize VA-API for MJPEG encoding: %s. JPEGs will be encoded in software if needed.\n", error.c_str()); } + encoder_thread = thread(&MJPEGEncoder::encoder_thread_func, this); + if (va_dpy != nullptr) { + va_receiver_thread = thread(&MJPEGEncoder::va_receiver_thread_func, this); + } + running = true; } @@ -191,6 +194,9 @@ void MJPEGEncoder::stop() should_quit = true; any_frames_to_be_encoded.notify_all(); encoder_thread.join(); + if (va_dpy != nullptr) { + va_receiver_thread.join(); + } } unique_ptr MJPEGEncoder::try_open_va(const string &va_display, string *error, VAConfigID *config_id) @@ -258,6 +264,10 @@ void MJPEGEncoder::upload_frame(int64_t pts, unsigned card_index, RefCountedFram } lock_guard lock(mu); + if (frames_to_be_encoded.size() + frames_encoding.size() > 10) { + fprintf(stderr, "WARNING: MJPEG encoding doesn't keep up, discarding frame.\n"); + return; + } frames_to_be_encoded.push(QueuedFrame{ pts, card_index, frame, video_format, y_offset, cbcr_offset }); any_frames_to_be_encoded.notify_all(); } @@ -270,27 +280,23 @@ void MJPEGEncoder::encoder_thread_func() posix_memalign((void **)&tmp_cb, 4096, 4096 * 8); posix_memalign((void **)&tmp_cr, 4096, 4096 * 8); - unique_lock lock(mu); for (;;) { - any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; }); - if (should_quit) break; - QueuedFrame qf = move(frames_to_be_encoded.front()); - frames_to_be_encoded.pop(); - - vector 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); + QueuedFrame qf; + { + unique_lock lock(mu); + any_frames_to_be_encoded.wait(lock, [this] { return !frames_to_be_encoded.empty() || should_quit; }); + if (should_quit) break; + qf = move(frames_to_be_encoded.front()); + frames_to_be_encoded.pop(); + } + + if (va_dpy != nullptr) { + // Will call back in the receiver thread. + encode_jpeg_va(move(qf)); + } else { + // Encode synchronously, in the same thread. + vector jpeg = encode_jpeg_libjpeg(qf); + write_mjpeg_packet(qf.pts, qf.card_index, jpeg); } } @@ -300,6 +306,23 @@ void MJPEGEncoder::encoder_thread_func() free(tmp_cr); } +void MJPEGEncoder::write_mjpeg_packet(int64_t pts, unsigned card_index, const vector &jpeg) +{ + AVPacket pkt; + memset(&pkt, 0, sizeof(pkt)); + pkt.buf = nullptr; + pkt.data = const_cast(&jpeg[0]); + pkt.size = jpeg.size(); + pkt.stream_index = card_index; + pkt.flags = AV_PKT_FLAG_KEY; + pkt.pts = pkt.dts = 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) @@ -549,16 +572,7 @@ MJPEGEncoder::VAData MJPEGEncoder::get_va_data_for_resolution(unsigned width, un return ret; } -vector MJPEGEncoder::encode_jpeg(const QueuedFrame &qf) -{ - if (va_dpy != nullptr) { - return encode_jpeg_va(qf); - } else { - return encode_jpeg_libjpeg(qf); - } -} - -vector MJPEGEncoder::encode_jpeg_va(const QueuedFrame &qf) +void MJPEGEncoder::encode_jpeg_va(QueuedFrame &&qf) { unsigned width = qf.video_format.width; unsigned height = qf.video_format.height; @@ -643,20 +657,42 @@ vector MJPEGEncoder::encode_jpeg_va(const QueuedFrame &qf) 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"); + qf.resources = move(resources); + qf.resource_releaser = move(release); - VACodedBufferSegment *segment; - va_status = vaMapBuffer(va_dpy->va_dpy, resources.data_buffer, (void **)&segment); - CHECK_VASTATUS(va_status, "vaMapBuffer"); + lock_guard lock(mu); + frames_encoding.push(move(qf)); + any_frames_encoding.notify_all(); +} - const char *coded_buf = reinterpret_cast(segment->buf); - vector jpeg(coded_buf, coded_buf + segment->size); +void MJPEGEncoder::va_receiver_thread_func() +{ + pthread_setname_np(pthread_self(), "MJPEG_Receive"); + for (;;) { + QueuedFrame qf; + { + unique_lock lock(mu); + any_frames_encoding.wait(lock, [this] { return !frames_encoding.empty() || should_quit; }); + if (should_quit) return; + qf = move(frames_encoding.front()); + frames_encoding.pop(); + } - va_status = vaUnmapBuffer(va_dpy->va_dpy, resources.data_buffer); - CHECK_VASTATUS(va_status, "vaUnmapBuffer"); + VAStatus va_status = vaSyncSurface(va_dpy->va_dpy, qf.resources.surface); + CHECK_VASTATUS(va_status, "vaSyncSurface"); - return jpeg; + VACodedBufferSegment *segment; + va_status = vaMapBuffer(va_dpy->va_dpy, qf.resources.data_buffer, (void **)&segment); + CHECK_VASTATUS(va_status, "vaMapBuffer"); + + const char *coded_buf = reinterpret_cast(segment->buf); + vector jpeg(coded_buf, coded_buf + segment->size); + + va_status = vaUnmapBuffer(va_dpy->va_dpy, qf.resources.data_buffer); + CHECK_VASTATUS(va_status, "vaUnmapBuffer"); + + write_mjpeg_packet(qf.pts, qf.card_index, jpeg); + } } vector MJPEGEncoder::encode_jpeg_libjpeg(const QueuedFrame &qf) diff --git a/nageru/mjpeg_encoder.h b/nageru/mjpeg_encoder.h index f3c9d99..3630d9f 100644 --- a/nageru/mjpeg_encoder.h +++ b/nageru/mjpeg_encoder.h @@ -37,31 +37,86 @@ public: private: static constexpr int quality = 90; + struct VAResources { + unsigned width, height; + VASurfaceID surface; + VAContextID context; + VABufferID data_buffer; + }; + + // RAII wrapper to release VAResources on return (even on error). + class ReleaseVAResources { + public: + ReleaseVAResources() : committed(true) {} + + ReleaseVAResources(MJPEGEncoder *mjpeg, const VAResources &resources) + : mjpeg(mjpeg), resources(resources) {} + + ReleaseVAResources(ReleaseVAResources &) = delete; + + ReleaseVAResources(ReleaseVAResources &&other) + : mjpeg(other.mjpeg), resources(other.resources), committed(other.committed) { + other.commit(); + } + + ReleaseVAResources &operator= (ReleaseVAResources &) = delete; + + ReleaseVAResources &operator= (ReleaseVAResources &&other) { + if (!committed) { + mjpeg->release_va_resources(resources); + } + mjpeg = other.mjpeg; + resources = std::move(other.resources); + committed = other.committed; + other.commit(); + return *this; + } + + ~ReleaseVAResources() + { + if (!committed) { + mjpeg->release_va_resources(resources); + } + } + + void commit() { committed = true; } + + private: + MJPEGEncoder *mjpeg = nullptr; + VAResources resources; + bool committed = false; + }; + struct QueuedFrame { int64_t pts; unsigned card_index; RefCountedFrame frame; bmusb::VideoFormat video_format; size_t y_offset, cbcr_offset; + + // Only for frames in the process of being encoded by VA-API. + VAResources resources; + ReleaseVAResources resource_releaser; }; void encoder_thread_func(); - std::vector encode_jpeg(const QueuedFrame &qf); - std::vector encode_jpeg_va(const QueuedFrame &qf); + void va_receiver_thread_func(); + void encode_jpeg_va(QueuedFrame &&qf); std::vector encode_jpeg_libjpeg(const QueuedFrame &qf); + void write_mjpeg_packet(int64_t pts, unsigned card_index, const std::vector &jpeg); void init_jpeg_422(unsigned width, unsigned height, VectorDestinationManager *dest, jpeg_compress_struct *cinfo); std::vector 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::thread encoder_thread, va_receiver_thread; std::mutex mu; std::queue frames_to_be_encoded; // Under mu. - std::condition_variable any_frames_to_be_encoded; + std::condition_variable any_frames_to_be_encoded; // Governs changes in both frames_to_be_encoded and frames_under_encoding - std::queue frames_encoding; // Under mu. + std::queue frames_encoding; // Under mu. Used for VA-API only. std::condition_variable any_frames_encoding; AVFormatContextWithCloser avctx; @@ -83,40 +138,14 @@ private: std::map, 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 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 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. + uint8_t *tmp_y, *tmp_cbcr, *tmp_cb, *tmp_cr; // Private to the encoder thread. Used by the libjpeg backend only. }; #endif // !defined(_MJPEG_ENCODER_H) -- 2.39.2