X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=quicksync_encoder.cpp;h=7edb3cc03fccc67258f5ecc4066f7f4ab14a2d33;hb=dc7db20e764306a33dbf80cb50da7822a79cfda2;hp=cc4020f69508ffe9d5706be7a88be8b7cc240da9;hpb=c93d4d8a7e35f5e53868a938f1a216792c391598;p=nageru diff --git a/quicksync_encoder.cpp b/quicksync_encoder.cpp index cc4020f..7edb3cc 100644 --- a/quicksync_encoder.cpp +++ b/quicksync_encoder.cpp @@ -18,9 +18,11 @@ #include #include #include +#include #include #include #include +#include #include #include #include @@ -42,14 +44,19 @@ extern "C" { #include "audio_encoder.h" #include "context.h" #include "defs.h" +#include "disk_space_estimator.h" #include "ffmpeg_raii.h" #include "flags.h" #include "mux.h" +#include "print_latency.h" +#include "quicksync_encoder_impl.h" #include "ref_counted_frame.h" #include "timebase.h" #include "x264_encoder.h" using namespace std; +using namespace std::chrono; +using namespace std::placeholders; class QOpenGLContext; class QSurface; @@ -91,9 +98,6 @@ class QSurface; #define PROFILE_IDC_HIGH 100 #define BITSTREAM_ALLOCATE_STEPPING 4096 -#define SURFACE_NUM 16 /* 16 surfaces for source YUV */ -#define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM -#define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM static constexpr unsigned int MaxFrameNum = (2<<16); static constexpr unsigned int MaxPicOrderCntLsb = (2<<8); @@ -112,61 +116,8 @@ static constexpr int rc_default_modes[] = { // Priority list of modes. #define SRC_SURFACE_FREE 0 #define SRC_SURFACE_IN_ENCODING 1 -struct __bitstream { - unsigned int *buffer; - int bit_offset; - int max_size_in_dword; -}; -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); - - struct Frame { - int64_t pts, duration; - uint8_t *data; - - // Invert to get the smallest pts first. - bool operator< (const Frame &other) const { return pts > other.pts; } - }; - - // 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). - Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data); - - // The same as reorder_frame, but without inserting anything. Used to empty the queue. - Frame get_first_frame(); - - bool empty() const { return frames.empty(); } - -private: - unsigned queue_length; - int width, height; - - priority_queue frames; - stack 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> owner; -}; - FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height) : queue_length(queue_length), width(width), height(height) { @@ -176,22 +127,22 @@ FrameReorderer::FrameReorderer(unsigned queue_length, int width, int height) } } -FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data) +FrameReorderer::Frame FrameReorderer::reorder_frame(int64_t pts, int64_t duration, uint8_t *data, const ReceivedTimestamps &received_ts) { if (queue_length == 0) { - return Frame{pts, duration, data}; + return Frame{pts, duration, data, received_ts}; } assert(!freelist.empty()); uint8_t *storage = freelist.top(); freelist.pop(); memcpy(storage, data, width * height * 2); - frames.push(Frame{pts, duration, storage}); + frames.push(Frame{pts, duration, storage, received_ts}); if (frames.size() >= queue_length) { return get_first_frame(); } else { - return Frame{-1, -1, nullptr}; + return Frame{-1, -1, nullptr, steady_clock::time_point::min(), steady_clock::time_point::min()}; } } @@ -204,158 +155,6 @@ FrameReorderer::Frame FrameReorderer::get_first_frame() return storage; } -class QuickSyncEncoderImpl { -public: - QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder); - ~QuickSyncEncoderImpl(); - void add_audio(int64_t pts, vector audio); - bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex); - RefCountedGLsync end_frame(int64_t pts, int64_t duration, const vector &input_frames); - void shutdown(); - void release_gl_resources(); - void set_stream_mux(Mux *mux) - { - stream_mux = mux; - } - - // So we never get negative dts. - int64_t global_delay() const { - return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS); - } - -private: - struct storage_task { - unsigned long long display_order; - int frame_type; - vector audio; - int64_t pts, dts, duration; - }; - struct PendingFrame { - RefCountedGLsync fence; - vector input_frames; - int64_t pts, duration; - }; - - void open_output_file(const std::string &filename); - 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, int64_t duration, 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, int64_t duration); - void storage_task_thread(); - void storage_task_enqueue(storage_task task); - void save_codeddata(storage_task task); - int render_packedsequence(); - int render_packedpicture(); - void render_packedslice(); - int render_sequence(); - int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num); - void sps_rbsp(bitstream *bs); - void pps_rbsp(bitstream *bs); - int build_packed_pic_buffer(unsigned char **header_buffer); - int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type); - void slice_header(bitstream *bs); - int build_packed_seq_buffer(unsigned char **header_buffer); - int build_packed_slice_buffer(unsigned char **header_buffer); - int init_va(const string &va_display); - int deinit_va(); - void enable_zerocopy_if_possible(); - VADisplay va_open_display(const string &va_display); - void va_close_display(VADisplay va_dpy); - int setup_encode(); - void release_encode(); - void update_ReferenceFrames(int frame_type); - int update_RefPicList(int frame_type); - - bool is_shutdown = false; - bool has_released_gl_resources = false; - bool use_zerocopy; - int drm_fd = -1; - - thread encode_thread, storage_thread; - - mutex storage_task_queue_mutex; - condition_variable storage_task_queue_changed; - int srcsurface_status[SURFACE_NUM]; // protected by storage_task_queue_mutex - queue storage_task_queue; // protected by storage_task_queue_mutex - bool storage_thread_should_quit = false; // protected by storage_task_queue_mutex - - mutex frame_queue_mutex; - condition_variable frame_queue_nonempty; - bool encode_thread_should_quit = false; // under frame_queue_mutex - - int current_storage_frame; - - map pending_video_frames; // under frame_queue_mutex - movit::ResourcePool *resource_pool; - QSurface *surface; - - unique_ptr file_audio_encoder; - - unique_ptr reorderer; - X264Encoder *x264_encoder; // nullptr if not using x264. - - Mux* stream_mux = nullptr; // To HTTP. - unique_ptr file_mux; // To local disk. - - Display *x11_display = nullptr; - - // Encoder parameters - VADisplay va_dpy; - VAProfile h264_profile = (VAProfile)~0; - VAConfigAttrib config_attrib[VAConfigAttribTypeMax]; - int config_attrib_num = 0, enc_packed_header_idx; - - struct GLSurface { - VASurfaceID src_surface, ref_surface; - VABufferID coded_buf; - - VAImage surface_image; - GLuint y_tex, cbcr_tex; - - // Only if use_zerocopy == true. - EGLImage y_egl_image, cbcr_egl_image; - - // Only if use_zerocopy == false. - GLuint pbo; - uint8_t *y_ptr, *cbcr_ptr; - size_t y_offset, cbcr_offset; - }; - GLSurface gl_surfaces[SURFACE_NUM]; - - VAConfigID config_id; - VAContextID context_id; - VAEncSequenceParameterBufferH264 seq_param; - VAEncPictureParameterBufferH264 pic_param; - VAEncSliceParameterBufferH264 slice_param; - VAPictureH264 CurrentCurrPic; - VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2]; - - // Static quality settings. - static constexpr unsigned int frame_bitrate = 15000000 / 60; // Doesn't really matter; only initial_qp does. - static constexpr unsigned int num_ref_frames = 2; - static constexpr int initial_qp = 15; - static constexpr int minimal_qp = 0; - static constexpr int intra_period = 30; - static constexpr int intra_idr_period = MAX_FPS; // About a second; more at lower frame rates. Not ideal. - - // Quality settings that are meant to be static, but might be overridden - // by the profile. - int constraint_set_flag = 0; - int h264_packedheader = 0; /* support pack header? */ - int h264_maxref = (1<<16|1); - int h264_entropy_mode = 1; /* cabac */ - int ip_period = 3; - - int rc_mode = -1; - unsigned int current_frame_num = 0; - unsigned int numShortTerm = 0; - - int frame_width; - int frame_height; - int frame_width_mbaligned; - int frame_height_mbaligned; -}; // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly, // but if we don't delete it here, we get leaks. The GStreamer implementation @@ -1625,6 +1424,10 @@ void QuickSyncEncoderImpl::save_codeddata(storage_task task) } vaUnmapBuffer(va_dpy, gl_surfaces[task.display_order % SURFACE_NUM].coded_buf); + static int frameno = 0; + print_latency("Current QuickSync latency (video inputs → disk mux):", + task.received_ts, (task.frame_type == FRAME_B), &frameno); + { // Add video. AVPacket pkt; @@ -1728,8 +1531,8 @@ int QuickSyncEncoderImpl::deinit_va() return 0; } -QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder) - : current_storage_frame(0), resource_pool(resource_pool), surface(surface), x264_encoder(x264_encoder), frame_width(width), frame_height(height) +QuickSyncEncoderImpl::QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator) + : current_storage_frame(0), resource_pool(resource_pool), surface(surface), x264_encoder(x264_encoder), frame_width(width), frame_height(height), disk_space_estimator(disk_space_estimator) { file_audio_encoder.reset(new AudioEncoder(AUDIO_OUTPUT_CODEC_NAME, DEFAULT_AUDIO_OUTPUT_BIT_RATE, oformat)); open_output_file(filename); @@ -1951,7 +1754,8 @@ void QuickSyncEncoderImpl::open_output_file(const std::string &filename) string video_extradata = ""; // FIXME: See other comment about global headers. AVCodecParametersWithDeleter audio_codecpar = file_audio_encoder->get_codec_parameters(); - file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE)); + file_mux.reset(new Mux(avctx, frame_width, frame_height, Mux::CODEC_H264, video_extradata, audio_codecpar.get(), TIMEBASE, + std::bind(&DiskSpaceEstimator::report_write, disk_space_estimator, filename, _1))); } void QuickSyncEncoderImpl::encode_thread_func() @@ -2028,7 +1832,7 @@ void QuickSyncEncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data); } else { assert(global_flags.x264_video_to_http); - x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data); + x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data, output_frame.received_ts); } } } @@ -2075,6 +1879,16 @@ void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame } while (sync_status == GL_TIMEOUT_EXPIRED); assert(sync_status != GL_WAIT_FAILED); + // Find min and max timestamp of all input frames that have a timestamp. + steady_clock::time_point min_ts = steady_clock::time_point::max(), max_ts = steady_clock::time_point::min(); + for (const RefCountedFrame &input_frame : frame.input_frames) { + if (input_frame && input_frame->received_timestamp > steady_clock::time_point::min()) { + min_ts = min(min_ts, input_frame->received_timestamp); + max_ts = max(max_ts, input_frame->received_timestamp); + } + } + const ReceivedTimestamps received_ts{ min_ts, max_ts }; + // Release back any input frames we needed to render this frame. frame.input_frames.clear(); @@ -2103,18 +1917,22 @@ void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame global_flags.x264_video_to_http) { // Add uncompressed video. (Note that pts == dts here.) // Delay needs to match audio. - FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast(surf->y_ptr)); + FrameReorderer::Frame output_frame = reorderer->reorder_frame(pts + global_delay(), duration, reinterpret_cast(surf->y_ptr), received_ts); if (output_frame.data != nullptr) { if (global_flags.uncompressed_video_to_http) { add_packet_for_uncompressed_frame(output_frame.pts, output_frame.duration, output_frame.data); } else { assert(global_flags.x264_video_to_http); - x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data); + x264_encoder->add_frame(output_frame.pts, output_frame.duration, output_frame.data, output_frame.received_ts); } } } } + static int frameno = 0; + print_latency("Current mixer latency (video inputs → ready for encode):", + received_ts, (frame_type == FRAME_B), &frameno); + va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id); CHECK_VASTATUS(va_status, "vaDestroyImage"); @@ -2150,14 +1968,15 @@ void QuickSyncEncoderImpl::encode_frame(QuickSyncEncoderImpl::PendingFrame frame tmp.pts = pts; tmp.dts = dts; tmp.duration = duration; + tmp.received_ts = received_ts; storage_task_enqueue(move(tmp)); update_ReferenceFrames(frame_type); } // Proxy object. -QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder) - : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder)) {} +QuickSyncEncoder::QuickSyncEncoder(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const string &va_display, int width, int height, AVOutputFormat *oformat, X264Encoder *x264_encoder, DiskSpaceEstimator *disk_space_estimator) + : impl(new QuickSyncEncoderImpl(filename, resource_pool, surface, va_display, width, height, oformat, x264_encoder, disk_space_estimator)) {} // Must be defined here because unique_ptr<> destructor needs to know the impl. QuickSyncEncoder::~QuickSyncEncoder() {}