git.sesse.net Git - nageru/blob - quicksync_encoder_impl.h

   1 #ifndef _QUICKSYNC_ENCODER_IMPL_H
   2 #define _QUICKSYNC_ENCODER_IMPL_H 1
   3
   4 #include <epoxy/egl.h>
   5 #include <va/va.h>
   6
   7 #include <condition_variable>
   8 #include <map>
   9 #include <memory>
  10 #include <mutex>
  11 #include <queue>
  12 #include <string>
  13 #include <stack>
  14 #include <thread>
  15
  16 #include "audio_encoder.h"
  17 #include "defs.h"
  18 #include "timebase.h"
  19 #include "print_latency.h"
  20
  21 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  22 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  23 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  24
  25 struct __bitstream {
  26     unsigned int *buffer;
  27     int bit_offset;
  28     int max_size_in_dword;
  29 };
  30 typedef struct __bitstream bitstream;
  31
  32 // H.264 video comes out in encoding order (e.g. with two B-frames:
  33 // 0, 3, 1, 2, 6, 4, 5, etc.), but uncompressed video needs to
  34 // come in the right order. Since we do everything, including waiting
  35 // for the frames to come out of OpenGL, in encoding order, we need
  36 // a reordering buffer for uncompressed frames so that they come out
  37 // correctly. We go the super-lazy way of not making it understand
  38 // anything about the true order (which introduces some extra latency,
  39 // though); we know that for N B-frames we need at most (N-1) frames
  40 // in the reorder buffer, and can just sort on that.
  41 //
  42 // The class also deals with keeping a freelist as needed.
  43 class FrameReorderer {
  44 public:
  45         FrameReorderer(unsigned queue_length, int width, int height);
  46
  47         struct Frame {
  48                 int64_t pts, duration;
  49                 uint8_t *data;
  50                 ReceivedTimestamps received_ts;
  51
  52                 // Invert to get the smallest pts first.
  53                 bool operator< (const Frame &other) const { return pts > other.pts; }
  54         };
  55
  56         // Returns the next frame to insert with its pts, if any. Otherwise -1 and nullptr.
  57         // Does _not_ take ownership of data; a copy is taken if needed.
  58         // The returned pointer is valid until the next call to reorder_frame, or destruction.
  59         // As a special case, if queue_length == 0, will just return pts and data (no reordering needed).
  60         Frame reorder_frame(int64_t pts, int64_t duration, uint8_t *data, const ReceivedTimestamps &received_ts);
  61
  62         // The same as reorder_frame, but without inserting anything. Used to empty the queue.
  63         Frame get_first_frame();
  64
  65         bool empty() const { return frames.empty(); }
  66
  67 private:
  68         unsigned queue_length;
  69         int width, height;
  70
  71         std::priority_queue<Frame> frames;
  72         std::stack<uint8_t *> freelist;  // Includes the last value returned from reorder_frame.
  73
  74         // Owns all the pointers. Normally, freelist and frames could do this themselves,
  75         // except priority_queue doesn't work well with movable-only types.
  76         std::vector<std::unique_ptr<uint8_t[]>> owner;
  77 };
  78
  79
  80 class QuickSyncEncoderImpl {
  81 public:
  82         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);
  83         ~QuickSyncEncoderImpl();
  84         void add_audio(int64_t pts, std::vector<float> audio);
  85         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
  86         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const std::vector<RefCountedFrame> &input_frames);
  87         void shutdown();
  88         void release_gl_resources();
  89         void set_stream_mux(Mux *mux)
  90         {
  91                 stream_mux = mux;
  92         }
  93
  94         // So we never get negative dts.
  95         int64_t global_delay() const {
  96                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
  97         }
  98
  99 private:
 100         struct storage_task {
 101                 unsigned long long display_order;
 102                 int frame_type;
 103                 std::vector<float> audio;
 104                 int64_t pts, dts, duration;
 105                 ReceivedTimestamps received_ts;
 106         };
 107         struct PendingFrame {
 108                 RefCountedGLsync fence;
 109                 std::vector<RefCountedFrame> input_frames;
 110                 int64_t pts, duration;
 111         };
 112
 113         void open_output_file(const std::string &filename);
 114         void encode_thread_func();
 115         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 116         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 117         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 118                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 119         void storage_task_thread();
 120         void storage_task_enqueue(storage_task task);
 121         void save_codeddata(storage_task task);
 122         int render_packedsequence();
 123         int render_packedpicture();
 124         void render_packedslice();
 125         int render_sequence();
 126         int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
 127         void sps_rbsp(bitstream *bs);
 128         void pps_rbsp(bitstream *bs);
 129         int build_packed_pic_buffer(unsigned char **header_buffer);
 130         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 131         void slice_header(bitstream *bs);
 132         int build_packed_seq_buffer(unsigned char **header_buffer);
 133         int build_packed_slice_buffer(unsigned char **header_buffer);
 134         int init_va(const std::string &va_display);
 135         int deinit_va();
 136         void enable_zerocopy_if_possible();
 137         VADisplay va_open_display(const std::string &va_display);
 138         void va_close_display(VADisplay va_dpy);
 139         int setup_encode();
 140         void release_encode();
 141         void update_ReferenceFrames(int frame_type);
 142         int update_RefPicList(int frame_type);
 143
 144         bool is_shutdown = false;
 145         bool has_released_gl_resources = false;
 146         bool use_zerocopy;
 147         int drm_fd = -1;
 148
 149         std::thread encode_thread, storage_thread;
 150
 151         std::mutex storage_task_queue_mutex;
 152         std::condition_variable storage_task_queue_changed;
 153         int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
 154         std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 155         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 156
 157         std::mutex frame_queue_mutex;
 158         std::condition_variable frame_queue_nonempty;
 159         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 160
 161         int current_storage_frame;
 162
 163         std::map<int, PendingFrame> pending_video_frames;  // under frame_queue_mutex
 164         movit::ResourcePool *resource_pool;
 165         QSurface *surface;
 166
 167         std::unique_ptr<AudioEncoder> file_audio_encoder;
 168
 169         std::unique_ptr<FrameReorderer> reorderer;
 170         X264Encoder *x264_encoder;  // nullptr if not using x264.
 171
 172         Mux* stream_mux = nullptr;  // To HTTP.
 173         std::unique_ptr<Mux> file_mux;  // To local disk.
 174
 175         Display *x11_display = nullptr;
 176
 177         // Encoder parameters
 178         VADisplay va_dpy;
 179         VAProfile h264_profile = (VAProfile)~0;
 180         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 181         int config_attrib_num = 0, enc_packed_header_idx;
 182
 183         struct GLSurface {
 184                 VASurfaceID src_surface, ref_surface;
 185                 VABufferID coded_buf;
 186
 187                 VAImage surface_image;
 188                 GLuint y_tex, cbcr_tex;
 189
 190                 // Only if use_zerocopy == true.
 191                 EGLImage y_egl_image, cbcr_egl_image;
 192
 193                 // Only if use_zerocopy == false.
 194                 GLuint pbo;
 195                 uint8_t *y_ptr, *cbcr_ptr;
 196                 size_t y_offset, cbcr_offset;
 197         };
 198         GLSurface gl_surfaces[SURFACE_NUM];
 199
 200         VAConfigID config_id;
 201         VAContextID context_id;
 202         VAEncSequenceParameterBufferH264 seq_param;
 203         VAEncPictureParameterBufferH264 pic_param;
 204         VAEncSliceParameterBufferH264 slice_param;
 205         VAPictureH264 CurrentCurrPic;
 206         VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 207
 208         // Static quality settings.
 209         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 210         static constexpr unsigned int num_ref_frames = 2;
 211         static constexpr int initial_qp = 15;
 212         static constexpr int minimal_qp = 0;
 213         static constexpr int intra_period = 30;
 214         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 215
 216         // Quality settings that are meant to be static, but might be overridden
 217         // by the profile.
 218         int constraint_set_flag = 0;
 219         int h264_packedheader = 0; /* support pack header? */
 220         int h264_maxref = (1<<16|1);
 221         int h264_entropy_mode = 1; /* cabac */
 222         int ip_period = 3;
 223
 224         int rc_mode = -1;
 225         unsigned int current_frame_num = 0;
 226         unsigned int numShortTerm = 0;
 227
 228         int frame_width;
 229         int frame_height;
 230         int frame_width_mbaligned;
 231         int frame_height_mbaligned;
 232
 233         DiskSpaceEstimator *disk_space_estimator;
 234 };
 235
 236 #endif  // !defined(_QUICKSYNC_ENCODER_IMPL_H)