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 #include <unordered_map>
  16
  17 #include "audio_encoder.h"
  18 #include "defs.h"
  19 #include "timebase.h"
  20 #include "print_latency.h"
  21
  22 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  23 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  24 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  25
  26 struct __bitstream {
  27     unsigned int *buffer;
  28     int bit_offset;
  29     int max_size_in_dword;
  30 };
  31 typedef struct __bitstream bitstream;
  32
  33 class QuickSyncEncoderImpl {
  34 public:
  35         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);
  36         ~QuickSyncEncoderImpl();
  37         void add_audio(int64_t pts, std::vector<float> audio);
  38         bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
  39         RefCountedGLsync end_frame(int64_t pts, int64_t duration, const std::vector<RefCountedFrame> &input_frames);
  40         void shutdown();
  41         void release_gl_resources();
  42         void set_stream_mux(Mux *mux)
  43         {
  44                 stream_mux = mux;
  45         }
  46
  47         // So we never get negative dts.
  48         int64_t global_delay() const {
  49                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
  50         }
  51
  52 private:
  53         struct storage_task {
  54                 unsigned long long display_order;
  55                 int frame_type;
  56                 std::vector<float> audio;
  57                 int64_t pts, dts, duration;
  58                 ReceivedTimestamps received_ts;
  59                 std::vector<size_t> ref_display_frame_numbers;
  60         };
  61         struct PendingFrame {
  62                 RefCountedGLsync fence;
  63                 std::vector<RefCountedFrame> input_frames;
  64                 int64_t pts, duration;
  65         };
  66         struct GLSurface {
  67                 VASurfaceID src_surface, ref_surface;
  68                 VABufferID coded_buf;
  69
  70                 VAImage surface_image;
  71                 GLuint y_tex, cbcr_tex;
  72
  73                 // Only if use_zerocopy == true.
  74                 EGLImage y_egl_image, cbcr_egl_image;
  75
  76                 // Only if use_zerocopy == false.
  77                 GLuint pbo;
  78                 uint8_t *y_ptr, *cbcr_ptr;
  79                 size_t y_offset, cbcr_offset;
  80
  81                 // Surfaces can be busy (have refcount > 0) for a variety of
  82                 // reasons: First of all because they belong to a frame that's
  83                 // under encoding. But also reference frames take refcounts;
  84                 // while a frame is being encoded, all its reference frames
  85                 // also have increased refcounts so that they are not dropped.
  86                 // Similarly, just being in <reference_frames> increases the
  87                 // refcount. Until it is back to zero, the surface cannot be given
  88                 // out for encoding another frame. Use release_gl_surface()
  89                 // to reduce the refcount, which will free the surface if
  90                 // the refcount reaches zero.
  91                 //
  92                 // Protected by storage_task_queue_mutex.
  93                 int refcount = 0;
  94         };
  95
  96         void open_output_file(const std::string &filename);
  97         void encode_thread_func();
  98         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
  99         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 100         void pass_frame(PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration);
 101         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 102                           int frame_type, int64_t pts, int64_t dts, int64_t duration);
 103         void storage_task_thread();
 104         void storage_task_enqueue(storage_task task);
 105         void save_codeddata(GLSurface *surf, storage_task task);
 106         int render_packedsequence();
 107         int render_packedpicture();
 108         void render_packedslice();
 109         int render_sequence();
 110         int render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num);
 111         void sps_rbsp(bitstream *bs);
 112         void pps_rbsp(bitstream *bs);
 113         int build_packed_pic_buffer(unsigned char **header_buffer);
 114         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 115         void slice_header(bitstream *bs);
 116         int build_packed_seq_buffer(unsigned char **header_buffer);
 117         int build_packed_slice_buffer(unsigned char **header_buffer);
 118         int init_va(const std::string &va_display);
 119         int deinit_va();
 120         void enable_zerocopy_if_possible();
 121         VADisplay va_open_display(const std::string &va_display);
 122         void va_close_display(VADisplay va_dpy);
 123         int setup_encode();
 124         void release_encode();
 125         void update_ReferenceFrames(int current_display_frame, int frame_type);
 126         void update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2]);
 127         void update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2]);
 128         GLSurface *allocate_gl_surface();
 129         void release_gl_surface(size_t display_frame_num);
 130
 131         bool is_shutdown = false;
 132         bool has_released_gl_resources = false;
 133         bool use_zerocopy;
 134         int drm_fd = -1;
 135
 136         std::thread encode_thread, storage_thread;
 137
 138         std::mutex storage_task_queue_mutex;
 139         std::condition_variable storage_task_queue_changed;
 140         std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 141         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 142
 143         std::mutex frame_queue_mutex;
 144         std::condition_variable frame_queue_nonempty;
 145         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 146
 147         int current_storage_frame;
 148
 149         std::queue<PendingFrame> pending_video_frames;  // under frame_queue_mutex
 150         movit::ResourcePool *resource_pool;
 151         QSurface *surface;
 152
 153         // Frames that are done rendering and passed on to x264 (if enabled),
 154         // but have not been encoded by Quick Sync yet, and thus also not freed.
 155         // The key is the display frame number.
 156         std::map<int, PendingFrame> reorder_buffer;
 157         int quicksync_encoding_frame_num = 0;
 158
 159         std::unique_ptr<AudioEncoder> file_audio_encoder;
 160
 161         X264Encoder *x264_encoder;  // nullptr if not using x264.
 162
 163         Mux* stream_mux = nullptr;  // To HTTP.
 164         std::unique_ptr<Mux> file_mux;  // To local disk.
 165
 166         Display *x11_display = nullptr;
 167
 168         // Encoder parameters
 169         VADisplay va_dpy;
 170         VAProfile h264_profile = (VAProfile)~0;
 171         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 172         int config_attrib_num = 0, enc_packed_header_idx;
 173
 174         GLSurface gl_surfaces[SURFACE_NUM];
 175
 176         // For all frames in encoding (refcount > 0), a pointer into gl_surfaces
 177         // for the surface used for that frame. Protected by storage_task_queue_mutex.
 178         // The key is display frame number.
 179         std::unordered_map<size_t, GLSurface *> surface_for_frame;
 180
 181         VAConfigID config_id;
 182         VAContextID context_id;
 183         VAEncSequenceParameterBufferH264 seq_param;
 184         VAEncPictureParameterBufferH264 pic_param;
 185         VAEncSliceParameterBufferH264 slice_param;
 186         VAPictureH264 CurrentCurrPic;
 187
 188         struct ReferenceFrame {
 189                 VAPictureH264 pic;
 190                 int display_number;  // To track reference counts.
 191         };
 192         std::deque<ReferenceFrame> reference_frames;
 193
 194         // Static quality settings.
 195         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 196         static constexpr unsigned int num_ref_frames = 2;
 197         static constexpr int initial_qp = 15;
 198         static constexpr int minimal_qp = 0;
 199         static constexpr int intra_period = 30;
 200         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 201
 202         // Quality settings that are meant to be static, but might be overridden
 203         // by the profile.
 204         int constraint_set_flag = 0;
 205         int h264_packedheader = 0; /* support pack header? */
 206         int h264_maxref = (1<<16|1);
 207         int h264_entropy_mode = 1; /* cabac */
 208         int ip_period = 3;
 209
 210         int rc_mode = -1;
 211         unsigned int current_ref_frame_num = 0;  // Encoding frame order within this GOP, sans B-frames.
 212
 213         int frame_width;
 214         int frame_height;
 215         int frame_width_mbaligned;
 216         int frame_height_mbaligned;
 217
 218         DiskSpaceEstimator *disk_space_estimator;
 219 };
 220
 221 #endif  // !defined(_QUICKSYNC_ENCODER_IMPL_H)