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 <movit/image_format.h>
   6 #include <va/va.h>
   7
   8 #include <condition_variable>
   9 #include <map>
  10 #include <memory>
  11 #include <mutex>
  12 #include <queue>
  13 #include <string>
  14 #include <stack>
  15 #include <thread>
  16 #include <unordered_map>
  17
  18 #include "audio_encoder.h"
  19 #include "defs.h"
  20 #include "timebase.h"
  21 #include "print_latency.h"
  22
  23 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  24 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  25 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  26
  27 struct __bitstream {
  28     unsigned int *buffer;
  29     int bit_offset;
  30     int max_size_in_dword;
  31 };
  32 typedef struct __bitstream bitstream;
  33
  34 class QuickSyncEncoderImpl {
  35 public:
  36         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);
  37         ~QuickSyncEncoderImpl();
  38         void add_audio(int64_t pts, std::vector<float> audio);
  39         bool is_zerocopy() const;
  40         bool begin_frame(int64_t pts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients, const std::vector<RefCountedFrame> &input_frames, GLuint *y_tex, GLuint *cbcr_tex);
  41         RefCountedGLsync end_frame();
  42         void shutdown();
  43         void close_file();
  44         void release_gl_resources();
  45         void set_stream_mux(Mux *mux)
  46         {
  47                 stream_mux = mux;
  48         }
  49
  50         // So we never get negative dts.
  51         int64_t global_delay() const {
  52                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
  53         }
  54
  55 private:
  56         struct storage_task {
  57                 unsigned long long display_order;
  58                 int frame_type;
  59                 std::vector<float> audio;
  60                 int64_t pts, dts, duration;
  61                 movit::YCbCrLumaCoefficients ycbcr_coefficients;
  62                 ReceivedTimestamps received_ts;
  63                 std::vector<size_t> ref_display_frame_numbers;
  64         };
  65         struct PendingFrame {
  66                 RefCountedGLsync fence;
  67                 std::vector<RefCountedFrame> input_frames;
  68                 int64_t pts, duration;
  69                 movit::YCbCrLumaCoefficients ycbcr_coefficients;
  70         };
  71         struct GLSurface {
  72                 // Only if x264_video_to_disk == false.
  73                 VASurfaceID src_surface, ref_surface;
  74                 VABufferID coded_buf;
  75                 VAImage surface_image;
  76
  77                 // Only if use_zerocopy == true (which implies x264_video_to_disk == false).
  78                 GLuint y_tex, cbcr_tex;
  79                 EGLImage y_egl_image, cbcr_egl_image;
  80
  81                 // Only if use_zerocopy == false.
  82                 GLuint pbo;
  83                 uint8_t *y_ptr, *cbcr_ptr;
  84                 size_t y_offset, cbcr_offset;
  85
  86                 // Surfaces can be busy (have refcount > 0) for a variety of
  87                 // reasons: First of all because they belong to a frame that's
  88                 // under encoding. But also reference frames take refcounts;
  89                 // while a frame is being encoded, all its reference frames
  90                 // also have increased refcounts so that they are not dropped.
  91                 // Similarly, just being in <reference_frames> increases the
  92                 // refcount. Until it is back to zero, the surface cannot be given
  93                 // out for encoding another frame. Use release_gl_surface()
  94                 // to reduce the refcount, which will free the surface if
  95                 // the refcount reaches zero.
  96                 //
  97                 // Protected by storage_task_queue_mutex.
  98                 int refcount = 0;
  99         };
 100
 101         void open_output_file(const std::string &filename);
 102         void encode_thread_func();
 103         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 104         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 105         void pass_frame(PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration);
 106         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 107                           int frame_type, int64_t pts, int64_t dts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients);
 108         void storage_task_thread();
 109         void storage_task_enqueue(storage_task task);
 110         void save_codeddata(GLSurface *surf, storage_task task);
 111         int render_packedsequence(movit::YCbCrLumaCoefficients ycbcr_coefficients);
 112         int render_packedpicture();
 113         void render_packedslice();
 114         int render_sequence();
 115         int render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num);
 116         void sps_rbsp(movit::YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs);
 117         void pps_rbsp(bitstream *bs);
 118         int build_packed_pic_buffer(unsigned char **header_buffer);
 119         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 120         void slice_header(bitstream *bs);
 121         int build_packed_seq_buffer(movit::YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer);
 122         int build_packed_slice_buffer(unsigned char **header_buffer);
 123         int init_va(const std::string &va_display);
 124         int deinit_va();
 125         void enable_zerocopy_if_possible();
 126         VADisplay va_open_display(const std::string &va_display);
 127         void va_close_display(VADisplay va_dpy);
 128         int setup_encode();
 129         void release_encode();
 130         void update_ReferenceFrames(int current_display_frame, int frame_type);
 131         void update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2]);
 132         void update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2]);
 133         GLSurface *allocate_gl_surface();
 134         void release_gl_surface(size_t display_frame_num);
 135
 136         bool is_shutdown = false;
 137         bool has_released_gl_resources = false;
 138         std::atomic<bool> use_zerocopy;
 139         int drm_fd = -1;
 140
 141         std::thread encode_thread, storage_thread;
 142
 143         std::mutex storage_task_queue_mutex;
 144         std::condition_variable storage_task_queue_changed;
 145         std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 146         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 147
 148         std::mutex frame_queue_mutex;
 149         std::condition_variable frame_queue_nonempty;
 150         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 151
 152         int current_storage_frame;
 153
 154         PendingFrame current_video_frame;  // Used only between begin_frame() and end_frame().
 155         std::queue<PendingFrame> pending_video_frames;  // under frame_queue_mutex
 156         movit::ResourcePool *resource_pool;
 157         QSurface *surface;
 158
 159         // Frames that are done rendering and passed on to x264 (if enabled),
 160         // but have not been encoded by Quick Sync yet, and thus also not freed.
 161         // The key is the display frame number.
 162         std::map<int, PendingFrame> reorder_buffer;
 163         int quicksync_encoding_frame_num = 0;
 164
 165         std::mutex file_audio_encoder_mutex;
 166         std::unique_ptr<AudioEncoder> file_audio_encoder;
 167
 168         X264Encoder *x264_encoder;  // nullptr if not using x264.
 169
 170         Mux* stream_mux = nullptr;  // To HTTP.
 171         std::unique_ptr<Mux> file_mux;  // To local disk.
 172
 173         Display *x11_display = nullptr;
 174
 175         // Encoder parameters
 176         VADisplay va_dpy;
 177         VAProfile h264_profile = (VAProfile)~0;
 178         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 179         int config_attrib_num = 0, enc_packed_header_idx;
 180
 181         GLSurface gl_surfaces[SURFACE_NUM];
 182
 183         // For all frames in encoding (refcount > 0), a pointer into gl_surfaces
 184         // for the surface used for that frame. Protected by storage_task_queue_mutex.
 185         // The key is display frame number.
 186         std::unordered_map<size_t, GLSurface *> surface_for_frame;
 187
 188         VAConfigID config_id;
 189         VAContextID context_id;
 190         VAEncSequenceParameterBufferH264 seq_param;
 191         VAEncPictureParameterBufferH264 pic_param;
 192         VAEncSliceParameterBufferH264 slice_param;
 193         VAPictureH264 CurrentCurrPic;
 194
 195         struct ReferenceFrame {
 196                 VAPictureH264 pic;
 197                 int display_number;  // To track reference counts.
 198         };
 199         std::deque<ReferenceFrame> reference_frames;
 200
 201         // Static quality settings.
 202         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 203         static constexpr unsigned int num_ref_frames = 2;
 204         static constexpr int initial_qp = 15;
 205         static constexpr int minimal_qp = 0;
 206         static constexpr int intra_period = 30;
 207         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 208
 209         // Quality settings that are meant to be static, but might be overridden
 210         // by the profile.
 211         int constraint_set_flag = 0;
 212         int h264_packedheader = 0; /* support pack header? */
 213         int h264_maxref = (1<<16|1);
 214         int h264_entropy_mode = 1; /* cabac */
 215         int ip_period = 3;
 216
 217         int rc_mode = -1;
 218         unsigned int current_ref_frame_num = 0;  // Encoding frame order within this GOP, sans B-frames.
 219
 220         int frame_width;
 221         int frame_height;
 222         int frame_width_mbaligned;
 223         int frame_height_mbaligned;
 224
 225         DiskSpaceEstimator *disk_space_estimator;
 226 };
 227
 228 #endif  // !defined(_QUICKSYNC_ENCODER_IMPL_H)