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 #include "ref_counted_gl_sync.h"
  23 #include "va_display_with_cleanup.h"
  24
  25 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  26 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  27 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  28
  29 struct __bitstream {
  30     unsigned int *buffer;
  31     int bit_offset;
  32     int max_size_in_dword;
  33 };
  34 typedef struct __bitstream bitstream;
  35
  36 namespace movit {
  37 class ResourcePool;
  38 }
  39 class DiskSpaceEstimator;
  40 class QSurface;
  41 class X264Encoder;
  42
  43 class QuickSyncEncoderImpl {
  44 public:
  45         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);
  46         ~QuickSyncEncoderImpl();
  47         void add_audio(int64_t pts, std::vector<float> audio);
  48         bool is_zerocopy() const;
  49         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);
  50         RefCountedGLsync end_frame();
  51         void shutdown();
  52         void close_file();
  53         void release_gl_resources();
  54         void set_stream_mux(Mux *mux)
  55         {
  56                 stream_mux = mux;
  57         }
  58
  59         // So we never get negative dts.
  60         int64_t global_delay() const {
  61                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
  62         }
  63
  64 private:
  65         struct storage_task {
  66                 unsigned long long display_order;
  67                 int frame_type;
  68                 std::vector<float> audio;
  69                 int64_t pts, dts, duration;
  70                 movit::YCbCrLumaCoefficients ycbcr_coefficients;
  71                 ReceivedTimestamps received_ts;
  72                 std::vector<size_t> ref_display_frame_numbers;
  73         };
  74         struct PendingFrame {
  75                 RefCountedGLsync fence;
  76                 std::vector<RefCountedFrame> input_frames;
  77                 int64_t pts, duration;
  78                 movit::YCbCrLumaCoefficients ycbcr_coefficients;
  79         };
  80         struct GLSurface {
  81                 // Only if x264_video_to_disk == false.
  82                 VASurfaceID src_surface, ref_surface;
  83                 VABufferID coded_buf;
  84                 VAImage surface_image;
  85
  86                 // Only if use_zerocopy == true (which implies x264_video_to_disk == false).
  87                 GLuint y_tex, cbcr_tex;
  88                 EGLImage y_egl_image, cbcr_egl_image;
  89
  90                 // Only if use_zerocopy == false.
  91                 GLuint pbo;
  92                 uint8_t *y_ptr, *cbcr_ptr;
  93                 size_t y_offset, cbcr_offset;
  94
  95                 // Surfaces can be busy (have refcount > 0) for a variety of
  96                 // reasons: First of all because they belong to a frame that's
  97                 // under encoding. But also reference frames take refcounts;
  98                 // while a frame is being encoded, all its reference frames
  99                 // also have increased refcounts so that they are not dropped.
 100                 // Similarly, just being in <reference_frames> increases the
 101                 // refcount. Until it is back to zero, the surface cannot be given
 102                 // out for encoding another frame. Use release_gl_surface()
 103                 // to reduce the refcount, which will free the surface if
 104                 // the refcount reaches zero.
 105                 //
 106                 // Protected by storage_task_queue_mutex.
 107                 int refcount = 0;
 108         };
 109
 110         void open_output_file(const std::string &filename);
 111         void encode_thread_func();
 112         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 113         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 114         void pass_frame(PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration);
 115         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 116                           int frame_type, int64_t pts, int64_t dts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients);
 117         void storage_task_thread();
 118         void storage_task_enqueue(storage_task task);
 119         void save_codeddata(GLSurface *surf, storage_task task);
 120         int render_packedsequence(movit::YCbCrLumaCoefficients ycbcr_coefficients);
 121         int render_packedpicture();
 122         void render_packedslice();
 123         int render_sequence();
 124         int render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num);
 125         void sps_rbsp(movit::YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs);
 126         void pps_rbsp(bitstream *bs);
 127         int build_packed_pic_buffer(unsigned char **header_buffer);
 128         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 129         void slice_header(bitstream *bs);
 130         int build_packed_seq_buffer(movit::YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer);
 131         int build_packed_slice_buffer(unsigned char **header_buffer);
 132         int init_va(const std::string &va_display);
 133         void enable_zerocopy_if_possible();
 134         int setup_encode();
 135         void release_encode();
 136         void update_ReferenceFrames(int current_display_frame, int frame_type);
 137         void update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2]);
 138         void update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2]);
 139         GLSurface *allocate_gl_surface();
 140         void release_gl_surface(size_t display_frame_num);
 141
 142         bool is_shutdown = false;
 143         bool has_released_gl_resources = false;
 144         std::atomic<bool> use_zerocopy{false};
 145
 146         std::thread encode_thread, storage_thread;
 147
 148         std::mutex storage_task_queue_mutex;
 149         std::condition_variable storage_task_queue_changed;
 150         std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 151         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 152
 153         std::mutex frame_queue_mutex;
 154         std::condition_variable frame_queue_nonempty;
 155         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 156
 157         int current_storage_frame;
 158
 159         PendingFrame current_video_frame;  // Used only between begin_frame() and end_frame().
 160         std::queue<PendingFrame> pending_video_frames;  // under frame_queue_mutex
 161         movit::ResourcePool *resource_pool;
 162         QSurface *surface;
 163
 164         // Frames that are done rendering and passed on to x264 (if enabled),
 165         // but have not been encoded by Quick Sync yet, and thus also not freed.
 166         // The key is the display frame number.
 167         std::map<int, PendingFrame> reorder_buffer;
 168         int quicksync_encoding_frame_num = 0;
 169
 170         std::mutex file_audio_encoder_mutex;
 171         std::unique_ptr<AudioEncoder> file_audio_encoder;
 172
 173         X264Encoder *x264_encoder;  // nullptr if not using x264.
 174
 175         Mux* stream_mux = nullptr;  // To HTTP.
 176         std::unique_ptr<Mux> file_mux;  // To local disk.
 177
 178         // Encoder parameters
 179         std::unique_ptr<VADisplayWithCleanup> va_dpy;
 180         VAProfile h264_profile = (VAProfile)~0;
 181         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 182         int config_attrib_num = 0, enc_packed_header_idx;
 183
 184         GLSurface gl_surfaces[SURFACE_NUM];
 185
 186         // For all frames in encoding (refcount > 0), a pointer into gl_surfaces
 187         // for the surface used for that frame. Protected by storage_task_queue_mutex.
 188         // The key is display frame number.
 189         std::unordered_map<size_t, GLSurface *> surface_for_frame;
 190
 191         VAConfigID config_id;
 192         VAContextID context_id;
 193         VAEncSequenceParameterBufferH264 seq_param;
 194         VAEncPictureParameterBufferH264 pic_param;
 195         VAEncSliceParameterBufferH264 slice_param;
 196         VAPictureH264 CurrentCurrPic;
 197
 198         struct ReferenceFrame {
 199                 VAPictureH264 pic;
 200                 int display_number;  // To track reference counts.
 201         };
 202         std::deque<ReferenceFrame> reference_frames;
 203
 204         // Static quality settings.
 205         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 206         static constexpr unsigned int num_ref_frames = 2;
 207         static constexpr int initial_qp = 15;
 208         static constexpr int minimal_qp = 0;
 209         static constexpr int intra_period = 30;
 210         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 211
 212         // Quality settings that are meant to be static, but might be overridden
 213         // by the profile.
 214         int constraint_set_flag = 0;
 215         int h264_packedheader = 0; /* support pack header? */
 216         int h264_maxref = (1<<16|1);
 217         int h264_entropy_mode = 1; /* cabac */
 218         int ip_period = 3;
 219
 220         unsigned int current_ref_frame_num = 0;  // Encoding frame order within this GOP, sans B-frames.
 221
 222         int frame_width;
 223         int frame_height;
 224         int frame_width_mbaligned;
 225         int frame_height_mbaligned;
 226
 227         DiskSpaceEstimator *disk_space_estimator;
 228 };
 229
 230 #endif  // !defined(_QUICKSYNC_ENCODER_IMPL_H)