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

   1 #ifndef _QUICKSYNC_ENCODER_IMPL_H
   2 #define _QUICKSYNC_ENCODER_IMPL_H 1
   3
   4 #include <stddef.h>
   5 #include <stdint.h>
   6 #include <epoxy/egl.h>
   7 #include <epoxy/gl.h>
   8 #include <movit/image_format.h>
   9 #include <va/va.h>
  10 #include <va/va_enc_h264.h>
  11
  12 #include <atomic>
  13 #include <condition_variable>
  14 #include <deque>
  15 #include <map>
  16 #include <memory>
  17 #include <mutex>
  18 #include <queue>
  19 #include <string>
  20 #include <thread>
  21 #include <unordered_map>
  22 #include <vector>
  23
  24 #include "audio_encoder.h"
  25 #include "defs.h"
  26 #include "shared/timebase.h"
  27 #include "print_latency.h"
  28 #include "ref_counted_frame.h"
  29 #include "shared/ref_counted_gl_sync.h"
  30 #include "shared/va_display.h"
  31 #include "v4l_output.h"
  32
  33 extern "C" {
  34 #include <libavformat/avformat.h>
  35 }
  36
  37 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
  38 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  39 #define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
  40
  41 struct __bitstream {
  42     unsigned int *buffer;
  43     int bit_offset;
  44     int max_size_in_dword;
  45 };
  46 typedef struct __bitstream bitstream;
  47
  48 namespace movit {
  49 class ResourcePool;
  50 }
  51 class DiskSpaceEstimator;
  52 class QSurface;
  53 class VideoCodecInterface;
  54
  55 class QuickSyncEncoderImpl {
  56 public:
  57         QuickSyncEncoderImpl(const std::string &filename, movit::ResourcePool *resource_pool, QSurface *surface, const std::string &va_display, int width, int height, const AVOutputFormat *oformat, VideoCodecInterface *http_encoder, VideoCodecInterface *disk_encoder, DiskSpaceEstimator *disk_space_estimator);
  58         ~QuickSyncEncoderImpl();
  59         void add_audio(int64_t pts, std::vector<float> audio);
  60         bool is_zerocopy() const;
  61         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);
  62         RefCountedGLsync end_frame();
  63         void shutdown();
  64         void close_file();
  65         void release_gl_resources();
  66         void set_http_mux(Mux *mux)
  67         {
  68                 http_mux = mux;
  69         }
  70         void set_srt_mux(Mux *mux)
  71         {
  72                 srt_mux = mux;
  73         }
  74
  75         // So we never get negative dts.
  76         int64_t global_delay() const {
  77                 return int64_t(ip_period - 1) * (TIMEBASE / TYPICAL_FPS);
  78         }
  79
  80 private:
  81         struct storage_task {
  82                 unsigned long long display_order;
  83                 int frame_type;
  84                 std::vector<float> audio;
  85                 int64_t pts, dts, duration;
  86                 movit::YCbCrLumaCoefficients ycbcr_coefficients;
  87                 ReceivedTimestamps received_ts;
  88                 std::vector<size_t> ref_display_frame_numbers;
  89         };
  90         struct PendingFrame {
  91                 RefCountedGLsync fence;
  92                 std::vector<RefCountedFrame> input_frames;
  93                 int64_t pts, duration;
  94                 movit::YCbCrLumaCoefficients ycbcr_coefficients;
  95         };
  96         struct GLSurface {
  97                 // Only if x264_video_to_disk == false.
  98                 VASurfaceID src_surface, ref_surface;
  99                 VABufferID coded_buf;
 100                 VAImage surface_image;
 101
 102                 // Only if use_zerocopy == true (which implies x264_video_to_disk == false).
 103                 GLuint y_tex, cbcr_tex;
 104                 EGLImage y_egl_image, cbcr_egl_image;
 105
 106                 // Only if use_zerocopy == false.
 107                 GLuint pbo;
 108                 uint8_t *y_ptr, *cbcr_ptr;
 109                 size_t y_offset, cbcr_offset;
 110
 111                 // Surfaces can be busy (have refcount > 0) for a variety of
 112                 // reasons: First of all because they belong to a frame that's
 113                 // under encoding. But also reference frames take refcounts;
 114                 // while a frame is being encoded, all its reference frames
 115                 // also have increased refcounts so that they are not dropped.
 116                 // Similarly, just being in <reference_frames> increases the
 117                 // refcount. Until it is back to zero, the surface cannot be given
 118                 // out for encoding another frame. Use release_gl_surface()
 119                 // to reduce the refcount, which will free the surface if
 120                 // the refcount reaches zero.
 121                 //
 122                 // Protected by storage_task_queue_mutex.
 123                 int refcount = 0;
 124         };
 125
 126         void open_output_file(const std::string &filename);
 127         void encode_thread_func();
 128         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 129         void pass_frame(PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration);
 130         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 131                           int frame_type, int64_t pts, int64_t dts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients);
 132         void storage_task_thread();
 133         void storage_task_enqueue(storage_task task);
 134         void save_codeddata(GLSurface *surf, storage_task task);
 135         int render_packedsequence(movit::YCbCrLumaCoefficients ycbcr_coefficients);
 136         int render_packedpicture();
 137         void render_packedslice();
 138         int render_sequence();
 139         int render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num);
 140         void sps_rbsp(movit::YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs);
 141         void pps_rbsp(bitstream *bs);
 142         int build_packed_pic_buffer(unsigned char **header_buffer);
 143         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 144         void slice_header(bitstream *bs);
 145         int build_packed_seq_buffer(movit::YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer);
 146         int build_packed_slice_buffer(unsigned char **header_buffer);
 147         int init_va(const std::string &va_display);
 148         void enable_zerocopy_if_possible();
 149         int setup_encode();
 150         void release_encode();
 151         void update_ReferenceFrames(int current_display_frame, int frame_type);
 152         void update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2]);
 153         void update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2]);
 154         GLSurface *allocate_gl_surface();
 155         void release_gl_surface(size_t display_frame_num);
 156
 157         bool is_shutdown = false;
 158         bool has_released_gl_resources = false;
 159         std::atomic<bool> use_zerocopy{false};
 160
 161         std::thread encode_thread, storage_thread;
 162
 163         std::mutex storage_task_queue_mutex;
 164         std::condition_variable storage_task_queue_changed;
 165         std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 166         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 167
 168         std::mutex frame_queue_mutex;
 169         std::condition_variable frame_queue_nonempty;
 170         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 171
 172         int current_storage_frame;
 173
 174         PendingFrame current_video_frame;  // Used only between begin_frame() and end_frame().
 175         std::queue<PendingFrame> pending_video_frames;  // under frame_queue_mutex
 176         movit::ResourcePool *resource_pool;
 177         QSurface *surface;
 178
 179         // Frames that are done rendering and passed on to x264 (if enabled),
 180         // but have not been encoded by Quick Sync yet, and thus also not freed.
 181         // The key is the display frame number.
 182         std::map<int, PendingFrame> reorder_buffer;
 183         int quicksync_encoding_frame_num = 0;
 184
 185         std::mutex file_audio_encoder_mutex;
 186         std::unique_ptr<AudioEncoder> file_audio_encoder;
 187
 188         VideoCodecInterface *http_encoder;  // nullptr if not using x264/SVT-AV1.
 189         VideoCodecInterface *disk_encoder;
 190         std::unique_ptr<V4LOutput> v4l_output;  // nullptr if not using V4L2 output.
 191
 192         Mux* http_mux = nullptr;  // To the HTTP server.
 193         Mux* srt_mux = nullptr;  // To the remote SRT endpoint, if any.
 194         std::unique_ptr<Mux> file_mux;  // To local disk.
 195
 196         // Encoder parameters
 197         std::unique_ptr<VADisplayWithCleanup> va_dpy;
 198         VAProfile h264_profile = (VAProfile)~0;
 199         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 200         int config_attrib_num = 0, enc_packed_header_idx;
 201
 202         GLSurface gl_surfaces[SURFACE_NUM];
 203
 204         // For all frames in encoding (refcount > 0), a pointer into gl_surfaces
 205         // for the surface used for that frame. Protected by storage_task_queue_mutex.
 206         // The key is display frame number.
 207         std::unordered_map<size_t, GLSurface *> surface_for_frame;
 208
 209         VAConfigID config_id;
 210         VAContextID context_id;
 211         VAEncSequenceParameterBufferH264 seq_param;
 212         VAEncPictureParameterBufferH264 pic_param;
 213         VAEncSliceParameterBufferH264 slice_param;
 214         VAPictureH264 CurrentCurrPic;
 215
 216         struct ReferenceFrame {
 217                 VAPictureH264 pic;
 218                 int display_number;  // To track reference counts.
 219         };
 220         std::deque<ReferenceFrame> reference_frames;
 221
 222         // Static quality settings.
 223         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 224         static constexpr unsigned int num_ref_frames = 2;
 225         static constexpr int initial_qp = 15;
 226         static constexpr int minimal_qp = 0;
 227         static constexpr int intra_period = 30;
 228         static constexpr int intra_idr_period = TYPICAL_FPS;  // About a second; more at lower frame rates. Not ideal.
 229
 230         // Quality settings that are meant to be static, but might be overridden
 231         // by the profile.
 232         int constraint_set_flag = 0;
 233         int h264_packedheader = 0; /* support pack header? */
 234         int h264_maxref = (1<<16|1);
 235         int h264_entropy_mode = 1; /* cabac */
 236         int ip_period = 3;
 237
 238         unsigned int current_ref_frame_num = 0;  // Encoding frame order within this GOP, sans B-frames.
 239
 240         int frame_width;
 241         int frame_height;
 242         int frame_width_mbaligned;
 243         int frame_height_mbaligned;
 244
 245         DiskSpaceEstimator *disk_space_estimator;
 246 };
 247
 248 #endif  // !defined(_QUICKSYNC_ENCODER_IMPL_H)