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