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 struct VADisplayWithCleanup {
  35         ~VADisplayWithCleanup();
  36
  37         VADisplay va_dpy;
  38         Display *x11_display = nullptr;
  39         bool can_use_zerocopy = true;
  40         int drm_fd = -1;
  41 };
  42 std::unique_ptr<VADisplayWithCleanup> va_open_display(const std::string &va_display);  // Can return nullptr on failure.
  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, AVOutputFormat *oformat, X264Encoder *x264_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_stream_mux(Mux *mux)
  56         {
  57                 stream_mux = mux;
  58         }
  59
  60         // So we never get negative dts.
  61         int64_t global_delay() const {
  62                 return int64_t(ip_period - 1) * (TIMEBASE / MAX_FPS);
  63         }
  64
  65 private:
  66         struct storage_task {
  67                 unsigned long long display_order;
  68                 int frame_type;
  69                 std::vector<float> audio;
  70                 int64_t pts, dts, duration;
  71                 movit::YCbCrLumaCoefficients ycbcr_coefficients;
  72                 ReceivedTimestamps received_ts;
  73                 std::vector<size_t> ref_display_frame_numbers;
  74         };
  75         struct PendingFrame {
  76                 RefCountedGLsync fence;
  77                 std::vector<RefCountedFrame> input_frames;
  78                 int64_t pts, duration;
  79                 movit::YCbCrLumaCoefficients ycbcr_coefficients;
  80         };
  81         struct GLSurface {
  82                 // Only if x264_video_to_disk == false.
  83                 VASurfaceID src_surface, ref_surface;
  84                 VABufferID coded_buf;
  85                 VAImage surface_image;
  86
  87                 // Only if use_zerocopy == true (which implies x264_video_to_disk == false).
  88                 GLuint y_tex, cbcr_tex;
  89                 EGLImage y_egl_image, cbcr_egl_image;
  90
  91                 // Only if use_zerocopy == false.
  92                 GLuint pbo;
  93                 uint8_t *y_ptr, *cbcr_ptr;
  94                 size_t y_offset, cbcr_offset;
  95
  96                 // Surfaces can be busy (have refcount > 0) for a variety of
  97                 // reasons: First of all because they belong to a frame that's
  98                 // under encoding. But also reference frames take refcounts;
  99                 // while a frame is being encoded, all its reference frames
 100                 // also have increased refcounts so that they are not dropped.
 101                 // Similarly, just being in <reference_frames> increases the
 102                 // refcount. Until it is back to zero, the surface cannot be given
 103                 // out for encoding another frame. Use release_gl_surface()
 104                 // to reduce the refcount, which will free the surface if
 105                 // the refcount reaches zero.
 106                 //
 107                 // Protected by storage_task_queue_mutex.
 108                 int refcount = 0;
 109         };
 110
 111         void open_output_file(const std::string &filename);
 112         void encode_thread_func();
 113         void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts);
 114         void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
 115         void pass_frame(PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration);
 116         void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
 117                           int frame_type, int64_t pts, int64_t dts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients);
 118         void storage_task_thread();
 119         void storage_task_enqueue(storage_task task);
 120         void save_codeddata(GLSurface *surf, storage_task task);
 121         int render_packedsequence(movit::YCbCrLumaCoefficients ycbcr_coefficients);
 122         int render_packedpicture();
 123         void render_packedslice();
 124         int render_sequence();
 125         int render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num);
 126         void sps_rbsp(movit::YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs);
 127         void pps_rbsp(bitstream *bs);
 128         int build_packed_pic_buffer(unsigned char **header_buffer);
 129         int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
 130         void slice_header(bitstream *bs);
 131         int build_packed_seq_buffer(movit::YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer);
 132         int build_packed_slice_buffer(unsigned char **header_buffer);
 133         int init_va(const std::string &va_display);
 134         void enable_zerocopy_if_possible();
 135         int setup_encode();
 136         void release_encode();
 137         void update_ReferenceFrames(int current_display_frame, int frame_type);
 138         void update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2]);
 139         void update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2]);
 140         GLSurface *allocate_gl_surface();
 141         void release_gl_surface(size_t display_frame_num);
 142
 143         bool is_shutdown = false;
 144         bool has_released_gl_resources = false;
 145         std::atomic<bool> use_zerocopy{false};
 146
 147         std::thread encode_thread, storage_thread;
 148
 149         std::mutex storage_task_queue_mutex;
 150         std::condition_variable storage_task_queue_changed;
 151         std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
 152         bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 153
 154         std::mutex frame_queue_mutex;
 155         std::condition_variable frame_queue_nonempty;
 156         bool encode_thread_should_quit = false;  // under frame_queue_mutex
 157
 158         int current_storage_frame;
 159
 160         PendingFrame current_video_frame;  // Used only between begin_frame() and end_frame().
 161         std::queue<PendingFrame> pending_video_frames;  // under frame_queue_mutex
 162         movit::ResourcePool *resource_pool;
 163         QSurface *surface;
 164
 165         // Frames that are done rendering and passed on to x264 (if enabled),
 166         // but have not been encoded by Quick Sync yet, and thus also not freed.
 167         // The key is the display frame number.
 168         std::map<int, PendingFrame> reorder_buffer;
 169         int quicksync_encoding_frame_num = 0;
 170
 171         std::mutex file_audio_encoder_mutex;
 172         std::unique_ptr<AudioEncoder> file_audio_encoder;
 173
 174         X264Encoder *x264_encoder;  // nullptr if not using x264.
 175
 176         Mux* stream_mux = nullptr;  // To HTTP.
 177         std::unique_ptr<Mux> file_mux;  // To local disk.
 178
 179         // Encoder parameters
 180         std::unique_ptr<VADisplayWithCleanup> va_dpy;
 181         VAProfile h264_profile = (VAProfile)~0;
 182         VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
 183         int config_attrib_num = 0, enc_packed_header_idx;
 184
 185         GLSurface gl_surfaces[SURFACE_NUM];
 186
 187         // For all frames in encoding (refcount > 0), a pointer into gl_surfaces
 188         // for the surface used for that frame. Protected by storage_task_queue_mutex.
 189         // The key is display frame number.
 190         std::unordered_map<size_t, GLSurface *> surface_for_frame;
 191
 192         VAConfigID config_id;
 193         VAContextID context_id;
 194         VAEncSequenceParameterBufferH264 seq_param;
 195         VAEncPictureParameterBufferH264 pic_param;
 196         VAEncSliceParameterBufferH264 slice_param;
 197         VAPictureH264 CurrentCurrPic;
 198
 199         struct ReferenceFrame {
 200                 VAPictureH264 pic;
 201                 int display_number;  // To track reference counts.
 202         };
 203         std::deque<ReferenceFrame> reference_frames;
 204
 205         // Static quality settings.
 206         static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
 207         static constexpr unsigned int num_ref_frames = 2;
 208         static constexpr int initial_qp = 15;
 209         static constexpr int minimal_qp = 0;
 210         static constexpr int intra_period = 30;
 211         static constexpr int intra_idr_period = MAX_FPS;  // About a second; more at lower frame rates. Not ideal.
 212
 213         // Quality settings that are meant to be static, but might be overridden
 214         // by the profile.
 215         int constraint_set_flag = 0;
 216         int h264_packedheader = 0; /* support pack header? */
 217         int h264_maxref = (1<<16|1);
 218         int h264_entropy_mode = 1; /* cabac */
 219         int ip_period = 3;
 220
 221         unsigned int current_ref_frame_num = 0;  // Encoding frame order within this GOP, sans B-frames.
 222
 223         int frame_width;
 224         int frame_height;
 225         int frame_width_mbaligned;
 226         int frame_height_mbaligned;
 227
 228         DiskSpaceEstimator *disk_space_estimator;
 229 };
 230
 231 #endif  // !defined(_QUICKSYNC_ENCODER_IMPL_H)