]> git.sesse.net Git - nageru/blobdiff - quicksync_encoder_impl.h
Add functionality for MJPEG export.
[nageru] / quicksync_encoder_impl.h
index 72eb3b68d97bc4b5168082db9e0c0a8d6640f1cc..0317b6af0393723bb1ca38ac0100947d8c800c85 100644 (file)
@@ -2,6 +2,7 @@
 #define _QUICKSYNC_ENCODER_IMPL_H 1
 
 #include <epoxy/egl.h>
+#include <movit/image_format.h>
 #include <va/va.h>
 
 #include <condition_variable>
 #include <string>
 #include <stack>
 #include <thread>
+#include <unordered_map>
 
 #include "audio_encoder.h"
 #include "defs.h"
 #include "timebase.h"
 #include "print_latency.h"
+#include "ref_counted_gl_sync.h"
+#include "va_display_with_cleanup.h"
 
 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
 #define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
@@ -29,14 +33,23 @@ struct __bitstream {
 };
 typedef struct __bitstream bitstream;
 
+namespace movit {
+class ResourcePool;
+}
+class DiskSpaceEstimator;
+class QSurface;
+class X264Encoder;
+
 class QuickSyncEncoderImpl {
 public:
        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);
        ~QuickSyncEncoderImpl();
        void add_audio(int64_t pts, std::vector<float> audio);
-       bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex);
-       RefCountedGLsync end_frame(int64_t pts, int64_t duration, const std::vector<RefCountedFrame> &input_frames);
+       bool is_zerocopy() const;
+       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);
+       RefCountedGLsync end_frame();
        void shutdown();
+       void close_file();
        void release_gl_resources();
        void set_stream_mux(Mux *mux)
        {
@@ -54,12 +67,44 @@ private:
                int frame_type;
                std::vector<float> audio;
                int64_t pts, dts, duration;
+               movit::YCbCrLumaCoefficients ycbcr_coefficients;
                ReceivedTimestamps received_ts;
+               std::vector<size_t> ref_display_frame_numbers;
        };
        struct PendingFrame {
                RefCountedGLsync fence;
                std::vector<RefCountedFrame> input_frames;
                int64_t pts, duration;
+               movit::YCbCrLumaCoefficients ycbcr_coefficients;
+       };
+       struct GLSurface {
+               // Only if x264_video_to_disk == false.
+               VASurfaceID src_surface, ref_surface;
+               VABufferID coded_buf;
+               VAImage surface_image;
+
+               // Only if use_zerocopy == true (which implies x264_video_to_disk == false).
+               GLuint y_tex, cbcr_tex;
+               EGLImage y_egl_image, cbcr_egl_image;
+
+               // Only if use_zerocopy == false.
+               GLuint pbo;
+               uint8_t *y_ptr, *cbcr_ptr;
+               size_t y_offset, cbcr_offset;
+
+               // Surfaces can be busy (have refcount > 0) for a variety of
+               // reasons: First of all because they belong to a frame that's
+               // under encoding. But also reference frames take refcounts;
+               // while a frame is being encoded, all its reference frames
+               // also have increased refcounts so that they are not dropped.
+               // Similarly, just being in <reference_frames> increases the
+               // refcount. Until it is back to zero, the surface cannot be given
+               // out for encoding another frame. Use release_gl_surface()
+               // to reduce the refcount, which will free the surface if
+               // the refcount reaches zero.
+               //
+               // Protected by storage_task_queue_mutex.
+               int refcount = 0;
        };
 
        void open_output_file(const std::string &filename);
@@ -68,42 +113,40 @@ private:
        void add_packet_for_uncompressed_frame(int64_t pts, int64_t duration, const uint8_t *data);
        void pass_frame(PendingFrame frame, int display_frame_num, int64_t pts, int64_t duration);
        void encode_frame(PendingFrame frame, int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num,
-                         int frame_type, int64_t pts, int64_t dts, int64_t duration);
+                         int frame_type, int64_t pts, int64_t dts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients);
        void storage_task_thread();
        void storage_task_enqueue(storage_task task);
-       void save_codeddata(storage_task task);
-       int render_packedsequence();
+       void save_codeddata(GLSurface *surf, storage_task task);
+       int render_packedsequence(movit::YCbCrLumaCoefficients ycbcr_coefficients);
        int render_packedpicture();
        void render_packedslice();
        int render_sequence();
-       int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num);
-       void sps_rbsp(bitstream *bs);
+       int render_picture(GLSurface *surf, int frame_type, int display_frame_num, int gop_start_display_frame_num);
+       void sps_rbsp(movit::YCbCrLumaCoefficients ycbcr_coefficients, bitstream *bs);
        void pps_rbsp(bitstream *bs);
        int build_packed_pic_buffer(unsigned char **header_buffer);
        int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type);
        void slice_header(bitstream *bs);
-       int build_packed_seq_buffer(unsigned char **header_buffer);
+       int build_packed_seq_buffer(movit::YCbCrLumaCoefficients ycbcr_coefficients, unsigned char **header_buffer);
        int build_packed_slice_buffer(unsigned char **header_buffer);
        int init_va(const std::string &va_display);
-       int deinit_va();
        void enable_zerocopy_if_possible();
-       VADisplay va_open_display(const std::string &va_display);
-       void va_close_display(VADisplay va_dpy);
        int setup_encode();
        void release_encode();
-       void update_ReferenceFrames(int frame_type);
-       int update_RefPicList(int frame_type);
+       void update_ReferenceFrames(int current_display_frame, int frame_type);
+       void update_RefPicList_P(VAPictureH264 RefPicList0_P[MAX_NUM_REF2]);
+       void update_RefPicList_B(VAPictureH264 RefPicList0_B[MAX_NUM_REF2], VAPictureH264 RefPicList1_B[MAX_NUM_REF2]);
+       GLSurface *allocate_gl_surface();
+       void release_gl_surface(size_t display_frame_num);
 
        bool is_shutdown = false;
        bool has_released_gl_resources = false;
-       bool use_zerocopy;
-       int drm_fd = -1;
+       std::atomic<bool> use_zerocopy{false};
 
        std::thread encode_thread, storage_thread;
 
        std::mutex storage_task_queue_mutex;
        std::condition_variable storage_task_queue_changed;
-       int srcsurface_status[SURFACE_NUM];  // protected by storage_task_queue_mutex
        std::queue<storage_task> storage_task_queue;  // protected by storage_task_queue_mutex
        bool storage_thread_should_quit = false;  // protected by storage_task_queue_mutex
 
@@ -113,6 +156,7 @@ private:
 
        int current_storage_frame;
 
+       PendingFrame current_video_frame;  // Used only between begin_frame() and end_frame().
        std::queue<PendingFrame> pending_video_frames;  // under frame_queue_mutex
        movit::ResourcePool *resource_pool;
        QSurface *surface;
@@ -123,6 +167,7 @@ private:
        std::map<int, PendingFrame> reorder_buffer;
        int quicksync_encoding_frame_num = 0;
 
+       std::mutex file_audio_encoder_mutex;
        std::unique_ptr<AudioEncoder> file_audio_encoder;
 
        X264Encoder *x264_encoder;  // nullptr if not using x264.
@@ -130,38 +175,31 @@ private:
        Mux* stream_mux = nullptr;  // To HTTP.
        std::unique_ptr<Mux> file_mux;  // To local disk.
 
-       Display *x11_display = nullptr;
-
        // Encoder parameters
-       VADisplay va_dpy;
+       std::unique_ptr<VADisplayWithCleanup> va_dpy;
        VAProfile h264_profile = (VAProfile)~0;
        VAConfigAttrib config_attrib[VAConfigAttribTypeMax];
        int config_attrib_num = 0, enc_packed_header_idx;
 
-       struct GLSurface {
-               VASurfaceID src_surface, ref_surface;
-               VABufferID coded_buf;
-
-               VAImage surface_image;
-               GLuint y_tex, cbcr_tex;
-
-               // Only if use_zerocopy == true.
-               EGLImage y_egl_image, cbcr_egl_image;
-
-               // Only if use_zerocopy == false.
-               GLuint pbo;
-               uint8_t *y_ptr, *cbcr_ptr;
-               size_t y_offset, cbcr_offset;
-       };
        GLSurface gl_surfaces[SURFACE_NUM];
 
+       // For all frames in encoding (refcount > 0), a pointer into gl_surfaces
+       // for the surface used for that frame. Protected by storage_task_queue_mutex.
+       // The key is display frame number.
+       std::unordered_map<size_t, GLSurface *> surface_for_frame;
+
        VAConfigID config_id;
        VAContextID context_id;
        VAEncSequenceParameterBufferH264 seq_param;
        VAEncPictureParameterBufferH264 pic_param;
        VAEncSliceParameterBufferH264 slice_param;
        VAPictureH264 CurrentCurrPic;
-       VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
+
+       struct ReferenceFrame {
+               VAPictureH264 pic;
+               int display_number;  // To track reference counts.
+       };
+       std::deque<ReferenceFrame> reference_frames;
 
        // Static quality settings.
        static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
@@ -179,9 +217,7 @@ private:
        int h264_entropy_mode = 1; /* cabac */
        int ip_period = 3;
 
-       int rc_mode = -1;
-       unsigned int current_frame_num = 0;
-       unsigned int numShortTerm = 0;
+       unsigned int current_ref_frame_num = 0;  // Encoding frame order within this GOP, sans B-frames.
 
        int frame_width;
        int frame_height;