--- /dev/null
+// An AV1 encoder using the SVT-AV1 encoder library. (libaom does not seem
+// to be suitable for real-time streaming as of 2022, as it is not using
+// threads particularly efficiently.) AV1 is a newer format than H.264,
+// obviously both for better and for worse: Higher coding efficiency
+// (with sufficient amount of cores, SVT-AV1 is even better than x264
+// at higher frame rates), but generally smaller ecosystem, no speed
+// control, less sophisticated rate control, etc..
+//
+// We don't support storing AV1 to disk currently, only to HTTP.
+// We also don't support hardware AV1 encoding, as hardware supporting it
+// is very rare currently.
+//
+// Since SVT-AV1 does not support VFR, you need to declare the frame rate
+// up-front, using the --av1-framerate flag. If the given frame rate is
+// too high (ie., you are producing frames too slowly), rate control will get
+// confused and use too little bitrate. If it is too low, Nageru will need to
+// drop frames before input to the encoder.
+
+#ifndef _AV1_ENCODER_H
+#define _AV1_ENCODER_H 1
+
+#include <sched.h>
+#include <stdint.h>
+#include <atomic>
+#include <condition_variable>
+#include <memory>
+#include <mutex>
+#include <queue>
+#include <string>
+#include <thread>
+#include <unordered_map>
+#include <vector>
+
+extern "C" {
+#include <libavformat/avformat.h>
+}
+
+#include <movit/image_format.h>
+
+#include "defs.h"
+#include "shared/metrics.h"
+#include "print_latency.h"
+#include "video_codec_interface.h"
+
+class Mux;
+struct EbBufferHeaderType;
+struct EbComponentType;
+
+class AV1Encoder : public VideoCodecInterface {
+public:
+ AV1Encoder(const AVOutputFormat *oformat); // Does not take ownership.
+
+ // Called after the last frame. Will block; once this returns,
+ // the last data is flushed.
+ ~AV1Encoder() override;
+
+ // Must be called before first frame. Does not take ownership.
+ void add_mux(Mux *mux) override { muxes.push_back(mux); }
+
+ // <data> is taken to be raw NV12 data of WIDTHxHEIGHT resolution.
+ // Does not block.
+ void add_frame(int64_t pts, int64_t duration, movit::YCbCrLumaCoefficients ycbcr_coefficients, const uint8_t *data, const ReceivedTimestamps &received_ts) override;
+
+ std::string get_global_headers() const override {
+ while (!av1_init_done) {
+ sched_yield();
+ }
+ return global_headers;
+ }
+
+private:
+ struct QueuedFrame {
+ int64_t pts, duration;
+ movit::YCbCrLumaCoefficients ycbcr_coefficients;
+ uint8_t *data;
+ ReceivedTimestamps received_ts;
+ };
+ void encoder_thread_func();
+ void init_av1();
+ void encode_frame(QueuedFrame qf);
+ void process_packet(EbBufferHeaderType *buf); // Takes ownership.
+
+ // One big memory chunk of all 50 (or whatever) frames, allocated in
+ // the constructor. All data functions just use pointers into this
+ // pool.
+ std::unique_ptr<uint8_t[]> frame_pool;
+
+ std::vector<Mux *> muxes;
+ const bool wants_global_headers;
+
+ std::string global_headers;
+
+ std::thread encoder_thread;
+ std::atomic<bool> av1_init_done{false};
+ std::atomic<bool> should_quit{false};
+ EbComponentType *encoder;
+
+ int64_t last_pts = -1;
+
+ // Protects everything below it.
+ std::mutex mu;
+
+ // Frames that are not being encoded or waiting to be encoded,
+ // so that add_frame() can use new ones.
+ // TODO: Do we actually need this queue, or is SVT-AV1's queue
+ // non-blocking so that we can drop it?
+ std::queue<uint8_t *> free_frames;
+
+ // Frames that are waiting to be encoded (ie., add_frame() has been
+ // called, but they are not picked up for encoding yet).
+ std::queue<QueuedFrame> queued_frames;
+
+ // Whenever the state of <queued_frames> changes.
+ std::condition_variable queued_frames_nonempty;
+
+ // Key is the pts of the frame.
+ std::unordered_map<int64_t, ReceivedTimestamps> frames_being_encoded;
+};
+
+#endif // !defined(_AV1_ENCODER_H)