X-Git-Url: https://git.sesse.net/?a=blobdiff_plain;f=h264encode.cpp;h=74bf947f84b06734eb0ffdff475fe6b3f496add4;hb=a9313d3eb914641cbdf8f2426d1fa79b45b55107;hp=3520158857c28bd02af855f67adfbaa7d5cc1f5d;hpb=b300bb0e8f761d39dba2fedaad28ec5537059cf2;p=nageru diff --git a/h264encode.cpp b/h264encode.cpp index 3520158..74bf947 100644 --- a/h264encode.cpp +++ b/h264encode.cpp @@ -6,11 +6,14 @@ #include #include #include +extern "C" { #include +#include #include #include #include #include +} #include #include #include @@ -21,6 +24,7 @@ #include #include #include +#include #include #include #include @@ -131,10 +135,7 @@ static int rc_default_modes[] = { VA_RC_VCM, VA_RC_NONE, }; -static unsigned long long current_frame_display = 0; -static unsigned long long current_IDR_display = 0; static unsigned int current_frame_num = 0; -static int current_frame_type; static int misc_priv_type = 0; static int misc_priv_value = 0; @@ -152,6 +153,60 @@ typedef struct __bitstream bitstream; using namespace std; +class H264EncoderImpl { +public: + H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd); + ~H264EncoderImpl(); + void add_audio(int64_t pts, std::vector audio); // Needs to come before end_frame() of same pts. + bool begin_frame(GLuint *y_tex, GLuint *cbcr_tex); + void end_frame(RefCountedGLsync fence, int64_t pts, const std::vector &input_frames); + +private: + struct storage_task { + unsigned long long display_order; + int frame_type; + std::vector audio; + int64_t pts, dts; + }; + struct PendingFrame { + RefCountedGLsync fence; + std::vector input_frames; + int64_t pts; + }; + + void encode_thread_func(); + void encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts); + 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); + void storage_task_thread(); + void storage_task_enqueue(storage_task task); + void save_codeddata(storage_task task); + + 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_queue; // protected by storage_task_queue_mutex + bool storage_thread_should_quit = false; // protected by storage_task_queue_mutex + + std::mutex frame_queue_mutex; + std::condition_variable frame_queue_nonempty; + bool encode_thread_should_quit = false; // under frame_queue_mutex + + //int frame_width, frame_height; + //int ; + int current_storage_frame; + + std::map pending_video_frames; // under frame_queue_mutex + std::map> pending_audio_frames; // under frame_queue_mutex + QSurface *surface; + + AVCodecContext *context_audio; + HTTPD *httpd; +}; + + // Supposedly vaRenderPicture() is supposed to destroy the buffer implicitly, // but if we don't delete it here, we get leaks. The GStreamer implementation // does the same. @@ -662,9 +717,9 @@ static int build_packed_slice_buffer(unsigned char **header_buffer) #define FRAME_I 2 #define FRAME_IDR 7 void encoding2display_order( - unsigned long long encoding_order, int intra_period, + int encoding_order, int intra_period, int intra_idr_period, int ip_period, - unsigned long long *displaying_order, + int *displaying_order, int *frame_type, int *pts_lag) { int encoding_order_gop = 0; @@ -1209,12 +1264,12 @@ static void sort_two(VAPictureH264 ref[], int left, int right, unsigned int key, sort_one(ref, j+1, right, list1_ascending, frame_idx); } -static int update_ReferenceFrames(void) +static void update_ReferenceFrames(int frame_type) { int i; - if (current_frame_type == FRAME_B) - return 0; + if (frame_type == FRAME_B) + return; CurrentCurrPic.flags = VA_PICTURE_H264_SHORT_TERM_REFERENCE; numShortTerm++; @@ -1224,25 +1279,22 @@ static int update_ReferenceFrames(void) ReferenceFrames[i] = ReferenceFrames[i-1]; ReferenceFrames[0] = CurrentCurrPic; - if (current_frame_type != FRAME_B) - current_frame_num++; + current_frame_num++; if (current_frame_num > MaxFrameNum) current_frame_num = 0; - - return 0; } -static int update_RefPicList(void) +static int update_RefPicList(int frame_type) { unsigned int current_poc = CurrentCurrPic.TopFieldOrderCnt; - if (current_frame_type == FRAME_P) { + if (frame_type == FRAME_P) { memcpy(RefPicList0_P, ReferenceFrames, numShortTerm * sizeof(VAPictureH264)); sort_one(RefPicList0_P, 0, numShortTerm-1, 0, 1); } - if (current_frame_type == FRAME_B) { + if (frame_type == FRAME_B) { memcpy(RefPicList0_B, ReferenceFrames, numShortTerm * sizeof(VAPictureH264)); sort_two(RefPicList0_B, 0, numShortTerm-1, current_poc, 0, 1, 0, 1); @@ -1336,13 +1388,13 @@ static int render_sequence(void) return 0; } -static int calc_poc(int pic_order_cnt_lsb) +static int calc_poc(int pic_order_cnt_lsb, int frame_type) { static int PicOrderCntMsb_ref = 0, pic_order_cnt_lsb_ref = 0; int prevPicOrderCntMsb, prevPicOrderCntLsb; int PicOrderCntMsb, TopFieldOrderCnt; - if (current_frame_type == FRAME_IDR) + if (frame_type == FRAME_IDR) prevPicOrderCntMsb = prevPicOrderCntLsb = 0; else { prevPicOrderCntMsb = PicOrderCntMsb_ref; @@ -1360,7 +1412,7 @@ static int calc_poc(int pic_order_cnt_lsb) TopFieldOrderCnt = PicOrderCntMsb + pic_order_cnt_lsb; - if (current_frame_type != FRAME_B) { + if (frame_type != FRAME_B) { PicOrderCntMsb_ref = PicOrderCntMsb; pic_order_cnt_lsb_ref = pic_order_cnt_lsb; } @@ -1368,42 +1420,31 @@ static int calc_poc(int pic_order_cnt_lsb) return TopFieldOrderCnt; } -static int render_picture(void) +static int render_picture(int frame_type, int display_frame_num, int gop_start_display_frame_num) { VABufferID pic_param_buf; VAStatus va_status; int i = 0; - pic_param.CurrPic.picture_id = gl_surfaces[current_frame_display % SURFACE_NUM].ref_surface; + pic_param.CurrPic.picture_id = gl_surfaces[display_frame_num % SURFACE_NUM].ref_surface; pic_param.CurrPic.frame_idx = current_frame_num; pic_param.CurrPic.flags = 0; - pic_param.CurrPic.TopFieldOrderCnt = calc_poc((current_frame_display - current_IDR_display) % MaxPicOrderCntLsb); + pic_param.CurrPic.TopFieldOrderCnt = calc_poc((display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb, frame_type); pic_param.CurrPic.BottomFieldOrderCnt = pic_param.CurrPic.TopFieldOrderCnt; CurrentCurrPic = pic_param.CurrPic; - if (getenv("TO_DEL")) { /* set RefPicList into ReferenceFrames */ - update_RefPicList(); /* calc RefPicList */ - memset(pic_param.ReferenceFrames, 0xff, 16 * sizeof(VAPictureH264)); /* invalid all */ - if (current_frame_type == FRAME_P) { - pic_param.ReferenceFrames[0] = RefPicList0_P[0]; - } else if (current_frame_type == FRAME_B) { - pic_param.ReferenceFrames[0] = RefPicList0_B[0]; - pic_param.ReferenceFrames[1] = RefPicList1_B[0]; - } - } else { - memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264)); - for (i = numShortTerm; i < SURFACE_NUM; i++) { - pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE; - pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID; - } + memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264)); + for (i = numShortTerm; i < SURFACE_NUM; i++) { + pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE; + pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID; } - pic_param.pic_fields.bits.idr_pic_flag = (current_frame_type == FRAME_IDR); - pic_param.pic_fields.bits.reference_pic_flag = (current_frame_type != FRAME_B); + pic_param.pic_fields.bits.idr_pic_flag = (frame_type == FRAME_IDR); + pic_param.pic_fields.bits.reference_pic_flag = (frame_type != FRAME_B); pic_param.pic_fields.bits.entropy_coding_mode_flag = h264_entropy_mode; pic_param.pic_fields.bits.deblocking_filter_control_present_flag = 1; pic_param.frame_num = current_frame_num; - pic_param.coded_buf = gl_surfaces[current_frame_display % SURFACE_NUM].coded_buf; + pic_param.coded_buf = gl_surfaces[display_frame_num % SURFACE_NUM].coded_buf; pic_param.last_picture = false; // FIXME pic_param.pic_init_qp = initial_qp; @@ -1524,22 +1565,22 @@ static void render_packedslice() free(packedslice_buffer); } -static int render_slice(int encoding_frame_num) +static int render_slice(int encoding_frame_num, int display_frame_num, int gop_start_display_frame_num, int frame_type) { VABufferID slice_param_buf; VAStatus va_status; int i; - update_RefPicList(); + update_RefPicList(frame_type); /* one frame, one slice */ slice_param.macroblock_address = 0; slice_param.num_macroblocks = frame_width_mbaligned * frame_height_mbaligned/(16*16); /* Measured by MB */ - slice_param.slice_type = (current_frame_type == FRAME_IDR)?2:current_frame_type; - if (current_frame_type == FRAME_IDR) { + slice_param.slice_type = (frame_type == FRAME_IDR)?2:frame_type; + if (frame_type == FRAME_IDR) { if (encoding_frame_num != 0) ++slice_param.idr_pic_id; - } else if (current_frame_type == FRAME_P) { + } else if (frame_type == FRAME_P) { int refpiclist0_max = h264_maxref & 0xffff; memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264)); @@ -1547,7 +1588,7 @@ static int render_slice(int encoding_frame_num) slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE; slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID; } - } else if (current_frame_type == FRAME_B) { + } else if (frame_type == FRAME_B) { int refpiclist0_max = h264_maxref & 0xffff; int refpiclist1_max = (h264_maxref >> 16) & 0xffff; @@ -1567,7 +1608,7 @@ static int render_slice(int encoding_frame_num) slice_param.slice_alpha_c0_offset_div2 = 0; slice_param.slice_beta_offset_div2 = 0; slice_param.direct_spatial_mv_pred_flag = 1; - slice_param.pic_order_cnt_lsb = (current_frame_display - current_IDR_display) % MaxPicOrderCntLsb; + slice_param.pic_order_cnt_lsb = (display_frame_num - gop_start_display_frame_num) % MaxPicOrderCntLsb; if (h264_packedheader && @@ -1585,7 +1626,7 @@ static int render_slice(int encoding_frame_num) -void H264Encoder::save_codeddata(storage_task task) +void H264EncoderImpl::save_codeddata(storage_task task) { VACodedBufferSegment *buf_list = NULL; VAStatus va_status; @@ -1627,8 +1668,8 @@ void H264Encoder::save_codeddata(storage_task task) vector audio; { unique_lock lock(frame_queue_mutex); - frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || !pending_audio_frames.empty(); }); - if (copy_thread_should_quit) return; + frame_queue_nonempty.wait(lock, [this]{ return storage_thread_should_quit || !pending_audio_frames.empty(); }); + if (storage_thread_should_quit && pending_audio_frames.empty()) return; auto it = pending_audio_frames.begin(); if (it->first > task.pts) break; audio_pts = it->first; @@ -1696,7 +1737,7 @@ void H264Encoder::save_codeddata(storage_task task) // this is weird. but it seems to put a new frame onto the queue -void H264Encoder::storage_task_enqueue(storage_task task) +void H264EncoderImpl::storage_task_enqueue(storage_task task) { unique_lock lock(storage_task_queue_mutex); storage_task_queue.push(move(task)); @@ -1704,7 +1745,7 @@ void H264Encoder::storage_task_enqueue(storage_task task) storage_task_queue_changed.notify_all(); } -void H264Encoder::storage_task_thread() +void H264EncoderImpl::storage_task_thread() { for ( ;; ) { storage_task current; @@ -1712,7 +1753,7 @@ void H264Encoder::storage_task_thread() // wait until there's an encoded frame unique_lock lock(storage_task_queue_mutex); storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || !storage_task_queue.empty(); }); - if (storage_thread_should_quit) return; + if (storage_thread_should_quit && storage_task_queue.empty()) return; current = move(storage_task_queue.front()); storage_task_queue.pop(); } @@ -1758,7 +1799,7 @@ static int deinit_va() } -H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd) +H264EncoderImpl::H264EncoderImpl(QSurface *surface, int width, int height, HTTPD *httpd) : current_storage_frame(0), surface(surface), httpd(httpd) { AVCodec *codec_audio = avcodec_find_encoder(AUDIO_OUTPUT_CODEC); @@ -1792,9 +1833,9 @@ H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd) memset(&pic_param, 0, sizeof(pic_param)); memset(&slice_param, 0, sizeof(slice_param)); - storage_thread = thread(&H264Encoder::storage_task_thread, this); + storage_thread = thread(&H264EncoderImpl::storage_task_thread, this); - copy_thread = thread([this]{ + encode_thread = thread([this]{ //SDL_GL_MakeCurrent(window, context); QOpenGLContext *context = create_context(this->surface); eglBindAPI(EGL_OPENGL_API); @@ -1803,30 +1844,31 @@ H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd) eglGetError()); exit(1); } - copy_thread_func(); + encode_thread_func(); }); } -H264Encoder::~H264Encoder() +H264EncoderImpl::~H264EncoderImpl() { { - unique_lock lock(storage_task_queue_mutex); - storage_thread_should_quit = true; - storage_task_queue_changed.notify_all(); + unique_lock lock(frame_queue_mutex); + encode_thread_should_quit = true; + frame_queue_nonempty.notify_all(); } + encode_thread.join(); { - unique_lock lock(frame_queue_mutex); - copy_thread_should_quit = true; + unique_lock lock(storage_task_queue_mutex); + storage_thread_should_quit = true; frame_queue_nonempty.notify_all(); + storage_task_queue_changed.notify_all(); } storage_thread.join(); - copy_thread.join(); release_encode(); deinit_va(); } -bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) +bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) { { // Wait until this frame slot is done encoding. @@ -1890,7 +1932,7 @@ bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) return true; } -void H264Encoder::add_audio(int64_t pts, vector audio) +void H264EncoderImpl::add_audio(int64_t pts, vector audio) { { unique_lock lock(frame_queue_mutex); @@ -1899,7 +1941,7 @@ void H264Encoder::add_audio(int64_t pts, vector audio) frame_queue_nonempty.notify_all(); } -void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector &input_frames) +void H264EncoderImpl::end_frame(RefCountedGLsync fence, int64_t pts, const vector &input_frames) { { unique_lock lock(frame_queue_mutex); @@ -1909,28 +1951,38 @@ void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector lock(frame_queue_mutex); - frame_queue_nonempty.wait(lock, [this]{ return copy_thread_should_quit || pending_video_frames.count(current_frame_display) != 0; }); - if (copy_thread_should_quit) { + frame_queue_nonempty.wait(lock, [this, display_frame_num]{ + return encode_thread_should_quit || pending_video_frames.count(display_frame_num) != 0; + }); + if (encode_thread_should_quit && pending_video_frames.count(display_frame_num) == 0) { + // We have queued frames that were supposed to be B-frames, + // but will be no P-frame to encode them against. Encode them all + // as P-frames instead. Note that this happens under the mutex, + // but nobody else uses it at this point, since we're shutting down, + // so there's no contention. + encode_remaining_frames_as_p(encoding_frame_num, gop_start_display_frame_num, last_dts); return; } else { - frame = move(pending_video_frames[current_frame_display]); - pending_video_frames.erase(current_frame_display); + frame = move(pending_video_frames[display_frame_num]); + pending_video_frames.erase(display_frame_num); } } @@ -1944,11 +1996,29 @@ void H264Encoder::copy_thread_func() } last_dts = dts; - encode_frame(frame, encoding_frame_num, frame.pts, dts); + encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts); } } -void H264Encoder::encode_frame(H264Encoder::PendingFrame frame, int encoding_frame_num, int64_t pts, int64_t dts) +void H264EncoderImpl::encode_remaining_frames_as_p(int encoding_frame_num, int gop_start_display_frame_num, int64_t last_dts) +{ + if (pending_video_frames.empty()) { + return; + } + + for (auto &pending_frame : pending_video_frames) { + int display_frame_num = pending_frame.first; + assert(display_frame_num > 0); + PendingFrame frame = move(pending_frame.second); + int64_t dts = last_dts + (TIMEBASE / MAX_FPS); + printf("Finalizing encode: Encoding leftover frame %d as P-frame instead of B-frame.\n", display_frame_num); + encode_frame(frame, encoding_frame_num++, display_frame_num, gop_start_display_frame_num, FRAME_P, frame.pts, dts); + last_dts = dts; + } +} + +void H264EncoderImpl::encode_frame(H264EncoderImpl::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) { // Wait for the GPU to be done with the frame. glClientWaitSync(frame.fence.get(), 0, 0); @@ -1957,7 +2027,7 @@ void H264Encoder::encode_frame(H264Encoder::PendingFrame frame, int encoding_fra frame.input_frames.clear(); // Unmap the image. - GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM]; + GLSurface *surf = &gl_surfaces[display_frame_num % SURFACE_NUM]; eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image); eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image); VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf); @@ -1971,18 +2041,18 @@ void H264Encoder::encode_frame(H264Encoder::PendingFrame frame, int encoding_fra va_status = vaBeginPicture(va_dpy, context_id, surface); CHECK_VASTATUS(va_status, "vaBeginPicture"); - if (current_frame_type == FRAME_IDR) { + if (frame_type == FRAME_IDR) { render_sequence(); - render_picture(); + render_picture(frame_type, display_frame_num, gop_start_display_frame_num); if (h264_packedheader) { render_packedsequence(); render_packedpicture(); } } else { //render_sequence(); - render_picture(); + render_picture(frame_type, display_frame_num, gop_start_display_frame_num); } - render_slice(encoding_frame_num); + render_slice(encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type); va_status = vaEndPicture(va_dpy, context_id); CHECK_VASTATUS(va_status, "vaEndPicture"); @@ -1990,11 +2060,35 @@ void H264Encoder::encode_frame(H264Encoder::PendingFrame frame, int encoding_fra // so now the data is done encoding (well, async job kicked off)... // we send that to the storage thread storage_task tmp; - tmp.display_order = current_frame_display; - tmp.frame_type = current_frame_type; + tmp.display_order = display_frame_num; + tmp.frame_type = frame_type; tmp.pts = pts; tmp.dts = dts; storage_task_enqueue(move(tmp)); - update_ReferenceFrames(); + update_ReferenceFrames(frame_type); } + +// Proxy object. +H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd) + : impl(new H264EncoderImpl(surface, width, height, httpd)) {} + +// Must be defined here because unique_ptr<> destructor needs to know the impl. +H264Encoder::~H264Encoder() {} + +void H264Encoder::add_audio(int64_t pts, std::vector audio) +{ + impl->add_audio(pts, audio); +} + +bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex) +{ + return impl->begin_frame(y_tex, cbcr_tex); +} + +void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector &input_frames) +{ + impl->end_frame(fence, pts, input_frames); +} + +// Real class.