]> git.sesse.net Git - nageru/blobdiff - h264encode.cpp
If we have waiting B-frames at the end of the encode, encode them as such.
[nageru] / h264encode.cpp
index d263b3a44105d1672cb15b4b31d7f25621f666ae..d0da39d08c0925e79ab9c9cef418257c3b95bb2e 100644 (file)
@@ -33,6 +33,8 @@
 #include "httpd.h"
 #include "timebase.h"
 
+using namespace std;
+
 class QOpenGLContext;
 class QSurface;
 
@@ -129,11 +131,7 @@ static  int rc_default_modes[] = {
     VA_RC_VCM,
     VA_RC_NONE,
 };
-static  unsigned long long current_frame_encoding = 0;
-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;
@@ -661,9 +659,9 @@ 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;
@@ -1208,12 +1206,12 @@ 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++;
@@ -1223,25 +1221,22 @@ 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);
@@ -1335,13 +1330,13 @@ 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;
@@ -1359,7 +1354,7 @@ 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;
     }
@@ -1367,42 +1362,31 @@ 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;
 
@@ -1523,22 +1507,22 @@ static void render_packedslice()
     free(packedslice_buffer);
 }
 
-static int render_slice(void)
+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) {
-        if (current_frame_encoding != 0)
+    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));
 
@@ -1546,7 +1530,7 @@ static int render_slice(void)
             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;
 
@@ -1566,7 +1550,7 @@ static int render_slice(void)
     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 &&
@@ -1623,11 +1607,11 @@ void H264Encoder::save_codeddata(storage_task task)
     // Encode and add all audio frames up to and including the pts of this video frame.
     for ( ;; ) {
         int64_t audio_pts;
-        std::vector<float> audio;
+        vector<float> audio;
         {
              unique_lock<mutex> 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;
@@ -1697,7 +1681,7 @@ 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)
 {
-       std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
+       unique_lock<mutex> lock(storage_task_queue_mutex);
        storage_task_queue.push(move(task));
        srcsurface_status[task.display_order % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
        storage_task_queue_changed.notify_all();
@@ -1709,9 +1693,9 @@ void H264Encoder::storage_task_thread()
                storage_task current;
                {
                        // wait until there's an encoded frame  
-                       std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
+                       unique_lock<mutex> 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();
                }
@@ -1724,7 +1708,7 @@ void H264Encoder::storage_task_thread()
                save_codeddata(move(current));
 
                {
-                       std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
+                       unique_lock<mutex> lock(storage_task_queue_mutex);
                        srcsurface_status[current.display_order % SURFACE_NUM] = SRC_SURFACE_FREE;
                        storage_task_queue_changed.notify_all();
                }
@@ -1778,7 +1762,6 @@ H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
        frame_width_mbaligned = (frame_width + 15) & (~15);
        frame_height_mbaligned = (frame_height + 15) & (~15);
         frame_bitrate = 15000000;  // / 60;
-       current_frame_encoding = 0;
 
        //print_input();
 
@@ -1792,9 +1775,9 @@ 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 = std::thread(&H264Encoder::storage_task_thread, this);
+       storage_thread = thread(&H264Encoder::storage_task_thread, this);
 
-       copy_thread = std::thread([this]{
+       encode_thread = thread([this]{
                //SDL_GL_MakeCurrent(window, context);
                QOpenGLContext *context = create_context(this->surface);
                eglBindAPI(EGL_OPENGL_API);
@@ -1803,24 +1786,25 @@ H264Encoder::H264Encoder(QSurface *surface, int width, int height, HTTPD *httpd)
                                eglGetError());
                        exit(1);
                }
-               copy_thread_func();
+               encode_thread_func();
        });
 }
 
 H264Encoder::~H264Encoder()
 {
        {
-               unique_lock<mutex> lock(storage_task_queue_mutex);
-               storage_thread_should_quit = true;
-               storage_task_queue_changed.notify_all();
+               unique_lock<mutex> lock(frame_queue_mutex);
+               encode_thread_should_quit = true;
+               frame_queue_nonempty.notify_all();
        }
+       encode_thread.join();
        {
-               unique_lock<mutex> lock(frame_queue_mutex);
-               copy_thread_should_quit = true;
+               unique_lock<mutex> 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();
@@ -1830,7 +1814,7 @@ bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
 {
        {
                // Wait until this frame slot is done encoding.
-               std::unique_lock<std::mutex> lock(storage_task_queue_mutex);
+               unique_lock<mutex> lock(storage_task_queue_mutex);
                storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
                if (storage_thread_should_quit) return false;
        }
@@ -1890,7 +1874,7 @@ bool H264Encoder::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
        return true;
 }
 
-void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
+void H264Encoder::add_audio(int64_t pts, vector<float> audio)
 {
        {
                unique_lock<mutex> lock(frame_queue_mutex);
@@ -1899,7 +1883,7 @@ void H264Encoder::add_audio(int64_t pts, std::vector<float> audio)
        frame_queue_nonempty.notify_all();
 }
 
-void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vector<RefCountedFrame> &input_frames)
+void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const vector<RefCountedFrame> &input_frames)
 {
        {
                unique_lock<mutex> lock(frame_queue_mutex);
@@ -1909,86 +1893,120 @@ void H264Encoder::end_frame(RefCountedGLsync fence, int64_t pts, const std::vect
        frame_queue_nonempty.notify_all();
 }
 
-void H264Encoder::copy_thread_func()
+void H264Encoder::encode_thread_func()
 {
        int64_t last_dts = -1;
-       for ( ;; ) {
+       int gop_start_display_frame_num = 0;
+       for (int encoding_frame_num = 0; ; ++encoding_frame_num) {
                PendingFrame frame;
                int pts_lag;
-               encoding2display_order(current_frame_encoding, intra_period, intra_idr_period, ip_period,
-                                      &current_frame_display, &current_frame_type, &pts_lag);
-               if (current_frame_type == FRAME_IDR) {
+               int frame_type, display_frame_num;
+               encoding2display_order(encoding_frame_num, intra_period, intra_idr_period, ip_period,
+                                      &display_frame_num, &frame_type, &pts_lag);
+               if (frame_type == FRAME_IDR) {
                        numShortTerm = 0;
                        current_frame_num = 0;
-                       current_IDR_display = current_frame_display;
+                       gop_start_display_frame_num = display_frame_num;
                }
 
                {
                        unique_lock<mutex> 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) return;
-                       frame = move(pending_video_frames[current_frame_display]);
-                       pending_video_frames.erase(current_frame_display);
-               }
-
-               // Wait for the GPU to be done with the frame.
-               glClientWaitSync(frame.fence.get(), 0, 0);
-
-               // Release back any input frames we needed to render this frame.
-               frame.input_frames.clear();
-
-               // Unmap the image.
-               GLSurface *surf = &gl_surfaces[current_frame_display % SURFACE_NUM];
-               eglDestroyImageKHR(eglGetCurrentDisplay(), surf->y_egl_image);
-               eglDestroyImageKHR(eglGetCurrentDisplay(), surf->cbcr_egl_image);
-               VAStatus va_status = vaReleaseBufferHandle(va_dpy, surf->surface_image.buf);
-               CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
-               va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
-               CHECK_VASTATUS(va_status, "vaDestroyImage");
-
-               VASurfaceID surface = surf->src_surface;
-
-               // Schedule the frame for encoding.
-               va_status = vaBeginPicture(va_dpy, context_id, surface);
-               CHECK_VASTATUS(va_status, "vaBeginPicture");
-
-               if (current_frame_type == FRAME_IDR) {
-                       render_sequence();
-                       render_picture();            
-                       if (h264_packedheader) {
-                               render_packedsequence();
-                               render_packedpicture();
+                       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[display_frame_num]);
+                               pending_video_frames.erase(display_frame_num);
                        }
-               } else {
-                       //render_sequence();
-                       render_picture();
                }
-               render_slice();
-               
-               va_status = vaEndPicture(va_dpy, context_id);
-               CHECK_VASTATUS(va_status, "vaEndPicture");
 
-               // Determine the pts and dts of this frame.
-               int64_t pts = frame.pts;
+               // Determine the dts of this frame.
                int64_t dts;
                if (pts_lag == -1) {
                        assert(last_dts != -1);
                        dts = last_dts + (TIMEBASE / MAX_FPS);
                } else {
-                       dts = pts - pts_lag;
+                       dts = frame.pts - pts_lag;
                }
                last_dts = dts;
 
-               // 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.pts = pts;
-               tmp.dts = dts;
-               storage_task_enqueue(move(tmp));
-               
-               update_ReferenceFrames();
-               ++current_frame_encoding;
+               encode_frame(frame, encoding_frame_num, display_frame_num, gop_start_display_frame_num, frame_type, frame.pts, dts);
        }
 }
+
+void H264Encoder::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 H264Encoder::encode_frame(H264Encoder::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);
+
+       // Release back any input frames we needed to render this frame.
+       frame.input_frames.clear();
+
+       // Unmap the image.
+       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);
+       CHECK_VASTATUS(va_status, "vaReleaseBufferHandle");
+       va_status = vaDestroyImage(va_dpy, surf->surface_image.image_id);
+       CHECK_VASTATUS(va_status, "vaDestroyImage");
+
+       VASurfaceID surface = surf->src_surface;
+
+       // Schedule the frame for encoding.
+       va_status = vaBeginPicture(va_dpy, context_id, surface);
+       CHECK_VASTATUS(va_status, "vaBeginPicture");
+
+       if (frame_type == FRAME_IDR) {
+               render_sequence();
+               render_picture(frame_type, display_frame_num, gop_start_display_frame_num);
+               if (h264_packedheader) {
+                       render_packedsequence();
+                       render_packedpicture();
+               }
+       } else {
+               //render_sequence();
+               render_picture(frame_type, display_frame_num, gop_start_display_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");
+
+       // 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 = display_frame_num;
+       tmp.frame_type = frame_type;
+       tmp.pts = pts;
+       tmp.dts = dts;
+       storage_task_enqueue(move(tmp));
+
+       update_ReferenceFrames(frame_type);
+}