Fix an issue in the H264 encoder where a surface could be rendered to while it was...

[nageru] / h264encode.cpp

diff --git a/h264encode.cpp b/h264encode.cpp
index 784531663c3e6f7191f1847130b3291d34ee7203..6a402f45765e4932f1c0dcaf57208554d19f6969 100644 (file)
--- a/h264encode.cpp
+++ b/h264encode.cpp
@@ -84,6 +84,8 @@ class QSurface;
    
 #define BITSTREAM_ALLOCATE_STEPPING     4096
 #define SURFACE_NUM 16 /* 16 surfaces for source YUV */
+#define MAX_NUM_REF1 16 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
+#define MAX_NUM_REF2 32 // Seemingly a hardware-fixed value, not related to SURFACE_NUM
 
 static constexpr unsigned int MaxFrameNum = (2<<16);
 static constexpr unsigned int MaxPicOrderCntLsb = (2<<8);
@@ -217,7 +219,7 @@ private:
        VAEncPictureParameterBufferH264 pic_param;
        VAEncSliceParameterBufferH264 slice_param;
        VAPictureH264 CurrentCurrPic;
-       VAPictureH264 ReferenceFrames[16], RefPicList0_P[32], RefPicList0_B[32], RefPicList1_B[32];
+       VAPictureH264 ReferenceFrames[MAX_NUM_REF1], RefPicList0_P[MAX_NUM_REF2], RefPicList0_B[MAX_NUM_REF2], RefPicList1_B[MAX_NUM_REF2];
 
        // Static quality settings.
        static constexpr unsigned int frame_bitrate = 15000000 / 60;  // Doesn't really matter; only initial_qp does.
@@ -924,7 +926,7 @@ int H264EncoderImpl::init_va(const string &va_display)
     
     if (support_encode == 0) {
         printf("Can't find VAEntrypointEncSlice for H264 profiles. If you are using a non-Intel GPU\n");
-        printf("but have one in your system, try launching Nageru with --va-display /dev/dri/card0\n");
+        printf("but have one in your system, try launching Nageru with --va-display /dev/dri/renderD128\n");
         printf("to use VA-API against DRM instead of X11.\n");
         exit(1);
     } else {
@@ -1298,7 +1300,7 @@ int H264EncoderImpl::render_picture(int frame_type, int display_frame_num, int g
     CurrentCurrPic = pic_param.CurrPic;
 
     memcpy(pic_param.ReferenceFrames, ReferenceFrames, numShortTerm*sizeof(VAPictureH264));
-    for (i = numShortTerm; i < SURFACE_NUM; i++) {
+    for (i = numShortTerm; i < MAX_NUM_REF1; i++) {
         pic_param.ReferenceFrames[i].picture_id = VA_INVALID_SURFACE;
         pic_param.ReferenceFrames[i].flags = VA_PICTURE_H264_INVALID;
     }
@@ -1448,7 +1450,7 @@ int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num,
         int refpiclist0_max = h264_maxref & 0xffff;
         memcpy(slice_param.RefPicList0, RefPicList0_P, refpiclist0_max*sizeof(VAPictureH264));
 
-        for (i = refpiclist0_max; i < 32; i++) {
+        for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
         }
@@ -1457,13 +1459,13 @@ int H264EncoderImpl::render_slice(int encoding_frame_num, int display_frame_num,
         int refpiclist1_max = (h264_maxref >> 16) & 0xffff;
 
         memcpy(slice_param.RefPicList0, RefPicList0_B, refpiclist0_max*sizeof(VAPictureH264));
-        for (i = refpiclist0_max; i < 32; i++) {
+        for (i = refpiclist0_max; i < MAX_NUM_REF2; i++) {
             slice_param.RefPicList0[i].picture_id = VA_INVALID_SURFACE;
             slice_param.RefPicList0[i].flags = VA_PICTURE_H264_INVALID;
         }
 
         memcpy(slice_param.RefPicList1, RefPicList1_B, refpiclist1_max*sizeof(VAPictureH264));
-        for (i = refpiclist1_max; i < 32; i++) {
+        for (i = refpiclist1_max; i < MAX_NUM_REF2; i++) {
             slice_param.RefPicList1[i].picture_id = VA_INVALID_SURFACE;
             slice_param.RefPicList1[i].flags = VA_PICTURE_H264_INVALID;
         }
@@ -1538,7 +1540,7 @@ void H264EncoderImpl::save_codeddata(storage_task task)
              pending_audio_frames.erase(it); 
         }
 
-        AVFrame *frame = avcodec_alloc_frame();
+        AVFrame *frame = av_frame_alloc();
         frame->nb_samples = audio.size() / 2;
         frame->format = AV_SAMPLE_FMT_S32;
         frame->channel_layout = AV_CH_LAYOUT_STEREO;
@@ -1570,7 +1572,7 @@ void H264EncoderImpl::save_codeddata(storage_task task)
             httpd->add_packet(pkt, audio_pts + global_delay, audio_pts + global_delay);
         }
         // TODO: Delayed frames.
-        avcodec_free_frame(&frame);
+        av_frame_unref(frame);
         av_free_packet(&pkt);
         if (audio_pts == task.pts) break;
     }
@@ -1601,7 +1603,6 @@ void H264EncoderImpl::storage_task_enqueue(storage_task task)
 {
        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();
 }
 
@@ -1725,7 +1726,12 @@ bool H264EncoderImpl::begin_frame(GLuint *y_tex, GLuint *cbcr_tex)
        {
                // Wait until this frame slot is done encoding.
                unique_lock<mutex> lock(storage_task_queue_mutex);
+               if (srcsurface_status[current_storage_frame % SURFACE_NUM] != SRC_SURFACE_FREE) {
+                       fprintf(stderr, "Warning: Slot %d (for frame %d) is still encoding, rendering has to wait for H.264 encoder\n",
+                               current_storage_frame % SURFACE_NUM, current_storage_frame);
+               }
                storage_task_queue_changed.wait(lock, [this]{ return storage_thread_should_quit || (srcsurface_status[current_storage_frame % SURFACE_NUM] == SRC_SURFACE_FREE); });
+               srcsurface_status[current_storage_frame % SURFACE_NUM] = SRC_SURFACE_IN_ENCODING;
                if (storage_thread_should_quit) return false;
        }