ilclient_set_fill_buffer_done_callback(decoder->client, fill_buffer_done, decoder);

static void fill_buffer_done(void* data, COMPONENT_T* comp) {
	H264_decoder *decoder = (H264_decoder *)data;
	if (decoder->outbuf) {
		//printf("buf size %d addr %p\n",decoder->outbuf->nFilledLen,decoder->outbuf->pBuffer);
	}
	/* Signal complete event. */
	pthread_mutex_lock(&fill_buffer_done_mutex);
	fill_buffer_done_val = 1;
	pthread_cond_signal(&fill_buffer_done_cond);
	pthread_mutex_unlock(&fill_buffer_done_mutex);
}
static int rpi_h264_end_frame(AVCodecContext *avctx) {
	struct timeval start,end,start2,end2;
	OMXContext *hwctx = avctx->hwaccel_context; // made in vd_ffmpeg.c
	H264Context *h = avctx->priv_data;
	Picture *pic = h->cur_pic_ptr;
	OMX_ERRORTYPE err;
	char name[100];
	struct rpi_picture_context *pic_ctx = pic->hwaccel_picture_private; // allocated by lib, based on priv_data_size

	gettimeofday(&start,NULL);

	if (decoder->outbuf != NULL) {
		gettimeofday(&start2,NULL);
		err = OMX_FillThisBuffer(decoder->image_resize->handle, decoder->outbuf);
		if (err != OMX_ErrorNone) {
			printf("error filling buffer: %x\n",err);
		} else {
			/* Wait for fill_buffer_done. */
			pthread_mutex_lock(&fill_buffer_done_mutex);
			while (fill_buffer_done_val == 0) {
				pthread_cond_wait(&fill_buffer_done_cond, &fill_buffer_done_mutex);
			}
			fill_buffer_done_val = 0;
			pthread_mutex_unlock(&fill_buffer_done_mutex);
		}
		gettimeofday(&end2,NULL);
	} else {
		//puts("no outbuf found");
	}
	if (hwctx != NULL) {
		// call the vo module in mplayer
		hwctx->render(avctx, &pic->f, pic_ctx->first);
	} else {
		if (decoder->outbuf != NULL) {
			frames++;
#if 0
			snprintf(name,100,"/media/videos/4tb/rpi/frame%d.yuv",frames);
			FILE *frameout = fopen(name,"w");
			fwrite(decoder->outbuf->pBuffer,decoder->outbuf->nFilledLen,1,frameout);
			fclose(frameout);
#endif
			// start over, with examples from Daemon404
			int p;
			for (p=0; p < 3; p++) {
				uint8_t *buf = pic->f.data[p];
				uint8_t *src;
				int stride = decoder->stride >> !!p; // 10 17:44:25  * Daemon404 runs in shame
				int height = decoder->sliceHeight >> !!p;
				if (p==0) src = decoder->outbuf->pBuffer;
				else if (p==1) src = decoder->outbuf->pBuffer + (decoder->stride*decoder->sliceHeight);
				else if (p==2) src = decoder->outbuf->pBuffer + (decoder->stride*decoder->sliceHeight) + (stride*height);
				if (stride != pic->f.linesize[p]) {
					printf("stride mismatch between gpu and ffmpeg, %d %d\n",stride,pic->f.linesize[p]);
					exit(-2);
				}
				//printf("buf:%p src:%p stride:%d plane:%d linesize:%d height:%d\n",buf,src,stride,p,pic->f.linesize[p],height);
				memcpy(buf,src,stride * height);
			}
			gettimeofday(&end,NULL);
			float total = end.tv_sec - start.tv_sec;
			total += (float)(end.tv_usec - start.tv_usec) / 1000000.0;
			float total2 = end2.tv_sec - start2.tv_sec;
			total2 += (float)(end2.tv_usec - start2.tv_usec) / 1000000.0;
			printf("%f %f %3d size:%d offset:%d\n",total,total2,frames,decoder->outbuf->nFilledLen,decoder->outbuf->nOffset);
		}
	}
	//if (frames > 200) exit(-2);
	return 0;
}