diff --git a/samples/videoDecodePerf/videodecodeperf.cpp b/samples/videoDecodePerf/videodecodeperf.cpp
index 81457f3bdf..9927671057 100644
--- a/samples/videoDecodePerf/videodecodeperf.cpp
+++ b/samples/videoDecodePerf/videodecodeperf.cpp
@@ -37,7 +37,7 @@ THE SOFTWARE.
 #include "roc_video_dec.h"
 #include "common.h"
 
-void DecProc(RocVideoDecoder *p_dec, VideoDemuxer *demuxer, int *pn_frame, int *pn_pic_dec, double *pn_fps, double *pn_fps_dec, int max_num_frames) {
+void DecProc(RocVideoDecoder *p_dec, VideoDemuxer *demuxer, int *pn_frame, int *pn_pic_dec, double *pn_fps, double *pn_fps_dec, int max_num_frames, OutputSurfaceMemoryType mem_type) {
     int n_video_bytes = 0, n_frame_returned = 0, n_frame = 0;
     int n_pic_decoded = 0, decoded_pics = 0;
     uint8_t *p_video = nullptr;
@@ -54,6 +54,9 @@ void DecProc(RocVideoDecoder *p_dec, VideoDemuxer *demuxer, int *pn_frame, int *
             break;
         }
     } while (n_video_bytes);
+    if (mem_type == OUT_SURFACE_MEM_NOT_MAPPED) {
+        p_dec->WaitForDecodeCompletion();
+    }
 
     auto end_time = std::chrono::high_resolution_clock::now();
     auto time_per_decode = std::chrono::duration<double, std::milli>(end_time - start_time).count();
@@ -76,6 +79,7 @@ void ShowHelpAndExit(const char *option = NULL) {
     << "-t Number of threads (>= 1) - optional; default: 1" << std::endl
     << "-d Device ID (>= 0) - optional; default: 0" << std::endl
     << "-z Force zero latency (decoded frames will be flushed out for display immediately) - optional" << std::endl
+    << "-disp_delay -specify the number of frames to be delayed for display; optional" << std::endl
     << "-m Memory type (integer values between 0 to 3: specifies where to store the decoded output:" << std::endl
     << "                                               0 = decoded output will be in internal interopped memory," << std::endl
     << "                                               1 = decoded output will be copied to a separate device memory," << std::endl
@@ -249,7 +253,7 @@ int main(int argc, char **argv) {
         }
 
         for (int i = 0; i < n_thread; i++) {
-            v_thread.push_back(std::thread(DecProc, v_viddec[i].get(), v_demuxer[i].get(), &v_frame[i], &v_frame_dec[i], &v_fps[i], &v_fps_dec[i], max_num_frames));
+            v_thread.push_back(std::thread(DecProc, v_viddec[i].get(), v_demuxer[i].get(), &v_frame[i], &v_frame_dec[i], &v_fps[i], &v_fps_dec[i], max_num_frames, mem_type));
         }
 
         for (int i = 0; i < n_thread; i++) {
diff --git a/utils/rocvideodecode/roc_video_dec.cpp b/utils/rocvideodecode/roc_video_dec.cpp
index 54b1aecba1..a7648f1430 100644
--- a/utils/rocvideodecode/roc_video_dec.cpp
+++ b/utils/rocvideodecode/roc_video_dec.cpp
@@ -643,6 +643,7 @@ int RocVideoDecoder::HandlePictureDecode(RocdecPicParams *pPicParams) {
     }
     pic_num_in_dec_order_[pPicParams->curr_pic_idx] = decode_poc_++;
     ROCDEC_API_CALL(rocDecDecodeFrame(roc_decoder_, pPicParams));
+    last_decode_surf_idx_ = pPicParams->curr_pic_idx;
     decoded_pic_cnt_++;
     if (b_force_zero_latency_ && ((!pPicParams->field_pic_flag) || (pPicParams->second_field))) {
         RocdecParserDispInfo disp_info;
@@ -1189,3 +1190,11 @@ bool RocVideoDecoder::CodecSupported(int device_id, rocDecVideoCodec codec_id, u
     }
     return true;
 }
+
+void RocVideoDecoder::WaitForDecodeCompletion() {
+    RocdecDecodeStatus dec_status;
+    memset(&dec_status, 0, sizeof(dec_status));
+    do {
+        rocDecStatus result = rocDecGetDecodeStatus(roc_decoder_, last_decode_surf_idx_, &dec_status);
+    } while (dec_status.decode_status == rocDecodeStatus_InProgress);
+}
\ No newline at end of file
diff --git a/utils/rocvideodecode/roc_video_dec.h b/utils/rocvideodecode/roc_video_dec.h
index 797c245583..aec4125820 100644
--- a/utils/rocvideodecode/roc_video_dec.h
+++ b/utils/rocvideodecode/roc_video_dec.h
@@ -336,7 +336,7 @@ class RocVideoDecoder {
         /**
          * @brief Helper funtion to close a existing file and dump to new file in case of multiple files using same decoder
         */
-       void ResetSaveFrameToFile();
+        void ResetSaveFrameToFile();
 
         /**
          * @brief Helper function to start MD5 calculation
@@ -366,6 +366,10 @@ class RocVideoDecoder {
          */
         int32_t GetNumOfFlushedFrames() { return num_frames_flushed_during_reconfig_;}
 
+        /*! \brief Function to wait for the decode completion of the last submitted picture
+         */
+        void WaitForDecodeCompletion();
+
         // Session overhead refers to decoder initialization and deinitialization time
         void AddDecoderSessionOverHead(std::thread::id session_id, double duration) { session_overhead_[session_id] += duration; }
         double GetDecoderSessionOverHead(std::thread::id session_id) {
@@ -480,6 +484,7 @@ class RocVideoDecoder {
         int decoded_pic_cnt_ = 0;
         int decode_poc_ = 0, pic_num_in_dec_order_[MAX_FRAME_NUM];
         int num_alloced_frames_ = 0;
+        int last_decode_surf_idx_ = 0;
         std::ostringstream input_video_info_str_;
         int bitdepth_minus_8_ = 0;
         uint32_t byte_per_pixel_ = 1;