diff --git a/chip/ps4_videodrv.pas b/chip/ps4_videodrv.pas
index e3cb8930..4f28ec1b 100644
--- a/chip/ps4_videodrv.pas
+++ b/chip/ps4_videodrv.pas
@@ -114,7 +114,15 @@ type
   Interrupt:Boolean;
  end;
 
- TvMicroEngineType=(metCmdBuffer,metFlip,metEop);
+ PvMeWaitMemInfo=^TvMeWaitMemInfo;
+ TvMeWaitMemInfo=record
+  adr:Pointer;
+  ref:DWORD;
+  mask:DWORD;
+  cFunc:Byte;
+ end;
+
+ TvMicroEngineType=(metCmdBuffer,metFlip,metEop,metWaitMem);
 
  PvMicroEngineNode=^TvMicroEngineNode;
  TvMicroEngineNode=record
@@ -126,6 +134,7 @@ type
    0:(CmdBuffer:TvCmdBuffer);
    1:(FlipInfo:TvMeFlipInfo);
    2:(EopInfo:TvMeEopInfo);
+   3:(WaitMem:TvMeWaitMemInfo);
  end;
 
  TvMicroEngine=object
@@ -136,6 +145,7 @@ type
   Procedure PushCmd(var Cmd:TvCmdBuffer);
   Procedure PushFlip(var qcInfo:TqcFlipInfo;FlipLData:DWORD;FlipLabel:PDWORD;Interrupt:Boolean);
   Procedure PushEop(adr:Pointer;data:QWORD;dataSel:Byte;Interrupt:Boolean);
+  Procedure PushWaitMem(adr:Pointer;ref,mask:DWORD;cFunc:Byte);
  end;
 
 var
@@ -145,6 +155,8 @@ var
  GFXRing:TvCmdRing;
  GFXMicroEngine:TvMicroEngine;
 
+ FIdleEvent:PRTLEvent=nil;
+
  GPU_REGS:TGPU_REGS;
 
  FCmdPool:TvCmdPool;
@@ -250,6 +262,22 @@ begin
  Queue.Push(node);
 end;
 
+Procedure TvMicroEngine.PushWaitMem(adr:Pointer;ref,mask:DWORD;cFunc:Byte);
+var
+ node:PvMicroEngineNode;
+begin
+ node:=AllocMem(SizeOf(TvMicroEngineNode));
+ if (node=nil) then Exit;
+
+ node^.mode         :=metWaitMem;
+ node^.WaitMem.adr  :=adr;
+ node^.WaitMem.ref  :=ref;
+ node^.WaitMem.mask :=mask;
+ node^.WaitMem.cFunc:=cFunc;
+
+ Queue.Push(node);
+end;
+
 procedure gfx_cp_parser(node:PvSubmitInfo);          forward;
 function  gfx_submit(CmdBuffer:TvCmdBuffer):Boolean; forward;
 function  gfx_test(CmdBuffer:TvCmdBuffer):Boolean;   forward;
@@ -296,6 +324,25 @@ begin
  end;
 end;
 
+Function me_test_mem(node:PvMeWaitMemInfo):Boolean;
+var
+ val,ref:DWORD;
+begin
+ val:=PDWORD(node^.adr)^ and node^.mask;
+ ref:=node^.ref;
+ Case node^.cFunc of
+  WAIT_REG_MEM_FUNC_ALWAYS       :Result:=True;
+  WAIT_REG_MEM_FUNC_LESS         :Result:=(val<ref);
+  WAIT_REG_MEM_FUNC_LESS_EQUAL   :Result:=(val<=ref);
+  WAIT_REG_MEM_FUNC_EQUAL        :Result:=(val=ref);
+  WAIT_REG_MEM_FUNC_NOT_EQUAL    :Result:=(val<>ref);
+  WAIT_REG_MEM_FUNC_GREATER_EQUAL:Result:=(val>ref);
+  WAIT_REG_MEM_FUNC_GREATER      :Result:=(val>=ref);
+  else
+   Assert(false);
+ end;
+end;
+
 Function me_node_test(node:PvMicroEngineNode):Boolean;
 begin
  Result:=True;
@@ -306,7 +353,14 @@ begin
    begin
     Result:=gfx_test(node^.CmdBuffer);
    end;
-  else;
+  metFlip:Result:=True;
+  metEop :Result:=True;
+  metWaitMem:
+   begin
+    Result:=me_test_mem(@node^.WaitMem);
+   end;
+  else
+   Assert(false);
  end;
 
 end;
@@ -329,6 +383,9 @@ begin
    begin
     Result:=me_eop(@node^.EopInfo);
    end;
+  metWaitMem:;
+  else
+   Assert(false);
  end;
 end;
 
@@ -351,7 +408,7 @@ begin
   begin
    if not me_node_test(GFXMicroEngine.Current) then
    begin
-    time:=-1000000;
+    time:=-100000;
     NtDelayExecution(True,@time);
     Continue;
    end;
@@ -366,6 +423,7 @@ begin
 
   if not work_do then
   begin
+   RTLEventSetEvent(FIdleEvent);
    time:=Int64(NT_INFINITE);
    NtDelayExecution(True,@time);
   end;
@@ -383,6 +441,9 @@ begin
   GFXRing.Init;
   GFXMicroEngine.Init;
 
+  FIdleEvent:=RTLEventCreate;
+  RTLEventSetEvent(FIdleEvent);
+
   t:=BeginThread(@GFX_thread);
 
   _gfx_handle:=t;
@@ -521,14 +582,19 @@ begin
   node^.Flip:=Flip^;
  end;
 
+ RTLEventResetEvent(FIdleEvent);
  GFXRing.Queue.Push(node);
  NtQueueApcThread(_gfx_handle,@_apc_null,0,nil,0);
 end;
 
 procedure vSubmitDone;
 begin
+ if (FIdleEvent<>nil) then
+ begin
+  RTLEventWaitFor(FIdleEvent);
+ end;
  //Sleep(100);
- Device.WaitIdle;
+ //Device.WaitIdle;
 end;
 
 
@@ -819,20 +885,22 @@ begin
 end;
 
 procedure onWaitRegMem(pm4Hdr:PM4_TYPE_3_HEADER;Body:PPM4CMDWAITREGMEM);
+var
+ adr:Pointer;
 begin
   {$ifdef ww}
  Case Body^.engine of
-  0: //ME
+  WAIT_REG_MEM_ENGINE_ME:
     Case Body^.memSpace of
-     0:Writeln(' waitOnRegister');
-     1:Writeln(' waitOnAddress');
+     WAIT_REG_MEM_SPACE_REGISTER:Writeln(' waitOnRegister');
+     WAIT_REG_MEM_SPACE_MEMORY  :Writeln(' waitOnAddress');
      else
       Assert(false);
     end;
-  1: //PFP
+  WAIT_REG_MEM_ENGINE_PFP:
     Case Body^.memSpace of
-     0:Writeln(' waitOnRegisterAndStall');
-     1:Writeln(' waitOnAddressAndStall');
+     WAIT_REG_MEM_SPACE_REGISTER:Writeln(' waitOnRegisterAndStall');
+     WAIT_REG_MEM_SPACE_MEMORY  :Writeln(' waitOnAddressAndStall');
      else
       Assert(false);
     end;
@@ -840,6 +908,26 @@ begin
    Assert(false);
  end;
  {$endif}
+
+ Case Body^.engine of
+  WAIT_REG_MEM_ENGINE_ME:
+    Case Body^.memSpace of
+     WAIT_REG_MEM_SPACE_MEMORY:
+       begin
+
+        QWORD(adr):=QWORD(Body^.pollAddressLo) or (QWORD(Body^.pollAddressHi) shl $20);
+
+        GFXMicroEngine.PushCmd(GFXRing.CmdBuffer);
+        GFXMicroEngine.PushWaitMem(adr,Body^.reference,Body^.mask,Body^.compareFunc);
+
+       end;
+     else
+      Assert(false);
+    end;
+  else
+   Assert(false);
+ end;
+
 end;
 
 procedure onPm40(pm4Hdr:PM4_TYPE_0_HEADER;Body:PDWORD);
diff --git a/ps4_libscegnmdriver.pas b/ps4_libscegnmdriver.pas
index 7f2f112d..33ee1bff 100644
--- a/ps4_libscegnmdriver.pas
+++ b/ps4_libscegnmdriver.pas
@@ -519,16 +519,15 @@ end;
 // called in waitUntilSafeForRendering
 function ps4_sceGnmInsertWaitFlipDone(cmdBuffer:PDWORD;numDwords:DWORD;videoOutHandle,displayBufferIndex:Integer):Integer; SysV_ABI_CDecl;
 var
- addr:Pointer;
+ addr:PInt64;
 begin
  Result:=-1;
  if (numDwords<>7) then Exit;
- //sceVideoOutGetBufferLabelAddress(videoOutHandle,&base);
- //addr = base + (ulong)(uint)displayBufferIndex * 8;
- _sig_lock;
- addr:=_VideoOutGetBufferAdr(videoOutHandle,displayBufferIndex);
- _sig_unlock;
- if (addr=nil) then Exit;
+
+ Result:=ps4_sceVideoOutGetBufferLabelAddress(videoOutHandle,@addr);
+ if (Result<>0) then Exit(-1);
+
+ addr:=@addr[displayBufferIndex];
 
  cmdBuffer[0]:=$c0053c00; //IT_WAIT_REG_MEM
  cmdBuffer[1]:=$13;
diff --git a/ps4_libscevideoout.pas b/ps4_libscevideoout.pas
index f14558a2..78cd5d39 100644
--- a/ps4_libscevideoout.pas
+++ b/ps4_libscevideoout.pas
@@ -214,7 +214,10 @@ const
  SCE_VIDEO_OUT_ERROR_UNKNOWN                          =-2144796418; // 0x802900FE */
  SCE_VIDEO_OUT_ERROR_ENOMEM                           =-2144792564; // 0x8029100C */
 
-function _VideoOutGetBufferAdr(hVideo:Integer;bufferIndex:Integer):Pointer;
+type
+ PPInt64=^PInt64;
+
+function ps4_sceVideoOutGetBufferLabelAddress(hVideo:Integer;ptr:PPInt64):Integer; SysV_ABI_CDecl;
 
 Procedure App_Run;
 
@@ -234,6 +237,7 @@ implementation
 uses
  vFlip,
  sys_signal,
+ sys_kernel,
  sys_time,
  ps4_time,
  spinlock,
@@ -377,6 +381,8 @@ type
    FcurrentBuffer:Longint;
   end;
 
+  FLabels:array[0..15] of Int64;
+
   //(MAIN port: 1 to 16, AUX port: 1 to 8)
   FBuffers:record
    lock:Pointer;
@@ -449,7 +455,7 @@ begin
    Acqure;
    Exit(n);
   end;
-  System.ThreadSwitch;
+  SwYieldExecution;
  until false;
 end;
 
@@ -939,20 +945,24 @@ begin
  H.Release;
 end;
 
-function _VideoOutGetBufferAdr(hVideo:Integer;bufferIndex:Integer):Pointer;
+function ps4_sceVideoOutGetBufferLabelAddress(hVideo:Integer;ptr:PPInt64):Integer; SysV_ABI_CDecl;
 var
  H:TVideoOut;
 begin
- Result:=nil;
+ if (ptr=nil) then Exit(SCE_VIDEO_OUT_ERROR_INVALID_ADDRESS);
+
+ _sig_lock;
  H:=TVideoOut(FVideoOutMap.Acqure(hVideo));
- if (H=nil) then Exit;
- if (bufferIndex>=0) and (bufferIndex<SCE_VIDEO_OUT_BUFFER_NUM_MAX) then
- begin
-  spin_lock(H.FBuffers.lock);
-  Result:=H.FBuffers.addr[bufferIndex];
-  spin_unlock(H.FBuffers.lock);
- end;
+ _sig_unlock;
+
+ if (H=nil) then Exit(SCE_VIDEO_OUT_ERROR_INVALID_HANDLE);
+ ptr^:=@H.FLabels;
+
+ _sig_lock;
  H.Release;
+ _sig_unlock;
+
+ Result:=0;
 end;
 
 function _on_after(node:PKEventNode;data:Pointer):Boolean;
@@ -1038,6 +1048,7 @@ begin
    //ps4_usleep(150);
   end else
   begin
+
    spin_lock(FBuffers.lock);
    addr:=FBuffers.addr[bufferIndex];
    attr:=FBuffers.attr[bufferIndex];
@@ -1076,18 +1087,17 @@ begin
   //post_event_flip(flipArg);
  end;
 
+ node^.wait:=1;
+ free_node(node);
+
  //FlipRate:=20;
  if (flipMode=SCE_VIDEO_OUT_FLIP_MODE_VSYNC) then
  if (FlipRate<>0) then
  begin
-  time:=(1000000 div (FlipRate+1)); //+5 selected empirically
-  //time:=time-1300;
-  //time:=time-1300;
+  time:=(1000000 div (FlipRate{+1}));
 
   elap:=SwTimePassedUnits(FlipStatus.Ftsc_flips);
-  elap:=(elap+9) div 10;
-
-  //elap:=elap+(elap div 100)*14;
+  elap:=((elap+9) div 10)+120;
 
   if (elap<time) then
   begin
@@ -1106,13 +1116,13 @@ begin
   //t2:=(t2+9) div 10;
   //Writeln('elap=',elap,' time=',time,' usleep=',t2);
 
-  if (FGpuFlip<>nil) then
+  if (FGpuFlip<>nil) and (bufferIndex<>SCE_VIDEO_OUT_BUFFER_INDEX_BLANK) then
   begin
-   FGpuFlip.IsComplite(FlipStatus.FcurrentBuffer);
-   //While (not FGpuFlip.IsComplite(FcurrentBuffer)) do
-   //begin
-   // ps4_usleep(150);
-   //end;
+   //FGpuFlip.IsComplite(bufferIndex);
+   While (not FGpuFlip.IsComplite(bufferIndex)) do
+   begin
+    ps4_usleep(150);
+   end;
   end;
 
  end;
@@ -1122,6 +1132,8 @@ begin
   post_event_vblank(flipArg);
  end else
  begin
+  System.InterlockedDecrement64(FLabels[bufferIndex]);
+
   post_event_flip(flipArg);
   post_event_vblank(flipArg);
  end;
@@ -1155,8 +1167,8 @@ begin
   end;
  end;
 
- node^.wait:=1;
- free_node(node);
+ //node^.wait:=1;
+ //free_node(node);
 
 end;
 
@@ -1179,9 +1191,7 @@ begin
  _sig_unlock;
  if (H=nil) then Exit(SCE_VIDEO_OUT_ERROR_INVALID_HANDLE);
 
- _sig_lock;
  node:=H.alloc_node;
- _sig_unlock;
 
  if (node=nil) then
  begin
@@ -1189,6 +1199,11 @@ begin
   Exit(SCE_VIDEO_OUT_ERROR_FLIP_QUEUE_FULL);
  end;
 
+ if (bufferIndex<>SCE_VIDEO_OUT_BUFFER_INDEX_BLANK) then
+ begin
+  System.InterlockedIncrement64(H.FLabels[bufferIndex]);
+ end;
+
  node^.Parent:=@H.sceVideoOutSubmitFlip;
  node^.u.bufferIndex:=bufferIndex;
  node^.u.flipMode   :=flipMode;
@@ -1242,6 +1257,11 @@ begin
   Exit(SCE_VIDEO_OUT_ERROR_FLIP_QUEUE_FULL);
  end;
 
+ if (Flip^.bufferIndex<>SCE_VIDEO_OUT_BUFFER_INDEX_BLANK) then
+ begin
+  System.InterlockedIncrement64(H.FLabels[Flip^.bufferIndex]);
+ end;
+
  node^.Parent:=@H.sceVideoOutSubmitFlip;
  node^.u.bufferIndex:=Flip^.bufferIndex;
  node^.u.flipMode   :=Flip^.flipMode;
diff --git a/vulkan/vFlip.pas b/vulkan/vFlip.pas
index 8bee4bfa..476add32 100644
--- a/vulkan/vFlip.pas
+++ b/vulkan/vFlip.pas
@@ -618,23 +618,35 @@ begin
 
  buf:=@FBuffers[FcurrentBuffer];
 
- if (Ffilp_shader=nil) then Exit;
-
- if not buf^.IsPrepare then Exit;
-
- if (FSwapChain=nil) then recreateSwapChain;
- if (FSwapChain.FHandle=VK_NULL_HANDLE) then
+ if (Ffilp_shader=nil) then
  begin
-  recreateSwapChain;
-  if (FSwapChain.FHandle=VK_NULL_HANDLE) then Exit;
+  buf^.cmdbuf.ret:=1;
+  Exit;
  end;
 
- if not FSetLayout.Compile then Exit;
- if not FLayout.Compile    then Exit;
- if not FPipelineFlip.Compile  then Exit;
- if not FSetsPool.Compile  then Exit;
+ if (not buf^.IsPrepare)        or
+    (not FSetLayout.Compile)    or
+    (not FLayout.Compile)       or
+    (not FPipelineFlip.Compile) or
+    (not FSetsPool.Compile) then
+ begin
+  buf^.cmdbuf.ret:=2;
+  Exit;
+ end;
 
  repeat
+
+  if (FSwapChain=nil) then recreateSwapChain;
+  if (FSwapChain.FHandle=VK_NULL_HANDLE) then
+  begin
+   recreateSwapChain;
+   if (FSwapChain.FHandle=VK_NULL_HANDLE) then
+   begin
+    buf^.cmdbuf.ret:=3;
+    Exit;
+   end;
+  end;
+
   FixCurrentFrame;
   imageAvailableSemaphore:=FimageAvailableSemaphore[FcurrentFrame];
   renderFinishedSemaphore:=FrenderFinishedSemaphore[FcurrentFrame];
@@ -655,6 +667,7 @@ begin
      Exit;
     end;
   end;
+
  until false;
  SwapImage:=FSwapChain.FImages[imageIndex];
 
diff --git a/vulkan/vImageManager.pas b/vulkan/vImageManager.pas
index 18852729..54a31bf7 100644
--- a/vulkan/vImageManager.pas
+++ b/vulkan/vImageManager.pas
@@ -436,6 +436,8 @@ begin
  if (cmd=nil) then Exit;
  if (not cmd.BeginCmdBuffer) then Exit;
 
+ rwlock_wrlock(lock);
+
  if Barrier.Push(cmd.cmdbuf,
                  FHandle,
                  GetSubresRange,
@@ -445,6 +447,8 @@ begin
  begin
   Inc(cmd.cmd_count);
  end;
+
+ rwlock_unlock(lock);
 end;
 
 procedure TvHostImage2.PushBarrier(cmd:TvCustomCmdBuffer;