diff --git a/src/util/d3d11_device.cpp b/src/util/d3d11_device.cpp
index 9f50261d7..4c9db05c5 100644
--- a/src/util/d3d11_device.cpp
+++ b/src/util/d3d11_device.cpp
@@ -509,12 +509,14 @@ std::string D3D11Device::GetDriverInfo() const
 void D3D11Device::FlushCommands()
 {
   m_context->Flush();
+  EndTimestampQuery();
   TrimTexturePool();
 }
 
 void D3D11Device::WaitForGPUIdle()
 {
   m_context->Flush();
+  EndTimestampQuery();
   TrimTexturePool();
 }
 
@@ -671,7 +673,10 @@ GPUDevice::PresentResult D3D11Device::BeginPresent(GPUSwapChain* swap_chain, u32
   // in this configuration. It does reduce accuracy a little, but better than seeing 100% all of
   // the time, when it's more like a couple of percent.
   if (SC == m_main_swap_chain.get() && SC->GetVSyncMode() == GPUVSyncMode::FIFO && m_gpu_timing_enabled)
+  {
     PopTimestampQuery();
+    EndTimestampQuery();
+  }
 
   m_context->ClearRenderTargetView(SC->GetRTV(), GSVector4::rgba32(clear_color).F32);
   m_context->OMSetRenderTargets(1, SC->GetRTVArray(), nullptr);
@@ -690,7 +695,10 @@ void D3D11Device::EndPresent(GPUSwapChain* swap_chain, bool explicit_present, u6
   DebugAssert(m_num_current_render_targets == 0 && !m_current_depth_target);
 
   if (SC == m_main_swap_chain.get() && SC->GetVSyncMode() != GPUVSyncMode::FIFO && m_gpu_timing_enabled)
+  {
     PopTimestampQuery();
+    EndTimestampQuery();
+  }
 
   const UINT sync_interval = static_cast<UINT>(SC->GetVSyncMode() == GPUVSyncMode::FIFO);
   const UINT flags =
@@ -698,7 +706,7 @@ void D3D11Device::EndPresent(GPUSwapChain* swap_chain, bool explicit_present, u6
   SC->GetSwapChain()->Present(sync_interval, flags);
 
   if (m_gpu_timing_enabled)
-    KickTimestampQuery();
+    StartTimestampQuery();
 
   TrimTexturePool();
 }
@@ -724,7 +732,7 @@ bool D3D11Device::CreateTimestampQueries()
     }
   }
 
-  KickTimestampQuery();
+  StartTimestampQuery();
   return true;
 }
 
@@ -783,7 +791,10 @@ void D3D11Device::PopTimestampQuery()
       }
     }
   }
+}
 
+void D3D11Device::EndTimestampQuery()
+{
   if (m_timestamp_query_started)
   {
     m_context->End(m_timestamp_queries[m_write_timestamp_query][2].Get());
@@ -794,7 +805,7 @@ void D3D11Device::PopTimestampQuery()
   }
 }
 
-void D3D11Device::KickTimestampQuery()
+void D3D11Device::StartTimestampQuery()
 {
   if (m_timestamp_query_started || !m_timestamp_queries[0][0] || m_waiting_timestamp_queries == NUM_TIMESTAMP_QUERIES)
     return;
@@ -815,7 +826,7 @@ bool D3D11Device::SetGPUTimingEnabled(bool enabled)
     if (!CreateTimestampQueries())
       return false;
 
-    KickTimestampQuery();
+    StartTimestampQuery();
     return true;
   }
   else
diff --git a/src/util/d3d11_device.h b/src/util/d3d11_device.h
index c279fdf53..bc6edd829 100644
--- a/src/util/d3d11_device.h
+++ b/src/util/d3d11_device.h
@@ -161,7 +161,8 @@ private:
   bool CreateTimestampQueries();
   void DestroyTimestampQueries();
   void PopTimestampQuery();
-  void KickTimestampQuery();
+  void EndTimestampQuery();
+  void StartTimestampQuery();
 
   ComPtr<ID3D11Device1> m_device;
   ComPtr<ID3D11DeviceContext1> m_context;
diff --git a/src/util/opengl_device.cpp b/src/util/opengl_device.cpp
index 4b181ffbc..80bbe24f5 100644
--- a/src/util/opengl_device.cpp
+++ b/src/util/opengl_device.cpp
@@ -656,13 +656,15 @@ std::string OpenGLDevice::GetDriverInfo() const
 void OpenGLDevice::FlushCommands()
 {
   glFlush();
+  EndTimestampQuery();
   TrimTexturePool();
 }
 
 void OpenGLDevice::WaitForGPUIdle()
 {
-  // Could be glFinish(), but I'm afraid for mobile drivers...
-  glFlush();
+  glFinish();
+  EndTimestampQuery();
+  TrimTexturePool();
 }
 
 void OpenGLDevice::RenderBlankFrame()
@@ -818,12 +820,15 @@ void OpenGLDevice::EndPresent(GPUSwapChain* swap_chain, bool explicit_present, u
   DebugAssert(m_current_fbo == 0);
 
   if (swap_chain == m_main_swap_chain.get() && m_gpu_timing_enabled)
+  {
     PopTimestampQuery();
+    EndTimestampQuery();
+  }
 
   m_gl_context->SwapBuffers();
 
   if (swap_chain == m_main_swap_chain.get() && m_gpu_timing_enabled)
-    KickTimestampQuery();
+    StartTimestampQuery();
 
   TrimTexturePool();
 }
@@ -839,7 +844,7 @@ void OpenGLDevice::CreateTimestampQueries()
   const auto GenQueries = gles ? glGenQueriesEXT : glGenQueries;
 
   GenQueries(static_cast<u32>(m_timestamp_queries.size()), m_timestamp_queries.data());
-  KickTimestampQuery();
+  StartTimestampQuery();
 }
 
 void OpenGLDevice::DestroyTimestampQueries()
@@ -866,8 +871,7 @@ void OpenGLDevice::DestroyTimestampQueries()
 
 void OpenGLDevice::PopTimestampQuery()
 {
-  const bool gles = m_gl_context->IsGLES();
-
+  const bool gles = IsGLES();
   if (gles)
   {
     GLint disjoint = 0;
@@ -885,11 +889,10 @@ void OpenGLDevice::PopTimestampQuery()
     }
   }
 
+  const auto GetQueryObjectiv = gles ? glGetQueryObjectivEXT : glGetQueryObjectiv;
+  const auto GetQueryObjectui64v = gles ? glGetQueryObjectui64vEXT : glGetQueryObjectui64v;
   while (m_waiting_timestamp_queries > 0)
   {
-    const auto GetQueryObjectiv = gles ? glGetQueryObjectivEXT : glGetQueryObjectiv;
-    const auto GetQueryObjectui64v = gles ? glGetQueryObjectui64vEXT : glGetQueryObjectui64v;
-
     GLint available = 0;
     GetQueryObjectiv(m_timestamp_queries[m_read_timestamp_query], GL_QUERY_RESULT_AVAILABLE, &available);
     if (!available)
@@ -901,10 +904,25 @@ void OpenGLDevice::PopTimestampQuery()
     m_read_timestamp_query = (m_read_timestamp_query + 1) % NUM_TIMESTAMP_QUERIES;
     m_waiting_timestamp_queries--;
   }
+}
 
+void OpenGLDevice::StartTimestampQuery()
+{
+  if (m_timestamp_query_started || m_waiting_timestamp_queries == NUM_TIMESTAMP_QUERIES)
+    return;
+
+  const bool gles = IsGLES();
+  const auto BeginQuery = gles ? glBeginQueryEXT : glBeginQuery;
+
+  BeginQuery(GL_TIME_ELAPSED, m_timestamp_queries[m_write_timestamp_query]);
+  m_timestamp_query_started = true;
+}
+
+void OpenGLDevice::EndTimestampQuery()
+{
   if (m_timestamp_query_started)
   {
-    const auto EndQuery = gles ? glEndQueryEXT : glEndQuery;
+    const auto EndQuery = IsGLES() ? glEndQueryEXT : glEndQuery;
     EndQuery(GL_TIME_ELAPSED);
 
     m_write_timestamp_query = (m_write_timestamp_query + 1) % NUM_TIMESTAMP_QUERIES;
@@ -913,18 +931,6 @@ void OpenGLDevice::PopTimestampQuery()
   }
 }
 
-void OpenGLDevice::KickTimestampQuery()
-{
-  if (m_timestamp_query_started || m_waiting_timestamp_queries == NUM_TIMESTAMP_QUERIES)
-    return;
-
-  const bool gles = m_gl_context->IsGLES();
-  const auto BeginQuery = gles ? glBeginQueryEXT : glBeginQuery;
-
-  BeginQuery(GL_TIME_ELAPSED, m_timestamp_queries[m_write_timestamp_query]);
-  m_timestamp_query_started = true;
-}
-
 bool OpenGLDevice::SetGPUTimingEnabled(bool enabled)
 {
   if (m_gpu_timing_enabled == enabled)
diff --git a/src/util/opengl_device.h b/src/util/opengl_device.h
index 5846b34b0..68b5352c3 100644
--- a/src/util/opengl_device.h
+++ b/src/util/opengl_device.h
@@ -177,7 +177,8 @@ private:
   void CreateTimestampQueries();
   void DestroyTimestampQueries();
   void PopTimestampQuery();
-  void KickTimestampQuery();
+  void StartTimestampQuery();
+  void EndTimestampQuery();
 
   GLuint CreateProgramFromPipelineCache(const OpenGLPipeline::ProgramCacheItem& it,
                                         const GPUPipeline::GraphicsConfig& plconfig);