diff --git a/qrenderdoc/Code/CaptureContext.cpp b/qrenderdoc/Code/CaptureContext.cpp
index 0c0c9e4b8..db3604dbe 100644
--- a/qrenderdoc/Code/CaptureContext.cpp
+++ b/qrenderdoc/Code/CaptureContext.cpp
@@ -66,6 +66,8 @@
 
 CaptureContext::CaptureContext(PersistantConfig &cfg) : m_Config(cfg)
 {
+  RENDERDOC_PROFILEFUNCTION();
+
   m_CaptureLoaded = false;
   m_LoadInProgress = false;
 
@@ -699,6 +701,8 @@ void CaptureContext::CleanMenu(QAction *action)
 void CaptureContext::LoadCapture(const rdcstr &captureFile, const ReplayOptions &opts,
                                  const rdcstr &origFilename, bool temporary, bool local)
 {
+  RENDERDOC_PROFILEFUNCTION();
+
   CloseCapture();
 
   PointerTypeRegistry::Init();
@@ -1458,6 +1462,8 @@ void CaptureContext::ExportCapture(const CaptureFileFormat &fmt, const rdcstr &e
 void CaptureContext::SetEventID(const rdcarray<ICaptureViewer *> &exclude, uint32_t selectedEventID,
                                 uint32_t eventId, bool force)
 {
+  RENDERDOC_PROFILEFUNCTION();
+
   if(!IsCaptureLoaded())
     return;
 
diff --git a/qrenderdoc/Windows/PipelineState/PipelineStateViewer.cpp b/qrenderdoc/Windows/PipelineState/PipelineStateViewer.cpp
index d8f0af76d..34b0195c0 100644
--- a/qrenderdoc/Windows/PipelineState/PipelineStateViewer.cpp
+++ b/qrenderdoc/Windows/PipelineState/PipelineStateViewer.cpp
@@ -131,6 +131,8 @@ void PipelineStateViewer::OnCaptureClosed()
 
 void PipelineStateViewer::OnEventChanged(uint32_t eventId)
 {
+  RENDERDOC_PROFILEFUNCTION();
+
   if(m_Ctx.APIProps().pipelineType == GraphicsAPI::D3D11)
     setToD3D11();
   else if(m_Ctx.APIProps().pipelineType == GraphicsAPI::D3D12)
diff --git a/renderdoc/driver/d3d11/d3d11_device.cpp b/renderdoc/driver/d3d11/d3d11_device.cpp
index 7dd0bec89..e8ff0353e 100644
--- a/renderdoc/driver/d3d11/d3d11_device.cpp
+++ b/renderdoc/driver/d3d11/d3d11_device.cpp
@@ -1489,6 +1489,7 @@ void WrappedID3D11Device::ReplayLog(uint32_t startEventID, uint32_t endEventID,
 
   if(!partial)
   {
+    RENDERDOC_PROFILEREGION("ApplyInitialContents");
     D3D11MarkerRegion apply("!!!!RenderDoc Internal: ApplyInitialContents");
     GetResourceManager()->ApplyInitialContents();
   }
diff --git a/renderdoc/driver/d3d12/d3d12_device.cpp b/renderdoc/driver/d3d12/d3d12_device.cpp
index dd353637c..34215095e 100644
--- a/renderdoc/driver/d3d12/d3d12_device.cpp
+++ b/renderdoc/driver/d3d12/d3d12_device.cpp
@@ -1271,6 +1271,8 @@ ID3D12Resource *WrappedID3D12Device::GetUploadBuffer(uint64_t chunkOffset, uint6
 
 void WrappedID3D12Device::ApplyInitialContents()
 {
+  RENDERDOC_PROFILEFUNCTION();
+
   initStateCurBatch = 0;
   initStateCurList = NULL;
 
diff --git a/renderdoc/driver/gl/gl_driver.cpp b/renderdoc/driver/gl/gl_driver.cpp
index dc2aeee15..7a9441211 100644
--- a/renderdoc/driver/gl/gl_driver.cpp
+++ b/renderdoc/driver/gl/gl_driver.cpp
@@ -5636,6 +5636,7 @@ void WrappedOpenGL::ReplayLog(uint32_t startEventID, uint32_t endEventID, Replay
 
   if(!partial)
   {
+    RENDERDOC_PROFILEREGION("ApplyInitialContents");
     GLMarkerRegion apply("!!!!RenderDoc Internal: ApplyInitialContents");
     GetResourceManager()->ApplyInitialContents();
 
diff --git a/renderdoc/driver/vulkan/vk_core.cpp b/renderdoc/driver/vulkan/vk_core.cpp
index bbcbb59fe..532ce9df8 100644
--- a/renderdoc/driver/vulkan/vk_core.cpp
+++ b/renderdoc/driver/vulkan/vk_core.cpp
@@ -254,6 +254,7 @@ void WrappedVulkan::AddFreeCommandBuffer(VkCommandBuffer cmd)
 void WrappedVulkan::SubmitCmds(VkSemaphore *unwrappedWaitSemaphores,
                                VkPipelineStageFlags *waitStageMask, uint32_t waitSemaphoreCount)
 {
+  RENDERDOC_PROFILEFUNCTION();
   // nothing to do
   if(m_InternalCmds.pendingcmds.empty())
     return;
@@ -333,6 +334,7 @@ void WrappedVulkan::SubmitSemaphores()
 
 void WrappedVulkan::FlushQ()
 {
+  RENDERDOC_PROFILEFUNCTION();
   // VKTODOLOW could do away with the need for this function by keeping
   // commands until N presents later, or something, or checking on fences.
   // If we do so, then check each use for FlushQ to see if it needs a
@@ -2695,6 +2697,7 @@ ReplayStatus WrappedVulkan::ContextReplayLog(CaptureState readType, uint32_t sta
 
 void WrappedVulkan::ApplyInitialContents()
 {
+  RENDERDOC_PROFILEFUNCTION();
   VkMarkerRegion region("ApplyInitialContents");
 
   // check that we have all external queues necessary
diff --git a/renderdoc/replay/replay_controller.cpp b/renderdoc/replay/replay_controller.cpp
index 1aa1cce21..558055d9a 100644
--- a/renderdoc/replay/replay_controller.cpp
+++ b/renderdoc/replay/replay_controller.cpp
@@ -91,6 +91,7 @@ ReplayController::~ReplayController()
 void ReplayController::SetFrameEvent(uint32_t eventId, bool force)
 {
   CHECK_REPLAY_THREAD();
+  RENDERDOC_PROFILEFUNCTION();
 
   if(eventId != m_EventID || force)
   {
@@ -165,6 +166,8 @@ rdcstr ReplayController::DisassembleShader(ResourceId pipeline, const ShaderRefl
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   if(refl == NULL)
     return "; Error: No shader specified";
 
@@ -419,6 +422,8 @@ rdcarray<CounterResult> ReplayController::FetchCounters(const rdcarray<GPUCounte
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   return m_pDevice->FetchCounters(counters);
 }
 
@@ -533,6 +538,7 @@ bytebuf ReplayController::GetBufferData(ResourceId buff, uint64_t offset, uint64
 bytebuf ReplayController::GetTextureData(ResourceId tex, const Subresource &sub)
 {
   CHECK_REPLAY_THREAD();
+  RENDERDOC_PROFILEFUNCTION();
 
   bytebuf ret;
 
@@ -552,6 +558,7 @@ bytebuf ReplayController::GetTextureData(ResourceId tex, const Subresource &sub)
 bool ReplayController::SaveTexture(const TextureSave &saveData, const char *path)
 {
   CHECK_REPLAY_THREAD();
+  RENDERDOC_PROFILEFUNCTION();
 
   TextureSave sd = saveData;    // mutable copy
   ResourceId liveid = m_pDevice->GetLiveID(sd.resourceId);
@@ -1414,6 +1421,8 @@ rdcarray<PixelModification> ReplayController::PixelHistory(ResourceId target, ui
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   rdcarray<PixelModification> ret;
 
   Subresource subresource = sub;
@@ -1541,6 +1550,7 @@ PixelValue ReplayController::PickPixel(ResourceId tex, uint32_t x, uint32_t y,
                                        const Subresource &sub, CompType typeCast)
 {
   CHECK_REPLAY_THREAD();
+  RENDERDOC_PROFILEFUNCTION();
 
   PixelValue ret;
 
@@ -1587,6 +1597,8 @@ ShaderDebugTrace *ReplayController::DebugVertex(uint32_t vertid, uint32_t instid
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   ShaderDebugTrace *ret = m_pDevice->DebugVertex(m_EventID, vertid, instid, idx, view);
 
   SetFrameEvent(m_EventID, true);
@@ -1599,6 +1611,8 @@ ShaderDebugTrace *ReplayController::DebugPixel(uint32_t x, uint32_t y, uint32_t
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   ShaderDebugTrace *ret = m_pDevice->DebugPixel(m_EventID, x, y, sample, primitive);
 
   SetFrameEvent(m_EventID, true);
@@ -1610,6 +1624,8 @@ ShaderDebugTrace *ReplayController::DebugThread(const uint32_t groupid[3], const
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   ShaderDebugTrace *ret = m_pDevice->DebugThread(m_EventID, groupid, threadid);
 
   SetFrameEvent(m_EventID, true);
@@ -1621,6 +1637,8 @@ rdcarray<ShaderDebugState> ReplayController::ContinueDebug(ShaderDebugger *debug
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   rdcarray<ShaderDebugState> ret = m_pDevice->ContinueDebug(debugger);
 
   return ret;
@@ -1643,6 +1661,8 @@ rdcarray<ShaderVariable> ReplayController::GetCBufferVariableContents(
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   bytebuf data;
   if(buffer != ResourceId())
   {
@@ -1857,6 +1877,8 @@ rdcpair<ResourceId, rdcstr> ReplayController::BuildTargetShader(
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   if(source.empty())
     return rdcpair<ResourceId, rdcstr>(ResourceId(), "0-byte shader is not valid");
 
@@ -1975,6 +1997,8 @@ ReplayStatus ReplayController::CreateDevice(RDCFile *rdc, const ReplayOptions &o
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   IReplayDriver *driver = NULL;
   ReplayStatus status = RenderDoc::Inst().CreateReplayDriver(rdc, opts, &driver);
 
@@ -2006,6 +2030,8 @@ ReplayStatus ReplayController::PostCreateInit(IReplayDriver *device, RDCFile *rd
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   m_pDevice = device;
 
   m_APIProps = m_pDevice->GetAPIProperties();
@@ -2054,6 +2080,8 @@ void ReplayController::FetchPipelineState(uint32_t eventId)
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   m_pDevice->SavePipelineState(eventId);
 
   m_D3D11PipelineState = m_pDevice->GetD3D11PipelineState();
diff --git a/renderdoc/replay/replay_output.cpp b/renderdoc/replay/replay_output.cpp
index 3b75515ce..b964d2fd1 100644
--- a/renderdoc/replay/replay_output.cpp
+++ b/renderdoc/replay/replay_output.cpp
@@ -389,6 +389,8 @@ rdcpair<uint32_t, uint32_t> ReplayOutput::PickVertex(uint32_t x, uint32_t y)
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   DrawcallDescription *draw = m_pRenderer->GetDrawcallByEID(m_EventID);
 
   const rdcpair<uint32_t, uint32_t> errorReturn = {~0U, ~0U};
@@ -566,6 +568,8 @@ void ReplayOutput::Display()
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   if(m_pDevice->CheckResizeOutputWindow(m_MainOutput.outputID))
   {
     m_pDevice->GetOutputWindowDimensions(m_MainOutput.outputID, m_Width, m_Height);
@@ -680,6 +684,8 @@ void ReplayOutput::DisplayTex()
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   DrawcallDescription *draw = m_pRenderer->GetDrawcallByEID(m_EventID);
 
   if(m_MainOutput.outputID == 0)
@@ -767,6 +773,8 @@ void ReplayOutput::DisplayMesh()
 {
   CHECK_REPLAY_THREAD();
 
+  RENDERDOC_PROFILEFUNCTION();
+
   DrawcallDescription *draw = m_pRenderer->GetDrawcallByEID(m_EventID);
 
   if(draw == NULL || m_MainOutput.outputID == 0 || m_Width <= 0 || m_Height <= 0 ||