diff --git a/renderdoc/driver/shaders/spirv/spirv_debug.cpp b/renderdoc/driver/shaders/spirv/spirv_debug.cpp
index 6b3f994af..bc1cc9cf5 100644
--- a/renderdoc/driver/shaders/spirv/spirv_debug.cpp
+++ b/renderdoc/driver/shaders/spirv/spirv_debug.cpp
@@ -5142,3 +5142,16 @@ rdcstr DoStringise(const rdcspv::StepThreadMode &el)
   }
   END_ENUM_STRINGISE();
 };
+
+template <>
+rdcstr DoStringise(const rdcspv::DeviceOpResult &el)
+{
+  BEGIN_ENUM_STRINGISE(rdcspv::DeviceOpResult)
+  {
+    STRINGISE_ENUM_CLASS(Unknown)
+    STRINGISE_ENUM_CLASS(Succeeded)
+    STRINGISE_ENUM_CLASS(Failed)
+    STRINGISE_ENUM_CLASS(NeedsDevice)
+  }
+  END_ENUM_STRINGISE();
+};
diff --git a/renderdoc/driver/shaders/spirv/spirv_debug.h b/renderdoc/driver/shaders/spirv/spirv_debug.h
index ee329b60c..e3807dc49 100644
--- a/renderdoc/driver/shaders/spirv/spirv_debug.h
+++ b/renderdoc/driver/shaders/spirv/spirv_debug.h
@@ -58,6 +58,14 @@ enum class ThreadProperty : uint32_t
 
 ITERABLE_OPERATORS(ThreadProperty);
 
+enum class DeviceOpResult : uint32_t
+{
+  Unknown,
+  Succeeded,
+  Failed,
+  NeedsDevice,
+};
+
 inline void AtomicStore(int32_t *var, int32_t newVal)
 {
   int32_t oldVal = *var;
@@ -372,6 +380,7 @@ struct ThreadState
     AtomicStore(&atomic_isSimulationStepActive, 1);
     AtomicStore(&atomic_stepNeedsGpuSampleGatherOp, 0);
     AtomicStore(&atomic_stepNeedsGpuMathOp, 0);
+    AtomicStore(&atomic_stepNeedsDeviceThread, 0);
   }
   void SetStepNeedsGpuSampleGatherOp()
   {
@@ -388,6 +397,12 @@ struct ThreadState
     SetPendingResultStatus(PendingResultStatus::Pending);
   }
   bool StepNeedsGpuMathOp() const { return (AtomicLoad(&atomic_stepNeedsGpuMathOp) == 1); }
+  void SetStepNeedsDeviceThread()
+  {
+    AtomicStore(&atomic_stepNeedsDeviceThread, 1);
+    SetPendingResultStatus(PendingResultStatus::Pending);
+  }
+  bool StepNeedsDeviceThread() const { return (AtomicLoad(&atomic_stepNeedsDeviceThread) == 1); }
   const GpuMathOperation &GetQueuedGpuMathOp() const
   {
     RDCASSERT(AtomicLoad(&atomic_stepNeedsGpuMathOp));
@@ -455,6 +470,7 @@ private:
   int32_t atomic_pendingResultStatus = (int32_t)PendingResultStatus::Unknown;
   int32_t atomic_stepNeedsGpuSampleGatherOp = 0;
   int32_t atomic_stepNeedsGpuMathOp = 0;
+  int32_t atomic_stepNeedsDeviceThread = 0;
   int32_t atomic_isSimulationStepActive = 0;
 };
 
@@ -757,6 +773,9 @@ private:
   void ClampScalars(const ShaderVariable &var, uint8_t &scalar0) const;
   void ClampScalars(const ShaderVariable &var, uint8_t &scalar0, uint8_t &scalar1) const;
 
+  void QueueDeviceThreadStep(uint32_t lane);
+  void ProcessQueuedDeviceThreadSteps();
+
   void QueueJob(uint32_t lane, rdcarray<ShaderDebugState> *ret);
   void StepThread(uint32_t lane, StepThreadMode stepMode, rdcarray<ShaderDebugState> *ret);
   void InternalStepThread(uint32_t lane, rdcarray<ShaderDebugState> *ret);
@@ -772,6 +791,7 @@ private:
   mutable rdcarray<DebugMessage> queuedDebugMessages;
   rdcarray<bool> queuedGpuMathOps;
   rdcarray<bool> queuedGpuSampleGatherOps;
+  rdcarray<bool> queuedDeviceThreadSteps;
   rdcarray<ShaderDebugState> *shaderChangesReturn;
 
   bool retireIDs = true;
@@ -792,3 +812,4 @@ void AssignValue(ShaderVariable &dst, const ShaderVariable &src);
 
 DECLARE_REFLECTION_ENUM(rdcspv::ThreadState::PendingResultStatus);
 DECLARE_REFLECTION_ENUM(rdcspv::StepThreadMode);
+DECLARE_REFLECTION_ENUM(rdcspv::DeviceOpResult);
diff --git a/renderdoc/driver/shaders/spirv/spirv_debug_setup.cpp b/renderdoc/driver/shaders/spirv/spirv_debug_setup.cpp
index d0018b9d5..1cb6e3a7c 100644
--- a/renderdoc/driver/shaders/spirv/spirv_debug_setup.cpp
+++ b/renderdoc/driver/shaders/spirv/spirv_debug_setup.cpp
@@ -1024,12 +1024,14 @@ ShaderDebugTrace *Debugger::BeginDebug(DebugAPIWrapper *api, const ShaderStage s
   stage = shaderStage;
   apiWrapper = api;
 
+  queuedDeviceThreadSteps.resize(threadsInWorkgroup);
   queuedGpuMathOps.resize(threadsInWorkgroup);
   queuedGpuSampleGatherOps.resize(threadsInWorkgroup);
   pendingLanes.resize(threadsInWorkgroup);
   for(uint32_t i = 0; i < threadsInWorkgroup; i++)
   {
     workgroup.push_back(ThreadState(*this, global));
+    queuedDeviceThreadSteps[i] = false;
     queuedGpuMathOps[i] = false;
     queuedGpuSampleGatherOps[i] = false;
     pendingLanes[i] = false;
@@ -2693,6 +2695,7 @@ rdcarray<ShaderDebugState> Debugger::ContinueDebug()
     do
     {
       ProcessQueuedDebugMessages();
+      ProcessQueuedDeviceThreadSteps();
       // Convert the simulation threads queued operations into pending operations i.e. GPU commands
       ProcessQueuedOps();
       // Sync any pending GPU operations and set the results to the pending threads
@@ -4669,6 +4672,8 @@ void Debugger::StepThread(uint32_t lane, StepThreadMode stepMode, rdcarray<Shade
       break;
     else if(thread.StepNeedsGpuMathOp())
       break;
+    else if(thread.StepNeedsDeviceThread())
+      break;
 
     if(isActiveThread)
       curActiveSteps++;
@@ -4695,6 +4700,12 @@ void Debugger::StepThread(uint32_t lane, StepThreadMode stepMode, rdcarray<Shade
     QueueGpuMathOp(lane);
     return;
   }
+  if(thread.StepNeedsDeviceThread())
+  {
+    RDCASSERT(!simulateStep);
+    QueueDeviceThreadStep(lane);
+    return;
+  }
   RDCASSERT(!thread.IsPendingResultPending());
   thread.SetSimulationStepCompleted();
 }
@@ -4744,6 +4755,8 @@ void Debugger::InternalStepThread(uint32_t lane, rdcarray<ShaderDebugState> *ret
       return;
     if(thread.StepNeedsGpuMathOp())
       return;
+    if(thread.StepNeedsDeviceThread())
+      return;
 
     if(!thread.IsPendingResultPending())
     {
@@ -4821,6 +4834,8 @@ void Debugger::InternalStepThread(uint32_t lane, rdcarray<ShaderDebugState> *ret
       return;
     if(thread.StepNeedsGpuMathOp())
       return;
+    if(thread.StepNeedsDeviceThread())
+      return;
   }
 }
 
@@ -4856,6 +4871,32 @@ void Debugger::AddDebugMessage(MessageCategory c, MessageSeverity sv, MessageSou
   queuedDebugMessages.push_back({c, sv, src, d});
 }
 
+// Can be called from any thread
+void Debugger::QueueDeviceThreadStep(uint32_t lane)
+{
+  ThreadState &thread = workgroup[lane];
+  RDCASSERT(thread.IsSimulationStepActive());
+  thread.SetStepQueued();
+  RDCASSERT(!queuedDeviceThreadSteps[lane]);
+  queuedDeviceThreadSteps[lane] = true;
+}
+
+// Must be called from the replay manager thread (the debugger thread)
+void Debugger::ProcessQueuedDeviceThreadSteps()
+{
+  CHECK_DEBUGGER_THREAD();
+  for(uint32_t lane = 0; lane < queuedDeviceThreadSteps.size(); ++lane)
+  {
+    if(queuedDeviceThreadSteps[lane])
+    {
+      queuedDeviceThreadSteps[lane] = false;
+      ThreadState &thread = workgroup[lane];
+      thread.SetPendingResultUnknown();
+      RDCASSERT(thread.IsSimulationStepActive());
+      StepThread(lane, StepThreadMode::QUEUE_MULTIPLE_STEPS, shaderChangesReturn);
+    }
+  }
+}
 };    // namespace rdcspv
 
 #if ENABLED(ENABLE_UNIT_TESTS)