2 files changed, 65 insertions, 11 deletions
diff --git a/source/slang/slang-ir-autodiff.cpp b/source/slang/slang-ir-autodiff.cpp
index 1909f860c..ddffd0e21 100644
--- a/source/slang/slang-ir-autodiff.cpp
+++ b/source/slang/slang-ir-autodiff.cpp
@@ -693,16 +693,19 @@ struct AutoDiffPass : public InstPassBase
 
     bool lowerIntermediateContextType(IRBuilder* builder)
     {
-        bool changed = false;
+        bool result = false;
         OrderedHashSet<IRInst*> loweredIntermediateTypes;
 
         // Replace all `BackwardDiffIntermediateContextType` insts with the struct type
         // that we generated during backward diff pass.
-        processAllInsts([&](IRInst* inst)
-            {
-                switch (inst->getOp())
+        for (;;)
+        {
+            bool changed = false;
+            processAllInsts([&](IRInst* inst)
                 {
-                case kIROp_BackwardDiffIntermediateContextType:
+                    switch (inst->getOp())
+                    {
+                    case kIROp_BackwardDiffIntermediateContextType:
                     {
                         auto differentiateInst = as<IRBackwardDiffIntermediateContextType>(inst);
 
@@ -719,15 +722,18 @@ struct AutoDiffPass : public InstPassBase
                         }
                     }
                     break;
-                default:
-                    break;
-                }
-            });
-
+                    default:
+                        break;
+                    }
+                });
+            result |= changed;
+            if (!changed)
+                break;
+        }
         // Now we generate the differential type for the intermediate context type
         // to allow higher order differentiation.
         generateDifferentialImplementationForContextType(loweredIntermediateTypes);
-        return changed;
+        return result;
     }
 
     // Utility function for topology sorting the intermediate context types.
diff --git a/tests/autodiff/cuda-kernel-export-2.slang b/tests/autodiff/cuda-kernel-export-2.slang
new file mode 100644
index 000000000..9cbb4e881
--- /dev/null
+++ b/tests/autodiff/cuda-kernel-export-2.slang
@@ -0,0 +1,48 @@
+//DISABLE_TEST:SIMPLE: -target cuda -line-directive-mode none
+
+// Verify that we can output a cuda device function with [CudaDeviceExport].
+// Disabled until we have FileCheck.
+
+
+//////////////////////////////////////////////////////////////////////////
+// Lambda GGX
+//////////////////////////////////////////////////////////////////////////
+
+[CudaDeviceExport]
+[BackwardDifferentiable]
+float lambdaGGX(const float alphaSqr, const float cosTheta)
+{
+    const float SPECULAR_EPSILON = 1e-4f;
+    float _cosTheta = clamp(cosTheta, SPECULAR_EPSILON, 1.0f - SPECULAR_EPSILON);
+    float cosThetaSqr = _cosTheta * _cosTheta;
+    float tanThetaSqr = (1.0 - cosThetaSqr) / cosThetaSqr;
+    return 0.5f * (sqrt(1.0f + alphaSqr * tanThetaSqr) - 1.0f);
+}
+
+[CudaDeviceExport]
+void lambdaGGX_bwd(inout DifferentialPair<float> alphaSqr, inout DifferentialPair<float> cosTheta, const float d_out)
+{
+    __bwd_diff(lambdaGGX)(alphaSqr, cosTheta, d_out);
+}
+
+//////////////////////////////////////////////////////////////////////////
+// Masking Smith
+//////////////////////////////////////////////////////////////////////////
+
+[CudaDeviceExport]
+[BackwardDifferentiable]
+float maskingSmithGGXCorrelated(const float alphaSqr, const float cosThetaI, const float cosThetaO)
+{
+    float lambdaI = lambdaGGX(alphaSqr, cosThetaI);
+    float lambdaO = lambdaGGX(alphaSqr, cosThetaO);
+    return 1.0f / (1.0f + lambdaI + lambdaO);
+}
+
+[CudaDeviceExport]
+void maskingSmithGGXCorrelated_bwd(inout DifferentialPair<float> alphaSqr,
+                                   inout DifferentialPair<float> cosThetaI,
+                                   inout DifferentialPair<float> cosThetaO,
+                                   const float d_out)
+{
+    __bwd_diff(maskingSmithGGXCorrelated)(alphaSqr, cosThetaI, cosThetaO, d_out);
+}