Add support for `kIROp_FloatCast` (#2824)

* Add support for `kIROp_FloatCast`

* Update float-cast.slang

* Added flag to d3d options

3 files changed, 91 insertions, 0 deletions

diff --git a/source/slang/slang-ir-autodiff-transpose.h b/source/slang/slang-ir-autodiff-transpose.h
index c7ac8c357..55f042352 100644
--- a/source/slang/slang-ir-autodiff-transpose.h
+++ b/source/slang/slang-ir-autodiff-transpose.h
@@ -1744,6 +1744,9 @@ struct DiffTransposePass
             case kIROp_UpdateElement:
                 return transposeUpdateElement(builder, fwdInst, revValue);
 
+            case kIROp_FloatCast:
+                return transposeFloatCast(builder, fwdInst, revValue);
+
             case kIROp_LoadReverseGradient:
             case kIROp_ReverseGradientDiffPairRef:
             case kIROp_DefaultConstruct:
@@ -2228,6 +2231,22 @@ struct DiffTransposePass
         return TranspositionResult(gradients);
     }
 
+    TranspositionResult transposeFloatCast(IRBuilder* builder, IRInst* fwdInst, IRInst* revValue)
+    {
+        // (A = cast<T, U>(B)) -> (dB += cast<U, T>(dA))
+        return TranspositionResult(
+            List<RevGradient>(
+                RevGradient(
+                    RevGradient::Flavor::Simple,
+                    fwdInst->getOperand(0),
+                    builder->emitIntrinsicInst(
+                        fwdInst->getOperand(0)->getDataType(),
+                        kIROp_FloatCast,
+                        1,
+                        &revValue),
+                    fwdInst)));
+    }
+
     // Gather all reverse-mode gradients for a Load inst, aggregate them and store them in the ptr.
     // 
     void accumulateGradientsForLoad(IRBuilder* builder, IRLoad* revLoad)
diff --git a/tests/autodiff/float-cast.slang b/tests/autodiff/float-cast.slang
new file mode 100644
index 000000000..2c9797fa6
--- /dev/null
+++ b/tests/autodiff/float-cast.slang
@@ -0,0 +1,66 @@
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type -render-features half
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type -render-features half
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
+RWStructuredBuffer<float> outputBuffer;
+
+typedef DifferentialPair<float> dpfloat;
+typedef DifferentialPair<double> dpdouble;
+typedef DifferentialPair<half> dphalf;
+typedef float.Differential dfloat;
+typedef double.Differential ddouble;
+typedef half.Differential dhalf;
+
+[BackwardDifferentiable]
+float f(double x)
+{
+    return (float)x;
+}
+
+[BackwardDifferentiable]
+float h(float x, float y)
+{
+    float m = x + y;
+    float n = x - y;
+    return m * n + 2 * x * y;
+}
+
+[BackwardDifferentiable]
+float j(half x, half y)
+{
+    float m = x / y;
+    return m * y;
+}
+
+[numthreads(1, 1, 1)]
+void computeMain(uint3 dispatchThreadID: SV_DispatchThreadID)
+{
+    {
+        dpdouble dpa = dpdouble(2.0, 1.0);
+
+        __bwd_diff(f)(dpa, 1.0);
+
+        outputBuffer[0] = (float)dpa.d; // Expect: 1
+    }
+
+    {
+        dpfloat dpa = dpfloat(2.0, 1.0);
+        dpfloat dpb = dpfloat(1.5, 1.0);
+
+        __bwd_diff(h)(dpa, dpb, 1.0);
+
+        outputBuffer[1] = dpa.d; // Expect: (2 * 2.0) + (2 * 1.5) = 7.0
+        outputBuffer[2] = dpb.d; // Expect: -(2 * 1.5) + (2 * 2.0) = 1.0
+    }
+
+    {
+        dphalf dpa = dphalf((half)2.0, (half)1.0);
+        dphalf dpb = dphalf((half)1.5, (half)1.0);
+
+        __bwd_diff(j)(dpa, dpb, 1.0);
+
+        outputBuffer[3] = dpa.d; // Expect: 1
+        outputBuffer[4] = dpb.d; // Expect: 0
+    }
+
+}
diff --git a/tests/autodiff/float-cast.slang.expected.txt b/tests/autodiff/float-cast.slang.expected.txt
new file mode 100644
index 000000000..d9726be10
--- /dev/null
+++ b/tests/autodiff/float-cast.slang.expected.txt
@@ -0,0 +1,6 @@
+type: float
+1.000000
+7.000000
+1.000000
+1.000000
+0.000000