Implement derivatives for HLSL intrinsics. (#2684)

* Implement derivatives for HLSL intrinsics.

* Vector intrinsics.

* Add all intrinsics.

---------

Co-authored-by: Yong He <yhe@nvidia.com>

4 files changed, 93 insertions, 0 deletions

diff --git a/tests/autodiff-dstdlib/vector-cross.slang b/tests/autodiff-dstdlib/vector-cross.slang
new file mode 100644
index 000000000..be08894cb
--- /dev/null
+++ b/tests/autodiff-dstdlib/vector-cross.slang
@@ -0,0 +1,40 @@
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0 0 0 0 0 0 0 0], stride=4):out,name=outputBuffer
+RWStructuredBuffer<float> outputBuffer;
+
+[BackwardDifferentiable]
+float3 crossImpl(float3 x, float3 y)
+{
+    return float3(x.y * y.z - y.y * x.z,
+                  x.z * y.x - y.z * x.x,
+                  x.x * y.y - y.x * x.y);
+}
+
+[numthreads(1, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    {
+        let x = float3(-0.51, 0.74, 0.86);
+        let y = float3(1.43, -0.92, 4.36);
+        let dOut = float3(3.41, 6.55, 2.39);
+        var dpx = diffPair(x);
+        var dpy = diffPair(y);
+        __bwd_diff(cross)(dpx, dpy, dOut);
+        outputBuffer[0] = dpx.d[0]; 
+        outputBuffer[1] = dpx.d[1]; 
+        outputBuffer[2] = dpx.d[2]; 
+        outputBuffer[3] = dpy.d[0]; 
+        outputBuffer[4] = dpy.d[1]; 
+        outputBuffer[5] = dpy.d[2]; 
+
+        __bwd_diff(crossImpl)(dpx, dpy, dOut);
+        outputBuffer[6] = dpx.d[0];
+        outputBuffer[7] = dpx.d[1];
+        outputBuffer[8] = dpx.d[2];
+        outputBuffer[9] = dpy.d[0];
+        outputBuffer[10] = dpy.d[1];
+        outputBuffer[11] = dpy.d[2];
+    }
+}
diff --git a/tests/autodiff-dstdlib/vector-cross.slang.expected.txt b/tests/autodiff-dstdlib/vector-cross.slang.expected.txt
new file mode 100644
index 000000000..9d472f078
--- /dev/null
+++ b/tests/autodiff-dstdlib/vector-cross.slang.expected.txt
@@ -0,0 +1,13 @@
+type: float
+-30.756802
+11.449901
+12.503700
+3.864400
+-4.151500
+5.863900
+-30.756804
+11.449901
+12.503700
+3.864400
+-4.151500
+5.863900
diff --git a/tests/autodiff-dstdlib/vector-length.slang b/tests/autodiff-dstdlib/vector-length.slang
new file mode 100644
index 000000000..c5064e54e
--- /dev/null
+++ b/tests/autodiff-dstdlib/vector-length.slang
@@ -0,0 +1,36 @@
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
+
+//TEST_INPUT:ubuffer(data=[0 0 0], stride=4):out,name=outputBuffer
+RWStructuredBuffer<float> outputBuffer;
+
+float lengthDiffX(float3 x)
+{
+    return (length(float3(x.x + 0.001, x.yz)) - length(float3(x.x - 0.001, x.yz))) / 0.002;
+}
+float lengthDiffY(float3 x)
+{
+    return (length(float3(x.x, x.y + 0.001, x.z)) - length(float3(x.x, x.y - 0.001, x.z))) / 0.002;
+}
+float lengthDiffZ(float3 x)
+{
+    return (length(float3(x.xy, x.z + 0.001)) - length(float3(x.xy, x.z - 0.001))) / 0.002;
+}
+
+[numthreads(1, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    {
+        let x = float3(12, 23, 31);
+        var dpx = diffPair(x);
+        __bwd_diff(length)(dpx, 1.0);
+        outputBuffer[0] = dpx.d[0];
+        outputBuffer[1] = dpx.d[1];
+        outputBuffer[2] = dpx.d[2];
+
+        // for reference:
+        //outputBuffer[3] = lengthDiffX(x);
+        //outputBuffer[4] = lengthDiffY(x);
+        //outputBuffer[5] = lengthDiffZ(x);
+    }
+}
diff --git a/tests/autodiff-dstdlib/vector-length.slang.expected.txt b/tests/autodiff-dstdlib/vector-length.slang.expected.txt
new file mode 100644
index 000000000..3c3f3727d
--- /dev/null
+++ b/tests/autodiff-dstdlib/vector-length.slang.expected.txt
@@ -0,0 +1,4 @@
+type: float
+0.296862
+0.568986
+0.766895