Overhaul `transposeParameterBlock` to support `inout` params. (#2621)

* Overhaul `transposeParameterBlock` to support `inout` params.

* Small bug fixes.

* Bug fix on differentiable intrinsic specialization.

* Fixes.

* Run autodiff tests on CPU.

* Clean up.

* More bug fixes.,

* Add test coverage on inout param.

* Fix language server hinting for transcribed mutable params.

---------

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

15 files changed, 194 insertions, 16 deletions

diff --git a/tests/autodiff-dstdlib/dstdlib-elementary-bwd.slang b/tests/autodiff-dstdlib/dstdlib-elementary-bwd.slang
index ce282eb77..379e2c3ef 100644
--- a/tests/autodiff-dstdlib/dstdlib-elementary-bwd.slang
+++ b/tests/autodiff-dstdlib/dstdlib-elementary-bwd.slang
@@ -1,7 +1,8 @@
 //TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
 //TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
 
-//TEST_INPUT:ubuffer(data=[0 0 0 0], stride=4):out,name=outputBuffer
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;
 
 typedef DifferentialPair<float> dpfloat;
@@ -14,6 +15,12 @@ float diffExp(float x)
     return exp(x);
 }
 
+[BackwardDifferentiable]
+float diffSin(float x)
+{
+    return sin(x);
+}
+
 [numthreads(1, 1, 1)]
 void computeMain(uint3 dispatchThreadID: SV_DispatchThreadID)
 {
@@ -30,4 +37,10 @@ void computeMain(uint3 dispatchThreadID: SV_DispatchThreadID)
         outputBuffer[2] = dpx.p; // Expect: 0.000000
         outputBuffer[3] = dpx.d; // Expect: 1.000000
     }
+
+    {
+        dpfloat dpx = dpfloat(float.getPi(), 1.0);
+        __bwd_diff(diffSin)(dpx, 1.0);
+        outputBuffer[4] = dpx.d; // Expect: -1.000000
+    }
 }
diff --git a/tests/autodiff-dstdlib/dstdlib-elementary-bwd.slang.expected.txt b/tests/autodiff-dstdlib/dstdlib-elementary-bwd.slang.expected.txt
index 49e34b0bf..a4b804cb8 100644
--- a/tests/autodiff-dstdlib/dstdlib-elementary-bwd.slang.expected.txt
+++ b/tests/autodiff-dstdlib/dstdlib-elementary-bwd.slang.expected.txt
@@ -2,4 +2,5 @@ type: float
 2.000000
 7.389056
 0.000000
-1.000000
\ No newline at end of file
+1.000000
+-1.00000
\ No newline at end of file
diff --git a/tests/autodiff/reverse-addr-eliminate.slang b/tests/autodiff/reverse-addr-eliminate.slang
index e23e83e6a..49f34a6e3 100644
--- a/tests/autodiff/reverse-addr-eliminate.slang
+++ b/tests/autodiff/reverse-addr-eliminate.slang
@@ -1,5 +1,6 @@
 //TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
 //TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
 
 //TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;
diff --git a/tests/autodiff/reverse-array-out-param.slang b/tests/autodiff/reverse-array-out-param.slang
index e1c822489..c7965ca5d 100644
--- a/tests/autodiff/reverse-array-out-param.slang
+++ b/tests/autodiff/reverse-array-out-param.slang
@@ -1,5 +1,6 @@
 //TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
 //TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
 
 //TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;
@@ -39,17 +40,4 @@ void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
 
     outputBuffer[0] = dpd.d.m;
     outputBuffer[1] = dpd.d.n;
-}
-
-[BackwardDifferentiable]
-void inner(float x, out float y)
-{
-    y = 2 * x;
-}
-
-[BackwardDifferentiable]
-void outter(float x, out float y)
-{
-    inner(x, y);
-    inner(x, y);
-}
+}
\ No newline at end of file
diff --git a/tests/autodiff/reverse-inout-param-0.slang b/tests/autodiff/reverse-inout-param-0.slang
new file mode 100644
index 000000000..13bfb6277
--- /dev/null
+++ b/tests/autodiff/reverse-inout-param-0.slang
@@ -0,0 +1,31 @@
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
+RWStructuredBuffer<float> outputBuffer;
+
+[BackwardDifferentiable]
+void g(float x, out float y)
+{
+    y = (x + 1) * (x + 1);
+}
+
+[BackwardDifferentiable]
+void f(float x, out float y)
+{
+    g(x, y);
+    g(x, y);
+}
+
+[numthreads(1, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    var x = diffPair(5.0);
+    float yDiffOut = 1.0;
+
+    __bwd_diff(f)(x, yDiffOut);
+
+    outputBuffer[0] = x.p; // should be 5, since bwd_diff does not write back new primal val.
+    outputBuffer[1] = x.d; // 12.0
+}
\ No newline at end of file
diff --git a/tests/autodiff/reverse-inout-param-0.slang.expected.txt b/tests/autodiff/reverse-inout-param-0.slang.expected.txt
new file mode 100644
index 000000000..a11719af3
--- /dev/null
+++ b/tests/autodiff/reverse-inout-param-0.slang.expected.txt
@@ -0,0 +1,7 @@
+type: float
+5.000000
+12.000000
+0.000000
+0.000000
+0.000000
+
diff --git a/tests/autodiff/reverse-inout-param-2.slang b/tests/autodiff/reverse-inout-param-2.slang
new file mode 100644
index 000000000..1fa64751b
--- /dev/null
+++ b/tests/autodiff/reverse-inout-param-2.slang
@@ -0,0 +1,72 @@
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
+RWStructuredBuffer<float> outputBuffer;
+
+struct D : IDifferentiable
+{
+    float m;
+    float n;
+}
+
+struct ND
+{
+    float nd;
+}
+
+[BackwardDifferentiable]
+void g(
+    inout no_diff D p,
+    out no_diff D po,
+    out ND v1,
+    inout ND v2,
+    float x,
+    out float y)
+{
+    v1 = v2;
+    v2.nd = v2.nd + 1.0;
+    p.n = v1.nd + 1.0;
+    p.m = v2.nd + 1.0 + x; // == v2.nd + 2 + x == 1 + 2 + x == 3+x
+    po = p;
+    po.m += 1.0; // == 4+x
+    y = p.m * x; // == (3+x)*x
+}
+
+[BackwardDifferentiable]
+void f(inout no_diff D p, out no_diff D p0, out ND v1, inout ND v2, float x, out float y)
+{
+    // v2.nd is 3.
+    g(p, p0, v1, v2, x, y);
+    // v2.nd is now 4, now g is equivalent to (4+x)*x.
+    g(p, p0, v1, v2, x, y);
+}
+
+[ForwardDifferentiable]
+float f_ref(float x)
+{
+    return (3 + 3 * x + x * x) * (3 * x + x * x);
+}
+
+[numthreads(1, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    D p;
+    p.m = 1.0;
+    p.n = 2.0;
+
+    ND v2 = { 1.0 };
+
+    var x = diffPair(5.0);
+    float yDiffOut = 1.0;
+
+    __bwd_diff(f)(p, v2, x, yDiffOut);
+
+    // (3+((3+x)*x))*((3+x)*x) = (3+3x+x^2)*(3x+x^2)
+    outputBuffer[0] = x.p; // should be 5, since bwd_diff does not write back new primal val.
+    outputBuffer[1] = x.d; // 14
+    outputBuffer[2] = p.m; // 1.0
+    outputBuffer[3] = p.n; // 2.0
+    outputBuffer[4] = v2.nd; // 1.0
+}
\ No newline at end of file
diff --git a/tests/autodiff/reverse-inout-param-2.slang.expected.txt b/tests/autodiff/reverse-inout-param-2.slang.expected.txt
new file mode 100644
index 000000000..65933cc7d
--- /dev/null
+++ b/tests/autodiff/reverse-inout-param-2.slang.expected.txt
@@ -0,0 +1,7 @@
+type: float
+5.000000
+14.000000
+1.000000
+2.000000
+1.000000
+
diff --git a/tests/autodiff/reverse-inout-param.slang b/tests/autodiff/reverse-inout-param.slang
new file mode 100644
index 000000000..7d7f4cb05
--- /dev/null
+++ b/tests/autodiff/reverse-inout-param.slang
@@ -0,0 +1,47 @@
+//TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
+//TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
+RWStructuredBuffer<float> outputBuffer;
+
+struct D : IDifferentiable
+{
+    float n;
+    float m;
+}
+
+[BackwardDifferentiable]
+void g(no_diff float p, inout float x)
+{
+    x = p * ((x+1)*(x+1));
+}
+
+[BackwardDifferentiable]
+void f(no_diff float p, inout float x)
+{
+    g(p, x);
+    g(p, x);
+}
+[BackwardDifferentiable]
+float f_ref(no_diff float p, float x)
+{
+    float y1 = p * (x+1)*(x+1);
+    float y2 = p * (y1+1)*(y1+1);
+    return y2;
+}
+
+[numthreads(1, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID)
+{
+    var x = diffPair(2.0, 1.0);
+
+    __bwd_diff(f)(3.0, x);
+
+    outputBuffer[0] = x.p; // should be 2, since bwd_diff does not write back new primal val.
+    outputBuffer[1] = x.d; // 3024
+
+    var refVal = __fwd_diff(f_ref)(3.0, diffPair(2.0, 1.0)).d;
+    outputBuffer[2] = refVal; // 3024
+
+}
\ No newline at end of file
diff --git a/tests/autodiff/reverse-inout-param.slang.expected.txt b/tests/autodiff/reverse-inout-param.slang.expected.txt
new file mode 100644
index 000000000..2df174e2f
--- /dev/null
+++ b/tests/autodiff/reverse-inout-param.slang.expected.txt
@@ -0,0 +1,6 @@
+type: float
+2.000000
+3024.000000
+3024.000000
+0.000000
+0.000000
diff --git a/tests/autodiff/reverse-loop.slang b/tests/autodiff/reverse-loop.slang
index 6ae207080..46d707548 100644
--- a/tests/autodiff/reverse-loop.slang
+++ b/tests/autodiff/reverse-loop.slang
@@ -1,6 +1,7 @@
 //TEST_IGNORE_FILE:
 //TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
 //TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
 
 //TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;
diff --git a/tests/autodiff/reverse-multi-return.slang b/tests/autodiff/reverse-multi-return.slang
index ee8bb9a4c..21f72619e 100644
--- a/tests/autodiff/reverse-multi-return.slang
+++ b/tests/autodiff/reverse-multi-return.slang
@@ -1,5 +1,6 @@
 //TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
 //TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
 
 //TEST_INPUT:ubuffer(data=[0 0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;
diff --git a/tests/autodiff/reverse-nested-calls.slang b/tests/autodiff/reverse-nested-calls.slang
index 55d22e191..caf2df6f8 100644
--- a/tests/autodiff/reverse-nested-calls.slang
+++ b/tests/autodiff/reverse-nested-calls.slang
@@ -1,5 +1,6 @@
 //TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
 //TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
 
 //TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;
diff --git a/tests/autodiff/reverse-single-iter-loop.slang b/tests/autodiff/reverse-single-iter-loop.slang
index 20c26e000..3597f0ec8 100644
--- a/tests/autodiff/reverse-single-iter-loop.slang
+++ b/tests/autodiff/reverse-single-iter-loop.slang
@@ -1,5 +1,6 @@
 //TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
 //TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
 
 //TEST_INPUT:ubuffer(data=[0 0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;
diff --git a/tests/autodiff/reverse-struct-multi-write.slang b/tests/autodiff/reverse-struct-multi-write.slang
index dd12c7d3d..2a63141cb 100644
--- a/tests/autodiff/reverse-struct-multi-write.slang
+++ b/tests/autodiff/reverse-struct-multi-write.slang
@@ -1,6 +1,7 @@
 
 //TEST(compute):COMPARE_COMPUTE_EX:-slang -compute -shaderobj -output-using-type
 //TEST(compute, vulkan):COMPARE_COMPUTE_EX:-vk -compute -shaderobj -output-using-type
+//TEST(compute):COMPARE_COMPUTE_EX:-cpu -compute -output-using-type -shaderobj
 
 //TEST_INPUT:ubuffer(data=[0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;