Merge branch 'master' into fix-optix-raytracing-test

Author: ksavoie-nv

source/slang/diff.meta.slang

@@ -1,4 +1,3 @@
-
 /// `[ForwardDerivative(fwdFn)]` attribute can be used to provide a forward-mode
 /// derivative implementation.
 /// Invoking `fwd_diff(decoratedFn)` will place a call to `fwdFn` instead of synthesizing
@@ -80,7 +79,6 @@ attribute_syntax [ForwardDerivative(function)] : ForwardDerivativeAttribute;
 /// For member functions, or functions nested inside namespaces, `bwdFn` may need to be a fully qualified 
 /// name.
 ///
-///
 __attributeTarget(FunctionDeclBase)
 attribute_syntax [BackwardDerivative(function)] : BackwardDerivativeAttribute;
 
@@ -2150,17 +2148,19 @@ DifferentialPair<T> __d_clamp(DifferentialPair<T> dpx, DifferentialPair<T> dpMin
 {
     return DifferentialPair<T>(
         clamp(dpx.p, dpMin.p, dpMax.p),
-        dpx.p < dpMin.p ? dpMin.d : (dpx.p > dpMax.p ? dpMax.d : dpx.d));
+        (dpx.p >= dpMin.p && dpx.p <= dpMax.p) ? dpx.d : (dpx.p < dpMin.p ? dpMin.d : dpMax.d));
 }
 __generic<T : __BuiltinFloatingPointType>
 [BackwardDifferentiable]
 [PreferRecompute]
 [BackwardDerivativeOf(clamp)]
 void __d_clamp(inout DifferentialPair<T> dpx, inout DifferentialPair<T> dpMin, inout DifferentialPair<T> dpMax, T.Differential dOut)
 {
-    dpx = diffPair(dpx.p, dpx.p > dpMin.p && dpx.p < dpMax.p ? dOut : T.dzero());
-    dpMin = diffPair(dpMin.p, dpx.p <= dpMin.p ? dOut : T.dzero());
-    dpMax = diffPair(dpMax.p, dpx.p >= dpMax.p ? dOut : T.dzero());
+    // Propagate the derivative to x if x is within [min, max] (including the boundaries).
+    dpx = diffPair(dpx.p, (dpx.p >= dpMin.p && dpx.p <= dpMax.p) ? dOut : T.dzero());
+    // If x is strictly below min or above max, the gradient is instead applied to the clamp bounds
+    dpMin = diffPair(dpMin.p, dpx.p < dpMin.p ? dOut : T.dzero());
+    dpMax = diffPair(dpMax.p, dpx.p > dpMax.p ? dOut : T.dzero());
 }
 VECTOR_MATRIX_TERNARY_DIFF_IMPL(clamp)
 

source/slang/slang-emit-hlsl.cpp

@@ -1252,23 +1252,30 @@ void HLSLSourceEmitter::emitSimpleValueImpl(IRInst* inst)
             {
             case IRConstant::FloatKind::Nan:
                 {
-                    m_writer->emit("(0.0 / 0.0)");
+                    m_writer->emit("(0.0f / 0.0f)");
                     return;
                 }
             case IRConstant::FloatKind::PositiveInfinity:
                 {
-                    m_writer->emit("(1.0 / 0.0)");
+                    m_writer->emit("(1.0f / 0.0f)");
                     return;
                 }
             case IRConstant::FloatKind::NegativeInfinity:
                 {
-                    m_writer->emit("(-1.0 / 0.0)");
+                    m_writer->emit("(-1.0f / 0.0f)");
                     return;
                 }
             default:
-                break;
+                {
+                    m_writer->emit(constantInst->value.floatVal);
+                    // Add 'f' suffix for 32-bit float literals to ensure DXC treats them as float
+                    if (constantInst->getDataType()->getOp() == kIROp_FloatType)
+                    {
+                        m_writer->emit("f");
+                    }
+                    return;
+                }
             }
-            break;
         }
 
     default:

source/slang/slang-ir-lower-binding-query.cpp

@@ -288,7 +288,118 @@ struct BindingQueryLoweringContext : public WorkListPass
     //
     OpaqueValueInfo computeOpaqueValueInfo(IRInst* opaqueValue)
     {
-        if (auto globalParam = as<IRGlobalParam>(opaqueValue))
+        if (auto getElement = as<IRGetElement>(opaqueValue))
+        {
+            IRInst* baseInst = getElement->getBase();
+            IRInst* indexInst = getElement->getIndex();
+
+            IRInst* elementType = getElement->getDataType();
+
+            // TODO(JS): This a hack to make this work for arrays of resource type.
+            // It won't work in the general case as it stands because we would need
+            // to propogate layout kind types needed at usage sites.
+            // Without knowing the resource kind that is being processed it's not possible
+            // to accumulate the calculation.
+            //
+            // So presumably we need to request a binding query for a specific resource kind.
+            // We could do this by making the type of the binding query hold the type.
+
+            // We need to add instructions which will work out the binding for the base
+            OpaqueValueInfo baseInfo = findOrComputeOpaqueValueInfo(baseInst);
+
+            // If we couldn't find it we are done
+            if (baseInfo.registerIndex == nullptr || baseInfo.registerSpace == nullptr)
+            {
+                return baseInfo;
+            }
+
+
+            LayoutResourceKind kind = LayoutResourceKind::None;
+            Index stride = 1;
+
+            if (auto resourceType = as<IRResourceType>(elementType))
+            {
+                const auto shape = resourceType->getShape();
+
+                switch (shape)
+                {
+                case SLANG_TEXTURE_1D:
+                case SLANG_TEXTURE_2D:
+                case SLANG_TEXTURE_3D:
+                case SLANG_TEXTURE_CUBE:
+                case SLANG_STRUCTURED_BUFFER:
+                case SLANG_BYTE_ADDRESS_BUFFER:
+                case SLANG_TEXTURE_BUFFER:
+                    {
+                        const auto access = resourceType->getAccess();
+                        bool isReadOnly = (access == SLANG_RESOURCE_ACCESS_READ);
+
+                        kind = isReadOnly ? LayoutResourceKind::ShaderResource
+                                          : LayoutResourceKind::UnorderedAccess;
+                        break;
+                    }
+                default:
+                    break;
+                }
+            }
+            else if (as<IRSamplerStateTypeBase>(elementType))
+            {
+                kind = LayoutResourceKind::SamplerState;
+            }
+            else if (as<IRConstantBufferType>(elementType))
+            {
+                kind = LayoutResourceKind::ConstantBuffer;
+            }
+
+            if (kind == LayoutResourceKind::None)
+            {
+                // Can't determine the kind
+                return OpaqueValueInfo();
+            }
+
+            // If the element type has type layout we can try and use that
+            if (auto layoutDecoration = elementType->findDecoration<IRLayoutDecoration>())
+            {
+                // We have to calculate
+                if (auto elementTypeLayout = as<IRTypeLayout>(layoutDecoration->getLayout()))
+                {
+                    IRTypeSizeAttr* sizeAttr = elementTypeLayout->findSizeAttr(kind);
+                    sizeAttr = sizeAttr ? sizeAttr
+                                        : elementTypeLayout->findSizeAttr(
+                                              LayoutResourceKind::DescriptorTableSlot);
+
+                    if (!sizeAttr)
+                    {
+                        // Couldn't work it out
+                        return OpaqueValueInfo();
+                    }
+
+                    // TODO(JS): Perhaps we have to do something else if not finite?
+                    stride = sizeAttr->getFiniteSize();
+                }
+            }
+
+            SLANG_UNUSED(indexInst);
+
+            // Okay we need to create an instruction which is
+            // base + stride * index
+
+            IRBuilder builder(module);
+
+            builder.setInsertBefore(opaqueValue);
+
+            auto calcRegisterInst = builder.emitAdd(
+                indexType,
+                builder.emitMul(indexType, builder.getIntValue(indexType, stride), indexInst),
+                baseInfo.registerIndex);
+
+            OpaqueValueInfo finalInfo;
+            finalInfo.registerIndex = calcRegisterInst;
+            finalInfo.registerSpace = baseInfo.registerSpace;
+
+            return finalInfo;
+        }
+        else if (auto globalParam = as<IRGlobalParam>(opaqueValue))
         {
             // The simple/base case is when we have a global shader
             // parameter that has layout information attached.

source/slang/slang-ir-specialize-resources.cpp

@@ -1359,6 +1359,10 @@ bool isIllegalGLSLParameterType(IRType* type)
         return true;
     if (as<IRDynamicResourceType>(type))
         return true;
+    if (as<IRHLSLInputPatchType>(type))
+        return true;
+    if (as<IRHLSLOutputPatchType>(type))
+        return true;
     return false;
 }
 

source/slang/slang-ir.cpp

@@ -5259,6 +5259,11 @@ IRInst* IRBuilder::emitElementAddress(IRInst* basePtr, IRInst* index)
         SLANG_ASSERT(as<IRIntLit>(index));
         type = (IRType*)tupleType->getOperand(getIntVal(index));
     }
+    else if (auto hlslInputPatchType = as<IRHLSLInputPatchType>(valueType))
+    {
+        type = hlslInputPatchType->getElementType();
+    }
+
     SLANG_RELEASE_ASSERT(type);
     auto inst = createInst<IRGetElementPtr>(
         this,

tests/autodiff-dstdlib/dstdlib-clamp.slang

@@ -1,7 +1,7 @@
 //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 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0], stride=4):out,name=outputBuffer
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0], stride=4):out,name=outputBuffer
 RWStructuredBuffer<float> outputBuffer;
 
 typedef DifferentialPair<float> dpfloat;
@@ -178,4 +178,46 @@ void computeMain(uint3 dispatchThreadID: SV_DispatchThreadID)
         outputBuffer[28] = dpmax.d.y; // Expected: 0.0
         outputBuffer[29] = dpmax.d.z; // Expected: 0.3
     }
+
+    // New tests: Forward-mode tests for derivative propagation at the edges with clamp(x, 0, 1)
+    {
+        // Lower edge: x exactly = 0
+        dpfloat dpx = dpfloat(0.0, 0.4);
+        dpfloat dpmin = dpfloat(0.0, 0.8);
+        dpfloat dpmax = dpfloat(1.0, 0.5);
+        dpfloat res = fwd_diff(_clamp)(dpx, dpmin, dpmax);
+        outputBuffer[30] = res.d; // Expected: 0.4 (propagated from x)
+    }
+
+    {
+        // Upper edge: x exactly = 1
+        dpfloat dpx = dpfloat(1.0, 0.7);
+        dpfloat dpmin = dpfloat(0.0, 0.8);
+        dpfloat dpmax = dpfloat(1.0, 0.9);
+        dpfloat res = fwd_diff(_clamp)(dpx, dpmin, dpmax);
+        outputBuffer[31] = res.d; // Expected: 0.7 (propagated from x)
+    }
+
+    // Reverse-mode tests for derivative propagation at the edges with clamp(x, 0, 1)
+    {
+        // Lower edge: x exactly = 0
+        dpfloat dpx = dpfloat(0.0, 0.0);
+        dpfloat dpmin = dpfloat(0.0, 0.0);
+        dpfloat dpmax = dpfloat(1.0, 0.0);
+        bwd_diff(_clamp)(dpx, dpmin, dpmax, 1.0);
+        outputBuffer[32] = dpx.d;   // Expected: 1.0 (propagated from x)
+        outputBuffer[33] = dpmin.d; // Expected: 0.0
+        outputBuffer[34] = dpmax.d; // Expected: 0.0
+    }
+
+    {
+        // Upper edge: x exactly = 1
+        dpfloat dpx = dpfloat(1.0, 0.0);
+        dpfloat dpmin = dpfloat(0.0, 0.0);
+        dpfloat dpmax = dpfloat(1.0, 0.0);
+        bwd_diff(_clamp)(dpx, dpmin, dpmax, 1.0);
+        outputBuffer[35] = dpx.d;   // Expected: 1.0 (propagated from x)
+        outputBuffer[36] = dpmin.d; // Expected: 0.0
+        outputBuffer[37] = dpmax.d; // Expected: 0.0
+    }
 }

tests/autodiff-dstdlib/dstdlib-clamp.slang.expected.txt

@@ -29,3 +29,11 @@ type: float
 0.000000
 0.000000
 0.300000
+0.400000
+0.700000
+1.000000
+0.000000
+0.000000
+1.000000
+0.000000
+0.000000

tests/expected-failure-github.txt

@@ -1,9 +1,7 @@
-tests/language-feature/spirv-asm/imageoperands-warning.slang (vk)
 tests/language-feature/saturated-cooperation/simple.slang (vk)
 tests/language-feature/saturated-cooperation/fuse3.slang (vk)
 tests/language-feature/saturated-cooperation/fuse-product.slang (vk)
 tests/language-feature/saturated-cooperation/fuse.slang (vk)
-tests/bugs/byte-address-buffer-interlocked-add-f32.slang (vk)
 tests/render/render0.hlsl (mtl)
 tests/render/multiple-stage-io-locations.slang (mtl)
 tests/render/nointerpolation.hlsl (mtl)

tests/gpu-feature/texture/query/footprint/nv-shader-texture-footprint-array.slang

@@ -0,0 +1,76 @@
+//DISABLE_TEST:SIMPLE:-target dxil-assembly -entry fragmentMain -profile sm_6_5 -stage fragment -DNV_SHADER_EXTN_SLOT=u0 
+
+//TEST:SIMPLE(filecheck=SPIRV):-target spirv-assembly -entry fragmentMain -stage fragment
+//TEST:SIMPLE(filecheck=DXIL):-target dxil-assembly -entry fragmentMain -profile sm_6_5 -stage fragment -DNV_SHADER_EXTN_SLOT=u0 
+//TEST:SIMPLE(filecheck=HLSL):-target hlsl -entry fragmentMain -stage fragment
+
+//DISABLED_TEST:SIMPLE:-target spirv-assembly -entry fragmentMain -stage fragment
+//DISABLED_TEST:SIMPLE:-target dxil-assembly -entry fragmentMain -stage fragment
+//DISABLED_TEST:SIMPLE:-target hlsl -entry fragmentMain -stage fragment
+
+uniform Texture2D textures[] : register(t2, space10);
+uniform SamplerState sampler;
+uniform RWStructuredBuffer<uint> outputBuffer;
+
+static Texture2D _getBindlessTexture2d(uint texIdx)
+{
+    return textures[NonUniformResourceIndex(texIdx)];
+}
+
+void accumulate(inout uint r, uint u)
+{
+    r = r ^ u;
+}
+
+void accumulate(inout uint r, bool b)
+{
+    accumulate(r, uint(b));
+}
+
+void accumulate(inout uint r, uint2 u)
+{
+    accumulate(r, u.x);
+    accumulate(r, u.y);
+}
+
+void accumulate(inout uint r, uint3 u)
+{
+    accumulate(r, u.x);
+    accumulate(r, u.y);
+    accumulate(r, u.z);
+}
+
+void accumulate(inout uint r, TextureFootprint2D f)
+{
+    accumulate(r, f.anchor);
+    accumulate(r, f.offset);
+    accumulate(r, f.mask);
+    accumulate(r, f.lod);
+    accumulate(r, f.granularity);
+    accumulate(r, f.isSingleLevel);
+}
+
+cbuffer Uniforms
+{
+    uniform float2 coords;
+    uniform uint granularity;
+};
+
+void fragmentMain(
+    float v : VARYING)
+{
+    uint index = uint(v);
+    uint r = 0;
+
+    accumulate(r, _getBindlessTexture2d(index).queryFootprintCoarse(granularity, sampler, coords));
+
+// SPIRV: Extension  "SPV_NV_shader_image_footprint"
+// SPIRV: ImageSampleFootprintNV
+
+// DXIL: struct struct.NvShaderExtnStruct
+
+// HLSL: NvFootprintCoarse
+
+    outputBuffer[index] = r;
+}
+

tests/hlsl/float-literal-suffix.slang

@@ -0,0 +1,10 @@
+//TEST:SIMPLE(filecheck=HLSL):-target hlsl -profile ps_6_6 -entry fragmentMain
+//TEST:SIMPLE(filecheck=DXIL):-target dxil -profile ps_6_6 -entry fragmentMain
+
+float4 fragmentMain(float2 uv : TEXCOORD) : SV_Target
+{
+    //HLSL:, ddx({{.*}}1.5f)
+    //DXIL: = call float @dx.op.unary.f32(i32 {{.*}} ; DerivCoarseX(value)
+    float val = 1.5;
+    return float4(1.0, ddx(val), 0.0, 1.0);
+}