Implement Quad Control intrinsics (#5981)

docs/user-guide/a3-02-reference-capability-atoms.md

@@ -363,6 +363,12 @@ Extensions
 `SPV_EXT_demote_to_helper_invocation`
 > Represents the SPIR-V extension for demoting to helper invocation.
 
+`SPV_KHR_maximal_reconvergence`
+> Represents the SPIR-V extension for maximal reconvergence.
+
+`SPV_KHR_quad_control`
+> Represents the SPIR-V extension for quad group control.
+
 `SPV_KHR_fragment_shader_barycentric`
 > Represents the SPIR-V extension for fragment shader barycentric.
 
@@ -503,6 +509,12 @@ Extensions
 `spvDemoteToHelperInvocation`
 > Represents the SPIR-V capability for demoting to helper invocation.
 
+`spvMaximalReconvergenceKHR`
+> Represents the SPIR-V capability for maximal reconvergence.
+
+`spvQuadControlKHR`
+> Represents the SPIR-V capability for quad group control.
+
 `GL_EXT_buffer_reference`
 > Represents the GL_EXT_buffer_reference extension.
 
@@ -515,6 +527,12 @@ Extensions
 `GL_EXT_demote_to_helper_invocation`
 > Represents the GL_EXT_demote_to_helper_invocation extension.
 
+`GL_EXT_maximal_reconvergence`
+> Represents the GL_EXT_maximal_reconvergence extension.
+
+`GL_EXT_shader_quad_control`
+> Represents the GL_EXT_shader_quad_control extension.
+
 `GL_EXT_fragment_shader_barycentric`
 > Represents the GL_EXT_fragment_shader_barycentric extension.
 
@@ -1078,6 +1096,9 @@ Compound Capabilities
 `helper_lane`
 > Capabilities required to enable helper-lane demotion
 
+`quad_control`
+> Capabilities required to enable quad group control
+
 `breakpoint`
 > Capabilities required to enable shader breakpoints
 

source/slang/core.meta.slang

@@ -3244,6 +3244,12 @@ __Addr<T> __getLegalizedSPIRVGlobalParamAddr(T val);
 __intrinsic_op($(kIROp_RequireComputeDerivative))
 void __requireComputeDerivative();
 
+__intrinsic_op($(kIROp_RequireMaximallyReconverges))
+void __requireMaximallyReconverges();
+
+__intrinsic_op($(kIROp_RequireQuadDerivatives))
+void __requireQuadDerivatives();
+
 //@ public:
 /// @category misc_types 
 enum MemoryOrder
@@ -3978,6 +3984,21 @@ attribute_syntax [DerivativeGroupQuad] : DerivativeGroupQuadAttribute;
 __attributeTarget(FuncDecl)
 attribute_syntax [DerivativeGroupLinear] : DerivativeGroupLinearAttribute;
 
+/// Emits `MaximallyReconvergesKHR` execution mode when producing SPIR-V.
+/// This attribute has no effect on other targets.
+__attributeTarget(FuncDecl)
+attribute_syntax [MaximallyReconverges] : MaximallyReconvergesAttribute;
+
+/// Emits `QuadDerivativesKHR` execution mode when producing SPIR-V.
+/// This attribute has no effect on other targets.
+__attributeTarget(FuncDecl)
+attribute_syntax [QuadDerivatives] : QuadDerivativesAttribute;
+
+/// Emits `RequireFullQuadsKHR` execution mode when producing SPIR-V.
+/// This attribute has no effect on other targets.
+__attributeTarget(FuncDecl)
+attribute_syntax [RequireFullQuads] : RequireFullQuadsAttribute;
+
 __generic<T>
 typealias NodePayloadPtr = Ptr<T, $( (uint64_t)AddressSpace::NodePayloadAMDX)>;
 

source/slang/hlsl.meta.slang

@@ -15687,6 +15687,63 @@ bool IsHelperLane()
     }
 }
 
+//
+// Quad Control intrinsics
+//
+// For SPIRV and GLSL targets, the behavior is taken from Vulkan's `VK_KHR_shader_quad_control` spec.
+// QuadAny/QuadAll will map to OpGroupNonUniformQuadAny/All, and using either of these functions will 
+// result in the QuadDerivativesKHR execution mode being used. If MaximallyReconvergesKHR is not already 
+// specified by other means, it will be added when using either of QuadAny/QuadAll,
+//
+
+//@public:
+/// Returns true if `expr` is true in any lane of the current quad.
+__glsl_extension(GL_KHR_shader_subgroup_vote)
+__glsl_extension(GL_EXT_maximal_reconvergence)
+__glsl_extension(GL_EXT_shader_quad_control)
+[ForceInline]
+[require(glsl_hlsl_metal_spirv, quad_control)]
+bool QuadAny(bool expr)
+{
+    __requireMaximallyReconverges();
+    __requireQuadDerivatives();
+    __target_switch
+    {
+    case hlsl: __intrinsic_asm "QuadAny";
+    case glsl: __intrinsic_asm "subgroupQuadAny";
+    case metal: __intrinsic_asm "quad_any";
+    case spirv:
+        return spirv_asm
+        {
+            result:$$bool = OpGroupNonUniformQuadAnyKHR $expr;
+        };
+    }
+}
+
+//@public:
+/// Returns true if `expr` is true in all lanes of the current quad.
+__glsl_extension(GL_KHR_shader_subgroup_vote)
+__glsl_extension(GL_EXT_maximal_reconvergence)
+__glsl_extension(GL_EXT_shader_quad_control)
+[ForceInline]
+[require(glsl_hlsl_metal_spirv, quad_control)]
+bool QuadAll(bool expr)
+{
+    __requireMaximallyReconverges();
+    __requireQuadDerivatives();
+    __target_switch
+    {
+    case hlsl: __intrinsic_asm "QuadAll";
+    case glsl: __intrinsic_asm "subgroupQuadAll";
+    case metal: __intrinsic_asm "quad_all";
+    case spirv:
+        return spirv_asm
+        {
+            result:$$bool = OpGroupNonUniformQuadAllKHR $expr;
+        };
+    }
+}
+
 // `typedef`s to help with the fact that HLSL has been sorta-kinda case insensitive at various points
 //@hidden:
 typedef Texture2D texture2D;

source/slang/slang-ast-modifier.h

@@ -1664,6 +1664,21 @@ class DerivativeGroupLinearAttribute : public Attribute
     SLANG_AST_CLASS(DerivativeGroupLinearAttribute)
 };
 
+class MaximallyReconvergesAttribute : public Attribute
+{
+    SLANG_AST_CLASS(MaximallyReconvergesAttribute)
+};
+
+class QuadDerivativesAttribute : public Attribute
+{
+    SLANG_AST_CLASS(QuadDerivativesAttribute)
+};
+
+class RequireFullQuadsAttribute : public Attribute
+{
+    SLANG_AST_CLASS(RequireFullQuadsAttribute)
+};
+
 /// A `[payload]` attribute indicates that a `struct` type will be used as
 /// a ray payload for `TraceRay()` calls, and thus also as input/output
 /// for shaders in the ray tracing pipeline that might be invoked for

source/slang/slang-capabilities.capdef

@@ -466,6 +466,14 @@ def SPV_EXT_mesh_shader : _spirv_1_4;
 /// [EXT]
 def SPV_EXT_demote_to_helper_invocation : _spirv_1_4;
 
+/// Represents the SPIR-V extension for maximal reconvergence.
+/// [EXT]
+def SPV_KHR_maximal_reconvergence : _spirv_1_0;
+
+/// Represents the SPIR-V extension for quad group control.
+/// [EXT]
+def SPV_KHR_quad_control : _spirv_1_3;
+
 /// Represents the SPIR-V extension for fragment shader barycentric.
 /// [EXT]
 def SPV_KHR_fragment_shader_barycentric : _spirv_1_0;
@@ -654,6 +662,14 @@ def spvDemoteToHelperInvocationEXT : SPV_EXT_demote_to_helper_invocation;
 /// [EXT]
 def spvDemoteToHelperInvocation : spvDemoteToHelperInvocationEXT;
 
+/// Represents the SPIR-V capability for maximal reconvergence.
+/// [EXT]
+def spvMaximalReconvergenceKHR : SPV_KHR_maximal_reconvergence;
+
+/// Represents the SPIR-V capability for quad group control.
+/// [EXT]
+def spvQuadControlKHR : SPV_KHR_quad_control;
+
 // The following capabilities all pertain to how ray tracing shaders are translated
 // to GLSL, where there are two different extensions that can provide the core
 // functionality of `TraceRay` and the related operations.
@@ -691,6 +707,8 @@ def _GL_EXT_shader_image_load_store : _GLSL_130;
 def _GL_EXT_shader_realtime_clock : glsl;
 def _GL_EXT_texture_query_lod : glsl;
 def _GL_EXT_texture_shadow_lod : _GLSL_130;
+def _GL_EXT_maximal_reconvergence : _GLSL_140;
+def _GL_EXT_shader_quad_control : _GLSL_140;
 
 def _GL_ARB_derivative_control : _GLSL_400;
 def _GL_ARB_fragment_shader_interlock : _GLSL_450;
@@ -746,6 +764,14 @@ alias GL_EXT_debug_printf = _GL_EXT_debug_printf | SPV_KHR_non_semantic_info;
 /// [EXT]
 alias GL_EXT_demote_to_helper_invocation = _GL_EXT_demote_to_helper_invocation | spvDemoteToHelperInvocationEXT;
 
+/// Represents the GL_EXT_maximal_reconvergence extension.
+/// [EXT]
+alias GL_EXT_maximal_reconvergence = _GL_EXT_maximal_reconvergence | spvMaximalReconvergenceKHR;
+
+/// Represents the GL_EXT_shader_quad_control extension.
+/// [EXT]
+alias GL_EXT_shader_quad_control = _GL_EXT_shader_quad_control | spvQuadControlKHR;
+
 /// Represents the GL_EXT_fragment_shader_barycentric extension.
 /// [EXT]
 alias GL_EXT_fragment_shader_barycentric = _GL_EXT_fragment_shader_barycentric | spvFragmentBarycentricKHR;
@@ -1925,6 +1951,13 @@ alias helper_lane = _sm_6_0 + fragment
                    | metal + fragment
                    ;
 
+/// Capabilities required to enable quad group control
+/// [Compound]
+alias quad_control = _sm_6_7
+                    | GL_EXT_shader_quad_control + GL_EXT_maximal_reconvergence + GL_KHR_shader_subgroup_vote
+                    | metal
+                    ;
+
 /// Capabilities required to enable shader breakpoints
 /// [Compound]
 alias breakpoint = GL_EXT_debug_printf | hlsl | _cuda_sm_8_0 | cpp;

source/slang/slang-emit-c-like.cpp

@@ -129,14 +129,6 @@ void CLikeSourceEmitter::emitPreModuleImpl()
         m_writer->emit("\n");
     }
 }
-void CLikeSourceEmitter::emitPostModuleImpl()
-{
-    if (m_requiredAfter.requireComputeDerivatives.getLength() > 0)
-    {
-        m_writer->emit(m_requiredAfter.requireComputeDerivatives);
-        m_writer->emit("\n");
-    }
-}
 
 //
 // Types

source/slang/slang-emit-c-like.h

@@ -470,7 +470,6 @@ class CLikeSourceEmitter : public SourceEmitterBase
     void emitFrontMatter(TargetRequest* targetReq) { emitFrontMatterImpl(targetReq); }
 
     void emitPreModule() { emitPreModuleImpl(); }
-    void emitPostModule() { emitPostModuleImpl(); }
     void emitModule(IRModule* module, DiagnosticSink* sink)
     {
         m_irModule = module;
@@ -555,7 +554,6 @@ class CLikeSourceEmitter : public SourceEmitterBase
     /// For example on targets that don't have built in vector/matrix support, this is where
     /// the appropriate generated declarations occur.
     virtual void emitPreModuleImpl();
-    virtual void emitPostModuleImpl();
 
     virtual void emitSimpleTypeAndDeclaratorImpl(IRType* type, DeclaratorInfo* declarator);
     void emitSimpleTypeAndDeclarator(IRType* type, DeclaratorInfo* declarator)
@@ -736,10 +734,6 @@ class CLikeSourceEmitter : public SourceEmitterBase
     Dictionary<IRInst*, String> m_mapInstToName;
 
     OrderedHashSet<IRStringLit*> m_requiredPreludes;
-    struct RequiredAfter
-    {
-        String requireComputeDerivatives;
-    } m_requiredAfter;
 
     Dictionary<const char*, IRStringLit*> m_builtinPreludes;
 };