[clang] [llvm] [HLSL][SPIRV] Implement `WaveActiveMax` intrinsic (PR #112991)

Finn Plummer via cfe-commits cfe-commits at lists.llvm.org
Fri Oct 18 15:36:26 PDT 2024


https://github.com/inbelic updated https://github.com/llvm/llvm-project/pull/112991

>From b18e40978cae2c0d9ba9aeb61cf7294809b75012 Mon Sep 17 00:00:00 2001
From: Finn Plummer <canadienfinn at gmail.com>
Date: Fri, 18 Oct 2024 14:41:57 -0700
Subject: [PATCH 1/2] [HLSL][SPIRV] Implement `WaveActiveMax` intrinsic

  - add clang builtin to Builtins.td
  - link builtin in hlsl_intrinsics
  - add codegen for spirv intrinsic and two directx intrinsics to retain
    signedness information of the operands in CGBuiltin.cpp
  - add semantic analysis in SemaHLSL.cpp
  - add lowering of spirv intrinsic to spirv backend in
    SPIRVInstructionSelector.cpp

  - add test cases to illustrate passes

Note that this defines the dx intrinsics but does not implement
the DirectX lowering to DXIL. This will be implemented in a second pr
when the dependent pr merges.
---
 clang/include/clang/Basic/Builtins.td         |  6 ++
 .../clang/Basic/DiagnosticSemaKinds.td        |  3 +
 clang/lib/CodeGen/CGBuiltin.cpp               | 36 +++++++
 clang/lib/Headers/hlsl/hlsl_intrinsics.h      | 99 +++++++++++++++++++
 clang/lib/Sema/SemaHLSL.cpp                   | 31 ++++++
 .../CodeGenHLSL/builtins/WaveActiveMax.hlsl   | 48 +++++++++
 .../BuiltIns/WaveActiveMax-errors.hlsl        | 28 ++++++
 llvm/include/llvm/IR/IntrinsicsDirectX.td     |  2 +
 llvm/include/llvm/IR/IntrinsicsSPIRV.td       |  2 +
 .../Target/SPIRV/SPIRVInstructionSelector.cpp | 34 +++++++
 .../SPIRV/hlsl-intrinsics/WaveActiveMax.ll    | 51 ++++++++++
 11 files changed, 340 insertions(+)
 create mode 100644 clang/test/CodeGenHLSL/builtins/WaveActiveMax.hlsl
 create mode 100644 clang/test/SemaHLSL/BuiltIns/WaveActiveMax-errors.hlsl
 create mode 100644 llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WaveActiveMax.ll

diff --git a/clang/include/clang/Basic/Builtins.td b/clang/include/clang/Basic/Builtins.td
index 382fb6b7a3c031..ae2068c7a153f7 100644
--- a/clang/include/clang/Basic/Builtins.td
+++ b/clang/include/clang/Basic/Builtins.td
@@ -4749,6 +4749,12 @@ def HLSLWaveActiveCountBits : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "unsigned int(bool)";
 }
 
+def HLSLWaveActiveMax : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_wave_active_max"];
+  let Attributes = [NoThrow, Const];
+  let Prototype = "void(...)";
+}
+
 def HLSLWaveGetLaneIndex : LangBuiltin<"HLSL_LANG"> {
   let Spellings = ["__builtin_hlsl_wave_get_lane_index"];
   let Attributes = [NoThrow, Const];
diff --git a/clang/include/clang/Basic/DiagnosticSemaKinds.td b/clang/include/clang/Basic/DiagnosticSemaKinds.td
index c458a62d9be48c..95093723c2f495 100644
--- a/clang/include/clang/Basic/DiagnosticSemaKinds.td
+++ b/clang/include/clang/Basic/DiagnosticSemaKinds.td
@@ -9232,6 +9232,9 @@ def err_typecheck_expect_scalar_or_vector : Error<
   "a vector of such type is required">;
 def err_typecheck_expect_any_scalar_or_vector : Error<
   "invalid operand of type %0 where a scalar or vector is required">;
+def err_typecheck_expect_scalar_or_vector_not_type : Error<
+  "invalid operand of type %0 where %1 or "
+  "a vector of such type is not allowed">;
 def err_typecheck_expect_flt_or_vector : Error<
   "invalid operand of type %0 where floating, complex or "
   "a vector of such types is required">;
diff --git a/clang/lib/CodeGen/CGBuiltin.cpp b/clang/lib/CodeGen/CGBuiltin.cpp
index 12f99d9f1178a9..dd17b31f3d7168 100644
--- a/clang/lib/CodeGen/CGBuiltin.cpp
+++ b/clang/lib/CodeGen/CGBuiltin.cpp
@@ -18638,6 +18638,25 @@ static Intrinsic::ID getDotProductIntrinsic(CGHLSLRuntime &RT, QualType QT) {
   return RT.getUDotIntrinsic();
 }
 
+// Return wave active max that corresponds to the QT scalar type
+static Intrinsic::ID getWaveActiveMaxIntrinsic(llvm::Triple::ArchType Arch,
+                                               CGHLSLRuntime &RT, QualType QT) {
+  switch (Arch) {
+  case llvm::Triple::spirv:
+    if (QT->isUnsignedIntegerType())
+      return llvm::Intrinsic::spv_wave_active_umax;
+    return llvm::Intrinsic::spv_wave_active_max;
+  case llvm::Triple::dxil: {
+    if (QT->isUnsignedIntegerType())
+      return llvm::Intrinsic::dx_wave_active_umax;
+    return llvm::Intrinsic::dx_wave_active_max;
+  }
+  default:
+    llvm_unreachable("Intrinsic WaveActiveMax"
+                     " not supported by target architecture");
+  }
+}
+
 Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
                                             const CallExpr *E,
                                             ReturnValueSlot ReturnValue) {
@@ -18883,6 +18902,23 @@ case Builtin::BI__builtin_hlsl_elementwise_isinf: {
         /*ReturnType=*/Op0->getType(), CGM.getHLSLRuntime().getStepIntrinsic(),
         ArrayRef<Value *>{Op0, Op1}, nullptr, "hlsl.step");
   }
+  case Builtin::BI__builtin_hlsl_wave_active_max: {
+    // Due to the use of variadic arguments, explicitly retreive argument
+    Value *OpExpr = EmitScalarExpr(E->getArg(0));
+    llvm::FunctionType *FT = llvm::FunctionType::get(
+        OpExpr->getType(), ArrayRef{OpExpr->getType()}, false);
+    Intrinsic::ID IID = getWaveActiveMaxIntrinsic(
+        getTarget().getTriple().getArch(), CGM.getHLSLRuntime(),
+        E->getArg(0)->getType());
+
+    // Get overloaded name
+    std::string Name =
+        Intrinsic::getName(IID, ArrayRef{OpExpr->getType()}, &CGM.getModule());
+    return EmitRuntimeCall(CGM.CreateRuntimeFunction(FT, Name, {},
+                                                     /*Local=*/false,
+                                                     /*AssumeConvergent=*/true),
+                           ArrayRef{OpExpr}, "hlsl.wave.active.max");
+  }
   case Builtin::BI__builtin_hlsl_wave_get_lane_index: {
     // We don't define a SPIR-V intrinsic, instead it is a SPIR-V built-in
     // defined in SPIRVBuiltins.td. So instead we manually get the matching name
diff --git a/clang/lib/Headers/hlsl/hlsl_intrinsics.h b/clang/lib/Headers/hlsl/hlsl_intrinsics.h
index 30dce60b3ff702..c0723133455a02 100644
--- a/clang/lib/Headers/hlsl/hlsl_intrinsics.h
+++ b/clang/lib/Headers/hlsl/hlsl_intrinsics.h
@@ -2177,6 +2177,105 @@ __attribute__((convergent)) double3 WaveReadLaneAt(double3, int32_t);
 _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_read_lane_at)
 __attribute__((convergent)) double4 WaveReadLaneAt(double4, int32_t);
 
+//===----------------------------------------------------------------------===//
+// WaveActiveMax builtins
+//===----------------------------------------------------------------------===//
+
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) half WaveActiveMax(half);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) half2 WaveActiveMax(half2);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) half3 WaveActiveMax(half3);
+_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) half4 WaveActiveMax(half4);
+
+#ifdef __HLSL_ENABLE_16_BIT
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int16_t WaveActiveMax(int16_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int16_t2 WaveActiveMax(int16_t2);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int16_t3 WaveActiveMax(int16_t3);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int16_t4 WaveActiveMax(int16_t4);
+
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint16_t WaveActiveMax(uint16_t);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint16_t2 WaveActiveMax(uint16_t2);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint16_t3 WaveActiveMax(uint16_t3);
+_HLSL_AVAILABILITY(shadermodel, 6.0)
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint16_t4 WaveActiveMax(uint16_t4);
+#endif
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int WaveActiveMax(int);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int2 WaveActiveMax(int2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int3 WaveActiveMax(int3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int4 WaveActiveMax(int4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint WaveActiveMax(uint);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint2 WaveActiveMax(uint2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint3 WaveActiveMax(uint3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint4 WaveActiveMax(uint4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int64_t WaveActiveMax(int64_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int64_t2 WaveActiveMax(int64_t2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int64_t3 WaveActiveMax(int64_t3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) int64_t4 WaveActiveMax(int64_t4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint64_t WaveActiveMax(uint64_t);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint64_t2 WaveActiveMax(uint64_t2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint64_t3 WaveActiveMax(uint64_t3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) uint64_t4 WaveActiveMax(uint64_t4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) float WaveActiveMax(float);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) float2 WaveActiveMax(float2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) float3 WaveActiveMax(float3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) float4 WaveActiveMax(float4);
+
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) double WaveActiveMax(double);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) double2 WaveActiveMax(double2);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) double3 WaveActiveMax(double3);
+_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_max)
+__attribute((convergent)) double4 WaveActiveMax(double4);
+
 //===----------------------------------------------------------------------===//
 // sign builtins
 //===----------------------------------------------------------------------===//
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index 698fdbed0484e5..0a214736dd3e9b 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -1767,6 +1767,23 @@ static bool CheckAnyScalarOrVector(Sema *S, CallExpr *TheCall,
   return false;
 }
 
+static bool CheckNotScalarType(Sema *S, CallExpr *TheCall, QualType Scalar,
+                               unsigned ArgIndex) {
+  assert(TheCall->getNumArgs() >= ArgIndex);
+  QualType ArgType = TheCall->getArg(ArgIndex)->getType();
+  auto *VTy = ArgType->getAs<VectorType>();
+  // is the scalar or vector<scalar>
+  if (S->Context.hasSameUnqualifiedType(ArgType, Scalar) ||
+      (VTy &&
+       S->Context.hasSameUnqualifiedType(VTy->getElementType(), Scalar))) {
+    S->Diag(TheCall->getArg(0)->getBeginLoc(),
+            diag::err_typecheck_expect_scalar_or_vector_not_type)
+        << ArgType << Scalar;
+    return true;
+  }
+  return false;
+}
+
 static bool CheckBoolSelect(Sema *S, CallExpr *TheCall) {
   assert(TheCall->getNumArgs() == 3);
   Expr *Arg1 = TheCall->getArg(1);
@@ -2002,6 +2019,20 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
     TheCall->setType(ArgTyA);
     break;
   }
+  case Builtin::BI__builtin_hlsl_wave_active_max: {
+    if (SemaRef.checkArgCount(TheCall, 1))
+      return true;
+
+    // Ensure input expr type is a scalar/vector and the same as the return type
+    if (CheckAnyScalarOrVector(&SemaRef, TheCall, 0))
+      return true;
+    if (CheckNotScalarType(&SemaRef, TheCall, getASTContext().BoolTy, 0))
+      return true;
+    ExprResult Expr = TheCall->getArg(0);
+    QualType ArgTyExpr = Expr.get()->getType();
+    TheCall->setType(ArgTyExpr);
+    break;
+  }
   // Note these are llvm builtins that we want to catch invalid intrinsic
   // generation. Normal handling of these builitns will occur elsewhere.
   case Builtin::BI__builtin_elementwise_bitreverse: {
diff --git a/clang/test/CodeGenHLSL/builtins/WaveActiveMax.hlsl b/clang/test/CodeGenHLSL/builtins/WaveActiveMax.hlsl
new file mode 100644
index 00000000000000..4141bb6f1b9f24
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/WaveActiveMax.hlsl
@@ -0,0 +1,48 @@
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -triple \
+// RUN:   dxil-pc-shadermodel6.3-compute %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-DXIL
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -triple \
+// RUN:   spirv-pc-vulkan-compute %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,CHECK-SPIRV
+
+// Test basic lowering to runtime function call.
+
+// CHECK-LABEL: test_int
+int test_int(int expr) {
+  // CHECK-SPIRV: %[[#entry_tok:]] = call token @llvm.experimental.convergence.entry()
+  // CHECK-SPIRV:  %[[RET:.*]] = call [[TY:.*]] @llvm.spv.wave.active.max.i32([[TY]] %[[#]]) [ "convergencectrl"(token %[[#entry_tok]]) ]
+  // CHECK-DXIL:  %[[RET:.*]] = call [[TY:.*]] @llvm.dx.wave.active.max.i32([[TY]] %[[#]])
+  // CHECK:  ret [[TY]] %[[RET]]
+  return WaveActiveMax(expr);
+}
+
+// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.active.max.i32([[TY]]) #[[#attr:]]
+// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.active.max.i32([[TY]]) #[[#attr:]]
+
+// CHECK-LABEL: test_uint64_t
+uint64_t test_uint64_t(uint64_t expr) {
+  // CHECK-SPIRV: %[[#entry_tok1:]] = call token @llvm.experimental.convergence.entry()
+  // CHECK-SPIRV:  %[[RET:.*]] = call [[TY1:.*]] @llvm.spv.wave.active.umax.i64([[TY1]] %[[#]]) [ "convergencectrl"(token %[[#entry_tok1]]) ]
+  // CHECK-DXIL:  %[[RET:.*]] = call [[TY1:.*]] @llvm.dx.wave.active.umax.i64([[TY1]] %[[#]])
+  // CHECK:  ret [[TY1]] %[[RET]]
+  return WaveActiveMax(expr);
+}
+
+// CHECK-DXIL: declare [[TY1]] @llvm.dx.wave.active.umax.i64([[TY1]]) #[[#attr:]]
+// CHECK-SPIRV: declare [[TY1]] @llvm.spv.wave.active.umax.i64([[TY1]]) #[[#attr:]]
+
+// Test basic lowering to runtime function call with array and float value.
+
+// CHECK-LABEL: test_floatv4
+float4 test_floatv4(float4 expr) {
+  // CHECK-SPIRV: %[[#entry_tok2:]] = call token @llvm.experimental.convergence.entry()
+  // CHECK-SPIRV:  %[[RET1:.*]] = call [[TY2:.*]] @llvm.spv.wave.active.max.v4f32([[TY2]] %[[#]]) [ "convergencectrl"(token %[[#entry_tok2]]) ]
+  // CHECK-DXIL:  %[[RET1:.*]] = call [[TY2:.*]] @llvm.dx.wave.active.max.v4f32([[TY2]] %[[#]])
+  // CHECK:  ret [[TY2]] %[[RET1]]
+  return WaveActiveMax(expr);
+}
+
+// CHECK-DXIL: declare [[TY2]] @llvm.dx.wave.active.max.v4f32([[TY2]]) #[[#attr]]
+// CHECK-SPIRV: declare [[TY2]] @llvm.spv.wave.active.max.v4f32([[TY2]]) #[[#attr]]
+
+// CHECK: attributes #[[#attr]] = {{{.*}} convergent {{.*}}}
diff --git a/clang/test/SemaHLSL/BuiltIns/WaveActiveMax-errors.hlsl b/clang/test/SemaHLSL/BuiltIns/WaveActiveMax-errors.hlsl
new file mode 100644
index 00000000000000..e2670b46794dda
--- /dev/null
+++ b/clang/test/SemaHLSL/BuiltIns/WaveActiveMax-errors.hlsl
@@ -0,0 +1,28 @@
+// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.6-library %s -emit-llvm-only -disable-llvm-passes -verify
+
+int test_too_few_arg() {
+  return __builtin_hlsl_wave_active_max();
+  // expected-error at -1 {{too few arguments to function call, expected 1, have 0}}
+}
+
+float2 test_too_many_arg(float2 p0) {
+  return __builtin_hlsl_wave_active_max(p0, p0);
+  // expected-error at -1 {{too many arguments to function call, expected 1, have 2}}
+}
+
+bool test_expr_bool_type_check(bool p0) {
+  return __builtin_hlsl_wave_active_max(p0);
+  // expected-error at -1 {{invalid operand of type 'bool' where 'bool' or a vector of such type is not allowed}}
+}
+
+bool2 test_expr_bool_vec_type_check(bool2 p0) {
+  return __builtin_hlsl_wave_active_max(p0);
+  // expected-error at -1 {{invalid operand of type 'bool2' (aka 'vector<bool, 2>') where 'bool' or a vector of such type is not allowed}}
+}
+
+struct S { float f; };
+
+S test_expr_struct_type_check(S p0) {
+  return __builtin_hlsl_wave_active_max(p0);
+  // expected-error at -1 {{invalid operand of type 'S' where a scalar or vector is required}}
+}
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index 27a437a83be6dd..01ced0317f7f4b 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -85,6 +85,8 @@ def int_dx_umad : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLV
 def int_dx_normalize : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty], [IntrNoMem]>;
 def int_dx_rsqrt  : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_wave_getlaneindex : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrConvergent, IntrNoMem]>;
+def int_dx_wave_active_max : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>;
+def int_dx_wave_active_umax : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>;
 def int_dx_wave_is_first_lane : DefaultAttrsIntrinsic<[llvm_i1_ty], [], [IntrConvergent]>;
 def int_dx_wave_readlane : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrConvergent, IntrNoMem]>;
 def int_dx_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty], [IntrNoMem]>;
diff --git a/llvm/include/llvm/IR/IntrinsicsSPIRV.td b/llvm/include/llvm/IR/IntrinsicsSPIRV.td
index 6df2eb156a0774..6ebc3da11d6f8e 100644
--- a/llvm/include/llvm/IR/IntrinsicsSPIRV.td
+++ b/llvm/include/llvm/IR/IntrinsicsSPIRV.td
@@ -83,6 +83,8 @@ let TargetPrefix = "spv" in {
     DefaultAttrsIntrinsic<[LLVMVectorElementType<0>],
     [llvm_anyint_ty, LLVMScalarOrSameVectorWidth<0, LLVMVectorElementType<0>>],
     [IntrNoMem, Commutative] >;
+  def int_spv_wave_active_max : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>;
+  def int_spv_wave_active_umax : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>;
   def int_spv_wave_is_first_lane : DefaultAttrsIntrinsic<[llvm_i1_ty], [], [IntrConvergent]>;
   def int_spv_wave_readlane : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrConvergent, IntrNoMem]>;
   def int_spv_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty], [IntrNoMem]>;
diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
index d9377fe4b91a1a..6b9d5cab6db879 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
@@ -164,6 +164,9 @@ class SPIRVInstructionSelector : public InstructionSelector {
   bool selectIntegerDot(Register ResVReg, const SPIRVType *ResType,
                         MachineInstr &I) const;
 
+  bool selectWaveActiveMax(Register ResVReg, const SPIRVType *ResType,
+                           MachineInstr &I, bool IsSigned) const;
+
   void renderImm32(MachineInstrBuilder &MIB, const MachineInstr &I,
                    int OpIdx) const;
   void renderFImm64(MachineInstrBuilder &MIB, const MachineInstr &I,
@@ -1782,6 +1785,33 @@ bool SPIRVInstructionSelector::selectWaveReadLaneAt(Register ResVReg,
       .addUse(I.getOperand(3).getReg());
 }
 
+bool SPIRVInstructionSelector::selectWaveActiveMax(Register ResVReg,
+                                                   const SPIRVType *ResType,
+                                                   MachineInstr &I,
+                                                   bool IsSigned) const {
+  assert(I.getNumOperands() == 3);
+  assert(I.getOperand(2).isReg());
+  MachineBasicBlock &BB = *I.getParent();
+  Register InputRegister = I.getOperand(2).getReg();
+  SPIRVType *InputType = GR.getSPIRVTypeForVReg(InputRegister);
+
+  if (!InputType)
+    report_fatal_error("Input Type could not be determined.");
+
+  SPIRVType *IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
+  // Retreive the operation to use based on input type
+  bool IsFloatTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
+  auto Opcode =
+      IsFloatTy ? SPIRV::OpGroupNonUniformFMax : (IsSigned ?
+               SPIRV::OpGroupNonUniformSMax : SPIRV::OpGroupNonUniformUMax);
+  return BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
+      .addDef(ResVReg)
+      .addUse(GR.getSPIRVTypeID(ResType))
+      .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII))
+      .addImm(SPIRV::GroupOperation::Reduce)
+      .addUse(I.getOperand(2).getReg());
+}
+
 bool SPIRVInstructionSelector::selectBitreverse(Register ResVReg,
                                                 const SPIRVType *ResType,
                                                 MachineInstr &I) const {
@@ -2559,6 +2589,10 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg,
   } break;
   case Intrinsic::spv_saturate:
     return selectSaturate(ResVReg, ResType, I);
+  case Intrinsic::spv_wave_active_umax:
+    return selectWaveActiveMax(ResVReg, ResType, I, false);
+  case Intrinsic::spv_wave_active_max:
+    return selectWaveActiveMax(ResVReg, ResType, I, true);
   case Intrinsic::spv_wave_is_first_lane: {
     SPIRVType *IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
     return BuildMI(BB, I, I.getDebugLoc(),
diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WaveActiveMax.ll b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WaveActiveMax.ll
new file mode 100644
index 00000000000000..d737533c2ba04d
--- /dev/null
+++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WaveActiveMax.ll
@@ -0,0 +1,51 @@
+; RUN: llc -verify-machineinstrs -O0 -mtriple=spirv32v1.3-vulkan-unknown %s -o - | FileCheck %s
+; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv32v1.3-vulkan-unknown %s -o - -filetype=obj | spirv-val %}
+
+; Test lowering to spir-v backend for various types and scalar/vector
+
+; CHECK-DAG:   %[[#f16:]] = OpTypeFloat 16
+; CHECK-DAG:   %[[#f32:]] = OpTypeFloat 32
+; CHECK-DAG:   %[[#uint:]] = OpTypeInt 32 0
+; CHECK-DAG:   %[[#v4_half:]] = OpTypeVector %[[#f16]] 4
+; CHECK-DAG:   %[[#scope:]] = OpConstant %[[#uint]] 3
+
+; CHECK-LABEL: Begin function test_float
+; CHECK:   %[[#fexpr:]] = OpFunctionParameter %[[#f32]]
+define float @test_float(float %fexpr) {
+entry:
+; CHECK:   %[[#fret:]] = OpGroupNonUniformFMax %[[#f32]] %[[#scope]] Reduce %[[#fexpr]]
+  %0 = call float @llvm.spv.wave.active.max.f32(float %fexpr)
+  ret float %0
+}
+
+; CHECK-LABEL: Begin function test_int
+; CHECK:   %[[#iexpr:]] = OpFunctionParameter %[[#uint]]
+define i32 @test_int(i32 %iexpr) {
+entry:
+; CHECK:   %[[#iret:]] = OpGroupNonUniformSMax %[[#uint]] %[[#scope]] Reduce %[[#iexpr]]
+  %0 = call i32 @llvm.spv.wave.active.max.i32(i32 %iexpr)
+  ret i32 %0
+}
+
+; CHECK-LABEL: Begin function test_uint
+; CHECK:   %[[#iexpr:]] = OpFunctionParameter %[[#uint]]
+define i32 @test_uint(i32 %iexpr) {
+entry:
+; CHECK:   %[[#iret:]] = OpGroupNonUniformUMax %[[#uint]] %[[#scope]] Reduce %[[#iexpr]]
+  %0 = call i32 @llvm.spv.wave.active.umax.i32(i32 %iexpr)
+  ret i32 %0
+}
+
+; CHECK-LABEL: Begin function test_vhalf
+; CHECK:   %[[#vbexpr:]] = OpFunctionParameter %[[#v4_half]]
+define <4 x half> @test_vhalf(<4 x half> %vbexpr) {
+entry:
+; CHECK:   %[[#vhalfret:]] = OpGroupNonUniformFMax %[[#v4_half]] %[[#scope]] Reduce %[[#vbexpr]]
+  %0 = call <4 x half> @llvm.spv.wave.active.max.v4half(<4 x half> %vbexpr)
+  ret <4 x half> %0
+}
+
+declare float @llvm.spv.wave.active.max.f32(float)
+declare i32 @llvm.spv.wave.active.max.i32(i32)
+declare i32 @llvm.spv.wave.active.umax.i32(i32)
+declare <4 x half> @llvm.spv.wave.active.max.v4half(<4 x half>)

>From 6394b9f5076aa6c4a96a0e963b7ee3393a9ee6fe Mon Sep 17 00:00:00 2001
From: Finn Plummer <canadienfinn at gmail.com>
Date: Fri, 18 Oct 2024 15:36:02 -0700
Subject: [PATCH 2/2] clang

---
 llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
index 6b9d5cab6db879..b7a8fc91f96846 100644
--- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
+++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp
@@ -1801,9 +1801,9 @@ bool SPIRVInstructionSelector::selectWaveActiveMax(Register ResVReg,
   SPIRVType *IntTy = GR.getOrCreateSPIRVIntegerType(32, I, TII);
   // Retreive the operation to use based on input type
   bool IsFloatTy = GR.isScalarOrVectorOfType(InputRegister, SPIRV::OpTypeFloat);
-  auto Opcode =
-      IsFloatTy ? SPIRV::OpGroupNonUniformFMax : (IsSigned ?
-               SPIRV::OpGroupNonUniformSMax : SPIRV::OpGroupNonUniformUMax);
+  auto Opcode = IsFloatTy ? SPIRV::OpGroupNonUniformFMax
+                          : (IsSigned ? SPIRV::OpGroupNonUniformSMax
+                                      : SPIRV::OpGroupNonUniformUMax);
   return BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode))
       .addDef(ResVReg)
       .addUse(GR.getSPIRVTypeID(ResType))



More information about the cfe-commits mailing list