https://github.com/kcloudy0717 updated https://github.com/llvm/llvm-project/pull/167946
>From 1c2948076a74a16ff11c3796420556127418b0b3 Mon Sep 17 00:00:00 2001 From: Kai Huang <[email protected]> Date: Fri, 14 Nov 2025 02:14:23 +0800 Subject: [PATCH] [HLSL][DXIL][SPRIV] WavePrefixSum intrinsic support This commit adds WavePrefixSum intrinsic support to HLSL. --- clang/include/clang/Basic/Builtins.td | 6 + clang/lib/CodeGen/CGHLSLBuiltins.cpp | 26 ++++ .../lib/Headers/hlsl/hlsl_alias_intrinsics.h | 99 ++++++++++++ clang/lib/Sema/SemaHLSL.cpp | 35 ++++- .../CodeGenHLSL/builtins/WavePrefixSum.hlsl | 46 ++++++ .../BuiltIns/WavePrefixSum-errors.hlsl | 28 ++++ llvm/include/llvm/IR/IntrinsicsDirectX.td | 2 + llvm/include/llvm/IR/IntrinsicsSPIRV.td | 1 + llvm/lib/Target/DirectX/DXIL.td | 24 +++ llvm/lib/Target/DirectX/DXILShaderFlags.cpp | 3 + .../DirectX/DirectXTargetTransformInfo.cpp | 2 + .../Target/SPIRV/SPIRVInstructionSelector.cpp | 47 ++++++ .../CodeGen/DirectX/ShaderFlags/wave-ops.ll | 14 ++ llvm/test/CodeGen/DirectX/WavePrefixSum.ll | 143 ++++++++++++++++++ .../SPIRV/hlsl-intrinsics/WavePrefixSum.ll | 41 +++++ 15 files changed, 514 insertions(+), 3 deletions(-) create mode 100644 clang/test/CodeGenHLSL/builtins/WavePrefixSum.hlsl create mode 100644 clang/test/SemaHLSL/BuiltIns/WavePrefixSum-errors.hlsl create mode 100644 llvm/test/CodeGen/DirectX/WavePrefixSum.ll create mode 100644 llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WavePrefixSum.ll diff --git a/clang/include/clang/Basic/Builtins.td b/clang/include/clang/Basic/Builtins.td index a410a138836eb..afa61f19017ef 100644 --- a/clang/include/clang/Basic/Builtins.td +++ b/clang/include/clang/Basic/Builtins.td @@ -5151,6 +5151,12 @@ def HLSLWaveGetLaneCount : LangBuiltin<"HLSL_LANG"> { let Prototype = "unsigned int()"; } +def HLSLWavePrefixSum : LangBuiltin<"HLSL_LANG"> { + let Spellings = ["__builtin_hlsl_wave_prefix_sum"]; + let Attributes = [NoThrow, Const]; + let Prototype = "void(...)"; +} + def HLSLClamp : LangBuiltin<"HLSL_LANG"> { let Spellings = ["__builtin_hlsl_elementwise_clamp"]; let Attributes = [NoThrow, Const, CustomTypeChecking]; diff --git a/clang/lib/CodeGen/CGHLSLBuiltins.cpp b/clang/lib/CodeGen/CGHLSLBuiltins.cpp index b3c0c90faeb02..e8bf6198a9382 100644 --- a/clang/lib/CodeGen/CGHLSLBuiltins.cpp +++ b/clang/lib/CodeGen/CGHLSLBuiltins.cpp @@ -397,6 +397,23 @@ static Intrinsic::ID getPrefixCountBitsIntrinsic(llvm::Triple::ArchType Arch) { } } +// Return wave prefix sum that corresponds to the QT scalar type +static Intrinsic::ID getWavePrefixSumIntrinsic(llvm::Triple::ArchType Arch, + CGHLSLRuntime &RT, QualType QT) { + switch (Arch) { + case llvm::Triple::spirv: + return Intrinsic::spv_wave_prefix_sum; + case llvm::Triple::dxil: { + if (QT->isUnsignedIntegerType()) + return Intrinsic::dx_wave_prefix_usum; + return Intrinsic::dx_wave_prefix_sum; + } + default: + llvm_unreachable("Intrinsic WavePrefixSum" + " not supported by target architecture"); + } +} + // Returns the mangled name for a builtin function that the SPIR-V backend // will expand into a spec Constant. static std::string getSpecConstantFunctionName(clang::QualType SpecConstantType, @@ -1070,6 +1087,15 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID, {OpExpr->getType()}), ArrayRef{OpExpr, OpIndex}, "hlsl.wave.readlane"); } + case Builtin::BI__builtin_hlsl_wave_prefix_sum: { + Value *OpExpr = EmitScalarExpr(E->getArg(0)); + Intrinsic::ID IID = getWavePrefixSumIntrinsic( + getTarget().getTriple().getArch(), CGM.getHLSLRuntime(), + E->getArg(0)->getType()); + return EmitRuntimeCall(Intrinsic::getOrInsertDeclaration( + &CGM.getModule(), IID, {OpExpr->getType()}), + ArrayRef{OpExpr}, "hlsl.wave.prefix.sum"); + } case Builtin::BI__builtin_hlsl_elementwise_sign: { auto *Arg0 = E->getArg(0); Value *Op0 = EmitScalarExpr(Arg0); diff --git a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h index eb51244467bec..aaad8f94e23af 100644 --- a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h +++ b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h @@ -2908,6 +2908,105 @@ __attribute__((convergent)) double3 WaveActiveSum(double3); _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_sum) __attribute__((convergent)) double4 WaveActiveSum(double4); +//===----------------------------------------------------------------------===// +// WavePrefixSum builtins +//===----------------------------------------------------------------------===// + +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) half WavePrefixSum(half); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) half2 WavePrefixSum(half2); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) half3 WavePrefixSum(half3); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) half4 WavePrefixSum(half4); + +#ifdef __HLSL_ENABLE_16_BIT +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int16_t WavePrefixSum(int16_t); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int16_t2 WavePrefixSum(int16_t2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int16_t3 WavePrefixSum(int16_t3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int16_t4 WavePrefixSum(int16_t4); + +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint16_t WavePrefixSum(uint16_t); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint16_t2 WavePrefixSum(uint16_t2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint16_t3 WavePrefixSum(uint16_t3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint16_t4 WavePrefixSum(uint16_t4); +#endif + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int WavePrefixSum(int); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int2 WavePrefixSum(int2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int3 WavePrefixSum(int3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int4 WavePrefixSum(int4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint WavePrefixSum(uint); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint2 WavePrefixSum(uint2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint3 WavePrefixSum(uint3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint4 WavePrefixSum(uint4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int64_t WavePrefixSum(int64_t); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int64_t2 WavePrefixSum(int64_t2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int64_t3 WavePrefixSum(int64_t3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) int64_t4 WavePrefixSum(int64_t4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint64_t WavePrefixSum(uint64_t); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint64_t2 WavePrefixSum(uint64_t2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint64_t3 WavePrefixSum(uint64_t3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) uint64_t4 WavePrefixSum(uint64_t4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) float WavePrefixSum(float); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) float2 WavePrefixSum(float2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) float3 WavePrefixSum(float3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) float4 WavePrefixSum(float4); + +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) double WavePrefixSum(double); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) double2 WavePrefixSum(double2); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) double3 WavePrefixSum(double3); +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_prefix_sum) +__attribute__((convergent)) double4 WavePrefixSum(double4); + //===----------------------------------------------------------------------===// // sign builtins //===----------------------------------------------------------------------===// diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp index a880cb2271270..45f62b78e71ca 100644 --- a/clang/lib/Sema/SemaHLSL.cpp +++ b/clang/lib/Sema/SemaHLSL.cpp @@ -3189,10 +3189,13 @@ static bool CheckAnyScalarOrVector(Sema *S, CallExpr *TheCall, return false; } -static bool CheckWaveActive(Sema *S, CallExpr *TheCall) { +// Check that the argument is not a bool or vector<bool> +// Returns true on error +static bool CheckNotBoolScalarOrVector(Sema *S, CallExpr *TheCall, + unsigned ArgIndex) { QualType BoolType = S->getASTContext().BoolTy; - assert(TheCall->getNumArgs() >= 1); - QualType ArgType = TheCall->getArg(0)->getType(); + assert(ArgIndex < TheCall->getNumArgs()); + QualType ArgType = TheCall->getArg(ArgIndex)->getType(); auto *VTy = ArgType->getAs<VectorType>(); // is the bool or vector<bool> if (S->Context.hasSameUnqualifiedType(ArgType, BoolType) || @@ -3206,6 +3209,18 @@ static bool CheckWaveActive(Sema *S, CallExpr *TheCall) { return false; } +static bool CheckWaveActive(Sema *S, CallExpr *TheCall) { + if (CheckNotBoolScalarOrVector(S, TheCall, 0)) + return true; + return false; +} + +static bool CheckWavePrefix(Sema *S, CallExpr *TheCall) { + if (CheckNotBoolScalarOrVector(S, TheCall, 0)) + return true; + return false; +} + static bool CheckBoolSelect(Sema *S, CallExpr *TheCall) { assert(TheCall->getNumArgs() == 3); Expr *Arg1 = TheCall->getArg(1); @@ -3741,6 +3756,20 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) { return true; break; } + case Builtin::BI__builtin_hlsl_wave_prefix_sum: { + 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 (CheckWavePrefix(&SemaRef, TheCall)) + return true; + ExprResult Expr = TheCall->getArg(0); + QualType ArgTyExpr = Expr.get()->getType(); + TheCall->setType(ArgTyExpr); + break; + } case Builtin::BI__builtin_hlsl_elementwise_splitdouble: { if (SemaRef.checkArgCount(TheCall, 3)) return true; diff --git a/clang/test/CodeGenHLSL/builtins/WavePrefixSum.hlsl b/clang/test/CodeGenHLSL/builtins/WavePrefixSum.hlsl new file mode 100644 index 0000000000000..f22aa69ba45d5 --- /dev/null +++ b/clang/test/CodeGenHLSL/builtins/WavePrefixSum.hlsl @@ -0,0 +1,46 @@ +// 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: %[[RET:.*]] = call spir_func [[TY:.*]] @llvm.spv.wave.prefix.sum.i32([[TY]] %[[#]]) + // CHECK-DXIL: %[[RET:.*]] = call [[TY:.*]] @llvm.dx.wave.prefix.sum.i32([[TY]] %[[#]]) + // CHECK: ret [[TY]] %[[RET]] + return WavePrefixSum(expr); +} + +// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.prefix.sum.i32([[TY]]) #[[#attr:]] +// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.prefix.sum.i32([[TY]]) #[[#attr:]] + +// CHECK-LABEL: test_uint64_t +uint64_t test_uint64_t(uint64_t expr) { + // CHECK-SPIRV: %[[RET:.*]] = call spir_func [[TY:.*]] @llvm.spv.wave.prefix.sum.i64([[TY]] %[[#]]) + // CHECK-DXIL: %[[RET:.*]] = call [[TY:.*]] @llvm.dx.wave.prefix.usum.i64([[TY]] %[[#]]) + // CHECK: ret [[TY]] %[[RET]] + return WavePrefixSum(expr); +} + +// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.prefix.usum.i64([[TY]]) #[[#attr:]] +// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.prefix.sum.i64([[TY]]) #[[#attr:]] + +// Test basic lowering to runtime function call with array and float value. + +// CHECK-LABEL: test_floatv4 +float4 test_floatv4(float4 expr) { + // CHECK-SPIRV: %[[RET1:.*]] = call reassoc nnan ninf nsz arcp afn spir_func [[TY1:.*]] @llvm.spv.wave.prefix.sum.v4f32([[TY1]] %[[#]] + // CHECK-DXIL: %[[RET1:.*]] = call reassoc nnan ninf nsz arcp afn [[TY1:.*]] @llvm.dx.wave.prefix.sum.v4f32([[TY1]] %[[#]]) + // CHECK: ret [[TY1]] %[[RET1]] + return WavePrefixSum(expr); +} + +// CHECK-DXIL: declare [[TY1]] @llvm.dx.wave.prefix.sum.v4f32([[TY1]]) #[[#attr]] +// CHECK-SPIRV: declare [[TY1]] @llvm.spv.wave.prefix.sum.v4f32([[TY1]]) #[[#attr]] + +// CHECK: attributes #[[#attr]] = {{{.*}} convergent {{.*}}} + diff --git a/clang/test/SemaHLSL/BuiltIns/WavePrefixSum-errors.hlsl b/clang/test/SemaHLSL/BuiltIns/WavePrefixSum-errors.hlsl new file mode 100644 index 0000000000000..1e575c94e67a5 --- /dev/null +++ b/clang/test/SemaHLSL/BuiltIns/WavePrefixSum-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_prefix_sum(); + // expected-error@-1 {{too few arguments to function call, expected 1, have 0}} +} + +float2 test_too_many_arg(float2 p0) { + return __builtin_hlsl_wave_prefix_sum(p0, p0); + // expected-error@-1 {{too many arguments to function call, expected 1, have 2}} +} + +bool test_expr_bool_type_check(bool p0) { + return __builtin_hlsl_wave_prefix_sum(p0); + // expected-error@-1 {{invalid operand of type 'bool'}} +} + +bool2 test_expr_bool_vec_type_check(bool2 p0) { + return __builtin_hlsl_wave_prefix_sum(p0); + // expected-error@-1 {{invalid operand of type 'bool2' (aka 'vector<bool, 2>')}} +} + +struct S { float f; }; + +S test_expr_struct_type_check(S p0) { + return __builtin_hlsl_wave_prefix_sum(p0); + // expected-error@-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 f3494450fa88f..88732bfa5a892 100644 --- a/llvm/include/llvm/IR/IntrinsicsDirectX.td +++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td @@ -180,6 +180,8 @@ def int_dx_wave_is_first_lane : DefaultAttrsIntrinsic<[llvm_i1_ty], [], [IntrCon def int_dx_wave_readlane : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrConvergent, IntrNoMem]>; def int_dx_wave_get_lane_count : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrConvergent]>; +def int_dx_wave_prefix_sum : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; +def int_dx_wave_prefix_usum : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; def int_dx_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty], [IntrNoMem]>; def int_dx_step : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty, LLVMMatchType<0>], [IntrNoMem]>; def int_dx_splitdouble : DefaultAttrsIntrinsic<[llvm_anyint_ty, LLVMMatchType<0>], diff --git a/llvm/include/llvm/IR/IntrinsicsSPIRV.td b/llvm/include/llvm/IR/IntrinsicsSPIRV.td index 873033c426873..55db1f6b69347 100644 --- a/llvm/include/llvm/IR/IntrinsicsSPIRV.td +++ b/llvm/include/llvm/IR/IntrinsicsSPIRV.td @@ -132,6 +132,7 @@ def int_spv_rsqrt : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty] def int_spv_wave_readlane : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrConvergent, IntrNoMem]>; def int_spv_wave_get_lane_count : DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrConvergent]>; + def int_spv_wave_prefix_sum : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; def int_spv_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty], [IntrNoMem]>; def int_spv_radians : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty], [IntrNoMem]>; def int_spv_group_memory_barrier_with_group_sync : ClangBuiltin<"__builtin_spirv_group_barrier">, diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td index 4c1456fd5abb8..8751484496395 100644 --- a/llvm/lib/Target/DirectX/DXIL.td +++ b/llvm/lib/Target/DirectX/DXIL.td @@ -1124,6 +1124,30 @@ def WaveActiveOp : DXILOp<119, waveActiveOp> { let attributes = [Attributes<DXIL1_0, []>]; } +def WavePrefixOp : DXILOp<121, wavePrefixOp> { + let Doc = "returns partial result of the computation in the corresponding lane"; + let intrinsics = [ + IntrinSelect<int_dx_wave_prefix_sum, + [ + IntrinArgIndex<0>, IntrinArgI8<WaveOpKind_Sum>, + IntrinArgI8<SignedOpKind_Signed> + ]>, + IntrinSelect<int_dx_wave_prefix_usum, + [ + IntrinArgIndex<0>, IntrinArgI8<WaveOpKind_Sum>, + IntrinArgI8<SignedOpKind_Unsigned> + ]>, + ]; + + let arguments = [OverloadTy, Int8Ty, Int8Ty]; + let result = OverloadTy; + let overloads = [ + Overloads<DXIL1_0, [HalfTy, FloatTy, DoubleTy, Int16Ty, Int32Ty, Int64Ty]> + ]; + let stages = [Stages<DXIL1_0, [all_stages]>]; + let attributes = [Attributes<DXIL1_0, []>]; +} + def WavePrefixBitCount : DXILOp<136, wavePrefixOp> { let Doc = "returns the count of bits of Expr set to 1 on prior lanes"; let intrinsics = [IntrinSelect<int_dx_wave_prefix_bit_count>]; diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp index 57a6097aedcc0..54cf56f92277b 100644 --- a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp +++ b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp @@ -97,6 +97,9 @@ static bool checkWaveOps(Intrinsic::ID IID) { case Intrinsic::dx_wave_reduce_umax: case Intrinsic::dx_wave_reduce_min: case Intrinsic::dx_wave_reduce_umin: + // Wave Prefix Op Variants + case Intrinsic::dx_wave_prefix_sum: + case Intrinsic::dx_wave_prefix_usum: return true; } } diff --git a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp index 23f45b5fe2270..4d6f6d6472b25 100644 --- a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp +++ b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp @@ -61,9 +61,11 @@ bool DirectXTTIImpl::isTargetIntrinsicTriviallyScalarizable( case Intrinsic::dx_wave_reduce_max: case Intrinsic::dx_wave_reduce_min: case Intrinsic::dx_wave_reduce_sum: + case Intrinsic::dx_wave_prefix_sum: case Intrinsic::dx_wave_reduce_umax: case Intrinsic::dx_wave_reduce_umin: case Intrinsic::dx_wave_reduce_usum: + case Intrinsic::dx_wave_prefix_usum: case Intrinsic::dx_imad: case Intrinsic::dx_umad: case Intrinsic::dx_ddx_coarse: diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp index 0d0806def7537..72089ce0107ba 100644 --- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp +++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp @@ -241,6 +241,14 @@ class SPIRVInstructionSelector : public InstructionSelector { bool selectWaveReduceSum(Register ResVReg, const SPIRVType *ResType, MachineInstr &I) const; + template <typename PickOpcodeFn> + bool selectWaveExclusiveScan(Register ResVReg, const SPIRVType *ResType, + MachineInstr &I, bool IsUnsigned, + PickOpcodeFn &&PickOpcode) const; + + bool selectWaveExclusiveScanSum(Register ResVReg, const SPIRVType *ResType, + MachineInstr &I) const; + bool selectConst(Register ResVReg, const SPIRVType *ResType, MachineInstr &I) const; @@ -2852,6 +2860,43 @@ bool SPIRVInstructionSelector::selectWaveReduce( .constrainAllUses(TII, TRI, RBI); } +bool SPIRVInstructionSelector::selectWaveExclusiveScanSum( + Register ResVReg, const SPIRVType *ResType, MachineInstr &I) const { + return selectWaveExclusiveScan(ResVReg, ResType, I, /*IsUnsigned*/ false, + [&](Register InputRegister, bool IsUnsigned) { + bool IsFloatTy = GR.isScalarOrVectorOfType( + InputRegister, SPIRV::OpTypeFloat); + return IsFloatTy + ? SPIRV::OpGroupNonUniformFAdd + : SPIRV::OpGroupNonUniformIAdd; + }); +} + +template <typename PickOpcodeFn> +bool SPIRVInstructionSelector::selectWaveExclusiveScan( + Register ResVReg, const SPIRVType *ResType, MachineInstr &I, + bool IsUnsigned, PickOpcodeFn &&PickOpcode) 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); + const unsigned Opcode = PickOpcode(InputRegister, IsUnsigned); + return BuildMI(BB, I, I.getDebugLoc(), TII.get(Opcode)) + .addDef(ResVReg) + .addUse(GR.getSPIRVTypeID(ResType)) + .addUse(GR.getOrCreateConstInt(SPIRV::Scope::Subgroup, I, IntTy, TII, + !STI.isShader())) + .addImm(SPIRV::GroupOperation::ExclusiveScan) + .addUse(I.getOperand(2).getReg()) + .constrainAllUses(TII, TRI, RBI); +} + bool SPIRVInstructionSelector::selectBitreverse(Register ResVReg, const SPIRVType *ResType, MachineInstr &I) const { @@ -3938,6 +3983,8 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg, case Intrinsic::spv_wave_readlane: return selectWaveOpInst(ResVReg, ResType, I, SPIRV::OpGroupNonUniformShuffle); + case Intrinsic::spv_wave_prefix_sum: + return selectWaveExclusiveScanSum(ResVReg, ResType, I); case Intrinsic::spv_step: return selectExtInst(ResVReg, ResType, I, CL::step, GL::Step); case Intrinsic::spv_radians: diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll b/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll index c1eae0429baad..12dca7a78cc27 100644 --- a/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll +++ b/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll @@ -111,3 +111,17 @@ entry: %0 = call i32 @llvm.dx.wave.prefix.bit.count(i1 %expr) ret void } + +define noundef i32 @wave_prefix_sum(i32 noundef %x) { +entry: + ; CHECK: Function wave_prefix_sum : [[WAVE_FLAG]] + %ret = call i32 @llvm.dx.wave.prefix.sum.i32(i32 %x) + ret i32 %ret +} + +define noundef i32 @wave_prefix_usum(i32 noundef %x) { +entry: + ; CHECK: Function wave_prefix_usum : [[WAVE_FLAG]] + %ret = call i32 @llvm.dx.wave.prefix.usum.i32(i32 %x) + ret i32 %ret +} diff --git a/llvm/test/CodeGen/DirectX/WavePrefixSum.ll b/llvm/test/CodeGen/DirectX/WavePrefixSum.ll new file mode 100644 index 0000000000000..ed8c2b2b85465 --- /dev/null +++ b/llvm/test/CodeGen/DirectX/WavePrefixSum.ll @@ -0,0 +1,143 @@ +; RUN: opt -S -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s + +; Test that for scalar values, WavePrefixSum maps down to the DirectX op + +define noundef half @wave_prefix_sum_half(half noundef %expr) { +entry: +; CHECK: call half @dx.op.wavePrefixOp.f16(i32 121, half %expr, i8 0, i8 0) + %ret = call half @llvm.dx.wave.prefix.sum.f16(half %expr) + ret half %ret +} + +define noundef float @wave_prefix_sum_float(float noundef %expr) { +entry: +; CHECK: call float @dx.op.wavePrefixOp.f32(i32 121, float %expr, i8 0, i8 0) + %ret = call float @llvm.dx.wave.prefix.sum.f32(float %expr) + ret float %ret +} + +define noundef double @wave_prefix_sum_double(double noundef %expr) { +entry: +; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr, i8 0, i8 0) + %ret = call double @llvm.dx.wave.prefix.sum.f64(double %expr) + ret double %ret +} + +define noundef i16 @wave_prefix_sum_i16(i16 noundef %expr) { +entry: +; CHECK: call i16 @dx.op.wavePrefixOp.i16(i32 121, i16 %expr, i8 0, i8 0) + %ret = call i16 @llvm.dx.wave.prefix.sum.i16(i16 %expr) + ret i16 %ret +} + +define noundef i32 @wave_prefix_sum_i32(i32 noundef %expr) { +entry: +; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr, i8 0, i8 0) + %ret = call i32 @llvm.dx.wave.prefix.sum.i32(i32 %expr) + ret i32 %ret +} + +define noundef i64 @wave_prefix_sum_i64(i64 noundef %expr) { +entry: +; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr, i8 0, i8 0) + %ret = call i64 @llvm.dx.wave.prefix.sum.i64(i64 %expr) + ret i64 %ret +} + +define noundef i16 @wave_prefix_usum_i16(i16 noundef %expr) { +entry: +; CHECK: call i16 @dx.op.wavePrefixOp.i16(i32 121, i16 %expr, i8 0, i8 1) + %ret = call i16 @llvm.dx.wave.prefix.usum.i16(i16 %expr) + ret i16 %ret +} + +define noundef i32 @wave_prefix_usum_i32(i32 noundef %expr) { +entry: +; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr, i8 0, i8 1) + %ret = call i32 @llvm.dx.wave.prefix.usum.i32(i32 %expr) + ret i32 %ret +} + +define noundef i64 @wave_prefix_usum_i64(i64 noundef %expr) { +entry: +; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr, i8 0, i8 1) + %ret = call i64 @llvm.dx.wave.prefix.usum.i64(i64 %expr) + ret i64 %ret +} + +declare half @llvm.dx.wave.prefix.sum.f16(half) +declare float @llvm.dx.wave.prefix.sum.f32(float) +declare double @llvm.dx.wave.prefix.sum.f64(double) + +declare i16 @llvm.dx.wave.prefix.sum.i16(i16) +declare i32 @llvm.dx.wave.prefix.sum.i32(i32) +declare i64 @llvm.dx.wave.prefix.sum.i64(i64) + +declare i16 @llvm.dx.wave.prefix.usum.i16(i16) +declare i32 @llvm.dx.wave.prefix.usum.i32(i32) +declare i64 @llvm.dx.wave.prefix.usum.i64(i64) + +; Test that for vector values, WavePrefixSum scalarizes and maps down to the +; DirectX op + +define noundef <2 x half> @wave_prefix_sum_v2half(<2 x half> noundef %expr) { +entry: +; CHECK: call half @dx.op.wavePrefixOp.f16(i32 121, half %expr.i0, i8 0, i8 0) +; CHECK: call half @dx.op.wavePrefixOp.f16(i32 121, half %expr.i1, i8 0, i8 0) + %ret = call <2 x half> @llvm.dx.wave.prefix.sum.v2f16(<2 x half> %expr) + ret <2 x half> %ret +} + +define noundef <3 x i32> @wave_prefix_sum_v3i32(<3 x i32> noundef %expr) { +entry: +; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i0, i8 0, i8 0) +; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i1, i8 0, i8 0) +; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i2, i8 0, i8 0) + %ret = call <3 x i32> @llvm.dx.wave.prefix.sum.v3i32(<3 x i32> %expr) + ret <3 x i32> %ret +} + +define noundef <4 x double> @wave_prefix_sum_v4f64(<4 x double> noundef %expr) { +entry: +; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr.i0, i8 0, i8 0) +; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr.i1, i8 0, i8 0) +; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr.i2, i8 0, i8 0) +; CHECK: call double @dx.op.wavePrefixOp.f64(i32 121, double %expr.i3, i8 0, i8 0) + %ret = call <4 x double> @llvm.dx.wave.prefix.sum.v464(<4 x double> %expr) + ret <4 x double> %ret +} + +declare <2 x half> @llvm.dx.wave.prefix.sum.v2f16(<2 x half>) +declare <3 x i32> @llvm.dx.wave.prefix.sum.v3i32(<3 x i32>) +declare <4 x double> @llvm.dx.wave.prefix.sum.v4f64(<4 x double>) + +define noundef <2 x i16> @wave_prefix_usum_v2i16(<2 x i16> noundef %expr) { +entry: +; CHECK: call i16 @dx.op.wavePrefixOp.i16(i32 121, i16 %expr.i0, i8 0, i8 1) +; CHECK: call i16 @dx.op.wavePrefixOp.i16(i32 121, i16 %expr.i1, i8 0, i8 1) + %ret = call <2 x i16> @llvm.dx.wave.prefix.usum.v2f16(<2 x i16> %expr) + ret <2 x i16> %ret +} + +define noundef <3 x i32> @wave_prefix_usum_v3i32(<3 x i32> noundef %expr) { +entry: +; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i0, i8 0, i8 1) +; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i1, i8 0, i8 1) +; CHECK: call i32 @dx.op.wavePrefixOp.i32(i32 121, i32 %expr.i2, i8 0, i8 1) + %ret = call <3 x i32> @llvm.dx.wave.prefix.usum.v3i32(<3 x i32> %expr) + ret <3 x i32> %ret +} + +define noundef <4 x i64> @wave_prefix_usum_v4f64(<4 x i64> noundef %expr) { +entry: +; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr.i0, i8 0, i8 1) +; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr.i1, i8 0, i8 1) +; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr.i2, i8 0, i8 1) +; CHECK: call i64 @dx.op.wavePrefixOp.i64(i32 121, i64 %expr.i3, i8 0, i8 1) + %ret = call <4 x i64> @llvm.dx.wave.prefix.usum.v464(<4 x i64> %expr) + ret <4 x i64> %ret +} + +declare <2 x i16> @llvm.dx.wave.prefix.usum.v2f16(<2 x i16>) +declare <3 x i32> @llvm.dx.wave.prefix.usum.v3i32(<3 x i32>) +declare <4 x i64> @llvm.dx.wave.prefix.usum.v4f64(<4 x i64>) diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WavePrefixSum.ll b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WavePrefixSum.ll new file mode 100644 index 0000000000000..5fb82fd9ebf19 --- /dev/null +++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WavePrefixSum.ll @@ -0,0 +1,41 @@ +; RUN: llc -verify-machineinstrs -O0 -mtriple=spirv-vulkan-unknown %s -o - | FileCheck %s +; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv-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:]] = OpGroupNonUniformFAdd %[[#f32]] %[[#scope]] ExclusiveScan %[[#fexpr]] + %0 = call float @llvm.spv.wave.prefix.sum.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:]] = OpGroupNonUniformIAdd %[[#uint]] %[[#scope]] ExclusiveScan %[[#iexpr]] + %0 = call i32 @llvm.spv.wave.prefix.sum.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:]] = OpGroupNonUniformFAdd %[[#v4_half]] %[[#scope]] ExclusiveScan %[[#vbexpr]] + %0 = call <4 x half> @llvm.spv.wave.prefix.sum.v4half(<4 x half> %vbexpr) + ret <4 x half> %0 +} + +declare float @llvm.spv.wave.prefix.sum.f32(float) +declare i32 @llvm.spv.wave.prefix.sum.i32(i32) +declare <4 x half> @llvm.spv.wave.prefix.sum.v4half(<4 x half>) _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
