diff --git a/clang/include/clang/Basic/Builtins.td b/clang/include/clang/Basic/Builtins.td index 792e2e07ec594..6ed6e0b38e31b 100644 --- a/clang/include/clang/Basic/Builtins.td +++ b/clang/include/clang/Basic/Builtins.td @@ -5017,6 +5017,12 @@ def HLSLWaveActiveSum : LangBuiltin<"HLSL_LANG"> { let Prototype = "void (...)"; } +def HLSLWaveActiveProduct : LangBuiltin<"HLSL_LANG"> { + let Spellings = ["__builtin_hlsl_wave_active_product"]; + 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/lib/CodeGen/CGHLSLBuiltins.cpp b/clang/lib/CodeGen/CGHLSLBuiltins.cpp index 4f2f5a761f197..68c2de759cebd 100644 --- a/clang/lib/CodeGen/CGHLSLBuiltins.cpp +++ b/clang/lib/CodeGen/CGHLSLBuiltins.cpp @@ -196,6 +196,23 @@ static Intrinsic::ID getWaveActiveSumIntrinsic(llvm::Triple::ArchType Arch, } } +// Return wave active product that corresponds to the QT scalar type +static Intrinsic::ID getWaveActiveProductIntrinsic(llvm::Triple::ArchType Arch, + CGHLSLRuntime &RT, QualType QT) { + switch (Arch) { + case llvm::Triple::spirv: + return Intrinsic::spv_wave_reduce_product; + case llvm::Triple::dxil: { + if (QT->isUnsignedIntegerType()) + return Intrinsic::dx_wave_reduce_uproduct; + return Intrinsic::dx_wave_reduce_product; + } + default: + llvm_unreachable("Intrinsic WaveActiveProduct" + " not supported by target architecture"); + } +} + // Return wave active sum that corresponds to the QT scalar type static Intrinsic::ID getWaveActiveMaxIntrinsic(llvm::Triple::ArchType Arch, CGHLSLRuntime &RT, QualType QT) { @@ -708,6 +725,17 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID, &CGM.getModule(), IID, {OpExpr->getType()}), ArrayRef{OpExpr}, "hlsl.wave.active.sum"); } + case Builtin::BI__builtin_hlsl_wave_active_product: { + // Due to the use of variadic arguments, explicitly retreive argument + Value *OpExpr = EmitScalarExpr(E->getArg(0)); + Intrinsic::ID IID = getWaveActiveProductIntrinsic( + getTarget().getTriple().getArch(), CGM.getHLSLRuntime(), + E->getArg(0)->getType()); + + return EmitRuntimeCall(Intrinsic::getOrInsertDeclaration( + &CGM.getModule(), IID, {OpExpr->getType()}), + ArrayRef{OpExpr}, "hlsl.wave.active.product"); + } case Builtin::BI__builtin_hlsl_wave_active_max: { // Due to the use of variadic arguments, explicitly retreive argument Value *OpExpr = EmitScalarExpr(E->getArg(0)); diff --git a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h index d973371312701..0eb1b9075bf8d 100644 --- a/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h +++ b/clang/lib/Headers/hlsl/hlsl_alias_intrinsics.h @@ -2696,6 +2696,130 @@ __attribute__((convergent)) double3 WaveActiveSum(double3); _HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_sum) __attribute__((convergent)) double4 WaveActiveSum(double4); +//===----------------------------------------------------------------------===// +// WaveActiveProduct builtins +//===----------------------------------------------------------------------===// + +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) half WaveActiveProduct(half); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) half2 WaveActiveProduct(half2); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) half3 WaveActiveProduct(half3); +_HLSL_16BIT_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) half4 WaveActiveProduct(half4); + +#ifdef __HLSL_ENABLE_16_BIT +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int16_t WaveActiveProduct(int16_t); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int16_t2 WaveActiveProduct(int16_t2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int16_t3 WaveActiveProduct(int16_t3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int16_t4 WaveActiveProduct(int16_t4); + +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint16_t WaveActiveProduct(uint16_t); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint16_t2 WaveActiveProduct(uint16_t2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint16_t3 WaveActiveProduct(uint16_t3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint16_t4 WaveActiveProduct(uint16_t4); +#endif + +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int WaveActiveProduct(int); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int2 WaveActiveProduct(int2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int3 WaveActiveProduct(int3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int4 WaveActiveProduct(int4); + +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint WaveActiveProduct(uint); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint2 WaveActiveProduct(uint2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint3 WaveActiveProduct(uint3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint4 WaveActiveProduct(uint4); + +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int64_t WaveActiveProduct(int64_t); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int64_t2 WaveActiveProduct(int64_t2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int64_t3 WaveActiveProduct(int64_t3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) int64_t4 WaveActiveProduct(int64_t4); + +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint64_t WaveActiveProduct(uint64_t); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint64_t2 WaveActiveProduct(uint64_t2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint64_t3 WaveActiveProduct(uint64_t3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) uint64_t4 WaveActiveProduct(uint64_t4); + +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) float WaveActiveProduct(float); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) float2 WaveActiveProduct(float2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) float3 WaveActiveProduct(float3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) float4 WaveActiveProduct(float4); + +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) double WaveActiveProduct(double); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) double2 WaveActiveProduct(double2); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) double3 WaveActiveProduct(double3); +_HLSL_AVAILABILITY(shadermodel, 6.0) +_HLSL_BUILTIN_ALIAS(__builtin_hlsl_wave_active_product) +__attribute__((convergent)) double4 WaveActiveProduct(double4); + + //===----------------------------------------------------------------------===// // sign builtins //===----------------------------------------------------------------------===// diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp index f34706677b59f..ed3c8a406e0f8 100644 --- a/clang/lib/Sema/SemaHLSL.cpp +++ b/clang/lib/Sema/SemaHLSL.cpp @@ -3197,7 +3197,8 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) { break; } case Builtin::BI__builtin_hlsl_wave_active_max: - case Builtin::BI__builtin_hlsl_wave_active_sum: { + case Builtin::BI__builtin_hlsl_wave_active_sum: + case Builtin::BI__builtin_hlsl_wave_active_product: { if (SemaRef.checkArgCount(TheCall, 1)) return true; diff --git a/clang/test/CodeGenHLSL/builtins/WaveActiveProduct.hlsl b/clang/test/CodeGenHLSL/builtins/WaveActiveProduct.hlsl new file mode 100644 index 0000000000000..9de5971a80d97 --- /dev/null +++ b/clang/test/CodeGenHLSL/builtins/WaveActiveProduct.hlsl @@ -0,0 +1,45 @@ +// 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.reduce.product.i32([[TY]] %[[#]]) + // CHECK-DXIL: %[[RET:.*]] = call [[TY:.*]] @llvm.dx.wave.reduce.product.i32([[TY]] %[[#]]) + // CHECK: ret [[TY]] %[[RET]] + return WaveActiveProduct(expr); +} + +// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.reduce.product.i32([[TY]]) #[[#attr:]] +// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.reduce.product.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.reduce.product.i64([[TY]] %[[#]]) + // CHECK-DXIL: %[[RET:.*]] = call [[TY:.*]] @llvm.dx.wave.reduce.uproduct.i64([[TY]] %[[#]]) + // CHECK: ret [[TY]] %[[RET]] + return WaveActiveProduct(expr); +} + +// CHECK-DXIL: declare [[TY]] @llvm.dx.wave.reduce.uproduct.i64([[TY]]) #[[#attr:]] +// CHECK-SPIRV: declare [[TY]] @llvm.spv.wave.reduce.product.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.reduce.product.v4f32([[TY1]] %[[#]] + // CHECK-DXIL: %[[RET1:.*]] = call reassoc nnan ninf nsz arcp afn [[TY1:.*]] @llvm.dx.wave.reduce.product.v4f32([[TY1]] %[[#]]) + // CHECK: ret [[TY1]] %[[RET1]] + return WaveActiveProduct(expr); +} + +// CHECK-DXIL: declare [[TY1]] @llvm.dx.wave.reduce.product.v4f32([[TY1]]) #[[#attr]] +// CHECK-SPIRV: declare [[TY1]] @llvm.spv.wave.reduce.product.v4f32([[TY1]]) #[[#attr]] + +// CHECK: attributes #[[#attr]] = {{{.*}} convergent {{.*}}} diff --git a/clang/test/SemaHLSL/BuiltIns/WaveActiveProduct-errors.hlsl b/clang/test/SemaHLSL/BuiltIns/WaveActiveProduct-errors.hlsl new file mode 100644 index 0000000000000..43ad02b35fc1c --- /dev/null +++ b/clang/test/SemaHLSL/BuiltIns/WaveActiveProduct-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_product(); + // expected-error@-1 {{too few arguments to function call, expected 1, have 0}} +} + +float2 test_too_many_arg(float2 p0) { + return __builtin_hlsl_wave_active_product(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_active_product(p0); + // expected-error@-1 {{invalid operand of type 'bool'}} +} + +bool2 test_expr_bool_vec_type_check(bool2 p0) { + return __builtin_hlsl_wave_active_product(p0); + // expected-error@-1 {{invalid operand of type 'bool2' (aka 'vector')}} +} + +struct S { float f; }; + +S test_expr_struct_type_check(S p0) { + return __builtin_hlsl_wave_active_product(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 3b7077c52db21..8906548cfbbe4 100644 --- a/llvm/include/llvm/IR/IntrinsicsDirectX.td +++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td @@ -155,6 +155,8 @@ def int_dx_wave_reduce_max : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType def int_dx_wave_reduce_umax : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; def int_dx_wave_reduce_sum : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; def int_dx_wave_reduce_usum : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; +def int_dx_wave_reduce_product : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; +def int_dx_wave_reduce_uproduct : 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_wave_get_lane_count diff --git a/llvm/include/llvm/IR/IntrinsicsSPIRV.td b/llvm/include/llvm/IR/IntrinsicsSPIRV.td index 49a182be98acd..a0e387a504b75 100644 --- a/llvm/include/llvm/IR/IntrinsicsSPIRV.td +++ b/llvm/include/llvm/IR/IntrinsicsSPIRV.td @@ -123,6 +123,7 @@ def int_spv_rsqrt : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty] def int_spv_wave_reduce_umax : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; def int_spv_wave_reduce_max : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; def int_spv_wave_reduce_sum : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>], [IntrConvergent, IntrNoMem]>; + def int_spv_wave_reduce_product : DefaultAttrsIntrinsic<[llvm_any_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_wave_get_lane_count @@ -136,7 +137,7 @@ def int_spv_rsqrt : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty] def int_spv_sclamp : DefaultAttrsIntrinsic<[llvm_anyint_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>; def int_spv_nclamp : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>; - // Create resource handle given the binding information. Returns a + // Create resource handle given the binding information. Returns a // type appropriate for the kind of resource given the set id, binding id, // array size of the binding, as well as an index and an indicator // whether that index may be non-uniform. diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td index 44c48305f2832..22c427ccdefae 100644 --- a/llvm/lib/Target/DirectX/DXIL.td +++ b/llvm/lib/Target/DirectX/DXIL.td @@ -1048,6 +1048,16 @@ def WaveActiveOp : DXILOp<119, waveActiveOp> { IntrinArgIndex<0>, IntrinArgI8, IntrinArgI8 ]>, + IntrinSelect, IntrinArgI8, + IntrinArgI8 + ]>, + IntrinSelect, IntrinArgI8, + IntrinArgI8 + ]>, IntrinSelect, IntrinArgI8, diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp index e7e7f2ce66ae8..0c0daa3aadd17 100644 --- a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp +++ b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp @@ -92,6 +92,8 @@ static bool checkWaveOps(Intrinsic::ID IID) { // Wave Active Op Variants case Intrinsic::dx_wave_reduce_sum: case Intrinsic::dx_wave_reduce_usum: + case Intrinsic::dx_wave_reduce_product: + case Intrinsic::dx_wave_reduce_uproduct: case Intrinsic::dx_wave_reduce_max: case Intrinsic::dx_wave_reduce_umax: return true; diff --git a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp index 68fd3e0bc74c7..a863476353e59 100644 --- a/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp +++ b/llvm/lib/Target/DirectX/DirectXTargetTransformInfo.cpp @@ -56,8 +56,10 @@ bool DirectXTTIImpl::isTargetIntrinsicTriviallyScalarizable( case Intrinsic::dx_wave_readlane: case Intrinsic::dx_wave_reduce_max: case Intrinsic::dx_wave_reduce_sum: + case Intrinsic::dx_wave_reduce_product: case Intrinsic::dx_wave_reduce_umax: case Intrinsic::dx_wave_reduce_usum: + case Intrinsic::dx_wave_reduce_uproduct: case Intrinsic::dx_imad: case Intrinsic::dx_umad: return true; diff --git a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp index a0cff4d82b500..c02b65c68983c 100644 --- a/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp +++ b/llvm/lib/Target/SPIRV/SPIRVInstructionSelector.cpp @@ -225,6 +225,9 @@ class SPIRVInstructionSelector : public InstructionSelector { bool selectWaveReduceSum(Register ResVReg, const SPIRVType *ResType, MachineInstr &I) const; + bool selectWaveReduceProduct(Register ResVReg, const SPIRVType *ResType, + MachineInstr &I) const; + bool selectConst(Register ResVReg, const SPIRVType *ResType, MachineInstr &I) const; @@ -2482,6 +2485,32 @@ bool SPIRVInstructionSelector::selectWaveReduceSum(Register ResVReg, .addUse(I.getOperand(2).getReg()); } +bool SPIRVInstructionSelector::selectWaveReduceProduct(Register ResVReg, + const SPIRVType *ResType, + MachineInstr &I) 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::OpGroupNonUniformFMul : SPIRV::OpGroupNonUniformIMul; + 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::Reduce) + .addUse(I.getOperand(2).getReg()); +} + bool SPIRVInstructionSelector::selectBitreverse(Register ResVReg, const SPIRVType *ResType, MachineInstr &I) const { @@ -3433,6 +3462,8 @@ bool SPIRVInstructionSelector::selectIntrinsic(Register ResVReg, return selectWaveReduceMax(ResVReg, ResType, I, /*IsUnsigned*/ false); case Intrinsic::spv_wave_reduce_sum: return selectWaveReduceSum(ResVReg, ResType, I); + case Intrinsic::spv_wave_reduce_product: + return selectWaveReduceProduct(ResVReg, ResType, I); case Intrinsic::spv_wave_readlane: return selectWaveOpInst(ResVReg, ResType, I, SPIRV::OpGroupNonUniformShuffle); diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll b/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll index 7a876f67615cd..40f67cbd7608a 100644 --- a/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll +++ b/llvm/test/CodeGen/DirectX/ShaderFlags/wave-ops.ll @@ -62,6 +62,20 @@ entry: ret i32 %ret } +define noundef i32 @wave_reduce_product(i32 noundef %x) { +entry: + ; CHECK: Function wave_reduce_product : [[WAVE_FLAG]] + %ret = call i32 @llvm.dx.wave.reduce.product.i32(i32 %x) + ret i32 %ret +} + +define noundef i32 @wave_reduce_uproduct(i32 noundef %x) { +entry: + ; CHECK: Function wave_reduce_uproduct : [[WAVE_FLAG]] + %ret = call i32 @llvm.dx.wave.reduce.product.i32(i32 %x) + ret i32 %ret +} + define noundef i32 @wave_reduce_max(i32 noundef %x) { entry: ; CHECK: Function wave_reduce_max : [[WAVE_FLAG]] diff --git a/llvm/test/CodeGen/DirectX/WaveActiveProduct.ll b/llvm/test/CodeGen/DirectX/WaveActiveProduct.ll new file mode 100644 index 0000000000000..373d8972c3435 --- /dev/null +++ b/llvm/test/CodeGen/DirectX/WaveActiveProduct.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, WaveAcitveProduct maps down to the DirectX op + +define noundef half @wave_active_product_half(half noundef %expr) { +entry: +; CHECK: call half @dx.op.waveActiveOp.f16(i32 119, half %expr, i8 1, i8 0) + %ret = call half @llvm.dx.wave.reduce.product.f16(half %expr) + ret half %ret +} + +define noundef float @wave_active_product_float(float noundef %expr) { +entry: +; CHECK: call float @dx.op.waveActiveOp.f32(i32 119, float %expr, i8 1, i8 0) + %ret = call float @llvm.dx.wave.reduce.product.f32(float %expr) + ret float %ret +} + +define noundef double @wave_active_product_double(double noundef %expr) { +entry: +; CHECK: call double @dx.op.waveActiveOp.f64(i32 119, double %expr, i8 1, i8 0) + %ret = call double @llvm.dx.wave.reduce.product.f64(double %expr) + ret double %ret +} + +define noundef i16 @wave_active_product_i16(i16 noundef %expr) { +entry: +; CHECK: call i16 @dx.op.waveActiveOp.i16(i32 119, i16 %expr, i8 1, i8 0) + %ret = call i16 @llvm.dx.wave.reduce.product.i16(i16 %expr) + ret i16 %ret +} + +define noundef i32 @wave_active_product_i32(i32 noundef %expr) { +entry: +; CHECK: call i32 @dx.op.waveActiveOp.i32(i32 119, i32 %expr, i8 1, i8 0) + %ret = call i32 @llvm.dx.wave.reduce.product.i32(i32 %expr) + ret i32 %ret +} + +define noundef i64 @wave_active_product_i64(i64 noundef %expr) { +entry: +; CHECK: call i64 @dx.op.waveActiveOp.i64(i32 119, i64 %expr, i8 1, i8 0) + %ret = call i64 @llvm.dx.wave.reduce.product.i64(i64 %expr) + ret i64 %ret +} + +define noundef i16 @wave_active_uproduct_i16(i16 noundef %expr) { +entry: +; CHECK: call i16 @dx.op.waveActiveOp.i16(i32 119, i16 %expr, i8 1, i8 1) + %ret = call i16 @llvm.dx.wave.reduce.uproduct.i16(i16 %expr) + ret i16 %ret +} + +define noundef i32 @wave_active_uproduct_i32(i32 noundef %expr) { +entry: +; CHECK: call i32 @dx.op.waveActiveOp.i32(i32 119, i32 %expr, i8 1, i8 1) + %ret = call i32 @llvm.dx.wave.reduce.uproduct.i32(i32 %expr) + ret i32 %ret +} + +define noundef i64 @wave_active_uproduct_i64(i64 noundef %expr) { +entry: +; CHECK: call i64 @dx.op.waveActiveOp.i64(i32 119, i64 %expr, i8 1, i8 1) + %ret = call i64 @llvm.dx.wave.reduce.uproduct.i64(i64 %expr) + ret i64 %ret +} + +declare half @llvm.dx.wave.reduce.product.f16(half) +declare float @llvm.dx.wave.reduce.product.f32(float) +declare double @llvm.dx.wave.reduce.product.f64(double) + +declare i16 @llvm.dx.wave.reduce.product.i16(i16) +declare i32 @llvm.dx.wave.reduce.product.i32(i32) +declare i64 @llvm.dx.wave.reduce.product.i64(i64) + +declare i16 @llvm.dx.wave.reduce.uproduct.i16(i16) +declare i32 @llvm.dx.wave.reduce.uproduct.i32(i32) +declare i64 @llvm.dx.wave.reduce.uproduct.i64(i64) + +; Test that for vector values, WaveAcitveProduct scalarizes and maps down to the +; DirectX op + +define noundef <2 x half> @wave_active_product_v2half(<2 x half> noundef %expr) { +entry: +; CHECK: call half @dx.op.waveActiveOp.f16(i32 119, half %expr.i0, i8 1, i8 0) +; CHECK: call half @dx.op.waveActiveOp.f16(i32 119, half %expr.i1, i8 1, i8 0) + %ret = call <2 x half> @llvm.dx.wave.reduce.product.v2f16(<2 x half> %expr) + ret <2 x half> %ret +} + +define noundef <3 x i32> @wave_active_product_v3i32(<3 x i32> noundef %expr) { +entry: +; CHECK: call i32 @dx.op.waveActiveOp.i32(i32 119, i32 %expr.i0, i8 1, i8 0) +; CHECK: call i32 @dx.op.waveActiveOp.i32(i32 119, i32 %expr.i1, i8 1, i8 0) +; CHECK: call i32 @dx.op.waveActiveOp.i32(i32 119, i32 %expr.i2, i8 1, i8 0) + %ret = call <3 x i32> @llvm.dx.wave.reduce.product.v3i32(<3 x i32> %expr) + ret <3 x i32> %ret +} + +define noundef <4 x double> @wave_active_product_v4f64(<4 x double> noundef %expr) { +entry: +; CHECK: call double @dx.op.waveActiveOp.f64(i32 119, double %expr.i0, i8 1, i8 0) +; CHECK: call double @dx.op.waveActiveOp.f64(i32 119, double %expr.i1, i8 1, i8 0) +; CHECK: call double @dx.op.waveActiveOp.f64(i32 119, double %expr.i2, i8 1, i8 0) +; CHECK: call double @dx.op.waveActiveOp.f64(i32 119, double %expr.i3, i8 1, i8 0) + %ret = call <4 x double> @llvm.dx.wave.reduce.product.v464(<4 x double> %expr) + ret <4 x double> %ret +} + +declare <2 x half> @llvm.dx.wave.reduce.product.v2f16(<2 x half>) +declare <3 x i32> @llvm.dx.wave.reduce.product.v3i32(<3 x i32>) +declare <4 x double> @llvm.dx.wave.reduce.product.v4f64(<4 x double>) + +define noundef <2 x i16> @wave_active_uproduct_v2i16(<2 x i16> noundef %expr) { +entry: +; CHECK: call i16 @dx.op.waveActiveOp.i16(i32 119, i16 %expr.i0, i8 1, i8 1) +; CHECK: call i16 @dx.op.waveActiveOp.i16(i32 119, i16 %expr.i1, i8 1, i8 1) + %ret = call <2 x i16> @llvm.dx.wave.reduce.uproduct.v2f16(<2 x i16> %expr) + ret <2 x i16> %ret +} + +define noundef <3 x i32> @wave_active_uproduct_v3i32(<3 x i32> noundef %expr) { +entry: +; CHECK: call i32 @dx.op.waveActiveOp.i32(i32 119, i32 %expr.i0, i8 1, i8 1) +; CHECK: call i32 @dx.op.waveActiveOp.i32(i32 119, i32 %expr.i1, i8 1, i8 1) +; CHECK: call i32 @dx.op.waveActiveOp.i32(i32 119, i32 %expr.i2, i8 1, i8 1) + %ret = call <3 x i32> @llvm.dx.wave.reduce.uproduct.v3i32(<3 x i32> %expr) + ret <3 x i32> %ret +} + +define noundef <4 x i64> @wave_active_uproduct_v4f64(<4 x i64> noundef %expr) { +entry: +; CHECK: call i64 @dx.op.waveActiveOp.i64(i32 119, i64 %expr.i0, i8 1, i8 1) +; CHECK: call i64 @dx.op.waveActiveOp.i64(i32 119, i64 %expr.i1, i8 1, i8 1) +; CHECK: call i64 @dx.op.waveActiveOp.i64(i32 119, i64 %expr.i2, i8 1, i8 1) +; CHECK: call i64 @dx.op.waveActiveOp.i64(i32 119, i64 %expr.i3, i8 1, i8 1) + %ret = call <4 x i64> @llvm.dx.wave.reduce.uproduct.v464(<4 x i64> %expr) + ret <4 x i64> %ret +} + +declare <2 x i16> @llvm.dx.wave.reduce.uproduct.v2f16(<2 x i16>) +declare <3 x i32> @llvm.dx.wave.reduce.uproduct.v3i32(<3 x i32>) +declare <4 x i64> @llvm.dx.wave.reduce.uproduct.v4f64(<4 x i64>) diff --git a/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WaveActiveProduct.ll b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WaveActiveProduct.ll new file mode 100644 index 0000000000000..e7633fcd4c9ea --- /dev/null +++ b/llvm/test/CodeGen/SPIRV/hlsl-intrinsics/WaveActiveProduct.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:]] = OpGroupNonUniformFMul %[[#f32]] %[[#scope]] Reduce %[[#fexpr]] + %0 = call float @llvm.spv.wave.reduce.product.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:]] = OpGroupNonUniformIMul %[[#uint]] %[[#scope]] Reduce %[[#iexpr]] + %0 = call i32 @llvm.spv.wave.reduce.product.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:]] = OpGroupNonUniformFMul %[[#v4_half]] %[[#scope]] Reduce %[[#vbexpr]] + %0 = call <4 x half> @llvm.spv.wave.reduce.product.v4half(<4 x half> %vbexpr) + ret <4 x half> %0 +} + +declare float @llvm.spv.wave.reduce.product.f32(float) +declare i32 @llvm.spv.wave.reduce.product.i32(i32) +declare <4 x half> @llvm.spv.wave.reduce.product.v4half(<4 x half>)