153 lines · plain
1; RUN: llc -O0 -mtriple=spirv64-unknown-unknown %s -o - | FileCheck %s --check-prefix=CHECK-SPIRV2; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv64-unknown-unknown %s -o - -filetype=obj | spirv-val --target-env spv1.4 %}3 4; RUN: llc -O0 -mtriple=spirv32-unknown-unknown %s -o - | FileCheck %s --check-prefix=CHECK-SPIRV5; RUN: %if spirv-tools %{ llc -O0 -mtriple=spirv32-unknown-unknown %s -o - -filetype=obj | spirv-val --target-env spv1.4 %}6 7; CHECK-SPIRV: OpCapability Groups8; CHECK-SPIRV-DAG: %[[#Int32Ty:]] = OpTypeInt 32 09; CHECK-SPIRV-DAG: %[[#Int64Ty:]] = OpTypeInt 64 010; CHECK-SPIRV-DAG: %[[#Float32Ty:]] = OpTypeFloat 3211; CHECK-SPIRV-DAG: %[[#Vec2Int32Ty:]] = OpTypeVector %[[#Int32Ty]] 212; CHECK-SPIRV-DAG: %[[#Vec3Int32Ty:]] = OpTypeVector %[[#Int32Ty]] 313; CHECK-SPIRV-DAG: %[[#Vec2Int64Ty:]] = OpTypeVector %[[#Int64Ty]] 214; CHECK-SPIRV-DAG: %[[#C2:]] = OpConstant %[[#Int32Ty]] 215 16; CHECK-SPIRV: OpFunction17; CHECK-SPIRV: %[[#Val:]] = OpFunctionParameter %[[#Int32Ty]]18; CHECK-SPIRV: %[[#X:]] = OpFunctionParameter %[[#Int32Ty]]19; CHECK-SPIRV: %[[#Y:]] = OpFunctionParameter %[[#Int32Ty]]20; CHECK-SPIRV: %[[#Z:]] = OpFunctionParameter %[[#Int32Ty]]21; CHECK-SPIRV: %[[#]] = OpGroupBroadcast %[[#Int32Ty]] %[[#C2]] %[[#Val]] %[[#X]]22; CHECK-SPIRV: %[[#XY:]] = OpCompositeConstruct %[[#Vec2Int32Ty]] %[[#X]] %[[#Y]]23; CHECK-SPIRV: %[[#]] = OpGroupBroadcast %[[#Int32Ty]] %[[#C2]] %[[#Val]] %[[#XY]]24; CHECK-SPIRV: %[[#XYZ:]] = OpCompositeConstruct %[[#Vec3Int32Ty]] %[[#X]] %[[#Y]] %[[#Z]]25; CHECK-SPIRV: %[[#]] = OpGroupBroadcast %[[#Int32Ty]] %[[#C2]] %[[#Val]] %[[#XYZ]]26define spir_kernel void @test_broadcast_xyz(i32 noundef %a, i32 noundef %x, i32 noundef %y, i32 noundef %z) {27entry:28 %call1 = call spir_func i32 @_Z20work_group_broadcastjj(i32 noundef %a, i32 noundef %x)29 %call2 = call spir_func i32 @_Z20work_group_broadcastjj(i32 noundef %a, i32 noundef %x, i32 noundef %y)30 %call3 = call spir_func i32 @_Z20work_group_broadcastjj(i32 noundef %a, i32 noundef %x, i32 noundef %y, i32 noundef %z)31 ret void32}33 34declare spir_func i32 @_Z20work_group_broadcastjj(i32, i32)35declare spir_func i32 @_Z20work_group_broadcastjjj(i32, i32, i32)36declare spir_func i32 @_Z20work_group_broadcastjjjj(i32, i32, i32, i32)37 38; CHECK-SPIRV: OpFunction39; CHECK-SPIRV: OpInBoundsPtrAccessChain40; CHECK-SPIRV: %[[#LoadedVal:]] = OpLoad %[[#Float32Ty]] %[[#]]41; CHECK-SPIRV: %[[#IdX:]] = OpCompositeExtract %[[#Int64Ty]] %[[#]] 042; CHECK-SPIRV: %[[#IdY:]] = OpCompositeExtract %[[#Int64Ty]] %[[#]] 143; CHECK-SPIRV: %[[#LocIdsVec:]] = OpCompositeConstruct %[[#Vec2Int64Ty]] %[[#IdX]] %[[#IdY]]44; CHECK-SPIRV: %[[#]] = OpGroupBroadcast %[[#Float32Ty]] %[[#C2]] %[[#LoadedVal]] %[[#LocIdsVec]]45define spir_kernel void @test_wg_broadcast_2D(ptr addrspace(1) %input, ptr addrspace(1) %output) #0 !kernel_arg_addr_space !7 !kernel_arg_access_qual !8 !kernel_arg_type !9 !kernel_arg_type_qual !10 !kernel_arg_base_type !9 !spirv.ParameterDecorations !11 {46entry:47 %0 = call spir_func i64 @_Z13get_global_idj(i32 0) #148 %1 = insertelement <3 x i64> undef, i64 %0, i32 049 %2 = call spir_func i64 @_Z13get_global_idj(i32 1) #150 %3 = insertelement <3 x i64> %1, i64 %2, i32 151 %4 = call spir_func i64 @_Z13get_global_idj(i32 2) #152 %5 = insertelement <3 x i64> %3, i64 %4, i32 253 %call = extractelement <3 x i64> %5, i32 054 %6 = call spir_func i64 @_Z13get_global_idj(i32 0) #155 %7 = insertelement <3 x i64> undef, i64 %6, i32 056 %8 = call spir_func i64 @_Z13get_global_idj(i32 1) #157 %9 = insertelement <3 x i64> %7, i64 %8, i32 158 %10 = call spir_func i64 @_Z13get_global_idj(i32 2) #159 %11 = insertelement <3 x i64> %9, i64 %10, i32 260 %call1 = extractelement <3 x i64> %11, i32 161 %12 = call spir_func i64 @_Z12get_group_idj(i32 0) #162 %13 = insertelement <3 x i64> undef, i64 %12, i32 063 %14 = call spir_func i64 @_Z12get_group_idj(i32 1) #164 %15 = insertelement <3 x i64> %13, i64 %14, i32 165 %16 = call spir_func i64 @_Z12get_group_idj(i32 2) #166 %17 = insertelement <3 x i64> %15, i64 %16, i32 267 %call2 = extractelement <3 x i64> %17, i32 068 %18 = call spir_func i64 @_Z14get_local_sizej(i32 0) #169 %19 = insertelement <3 x i64> undef, i64 %18, i32 070 %20 = call spir_func i64 @_Z14get_local_sizej(i32 1) #171 %21 = insertelement <3 x i64> %19, i64 %20, i32 172 %22 = call spir_func i64 @_Z14get_local_sizej(i32 2) #173 %23 = insertelement <3 x i64> %21, i64 %22, i32 274 %call3 = extractelement <3 x i64> %23, i32 075 %rem = urem i64 %call2, %call376 %24 = call spir_func i64 @_Z12get_group_idj(i32 0) #177 %25 = insertelement <3 x i64> undef, i64 %24, i32 078 %26 = call spir_func i64 @_Z12get_group_idj(i32 1) #179 %27 = insertelement <3 x i64> %25, i64 %26, i32 180 %28 = call spir_func i64 @_Z12get_group_idj(i32 2) #181 %29 = insertelement <3 x i64> %27, i64 %28, i32 282 %call4 = extractelement <3 x i64> %29, i32 183 %30 = call spir_func i64 @_Z14get_local_sizej(i32 0) #184 %31 = insertelement <3 x i64> undef, i64 %30, i32 085 %32 = call spir_func i64 @_Z14get_local_sizej(i32 1) #186 %33 = insertelement <3 x i64> %31, i64 %32, i32 187 %34 = call spir_func i64 @_Z14get_local_sizej(i32 2) #188 %35 = insertelement <3 x i64> %33, i64 %34, i32 289 %call5 = extractelement <3 x i64> %35, i32 190 %rem6 = urem i64 %call4, %call591 %36 = call spir_func i64 @_Z15get_global_sizej(i32 0) #192 %37 = insertelement <3 x i64> undef, i64 %36, i32 093 %38 = call spir_func i64 @_Z15get_global_sizej(i32 1) #194 %39 = insertelement <3 x i64> %37, i64 %38, i32 195 %40 = call spir_func i64 @_Z15get_global_sizej(i32 2) #196 %41 = insertelement <3 x i64> %39, i64 %40, i32 297 %call7 = extractelement <3 x i64> %41, i32 098 %mul = mul i64 %call1, %call799 %add = add i64 %mul, %call100 %arrayidx = getelementptr inbounds float, ptr addrspace(1) %input, i64 %add101 %42 = load float, ptr addrspace(1) %arrayidx, align 4102 %.splatinsert = insertelement <2 x i64> undef, i64 %rem, i32 0103 %.splat = shufflevector <2 x i64> %.splatinsert, <2 x i64> undef, <2 x i32> zeroinitializer104 %43 = insertelement <2 x i64> %.splat, i64 %rem6, i32 1105 %44 = extractelement <2 x i64> %43, i32 0106 %45 = extractelement <2 x i64> %43, i32 1107 %call8 = call spir_func float @_Z20work_group_broadcastfmm(float %42, i64 %44, i64 %45) #2108 %arrayidx9 = getelementptr inbounds float, ptr addrspace(1) %output, i64 %add109 store float %call8, ptr addrspace(1) %arrayidx9, align 4110 ret void111}112 113; Function Attrs: nounwind willreturn memory(none)114declare spir_func i64 @_Z13get_global_idj(i32) #1115 116; Function Attrs: nounwind willreturn memory(none)117declare spir_func i64 @_Z12get_group_idj(i32) #1118 119; Function Attrs: nounwind willreturn memory(none)120declare spir_func i64 @_Z14get_local_sizej(i32) #1121 122; Function Attrs: nounwind willreturn memory(none)123declare spir_func i64 @_Z15get_global_sizej(i32) #1124 125; Function Attrs: convergent nounwind126declare spir_func float @_Z20work_group_broadcastfmm(float, i64, i64) #2127 128attributes #0 = { nounwind }129attributes #1 = { nounwind willreturn memory(none) }130attributes #2 = { convergent nounwind }131 132!spirv.MemoryModel = !{!0}133!opencl.enable.FP_CONTRACT = !{}134!spirv.Source = !{!1}135!opencl.spir.version = !{!2}136!opencl.ocl.version = !{!3}137!opencl.used.extensions = !{!4}138!opencl.used.optional.core.features = !{!5}139!spirv.Generator = !{!6}140 141!0 = !{i32 2, i32 2}142!1 = !{i32 3, i32 300000}143!2 = !{i32 2, i32 0}144!3 = !{i32 3, i32 0}145!4 = !{!"cl_khr_subgroups"}146!5 = !{}147!6 = !{i16 6, i16 14}148!7 = !{i32 1, i32 1}149!8 = !{!"none", !"none"}150!9 = !{!"float*", !"float*"}151!10 = !{!"", !""}152!11 = !{!5, !5}153