1 //===------- Utils.cpp - OpenMP device runtime utility functions -- C++ -*-===// 2 // 3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 4 // See https://llvm.org/LICENSE.txt for license information. 5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 6 // 7 //===----------------------------------------------------------------------===// 8 // 9 // 10 //===----------------------------------------------------------------------===// 11 12 #include "Utils.h" 13 14 #include "Interface.h" 15 #include "Mapping.h" 16 17 #pragma omp declare target 18 19 using namespace _OMP; 20 21 namespace _OMP { 22 /// Helper to keep code alive without introducing a performance penalty. 23 __attribute__((used, weak, optnone)) void keepAlive() { 24 __kmpc_get_hardware_thread_id_in_block(); 25 __kmpc_get_hardware_num_threads_in_block(); 26 __kmpc_barrier_simple_spmd(nullptr, 0); 27 } 28 } // namespace _OMP 29 30 namespace impl { 31 32 /// AMDGCN Implementation 33 /// 34 ///{ 35 #pragma omp begin declare variant match(device = {arch(amdgcn)}) 36 37 void Unpack(uint64_t Val, uint32_t *LowBits, uint32_t *HighBits) { 38 *LowBits = (uint32_t)(Val & UINT64_C(0x00000000FFFFFFFF)); 39 *HighBits = (uint32_t)((Val & UINT64_C(0xFFFFFFFF00000000)) >> 32); 40 } 41 42 uint64_t Pack(uint32_t LowBits, uint32_t HighBits) { 43 return (((uint64_t)HighBits) << 32) | (uint64_t)LowBits; 44 } 45 46 #pragma omp end declare variant 47 48 /// NVPTX Implementation 49 /// 50 ///{ 51 #pragma omp begin declare variant match( \ 52 device = {arch(nvptx, nvptx64)}, implementation = {extension(match_any)}) 53 54 void Unpack(uint64_t Val, uint32_t *LowBits, uint32_t *HighBits) { 55 uint32_t LowBitsLocal, HighBitsLocal; 56 asm("mov.b64 {%0,%1}, %2;" 57 : "=r"(LowBitsLocal), "=r"(HighBitsLocal) 58 : "l"(Val)); 59 *LowBits = LowBitsLocal; 60 *HighBits = HighBitsLocal; 61 } 62 63 uint64_t Pack(uint32_t LowBits, uint32_t HighBits) { 64 uint64_t Val; 65 asm("mov.b64 %0, {%1,%2};" : "=l"(Val) : "r"(LowBits), "r"(HighBits)); 66 return Val; 67 } 68 69 #pragma omp end declare variant 70 71 /// AMDGCN Implementation 72 /// 73 ///{ 74 #pragma omp begin declare variant match(device = {arch(amdgcn)}) 75 76 int32_t shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) { 77 int Width = mapping::getWarpSize(); 78 int Self = mapping::getgetThreadIdInWarp(); 79 int Index = SrcLane + (Self & ~(Width - 1)); 80 return __builtin_amdgcn_ds_bpermute(Index << 2, Var); 81 } 82 83 int32_t shuffleDown(uint64_t Mask, int32_t Var, uint32_t LaneDelta, 84 int32_t Width) { 85 int Self = mapping::getThreadIdInWarp(); 86 int Index = Self + LaneDelta; 87 Index = (int)(LaneDelta + (Self & (Width - 1))) >= Width ? Self : Index; 88 return __builtin_amdgcn_ds_bpermute(Index << 2, Var); 89 } 90 91 #pragma omp end declare variant 92 ///} 93 94 /// NVPTX Implementation 95 /// 96 ///{ 97 #pragma omp begin declare variant match( \ 98 device = {arch(nvptx, nvptx64)}, implementation = {extension(match_any)}) 99 100 int32_t shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) { 101 return __nvvm_shfl_sync_idx_i32(Mask, Var, SrcLane, 0x1f); 102 } 103 104 int32_t shuffleDown(uint64_t Mask, int32_t Var, uint32_t Delta, int32_t Width) { 105 int32_t T = ((mapping::getWarpSize() - Width) << 8) | 0x1f; 106 return __nvvm_shfl_sync_down_i32(Mask, Var, Delta, T); 107 } 108 109 #pragma omp end declare variant 110 } // namespace impl 111 112 uint64_t utils::pack(uint32_t LowBits, uint32_t HighBits) { 113 return impl::Pack(LowBits, HighBits); 114 } 115 116 void utils::unpack(uint64_t Val, uint32_t &LowBits, uint32_t &HighBits) { 117 impl::Unpack(Val, &LowBits, &HighBits); 118 } 119 120 int32_t utils::shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) { 121 return impl::shuffle(Mask, Var, SrcLane); 122 } 123 124 int32_t utils::shuffleDown(uint64_t Mask, int32_t Var, uint32_t Delta, 125 int32_t Width) { 126 return impl::shuffleDown(Mask, Var, Delta, Width); 127 } 128 129 extern "C" { 130 int32_t __kmpc_shuffle_int32(int32_t Val, int16_t Delta, int16_t SrcLane) { 131 return impl::shuffleDown(lanes::All, Val, Delta, SrcLane); 132 } 133 134 int64_t __kmpc_shuffle_int64(int64_t Val, int16_t Delta, int16_t Width) { 135 uint32_t lo, hi; 136 utils::unpack(Val, lo, hi); 137 hi = impl::shuffleDown(lanes::All, hi, Delta, Width); 138 lo = impl::shuffleDown(lanes::All, lo, Delta, Width); 139 return utils::pack(lo, hi); 140 } 141 } 142 143 #pragma omp end declare target 144