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 "Debug.h"
15 #include "Interface.h"
16 #include "Mapping.h"
17 
18 #pragma omp begin declare target device_type(nohost)
19 
20 using namespace _OMP;
21 
22 namespace _OMP {
23 /// Helper to keep code alive without introducing a performance penalty.
24 __attribute__((weak, optnone, cold)) KEEP_ALIVE void keepAlive() {
25   __kmpc_get_hardware_thread_id_in_block();
26   __kmpc_get_hardware_num_threads_in_block();
27   __kmpc_get_warp_size();
28   __kmpc_barrier_simple_spmd(nullptr, 0);
29   __kmpc_barrier_simple_generic(nullptr, 0);
30 }
31 } // namespace _OMP
32 
33 namespace impl {
34 
35 void Unpack(uint64_t Val, uint32_t *LowBits, uint32_t *HighBits);
36 uint64_t Pack(uint32_t LowBits, uint32_t HighBits);
37 
38 /// AMDGCN Implementation
39 ///
40 ///{
41 #pragma omp begin declare variant match(device = {arch(amdgcn)})
42 
43 void Unpack(uint64_t Val, uint32_t *LowBits, uint32_t *HighBits) {
44   static_assert(sizeof(unsigned long) == 8, "");
45   *LowBits = (uint32_t)(Val & 0x00000000FFFFFFFFUL);
46   *HighBits = (uint32_t)((Val & 0xFFFFFFFF00000000UL) >> 32);
47 }
48 
49 uint64_t Pack(uint32_t LowBits, uint32_t HighBits) {
50   return (((uint64_t)HighBits) << 32) | (uint64_t)LowBits;
51 }
52 
53 #pragma omp end declare variant
54 
55 /// NVPTX Implementation
56 ///
57 ///{
58 #pragma omp begin declare variant match(                                       \
59     device = {arch(nvptx, nvptx64)}, implementation = {extension(match_any)})
60 
61 void Unpack(uint64_t Val, uint32_t *LowBits, uint32_t *HighBits) {
62   uint32_t LowBitsLocal, HighBitsLocal;
63   asm("mov.b64 {%0,%1}, %2;"
64       : "=r"(LowBitsLocal), "=r"(HighBitsLocal)
65       : "l"(Val));
66   *LowBits = LowBitsLocal;
67   *HighBits = HighBitsLocal;
68 }
69 
70 uint64_t Pack(uint32_t LowBits, uint32_t HighBits) {
71   uint64_t Val;
72   asm("mov.b64 %0, {%1,%2};" : "=l"(Val) : "r"(LowBits), "r"(HighBits));
73   return Val;
74 }
75 
76 #pragma omp end declare variant
77 
78 int32_t shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane);
79 int32_t shuffleDown(uint64_t Mask, int32_t Var, uint32_t LaneDelta,
80                     int32_t Width);
81 
82 /// AMDGCN Implementation
83 ///
84 ///{
85 #pragma omp begin declare variant match(device = {arch(amdgcn)})
86 
87 int32_t shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) {
88   int Width = mapping::getWarpSize();
89   int Self = mapping::getThreadIdInWarp();
90   int Index = SrcLane + (Self & ~(Width - 1));
91   return __builtin_amdgcn_ds_bpermute(Index << 2, Var);
92 }
93 
94 int32_t shuffleDown(uint64_t Mask, int32_t Var, uint32_t LaneDelta,
95                     int32_t Width) {
96   int Self = mapping::getThreadIdInWarp();
97   int Index = Self + LaneDelta;
98   Index = (int)(LaneDelta + (Self & (Width - 1))) >= Width ? Self : Index;
99   return __builtin_amdgcn_ds_bpermute(Index << 2, Var);
100 }
101 
102 #pragma omp end declare variant
103 ///}
104 
105 /// NVPTX Implementation
106 ///
107 ///{
108 #pragma omp begin declare variant match(                                       \
109     device = {arch(nvptx, nvptx64)}, implementation = {extension(match_any)})
110 
111 int32_t shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) {
112   return __nvvm_shfl_sync_idx_i32(Mask, Var, SrcLane, 0x1f);
113 }
114 
115 int32_t shuffleDown(uint64_t Mask, int32_t Var, uint32_t Delta, int32_t Width) {
116   int32_t T = ((mapping::getWarpSize() - Width) << 8) | 0x1f;
117   return __nvvm_shfl_sync_down_i32(Mask, Var, Delta, T);
118 }
119 
120 #pragma omp end declare variant
121 } // namespace impl
122 
123 uint64_t utils::pack(uint32_t LowBits, uint32_t HighBits) {
124   return impl::Pack(LowBits, HighBits);
125 }
126 
127 void utils::unpack(uint64_t Val, uint32_t &LowBits, uint32_t &HighBits) {
128   impl::Unpack(Val, &LowBits, &HighBits);
129 }
130 
131 int32_t utils::shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) {
132   return impl::shuffle(Mask, Var, SrcLane);
133 }
134 
135 int32_t utils::shuffleDown(uint64_t Mask, int32_t Var, uint32_t Delta,
136                            int32_t Width) {
137   return impl::shuffleDown(Mask, Var, Delta, Width);
138 }
139 
140 extern "C" {
141 int32_t __kmpc_shuffle_int32(int32_t Val, int16_t Delta, int16_t SrcLane) {
142   FunctionTracingRAII();
143   return impl::shuffleDown(lanes::All, Val, Delta, SrcLane);
144 }
145 
146 int64_t __kmpc_shuffle_int64(int64_t Val, int16_t Delta, int16_t Width) {
147   FunctionTracingRAII();
148   uint32_t lo, hi;
149   utils::unpack(Val, lo, hi);
150   hi = impl::shuffleDown(lanes::All, hi, Delta, Width);
151   lo = impl::shuffleDown(lanes::All, lo, Delta, Width);
152   return utils::pack(lo, hi);
153 }
154 }
155 
156 #pragma omp end declare target
157