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_barrier_simple_spmd(nullptr, 0);
26 }
27 } // namespace _OMP
28 
29 namespace impl {
30 
31 /// AMDGCN Implementation
32 ///
33 ///{
34 #pragma omp begin declare variant match(device = {arch(amdgcn)})
35 
36 void Unpack(uint64_t Val, uint32_t *LowBits, uint32_t *HighBits) {
37   *LowBits = (uint32_t)(Val & UINT64_C(0x00000000FFFFFFFF));
38   *HighBits = (uint32_t)((Val & UINT64_C(0xFFFFFFFF00000000)) >> 32);
39 }
40 
41 uint64_t Pack(uint32_t LowBits, uint32_t HighBits) {
42   return (((uint64_t)HighBits) << 32) | (uint64_t)LowBits;
43 }
44 
45 #pragma omp end declare variant
46 
47 /// NVPTX Implementation
48 ///
49 ///{
50 #pragma omp begin declare variant match(                                       \
51     device = {arch(nvptx, nvptx64)}, implementation = {extension(match_any)})
52 
53 void Unpack(uint64_t Val, uint32_t *LowBits, uint32_t *HighBits) {
54   uint32_t LowBitsLocal, HighBitsLocal;
55   asm("mov.b64 {%0,%1}, %2;"
56       : "=r"(LowBitsLocal), "=r"(HighBitsLocal)
57       : "l"(Val));
58   *LowBits = LowBitsLocal;
59   *HighBits = HighBitsLocal;
60 }
61 
62 uint64_t Pack(uint32_t LowBits, uint32_t HighBits) {
63   uint64_t Val;
64   asm("mov.b64 %0, {%1,%2};" : "=l"(Val) : "r"(LowBits), "r"(HighBits));
65   return Val;
66 }
67 
68 #pragma omp end declare variant
69 
70 /// AMDGCN Implementation
71 ///
72 ///{
73 #pragma omp begin declare variant match(device = {arch(amdgcn)})
74 
75 int32_t shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) {
76   int Width = mapping::getWarpSize();
77   int Self = mapping::getgetThreadIdInWarp();
78   int Index = SrcLane + (Self & ~(Width - 1));
79   return __builtin_amdgcn_ds_bpermute(Index << 2, Var);
80 }
81 
82 int32_t shuffleDown(uint64_t Mask, int32_t Var, uint32_t LaneDelta,
83                     int32_t Width) {
84   int Self = mapping::getThreadIdInWarp();
85   int Index = Self + LaneDelta;
86   Index = (int)(LaneDelta + (Self & (Width - 1))) >= Width ? Self : Index;
87   return __builtin_amdgcn_ds_bpermute(Index << 2, Var);
88 }
89 
90 #pragma omp end declare variant
91 ///}
92 
93 /// NVPTX Implementation
94 ///
95 ///{
96 #pragma omp begin declare variant match(                                       \
97     device = {arch(nvptx, nvptx64)}, implementation = {extension(match_any)})
98 
99 int32_t shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) {
100   return __nvvm_shfl_sync_idx_i32(Mask, Var, SrcLane, 0x1f);
101 }
102 
103 int32_t shuffleDown(uint64_t Mask, int32_t Var, uint32_t Delta, int32_t Width) {
104   int32_t T = ((mapping::getWarpSize() - Width) << 8) | 0x1f;
105   return __nvvm_shfl_sync_down_i32(Mask, Var, Delta, T);
106 }
107 
108 #pragma omp end declare variant
109 } // namespace impl
110 
111 uint64_t utils::pack(uint32_t LowBits, uint32_t HighBits) {
112   return impl::Pack(LowBits, HighBits);
113 }
114 
115 void utils::unpack(uint64_t Val, uint32_t &LowBits, uint32_t &HighBits) {
116   impl::Unpack(Val, &LowBits, &HighBits);
117 }
118 
119 int32_t utils::shuffle(uint64_t Mask, int32_t Var, int32_t SrcLane) {
120   return impl::shuffle(Mask, Var, SrcLane);
121 }
122 
123 int32_t utils::shuffleDown(uint64_t Mask, int32_t Var, uint32_t Delta,
124                            int32_t Width) {
125   return impl::shuffleDown(Mask, Var, Delta, Width);
126 }
127 
128 extern "C" {
129 int32_t __kmpc_shuffle_int32(int32_t Val, int16_t Delta, int16_t SrcLane) {
130   return impl::shuffleDown(lanes::All, Val, Delta, SrcLane);
131 }
132 
133 int64_t __kmpc_shuffle_int64(int64_t Val, int16_t Delta, int16_t Width) {
134   uint32_t lo, hi;
135   utils::unpack(Val, lo, hi);
136   hi = impl::shuffleDown(lanes::All, hi, Delta, Width);
137   lo = impl::shuffleDown(lanes::All, lo, Delta, Width);
138   return utils::pack(lo, hi);
139 }
140 }
141 
142 #pragma omp end declare target
143