1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
| //===---- omptarget-nvptxi.h - NVPTX OpenMP GPU initialization --- CUDA -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file contains the declarations of all library macros, types,
// and functions.
//
//===----------------------------------------------------------------------===//
////////////////////////////////////////////////////////////////////////////////
// Task Descriptor
////////////////////////////////////////////////////////////////////////////////
INLINE omp_sched_t omptarget_nvptx_TaskDescr::GetRuntimeSched() const {
// sched starts from 1..4; encode it as 0..3; so add 1 here
uint8_t rc = (items.flags & TaskDescr_SchedMask) + 1;
return (omp_sched_t)rc;
}
INLINE void omptarget_nvptx_TaskDescr::SetRuntimeSched(omp_sched_t sched) {
// sched starts from 1..4; encode it as 0..3; so sub 1 here
uint8_t val = ((uint8_t)sched) - 1;
// clear current sched
items.flags &= ~TaskDescr_SchedMask;
// set new sched
items.flags |= val;
}
INLINE void
omptarget_nvptx_TaskDescr::InitLevelZeroTaskDescr() {
// slow method
// flag:
// default sched is static,
// dyn is off (unused now anyway, but may need to sample from host ?)
// not in parallel
items.flags = 0;
items.threadId = 0; // is master
items.runtimeChunkSize = 1; // prefered chunking statik with chunk 1
}
// This is called when all threads are started together in SPMD mode.
// OMP directives include target parallel, target distribute parallel for, etc.
INLINE void omptarget_nvptx_TaskDescr::InitLevelOneTaskDescr(
omptarget_nvptx_TaskDescr *parentTaskDescr) {
// slow method
// flag:
// default sched is static,
// dyn is off (unused now anyway, but may need to sample from host ?)
// in L1 parallel
items.flags =
TaskDescr_InPar | TaskDescr_IsParConstr; // set flag to parallel
items.threadId =
GetThreadIdInBlock(); // get ids from cuda (only called for 1st level)
items.runtimeChunkSize = 1; // prefered chunking statik with chunk 1
prev = parentTaskDescr;
}
INLINE void omptarget_nvptx_TaskDescr::CopyData(
omptarget_nvptx_TaskDescr *sourceTaskDescr) {
items = sourceTaskDescr->items;
}
INLINE void
omptarget_nvptx_TaskDescr::Copy(omptarget_nvptx_TaskDescr *sourceTaskDescr) {
CopyData(sourceTaskDescr);
prev = sourceTaskDescr->prev;
}
INLINE void omptarget_nvptx_TaskDescr::CopyParent(
omptarget_nvptx_TaskDescr *parentTaskDescr) {
CopyData(parentTaskDescr);
prev = parentTaskDescr;
}
INLINE void omptarget_nvptx_TaskDescr::CopyForExplicitTask(
omptarget_nvptx_TaskDescr *parentTaskDescr) {
CopyParent(parentTaskDescr);
items.flags = items.flags & ~TaskDescr_IsParConstr;
ASSERT0(LT_FUSSY, IsTaskConstruct(), "expected task");
}
INLINE void omptarget_nvptx_TaskDescr::CopyToWorkDescr(
omptarget_nvptx_TaskDescr *masterTaskDescr) {
CopyParent(masterTaskDescr);
// overrwrite specific items;
items.flags |=
TaskDescr_InPar | TaskDescr_IsParConstr; // set flag to parallel
}
INLINE void omptarget_nvptx_TaskDescr::CopyFromWorkDescr(
omptarget_nvptx_TaskDescr *workTaskDescr) {
Copy(workTaskDescr);
//
// overrwrite specific items;
//
// The threadID should be GetThreadIdInBlock() % GetMasterThreadID().
// This is so that the serial master (first lane in the master warp)
// gets a threadId of 0.
// However, we know that this function is always called in a parallel
// region where only workers are active. The serial master thread
// never enters this region. When a parallel region is executed serially,
// the threadId is set to 0 elsewhere and the kmpc_serialized_* functions
// are called, which never activate this region.
items.threadId =
GetThreadIdInBlock(); // get ids from cuda (only called for 1st level)
}
INLINE void omptarget_nvptx_TaskDescr::CopyConvergentParent(
omptarget_nvptx_TaskDescr *parentTaskDescr, uint16_t tid, uint16_t tnum) {
CopyParent(parentTaskDescr);
items.flags |= TaskDescr_InParL2P; // In L2+ parallelism
items.threadId = tid;
}
INLINE void omptarget_nvptx_TaskDescr::SaveLoopData() {
loopData.loopUpperBound =
omptarget_nvptx_threadPrivateContext->LoopUpperBound(items.threadId);
loopData.nextLowerBound =
omptarget_nvptx_threadPrivateContext->NextLowerBound(items.threadId);
loopData.schedule =
omptarget_nvptx_threadPrivateContext->ScheduleType(items.threadId);
loopData.chunk = omptarget_nvptx_threadPrivateContext->Chunk(items.threadId);
loopData.stride =
omptarget_nvptx_threadPrivateContext->Stride(items.threadId);
}
INLINE void omptarget_nvptx_TaskDescr::RestoreLoopData() const {
omptarget_nvptx_threadPrivateContext->Chunk(items.threadId) = loopData.chunk;
omptarget_nvptx_threadPrivateContext->LoopUpperBound(items.threadId) =
loopData.loopUpperBound;
omptarget_nvptx_threadPrivateContext->NextLowerBound(items.threadId) =
loopData.nextLowerBound;
omptarget_nvptx_threadPrivateContext->Stride(items.threadId) =
loopData.stride;
omptarget_nvptx_threadPrivateContext->ScheduleType(items.threadId) =
loopData.schedule;
}
////////////////////////////////////////////////////////////////////////////////
// Thread Private Context
////////////////////////////////////////////////////////////////////////////////
INLINE omptarget_nvptx_TaskDescr *
omptarget_nvptx_ThreadPrivateContext::GetTopLevelTaskDescr(int tid) const {
ASSERT0(
LT_FUSSY, tid < MAX_THREADS_PER_TEAM,
"Getting top level, tid is larger than allocated data structure size");
return topTaskDescr[tid];
}
INLINE void
omptarget_nvptx_ThreadPrivateContext::InitThreadPrivateContext(int tid) {
// levelOneTaskDescr is init when starting the parallel region
// top task descr is NULL (team master version will be fixed separately)
topTaskDescr[tid] = NULL;
// no num threads value has been pushed
nextRegion.tnum[tid] = 0;
// the following don't need to be init here; they are init when using dyn
// sched
// current_Event, events_Number, chunk, num_Iterations, schedule
}
////////////////////////////////////////////////////////////////////////////////
// Team Descriptor
////////////////////////////////////////////////////////////////////////////////
INLINE void omptarget_nvptx_TeamDescr::InitTeamDescr() {
levelZeroTaskDescr.InitLevelZeroTaskDescr();
}
////////////////////////////////////////////////////////////////////////////////
// Get private data structure for thread
////////////////////////////////////////////////////////////////////////////////
// Utility routines for CUDA threads
INLINE omptarget_nvptx_TeamDescr &getMyTeamDescriptor() {
return omptarget_nvptx_threadPrivateContext->TeamContext();
}
INLINE omptarget_nvptx_WorkDescr &getMyWorkDescriptor() {
omptarget_nvptx_TeamDescr &currTeamDescr = getMyTeamDescriptor();
return currTeamDescr.WorkDescr();
}
INLINE omptarget_nvptx_TaskDescr *getMyTopTaskDescriptor(int threadId) {
return omptarget_nvptx_threadPrivateContext->GetTopLevelTaskDescr(threadId);
}
INLINE omptarget_nvptx_TaskDescr *
getMyTopTaskDescriptor(bool isSPMDExecutionMode) {
return getMyTopTaskDescriptor(GetLogicalThreadIdInBlock(isSPMDExecutionMode));
}
////////////////////////////////////////////////////////////////////////////////
// Memory management runtime functions.
////////////////////////////////////////////////////////////////////////////////
INLINE void omptarget_nvptx_SimpleMemoryManager::Release() {
ASSERT0(LT_FUSSY, usedSlotIdx < MAX_SM,
"SlotIdx is too big or uninitialized.");
ASSERT0(LT_FUSSY, usedMemIdx < OMP_STATE_COUNT,
"MemIdx is too big or uninitialized.");
MemDataTy &MD = MemData[usedSlotIdx];
atomicExch((unsigned *)&MD.keys[usedMemIdx], 0);
}
INLINE const void *omptarget_nvptx_SimpleMemoryManager::Acquire(const void *buf,
size_t size) {
ASSERT0(LT_FUSSY, usedSlotIdx < MAX_SM,
"SlotIdx is too big or uninitialized.");
const unsigned sm = usedSlotIdx;
MemDataTy &MD = MemData[sm];
unsigned i = hash(GetBlockIdInKernel());
while (atomicCAS((unsigned *)&MD.keys[i], 0, 1) != 0) {
i = hash(i + 1);
}
usedSlotIdx = sm;
usedMemIdx = i;
return static_cast<const char *>(buf) + (sm * OMP_STATE_COUNT + i) * size;
}
|