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
|
// Copyright 2012 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Parallel for algorithm.
#include "runtime.h"
#include "arch_GOARCH.h"
struct ParForThread
{
// the thread's iteration space [32lsb, 32msb)
uint64 pos;
// stats
uint64 nsteal;
uint64 nstealcnt;
uint64 nprocyield;
uint64 nosyield;
uint64 nsleep;
byte pad[CacheLineSize];
};
ParFor*
runtime·parforalloc(uint32 nthrmax)
{
ParFor *desc;
// The ParFor object is followed by CacheLineSize padding
// and then nthrmax ParForThread.
desc = (ParFor*)runtime·malloc(sizeof(ParFor) + CacheLineSize + nthrmax * sizeof(ParForThread));
desc->thr = (ParForThread*)((byte*)(desc+1) + CacheLineSize);
desc->nthrmax = nthrmax;
return desc;
}
// For testing from Go
// func parforalloc2(nthrmax uint32) *ParFor
void
runtime·parforalloc2(uint32 nthrmax, ParFor *desc)
{
desc = runtime·parforalloc(nthrmax);
FLUSH(&desc);
}
void
runtime·parforsetup(ParFor *desc, uint32 nthr, uint32 n, void *ctx, bool wait, void (*body)(ParFor*, uint32))
{
uint32 i, begin, end;
uint64 *pos;
if(desc == nil || nthr == 0 || nthr > desc->nthrmax || body == nil) {
runtime·printf("desc=%p nthr=%d count=%d body=%p\n", desc, nthr, n, body);
runtime·throw("parfor: invalid args");
}
desc->body = body;
desc->done = 0;
desc->nthr = nthr;
desc->thrseq = 0;
desc->cnt = n;
desc->ctx = ctx;
desc->wait = wait;
desc->nsteal = 0;
desc->nstealcnt = 0;
desc->nprocyield = 0;
desc->nosyield = 0;
desc->nsleep = 0;
for(i=0; i<nthr; i++) {
begin = (uint64)n*i / nthr;
end = (uint64)n*(i+1) / nthr;
pos = &desc->thr[i].pos;
if(((uintptr)pos & 7) != 0)
runtime·throw("parforsetup: pos is not aligned");
*pos = (uint64)begin | (((uint64)end)<<32);
}
}
// For testing from Go
// func parforsetup2(desc *ParFor, nthr, n uint32, ctx *byte, wait bool, body func(*ParFor, uint32))
void
runtime·parforsetup2(ParFor *desc, uint32 nthr, uint32 n, void *ctx, bool wait, void *body)
{
runtime·parforsetup(desc, nthr, n, ctx, wait, *(void(**)(ParFor*, uint32))body);
}
void
runtime·parfordo(ParFor *desc)
{
ParForThread *me;
uint32 tid, begin, end, begin2, try, victim, i;
uint64 *mypos, *victimpos, pos, newpos;
void (*body)(ParFor*, uint32);
bool idle;
// Obtain 0-based thread index.
tid = runtime·xadd(&desc->thrseq, 1) - 1;
if(tid >= desc->nthr) {
runtime·printf("tid=%d nthr=%d\n", tid, desc->nthr);
runtime·throw("parfor: invalid tid");
}
// If single-threaded, just execute the for serially.
if(desc->nthr==1) {
for(i=0; i<desc->cnt; i++)
desc->body(desc, i);
return;
}
body = desc->body;
me = &desc->thr[tid];
mypos = &me->pos;
for(;;) {
for(;;) {
// While there is local work,
// bump low index and execute the iteration.
pos = runtime·xadd64(mypos, 1);
begin = (uint32)pos-1;
end = (uint32)(pos>>32);
if(begin < end) {
body(desc, begin);
continue;
}
break;
}
// Out of work, need to steal something.
idle = false;
for(try=0;; try++) {
// If we don't see any work for long enough,
// increment the done counter...
if(try > desc->nthr*4 && !idle) {
idle = true;
runtime·xadd(&desc->done, 1);
}
// ...if all threads have incremented the counter,
// we are done.
if(desc->done + !idle == desc->nthr) {
if(!idle)
runtime·xadd(&desc->done, 1);
goto exit;
}
// Choose a random victim for stealing.
victim = runtime·fastrand1() % (desc->nthr-1);
if(victim >= tid)
victim++;
victimpos = &desc->thr[victim].pos;
for(;;) {
// See if it has any work.
pos = runtime·atomicload64(victimpos);
begin = (uint32)pos;
end = (uint32)(pos>>32);
if(begin+1 >= end) {
begin = end = 0;
break;
}
if(idle) {
runtime·xadd(&desc->done, -1);
idle = false;
}
begin2 = begin + (end-begin)/2;
newpos = (uint64)begin | (uint64)begin2<<32;
if(runtime·cas64(victimpos, pos, newpos)) {
begin = begin2;
break;
}
}
if(begin < end) {
// Has successfully stolen some work.
if(idle)
runtime·throw("parfor: should not be idle");
runtime·atomicstore64(mypos, (uint64)begin | (uint64)end<<32);
me->nsteal++;
me->nstealcnt += end-begin;
break;
}
// Backoff.
if(try < desc->nthr) {
// nothing
} else if (try < 4*desc->nthr) {
me->nprocyield++;
runtime·procyield(20);
// If a caller asked not to wait for the others, exit now
// (assume that most work is already done at this point).
} else if (!desc->wait) {
if(!idle)
runtime·xadd(&desc->done, 1);
goto exit;
} else if (try < 6*desc->nthr) {
me->nosyield++;
runtime·osyield();
} else {
me->nsleep++;
runtime·usleep(1);
}
}
}
exit:
runtime·xadd64(&desc->nsteal, me->nsteal);
runtime·xadd64(&desc->nstealcnt, me->nstealcnt);
runtime·xadd64(&desc->nprocyield, me->nprocyield);
runtime·xadd64(&desc->nosyield, me->nosyield);
runtime·xadd64(&desc->nsleep, me->nsleep);
me->nsteal = 0;
me->nstealcnt = 0;
me->nprocyield = 0;
me->nosyield = 0;
me->nsleep = 0;
}
// For testing from Go
// func parforiters(desc *ParFor, tid uintptr) (uintptr, uintptr)
void
runtime·parforiters(ParFor *desc, uintptr tid, uintptr start, uintptr end)
{
start = (uint32)desc->thr[tid].pos;
end = (uint32)(desc->thr[tid].pos>>32);
FLUSH(&start);
FLUSH(&end);
}
|
