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
227
228
229
230
231
232
233
234
|
/*
-- MAGMA (version 2.9.0) --
Univ. of Tennessee, Knoxville
Univ. of California, Berkeley
Univ. of Colorado, Denver
@date January 2025
@author Mark Gates
*/
#include "magma_internal.h"
/***************************************************************************//**
Returns version of MAGMA, as defined by
MAGMA_VERSION_MAJOR, MAGMA_VERSION_MINOR, MAGMA_VERSION_MICRO constants.
@param[out] major Set to major version number.
@param[out] minor Set to minor version number.
@param[out] micro Set to micro version number.
@ingroup magma_util
*******************************************************************************/
extern "C"
void magma_version( magma_int_t* major, magma_int_t* minor, magma_int_t* micro )
{
if ( major != NULL && minor != NULL && micro != NULL ) {
*major = MAGMA_VERSION_MAJOR;
*minor = MAGMA_VERSION_MINOR;
*micro = MAGMA_VERSION_MICRO;
}
}
/***************************************************************************//**
Determines the number of GPUs to use, based on $MAGMA_NUM_GPUS environment
variable, and limited to actual number of GPUs available.
If $MAGMA_NUM_GPUS is not set, uses 1.
@return Number of GPUs to use.
@ingroup magma_util
*******************************************************************************/
extern "C"
magma_int_t magma_num_gpus( void )
{
const char *ngpu_str = getenv("MAGMA_NUM_GPUS");
magma_int_t ngpu = 1;
if ( ngpu_str != NULL ) {
char* endptr;
ngpu = strtol( ngpu_str, &endptr, 10 );
magma_int_t ndevices;
magma_device_t devices[ MagmaMaxGPUs ];
magma_getdevices( devices, MagmaMaxGPUs, &ndevices );
// if *endptr == '\0' then entire string was valid number (or empty)
if ( ngpu < 1 || *endptr != '\0' ) {
ngpu = 1;
fprintf( stderr, "$MAGMA_NUM_GPUS='%s' is an invalid number; using %lld GPU.\n",
ngpu_str, (long long) ngpu );
}
else if ( ngpu > MagmaMaxGPUs || ngpu > ndevices ) {
ngpu = min( ndevices, MagmaMaxGPUs );
fprintf( stderr, "$MAGMA_NUM_GPUS='%s' exceeds MagmaMaxGPUs=%d or available GPUs=%lld; using %lld GPUs.\n",
ngpu_str, MagmaMaxGPUs, (long long) ndevices, (long long) ngpu );
}
assert( 1 <= ngpu && ngpu <= ndevices );
}
return ngpu;
}
/***************************************************************************//**
Auxiliary function: ipiv(i) indicates that row i has been swapped with
ipiv(i) from top to bottom. This function rearranges ipiv into newipiv
where row i has to be moved to newipiv(i). The new pivoting allows for
parallel processing vs the original one assumes a specific ordering and
has to be done sequentially.
@ingroup magma_internal
*******************************************************************************/
extern "C"
void magma_swp2pswp( magma_trans_t trans, magma_int_t n, magma_int_t *ipiv, magma_int_t *newipiv)
{
magma_int_t i, newind, ind;
for (i=0; i < n; i++)
newipiv[i] = -1;
if (trans == MagmaNoTrans) {
for (i=0; i < n; i++) {
newind = ipiv[i] - 1;
if (newipiv[newind] == -1) {
if (newipiv[i] == -1) {
newipiv[i] = newind;
if (newind > i)
newipiv[newind]= i;
}
else {
ind = newipiv[i];
newipiv[i] = newind;
if (newind > i)
newipiv[newind]= ind;
}
}
else {
if (newipiv[i] == -1) {
if (newind > i) {
ind = newipiv[newind];
newipiv[newind] = i;
newipiv[i] = ind;
}
else
newipiv[i] = newipiv[newind];
}
else {
ind = newipiv[i];
newipiv[i] = newipiv[newind];
if (newind > i)
newipiv[newind] = ind;
}
}
}
}
else {
// Transpose
for (i=n-1; i >= 0; i--) {
newind = ipiv[i] - 1;
if (newipiv[newind] == -1) {
if (newipiv[i] == -1) {
newipiv[i] = newind;
if (newind > i)
newipiv[newind]= i;
}
else {
ind = newipiv[i];
newipiv[i] = newind;
if (newind > i)
newipiv[newind]= ind;
}
}
else {
if (newipiv[i] == -1) {
if (newind > i) {
ind = newipiv[newind];
newipiv[newind] = i;
newipiv[i] = ind;
}
else
newipiv[i] = newipiv[newind];
}
else {
ind = newipiv[i];
newipiv[i] = newipiv[newind];
if (newind > i)
newipiv[newind] = ind;
}
}
}
}
}
/***************************************************************************//**
Convert global indices [j0, j1) to local indices [dj0, dj1) on GPU dev,
according to 1D block cyclic distribution.
Note j0 and dj0 are inclusive, while j1 and dj1 are exclusive.
This is consistent with the C++ container notion of first and last.
Example with n = 75, nb = 10, ngpu = 3.
Distribution of columns (ranges are inclusive):
local dj: 0- 9, 10-19, 20-29
-----------------------------------------------
dev 0: 3 blocks, global j: 0- 9, 30-39, 60-69
dev 1: 3 blocks, global j: 10-19, 40-49, 70-74 (partial)
dev 2: 2 block, global j: 20-29, 50-59
Calls return:
input global j=13-68 inclusive => output
nb=10, ngpu=3, dev=0, j0=13, j1=69 => dj0=10, dj1=29 (i.e., global j= 30-39, 60-68)
nb=10, ngpu=3, dev=1, j0=13, j1=69 => dj0= 3, dj1=20 (i.e., global j=13-19, 40-49)
nb=10, ngpu=3, dev=2, j0=13, j1=69 => dj0= 0, dj1=20 (i.e., global j=20-29, 50-59)
input global j=13-69 inclusive => output
nb=10, ngpu=3, dev=0, j0=13, j1=70 => dj0=10, dj1=30 (i.e., global j= 30-39, 60-69)
nb=10, ngpu=3, dev=1, j0=13, j1=70 => dj0= 3, dj1=20 (i.e., global j=13-19, 40-49)
nb=10, ngpu=3, dev=2, j0=13, j1=70 => dj0= 0, dj1=20 (i.e., global j=20-29, 50-59)
input global j=13-70 inclusive => output
nb=10, ngpu=3, dev=0, j0=13, j1=71 => dj0=10, dj1=30 (i.e., global j= 30-39, 60-69)
nb=10, ngpu=3, dev=1, j0=13, j1=71 => dj0= 3, dj1=21 (i.e., global j=13-19, 40-49, 70)
nb=10, ngpu=3, dev=2, j0=13, j1=71 => dj0= 0, dj1=20 (i.e., global j=20-29, 50-59)
input global j=13-71 inclusive => output
nb=10, ngpu=3, dev=0, j0=13, j1=72 => dj0=10, dj1=30 (i.e., global j= 30-39, 60-69)
nb=10, ngpu=3, dev=1, j0=13, j1=72 => dj0= 3, dj1=22 (i.e., global j=13-19, 40-49, 70-71)
nb=10, ngpu=3, dev=2, j0=13, j1=72 => dj0= 0, dj1=20 (i.e., global j=20-29, 50-59)
@ingroup magma_internal
*******************************************************************************/
extern "C"
void magma_indices_1D_bcyclic( magma_int_t nb, magma_int_t ngpu, magma_int_t dev,
magma_int_t j0, magma_int_t j1,
magma_int_t* dj0, magma_int_t* dj1 )
{
// on GPU jdev, which contains j0, dj0 maps to j0.
// on other GPUs, dj0 is start of the block on that GPU after j0's block.
magma_int_t jblock = (j0 / nb) / ngpu;
magma_int_t jdev = (j0 / nb) % ngpu;
if ( dev < jdev ) {
jblock += 1;
}
*dj0 = jblock*nb;
if ( dev == jdev ) {
*dj0 += (j0 % nb);
}
// on GPU jdev, which contains j1-1, dj1 maps to j1.
// on other GPUs, dj1 is end of the block on that GPU before j1's block.
// j1 points to element after end (e.g., n), so subtract 1 to get last
// element, compute index, then add 1 to get just after that index again.
j1 -= 1;
jblock = (j1 / nb) / ngpu;
jdev = (j1 / nb) % ngpu;
if ( dev > jdev ) {
jblock -= 1;
}
if ( dev == jdev ) {
*dj1 = jblock*nb + (j1 % nb) + 1;
}
else {
*dj1 = jblock*nb + nb;
}
}
|
