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
|
/*
* Copyright (c) 2003 Matteo Frigo
* Copyright (c) 2003 Massachusetts Institute of Technology
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*
*/
/* $Id: vrank3-transpose.c,v 1.1 2008/10/17 06:11:08 scuri Exp $ */
/* rank-0, vector-rank-3, square transposition */
#include "dft.h"
/* transposition routine. TODO: optimize? */
static void t(R *rA, R *iA, int n, int is, int js, int vn, int vs)
{
int i, j, iv;
int im = iA - rA;
for (i = 1; i < n; ++i) {
for (j = 0; j < i; ++j) {
R *p0 = rA + i * is + j * js;
R *p1 = rA + j * is + i * js;
for (iv = 0; iv < vn; ++iv) {
R ar = p0[0], ai = p0[im];
R br = p1[0], bi = p1[im];
p1[0] = ar; p1[im] = ai; p1 += vs;
p0[0] = br; p0[im] = bi; p0 += vs;
}
}
}
}
typedef solver S;
typedef struct {
plan_dft super;
int n, vl;
int s0, s1, vs;
int m;
int offset;
int nd, md, d; /* d = gcd(n,m), nd = n / d, md = m / d */
} P;
static void apply(const plan *ego_, R *ri, R *ii, R *ro, R *io)
{
const P *ego = (const P *) ego_;
UNUSED(ro);
UNUSED(io);
A(ego->n == ego->m);
t(ri, ii, ego->n, ego->s0, ego->s1, ego->vl, ego->vs);
}
static void apply_general(const plan *ego_, R *ri, R *ii, R *ro, R *io)
{
const P *ego = (const P *) ego_;
int nd = ego->nd, md = ego->md, d = ego->d, vl = ego->vl;
R *buf = (R *)MALLOC((sizeof(R) * 2) * vl * nd * md * d, BUFFERS);
UNUSED(ii); UNUSED(ro); UNUSED(io);
X(transpose)(ri + ego->offset, nd, md, d, 2*vl, buf);
X(ifree)(buf);
}
static void apply_slow(const plan *ego_, R *ri, R *ii, R *ro, R *io)
{
const P *ego = (const P *) ego_;
int n = ego->n, m = ego->m, vl = ego->vl;
R *buf = (R *)MALLOC((sizeof(R) * 4) * vl, BUFFERS);
int move_size = (n + m) / 2;
char *move;
UNUSED(ii); UNUSED(ro); UNUSED(io);
STACK_MALLOC(char *, move, move_size);
X(transpose_slow)(ri + ego->offset, n, m, 2*vl, move, move_size, buf);
STACK_FREE(move);
X(ifree)(buf);
}
static int pickdim(const tensor *s, int *pdim0, int *pdim1, int *pdim2,
R *ri, R *ii)
{
int dim0, dim1;
for (dim0 = 0; dim0 < s->rnk; ++dim0)
for (dim1 = dim0 + 1; dim1 < s->rnk; ++dim1) {
int dim2 = 3 - dim0 - dim1;
if (s->dims[dim2].is == s->dims[dim2].os
&& X(transposable)(s->dims + dim0, s->dims + dim1,
s->dims[dim2].n, s->dims[dim2].is,
ri, ii)) {
*pdim0 = dim0;
*pdim1 = dim1;
*pdim2 = dim2;
return 1;
}
}
return 0;
}
static int applicable0(const problem *p_, int *dim0, int *dim1, int *dim2)
{
if (DFTP(p_)) {
const problem_dft *p = (const problem_dft *)p_;
return (1
&& p->ri == p->ro
&& p->sz->rnk == 0
&& p->vecsz->rnk == 3
&& pickdim(p->vecsz, dim0, dim1, dim2, p->ri, p->ii)
);
}
return 0;
}
static int applicable(const problem *p_, const planner *plnr,
int *dim0, int *dim1, int *dim2)
{
const problem_dft *p;
if (!applicable0(p_, dim0, dim1, dim2))
return 0;
p = (const problem_dft *) p_;
if (NO_UGLYP(plnr))
if (p->vecsz->dims[*dim2].is > X(imax)(p->vecsz->dims[*dim0].is,
p->vecsz->dims[*dim0].os))
/* loops are in the wrong order for locality */
return 0;
if (NO_UGLYP(plnr) && p->vecsz->dims[*dim0].n != p->vecsz->dims[*dim1].n)
return 0;
return 1;
}
static void print(const plan *ego_, printer *p)
{
const P *ego = (const P *) ego_;
p->print(p, "(dft-transpose-%dx%d%v)", ego->n, ego->m, ego->vl);
}
static plan *mkplan(const solver *ego, const problem *p_, planner *plnr)
{
const problem_dft *p;
P *pln;
const iodim *d;
int dim0, dim1, dim2;
int vl;
static const plan_adt padt = {
X(dft_solve), X(null_awake), print, X(plan_null_destroy)
};
UNUSED(plnr);
UNUSED(ego);
if (!applicable(p_, plnr, &dim0, &dim1, &dim2))
return (plan *) 0;
p = (const problem_dft *) p_;
d = p->vecsz->dims;
vl = d[dim2].n;
pln = MKPLAN_DFT(P, &padt,
X(transpose_simplep)(d+dim0, d+dim1,
vl, p->vecsz->dims[dim2].is,
p->ri, p->ii) ? apply :
(X(transpose_slowp)(d+dim0, d+dim1, 2*vl) ? apply_slow
: apply_general));
X(transpose_dims)(d+dim0, d+dim1,
&pln->n, &pln->m, &pln->d, &pln->nd, &pln->md);
pln->offset = (p->ri - p->ii == 1) ? -1 : 0;
pln->s0 = d[dim0].is;
pln->s1 = d[dim0].os;
pln->vl = vl;
pln->vs = d[dim2].is; /* == os */
/* pln->vl * (4 loads + 4 stores) * (pln->n \choose 2)
(FIXME? underestimate for non-square) */
X(ops_other)(4 * pln->vl * pln->n * (pln->m - 1), &pln->super.super.ops);
return &(pln->super.super);
}
static solver *mksolver(void)
{
static const solver_adt sadt = { mkplan };
return MKSOLVER(S, &sadt);
}
void X(dft_vrank3_transpose_register)(planner *p)
{
REGISTER_SOLVER(p, mksolver());
}
|
