Mercurial > hg > sv-dependency-builds

annotate src/fftw-3.3.5/kernel/cpy2d.c @ 84:08ae793730bd

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Add null config files
author Chris Cannam
date Mon, 02 Mar 2020 14:03:47 +0000
parents 2cd0e3b3e1fd
children
rev   line source
Chris@42 1 /*
Chris@42 2 * Copyright (c) 2003, 2007-14 Matteo Frigo
Chris@42 3 * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
Chris@42 4 *
Chris@42 5 * This program is free software; you can redistribute it and/or modify
Chris@42 6 * it under the terms of the GNU General Public License as published by
Chris@42 7 * the Free Software Foundation; either version 2 of the License, or
Chris@42 8 * (at your option) any later version.
Chris@42 9 *
Chris@42 10 * This program is distributed in the hope that it will be useful,
Chris@42 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@42 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
Chris@42 13 * GNU General Public License for more details.
Chris@42 14 *
Chris@42 15 * You should have received a copy of the GNU General Public License
Chris@42 16 * along with this program; if not, write to the Free Software
Chris@42 17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
Chris@42 18 *
Chris@42 19 */
Chris@42 20
Chris@42 21 /* out of place 2D copy routines */
Chris@42 22 #include "ifftw.h"
Chris@42 23
Chris@42 24 #if defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64)
Chris@42 25 # ifdef HAVE_XMMINTRIN_H
Chris@42 26 # include <xmmintrin.h>
Chris@42 27 # define WIDE_TYPE __m128
Chris@42 28 # endif
Chris@42 29 #endif
Chris@42 30
Chris@42 31 #ifndef WIDE_TYPE
Chris@42 32 /* fall back to double, which means that WIDE_TYPE will be unused */
Chris@42 33 # define WIDE_TYPE double
Chris@42 34 #endif
Chris@42 35
Chris@42 36 void X(cpy2d)(R *I, R *O,
Chris@42 37 INT n0, INT is0, INT os0,
Chris@42 38 INT n1, INT is1, INT os1,
Chris@42 39 INT vl)
Chris@42 40 {
Chris@42 41 INT i0, i1, v;
Chris@42 42
Chris@42 43 switch (vl) {
Chris@42 44 case 1:
Chris@42 45 for (i1 = 0; i1 < n1; ++i1)
Chris@42 46 for (i0 = 0; i0 < n0; ++i0) {
Chris@42 47 R x0 = I[i0 * is0 + i1 * is1];
Chris@42 48 O[i0 * os0 + i1 * os1] = x0;
Chris@42 49 }
Chris@42 50 break;
Chris@42 51 case 2:
Chris@42 52 if (1
Chris@42 53 && (2 * sizeof(R) == sizeof(WIDE_TYPE))
Chris@42 54 && (sizeof(WIDE_TYPE) > sizeof(double))
Chris@42 55 && (((size_t)I) % sizeof(WIDE_TYPE) == 0)
Chris@42 56 && (((size_t)O) % sizeof(WIDE_TYPE) == 0)
Chris@42 57 && ((is0 & 1) == 0)
Chris@42 58 && ((is1 & 1) == 0)
Chris@42 59 && ((os0 & 1) == 0)
Chris@42 60 && ((os1 & 1) == 0)) {
Chris@42 61 /* copy R[2] as WIDE_TYPE if WIDE_TYPE is large
Chris@42 62 enough to hold R[2], and if the input is
Chris@42 63 properly aligned. This is a win when R==double
Chris@42 64 and WIDE_TYPE is 128 bits. */
Chris@42 65 for (i1 = 0; i1 < n1; ++i1)
Chris@42 66 for (i0 = 0; i0 < n0; ++i0) {
Chris@42 67 *(WIDE_TYPE *)&O[i0 * os0 + i1 * os1] =
Chris@42 68 *(WIDE_TYPE *)&I[i0 * is0 + i1 * is1];
Chris@42 69 }
Chris@42 70 } else if (1
Chris@42 71 && (2 * sizeof(R) == sizeof(double))
Chris@42 72 && (((size_t)I) % sizeof(double) == 0)
Chris@42 73 && (((size_t)O) % sizeof(double) == 0)
Chris@42 74 && ((is0 & 1) == 0)
Chris@42 75 && ((is1 & 1) == 0)
Chris@42 76 && ((os0 & 1) == 0)
Chris@42 77 && ((os1 & 1) == 0)) {
Chris@42 78 /* copy R[2] as double if double is large enough to
Chris@42 79 hold R[2], and if the input is properly aligned.
Chris@42 80 This case applies when R==float */
Chris@42 81 for (i1 = 0; i1 < n1; ++i1)
Chris@42 82 for (i0 = 0; i0 < n0; ++i0) {
Chris@42 83 *(double *)&O[i0 * os0 + i1 * os1] =
Chris@42 84 *(double *)&I[i0 * is0 + i1 * is1];
Chris@42 85 }
Chris@42 86 } else {
Chris@42 87 for (i1 = 0; i1 < n1; ++i1)
Chris@42 88 for (i0 = 0; i0 < n0; ++i0) {
Chris@42 89 R x0 = I[i0 * is0 + i1 * is1];
Chris@42 90 R x1 = I[i0 * is0 + i1 * is1 + 1];
Chris@42 91 O[i0 * os0 + i1 * os1] = x0;
Chris@42 92 O[i0 * os0 + i1 * os1 + 1] = x1;
Chris@42 93 }
Chris@42 94 }
Chris@42 95 break;
Chris@42 96 default:
Chris@42 97 for (i1 = 0; i1 < n1; ++i1)
Chris@42 98 for (i0 = 0; i0 < n0; ++i0)
Chris@42 99 for (v = 0; v < vl; ++v) {
Chris@42 100 R x0 = I[i0 * is0 + i1 * is1 + v];
Chris@42 101 O[i0 * os0 + i1 * os1 + v] = x0;
Chris@42 102 }
Chris@42 103 break;
Chris@42 104 }
Chris@42 105 }
Chris@42 106
Chris@42 107 /* like cpy2d, but read input contiguously if possible */
Chris@42 108 void X(cpy2d_ci)(R *I, R *O,
Chris@42 109 INT n0, INT is0, INT os0,
Chris@42 110 INT n1, INT is1, INT os1,
Chris@42 111 INT vl)
Chris@42 112 {
Chris@42 113 if (IABS(is0) < IABS(is1)) /* inner loop is for n0 */
Chris@42 114 X(cpy2d) (I, O, n0, is0, os0, n1, is1, os1, vl);
Chris@42 115 else
Chris@42 116 X(cpy2d) (I, O, n1, is1, os1, n0, is0, os0, vl);
Chris@42 117 }
Chris@42 118
Chris@42 119 /* like cpy2d, but write output contiguously if possible */
Chris@42 120 void X(cpy2d_co)(R *I, R *O,
Chris@42 121 INT n0, INT is0, INT os0,
Chris@42 122 INT n1, INT is1, INT os1,
Chris@42 123 INT vl)
Chris@42 124 {
Chris@42 125 if (IABS(os0) < IABS(os1)) /* inner loop is for n0 */
Chris@42 126 X(cpy2d) (I, O, n0, is0, os0, n1, is1, os1, vl);
Chris@42 127 else
Chris@42 128 X(cpy2d) (I, O, n1, is1, os1, n0, is0, os0, vl);
Chris@42 129 }
Chris@42 130
Chris@42 131
Chris@42 132 /* tiled copy routines */
Chris@42 133 struct cpy2d_closure {
Chris@42 134 R *I, *O;
Chris@42 135 INT is0, os0, is1, os1, vl;
Chris@42 136 R *buf;
Chris@42 137 };
Chris@42 138
Chris@42 139 static void dotile(INT n0l, INT n0u, INT n1l, INT n1u, void *args)
Chris@42 140 {
Chris@42 141 struct cpy2d_closure *k = (struct cpy2d_closure *)args;
Chris@42 142 X(cpy2d)(k->I + n0l * k->is0 + n1l * k->is1,
Chris@42 143 k->O + n0l * k->os0 + n1l * k->os1,
Chris@42 144 n0u - n0l, k->is0, k->os0,
Chris@42 145 n1u - n1l, k->is1, k->os1,
Chris@42 146 k->vl);
Chris@42 147 }
Chris@42 148
Chris@42 149 static void dotile_buf(INT n0l, INT n0u, INT n1l, INT n1u, void *args)
Chris@42 150 {
Chris@42 151 struct cpy2d_closure *k = (struct cpy2d_closure *)args;
Chris@42 152
Chris@42 153 /* copy from I to buf */
Chris@42 154 X(cpy2d_ci)(k->I + n0l * k->is0 + n1l * k->is1,
Chris@42 155 k->buf,
Chris@42 156 n0u - n0l, k->is0, k->vl,
Chris@42 157 n1u - n1l, k->is1, k->vl * (n0u - n0l),
Chris@42 158 k->vl);
Chris@42 159
Chris@42 160 /* copy from buf to O */
Chris@42 161 X(cpy2d_co)(k->buf,
Chris@42 162 k->O + n0l * k->os0 + n1l * k->os1,
Chris@42 163 n0u - n0l, k->vl, k->os0,
Chris@42 164 n1u - n1l, k->vl * (n0u - n0l), k->os1,
Chris@42 165 k->vl);
Chris@42 166 }
Chris@42 167
Chris@42 168
Chris@42 169 void X(cpy2d_tiled)(R *I, R *O,
Chris@42 170 INT n0, INT is0, INT os0,
Chris@42 171 INT n1, INT is1, INT os1, INT vl)
Chris@42 172 {
Chris@42 173 INT tilesz = X(compute_tilesz)(vl,
Chris@42 174 1 /* input array */
Chris@42 175 + 1 /* ouput array */);
Chris@42 176 struct cpy2d_closure k;
Chris@42 177 k.I = I;
Chris@42 178 k.O = O;
Chris@42 179 k.is0 = is0;
Chris@42 180 k.os0 = os0;
Chris@42 181 k.is1 = is1;
Chris@42 182 k.os1 = os1;
Chris@42 183 k.vl = vl;
Chris@42 184 k.buf = 0; /* unused */
Chris@42 185 X(tile2d)(0, n0, 0, n1, tilesz, dotile, &k);
Chris@42 186 }
Chris@42 187
Chris@42 188 void X(cpy2d_tiledbuf)(R *I, R *O,
Chris@42 189 INT n0, INT is0, INT os0,
Chris@42 190 INT n1, INT is1, INT os1, INT vl)
Chris@42 191 {
Chris@42 192 R buf[CACHESIZE / (2 * sizeof(R))];
Chris@42 193 /* input and buffer in cache, or
Chris@42 194 output and buffer in cache */
Chris@42 195 INT tilesz = X(compute_tilesz)(vl, 2);
Chris@42 196 struct cpy2d_closure k;
Chris@42 197 k.I = I;
Chris@42 198 k.O = O;
Chris@42 199 k.is0 = is0;
Chris@42 200 k.os0 = os0;
Chris@42 201 k.is1 = is1;
Chris@42 202 k.os1 = os1;
Chris@42 203 k.vl = vl;
Chris@42 204 k.buf = buf;
Chris@42 205 A(tilesz * tilesz * vl * sizeof(R) <= sizeof(buf));
Chris@42 206 X(tile2d)(0, n0, 0, n1, tilesz, dotile_buf, &k);
Chris@42 207 }