1 /*******************************************************************************
2 Copyright (c) 2016, The OpenBLAS Project
3 All rights reserved.
4 Redistribution and use in source and binary forms, with or without
5 modification, are permitted provided that the following conditions are
6 met:
7 1. Redistributions of source code must retain the above copyright
8 notice, this list of conditions and the following disclaimer.
9 2. Redistributions in binary form must reproduce the above copyright
10 notice, this list of conditions and the following disclaimer in
11 the documentation and/or other materials provided with the
12 distribution.
13 3. Neither the name of the OpenBLAS project nor the names of
14 its contributors may be used to endorse or promote products
15 derived from this software without specific prior written permission.
16 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
17 AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 ARE DISCLAIMED. IN NO EVENT SHALL THE OPENBLAS PROJECT OR CONTRIBUTORS BE
20 LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21 DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
22 SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
23 CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
24 OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE
25 USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
26 *******************************************************************************/
27
28 #include "common.h"
29 #include "macros_msa.h"
30
CNAME(BLASLONG n,BLASLONG dummy0,BLASLONG dummy1,FLOAT dummy3,FLOAT dummy4,FLOAT * srcx,BLASLONG inc_x,FLOAT * srcy,BLASLONG inc_y,FLOAT * dummy,BLASLONG dummy2)31 int CNAME(BLASLONG n, BLASLONG dummy0, BLASLONG dummy1, FLOAT dummy3,
32 FLOAT dummy4, FLOAT *srcx, BLASLONG inc_x, FLOAT *srcy,
33 BLASLONG inc_y, FLOAT *dummy, BLASLONG dummy2)
34 {
35 BLASLONG i = 0, pref_offsetx, pref_offsety;
36 FLOAT *px, *py;
37 BLASLONG inc_x2, inc_y2;
38 FLOAT x0, x1, x2, x3, x4, x5, x6, x7;
39 FLOAT y0, y1, y2, y3, y4, y5, y6, y7;
40 v4f32 xv0, xv1, xv2, xv3, xv4, xv5, xv6, xv7;
41 v4f32 yv0, yv1, yv2, yv3, yv4, yv5, yv6, yv7;
42
43 if (n < 0) return (0);
44
45 pref_offsetx = (BLASLONG)srcx & (L1_DATA_LINESIZE - 1);
46 if (pref_offsetx > 0)
47 {
48 pref_offsetx = L1_DATA_LINESIZE - pref_offsetx;
49 pref_offsetx = pref_offsetx / sizeof(FLOAT);
50 }
51
52 pref_offsety = (BLASLONG)srcy & (L1_DATA_LINESIZE - 1);
53 if (pref_offsety > 0)
54 {
55 pref_offsety = L1_DATA_LINESIZE - pref_offsety;
56 pref_offsety = pref_offsety / sizeof(FLOAT);
57 }
58
59 px = srcx;
60 py = srcy;
61
62 inc_x2 = 2 * inc_x;
63 inc_y2 = 2 * inc_y;
64
65 if ((1 == inc_x) && (1 == inc_y))
66 {
67 if (n >> 4)
68 {
69 LD_SP8_INC(px, 4, xv0, xv1, xv2, xv3, xv4, xv5, xv6, xv7);
70
71 for (i = (n >> 4) - 1; i--;)
72 {
73 PREFETCH(px + pref_offsetx + 32);
74 PREFETCH(px + pref_offsetx + 40);
75 PREFETCH(px + pref_offsetx + 48);
76 PREFETCH(px + pref_offsetx + 56);
77
78 PREFETCH(py + pref_offsety + 32);
79 PREFETCH(py + pref_offsety + 40);
80 PREFETCH(py + pref_offsety + 48);
81 PREFETCH(py + pref_offsety + 56);
82
83 yv0 = LD_SP(py); py += 4;
84 ST_SP(xv0, srcy); srcy += 4;
85 yv1 = LD_SP(py); py += 4;
86 ST_SP(xv1, srcy); srcy += 4;
87 yv2 = LD_SP(py); py += 4;
88 ST_SP(xv2, srcy); srcy += 4;
89 yv3 = LD_SP(py); py += 4;
90 ST_SP(xv3, srcy); srcy += 4;
91 yv4 = LD_SP(py); py += 4;
92 ST_SP(xv4, srcy); srcy += 4;
93 yv5 = LD_SP(py); py += 4;
94 ST_SP(xv5, srcy); srcy += 4;
95 yv6 = LD_SP(py); py += 4;
96 ST_SP(xv6, srcy); srcy += 4;
97 yv7 = LD_SP(py); py += 4;
98 ST_SP(xv7, srcy); srcy += 4;
99
100 xv0 = LD_SP(px); px += 4;
101 ST_SP(yv0, srcx); srcx += 4;
102 xv1 = LD_SP(px); px += 4;
103 ST_SP(yv1, srcx); srcx += 4;
104 xv2 = LD_SP(px); px += 4;
105 ST_SP(yv2, srcx); srcx += 4;
106 xv3 = LD_SP(px); px += 4;
107 ST_SP(yv3, srcx); srcx += 4;
108 xv4 = LD_SP(px); px += 4;
109 ST_SP(yv4, srcx); srcx += 4;
110 xv5 = LD_SP(px); px += 4;
111 ST_SP(yv5, srcx); srcx += 4;
112 xv6 = LD_SP(px); px += 4;
113 ST_SP(yv6, srcx); srcx += 4;
114 xv7 = LD_SP(px); px += 4;
115 ST_SP(yv7, srcx); srcx += 4;
116 }
117
118 LD_SP8_INC(py, 4, yv0, yv1, yv2, yv3, yv4, yv5, yv6, yv7);
119 ST_SP8_INC(xv0, xv1, xv2, xv3, xv4, xv5, xv6, xv7, srcy, 4);
120 ST_SP8_INC(yv0, yv1, yv2, yv3, yv4, yv5, yv6, yv7, srcx, 4);
121 }
122
123 if (n & 15)
124 {
125 if ((n & 8) && (n & 4) && (n & 2))
126 {
127 LD_SP7_INC(px, 4, xv0, xv1, xv2, xv3, xv4, xv5, xv6);
128 LD_SP7_INC(py, 4, yv0, yv1, yv2, yv3, yv4, yv5, yv6);
129 ST_SP7_INC(xv0, xv1, xv2, xv3, xv4, xv5, xv6, srcy, 4);
130 ST_SP7_INC(yv0, yv1, yv2, yv3, yv4, yv5, yv6, srcx, 4);
131 }
132 else if ((n & 8) && (n & 4))
133 {
134 LD_SP6_INC(px, 4, xv0, xv1, xv2, xv3, xv4, xv5);
135 LD_SP6_INC(py, 4, yv0, yv1, yv2, yv3, yv4, yv5);
136 ST_SP6_INC(xv0, xv1, xv2, xv3, xv4, xv5, srcy, 4);
137 ST_SP6_INC(yv0, yv1, yv2, yv3, yv4, yv5, srcx, 4);
138 }
139 else if ((n & 8) && (n & 2))
140 {
141 LD_SP5_INC(px, 4, xv0, xv1, xv2, xv3, xv4);
142 LD_SP5_INC(py, 4, yv0, yv1, yv2, yv3, yv4);
143 ST_SP5_INC(xv0, xv1, xv2, xv3, xv4, srcy, 4);
144 ST_SP5_INC(yv0, yv1, yv2, yv3, yv4, srcx, 4);
145 }
146 else if ((n & 4) && (n & 2))
147 {
148 LD_SP3_INC(px, 4, xv0, xv1, xv2);
149 LD_SP3_INC(py, 4, yv0, yv1, yv2);
150 ST_SP3_INC(xv0, xv1, xv2, srcy, 4);
151 ST_SP3_INC(yv0, yv1, yv2, srcx, 4);
152 }
153 else if (n & 8)
154 {
155 LD_SP4_INC(px, 4, xv0, xv1, xv2, xv3);
156 LD_SP4_INC(py, 4, yv0, yv1, yv2, yv3);
157 ST_SP4_INC(xv0, xv1, xv2, xv3, srcy, 4);
158 ST_SP4_INC(yv0, yv1, yv2, yv3, srcx, 4);
159 }
160 else if (n & 4)
161 {
162 LD_SP2_INC(px, 4, xv0, xv1);
163 LD_SP2_INC(py, 4, yv0, yv1);
164 ST_SP2_INC(xv0, xv1, srcy, 4);
165 ST_SP2_INC(yv0, yv1, srcx, 4);
166 }
167 else if (n & 2)
168 {
169 xv0 = LD_SP(px);
170 yv0 = LD_SP(py);
171
172 px += 4;
173 py += 4;
174
175 ST_SP(xv0, srcy);
176 ST_SP(yv0, srcx);
177
178 srcx += 4;
179 srcy += 4;
180 }
181
182 if (n & 1)
183 {
184 LD_GP2_INC(px, 1, x0, x1);
185 LD_GP2_INC(py, 1, y0, y1);
186 ST_GP2_INC(x0, x1, srcy, 1);
187 ST_GP2_INC(y0, y1, srcx, 1);
188 }
189 }
190 }
191 else
192 {
193 for (i = (n >> 2); i--;)
194 {
195 x0 = srcx[0 * inc_x2];
196 x1 = srcx[0 * inc_x2 + 1];
197 x2 = srcx[1 * inc_x2];
198 x3 = srcx[1 * inc_x2 + 1];
199 x4 = srcx[2 * inc_x2];
200 x5 = srcx[2 * inc_x2 + 1];
201 x6 = srcx[3 * inc_x2];
202 x7 = srcx[3 * inc_x2 + 1];
203
204 y0 = srcy[0 * inc_y2];
205 y1 = srcy[0 * inc_y2 + 1];
206 y2 = srcy[1 * inc_y2];
207 y3 = srcy[1 * inc_y2 + 1];
208 y4 = srcy[2 * inc_y2];
209 y5 = srcy[2 * inc_y2 + 1];
210 y6 = srcy[3 * inc_y2];
211 y7 = srcy[3 * inc_y2 + 1];
212
213 srcx[0 * inc_x2] = y0;
214 srcx[0 * inc_x2 + 1] = y1;
215 srcx[1 * inc_x2] = y2;
216 srcx[1 * inc_x2 + 1] = y3;
217 srcx[2 * inc_x2] = y4;
218 srcx[2 * inc_x2 + 1] = y5;
219 srcx[3 * inc_x2] = y6;
220 srcx[3 * inc_x2 + 1] = y7;
221
222 srcy[0 * inc_y2] = x0;
223 srcy[0 * inc_y2 + 1] = x1;
224 srcy[1 * inc_y2] = x2;
225 srcy[1 * inc_y2 + 1] = x3;
226 srcy[2 * inc_y2] = x4;
227 srcy[2 * inc_y2 + 1] = x5;
228 srcy[3 * inc_y2] = x6;
229 srcy[3 * inc_y2 + 1] = x7;
230
231 srcx += 4 * inc_x2;
232 srcy += 4 * inc_y2;
233 }
234
235 if (n & 2)
236 {
237 x0 = srcx[0 * inc_x2];
238 x1 = srcx[0 * inc_x2 + 1];
239 x2 = srcx[1 * inc_x2];
240 x3 = srcx[1 * inc_x2 + 1];
241
242 y0 = srcy[0 * inc_y2];
243 y1 = srcy[0 * inc_y2 + 1];
244 y2 = srcy[1 * inc_y2];
245 y3 = srcy[1 * inc_y2 + 1];
246
247 srcx[0 * inc_x2] = y0;
248 srcx[0 * inc_x2 + 1] = y1;
249 srcx[1 * inc_x2] = y2;
250 srcx[1 * inc_x2 + 1] = y3;
251
252 srcy[0 * inc_y2] = x0;
253 srcy[0 * inc_y2 + 1] = x1;
254 srcy[1 * inc_y2] = x2;
255 srcy[1 * inc_y2 + 1] = x3;
256
257 srcx += 2 * inc_x2;
258 srcy += 2 * inc_y2;
259 }
260
261 if (n & 1)
262 {
263 x0 = srcx[0 * inc_x2];
264 x1 = srcx[0 * inc_x2 + 1];
265
266 y0 = srcy[0 * inc_y2];
267 y1 = srcy[0 * inc_y2 + 1];
268
269 srcx[0 * inc_x2] = y0;
270 srcx[0 * inc_x2 + 1] = y1;
271
272 srcy[0 * inc_y2] = x0;
273 srcy[0 * inc_y2 + 1] = x1;
274
275 srcx += inc_x2;
276 srcy += inc_y2;
277 }
278 }
279
280 return (0);
281 }
282