1 /*
2 * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include "vpx_dsp/mips/common_dspr2.h"
12
13 #if HAVE_DSPR2
vpx_h_predictor_16x16_dspr2(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)14 void vpx_h_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride,
15 const uint8_t *above, const uint8_t *left) {
16 int32_t tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
17 int32_t tmp9, tmp10, tmp11, tmp12, tmp13, tmp14, tmp15, tmp16;
18 (void)above;
19
20 __asm__ __volatile__(
21 "lb %[tmp1], (%[left]) \n\t"
22 "lb %[tmp2], 1(%[left]) \n\t"
23 "lb %[tmp3], 2(%[left]) \n\t"
24 "lb %[tmp4], 3(%[left]) \n\t"
25 "lb %[tmp5], 4(%[left]) \n\t"
26 "lb %[tmp6], 5(%[left]) \n\t"
27 "lb %[tmp7], 6(%[left]) \n\t"
28 "lb %[tmp8], 7(%[left]) \n\t"
29 "lb %[tmp9], 8(%[left]) \n\t"
30 "lb %[tmp10], 9(%[left]) \n\t"
31 "lb %[tmp11], 10(%[left]) \n\t"
32 "lb %[tmp12], 11(%[left]) \n\t"
33 "lb %[tmp13], 12(%[left]) \n\t"
34 "lb %[tmp14], 13(%[left]) \n\t"
35 "lb %[tmp15], 14(%[left]) \n\t"
36 "lb %[tmp16], 15(%[left]) \n\t"
37
38 "replv.qb %[tmp1], %[tmp1] \n\t"
39 "replv.qb %[tmp2], %[tmp2] \n\t"
40 "replv.qb %[tmp3], %[tmp3] \n\t"
41 "replv.qb %[tmp4], %[tmp4] \n\t"
42 "replv.qb %[tmp5], %[tmp5] \n\t"
43 "replv.qb %[tmp6], %[tmp6] \n\t"
44 "replv.qb %[tmp7], %[tmp7] \n\t"
45 "replv.qb %[tmp8], %[tmp8] \n\t"
46 "replv.qb %[tmp9], %[tmp9] \n\t"
47 "replv.qb %[tmp10], %[tmp10] \n\t"
48 "replv.qb %[tmp11], %[tmp11] \n\t"
49 "replv.qb %[tmp12], %[tmp12] \n\t"
50 "replv.qb %[tmp13], %[tmp13] \n\t"
51 "replv.qb %[tmp14], %[tmp14] \n\t"
52 "replv.qb %[tmp15], %[tmp15] \n\t"
53 "replv.qb %[tmp16], %[tmp16] \n\t"
54
55 "sw %[tmp1], (%[dst]) \n\t"
56 "sw %[tmp1], 4(%[dst]) \n\t"
57 "sw %[tmp1], 8(%[dst]) \n\t"
58 "sw %[tmp1], 12(%[dst]) \n\t"
59
60 "add %[dst], %[dst], %[stride] \n\t"
61 "sw %[tmp2], (%[dst]) \n\t"
62 "sw %[tmp2], 4(%[dst]) \n\t"
63 "sw %[tmp2], 8(%[dst]) \n\t"
64 "sw %[tmp2], 12(%[dst]) \n\t"
65
66 "add %[dst], %[dst], %[stride] \n\t"
67 "sw %[tmp3], (%[dst]) \n\t"
68 "sw %[tmp3], 4(%[dst]) \n\t"
69 "sw %[tmp3], 8(%[dst]) \n\t"
70 "sw %[tmp3], 12(%[dst]) \n\t"
71
72 "add %[dst], %[dst], %[stride] \n\t"
73 "sw %[tmp4], (%[dst]) \n\t"
74 "sw %[tmp4], 4(%[dst]) \n\t"
75 "sw %[tmp4], 8(%[dst]) \n\t"
76 "sw %[tmp4], 12(%[dst]) \n\t"
77
78 "add %[dst], %[dst], %[stride] \n\t"
79 "sw %[tmp5], (%[dst]) \n\t"
80 "sw %[tmp5], 4(%[dst]) \n\t"
81 "sw %[tmp5], 8(%[dst]) \n\t"
82 "sw %[tmp5], 12(%[dst]) \n\t"
83
84 "add %[dst], %[dst], %[stride] \n\t"
85 "sw %[tmp6], (%[dst]) \n\t"
86 "sw %[tmp6], 4(%[dst]) \n\t"
87 "sw %[tmp6], 8(%[dst]) \n\t"
88 "sw %[tmp6], 12(%[dst]) \n\t"
89
90 "add %[dst], %[dst], %[stride] \n\t"
91 "sw %[tmp7], (%[dst]) \n\t"
92 "sw %[tmp7], 4(%[dst]) \n\t"
93 "sw %[tmp7], 8(%[dst]) \n\t"
94 "sw %[tmp7], 12(%[dst]) \n\t"
95
96 "add %[dst], %[dst], %[stride] \n\t"
97 "sw %[tmp8], (%[dst]) \n\t"
98 "sw %[tmp8], 4(%[dst]) \n\t"
99 "sw %[tmp8], 8(%[dst]) \n\t"
100 "sw %[tmp8], 12(%[dst]) \n\t"
101
102 "add %[dst], %[dst], %[stride] \n\t"
103 "sw %[tmp9], (%[dst]) \n\t"
104 "sw %[tmp9], 4(%[dst]) \n\t"
105 "sw %[tmp9], 8(%[dst]) \n\t"
106 "sw %[tmp9], 12(%[dst]) \n\t"
107
108 "add %[dst], %[dst], %[stride] \n\t"
109 "sw %[tmp10], (%[dst]) \n\t"
110 "sw %[tmp10], 4(%[dst]) \n\t"
111 "sw %[tmp10], 8(%[dst]) \n\t"
112 "sw %[tmp10], 12(%[dst]) \n\t"
113
114 "add %[dst], %[dst], %[stride] \n\t"
115 "sw %[tmp11], (%[dst]) \n\t"
116 "sw %[tmp11], 4(%[dst]) \n\t"
117 "sw %[tmp11], 8(%[dst]) \n\t"
118 "sw %[tmp11], 12(%[dst]) \n\t"
119
120 "add %[dst], %[dst], %[stride] \n\t"
121 "sw %[tmp12], (%[dst]) \n\t"
122 "sw %[tmp12], 4(%[dst]) \n\t"
123 "sw %[tmp12], 8(%[dst]) \n\t"
124 "sw %[tmp12], 12(%[dst]) \n\t"
125
126 "add %[dst], %[dst], %[stride] \n\t"
127 "sw %[tmp13], (%[dst]) \n\t"
128 "sw %[tmp13], 4(%[dst]) \n\t"
129 "sw %[tmp13], 8(%[dst]) \n\t"
130 "sw %[tmp13], 12(%[dst]) \n\t"
131
132 "add %[dst], %[dst], %[stride] \n\t"
133 "sw %[tmp14], (%[dst]) \n\t"
134 "sw %[tmp14], 4(%[dst]) \n\t"
135 "sw %[tmp14], 8(%[dst]) \n\t"
136 "sw %[tmp14], 12(%[dst]) \n\t"
137
138 "add %[dst], %[dst], %[stride] \n\t"
139 "sw %[tmp15], (%[dst]) \n\t"
140 "sw %[tmp15], 4(%[dst]) \n\t"
141 "sw %[tmp15], 8(%[dst]) \n\t"
142 "sw %[tmp15], 12(%[dst]) \n\t"
143
144 "add %[dst], %[dst], %[stride] \n\t"
145 "sw %[tmp16], (%[dst]) \n\t"
146 "sw %[tmp16], 4(%[dst]) \n\t"
147 "sw %[tmp16], 8(%[dst]) \n\t"
148 "sw %[tmp16], 12(%[dst]) \n\t"
149
150 : [tmp1] "=&r"(tmp1), [tmp2] "=&r"(tmp2), [tmp3] "=&r"(tmp3),
151 [tmp4] "=&r"(tmp4), [tmp5] "=&r"(tmp5), [tmp7] "=&r"(tmp7),
152 [tmp6] "=&r"(tmp6), [tmp8] "=&r"(tmp8), [tmp9] "=&r"(tmp9),
153 [tmp10] "=&r"(tmp10), [tmp11] "=&r"(tmp11), [tmp12] "=&r"(tmp12),
154 [tmp13] "=&r"(tmp13), [tmp14] "=&r"(tmp14), [tmp15] "=&r"(tmp15),
155 [tmp16] "=&r"(tmp16)
156 : [left] "r"(left), [dst] "r"(dst), [stride] "r"(stride));
157 }
158
vpx_dc_predictor_16x16_dspr2(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)159 void vpx_dc_predictor_16x16_dspr2(uint8_t *dst, ptrdiff_t stride,
160 const uint8_t *above, const uint8_t *left) {
161 int32_t expected_dc;
162 int32_t average;
163 int32_t tmp, above1, above_l1, above_r1, left1, left_r1, left_l1;
164 int32_t above2, left2;
165
166 __asm__ __volatile__(
167 "lw %[above1], (%[above]) \n\t"
168 "lw %[above2], 4(%[above]) \n\t"
169 "lw %[left1], (%[left]) \n\t"
170 "lw %[left2], 4(%[left]) \n\t"
171
172 "preceu.ph.qbl %[above_l1], %[above1] \n\t"
173 "preceu.ph.qbr %[above_r1], %[above1] \n\t"
174 "preceu.ph.qbl %[left_l1], %[left1] \n\t"
175 "preceu.ph.qbr %[left_r1], %[left1] \n\t"
176
177 "addu.ph %[average], %[above_r1], %[above_l1] \n\t"
178 "addu.ph %[average], %[average], %[left_l1] \n\t"
179 "addu.ph %[average], %[average], %[left_r1] \n\t"
180
181 "preceu.ph.qbl %[above_l1], %[above2] \n\t"
182 "preceu.ph.qbr %[above_r1], %[above2] \n\t"
183 "preceu.ph.qbl %[left_l1], %[left2] \n\t"
184 "preceu.ph.qbr %[left_r1], %[left2] \n\t"
185
186 "addu.ph %[average], %[average], %[above_l1] \n\t"
187 "addu.ph %[average], %[average], %[above_r1] \n\t"
188 "addu.ph %[average], %[average], %[left_l1] \n\t"
189 "addu.ph %[average], %[average], %[left_r1] \n\t"
190
191 "lw %[above1], 8(%[above]) \n\t"
192 "lw %[above2], 12(%[above]) \n\t"
193 "lw %[left1], 8(%[left]) \n\t"
194 "lw %[left2], 12(%[left]) \n\t"
195
196 "preceu.ph.qbl %[above_l1], %[above1] \n\t"
197 "preceu.ph.qbr %[above_r1], %[above1] \n\t"
198 "preceu.ph.qbl %[left_l1], %[left1] \n\t"
199 "preceu.ph.qbr %[left_r1], %[left1] \n\t"
200
201 "addu.ph %[average], %[average], %[above_l1] \n\t"
202 "addu.ph %[average], %[average], %[above_r1] \n\t"
203 "addu.ph %[average], %[average], %[left_l1] \n\t"
204 "addu.ph %[average], %[average], %[left_r1] \n\t"
205
206 "preceu.ph.qbl %[above_l1], %[above2] \n\t"
207 "preceu.ph.qbr %[above_r1], %[above2] \n\t"
208 "preceu.ph.qbl %[left_l1], %[left2] \n\t"
209 "preceu.ph.qbr %[left_r1], %[left2] \n\t"
210
211 "addu.ph %[average], %[average], %[above_l1] \n\t"
212 "addu.ph %[average], %[average], %[above_r1] \n\t"
213 "addu.ph %[average], %[average], %[left_l1] \n\t"
214 "addu.ph %[average], %[average], %[left_r1] \n\t"
215
216 "addiu %[average], %[average], 16 \n\t"
217 "srl %[tmp], %[average], 16 \n\t"
218 "addu.ph %[average], %[tmp], %[average] \n\t"
219 "srl %[expected_dc], %[average], 5 \n\t"
220 "replv.qb %[expected_dc], %[expected_dc] \n\t"
221
222 "sw %[expected_dc], (%[dst]) \n\t"
223 "sw %[expected_dc], 4(%[dst]) \n\t"
224 "sw %[expected_dc], 8(%[dst]) \n\t"
225 "sw %[expected_dc], 12(%[dst]) \n\t"
226
227 "add %[dst], %[dst], %[stride] \n\t"
228 "sw %[expected_dc], (%[dst]) \n\t"
229 "sw %[expected_dc], 4(%[dst]) \n\t"
230 "sw %[expected_dc], 8(%[dst]) \n\t"
231 "sw %[expected_dc], 12(%[dst]) \n\t"
232
233 "add %[dst], %[dst], %[stride] \n\t"
234 "sw %[expected_dc], (%[dst]) \n\t"
235 "sw %[expected_dc], 4(%[dst]) \n\t"
236 "sw %[expected_dc], 8(%[dst]) \n\t"
237 "sw %[expected_dc], 12(%[dst]) \n\t"
238
239 "add %[dst], %[dst], %[stride] \n\t"
240 "sw %[expected_dc], (%[dst]) \n\t"
241 "sw %[expected_dc], 4(%[dst]) \n\t"
242 "sw %[expected_dc], 8(%[dst]) \n\t"
243 "sw %[expected_dc], 12(%[dst]) \n\t"
244
245 "add %[dst], %[dst], %[stride] \n\t"
246 "sw %[expected_dc], (%[dst]) \n\t"
247 "sw %[expected_dc], 4(%[dst]) \n\t"
248 "sw %[expected_dc], 8(%[dst]) \n\t"
249 "sw %[expected_dc], 12(%[dst]) \n\t"
250
251 "add %[dst], %[dst], %[stride] \n\t"
252 "sw %[expected_dc], (%[dst]) \n\t"
253 "sw %[expected_dc], 4(%[dst]) \n\t"
254 "sw %[expected_dc], 8(%[dst]) \n\t"
255 "sw %[expected_dc], 12(%[dst]) \n\t"
256
257 "add %[dst], %[dst], %[stride] \n\t"
258 "sw %[expected_dc], (%[dst]) \n\t"
259 "sw %[expected_dc], 4(%[dst]) \n\t"
260 "sw %[expected_dc], 8(%[dst]) \n\t"
261 "sw %[expected_dc], 12(%[dst]) \n\t"
262
263 "add %[dst], %[dst], %[stride] \n\t"
264 "sw %[expected_dc], (%[dst]) \n\t"
265 "sw %[expected_dc], 4(%[dst]) \n\t"
266 "sw %[expected_dc], 8(%[dst]) \n\t"
267 "sw %[expected_dc], 12(%[dst]) \n\t"
268
269 "add %[dst], %[dst], %[stride] \n\t"
270 "sw %[expected_dc], (%[dst]) \n\t"
271 "sw %[expected_dc], 4(%[dst]) \n\t"
272 "sw %[expected_dc], 8(%[dst]) \n\t"
273 "sw %[expected_dc], 12(%[dst]) \n\t"
274
275 "add %[dst], %[dst], %[stride] \n\t"
276 "sw %[expected_dc], (%[dst]) \n\t"
277 "sw %[expected_dc], 4(%[dst]) \n\t"
278 "sw %[expected_dc], 8(%[dst]) \n\t"
279 "sw %[expected_dc], 12(%[dst]) \n\t"
280
281 "add %[dst], %[dst], %[stride] \n\t"
282 "sw %[expected_dc], (%[dst]) \n\t"
283 "sw %[expected_dc], 4(%[dst]) \n\t"
284 "sw %[expected_dc], 8(%[dst]) \n\t"
285 "sw %[expected_dc], 12(%[dst]) \n\t"
286
287 "add %[dst], %[dst], %[stride] \n\t"
288 "sw %[expected_dc], (%[dst]) \n\t"
289 "sw %[expected_dc], 4(%[dst]) \n\t"
290 "sw %[expected_dc], 8(%[dst]) \n\t"
291 "sw %[expected_dc], 12(%[dst]) \n\t"
292
293 "add %[dst], %[dst], %[stride] \n\t"
294 "sw %[expected_dc], (%[dst]) \n\t"
295 "sw %[expected_dc], 4(%[dst]) \n\t"
296 "sw %[expected_dc], 8(%[dst]) \n\t"
297 "sw %[expected_dc], 12(%[dst]) \n\t"
298
299 "add %[dst], %[dst], %[stride] \n\t"
300 "sw %[expected_dc], (%[dst]) \n\t"
301 "sw %[expected_dc], 4(%[dst]) \n\t"
302 "sw %[expected_dc], 8(%[dst]) \n\t"
303 "sw %[expected_dc], 12(%[dst]) \n\t"
304
305 "add %[dst], %[dst], %[stride] \n\t"
306 "sw %[expected_dc], (%[dst]) \n\t"
307 "sw %[expected_dc], 4(%[dst]) \n\t"
308 "sw %[expected_dc], 8(%[dst]) \n\t"
309 "sw %[expected_dc], 12(%[dst]) \n\t"
310
311 "add %[dst], %[dst], %[stride] \n\t"
312 "sw %[expected_dc], (%[dst]) \n\t"
313 "sw %[expected_dc], 4(%[dst]) \n\t"
314 "sw %[expected_dc], 8(%[dst]) \n\t"
315 "sw %[expected_dc], 12(%[dst]) \n\t"
316
317 : [left1] "=&r"(left1), [above1] "=&r"(above1), [left_l1] "=&r"(left_l1),
318 [above_l1] "=&r"(above_l1), [left_r1] "=&r"(left_r1),
319 [above_r1] "=&r"(above_r1), [above2] "=&r"(above2),
320 [left2] "=&r"(left2), [average] "=&r"(average), [tmp] "=&r"(tmp),
321 [expected_dc] "=&r"(expected_dc)
322 : [above] "r"(above), [left] "r"(left), [dst] "r"(dst),
323 [stride] "r"(stride));
324 }
325 #endif // #if HAVE_DSPR2
326