1 /*
2 * Copyright (c) 2013 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 #include <stdlib.h>
11
12 #include "./vp9_rtcd.h"
13 #include "vp9/common/vp9_common.h"
14 #include "vp9/common/mips/dspr2/vp9_common_dspr2.h"
15
16 #if HAVE_DSPR2
vp9_h_predictor_8x8_dspr2(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)17 void vp9_h_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
18 const uint8_t *above, const uint8_t *left) {
19 int32_t tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8;
20
21 __asm__ __volatile__ (
22 "lb %[tmp1], (%[left]) \n\t"
23 "lb %[tmp2], 1(%[left]) \n\t"
24 "lb %[tmp3], 2(%[left]) \n\t"
25 "lb %[tmp4], 3(%[left]) \n\t"
26 "lb %[tmp5], 4(%[left]) \n\t"
27 "lb %[tmp6], 5(%[left]) \n\t"
28 "lb %[tmp7], 6(%[left]) \n\t"
29 "lb %[tmp8], 7(%[left]) \n\t"
30
31 "replv.qb %[tmp1], %[tmp1] \n\t"
32 "replv.qb %[tmp2], %[tmp2] \n\t"
33 "replv.qb %[tmp3], %[tmp3] \n\t"
34 "replv.qb %[tmp4], %[tmp4] \n\t"
35 "replv.qb %[tmp5], %[tmp5] \n\t"
36 "replv.qb %[tmp6], %[tmp6] \n\t"
37 "replv.qb %[tmp7], %[tmp7] \n\t"
38 "replv.qb %[tmp8], %[tmp8] \n\t"
39
40 "sw %[tmp1], (%[dst]) \n\t"
41 "sw %[tmp1], 4(%[dst]) \n\t"
42 "add %[dst], %[dst], %[stride] \n\t"
43 "sw %[tmp2], (%[dst]) \n\t"
44 "sw %[tmp2], 4(%[dst]) \n\t"
45 "add %[dst], %[dst], %[stride] \n\t"
46 "sw %[tmp3], (%[dst]) \n\t"
47 "sw %[tmp3], 4(%[dst]) \n\t"
48 "add %[dst], %[dst], %[stride] \n\t"
49 "sw %[tmp4], (%[dst]) \n\t"
50 "sw %[tmp4], 4(%[dst]) \n\t"
51 "add %[dst], %[dst], %[stride] \n\t"
52 "sw %[tmp5], (%[dst]) \n\t"
53 "sw %[tmp5], 4(%[dst]) \n\t"
54 "add %[dst], %[dst], %[stride] \n\t"
55 "sw %[tmp6], (%[dst]) \n\t"
56 "sw %[tmp6], 4(%[dst]) \n\t"
57 "add %[dst], %[dst], %[stride] \n\t"
58 "sw %[tmp7], (%[dst]) \n\t"
59 "sw %[tmp7], 4(%[dst]) \n\t"
60 "add %[dst], %[dst], %[stride] \n\t"
61 "sw %[tmp8], (%[dst]) \n\t"
62 "sw %[tmp8], 4(%[dst]) \n\t"
63
64 : [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2),
65 [tmp3] "=&r" (tmp3), [tmp4] "=&r" (tmp4),
66 [tmp5] "=&r" (tmp5), [tmp7] "=&r" (tmp7),
67 [tmp6] "=&r" (tmp6), [tmp8] "=&r" (tmp8)
68 : [left] "r" (left), [dst] "r" (dst),
69 [stride] "r" (stride)
70 );
71 }
72
vp9_dc_predictor_8x8_dspr2(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)73 void vp9_dc_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
74 const uint8_t *above, const uint8_t *left) {
75 int32_t expected_dc;
76 int32_t average;
77 int32_t tmp, above1, above_l1, above_r1, left1, left_r1, left_l1;
78 int32_t above2, above_l2, above_r2, left2, left_r2, left_l2;
79
80 __asm__ __volatile__ (
81 "lw %[above1], (%[above]) \n\t"
82 "lw %[above2], 4(%[above]) \n\t"
83 "lw %[left1], (%[left]) \n\t"
84 "lw %[left2], 4(%[left]) \n\t"
85
86 "preceu.ph.qbl %[above_l1], %[above1] \n\t"
87 "preceu.ph.qbr %[above_r1], %[above1] \n\t"
88 "preceu.ph.qbl %[left_l1], %[left1] \n\t"
89 "preceu.ph.qbr %[left_r1], %[left1] \n\t"
90
91 "preceu.ph.qbl %[above_l2], %[above2] \n\t"
92 "preceu.ph.qbr %[above_r2], %[above2] \n\t"
93 "preceu.ph.qbl %[left_l2], %[left2] \n\t"
94 "preceu.ph.qbr %[left_r2], %[left2] \n\t"
95
96 "addu.ph %[average], %[above_r1], %[above_l1] \n\t"
97 "addu.ph %[average], %[average], %[left_l1] \n\t"
98 "addu.ph %[average], %[average], %[left_r1] \n\t"
99
100 "addu.ph %[average], %[average], %[above_l2] \n\t"
101 "addu.ph %[average], %[average], %[above_r2] \n\t"
102 "addu.ph %[average], %[average], %[left_l2] \n\t"
103 "addu.ph %[average], %[average], %[left_r2] \n\t"
104
105 "addiu %[average], %[average], 8 \n\t"
106
107 "srl %[tmp], %[average], 16 \n\t"
108 "addu.ph %[average], %[tmp], %[average] \n\t"
109 "srl %[expected_dc], %[average], 4 \n\t"
110 "replv.qb %[expected_dc], %[expected_dc] \n\t"
111
112 "sw %[expected_dc], (%[dst]) \n\t"
113 "sw %[expected_dc], 4(%[dst]) \n\t"
114
115 "add %[dst], %[dst], %[stride] \n\t"
116 "sw %[expected_dc], (%[dst]) \n\t"
117 "sw %[expected_dc], 4(%[dst]) \n\t"
118
119 "add %[dst], %[dst], %[stride] \n\t"
120 "sw %[expected_dc], (%[dst]) \n\t"
121 "sw %[expected_dc], 4(%[dst]) \n\t"
122
123 "add %[dst], %[dst], %[stride] \n\t"
124 "sw %[expected_dc], (%[dst]) \n\t"
125 "sw %[expected_dc], 4(%[dst]) \n\t"
126
127 "add %[dst], %[dst], %[stride] \n\t"
128 "sw %[expected_dc], (%[dst]) \n\t"
129 "sw %[expected_dc], 4(%[dst]) \n\t"
130
131 "add %[dst], %[dst], %[stride] \n\t"
132 "sw %[expected_dc], (%[dst]) \n\t"
133 "sw %[expected_dc], 4(%[dst]) \n\t"
134
135 "add %[dst], %[dst], %[stride] \n\t"
136 "sw %[expected_dc], (%[dst]) \n\t"
137 "sw %[expected_dc], 4(%[dst]) \n\t"
138
139 "add %[dst], %[dst], %[stride] \n\t"
140 "sw %[expected_dc], (%[dst]) \n\t"
141 "sw %[expected_dc], 4(%[dst]) \n\t"
142
143 : [above1] "=&r" (above1), [above_l1] "=&r" (above_l1),
144 [above_r1] "=&r" (above_r1), [left1] "=&r" (left1),
145 [left_l1] "=&r" (left_l1), [left_r1] "=&r" (left_r1),
146 [above2] "=&r" (above2), [above_l2] "=&r" (above_l2),
147 [above_r2] "=&r" (above_r2), [left2] "=&r" (left2),
148 [left_l2] "=&r" (left_l2), [left_r2] "=&r" (left_r2),
149 [average] "=&r" (average), [tmp] "=&r" (tmp),
150 [expected_dc] "=&r" (expected_dc)
151 : [above] "r" (above), [left] "r" (left), [dst] "r" (dst),
152 [stride] "r" (stride)
153 );
154 }
155
vp9_tm_predictor_8x8_dspr2(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)156 void vp9_tm_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
157 const uint8_t *above, const uint8_t *left) {
158 int32_t abovel, abover;
159 int32_t abovel_1, abover_1;
160 int32_t left0;
161 int32_t res0, res1, res2, res3;
162 int32_t reshw;
163 int32_t top_left;
164 uint8_t *cm = vp9_ff_cropTbl;
165
166 __asm__ __volatile__ (
167 "ulw %[reshw], (%[above]) \n\t"
168 "ulw %[top_left], 4(%[above]) \n\t"
169
170 "lbu %[left0], (%[left]) \n\t"
171
172 "preceu.ph.qbl %[abovel], %[reshw] \n\t"
173 "preceu.ph.qbr %[abover], %[reshw] \n\t"
174 "preceu.ph.qbl %[abovel_1], %[top_left] \n\t"
175 "preceu.ph.qbr %[abover_1], %[top_left] \n\t"
176
177 "lbu %[top_left], -1(%[above]) \n\t"
178 "replv.ph %[left0], %[left0] \n\t"
179
180 "replv.ph %[top_left], %[top_left] \n\t"
181
182 "addu.ph %[reshw], %[abovel], %[left0] \n\t"
183 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
184
185 "sll %[res2], %[reshw], 16 \n\t"
186 "sra %[res2], %[res2], 16 \n\t"
187 "sra %[res3], %[reshw], 16 \n\t"
188
189 "addu.ph %[reshw], %[abover], %[left0] \n\t"
190 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
191
192 "sll %[res0], %[reshw], 16 \n\t"
193 "sra %[res0], %[res0], 16 \n\t"
194 "sra %[res1], %[reshw], 16 \n\t"
195
196 "lbux %[res0], %[res0](%[cm]) \n\t"
197 "lbux %[res1], %[res1](%[cm]) \n\t"
198 "lbux %[res2], %[res2](%[cm]) \n\t"
199 "lbux %[res3], %[res3](%[cm]) \n\t"
200
201 "sb %[res0], (%[dst]) \n\t"
202 "sb %[res1], 1(%[dst]) \n\t"
203 "sb %[res2], 2(%[dst]) \n\t"
204 "sb %[res3], 3(%[dst]) \n\t"
205
206 "addu.ph %[reshw], %[abovel_1], %[left0] \n\t"
207 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
208
209 "sll %[res2], %[reshw], 16 \n\t"
210 "sra %[res2], %[res2], 16 \n\t"
211 "sra %[res3], %[reshw], 16 \n\t"
212
213 "addu.ph %[reshw], %[abover_1], %[left0] \n\t"
214 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
215
216 "sll %[res0], %[reshw], 16 \n\t"
217 "sra %[res0], %[res0], 16 \n\t"
218 "sra %[res1], %[reshw], 16 \n\t"
219
220 "lbu %[left0], 1(%[left]) \n\t"
221
222 "lbux %[res0], %[res0](%[cm]) \n\t"
223 "lbux %[res1], %[res1](%[cm]) \n\t"
224 "lbux %[res2], %[res2](%[cm]) \n\t"
225 "lbux %[res3], %[res3](%[cm]) \n\t"
226
227 "sb %[res0], 4(%[dst]) \n\t"
228 "sb %[res1], 5(%[dst]) \n\t"
229 "sb %[res2], 6(%[dst]) \n\t"
230 "sb %[res3], 7(%[dst]) \n\t"
231
232 "replv.ph %[left0], %[left0] \n\t"
233 "add %[dst], %[dst], %[stride] \n\t"
234
235 "addu.ph %[reshw], %[abovel], %[left0] \n\t"
236 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
237
238 "sll %[res2], %[reshw], 16 \n\t"
239 "sra %[res2], %[res2], 16 \n\t"
240 "sra %[res3], %[reshw], 16 \n\t"
241
242 "addu.ph %[reshw], %[abover], %[left0] \n\t"
243 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
244
245 "sll %[res0], %[reshw], 16 \n\t"
246 "sra %[res0], %[res0], 16 \n\t"
247 "sra %[res1], %[reshw], 16 \n\t"
248
249 "lbux %[res0], %[res0](%[cm]) \n\t"
250 "lbux %[res1], %[res1](%[cm]) \n\t"
251 "lbux %[res2], %[res2](%[cm]) \n\t"
252 "lbux %[res3], %[res3](%[cm]) \n\t"
253
254 "sb %[res0], (%[dst]) \n\t"
255 "sb %[res1], 1(%[dst]) \n\t"
256 "sb %[res2], 2(%[dst]) \n\t"
257 "sb %[res3], 3(%[dst]) \n\t"
258
259 "addu.ph %[reshw], %[abovel_1], %[left0] \n\t"
260 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
261
262 "sll %[res2], %[reshw], 16 \n\t"
263 "sra %[res2], %[res2], 16 \n\t"
264 "sra %[res3], %[reshw], 16 \n\t"
265
266 "addu.ph %[reshw], %[abover_1], %[left0] \n\t"
267 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
268
269 "sll %[res0], %[reshw], 16 \n\t"
270 "sra %[res0], %[res0], 16 \n\t"
271 "sra %[res1], %[reshw], 16 \n\t"
272
273 "lbu %[left0], 2(%[left]) \n\t"
274
275 "lbux %[res0], %[res0](%[cm]) \n\t"
276 "lbux %[res1], %[res1](%[cm]) \n\t"
277 "lbux %[res2], %[res2](%[cm]) \n\t"
278 "lbux %[res3], %[res3](%[cm]) \n\t"
279
280 "sb %[res0], 4(%[dst]) \n\t"
281 "sb %[res1], 5(%[dst]) \n\t"
282 "sb %[res2], 6(%[dst]) \n\t"
283 "sb %[res3], 7(%[dst]) \n\t"
284
285 "replv.ph %[left0], %[left0] \n\t"
286 "add %[dst], %[dst], %[stride] \n\t"
287
288 "addu.ph %[reshw], %[abovel], %[left0] \n\t"
289 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
290
291 "sll %[res2], %[reshw], 16 \n\t"
292 "sra %[res2], %[res2], 16 \n\t"
293 "sra %[res3], %[reshw], 16 \n\t"
294
295 "addu.ph %[reshw], %[abover], %[left0] \n\t"
296 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
297
298 "sll %[res0], %[reshw], 16 \n\t"
299 "sra %[res0], %[res0], 16 \n\t"
300 "sra %[res1], %[reshw], 16 \n\t"
301
302 "lbux %[res0], %[res0](%[cm]) \n\t"
303 "lbux %[res1], %[res1](%[cm]) \n\t"
304 "lbux %[res2], %[res2](%[cm]) \n\t"
305 "lbux %[res3], %[res3](%[cm]) \n\t"
306
307 "sb %[res0], (%[dst]) \n\t"
308 "sb %[res1], 1(%[dst]) \n\t"
309 "sb %[res2], 2(%[dst]) \n\t"
310 "sb %[res3], 3(%[dst]) \n\t"
311
312 "addu.ph %[reshw], %[abovel_1], %[left0] \n\t"
313 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
314
315 "sll %[res2], %[reshw], 16 \n\t"
316 "sra %[res2], %[res2], 16 \n\t"
317 "sra %[res3], %[reshw], 16 \n\t"
318
319 "addu.ph %[reshw], %[abover_1], %[left0] \n\t"
320 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
321
322 "sll %[res0], %[reshw], 16 \n\t"
323 "sra %[res0], %[res0], 16 \n\t"
324 "sra %[res1], %[reshw], 16 \n\t"
325
326 "lbu %[left0], 3(%[left]) \n\t"
327
328 "lbux %[res0], %[res0](%[cm]) \n\t"
329 "lbux %[res1], %[res1](%[cm]) \n\t"
330 "lbux %[res2], %[res2](%[cm]) \n\t"
331 "lbux %[res3], %[res3](%[cm]) \n\t"
332
333 "sb %[res0], 4(%[dst]) \n\t"
334 "sb %[res1], 5(%[dst]) \n\t"
335 "sb %[res2], 6(%[dst]) \n\t"
336 "sb %[res3], 7(%[dst]) \n\t"
337
338 "replv.ph %[left0], %[left0] \n\t"
339 "add %[dst], %[dst], %[stride] \n\t"
340
341 "addu.ph %[reshw], %[abovel], %[left0] \n\t"
342 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
343
344 "sll %[res2], %[reshw], 16 \n\t"
345 "sra %[res2], %[res2], 16 \n\t"
346 "sra %[res3], %[reshw], 16 \n\t"
347
348 "addu.ph %[reshw], %[abover], %[left0] \n\t"
349 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
350
351 "sll %[res0], %[reshw], 16 \n\t"
352 "sra %[res0], %[res0], 16 \n\t"
353 "sra %[res1], %[reshw], 16 \n\t"
354
355 "lbux %[res0], %[res0](%[cm]) \n\t"
356 "lbux %[res1], %[res1](%[cm]) \n\t"
357 "lbux %[res2], %[res2](%[cm]) \n\t"
358 "lbux %[res3], %[res3](%[cm]) \n\t"
359
360 "sb %[res0], (%[dst]) \n\t"
361 "sb %[res1], 1(%[dst]) \n\t"
362 "sb %[res2], 2(%[dst]) \n\t"
363 "sb %[res3], 3(%[dst]) \n\t"
364
365 "addu.ph %[reshw], %[abovel_1], %[left0] \n\t"
366 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
367
368 "sll %[res2], %[reshw], 16 \n\t"
369 "sra %[res2], %[res2], 16 \n\t"
370 "sra %[res3], %[reshw], 16 \n\t"
371
372 "addu.ph %[reshw], %[abover_1], %[left0] \n\t"
373 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
374
375 "sll %[res0], %[reshw], 16 \n\t"
376 "sra %[res0], %[res0], 16 \n\t"
377 "sra %[res1], %[reshw], 16 \n\t"
378
379 "lbu %[left0], 4(%[left]) \n\t"
380
381 "lbux %[res0], %[res0](%[cm]) \n\t"
382 "lbux %[res1], %[res1](%[cm]) \n\t"
383 "lbux %[res2], %[res2](%[cm]) \n\t"
384 "lbux %[res3], %[res3](%[cm]) \n\t"
385
386 "sb %[res0], 4(%[dst]) \n\t"
387 "sb %[res1], 5(%[dst]) \n\t"
388 "sb %[res2], 6(%[dst]) \n\t"
389 "sb %[res3], 7(%[dst]) \n\t"
390
391 "replv.ph %[left0], %[left0] \n\t"
392 "add %[dst], %[dst], %[stride] \n\t"
393
394 "addu.ph %[reshw], %[abovel], %[left0] \n\t"
395 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
396
397 "sll %[res2], %[reshw], 16 \n\t"
398 "sra %[res2], %[res2], 16 \n\t"
399 "sra %[res3], %[reshw], 16 \n\t"
400
401 "addu.ph %[reshw], %[abover], %[left0] \n\t"
402 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
403
404 "sll %[res0], %[reshw], 16 \n\t"
405 "sra %[res0], %[res0], 16 \n\t"
406 "sra %[res1], %[reshw], 16 \n\t"
407
408 "lbux %[res0], %[res0](%[cm]) \n\t"
409 "lbux %[res1], %[res1](%[cm]) \n\t"
410 "lbux %[res2], %[res2](%[cm]) \n\t"
411 "lbux %[res3], %[res3](%[cm]) \n\t"
412
413 "sb %[res0], (%[dst]) \n\t"
414 "sb %[res1], 1(%[dst]) \n\t"
415 "sb %[res2], 2(%[dst]) \n\t"
416 "sb %[res3], 3(%[dst]) \n\t"
417
418 "addu.ph %[reshw], %[abovel_1], %[left0] \n\t"
419 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
420
421 "sll %[res2], %[reshw], 16 \n\t"
422 "sra %[res2], %[res2], 16 \n\t"
423 "sra %[res3], %[reshw], 16 \n\t"
424
425 "addu.ph %[reshw], %[abover_1], %[left0] \n\t"
426 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
427
428 "sll %[res0], %[reshw], 16 \n\t"
429 "sra %[res0], %[res0], 16 \n\t"
430 "sra %[res1], %[reshw], 16 \n\t"
431
432 "lbu %[left0], 5(%[left]) \n\t"
433
434 "lbux %[res0], %[res0](%[cm]) \n\t"
435 "lbux %[res1], %[res1](%[cm]) \n\t"
436 "lbux %[res2], %[res2](%[cm]) \n\t"
437 "lbux %[res3], %[res3](%[cm]) \n\t"
438
439 "sb %[res0], 4(%[dst]) \n\t"
440 "sb %[res1], 5(%[dst]) \n\t"
441 "sb %[res2], 6(%[dst]) \n\t"
442 "sb %[res3], 7(%[dst]) \n\t"
443
444 "replv.ph %[left0], %[left0] \n\t"
445 "add %[dst], %[dst], %[stride] \n\t"
446
447 "addu.ph %[reshw], %[abovel], %[left0] \n\t"
448 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
449
450 "sll %[res2], %[reshw], 16 \n\t"
451 "sra %[res2], %[res2], 16 \n\t"
452 "sra %[res3], %[reshw], 16 \n\t"
453
454 "addu.ph %[reshw], %[abover], %[left0] \n\t"
455 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
456
457 "sll %[res0], %[reshw], 16 \n\t"
458 "sra %[res0], %[res0], 16 \n\t"
459 "sra %[res1], %[reshw], 16 \n\t"
460
461 "lbux %[res0], %[res0](%[cm]) \n\t"
462 "lbux %[res1], %[res1](%[cm]) \n\t"
463 "lbux %[res2], %[res2](%[cm]) \n\t"
464 "lbux %[res3], %[res3](%[cm]) \n\t"
465
466 "sb %[res0], (%[dst]) \n\t"
467 "sb %[res1], 1(%[dst]) \n\t"
468 "sb %[res2], 2(%[dst]) \n\t"
469 "sb %[res3], 3(%[dst]) \n\t"
470
471 "addu.ph %[reshw], %[abovel_1], %[left0] \n\t"
472 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
473
474 "sll %[res2], %[reshw], 16 \n\t"
475 "sra %[res2], %[res2], 16 \n\t"
476 "sra %[res3], %[reshw], 16 \n\t"
477
478 "addu.ph %[reshw], %[abover_1], %[left0] \n\t"
479 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
480
481 "sll %[res0], %[reshw], 16 \n\t"
482 "sra %[res0], %[res0], 16 \n\t"
483 "sra %[res1], %[reshw], 16 \n\t"
484
485 "lbu %[left0], 6(%[left]) \n\t"
486
487 "lbux %[res0], %[res0](%[cm]) \n\t"
488 "lbux %[res1], %[res1](%[cm]) \n\t"
489 "lbux %[res2], %[res2](%[cm]) \n\t"
490 "lbux %[res3], %[res3](%[cm]) \n\t"
491
492 "sb %[res0], 4(%[dst]) \n\t"
493 "sb %[res1], 5(%[dst]) \n\t"
494 "sb %[res2], 6(%[dst]) \n\t"
495 "sb %[res3], 7(%[dst]) \n\t"
496
497 "replv.ph %[left0], %[left0] \n\t"
498 "add %[dst], %[dst], %[stride] \n\t"
499
500 "addu.ph %[reshw], %[abovel], %[left0] \n\t"
501 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
502
503 "sll %[res2], %[reshw], 16 \n\t"
504 "sra %[res2], %[res2], 16 \n\t"
505 "sra %[res3], %[reshw], 16 \n\t"
506
507 "addu.ph %[reshw], %[abover], %[left0] \n\t"
508 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
509
510 "sll %[res0], %[reshw], 16 \n\t"
511 "sra %[res0], %[res0], 16 \n\t"
512 "sra %[res1], %[reshw], 16 \n\t"
513
514 "lbux %[res0], %[res0](%[cm]) \n\t"
515 "lbux %[res1], %[res1](%[cm]) \n\t"
516 "lbux %[res2], %[res2](%[cm]) \n\t"
517 "lbux %[res3], %[res3](%[cm]) \n\t"
518
519 "sb %[res0], (%[dst]) \n\t"
520 "sb %[res1], 1(%[dst]) \n\t"
521 "sb %[res2], 2(%[dst]) \n\t"
522 "sb %[res3], 3(%[dst]) \n\t"
523
524 "addu.ph %[reshw], %[abovel_1], %[left0] \n\t"
525 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
526
527 "sll %[res2], %[reshw], 16 \n\t"
528 "sra %[res2], %[res2], 16 \n\t"
529 "sra %[res3], %[reshw], 16 \n\t"
530
531 "addu.ph %[reshw], %[abover_1], %[left0] \n\t"
532 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
533
534 "sll %[res0], %[reshw], 16 \n\t"
535 "sra %[res0], %[res0], 16 \n\t"
536 "sra %[res1], %[reshw], 16 \n\t"
537
538 "lbu %[left0], 7(%[left]) \n\t"
539
540 "lbux %[res0], %[res0](%[cm]) \n\t"
541 "lbux %[res1], %[res1](%[cm]) \n\t"
542 "lbux %[res2], %[res2](%[cm]) \n\t"
543 "lbux %[res3], %[res3](%[cm]) \n\t"
544
545 "sb %[res0], 4(%[dst]) \n\t"
546 "sb %[res1], 5(%[dst]) \n\t"
547 "sb %[res2], 6(%[dst]) \n\t"
548 "sb %[res3], 7(%[dst]) \n\t"
549
550 "replv.ph %[left0], %[left0] \n\t"
551 "add %[dst], %[dst], %[stride] \n\t"
552
553 "addu.ph %[reshw], %[abovel], %[left0] \n\t"
554 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
555
556 "sll %[res2], %[reshw], 16 \n\t"
557 "sra %[res2], %[res2], 16 \n\t"
558 "sra %[res3], %[reshw], 16 \n\t"
559
560 "addu.ph %[reshw], %[abover], %[left0] \n\t"
561 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
562
563 "sll %[res0], %[reshw], 16 \n\t"
564 "sra %[res0], %[res0], 16 \n\t"
565 "sra %[res1], %[reshw], 16 \n\t"
566
567 "lbux %[res0], %[res0](%[cm]) \n\t"
568 "lbux %[res1], %[res1](%[cm]) \n\t"
569 "lbux %[res2], %[res2](%[cm]) \n\t"
570 "lbux %[res3], %[res3](%[cm]) \n\t"
571
572 "sb %[res0], (%[dst]) \n\t"
573 "sb %[res1], 1(%[dst]) \n\t"
574 "sb %[res2], 2(%[dst]) \n\t"
575 "sb %[res3], 3(%[dst]) \n\t"
576
577 "addu.ph %[reshw], %[abovel_1], %[left0] \n\t"
578 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
579
580 "sll %[res2], %[reshw], 16 \n\t"
581 "sra %[res2], %[res2], 16 \n\t"
582 "sra %[res3], %[reshw], 16 \n\t"
583
584 "addu.ph %[reshw], %[abover_1], %[left0] \n\t"
585 "subu.ph %[reshw], %[reshw], %[top_left] \n\t"
586
587 "sll %[res0], %[reshw], 16 \n\t"
588 "sra %[res0], %[res0], 16 \n\t"
589 "sra %[res1], %[reshw], 16 \n\t"
590
591 "lbux %[res0], %[res0](%[cm]) \n\t"
592 "lbux %[res1], %[res1](%[cm]) \n\t"
593 "lbux %[res2], %[res2](%[cm]) \n\t"
594 "lbux %[res3], %[res3](%[cm]) \n\t"
595
596 "sb %[res0], 4(%[dst]) \n\t"
597 "sb %[res1], 5(%[dst]) \n\t"
598 "sb %[res2], 6(%[dst]) \n\t"
599 "sb %[res3], 7(%[dst]) \n\t"
600
601 : [abovel] "=&r" (abovel), [abover] "=&r" (abover),
602 [abovel_1] "=&r" (abovel_1), [abover_1] "=&r" (abover_1),
603 [left0] "=&r" (left0), [res2] "=&r" (res2), [res3] "=&r" (res3),
604 [res0] "=&r" (res0), [res1] "=&r" (res1),
605 [reshw] "=&r" (reshw), [top_left] "=&r" (top_left)
606 : [above] "r" (above), [left] "r" (left),
607 [dst] "r" (dst), [stride] "r" (stride), [cm] "r" (cm)
608 );
609 }
610 #endif // #if HAVE_DSPR2
611