intrapred8_dspr2.c revision 7ce0a1d1337c01056ba24006efab21f00e179e04 
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
14void vpx_h_predictor_8x8_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
18  __asm__ __volatile__ (
19      "lb         %[tmp1],      (%[left])                   \n\t"
20      "lb         %[tmp2],      1(%[left])                  \n\t"
21      "lb         %[tmp3],      2(%[left])                  \n\t"
22      "lb         %[tmp4],      3(%[left])                  \n\t"
23      "lb         %[tmp5],      4(%[left])                  \n\t"
24      "lb         %[tmp6],      5(%[left])                  \n\t"
25      "lb         %[tmp7],      6(%[left])                  \n\t"
26      "lb         %[tmp8],      7(%[left])                  \n\t"
27
28      "replv.qb   %[tmp1],      %[tmp1]                     \n\t"
29      "replv.qb   %[tmp2],      %[tmp2]                     \n\t"
30      "replv.qb   %[tmp3],      %[tmp3]                     \n\t"
31      "replv.qb   %[tmp4],      %[tmp4]                     \n\t"
32      "replv.qb   %[tmp5],      %[tmp5]                     \n\t"
33      "replv.qb   %[tmp6],      %[tmp6]                     \n\t"
34      "replv.qb   %[tmp7],      %[tmp7]                     \n\t"
35      "replv.qb   %[tmp8],      %[tmp8]                     \n\t"
36
37      "sw         %[tmp1],      (%[dst])                    \n\t"
38      "sw         %[tmp1],      4(%[dst])                   \n\t"
39      "add        %[dst],       %[dst],         %[stride]   \n\t"
40      "sw         %[tmp2],      (%[dst])                    \n\t"
41      "sw         %[tmp2],      4(%[dst])                   \n\t"
42      "add        %[dst],       %[dst],         %[stride]   \n\t"
43      "sw         %[tmp3],      (%[dst])                    \n\t"
44      "sw         %[tmp3],      4(%[dst])                   \n\t"
45      "add        %[dst],       %[dst],         %[stride]   \n\t"
46      "sw         %[tmp4],      (%[dst])                    \n\t"
47      "sw         %[tmp4],      4(%[dst])                   \n\t"
48      "add        %[dst],       %[dst],         %[stride]   \n\t"
49      "sw         %[tmp5],      (%[dst])                    \n\t"
50      "sw         %[tmp5],      4(%[dst])                   \n\t"
51      "add        %[dst],       %[dst],         %[stride]   \n\t"
52      "sw         %[tmp6],      (%[dst])                    \n\t"
53      "sw         %[tmp6],      4(%[dst])                   \n\t"
54      "add        %[dst],       %[dst],         %[stride]   \n\t"
55      "sw         %[tmp7],      (%[dst])                    \n\t"
56      "sw         %[tmp7],      4(%[dst])                   \n\t"
57      "add        %[dst],       %[dst],         %[stride]   \n\t"
58      "sw         %[tmp8],      (%[dst])                    \n\t"
59      "sw         %[tmp8],      4(%[dst])                   \n\t"
60
61      : [tmp1] "=&r" (tmp1), [tmp2] "=&r" (tmp2),
62        [tmp3] "=&r" (tmp3), [tmp4] "=&r" (tmp4),
63        [tmp5] "=&r" (tmp5), [tmp7] "=&r" (tmp7),
64        [tmp6] "=&r" (tmp6), [tmp8] "=&r" (tmp8)
65      : [left] "r" (left), [dst] "r" (dst),
66        [stride] "r" (stride)
67  );
68}
69
70void vpx_dc_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
71                                const uint8_t *above, const uint8_t *left) {
72  int32_t  expected_dc;
73  int32_t  average;
74  int32_t  tmp, above1, above_l1, above_r1, left1, left_r1, left_l1;
75  int32_t  above2, above_l2, above_r2, left2, left_r2, left_l2;
76
77  __asm__ __volatile__ (
78      "lw              %[above1],         (%[above])                      \n\t"
79      "lw              %[above2],         4(%[above])                     \n\t"
80      "lw              %[left1],          (%[left])                       \n\t"
81      "lw              %[left2],          4(%[left])                      \n\t"
82
83      "preceu.ph.qbl   %[above_l1],       %[above1]                       \n\t"
84      "preceu.ph.qbr   %[above_r1],       %[above1]                       \n\t"
85      "preceu.ph.qbl   %[left_l1],        %[left1]                        \n\t"
86      "preceu.ph.qbr   %[left_r1],        %[left1]                        \n\t"
87
88      "preceu.ph.qbl   %[above_l2],       %[above2]                       \n\t"
89      "preceu.ph.qbr   %[above_r2],       %[above2]                       \n\t"
90      "preceu.ph.qbl   %[left_l2],        %[left2]                        \n\t"
91      "preceu.ph.qbr   %[left_r2],        %[left2]                        \n\t"
92
93      "addu.ph         %[average],        %[above_r1],      %[above_l1]   \n\t"
94      "addu.ph         %[average],        %[average],       %[left_l1]    \n\t"
95      "addu.ph         %[average],        %[average],       %[left_r1]    \n\t"
96
97      "addu.ph         %[average],        %[average],       %[above_l2]   \n\t"
98      "addu.ph         %[average],        %[average],       %[above_r2]   \n\t"
99      "addu.ph         %[average],        %[average],       %[left_l2]    \n\t"
100      "addu.ph         %[average],        %[average],       %[left_r2]    \n\t"
101
102      "addiu           %[average],        %[average],       8             \n\t"
103
104      "srl             %[tmp],            %[average],       16            \n\t"
105      "addu.ph         %[average],        %[tmp],           %[average]    \n\t"
106      "srl             %[expected_dc],    %[average],       4             \n\t"
107      "replv.qb        %[expected_dc],    %[expected_dc]                  \n\t"
108
109      "sw              %[expected_dc],    (%[dst])                        \n\t"
110      "sw              %[expected_dc],    4(%[dst])                       \n\t"
111
112      "add             %[dst],             %[dst],          %[stride]     \n\t"
113      "sw              %[expected_dc],    (%[dst])                        \n\t"
114      "sw              %[expected_dc],    4(%[dst])                       \n\t"
115
116      "add             %[dst],             %[dst],          %[stride]     \n\t"
117      "sw              %[expected_dc],    (%[dst])                        \n\t"
118      "sw              %[expected_dc],    4(%[dst])                       \n\t"
119
120      "add             %[dst],             %[dst],          %[stride]     \n\t"
121      "sw              %[expected_dc],    (%[dst])                        \n\t"
122      "sw              %[expected_dc],    4(%[dst])                       \n\t"
123
124      "add             %[dst],             %[dst],          %[stride]     \n\t"
125      "sw              %[expected_dc],    (%[dst])                        \n\t"
126      "sw              %[expected_dc],    4(%[dst])                       \n\t"
127
128      "add             %[dst],             %[dst],          %[stride]     \n\t"
129      "sw              %[expected_dc],    (%[dst])                        \n\t"
130      "sw              %[expected_dc],    4(%[dst])                       \n\t"
131
132      "add             %[dst],             %[dst],          %[stride]     \n\t"
133      "sw              %[expected_dc],    (%[dst])                        \n\t"
134      "sw              %[expected_dc],    4(%[dst])                       \n\t"
135
136      "add             %[dst],             %[dst],          %[stride]     \n\t"
137      "sw              %[expected_dc],    (%[dst])                        \n\t"
138      "sw              %[expected_dc],    4(%[dst])                       \n\t"
139
140      : [above1] "=&r" (above1), [above_l1] "=&r" (above_l1),
141        [above_r1] "=&r" (above_r1), [left1] "=&r" (left1),
142        [left_l1] "=&r" (left_l1), [left_r1] "=&r" (left_r1),
143        [above2] "=&r" (above2), [above_l2] "=&r" (above_l2),
144        [above_r2] "=&r" (above_r2), [left2] "=&r" (left2),
145        [left_l2] "=&r" (left_l2), [left_r2] "=&r" (left_r2),
146        [average] "=&r" (average), [tmp] "=&r" (tmp),
147        [expected_dc] "=&r" (expected_dc)
148      : [above] "r" (above), [left] "r" (left), [dst] "r" (dst),
149        [stride] "r" (stride)
150  );
151}
152
153void vpx_tm_predictor_8x8_dspr2(uint8_t *dst, ptrdiff_t stride,
154                                const uint8_t *above, const uint8_t *left) {
155  int32_t   abovel, abover;
156  int32_t   abovel_1, abover_1;
157  int32_t   left0;
158  int32_t   res0, res1, res2, res3;
159  int32_t   reshw;
160  int32_t   top_left;
161  uint8_t   *cm = vpx_ff_cropTbl;
162
163  __asm__ __volatile__ (
164      "ulw             %[reshw],       (%[above])                         \n\t"
165      "ulw             %[top_left],    4(%[above])                        \n\t"
166
167      "lbu             %[left0],       (%[left])                          \n\t"
168
169      "preceu.ph.qbl   %[abovel],      %[reshw]                           \n\t"
170      "preceu.ph.qbr   %[abover],      %[reshw]                           \n\t"
171      "preceu.ph.qbl   %[abovel_1],    %[top_left]                        \n\t"
172      "preceu.ph.qbr   %[abover_1],    %[top_left]                        \n\t"
173
174      "lbu             %[top_left],    -1(%[above])                       \n\t"
175      "replv.ph        %[left0],       %[left0]                           \n\t"
176
177      "replv.ph        %[top_left],    %[top_left]                        \n\t"
178
179      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
180      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
181
182      "sll             %[res2],        %[reshw],            16            \n\t"
183      "sra             %[res2],        %[res2],             16            \n\t"
184      "sra             %[res3],        %[reshw],            16            \n\t"
185
186      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
187      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
188
189      "sll             %[res0],        %[reshw],            16            \n\t"
190      "sra             %[res0],        %[res0],             16            \n\t"
191      "sra             %[res1],        %[reshw],            16            \n\t"
192
193      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
194      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
195      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
196      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
197
198      "sb              %[res0],        (%[dst])                           \n\t"
199      "sb              %[res1],        1(%[dst])                          \n\t"
200      "sb              %[res2],        2(%[dst])                          \n\t"
201      "sb              %[res3],        3(%[dst])                          \n\t"
202
203      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
204      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
205
206      "sll             %[res2],        %[reshw],            16            \n\t"
207      "sra             %[res2],        %[res2],             16            \n\t"
208      "sra             %[res3],        %[reshw],            16            \n\t"
209
210      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
211      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
212
213      "sll             %[res0],        %[reshw],            16            \n\t"
214      "sra             %[res0],        %[res0],             16            \n\t"
215      "sra             %[res1],        %[reshw],            16            \n\t"
216
217      "lbu             %[left0],       1(%[left])                         \n\t"
218
219      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
220      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
221      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
222      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
223
224      "sb              %[res0],        4(%[dst])                          \n\t"
225      "sb              %[res1],        5(%[dst])                          \n\t"
226      "sb              %[res2],        6(%[dst])                          \n\t"
227      "sb              %[res3],        7(%[dst])                          \n\t"
228
229      "replv.ph        %[left0],       %[left0]                           \n\t"
230      "add             %[dst],          %[dst],             %[stride]     \n\t"
231
232      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
233      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
234
235      "sll             %[res2],        %[reshw],            16            \n\t"
236      "sra             %[res2],        %[res2],             16            \n\t"
237      "sra             %[res3],        %[reshw],            16            \n\t"
238
239      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
240      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
241
242      "sll             %[res0],        %[reshw],            16            \n\t"
243      "sra             %[res0],        %[res0],             16            \n\t"
244      "sra             %[res1],        %[reshw],            16            \n\t"
245
246      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
247      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
248      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
249      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
250
251      "sb              %[res0],        (%[dst])                           \n\t"
252      "sb              %[res1],        1(%[dst])                          \n\t"
253      "sb              %[res2],        2(%[dst])                          \n\t"
254      "sb              %[res3],        3(%[dst])                          \n\t"
255
256      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
257      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
258
259      "sll             %[res2],        %[reshw],            16            \n\t"
260      "sra             %[res2],        %[res2],             16            \n\t"
261      "sra             %[res3],        %[reshw],            16            \n\t"
262
263      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
264      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
265
266      "sll             %[res0],        %[reshw],            16            \n\t"
267      "sra             %[res0],        %[res0],             16            \n\t"
268      "sra             %[res1],        %[reshw],            16            \n\t"
269
270      "lbu             %[left0],       2(%[left])                         \n\t"
271
272      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
273      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
274      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
275      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
276
277      "sb              %[res0],        4(%[dst])                          \n\t"
278      "sb              %[res1],        5(%[dst])                          \n\t"
279      "sb              %[res2],        6(%[dst])                          \n\t"
280      "sb              %[res3],        7(%[dst])                          \n\t"
281
282      "replv.ph        %[left0],       %[left0]                           \n\t"
283      "add             %[dst],          %[dst],             %[stride]     \n\t"
284
285      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
286      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
287
288      "sll             %[res2],        %[reshw],            16            \n\t"
289      "sra             %[res2],        %[res2],             16            \n\t"
290      "sra             %[res3],        %[reshw],            16            \n\t"
291
292      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
293      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
294
295      "sll             %[res0],        %[reshw],            16            \n\t"
296      "sra             %[res0],        %[res0],             16            \n\t"
297      "sra             %[res1],        %[reshw],            16            \n\t"
298
299      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
300      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
301      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
302      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
303
304      "sb              %[res0],        (%[dst])                           \n\t"
305      "sb              %[res1],        1(%[dst])                          \n\t"
306      "sb              %[res2],        2(%[dst])                          \n\t"
307      "sb              %[res3],        3(%[dst])                          \n\t"
308
309      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
310      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
311
312      "sll             %[res2],        %[reshw],            16            \n\t"
313      "sra             %[res2],        %[res2],             16            \n\t"
314      "sra             %[res3],        %[reshw],            16            \n\t"
315
316      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
317      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
318
319      "sll             %[res0],        %[reshw],            16            \n\t"
320      "sra             %[res0],        %[res0],             16            \n\t"
321      "sra             %[res1],        %[reshw],            16            \n\t"
322
323      "lbu             %[left0],       3(%[left])                         \n\t"
324
325      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
326      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
327      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
328      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
329
330      "sb              %[res0],        4(%[dst])                          \n\t"
331      "sb              %[res1],        5(%[dst])                          \n\t"
332      "sb              %[res2],        6(%[dst])                          \n\t"
333      "sb              %[res3],        7(%[dst])                          \n\t"
334
335      "replv.ph        %[left0],       %[left0]                           \n\t"
336      "add             %[dst],          %[dst],             %[stride]     \n\t"
337
338      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
339      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
340
341      "sll             %[res2],        %[reshw],            16            \n\t"
342      "sra             %[res2],        %[res2],             16            \n\t"
343      "sra             %[res3],        %[reshw],            16            \n\t"
344
345      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
346      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
347
348      "sll             %[res0],        %[reshw],            16            \n\t"
349      "sra             %[res0],        %[res0],             16            \n\t"
350      "sra             %[res1],        %[reshw],            16            \n\t"
351
352      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
353      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
354      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
355      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
356
357      "sb              %[res0],        (%[dst])                           \n\t"
358      "sb              %[res1],        1(%[dst])                          \n\t"
359      "sb              %[res2],        2(%[dst])                          \n\t"
360      "sb              %[res3],        3(%[dst])                          \n\t"
361
362      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
363      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
364
365      "sll             %[res2],        %[reshw],            16            \n\t"
366      "sra             %[res2],        %[res2],             16            \n\t"
367      "sra             %[res3],        %[reshw],            16            \n\t"
368
369      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
370      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
371
372      "sll             %[res0],        %[reshw],            16            \n\t"
373      "sra             %[res0],        %[res0],             16            \n\t"
374      "sra             %[res1],        %[reshw],            16            \n\t"
375
376      "lbu             %[left0],       4(%[left])                         \n\t"
377
378      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
379      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
380      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
381      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
382
383      "sb              %[res0],        4(%[dst])                          \n\t"
384      "sb              %[res1],        5(%[dst])                          \n\t"
385      "sb              %[res2],        6(%[dst])                          \n\t"
386      "sb              %[res3],        7(%[dst])                          \n\t"
387
388      "replv.ph        %[left0],       %[left0]                           \n\t"
389      "add             %[dst],          %[dst],             %[stride]     \n\t"
390
391      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
392      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
393
394      "sll             %[res2],        %[reshw],            16            \n\t"
395      "sra             %[res2],        %[res2],             16            \n\t"
396      "sra             %[res3],        %[reshw],            16            \n\t"
397
398      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
399      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
400
401      "sll             %[res0],        %[reshw],            16            \n\t"
402      "sra             %[res0],        %[res0],             16            \n\t"
403      "sra             %[res1],        %[reshw],            16            \n\t"
404
405      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
406      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
407      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
408      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
409
410      "sb              %[res0],        (%[dst])                           \n\t"
411      "sb              %[res1],        1(%[dst])                          \n\t"
412      "sb              %[res2],        2(%[dst])                          \n\t"
413      "sb              %[res3],        3(%[dst])                          \n\t"
414
415      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
416      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
417
418      "sll             %[res2],        %[reshw],            16            \n\t"
419      "sra             %[res2],        %[res2],             16            \n\t"
420      "sra             %[res3],        %[reshw],            16            \n\t"
421
422      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
423      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
424
425      "sll             %[res0],        %[reshw],            16            \n\t"
426      "sra             %[res0],        %[res0],             16            \n\t"
427      "sra             %[res1],        %[reshw],            16            \n\t"
428
429      "lbu             %[left0],       5(%[left])                         \n\t"
430
431      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
432      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
433      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
434      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
435
436      "sb              %[res0],        4(%[dst])                          \n\t"
437      "sb              %[res1],        5(%[dst])                          \n\t"
438      "sb              %[res2],        6(%[dst])                          \n\t"
439      "sb              %[res3],        7(%[dst])                          \n\t"
440
441      "replv.ph        %[left0],       %[left0]                           \n\t"
442      "add             %[dst],          %[dst],             %[stride]     \n\t"
443
444      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
445      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
446
447      "sll             %[res2],        %[reshw],            16            \n\t"
448      "sra             %[res2],        %[res2],             16            \n\t"
449      "sra             %[res3],        %[reshw],            16            \n\t"
450
451      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
452      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
453
454      "sll             %[res0],        %[reshw],            16            \n\t"
455      "sra             %[res0],        %[res0],             16            \n\t"
456      "sra             %[res1],        %[reshw],            16            \n\t"
457
458      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
459      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
460      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
461      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
462
463      "sb              %[res0],        (%[dst])                           \n\t"
464      "sb              %[res1],        1(%[dst])                          \n\t"
465      "sb              %[res2],        2(%[dst])                          \n\t"
466      "sb              %[res3],        3(%[dst])                          \n\t"
467
468      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
469      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
470
471      "sll             %[res2],        %[reshw],            16            \n\t"
472      "sra             %[res2],        %[res2],             16            \n\t"
473      "sra             %[res3],        %[reshw],            16            \n\t"
474
475      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
476      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
477
478      "sll             %[res0],        %[reshw],            16            \n\t"
479      "sra             %[res0],        %[res0],             16            \n\t"
480      "sra             %[res1],        %[reshw],            16            \n\t"
481
482      "lbu             %[left0],       6(%[left])                         \n\t"
483
484      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
485      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
486      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
487      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
488
489      "sb              %[res0],        4(%[dst])                          \n\t"
490      "sb              %[res1],        5(%[dst])                          \n\t"
491      "sb              %[res2],        6(%[dst])                          \n\t"
492      "sb              %[res3],        7(%[dst])                          \n\t"
493
494      "replv.ph        %[left0],       %[left0]                           \n\t"
495      "add             %[dst],          %[dst],             %[stride]     \n\t"
496
497      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
498      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
499
500      "sll             %[res2],        %[reshw],            16            \n\t"
501      "sra             %[res2],        %[res2],             16            \n\t"
502      "sra             %[res3],        %[reshw],            16            \n\t"
503
504      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
505      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
506
507      "sll             %[res0],        %[reshw],            16            \n\t"
508      "sra             %[res0],        %[res0],             16            \n\t"
509      "sra             %[res1],        %[reshw],            16            \n\t"
510
511      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
512      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
513      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
514      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
515
516      "sb              %[res0],        (%[dst])                           \n\t"
517      "sb              %[res1],        1(%[dst])                          \n\t"
518      "sb              %[res2],        2(%[dst])                          \n\t"
519      "sb              %[res3],        3(%[dst])                          \n\t"
520
521      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
522      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
523
524      "sll             %[res2],        %[reshw],            16            \n\t"
525      "sra             %[res2],        %[res2],             16            \n\t"
526      "sra             %[res3],        %[reshw],            16            \n\t"
527
528      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
529      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
530
531      "sll             %[res0],        %[reshw],            16            \n\t"
532      "sra             %[res0],        %[res0],             16            \n\t"
533      "sra             %[res1],        %[reshw],            16            \n\t"
534
535      "lbu             %[left0],       7(%[left])                         \n\t"
536
537      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
538      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
539      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
540      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
541
542      "sb              %[res0],        4(%[dst])                          \n\t"
543      "sb              %[res1],        5(%[dst])                          \n\t"
544      "sb              %[res2],        6(%[dst])                          \n\t"
545      "sb              %[res3],        7(%[dst])                          \n\t"
546
547      "replv.ph        %[left0],       %[left0]                           \n\t"
548      "add             %[dst],          %[dst],             %[stride]     \n\t"
549
550      "addu.ph         %[reshw],       %[abovel],           %[left0]      \n\t"
551      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
552
553      "sll             %[res2],        %[reshw],            16            \n\t"
554      "sra             %[res2],        %[res2],             16            \n\t"
555      "sra             %[res3],        %[reshw],            16            \n\t"
556
557      "addu.ph         %[reshw],       %[abover],           %[left0]      \n\t"
558      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
559
560      "sll             %[res0],        %[reshw],            16            \n\t"
561      "sra             %[res0],        %[res0],             16            \n\t"
562      "sra             %[res1],        %[reshw],            16            \n\t"
563
564      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
565      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
566      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
567      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
568
569      "sb              %[res0],        (%[dst])                           \n\t"
570      "sb              %[res1],        1(%[dst])                          \n\t"
571      "sb              %[res2],        2(%[dst])                          \n\t"
572      "sb              %[res3],        3(%[dst])                          \n\t"
573
574      "addu.ph         %[reshw],       %[abovel_1],         %[left0]      \n\t"
575      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
576
577      "sll             %[res2],        %[reshw],            16            \n\t"
578      "sra             %[res2],        %[res2],             16            \n\t"
579      "sra             %[res3],        %[reshw],            16            \n\t"
580
581      "addu.ph         %[reshw],       %[abover_1],         %[left0]      \n\t"
582      "subu.ph         %[reshw],       %[reshw],            %[top_left]   \n\t"
583
584      "sll             %[res0],        %[reshw],            16            \n\t"
585      "sra             %[res0],        %[res0],             16            \n\t"
586      "sra             %[res1],        %[reshw],            16            \n\t"
587
588      "lbux            %[res0],        %[res0](%[cm])                     \n\t"
589      "lbux            %[res1],        %[res1](%[cm])                     \n\t"
590      "lbux            %[res2],        %[res2](%[cm])                     \n\t"
591      "lbux            %[res3],        %[res3](%[cm])                     \n\t"
592
593      "sb              %[res0],        4(%[dst])                          \n\t"
594      "sb              %[res1],        5(%[dst])                          \n\t"
595      "sb              %[res2],        6(%[dst])                          \n\t"
596      "sb              %[res3],        7(%[dst])                          \n\t"
597
598      : [abovel] "=&r" (abovel), [abover] "=&r" (abover),
599        [abovel_1] "=&r" (abovel_1), [abover_1] "=&r" (abover_1),
600        [left0] "=&r" (left0), [res2] "=&r" (res2), [res3] "=&r" (res3),
601        [res0] "=&r" (res0), [res1] "=&r" (res1),
602        [reshw] "=&r" (reshw), [top_left] "=&r" (top_left)
603      : [above] "r" (above), [left] "r" (left),
604        [dst] "r" (dst), [stride] "r" (stride), [cm] "r" (cm)
605  );
606}
607#endif  // #if HAVE_DSPR2
608