2 * Loongson SIMD optimized h264dsp
4 * Copyright (c) 2015 Loongson Technology Corporation Limited
5 * Copyright (c) 2015 Zhou Xiaoyong <zhouxiaoyong@loongson.cn>
6 * Zhang Shuangshuang <zhangshuangshuang@ict.ac.cn>
9 * This file is part of FFmpeg.
11 * FFmpeg is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU Lesser General Public
13 * License as published by the Free Software Foundation; either
14 * version 2.1 of the License, or (at your option) any later version.
16 * FFmpeg is distributed in the hope that it will be useful,
17 * but WITHOUT ANY WARRANTY; without even the implied warranty of
18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
19 * Lesser General Public License for more details.
21 * You should have received a copy of the GNU Lesser General Public
22 * License along with FFmpeg; if not, write to the Free Software
23 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
26 #include "libavcodec/bit_depth_template.c"
27 #include "h264dsp_mips.h"
28 #include "libavutil/mips/mmiutils.h"
30 void ff_h264_add_pixels4_8_mmi(uint8_t *dst, int16_t *src, int stride)
36 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
37 MMI_LDC1(%[ftmp1], %[src], 0x00)
38 MMI_LDC1(%[ftmp2], %[src], 0x08)
39 MMI_LDC1(%[ftmp3], %[src], 0x10)
40 MMI_LDC1(%[ftmp4], %[src], 0x18)
41 MMI_ULWC1(%[ftmp5], %[dst0], 0x00)
42 MMI_ULWC1(%[ftmp6], %[dst1], 0x00)
43 MMI_ULWC1(%[ftmp7], %[dst2], 0x00)
44 MMI_ULWC1(%[ftmp8], %[dst3], 0x00)
45 "punpcklbh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
46 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
47 "punpcklbh %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
48 "punpcklbh %[ftmp8], %[ftmp8], %[ftmp0] \n\t"
49 "paddh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
50 "paddh %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
51 "paddh %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
52 "paddh %[ftmp4], %[ftmp4], %[ftmp8] \n\t"
53 "packushb %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
54 "packushb %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
55 "packushb %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
56 "packushb %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
57 MMI_SWC1(%[ftmp1], %[dst0], 0x00)
58 MMI_SWC1(%[ftmp2], %[dst1], 0x00)
59 MMI_SWC1(%[ftmp3], %[dst2], 0x00)
60 MMI_SWC1(%[ftmp4], %[dst3], 0x00)
62 /* memset(src, 0, 32); */
63 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
64 "gssqc1 %[ftmp0], %[ftmp0], 0x00(%[src]) \n\t"
65 "gssqc1 %[ftmp0], %[ftmp0], 0x10(%[src]) \n\t"
66 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
67 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
68 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
69 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
72 : [dst0]"r"(dst), [dst1]"r"(dst+stride),
73 [dst2]"r"(dst+2*stride), [dst3]"r"(dst+3*stride),
80 void ff_h264_idct_add_8_mmi(uint8_t *dst, int16_t *block, int stride)
88 "dli %[tmp0], 0x01 \n\t"
89 MMI_LDC1(%[ftmp0], %[block], 0x00)
90 "mtc1 %[tmp0], %[ftmp8] \n\t"
91 MMI_LDC1(%[ftmp1], %[block], 0x08)
92 "dli %[tmp0], 0x06 \n\t"
93 MMI_LDC1(%[ftmp2], %[block], 0x10)
94 "mtc1 %[tmp0], %[ftmp9] \n\t"
95 "psrah %[ftmp4], %[ftmp1], %[ftmp8] \n\t"
96 MMI_LDC1(%[ftmp3], %[block], 0x18)
97 "psrah %[ftmp5], %[ftmp3], %[ftmp8] \n\t"
98 "psubh %[ftmp4], %[ftmp4], %[ftmp3] \n\t"
99 "paddh %[ftmp5], %[ftmp5], %[ftmp1] \n\t"
100 "paddh %[ftmp10], %[ftmp2], %[ftmp0] \n\t"
101 "psubh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
102 "paddh %[ftmp11], %[ftmp5], %[ftmp10] \n\t"
103 "psubh %[ftmp2], %[ftmp10], %[ftmp5] \n\t"
104 "paddh %[ftmp10], %[ftmp4], %[ftmp0] \n\t"
105 "psubh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
106 "punpckhhw %[ftmp1], %[ftmp11], %[ftmp10] \n\t"
107 "punpcklhw %[ftmp5], %[ftmp11], %[ftmp10] \n\t"
108 "punpckhhw %[ftmp4], %[ftmp0], %[ftmp2] \n\t"
109 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
110 "punpckhwd %[ftmp2], %[ftmp5], %[ftmp0] \n\t"
111 "punpcklwd %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
112 "punpcklwd %[ftmp10], %[ftmp1], %[ftmp4] \n\t"
113 "punpckhwd %[ftmp0], %[ftmp1], %[ftmp4] \n\t"
114 "paddh %[ftmp5], %[ftmp5], %[ff_pw_32] \n\t"
115 "psrah %[ftmp4], %[ftmp2], %[ftmp8] \n\t"
116 "psrah %[ftmp3], %[ftmp0], %[ftmp8] \n\t"
117 "psubh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
118 "paddh %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
119 "paddh %[ftmp1], %[ftmp10], %[ftmp5] \n\t"
120 "psubh %[ftmp5], %[ftmp5], %[ftmp10] \n\t"
121 "paddh %[ftmp10], %[ftmp3], %[ftmp1] \n\t"
122 "psubh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
123 "paddh %[ftmp11], %[ftmp4], %[ftmp5] \n\t"
124 "xor %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
125 "psubh %[ftmp5], %[ftmp5], %[ftmp4] \n\t"
126 MMI_SDC1(%[ftmp7], %[block], 0x00)
127 MMI_SDC1(%[ftmp7], %[block], 0x08)
128 MMI_SDC1(%[ftmp7], %[block], 0x10)
129 MMI_SDC1(%[ftmp7], %[block], 0x18)
130 MMI_ULWC1(%[ftmp2], %[dst], 0x00)
131 "psrah %[ftmp3], %[ftmp10], %[ftmp9] \n\t"
132 MMI_LWXC1(%[ftmp0], %[dst], %[stride], 0x00)
133 "psrah %[ftmp4], %[ftmp11], %[ftmp9] \n\t"
134 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
135 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
136 "paddh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
137 "paddh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
138 "packushb %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
139 "packushb %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
140 MMI_SWC1(%[ftmp2], %[dst], 0x00)
141 MMI_SWXC1(%[ftmp0], %[dst], %[stride], 0x00)
142 PTR_ADDU "%[dst], %[dst], %[stride] \n\t"
143 PTR_ADDU "%[dst], %[dst], %[stride] \n\t"
144 MMI_ULWC1(%[ftmp2], %[dst], 0x00)
145 "psrah %[ftmp5], %[ftmp5], %[ftmp9] \n\t"
146 MMI_LWXC1(%[ftmp0], %[dst], %[stride], 0x00)
147 "psrah %[ftmp1], %[ftmp1], %[ftmp9] \n\t"
148 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
149 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
150 "paddh %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
151 "paddh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
152 "packushb %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
153 MMI_SWC1(%[ftmp2], %[dst], 0x00)
154 "packushb %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
155 MMI_SWXC1(%[ftmp0], %[dst], %[stride], 0x00)
157 /* memset(block, 0, 32) */
158 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
159 "gssqc1 %[ftmp0], %[ftmp0], 0x00(%[block]) \n\t"
160 "gssqc1 %[ftmp0], %[ftmp0], 0x10(%[block]) \n\t"
161 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
162 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
163 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
164 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
165 [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),
166 [ftmp10]"=&f"(ftmp[10]), [ftmp11]"=&f"(ftmp[11]),
170 : [dst]"r"(dst), [block]"r"(block),
171 [stride]"r"((mips_reg)stride), [ff_pw_32]"f"(ff_pw_32)
177 void ff_h264_idct8_add_8_mmi(uint8_t *dst, int16_t *block, int stride)
186 "lhu %[tmp0], 0x00(%[block]) \n\t"
187 PTR_ADDI "$29, $29, -0x20 \n\t"
188 PTR_ADDIU "%[tmp0], %[tmp0], 0x20 \n\t"
189 MMI_LDC1(%[ftmp1], %[block], 0x10)
190 "sh %[tmp0], 0x00(%[block]) \n\t"
191 MMI_LDC1(%[ftmp2], %[block], 0x20)
192 "dli %[tmp0], 0x01 \n\t"
193 MMI_LDC1(%[ftmp3], %[block], 0x30)
194 "mtc1 %[tmp0], %[ftmp8] \n\t"
195 MMI_LDC1(%[ftmp5], %[block], 0x50)
196 MMI_LDC1(%[ftmp6], %[block], 0x60)
197 MMI_LDC1(%[ftmp7], %[block], 0x70)
198 "mov.d %[ftmp0], %[ftmp1] \n\t"
199 "psrah %[ftmp1], %[ftmp1], %[ftmp8] \n\t"
200 "psrah %[ftmp4], %[ftmp5], %[ftmp8] \n\t"
201 "paddh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
202 "paddh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
203 "paddh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
204 "paddh %[ftmp4], %[ftmp4], %[ftmp7] \n\t"
205 "paddh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
206 "psubh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
207 "psubh %[ftmp0], %[ftmp0], %[ftmp3] \n\t"
208 "psubh %[ftmp5], %[ftmp5], %[ftmp3] \n\t"
209 "psrah %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
210 "paddh %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
211 "psubh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
212 "psrah %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
213 "psubh %[ftmp0], %[ftmp0], %[ftmp3] \n\t"
214 "dli %[tmp0], 0x02 \n\t"
215 "psubh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
216 "mtc1 %[tmp0], %[ftmp9] \n\t"
217 "mov.d %[ftmp7], %[ftmp1] \n\t"
218 "psrah %[ftmp1], %[ftmp1], %[ftmp9] \n\t"
219 "psrah %[ftmp3], %[ftmp4], %[ftmp9] \n\t"
220 "paddh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
221 "psrah %[ftmp0], %[ftmp0], %[ftmp9] \n\t"
222 "paddh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
223 "psrah %[ftmp5], %[ftmp5], %[ftmp9] \n\t"
224 "psubh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
225 "psubh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
226 "mov.d %[ftmp5], %[ftmp6] \n\t"
227 "psrah %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
228 "psrah %[ftmp4], %[ftmp2], %[ftmp8] \n\t"
229 "paddh %[ftmp6], %[ftmp6], %[ftmp2] \n\t"
230 "psubh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
231 MMI_LDC1(%[ftmp2], %[block], 0x00)
232 MMI_LDC1(%[ftmp5], %[block], 0x40)
233 "paddh %[ftmp5], %[ftmp5], %[ftmp2] \n\t"
234 "paddh %[ftmp2], %[ftmp2], %[ftmp2] \n\t"
235 "paddh %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
236 "psubh %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
237 "paddh %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
238 "paddh %[ftmp4], %[ftmp4], %[ftmp2] \n\t"
239 "psubh %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
240 "paddh %[ftmp2], %[ftmp2], %[ftmp2] \n\t"
241 "paddh %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
242 "psubh %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
243 "paddh %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
244 "paddh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
245 "psubh %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
246 "paddh %[ftmp4], %[ftmp4], %[ftmp4] \n\t"
247 "paddh %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
248 "psubh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
249 "paddh %[ftmp2], %[ftmp2], %[ftmp2] \n\t"
250 "paddh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
251 "psubh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
252 "paddh %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
253 MMI_SDC1(%[ftmp6], %[block], 0x00)
254 "psubh %[ftmp5], %[ftmp5], %[ftmp1] \n\t"
255 "punpckhhw %[ftmp6], %[ftmp7], %[ftmp0] \n\t"
256 "punpcklhw %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
257 "punpckhhw %[ftmp0], %[ftmp3], %[ftmp1] \n\t"
258 "punpcklhw %[ftmp3], %[ftmp3], %[ftmp1] \n\t"
259 "punpckhwd %[ftmp1], %[ftmp7], %[ftmp3] \n\t"
260 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
261 "punpckhwd %[ftmp3], %[ftmp6], %[ftmp0] \n\t"
262 "punpcklwd %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
263 MMI_LDC1(%[ftmp0], %[block], 0x00)
264 MMI_SDC1(%[ftmp7], $29, 0x00)
265 MMI_SDC1(%[ftmp1], $29, 0x10)
266 "dmfc1 %[tmp1], %[ftmp6] \n\t"
267 "dmfc1 %[tmp3], %[ftmp3] \n\t"
268 "punpckhhw %[ftmp3], %[ftmp5], %[ftmp2] \n\t"
269 "punpcklhw %[ftmp5], %[ftmp5], %[ftmp2] \n\t"
270 "punpckhhw %[ftmp2], %[ftmp4], %[ftmp0] \n\t"
271 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
272 "punpckhwd %[ftmp0], %[ftmp5], %[ftmp4] \n\t"
273 "punpcklwd %[ftmp5], %[ftmp5], %[ftmp4] \n\t"
274 "punpckhwd %[ftmp4], %[ftmp3], %[ftmp2] \n\t"
275 "punpcklwd %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
276 MMI_SDC1(%[ftmp5], $29, 0x08)
277 MMI_SDC1(%[ftmp0], $29, 0x18)
278 "dmfc1 %[tmp2], %[ftmp3] \n\t"
279 "dmfc1 %[tmp4], %[ftmp4] \n\t"
280 MMI_LDC1(%[ftmp1], %[block], 0x18)
281 MMI_LDC1(%[ftmp6], %[block], 0x28)
282 MMI_LDC1(%[ftmp2], %[block], 0x38)
283 MMI_LDC1(%[ftmp0], %[block], 0x58)
284 MMI_LDC1(%[ftmp3], %[block], 0x68)
285 MMI_LDC1(%[ftmp4], %[block], 0x78)
286 "mov.d %[ftmp7], %[ftmp1] \n\t"
287 "psrah %[ftmp5], %[ftmp0], %[ftmp8] \n\t"
288 "psrah %[ftmp1], %[ftmp1], %[ftmp8] \n\t"
289 "paddh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
290 "paddh %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
291 "paddh %[ftmp5], %[ftmp5], %[ftmp4] \n\t"
292 "paddh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
293 "psubh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
294 "paddh %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
295 "psubh %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
296 "psubh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
297 "psrah %[ftmp2], %[ftmp2], %[ftmp8] \n\t"
298 "paddh %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
299 "psubh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
300 "psrah %[ftmp4], %[ftmp4], %[ftmp8] \n\t"
301 "psubh %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
302 "psubh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
303 "mov.d %[ftmp4], %[ftmp1] \n\t"
304 "psrah %[ftmp2], %[ftmp5], %[ftmp9] \n\t"
305 "psrah %[ftmp1], %[ftmp1], %[ftmp9] \n\t"
306 "paddh %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
307 "psrah %[ftmp7], %[ftmp7], %[ftmp9] \n\t"
308 "paddh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
309 "psrah %[ftmp0], %[ftmp0], %[ftmp9] \n\t"
310 "psubh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
311 "psubh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
312 "mov.d %[ftmp0], %[ftmp3] \n\t"
313 "psrah %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
314 "psrah %[ftmp5], %[ftmp6], %[ftmp8] \n\t"
315 "paddh %[ftmp3], %[ftmp3], %[ftmp6] \n\t"
316 "psubh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
317 MMI_LDC1(%[ftmp6], %[block], 0x08)
318 MMI_LDC1(%[ftmp0], %[block], 0x48)
319 "paddh %[ftmp0], %[ftmp0], %[ftmp6] \n\t"
320 "paddh %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
321 "paddh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
322 "psubh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
323 "paddh %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
324 "paddh %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
325 "psubh %[ftmp0], %[ftmp0], %[ftmp3] \n\t"
326 "paddh %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
327 "paddh %[ftmp4], %[ftmp4], %[ftmp3] \n\t"
328 "psubh %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
329 "paddh %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
330 "paddh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
331 "psubh %[ftmp3], %[ftmp3], %[ftmp4] \n\t"
332 "paddh %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
333 "paddh %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
334 "psubh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
335 "paddh %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
336 "paddh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
337 "psubh %[ftmp6], %[ftmp6], %[ftmp2] \n\t"
338 "paddh %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
339 MMI_SDC1(%[ftmp3], %[block], 0x08)
340 "psubh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
341 "punpckhhw %[ftmp3], %[ftmp4], %[ftmp7] \n\t"
342 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp7] \n\t"
343 "punpckhhw %[ftmp7], %[ftmp2], %[ftmp1] \n\t"
344 "punpcklhw %[ftmp2], %[ftmp2], %[ftmp1] \n\t"
345 "punpckhwd %[ftmp1], %[ftmp4], %[ftmp2] \n\t"
346 "punpcklwd %[ftmp4], %[ftmp4], %[ftmp2] \n\t"
347 "punpckhwd %[ftmp2], %[ftmp3], %[ftmp7] \n\t"
348 "punpcklwd %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
349 MMI_LDC1(%[ftmp7], %[block], 0x08)
350 "dmfc1 %[tmp5], %[ftmp4] \n\t"
351 "mov.d %[ftmp10], %[ftmp1] \n\t"
352 "mov.d %[ftmp12], %[ftmp3] \n\t"
353 "mov.d %[ftmp14], %[ftmp2] \n\t"
354 "punpckhhw %[ftmp2], %[ftmp0], %[ftmp6] \n\t"
355 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp6] \n\t"
356 "punpckhhw %[ftmp6], %[ftmp5], %[ftmp7] \n\t"
357 "punpcklhw %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
358 "punpckhwd %[ftmp7], %[ftmp0], %[ftmp5] \n\t"
359 "punpcklwd %[ftmp0], %[ftmp0], %[ftmp5] \n\t"
360 "punpckhwd %[ftmp5], %[ftmp2], %[ftmp6] \n\t"
361 "punpcklwd %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
362 "dmfc1 %[tmp6], %[ftmp0] \n\t"
363 "mov.d %[ftmp11], %[ftmp7] \n\t"
364 "mov.d %[ftmp13], %[ftmp2] \n\t"
365 "mov.d %[ftmp15], %[ftmp5] \n\t"
366 PTR_ADDIU "%[addr0], %[dst], 0x04 \n\t"
367 "mov.d %[ftmp7], %[ftmp10] \n\t"
368 "dmtc1 %[tmp3], %[ftmp6] \n\t"
369 MMI_LDC1(%[ftmp1], $29, 0x10)
370 "dmtc1 %[tmp1], %[ftmp3] \n\t"
371 "mov.d %[ftmp4], %[ftmp1] \n\t"
372 "psrah %[ftmp1], %[ftmp1], %[ftmp8] \n\t"
373 "psrah %[ftmp0], %[ftmp7], %[ftmp8] \n\t"
374 "paddh %[ftmp1], %[ftmp1], %[ftmp4] \n\t"
375 "paddh %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
376 "paddh %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
377 "paddh %[ftmp0], %[ftmp0], %[ftmp14] \n\t"
378 "paddh %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
379 "psubh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
380 "psubh %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
381 "psubh %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
382 "psrah %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
383 "paddh %[ftmp4], %[ftmp4], %[ftmp14] \n\t"
384 "psubh %[ftmp7], %[ftmp7], %[ftmp14] \n\t"
385 "psrah %[ftmp5], %[ftmp14], %[ftmp8] \n\t"
386 "psubh %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
387 "psubh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
388 "mov.d %[ftmp5], %[ftmp1] \n\t"
389 "psrah %[ftmp1], %[ftmp1], %[ftmp9] \n\t"
390 "psrah %[ftmp6], %[ftmp0], %[ftmp9] \n\t"
391 "paddh %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
392 "paddh %[ftmp6], %[ftmp6], %[ftmp4] \n\t"
393 "psrah %[ftmp4], %[ftmp4], %[ftmp9] \n\t"
394 "psrah %[ftmp7], %[ftmp7], %[ftmp9] \n\t"
395 "psubh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
396 "psubh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
397 "mov.d %[ftmp7], %[ftmp12] \n\t"
398 "psrah %[ftmp2], %[ftmp12], %[ftmp8] \n\t"
399 "psrah %[ftmp0], %[ftmp3], %[ftmp8] \n\t"
400 "paddh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
401 "psubh %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
402 MMI_LDC1(%[ftmp3], $29, 0x00)
403 "dmtc1 %[tmp5], %[ftmp7] \n\t"
404 "paddh %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
405 "paddh %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
406 "paddh %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
407 "psubh %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
408 "paddh %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
409 "paddh %[ftmp0], %[ftmp0], %[ftmp3] \n\t"
410 "psubh %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
411 "paddh %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
412 "paddh %[ftmp5], %[ftmp5], %[ftmp2] \n\t"
413 "psubh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
414 "paddh %[ftmp2], %[ftmp2], %[ftmp2] \n\t"
415 "paddh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
416 "psubh %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
417 "paddh %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
418 "paddh %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
419 "psubh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
420 "paddh %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
421 "paddh %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
422 "psubh %[ftmp3], %[ftmp3], %[ftmp6] \n\t"
423 "paddh %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
424 MMI_SDC1(%[ftmp3], $29, 0x00)
425 "psubh %[ftmp7], %[ftmp7], %[ftmp1] \n\t"
426 MMI_SDC1(%[ftmp0], $29, 0x10)
427 "dmfc1 %[tmp1], %[ftmp2] \n\t"
428 "xor %[ftmp2], %[ftmp2], %[ftmp2] \n\t"
429 MMI_SDC1(%[ftmp2], %[block], 0x00)
430 MMI_SDC1(%[ftmp2], %[block], 0x08)
431 MMI_SDC1(%[ftmp2], %[block], 0x10)
432 MMI_SDC1(%[ftmp2], %[block], 0x18)
433 MMI_SDC1(%[ftmp2], %[block], 0x20)
434 MMI_SDC1(%[ftmp2], %[block], 0x28)
435 MMI_SDC1(%[ftmp2], %[block], 0x30)
436 MMI_SDC1(%[ftmp2], %[block], 0x38)
437 MMI_SDC1(%[ftmp2], %[block], 0x40)
438 MMI_SDC1(%[ftmp2], %[block], 0x48)
439 MMI_SDC1(%[ftmp2], %[block], 0x50)
440 MMI_SDC1(%[ftmp2], %[block], 0x58)
441 MMI_SDC1(%[ftmp2], %[block], 0x60)
442 MMI_SDC1(%[ftmp2], %[block], 0x68)
443 MMI_SDC1(%[ftmp2], %[block], 0x70)
444 MMI_SDC1(%[ftmp2], %[block], 0x78)
445 "dli %[tmp3], 0x06 \n\t"
446 "mtc1 %[tmp3], %[ftmp10] \n\t"
447 MMI_ULWC1(%[ftmp3], %[dst], 0x00)
448 MMI_LWXC1(%[ftmp0], %[dst], %[stride], 0x00)
449 "psrah %[ftmp5], %[ftmp5], %[ftmp10] \n\t"
450 "psrah %[ftmp4], %[ftmp4], %[ftmp10] \n\t"
451 "punpcklbh %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
452 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
453 "paddh %[ftmp3], %[ftmp3], %[ftmp5] \n\t"
454 "paddh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
455 "packushb %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
456 "packushb %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
457 MMI_SWC1(%[ftmp3], %[dst], 0x00)
458 MMI_SWXC1(%[ftmp0], %[dst], %[stride], 0x00)
459 PTR_ADDU "%[dst], %[dst], %[stride] \n\t"
460 PTR_ADDU "%[dst], %[dst], %[stride] \n\t"
461 MMI_ULWC1(%[ftmp3], %[dst], 0x00)
462 MMI_LWXC1(%[ftmp0], %[dst], %[stride], 0x00)
463 "psrah %[ftmp6], %[ftmp6], %[ftmp10] \n\t"
464 "psrah %[ftmp1], %[ftmp1], %[ftmp10] \n\t"
465 "punpcklbh %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
466 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
467 "paddh %[ftmp3], %[ftmp3], %[ftmp6] \n\t"
468 "paddh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
469 "packushb %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
470 "packushb %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
471 MMI_SWC1(%[ftmp3], %[dst], 0x00)
472 MMI_SWXC1(%[ftmp0], %[dst], %[stride], 0x00)
473 MMI_LDC1(%[ftmp5], $29, 0x00)
474 MMI_LDC1(%[ftmp4], $29, 0x10)
475 "dmtc1 %[tmp1], %[ftmp6] \n\t"
476 PTR_ADDU "%[dst], %[dst], %[stride] \n\t"
477 PTR_ADDU "%[dst], %[dst], %[stride] \n\t"
478 MMI_ULWC1(%[ftmp3], %[dst], 0x00)
479 MMI_LWXC1(%[ftmp0], %[dst], %[stride], 0x00)
480 "psrah %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
481 "psrah %[ftmp5], %[ftmp5], %[ftmp10] \n\t"
482 "punpcklbh %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
483 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
484 "paddh %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
485 "paddh %[ftmp0], %[ftmp0], %[ftmp5] \n\t"
486 "packushb %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
487 "packushb %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
488 MMI_SWC1(%[ftmp3], %[dst], 0x00)
489 MMI_SWXC1(%[ftmp0], %[dst], %[stride], 0x00)
490 PTR_ADDU "%[dst], %[dst], %[stride] \n\t"
491 PTR_ADDU "%[dst], %[dst], %[stride] \n\t"
492 MMI_ULWC1(%[ftmp3], %[dst], 0x00)
493 MMI_LWXC1(%[ftmp0], %[dst], %[stride], 0x00)
494 "psrah %[ftmp4], %[ftmp4], %[ftmp10] \n\t"
495 "psrah %[ftmp6], %[ftmp6], %[ftmp10] \n\t"
496 "punpcklbh %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
497 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
498 "paddh %[ftmp3], %[ftmp3], %[ftmp4] \n\t"
499 "paddh %[ftmp0], %[ftmp0], %[ftmp6] \n\t"
500 "packushb %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
501 "packushb %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
502 MMI_SWC1(%[ftmp3], %[dst], 0x00)
503 MMI_SWXC1(%[ftmp0], %[dst], %[stride], 0x00)
504 "dmtc1 %[tmp4], %[ftmp1] \n\t"
505 "dmtc1 %[tmp2], %[ftmp6] \n\t"
506 MMI_LDC1(%[ftmp4], $29, 0x18)
507 "mov.d %[ftmp5], %[ftmp4] \n\t"
508 "psrah %[ftmp4], %[ftmp4], %[ftmp8] \n\t"
509 "psrah %[ftmp7], %[ftmp11], %[ftmp8] \n\t"
510 "paddh %[ftmp7], %[ftmp7], %[ftmp11] \n\t"
511 "paddh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
512 "paddh %[ftmp7], %[ftmp7], %[ftmp15] \n\t"
513 "paddh %[ftmp4], %[ftmp4], %[ftmp11] \n\t"
514 "psubh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
515 "paddh %[ftmp4], %[ftmp4], %[ftmp1] \n\t"
516 "psubh %[ftmp5], %[ftmp5], %[ftmp1] \n\t"
517 "psubh %[ftmp3], %[ftmp11], %[ftmp1] \n\t"
518 "psrah %[ftmp1], %[ftmp1], %[ftmp8] \n\t"
519 "paddh %[ftmp5], %[ftmp5], %[ftmp15] \n\t"
520 "psubh %[ftmp3], %[ftmp3], %[ftmp15] \n\t"
521 "psrah %[ftmp2], %[ftmp15], %[ftmp8] \n\t"
522 "psubh %[ftmp5], %[ftmp5], %[ftmp1] \n\t"
523 "psubh %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
524 "mov.d %[ftmp2], %[ftmp4] \n\t"
525 "psrah %[ftmp4], %[ftmp4], %[ftmp9] \n\t"
526 "psrah %[ftmp1], %[ftmp7], %[ftmp9] \n\t"
527 "paddh %[ftmp4], %[ftmp4], %[ftmp3] \n\t"
528 "paddh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
529 "psrah %[ftmp5], %[ftmp5], %[ftmp9] \n\t"
530 "psrah %[ftmp3], %[ftmp3], %[ftmp9] \n\t"
531 "psubh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
532 "psubh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
533 "mov.d %[ftmp3], %[ftmp13] \n\t"
534 "psrah %[ftmp0], %[ftmp13], %[ftmp8] \n\t"
535 "psrah %[ftmp7], %[ftmp6], %[ftmp8] \n\t"
536 "paddh %[ftmp0], %[ftmp0], %[ftmp6] \n\t"
537 "psubh %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
538 MMI_LDC1(%[ftmp6], $29, 0x08)
539 "dmtc1 %[tmp6], %[ftmp3] \n\t"
540 "paddh %[ftmp3], %[ftmp3], %[ftmp6] \n\t"
541 "paddh %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
542 "paddh %[ftmp0], %[ftmp0], %[ftmp3] \n\t"
543 "psubh %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
544 "paddh %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
545 "paddh %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
546 "psubh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
547 "paddh %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
548 "paddh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
549 "psubh %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
550 "paddh %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
551 "paddh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
552 "psubh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
553 "paddh %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
554 "paddh %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
555 "psubh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
556 "paddh %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
557 "paddh %[ftmp4], %[ftmp4], %[ftmp3] \n\t"
558 "psubh %[ftmp6], %[ftmp6], %[ftmp1] \n\t"
559 "paddh %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
560 MMI_SDC1(%[ftmp6], $29, 0x08)
561 "psubh %[ftmp3], %[ftmp3], %[ftmp4] \n\t"
562 MMI_SDC1(%[ftmp7], $29, 0x18)
563 "dmfc1 %[tmp2], %[ftmp0] \n\t"
564 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
565 MMI_ULWC1(%[ftmp6], %[addr0], 0x00)
566 MMI_LWXC1(%[ftmp7], %[addr0], %[stride], 0x00)
567 "psrah %[ftmp2], %[ftmp2], %[ftmp10] \n\t"
568 "psrah %[ftmp5], %[ftmp5], %[ftmp10] \n\t"
569 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
570 "punpcklbh %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
571 "paddh %[ftmp6], %[ftmp6], %[ftmp2] \n\t"
572 "paddh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
573 "packushb %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
574 "packushb %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
575 MMI_SWC1(%[ftmp6], %[addr0], 0x00)
576 MMI_SWXC1(%[ftmp7], %[addr0], %[stride], 0x00)
577 PTR_ADDU "%[addr0], %[addr0], %[stride] \n\t"
578 PTR_ADDU "%[addr0], %[addr0], %[stride] \n\t"
579 MMI_ULWC1(%[ftmp6], %[addr0], 0x00)
580 MMI_LWXC1(%[ftmp7], %[addr0], %[stride], 0x00)
581 "psrah %[ftmp1], %[ftmp1], %[ftmp10] \n\t"
582 "psrah %[ftmp4], %[ftmp4], %[ftmp10] \n\t"
583 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
584 "punpcklbh %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
585 "paddh %[ftmp6], %[ftmp6], %[ftmp1] \n\t"
586 "paddh %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
587 "packushb %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
588 "packushb %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
589 MMI_SWC1(%[ftmp6], %[addr0], 0x00)
590 MMI_SWXC1(%[ftmp7], %[addr0], %[stride], 0x00)
591 MMI_LDC1(%[ftmp2], $29, 0x08)
592 MMI_LDC1(%[ftmp5], $29, 0x18)
593 PTR_ADDU "%[addr0], %[addr0], %[stride] \n\t"
594 "dmtc1 %[tmp2], %[ftmp1] \n\t"
595 PTR_ADDU "%[addr0], %[addr0], %[stride] \n\t"
596 MMI_ULWC1(%[ftmp6], %[addr0], 0x00)
597 MMI_LWXC1(%[ftmp7], %[addr0], %[stride], 0x00)
598 "psrah %[ftmp3], %[ftmp3], %[ftmp10] \n\t"
599 "psrah %[ftmp2], %[ftmp2], %[ftmp10] \n\t"
600 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
601 "punpcklbh %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
602 "paddh %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
603 "paddh %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
604 "packushb %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
605 "packushb %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
606 MMI_SWC1(%[ftmp6], %[addr0], 0x00)
607 MMI_SWXC1(%[ftmp7], %[addr0], %[stride], 0x00)
608 PTR_ADDU "%[addr0], %[addr0], %[stride] \n\t"
609 PTR_ADDU "%[addr0], %[addr0], %[stride] \n\t"
610 MMI_ULWC1(%[ftmp6], %[addr0], 0x00)
611 MMI_LWXC1(%[ftmp7], %[addr0], %[stride], 0x00)
612 "psrah %[ftmp5], %[ftmp5], %[ftmp10] \n\t"
613 "psrah %[ftmp1], %[ftmp1], %[ftmp10] \n\t"
614 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
615 "punpcklbh %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
616 "paddh %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
617 "paddh %[ftmp7], %[ftmp7], %[ftmp1] \n\t"
618 "packushb %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
619 "packushb %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
620 MMI_SWC1(%[ftmp6], %[addr0], 0x00)
621 MMI_SWXC1(%[ftmp7], %[addr0], %[stride], 0x00)
622 PTR_ADDIU "$29, $29, 0x20 \n\t"
624 /* memset(block, 0, 128) */
625 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
626 "gssqc1 %[ftmp0], %[ftmp0], 0x00(%[block]) \n\t"
627 "gssqc1 %[ftmp0], %[ftmp0], 0x10(%[block]) \n\t"
628 "gssqc1 %[ftmp0], %[ftmp0], 0x20(%[block]) \n\t"
629 "gssqc1 %[ftmp0], %[ftmp0], 0x30(%[block]) \n\t"
630 "gssqc1 %[ftmp0], %[ftmp0], 0x40(%[block]) \n\t"
631 "gssqc1 %[ftmp0], %[ftmp0], 0x50(%[block]) \n\t"
632 "gssqc1 %[ftmp0], %[ftmp0], 0x60(%[block]) \n\t"
633 "gssqc1 %[ftmp0], %[ftmp0], 0x70(%[block]) \n\t"
634 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
635 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
636 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
637 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
638 [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),
639 [ftmp10]"=&f"(ftmp[10]), [ftmp11]"=&f"(ftmp[11]),
640 [ftmp12]"=&f"(ftmp[12]), [ftmp13]"=&f"(ftmp[13]),
641 [ftmp14]"=&f"(ftmp[14]), [ftmp15]"=&f"(ftmp[15]),
642 [tmp0]"=&r"(tmp[0]), [tmp1]"=&r"(tmp[1]),
643 [tmp2]"=&r"(tmp[2]), [tmp3]"=&r"(tmp[3]),
644 [tmp4]"=&r"(tmp[4]), [tmp5]"=&r"(tmp[5]),
648 [addr0]"=&r"(addr[0])
649 : [dst]"r"(dst), [block]"r"(block),
650 [stride]"r"((mips_reg)stride)
656 void ff_h264_idct_dc_add_8_mmi(uint8_t *dst, int16_t *block, int stride)
658 int dc = (block[0] + 32) >> 6;
665 "mtc1 %[dc], %[ftmp5] \n\t"
666 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
667 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
668 MMI_ULWC1(%[ftmp1], %[dst0], 0x00)
669 MMI_ULWC1(%[ftmp2], %[dst1], 0x00)
670 MMI_ULWC1(%[ftmp3], %[dst2], 0x00)
671 MMI_ULWC1(%[ftmp4], %[dst3], 0x00)
672 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
673 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
674 "punpcklbh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
675 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
676 "paddsh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
677 "paddsh %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
678 "paddsh %[ftmp3], %[ftmp3], %[ftmp5] \n\t"
679 "paddsh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
680 "packushb %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
681 "packushb %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
682 "packushb %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
683 "packushb %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
684 MMI_SWC1(%[ftmp1], %[dst0], 0x00)
685 MMI_SWC1(%[ftmp2], %[dst1], 0x00)
686 MMI_SWC1(%[ftmp3], %[dst2], 0x00)
687 MMI_SWC1(%[ftmp4], %[dst3], 0x00)
688 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
689 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
690 [ftmp4]"=&f"(ftmp[4]),
692 [ftmp5]"=&f"(ftmp[5])
693 : [dst0]"r"(dst), [dst1]"r"(dst+stride),
694 [dst2]"r"(dst+2*stride), [dst3]"r"(dst+3*stride),
700 void ff_h264_idct8_dc_add_8_mmi(uint8_t *dst, int16_t *block, int stride)
702 int dc = (block[0] + 32) >> 6;
709 "mtc1 %[dc], %[ftmp5] \n\t"
710 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
711 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
712 MMI_LDC1(%[ftmp1], %[dst0], 0x00)
713 MMI_LDC1(%[ftmp2], %[dst1], 0x00)
714 MMI_LDC1(%[ftmp3], %[dst2], 0x00)
715 MMI_LDC1(%[ftmp4], %[dst3], 0x00)
716 "punpckhbh %[ftmp6], %[ftmp1], %[ftmp0] \n\t"
717 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
718 "punpckhbh %[ftmp7], %[ftmp2], %[ftmp0] \n\t"
719 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
720 "punpckhbh %[ftmp8], %[ftmp3], %[ftmp0] \n\t"
721 "punpcklbh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
722 "punpckhbh %[ftmp9], %[ftmp4], %[ftmp0] \n\t"
723 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
724 "paddsh %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
725 "paddsh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
726 "paddsh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
727 "paddsh %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
728 "paddsh %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
729 "paddsh %[ftmp3], %[ftmp3], %[ftmp5] \n\t"
730 "paddsh %[ftmp9], %[ftmp9], %[ftmp5] \n\t"
731 "paddsh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
732 "packushb %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
733 "packushb %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
734 "packushb %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
735 "packushb %[ftmp4], %[ftmp4], %[ftmp9] \n\t"
736 MMI_SDC1(%[ftmp1], %[dst0], 0x00)
737 MMI_SDC1(%[ftmp2], %[dst1], 0x00)
738 MMI_SDC1(%[ftmp3], %[dst2], 0x00)
739 MMI_SDC1(%[ftmp4], %[dst3], 0x00)
741 MMI_LDC1(%[ftmp1], %[dst4], 0x00)
742 MMI_LDC1(%[ftmp2], %[dst5], 0x00)
743 MMI_LDC1(%[ftmp3], %[dst6], 0x00)
744 MMI_LDC1(%[ftmp4], %[dst7], 0x00)
745 "punpckhbh %[ftmp6], %[ftmp1], %[ftmp0] \n\t"
746 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
747 "punpckhbh %[ftmp7], %[ftmp2], %[ftmp0] \n\t"
748 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
749 "punpckhbh %[ftmp8], %[ftmp3], %[ftmp0] \n\t"
750 "punpcklbh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
751 "punpckhbh %[ftmp9], %[ftmp4], %[ftmp0] \n\t"
752 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
753 "paddsh %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
754 "paddsh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
755 "paddsh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
756 "paddsh %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
757 "paddsh %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
758 "paddsh %[ftmp3], %[ftmp3], %[ftmp5] \n\t"
759 "paddsh %[ftmp9], %[ftmp9], %[ftmp5] \n\t"
760 "paddsh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
761 "packushb %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
762 "packushb %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
763 "packushb %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
764 "packushb %[ftmp4], %[ftmp4], %[ftmp9] \n\t"
765 MMI_SDC1(%[ftmp1], %[dst4], 0x00)
766 MMI_SDC1(%[ftmp2], %[dst5], 0x00)
767 MMI_SDC1(%[ftmp3], %[dst6], 0x00)
768 MMI_SDC1(%[ftmp4], %[dst7], 0x00)
769 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
770 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
771 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
772 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
773 [ftmp8]"=&f"(ftmp[8]),
775 [ftmp9]"=&f"(ftmp[9])
776 : [dst0]"r"(dst), [dst1]"r"(dst+stride),
777 [dst2]"r"(dst+2*stride), [dst3]"r"(dst+3*stride),
778 [dst4]"r"(dst+4*stride), [dst5]"r"(dst+5*stride),
779 [dst6]"r"(dst+6*stride), [dst7]"r"(dst+7*stride),
785 void ff_h264_idct_add16_8_mmi(uint8_t *dst, const int *block_offset,
786 int16_t *block, int stride, const uint8_t nnzc[15*8])
790 int nnz = nnzc[ scan8[i] ];
792 if(nnz==1 && ((int16_t*)block)[i*16])
793 ff_h264_idct_dc_add_8_mmi(dst + block_offset[i], block + i*16,
796 ff_h264_idct_add_8_mmi(dst + block_offset[i], block + i*16,
802 void ff_h264_idct_add16intra_8_mmi(uint8_t *dst, const int *block_offset,
803 int16_t *block, int stride, const uint8_t nnzc[15*8])
808 ff_h264_idct_add_8_mmi(dst + block_offset[i], block + i*16, stride);
809 else if(((int16_t*)block)[i*16])
810 ff_h264_idct_dc_add_8_mmi(dst + block_offset[i], block + i*16,
815 void ff_h264_idct8_add4_8_mmi(uint8_t *dst, const int *block_offset,
816 int16_t *block, int stride, const uint8_t nnzc[15*8])
819 for(i=0; i<16; i+=4){
820 int nnz = nnzc[ scan8[i] ];
822 if(nnz==1 && ((int16_t*)block)[i*16])
823 ff_h264_idct8_dc_add_8_mmi(dst + block_offset[i],
824 block + i*16, stride);
826 ff_h264_idct8_add_8_mmi(dst + block_offset[i], block + i*16,
832 void ff_h264_idct_add8_8_mmi(uint8_t **dest, const int *block_offset,
833 int16_t *block, int stride, const uint8_t nnzc[15*8])
837 for(i=j*16; i<j*16+4; i++){
839 ff_h264_idct_add_8_mmi(dest[j-1] + block_offset[i],
840 block + i*16, stride);
841 else if(((int16_t*)block)[i*16])
842 ff_h264_idct_dc_add_8_mmi(dest[j-1] + block_offset[i],
843 block + i*16, stride);
848 void ff_h264_idct_add8_422_8_mmi(uint8_t **dest, const int *block_offset,
849 int16_t *block, int stride, const uint8_t nnzc[15*8])
854 for(i=j*16; i<j*16+4; i++){
856 ff_h264_idct_add_8_mmi(dest[j-1] + block_offset[i],
857 block + i*16, stride);
858 else if(((int16_t*)block)[i*16])
859 ff_h264_idct_dc_add_8_mmi(dest[j-1] + block_offset[i],
860 block + i*16, stride);
865 for(i=j*16+4; i<j*16+8; i++){
866 if(nnzc[ scan8[i+4] ])
867 ff_h264_idct_add_8_mmi(dest[j-1] + block_offset[i+4],
868 block + i*16, stride);
869 else if(((int16_t*)block)[i*16])
870 ff_h264_idct_dc_add_8_mmi(dest[j-1] + block_offset[i+4],
871 block + i*16, stride);
876 void ff_h264_luma_dc_dequant_idct_8_mmi(int16_t *output, int16_t *input,
884 ".set noreorder \n\t"
885 "dli %[tmp0], 0x08 \n\t"
886 MMI_LDC1(%[ftmp3], %[input], 0x18)
887 "mtc1 %[tmp0], %[ftmp8] \n\t"
888 MMI_LDC1(%[ftmp2], %[input], 0x10)
889 "dli %[tmp0], 0x20 \n\t"
890 MMI_LDC1(%[ftmp1], %[input], 0x08)
891 "mtc1 %[tmp0], %[ftmp9] \n\t"
892 MMI_LDC1(%[ftmp0], %[input], 0x00)
893 "mov.d %[ftmp4], %[ftmp3] \n\t"
894 "paddh %[ftmp3], %[ftmp3], %[ftmp2] \n\t"
895 "psubh %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
896 "mov.d %[ftmp4], %[ftmp1] \n\t"
897 "paddh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
898 "psubh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
899 "mov.d %[ftmp4], %[ftmp3] \n\t"
900 "paddh %[ftmp3], %[ftmp3], %[ftmp1] \n\t"
901 "psubh %[ftmp1], %[ftmp1], %[ftmp4] \n\t"
902 "mov.d %[ftmp4], %[ftmp2] \n\t"
903 "paddh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
904 "psubh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
905 "mov.d %[ftmp4], %[ftmp3] \n\t"
906 "punpcklhw %[ftmp3], %[ftmp3], %[ftmp1] \n\t"
907 "punpckhhw %[ftmp4], %[ftmp4], %[ftmp1] \n\t"
908 "punpckhhw %[ftmp1], %[ftmp0], %[ftmp2] \n\t"
909 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
910 "punpckhwd %[ftmp2], %[ftmp3], %[ftmp0] \n\t"
911 "punpcklwd %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
912 "mov.d %[ftmp0], %[ftmp4] \n\t"
913 "punpcklwd %[ftmp4], %[ftmp4], %[ftmp1] \n\t"
914 "punpckhwd %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
915 "mov.d %[ftmp1], %[ftmp0] \n\t"
916 "paddh %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
917 "psubh %[ftmp4], %[ftmp4], %[ftmp1] \n\t"
918 "mov.d %[ftmp1], %[ftmp2] \n\t"
919 "paddh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
920 "psubh %[ftmp3], %[ftmp3], %[ftmp1] \n\t"
921 "mov.d %[ftmp1], %[ftmp0] \n\t"
922 "paddh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
923 "psubh %[ftmp2], %[ftmp2], %[ftmp1] \n\t"
924 "mov.d %[ftmp1], %[ftmp4] \n\t"
925 "daddi %[tmp0], %[qmul], -0x7fff \n\t"
926 "paddh %[ftmp4], %[ftmp4], %[ftmp3] \n\t"
927 "bgtz %[tmp0], 1f \n\t"
928 "psubh %[ftmp3], %[ftmp3], %[ftmp1] \n\t"
929 "ori %[tmp0], $0, 0x80 \n\t"
930 "dsll %[tmp0], %[tmp0], 0x10 \n\t"
931 "punpckhhw %[ftmp1], %[ftmp0], %[ff_pw_1] \n\t"
932 "daddu %[qmul], %[qmul], %[tmp0] \n\t"
933 "punpcklhw %[ftmp0], %[ftmp0], %[ff_pw_1] \n\t"
934 "punpckhhw %[ftmp5], %[ftmp2], %[ff_pw_1] \n\t"
935 "punpcklhw %[ftmp2], %[ftmp2], %[ff_pw_1] \n\t"
936 "mtc1 %[qmul], %[ftmp7] \n\t"
937 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
938 "pmaddhw %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
939 "pmaddhw %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
940 "pmaddhw %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
941 "pmaddhw %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
942 "psraw %[ftmp0], %[ftmp0], %[ftmp8] \n\t"
943 "psraw %[ftmp2], %[ftmp2], %[ftmp8] \n\t"
944 "psraw %[ftmp1], %[ftmp1], %[ftmp8] \n\t"
945 "psraw %[ftmp5], %[ftmp5], %[ftmp8] \n\t"
946 "packsswh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
947 "packsswh %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
948 "dmfc1 %[tmp1], %[ftmp0] \n\t"
949 "dsrl %[ftmp0], %[ftmp0], %[ftmp9] \n\t"
950 "mfc1 %[input], %[ftmp0] \n\t"
951 "sh %[tmp1], 0x00(%[output]) \n\t"
952 "sh %[input], 0x80(%[output]) \n\t"
953 "dsrl %[tmp1], %[tmp1], 0x10 \n\t"
954 PTR_SRL "%[input], %[input], 0x10 \n\t"
955 "sh %[tmp1], 0x20(%[output]) \n\t"
956 "sh %[input], 0xa0(%[output]) \n\t"
957 "dmfc1 %[tmp1], %[ftmp2] \n\t"
958 "dsrl %[ftmp2], %[ftmp2], %[ftmp9] \n\t"
959 "mfc1 %[input], %[ftmp2] \n\t"
960 "sh %[tmp1], 0x40(%[output]) \n\t"
961 "sh %[input], 0xc0(%[output]) \n\t"
962 "dsrl %[tmp1], %[tmp1], 0x10 \n\t"
963 PTR_SRL "%[input], %[input], 0x10 \n\t"
964 "sh %[tmp1], 0x60(%[output]) \n\t"
965 "sh %[input], 0xe0(%[output]) \n\t"
966 "punpckhhw %[ftmp1], %[ftmp3], %[ff_pw_1] \n\t"
967 "punpcklhw %[ftmp3], %[ftmp3], %[ff_pw_1] \n\t"
968 "punpckhhw %[ftmp5], %[ftmp4], %[ff_pw_1] \n\t"
969 "punpcklhw %[ftmp4], %[ftmp4], %[ff_pw_1] \n\t"
970 "mtc1 %[qmul], %[ftmp7] \n\t"
971 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
972 "pmaddhw %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
973 "pmaddhw %[ftmp4], %[ftmp4], %[ftmp7] \n\t"
974 "pmaddhw %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
975 "pmaddhw %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
976 "psraw %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
977 "psraw %[ftmp4], %[ftmp4], %[ftmp8] \n\t"
978 "psraw %[ftmp1], %[ftmp1], %[ftmp8] \n\t"
979 "psraw %[ftmp5], %[ftmp5], %[ftmp8] \n\t"
980 "packsswh %[ftmp3], %[ftmp3], %[ftmp1] \n\t"
981 "packsswh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
982 "dmfc1 %[tmp1], %[ftmp3] \n\t"
983 "dsrl %[ftmp3], %[ftmp3], %[ftmp9] \n\t"
984 "mfc1 %[input], %[ftmp3] \n\t"
985 "sh %[tmp1], 0x100(%[output]) \n\t"
986 "sh %[input], 0x180(%[output]) \n\t"
987 "dsrl %[tmp1], %[tmp1], 0x10 \n\t"
988 PTR_SRL "%[input], %[input], 0x10 \n\t"
989 "sh %[tmp1], 0x120(%[output]) \n\t"
990 "sh %[input], 0x1a0(%[output]) \n\t"
991 "dmfc1 %[tmp1], %[ftmp4] \n\t"
992 "dsrl %[ftmp4], %[ftmp4], %[ftmp9] \n\t"
993 "mfc1 %[input], %[ftmp4] \n\t"
994 "sh %[tmp1], 0x140(%[output]) \n\t"
995 "sh %[input], 0x1c0(%[output]) \n\t"
996 "dsrl %[tmp1], %[tmp1], 0x10 \n\t"
997 PTR_SRL "%[input], %[input], 0x10 \n\t"
998 "sh %[tmp1], 0x160(%[output]) \n\t"
1000 "sh %[input], 0x1e0(%[output]) \n\t"
1002 "ori %[tmp0], $0, 0x1f \n\t"
1004 "clz %[tmp1], %[qmul] \n\t"
1005 #elif HAVE_LOONGSON2
1007 "ori %[input], $0, 0x07 \n\t"
1008 "dsubu %[tmp1], %[tmp0], %[tmp1] \n\t"
1009 "ori %[tmp0], $0, 0x80 \n\t"
1010 "dsll %[tmp0], %[tmp0], 0x10 \n\t"
1011 "daddu %[qmul], %[qmul], %[tmp0] \n\t"
1012 "dsubu %[tmp0], %[tmp1], %[input] \n\t"
1013 "movn %[tmp1], %[input], %[tmp0] \n\t"
1014 PTR_ADDIU "%[input], %[input], 0x01 \n\t"
1015 "andi %[tmp0], %[tmp1], 0xff \n\t"
1016 "srlv %[qmul], %[qmul], %[tmp0] \n\t"
1017 PTR_SUBU "%[input], %[input], %[tmp1] \n\t"
1018 "mtc1 %[input], %[ftmp6] \n\t"
1019 "punpckhhw %[ftmp1], %[ftmp0], %[ff_pw_1] \n\t"
1020 "punpcklhw %[ftmp0], %[ftmp0], %[ff_pw_1] \n\t"
1021 "punpckhhw %[ftmp5], %[ftmp2], %[ff_pw_1] \n\t"
1022 "punpcklhw %[ftmp2], %[ftmp2], %[ff_pw_1] \n\t"
1023 "mtc1 %[qmul], %[ftmp7] \n\t"
1024 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
1025 "pmaddhw %[ftmp0], %[ftmp0], %[ftmp7] \n\t"
1026 "pmaddhw %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
1027 "pmaddhw %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
1028 "pmaddhw %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1029 "psraw %[ftmp0], %[ftmp0], %[ftmp6] \n\t"
1030 "psraw %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
1031 "psraw %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
1032 "psraw %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1033 "packsswh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
1034 "packsswh %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
1035 "dmfc1 %[tmp1], %[ftmp0] \n\t"
1036 "dsrl %[ftmp0], %[ftmp0], %[ftmp9] \n\t"
1037 "sh %[tmp1], 0x00(%[output]) \n\t"
1038 "mfc1 %[input], %[ftmp0] \n\t"
1039 "dsrl %[tmp1], %[tmp1], 0x10 \n\t"
1040 "sh %[input], 0x80(%[output]) \n\t"
1041 "sh %[tmp1], 0x20(%[output]) \n\t"
1042 PTR_SRL "%[input], %[input], 0x10 \n\t"
1043 "dmfc1 %[tmp1], %[ftmp2] \n\t"
1044 "sh %[input], 0xa0(%[output]) \n\t"
1045 "dsrl %[ftmp2], %[ftmp2], %[ftmp9] \n\t"
1046 "sh %[tmp1], 0x40(%[output]) \n\t"
1047 "mfc1 %[input], %[ftmp2] \n\t"
1048 "dsrl %[tmp1], %[tmp1], 0x10 \n\t"
1049 "sh %[input], 0xc0(%[output]) \n\t"
1050 "sh %[tmp1], 0x60(%[output]) \n\t"
1051 PTR_SRL "%[input], %[input], 0x10 \n\t"
1052 "sh %[input], 0xe0(%[output]) \n\t"
1053 "punpckhhw %[ftmp1], %[ftmp3], %[ff_pw_1] \n\t"
1054 "punpcklhw %[ftmp3], %[ftmp3], %[ff_pw_1] \n\t"
1055 "punpckhhw %[ftmp5], %[ftmp4], %[ff_pw_1] \n\t"
1056 "punpcklhw %[ftmp4], %[ftmp4], %[ff_pw_1] \n\t"
1057 "mtc1 %[qmul], %[ftmp7] \n\t"
1058 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
1059 "pmaddhw %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
1060 "pmaddhw %[ftmp4], %[ftmp4], %[ftmp7] \n\t"
1061 "pmaddhw %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
1062 "pmaddhw %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1063 "psraw %[ftmp3], %[ftmp3], %[ftmp6] \n\t"
1064 "psraw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
1065 "psraw %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
1066 "psraw %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1067 "packsswh %[ftmp3], %[ftmp3], %[ftmp1] \n\t"
1068 "packsswh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1069 "dmfc1 %[tmp1], %[ftmp3] \n\t"
1070 "dsrl %[ftmp3], %[ftmp3], %[ftmp9] \n\t"
1071 "mfc1 %[input], %[ftmp3] \n\t"
1072 "sh %[tmp1], 0x100(%[output]) \n\t"
1073 "sh %[input], 0x180(%[output]) \n\t"
1074 "dsrl %[tmp1], %[tmp1], 0x10 \n\t"
1075 PTR_SRL "%[input], %[input], 0x10 \n\t"
1076 "sh %[tmp1], 0x120(%[output]) \n\t"
1077 "sh %[input], 0x1a0(%[output]) \n\t"
1078 "dmfc1 %[tmp1], %[ftmp4] \n\t"
1079 "dsrl %[ftmp4], %[ftmp4], %[ftmp9] \n\t"
1080 "mfc1 %[input], %[ftmp4] \n\t"
1081 "sh %[tmp1], 0x140(%[output]) \n\t"
1082 "sh %[input], 0x1c0(%[output]) \n\t"
1083 "dsrl %[tmp1], %[tmp1], 0x10 \n\t"
1084 PTR_SRL "%[input], %[input], 0x10 \n\t"
1085 "sh %[tmp1], 0x160(%[output]) \n\t"
1086 "sh %[input], 0x1e0(%[output]) \n\t"
1089 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1090 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1091 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1092 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
1093 [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),
1094 [tmp0]"=&r"(tmp[0]), [tmp1]"=&r"(tmp[1]),
1096 [output]"+&r"(output), [input]"+&r"(input),
1098 : [ff_pw_1]"f"(ff_pw_1)
1103 void ff_h264_chroma422_dc_dequant_idct_8_mmi(int16_t *block, int qmul)
1108 temp[0] = block[0] + block[16];
1109 temp[1] = block[0] - block[16];
1110 temp[2] = block[32] + block[48];
1111 temp[3] = block[32] - block[48];
1112 temp[4] = block[64] + block[80];
1113 temp[5] = block[64] - block[80];
1114 temp[6] = block[96] + block[112];
1115 temp[7] = block[96] - block[112];
1117 t[0] = temp[0] + temp[4] + temp[2] + temp[6];
1118 t[1] = temp[0] - temp[4] + temp[2] - temp[6];
1119 t[2] = temp[0] - temp[4] - temp[2] + temp[6];
1120 t[3] = temp[0] + temp[4] - temp[2] - temp[6];
1121 t[4] = temp[1] + temp[5] + temp[3] + temp[7];
1122 t[5] = temp[1] - temp[5] + temp[3] - temp[7];
1123 t[6] = temp[1] - temp[5] - temp[3] + temp[7];
1124 t[7] = temp[1] + temp[5] - temp[3] - temp[7];
1126 block[ 0]= (t[0]*qmul + 128) >> 8;
1127 block[ 32]= (t[1]*qmul + 128) >> 8;
1128 block[ 64]= (t[2]*qmul + 128) >> 8;
1129 block[ 96]= (t[3]*qmul + 128) >> 8;
1130 block[ 16]= (t[4]*qmul + 128) >> 8;
1131 block[ 48]= (t[5]*qmul + 128) >> 8;
1132 block[ 80]= (t[6]*qmul + 128) >> 8;
1133 block[112]= (t[7]*qmul + 128) >> 8;
1136 void ff_h264_chroma_dc_dequant_idct_8_mmi(int16_t *block, int qmul)
1140 d = block[0] - block[16];
1141 a = block[0] + block[16];
1142 b = block[32] - block[48];
1143 c = block[32] + block[48];
1144 block[0] = ((a+c)*qmul) >> 7;
1145 block[16]= ((d+b)*qmul) >> 7;
1146 block[32]= ((a-c)*qmul) >> 7;
1147 block[48]= ((d-b)*qmul) >> 7;
1150 void ff_h264_weight_pixels16_8_mmi(uint8_t *block, ptrdiff_t stride, int height,
1151 int log2_denom, int weight, int offset)
1157 offset <<= log2_denom;
1160 offset += 1 << (log2_denom - 1);
1162 for (y=0; y<height; y++, block+=stride) {
1164 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1165 MMI_LDC1(%[ftmp1], %[block0], 0x00)
1166 MMI_LDC1(%[ftmp2], %[block1], 0x00)
1167 "mtc1 %[weight], %[ftmp3] \n\t"
1168 "mtc1 %[offset], %[ftmp4] \n\t"
1169 "mtc1 %[log2_denom], %[ftmp5] \n\t"
1170 "pshufh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
1171 "pshufh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
1172 "punpckhbh %[ftmp6], %[ftmp1], %[ftmp0] \n\t"
1173 "punpckhbh %[ftmp7], %[ftmp2], %[ftmp0] \n\t"
1174 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1175 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
1176 "pmullh %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
1177 "pmullh %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
1178 "pmullh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
1179 "pmullh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
1180 "paddsh %[ftmp6], %[ftmp6], %[ftmp4] \n\t"
1181 "paddsh %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
1182 "paddsh %[ftmp1], %[ftmp1], %[ftmp4] \n\t"
1183 "paddsh %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
1184 "psrah %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
1185 "psrah %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
1186 "psrah %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
1187 "psrah %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
1188 "packushb %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
1189 "packushb %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
1190 MMI_SDC1(%[ftmp1], %[block0], 0x00)
1191 MMI_SDC1(%[ftmp2], %[block1], 0x00)
1192 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1193 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1194 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1195 [ftmp6]"=&f"(ftmp[6]),
1197 [ftmp7]"=&f"(ftmp[7])
1198 : [block0]"r"(block), [block1]"r"(block+8),
1199 [weight]"r"(weight), [offset]"r"(offset),
1200 [log2_denom]"r"(log2_denom)
1206 void ff_h264_biweight_pixels16_8_mmi(uint8_t *dst, uint8_t *src,
1207 ptrdiff_t stride, int height, int log2_denom, int weightd, int weights,
1214 offset = ((offset + 1) | 1) << log2_denom;
1216 for (y=0; y<height; y++, dst+=stride, src+=stride) {
1218 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1219 MMI_LDC1(%[ftmp1], %[src0], 0x00)
1220 MMI_LDC1(%[ftmp2], %[dst0], 0x00)
1221 "mtc1 %[weights], %[ftmp3] \n\t"
1222 "mtc1 %[weightd], %[ftmp4] \n\t"
1223 "mtc1 %[offset], %[ftmp5] \n\t"
1224 "mtc1 %[log2_denom], %[ftmp6] \n\t"
1225 "pshufh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
1226 "pshufh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
1227 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
1228 "punpckhbh %[ftmp7], %[ftmp1], %[ftmp0] \n\t"
1229 "punpckhbh %[ftmp8], %[ftmp2], %[ftmp0] \n\t"
1230 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1231 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
1232 "pmullh %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
1233 "pmullh %[ftmp8], %[ftmp8], %[ftmp4] \n\t"
1234 "pmullh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
1235 "pmullh %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
1236 "paddsh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
1237 "paddsh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
1238 "paddsh %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1239 "paddsh %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
1240 "psrah %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
1241 "psrah %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
1242 "packushb %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
1243 MMI_SDC1(%[ftmp1], %[dst0], 0x00)
1244 MMI_LDC1(%[ftmp1], %[src1], 0x00)
1245 MMI_LDC1(%[ftmp2], %[dst1], 0x00)
1246 "punpckhbh %[ftmp7], %[ftmp1], %[ftmp0] \n\t"
1247 "punpckhbh %[ftmp8], %[ftmp2], %[ftmp0] \n\t"
1248 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1249 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
1250 "pmullh %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
1251 "pmullh %[ftmp8], %[ftmp8], %[ftmp4] \n\t"
1252 "pmullh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
1253 "pmullh %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
1254 "paddsh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
1255 "paddsh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
1256 "paddsh %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1257 "paddsh %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
1258 "psrah %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
1259 "psrah %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
1260 "packushb %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
1261 MMI_SDC1(%[ftmp1], %[dst1], 0x00)
1262 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1263 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1264 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1265 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
1267 [ftmp8]"=&f"(ftmp[8])
1268 : [dst0]"r"(dst), [dst1]"r"(dst+8),
1269 [src0]"r"(src), [src1]"r"(src+8),
1270 [weights]"r"(weights), [weightd]"r"(weightd),
1271 [offset]"r"(offset), [log2_denom]"r"(log2_denom+1)
1277 void ff_h264_weight_pixels8_8_mmi(uint8_t *block, ptrdiff_t stride, int height,
1278 int log2_denom, int weight, int offset)
1284 offset <<= log2_denom;
1287 offset += 1 << (log2_denom - 1);
1289 for (y=0; y<height; y++, block+=stride) {
1291 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1292 MMI_LDC1(%[ftmp1], %[block], 0x00)
1293 "mtc1 %[weight], %[ftmp2] \n\t"
1294 "mtc1 %[offset], %[ftmp3] \n\t"
1295 "mtc1 %[log2_denom], %[ftmp5] \n\t"
1296 "pshufh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
1297 "pshufh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
1298 "punpckhbh %[ftmp4], %[ftmp1], %[ftmp0] \n\t"
1299 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1300 "pmullh %[ftmp4], %[ftmp4], %[ftmp2] \n\t"
1301 "pmullh %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
1302 "paddsh %[ftmp4], %[ftmp4], %[ftmp3] \n\t"
1303 "paddsh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
1304 "psrah %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1305 "psrah %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
1306 "packushb %[ftmp1], %[ftmp1], %[ftmp4] \n\t"
1307 MMI_SDC1(%[ftmp1], %[block], 0x00)
1308 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1309 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1310 [ftmp4]"=&f"(ftmp[4]),
1312 [ftmp5]"=&f"(ftmp[5])
1313 : [block]"r"(block), [weight]"r"(weight),
1314 [offset]"r"(offset), [log2_denom]"r"(log2_denom)
1320 void ff_h264_biweight_pixels8_8_mmi(uint8_t *dst, uint8_t *src,
1321 ptrdiff_t stride, int height, int log2_denom, int weightd, int weights,
1328 offset = ((offset + 1) | 1) << log2_denom;
1330 for (y=0; y<height; y++, dst+=stride, src+=stride) {
1332 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1333 MMI_LDC1(%[ftmp1], %[src], 0x00)
1334 MMI_LDC1(%[ftmp2], %[dst], 0x00)
1335 "mtc1 %[weights], %[ftmp3] \n\t"
1336 "mtc1 %[weightd], %[ftmp4] \n\t"
1337 "mtc1 %[offset], %[ftmp5] \n\t"
1338 "mtc1 %[log2_denom], %[ftmp6] \n\t"
1339 "pshufh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
1340 "pshufh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
1341 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
1342 "punpckhbh %[ftmp7], %[ftmp1], %[ftmp0] \n\t"
1343 "punpckhbh %[ftmp8], %[ftmp2], %[ftmp0] \n\t"
1344 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1345 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
1346 "pmullh %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
1347 "pmullh %[ftmp8], %[ftmp8], %[ftmp4] \n\t"
1348 "pmullh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
1349 "pmullh %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
1350 "paddsh %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
1351 "paddsh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
1352 "paddsh %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1353 "paddsh %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
1354 "psrah %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
1355 "psrah %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
1356 "packushb %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
1357 MMI_SDC1(%[ftmp1], %[dst], 0x00)
1358 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1359 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1360 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1361 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
1363 [ftmp8]"=&f"(ftmp[8])
1364 : [dst]"r"(dst), [src]"r"(src),
1365 [weights]"r"(weights), [weightd]"r"(weightd),
1366 [offset]"r"(offset), [log2_denom]"r"(log2_denom+1)
1372 void ff_h264_weight_pixels4_8_mmi(uint8_t *block, ptrdiff_t stride, int height,
1373 int log2_denom, int weight, int offset)
1379 offset <<= log2_denom;
1382 offset += 1 << (log2_denom - 1);
1384 for (y=0; y<height; y++, block+=stride) {
1386 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1387 MMI_ULWC1(%[ftmp1], %[block], 0x00)
1388 "mtc1 %[weight], %[ftmp2] \n\t"
1389 "mtc1 %[offset], %[ftmp3] \n\t"
1390 "mtc1 %[log2_denom], %[ftmp4] \n\t"
1391 "pshufh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
1392 "pshufh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
1393 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1394 "pmullh %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
1395 "paddsh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
1396 "psrah %[ftmp1], %[ftmp1], %[ftmp4] \n\t"
1397 "packushb %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1398 MMI_SWC1(%[ftmp1], %[block], 0x00)
1399 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1400 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1402 [ftmp4]"=&f"(ftmp[4])
1403 : [block]"r"(block), [weight]"r"(weight),
1404 [offset]"r"(offset), [log2_denom]"r"(log2_denom)
1410 void ff_h264_biweight_pixels4_8_mmi(uint8_t *dst, uint8_t *src,
1411 ptrdiff_t stride, int height, int log2_denom, int weightd, int weights,
1418 offset = ((offset + 1) | 1) << log2_denom;
1420 for (y=0; y<height; y++, dst+=stride, src+=stride) {
1422 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1423 MMI_ULWC1(%[ftmp1], %[src], 0x00)
1424 MMI_ULWC1(%[ftmp2], %[dst], 0x00)
1425 "mtc1 %[weight], %[ftmp3] \n\t"
1426 "mtc1 %[weightd], %[ftmp4] \n\t"
1427 "mtc1 %[offset], %[ftmp5] \n\t"
1428 "mtc1 %[log2_denom], %[ftmp6] \n\t"
1429 "pshufh %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
1430 "pshufh %[ftmp4], %[ftmp4], %[ftmp0] \n\t"
1431 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
1432 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1433 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
1434 "pmullh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
1435 "pmullh %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
1436 "paddsh %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
1437 "paddsh %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
1438 "psrah %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
1439 "packushb %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
1440 MMI_SWC1(%[ftmp1], %[dst], 0x00)
1441 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1442 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1443 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1445 [ftmp6]"=&f"(ftmp[6])
1446 : [dst]"r"(dst), [src]"r"(src),
1447 [weight]"r"(weights), [weightd]"r"(weightd),
1448 [offset]"r"(offset), [log2_denom]"r"(log2_denom+1)
1454 void ff_deblock_v8_luma_8_mmi(uint8_t *pix, int stride, int alpha, int beta,
1464 PTR_ADDU "%[addr0], %[stride], %[stride] \n\t"
1465 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1466 PTR_ADDU "%[addr1], %[stride], %[addr0] \n\t"
1467 "addi %[alpha], %[alpha], -0x01 \n\t"
1468 PTR_SUBU "%[addr1], $0, %[addr1] \n\t"
1469 "addi %[beta], %[beta], -0x01 \n\t"
1470 PTR_ADDU "%[addr1], %[addr1], %[pix] \n\t"
1471 MMI_LDC1(%[ftmp3], %[pix], 0x00)
1472 MMI_LDXC1(%[ftmp1], %[addr1], %[stride], 0x00)
1473 MMI_LDXC1(%[ftmp2], %[addr1], %[addr0], 0x00)
1474 MMI_LDXC1(%[ftmp4], %[pix], %[stride], 0x00)
1475 "mtc1 %[alpha], %[ftmp5] \n\t"
1476 "mtc1 %[beta], %[ftmp6] \n\t"
1477 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
1478 "pshufh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
1479 "packushb %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
1480 "packushb %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
1481 "psubusb %[ftmp7], %[ftmp3], %[ftmp2] \n\t"
1482 "psubusb %[ftmp8], %[ftmp2], %[ftmp3] \n\t"
1483 "or %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1484 "psubusb %[ftmp7], %[ftmp2], %[ftmp1] \n\t"
1485 "psubusb %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1486 "psubusb %[ftmp5], %[ftmp1], %[ftmp2] \n\t"
1487 "or %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1488 "psubusb %[ftmp7], %[ftmp3], %[ftmp4] \n\t"
1489 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1490 "or %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1491 "psubusb %[ftmp5], %[ftmp4], %[ftmp3] \n\t"
1492 "or %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1493 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1494 "or %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1495 "pcmpeqb %[ftmp8], %[ftmp8], %[ftmp0] \n\t"
1496 "pcmpeqb %[ftmp4], %[ftmp4], %[ftmp4] \n\t"
1497 MMI_ULWC1(%[ftmp5], %[tc0], 0x00)
1498 "punpcklbh %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
1499 "punpcklbh %[ftmp9], %[ftmp5], %[ftmp5] \n\t"
1500 "pcmpgtb %[ftmp5], %[ftmp9], %[ftmp4] \n\t"
1501 MMI_LDC1(%[ftmp4], %[addr1], 0x00)
1502 "and %[ftmp10], %[ftmp5], %[ftmp8] \n\t"
1503 "psubusb %[ftmp8], %[ftmp4], %[ftmp2] \n\t"
1504 "psubusb %[ftmp7], %[ftmp2], %[ftmp4] \n\t"
1505 "psubusb %[ftmp8], %[ftmp8], %[ftmp6] \n\t"
1506 "psubusb %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
1507 "pcmpeqb %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1508 "and %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
1509 "and %[ftmp5], %[ftmp10], %[ftmp9] \n\t"
1510 "psubb %[ftmp8], %[ftmp5], %[ftmp7] \n\t"
1511 "and %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
1512 "pavgb %[ftmp5], %[ftmp2], %[ftmp3] \n\t"
1513 MMI_LDC1(%[ftmp11], %[addr1], 0x00)
1514 "pavgb %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1515 "xor %[ftmp5], %[ftmp5], %[ftmp11] \n\t"
1516 "and %[ftmp5], %[ftmp5], %[ff_pb_1] \n\t"
1517 "psubusb %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1518 "psubusb %[ftmp5], %[ftmp1], %[ftmp7] \n\t"
1519 "paddusb %[ftmp7], %[ftmp7], %[ftmp1] \n\t"
1520 "pmaxub %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1521 "pminub %[ftmp4], %[ftmp4], %[ftmp7] \n\t"
1522 MMI_SDXC1(%[ftmp4], %[addr1], %[stride], 0x00)
1523 MMI_LDXC1(%[ftmp5], %[pix], %[addr0], 0x00)
1524 "psubusb %[ftmp4], %[ftmp5], %[ftmp3] \n\t"
1525 "psubusb %[ftmp7], %[ftmp3], %[ftmp5] \n\t"
1526 "psubusb %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
1527 "psubusb %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
1528 "pcmpeqb %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
1529 "and %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
1530 "psubb %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1531 "and %[ftmp6], %[ftmp9], %[ftmp7] \n\t"
1532 MMI_LDXC1(%[ftmp4], %[pix], %[stride], 0x00)
1533 "pavgb %[ftmp7], %[ftmp2], %[ftmp3] \n\t"
1534 MMI_LDXC1(%[ftmp11], %[pix], %[addr0], 0x00)
1535 "pavgb %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1536 "xor %[ftmp7], %[ftmp7], %[ftmp11] \n\t"
1537 "and %[ftmp7], %[ftmp7], %[ff_pb_1] \n\t"
1538 "psubusb %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1539 "psubusb %[ftmp7], %[ftmp4], %[ftmp6] \n\t"
1540 "paddusb %[ftmp6], %[ftmp6], %[ftmp4] \n\t"
1541 "pmaxub %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1542 "pminub %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1543 MMI_SDXC1(%[ftmp5], %[pix], %[stride], 0x00)
1544 "xor %[ftmp6], %[ftmp2], %[ftmp3] \n\t"
1545 "pcmpeqb %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
1546 "and %[ftmp6], %[ftmp6], %[ff_pb_1] \n\t"
1547 "xor %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1548 "xor %[ftmp5], %[ftmp5], %[ftmp2] \n\t"
1549 "pavgb %[ftmp4], %[ftmp4], %[ftmp1] \n\t"
1550 "pavgb %[ftmp4], %[ftmp4], %[ff_pb_3] \n\t"
1551 "pavgb %[ftmp5], %[ftmp5], %[ftmp3] \n\t"
1552 "pavgb %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
1553 "paddusb %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1554 "psubusb %[ftmp7], %[ff_pb_A1], %[ftmp4] \n\t"
1555 "psubusb %[ftmp4], %[ftmp4], %[ff_pb_A1] \n\t"
1556 "pminub %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1557 "pminub %[ftmp4], %[ftmp4], %[ftmp8] \n\t"
1558 "psubusb %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
1559 "psubusb %[ftmp3], %[ftmp3], %[ftmp4] \n\t"
1560 "paddusb %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
1561 "paddusb %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
1562 MMI_SDXC1(%[ftmp2], %[addr1], %[addr0], 0x00)
1563 MMI_SDC1(%[ftmp3], %[pix], 0x00)
1564 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1565 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1566 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1567 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
1568 [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),
1569 [ftmp10]"=&f"(ftmp[10]), [ftmp11]"=&f"(ftmp[11]),
1573 [addr0]"=&r"(addr[0]), [addr1]"=&r"(addr[1])
1574 : [pix]"r"(pix), [stride]"r"((mips_reg)stride),
1575 [alpha]"r"((mips_reg)alpha), [beta]"r"((mips_reg)beta),
1576 [tc0]"r"(tc0), [ff_pb_1]"f"(ff_pb_1),
1577 [ff_pb_3]"f"(ff_pb_3), [ff_pb_A1]"f"(ff_pb_A1)
1582 static void deblock_v8_luma_intra_8_mmi(uint8_t *pix, int stride, int alpha,
1585 DECLARE_ALIGNED(8, const uint64_t, stack[0x0a]);
1593 "ori %[tmp0], $0, 0x01 \n\t"
1594 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1595 "mtc1 %[tmp0], %[ftmp9] \n\t"
1596 PTR_SLL "%[addr0], %[stride], 0x02 \n\t"
1597 PTR_ADDU "%[addr2], %[stride], %[stride] \n\t"
1598 PTR_ADDIU "%[alpha], %[alpha], -0x01 \n\t"
1599 PTR_SLL "%[ftmp11], %[ftmp9], %[ftmp9] \n\t"
1600 "bltz %[alpha], 1f \n\t"
1601 PTR_ADDU "%[addr1], %[addr2], %[stride] \n\t"
1602 PTR_ADDIU "%[beta], %[beta], -0x01 \n\t"
1603 "bltz %[beta], 1f \n\t"
1604 PTR_SUBU "%[addr0], $0, %[addr0] \n\t"
1605 PTR_ADDU "%[addr0], %[addr0], %[pix] \n\t"
1606 MMI_LDC1(%[ftmp3], %[pix], 0x00)
1607 MMI_LDXC1(%[ftmp1], %[addr0], %[addr2], 0x00)
1608 MMI_LDXC1(%[ftmp2], %[addr0], %[addr1], 0x00)
1609 MMI_LDXC1(%[ftmp4], %[pix], %[stride], 0x00)
1610 "mtc1 %[alpha], %[ftmp5] \n\t"
1611 "mtc1 %[beta], %[ftmp6] \n\t"
1612 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
1613 "pshufh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
1614 "packushb %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
1615 "psubusb %[ftmp7], %[ftmp3], %[ftmp2] \n\t"
1616 "psubusb %[ftmp8], %[ftmp2], %[ftmp3] \n\t"
1617 "packushb %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
1618 "or %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1619 MMI_SDC1(%[ftmp5], %[stack], 0x10)
1620 "psubusb %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1621 "psubusb %[ftmp7], %[ftmp2], %[ftmp1] \n\t"
1622 "psubusb %[ftmp5], %[ftmp1], %[ftmp2] \n\t"
1623 "or %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1624 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1625 "or %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1626 "psubusb %[ftmp7], %[ftmp3], %[ftmp4] \n\t"
1627 "psubusb %[ftmp5], %[ftmp4], %[ftmp3] \n\t"
1628 "or %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1629 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1630 "or %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1631 "xor %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
1632 MMI_LDC1(%[ftmp5], %[stack], 0x10)
1633 "pcmpeqb %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1634 "ldc1 %[ftmp10], %[ff_pb_1] \n\t"
1635 MMI_SDC1(%[ftmp8], %[stack], 0x20)
1636 "pavgb %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
1637 "psubusb %[ftmp8], %[ftmp3], %[ftmp2] \n\t"
1638 "pavgb %[ftmp5], %[ftmp5], %[ftmp10] \n\t"
1639 "psubusb %[ftmp7], %[ftmp2], %[ftmp3] \n\t"
1640 "psubusb %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1641 "psubusb %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
1642 MMI_LDC1(%[ftmp15], %[stack], 0x20)
1643 "pcmpeqb %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1644 "and %[ftmp7], %[ftmp7], %[ftmp15] \n\t"
1645 MMI_LDXC1(%[ftmp15], %[addr0], %[stride], 0x00)
1646 "psubusb %[ftmp8], %[ftmp15], %[ftmp2] \n\t"
1647 "psubusb %[ftmp5], %[ftmp2], %[ftmp15] \n\t"
1648 "psubusb %[ftmp8], %[ftmp8], %[ftmp6] \n\t"
1649 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1650 "pcmpeqb %[ftmp5], %[ftmp5], %[ftmp8] \n\t"
1651 "and %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1652 MMI_LDXC1(%[ftmp14], %[pix], %[addr2], 0x00)
1653 MMI_SDC1(%[ftmp5], %[stack], 0x30)
1654 "psubusb %[ftmp8], %[ftmp14], %[ftmp3] \n\t"
1655 "psubusb %[ftmp5], %[ftmp3], %[ftmp14] \n\t"
1656 "psubusb %[ftmp8], %[ftmp8], %[ftmp6] \n\t"
1657 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1658 "pcmpeqb %[ftmp5], %[ftmp5], %[ftmp8] \n\t"
1659 "and %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1660 MMI_SDC1(%[ftmp5], %[stack], 0x40)
1661 "pavgb %[ftmp5], %[ftmp15], %[ftmp1] \n\t"
1662 "pavgb %[ftmp6], %[ftmp2], %[ftmp3] \n\t"
1663 "pavgb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1664 MMI_SDC1(%[ftmp6], %[stack], 0x10)
1665 "paddb %[ftmp7], %[ftmp15], %[ftmp1] \n\t"
1666 "paddb %[ftmp8], %[ftmp2], %[ftmp3] \n\t"
1667 "paddb %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1668 "mov.d %[ftmp8], %[ftmp7] \n\t"
1669 MMI_SDC1(%[ftmp7], %[stack], 0x00)
1670 "psrlh %[ftmp7], %[ftmp7], %[ftmp9] \n\t"
1671 "pavgb %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
1672 "xor %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
1673 "and %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
1674 "psubb %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1675 "pavgb %[ftmp6], %[ftmp15], %[ftmp4] \n\t"
1676 "psubb %[ftmp7], %[ftmp15], %[ftmp4] \n\t"
1677 "paddb %[ftmp8], %[ftmp8], %[ftmp8] \n\t"
1678 "psubb %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1679 "and %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
1680 "psubb %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
1681 MMI_LDC1(%[ftmp13], %[stack], 0x10)
1682 "pavgb %[ftmp6], %[ftmp6], %[ftmp1] \n\t"
1683 "psrlh %[ftmp8], %[ftmp8], %[ftmp11] \n\t"
1684 "pavgb %[ftmp6], %[ftmp6], %[ftmp13] \n\t"
1685 "pavgb %[ftmp8], %[ftmp8], %[ftmp0] \n\t"
1686 "xor %[ftmp8], %[ftmp8], %[ftmp6] \n\t"
1687 "and %[ftmp8], %[ftmp8], %[ftmp10] \n\t"
1688 "psubb %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
1689 "xor %[ftmp8], %[ftmp2], %[ftmp4] \n\t"
1690 "pavgb %[ftmp7], %[ftmp2], %[ftmp4] \n\t"
1691 "and %[ftmp8], %[ftmp8], %[ftmp10] \n\t"
1692 "psubb %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1693 MMI_LDC1(%[ftmp13], %[stack], 0x30)
1694 "pavgb %[ftmp7], %[ftmp7], %[ftmp1] \n\t"
1695 MMI_LDC1(%[ftmp12], %[stack], 0x20)
1696 "xor %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
1697 "xor %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
1698 "and %[ftmp6], %[ftmp6], %[ftmp13] \n\t"
1699 "and %[ftmp7], %[ftmp7], %[ftmp12] \n\t"
1700 "xor %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
1701 "xor %[ftmp6], %[ftmp6], %[ftmp2] \n\t"
1702 MMI_SDXC1(%[ftmp6], %[addr0], %[addr1], 0x00)
1703 MMI_LDC1(%[ftmp6], %[addr0], 0x00)
1704 "paddb %[ftmp7], %[ftmp15], %[ftmp6] \n\t"
1705 "pavgb %[ftmp6], %[ftmp6], %[ftmp15] \n\t"
1706 MMI_LDC1(%[ftmp12], %[stack], 0x00)
1707 "pavgb %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
1708 "paddb %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
1709 "paddb %[ftmp7], %[ftmp7], %[ftmp12] \n\t"
1710 "psrlh %[ftmp7], %[ftmp7], %[ftmp11] \n\t"
1711 "pavgb %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
1712 "xor %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
1713 "and %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
1714 MMI_LDC1(%[ftmp12], %[stack], 0x30)
1715 "psubb %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
1716 "xor %[ftmp5], %[ftmp5], %[ftmp1] \n\t"
1717 "xor %[ftmp6], %[ftmp6], %[ftmp15] \n\t"
1718 "and %[ftmp5], %[ftmp5], %[ftmp12] \n\t"
1719 "and %[ftmp6], %[ftmp6], %[ftmp12] \n\t"
1720 "xor %[ftmp5], %[ftmp5], %[ftmp1] \n\t"
1721 "xor %[ftmp6], %[ftmp6], %[ftmp15] \n\t"
1722 MMI_SDXC1(%[ftmp5], %[addr0], %[addr2], 0x00)
1723 MMI_SDXC1(%[ftmp6], %[addr0], %[stride], 0x00)
1724 "pavgb %[ftmp5], %[ftmp14], %[ftmp4] \n\t"
1725 "pavgb %[ftmp6], %[ftmp3], %[ftmp2] \n\t"
1726 "pavgb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1727 MMI_SDC1(%[ftmp6], %[stack], 0x10)
1728 "paddb %[ftmp7], %[ftmp14], %[ftmp4] \n\t"
1729 "paddb %[ftmp8], %[ftmp3], %[ftmp2] \n\t"
1730 "paddb %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1731 "mov.d %[ftmp8], %[ftmp7] \n\t"
1732 MMI_SDC1(%[ftmp7], %[stack], 0x00)
1733 "psrlh %[ftmp7], %[ftmp7], %[ftmp9] \n\t"
1734 "pavgb %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
1735 "xor %[ftmp7], %[ftmp7], %[ftmp5] \n\t"
1736 "and %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
1737 "psubb %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1738 "pavgb %[ftmp6], %[ftmp14], %[ftmp1] \n\t"
1739 "paddb %[ftmp8], %[ftmp8], %[ftmp8] \n\t"
1740 "psubb %[ftmp7], %[ftmp14], %[ftmp1] \n\t"
1741 "psubb %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1742 "and %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
1743 "psubb %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
1744 MMI_LDC1(%[ftmp12], %[stack], 0x10)
1745 "pavgb %[ftmp6], %[ftmp6], %[ftmp4] \n\t"
1746 "pavgb %[ftmp6], %[ftmp6], %[ftmp12] \n\t"
1747 "psrlh %[ftmp8], %[ftmp8], %[ftmp11] \n\t"
1748 "pavgb %[ftmp8], %[ftmp8], %[ftmp0] \n\t"
1749 "xor %[ftmp8], %[ftmp8], %[ftmp6] \n\t"
1750 "and %[ftmp8], %[ftmp8], %[ftmp10] \n\t"
1751 "psubb %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
1752 "xor %[ftmp8], %[ftmp3], %[ftmp1] \n\t"
1753 "pavgb %[ftmp7], %[ftmp3], %[ftmp1] \n\t"
1754 "and %[ftmp8], %[ftmp8], %[ftmp10] \n\t"
1755 MMI_LDC1(%[ftmp12], %[stack], 0x40)
1756 "psubb %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1757 MMI_LDC1(%[ftmp13], %[stack], 0x20)
1758 "pavgb %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
1759 "xor %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
1760 "xor %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
1761 "and %[ftmp6], %[ftmp6], %[ftmp12] \n\t"
1762 "and %[ftmp7], %[ftmp7], %[ftmp13] \n\t"
1763 "xor %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
1764 "xor %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
1765 MMI_SDC1(%[ftmp6], %[pix], 0x00)
1766 MMI_LDXC1(%[ftmp6], %[pix], %[addr1], 0x00)
1767 "paddb %[ftmp7], %[ftmp14], %[ftmp6] \n\t"
1768 "pavgb %[ftmp6], %[ftmp6], %[ftmp14] \n\t"
1769 MMI_LDC1(%[ftmp12], %[stack], 0x00)
1770 "pavgb %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
1771 "paddb %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
1772 "paddb %[ftmp7], %[ftmp7], %[ftmp12] \n\t"
1773 "psrlh %[ftmp7], %[ftmp7], %[ftmp11] \n\t"
1774 "pavgb %[ftmp7], %[ftmp7], %[ftmp0] \n\t"
1775 "xor %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
1776 "and %[ftmp7], %[ftmp7], %[ftmp10] \n\t"
1777 MMI_LDC1(%[ftmp12], %[stack], 0x40)
1778 "psubb %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
1779 "xor %[ftmp5], %[ftmp5], %[ftmp4] \n\t"
1780 "xor %[ftmp6], %[ftmp6], %[ftmp14] \n\t"
1781 "and %[ftmp5], %[ftmp5], %[ftmp12] \n\t"
1782 "and %[ftmp6], %[ftmp6], %[ftmp12] \n\t"
1783 "xor %[ftmp5], %[ftmp5], %[ftmp4] \n\t"
1784 "xor %[ftmp6], %[ftmp6], %[ftmp14] \n\t"
1785 MMI_SDXC1(%[ftmp5], %[pix], %[stride], 0x00)
1786 MMI_SDXC1(%[ftmp6], %[pix], %[addr2], 0x00)
1788 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1789 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1790 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1791 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
1792 [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),
1793 [ftmp10]"=&f"(ftmp[10]), [ftmp11]"=&f"(ftmp[11]),
1794 [ftmp12]"=&f"(ftmp[12]), [ftmp13]"=&f"(ftmp[13]),
1795 [ftmp14]"=&f"(ftmp[14]), [ftmp15]"=&f"(ftmp[15]),
1796 [tmp0]"=&r"(tmp[0]),
1799 [addr0]"=&r"(addr[0]), [addr1]"=&r"(addr[1]),
1800 [addr2]"=&r"(addr[2]),
1801 [alpha]"+&r"(alpha), [beta]"+&r"(beta)
1802 : [pix]"r"(pix), [stride]"r"((mips_reg)stride),
1803 [stack]"r"(stack), [ff_pb_1]"m"(ff_pb_1)
1808 void ff_deblock_v_chroma_8_mmi(uint8_t *pix, ptrdiff_t stride, int alpha,
1809 int beta, int8_t *tc0)
1818 "addi %[alpha], %[alpha], -0x01 \n\t"
1819 "addi %[beta], %[beta], -0x01 \n\t"
1820 "or %[addr0], $0, %[pix] \n\t"
1821 PTR_SUBU "%[addr0], %[addr0], %[stride] \n\t"
1822 PTR_SUBU "%[addr0], %[addr0], %[stride] \n\t"
1823 MMI_LDC1(%[ftmp1], %[addr0], 0x00)
1824 MMI_LDXC1(%[ftmp2], %[addr0], %[stride], 0x00)
1825 MMI_LDC1(%[ftmp3], %[pix], 0x00)
1826 MMI_LDXC1(%[ftmp4], %[pix], %[stride], 0x00)
1828 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1829 "mtc1 %[alpha], %[ftmp5] \n\t"
1830 "mtc1 %[beta], %[ftmp6] \n\t"
1831 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
1832 "pshufh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
1833 "packushb %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
1834 "packushb %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
1835 "psubusb %[ftmp7], %[ftmp3], %[ftmp2] \n\t"
1836 "psubusb %[ftmp8], %[ftmp2], %[ftmp3] \n\t"
1837 "or %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1838 "psubusb %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1839 "psubusb %[ftmp7], %[ftmp2], %[ftmp1] \n\t"
1840 "psubusb %[ftmp5], %[ftmp1], %[ftmp2] \n\t"
1841 "or %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1842 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1843 "or %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1844 "psubusb %[ftmp7], %[ftmp3], %[ftmp4] \n\t"
1845 "psubusb %[ftmp5], %[ftmp4], %[ftmp3] \n\t"
1846 "or %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1847 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1848 "or %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1849 "xor %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
1850 "pcmpeqb %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1851 MMI_ULWC1(%[ftmp7], %[tc0], 0x00)
1852 "punpcklbh %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
1853 "and %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1854 "pcmpeqb %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
1855 "xor %[ftmp6], %[ftmp2], %[ftmp3] \n\t"
1856 "xor %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1857 "and %[ftmp6], %[ftmp6], %[ff_pb_1] \n\t"
1858 "pavgb %[ftmp4], %[ftmp4], %[ftmp1] \n\t"
1859 "xor %[ftmp5], %[ftmp5], %[ftmp2] \n\t"
1860 "pavgb %[ftmp4], %[ftmp4], %[ff_pb_3] \n\t"
1861 "pavgb %[ftmp5], %[ftmp5], %[ftmp3] \n\t"
1862 "pavgb %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
1863 "paddusb %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
1864 "psubusb %[ftmp7], %[ff_pb_A1], %[ftmp4] \n\t"
1865 "psubusb %[ftmp4], %[ftmp4], %[ff_pb_A1] \n\t"
1866 "pminub %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
1867 "pminub %[ftmp4], %[ftmp4], %[ftmp8] \n\t"
1868 "psubusb %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
1869 "psubusb %[ftmp3], %[ftmp3], %[ftmp4] \n\t"
1870 "paddusb %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
1871 "paddusb %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
1873 MMI_SDXC1(%[ftmp2], %[addr0], %[stride], 0x00)
1874 MMI_SDC1(%[ftmp3], %[pix], 0x00)
1875 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1876 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1877 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1878 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
1879 [ftmp8]"=&f"(ftmp[8]),
1883 [addr0]"=&r"(addr[0])
1884 : [pix]"r"(pix), [stride]"r"((mips_reg)stride),
1885 [alpha]"r"(alpha), [beta]"r"(beta),
1886 [tc0]"r"(tc0), [ff_pb_1]"f"(ff_pb_1),
1887 [ff_pb_3]"f"(ff_pb_3), [ff_pb_A1]"f"(ff_pb_A1)
1892 void ff_deblock_v_chroma_intra_8_mmi(uint8_t *pix, int stride, int alpha,
1901 "addi %[alpha], %[alpha], -0x01 \n\t"
1902 "addi %[beta], %[beta], -0x01 \n\t"
1903 "or %[addr0], $0, %[pix] \n\t"
1904 PTR_SUBU "%[addr0], %[addr0], %[stride] \n\t"
1905 PTR_SUBU "%[addr0], %[addr0], %[stride] \n\t"
1906 MMI_LDC1(%[ftmp1], %[addr0], 0x00)
1907 MMI_LDXC1(%[ftmp2], %[addr0], %[stride], 0x00)
1908 MMI_LDC1(%[ftmp3], %[pix], 0x00)
1909 MMI_LDXC1(%[ftmp4], %[pix], %[stride], 0x00)
1911 "xor %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
1912 "mtc1 %[alpha], %[ftmp5] \n\t"
1913 "mtc1 %[beta], %[ftmp6] \n\t"
1914 "pshufh %[ftmp5], %[ftmp5], %[ftmp0] \n\t"
1915 "pshufh %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
1916 "packushb %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
1917 "packushb %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
1918 "psubusb %[ftmp7], %[ftmp3], %[ftmp2] \n\t"
1919 "psubusb %[ftmp8], %[ftmp2], %[ftmp3] \n\t"
1920 "or %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1921 "psubusb %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1922 "psubusb %[ftmp7], %[ftmp2], %[ftmp1] \n\t"
1923 "psubusb %[ftmp5], %[ftmp1], %[ftmp2] \n\t"
1924 "or %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1925 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1926 "or %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1927 "psubusb %[ftmp7], %[ftmp3], %[ftmp4] \n\t"
1928 "psubusb %[ftmp5], %[ftmp4], %[ftmp3] \n\t"
1929 "or %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
1930 "psubusb %[ftmp5], %[ftmp5], %[ftmp6] \n\t"
1931 "or %[ftmp8], %[ftmp8], %[ftmp5] \n\t"
1932 "xor %[ftmp7], %[ftmp7], %[ftmp7] \n\t"
1933 "pcmpeqb %[ftmp8], %[ftmp8], %[ftmp7] \n\t"
1934 "mov.d %[ftmp6], %[ftmp2] \n\t"
1935 "mov.d %[ftmp7], %[ftmp3] \n\t"
1936 "xor %[ftmp5], %[ftmp2], %[ftmp4] \n\t"
1937 "and %[ftmp5], %[ftmp5], %[ff_pb_1] \n\t"
1938 "pavgb %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
1939 "psubusb %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
1940 "pavgb %[ftmp2], %[ftmp2], %[ftmp1] \n\t"
1941 "xor %[ftmp5], %[ftmp3], %[ftmp1] \n\t"
1942 "and %[ftmp5], %[ftmp5], %[ff_pb_1] \n\t"
1943 "pavgb %[ftmp3], %[ftmp3], %[ftmp1] \n\t"
1944 "psubusb %[ftmp3], %[ftmp3], %[ftmp5] \n\t"
1945 "pavgb %[ftmp3], %[ftmp3], %[ftmp4] \n\t"
1946 "psubb %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
1947 "psubb %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
1948 "and %[ftmp2], %[ftmp2], %[ftmp8] \n\t"
1949 "and %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
1950 "paddb %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
1951 "paddb %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
1953 MMI_SDXC1(%[ftmp2], %[addr0], %[stride], 0x00)
1954 MMI_SDC1(%[ftmp3], %[pix], 0x00)
1955 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
1956 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
1957 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
1958 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
1959 [ftmp8]"=&f"(ftmp[8]),
1962 [addr0]"=&r"(addr[0])
1963 : [pix]"r"(pix), [stride]"r"((mips_reg)stride),
1964 [alpha]"r"(alpha), [beta]"r"(beta),
1965 [ff_pb_1]"f"(ff_pb_1)
1970 void ff_deblock_h_chroma_8_mmi(uint8_t *pix, int stride, int alpha, int beta,
1978 "addi %[alpha], %[alpha], -0x01 \n\t"
1979 "addi %[beta], %[beta], -0x01 \n\t"
1980 PTR_ADDU "%[addr0], %[stride], %[stride] \n\t"
1981 PTR_ADDI "%[pix], %[pix], -0x02 \n\t"
1982 PTR_ADDU "%[addr1], %[addr0], %[stride] \n\t"
1983 PTR_ADDU "%[addr2], %[addr0], %[addr0] \n\t"
1984 "or %[addr5], $0, %[pix] \n\t"
1985 PTR_ADDU "%[pix], %[pix], %[addr1] \n\t"
1986 MMI_ULWC1(%[ftmp0], %[addr5], 0x00)
1987 PTR_ADDU "%[addr3], %[addr5], %[stride] \n\t"
1988 MMI_ULWC1(%[ftmp2], %[addr3], 0x00)
1989 PTR_ADDU "%[addr4], %[addr5], %[addr0] \n\t"
1990 MMI_ULWC1(%[ftmp1], %[addr4], 0x00)
1991 MMI_ULWC1(%[ftmp3], %[pix], 0x00)
1992 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
1993 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
1994 PTR_ADDU "%[addr3], %[pix], %[stride] \n\t"
1995 "punpckhhw %[ftmp2], %[ftmp0], %[ftmp1] \n\t"
1996 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
1997 MMI_ULWC1(%[ftmp4], %[addr3], 0x00)
1998 PTR_ADDU "%[addr4], %[pix], %[addr0] \n\t"
1999 MMI_ULWC1(%[ftmp6], %[addr4], 0x00)
2000 PTR_ADDU "%[addr3], %[pix], %[addr1] \n\t"
2001 MMI_ULWC1(%[ftmp5], %[addr3], 0x00)
2002 PTR_ADDU "%[addr4], %[pix], %[addr2] \n\t"
2003 MMI_ULWC1(%[ftmp7], %[addr4], 0x00)
2004 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2005 "punpcklbh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
2006 "mov.d %[ftmp6], %[ftmp4] \n\t"
2007 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2008 "punpckhhw %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
2009 "punpckhwd %[ftmp1], %[ftmp0], %[ftmp4] \n\t"
2010 "punpckhwd %[ftmp3], %[ftmp2], %[ftmp6] \n\t"
2011 "punpcklwd %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
2012 "punpcklwd %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
2013 "mov.d %[ftmp9], %[ftmp0] \n\t"
2014 "mov.d %[ftmp10], %[ftmp3] \n\t"
2016 "xor %[ftmp8], %[ftmp8], %[ftmp8] \n\t"
2017 "mtc1 %[alpha], %[ftmp4] \n\t"
2018 "mtc1 %[beta], %[ftmp5] \n\t"
2019 "pshufh %[ftmp4], %[ftmp4], %[ftmp8] \n\t"
2020 "pshufh %[ftmp5], %[ftmp5], %[ftmp8] \n\t"
2021 "packushb %[ftmp4], %[ftmp4], %[ftmp4] \n\t"
2022 "packushb %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
2023 "psubusb %[ftmp6], %[ftmp2], %[ftmp1] \n\t"
2024 "psubusb %[ftmp7], %[ftmp1], %[ftmp2] \n\t"
2025 "or %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
2026 "psubusb %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
2027 "psubusb %[ftmp6], %[ftmp1], %[ftmp0] \n\t"
2028 "psubusb %[ftmp4], %[ftmp0], %[ftmp1] \n\t"
2029 "or %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2030 "psubusb %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2031 "or %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
2032 "psubusb %[ftmp6], %[ftmp2], %[ftmp3] \n\t"
2033 "psubusb %[ftmp4], %[ftmp3], %[ftmp2] \n\t"
2034 "or %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2035 "psubusb %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2036 "or %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
2037 "xor %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
2038 "pcmpeqb %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
2039 MMI_ULWC1(%[ftmp6], %[tc0], 0x00)
2040 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
2041 "and %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
2042 "pcmpeqb %[ftmp4], %[ftmp4], %[ftmp4] \n\t"
2043 "xor %[ftmp5], %[ftmp1], %[ftmp2] \n\t"
2044 "xor %[ftmp3], %[ftmp3], %[ftmp4] \n\t"
2045 "and %[ftmp5], %[ftmp5], %[ff_pb_1] \n\t"
2046 "pavgb %[ftmp3], %[ftmp3], %[ftmp0] \n\t"
2047 "xor %[ftmp4], %[ftmp4], %[ftmp1] \n\t"
2048 "pavgb %[ftmp3], %[ftmp3], %[ff_pb_3] \n\t"
2049 "pavgb %[ftmp4], %[ftmp4], %[ftmp2] \n\t"
2050 "pavgb %[ftmp3], %[ftmp3], %[ftmp5] \n\t"
2051 "paddusb %[ftmp3], %[ftmp3], %[ftmp4] \n\t"
2052 "psubusb %[ftmp6], %[ff_pb_A1], %[ftmp3] \n\t"
2053 "psubusb %[ftmp3], %[ftmp3], %[ff_pb_A1] \n\t"
2054 "pminub %[ftmp6], %[ftmp6], %[ftmp7] \n\t"
2055 "pminub %[ftmp3], %[ftmp3], %[ftmp7] \n\t"
2056 "psubusb %[ftmp1], %[ftmp1], %[ftmp6] \n\t"
2057 "psubusb %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2058 "paddusb %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
2059 "paddusb %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
2061 "punpckhwd %[ftmp4], %[ftmp9], %[ftmp9] \n\t"
2062 "punpckhwd %[ftmp5], %[ftmp1], %[ftmp1] \n\t"
2063 "punpckhwd %[ftmp6], %[ftmp2], %[ftmp2] \n\t"
2064 "punpcklbh %[ftmp0], %[ftmp9], %[ftmp1] \n\t"
2065 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp10] \n\t"
2066 "punpcklhw %[ftmp1], %[ftmp0], %[ftmp2] \n\t"
2067 "punpckhhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2068 MMI_USWC1(%[ftmp1], %[addr5], 0x00)
2069 PTR_ADDU "%[addr3], %[addr5], %[stride] \n\t"
2070 "punpckhwd %[ftmp1], %[ftmp1], %[ftmp1] \n\t"
2071 MMI_USWC1(%[ftmp1], %[addr3], 0x00)
2072 PTR_ADDU "%[addr4], %[addr5], %[addr0] \n\t"
2073 MMI_USWC1(%[ftmp0], %[addr4], 0x00)
2074 "punpckhwd %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
2075 "punpckhwd %[ftmp3], %[ftmp10], %[ftmp10] \n\t"
2076 MMI_USWC1(%[ftmp0], %[pix], 0x00)
2077 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2078 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
2079 PTR_ADDU "%[addr3], %[pix], %[stride] \n\t"
2080 "punpcklhw %[ftmp5], %[ftmp4], %[ftmp6] \n\t"
2081 "punpckhhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2082 MMI_USWC1(%[ftmp5], %[addr3], 0x00)
2083 "punpckhwd %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
2084 PTR_ADDU "%[addr3], %[pix], %[addr0] \n\t"
2085 PTR_ADDU "%[addr4], %[pix], %[addr1] \n\t"
2086 MMI_USWC1(%[ftmp5], %[addr3], 0x00)
2087 MMI_USWC1(%[ftmp4], %[addr4], 0x00)
2088 PTR_ADDU "%[addr3], %[pix], %[addr2] \n\t"
2089 "punpckhwd %[ftmp9], %[ftmp4], %[ftmp4] \n\t"
2090 MMI_USWC1(%[ftmp9], %[addr3], 0x00)
2091 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
2092 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
2093 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
2094 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
2095 [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),
2096 [ftmp10]"=&f"(ftmp[10]),
2098 [addr0]"=&r"(addr[0]), [addr1]"=&r"(addr[1]),
2099 [addr2]"=&r"(addr[2]), [addr3]"=&r"(addr[3]),
2100 [addr4]"=&r"(addr[4]), [addr5]"=&r"(addr[5]),
2102 : [alpha]"r"(alpha), [beta]"r"(beta),
2103 [stride]"r"((mips_reg)stride), [tc0]"r"(tc0),
2104 [ff_pb_1]"f"(ff_pb_1), [ff_pb_3]"f"(ff_pb_3),
2105 [ff_pb_A1]"f"(ff_pb_A1)
2110 void ff_deblock_h_chroma_intra_8_mmi(uint8_t *pix, int stride, int alpha,
2118 "addi %[alpha], %[alpha], -0x01 \n\t"
2119 "addi %[beta], %[beta], -0x01 \n\t"
2120 PTR_ADDU "%[addr0], %[stride], %[stride] \n\t"
2121 PTR_ADDI "%[pix], %[pix], -0x02 \n\t"
2122 PTR_ADDU "%[addr1], %[addr0], %[stride] \n\t"
2123 PTR_ADDU "%[addr2], %[addr0], %[addr0] \n\t"
2124 "or %[addr5], $0, %[pix] \n\t"
2125 PTR_ADDU "%[pix], %[pix], %[addr1] \n\t"
2126 MMI_ULWC1(%[ftmp0], %[addr5], 0x00)
2127 PTR_ADDU "%[addr3], %[addr5], %[stride] \n\t"
2128 MMI_ULWC1(%[ftmp2], %[addr3], 0x00)
2129 PTR_ADDU "%[addr4], %[addr5], %[addr0] \n\t"
2130 MMI_ULWC1(%[ftmp1], %[addr4], 0x00)
2131 MMI_ULWC1(%[ftmp3], %[pix], 0x00)
2132 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2133 "punpcklbh %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
2134 PTR_ADDU "%[addr3], %[pix], %[stride] \n\t"
2135 "punpckhhw %[ftmp2], %[ftmp0], %[ftmp1] \n\t"
2136 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2137 MMI_ULWC1(%[ftmp4], %[addr3], 0x00)
2138 PTR_ADDU "%[addr4], %[pix], %[addr0] \n\t"
2139 MMI_ULWC1(%[ftmp6], %[addr4], 0x00)
2140 PTR_ADDU "%[addr3], %[pix], %[addr1] \n\t"
2141 MMI_ULWC1(%[ftmp5], %[addr3], 0x00)
2142 PTR_ADDU "%[addr4], %[pix], %[addr2] \n\t"
2143 MMI_ULWC1(%[ftmp7], %[addr4], 0x00)
2144 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2145 "punpcklbh %[ftmp5], %[ftmp5], %[ftmp7] \n\t"
2146 "mov.d %[ftmp6], %[ftmp4] \n\t"
2147 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2148 "punpckhhw %[ftmp6], %[ftmp6], %[ftmp5] \n\t"
2149 "punpckhwd %[ftmp1], %[ftmp0], %[ftmp4] \n\t"
2150 "punpckhwd %[ftmp3], %[ftmp2], %[ftmp6] \n\t"
2151 "punpcklwd %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
2152 "punpcklwd %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
2154 "xor %[ftmp8], %[ftmp8], %[ftmp8] \n\t"
2155 "mtc1 %[alpha], %[ftmp4] \n\t"
2156 "mtc1 %[beta], %[ftmp5] \n\t"
2157 "pshufh %[ftmp4], %[ftmp4], %[ftmp8] \n\t"
2158 "pshufh %[ftmp5], %[ftmp5], %[ftmp8] \n\t"
2159 "packushb %[ftmp4], %[ftmp4], %[ftmp4] \n\t"
2160 "packushb %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
2161 "psubusb %[ftmp6], %[ftmp2], %[ftmp1] \n\t"
2162 "psubusb %[ftmp7], %[ftmp1], %[ftmp2] \n\t"
2163 "or %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
2164 "psubusb %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
2165 "psubusb %[ftmp6], %[ftmp1], %[ftmp0] \n\t"
2166 "psubusb %[ftmp4], %[ftmp0], %[ftmp1] \n\t"
2167 "or %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2168 "psubusb %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2169 "or %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
2170 "psubusb %[ftmp6], %[ftmp2], %[ftmp3] \n\t"
2171 "psubusb %[ftmp4], %[ftmp3], %[ftmp2] \n\t"
2172 "or %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2173 "psubusb %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2174 "or %[ftmp7], %[ftmp7], %[ftmp4] \n\t"
2175 "xor %[ftmp6], %[ftmp6], %[ftmp6] \n\t"
2176 "pcmpeqb %[ftmp7], %[ftmp7], %[ftmp6] \n\t"
2177 "mov.d %[ftmp5], %[ftmp1] \n\t"
2178 "mov.d %[ftmp6], %[ftmp2] \n\t"
2179 "xor %[ftmp4], %[ftmp1], %[ftmp3] \n\t"
2180 "and %[ftmp4], %[ftmp4], %[ff_pb_1] \n\t"
2181 "pavgb %[ftmp1], %[ftmp1], %[ftmp3] \n\t"
2182 "psubusb %[ftmp1], %[ftmp1], %[ftmp4] \n\t"
2183 "pavgb %[ftmp1], %[ftmp1], %[ftmp0] \n\t"
2184 "xor %[ftmp4], %[ftmp2], %[ftmp0] \n\t"
2185 "and %[ftmp4], %[ftmp4], %[ff_pb_1] \n\t"
2186 "pavgb %[ftmp2], %[ftmp2], %[ftmp0] \n\t"
2187 "psubusb %[ftmp2], %[ftmp2], %[ftmp4] \n\t"
2188 "pavgb %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2189 "psubb %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
2190 "psubb %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
2191 "and %[ftmp1], %[ftmp1], %[ftmp7] \n\t"
2192 "and %[ftmp2], %[ftmp2], %[ftmp7] \n\t"
2193 "paddb %[ftmp1], %[ftmp1], %[ftmp5] \n\t"
2194 "paddb %[ftmp2], %[ftmp2], %[ftmp6] \n\t"
2196 "punpckhwd %[ftmp4], %[ftmp0], %[ftmp0] \n\t"
2197 "punpckhwd %[ftmp5], %[ftmp1], %[ftmp1] \n\t"
2198 "punpckhwd %[ftmp6], %[ftmp2], %[ftmp2] \n\t"
2199 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2200 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2201 "punpcklhw %[ftmp1], %[ftmp0], %[ftmp2] \n\t"
2202 "punpckhhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2203 MMI_USWC1(%[ftmp1], %[addr5], 0x00)
2204 PTR_ADDU "%[addr3], %[addr5], %[stride] \n\t"
2205 "punpckhwd %[ftmp1], %[ftmp1], %[ftmp1] \n\t"
2206 PTR_ADDU "%[addr4], %[addr5], %[addr0] \n\t"
2207 MMI_USWC1(%[ftmp1], %[addr3], 0x00)
2208 MMI_USWC1(%[ftmp0], %[addr4], 0x00)
2209 "punpckhwd %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
2210 "punpckhwd %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
2211 MMI_USWC1(%[ftmp0], %[pix], 0x00)
2212 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2213 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
2214 PTR_ADDU "%[addr3], %[pix], %[stride] \n\t"
2215 "punpcklhw %[ftmp5], %[ftmp4], %[ftmp6] \n\t"
2216 "punpckhhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2217 MMI_USWC1(%[ftmp5], %[addr3], 0x00)
2218 "punpckhwd %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
2219 PTR_ADDU "%[addr3], %[pix], %[addr0] \n\t"
2220 PTR_ADDU "%[addr4], %[pix], %[addr1] \n\t"
2221 MMI_USWC1(%[ftmp5], %[addr3], 0x00)
2222 PTR_ADDU "%[addr3], %[pix], %[addr2] \n\t"
2223 MMI_USWC1(%[ftmp4], %[addr4], 0x00)
2224 "punpckhwd %[ftmp9], %[ftmp4], %[ftmp4] \n\t"
2225 MMI_USWC1(%[ftmp9], %[addr3], 0x00)
2226 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
2227 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
2228 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
2229 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
2230 [ftmp8]"=&f"(ftmp[8]), [ftmp9]"=&f"(ftmp[9]),
2231 [ftmp10]"=&f"(ftmp[10]),
2233 [addr0]"=&r"(addr[0]), [addr1]"=&r"(addr[1]),
2234 [addr2]"=&r"(addr[2]), [addr3]"=&r"(addr[3]),
2235 [addr4]"=&r"(addr[4]), [addr5]"=&r"(addr[5]),
2237 : [alpha]"r"(alpha), [beta]"r"(beta),
2238 [stride]"r"((mips_reg)stride), [ff_pb_1]"f"(ff_pb_1)
2243 void ff_deblock_v_luma_8_mmi(uint8_t *pix, int stride, int alpha, int beta,
2246 if ((tc0[0] & tc0[1]) >= 0)
2247 ff_deblock_v8_luma_8_mmi(pix + 0, stride, alpha, beta, tc0);
2248 if ((tc0[2] & tc0[3]) >= 0)
2249 ff_deblock_v8_luma_8_mmi(pix + 8, stride, alpha, beta, tc0 + 2);
2252 void ff_deblock_v_luma_intra_8_mmi(uint8_t *pix, int stride, int alpha,
2255 deblock_v8_luma_intra_8_mmi(pix + 0, stride, alpha, beta);
2256 deblock_v8_luma_intra_8_mmi(pix + 8, stride, alpha, beta);
2259 void ff_deblock_h_luma_8_mmi(uint8_t *pix, int stride, int alpha, int beta,
2262 DECLARE_ALIGNED(8, const uint64_t, stack[0x0d]);
2269 PTR_ADDU "%[addr0], %[stride], %[stride] \n\t"
2270 PTR_ADDI "%[addr1], %[pix], -0x4 \n\t"
2271 PTR_ADDU "%[addr2], %[stride], %[addr0] \n\t"
2272 MMI_ULDC1(%[ftmp0], %[addr1], 0x00)
2273 PTR_ADDU "%[addr3], %[addr1], %[stride] \n\t"
2274 PTR_ADDU "%[addr4], %[addr1], %[addr2] \n\t"
2275 MMI_ULDC1(%[ftmp1], %[addr3], 0x00)
2276 PTR_ADDU "%[addr5], %[addr1], %[addr0] \n\t"
2277 MMI_ULDC1(%[ftmp2], %[addr5], 0x00)
2278 MMI_ULDC1(%[ftmp3], %[addr4], 0x00)
2279 PTR_ADDU "%[addr3], %[addr4], %[stride] \n\t"
2280 MMI_ULDC1(%[ftmp4], %[addr3], 0x00)
2281 PTR_ADDU "%[addr5], %[addr4], %[addr0] \n\t"
2282 MMI_ULDC1(%[ftmp5], %[addr5], 0x00)
2283 PTR_ADDU "%[addr3], %[addr4], %[addr2] \n\t"
2284 MMI_ULDC1(%[ftmp6], %[addr3], 0x00)
2285 PTR_ADDU "%[addr6], %[addr0], %[addr0] \n\t"
2286 "punpckhbh %[ftmp7], %[ftmp0], %[ftmp1] \n\t"
2287 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2288 "punpckhbh %[ftmp1], %[ftmp2], %[ftmp3] \n\t"
2289 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2290 "punpckhbh %[ftmp3], %[ftmp4], %[ftmp5] \n\t"
2291 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2292 PTR_ADDU "%[addr3], %[addr4], %[addr6] \n\t"
2293 MMI_SDC1(%[ftmp1], %[stack], 0x10)
2294 MMI_ULDC1(%[ftmp8], %[addr3], 0x00)
2295 PTR_ADDU "%[addr7], %[addr6], %[addr6] \n\t"
2296 "punpckhbh %[ftmp5], %[ftmp6], %[ftmp8] \n\t"
2297 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
2298 "punpckhhw %[ftmp1], %[ftmp0], %[ftmp2] \n\t"
2299 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2300 "punpckhhw %[ftmp2], %[ftmp4], %[ftmp6] \n\t"
2301 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2302 MMI_LDC1(%[ftmp8], %[stack], 0x10)
2303 "punpckhwd %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
2304 MMI_SDC1(%[ftmp0], %[stack], 0x00)
2305 "punpckhhw %[ftmp6], %[ftmp7], %[ftmp8] \n\t"
2306 "punpcklhw %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
2307 "punpckhhw %[ftmp0], %[ftmp3], %[ftmp5] \n\t"
2308 "punpcklhw %[ftmp3], %[ftmp3], %[ftmp5] \n\t"
2309 "punpcklwd %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
2310 "punpckhwd %[ftmp5], %[ftmp7], %[ftmp3] \n\t"
2311 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
2312 "punpckhwd %[ftmp3], %[ftmp1], %[ftmp2] \n\t"
2313 "punpcklwd %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
2314 MMI_SDC1(%[ftmp1], %[stack], 0x10)
2315 MMI_SDC1(%[ftmp3], %[stack], 0x20)
2316 MMI_SDC1(%[ftmp7], %[stack], 0x30)
2317 MMI_SDC1(%[ftmp5], %[stack], 0x40)
2318 MMI_SDC1(%[ftmp6], %[stack], 0x50)
2319 PTR_ADDU "%[addr1], %[addr1], %[addr7] \n\t"
2320 PTR_ADDU "%[addr4], %[addr4], %[addr7] \n\t"
2321 MMI_ULDC1(%[ftmp0], %[addr1], 0x00)
2322 PTR_ADDU "%[addr3], %[addr1], %[stride] \n\t"
2323 MMI_ULDC1(%[ftmp1], %[addr3], 0x00)
2324 PTR_ADDU "%[addr5], %[addr1], %[addr0] \n\t"
2325 MMI_ULDC1(%[ftmp2], %[addr5], 0x00)
2326 MMI_ULDC1(%[ftmp3], %[addr4], 0x00)
2327 PTR_ADDU "%[addr3], %[addr4], %[stride] \n\t"
2328 MMI_ULDC1(%[ftmp4], %[addr3], 0x00)
2329 PTR_ADDU "%[addr5], %[addr4], %[addr0] \n\t"
2330 MMI_ULDC1(%[ftmp5], %[addr5], 0x00)
2331 PTR_ADDU "%[addr3], %[addr4], %[addr2] \n\t"
2332 MMI_ULDC1(%[ftmp6], %[addr3], 0x00)
2333 "punpckhbh %[ftmp7], %[ftmp0], %[ftmp1] \n\t"
2334 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2335 "punpckhbh %[ftmp1], %[ftmp2], %[ftmp3] \n\t"
2336 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2337 "punpckhbh %[ftmp3], %[ftmp4], %[ftmp5] \n\t"
2338 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2339 PTR_ADDU "%[addr3], %[addr4], %[addr6] \n\t"
2340 MMI_SDC1(%[ftmp1], %[stack], 0x18)
2341 MMI_ULDC1(%[ftmp8], %[addr3], 0x00)
2342 "punpckhhw %[ftmp1], %[ftmp0], %[ftmp2] \n\t"
2343 "punpckhbh %[ftmp5], %[ftmp6], %[ftmp8] \n\t"
2344 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
2345 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2346 "punpckhhw %[ftmp2], %[ftmp4], %[ftmp6] \n\t"
2347 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2348 "punpckhwd %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
2349 MMI_LDC1(%[ftmp8], %[stack], 0x18)
2350 MMI_SDC1(%[ftmp0], %[stack], 0x08)
2351 "punpckhhw %[ftmp6], %[ftmp7], %[ftmp8] \n\t"
2352 "punpcklhw %[ftmp7], %[ftmp7], %[ftmp8] \n\t"
2353 "punpckhhw %[ftmp0], %[ftmp3], %[ftmp5] \n\t"
2354 "punpcklhw %[ftmp3], %[ftmp3], %[ftmp5] \n\t"
2355 "punpckhwd %[ftmp5], %[ftmp7], %[ftmp3] \n\t"
2356 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp3] \n\t"
2357 "punpckhwd %[ftmp3], %[ftmp1], %[ftmp2] \n\t"
2358 "punpcklwd %[ftmp1], %[ftmp1], %[ftmp2] \n\t"
2359 "punpcklwd %[ftmp6], %[ftmp6], %[ftmp0] \n\t"
2360 MMI_SDC1(%[ftmp1], %[stack], 0x18)
2361 MMI_SDC1(%[ftmp3], %[stack], 0x28)
2362 MMI_SDC1(%[ftmp7], %[stack], 0x38)
2363 MMI_SDC1(%[ftmp5], %[stack], 0x48)
2364 MMI_SDC1(%[ftmp6], %[stack], 0x58)
2365 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
2366 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
2367 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
2368 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
2369 [ftmp8]"=&f"(ftmp[8]),
2371 [addr0]"=&r"(addr[0]), [addr1]"=&r"(addr[1]),
2372 [addr2]"=&r"(addr[2]), [addr3]"=&r"(addr[3]),
2373 [addr4]"=&r"(addr[4]), [addr5]"=&r"(addr[5]),
2374 [addr6]"=&r"(addr[6]), [addr7]"=&r"(addr[7])
2375 : [pix]"r"(pix), [stride]"r"((mips_reg)stride),
2380 ff_deblock_v_luma_8_mmi((uint8_t *) &stack[6], 0x10, alpha, beta, tc0);
2383 PTR_ADDU "%[addr0], %[stride], %[stride] \n\t"
2384 PTR_ADDI "%[addr1], %[pix], -0x02 \n\t"
2385 PTR_ADDU "%[addr6], %[addr0], %[addr0] \n\t"
2386 PTR_ADDU "%[addr2], %[addr0], %[stride] \n\t"
2387 PTR_ADDU "%[addr7], %[addr6], %[addr6] \n\t"
2388 PTR_ADDU "%[addr4], %[addr1], %[addr2] \n\t"
2389 MMI_LDC1(%[ftmp0], %[stack], 0x10)
2390 MMI_LDC1(%[ftmp1], %[stack], 0x20)
2391 MMI_LDC1(%[ftmp2], %[stack], 0x30)
2392 MMI_LDC1(%[ftmp3], %[stack], 0x40)
2393 "punpckhwd %[ftmp4], %[ftmp0], %[ftmp0] \n\t"
2394 "punpckhwd %[ftmp5], %[ftmp1], %[ftmp1] \n\t"
2395 "punpckhwd %[ftmp6], %[ftmp2], %[ftmp2] \n\t"
2396 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2397 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2398 "punpcklhw %[ftmp1], %[ftmp0], %[ftmp2] \n\t"
2399 "punpckhhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2400 MMI_USWC1(%[ftmp1], %[addr1], 0x00)
2401 PTR_ADDU "%[addr3], %[addr1], %[stride] \n\t"
2402 "punpckhwd %[ftmp1], %[ftmp1], %[ftmp1] \n\t"
2403 PTR_ADDU "%[addr5], %[addr1], %[addr0] \n\t"
2404 MMI_USWC1(%[ftmp1], %[addr3], 0x00)
2405 MMI_USWC1(%[ftmp0], %[addr5], 0x00)
2406 "punpckhwd %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
2407 "punpckhwd %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
2408 MMI_USWC1(%[ftmp0], %[addr4], 0x00)
2409 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2410 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
2411 "punpcklhw %[ftmp5], %[ftmp4], %[ftmp6] \n\t"
2412 PTR_ADDU "%[addr3], %[addr4], %[stride] \n\t"
2413 "punpckhhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2414 MMI_USWC1(%[ftmp5], %[addr3], 0x00)
2415 PTR_ADDU "%[addr3], %[addr4], %[addr0] \n\t"
2416 "punpckhwd %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
2417 PTR_ADDU "%[addr5], %[addr4], %[addr2] \n\t"
2418 MMI_USWC1(%[ftmp5], %[addr3], 0x00)
2419 MMI_USWC1(%[ftmp4], %[addr5], 0x00)
2420 PTR_ADDU "%[addr3], %[addr4], %[addr6] \n\t"
2421 "punpckhwd %[ftmp4], %[ftmp4], %[ftmp4] \n\t"
2422 PTR_ADDU "%[addr1], %[addr1], %[addr7] \n\t"
2423 MMI_USWC1(%[ftmp4], %[addr3], 0x00)
2424 PTR_ADDU "%[addr4], %[addr4], %[addr7] \n\t"
2425 MMI_LDC1(%[ftmp0], %[stack], 0x18)
2426 MMI_LDC1(%[ftmp1], %[stack], 0x28)
2427 MMI_LDC1(%[ftmp2], %[stack], 0x38)
2428 MMI_LDC1(%[ftmp3], %[stack], 0x48)
2429 PTR_ADDU "%[addr0], %[stride], %[stride] \n\t"
2430 "punpckhwd %[ftmp4], %[ftmp0], %[ftmp0] \n\t"
2431 PTR_ADDU "%[addr6], %[addr0], %[addr0] \n\t"
2432 "punpckhwd %[ftmp5], %[ftmp1], %[ftmp1] \n\t"
2433 "punpckhwd %[ftmp6], %[ftmp2], %[ftmp2] \n\t"
2434 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2435 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2436 PTR_ADDU "%[addr3], %[addr1], %[stride] \n\t"
2437 "punpcklhw %[ftmp1], %[ftmp0], %[ftmp2] \n\t"
2438 "punpckhhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2439 MMI_USWC1(%[ftmp1], %[addr1], 0x00)
2440 "punpckhwd %[ftmp1], %[ftmp1], %[ftmp1] \n\t"
2441 PTR_ADDU "%[addr5], %[addr1], %[addr0] \n\t"
2442 MMI_USWC1(%[ftmp1], %[addr3], 0x00)
2443 MMI_USWC1(%[ftmp0], %[addr5], 0x00)
2444 "punpckhwd %[ftmp0], %[ftmp0], %[ftmp0] \n\t"
2445 "punpckhwd %[ftmp3], %[ftmp3], %[ftmp3] \n\t"
2446 MMI_USWC1(%[ftmp0], %[addr4], 0x00)
2447 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2448 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp3] \n\t"
2449 PTR_ADDU "%[addr3], %[addr4], %[stride] \n\t"
2450 "punpcklhw %[ftmp5], %[ftmp4], %[ftmp6] \n\t"
2451 "punpckhhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2452 MMI_USWC1(%[ftmp5], %[addr3], 0x00)
2453 PTR_ADDU "%[addr3], %[addr4], %[addr0] \n\t"
2454 "punpckhwd %[ftmp5], %[ftmp5], %[ftmp5] \n\t"
2455 PTR_ADDU "%[addr5], %[addr4], %[addr2] \n\t"
2456 MMI_USWC1(%[ftmp5], %[addr3], 0x00)
2457 MMI_USWC1(%[ftmp4], %[addr5], 0x00)
2458 PTR_ADDU "%[addr3], %[addr4], %[addr6] \n\t"
2459 "punpckhwd %[ftmp4], %[ftmp4], %[ftmp4] \n\t"
2460 MMI_USWC1(%[ftmp4], %[addr3], 0x00)
2461 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
2462 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
2463 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
2464 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
2465 [ftmp8]"=&f"(ftmp[8]),
2468 [addr0]"=&r"(addr[0]), [addr1]"=&r"(addr[1]),
2469 [addr2]"=&r"(addr[2]), [addr3]"=&r"(addr[3]),
2470 [addr4]"=&r"(addr[4]), [addr5]"=&r"(addr[5]),
2471 [addr6]"=&r"(addr[6]), [addr7]"=&r"(addr[7])
2472 : [pix]"r"(pix), [stride]"r"((mips_reg)stride),
2478 void ff_deblock_h_luma_intra_8_mmi(uint8_t *pix, int stride, int alpha,
2481 DECLARE_ALIGNED(8, const uint64_t, ptmp[0x11]);
2482 DECLARE_ALIGNED(8, const uint64_t, pdat[0x04]);
2488 PTR_ADDU "%[addr0], %[stride], %[stride] \n\t"
2489 PTR_ADDI "%[addr1], %[pix], -0x04 \n\t"
2490 PTR_ADDU "%[addr2], %[addr0], %[stride] \n\t"
2491 PTR_ADDU "%[addr3], %[addr0], %[addr0] \n\t"
2492 PTR_ADDU "%[addr4], %[addr1], %[addr2] \n\t"
2493 PTR_ADDU "%[addr5], %[addr1], %[stride] \n\t"
2494 MMI_ULDC1(%[ftmp0], %[addr1], 0x00)
2495 PTR_ADDU "%[addr6], %[addr1], %[addr0] \n\t"
2496 MMI_ULDC1(%[ftmp1], %[addr5], 0x00)
2497 MMI_ULDC1(%[ftmp2], %[addr6], 0x00)
2498 PTR_ADDU "%[addr5], %[addr4], %[stride] \n\t"
2499 MMI_ULDC1(%[ftmp3], %[addr4], 0x00)
2500 PTR_ADDU "%[addr6], %[addr4], %[addr0] \n\t"
2501 MMI_ULDC1(%[ftmp4], %[addr5], 0x00)
2502 PTR_ADDU "%[addr5], %[addr4], %[addr2] \n\t"
2503 MMI_ULDC1(%[ftmp5], %[addr6], 0x00)
2504 MMI_ULDC1(%[ftmp6], %[addr5], 0x00)
2505 PTR_ADDU "%[addr5], %[addr4], %[addr3] \n\t"
2506 "punpckhbh %[ftmp7], %[ftmp0], %[ftmp1] \n\t"
2507 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2508 "punpckhbh %[ftmp1], %[ftmp2], %[ftmp3] \n\t"
2509 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2510 "punpckhbh %[ftmp3], %[ftmp4], %[ftmp5] \n\t"
2511 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2512 MMI_ULDC1(%[ftmp8], %[addr5], 0x00)
2513 "punpckhbh %[ftmp5], %[ftmp6], %[ftmp8] \n\t"
2514 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
2515 MMI_SDC1(%[ftmp3], %[ptmp], 0x00)
2516 "punpckhhw %[ftmp3], %[ftmp0], %[ftmp2] \n\t"
2517 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2518 "punpckhhw %[ftmp2], %[ftmp4], %[ftmp6] \n\t"
2519 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2520 "punpckhhw %[ftmp6], %[ftmp7], %[ftmp1] \n\t"
2521 "punpcklhw %[ftmp7], %[ftmp7], %[ftmp1] \n\t"
2522 MMI_SDC1(%[ftmp2], %[ptmp], 0x20)
2523 MMI_LDC1(%[ftmp2], %[ptmp], 0x00)
2524 "punpckhhw %[ftmp1], %[ftmp2], %[ftmp5] \n\t"
2525 "punpcklhw %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
2526 "punpckhwd %[ftmp5], %[ftmp0], %[ftmp4] \n\t"
2527 "punpcklwd %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
2528 "punpckhwd %[ftmp4], %[ftmp7], %[ftmp2] \n\t"
2529 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
2530 MMI_SDC1(%[ftmp0], %[ptmp], 0x00)
2531 MMI_SDC1(%[ftmp5], %[ptmp], 0x10)
2532 MMI_SDC1(%[ftmp7], %[ptmp], 0x40)
2533 MMI_SDC1(%[ftmp4], %[ptmp], 0x50)
2534 MMI_LDC1(%[ftmp8], %[ptmp], 0x20)
2535 "punpckhwd %[ftmp0], %[ftmp3], %[ftmp8] \n\t"
2536 "punpcklwd %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
2537 "punpckhwd %[ftmp5], %[ftmp6], %[ftmp1] \n\t"
2538 "punpcklwd %[ftmp6], %[ftmp6], %[ftmp1] \n\t"
2539 PTR_ADDU "%[addr5], %[addr3], %[addr3] \n\t"
2540 MMI_SDC1(%[ftmp3], %[ptmp], 0x20)
2541 MMI_SDC1(%[ftmp0], %[ptmp], 0x30)
2542 MMI_SDC1(%[ftmp6], %[ptmp], 0x60)
2543 MMI_SDC1(%[ftmp5], %[ptmp], 0x70)
2544 PTR_ADDU "%[addr1], %[addr1], %[addr5] \n\t"
2545 PTR_ADDU "%[addr4], %[addr4], %[addr5] \n\t"
2546 PTR_ADDU "%[addr5], %[addr1], %[stride] \n\t"
2547 MMI_ULDC1(%[ftmp0], %[addr1], 0x00)
2548 PTR_ADDU "%[addr6], %[addr1], %[addr0] \n\t"
2549 MMI_ULDC1(%[ftmp1], %[addr5], 0x00)
2550 MMI_ULDC1(%[ftmp2], %[addr6], 0x00)
2551 PTR_ADDU "%[addr5], %[addr4], %[stride] \n\t"
2552 MMI_ULDC1(%[ftmp3], %[addr4], 0x00)
2553 PTR_ADDU "%[addr6], %[addr4], %[addr0] \n\t"
2554 MMI_ULDC1(%[ftmp4], %[addr5], 0x00)
2555 PTR_ADDU "%[addr5], %[addr4], %[addr2] \n\t"
2556 MMI_ULDC1(%[ftmp5], %[addr6], 0x00)
2557 MMI_ULDC1(%[ftmp6], %[addr5], 0x00)
2558 PTR_ADDU "%[addr5], %[addr4], %[addr3] \n\t"
2559 "punpckhbh %[ftmp7], %[ftmp0], %[ftmp1] \n\t"
2560 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2561 "punpckhbh %[ftmp1], %[ftmp2], %[ftmp3] \n\t"
2562 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2563 "punpckhbh %[ftmp3], %[ftmp4], %[ftmp5] \n\t"
2564 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2565 MMI_ULDC1(%[ftmp8], %[addr5], 0x00)
2566 "punpckhbh %[ftmp5], %[ftmp6], %[ftmp8] \n\t"
2567 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
2568 MMI_SDC1(%[ftmp3], %[ptmp], 0x08)
2569 "punpckhhw %[ftmp3], %[ftmp0], %[ftmp2] \n\t"
2570 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2571 "punpckhhw %[ftmp2], %[ftmp4], %[ftmp6] \n\t"
2572 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2573 "punpckhhw %[ftmp6], %[ftmp7], %[ftmp1] \n\t"
2574 "punpcklhw %[ftmp7], %[ftmp7], %[ftmp1] \n\t"
2575 MMI_SDC1(%[ftmp2], %[ptmp], 0x28)
2576 MMI_LDC1(%[ftmp2], %[ptmp], 0x08)
2577 "punpckhhw %[ftmp1], %[ftmp2], %[ftmp5] \n\t"
2578 "punpcklhw %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
2579 "punpckhwd %[ftmp5], %[ftmp0], %[ftmp4] \n\t"
2580 "punpcklwd %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
2581 "punpckhwd %[ftmp4], %[ftmp7], %[ftmp2] \n\t"
2582 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
2583 MMI_SDC1(%[ftmp0], %[ptmp], 0x08)
2584 MMI_SDC1(%[ftmp5], %[ptmp], 0x18)
2585 MMI_SDC1(%[ftmp7], %[ptmp], 0x48)
2586 MMI_SDC1(%[ftmp4], %[ptmp], 0x58)
2587 MMI_LDC1(%[ftmp8], %[ptmp], 0x28)
2588 "punpckhwd %[ftmp0], %[ftmp3], %[ftmp8] \n\t"
2589 "punpcklwd %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
2590 "punpckhwd %[ftmp5], %[ftmp6], %[ftmp1] \n\t"
2591 "punpcklwd %[ftmp6], %[ftmp6], %[ftmp1] \n\t"
2592 MMI_SDC1(%[ftmp3], %[ptmp], 0x28)
2593 MMI_SDC1(%[ftmp0], %[ptmp], 0x38)
2594 MMI_SDC1(%[ftmp6], %[ptmp], 0x68)
2595 MMI_SDC1(%[ftmp5], %[ptmp], 0x78)
2596 PTR_S "%[addr1], 0x00(%[pdat]) \n\t"
2597 PTR_S "%[addr2], 0x08(%[pdat]) \n\t"
2598 PTR_S "%[addr0], 0x10(%[pdat]) \n\t"
2599 PTR_S "%[addr3], 0x18(%[pdat]) \n\t"
2600 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
2601 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
2602 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
2603 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
2604 [ftmp8]"=&f"(ftmp[8]),
2606 [addr0]"=&r"(addr[0]), [addr1]"=&r"(addr[1]),
2607 [addr2]"=&r"(addr[2]), [addr3]"=&r"(addr[3]),
2608 [addr4]"=&r"(addr[4]), [addr5]"=&r"(addr[5]),
2609 [addr6]"=&r"(addr[6])
2610 : [pix]"r"(pix), [stride]"r"((mips_reg)stride),
2611 [ptmp]"r"(ptmp), [pdat]"r"(pdat)
2615 ff_deblock_v_luma_intra_8_mmi((uint8_t *) &ptmp[8], 0x10, alpha, beta);
2618 PTR_L "%[addr1], 0x00(%[pdat]) \n\t"
2619 PTR_L "%[addr2], 0x08(%[pdat]) \n\t"
2620 PTR_L "%[addr0], 0x10(%[pdat]) \n\t"
2621 PTR_L "%[addr3], 0x18(%[pdat]) \n\t"
2622 PTR_ADDU "%[addr4], %[addr1], %[addr2] \n\t"
2623 MMI_LDC1(%[ftmp0], %[ptmp], 0x08)
2624 MMI_LDC1(%[ftmp1], %[ptmp], 0x18)
2625 MMI_LDC1(%[ftmp2], %[ptmp], 0x28)
2626 MMI_LDC1(%[ftmp3], %[ptmp], 0x38)
2627 MMI_LDC1(%[ftmp4], %[ptmp], 0x48)
2628 MMI_LDC1(%[ftmp5], %[ptmp], 0x58)
2629 MMI_LDC1(%[ftmp6], %[ptmp], 0x68)
2630 "punpckhbh %[ftmp7], %[ftmp0], %[ftmp1] \n\t"
2631 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2632 "punpckhbh %[ftmp1], %[ftmp2], %[ftmp3] \n\t"
2633 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2634 "punpckhbh %[ftmp3], %[ftmp4], %[ftmp5] \n\t"
2635 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2636 MMI_LDC1(%[ftmp8], %[ptmp], 0x78)
2637 "punpckhbh %[ftmp5], %[ftmp6], %[ftmp8] \n\t"
2638 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
2639 MMI_USDC1(%[ftmp3], %[addr1], 0x00)
2640 PTR_ADDU "%[addr5], %[addr1], %[addr0] \n\t"
2641 "punpckhhw %[ftmp3], %[ftmp0], %[ftmp2] \n\t"
2642 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2643 "punpckhhw %[ftmp2], %[ftmp4], %[ftmp6] \n\t"
2644 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2645 "punpckhhw %[ftmp6], %[ftmp7], %[ftmp1] \n\t"
2646 "punpcklhw %[ftmp7], %[ftmp7], %[ftmp1] \n\t"
2647 MMI_USDC1(%[ftmp2], %[addr5], 0x00)
2648 MMI_ULDC1(%[ftmp2], %[addr1], 0x00)
2649 "punpckhhw %[ftmp1], %[ftmp2], %[ftmp5] \n\t"
2650 "punpcklhw %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
2651 "punpckhwd %[ftmp5], %[ftmp0], %[ftmp4] \n\t"
2652 "punpcklwd %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
2653 "punpckhwd %[ftmp4], %[ftmp7], %[ftmp2] \n\t"
2654 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
2655 PTR_ADDU "%[addr5], %[addr1], %[stride] \n\t"
2656 MMI_USDC1(%[ftmp0], %[addr1], 0x00)
2657 PTR_ADDU "%[addr6], %[addr4], %[stride] \n\t"
2658 MMI_USDC1(%[ftmp5], %[addr5], 0x00)
2659 PTR_ADDU "%[addr5], %[addr4], %[addr0] \n\t"
2660 MMI_USDC1(%[ftmp7], %[addr6], 0x00)
2661 PTR_ADDU "%[addr6], %[addr1], %[addr0] \n\t"
2662 MMI_USDC1(%[ftmp4], %[addr5], 0x00)
2663 MMI_ULDC1(%[ftmp8], %[addr6], 0x00)
2664 PTR_ADDU "%[addr5], %[addr1], %[addr0] \n\t"
2665 "punpckhwd %[ftmp0], %[ftmp3], %[ftmp8] \n\t"
2666 "punpcklwd %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
2667 "punpckhwd %[ftmp5], %[ftmp6], %[ftmp1] \n\t"
2668 "punpcklwd %[ftmp6], %[ftmp6], %[ftmp1] \n\t"
2669 MMI_USDC1(%[ftmp3], %[addr5], 0x00)
2670 PTR_ADDU "%[addr5], %[addr4], %[addr2] \n\t"
2671 MMI_USDC1(%[ftmp0], %[addr4], 0x00)
2672 PTR_ADDU "%[addr6], %[addr4], %[addr3] \n\t"
2673 MMI_USDC1(%[ftmp6], %[addr5], 0x00)
2674 PTR_ADDU "%[addr5], %[addr3], %[addr3] \n\t"
2675 MMI_USDC1(%[ftmp5], %[addr6], 0x00)
2676 PTR_SUBU "%[addr1], %[addr1], %[addr5] \n\t"
2677 PTR_SUBU "%[addr4], %[addr4], %[addr5] \n\t"
2678 MMI_LDC1(%[ftmp0], %[ptmp], 0x00)
2679 MMI_LDC1(%[ftmp1], %[ptmp], 0x10)
2680 MMI_LDC1(%[ftmp2], %[ptmp], 0x20)
2681 MMI_LDC1(%[ftmp3], %[ptmp], 0x30)
2682 MMI_LDC1(%[ftmp4], %[ptmp], 0x40)
2683 MMI_LDC1(%[ftmp5], %[ptmp], 0x50)
2684 MMI_LDC1(%[ftmp6], %[ptmp], 0x60)
2685 "punpckhbh %[ftmp7], %[ftmp0], %[ftmp1] \n\t"
2686 "punpcklbh %[ftmp0], %[ftmp0], %[ftmp1] \n\t"
2687 "punpckhbh %[ftmp1], %[ftmp2], %[ftmp3] \n\t"
2688 "punpcklbh %[ftmp2], %[ftmp2], %[ftmp3] \n\t"
2689 "punpckhbh %[ftmp3], %[ftmp4], %[ftmp5] \n\t"
2690 "punpcklbh %[ftmp4], %[ftmp4], %[ftmp5] \n\t"
2691 MMI_LDC1(%[ftmp8], %[ptmp], 0x70)
2692 "punpckhbh %[ftmp5], %[ftmp6], %[ftmp8] \n\t"
2693 "punpcklbh %[ftmp6], %[ftmp6], %[ftmp8] \n\t"
2694 MMI_USDC1(%[ftmp3], %[addr1], 0x00)
2695 PTR_ADDU "%[addr5], %[addr1], %[addr0] \n\t"
2696 "punpckhhw %[ftmp3], %[ftmp0], %[ftmp2] \n\t"
2697 "punpcklhw %[ftmp0], %[ftmp0], %[ftmp2] \n\t"
2698 "punpckhhw %[ftmp2], %[ftmp4], %[ftmp6] \n\t"
2699 "punpcklhw %[ftmp4], %[ftmp4], %[ftmp6] \n\t"
2700 "punpckhhw %[ftmp6], %[ftmp7], %[ftmp1] \n\t"
2701 "punpcklhw %[ftmp7], %[ftmp7], %[ftmp1] \n\t"
2702 MMI_USDC1(%[ftmp2], %[addr5], 0x00)
2703 MMI_ULDC1(%[ftmp2], %[addr1], 0x00)
2704 "punpckhhw %[ftmp1], %[ftmp2], %[ftmp5] \n\t"
2705 "punpcklhw %[ftmp2], %[ftmp2], %[ftmp5] \n\t"
2706 "punpckhwd %[ftmp5], %[ftmp0], %[ftmp4] \n\t"
2707 "punpcklwd %[ftmp0], %[ftmp0], %[ftmp4] \n\t"
2708 "punpckhwd %[ftmp4], %[ftmp7], %[ftmp2] \n\t"
2709 "punpcklwd %[ftmp7], %[ftmp7], %[ftmp2] \n\t"
2710 PTR_ADDU "%[addr5], %[addr1], %[stride] \n\t"
2711 MMI_USDC1(%[ftmp0], %[addr1], 0x00)
2712 PTR_ADDU "%[addr6], %[addr4], %[stride] \n\t"
2713 MMI_USDC1(%[ftmp5], %[addr5], 0x00)
2714 PTR_ADDU "%[addr5], %[addr4], %[addr0] \n\t"
2715 MMI_USDC1(%[ftmp7], %[addr6], 0x00)
2716 PTR_ADDU "%[addr6], %[addr1], %[addr0] \n\t"
2717 MMI_USDC1(%[ftmp4], %[addr5], 0x00)
2718 MMI_ULDC1(%[ftmp8], %[addr6], 0x00)
2719 PTR_ADDU "%[addr5], %[addr1], %[addr0] \n\t"
2720 "punpckhwd %[ftmp0], %[ftmp3], %[ftmp8] \n\t"
2721 "punpcklwd %[ftmp3], %[ftmp3], %[ftmp8] \n\t"
2722 "punpckhwd %[ftmp5], %[ftmp6], %[ftmp1] \n\t"
2723 "punpcklwd %[ftmp6], %[ftmp6], %[ftmp1] \n\t"
2724 MMI_USDC1(%[ftmp3], %[addr5], 0x00)
2725 PTR_ADDU "%[addr5], %[addr4], %[addr2] \n\t"
2726 MMI_USDC1(%[ftmp0], %[addr4], 0x00)
2727 PTR_ADDU "%[addr6], %[addr4], %[addr3] \n\t"
2728 MMI_USDC1(%[ftmp6], %[addr5], 0x00)
2729 MMI_USDC1(%[ftmp5], %[addr6], 0x00)
2730 : [ftmp0]"=&f"(ftmp[0]), [ftmp1]"=&f"(ftmp[1]),
2731 [ftmp2]"=&f"(ftmp[2]), [ftmp3]"=&f"(ftmp[3]),
2732 [ftmp4]"=&f"(ftmp[4]), [ftmp5]"=&f"(ftmp[5]),
2733 [ftmp6]"=&f"(ftmp[6]), [ftmp7]"=&f"(ftmp[7]),
2734 [ftmp8]"=&f"(ftmp[8]),
2736 [addr0]"=&r"(addr[0]), [addr1]"=&r"(addr[1]),
2737 [addr2]"=&r"(addr[2]), [addr3]"=&r"(addr[3]),
2738 [addr4]"=&r"(addr[4]), [addr5]"=&r"(addr[5]),
2739 [addr6]"=&r"(addr[6])
2740 : [pix]"r"(pix), [stride]"r"((mips_reg)stride),
2741 [ptmp]"r"(ptmp), [pdat]"r"(pdat)
projects / ffmpeg / blob