qpel.c 38.2 KB
Newer Older
1
/*
2 3 4 5 6
 * This is optimized for sh, which have post increment addressing (*p++).
 * Some CPU may be index (p[n]) faster than post increment (*p++).
 *
 * copyright (c) 2001-2003 BERO <bero@geocities.co.jp>
 *
7 8 9
 * This file is part of FFmpeg.
 *
 * FFmpeg is free software; you can redistribute it and/or
10 11
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
12
 * version 2.1 of the License, or (at your option) any later version.
13
 *
14
 * FFmpeg is distributed in the hope that it will be useful,
15 16 17 18 19
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
20
 * License along with FFmpeg; if not, write to the Free Software
21 22
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 */
23

24
#include "libavutil/common.h"
25
#include "libavcodec/copy_block.h"
26
#include "libavcodec/rnd_avg.h"
27

28 29 30 31
#define PIXOP2(OPNAME, OP) \
\
static inline void OPNAME ## _pixels4_l2_aligned(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
32
        do {\
33
                OP(LP(dst  ),rnd_avg32(LPC(src1  ),LPC(src2  )) ); \
34 35 36 37
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
38 39 40 41
}\
\
static inline void OPNAME ## _pixels4_l2_aligned2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
42
        do {\
43
                OP(LP(dst  ),rnd_avg32(AV_RN32(src1  ),LPC(src2  )) ); \
44 45 46 47
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
48 49 50 51
}\
\
static inline void OPNAME ## _no_rnd_pixels16_l2_aligned2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
52
        do {\
53 54 55 56
                OP(LP(dst  ),no_rnd_avg32(AV_RN32(src1  ),LPC(src2  )) ); \
                OP(LP(dst+4),no_rnd_avg32(AV_RN32(src1+4),LPC(src2+4)) ); \
                OP(LP(dst+8),no_rnd_avg32(AV_RN32(src1+8),LPC(src2+8)) ); \
                OP(LP(dst+12),no_rnd_avg32(AV_RN32(src1+12),LPC(src2+12)) ); \
57 58 59 60
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
61 62 63 64
}\
\
static inline void OPNAME ## _pixels16_l2_aligned2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
65
        do {\
66 67 68 69
                OP(LP(dst  ),rnd_avg32(AV_RN32(src1  ),LPC(src2  )) ); \
                OP(LP(dst+4),rnd_avg32(AV_RN32(src1+4),LPC(src2+4)) ); \
                OP(LP(dst+8),rnd_avg32(AV_RN32(src1+8),LPC(src2+8)) ); \
                OP(LP(dst+12),rnd_avg32(AV_RN32(src1+12),LPC(src2+12)) ); \
70 71 72 73
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
74 75 76 77
}\
\
static inline void OPNAME ## _no_rnd_pixels8_l2_aligned2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
78
        do { /* onlye src2 aligned */\
79 80
                OP(LP(dst  ),no_rnd_avg32(AV_RN32(src1  ),LPC(src2  )) ); \
                OP(LP(dst+4),no_rnd_avg32(AV_RN32(src1+4),LPC(src2+4)) ); \
81 82 83 84
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
85 86 87 88
}\
\
static inline void OPNAME ## _pixels8_l2_aligned2(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
89
        do {\
90 91
                OP(LP(dst  ),rnd_avg32(AV_RN32(src1  ),LPC(src2  )) ); \
                OP(LP(dst+4),rnd_avg32(AV_RN32(src1+4),LPC(src2+4)) ); \
92 93 94 95
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
96 97 98 99
}\
\
static inline void OPNAME ## _no_rnd_pixels8_l2_aligned(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
100
        do {\
101 102
                OP(LP(dst  ),no_rnd_avg32(LPC(src1  ),LPC(src2  )) ); \
                OP(LP(dst+4),no_rnd_avg32(LPC(src1+4),LPC(src2+4)) ); \
103 104 105 106
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
107 108 109 110
}\
\
static inline void OPNAME ## _pixels8_l2_aligned(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
111
        do {\
112 113
                OP(LP(dst  ),rnd_avg32(LPC(src1  ),LPC(src2  )) ); \
                OP(LP(dst+4),rnd_avg32(LPC(src1+4),LPC(src2+4)) ); \
114 115 116 117
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
118 119 120 121
}\
\
static inline void OPNAME ## _no_rnd_pixels16_l2_aligned(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
122
        do {\
123 124 125 126
                OP(LP(dst  ),no_rnd_avg32(LPC(src1  ),LPC(src2  )) ); \
                OP(LP(dst+4),no_rnd_avg32(LPC(src1+4),LPC(src2+4)) ); \
                OP(LP(dst+8),no_rnd_avg32(LPC(src1+8),LPC(src2+8)) ); \
                OP(LP(dst+12),no_rnd_avg32(LPC(src1+12),LPC(src2+12)) ); \
127 128 129 130
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
131 132 133 134
}\
\
static inline void OPNAME ## _pixels16_l2_aligned(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{\
135
        do {\
136 137 138 139
                OP(LP(dst  ),rnd_avg32(LPC(src1  ),LPC(src2  )) ); \
                OP(LP(dst+4),rnd_avg32(LPC(src1+4),LPC(src2+4)) ); \
                OP(LP(dst+8),rnd_avg32(LPC(src1+8),LPC(src2+8)) ); \
                OP(LP(dst+12),rnd_avg32(LPC(src1+12),LPC(src2+12)) ); \
140 141 142 143
                src1+=src_stride1; \
                src2+=src_stride2; \
                dst+=dst_stride; \
        } while(--h); \
144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
}\
\
static inline void OPNAME ## _no_rnd_pixels16_l2_aligned1(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{ OPNAME ## _no_rnd_pixels16_l2_aligned2(dst,src2,src1,dst_stride,src_stride2,src_stride1,h); } \
\
static inline void OPNAME ## _pixels16_l2_aligned1(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{ OPNAME ## _pixels16_l2_aligned2(dst,src2,src1,dst_stride,src_stride2,src_stride1,h); } \
\
static inline void OPNAME ## _no_rnd_pixels8_l2_aligned1(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{ OPNAME ## _no_rnd_pixels8_l2_aligned2(dst,src2,src1,dst_stride,src_stride2,src_stride1,h); } \
\
static inline void OPNAME ## _pixels8_l2_aligned1(uint8_t *dst, const uint8_t *src1, const uint8_t *src2, int dst_stride, int src_stride1, int src_stride2, int h) \
{ OPNAME ## _pixels8_l2_aligned2(dst,src2,src1,dst_stride,src_stride2,src_stride1,h); } \
\
static inline void OPNAME ## _pixels8_l4_aligned(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
159 160
        do { \
                uint32_t a0,a1,a2,a3; \
161 162
                UNPACK(a0,a1,LPC(src1),LPC(src2)); \
                UNPACK(a2,a3,LPC(src3),LPC(src4)); \
163
                OP(LP(dst),rnd_PACK(a0,a1,a2,a3)); \
164 165
                UNPACK(a0,a1,LPC(src1+4),LPC(src2+4)); \
                UNPACK(a2,a3,LPC(src3+4),LPC(src4+4)); \
166 167 168 169 170 171 172
                OP(LP(dst+4),rnd_PACK(a0,a1,a2,a3)); \
                src1+=src_stride1;\
                src2+=src_stride2;\
                src3+=src_stride3;\
                src4+=src_stride4;\
                dst+=dst_stride;\
        } while(--h); \
173 174 175
} \
\
static inline void OPNAME ## _no_rnd_pixels8_l4_aligned(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
176 177
        do { \
                uint32_t a0,a1,a2,a3; \
178 179
                UNPACK(a0,a1,LPC(src1),LPC(src2)); \
                UNPACK(a2,a3,LPC(src3),LPC(src4)); \
180
                OP(LP(dst),no_rnd_PACK(a0,a1,a2,a3)); \
181 182
                UNPACK(a0,a1,LPC(src1+4),LPC(src2+4)); \
                UNPACK(a2,a3,LPC(src3+4),LPC(src4+4)); \
183 184 185 186 187 188 189
                OP(LP(dst+4),no_rnd_PACK(a0,a1,a2,a3)); \
                src1+=src_stride1;\
                src2+=src_stride2;\
                src3+=src_stride3;\
                src4+=src_stride4;\
                dst+=dst_stride;\
        } while(--h); \
190 191 192
} \
\
static inline void OPNAME ## _pixels8_l4_aligned0(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
193 194
        do { \
                uint32_t a0,a1,a2,a3; /* src1 only not aligned */\
195 196
                UNPACK(a0,a1,AV_RN32(src1),LPC(src2)); \
                UNPACK(a2,a3,LPC(src3),LPC(src4)); \
197
                OP(LP(dst),rnd_PACK(a0,a1,a2,a3)); \
198 199
                UNPACK(a0,a1,AV_RN32(src1+4),LPC(src2+4)); \
                UNPACK(a2,a3,LPC(src3+4),LPC(src4+4)); \
200 201 202 203 204 205 206
                OP(LP(dst+4),rnd_PACK(a0,a1,a2,a3)); \
                src1+=src_stride1;\
                src2+=src_stride2;\
                src3+=src_stride3;\
                src4+=src_stride4;\
                dst+=dst_stride;\
        } while(--h); \
207 208 209
} \
\
static inline void OPNAME ## _no_rnd_pixels8_l4_aligned0(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
210 211
        do { \
                uint32_t a0,a1,a2,a3; \
212 213
                UNPACK(a0,a1,AV_RN32(src1),LPC(src2)); \
                UNPACK(a2,a3,LPC(src3),LPC(src4)); \
214
                OP(LP(dst),no_rnd_PACK(a0,a1,a2,a3)); \
215 216
                UNPACK(a0,a1,AV_RN32(src1+4),LPC(src2+4)); \
                UNPACK(a2,a3,LPC(src3+4),LPC(src4+4)); \
217 218 219 220 221 222 223
                OP(LP(dst+4),no_rnd_PACK(a0,a1,a2,a3)); \
                src1+=src_stride1;\
                src2+=src_stride2;\
                src3+=src_stride3;\
                src4+=src_stride4;\
                dst+=dst_stride;\
        } while(--h); \
224 225 226
} \
\
static inline void OPNAME ## _pixels16_l4_aligned(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
227 228
        do { \
                uint32_t a0,a1,a2,a3; \
229 230
                UNPACK(a0,a1,LPC(src1),LPC(src2)); \
                UNPACK(a2,a3,LPC(src3),LPC(src4)); \
231
                OP(LP(dst),rnd_PACK(a0,a1,a2,a3)); \
232 233
                UNPACK(a0,a1,LPC(src1+4),LPC(src2+4)); \
                UNPACK(a2,a3,LPC(src3+4),LPC(src4+4)); \
234
                OP(LP(dst+8),rnd_PACK(a0,a1,a2,a3)); \
235 236
                UNPACK(a0,a1,LPC(src1+8),LPC(src2+8)); \
                UNPACK(a2,a3,LPC(src3+8),LPC(src4+8)); \
237
                OP(LP(dst+8),rnd_PACK(a0,a1,a2,a3)); \
238 239
                UNPACK(a0,a1,LPC(src1+12),LPC(src2+12)); \
                UNPACK(a2,a3,LPC(src3+12),LPC(src4+12)); \
240 241 242 243 244 245 246
                OP(LP(dst+12),rnd_PACK(a0,a1,a2,a3)); \
                src1+=src_stride1;\
                src2+=src_stride2;\
                src3+=src_stride3;\
                src4+=src_stride4;\
                dst+=dst_stride;\
        } while(--h); \
247 248 249
} \
\
static inline void OPNAME ## _no_rnd_pixels16_l4_aligned(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
250 251
        do { \
                uint32_t a0,a1,a2,a3; \
252 253
                UNPACK(a0,a1,LPC(src1),LPC(src2)); \
                UNPACK(a2,a3,LPC(src3),LPC(src4)); \
254
                OP(LP(dst),no_rnd_PACK(a0,a1,a2,a3)); \
255 256
                UNPACK(a0,a1,LPC(src1+4),LPC(src2+4)); \
                UNPACK(a2,a3,LPC(src3+4),LPC(src4+4)); \
257
                OP(LP(dst+4),no_rnd_PACK(a0,a1,a2,a3)); \
258 259
                UNPACK(a0,a1,LPC(src1+8),LPC(src2+8)); \
                UNPACK(a2,a3,LPC(src3+8),LPC(src4+8)); \
260
                OP(LP(dst+8),no_rnd_PACK(a0,a1,a2,a3)); \
261 262
                UNPACK(a0,a1,LPC(src1+12),LPC(src2+12)); \
                UNPACK(a2,a3,LPC(src3+12),LPC(src4+12)); \
263 264 265 266 267 268 269
                OP(LP(dst+12),no_rnd_PACK(a0,a1,a2,a3)); \
                src1+=src_stride1;\
                src2+=src_stride2;\
                src3+=src_stride3;\
                src4+=src_stride4;\
                dst+=dst_stride;\
        } while(--h); \
270 271 272
} \
\
static inline void OPNAME ## _pixels16_l4_aligned0(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
273 274
        do { /* src1 is unaligned */\
                uint32_t a0,a1,a2,a3; \
275 276
                UNPACK(a0,a1,AV_RN32(src1),LPC(src2)); \
                UNPACK(a2,a3,LPC(src3),LPC(src4)); \
277
                OP(LP(dst),rnd_PACK(a0,a1,a2,a3)); \
278 279
                UNPACK(a0,a1,AV_RN32(src1+4),LPC(src2+4)); \
                UNPACK(a2,a3,LPC(src3+4),LPC(src4+4)); \
280
                OP(LP(dst+8),rnd_PACK(a0,a1,a2,a3)); \
281 282
                UNPACK(a0,a1,AV_RN32(src1+8),LPC(src2+8)); \
                UNPACK(a2,a3,LPC(src3+8),LPC(src4+8)); \
283
                OP(LP(dst+8),rnd_PACK(a0,a1,a2,a3)); \
284 285
                UNPACK(a0,a1,AV_RN32(src1+12),LPC(src2+12)); \
                UNPACK(a2,a3,LPC(src3+12),LPC(src4+12)); \
286 287 288 289 290 291 292
                OP(LP(dst+12),rnd_PACK(a0,a1,a2,a3)); \
                src1+=src_stride1;\
                src2+=src_stride2;\
                src3+=src_stride3;\
                src4+=src_stride4;\
                dst+=dst_stride;\
        } while(--h); \
293 294 295
} \
\
static inline void OPNAME ## _no_rnd_pixels16_l4_aligned0(uint8_t *dst, const uint8_t *src1, uint8_t *src2, uint8_t *src3, uint8_t *src4,int dst_stride, int src_stride1, int src_stride2,int src_stride3,int src_stride4, int h){\
296 297
        do { \
                uint32_t a0,a1,a2,a3; \
298 299
                UNPACK(a0,a1,AV_RN32(src1),LPC(src2)); \
                UNPACK(a2,a3,LPC(src3),LPC(src4)); \
300
                OP(LP(dst),no_rnd_PACK(a0,a1,a2,a3)); \
301 302
                UNPACK(a0,a1,AV_RN32(src1+4),LPC(src2+4)); \
                UNPACK(a2,a3,LPC(src3+4),LPC(src4+4)); \
303
                OP(LP(dst+4),no_rnd_PACK(a0,a1,a2,a3)); \
304 305
                UNPACK(a0,a1,AV_RN32(src1+8),LPC(src2+8)); \
                UNPACK(a2,a3,LPC(src3+8),LPC(src4+8)); \
306
                OP(LP(dst+8),no_rnd_PACK(a0,a1,a2,a3)); \
307 308
                UNPACK(a0,a1,AV_RN32(src1+12),LPC(src2+12)); \
                UNPACK(a2,a3,LPC(src3+12),LPC(src4+12)); \
309 310 311 312 313 314 315
                OP(LP(dst+12),no_rnd_PACK(a0,a1,a2,a3)); \
                src1+=src_stride1;\
                src2+=src_stride2;\
                src3+=src_stride3;\
                src4+=src_stride4;\
                dst+=dst_stride;\
        } while(--h); \
316 317 318
} \
\

Michael Niedermayer's avatar
Michael Niedermayer committed
319
#define op_avg(a, b) a = rnd_avg32(a,b)
320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365
#define op_put(a, b) a = b

PIXOP2(avg, op_avg)
PIXOP2(put, op_put)
#undef op_avg
#undef op_put

#define avg2(a,b) ((a+b+1)>>1)
#define avg4(a,b,c,d) ((a+b+c+d+2)>>2)


static void gmc1_c(uint8_t *dst, uint8_t *src, int stride, int h, int x16, int y16, int rounder)
{
    const int A=(16-x16)*(16-y16);
    const int B=(   x16)*(16-y16);
    const int C=(16-x16)*(   y16);
    const int D=(   x16)*(   y16);

    do {
        int t0,t1,t2,t3;
        uint8_t *s0 = src;
        uint8_t *s1 = src+stride;
        t0 = *s0++; t2 = *s1++;
        t1 = *s0++; t3 = *s1++;
        dst[0]= (A*t0 + B*t1 + C*t2 + D*t3 + rounder)>>8;
        t0 = *s0++; t2 = *s1++;
        dst[1]= (A*t1 + B*t0 + C*t3 + D*t2 + rounder)>>8;
        t1 = *s0++; t3 = *s1++;
        dst[2]= (A*t0 + B*t1 + C*t2 + D*t3 + rounder)>>8;
        t0 = *s0++; t2 = *s1++;
        dst[3]= (A*t1 + B*t0 + C*t3 + D*t2 + rounder)>>8;
        t1 = *s0++; t3 = *s1++;
        dst[4]= (A*t0 + B*t1 + C*t2 + D*t3 + rounder)>>8;
        t0 = *s0++; t2 = *s1++;
        dst[5]= (A*t1 + B*t0 + C*t3 + D*t2 + rounder)>>8;
        t1 = *s0++; t3 = *s1++;
        dst[6]= (A*t0 + B*t1 + C*t2 + D*t3 + rounder)>>8;
        t0 = *s0++; t2 = *s1++;
        dst[7]= (A*t1 + B*t0 + C*t3 + D*t2 + rounder)>>8;
        dst+= stride;
        src+= stride;
    }while(--h);
}

#define QPEL_MC(r, OPNAME, RND, OP) \
static void OPNAME ## mpeg4_qpel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
366
    uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392
    do {\
        uint8_t *s = src; \
        int src0,src1,src2,src3,src4,src5,src6,src7,src8;\
        src0= *s++;\
        src1= *s++;\
        src2= *s++;\
        src3= *s++;\
        src4= *s++;\
        OP(dst[0], (src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4));\
        src5= *s++;\
        OP(dst[1], (src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5));\
        src6= *s++;\
        OP(dst[2], (src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6));\
        src7= *s++;\
        OP(dst[3], (src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7));\
        src8= *s++;\
        OP(dst[4], (src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8));\
        OP(dst[5], (src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8));\
        OP(dst[6], (src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7));\
        OP(dst[7], (src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6));\
        dst+=dstStride;\
        src+=srcStride;\
    }while(--h);\
}\
\
static void OPNAME ## mpeg4_qpel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
393
    uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420
    int w=8;\
    do{\
        uint8_t *s = src, *d=dst;\
        int src0,src1,src2,src3,src4,src5,src6,src7,src8;\
        src0 = *s; s+=srcStride; \
        src1 = *s; s+=srcStride; \
        src2 = *s; s+=srcStride; \
        src3 = *s; s+=srcStride; \
        src4 = *s; s+=srcStride; \
        OP(*d, (src0+src1)*20 - (src0+src2)*6 + (src1+src3)*3 - (src2+src4));d+=dstStride;\
        src5 = *s; s+=srcStride; \
        OP(*d, (src1+src2)*20 - (src0+src3)*6 + (src0+src4)*3 - (src1+src5));d+=dstStride;\
        src6 = *s; s+=srcStride; \
        OP(*d, (src2+src3)*20 - (src1+src4)*6 + (src0+src5)*3 - (src0+src6));d+=dstStride;\
        src7 = *s; s+=srcStride; \
        OP(*d, (src3+src4)*20 - (src2+src5)*6 + (src1+src6)*3 - (src0+src7));d+=dstStride;\
        src8 = *s; \
        OP(*d, (src4+src5)*20 - (src3+src6)*6 + (src2+src7)*3 - (src1+src8));d+=dstStride;\
        OP(*d, (src5+src6)*20 - (src4+src7)*6 + (src3+src8)*3 - (src2+src8));d+=dstStride;\
        OP(*d, (src6+src7)*20 - (src5+src8)*6 + (src4+src8)*3 - (src3+src7));d+=dstStride;\
        OP(*d, (src7+src8)*20 - (src6+src8)*6 + (src5+src7)*3 - (src4+src6));\
        dst++;\
        src++;\
    }while(--w);\
}\
\
static void OPNAME ## mpeg4_qpel16_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
421
    uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464
    do {\
        uint8_t *s = src;\
        int src0,src1,src2,src3,src4,src5,src6,src7,src8;\
        int src9,src10,src11,src12,src13,src14,src15,src16;\
        src0= *s++;\
        src1= *s++;\
        src2= *s++;\
        src3= *s++;\
        src4= *s++;\
        OP(dst[ 0], (src0 +src1 )*20 - (src0 +src2 )*6 + (src1 +src3 )*3 - (src2 +src4 ));\
        src5= *s++;\
        OP(dst[ 1], (src1 +src2 )*20 - (src0 +src3 )*6 + (src0 +src4 )*3 - (src1 +src5 ));\
        src6= *s++;\
        OP(dst[ 2], (src2 +src3 )*20 - (src1 +src4 )*6 + (src0 +src5 )*3 - (src0 +src6 ));\
        src7= *s++;\
        OP(dst[ 3], (src3 +src4 )*20 - (src2 +src5 )*6 + (src1 +src6 )*3 - (src0 +src7 ));\
        src8= *s++;\
        OP(dst[ 4], (src4 +src5 )*20 - (src3 +src6 )*6 + (src2 +src7 )*3 - (src1 +src8 ));\
        src9= *s++;\
        OP(dst[ 5], (src5 +src6 )*20 - (src4 +src7 )*6 + (src3 +src8 )*3 - (src2 +src9 ));\
        src10= *s++;\
        OP(dst[ 6], (src6 +src7 )*20 - (src5 +src8 )*6 + (src4 +src9 )*3 - (src3 +src10));\
        src11= *s++;\
        OP(dst[ 7], (src7 +src8 )*20 - (src6 +src9 )*6 + (src5 +src10)*3 - (src4 +src11));\
        src12= *s++;\
        OP(dst[ 8], (src8 +src9 )*20 - (src7 +src10)*6 + (src6 +src11)*3 - (src5 +src12));\
        src13= *s++;\
        OP(dst[ 9], (src9 +src10)*20 - (src8 +src11)*6 + (src7 +src12)*3 - (src6 +src13));\
        src14= *s++;\
        OP(dst[10], (src10+src11)*20 - (src9 +src12)*6 + (src8 +src13)*3 - (src7 +src14));\
        src15= *s++;\
        OP(dst[11], (src11+src12)*20 - (src10+src13)*6 + (src9 +src14)*3 - (src8 +src15));\
        src16= *s++;\
        OP(dst[12], (src12+src13)*20 - (src11+src14)*6 + (src10+src15)*3 - (src9 +src16));\
        OP(dst[13], (src13+src14)*20 - (src12+src15)*6 + (src11+src16)*3 - (src10+src16));\
        OP(dst[14], (src14+src15)*20 - (src13+src16)*6 + (src12+src16)*3 - (src11+src15));\
        OP(dst[15], (src15+src16)*20 - (src14+src16)*6 + (src13+src15)*3 - (src12+src14));\
        dst+=dstStride;\
        src+=srcStride;\
    }while(--h);\
}\
\
static void OPNAME ## mpeg4_qpel16_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
465
    uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;\
466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508
    int w=16;\
    do {\
        uint8_t *s = src, *d=dst;\
        int src0,src1,src2,src3,src4,src5,src6,src7,src8;\
        int src9,src10,src11,src12,src13,src14,src15,src16;\
        src0 = *s; s+=srcStride; \
        src1 = *s; s+=srcStride; \
        src2 = *s; s+=srcStride; \
        src3 = *s; s+=srcStride; \
        src4 = *s; s+=srcStride; \
        OP(*d, (src0 +src1 )*20 - (src0 +src2 )*6 + (src1 +src3 )*3 - (src2 +src4 ));d+=dstStride;\
        src5 = *s; s+=srcStride; \
        OP(*d, (src1 +src2 )*20 - (src0 +src3 )*6 + (src0 +src4 )*3 - (src1 +src5 ));d+=dstStride;\
        src6 = *s; s+=srcStride; \
        OP(*d, (src2 +src3 )*20 - (src1 +src4 )*6 + (src0 +src5 )*3 - (src0 +src6 ));d+=dstStride;\
        src7 = *s; s+=srcStride; \
        OP(*d, (src3 +src4 )*20 - (src2 +src5 )*6 + (src1 +src6 )*3 - (src0 +src7 ));d+=dstStride;\
        src8 = *s; s+=srcStride; \
        OP(*d, (src4 +src5 )*20 - (src3 +src6 )*6 + (src2 +src7 )*3 - (src1 +src8 ));d+=dstStride;\
        src9 = *s; s+=srcStride; \
        OP(*d, (src5 +src6 )*20 - (src4 +src7 )*6 + (src3 +src8 )*3 - (src2 +src9 ));d+=dstStride;\
        src10 = *s; s+=srcStride; \
        OP(*d, (src6 +src7 )*20 - (src5 +src8 )*6 + (src4 +src9 )*3 - (src3 +src10));d+=dstStride;\
        src11 = *s; s+=srcStride; \
        OP(*d, (src7 +src8 )*20 - (src6 +src9 )*6 + (src5 +src10)*3 - (src4 +src11));d+=dstStride;\
        src12 = *s; s+=srcStride; \
        OP(*d, (src8 +src9 )*20 - (src7 +src10)*6 + (src6 +src11)*3 - (src5 +src12));d+=dstStride;\
        src13 = *s; s+=srcStride; \
        OP(*d, (src9 +src10)*20 - (src8 +src11)*6 + (src7 +src12)*3 - (src6 +src13));d+=dstStride;\
        src14 = *s; s+=srcStride; \
        OP(*d, (src10+src11)*20 - (src9 +src12)*6 + (src8 +src13)*3 - (src7 +src14));d+=dstStride;\
        src15 = *s; s+=srcStride; \
        OP(*d, (src11+src12)*20 - (src10+src13)*6 + (src9 +src14)*3 - (src8 +src15));d+=dstStride;\
        src16 = *s; \
        OP(*d, (src12+src13)*20 - (src11+src14)*6 + (src10+src15)*3 - (src9 +src16));d+=dstStride;\
        OP(*d, (src13+src14)*20 - (src12+src15)*6 + (src11+src16)*3 - (src10+src16));d+=dstStride;\
        OP(*d, (src14+src15)*20 - (src13+src16)*6 + (src12+src16)*3 - (src11+src15));d+=dstStride;\
        OP(*d, (src15+src16)*20 - (src14+src16)*6 + (src13+src15)*3 - (src12+src14));\
        dst++;\
        src++;\
    }while(--w);\
}\
\
509
static void OPNAME ## qpel8_mc00_sh4 (uint8_t *dst, uint8_t *src, int stride){\
510 511 512
    OPNAME ## pixels8_c(dst, src, stride, 8);\
}\
\
513
static void OPNAME ## qpel8_mc10_sh4(uint8_t *dst, uint8_t *src, int stride){\
514 515 516 517 518
    uint8_t half[64];\
    put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
    OPNAME ## pixels8_l2_aligned2(dst, src, half, stride, stride, 8, 8);\
}\
\
519
static void OPNAME ## qpel8_mc20_sh4(uint8_t *dst, uint8_t *src, int stride){\
520 521 522
    OPNAME ## mpeg4_qpel8_h_lowpass(dst, src, stride, stride, 8);\
}\
\
523
static void OPNAME ## qpel8_mc30_sh4(uint8_t *dst, uint8_t *src, int stride){\
524 525 526 527 528
    uint8_t half[64];\
    put ## RND ## mpeg4_qpel8_h_lowpass(half, src, 8, stride, 8);\
    OPNAME ## pixels8_l2_aligned2(dst, src+1, half, stride, stride, 8, 8);\
}\
\
529
static void OPNAME ## qpel8_mc01_sh4(uint8_t *dst, uint8_t *src, int stride){\
530 531 532 533 534 535 536
    uint8_t full[16*9];\
    uint8_t half[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
    OPNAME ## pixels8_l2_aligned(dst, full, half, stride, 16, 8, 8);\
}\
\
537
static void OPNAME ## qpel8_mc02_sh4(uint8_t *dst, uint8_t *src, int stride){\
538 539 540 541 542
    uint8_t full[16*9];\
    copy_block9(full, src, 16, stride, 9);\
    OPNAME ## mpeg4_qpel8_v_lowpass(dst, full, stride, 16);\
}\
\
543
static void OPNAME ## qpel8_mc03_sh4(uint8_t *dst, uint8_t *src, int stride){\
544 545 546 547 548 549
    uint8_t full[16*9];\
    uint8_t half[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(half, full, 8, 16);\
    OPNAME ## pixels8_l2_aligned(dst, full+16, half, stride, 16, 8, 8);\
}\
550
static void OPNAME ## qpel8_mc11_sh4(uint8_t *dst, uint8_t *src, int stride){\
551 552 553 554 555 556 557 558 559
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2_aligned(halfH, halfH, full, 8, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2_aligned(dst, halfH, halfHV, stride, 8, 8, 8);\
}\
560
static void OPNAME ## qpel8_mc31_sh4(uint8_t *dst, uint8_t *src, int stride){\
561 562 563 564 565 566 567 568 569
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2_aligned1(halfH, halfH, full+1, 8, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2_aligned(dst, halfH, halfHV, stride, 8, 8, 8);\
}\
570
static void OPNAME ## qpel8_mc13_sh4(uint8_t *dst, uint8_t *src, int stride){\
571 572 573 574 575 576 577 578 579
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2_aligned(halfH, halfH, full, 8, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2_aligned(dst, halfH+8, halfHV, stride, 8, 8, 8);\
}\
580
static void OPNAME ## qpel8_mc33_sh4(uint8_t *dst, uint8_t *src, int stride){\
581 582 583 584 585 586 587 588 589
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2_aligned1(halfH, halfH, full+1, 8, 8, 16, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2_aligned(dst, halfH+8, halfHV, stride, 8, 8, 8);\
}\
590
static void OPNAME ## qpel8_mc21_sh4(uint8_t *dst, uint8_t *src, int stride){\
591 592 593 594 595 596
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2_aligned(dst, halfH, halfHV, stride, 8, 8, 8);\
}\
597
static void OPNAME ## qpel8_mc23_sh4(uint8_t *dst, uint8_t *src, int stride){\
598 599 600 601 602 603
    uint8_t halfH[72];\
    uint8_t halfHV[64];\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
    put ## RND ## mpeg4_qpel8_v_lowpass(halfHV, halfH, 8, 8);\
    OPNAME ## pixels8_l2_aligned(dst, halfH+8, halfHV, stride, 8, 8, 8);\
}\
604
static void OPNAME ## qpel8_mc12_sh4(uint8_t *dst, uint8_t *src, int stride){\
605 606 607 608 609 610 611
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2_aligned(halfH, halfH, full, 8, 8, 16, 9);\
    OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
}\
612
static void OPNAME ## qpel8_mc32_sh4(uint8_t *dst, uint8_t *src, int stride){\
613 614 615 616 617 618 619
    uint8_t full[16*9];\
    uint8_t halfH[72];\
    copy_block9(full, src, 16, stride, 9);\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, full, 8, 16, 9);\
    put ## RND ## pixels8_l2_aligned1(halfH, halfH, full+1, 8, 8, 16, 9);\
    OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
}\
620
static void OPNAME ## qpel8_mc22_sh4(uint8_t *dst, uint8_t *src, int stride){\
621 622 623 624
    uint8_t halfH[72];\
    put ## RND ## mpeg4_qpel8_h_lowpass(halfH, src, 8, stride, 9);\
    OPNAME ## mpeg4_qpel8_v_lowpass(dst, halfH, stride, 8);\
}\
625
static void OPNAME ## qpel16_mc00_sh4 (uint8_t *dst, uint8_t *src, int stride){\
626 627 628
    OPNAME ## pixels16_c(dst, src, stride, 16);\
}\
\
629
static void OPNAME ## qpel16_mc10_sh4(uint8_t *dst, uint8_t *src, int stride){\
630 631 632 633 634
    uint8_t half[256];\
    put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
    OPNAME ## pixels16_l2_aligned2(dst, src, half, stride, stride, 16, 16);\
}\
\
635
static void OPNAME ## qpel16_mc20_sh4(uint8_t *dst, uint8_t *src, int stride){\
636 637 638
    OPNAME ## mpeg4_qpel16_h_lowpass(dst, src, stride, stride, 16);\
}\
\
639
static void OPNAME ## qpel16_mc30_sh4(uint8_t *dst, uint8_t *src, int stride){\
640 641 642 643 644
    uint8_t half[256];\
    put ## RND ## mpeg4_qpel16_h_lowpass(half, src, 16, stride, 16);\
    OPNAME ## pixels16_l2_aligned2(dst, src+1, half, stride, stride, 16, 16);\
}\
\
645
static void OPNAME ## qpel16_mc01_sh4(uint8_t *dst, uint8_t *src, int stride){\
646 647 648 649 650 651 652
    uint8_t full[24*17];\
    uint8_t half[256];\
    copy_block17(full, src, 24, stride, 17);\
    put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24);\
    OPNAME ## pixels16_l2_aligned(dst, full, half, stride, 24, 16, 16);\
}\
\
653
static void OPNAME ## qpel16_mc02_sh4(uint8_t *dst, uint8_t *src, int stride){\
654 655 656 657 658
    uint8_t full[24*17];\
    copy_block17(full, src, 24, stride, 17);\
    OPNAME ## mpeg4_qpel16_v_lowpass(dst, full, stride, 24);\
}\
\
659
static void OPNAME ## qpel16_mc03_sh4(uint8_t *dst, uint8_t *src, int stride){\
660 661 662 663 664 665
    uint8_t full[24*17];\
    uint8_t half[256];\
    copy_block17(full, src, 24, stride, 17);\
    put ## RND ## mpeg4_qpel16_v_lowpass(half, full, 16, 24);\
    OPNAME ## pixels16_l2_aligned(dst, full+24, half, stride, 24, 16, 16);\
}\
666
static void OPNAME ## qpel16_mc11_sh4(uint8_t *dst, uint8_t *src, int stride){\
667 668 669 670 671 672 673 674 675
    uint8_t full[24*17];\
    uint8_t halfH[272];\
    uint8_t halfHV[256];\
    copy_block17(full, src, 24, stride, 17);\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
    put ## RND ## pixels16_l2_aligned(halfH, halfH, full, 16, 16, 24, 17);\
    put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
    OPNAME ## pixels16_l2_aligned(dst, halfH, halfHV, stride, 16, 16, 16);\
}\
676
static void OPNAME ## qpel16_mc31_sh4(uint8_t *dst, uint8_t *src, int stride){\
677 678 679 680 681 682 683 684 685
    uint8_t full[24*17];\
    uint8_t halfH[272];\
    uint8_t halfHV[256];\
    copy_block17(full, src, 24, stride, 17);\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
    put ## RND ## pixels16_l2_aligned1(halfH, halfH, full+1, 16, 16, 24, 17);\
    put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
    OPNAME ## pixels16_l2_aligned(dst, halfH, halfHV, stride, 16, 16, 16);\
}\
686
static void OPNAME ## qpel16_mc13_sh4(uint8_t *dst, uint8_t *src, int stride){\
687 688 689 690 691 692 693 694 695
    uint8_t full[24*17];\
    uint8_t halfH[272];\
    uint8_t halfHV[256];\
    copy_block17(full, src, 24, stride, 17);\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
    put ## RND ## pixels16_l2_aligned(halfH, halfH, full, 16, 16, 24, 17);\
    put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
    OPNAME ## pixels16_l2_aligned(dst, halfH+16, halfHV, stride, 16, 16, 16);\
}\
696
static void OPNAME ## qpel16_mc33_sh4(uint8_t *dst, uint8_t *src, int stride){\
697 698 699 700 701 702 703 704 705
    uint8_t full[24*17];\
    uint8_t halfH[272];\
    uint8_t halfHV[256];\
    copy_block17(full, src, 24, stride, 17);\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
    put ## RND ## pixels16_l2_aligned1(halfH, halfH, full+1, 16, 16, 24, 17);\
    put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
    OPNAME ## pixels16_l2_aligned(dst, halfH+16, halfHV, stride, 16, 16, 16);\
}\
706
static void OPNAME ## qpel16_mc21_sh4(uint8_t *dst, uint8_t *src, int stride){\
707 708 709 710 711 712
    uint8_t halfH[272];\
    uint8_t halfHV[256];\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
    put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
    OPNAME ## pixels16_l2_aligned(dst, halfH, halfHV, stride, 16, 16, 16);\
}\
713
static void OPNAME ## qpel16_mc23_sh4(uint8_t *dst, uint8_t *src, int stride){\
714 715 716 717 718 719
    uint8_t halfH[272];\
    uint8_t halfHV[256];\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
    put ## RND ## mpeg4_qpel16_v_lowpass(halfHV, halfH, 16, 16);\
    OPNAME ## pixels16_l2_aligned(dst, halfH+16, halfHV, stride, 16, 16, 16);\
}\
720
static void OPNAME ## qpel16_mc12_sh4(uint8_t *dst, uint8_t *src, int stride){\
721 722 723 724 725 726 727
    uint8_t full[24*17];\
    uint8_t halfH[272];\
    copy_block17(full, src, 24, stride, 17);\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
    put ## RND ## pixels16_l2_aligned(halfH, halfH, full, 16, 16, 24, 17);\
    OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
}\
728
static void OPNAME ## qpel16_mc32_sh4(uint8_t *dst, uint8_t *src, int stride){\
729 730 731 732 733 734 735
    uint8_t full[24*17];\
    uint8_t halfH[272];\
    copy_block17(full, src, 24, stride, 17);\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, full, 16, 24, 17);\
    put ## RND ## pixels16_l2_aligned1(halfH, halfH, full+1, 16, 16, 24, 17);\
    OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
}\
736
static void OPNAME ## qpel16_mc22_sh4(uint8_t *dst, uint8_t *src, int stride){\
737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756
    uint8_t halfH[272];\
    put ## RND ## mpeg4_qpel16_h_lowpass(halfH, src, 16, stride, 17);\
    OPNAME ## mpeg4_qpel16_v_lowpass(dst, halfH, stride, 16);\
}

#define op_avg(a, b) a = (((a)+cm[((b) + 16)>>5]+1)>>1)
#define op_avg_no_rnd(a, b) a = (((a)+cm[((b) + 15)>>5])>>1)
#define op_put(a, b) a = cm[((b) + 16)>>5]
#define op_put_no_rnd(a, b) a = cm[((b) + 15)>>5]

QPEL_MC(0, put_       , _       , op_put)
QPEL_MC(1, put_no_rnd_, _no_rnd_, op_put_no_rnd)
QPEL_MC(0, avg_       , _       , op_avg)
//QPEL_MC(1, avg_no_rnd , _       , op_avg)
#undef op_avg
#undef op_avg_no_rnd
#undef op_put
#undef op_put_no_rnd

static void wmv2_mspel8_h_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){
757
    uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781

    do{
        int src_1,src0,src1,src2,src3,src4,src5,src6,src7,src8,src9;
        uint8_t *s = src;
        src_1 = s[-1];
        src0 = *s++;
        src1 = *s++;
        src2 = *s++;
        dst[0]= cm[(9*(src0 + src1) - (src_1 + src2) + 8)>>4];
        src3 = *s++;
        dst[1]= cm[(9*(src1 + src2) - (src0 + src3) + 8)>>4];
        src4 = *s++;
        dst[2]= cm[(9*(src2 + src3) - (src1 + src4) + 8)>>4];
        src5 = *s++;
        dst[3]= cm[(9*(src3 + src4) - (src2 + src5) + 8)>>4];
        src6 = *s++;
        dst[4]= cm[(9*(src4 + src5) - (src3 + src6) + 8)>>4];
        src7 = *s++;
        dst[5]= cm[(9*(src5 + src6) - (src4 + src7) + 8)>>4];
        src8 = *s++;
        dst[6]= cm[(9*(src6 + src7) - (src5 + src8) + 8)>>4];
        src9 = *s++;
        dst[7]= cm[(9*(src7 + src8) - (src6 + src9) + 8)>>4];
        dst+=dstStride;
782
        src+=srcStride;
783 784 785 786
    }while(--h);
}

static void wmv2_mspel8_v_lowpass(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int w){
787
    uint8_t *cm = ff_cropTbl + MAX_NEG_CROP;
788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815

    do{
        int src_1,src0,src1,src2,src3,src4,src5,src6,src7,src8,src9;
        uint8_t *s = src,*d = dst;
        src_1 = *(s-srcStride);
        src0 = *s; s+=srcStride;
        src1 = *s; s+=srcStride;
        src2 = *s; s+=srcStride;
        *d= cm[(9*(src0 + src1) - (src_1 + src2) + 8)>>4]; d+=dstStride;
        src3 = *s; s+=srcStride;
        *d= cm[(9*(src1 + src2) - (src0  + src3) + 8)>>4]; d+=dstStride;
        src4 = *s; s+=srcStride;
        *d= cm[(9*(src2 + src3) - (src1  + src4) + 8)>>4]; d+=dstStride;
        src5 = *s; s+=srcStride;
        *d= cm[(9*(src3 + src4) - (src2  + src5) + 8)>>4]; d+=dstStride;
        src6 = *s; s+=srcStride;
        *d= cm[(9*(src4 + src5) - (src3  + src6) + 8)>>4]; d+=dstStride;
        src7 = *s; s+=srcStride;
        *d= cm[(9*(src5 + src6) - (src4  + src7) + 8)>>4]; d+=dstStride;
        src8 = *s; s+=srcStride;
        *d= cm[(9*(src6 + src7) - (src5  + src8) + 8)>>4]; d+=dstStride;
        src9 = *s;
        *d= cm[(9*(src7 + src8) - (src6  + src9) + 8)>>4]; d+=dstStride;
        src++;
        dst++;
    }while(--w);
}

816
static void put_mspel8_mc00_sh4 (uint8_t *dst, uint8_t *src, int stride){
817 818 819
    put_pixels8_c(dst, src, stride, 8);
}

820
static void put_mspel8_mc10_sh4(uint8_t *dst, uint8_t *src, int stride){
821 822 823 824 825
    uint8_t half[64];
    wmv2_mspel8_h_lowpass(half, src, 8, stride, 8);
    put_pixels8_l2_aligned2(dst, src, half, stride, stride, 8, 8);
}

826
static void put_mspel8_mc20_sh4(uint8_t *dst, uint8_t *src, int stride){
827 828 829
    wmv2_mspel8_h_lowpass(dst, src, stride, stride, 8);
}

830
static void put_mspel8_mc30_sh4(uint8_t *dst, uint8_t *src, int stride){
831 832 833 834 835
    uint8_t half[64];
    wmv2_mspel8_h_lowpass(half, src, 8, stride, 8);
    put_pixels8_l2_aligned2(dst, src+1, half, stride, stride, 8, 8);
}

836
static void put_mspel8_mc02_sh4(uint8_t *dst, uint8_t *src, int stride){
837 838 839
    wmv2_mspel8_v_lowpass(dst, src, stride, stride, 8);
}

840
static void put_mspel8_mc12_sh4(uint8_t *dst, uint8_t *src, int stride){
841 842 843 844 845 846 847 848
    uint8_t halfH[88];
    uint8_t halfV[64];
    uint8_t halfHV[64];
    wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
    wmv2_mspel8_v_lowpass(halfV, src, 8, stride, 8);
    wmv2_mspel8_v_lowpass(halfHV, halfH+8, 8, 8, 8);
    put_pixels8_l2_aligned(dst, halfV, halfHV, stride, 8, 8, 8);
}
849
static void put_mspel8_mc32_sh4(uint8_t *dst, uint8_t *src, int stride){
850 851 852 853 854 855 856 857
    uint8_t halfH[88];
    uint8_t halfV[64];
    uint8_t halfHV[64];
    wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
    wmv2_mspel8_v_lowpass(halfV, src+1, 8, stride, 8);
    wmv2_mspel8_v_lowpass(halfHV, halfH+8, 8, 8, 8);
    put_pixels8_l2_aligned(dst, halfV, halfHV, stride, 8, 8, 8);
}
858
static void put_mspel8_mc22_sh4(uint8_t *dst, uint8_t *src, int stride){
859 860 861 862
    uint8_t halfH[88];
    wmv2_mspel8_h_lowpass(halfH, src-stride, 8, stride, 11);
    wmv2_mspel8_v_lowpass(dst, halfH+8, stride, 8, 8);
}