libswscale/ppc/swscale_altivec_template.c

   1 /*
   2  * AltiVec-enhanced yuv2yuvX
   3  *
   4  * Copyright (C) 2004 Romain Dolbeau <romain@dolbeau.org>
   5  * based on the equivalent C code in swscale.c
   6  *
   7  * This file is part of FFmpeg.
   8  *
   9  * FFmpeg is free software; you can redistribute it and/or
  10  * modify it under the terms of the GNU Lesser General Public
  11  * 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  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  17  * Lesser General Public License for more details.
  18  *
  19  * 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  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  22  */
  23
  24 #define vzero vec_splat_s32(0)
  25
  26 static inline void
  27 altivec_packIntArrayToCharArray(int *val, uint8_t* dest, int dstW) {
  28     register int i;
  29     vector unsigned int altivec_vectorShiftInt19 =
  30         vec_add(vec_splat_u32(10), vec_splat_u32(9));
  31     if ((unsigned long)dest % 16) {
  32         /* badly aligned store, we force store alignment */
  33         /* and will handle load misalignment on val w/ vec_perm */
  34         vector unsigned char perm1;
  35         vector signed int v1;
  36         for (i = 0 ; (i < dstW) &&
  37             (((unsigned long)dest + i) % 16) ; i++) {
  38                 int t = val[i] >> 19;
  39                 dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
  40         }
  41         perm1 = vec_lvsl(i << 2, val);
  42         v1 = vec_ld(i << 2, val);
  43         for ( ; i < (dstW - 15); i+=16) {
  44             int offset = i << 2;
  45             vector signed int v2 = vec_ld(offset + 16, val);
  46             vector signed int v3 = vec_ld(offset + 32, val);
  47             vector signed int v4 = vec_ld(offset + 48, val);
  48             vector signed int v5 = vec_ld(offset + 64, val);
  49             vector signed int v12 = vec_perm(v1, v2, perm1);
  50             vector signed int v23 = vec_perm(v2, v3, perm1);
  51             vector signed int v34 = vec_perm(v3, v4, perm1);
  52             vector signed int v45 = vec_perm(v4, v5, perm1);
  53
  54             vector signed int vA = vec_sra(v12, altivec_vectorShiftInt19);
  55             vector signed int vB = vec_sra(v23, altivec_vectorShiftInt19);
  56             vector signed int vC = vec_sra(v34, altivec_vectorShiftInt19);
  57             vector signed int vD = vec_sra(v45, altivec_vectorShiftInt19);
  58             vector unsigned short vs1 = vec_packsu(vA, vB);
  59             vector unsigned short vs2 = vec_packsu(vC, vD);
  60             vector unsigned char vf = vec_packsu(vs1, vs2);
  61             vec_st(vf, i, dest);
  62             v1 = v5;
  63         }
  64     } else { // dest is properly aligned, great
  65         for (i = 0; i < (dstW - 15); i+=16) {
  66             int offset = i << 2;
  67             vector signed int v1 = vec_ld(offset, val);
  68             vector signed int v2 = vec_ld(offset + 16, val);
  69             vector signed int v3 = vec_ld(offset + 32, val);
  70             vector signed int v4 = vec_ld(offset + 48, val);
  71             vector signed int v5 = vec_sra(v1, altivec_vectorShiftInt19);
  72             vector signed int v6 = vec_sra(v2, altivec_vectorShiftInt19);
  73             vector signed int v7 = vec_sra(v3, altivec_vectorShiftInt19);
  74             vector signed int v8 = vec_sra(v4, altivec_vectorShiftInt19);
  75             vector unsigned short vs1 = vec_packsu(v5, v6);
  76             vector unsigned short vs2 = vec_packsu(v7, v8);
  77             vector unsigned char vf = vec_packsu(vs1, vs2);
  78             vec_st(vf, i, dest);
  79         }
  80     }
  81     for ( ; i < dstW ; i++) {
  82         int t = val[i] >> 19;
  83         dest[i] = (t < 0) ? 0 : ((t > 255) ? 255 : t);
  84     }
  85 }
  86
  87 static inline void
  88 yuv2yuvX_altivec_real(const int16_t *lumFilter, int16_t **lumSrc, int lumFilterSize,
  89                       const int16_t *chrFilter, int16_t **chrSrc, int chrFilterSize,
  90                       uint8_t *dest, uint8_t *uDest, uint8_t *vDest, int dstW, int chrDstW)
  91 {
  92     const vector signed int vini = {(1 << 18), (1 << 18), (1 << 18), (1 << 18)};
  93     register int i, j;
  94     {
  95         DECLARE_ALIGNED(16, int, val[dstW]);
  96
  97         for (i = 0; i < (dstW -7); i+=4) {
  98             vec_st(vini, i << 2, val);
  99         }
 100         for (; i < dstW; i++) {
 101             val[i] = (1 << 18);
 102         }
 103
 104         for (j = 0; j < lumFilterSize; j++) {
 105             vector signed short l1, vLumFilter = vec_ld(j << 1, lumFilter);
 106             vector unsigned char perm, perm0 = vec_lvsl(j << 1, lumFilter);
 107             vLumFilter = vec_perm(vLumFilter, vLumFilter, perm0);
 108             vLumFilter = vec_splat(vLumFilter, 0); // lumFilter[j] is loaded 8 times in vLumFilter
 109
 110             perm = vec_lvsl(0, lumSrc[j]);
 111             l1 = vec_ld(0, lumSrc[j]);
 112
 113             for (i = 0; i < (dstW - 7); i+=8) {
 114                 int offset = i << 2;
 115                 vector signed short l2 = vec_ld((i << 1) + 16, lumSrc[j]);
 116
 117                 vector signed int v1 = vec_ld(offset, val);
 118                 vector signed int v2 = vec_ld(offset + 16, val);
 119
 120                 vector signed short ls = vec_perm(l1, l2, perm); // lumSrc[j][i] ... lumSrc[j][i+7]
 121
 122                 vector signed int i1 = vec_mule(vLumFilter, ls);
 123                 vector signed int i2 = vec_mulo(vLumFilter, ls);
 124
 125                 vector signed int vf1 = vec_mergeh(i1, i2);
 126                 vector signed int vf2 = vec_mergel(i1, i2); // lumSrc[j][i] * lumFilter[j] ... lumSrc[j][i+7] * lumFilter[j]
 127
 128                 vector signed int vo1 = vec_add(v1, vf1);
 129                 vector signed int vo2 = vec_add(v2, vf2);
 130
 131                 vec_st(vo1, offset, val);
 132                 vec_st(vo2, offset + 16, val);
 133
 134                 l1 = l2;
 135             }
 136             for ( ; i < dstW; i++) {
 137                 val[i] += lumSrc[j][i] * lumFilter[j];
 138             }
 139         }
 140         altivec_packIntArrayToCharArray(val, dest, dstW);
 141     }
 142     if (uDest != 0) {
 143         DECLARE_ALIGNED(16, int, u[chrDstW]);
 144         DECLARE_ALIGNED(16, int, v[chrDstW]);
 145
 146         for (i = 0; i < (chrDstW -7); i+=4) {
 147             vec_st(vini, i << 2, u);
 148             vec_st(vini, i << 2, v);
 149         }
 150         for (; i < chrDstW; i++) {
 151             u[i] = (1 << 18);
 152             v[i] = (1 << 18);
 153         }
 154
 155         for (j = 0; j < chrFilterSize; j++) {
 156             vector signed short l1, l1_V, vChrFilter = vec_ld(j << 1, chrFilter);
 157             vector unsigned char perm, perm0 = vec_lvsl(j << 1, chrFilter);
 158             vChrFilter = vec_perm(vChrFilter, vChrFilter, perm0);
 159             vChrFilter = vec_splat(vChrFilter, 0); // chrFilter[j] is loaded 8 times in vChrFilter
 160
 161             perm = vec_lvsl(0, chrSrc[j]);
 162             l1 = vec_ld(0, chrSrc[j]);
 163             l1_V = vec_ld(2048 << 1, chrSrc[j]);
 164
 165             for (i = 0; i < (chrDstW - 7); i+=8) {
 166                 int offset = i << 2;
 167                 vector signed short l2 = vec_ld((i << 1) + 16, chrSrc[j]);
 168                 vector signed short l2_V = vec_ld(((i + 2048) << 1) + 16, chrSrc[j]);
 169
 170                 vector signed int v1 = vec_ld(offset, u);
 171                 vector signed int v2 = vec_ld(offset + 16, u);
 172                 vector signed int v1_V = vec_ld(offset, v);
 173                 vector signed int v2_V = vec_ld(offset + 16, v);
 174
 175                 vector signed short ls = vec_perm(l1, l2, perm); // chrSrc[j][i] ... chrSrc[j][i+7]
 176                 vector signed short ls_V = vec_perm(l1_V, l2_V, perm); // chrSrc[j][i+2048] ... chrSrc[j][i+2055]
 177
 178                 vector signed int i1 = vec_mule(vChrFilter, ls);
 179                 vector signed int i2 = vec_mulo(vChrFilter, ls);
 180                 vector signed int i1_V = vec_mule(vChrFilter, ls_V);
 181                 vector signed int i2_V = vec_mulo(vChrFilter, ls_V);
 182
 183                 vector signed int vf1 = vec_mergeh(i1, i2);
 184                 vector signed int vf2 = vec_mergel(i1, i2); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
 185                 vector signed int vf1_V = vec_mergeh(i1_V, i2_V);
 186                 vector signed int vf2_V = vec_mergel(i1_V, i2_V); // chrSrc[j][i] * chrFilter[j] ... chrSrc[j][i+7] * chrFilter[j]
 187
 188                 vector signed int vo1 = vec_add(v1, vf1);
 189                 vector signed int vo2 = vec_add(v2, vf2);
 190                 vector signed int vo1_V = vec_add(v1_V, vf1_V);
 191                 vector signed int vo2_V = vec_add(v2_V, vf2_V);
 192
 193                 vec_st(vo1, offset, u);
 194                 vec_st(vo2, offset + 16, u);
 195                 vec_st(vo1_V, offset, v);
 196                 vec_st(vo2_V, offset + 16, v);
 197
 198                 l1 = l2;
 199                 l1_V = l2_V;
 200             }
 201             for ( ; i < chrDstW; i++) {
 202                 u[i] += chrSrc[j][i] * chrFilter[j];
 203                 v[i] += chrSrc[j][i + 2048] * chrFilter[j];
 204             }
 205         }
 206         altivec_packIntArrayToCharArray(u, uDest, chrDstW);
 207         altivec_packIntArrayToCharArray(v, vDest, chrDstW);
 208     }
 209 }
 210
 211 static inline void hScale_altivec_real(int16_t *dst, int dstW,
 212                                        const uint8_t *src, int srcW,
 213                                        int xInc, const int16_t *filter,
 214                                        const int16_t *filterPos, int filterSize)
 215 {
 216     register int i;
 217     DECLARE_ALIGNED(16, int, tempo[4]);
 218
 219     if (filterSize % 4) {
 220         for (i=0; i<dstW; i++) {
 221             register int j;
 222             register int srcPos = filterPos[i];
 223             register int val = 0;
 224             for (j=0; j<filterSize; j++) {
 225                 val += ((int)src[srcPos + j])*filter[filterSize*i + j];
 226             }
 227             dst[i] = FFMIN(val>>7, (1<<15)-1);
 228         }
 229     }
 230     else
 231     switch (filterSize) {
 232     case 4:
 233     {
 234     for (i=0; i<dstW; i++) {
 235         register int srcPos = filterPos[i];
 236
 237         vector unsigned char src_v0 = vec_ld(srcPos, src);
 238         vector unsigned char src_v1, src_vF;
 239         vector signed short src_v, filter_v;
 240         vector signed int val_vEven, val_s;
 241         if ((((int)src + srcPos)% 16) > 12) {
 242             src_v1 = vec_ld(srcPos + 16, src);
 243         }
 244         src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
 245
 246         src_v = // vec_unpackh sign-extends...
 247             (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
 248         // now put our elements in the even slots
 249         src_v = vec_mergeh(src_v, (vector signed short)vzero);
 250
 251         filter_v = vec_ld(i << 3, filter);
 252         // The 3 above is 2 (filterSize == 4) + 1 (sizeof(short) == 2).
 253
 254         // The neat trick: We only care for half the elements,
 255         // high or low depending on (i<<3)%16 (it's 0 or 8 here),
 256         // and we're going to use vec_mule, so we choose
 257         // carefully how to "unpack" the elements into the even slots.
 258         if ((i << 3) % 16)
 259             filter_v = vec_mergel(filter_v, (vector signed short)vzero);
 260         else
 261             filter_v = vec_mergeh(filter_v, (vector signed short)vzero);
 262
 263         val_vEven = vec_mule(src_v, filter_v);
 264         val_s = vec_sums(val_vEven, vzero);
 265         vec_st(val_s, 0, tempo);
 266         dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
 267     }
 268     }
 269     break;
 270
 271     case 8:
 272     {
 273     for (i=0; i<dstW; i++) {
 274         register int srcPos = filterPos[i];
 275
 276         vector unsigned char src_v0 = vec_ld(srcPos, src);
 277         vector unsigned char src_v1, src_vF;
 278         vector signed short src_v, filter_v;
 279         vector signed int val_v, val_s;
 280         if ((((int)src + srcPos)% 16) > 8) {
 281             src_v1 = vec_ld(srcPos + 16, src);
 282         }
 283         src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
 284
 285         src_v = // vec_unpackh sign-extends...
 286             (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
 287         filter_v = vec_ld(i << 4, filter);
 288         // the 4 above is 3 (filterSize == 8) + 1 (sizeof(short) == 2)
 289
 290         val_v = vec_msums(src_v, filter_v, (vector signed int)vzero);
 291         val_s = vec_sums(val_v, vzero);
 292         vec_st(val_s, 0, tempo);
 293         dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
 294     }
 295     }
 296     break;
 297
 298     case 16:
 299     {
 300         for (i=0; i<dstW; i++) {
 301             register int srcPos = filterPos[i];
 302
 303             vector unsigned char src_v0 = vec_ld(srcPos, src);
 304             vector unsigned char src_v1 = vec_ld(srcPos + 16, src);
 305             vector unsigned char src_vF = vec_perm(src_v0, src_v1, vec_lvsl(srcPos, src));
 306
 307             vector signed short src_vA = // vec_unpackh sign-extends...
 308                 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
 309             vector signed short src_vB = // vec_unpackh sign-extends...
 310                 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
 311
 312             vector signed short filter_v0 = vec_ld(i << 5, filter);
 313             vector signed short filter_v1 = vec_ld((i << 5) + 16, filter);
 314             // the 5 above are 4 (filterSize == 16) + 1 (sizeof(short) == 2)
 315
 316             vector signed int val_acc = vec_msums(src_vA, filter_v0, (vector signed int)vzero);
 317             vector signed int val_v = vec_msums(src_vB, filter_v1, val_acc);
 318
 319             vector signed int val_s = vec_sums(val_v, vzero);
 320
 321             vec_st(val_s, 0, tempo);
 322             dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
 323         }
 324     }
 325     break;
 326
 327     default:
 328     {
 329     for (i=0; i<dstW; i++) {
 330         register int j;
 331         register int srcPos = filterPos[i];
 332
 333         vector signed int val_s, val_v = (vector signed int)vzero;
 334         vector signed short filter_v0R = vec_ld(i * 2 * filterSize, filter);
 335         vector unsigned char permF = vec_lvsl((i * 2 * filterSize), filter);
 336
 337         vector unsigned char src_v0 = vec_ld(srcPos, src);
 338         vector unsigned char permS = vec_lvsl(srcPos, src);
 339
 340         for (j = 0 ; j < filterSize - 15; j += 16) {
 341             vector unsigned char src_v1 = vec_ld(srcPos + j + 16, src);
 342             vector unsigned char src_vF = vec_perm(src_v0, src_v1, permS);
 343
 344             vector signed short src_vA = // vec_unpackh sign-extends...
 345                 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
 346             vector signed short src_vB = // vec_unpackh sign-extends...
 347                 (vector signed short)(vec_mergel((vector unsigned char)vzero, src_vF));
 348
 349             vector signed short filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
 350             vector signed short filter_v2R = vec_ld((i * 2 * filterSize) + (j * 2) + 32, filter);
 351             vector signed short filter_v0  = vec_perm(filter_v0R, filter_v1R, permF);
 352             vector signed short filter_v1  = vec_perm(filter_v1R, filter_v2R, permF);
 353
 354             vector signed int val_acc = vec_msums(src_vA, filter_v0, val_v);
 355             val_v = vec_msums(src_vB, filter_v1, val_acc);
 356
 357             filter_v0R = filter_v2R;
 358             src_v0 = src_v1;
 359         }
 360
 361         if (j < filterSize-7) {
 362             // loading src_v0 is useless, it's already done above
 363             //vector unsigned char src_v0 = vec_ld(srcPos + j, src);
 364             vector unsigned char src_v1, src_vF;
 365             vector signed short src_v, filter_v1R, filter_v;
 366             if ((((int)src + srcPos)% 16) > 8) {
 367                 src_v1 = vec_ld(srcPos + j + 16, src);
 368             }
 369             src_vF = vec_perm(src_v0, src_v1, permS);
 370
 371             src_v = // vec_unpackh sign-extends...
 372                 (vector signed short)(vec_mergeh((vector unsigned char)vzero, src_vF));
 373             // loading filter_v0R is useless, it's already done above
 374             //vector signed short filter_v0R = vec_ld((i * 2 * filterSize) + j, filter);
 375             filter_v1R = vec_ld((i * 2 * filterSize) + (j * 2) + 16, filter);
 376             filter_v = vec_perm(filter_v0R, filter_v1R, permF);
 377
 378             val_v = vec_msums(src_v, filter_v, val_v);
 379         }
 380
 381         val_s = vec_sums(val_v, vzero);
 382
 383         vec_st(val_s, 0, tempo);
 384         dst[i] = FFMIN(tempo[3]>>7, (1<<15)-1);
 385     }
 386
 387     }
 388     }
 389 }
 390
 391 static inline int yv12toyuy2_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
 392                                               int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) {
 393     uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
 394     // yv12toyuy2(src[0], src[1], src[2], dst, c->srcW, srcSliceH, srcStride[0], srcStride[1], dstStride[0]);
 395     uint8_t *ysrc = src[0];
 396     uint8_t *usrc = src[1];
 397     uint8_t *vsrc = src[2];
 398     const int width = c->srcW;
 399     const int height = srcSliceH;
 400     const int lumStride = srcStride[0];
 401     const int chromStride = srcStride[1];
 402     const int dstStride = dstStride_a[0];
 403     const vector unsigned char yperm = vec_lvsl(0, ysrc);
 404     const int vertLumPerChroma = 2;
 405     register unsigned int y;
 406
 407     if (width&15) {
 408         yv12toyuy2(ysrc, usrc, vsrc, dst, c->srcW, srcSliceH, lumStride, chromStride, dstStride);
 409         return srcSliceH;
 410     }
 411
 412     /* This code assumes:
 413
 414     1) dst is 16 bytes-aligned
 415     2) dstStride is a multiple of 16
 416     3) width is a multiple of 16
 417     4) lum & chrom stride are multiples of 8
 418     */
 419
 420     for (y=0; y<height; y++) {
 421         int i;
 422         for (i = 0; i < width - 31; i+= 32) {
 423             const unsigned int j = i >> 1;
 424             vector unsigned char v_yA = vec_ld(i, ysrc);
 425             vector unsigned char v_yB = vec_ld(i + 16, ysrc);
 426             vector unsigned char v_yC = vec_ld(i + 32, ysrc);
 427             vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
 428             vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
 429             vector unsigned char v_uA = vec_ld(j, usrc);
 430             vector unsigned char v_uB = vec_ld(j + 16, usrc);
 431             vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
 432             vector unsigned char v_vA = vec_ld(j, vsrc);
 433             vector unsigned char v_vB = vec_ld(j + 16, vsrc);
 434             vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
 435             vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
 436             vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
 437             vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
 438             vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
 439             vector unsigned char v_yuy2_2 = vec_mergeh(v_y2, v_uv_b);
 440             vector unsigned char v_yuy2_3 = vec_mergel(v_y2, v_uv_b);
 441             vec_st(v_yuy2_0, (i << 1), dst);
 442             vec_st(v_yuy2_1, (i << 1) + 16, dst);
 443             vec_st(v_yuy2_2, (i << 1) + 32, dst);
 444             vec_st(v_yuy2_3, (i << 1) + 48, dst);
 445         }
 446         if (i < width) {
 447             const unsigned int j = i >> 1;
 448             vector unsigned char v_y1 = vec_ld(i, ysrc);
 449             vector unsigned char v_u = vec_ld(j, usrc);
 450             vector unsigned char v_v = vec_ld(j, vsrc);
 451             vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
 452             vector unsigned char v_yuy2_0 = vec_mergeh(v_y1, v_uv_a);
 453             vector unsigned char v_yuy2_1 = vec_mergel(v_y1, v_uv_a);
 454             vec_st(v_yuy2_0, (i << 1), dst);
 455             vec_st(v_yuy2_1, (i << 1) + 16, dst);
 456         }
 457         if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
 458             usrc += chromStride;
 459             vsrc += chromStride;
 460         }
 461         ysrc += lumStride;
 462         dst += dstStride;
 463     }
 464
 465     return srcSliceH;
 466 }
 467
 468 static inline int yv12touyvy_unscaled_altivec(SwsContext *c, uint8_t* src[], int srcStride[], int srcSliceY,
 469                                               int srcSliceH, uint8_t* dstParam[], int dstStride_a[]) {
 470     uint8_t *dst=dstParam[0] + dstStride_a[0]*srcSliceY;
 471     // yv12toyuy2(src[0], src[1], src[2], dst, c->srcW, srcSliceH, srcStride[0], srcStride[1], dstStride[0]);
 472     uint8_t *ysrc = src[0];
 473     uint8_t *usrc = src[1];
 474     uint8_t *vsrc = src[2];
 475     const int width = c->srcW;
 476     const int height = srcSliceH;
 477     const int lumStride = srcStride[0];
 478     const int chromStride = srcStride[1];
 479     const int dstStride = dstStride_a[0];
 480     const int vertLumPerChroma = 2;
 481     const vector unsigned char yperm = vec_lvsl(0, ysrc);
 482     register unsigned int y;
 483
 484     if (width&15) {
 485         yv12touyvy(ysrc, usrc, vsrc, dst, c->srcW, srcSliceH, lumStride, chromStride, dstStride);
 486         return srcSliceH;
 487     }
 488
 489     /* This code assumes:
 490
 491     1) dst is 16 bytes-aligned
 492     2) dstStride is a multiple of 16
 493     3) width is a multiple of 16
 494     4) lum & chrom stride are multiples of 8
 495     */
 496
 497     for (y=0; y<height; y++) {
 498         int i;
 499         for (i = 0; i < width - 31; i+= 32) {
 500             const unsigned int j = i >> 1;
 501             vector unsigned char v_yA = vec_ld(i, ysrc);
 502             vector unsigned char v_yB = vec_ld(i + 16, ysrc);
 503             vector unsigned char v_yC = vec_ld(i + 32, ysrc);
 504             vector unsigned char v_y1 = vec_perm(v_yA, v_yB, yperm);
 505             vector unsigned char v_y2 = vec_perm(v_yB, v_yC, yperm);
 506             vector unsigned char v_uA = vec_ld(j, usrc);
 507             vector unsigned char v_uB = vec_ld(j + 16, usrc);
 508             vector unsigned char v_u = vec_perm(v_uA, v_uB, vec_lvsl(j, usrc));
 509             vector unsigned char v_vA = vec_ld(j, vsrc);
 510             vector unsigned char v_vB = vec_ld(j + 16, vsrc);
 511             vector unsigned char v_v = vec_perm(v_vA, v_vB, vec_lvsl(j, vsrc));
 512             vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
 513             vector unsigned char v_uv_b = vec_mergel(v_u, v_v);
 514             vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
 515             vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
 516             vector unsigned char v_uyvy_2 = vec_mergeh(v_uv_b, v_y2);
 517             vector unsigned char v_uyvy_3 = vec_mergel(v_uv_b, v_y2);
 518             vec_st(v_uyvy_0, (i << 1), dst);
 519             vec_st(v_uyvy_1, (i << 1) + 16, dst);
 520             vec_st(v_uyvy_2, (i << 1) + 32, dst);
 521             vec_st(v_uyvy_3, (i << 1) + 48, dst);
 522         }
 523         if (i < width) {
 524             const unsigned int j = i >> 1;
 525             vector unsigned char v_y1 = vec_ld(i, ysrc);
 526             vector unsigned char v_u = vec_ld(j, usrc);
 527             vector unsigned char v_v = vec_ld(j, vsrc);
 528             vector unsigned char v_uv_a = vec_mergeh(v_u, v_v);
 529             vector unsigned char v_uyvy_0 = vec_mergeh(v_uv_a, v_y1);
 530             vector unsigned char v_uyvy_1 = vec_mergel(v_uv_a, v_y1);
 531             vec_st(v_uyvy_0, (i << 1), dst);
 532             vec_st(v_uyvy_1, (i << 1) + 16, dst);
 533         }
 534         if ((y&(vertLumPerChroma-1)) == vertLumPerChroma-1) {
 535             usrc += chromStride;
 536             vsrc += chromStride;
 537         }
 538         ysrc += lumStride;
 539         dst += dstStride;
 540     }
 541     return srcSliceH;
 542 }