git.sesse.net Git - ffmpeg/blob - libavcodec/ppc/h264dsp.c

   1 /*
   2  * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org>
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "config.h"
  22 #include "libavutil/attributes.h"
  23 #include "libavutil/cpu.h"
  24 #include "libavutil/intreadwrite.h"
  25 #include "libavutil/ppc/cpu.h"
  26 #include "libavutil/ppc/types_altivec.h"
  27 #include "libavutil/ppc/util_altivec.h"
  28 #include "libavcodec/h264data.h"
  29 #include "libavcodec/h264dsp.h"
  30
  31 #if HAVE_ALTIVEC
  32
  33 /****************************************************************************
  34  * IDCT transform:
  35  ****************************************************************************/
  36
  37 #define VEC_1D_DCT(vb0,vb1,vb2,vb3,va0,va1,va2,va3)               \
  38     /* 1st stage */                                               \
  39     vz0 = vec_add(vb0,vb2);       /* temp[0] = Y[0] + Y[2] */     \
  40     vz1 = vec_sub(vb0,vb2);       /* temp[1] = Y[0] - Y[2] */     \
  41     vz2 = vec_sra(vb1,vec_splat_u16(1));                          \
  42     vz2 = vec_sub(vz2,vb3);       /* temp[2] = Y[1].1/2 - Y[3] */ \
  43     vz3 = vec_sra(vb3,vec_splat_u16(1));                          \
  44     vz3 = vec_add(vb1,vz3);       /* temp[3] = Y[1] + Y[3].1/2 */ \
  45     /* 2nd stage: output */                                       \
  46     va0 = vec_add(vz0,vz3);       /* x[0] = temp[0] + temp[3] */  \
  47     va1 = vec_add(vz1,vz2);       /* x[1] = temp[1] + temp[2] */  \
  48     va2 = vec_sub(vz1,vz2);       /* x[2] = temp[1] - temp[2] */  \
  49     va3 = vec_sub(vz0,vz3)        /* x[3] = temp[0] - temp[3] */
  50
  51 #define VEC_TRANSPOSE_4(a0,a1,a2,a3,b0,b1,b2,b3) \
  52     b0 = vec_mergeh( a0, a0 ); \
  53     b1 = vec_mergeh( a1, a0 ); \
  54     b2 = vec_mergeh( a2, a0 ); \
  55     b3 = vec_mergeh( a3, a0 ); \
  56     a0 = vec_mergeh( b0, b2 ); \
  57     a1 = vec_mergel( b0, b2 ); \
  58     a2 = vec_mergeh( b1, b3 ); \
  59     a3 = vec_mergel( b1, b3 ); \
  60     b0 = vec_mergeh( a0, a2 ); \
  61     b1 = vec_mergel( a0, a2 ); \
  62     b2 = vec_mergeh( a1, a3 ); \
  63     b3 = vec_mergel( a1, a3 )
  64
  65 #if HAVE_BIGENDIAN
  66 #define vdst_load(d)              \
  67     vdst_orig = vec_ld(0, dst);   \
  68     vdst = vec_perm(vdst_orig, zero_u8v, vdst_mask);
  69 #else
  70 #define vdst_load(d) vdst = vec_vsx_ld(0, dst)
  71 #endif
  72
  73 #define VEC_LOAD_U8_ADD_S16_STORE_U8(va)                      \
  74     vdst_load();                                              \
  75     vdst_ss = (vec_s16) VEC_MERGEH(zero_u8v, vdst);           \
  76     va = vec_add(va, vdst_ss);                                \
  77     va_u8 = vec_packsu(va, zero_s16v);                        \
  78     va_u32 = vec_splat((vec_u32)va_u8, 0);                  \
  79     vec_ste(va_u32, element, (uint32_t*)dst);
  80
  81 static void h264_idct_add_altivec(uint8_t *dst, int16_t *block, int stride)
  82 {
  83     vec_s16 va0, va1, va2, va3;
  84     vec_s16 vz0, vz1, vz2, vz3;
  85     vec_s16 vtmp0, vtmp1, vtmp2, vtmp3;
  86     vec_u8 va_u8;
  87     vec_u32 va_u32;
  88     vec_s16 vdst_ss;
  89     const vec_u16 v6us = vec_splat_u16(6);
  90     vec_u8 vdst, vdst_orig;
  91     vec_u8 vdst_mask = vec_lvsl(0, dst);
  92     int element = ((unsigned long)dst & 0xf) >> 2;
  93     LOAD_ZERO;
  94
  95     block[0] += 32;  /* add 32 as a DC-level for rounding */
  96
  97     vtmp0 = vec_ld(0,block);
  98     vtmp1 = vec_sld(vtmp0, vtmp0, 8);
  99     vtmp2 = vec_ld(16,block);
 100     vtmp3 = vec_sld(vtmp2, vtmp2, 8);
 101     memset(block, 0, 16 * sizeof(int16_t));
 102
 103     VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
 104     VEC_TRANSPOSE_4(va0,va1,va2,va3,vtmp0,vtmp1,vtmp2,vtmp3);
 105     VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
 106
 107     va0 = vec_sra(va0,v6us);
 108     va1 = vec_sra(va1,v6us);
 109     va2 = vec_sra(va2,v6us);
 110     va3 = vec_sra(va3,v6us);
 111
 112     VEC_LOAD_U8_ADD_S16_STORE_U8(va0);
 113     dst += stride;
 114     VEC_LOAD_U8_ADD_S16_STORE_U8(va1);
 115     dst += stride;
 116     VEC_LOAD_U8_ADD_S16_STORE_U8(va2);
 117     dst += stride;
 118     VEC_LOAD_U8_ADD_S16_STORE_U8(va3);
 119 }
 120
 121 #define IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,  d0, d1, d2, d3, d4, d5, d6, d7) {\
 122     /*        a0  = SRC(0) + SRC(4); */ \
 123     vec_s16 a0v = vec_add(s0, s4);    \
 124     /*        a2  = SRC(0) - SRC(4); */ \
 125     vec_s16 a2v = vec_sub(s0, s4);    \
 126     /*        a4  =           (SRC(2)>>1) - SRC(6); */ \
 127     vec_s16 a4v = vec_sub(vec_sra(s2, onev), s6);    \
 128     /*        a6  =           (SRC(6)>>1) + SRC(2); */ \
 129     vec_s16 a6v = vec_add(vec_sra(s6, onev), s2);    \
 130     /*        b0  =         a0 + a6; */ \
 131     vec_s16 b0v = vec_add(a0v, a6v);  \
 132     /*        b2  =         a2 + a4; */ \
 133     vec_s16 b2v = vec_add(a2v, a4v);  \
 134     /*        b4  =         a2 - a4; */ \
 135     vec_s16 b4v = vec_sub(a2v, a4v);  \
 136     /*        b6  =         a0 - a6; */ \
 137     vec_s16 b6v = vec_sub(a0v, a6v);  \
 138     /* a1 =  SRC(5) - SRC(3) - SRC(7) - (SRC(7)>>1); */ \
 139     /*        a1 =             (SRC(5)-SRC(3)) -  (SRC(7)  +  (SRC(7)>>1)); */ \
 140     vec_s16 a1v = vec_sub( vec_sub(s5, s3), vec_add(s7, vec_sra(s7, onev)) ); \
 141     /* a3 =  SRC(7) + SRC(1) - SRC(3) - (SRC(3)>>1); */ \
 142     /*        a3 =             (SRC(7)+SRC(1)) -  (SRC(3)  +  (SRC(3)>>1)); */ \
 143     vec_s16 a3v = vec_sub( vec_add(s7, s1), vec_add(s3, vec_sra(s3, onev)) );\
 144     /* a5 =  SRC(7) - SRC(1) + SRC(5) + (SRC(5)>>1); */ \
 145     /*        a5 =             (SRC(7)-SRC(1)) +   SRC(5) +   (SRC(5)>>1); */ \
 146     vec_s16 a5v = vec_add( vec_sub(s7, s1), vec_add(s5, vec_sra(s5, onev)) );\
 147     /*        a7 =                SRC(5)+SRC(3) +  SRC(1) +   (SRC(1)>>1); */ \
 148     vec_s16 a7v = vec_add( vec_add(s5, s3), vec_add(s1, vec_sra(s1, onev)) );\
 149     /*        b1 =                  (a7>>2)  +  a1; */ \
 150     vec_s16 b1v = vec_add( vec_sra(a7v, twov), a1v); \
 151     /*        b3 =          a3 +        (a5>>2); */ \
 152     vec_s16 b3v = vec_add(a3v, vec_sra(a5v, twov)); \
 153     /*        b5 =                  (a3>>2)  -   a5; */ \
 154     vec_s16 b5v = vec_sub( vec_sra(a3v, twov), a5v); \
 155     /*        b7 =           a7 -        (a1>>2); */ \
 156     vec_s16 b7v = vec_sub( a7v, vec_sra(a1v, twov)); \
 157     /* DST(0,    b0 + b7); */ \
 158     d0 = vec_add(b0v, b7v); \
 159     /* DST(1,    b2 + b5); */ \
 160     d1 = vec_add(b2v, b5v); \
 161     /* DST(2,    b4 + b3); */ \
 162     d2 = vec_add(b4v, b3v); \
 163     /* DST(3,    b6 + b1); */ \
 164     d3 = vec_add(b6v, b1v); \
 165     /* DST(4,    b6 - b1); */ \
 166     d4 = vec_sub(b6v, b1v); \
 167     /* DST(5,    b4 - b3); */ \
 168     d5 = vec_sub(b4v, b3v); \
 169     /* DST(6,    b2 - b5); */ \
 170     d6 = vec_sub(b2v, b5v); \
 171     /* DST(7,    b0 - b7); */ \
 172     d7 = vec_sub(b0v, b7v); \
 173 }
 174
 175 #if HAVE_BIGENDIAN
 176 #define GET_2PERM(ldv, stv, d)  \
 177     ldv = vec_lvsl(0, d);       \
 178     stv = vec_lvsr(8, d);
 179 #define dstv_load(d)            \
 180     vec_u8 hv = vec_ld( 0, d ); \
 181     vec_u8 lv = vec_ld( 7, d);  \
 182     vec_u8 dstv   = vec_perm( hv, lv, (vec_u8)perm_ldv );
 183 #define dest_unligned_store(d)                                 \
 184     vec_u8 edgehv;                                             \
 185     vec_u8 bodyv  = vec_perm( idstsum8, idstsum8, perm_stv );  \
 186     vec_u8 edgelv = vec_perm( sel, zero_u8v, perm_stv );       \
 187     lv    = vec_sel( lv, bodyv, edgelv );                      \
 188     vec_st( lv, 7, d );                                        \
 189     hv    = vec_ld( 0, d );                                    \
 190     edgehv = vec_perm( zero_u8v, sel, perm_stv );              \
 191     hv    = vec_sel( hv, bodyv, edgehv );                      \
 192     vec_st( hv, 0, d );
 193 #else
 194
 195 #define GET_2PERM(ldv, stv, d) {}
 196 #define dstv_load(d) vec_u8 dstv = vec_vsx_ld(0, d)
 197 #define dest_unligned_store(d)\
 198     vec_u8 dst8 = vec_perm((vec_u8)idstsum8, dstv, vcprm(2,3,s2,s3));\
 199     vec_vsx_st(dst8, 0, d)
 200 #endif /* HAVE_BIGENDIAN */
 201
 202 #define ALTIVEC_STORE_SUM_CLIP(dest, idctv, perm_ldv, perm_stv, sel) { \
 203     /* unaligned load */                                       \
 204     dstv_load(dest);                                           \
 205     vec_s16 idct_sh6 = vec_sra(idctv, sixv);                 \
 206     vec_u16 dst16 = (vec_u16)VEC_MERGEH(zero_u8v, dstv);   \
 207     vec_s16 idstsum = vec_adds(idct_sh6, (vec_s16)dst16);  \
 208     vec_u8 idstsum8 = vec_packsu(zero_s16v, idstsum);        \
 209     /* unaligned store */                                      \
 210     dest_unligned_store(dest);\
 211 }
 212
 213 static void h264_idct8_add_altivec(uint8_t *dst, int16_t *dct, int stride)
 214 {
 215     vec_s16 s0, s1, s2, s3, s4, s5, s6, s7;
 216     vec_s16 d0, d1, d2, d3, d4, d5, d6, d7;
 217     vec_s16 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7;
 218
 219     vec_u8 perm_ldv, perm_stv;
 220     GET_2PERM(perm_ldv, perm_stv, dst);
 221
 222     const vec_u16 onev = vec_splat_u16(1);
 223     const vec_u16 twov = vec_splat_u16(2);
 224     const vec_u16 sixv = vec_splat_u16(6);
 225
 226     const vec_u8 sel = (vec_u8) {0,0,0,0,0,0,0,0,-1,-1,-1,-1,-1,-1,-1,-1};
 227     LOAD_ZERO;
 228
 229     dct[0] += 32; // rounding for the >>6 at the end
 230
 231     s0 = vec_ld(0x00, (int16_t*)dct);
 232     s1 = vec_ld(0x10, (int16_t*)dct);
 233     s2 = vec_ld(0x20, (int16_t*)dct);
 234     s3 = vec_ld(0x30, (int16_t*)dct);
 235     s4 = vec_ld(0x40, (int16_t*)dct);
 236     s5 = vec_ld(0x50, (int16_t*)dct);
 237     s6 = vec_ld(0x60, (int16_t*)dct);
 238     s7 = vec_ld(0x70, (int16_t*)dct);
 239     memset(dct, 0, 64 * sizeof(int16_t));
 240
 241     IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,
 242                      d0, d1, d2, d3, d4, d5, d6, d7);
 243
 244     TRANSPOSE8( d0,  d1,  d2,  d3,  d4,  d5,  d6, d7 );
 245
 246     IDCT8_1D_ALTIVEC(d0,  d1,  d2,  d3,  d4,  d5,  d6, d7,
 247                      idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7);
 248
 249     ALTIVEC_STORE_SUM_CLIP(&dst[0*stride], idct0, perm_ldv, perm_stv, sel);
 250     ALTIVEC_STORE_SUM_CLIP(&dst[1*stride], idct1, perm_ldv, perm_stv, sel);
 251     ALTIVEC_STORE_SUM_CLIP(&dst[2*stride], idct2, perm_ldv, perm_stv, sel);
 252     ALTIVEC_STORE_SUM_CLIP(&dst[3*stride], idct3, perm_ldv, perm_stv, sel);
 253     ALTIVEC_STORE_SUM_CLIP(&dst[4*stride], idct4, perm_ldv, perm_stv, sel);
 254     ALTIVEC_STORE_SUM_CLIP(&dst[5*stride], idct5, perm_ldv, perm_stv, sel);
 255     ALTIVEC_STORE_SUM_CLIP(&dst[6*stride], idct6, perm_ldv, perm_stv, sel);
 256     ALTIVEC_STORE_SUM_CLIP(&dst[7*stride], idct7, perm_ldv, perm_stv, sel);
 257 }
 258
 259 #if HAVE_BIGENDIAN
 260 #define DST_LD vec_ld
 261 #else
 262 #define DST_LD vec_vsx_ld
 263 #endif
 264 static av_always_inline void h264_idct_dc_add_internal(uint8_t *dst, int16_t *block, int stride, int size)
 265 {
 266     vec_s16 dc16;
 267     vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner;
 268     vec_s32 v_dc32;
 269     LOAD_ZERO;
 270     DECLARE_ALIGNED(16, int, dc);
 271     int i;
 272
 273     dc = (block[0] + 32) >> 6;
 274     block[0] = 0;
 275     v_dc32 = vec_lde(0, &dc);
 276     dc16 = VEC_SPLAT16((vec_s16)v_dc32, 1);
 277
 278     if (size == 4)
 279         dc16 = VEC_SLD16(dc16, zero_s16v, 8);
 280     dcplus = vec_packsu(dc16, zero_s16v);
 281     dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v);
 282
 283 #if HAVE_BIGENDIAN
 284     aligner = vec_lvsr(0, dst);
 285     dcplus = vec_perm(dcplus, dcplus, aligner);
 286     dcminus = vec_perm(dcminus, dcminus, aligner);
 287 #endif
 288
 289     for (i = 0; i < size; i += 4) {
 290         v0 = DST_LD(0, dst+0*stride);
 291         v1 = DST_LD(0, dst+1*stride);
 292         v2 = DST_LD(0, dst+2*stride);
 293         v3 = DST_LD(0, dst+3*stride);
 294
 295         v0 = vec_adds(v0, dcplus);
 296         v1 = vec_adds(v1, dcplus);
 297         v2 = vec_adds(v2, dcplus);
 298         v3 = vec_adds(v3, dcplus);
 299
 300         v0 = vec_subs(v0, dcminus);
 301         v1 = vec_subs(v1, dcminus);
 302         v2 = vec_subs(v2, dcminus);
 303         v3 = vec_subs(v3, dcminus);
 304
 305         VEC_ST(v0, 0, dst+0*stride);
 306         VEC_ST(v1, 0, dst+1*stride);
 307         VEC_ST(v2, 0, dst+2*stride);
 308         VEC_ST(v3, 0, dst+3*stride);
 309
 310         dst += 4*stride;
 311     }
 312 }
 313
 314 static void h264_idct_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
 315 {
 316     h264_idct_dc_add_internal(dst, block, stride, 4);
 317 }
 318
 319 static void h264_idct8_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
 320 {
 321     h264_idct_dc_add_internal(dst, block, stride, 8);
 322 }
 323
 324 static void h264_idct_add16_altivec(uint8_t *dst, const int *block_offset,
 325                                     int16_t *block, int stride,
 326                                     const uint8_t nnzc[15 * 8])
 327 {
 328     int i;
 329     for(i=0; i<16; i++){
 330         int nnz = nnzc[ scan8[i] ];
 331         if(nnz){
 332             if(nnz==1 && block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 333             else                      h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
 334         }
 335     }
 336 }
 337
 338 static void h264_idct_add16intra_altivec(uint8_t *dst, const int *block_offset,
 339                                          int16_t *block, int stride,
 340                                          const uint8_t nnzc[15 * 8])
 341 {
 342     int i;
 343     for(i=0; i<16; i++){
 344         if(nnzc[ scan8[i] ]) h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
 345         else if(block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 346     }
 347 }
 348
 349 static void h264_idct8_add4_altivec(uint8_t *dst, const int *block_offset,
 350                                     int16_t *block, int stride,
 351                                     const uint8_t nnzc[15 * 8])
 352 {
 353     int i;
 354     for(i=0; i<16; i+=4){
 355         int nnz = nnzc[ scan8[i] ];
 356         if(nnz){
 357             if(nnz==1 && block[i*16]) h264_idct8_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 358             else                      h264_idct8_add_altivec(dst + block_offset[i], block + i*16, stride);
 359         }
 360     }
 361 }
 362
 363 static void h264_idct_add8_altivec(uint8_t **dest, const int *block_offset,
 364                                    int16_t *block, int stride,
 365                                    const uint8_t nnzc[15 * 8])
 366 {
 367     int i, j;
 368     for (j = 1; j < 3; j++) {
 369         for(i = j * 16; i < j * 16 + 4; i++){
 370             if(nnzc[ scan8[i] ])
 371                 h264_idct_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
 372             else if(block[i*16])
 373                 h264_idct_dc_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
 374         }
 375     }
 376 }
 377
 378 #define transpose4x16(r0, r1, r2, r3) {      \
 379     register vec_u8 r4;                    \
 380     register vec_u8 r5;                    \
 381     register vec_u8 r6;                    \
 382     register vec_u8 r7;                    \
 383                                              \
 384     r4 = vec_mergeh(r0, r2);  /*0, 2 set 0*/ \
 385     r5 = vec_mergel(r0, r2);  /*0, 2 set 1*/ \
 386     r6 = vec_mergeh(r1, r3);  /*1, 3 set 0*/ \
 387     r7 = vec_mergel(r1, r3);  /*1, 3 set 1*/ \
 388                                              \
 389     r0 = vec_mergeh(r4, r6);  /*all set 0*/  \
 390     r1 = vec_mergel(r4, r6);  /*all set 1*/  \
 391     r2 = vec_mergeh(r5, r7);  /*all set 2*/  \
 392     r3 = vec_mergel(r5, r7);  /*all set 3*/  \
 393 }
 394
 395 static inline void write16x4(uint8_t *dst, int dst_stride,
 396                              register vec_u8 r0, register vec_u8 r1,
 397                              register vec_u8 r2, register vec_u8 r3) {
 398     DECLARE_ALIGNED(16, unsigned char, result)[64];
 399     uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
 400     int int_dst_stride = dst_stride/4;
 401
 402     vec_st(r0, 0, result);
 403     vec_st(r1, 16, result);
 404     vec_st(r2, 32, result);
 405     vec_st(r3, 48, result);
 406     /* FIXME: there has to be a better way!!!! */
 407     *dst_int = *src_int;
 408     *(dst_int+   int_dst_stride) = *(src_int + 1);
 409     *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
 410     *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
 411     *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
 412     *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
 413     *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
 414     *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
 415     *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
 416     *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
 417     *(dst_int+10*int_dst_stride) = *(src_int + 10);
 418     *(dst_int+11*int_dst_stride) = *(src_int + 11);
 419     *(dst_int+12*int_dst_stride) = *(src_int + 12);
 420     *(dst_int+13*int_dst_stride) = *(src_int + 13);
 421     *(dst_int+14*int_dst_stride) = *(src_int + 14);
 422     *(dst_int+15*int_dst_stride) = *(src_int + 15);
 423 }
 424
 425 /** @brief performs a 6x16 transpose of data in src, and stores it to dst
 426     @todo FIXME: see if we can't spare some vec_lvsl() by them factorizing
 427     out of unaligned_load() */
 428 #define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\
 429     register vec_u8 r0  = unaligned_load(0,             src);            \
 430     register vec_u8 r1  = unaligned_load(   src_stride, src);            \
 431     register vec_u8 r2  = unaligned_load(2* src_stride, src);            \
 432     register vec_u8 r3  = unaligned_load(3* src_stride, src);            \
 433     register vec_u8 r4  = unaligned_load(4* src_stride, src);            \
 434     register vec_u8 r5  = unaligned_load(5* src_stride, src);            \
 435     register vec_u8 r6  = unaligned_load(6* src_stride, src);            \
 436     register vec_u8 r7  = unaligned_load(7* src_stride, src);            \
 437     register vec_u8 r14 = unaligned_load(14*src_stride, src);            \
 438     register vec_u8 r15 = unaligned_load(15*src_stride, src);            \
 439                                                                            \
 440     r8  = unaligned_load( 8*src_stride, src);                              \
 441     r9  = unaligned_load( 9*src_stride, src);                              \
 442     r10 = unaligned_load(10*src_stride, src);                              \
 443     r11 = unaligned_load(11*src_stride, src);                              \
 444     r12 = unaligned_load(12*src_stride, src);                              \
 445     r13 = unaligned_load(13*src_stride, src);                              \
 446                                                                            \
 447     /*Merge first pairs*/                                                  \
 448     r0 = vec_mergeh(r0, r8);    /*0, 8*/                                   \
 449     r1 = vec_mergeh(r1, r9);    /*1, 9*/                                   \
 450     r2 = vec_mergeh(r2, r10);   /*2,10*/                                   \
 451     r3 = vec_mergeh(r3, r11);   /*3,11*/                                   \
 452     r4 = vec_mergeh(r4, r12);   /*4,12*/                                   \
 453     r5 = vec_mergeh(r5, r13);   /*5,13*/                                   \
 454     r6 = vec_mergeh(r6, r14);   /*6,14*/                                   \
 455     r7 = vec_mergeh(r7, r15);   /*7,15*/                                   \
 456                                                                            \
 457     /*Merge second pairs*/                                                 \
 458     r8  = vec_mergeh(r0, r4);   /*0,4, 8,12 set 0*/                        \
 459     r9  = vec_mergel(r0, r4);   /*0,4, 8,12 set 1*/                        \
 460     r10 = vec_mergeh(r1, r5);   /*1,5, 9,13 set 0*/                        \
 461     r11 = vec_mergel(r1, r5);   /*1,5, 9,13 set 1*/                        \
 462     r12 = vec_mergeh(r2, r6);   /*2,6,10,14 set 0*/                        \
 463     r13 = vec_mergel(r2, r6);   /*2,6,10,14 set 1*/                        \
 464     r14 = vec_mergeh(r3, r7);   /*3,7,11,15 set 0*/                        \
 465     r15 = vec_mergel(r3, r7);   /*3,7,11,15 set 1*/                        \
 466                                                                            \
 467     /*Third merge*/                                                        \
 468     r0 = vec_mergeh(r8,  r12);  /*0,2,4,6,8,10,12,14 set 0*/               \
 469     r1 = vec_mergel(r8,  r12);  /*0,2,4,6,8,10,12,14 set 1*/               \
 470     r2 = vec_mergeh(r9,  r13);  /*0,2,4,6,8,10,12,14 set 2*/               \
 471     r4 = vec_mergeh(r10, r14);  /*1,3,5,7,9,11,13,15 set 0*/               \
 472     r5 = vec_mergel(r10, r14);  /*1,3,5,7,9,11,13,15 set 1*/               \
 473     r6 = vec_mergeh(r11, r15);  /*1,3,5,7,9,11,13,15 set 2*/               \
 474     /* Don't need to compute 3 and 7*/                                     \
 475                                                                            \
 476     /*Final merge*/                                                        \
 477     r8  = vec_mergeh(r0, r4);   /*all set 0*/                              \
 478     r9  = vec_mergel(r0, r4);   /*all set 1*/                              \
 479     r10 = vec_mergeh(r1, r5);   /*all set 2*/                              \
 480     r11 = vec_mergel(r1, r5);   /*all set 3*/                              \
 481     r12 = vec_mergeh(r2, r6);   /*all set 4*/                              \
 482     r13 = vec_mergel(r2, r6);   /*all set 5*/                              \
 483     /* Don't need to compute 14 and 15*/                                   \
 484                                                                            \
 485 }
 486
 487 // out: o = |x-y| < a
 488 static inline vec_u8 diff_lt_altivec ( register vec_u8 x,
 489                                          register vec_u8 y,
 490                                          register vec_u8 a) {
 491
 492     register vec_u8 diff = vec_subs(x, y);
 493     register vec_u8 diffneg = vec_subs(y, x);
 494     register vec_u8 o = vec_or(diff, diffneg); /* |x-y| */
 495     o = (vec_u8)vec_cmplt(o, a);
 496     return o;
 497 }
 498
 499 static inline vec_u8 h264_deblock_mask ( register vec_u8 p0,
 500                                            register vec_u8 p1,
 501                                            register vec_u8 q0,
 502                                            register vec_u8 q1,
 503                                            register vec_u8 alpha,
 504                                            register vec_u8 beta) {
 505
 506     register vec_u8 mask;
 507     register vec_u8 tempmask;
 508
 509     mask = diff_lt_altivec(p0, q0, alpha);
 510     tempmask = diff_lt_altivec(p1, p0, beta);
 511     mask = vec_and(mask, tempmask);
 512     tempmask = diff_lt_altivec(q1, q0, beta);
 513     mask = vec_and(mask, tempmask);
 514
 515     return mask;
 516 }
 517
 518 // out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
 519 static inline vec_u8 h264_deblock_q1(register vec_u8 p0,
 520                                        register vec_u8 p1,
 521                                        register vec_u8 p2,
 522                                        register vec_u8 q0,
 523                                        register vec_u8 tc0) {
 524
 525     register vec_u8 average = vec_avg(p0, q0);
 526     register vec_u8 temp;
 527     register vec_u8 uncliped;
 528     register vec_u8 ones;
 529     register vec_u8 max;
 530     register vec_u8 min;
 531     register vec_u8 newp1;
 532
 533     temp = vec_xor(average, p2);
 534     average = vec_avg(average, p2);     /*avg(p2, avg(p0, q0)) */
 535     ones = vec_splat_u8(1);
 536     temp = vec_and(temp, ones);         /*(p2^avg(p0, q0)) & 1 */
 537     uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
 538     max = vec_adds(p1, tc0);
 539     min = vec_subs(p1, tc0);
 540     newp1 = vec_max(min, uncliped);
 541     newp1 = vec_min(max, newp1);
 542     return newp1;
 543 }
 544
 545 #define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) {                                           \
 546                                                                                                   \
 547     const vec_u8 A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4));                               \
 548                                                                                                   \
 549     register vec_u8 pq0bit = vec_xor(p0,q0);                                                    \
 550     register vec_u8 q1minus;                                                                    \
 551     register vec_u8 p0minus;                                                                    \
 552     register vec_u8 stage1;                                                                     \
 553     register vec_u8 stage2;                                                                     \
 554     register vec_u8 vec160;                                                                     \
 555     register vec_u8 delta;                                                                      \
 556     register vec_u8 deltaneg;                                                                   \
 557                                                                                                   \
 558     q1minus = vec_nor(q1, q1);                 /* 255 - q1 */                                     \
 559     stage1 = vec_avg(p1, q1minus);             /* (p1 - q1 + 256)>>1 */                           \
 560     stage2 = vec_sr(stage1, vec_splat_u8(1));  /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */     \
 561     p0minus = vec_nor(p0, p0);                 /* 255 - p0 */                                     \
 562     stage1 = vec_avg(q0, p0minus);             /* (q0 - p0 + 256)>>1 */                           \
 563     pq0bit = vec_and(pq0bit, vec_splat_u8(1));                                                    \
 564     stage2 = vec_avg(stage2, pq0bit);          /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */ \
 565     stage2 = vec_adds(stage2, stage1);         /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */  \
 566     vec160 = vec_ld(0, &A0v);                                                                     \
 567     deltaneg = vec_subs(vec160, stage2);       /* -d */                                           \
 568     delta = vec_subs(stage2, vec160);          /* d */                                            \
 569     deltaneg = vec_min(tc0masked, deltaneg);                                                      \
 570     delta = vec_min(tc0masked, delta);                                                            \
 571     p0 = vec_subs(p0, deltaneg);                                                                  \
 572     q0 = vec_subs(q0, delta);                                                                     \
 573     p0 = vec_adds(p0, delta);                                                                     \
 574     q0 = vec_adds(q0, deltaneg);                                                                  \
 575 }
 576
 577 #define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) {            \
 578     DECLARE_ALIGNED(16, unsigned char, temp)[16];                                             \
 579     register vec_u8 alphavec;                                                              \
 580     register vec_u8 betavec;                                                               \
 581     register vec_u8 mask;                                                                  \
 582     register vec_u8 p1mask;                                                                \
 583     register vec_u8 q1mask;                                                                \
 584     register vector signed   char tc0vec;                                                    \
 585     register vec_u8 finaltc0;                                                              \
 586     register vec_u8 tc0masked;                                                             \
 587     register vec_u8 newp1;                                                                 \
 588     register vec_u8 newq1;                                                                 \
 589                                                                                              \
 590     temp[0] = alpha;                                                                         \
 591     temp[1] = beta;                                                                          \
 592     alphavec = vec_ld(0, temp);                                                              \
 593     betavec = vec_splat(alphavec, 0x1);                                                      \
 594     alphavec = vec_splat(alphavec, 0x0);                                                     \
 595     mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */            \
 596                                                                                              \
 597     AV_COPY32(temp, tc0);                                                                    \
 598     tc0vec = vec_ld(0, (signed char*)temp);                                                  \
 599     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 600     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 601     mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1)));  /* if tc0[i] >= 0 */         \
 602     finaltc0 = vec_and((vec_u8)tc0vec, mask);     /* tc = tc0 */                           \
 603                                                                                              \
 604     p1mask = diff_lt_altivec(p2, p0, betavec);                                               \
 605     p1mask = vec_and(p1mask, mask);                             /* if ( |p2 - p0| < beta) */ \
 606     tc0masked = vec_and(p1mask, (vec_u8)tc0vec);                                           \
 607     finaltc0 = vec_sub(finaltc0, p1mask);                       /* tc++ */                   \
 608     newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked);                                      \
 609     /*end if*/                                                                               \
 610                                                                                              \
 611     q1mask = diff_lt_altivec(q2, q0, betavec);                                               \
 612     q1mask = vec_and(q1mask, mask);                             /* if ( |q2 - q0| < beta ) */\
 613     tc0masked = vec_and(q1mask, (vec_u8)tc0vec);                                           \
 614     finaltc0 = vec_sub(finaltc0, q1mask);                       /* tc++ */                   \
 615     newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked);                                      \
 616     /*end if*/                                                                               \
 617                                                                                              \
 618     h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0);                                            \
 619     p1 = newp1;                                                                              \
 620     q1 = newq1;                                                                              \
 621 }
 622
 623 static void h264_v_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
 624
 625     if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) {
 626         register vec_u8 p2 = vec_ld(-3*stride, pix);
 627         register vec_u8 p1 = vec_ld(-2*stride, pix);
 628         register vec_u8 p0 = vec_ld(-1*stride, pix);
 629         register vec_u8 q0 = vec_ld(0, pix);
 630         register vec_u8 q1 = vec_ld(stride, pix);
 631         register vec_u8 q2 = vec_ld(2*stride, pix);
 632         h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
 633         vec_st(p1, -2*stride, pix);
 634         vec_st(p0, -1*stride, pix);
 635         vec_st(q0, 0, pix);
 636         vec_st(q1, stride, pix);
 637     }
 638 }
 639
 640 static void h264_h_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
 641
 642     register vec_u8 line0, line1, line2, line3, line4, line5;
 643     if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0)
 644         return;
 645     readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
 646     h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
 647     transpose4x16(line1, line2, line3, line4);
 648     write16x4(pix-2, stride, line1, line2, line3, line4);
 649 }
 650
 651 static av_always_inline
 652 void weight_h264_W_altivec(uint8_t *block, int stride, int height,
 653                            int log2_denom, int weight, int offset, int w)
 654 {
 655     int y, aligned;
 656     vec_u8 vblock;
 657     vec_s16 vtemp, vweight, voffset, v0, v1;
 658     vec_u16 vlog2_denom;
 659     DECLARE_ALIGNED(16, int32_t, temp)[4];
 660     LOAD_ZERO;
 661
 662     offset <<= log2_denom;
 663     if(log2_denom) offset += 1<<(log2_denom-1);
 664     temp[0] = log2_denom;
 665     temp[1] = weight;
 666     temp[2] = offset;
 667
 668     vtemp = (vec_s16)vec_ld(0, temp);
 669 #if !HAVE_BIGENDIAN
 670     vtemp =(vec_s16)vec_perm(vtemp, vtemp, vcswapi2s(0,1,2,3));
 671 #endif
 672     vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
 673     vweight = vec_splat(vtemp, 3);
 674     voffset = vec_splat(vtemp, 5);
 675     aligned = !((unsigned long)block & 0xf);
 676
 677     for (y = 0; y < height; y++) {
 678         vblock = vec_ld(0, block);
 679
 680         v0 = (vec_s16)VEC_MERGEH(zero_u8v, vblock);
 681         v1 = (vec_s16)VEC_MERGEL(zero_u8v, vblock);
 682
 683         if (w == 16 || aligned) {
 684             v0 = vec_mladd(v0, vweight, zero_s16v);
 685             v0 = vec_adds(v0, voffset);
 686             v0 = vec_sra(v0, vlog2_denom);
 687         }
 688         if (w == 16 || !aligned) {
 689             v1 = vec_mladd(v1, vweight, zero_s16v);
 690             v1 = vec_adds(v1, voffset);
 691             v1 = vec_sra(v1, vlog2_denom);
 692         }
 693         vblock = vec_packsu(v0, v1);
 694         vec_st(vblock, 0, block);
 695
 696         block += stride;
 697     }
 698 }
 699
 700 static av_always_inline
 701 void biweight_h264_W_altivec(uint8_t *dst, uint8_t *src, int stride, int height,
 702                              int log2_denom, int weightd, int weights, int offset, int w)
 703 {
 704     int y, dst_aligned, src_aligned;
 705     vec_u8 vsrc, vdst;
 706     vec_s16 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3;
 707     vec_u16 vlog2_denom;
 708     DECLARE_ALIGNED(16, int32_t, temp)[4];
 709     LOAD_ZERO;
 710
 711     offset = ((offset + 1) | 1) << log2_denom;
 712     temp[0] = log2_denom+1;
 713     temp[1] = weights;
 714     temp[2] = weightd;
 715     temp[3] = offset;
 716
 717     vtemp = (vec_s16)vec_ld(0, temp);
 718 #if !HAVE_BIGENDIAN
 719     vtemp =(vec_s16)vec_perm(vtemp, vtemp, vcswapi2s(0,1,2,3));
 720 #endif
 721     vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
 722     vweights = vec_splat(vtemp, 3);
 723     vweightd = vec_splat(vtemp, 5);
 724     voffset = vec_splat(vtemp, 7);
 725     dst_aligned = !((unsigned long)dst & 0xf);
 726     src_aligned = !((unsigned long)src & 0xf);
 727
 728     for (y = 0; y < height; y++) {
 729         vdst = vec_ld(0, dst);
 730         vsrc = vec_ld(0, src);
 731
 732         v0 = (vec_s16)VEC_MERGEH(zero_u8v, vdst);
 733         v1 = (vec_s16)VEC_MERGEL(zero_u8v, vdst);
 734         v2 = (vec_s16)VEC_MERGEH(zero_u8v, vsrc);
 735         v3 = (vec_s16)VEC_MERGEL(zero_u8v, vsrc);
 736
 737         if (w == 8) {
 738             if (src_aligned)
 739                 v3 = v2;
 740             else
 741                 v2 = v3;
 742         }
 743
 744         if (w == 16 || dst_aligned) {
 745             v0 = vec_mladd(v0, vweightd, zero_s16v);
 746             v2 = vec_mladd(v2, vweights, zero_s16v);
 747
 748             v0 = vec_adds(v0, voffset);
 749             v0 = vec_adds(v0, v2);
 750             v0 = vec_sra(v0, vlog2_denom);
 751         }
 752         if (w == 16 || !dst_aligned) {
 753             v1 = vec_mladd(v1, vweightd, zero_s16v);
 754             v3 = vec_mladd(v3, vweights, zero_s16v);
 755
 756             v1 = vec_adds(v1, voffset);
 757             v1 = vec_adds(v1, v3);
 758             v1 = vec_sra(v1, vlog2_denom);
 759         }
 760         vdst = vec_packsu(v0, v1);
 761         vec_st(vdst, 0, dst);
 762
 763         dst += stride;
 764         src += stride;
 765     }
 766 }
 767
 768 #define H264_WEIGHT(W) \
 769 static void weight_h264_pixels ## W ## _altivec(uint8_t *block, int stride, int height, \
 770                                                 int log2_denom, int weight, int offset) \
 771 { \
 772     weight_h264_W_altivec(block, stride, height, log2_denom, weight, offset, W); \
 773 }\
 774 static void biweight_h264_pixels ## W ## _altivec(uint8_t *dst, uint8_t *src, int stride, int height, \
 775                                                   int log2_denom, int weightd, int weights, int offset) \
 776 { \
 777     biweight_h264_W_altivec(dst, src, stride, height, log2_denom, weightd, weights, offset, W); \
 778 }
 779
 780 H264_WEIGHT(16)
 781 H264_WEIGHT( 8)
 782 #endif /* HAVE_ALTIVEC */
 783
 784 av_cold void ff_h264dsp_init_ppc(H264DSPContext *c, const int bit_depth,
 785                                  const int chroma_format_idc)
 786 {
 787 #if HAVE_ALTIVEC
 788     if (!PPC_ALTIVEC(av_get_cpu_flags()))
 789         return;
 790
 791     if (bit_depth == 8) {
 792         c->h264_idct_add = h264_idct_add_altivec;
 793         if (chroma_format_idc <= 1)
 794             c->h264_idct_add8 = h264_idct_add8_altivec;
 795         c->h264_idct_add16      = h264_idct_add16_altivec;
 796         c->h264_idct_add16intra = h264_idct_add16intra_altivec;
 797         c->h264_idct_dc_add= h264_idct_dc_add_altivec;
 798         c->h264_idct8_dc_add = h264_idct8_dc_add_altivec;
 799         c->h264_idct8_add    = h264_idct8_add_altivec;
 800         c->h264_idct8_add4   = h264_idct8_add4_altivec;
 801         c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_altivec;
 802         c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_altivec;
 803
 804         c->weight_h264_pixels_tab[0]   = weight_h264_pixels16_altivec;
 805         c->weight_h264_pixels_tab[1]   = weight_h264_pixels8_altivec;
 806         c->biweight_h264_pixels_tab[0] = biweight_h264_pixels16_altivec;
 807         c->biweight_h264_pixels_tab[1] = biweight_h264_pixels8_altivec;
 808     }
 809 #endif /* HAVE_ALTIVEC */
 810 }