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
  23 #include <stdint.h>
  24 #include <string.h>
  25
  26 #include "libavutil/attributes.h"
  27 #include "libavutil/cpu.h"
  28 #include "libavutil/intreadwrite.h"
  29 #include "libavutil/mem_internal.h"
  30 #include "libavutil/ppc/cpu.h"
  31 #include "libavutil/ppc/util_altivec.h"
  32
  33 #include "libavcodec/h264dec.h"
  34 #include "libavcodec/h264dsp.h"
  35
  36 #if HAVE_ALTIVEC
  37
  38 /****************************************************************************
  39  * IDCT transform:
  40  ****************************************************************************/
  41
  42 #define VEC_1D_DCT(vb0,vb1,vb2,vb3,va0,va1,va2,va3)               \
  43     /* 1st stage */                                               \
  44     vz0 = vec_add(vb0,vb2);       /* temp[0] = Y[0] + Y[2] */     \
  45     vz1 = vec_sub(vb0,vb2);       /* temp[1] = Y[0] - Y[2] */     \
  46     vz2 = vec_sra(vb1,vec_splat_u16(1));                          \
  47     vz2 = vec_sub(vz2,vb3);       /* temp[2] = Y[1].1/2 - Y[3] */ \
  48     vz3 = vec_sra(vb3,vec_splat_u16(1));                          \
  49     vz3 = vec_add(vb1,vz3);       /* temp[3] = Y[1] + Y[3].1/2 */ \
  50     /* 2nd stage: output */                                       \
  51     va0 = vec_add(vz0,vz3);       /* x[0] = temp[0] + temp[3] */  \
  52     va1 = vec_add(vz1,vz2);       /* x[1] = temp[1] + temp[2] */  \
  53     va2 = vec_sub(vz1,vz2);       /* x[2] = temp[1] - temp[2] */  \
  54     va3 = vec_sub(vz0,vz3)        /* x[3] = temp[0] - temp[3] */
  55
  56 #define VEC_TRANSPOSE_4(a0,a1,a2,a3,b0,b1,b2,b3) \
  57     b0 = vec_mergeh( a0, a0 ); \
  58     b1 = vec_mergeh( a1, a0 ); \
  59     b2 = vec_mergeh( a2, a0 ); \
  60     b3 = vec_mergeh( a3, a0 ); \
  61     a0 = vec_mergeh( b0, b2 ); \
  62     a1 = vec_mergel( b0, b2 ); \
  63     a2 = vec_mergeh( b1, b3 ); \
  64     a3 = vec_mergel( b1, b3 ); \
  65     b0 = vec_mergeh( a0, a2 ); \
  66     b1 = vec_mergel( a0, a2 ); \
  67     b2 = vec_mergeh( a1, a3 ); \
  68     b3 = vec_mergel( a1, a3 )
  69
  70 #if HAVE_BIGENDIAN
  71 #define vdst_load(d)              \
  72     vdst_orig = vec_ld(0, dst);   \
  73     vdst = vec_perm(vdst_orig, zero_u8v, vdst_mask);
  74 #else
  75 #define vdst_load(d) vdst = vec_vsx_ld(0, dst)
  76 #endif
  77
  78 #define VEC_LOAD_U8_ADD_S16_STORE_U8(va)                      \
  79     vdst_load();                                              \
  80     vdst_ss = (vec_s16) VEC_MERGEH(zero_u8v, vdst);           \
  81     va = vec_add(va, vdst_ss);                                \
  82     va_u8 = vec_packsu(va, zero_s16v);                        \
  83     va_u32 = vec_splat((vec_u32)va_u8, 0);                  \
  84     vec_ste(va_u32, element, (uint32_t*)dst);
  85
  86 static void h264_idct_add_altivec(uint8_t *dst, int16_t *block, int stride)
  87 {
  88     vec_s16 va0, va1, va2, va3;
  89     vec_s16 vz0, vz1, vz2, vz3;
  90     vec_s16 vtmp0, vtmp1, vtmp2, vtmp3;
  91     vec_u8 va_u8;
  92     vec_u32 va_u32;
  93     vec_s16 vdst_ss;
  94     const vec_u16 v6us = vec_splat_u16(6);
  95     vec_u8 vdst, vdst_orig;
  96     vec_u8 vdst_mask = vec_lvsl(0, dst);
  97     int element = ((unsigned long)dst & 0xf) >> 2;
  98     LOAD_ZERO;
  99
 100     block[0] += 32;  /* add 32 as a DC-level for rounding */
 101
 102     vtmp0 = vec_ld(0,block);
 103     vtmp1 = vec_sld(vtmp0, vtmp0, 8);
 104     vtmp2 = vec_ld(16,block);
 105     vtmp3 = vec_sld(vtmp2, vtmp2, 8);
 106     memset(block, 0, 16 * sizeof(int16_t));
 107
 108     VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
 109     VEC_TRANSPOSE_4(va0,va1,va2,va3,vtmp0,vtmp1,vtmp2,vtmp3);
 110     VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
 111
 112     va0 = vec_sra(va0,v6us);
 113     va1 = vec_sra(va1,v6us);
 114     va2 = vec_sra(va2,v6us);
 115     va3 = vec_sra(va3,v6us);
 116
 117     VEC_LOAD_U8_ADD_S16_STORE_U8(va0);
 118     dst += stride;
 119     VEC_LOAD_U8_ADD_S16_STORE_U8(va1);
 120     dst += stride;
 121     VEC_LOAD_U8_ADD_S16_STORE_U8(va2);
 122     dst += stride;
 123     VEC_LOAD_U8_ADD_S16_STORE_U8(va3);
 124 }
 125
 126 #define IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,  d0, d1, d2, d3, d4, d5, d6, d7) {\
 127     /*        a0  = SRC(0) + SRC(4); */ \
 128     vec_s16 a0v = vec_add(s0, s4);    \
 129     /*        a2  = SRC(0) - SRC(4); */ \
 130     vec_s16 a2v = vec_sub(s0, s4);    \
 131     /*        a4  =           (SRC(2)>>1) - SRC(6); */ \
 132     vec_s16 a4v = vec_sub(vec_sra(s2, onev), s6);    \
 133     /*        a6  =           (SRC(6)>>1) + SRC(2); */ \
 134     vec_s16 a6v = vec_add(vec_sra(s6, onev), s2);    \
 135     /*        b0  =         a0 + a6; */ \
 136     vec_s16 b0v = vec_add(a0v, a6v);  \
 137     /*        b2  =         a2 + a4; */ \
 138     vec_s16 b2v = vec_add(a2v, a4v);  \
 139     /*        b4  =         a2 - a4; */ \
 140     vec_s16 b4v = vec_sub(a2v, a4v);  \
 141     /*        b6  =         a0 - a6; */ \
 142     vec_s16 b6v = vec_sub(a0v, a6v);  \
 143     /* a1 =  SRC(5) - SRC(3) - SRC(7) - (SRC(7)>>1); */ \
 144     /*        a1 =             (SRC(5)-SRC(3)) -  (SRC(7)  +  (SRC(7)>>1)); */ \
 145     vec_s16 a1v = vec_sub( vec_sub(s5, s3), vec_add(s7, vec_sra(s7, onev)) ); \
 146     /* a3 =  SRC(7) + SRC(1) - SRC(3) - (SRC(3)>>1); */ \
 147     /*        a3 =             (SRC(7)+SRC(1)) -  (SRC(3)  +  (SRC(3)>>1)); */ \
 148     vec_s16 a3v = vec_sub( vec_add(s7, s1), vec_add(s3, vec_sra(s3, onev)) );\
 149     /* a5 =  SRC(7) - SRC(1) + SRC(5) + (SRC(5)>>1); */ \
 150     /*        a5 =             (SRC(7)-SRC(1)) +   SRC(5) +   (SRC(5)>>1); */ \
 151     vec_s16 a5v = vec_add( vec_sub(s7, s1), vec_add(s5, vec_sra(s5, onev)) );\
 152     /*        a7 =                SRC(5)+SRC(3) +  SRC(1) +   (SRC(1)>>1); */ \
 153     vec_s16 a7v = vec_add( vec_add(s5, s3), vec_add(s1, vec_sra(s1, onev)) );\
 154     /*        b1 =                  (a7>>2)  +  a1; */ \
 155     vec_s16 b1v = vec_add( vec_sra(a7v, twov), a1v); \
 156     /*        b3 =          a3 +        (a5>>2); */ \
 157     vec_s16 b3v = vec_add(a3v, vec_sra(a5v, twov)); \
 158     /*        b5 =                  (a3>>2)  -   a5; */ \
 159     vec_s16 b5v = vec_sub( vec_sra(a3v, twov), a5v); \
 160     /*        b7 =           a7 -        (a1>>2); */ \
 161     vec_s16 b7v = vec_sub( a7v, vec_sra(a1v, twov)); \
 162     /* DST(0,    b0 + b7); */ \
 163     d0 = vec_add(b0v, b7v); \
 164     /* DST(1,    b2 + b5); */ \
 165     d1 = vec_add(b2v, b5v); \
 166     /* DST(2,    b4 + b3); */ \
 167     d2 = vec_add(b4v, b3v); \
 168     /* DST(3,    b6 + b1); */ \
 169     d3 = vec_add(b6v, b1v); \
 170     /* DST(4,    b6 - b1); */ \
 171     d4 = vec_sub(b6v, b1v); \
 172     /* DST(5,    b4 - b3); */ \
 173     d5 = vec_sub(b4v, b3v); \
 174     /* DST(6,    b2 - b5); */ \
 175     d6 = vec_sub(b2v, b5v); \
 176     /* DST(7,    b0 - b7); */ \
 177     d7 = vec_sub(b0v, b7v); \
 178 }
 179
 180 #if HAVE_BIGENDIAN
 181 #define GET_2PERM(ldv, stv, d)  \
 182     ldv = vec_lvsl(0, d);       \
 183     stv = vec_lvsr(8, d);
 184 #define dstv_load(d)            \
 185     vec_u8 hv = vec_ld( 0, d ); \
 186     vec_u8 lv = vec_ld( 7, d);  \
 187     vec_u8 dstv   = vec_perm( hv, lv, (vec_u8)perm_ldv );
 188 #define dest_unligned_store(d)                                 \
 189     vec_u8 edgehv;                                             \
 190     vec_u8 bodyv  = vec_perm( idstsum8, idstsum8, perm_stv );  \
 191     vec_u8 edgelv = vec_perm( sel, zero_u8v, perm_stv );       \
 192     lv    = vec_sel( lv, bodyv, edgelv );                      \
 193     vec_st( lv, 7, d );                                        \
 194     hv    = vec_ld( 0, d );                                    \
 195     edgehv = vec_perm( zero_u8v, sel, perm_stv );              \
 196     hv    = vec_sel( hv, bodyv, edgehv );                      \
 197     vec_st( hv, 0, d );
 198 #else
 199
 200 #define GET_2PERM(ldv, stv, d) {}
 201 #define dstv_load(d) vec_u8 dstv = vec_vsx_ld(0, d)
 202 #define dest_unligned_store(d)\
 203     vec_u8 dst8 = vec_perm((vec_u8)idstsum8, dstv, vcprm(2,3,s2,s3));\
 204     vec_vsx_st(dst8, 0, d)
 205 #endif /* HAVE_BIGENDIAN */
 206
 207 #define ALTIVEC_STORE_SUM_CLIP(dest, idctv, perm_ldv, perm_stv, sel) { \
 208     /* unaligned load */                                       \
 209     dstv_load(dest);                                           \
 210     vec_s16 idct_sh6 = vec_sra(idctv, sixv);                 \
 211     vec_u16 dst16 = (vec_u16)VEC_MERGEH(zero_u8v, dstv);   \
 212     vec_s16 idstsum = vec_adds(idct_sh6, (vec_s16)dst16);  \
 213     vec_u8 idstsum8 = vec_packsu(zero_s16v, idstsum);        \
 214     /* unaligned store */                                      \
 215     dest_unligned_store(dest);\
 216 }
 217
 218 static void h264_idct8_add_altivec(uint8_t *dst, int16_t *dct, int stride)
 219 {
 220     vec_s16 s0, s1, s2, s3, s4, s5, s6, s7;
 221     vec_s16 d0, d1, d2, d3, d4, d5, d6, d7;
 222     vec_s16 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7;
 223
 224     vec_u8 perm_ldv, perm_stv;
 225     GET_2PERM(perm_ldv, perm_stv, dst);
 226
 227     const vec_u16 onev = vec_splat_u16(1);
 228     const vec_u16 twov = vec_splat_u16(2);
 229     const vec_u16 sixv = vec_splat_u16(6);
 230
 231     const vec_u8 sel = (vec_u8) {0,0,0,0,0,0,0,0,-1,-1,-1,-1,-1,-1,-1,-1};
 232     LOAD_ZERO;
 233
 234     dct[0] += 32; // rounding for the >>6 at the end
 235
 236     s0 = vec_ld(0x00, (int16_t*)dct);
 237     s1 = vec_ld(0x10, (int16_t*)dct);
 238     s2 = vec_ld(0x20, (int16_t*)dct);
 239     s3 = vec_ld(0x30, (int16_t*)dct);
 240     s4 = vec_ld(0x40, (int16_t*)dct);
 241     s5 = vec_ld(0x50, (int16_t*)dct);
 242     s6 = vec_ld(0x60, (int16_t*)dct);
 243     s7 = vec_ld(0x70, (int16_t*)dct);
 244     memset(dct, 0, 64 * sizeof(int16_t));
 245
 246     IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,
 247                      d0, d1, d2, d3, d4, d5, d6, d7);
 248
 249     TRANSPOSE8( d0,  d1,  d2,  d3,  d4,  d5,  d6, d7 );
 250
 251     IDCT8_1D_ALTIVEC(d0,  d1,  d2,  d3,  d4,  d5,  d6, d7,
 252                      idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7);
 253
 254     ALTIVEC_STORE_SUM_CLIP(&dst[0*stride], idct0, perm_ldv, perm_stv, sel);
 255     ALTIVEC_STORE_SUM_CLIP(&dst[1*stride], idct1, perm_ldv, perm_stv, sel);
 256     ALTIVEC_STORE_SUM_CLIP(&dst[2*stride], idct2, perm_ldv, perm_stv, sel);
 257     ALTIVEC_STORE_SUM_CLIP(&dst[3*stride], idct3, perm_ldv, perm_stv, sel);
 258     ALTIVEC_STORE_SUM_CLIP(&dst[4*stride], idct4, perm_ldv, perm_stv, sel);
 259     ALTIVEC_STORE_SUM_CLIP(&dst[5*stride], idct5, perm_ldv, perm_stv, sel);
 260     ALTIVEC_STORE_SUM_CLIP(&dst[6*stride], idct6, perm_ldv, perm_stv, sel);
 261     ALTIVEC_STORE_SUM_CLIP(&dst[7*stride], idct7, perm_ldv, perm_stv, sel);
 262 }
 263
 264 #if HAVE_BIGENDIAN
 265 #define DST_LD vec_ld
 266 #else
 267 #define DST_LD vec_vsx_ld
 268 #endif
 269 static av_always_inline void h264_idct_dc_add_internal(uint8_t *dst, int16_t *block, int stride, int size)
 270 {
 271     vec_s16 dc16;
 272     vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner;
 273     vec_s32 v_dc32;
 274     LOAD_ZERO;
 275     DECLARE_ALIGNED(16, int, dc);
 276     int i;
 277
 278     dc = (block[0] + 32) >> 6;
 279     block[0] = 0;
 280     v_dc32 = vec_lde(0, &dc);
 281     dc16 = VEC_SPLAT16((vec_s16)v_dc32, 1);
 282
 283     if (size == 4)
 284         dc16 = VEC_SLD16(dc16, zero_s16v, 8);
 285     dcplus = vec_packsu(dc16, zero_s16v);
 286     dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v);
 287
 288 #if HAVE_BIGENDIAN
 289     aligner = vec_lvsr(0, dst);
 290     dcplus = vec_perm(dcplus, dcplus, aligner);
 291     dcminus = vec_perm(dcminus, dcminus, aligner);
 292 #endif
 293
 294     for (i = 0; i < size; i += 4) {
 295         v0 = DST_LD(0, dst+0*stride);
 296         v1 = DST_LD(0, dst+1*stride);
 297         v2 = DST_LD(0, dst+2*stride);
 298         v3 = DST_LD(0, dst+3*stride);
 299
 300         v0 = vec_adds(v0, dcplus);
 301         v1 = vec_adds(v1, dcplus);
 302         v2 = vec_adds(v2, dcplus);
 303         v3 = vec_adds(v3, dcplus);
 304
 305         v0 = vec_subs(v0, dcminus);
 306         v1 = vec_subs(v1, dcminus);
 307         v2 = vec_subs(v2, dcminus);
 308         v3 = vec_subs(v3, dcminus);
 309
 310         VEC_ST(v0, 0, dst+0*stride);
 311         VEC_ST(v1, 0, dst+1*stride);
 312         VEC_ST(v2, 0, dst+2*stride);
 313         VEC_ST(v3, 0, dst+3*stride);
 314
 315         dst += 4*stride;
 316     }
 317 }
 318
 319 static void h264_idct_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
 320 {
 321     h264_idct_dc_add_internal(dst, block, stride, 4);
 322 }
 323
 324 static void h264_idct8_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
 325 {
 326     h264_idct_dc_add_internal(dst, block, stride, 8);
 327 }
 328
 329 static void h264_idct_add16_altivec(uint8_t *dst, const int *block_offset,
 330                                     int16_t *block, int stride,
 331                                     const uint8_t nnzc[15 * 8])
 332 {
 333     int i;
 334     for(i=0; i<16; i++){
 335         int nnz = nnzc[ scan8[i] ];
 336         if(nnz){
 337             if(nnz==1 && block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 338             else                      h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
 339         }
 340     }
 341 }
 342
 343 static void h264_idct_add16intra_altivec(uint8_t *dst, const int *block_offset,
 344                                          int16_t *block, int stride,
 345                                          const uint8_t nnzc[15 * 8])
 346 {
 347     int i;
 348     for(i=0; i<16; i++){
 349         if(nnzc[ scan8[i] ]) h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
 350         else if(block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 351     }
 352 }
 353
 354 static void h264_idct8_add4_altivec(uint8_t *dst, const int *block_offset,
 355                                     int16_t *block, int stride,
 356                                     const uint8_t nnzc[15 * 8])
 357 {
 358     int i;
 359     for(i=0; i<16; i+=4){
 360         int nnz = nnzc[ scan8[i] ];
 361         if(nnz){
 362             if(nnz==1 && block[i*16]) h264_idct8_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 363             else                      h264_idct8_add_altivec(dst + block_offset[i], block + i*16, stride);
 364         }
 365     }
 366 }
 367
 368 static void h264_idct_add8_altivec(uint8_t **dest, const int *block_offset,
 369                                    int16_t *block, int stride,
 370                                    const uint8_t nnzc[15 * 8])
 371 {
 372     int i, j;
 373     for (j = 1; j < 3; j++) {
 374         for(i = j * 16; i < j * 16 + 4; i++){
 375             if(nnzc[ scan8[i] ])
 376                 h264_idct_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
 377             else if(block[i*16])
 378                 h264_idct_dc_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
 379         }
 380     }
 381 }
 382
 383 #define transpose4x16(r0, r1, r2, r3) {      \
 384     register vec_u8 r4;                    \
 385     register vec_u8 r5;                    \
 386     register vec_u8 r6;                    \
 387     register vec_u8 r7;                    \
 388                                              \
 389     r4 = vec_mergeh(r0, r2);  /*0, 2 set 0*/ \
 390     r5 = vec_mergel(r0, r2);  /*0, 2 set 1*/ \
 391     r6 = vec_mergeh(r1, r3);  /*1, 3 set 0*/ \
 392     r7 = vec_mergel(r1, r3);  /*1, 3 set 1*/ \
 393                                              \
 394     r0 = vec_mergeh(r4, r6);  /*all set 0*/  \
 395     r1 = vec_mergel(r4, r6);  /*all set 1*/  \
 396     r2 = vec_mergeh(r5, r7);  /*all set 2*/  \
 397     r3 = vec_mergel(r5, r7);  /*all set 3*/  \
 398 }
 399
 400 static inline void write16x4(uint8_t *dst, int dst_stride,
 401                              register vec_u8 r0, register vec_u8 r1,
 402                              register vec_u8 r2, register vec_u8 r3) {
 403     DECLARE_ALIGNED(16, unsigned char, result)[64];
 404     uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
 405     int int_dst_stride = dst_stride/4;
 406
 407     vec_st(r0, 0, result);
 408     vec_st(r1, 16, result);
 409     vec_st(r2, 32, result);
 410     vec_st(r3, 48, result);
 411     /* FIXME: there has to be a better way!!!! */
 412     *dst_int = *src_int;
 413     *(dst_int+   int_dst_stride) = *(src_int + 1);
 414     *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
 415     *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
 416     *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
 417     *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
 418     *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
 419     *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
 420     *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
 421     *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
 422     *(dst_int+10*int_dst_stride) = *(src_int + 10);
 423     *(dst_int+11*int_dst_stride) = *(src_int + 11);
 424     *(dst_int+12*int_dst_stride) = *(src_int + 12);
 425     *(dst_int+13*int_dst_stride) = *(src_int + 13);
 426     *(dst_int+14*int_dst_stride) = *(src_int + 14);
 427     *(dst_int+15*int_dst_stride) = *(src_int + 15);
 428 }
 429
 430 /** @brief performs a 6x16 transpose of data in src, and stores it to dst
 431     @todo FIXME: see if we can't spare some vec_lvsl() by them factorizing
 432     out of unaligned_load() */
 433 #define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\
 434     register vec_u8 r0  = unaligned_load(0,             src);            \
 435     register vec_u8 r1  = unaligned_load(   src_stride, src);            \
 436     register vec_u8 r2  = unaligned_load(2* src_stride, src);            \
 437     register vec_u8 r3  = unaligned_load(3* src_stride, src);            \
 438     register vec_u8 r4  = unaligned_load(4* src_stride, src);            \
 439     register vec_u8 r5  = unaligned_load(5* src_stride, src);            \
 440     register vec_u8 r6  = unaligned_load(6* src_stride, src);            \
 441     register vec_u8 r7  = unaligned_load(7* src_stride, src);            \
 442     register vec_u8 r14 = unaligned_load(14*src_stride, src);            \
 443     register vec_u8 r15 = unaligned_load(15*src_stride, src);            \
 444                                                                            \
 445     r8  = unaligned_load( 8*src_stride, src);                              \
 446     r9  = unaligned_load( 9*src_stride, src);                              \
 447     r10 = unaligned_load(10*src_stride, src);                              \
 448     r11 = unaligned_load(11*src_stride, src);                              \
 449     r12 = unaligned_load(12*src_stride, src);                              \
 450     r13 = unaligned_load(13*src_stride, src);                              \
 451                                                                            \
 452     /*Merge first pairs*/                                                  \
 453     r0 = vec_mergeh(r0, r8);    /*0, 8*/                                   \
 454     r1 = vec_mergeh(r1, r9);    /*1, 9*/                                   \
 455     r2 = vec_mergeh(r2, r10);   /*2,10*/                                   \
 456     r3 = vec_mergeh(r3, r11);   /*3,11*/                                   \
 457     r4 = vec_mergeh(r4, r12);   /*4,12*/                                   \
 458     r5 = vec_mergeh(r5, r13);   /*5,13*/                                   \
 459     r6 = vec_mergeh(r6, r14);   /*6,14*/                                   \
 460     r7 = vec_mergeh(r7, r15);   /*7,15*/                                   \
 461                                                                            \
 462     /*Merge second pairs*/                                                 \
 463     r8  = vec_mergeh(r0, r4);   /*0,4, 8,12 set 0*/                        \
 464     r9  = vec_mergel(r0, r4);   /*0,4, 8,12 set 1*/                        \
 465     r10 = vec_mergeh(r1, r5);   /*1,5, 9,13 set 0*/                        \
 466     r11 = vec_mergel(r1, r5);   /*1,5, 9,13 set 1*/                        \
 467     r12 = vec_mergeh(r2, r6);   /*2,6,10,14 set 0*/                        \
 468     r13 = vec_mergel(r2, r6);   /*2,6,10,14 set 1*/                        \
 469     r14 = vec_mergeh(r3, r7);   /*3,7,11,15 set 0*/                        \
 470     r15 = vec_mergel(r3, r7);   /*3,7,11,15 set 1*/                        \
 471                                                                            \
 472     /*Third merge*/                                                        \
 473     r0 = vec_mergeh(r8,  r12);  /*0,2,4,6,8,10,12,14 set 0*/               \
 474     r1 = vec_mergel(r8,  r12);  /*0,2,4,6,8,10,12,14 set 1*/               \
 475     r2 = vec_mergeh(r9,  r13);  /*0,2,4,6,8,10,12,14 set 2*/               \
 476     r4 = vec_mergeh(r10, r14);  /*1,3,5,7,9,11,13,15 set 0*/               \
 477     r5 = vec_mergel(r10, r14);  /*1,3,5,7,9,11,13,15 set 1*/               \
 478     r6 = vec_mergeh(r11, r15);  /*1,3,5,7,9,11,13,15 set 2*/               \
 479     /* Don't need to compute 3 and 7*/                                     \
 480                                                                            \
 481     /*Final merge*/                                                        \
 482     r8  = vec_mergeh(r0, r4);   /*all set 0*/                              \
 483     r9  = vec_mergel(r0, r4);   /*all set 1*/                              \
 484     r10 = vec_mergeh(r1, r5);   /*all set 2*/                              \
 485     r11 = vec_mergel(r1, r5);   /*all set 3*/                              \
 486     r12 = vec_mergeh(r2, r6);   /*all set 4*/                              \
 487     r13 = vec_mergel(r2, r6);   /*all set 5*/                              \
 488     /* Don't need to compute 14 and 15*/                                   \
 489                                                                            \
 490 }
 491
 492 // out: o = |x-y| < a
 493 static inline vec_u8 diff_lt_altivec ( register vec_u8 x,
 494                                          register vec_u8 y,
 495                                          register vec_u8 a) {
 496
 497     register vec_u8 diff = vec_subs(x, y);
 498     register vec_u8 diffneg = vec_subs(y, x);
 499     register vec_u8 o = vec_or(diff, diffneg); /* |x-y| */
 500     o = (vec_u8)vec_cmplt(o, a);
 501     return o;
 502 }
 503
 504 static inline vec_u8 h264_deblock_mask ( register vec_u8 p0,
 505                                            register vec_u8 p1,
 506                                            register vec_u8 q0,
 507                                            register vec_u8 q1,
 508                                            register vec_u8 alpha,
 509                                            register vec_u8 beta) {
 510
 511     register vec_u8 mask;
 512     register vec_u8 tempmask;
 513
 514     mask = diff_lt_altivec(p0, q0, alpha);
 515     tempmask = diff_lt_altivec(p1, p0, beta);
 516     mask = vec_and(mask, tempmask);
 517     tempmask = diff_lt_altivec(q1, q0, beta);
 518     mask = vec_and(mask, tempmask);
 519
 520     return mask;
 521 }
 522
 523 // out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
 524 static inline vec_u8 h264_deblock_q1(register vec_u8 p0,
 525                                        register vec_u8 p1,
 526                                        register vec_u8 p2,
 527                                        register vec_u8 q0,
 528                                        register vec_u8 tc0) {
 529
 530     register vec_u8 average = vec_avg(p0, q0);
 531     register vec_u8 temp;
 532     register vec_u8 unclipped;
 533     register vec_u8 ones;
 534     register vec_u8 max;
 535     register vec_u8 min;
 536     register vec_u8 newp1;
 537
 538     temp = vec_xor(average, p2);
 539     average = vec_avg(average, p2);     /*avg(p2, avg(p0, q0)) */
 540     ones = vec_splat_u8(1);
 541     temp = vec_and(temp, ones);         /*(p2^avg(p0, q0)) & 1 */
 542     unclipped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
 543     max = vec_adds(p1, tc0);
 544     min = vec_subs(p1, tc0);
 545     newp1 = vec_max(min, unclipped);
 546     newp1 = vec_min(max, newp1);
 547     return newp1;
 548 }
 549
 550 #define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) {                                           \
 551                                                                                                   \
 552     const vec_u8 A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4));                               \
 553                                                                                                   \
 554     register vec_u8 pq0bit = vec_xor(p0,q0);                                                    \
 555     register vec_u8 q1minus;                                                                    \
 556     register vec_u8 p0minus;                                                                    \
 557     register vec_u8 stage1;                                                                     \
 558     register vec_u8 stage2;                                                                     \
 559     register vec_u8 vec160;                                                                     \
 560     register vec_u8 delta;                                                                      \
 561     register vec_u8 deltaneg;                                                                   \
 562                                                                                                   \
 563     q1minus = vec_nor(q1, q1);                 /* 255 - q1 */                                     \
 564     stage1 = vec_avg(p1, q1minus);             /* (p1 - q1 + 256)>>1 */                           \
 565     stage2 = vec_sr(stage1, vec_splat_u8(1));  /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */     \
 566     p0minus = vec_nor(p0, p0);                 /* 255 - p0 */                                     \
 567     stage1 = vec_avg(q0, p0minus);             /* (q0 - p0 + 256)>>1 */                           \
 568     pq0bit = vec_and(pq0bit, vec_splat_u8(1));                                                    \
 569     stage2 = vec_avg(stage2, pq0bit);          /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */ \
 570     stage2 = vec_adds(stage2, stage1);         /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */  \
 571     vec160 = vec_ld(0, &A0v);                                                                     \
 572     deltaneg = vec_subs(vec160, stage2);       /* -d */                                           \
 573     delta = vec_subs(stage2, vec160);          /* d */                                            \
 574     deltaneg = vec_min(tc0masked, deltaneg);                                                      \
 575     delta = vec_min(tc0masked, delta);                                                            \
 576     p0 = vec_subs(p0, deltaneg);                                                                  \
 577     q0 = vec_subs(q0, delta);                                                                     \
 578     p0 = vec_adds(p0, delta);                                                                     \
 579     q0 = vec_adds(q0, deltaneg);                                                                  \
 580 }
 581
 582 #define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) {            \
 583     DECLARE_ALIGNED(16, unsigned char, temp)[16];                                             \
 584     register vec_u8 alphavec;                                                              \
 585     register vec_u8 betavec;                                                               \
 586     register vec_u8 mask;                                                                  \
 587     register vec_u8 p1mask;                                                                \
 588     register vec_u8 q1mask;                                                                \
 589     register vector signed   char tc0vec;                                                    \
 590     register vec_u8 finaltc0;                                                              \
 591     register vec_u8 tc0masked;                                                             \
 592     register vec_u8 newp1;                                                                 \
 593     register vec_u8 newq1;                                                                 \
 594                                                                                              \
 595     temp[0] = alpha;                                                                         \
 596     temp[1] = beta;                                                                          \
 597     alphavec = vec_ld(0, temp);                                                              \
 598     betavec = vec_splat(alphavec, 0x1);                                                      \
 599     alphavec = vec_splat(alphavec, 0x0);                                                     \
 600     mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */            \
 601                                                                                              \
 602     AV_COPY32(temp, tc0);                                                                    \
 603     tc0vec = vec_ld(0, (signed char*)temp);                                                  \
 604     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 605     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 606     mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1)));  /* if tc0[i] >= 0 */         \
 607     finaltc0 = vec_and((vec_u8)tc0vec, mask);     /* tc = tc0 */                           \
 608                                                                                              \
 609     p1mask = diff_lt_altivec(p2, p0, betavec);                                               \
 610     p1mask = vec_and(p1mask, mask);                             /* if ( |p2 - p0| < beta) */ \
 611     tc0masked = vec_and(p1mask, (vec_u8)tc0vec);                                           \
 612     finaltc0 = vec_sub(finaltc0, p1mask);                       /* tc++ */                   \
 613     newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked);                                      \
 614     /*end if*/                                                                               \
 615                                                                                              \
 616     q1mask = diff_lt_altivec(q2, q0, betavec);                                               \
 617     q1mask = vec_and(q1mask, mask);                             /* if ( |q2 - q0| < beta ) */\
 618     tc0masked = vec_and(q1mask, (vec_u8)tc0vec);                                           \
 619     finaltc0 = vec_sub(finaltc0, q1mask);                       /* tc++ */                   \
 620     newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked);                                      \
 621     /*end if*/                                                                               \
 622                                                                                              \
 623     h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0);                                            \
 624     p1 = newp1;                                                                              \
 625     q1 = newq1;                                                                              \
 626 }
 627
 628 static void h264_v_loop_filter_luma_altivec(uint8_t *pix, ptrdiff_t stride, int alpha, int beta, int8_t *tc0) {
 629
 630     if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) {
 631         register vec_u8 p2 = vec_ld(-3*stride, pix);
 632         register vec_u8 p1 = vec_ld(-2*stride, pix);
 633         register vec_u8 p0 = vec_ld(-1*stride, pix);
 634         register vec_u8 q0 = vec_ld(0, pix);
 635         register vec_u8 q1 = vec_ld(stride, pix);
 636         register vec_u8 q2 = vec_ld(2*stride, pix);
 637         h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
 638         vec_st(p1, -2*stride, pix);
 639         vec_st(p0, -1*stride, pix);
 640         vec_st(q0, 0, pix);
 641         vec_st(q1, stride, pix);
 642     }
 643 }
 644
 645 static void h264_h_loop_filter_luma_altivec(uint8_t *pix, ptrdiff_t stride, int alpha, int beta, int8_t *tc0) {
 646
 647     register vec_u8 line0, line1, line2, line3, line4, line5;
 648     if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0)
 649         return;
 650     readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
 651     h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
 652     transpose4x16(line1, line2, line3, line4);
 653     write16x4(pix-2, stride, line1, line2, line3, line4);
 654 }
 655
 656 static av_always_inline
 657 void weight_h264_W_altivec(uint8_t *block, int stride, int height,
 658                            int log2_denom, int weight, int offset, int w)
 659 {
 660     int y, aligned;
 661     vec_u8 vblock;
 662     vec_s16 vtemp, vweight, voffset, v0, v1;
 663     vec_u16 vlog2_denom;
 664     DECLARE_ALIGNED(16, int32_t, temp)[4];
 665     LOAD_ZERO;
 666
 667     offset <<= log2_denom;
 668     if(log2_denom) offset += 1<<(log2_denom-1);
 669     temp[0] = log2_denom;
 670     temp[1] = weight;
 671     temp[2] = offset;
 672
 673     vtemp = (vec_s16)vec_ld(0, temp);
 674 #if !HAVE_BIGENDIAN
 675     vtemp =(vec_s16)vec_perm(vtemp, vtemp, vcswapi2s(0,1,2,3));
 676 #endif
 677     vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
 678     vweight = vec_splat(vtemp, 3);
 679     voffset = vec_splat(vtemp, 5);
 680     aligned = !((unsigned long)block & 0xf);
 681
 682     for (y = 0; y < height; y++) {
 683         vblock = vec_ld(0, block);
 684
 685         v0 = (vec_s16)VEC_MERGEH(zero_u8v, vblock);
 686         v1 = (vec_s16)VEC_MERGEL(zero_u8v, vblock);
 687
 688         if (w == 16 || aligned) {
 689             v0 = vec_mladd(v0, vweight, zero_s16v);
 690             v0 = vec_adds(v0, voffset);
 691             v0 = vec_sra(v0, vlog2_denom);
 692         }
 693         if (w == 16 || !aligned) {
 694             v1 = vec_mladd(v1, vweight, zero_s16v);
 695             v1 = vec_adds(v1, voffset);
 696             v1 = vec_sra(v1, vlog2_denom);
 697         }
 698         vblock = vec_packsu(v0, v1);
 699         vec_st(vblock, 0, block);
 700
 701         block += stride;
 702     }
 703 }
 704
 705 static av_always_inline
 706 void biweight_h264_W_altivec(uint8_t *dst, uint8_t *src, int stride, int height,
 707                              int log2_denom, int weightd, int weights, int offset, int w)
 708 {
 709     int y, dst_aligned, src_aligned;
 710     vec_u8 vsrc, vdst;
 711     vec_s16 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3;
 712     vec_u16 vlog2_denom;
 713     DECLARE_ALIGNED(16, int32_t, temp)[4];
 714     LOAD_ZERO;
 715
 716     offset = ((offset + 1) | 1) << log2_denom;
 717     temp[0] = log2_denom+1;
 718     temp[1] = weights;
 719     temp[2] = weightd;
 720     temp[3] = offset;
 721
 722     vtemp = (vec_s16)vec_ld(0, temp);
 723 #if !HAVE_BIGENDIAN
 724     vtemp =(vec_s16)vec_perm(vtemp, vtemp, vcswapi2s(0,1,2,3));
 725 #endif
 726     vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
 727     vweights = vec_splat(vtemp, 3);
 728     vweightd = vec_splat(vtemp, 5);
 729     voffset = vec_splat(vtemp, 7);
 730     dst_aligned = !((unsigned long)dst & 0xf);
 731     src_aligned = !((unsigned long)src & 0xf);
 732
 733     for (y = 0; y < height; y++) {
 734         vdst = vec_ld(0, dst);
 735         vsrc = vec_ld(0, src);
 736
 737         v0 = (vec_s16)VEC_MERGEH(zero_u8v, vdst);
 738         v1 = (vec_s16)VEC_MERGEL(zero_u8v, vdst);
 739         v2 = (vec_s16)VEC_MERGEH(zero_u8v, vsrc);
 740         v3 = (vec_s16)VEC_MERGEL(zero_u8v, vsrc);
 741
 742         if (w == 8) {
 743             if (src_aligned)
 744                 v3 = v2;
 745             else
 746                 v2 = v3;
 747         }
 748
 749         if (w == 16 || dst_aligned) {
 750             v0 = vec_mladd(v0, vweightd, zero_s16v);
 751             v2 = vec_mladd(v2, vweights, zero_s16v);
 752
 753             v0 = vec_adds(v0, voffset);
 754             v0 = vec_adds(v0, v2);
 755             v0 = vec_sra(v0, vlog2_denom);
 756         }
 757         if (w == 16 || !dst_aligned) {
 758             v1 = vec_mladd(v1, vweightd, zero_s16v);
 759             v3 = vec_mladd(v3, vweights, zero_s16v);
 760
 761             v1 = vec_adds(v1, voffset);
 762             v1 = vec_adds(v1, v3);
 763             v1 = vec_sra(v1, vlog2_denom);
 764         }
 765         vdst = vec_packsu(v0, v1);
 766         vec_st(vdst, 0, dst);
 767
 768         dst += stride;
 769         src += stride;
 770     }
 771 }
 772
 773 #define H264_WEIGHT(W) \
 774 static void weight_h264_pixels ## W ## _altivec(uint8_t *block, ptrdiff_t stride, int height, \
 775                                                 int log2_denom, int weight, int offset) \
 776 { \
 777     weight_h264_W_altivec(block, stride, height, log2_denom, weight, offset, W); \
 778 }\
 779 static void biweight_h264_pixels ## W ## _altivec(uint8_t *dst, uint8_t *src, ptrdiff_t stride, int height, \
 780                                                   int log2_denom, int weightd, int weights, int offset) \
 781 { \
 782     biweight_h264_W_altivec(dst, src, stride, height, log2_denom, weightd, weights, offset, W); \
 783 }
 784
 785 H264_WEIGHT(16)
 786 H264_WEIGHT( 8)
 787 #endif /* HAVE_ALTIVEC */
 788
 789 av_cold void ff_h264dsp_init_ppc(H264DSPContext *c, const int bit_depth,
 790                                  const int chroma_format_idc)
 791 {
 792 #if HAVE_ALTIVEC
 793     if (!PPC_ALTIVEC(av_get_cpu_flags()))
 794         return;
 795
 796     if (bit_depth == 8) {
 797         c->h264_idct_add = h264_idct_add_altivec;
 798         if (chroma_format_idc <= 1)
 799             c->h264_idct_add8 = h264_idct_add8_altivec;
 800         c->h264_idct_add16      = h264_idct_add16_altivec;
 801         c->h264_idct_add16intra = h264_idct_add16intra_altivec;
 802         c->h264_idct_dc_add= h264_idct_dc_add_altivec;
 803         c->h264_idct8_dc_add = h264_idct8_dc_add_altivec;
 804         c->h264_idct8_add    = h264_idct8_add_altivec;
 805         c->h264_idct8_add4   = h264_idct8_add4_altivec;
 806         c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_altivec;
 807         c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_altivec;
 808
 809         c->weight_h264_pixels_tab[0]   = weight_h264_pixels16_altivec;
 810         c->weight_h264_pixels_tab[1]   = weight_h264_pixels8_altivec;
 811         c->biweight_h264_pixels_tab[0] = biweight_h264_pixels16_altivec;
 812         c->biweight_h264_pixels_tab[1] = biweight_h264_pixels8_altivec;
 813     }
 814 #endif /* HAVE_ALTIVEC */
 815 }