git.sesse.net Git - ffmpeg/blob - libavcodec/ppc/h264_altivec.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 "libavutil/cpu.h"
  22 #include "libavutil/intreadwrite.h"
  23 #include "libavutil/ppc/types_altivec.h"
  24 #include "libavutil/ppc/util_altivec.h"
  25 #include "libavcodec/h264data.h"
  26 #include "libavcodec/h264dsp.h"
  27
  28 /****************************************************************************
  29  * IDCT transform:
  30  ****************************************************************************/
  31
  32 #define VEC_1D_DCT(vb0,vb1,vb2,vb3,va0,va1,va2,va3)               \
  33     /* 1st stage */                                               \
  34     vz0 = vec_add(vb0,vb2);       /* temp[0] = Y[0] + Y[2] */     \
  35     vz1 = vec_sub(vb0,vb2);       /* temp[1] = Y[0] - Y[2] */     \
  36     vz2 = vec_sra(vb1,vec_splat_u16(1));                          \
  37     vz2 = vec_sub(vz2,vb3);       /* temp[2] = Y[1].1/2 - Y[3] */ \
  38     vz3 = vec_sra(vb3,vec_splat_u16(1));                          \
  39     vz3 = vec_add(vb1,vz3);       /* temp[3] = Y[1] + Y[3].1/2 */ \
  40     /* 2nd stage: output */                                       \
  41     va0 = vec_add(vz0,vz3);       /* x[0] = temp[0] + temp[3] */  \
  42     va1 = vec_add(vz1,vz2);       /* x[1] = temp[1] + temp[2] */  \
  43     va2 = vec_sub(vz1,vz2);       /* x[2] = temp[1] - temp[2] */  \
  44     va3 = vec_sub(vz0,vz3)        /* x[3] = temp[0] - temp[3] */
  45
  46 #define VEC_TRANSPOSE_4(a0,a1,a2,a3,b0,b1,b2,b3) \
  47     b0 = vec_mergeh( a0, a0 ); \
  48     b1 = vec_mergeh( a1, a0 ); \
  49     b2 = vec_mergeh( a2, a0 ); \
  50     b3 = vec_mergeh( a3, a0 ); \
  51     a0 = vec_mergeh( b0, b2 ); \
  52     a1 = vec_mergel( b0, b2 ); \
  53     a2 = vec_mergeh( b1, b3 ); \
  54     a3 = vec_mergel( b1, b3 ); \
  55     b0 = vec_mergeh( a0, a2 ); \
  56     b1 = vec_mergel( a0, a2 ); \
  57     b2 = vec_mergeh( a1, a3 ); \
  58     b3 = vec_mergel( a1, a3 )
  59
  60 #define VEC_LOAD_U8_ADD_S16_STORE_U8(va)                      \
  61     vdst_orig = vec_ld(0, dst);                               \
  62     vdst = vec_perm(vdst_orig, zero_u8v, vdst_mask);          \
  63     vdst_ss = (vec_s16) vec_mergeh(zero_u8v, vdst);         \
  64     va = vec_add(va, vdst_ss);                                \
  65     va_u8 = vec_packsu(va, zero_s16v);                        \
  66     va_u32 = vec_splat((vec_u32)va_u8, 0);                  \
  67     vec_ste(va_u32, element, (uint32_t*)dst);
  68
  69 static void ff_h264_idct_add_altivec(uint8_t *dst, int16_t *block, int stride)
  70 {
  71     vec_s16 va0, va1, va2, va3;
  72     vec_s16 vz0, vz1, vz2, vz3;
  73     vec_s16 vtmp0, vtmp1, vtmp2, vtmp3;
  74     vec_u8 va_u8;
  75     vec_u32 va_u32;
  76     vec_s16 vdst_ss;
  77     const vec_u16 v6us = vec_splat_u16(6);
  78     vec_u8 vdst, vdst_orig;
  79     vec_u8 vdst_mask = vec_lvsl(0, dst);
  80     int element = ((unsigned long)dst & 0xf) >> 2;
  81     LOAD_ZERO;
  82
  83     block[0] += 32;  /* add 32 as a DC-level for rounding */
  84
  85     vtmp0 = vec_ld(0,block);
  86     vtmp1 = vec_sld(vtmp0, vtmp0, 8);
  87     vtmp2 = vec_ld(16,block);
  88     vtmp3 = vec_sld(vtmp2, vtmp2, 8);
  89
  90     VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
  91     VEC_TRANSPOSE_4(va0,va1,va2,va3,vtmp0,vtmp1,vtmp2,vtmp3);
  92     VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3);
  93
  94     va0 = vec_sra(va0,v6us);
  95     va1 = vec_sra(va1,v6us);
  96     va2 = vec_sra(va2,v6us);
  97     va3 = vec_sra(va3,v6us);
  98
  99     VEC_LOAD_U8_ADD_S16_STORE_U8(va0);
 100     dst += stride;
 101     VEC_LOAD_U8_ADD_S16_STORE_U8(va1);
 102     dst += stride;
 103     VEC_LOAD_U8_ADD_S16_STORE_U8(va2);
 104     dst += stride;
 105     VEC_LOAD_U8_ADD_S16_STORE_U8(va3);
 106 }
 107
 108 #define IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,  d0, d1, d2, d3, d4, d5, d6, d7) {\
 109     /*        a0  = SRC(0) + SRC(4); */ \
 110     vec_s16 a0v = vec_add(s0, s4);    \
 111     /*        a2  = SRC(0) - SRC(4); */ \
 112     vec_s16 a2v = vec_sub(s0, s4);    \
 113     /*        a4  =           (SRC(2)>>1) - SRC(6); */ \
 114     vec_s16 a4v = vec_sub(vec_sra(s2, onev), s6);    \
 115     /*        a6  =           (SRC(6)>>1) + SRC(2); */ \
 116     vec_s16 a6v = vec_add(vec_sra(s6, onev), s2);    \
 117     /*        b0  =         a0 + a6; */ \
 118     vec_s16 b0v = vec_add(a0v, a6v);  \
 119     /*        b2  =         a2 + a4; */ \
 120     vec_s16 b2v = vec_add(a2v, a4v);  \
 121     /*        b4  =         a2 - a4; */ \
 122     vec_s16 b4v = vec_sub(a2v, a4v);  \
 123     /*        b6  =         a0 - a6; */ \
 124     vec_s16 b6v = vec_sub(a0v, a6v);  \
 125     /* a1 =  SRC(5) - SRC(3) - SRC(7) - (SRC(7)>>1); */ \
 126     /*        a1 =             (SRC(5)-SRC(3)) -  (SRC(7)  +  (SRC(7)>>1)); */ \
 127     vec_s16 a1v = vec_sub( vec_sub(s5, s3), vec_add(s7, vec_sra(s7, onev)) ); \
 128     /* a3 =  SRC(7) + SRC(1) - SRC(3) - (SRC(3)>>1); */ \
 129     /*        a3 =             (SRC(7)+SRC(1)) -  (SRC(3)  +  (SRC(3)>>1)); */ \
 130     vec_s16 a3v = vec_sub( vec_add(s7, s1), vec_add(s3, vec_sra(s3, onev)) );\
 131     /* a5 =  SRC(7) - SRC(1) + SRC(5) + (SRC(5)>>1); */ \
 132     /*        a5 =             (SRC(7)-SRC(1)) +   SRC(5) +   (SRC(5)>>1); */ \
 133     vec_s16 a5v = vec_add( vec_sub(s7, s1), vec_add(s5, vec_sra(s5, onev)) );\
 134     /*        a7 =                SRC(5)+SRC(3) +  SRC(1) +   (SRC(1)>>1); */ \
 135     vec_s16 a7v = vec_add( vec_add(s5, s3), vec_add(s1, vec_sra(s1, onev)) );\
 136     /*        b1 =                  (a7>>2)  +  a1; */ \
 137     vec_s16 b1v = vec_add( vec_sra(a7v, twov), a1v); \
 138     /*        b3 =          a3 +        (a5>>2); */ \
 139     vec_s16 b3v = vec_add(a3v, vec_sra(a5v, twov)); \
 140     /*        b5 =                  (a3>>2)  -   a5; */ \
 141     vec_s16 b5v = vec_sub( vec_sra(a3v, twov), a5v); \
 142     /*        b7 =           a7 -        (a1>>2); */ \
 143     vec_s16 b7v = vec_sub( a7v, vec_sra(a1v, twov)); \
 144     /* DST(0,    b0 + b7); */ \
 145     d0 = vec_add(b0v, b7v); \
 146     /* DST(1,    b2 + b5); */ \
 147     d1 = vec_add(b2v, b5v); \
 148     /* DST(2,    b4 + b3); */ \
 149     d2 = vec_add(b4v, b3v); \
 150     /* DST(3,    b6 + b1); */ \
 151     d3 = vec_add(b6v, b1v); \
 152     /* DST(4,    b6 - b1); */ \
 153     d4 = vec_sub(b6v, b1v); \
 154     /* DST(5,    b4 - b3); */ \
 155     d5 = vec_sub(b4v, b3v); \
 156     /* DST(6,    b2 - b5); */ \
 157     d6 = vec_sub(b2v, b5v); \
 158     /* DST(7,    b0 - b7); */ \
 159     d7 = vec_sub(b0v, b7v); \
 160 }
 161
 162 #define ALTIVEC_STORE_SUM_CLIP(dest, idctv, perm_ldv, perm_stv, sel) { \
 163     /* unaligned load */                                       \
 164     vec_u8 hv = vec_ld( 0, dest );                           \
 165     vec_u8 lv = vec_ld( 7, dest );                           \
 166     vec_u8 dstv   = vec_perm( hv, lv, (vec_u8)perm_ldv );  \
 167     vec_s16 idct_sh6 = vec_sra(idctv, sixv);                 \
 168     vec_u16 dst16 = (vec_u16)vec_mergeh(zero_u8v, dstv);   \
 169     vec_s16 idstsum = vec_adds(idct_sh6, (vec_s16)dst16);  \
 170     vec_u8 idstsum8 = vec_packsu(zero_s16v, idstsum);        \
 171     vec_u8 edgehv;                                           \
 172     /* unaligned store */                                      \
 173     vec_u8 bodyv  = vec_perm( idstsum8, idstsum8, perm_stv );\
 174     vec_u8 edgelv = vec_perm( sel, zero_u8v, perm_stv );     \
 175     lv    = vec_sel( lv, bodyv, edgelv );                      \
 176     vec_st( lv, 7, dest );                                     \
 177     hv    = vec_ld( 0, dest );                                 \
 178     edgehv = vec_perm( zero_u8v, sel, perm_stv );              \
 179     hv    = vec_sel( hv, bodyv, edgehv );                      \
 180     vec_st( hv, 0, dest );                                     \
 181  }
 182
 183 static void ff_h264_idct8_add_altivec( uint8_t *dst, int16_t *dct, int stride ) {
 184     vec_s16 s0, s1, s2, s3, s4, s5, s6, s7;
 185     vec_s16 d0, d1, d2, d3, d4, d5, d6, d7;
 186     vec_s16 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7;
 187
 188     vec_u8 perm_ldv = vec_lvsl(0, dst);
 189     vec_u8 perm_stv = vec_lvsr(8, dst);
 190
 191     const vec_u16 onev = vec_splat_u16(1);
 192     const vec_u16 twov = vec_splat_u16(2);
 193     const vec_u16 sixv = vec_splat_u16(6);
 194
 195     const vec_u8 sel = (vec_u8) {0,0,0,0,0,0,0,0,-1,-1,-1,-1,-1,-1,-1,-1};
 196     LOAD_ZERO;
 197
 198     dct[0] += 32; // rounding for the >>6 at the end
 199
 200     s0 = vec_ld(0x00, (int16_t*)dct);
 201     s1 = vec_ld(0x10, (int16_t*)dct);
 202     s2 = vec_ld(0x20, (int16_t*)dct);
 203     s3 = vec_ld(0x30, (int16_t*)dct);
 204     s4 = vec_ld(0x40, (int16_t*)dct);
 205     s5 = vec_ld(0x50, (int16_t*)dct);
 206     s6 = vec_ld(0x60, (int16_t*)dct);
 207     s7 = vec_ld(0x70, (int16_t*)dct);
 208
 209     IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7,
 210                      d0, d1, d2, d3, d4, d5, d6, d7);
 211
 212     TRANSPOSE8( d0,  d1,  d2,  d3,  d4,  d5,  d6, d7 );
 213
 214     IDCT8_1D_ALTIVEC(d0,  d1,  d2,  d3,  d4,  d5,  d6, d7,
 215                      idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7);
 216
 217     ALTIVEC_STORE_SUM_CLIP(&dst[0*stride], idct0, perm_ldv, perm_stv, sel);
 218     ALTIVEC_STORE_SUM_CLIP(&dst[1*stride], idct1, perm_ldv, perm_stv, sel);
 219     ALTIVEC_STORE_SUM_CLIP(&dst[2*stride], idct2, perm_ldv, perm_stv, sel);
 220     ALTIVEC_STORE_SUM_CLIP(&dst[3*stride], idct3, perm_ldv, perm_stv, sel);
 221     ALTIVEC_STORE_SUM_CLIP(&dst[4*stride], idct4, perm_ldv, perm_stv, sel);
 222     ALTIVEC_STORE_SUM_CLIP(&dst[5*stride], idct5, perm_ldv, perm_stv, sel);
 223     ALTIVEC_STORE_SUM_CLIP(&dst[6*stride], idct6, perm_ldv, perm_stv, sel);
 224     ALTIVEC_STORE_SUM_CLIP(&dst[7*stride], idct7, perm_ldv, perm_stv, sel);
 225 }
 226
 227 static av_always_inline void h264_idct_dc_add_internal(uint8_t *dst, int16_t *block, int stride, int size)
 228 {
 229     vec_s16 dc16;
 230     vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner;
 231     LOAD_ZERO;
 232     DECLARE_ALIGNED(16, int, dc);
 233     int i;
 234
 235     dc = (block[0] + 32) >> 6;
 236     dc16 = vec_splat((vec_s16) vec_lde(0, &dc), 1);
 237
 238     if (size == 4)
 239         dc16 = vec_sld(dc16, zero_s16v, 8);
 240     dcplus = vec_packsu(dc16, zero_s16v);
 241     dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v);
 242
 243     aligner = vec_lvsr(0, dst);
 244     dcplus = vec_perm(dcplus, dcplus, aligner);
 245     dcminus = vec_perm(dcminus, dcminus, aligner);
 246
 247     for (i = 0; i < size; i += 4) {
 248         v0 = vec_ld(0, dst+0*stride);
 249         v1 = vec_ld(0, dst+1*stride);
 250         v2 = vec_ld(0, dst+2*stride);
 251         v3 = vec_ld(0, dst+3*stride);
 252
 253         v0 = vec_adds(v0, dcplus);
 254         v1 = vec_adds(v1, dcplus);
 255         v2 = vec_adds(v2, dcplus);
 256         v3 = vec_adds(v3, dcplus);
 257
 258         v0 = vec_subs(v0, dcminus);
 259         v1 = vec_subs(v1, dcminus);
 260         v2 = vec_subs(v2, dcminus);
 261         v3 = vec_subs(v3, dcminus);
 262
 263         vec_st(v0, 0, dst+0*stride);
 264         vec_st(v1, 0, dst+1*stride);
 265         vec_st(v2, 0, dst+2*stride);
 266         vec_st(v3, 0, dst+3*stride);
 267
 268         dst += 4*stride;
 269     }
 270 }
 271
 272 static void h264_idct_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
 273 {
 274     h264_idct_dc_add_internal(dst, block, stride, 4);
 275 }
 276
 277 static void ff_h264_idct8_dc_add_altivec(uint8_t *dst, int16_t *block, int stride)
 278 {
 279     h264_idct_dc_add_internal(dst, block, stride, 8);
 280 }
 281
 282 static void ff_h264_idct_add16_altivec(uint8_t *dst, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]){
 283     int i;
 284     for(i=0; i<16; i++){
 285         int nnz = nnzc[ scan8[i] ];
 286         if(nnz){
 287             if(nnz==1 && block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 288             else                      ff_h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
 289         }
 290     }
 291 }
 292
 293 static void ff_h264_idct_add16intra_altivec(uint8_t *dst, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]){
 294     int i;
 295     for(i=0; i<16; i++){
 296         if(nnzc[ scan8[i] ]) ff_h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride);
 297         else if(block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 298     }
 299 }
 300
 301 static void ff_h264_idct8_add4_altivec(uint8_t *dst, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]){
 302     int i;
 303     for(i=0; i<16; i+=4){
 304         int nnz = nnzc[ scan8[i] ];
 305         if(nnz){
 306             if(nnz==1 && block[i*16]) ff_h264_idct8_dc_add_altivec(dst + block_offset[i], block + i*16, stride);
 307             else                      ff_h264_idct8_add_altivec   (dst + block_offset[i], block + i*16, stride);
 308         }
 309     }
 310 }
 311
 312 static void ff_h264_idct_add8_altivec(uint8_t **dest, const int *block_offset, int16_t *block, int stride, const uint8_t nnzc[15*8]){
 313     int i, j;
 314     for (j = 1; j < 3; j++) {
 315         for(i = j * 16; i < j * 16 + 4; i++){
 316             if(nnzc[ scan8[i] ])
 317                 ff_h264_idct_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
 318             else if(block[i*16])
 319                 h264_idct_dc_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride);
 320         }
 321     }
 322 }
 323
 324 #define transpose4x16(r0, r1, r2, r3) {      \
 325     register vec_u8 r4;                    \
 326     register vec_u8 r5;                    \
 327     register vec_u8 r6;                    \
 328     register vec_u8 r7;                    \
 329                                              \
 330     r4 = vec_mergeh(r0, r2);  /*0, 2 set 0*/ \
 331     r5 = vec_mergel(r0, r2);  /*0, 2 set 1*/ \
 332     r6 = vec_mergeh(r1, r3);  /*1, 3 set 0*/ \
 333     r7 = vec_mergel(r1, r3);  /*1, 3 set 1*/ \
 334                                              \
 335     r0 = vec_mergeh(r4, r6);  /*all set 0*/  \
 336     r1 = vec_mergel(r4, r6);  /*all set 1*/  \
 337     r2 = vec_mergeh(r5, r7);  /*all set 2*/  \
 338     r3 = vec_mergel(r5, r7);  /*all set 3*/  \
 339 }
 340
 341 static inline void write16x4(uint8_t *dst, int dst_stride,
 342                              register vec_u8 r0, register vec_u8 r1,
 343                              register vec_u8 r2, register vec_u8 r3) {
 344     DECLARE_ALIGNED(16, unsigned char, result)[64];
 345     uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
 346     int int_dst_stride = dst_stride/4;
 347
 348     vec_st(r0, 0, result);
 349     vec_st(r1, 16, result);
 350     vec_st(r2, 32, result);
 351     vec_st(r3, 48, result);
 352     /* FIXME: there has to be a better way!!!! */
 353     *dst_int = *src_int;
 354     *(dst_int+   int_dst_stride) = *(src_int + 1);
 355     *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
 356     *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
 357     *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
 358     *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
 359     *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
 360     *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
 361     *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
 362     *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
 363     *(dst_int+10*int_dst_stride) = *(src_int + 10);
 364     *(dst_int+11*int_dst_stride) = *(src_int + 11);
 365     *(dst_int+12*int_dst_stride) = *(src_int + 12);
 366     *(dst_int+13*int_dst_stride) = *(src_int + 13);
 367     *(dst_int+14*int_dst_stride) = *(src_int + 14);
 368     *(dst_int+15*int_dst_stride) = *(src_int + 15);
 369 }
 370
 371 /** @brief performs a 6x16 transpose of data in src, and stores it to dst
 372     @todo FIXME: see if we can't spare some vec_lvsl() by them factorizing
 373     out of unaligned_load() */
 374 #define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\
 375     register vec_u8 r0  = unaligned_load(0,             src);            \
 376     register vec_u8 r1  = unaligned_load(   src_stride, src);            \
 377     register vec_u8 r2  = unaligned_load(2* src_stride, src);            \
 378     register vec_u8 r3  = unaligned_load(3* src_stride, src);            \
 379     register vec_u8 r4  = unaligned_load(4* src_stride, src);            \
 380     register vec_u8 r5  = unaligned_load(5* src_stride, src);            \
 381     register vec_u8 r6  = unaligned_load(6* src_stride, src);            \
 382     register vec_u8 r7  = unaligned_load(7* src_stride, src);            \
 383     register vec_u8 r14 = unaligned_load(14*src_stride, src);            \
 384     register vec_u8 r15 = unaligned_load(15*src_stride, src);            \
 385                                                                            \
 386     r8  = unaligned_load( 8*src_stride, src);                              \
 387     r9  = unaligned_load( 9*src_stride, src);                              \
 388     r10 = unaligned_load(10*src_stride, src);                              \
 389     r11 = unaligned_load(11*src_stride, src);                              \
 390     r12 = unaligned_load(12*src_stride, src);                              \
 391     r13 = unaligned_load(13*src_stride, src);                              \
 392                                                                            \
 393     /*Merge first pairs*/                                                  \
 394     r0 = vec_mergeh(r0, r8);    /*0, 8*/                                   \
 395     r1 = vec_mergeh(r1, r9);    /*1, 9*/                                   \
 396     r2 = vec_mergeh(r2, r10);   /*2,10*/                                   \
 397     r3 = vec_mergeh(r3, r11);   /*3,11*/                                   \
 398     r4 = vec_mergeh(r4, r12);   /*4,12*/                                   \
 399     r5 = vec_mergeh(r5, r13);   /*5,13*/                                   \
 400     r6 = vec_mergeh(r6, r14);   /*6,14*/                                   \
 401     r7 = vec_mergeh(r7, r15);   /*7,15*/                                   \
 402                                                                            \
 403     /*Merge second pairs*/                                                 \
 404     r8  = vec_mergeh(r0, r4);   /*0,4, 8,12 set 0*/                        \
 405     r9  = vec_mergel(r0, r4);   /*0,4, 8,12 set 1*/                        \
 406     r10 = vec_mergeh(r1, r5);   /*1,5, 9,13 set 0*/                        \
 407     r11 = vec_mergel(r1, r5);   /*1,5, 9,13 set 1*/                        \
 408     r12 = vec_mergeh(r2, r6);   /*2,6,10,14 set 0*/                        \
 409     r13 = vec_mergel(r2, r6);   /*2,6,10,14 set 1*/                        \
 410     r14 = vec_mergeh(r3, r7);   /*3,7,11,15 set 0*/                        \
 411     r15 = vec_mergel(r3, r7);   /*3,7,11,15 set 1*/                        \
 412                                                                            \
 413     /*Third merge*/                                                        \
 414     r0 = vec_mergeh(r8,  r12);  /*0,2,4,6,8,10,12,14 set 0*/               \
 415     r1 = vec_mergel(r8,  r12);  /*0,2,4,6,8,10,12,14 set 1*/               \
 416     r2 = vec_mergeh(r9,  r13);  /*0,2,4,6,8,10,12,14 set 2*/               \
 417     r4 = vec_mergeh(r10, r14);  /*1,3,5,7,9,11,13,15 set 0*/               \
 418     r5 = vec_mergel(r10, r14);  /*1,3,5,7,9,11,13,15 set 1*/               \
 419     r6 = vec_mergeh(r11, r15);  /*1,3,5,7,9,11,13,15 set 2*/               \
 420     /* Don't need to compute 3 and 7*/                                     \
 421                                                                            \
 422     /*Final merge*/                                                        \
 423     r8  = vec_mergeh(r0, r4);   /*all set 0*/                              \
 424     r9  = vec_mergel(r0, r4);   /*all set 1*/                              \
 425     r10 = vec_mergeh(r1, r5);   /*all set 2*/                              \
 426     r11 = vec_mergel(r1, r5);   /*all set 3*/                              \
 427     r12 = vec_mergeh(r2, r6);   /*all set 4*/                              \
 428     r13 = vec_mergel(r2, r6);   /*all set 5*/                              \
 429     /* Don't need to compute 14 and 15*/                                   \
 430                                                                            \
 431 }
 432
 433 // out: o = |x-y| < a
 434 static inline vec_u8 diff_lt_altivec ( register vec_u8 x,
 435                                          register vec_u8 y,
 436                                          register vec_u8 a) {
 437
 438     register vec_u8 diff = vec_subs(x, y);
 439     register vec_u8 diffneg = vec_subs(y, x);
 440     register vec_u8 o = vec_or(diff, diffneg); /* |x-y| */
 441     o = (vec_u8)vec_cmplt(o, a);
 442     return o;
 443 }
 444
 445 static inline vec_u8 h264_deblock_mask ( register vec_u8 p0,
 446                                            register vec_u8 p1,
 447                                            register vec_u8 q0,
 448                                            register vec_u8 q1,
 449                                            register vec_u8 alpha,
 450                                            register vec_u8 beta) {
 451
 452     register vec_u8 mask;
 453     register vec_u8 tempmask;
 454
 455     mask = diff_lt_altivec(p0, q0, alpha);
 456     tempmask = diff_lt_altivec(p1, p0, beta);
 457     mask = vec_and(mask, tempmask);
 458     tempmask = diff_lt_altivec(q1, q0, beta);
 459     mask = vec_and(mask, tempmask);
 460
 461     return mask;
 462 }
 463
 464 // out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
 465 static inline vec_u8 h264_deblock_q1(register vec_u8 p0,
 466                                        register vec_u8 p1,
 467                                        register vec_u8 p2,
 468                                        register vec_u8 q0,
 469                                        register vec_u8 tc0) {
 470
 471     register vec_u8 average = vec_avg(p0, q0);
 472     register vec_u8 temp;
 473     register vec_u8 uncliped;
 474     register vec_u8 ones;
 475     register vec_u8 max;
 476     register vec_u8 min;
 477     register vec_u8 newp1;
 478
 479     temp = vec_xor(average, p2);
 480     average = vec_avg(average, p2);     /*avg(p2, avg(p0, q0)) */
 481     ones = vec_splat_u8(1);
 482     temp = vec_and(temp, ones);         /*(p2^avg(p0, q0)) & 1 */
 483     uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
 484     max = vec_adds(p1, tc0);
 485     min = vec_subs(p1, tc0);
 486     newp1 = vec_max(min, uncliped);
 487     newp1 = vec_min(max, newp1);
 488     return newp1;
 489 }
 490
 491 #define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) {                                           \
 492                                                                                                   \
 493     const vec_u8 A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4));                               \
 494                                                                                                   \
 495     register vec_u8 pq0bit = vec_xor(p0,q0);                                                    \
 496     register vec_u8 q1minus;                                                                    \
 497     register vec_u8 p0minus;                                                                    \
 498     register vec_u8 stage1;                                                                     \
 499     register vec_u8 stage2;                                                                     \
 500     register vec_u8 vec160;                                                                     \
 501     register vec_u8 delta;                                                                      \
 502     register vec_u8 deltaneg;                                                                   \
 503                                                                                                   \
 504     q1minus = vec_nor(q1, q1);                 /* 255 - q1 */                                     \
 505     stage1 = vec_avg(p1, q1minus);             /* (p1 - q1 + 256)>>1 */                           \
 506     stage2 = vec_sr(stage1, vec_splat_u8(1));  /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */     \
 507     p0minus = vec_nor(p0, p0);                 /* 255 - p0 */                                     \
 508     stage1 = vec_avg(q0, p0minus);             /* (q0 - p0 + 256)>>1 */                           \
 509     pq0bit = vec_and(pq0bit, vec_splat_u8(1));                                                    \
 510     stage2 = vec_avg(stage2, pq0bit);          /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */ \
 511     stage2 = vec_adds(stage2, stage1);         /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */  \
 512     vec160 = vec_ld(0, &A0v);                                                                     \
 513     deltaneg = vec_subs(vec160, stage2);       /* -d */                                           \
 514     delta = vec_subs(stage2, vec160);          /* d */                                            \
 515     deltaneg = vec_min(tc0masked, deltaneg);                                                      \
 516     delta = vec_min(tc0masked, delta);                                                            \
 517     p0 = vec_subs(p0, deltaneg);                                                                  \
 518     q0 = vec_subs(q0, delta);                                                                     \
 519     p0 = vec_adds(p0, delta);                                                                     \
 520     q0 = vec_adds(q0, deltaneg);                                                                  \
 521 }
 522
 523 #define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) {            \
 524     DECLARE_ALIGNED(16, unsigned char, temp)[16];                                             \
 525     register vec_u8 alphavec;                                                              \
 526     register vec_u8 betavec;                                                               \
 527     register vec_u8 mask;                                                                  \
 528     register vec_u8 p1mask;                                                                \
 529     register vec_u8 q1mask;                                                                \
 530     register vector signed   char tc0vec;                                                    \
 531     register vec_u8 finaltc0;                                                              \
 532     register vec_u8 tc0masked;                                                             \
 533     register vec_u8 newp1;                                                                 \
 534     register vec_u8 newq1;                                                                 \
 535                                                                                              \
 536     temp[0] = alpha;                                                                         \
 537     temp[1] = beta;                                                                          \
 538     alphavec = vec_ld(0, temp);                                                              \
 539     betavec = vec_splat(alphavec, 0x1);                                                      \
 540     alphavec = vec_splat(alphavec, 0x0);                                                     \
 541     mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */            \
 542                                                                                              \
 543     AV_COPY32(temp, tc0);                                                                    \
 544     tc0vec = vec_ld(0, (signed char*)temp);                                                  \
 545     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 546     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 547     mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1)));  /* if tc0[i] >= 0 */         \
 548     finaltc0 = vec_and((vec_u8)tc0vec, mask);     /* tc = tc0 */                           \
 549                                                                                              \
 550     p1mask = diff_lt_altivec(p2, p0, betavec);                                               \
 551     p1mask = vec_and(p1mask, mask);                             /* if ( |p2 - p0| < beta) */ \
 552     tc0masked = vec_and(p1mask, (vec_u8)tc0vec);                                           \
 553     finaltc0 = vec_sub(finaltc0, p1mask);                       /* tc++ */                   \
 554     newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked);                                      \
 555     /*end if*/                                                                               \
 556                                                                                              \
 557     q1mask = diff_lt_altivec(q2, q0, betavec);                                               \
 558     q1mask = vec_and(q1mask, mask);                             /* if ( |q2 - q0| < beta ) */\
 559     tc0masked = vec_and(q1mask, (vec_u8)tc0vec);                                           \
 560     finaltc0 = vec_sub(finaltc0, q1mask);                       /* tc++ */                   \
 561     newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked);                                      \
 562     /*end if*/                                                                               \
 563                                                                                              \
 564     h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0);                                            \
 565     p1 = newp1;                                                                              \
 566     q1 = newq1;                                                                              \
 567 }
 568
 569 static void h264_v_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
 570
 571     if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) {
 572         register vec_u8 p2 = vec_ld(-3*stride, pix);
 573         register vec_u8 p1 = vec_ld(-2*stride, pix);
 574         register vec_u8 p0 = vec_ld(-1*stride, pix);
 575         register vec_u8 q0 = vec_ld(0, pix);
 576         register vec_u8 q1 = vec_ld(stride, pix);
 577         register vec_u8 q2 = vec_ld(2*stride, pix);
 578         h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
 579         vec_st(p1, -2*stride, pix);
 580         vec_st(p0, -1*stride, pix);
 581         vec_st(q0, 0, pix);
 582         vec_st(q1, stride, pix);
 583     }
 584 }
 585
 586 static void h264_h_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
 587
 588     register vec_u8 line0, line1, line2, line3, line4, line5;
 589     if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0)
 590         return;
 591     readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
 592     h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
 593     transpose4x16(line1, line2, line3, line4);
 594     write16x4(pix-2, stride, line1, line2, line3, line4);
 595 }
 596
 597 static av_always_inline
 598 void weight_h264_W_altivec(uint8_t *block, int stride, int height,
 599                            int log2_denom, int weight, int offset, int w)
 600 {
 601     int y, aligned;
 602     vec_u8 vblock;
 603     vec_s16 vtemp, vweight, voffset, v0, v1;
 604     vec_u16 vlog2_denom;
 605     DECLARE_ALIGNED(16, int32_t, temp)[4];
 606     LOAD_ZERO;
 607
 608     offset <<= log2_denom;
 609     if(log2_denom) offset += 1<<(log2_denom-1);
 610     temp[0] = log2_denom;
 611     temp[1] = weight;
 612     temp[2] = offset;
 613
 614     vtemp = (vec_s16)vec_ld(0, temp);
 615     vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
 616     vweight = vec_splat(vtemp, 3);
 617     voffset = vec_splat(vtemp, 5);
 618     aligned = !((unsigned long)block & 0xf);
 619
 620     for (y = 0; y < height; y++) {
 621         vblock = vec_ld(0, block);
 622
 623         v0 = (vec_s16)vec_mergeh(zero_u8v, vblock);
 624         v1 = (vec_s16)vec_mergel(zero_u8v, vblock);
 625
 626         if (w == 16 || aligned) {
 627             v0 = vec_mladd(v0, vweight, zero_s16v);
 628             v0 = vec_adds(v0, voffset);
 629             v0 = vec_sra(v0, vlog2_denom);
 630         }
 631         if (w == 16 || !aligned) {
 632             v1 = vec_mladd(v1, vweight, zero_s16v);
 633             v1 = vec_adds(v1, voffset);
 634             v1 = vec_sra(v1, vlog2_denom);
 635         }
 636         vblock = vec_packsu(v0, v1);
 637         vec_st(vblock, 0, block);
 638
 639         block += stride;
 640     }
 641 }
 642
 643 static av_always_inline
 644 void biweight_h264_W_altivec(uint8_t *dst, uint8_t *src, int stride, int height,
 645                              int log2_denom, int weightd, int weights, int offset, int w)
 646 {
 647     int y, dst_aligned, src_aligned;
 648     vec_u8 vsrc, vdst;
 649     vec_s16 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3;
 650     vec_u16 vlog2_denom;
 651     DECLARE_ALIGNED(16, int32_t, temp)[4];
 652     LOAD_ZERO;
 653
 654     offset = ((offset + 1) | 1) << log2_denom;
 655     temp[0] = log2_denom+1;
 656     temp[1] = weights;
 657     temp[2] = weightd;
 658     temp[3] = offset;
 659
 660     vtemp = (vec_s16)vec_ld(0, temp);
 661     vlog2_denom = (vec_u16)vec_splat(vtemp, 1);
 662     vweights = vec_splat(vtemp, 3);
 663     vweightd = vec_splat(vtemp, 5);
 664     voffset = vec_splat(vtemp, 7);
 665     dst_aligned = !((unsigned long)dst & 0xf);
 666     src_aligned = !((unsigned long)src & 0xf);
 667
 668     for (y = 0; y < height; y++) {
 669         vdst = vec_ld(0, dst);
 670         vsrc = vec_ld(0, src);
 671
 672         v0 = (vec_s16)vec_mergeh(zero_u8v, vdst);
 673         v1 = (vec_s16)vec_mergel(zero_u8v, vdst);
 674         v2 = (vec_s16)vec_mergeh(zero_u8v, vsrc);
 675         v3 = (vec_s16)vec_mergel(zero_u8v, vsrc);
 676
 677         if (w == 8) {
 678             if (src_aligned)
 679                 v3 = v2;
 680             else
 681                 v2 = v3;
 682         }
 683
 684         if (w == 16 || dst_aligned) {
 685             v0 = vec_mladd(v0, vweightd, zero_s16v);
 686             v2 = vec_mladd(v2, vweights, zero_s16v);
 687
 688             v0 = vec_adds(v0, voffset);
 689             v0 = vec_adds(v0, v2);
 690             v0 = vec_sra(v0, vlog2_denom);
 691         }
 692         if (w == 16 || !dst_aligned) {
 693             v1 = vec_mladd(v1, vweightd, zero_s16v);
 694             v3 = vec_mladd(v3, vweights, zero_s16v);
 695
 696             v1 = vec_adds(v1, voffset);
 697             v1 = vec_adds(v1, v3);
 698             v1 = vec_sra(v1, vlog2_denom);
 699         }
 700         vdst = vec_packsu(v0, v1);
 701         vec_st(vdst, 0, dst);
 702
 703         dst += stride;
 704         src += stride;
 705     }
 706 }
 707
 708 #define H264_WEIGHT(W) \
 709 static void ff_weight_h264_pixels ## W ## _altivec(uint8_t *block, int stride, int height, \
 710                                                    int log2_denom, int weight, int offset){ \
 711     weight_h264_W_altivec(block, stride, height, log2_denom, weight, offset, W); \
 712 }\
 713 static void ff_biweight_h264_pixels ## W ## _altivec(uint8_t *dst, uint8_t *src, int stride, int height, \
 714                                                      int log2_denom, int weightd, int weights, int offset){ \
 715     biweight_h264_W_altivec(dst, src, stride, height, log2_denom, weightd, weights, offset, W); \
 716 }
 717
 718 H264_WEIGHT(16)
 719 H264_WEIGHT( 8)
 720
 721 void ff_h264dsp_init_ppc(H264DSPContext *c, const int bit_depth, const int chroma_format_idc)
 722 {
 723     if (av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC) {
 724     if (bit_depth == 8) {
 725         c->h264_idct_add = ff_h264_idct_add_altivec;
 726         if (chroma_format_idc == 1)
 727             c->h264_idct_add8 = ff_h264_idct_add8_altivec;
 728         c->h264_idct_add16 = ff_h264_idct_add16_altivec;
 729         c->h264_idct_add16intra = ff_h264_idct_add16intra_altivec;
 730         c->h264_idct_dc_add= h264_idct_dc_add_altivec;
 731         c->h264_idct8_dc_add = ff_h264_idct8_dc_add_altivec;
 732         c->h264_idct8_add = ff_h264_idct8_add_altivec;
 733         c->h264_idct8_add4 = ff_h264_idct8_add4_altivec;
 734         c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_altivec;
 735         c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_altivec;
 736
 737         c->weight_h264_pixels_tab[0] = ff_weight_h264_pixels16_altivec;
 738         c->weight_h264_pixels_tab[1] = ff_weight_h264_pixels8_altivec;
 739         c->biweight_h264_pixels_tab[0] = ff_biweight_h264_pixels16_altivec;
 740         c->biweight_h264_pixels_tab[1] = ff_biweight_h264_pixels8_altivec;
 741     }
 742     }
 743 }