git.sesse.net Git - x264/blob - common/ppc/deblock.c

   1 /*****************************************************************************
   2  * deblock.c: ppc deblocking
   3  *****************************************************************************
   4  * Copyright (C) 2007-2012 x264 project
   5  *
   6  * Authors: Guillaume Poirier <gpoirier@mplayerhq.hu>
   7  *
   8  * This program is free software; you can redistribute it and/or modify
   9  * it under the terms of the GNU General Public License as published by
  10  * the Free Software Foundation; either version 2 of the License, or
  11  * (at your option) any later version.
  12  *
  13  * This program is distributed in the hope that it will be useful,
  14  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  16  * GNU General Public License for more details.
  17  *
  18  * You should have received a copy of the GNU General Public License
  19  * along with this program; if not, write to the Free Software
  20  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
  21  *
  22  * This program is also available under a commercial proprietary license.
  23  * For more information, contact us at licensing@x264.com.
  24  *****************************************************************************/
  25
  26 #include "common/common.h"
  27 #include "ppccommon.h"
  28
  29 #if !HIGH_BIT_DEPTH
  30 #define transpose4x16(r0, r1, r2, r3)        \
  31 {                                            \
  32     register vec_u8_t r4;                    \
  33     register vec_u8_t r5;                    \
  34     register vec_u8_t r6;                    \
  35     register vec_u8_t r7;                    \
  36                                              \
  37     r4 = vec_mergeh(r0, r2);  /*0, 2 set 0*/ \
  38     r5 = vec_mergel(r0, r2);  /*0, 2 set 1*/ \
  39     r6 = vec_mergeh(r1, r3);  /*1, 3 set 0*/ \
  40     r7 = vec_mergel(r1, r3);  /*1, 3 set 1*/ \
  41                                              \
  42     r0 = vec_mergeh(r4, r6);  /*all set 0*/  \
  43     r1 = vec_mergel(r4, r6);  /*all set 1*/  \
  44     r2 = vec_mergeh(r5, r7);  /*all set 2*/  \
  45     r3 = vec_mergel(r5, r7);  /*all set 3*/  \
  46 }
  47
  48 static inline void write16x4( uint8_t *dst, int dst_stride,
  49                               register vec_u8_t r0, register vec_u8_t r1,
  50                               register vec_u8_t r2, register vec_u8_t r3 )
  51 {
  52     ALIGNED_16(unsigned char result[64]);
  53     uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
  54     int int_dst_stride = dst_stride >> 2;
  55
  56     vec_st(r0, 0, result);
  57     vec_st(r1, 16, result);
  58     vec_st(r2, 32, result);
  59     vec_st(r3, 48, result);
  60     /* FIXME: there has to be a better way!!!! */
  61     *dst_int = *src_int;
  62     *(dst_int+   int_dst_stride) = *(src_int + 1);
  63     *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
  64     *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
  65     *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
  66     *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
  67     *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
  68     *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
  69     *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
  70     *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
  71     *(dst_int+10*int_dst_stride) = *(src_int + 10);
  72     *(dst_int+11*int_dst_stride) = *(src_int + 11);
  73     *(dst_int+12*int_dst_stride) = *(src_int + 12);
  74     *(dst_int+13*int_dst_stride) = *(src_int + 13);
  75     *(dst_int+14*int_dst_stride) = *(src_int + 14);
  76     *(dst_int+15*int_dst_stride) = *(src_int + 15);
  77 }
  78
  79 /** \brief performs a 6x16 transpose of data in src, and stores it to dst */
  80 #define read_and_transpose16x6(src, src_stride, r8, r9, r10, r11, r12, r13)\
  81 {\
  82     register vec_u8_t r0, r1, r2, r3, r4, r5, r6, r7, r14, r15;\
  83     VEC_LOAD(src,                  r0, 16, vec_u8_t, pix );    \
  84     VEC_LOAD(src +    src_stride,  r1, 16, vec_u8_t, pix );    \
  85     VEC_LOAD(src +  2*src_stride,  r2, 16, vec_u8_t, pix );    \
  86     VEC_LOAD(src +  3*src_stride,  r3, 16, vec_u8_t, pix );    \
  87     VEC_LOAD(src +  4*src_stride,  r4, 16, vec_u8_t, pix );    \
  88     VEC_LOAD(src +  5*src_stride,  r5, 16, vec_u8_t, pix );    \
  89     VEC_LOAD(src +  6*src_stride,  r6, 16, vec_u8_t, pix );    \
  90     VEC_LOAD(src +  7*src_stride,  r7, 16, vec_u8_t, pix );    \
  91     VEC_LOAD(src + 14*src_stride, r14, 16, vec_u8_t, pix );    \
  92     VEC_LOAD(src + 15*src_stride, r15, 16, vec_u8_t, pix );    \
  93                                                                \
  94     VEC_LOAD(src + 8*src_stride,   r8, 16, vec_u8_t, pix );    \
  95     VEC_LOAD(src + 9*src_stride,   r9, 16, vec_u8_t, pix );    \
  96     VEC_LOAD(src + 10*src_stride, r10, 16, vec_u8_t, pix );    \
  97     VEC_LOAD(src + 11*src_stride, r11, 16, vec_u8_t, pix );    \
  98     VEC_LOAD(src + 12*src_stride, r12, 16, vec_u8_t, pix );    \
  99     VEC_LOAD(src + 13*src_stride, r13, 16, vec_u8_t, pix );    \
 100                                                                \
 101     /*Merge first pairs*/                                      \
 102     r0 = vec_mergeh(r0, r8);    /*0, 8*/                       \
 103     r1 = vec_mergeh(r1, r9);    /*1, 9*/                       \
 104     r2 = vec_mergeh(r2, r10);   /*2,10*/                       \
 105     r3 = vec_mergeh(r3, r11);   /*3,11*/                       \
 106     r4 = vec_mergeh(r4, r12);   /*4,12*/                       \
 107     r5 = vec_mergeh(r5, r13);   /*5,13*/                       \
 108     r6 = vec_mergeh(r6, r14);   /*6,14*/                       \
 109     r7 = vec_mergeh(r7, r15);   /*7,15*/                       \
 110                                                                \
 111     /*Merge second pairs*/                                     \
 112     r8  = vec_mergeh(r0, r4);   /*0,4, 8,12 set 0*/            \
 113     r9  = vec_mergel(r0, r4);   /*0,4, 8,12 set 1*/            \
 114     r10 = vec_mergeh(r1, r5);   /*1,5, 9,13 set 0*/            \
 115     r11 = vec_mergel(r1, r5);   /*1,5, 9,13 set 1*/            \
 116     r12 = vec_mergeh(r2, r6);   /*2,6,10,14 set 0*/            \
 117     r13 = vec_mergel(r2, r6);   /*2,6,10,14 set 1*/            \
 118     r14 = vec_mergeh(r3, r7);   /*3,7,11,15 set 0*/            \
 119     r15 = vec_mergel(r3, r7);   /*3,7,11,15 set 1*/            \
 120                                                                \
 121     /*Third merge*/                                            \
 122     r0 = vec_mergeh(r8, r12);   /*0,2,4,6,8,10,12,14 set 0*/   \
 123     r1 = vec_mergel(r8, r12);   /*0,2,4,6,8,10,12,14 set 1*/   \
 124     r2 = vec_mergeh(r9, r13);   /*0,2,4,6,8,10,12,14 set 2*/   \
 125     r4 = vec_mergeh(r10, r14);  /*1,3,5,7,9,11,13,15 set 0*/   \
 126     r5 = vec_mergel(r10, r14);  /*1,3,5,7,9,11,13,15 set 1*/   \
 127     r6 = vec_mergeh(r11, r15);  /*1,3,5,7,9,11,13,15 set 2*/   \
 128     /* Don't need to compute 3 and 7*/                         \
 129                                                                \
 130     /*Final merge*/                                            \
 131     r8  = vec_mergeh(r0, r4);   /*all set 0*/                  \
 132     r9  = vec_mergel(r0, r4);   /*all set 1*/                  \
 133     r10 = vec_mergeh(r1, r5);   /*all set 2*/                  \
 134     r11 = vec_mergel(r1, r5);   /*all set 3*/                  \
 135     r12 = vec_mergeh(r2, r6);   /*all set 4*/                  \
 136     r13 = vec_mergel(r2, r6);   /*all set 5*/                  \
 137     /* Don't need to compute 14 and 15*/                       \
 138                                                                \
 139 }
 140
 141 // out: o = |x-y| < a
 142 static inline vec_u8_t diff_lt_altivec( register vec_u8_t x, register vec_u8_t y, register vec_u8_t a )
 143 {
 144     register vec_u8_t diff = vec_subs(x, y);
 145     register vec_u8_t diffneg = vec_subs(y, x);
 146     register vec_u8_t o = vec_or(diff, diffneg); /* |x-y| */
 147     o = (vec_u8_t)vec_cmplt(o, a);
 148     return o;
 149 }
 150
 151 static inline vec_u8_t h264_deblock_mask( register vec_u8_t p0, register vec_u8_t p1, register vec_u8_t q0,
 152                                           register vec_u8_t q1, register vec_u8_t alpha, register vec_u8_t beta )
 153 {
 154     register vec_u8_t mask;
 155     register vec_u8_t tempmask;
 156
 157     mask = diff_lt_altivec(p0, q0, alpha);
 158     tempmask = diff_lt_altivec(p1, p0, beta);
 159     mask = vec_and(mask, tempmask);
 160     tempmask = diff_lt_altivec(q1, q0, beta);
 161     mask = vec_and(mask, tempmask);
 162
 163     return mask;
 164 }
 165
 166 // out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
 167 static inline vec_u8_t h264_deblock_q1( register vec_u8_t p0, register vec_u8_t p1, register vec_u8_t p2,
 168                                         register vec_u8_t q0, register vec_u8_t tc0 )
 169 {
 170
 171     register vec_u8_t average = vec_avg(p0, q0);
 172     register vec_u8_t temp;
 173     register vec_u8_t uncliped;
 174     register vec_u8_t ones;
 175     register vec_u8_t max;
 176     register vec_u8_t min;
 177     register vec_u8_t newp1;
 178
 179     temp = vec_xor(average, p2);
 180     average = vec_avg(average, p2);     /*avg(p2, avg(p0, q0)) */
 181     ones = vec_splat_u8(1);
 182     temp = vec_and(temp, ones);         /*(p2^avg(p0, q0)) & 1 */
 183     uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
 184     max = vec_adds(p1, tc0);
 185     min = vec_subs(p1, tc0);
 186     newp1 = vec_max(min, uncliped);
 187     newp1 = vec_min(max, newp1);
 188     return newp1;
 189 }
 190
 191 #define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked)                                           \
 192 {                                                                                               \
 193     const vec_u8_t A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4));                             \
 194                                                                                                 \
 195     register vec_u8_t pq0bit = vec_xor(p0,q0);                                                  \
 196     register vec_u8_t q1minus;                                                                  \
 197     register vec_u8_t p0minus;                                                                  \
 198     register vec_u8_t stage1;                                                                   \
 199     register vec_u8_t stage2;                                                                   \
 200     register vec_u8_t vec160;                                                                   \
 201     register vec_u8_t delta;                                                                    \
 202     register vec_u8_t deltaneg;                                                                 \
 203                                                                                                 \
 204     q1minus = vec_nor(q1, q1);                /* 255 - q1 */                                    \
 205     stage1 = vec_avg(p1, q1minus);            /* (p1 - q1 + 256)>>1 */                          \
 206     stage2 = vec_sr(stage1, vec_splat_u8(1)); /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */    \
 207     p0minus = vec_nor(p0, p0);                /* 255 - p0 */                                    \
 208     stage1 = vec_avg(q0, p0minus);            /* (q0 - p0 + 256)>>1 */                          \
 209     pq0bit = vec_and(pq0bit, vec_splat_u8(1));                                                  \
 210     stage2 = vec_avg(stage2, pq0bit);         /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */\
 211     stage2 = vec_adds(stage2, stage1);        /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */ \
 212     vec160 = vec_ld(0, &A0v);                                                                   \
 213     deltaneg = vec_subs(vec160, stage2);      /* -d */                                          \
 214     delta = vec_subs(stage2, vec160);         /*  d */                                          \
 215     deltaneg = vec_min(tc0masked, deltaneg);                                                    \
 216     delta = vec_min(tc0masked, delta);                                                          \
 217     p0 = vec_subs(p0, deltaneg);                                                                \
 218     q0 = vec_subs(q0, delta);                                                                   \
 219     p0 = vec_adds(p0, delta);                                                                   \
 220     q0 = vec_adds(q0, deltaneg);                                                                \
 221 }
 222
 223 #define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0)              \
 224 {                                                                                            \
 225     ALIGNED_16(unsigned char temp[16]);                                                      \
 226     register vec_u8_t alphavec;                                                              \
 227     register vec_u8_t betavec;                                                               \
 228     register vec_u8_t mask;                                                                  \
 229     register vec_u8_t p1mask;                                                                \
 230     register vec_u8_t q1mask;                                                                \
 231     register vec_s8_t tc0vec;                                                                \
 232     register vec_u8_t finaltc0;                                                              \
 233     register vec_u8_t tc0masked;                                                             \
 234     register vec_u8_t newp1;                                                                 \
 235     register vec_u8_t newq1;                                                                 \
 236                                                                                              \
 237     temp[0] = alpha;                                                                         \
 238     temp[1] = beta;                                                                          \
 239     alphavec = vec_ld(0, temp);                                                              \
 240     betavec = vec_splat(alphavec, 0x1);                                                      \
 241     alphavec = vec_splat(alphavec, 0x0);                                                     \
 242     mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */            \
 243                                                                                              \
 244     M32( temp ) = M32( tc0 );                                                                \
 245     tc0vec = vec_ld(0, (signed char*)temp);                                                  \
 246     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 247     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 248     mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1)));  /* if tc0[i] >= 0 */         \
 249     finaltc0 = vec_and((vec_u8_t)tc0vec, mask);                 /* tc = tc0 */               \
 250                                                                                              \
 251     p1mask = diff_lt_altivec(p2, p0, betavec);                                               \
 252     p1mask = vec_and(p1mask, mask);                             /* if( |p2 - p0| < beta) */  \
 253     tc0masked = vec_and(p1mask, (vec_u8_t)tc0vec);                                           \
 254     finaltc0 = vec_sub(finaltc0, p1mask);                       /* tc++ */                   \
 255     newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked);                                      \
 256     /*end if*/                                                                               \
 257                                                                                              \
 258     q1mask = diff_lt_altivec(q2, q0, betavec);                                               \
 259     q1mask = vec_and(q1mask, mask);                             /* if ( |q2 - q0| < beta ) */\
 260     tc0masked = vec_and(q1mask, (vec_u8_t)tc0vec);                                           \
 261     finaltc0 = vec_sub(finaltc0, q1mask);                       /* tc++ */                   \
 262     newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked);                                      \
 263     /*end if*/                                                                               \
 264                                                                                              \
 265     h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0);                                            \
 266     p1 = newp1;                                                                              \
 267     q1 = newq1;                                                                              \
 268 }
 269
 270 void x264_deblock_v_luma_altivec( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
 271 {
 272     if( (tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0 )
 273     {
 274         register vec_u8_t p2 = vec_ld(-3*stride, pix);
 275         register vec_u8_t p1 = vec_ld(-2*stride, pix);
 276         register vec_u8_t p0 = vec_ld(-1*stride, pix);
 277         register vec_u8_t q0 = vec_ld(0, pix);
 278         register vec_u8_t q1 = vec_ld(stride, pix);
 279         register vec_u8_t q2 = vec_ld(2*stride, pix);
 280         h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
 281         vec_st(p1, -2*stride, pix);
 282         vec_st(p0, -1*stride, pix);
 283         vec_st(q0, 0, pix);
 284         vec_st(q1, stride, pix);
 285     }
 286 }
 287
 288 void x264_deblock_h_luma_altivec( uint8_t *pix, intptr_t stride, int alpha, int beta, int8_t *tc0 )
 289 {
 290
 291     register vec_u8_t line0, line1, line2, line3, line4, line5;
 292     if( (tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0 )
 293         return;
 294     PREP_LOAD;
 295     vec_u8_t _pix_ = vec_lvsl(0, pix-3);
 296     read_and_transpose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
 297     h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
 298     transpose4x16(line1, line2, line3, line4);
 299     write16x4(pix-2, stride, line1, line2, line3, line4);
 300 }
 301 #endif // !HIGH_BIT_DEPTH