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

   1 /*****************************************************************************
   2 * deblock.c: Altivec-accelerated deblocking for h264 encoder
   3 *****************************************************************************
   4 * Copyright (C) 2007-2008 Guillaume Poirier <gpoirier@mplayerhq.hu>
   5 *
   6 * This program is free software; you can redistribute it and/or modify
   7 * it under the terms of the GNU General Public License as published by
   8 * the Free Software Foundation; either version 2 of the License, or
   9 * (at your option) any later version.
  10 *
  11 * This program 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
  14 * GNU General Public License for more details.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * along with this program; if not, write to the Free Software
  18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02111, USA.
  19 *****************************************************************************/
  20
  21 #include "common/common.h"
  22 #include "ppccommon.h"
  23
  24 #define transpose4x16(r0, r1, r2, r3)        \
  25 {                                            \
  26     register vec_u8_t r4;                    \
  27     register vec_u8_t r5;                    \
  28     register vec_u8_t r6;                    \
  29     register vec_u8_t r7;                    \
  30                                              \
  31     r4 = vec_mergeh(r0, r2);  /*0, 2 set 0*/ \
  32     r5 = vec_mergel(r0, r2);  /*0, 2 set 1*/ \
  33     r6 = vec_mergeh(r1, r3);  /*1, 3 set 0*/ \
  34     r7 = vec_mergel(r1, r3);  /*1, 3 set 1*/ \
  35                                              \
  36     r0 = vec_mergeh(r4, r6);  /*all set 0*/  \
  37     r1 = vec_mergel(r4, r6);  /*all set 1*/  \
  38     r2 = vec_mergeh(r5, r7);  /*all set 2*/  \
  39     r3 = vec_mergel(r5, r7);  /*all set 3*/  \
  40 }
  41
  42 static inline void write16x4( uint8_t *dst, int dst_stride,
  43                               register vec_u8_t r0, register vec_u8_t r1,
  44                               register vec_u8_t r2, register vec_u8_t r3 )
  45 {
  46     ALIGNED_16(unsigned char result[64]);
  47     uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
  48     int int_dst_stride = dst_stride >> 2;
  49
  50     vec_st(r0, 0, result);
  51     vec_st(r1, 16, result);
  52     vec_st(r2, 32, result);
  53     vec_st(r3, 48, result);
  54     /* FIXME: there has to be a better way!!!! */
  55     *dst_int = *src_int;
  56     *(dst_int+   int_dst_stride) = *(src_int + 1);
  57     *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
  58     *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
  59     *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
  60     *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
  61     *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
  62     *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
  63     *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
  64     *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
  65     *(dst_int+10*int_dst_stride) = *(src_int + 10);
  66     *(dst_int+11*int_dst_stride) = *(src_int + 11);
  67     *(dst_int+12*int_dst_stride) = *(src_int + 12);
  68     *(dst_int+13*int_dst_stride) = *(src_int + 13);
  69     *(dst_int+14*int_dst_stride) = *(src_int + 14);
  70     *(dst_int+15*int_dst_stride) = *(src_int + 15);
  71 }
  72
  73 /** \brief performs a 6x16 transpose of data in src, and stores it to dst */
  74 #define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13)\
  75 {\
  76     register vec_u8_t r0, r1, r2, r3, r4, r5, r6, r7, r14, r15;\
  77     VEC_LOAD(src,                  r0, 16, vec_u8_t, pix );    \
  78     VEC_LOAD(src +    src_stride,  r1, 16, vec_u8_t, pix );    \
  79     VEC_LOAD(src +  2*src_stride,  r2, 16, vec_u8_t, pix );    \
  80     VEC_LOAD(src +  3*src_stride,  r3, 16, vec_u8_t, pix );    \
  81     VEC_LOAD(src +  4*src_stride,  r4, 16, vec_u8_t, pix );    \
  82     VEC_LOAD(src +  5*src_stride,  r5, 16, vec_u8_t, pix );    \
  83     VEC_LOAD(src +  6*src_stride,  r6, 16, vec_u8_t, pix );    \
  84     VEC_LOAD(src +  7*src_stride,  r7, 16, vec_u8_t, pix );    \
  85     VEC_LOAD(src + 14*src_stride, r14, 16, vec_u8_t, pix );    \
  86     VEC_LOAD(src + 15*src_stride, r15, 16, vec_u8_t, pix );    \
  87                                                                \
  88     VEC_LOAD(src + 8*src_stride,   r8, 16, vec_u8_t, pix );    \
  89     VEC_LOAD(src + 9*src_stride,   r9, 16, vec_u8_t, pix );    \
  90     VEC_LOAD(src + 10*src_stride, r10, 16, vec_u8_t, pix );    \
  91     VEC_LOAD(src + 11*src_stride, r11, 16, vec_u8_t, pix );    \
  92     VEC_LOAD(src + 12*src_stride, r12, 16, vec_u8_t, pix );    \
  93     VEC_LOAD(src + 13*src_stride, r13, 16, vec_u8_t, pix );    \
  94                                                                \
  95     /*Merge first pairs*/                                      \
  96     r0 = vec_mergeh(r0, r8);    /*0, 8*/                       \
  97     r1 = vec_mergeh(r1, r9);    /*1, 9*/                       \
  98     r2 = vec_mergeh(r2, r10);   /*2,10*/                       \
  99     r3 = vec_mergeh(r3, r11);   /*3,11*/                       \
 100     r4 = vec_mergeh(r4, r12);   /*4,12*/                       \
 101     r5 = vec_mergeh(r5, r13);   /*5,13*/                       \
 102     r6 = vec_mergeh(r6, r14);   /*6,14*/                       \
 103     r7 = vec_mergeh(r7, r15);   /*7,15*/                       \
 104                                                                \
 105     /*Merge second pairs*/                                     \
 106     r8  = vec_mergeh(r0, r4);   /*0,4, 8,12 set 0*/            \
 107     r9  = vec_mergel(r0, r4);   /*0,4, 8,12 set 1*/            \
 108     r10 = vec_mergeh(r1, r5);   /*1,5, 9,13 set 0*/            \
 109     r11 = vec_mergel(r1, r5);   /*1,5, 9,13 set 1*/            \
 110     r12 = vec_mergeh(r2, r6);   /*2,6,10,14 set 0*/            \
 111     r13 = vec_mergel(r2, r6);   /*2,6,10,14 set 1*/            \
 112     r14 = vec_mergeh(r3, r7);   /*3,7,11,15 set 0*/            \
 113     r15 = vec_mergel(r3, r7);   /*3,7,11,15 set 1*/            \
 114                                                                \
 115     /*Third merge*/                                            \
 116     r0 = vec_mergeh(r8, r12);   /*0,2,4,6,8,10,12,14 set 0*/   \
 117     r1 = vec_mergel(r8, r12);   /*0,2,4,6,8,10,12,14 set 1*/   \
 118     r2 = vec_mergeh(r9, r13);   /*0,2,4,6,8,10,12,14 set 2*/   \
 119     r4 = vec_mergeh(r10, r14);  /*1,3,5,7,9,11,13,15 set 0*/   \
 120     r5 = vec_mergel(r10, r14);  /*1,3,5,7,9,11,13,15 set 1*/   \
 121     r6 = vec_mergeh(r11, r15);  /*1,3,5,7,9,11,13,15 set 2*/   \
 122     /* Don't need to compute 3 and 7*/                         \
 123                                                                \
 124     /*Final merge*/                                            \
 125     r8  = vec_mergeh(r0, r4);   /*all set 0*/                  \
 126     r9  = vec_mergel(r0, r4);   /*all set 1*/                  \
 127     r10 = vec_mergeh(r1, r5);   /*all set 2*/                  \
 128     r11 = vec_mergel(r1, r5);   /*all set 3*/                  \
 129     r12 = vec_mergeh(r2, r6);   /*all set 4*/                  \
 130     r13 = vec_mergel(r2, r6);   /*all set 5*/                  \
 131     /* Don't need to compute 14 and 15*/                       \
 132                                                                \
 133 }
 134
 135 // out: o = |x-y| < a
 136 static inline vec_u8_t diff_lt_altivec( register vec_u8_t x, register vec_u8_t y, register vec_u8_t a )
 137 {
 138     register vec_u8_t diff = vec_subs(x, y);
 139     register vec_u8_t diffneg = vec_subs(y, x);
 140     register vec_u8_t o = vec_or(diff, diffneg); /* |x-y| */
 141     o = (vec_u8_t)vec_cmplt(o, a);
 142     return o;
 143 }
 144
 145 static inline vec_u8_t h264_deblock_mask( register vec_u8_t p0, register vec_u8_t p1, register vec_u8_t q0,
 146                                           register vec_u8_t q1, register vec_u8_t alpha, register vec_u8_t beta )
 147 {
 148     register vec_u8_t mask;
 149     register vec_u8_t tempmask;
 150
 151     mask = diff_lt_altivec(p0, q0, alpha);
 152     tempmask = diff_lt_altivec(p1, p0, beta);
 153     mask = vec_and(mask, tempmask);
 154     tempmask = diff_lt_altivec(q1, q0, beta);
 155     mask = vec_and(mask, tempmask);
 156
 157     return mask;
 158 }
 159
 160 // out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
 161 static inline vec_u8_t h264_deblock_q1( register vec_u8_t p0, register vec_u8_t p1, register vec_u8_t p2,
 162                                         register vec_u8_t q0, register vec_u8_t tc0 )
 163 {
 164
 165     register vec_u8_t average = vec_avg(p0, q0);
 166     register vec_u8_t temp;
 167     register vec_u8_t uncliped;
 168     register vec_u8_t ones;
 169     register vec_u8_t max;
 170     register vec_u8_t min;
 171     register vec_u8_t newp1;
 172
 173     temp = vec_xor(average, p2);
 174     average = vec_avg(average, p2);     /*avg(p2, avg(p0, q0)) */
 175     ones = vec_splat_u8(1);
 176     temp = vec_and(temp, ones);         /*(p2^avg(p0, q0)) & 1 */
 177     uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
 178     max = vec_adds(p1, tc0);
 179     min = vec_subs(p1, tc0);
 180     newp1 = vec_max(min, uncliped);
 181     newp1 = vec_min(max, newp1);
 182     return newp1;
 183 }
 184
 185 #define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked)                                           \
 186 {                                                                                               \
 187     const vec_u8_t A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4));                             \
 188                                                                                                 \
 189     register vec_u8_t pq0bit = vec_xor(p0,q0);                                                  \
 190     register vec_u8_t q1minus;                                                                  \
 191     register vec_u8_t p0minus;                                                                  \
 192     register vec_u8_t stage1;                                                                   \
 193     register vec_u8_t stage2;                                                                   \
 194     register vec_u8_t vec160;                                                                   \
 195     register vec_u8_t delta;                                                                    \
 196     register vec_u8_t deltaneg;                                                                 \
 197                                                                                                 \
 198     q1minus = vec_nor(q1, q1);                /* 255 - q1 */                                    \
 199     stage1 = vec_avg(p1, q1minus);            /* (p1 - q1 + 256)>>1 */                          \
 200     stage2 = vec_sr(stage1, vec_splat_u8(1)); /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */    \
 201     p0minus = vec_nor(p0, p0);                /* 255 - p0 */                                    \
 202     stage1 = vec_avg(q0, p0minus);            /* (q0 - p0 + 256)>>1 */                          \
 203     pq0bit = vec_and(pq0bit, vec_splat_u8(1));                                                  \
 204     stage2 = vec_avg(stage2, pq0bit);         /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */\
 205     stage2 = vec_adds(stage2, stage1);        /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */ \
 206     vec160 = vec_ld(0, &A0v);                                                                   \
 207     deltaneg = vec_subs(vec160, stage2);      /* -d */                                          \
 208     delta = vec_subs(stage2, vec160);         /*  d */                                          \
 209     deltaneg = vec_min(tc0masked, deltaneg);                                                    \
 210     delta = vec_min(tc0masked, delta);                                                          \
 211     p0 = vec_subs(p0, deltaneg);                                                                \
 212     q0 = vec_subs(q0, delta);                                                                   \
 213     p0 = vec_adds(p0, delta);                                                                   \
 214     q0 = vec_adds(q0, deltaneg);                                                                \
 215 }
 216
 217 #define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0)              \
 218 {                                                                                            \
 219     ALIGNED_16(unsigned char temp[16]);                                                      \
 220     register vec_u8_t alphavec;                                                              \
 221     register vec_u8_t betavec;                                                               \
 222     register vec_u8_t mask;                                                                  \
 223     register vec_u8_t p1mask;                                                                \
 224     register vec_u8_t q1mask;                                                                \
 225     register vec_s8_t tc0vec;                                                                \
 226     register vec_u8_t finaltc0;                                                              \
 227     register vec_u8_t tc0masked;                                                             \
 228     register vec_u8_t newp1;                                                                 \
 229     register vec_u8_t newq1;                                                                 \
 230                                                                                              \
 231     temp[0] = alpha;                                                                         \
 232     temp[1] = beta;                                                                          \
 233     alphavec = vec_ld(0, temp);                                                              \
 234     betavec = vec_splat(alphavec, 0x1);                                                      \
 235     alphavec = vec_splat(alphavec, 0x0);                                                     \
 236     mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */            \
 237                                                                                              \
 238     M32( temp ) = M32( tc0 );                                                                \
 239     tc0vec = vec_ld(0, (signed char*)temp);                                                  \
 240     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 241     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 242     mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1)));  /* if tc0[i] >= 0 */         \
 243     finaltc0 = vec_and((vec_u8_t)tc0vec, mask);                 /* tc = tc0 */               \
 244                                                                                              \
 245     p1mask = diff_lt_altivec(p2, p0, betavec);                                               \
 246     p1mask = vec_and(p1mask, mask);                             /* if( |p2 - p0| < beta) */  \
 247     tc0masked = vec_and(p1mask, (vec_u8_t)tc0vec);                                           \
 248     finaltc0 = vec_sub(finaltc0, p1mask);                       /* tc++ */                   \
 249     newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked);                                      \
 250     /*end if*/                                                                               \
 251                                                                                              \
 252     q1mask = diff_lt_altivec(q2, q0, betavec);                                               \
 253     q1mask = vec_and(q1mask, mask);                             /* if ( |q2 - q0| < beta ) */\
 254     tc0masked = vec_and(q1mask, (vec_u8_t)tc0vec);                                           \
 255     finaltc0 = vec_sub(finaltc0, q1mask);                       /* tc++ */                   \
 256     newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked);                                      \
 257     /*end if*/                                                                               \
 258                                                                                              \
 259     h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0);                                            \
 260     p1 = newp1;                                                                              \
 261     q1 = newq1;                                                                              \
 262 }
 263
 264 void x264_deblock_v_luma_altivec( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
 265 {
 266     if( (tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0 )
 267     {
 268         register vec_u8_t p2 = vec_ld(-3*stride, pix);
 269         register vec_u8_t p1 = vec_ld(-2*stride, pix);
 270         register vec_u8_t p0 = vec_ld(-1*stride, pix);
 271         register vec_u8_t q0 = vec_ld(0, pix);
 272         register vec_u8_t q1 = vec_ld(stride, pix);
 273         register vec_u8_t q2 = vec_ld(2*stride, pix);
 274         h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
 275         vec_st(p1, -2*stride, pix);
 276         vec_st(p0, -1*stride, pix);
 277         vec_st(q0, 0, pix);
 278         vec_st(q1, stride, pix);
 279     }
 280 }
 281
 282 void x264_deblock_h_luma_altivec( uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0 )
 283 {
 284
 285     register vec_u8_t line0, line1, line2, line3, line4, line5;
 286     if( (tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0 )
 287         return;
 288     PREP_LOAD;
 289     vec_u8_t _pix_ = vec_lvsl(0, pix-3);
 290     readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
 291     h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
 292     transpose4x16(line1, line2, line3, line4);
 293     write16x4(pix-2, stride, line1, line2, line3, line4);
 294 }