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 x264 project
   5 *
   6 * Authors: Guillaume Poirier <gpoirier CHEZ mplayerhq POINT hu>
   7 *          (based on code written by Graham Booker for FFmpeg)
   8 *
   9 * This program is free software; you can redistribute it and/or modify
  10 * it under the terms of the GNU General Public License as published by
  11 * the Free Software Foundation; either version 2 of the License, or
  12 * (at your option) any later version.
  13 *
  14 * This program is distributed in the hope that it will be useful,
  15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17 * GNU General Public License for more details.
  18 *
  19 * You should have received a copy of the GNU General Public License
  20 * along with this program; if not, write to the Free Software
  21 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA  02111, USA.
  22 *****************************************************************************/
  23
  24 #if defined SYS_LINUX
  25 #include <altivec.h>
  26 #endif
  27
  28 #include "common/common.h"
  29 #include "ppccommon.h"
  30
  31 #define transpose4x16(r0, r1, r2, r3) {      \
  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     DECLARE_ALIGNED(unsigned char, result[64], 16);
  52     uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst;
  53     int int_dst_stride = dst_stride/4;
  54
  55     vec_st(r0, 0, result);
  56     vec_st(r1, 16, result);
  57     vec_st(r2, 32, result);
  58     vec_st(r3, 48, result);
  59     /* FIXME: there has to be a better way!!!! */
  60     *dst_int = *src_int;
  61     *(dst_int+   int_dst_stride) = *(src_int + 1);
  62     *(dst_int+ 2*int_dst_stride) = *(src_int + 2);
  63     *(dst_int+ 3*int_dst_stride) = *(src_int + 3);
  64     *(dst_int+ 4*int_dst_stride) = *(src_int + 4);
  65     *(dst_int+ 5*int_dst_stride) = *(src_int + 5);
  66     *(dst_int+ 6*int_dst_stride) = *(src_int + 6);
  67     *(dst_int+ 7*int_dst_stride) = *(src_int + 7);
  68     *(dst_int+ 8*int_dst_stride) = *(src_int + 8);
  69     *(dst_int+ 9*int_dst_stride) = *(src_int + 9);
  70     *(dst_int+10*int_dst_stride) = *(src_int + 10);
  71     *(dst_int+11*int_dst_stride) = *(src_int + 11);
  72     *(dst_int+12*int_dst_stride) = *(src_int + 12);
  73     *(dst_int+13*int_dst_stride) = *(src_int + 13);
  74     *(dst_int+14*int_dst_stride) = *(src_int + 14);
  75     *(dst_int+15*int_dst_stride) = *(src_int + 15);
  76 }
  77
  78 /** \brief performs a 6x16 transpose of data in src, and stores it to dst
  79     \todo FIXME: see if we can't spare some vec_lvsl() by them factorizing
  80     out of unaligned_load() */
  81 #define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\
  82     register vec_u8_t r0, r1, r2, r3, r4, r5, r6, r7, r14, r15;\
  83     VEC_LOAD(src,                  r0, 16, vec_u8_t);          \
  84     VEC_LOAD(src +    src_stride,  r1, 16, vec_u8_t);          \
  85     VEC_LOAD(src +  2*src_stride,  r2, 16, vec_u8_t);          \
  86     VEC_LOAD(src +  3*src_stride,  r3, 16, vec_u8_t);          \
  87     VEC_LOAD(src +  4*src_stride,  r4, 16, vec_u8_t);          \
  88     VEC_LOAD(src +  5*src_stride,  r5, 16, vec_u8_t);          \
  89     VEC_LOAD(src +  6*src_stride,  r6, 16, vec_u8_t);          \
  90     VEC_LOAD(src +  7*src_stride,  r7, 16, vec_u8_t);          \
  91     VEC_LOAD(src + 14*src_stride, r14, 16, vec_u8_t);          \
  92     VEC_LOAD(src + 15*src_stride, r15, 16, vec_u8_t);          \
  93                                                                \
  94     VEC_LOAD(src + 8*src_stride,   r8, 16, vec_u8_t);          \
  95     VEC_LOAD(src + 9*src_stride,   r9, 16, vec_u8_t);          \
  96     VEC_LOAD(src + 10*src_stride, r10, 16, vec_u8_t);          \
  97     VEC_LOAD(src + 11*src_stride, r11, 16, vec_u8_t);          \
  98     VEC_LOAD(src + 12*src_stride, r12, 16, vec_u8_t);          \
  99     VEC_LOAD(src + 13*src_stride, r13, 16, vec_u8_t);          \
 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,
 143                                          register vec_u8_t y,
 144                                          register vec_u8_t a) {
 145
 146     register vec_u8_t diff = vec_subs(x, y);
 147     register vec_u8_t diffneg = vec_subs(y, x);
 148     register vec_u8_t o = vec_or(diff, diffneg); /* |x-y| */
 149     o = (vec_u8_t)vec_cmplt(o, a);
 150     return o;
 151 }
 152
 153 static inline vec_u8_t h264_deblock_mask ( register vec_u8_t p0,
 154                                            register vec_u8_t p1,
 155                                            register vec_u8_t q0,
 156                                            register vec_u8_t q1,
 157                                            register vec_u8_t alpha,
 158                                            register vec_u8_t beta) {
 159
 160     register vec_u8_t mask;
 161     register vec_u8_t tempmask;
 162
 163     mask = diff_lt_altivec(p0, q0, alpha);
 164     tempmask = diff_lt_altivec(p1, p0, beta);
 165     mask = vec_and(mask, tempmask);
 166     tempmask = diff_lt_altivec(q1, q0, beta);
 167     mask = vec_and(mask, tempmask);
 168
 169     return mask;
 170 }
 171
 172 // out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0)
 173 static inline vec_u8_t h264_deblock_q1(register vec_u8_t p0,
 174                                        register vec_u8_t p1,
 175                                        register vec_u8_t p2,
 176                                        register vec_u8_t q0,
 177                                        register vec_u8_t tc0) {
 178
 179     register vec_u8_t average = vec_avg(p0, q0);
 180     register vec_u8_t temp;
 181     register vec_u8_t uncliped;
 182     register vec_u8_t ones;
 183     register vec_u8_t max;
 184     register vec_u8_t min;
 185     register vec_u8_t newp1;
 186
 187     temp = vec_xor(average, p2);
 188     average = vec_avg(average, p2);     /*avg(p2, avg(p0, q0)) */
 189     ones = vec_splat_u8(1);
 190     temp = vec_and(temp, ones);         /*(p2^avg(p0, q0)) & 1 */
 191     uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */
 192     max = vec_adds(p1, tc0);
 193     min = vec_subs(p1, tc0);
 194     newp1 = vec_max(min, uncliped);
 195     newp1 = vec_min(max, newp1);
 196     return newp1;
 197 }
 198
 199 #define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) {                                         \
 200                                                                                                 \
 201     const vec_u8_t A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4));                             \
 202                                                                                                 \
 203     register vec_u8_t pq0bit = vec_xor(p0,q0);                                                  \
 204     register vec_u8_t q1minus;                                                                  \
 205     register vec_u8_t p0minus;                                                                  \
 206     register vec_u8_t stage1;                                                                   \
 207     register vec_u8_t stage2;                                                                   \
 208     register vec_u8_t vec160;                                                                   \
 209     register vec_u8_t delta;                                                                    \
 210     register vec_u8_t deltaneg;                                                                 \
 211                                                                                                 \
 212     q1minus = vec_nor(q1, q1);                /* 255 - q1 */                                    \
 213     stage1 = vec_avg(p1, q1minus);            /* (p1 - q1 + 256)>>1 */                          \
 214     stage2 = vec_sr(stage1, vec_splat_u8(1)); /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */    \
 215     p0minus = vec_nor(p0, p0);                /* 255 - p0 */                                    \
 216     stage1 = vec_avg(q0, p0minus);            /* (q0 - p0 + 256)>>1 */                          \
 217     pq0bit = vec_and(pq0bit, vec_splat_u8(1));                                                  \
 218     stage2 = vec_avg(stage2, pq0bit);         /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */\
 219     stage2 = vec_adds(stage2, stage1);        /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */ \
 220     vec160 = vec_ld(0, &A0v);                                                                   \
 221     deltaneg = vec_subs(vec160, stage2);      /* -d */                                          \
 222     delta = vec_subs(stage2, vec160);         /*  d */                                          \
 223     deltaneg = vec_min(tc0masked, deltaneg);                                                    \
 224     delta = vec_min(tc0masked, delta);                                                          \
 225     p0 = vec_subs(p0, deltaneg);                                                                \
 226     q0 = vec_subs(q0, delta);                                                                   \
 227     p0 = vec_adds(p0, delta);                                                                   \
 228     q0 = vec_adds(q0, deltaneg);                                                                \
 229 }
 230
 231 #define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) {            \
 232     DECLARE_ALIGNED(unsigned char, temp[16], 16);                                            \
 233     register vec_u8_t alphavec;                                                              \
 234     register vec_u8_t betavec;                                                               \
 235     register vec_u8_t mask;                                                                  \
 236     register vec_u8_t p1mask;                                                                \
 237     register vec_u8_t q1mask;                                                                \
 238     register vec_s8_t tc0vec;                                                                \
 239     register vec_u8_t finaltc0;                                                              \
 240     register vec_u8_t tc0masked;                                                             \
 241     register vec_u8_t newp1;                                                                 \
 242     register vec_u8_t newq1;                                                                 \
 243                                                                                              \
 244     temp[0] = alpha;                                                                         \
 245     temp[1] = beta;                                                                          \
 246     alphavec = vec_ld(0, temp);                                                              \
 247     betavec = vec_splat(alphavec, 0x1);                                                      \
 248     alphavec = vec_splat(alphavec, 0x0);                                                     \
 249     mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */            \
 250                                                                                              \
 251     *((int *)temp) = *((int *)tc0);                                                          \
 252     tc0vec = vec_ld(0, (signed char*)temp);                                                  \
 253     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 254     tc0vec = vec_mergeh(tc0vec, tc0vec);                                                     \
 255     mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1)));  /* if tc0[i] >= 0 */         \
 256     finaltc0 = vec_and((vec_u8_t)tc0vec, mask);                 /* tc = tc0 */               \
 257                                                                                              \
 258     p1mask = diff_lt_altivec(p2, p0, betavec);                                               \
 259     p1mask = vec_and(p1mask, mask);                             /* if( |p2 - p0| < beta) */  \
 260     tc0masked = vec_and(p1mask, (vec_u8_t)tc0vec);                                           \
 261     finaltc0 = vec_sub(finaltc0, p1mask);                       /* tc++ */                   \
 262     newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked);                                      \
 263     /*end if*/                                                                               \
 264                                                                                              \
 265     q1mask = diff_lt_altivec(q2, q0, betavec);                                               \
 266     q1mask = vec_and(q1mask, mask);                             /* if ( |q2 - q0| < beta ) */\
 267     tc0masked = vec_and(q1mask, (vec_u8_t)tc0vec);                                           \
 268     finaltc0 = vec_sub(finaltc0, q1mask);                       /* tc++ */                   \
 269     newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked);                                      \
 270     /*end if*/                                                                               \
 271                                                                                              \
 272     h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0);                                            \
 273     p1 = newp1;                                                                              \
 274     q1 = newq1;                                                                              \
 275 }
 276
 277 void x264_deblock_v_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
 278
 279     if((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) {
 280         register vec_u8_t p2 = vec_ld(-3*stride, pix);
 281         register vec_u8_t p1 = vec_ld(-2*stride, pix);
 282         register vec_u8_t p0 = vec_ld(-1*stride, pix);
 283         register vec_u8_t q0 = vec_ld(0, pix);
 284         register vec_u8_t q1 = vec_ld(stride, pix);
 285         register vec_u8_t q2 = vec_ld(2*stride, pix);
 286         h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0);
 287         vec_st(p1, -2*stride, pix);
 288         vec_st(p0, -1*stride, pix);
 289         vec_st(q0, 0, pix);
 290         vec_st(q1, stride, pix);
 291     }
 292 }
 293
 294 void x264_deblock_h_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) {
 295
 296     register vec_u8_t line0, line1, line2, line3, line4, line5;
 297     if((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0)
 298         return;
 299     PREP_LOAD;
 300     readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5);
 301     h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0);
 302     transpose4x16(line1, line2, line3, line4);
 303     write16x4(pix-2, stride, line1, line2, line3, line4);
 304 }