1/* 2 * Copyright (c) 2004 Romain Dolbeau <romain@dolbeau.org> 3 * 4 * This file is part of FFmpeg. 5 * 6 * FFmpeg is free software; you can redistribute it and/or 7 * modify it under the terms of the GNU Lesser General Public 8 * License as published by the Free Software Foundation; either 9 * version 2.1 of the License, or (at your option) any later version. 10 * 11 * FFmpeg is distributed in the hope that it will be useful, 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 14 * Lesser General Public License for more details. 15 * 16 * You should have received a copy of the GNU Lesser General Public 17 * License along with FFmpeg; if not, write to the Free Software 18 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA 19 */ 20 21#include "config.h" 22#include "libavutil/attributes.h" 23#include "libavutil/cpu.h" 24#include "libavutil/intreadwrite.h" 25#include "libavutil/ppc/cpu.h" 26#include "libavutil/ppc/types_altivec.h" 27#include "libavutil/ppc/util_altivec.h" 28#include "libavcodec/h264data.h" 29#include "libavcodec/h264dsp.h" 30 31#if HAVE_ALTIVEC 32 33/**************************************************************************** 34 * IDCT transform: 35 ****************************************************************************/ 36 37#define VEC_1D_DCT(vb0,vb1,vb2,vb3,va0,va1,va2,va3) \ 38 /* 1st stage */ \ 39 vz0 = vec_add(vb0,vb2); /* temp[0] = Y[0] + Y[2] */ \ 40 vz1 = vec_sub(vb0,vb2); /* temp[1] = Y[0] - Y[2] */ \ 41 vz2 = vec_sra(vb1,vec_splat_u16(1)); \ 42 vz2 = vec_sub(vz2,vb3); /* temp[2] = Y[1].1/2 - Y[3] */ \ 43 vz3 = vec_sra(vb3,vec_splat_u16(1)); \ 44 vz3 = vec_add(vb1,vz3); /* temp[3] = Y[1] + Y[3].1/2 */ \ 45 /* 2nd stage: output */ \ 46 va0 = vec_add(vz0,vz3); /* x[0] = temp[0] + temp[3] */ \ 47 va1 = vec_add(vz1,vz2); /* x[1] = temp[1] + temp[2] */ \ 48 va2 = vec_sub(vz1,vz2); /* x[2] = temp[1] - temp[2] */ \ 49 va3 = vec_sub(vz0,vz3) /* x[3] = temp[0] - temp[3] */ 50 51#define VEC_TRANSPOSE_4(a0,a1,a2,a3,b0,b1,b2,b3) \ 52 b0 = vec_mergeh( a0, a0 ); \ 53 b1 = vec_mergeh( a1, a0 ); \ 54 b2 = vec_mergeh( a2, a0 ); \ 55 b3 = vec_mergeh( a3, a0 ); \ 56 a0 = vec_mergeh( b0, b2 ); \ 57 a1 = vec_mergel( b0, b2 ); \ 58 a2 = vec_mergeh( b1, b3 ); \ 59 a3 = vec_mergel( b1, b3 ); \ 60 b0 = vec_mergeh( a0, a2 ); \ 61 b1 = vec_mergel( a0, a2 ); \ 62 b2 = vec_mergeh( a1, a3 ); \ 63 b3 = vec_mergel( a1, a3 ) 64 65#define VEC_LOAD_U8_ADD_S16_STORE_U8(va) \ 66 vdst_orig = vec_ld(0, dst); \ 67 vdst = vec_perm(vdst_orig, zero_u8v, vdst_mask); \ 68 vdst_ss = (vec_s16) vec_mergeh(zero_u8v, vdst); \ 69 va = vec_add(va, vdst_ss); \ 70 va_u8 = vec_packsu(va, zero_s16v); \ 71 va_u32 = vec_splat((vec_u32)va_u8, 0); \ 72 vec_ste(va_u32, element, (uint32_t*)dst); 73 74static void h264_idct_add_altivec(uint8_t *dst, int16_t *block, int stride) 75{ 76 vec_s16 va0, va1, va2, va3; 77 vec_s16 vz0, vz1, vz2, vz3; 78 vec_s16 vtmp0, vtmp1, vtmp2, vtmp3; 79 vec_u8 va_u8; 80 vec_u32 va_u32; 81 vec_s16 vdst_ss; 82 const vec_u16 v6us = vec_splat_u16(6); 83 vec_u8 vdst, vdst_orig; 84 vec_u8 vdst_mask = vec_lvsl(0, dst); 85 int element = ((unsigned long)dst & 0xf) >> 2; 86 LOAD_ZERO; 87 88 block[0] += 32; /* add 32 as a DC-level for rounding */ 89 90 vtmp0 = vec_ld(0,block); 91 vtmp1 = vec_sld(vtmp0, vtmp0, 8); 92 vtmp2 = vec_ld(16,block); 93 vtmp3 = vec_sld(vtmp2, vtmp2, 8); 94 memset(block, 0, 16 * sizeof(int16_t)); 95 96 VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3); 97 VEC_TRANSPOSE_4(va0,va1,va2,va3,vtmp0,vtmp1,vtmp2,vtmp3); 98 VEC_1D_DCT(vtmp0,vtmp1,vtmp2,vtmp3,va0,va1,va2,va3); 99 100 va0 = vec_sra(va0,v6us); 101 va1 = vec_sra(va1,v6us); 102 va2 = vec_sra(va2,v6us); 103 va3 = vec_sra(va3,v6us); 104 105 VEC_LOAD_U8_ADD_S16_STORE_U8(va0); 106 dst += stride; 107 VEC_LOAD_U8_ADD_S16_STORE_U8(va1); 108 dst += stride; 109 VEC_LOAD_U8_ADD_S16_STORE_U8(va2); 110 dst += stride; 111 VEC_LOAD_U8_ADD_S16_STORE_U8(va3); 112} 113 114#define IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7, d0, d1, d2, d3, d4, d5, d6, d7) {\ 115 /* a0 = SRC(0) + SRC(4); */ \ 116 vec_s16 a0v = vec_add(s0, s4); \ 117 /* a2 = SRC(0) - SRC(4); */ \ 118 vec_s16 a2v = vec_sub(s0, s4); \ 119 /* a4 = (SRC(2)>>1) - SRC(6); */ \ 120 vec_s16 a4v = vec_sub(vec_sra(s2, onev), s6); \ 121 /* a6 = (SRC(6)>>1) + SRC(2); */ \ 122 vec_s16 a6v = vec_add(vec_sra(s6, onev), s2); \ 123 /* b0 = a0 + a6; */ \ 124 vec_s16 b0v = vec_add(a0v, a6v); \ 125 /* b2 = a2 + a4; */ \ 126 vec_s16 b2v = vec_add(a2v, a4v); \ 127 /* b4 = a2 - a4; */ \ 128 vec_s16 b4v = vec_sub(a2v, a4v); \ 129 /* b6 = a0 - a6; */ \ 130 vec_s16 b6v = vec_sub(a0v, a6v); \ 131 /* a1 = SRC(5) - SRC(3) - SRC(7) - (SRC(7)>>1); */ \ 132 /* a1 = (SRC(5)-SRC(3)) - (SRC(7) + (SRC(7)>>1)); */ \ 133 vec_s16 a1v = vec_sub( vec_sub(s5, s3), vec_add(s7, vec_sra(s7, onev)) ); \ 134 /* a3 = SRC(7) + SRC(1) - SRC(3) - (SRC(3)>>1); */ \ 135 /* a3 = (SRC(7)+SRC(1)) - (SRC(3) + (SRC(3)>>1)); */ \ 136 vec_s16 a3v = vec_sub( vec_add(s7, s1), vec_add(s3, vec_sra(s3, onev)) );\ 137 /* a5 = SRC(7) - SRC(1) + SRC(5) + (SRC(5)>>1); */ \ 138 /* a5 = (SRC(7)-SRC(1)) + SRC(5) + (SRC(5)>>1); */ \ 139 vec_s16 a5v = vec_add( vec_sub(s7, s1), vec_add(s5, vec_sra(s5, onev)) );\ 140 /* a7 = SRC(5)+SRC(3) + SRC(1) + (SRC(1)>>1); */ \ 141 vec_s16 a7v = vec_add( vec_add(s5, s3), vec_add(s1, vec_sra(s1, onev)) );\ 142 /* b1 = (a7>>2) + a1; */ \ 143 vec_s16 b1v = vec_add( vec_sra(a7v, twov), a1v); \ 144 /* b3 = a3 + (a5>>2); */ \ 145 vec_s16 b3v = vec_add(a3v, vec_sra(a5v, twov)); \ 146 /* b5 = (a3>>2) - a5; */ \ 147 vec_s16 b5v = vec_sub( vec_sra(a3v, twov), a5v); \ 148 /* b7 = a7 - (a1>>2); */ \ 149 vec_s16 b7v = vec_sub( a7v, vec_sra(a1v, twov)); \ 150 /* DST(0, b0 + b7); */ \ 151 d0 = vec_add(b0v, b7v); \ 152 /* DST(1, b2 + b5); */ \ 153 d1 = vec_add(b2v, b5v); \ 154 /* DST(2, b4 + b3); */ \ 155 d2 = vec_add(b4v, b3v); \ 156 /* DST(3, b6 + b1); */ \ 157 d3 = vec_add(b6v, b1v); \ 158 /* DST(4, b6 - b1); */ \ 159 d4 = vec_sub(b6v, b1v); \ 160 /* DST(5, b4 - b3); */ \ 161 d5 = vec_sub(b4v, b3v); \ 162 /* DST(6, b2 - b5); */ \ 163 d6 = vec_sub(b2v, b5v); \ 164 /* DST(7, b0 - b7); */ \ 165 d7 = vec_sub(b0v, b7v); \ 166} 167 168#define ALTIVEC_STORE_SUM_CLIP(dest, idctv, perm_ldv, perm_stv, sel) { \ 169 /* unaligned load */ \ 170 vec_u8 hv = vec_ld( 0, dest ); \ 171 vec_u8 lv = vec_ld( 7, dest ); \ 172 vec_u8 dstv = vec_perm( hv, lv, (vec_u8)perm_ldv ); \ 173 vec_s16 idct_sh6 = vec_sra(idctv, sixv); \ 174 vec_u16 dst16 = (vec_u16)vec_mergeh(zero_u8v, dstv); \ 175 vec_s16 idstsum = vec_adds(idct_sh6, (vec_s16)dst16); \ 176 vec_u8 idstsum8 = vec_packsu(zero_s16v, idstsum); \ 177 vec_u8 edgehv; \ 178 /* unaligned store */ \ 179 vec_u8 bodyv = vec_perm( idstsum8, idstsum8, perm_stv );\ 180 vec_u8 edgelv = vec_perm( sel, zero_u8v, perm_stv ); \ 181 lv = vec_sel( lv, bodyv, edgelv ); \ 182 vec_st( lv, 7, dest ); \ 183 hv = vec_ld( 0, dest ); \ 184 edgehv = vec_perm( zero_u8v, sel, perm_stv ); \ 185 hv = vec_sel( hv, bodyv, edgehv ); \ 186 vec_st( hv, 0, dest ); \ 187 } 188 189static void h264_idct8_add_altivec(uint8_t *dst, int16_t *dct, int stride) 190{ 191 vec_s16 s0, s1, s2, s3, s4, s5, s6, s7; 192 vec_s16 d0, d1, d2, d3, d4, d5, d6, d7; 193 vec_s16 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7; 194 195 vec_u8 perm_ldv = vec_lvsl(0, dst); 196 vec_u8 perm_stv = vec_lvsr(8, dst); 197 198 const vec_u16 onev = vec_splat_u16(1); 199 const vec_u16 twov = vec_splat_u16(2); 200 const vec_u16 sixv = vec_splat_u16(6); 201 202 const vec_u8 sel = (vec_u8) {0,0,0,0,0,0,0,0,-1,-1,-1,-1,-1,-1,-1,-1}; 203 LOAD_ZERO; 204 205 dct[0] += 32; // rounding for the >>6 at the end 206 207 s0 = vec_ld(0x00, (int16_t*)dct); 208 s1 = vec_ld(0x10, (int16_t*)dct); 209 s2 = vec_ld(0x20, (int16_t*)dct); 210 s3 = vec_ld(0x30, (int16_t*)dct); 211 s4 = vec_ld(0x40, (int16_t*)dct); 212 s5 = vec_ld(0x50, (int16_t*)dct); 213 s6 = vec_ld(0x60, (int16_t*)dct); 214 s7 = vec_ld(0x70, (int16_t*)dct); 215 memset(dct, 0, 64 * sizeof(int16_t)); 216 217 IDCT8_1D_ALTIVEC(s0, s1, s2, s3, s4, s5, s6, s7, 218 d0, d1, d2, d3, d4, d5, d6, d7); 219 220 TRANSPOSE8( d0, d1, d2, d3, d4, d5, d6, d7 ); 221 222 IDCT8_1D_ALTIVEC(d0, d1, d2, d3, d4, d5, d6, d7, 223 idct0, idct1, idct2, idct3, idct4, idct5, idct6, idct7); 224 225 ALTIVEC_STORE_SUM_CLIP(&dst[0*stride], idct0, perm_ldv, perm_stv, sel); 226 ALTIVEC_STORE_SUM_CLIP(&dst[1*stride], idct1, perm_ldv, perm_stv, sel); 227 ALTIVEC_STORE_SUM_CLIP(&dst[2*stride], idct2, perm_ldv, perm_stv, sel); 228 ALTIVEC_STORE_SUM_CLIP(&dst[3*stride], idct3, perm_ldv, perm_stv, sel); 229 ALTIVEC_STORE_SUM_CLIP(&dst[4*stride], idct4, perm_ldv, perm_stv, sel); 230 ALTIVEC_STORE_SUM_CLIP(&dst[5*stride], idct5, perm_ldv, perm_stv, sel); 231 ALTIVEC_STORE_SUM_CLIP(&dst[6*stride], idct6, perm_ldv, perm_stv, sel); 232 ALTIVEC_STORE_SUM_CLIP(&dst[7*stride], idct7, perm_ldv, perm_stv, sel); 233} 234 235static av_always_inline void h264_idct_dc_add_internal(uint8_t *dst, int16_t *block, int stride, int size) 236{ 237 vec_s16 dc16; 238 vec_u8 dcplus, dcminus, v0, v1, v2, v3, aligner; 239 LOAD_ZERO; 240 DECLARE_ALIGNED(16, int, dc); 241 int i; 242 243 dc = (block[0] + 32) >> 6; 244 block[0] = 0; 245 dc16 = vec_splat((vec_s16) vec_lde(0, &dc), 1); 246 247 if (size == 4) 248 dc16 = vec_sld(dc16, zero_s16v, 8); 249 dcplus = vec_packsu(dc16, zero_s16v); 250 dcminus = vec_packsu(vec_sub(zero_s16v, dc16), zero_s16v); 251 252 aligner = vec_lvsr(0, dst); 253 dcplus = vec_perm(dcplus, dcplus, aligner); 254 dcminus = vec_perm(dcminus, dcminus, aligner); 255 256 for (i = 0; i < size; i += 4) { 257 v0 = vec_ld(0, dst+0*stride); 258 v1 = vec_ld(0, dst+1*stride); 259 v2 = vec_ld(0, dst+2*stride); 260 v3 = vec_ld(0, dst+3*stride); 261 262 v0 = vec_adds(v0, dcplus); 263 v1 = vec_adds(v1, dcplus); 264 v2 = vec_adds(v2, dcplus); 265 v3 = vec_adds(v3, dcplus); 266 267 v0 = vec_subs(v0, dcminus); 268 v1 = vec_subs(v1, dcminus); 269 v2 = vec_subs(v2, dcminus); 270 v3 = vec_subs(v3, dcminus); 271 272 vec_st(v0, 0, dst+0*stride); 273 vec_st(v1, 0, dst+1*stride); 274 vec_st(v2, 0, dst+2*stride); 275 vec_st(v3, 0, dst+3*stride); 276 277 dst += 4*stride; 278 } 279} 280 281static void h264_idct_dc_add_altivec(uint8_t *dst, int16_t *block, int stride) 282{ 283 h264_idct_dc_add_internal(dst, block, stride, 4); 284} 285 286static void h264_idct8_dc_add_altivec(uint8_t *dst, int16_t *block, int stride) 287{ 288 h264_idct_dc_add_internal(dst, block, stride, 8); 289} 290 291static void h264_idct_add16_altivec(uint8_t *dst, const int *block_offset, 292 int16_t *block, int stride, 293 const uint8_t nnzc[15 * 8]) 294{ 295 int i; 296 for(i=0; i<16; i++){ 297 int nnz = nnzc[ scan8[i] ]; 298 if(nnz){ 299 if(nnz==1 && block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride); 300 else h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride); 301 } 302 } 303} 304 305static void h264_idct_add16intra_altivec(uint8_t *dst, const int *block_offset, 306 int16_t *block, int stride, 307 const uint8_t nnzc[15 * 8]) 308{ 309 int i; 310 for(i=0; i<16; i++){ 311 if(nnzc[ scan8[i] ]) h264_idct_add_altivec(dst + block_offset[i], block + i*16, stride); 312 else if(block[i*16]) h264_idct_dc_add_altivec(dst + block_offset[i], block + i*16, stride); 313 } 314} 315 316static void h264_idct8_add4_altivec(uint8_t *dst, const int *block_offset, 317 int16_t *block, int stride, 318 const uint8_t nnzc[15 * 8]) 319{ 320 int i; 321 for(i=0; i<16; i+=4){ 322 int nnz = nnzc[ scan8[i] ]; 323 if(nnz){ 324 if(nnz==1 && block[i*16]) h264_idct8_dc_add_altivec(dst + block_offset[i], block + i*16, stride); 325 else h264_idct8_add_altivec(dst + block_offset[i], block + i*16, stride); 326 } 327 } 328} 329 330static void h264_idct_add8_altivec(uint8_t **dest, const int *block_offset, 331 int16_t *block, int stride, 332 const uint8_t nnzc[15 * 8]) 333{ 334 int i, j; 335 for (j = 1; j < 3; j++) { 336 for(i = j * 16; i < j * 16 + 4; i++){ 337 if(nnzc[ scan8[i] ]) 338 h264_idct_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride); 339 else if(block[i*16]) 340 h264_idct_dc_add_altivec(dest[j-1] + block_offset[i], block + i*16, stride); 341 } 342 } 343} 344 345#define transpose4x16(r0, r1, r2, r3) { \ 346 register vec_u8 r4; \ 347 register vec_u8 r5; \ 348 register vec_u8 r6; \ 349 register vec_u8 r7; \ 350 \ 351 r4 = vec_mergeh(r0, r2); /*0, 2 set 0*/ \ 352 r5 = vec_mergel(r0, r2); /*0, 2 set 1*/ \ 353 r6 = vec_mergeh(r1, r3); /*1, 3 set 0*/ \ 354 r7 = vec_mergel(r1, r3); /*1, 3 set 1*/ \ 355 \ 356 r0 = vec_mergeh(r4, r6); /*all set 0*/ \ 357 r1 = vec_mergel(r4, r6); /*all set 1*/ \ 358 r2 = vec_mergeh(r5, r7); /*all set 2*/ \ 359 r3 = vec_mergel(r5, r7); /*all set 3*/ \ 360} 361 362static inline void write16x4(uint8_t *dst, int dst_stride, 363 register vec_u8 r0, register vec_u8 r1, 364 register vec_u8 r2, register vec_u8 r3) { 365 DECLARE_ALIGNED(16, unsigned char, result)[64]; 366 uint32_t *src_int = (uint32_t *)result, *dst_int = (uint32_t *)dst; 367 int int_dst_stride = dst_stride/4; 368 369 vec_st(r0, 0, result); 370 vec_st(r1, 16, result); 371 vec_st(r2, 32, result); 372 vec_st(r3, 48, result); 373 /* FIXME: there has to be a better way!!!! */ 374 *dst_int = *src_int; 375 *(dst_int+ int_dst_stride) = *(src_int + 1); 376 *(dst_int+ 2*int_dst_stride) = *(src_int + 2); 377 *(dst_int+ 3*int_dst_stride) = *(src_int + 3); 378 *(dst_int+ 4*int_dst_stride) = *(src_int + 4); 379 *(dst_int+ 5*int_dst_stride) = *(src_int + 5); 380 *(dst_int+ 6*int_dst_stride) = *(src_int + 6); 381 *(dst_int+ 7*int_dst_stride) = *(src_int + 7); 382 *(dst_int+ 8*int_dst_stride) = *(src_int + 8); 383 *(dst_int+ 9*int_dst_stride) = *(src_int + 9); 384 *(dst_int+10*int_dst_stride) = *(src_int + 10); 385 *(dst_int+11*int_dst_stride) = *(src_int + 11); 386 *(dst_int+12*int_dst_stride) = *(src_int + 12); 387 *(dst_int+13*int_dst_stride) = *(src_int + 13); 388 *(dst_int+14*int_dst_stride) = *(src_int + 14); 389 *(dst_int+15*int_dst_stride) = *(src_int + 15); 390} 391 392/** @brief performs a 6x16 transpose of data in src, and stores it to dst 393 @todo FIXME: see if we can't spare some vec_lvsl() by them factorizing 394 out of unaligned_load() */ 395#define readAndTranspose16x6(src, src_stride, r8, r9, r10, r11, r12, r13) {\ 396 register vec_u8 r0 = unaligned_load(0, src); \ 397 register vec_u8 r1 = unaligned_load( src_stride, src); \ 398 register vec_u8 r2 = unaligned_load(2* src_stride, src); \ 399 register vec_u8 r3 = unaligned_load(3* src_stride, src); \ 400 register vec_u8 r4 = unaligned_load(4* src_stride, src); \ 401 register vec_u8 r5 = unaligned_load(5* src_stride, src); \ 402 register vec_u8 r6 = unaligned_load(6* src_stride, src); \ 403 register vec_u8 r7 = unaligned_load(7* src_stride, src); \ 404 register vec_u8 r14 = unaligned_load(14*src_stride, src); \ 405 register vec_u8 r15 = unaligned_load(15*src_stride, src); \ 406 \ 407 r8 = unaligned_load( 8*src_stride, src); \ 408 r9 = unaligned_load( 9*src_stride, src); \ 409 r10 = unaligned_load(10*src_stride, src); \ 410 r11 = unaligned_load(11*src_stride, src); \ 411 r12 = unaligned_load(12*src_stride, src); \ 412 r13 = unaligned_load(13*src_stride, src); \ 413 \ 414 /*Merge first pairs*/ \ 415 r0 = vec_mergeh(r0, r8); /*0, 8*/ \ 416 r1 = vec_mergeh(r1, r9); /*1, 9*/ \ 417 r2 = vec_mergeh(r2, r10); /*2,10*/ \ 418 r3 = vec_mergeh(r3, r11); /*3,11*/ \ 419 r4 = vec_mergeh(r4, r12); /*4,12*/ \ 420 r5 = vec_mergeh(r5, r13); /*5,13*/ \ 421 r6 = vec_mergeh(r6, r14); /*6,14*/ \ 422 r7 = vec_mergeh(r7, r15); /*7,15*/ \ 423 \ 424 /*Merge second pairs*/ \ 425 r8 = vec_mergeh(r0, r4); /*0,4, 8,12 set 0*/ \ 426 r9 = vec_mergel(r0, r4); /*0,4, 8,12 set 1*/ \ 427 r10 = vec_mergeh(r1, r5); /*1,5, 9,13 set 0*/ \ 428 r11 = vec_mergel(r1, r5); /*1,5, 9,13 set 1*/ \ 429 r12 = vec_mergeh(r2, r6); /*2,6,10,14 set 0*/ \ 430 r13 = vec_mergel(r2, r6); /*2,6,10,14 set 1*/ \ 431 r14 = vec_mergeh(r3, r7); /*3,7,11,15 set 0*/ \ 432 r15 = vec_mergel(r3, r7); /*3,7,11,15 set 1*/ \ 433 \ 434 /*Third merge*/ \ 435 r0 = vec_mergeh(r8, r12); /*0,2,4,6,8,10,12,14 set 0*/ \ 436 r1 = vec_mergel(r8, r12); /*0,2,4,6,8,10,12,14 set 1*/ \ 437 r2 = vec_mergeh(r9, r13); /*0,2,4,6,8,10,12,14 set 2*/ \ 438 r4 = vec_mergeh(r10, r14); /*1,3,5,7,9,11,13,15 set 0*/ \ 439 r5 = vec_mergel(r10, r14); /*1,3,5,7,9,11,13,15 set 1*/ \ 440 r6 = vec_mergeh(r11, r15); /*1,3,5,7,9,11,13,15 set 2*/ \ 441 /* Don't need to compute 3 and 7*/ \ 442 \ 443 /*Final merge*/ \ 444 r8 = vec_mergeh(r0, r4); /*all set 0*/ \ 445 r9 = vec_mergel(r0, r4); /*all set 1*/ \ 446 r10 = vec_mergeh(r1, r5); /*all set 2*/ \ 447 r11 = vec_mergel(r1, r5); /*all set 3*/ \ 448 r12 = vec_mergeh(r2, r6); /*all set 4*/ \ 449 r13 = vec_mergel(r2, r6); /*all set 5*/ \ 450 /* Don't need to compute 14 and 15*/ \ 451 \ 452} 453 454// out: o = |x-y| < a 455static inline vec_u8 diff_lt_altivec ( register vec_u8 x, 456 register vec_u8 y, 457 register vec_u8 a) { 458 459 register vec_u8 diff = vec_subs(x, y); 460 register vec_u8 diffneg = vec_subs(y, x); 461 register vec_u8 o = vec_or(diff, diffneg); /* |x-y| */ 462 o = (vec_u8)vec_cmplt(o, a); 463 return o; 464} 465 466static inline vec_u8 h264_deblock_mask ( register vec_u8 p0, 467 register vec_u8 p1, 468 register vec_u8 q0, 469 register vec_u8 q1, 470 register vec_u8 alpha, 471 register vec_u8 beta) { 472 473 register vec_u8 mask; 474 register vec_u8 tempmask; 475 476 mask = diff_lt_altivec(p0, q0, alpha); 477 tempmask = diff_lt_altivec(p1, p0, beta); 478 mask = vec_and(mask, tempmask); 479 tempmask = diff_lt_altivec(q1, q0, beta); 480 mask = vec_and(mask, tempmask); 481 482 return mask; 483} 484 485// out: newp1 = clip((p2 + ((p0 + q0 + 1) >> 1)) >> 1, p1-tc0, p1+tc0) 486static inline vec_u8 h264_deblock_q1(register vec_u8 p0, 487 register vec_u8 p1, 488 register vec_u8 p2, 489 register vec_u8 q0, 490 register vec_u8 tc0) { 491 492 register vec_u8 average = vec_avg(p0, q0); 493 register vec_u8 temp; 494 register vec_u8 uncliped; 495 register vec_u8 ones; 496 register vec_u8 max; 497 register vec_u8 min; 498 register vec_u8 newp1; 499 500 temp = vec_xor(average, p2); 501 average = vec_avg(average, p2); /*avg(p2, avg(p0, q0)) */ 502 ones = vec_splat_u8(1); 503 temp = vec_and(temp, ones); /*(p2^avg(p0, q0)) & 1 */ 504 uncliped = vec_subs(average, temp); /*(p2+((p0+q0+1)>>1))>>1 */ 505 max = vec_adds(p1, tc0); 506 min = vec_subs(p1, tc0); 507 newp1 = vec_max(min, uncliped); 508 newp1 = vec_min(max, newp1); 509 return newp1; 510} 511 512#define h264_deblock_p0_q0(p0, p1, q0, q1, tc0masked) { \ 513 \ 514 const vec_u8 A0v = vec_sl(vec_splat_u8(10), vec_splat_u8(4)); \ 515 \ 516 register vec_u8 pq0bit = vec_xor(p0,q0); \ 517 register vec_u8 q1minus; \ 518 register vec_u8 p0minus; \ 519 register vec_u8 stage1; \ 520 register vec_u8 stage2; \ 521 register vec_u8 vec160; \ 522 register vec_u8 delta; \ 523 register vec_u8 deltaneg; \ 524 \ 525 q1minus = vec_nor(q1, q1); /* 255 - q1 */ \ 526 stage1 = vec_avg(p1, q1minus); /* (p1 - q1 + 256)>>1 */ \ 527 stage2 = vec_sr(stage1, vec_splat_u8(1)); /* (p1 - q1 + 256)>>2 = 64 + (p1 - q1) >> 2 */ \ 528 p0minus = vec_nor(p0, p0); /* 255 - p0 */ \ 529 stage1 = vec_avg(q0, p0minus); /* (q0 - p0 + 256)>>1 */ \ 530 pq0bit = vec_and(pq0bit, vec_splat_u8(1)); \ 531 stage2 = vec_avg(stage2, pq0bit); /* 32 + ((q0 - p0)&1 + (p1 - q1) >> 2 + 1) >> 1 */ \ 532 stage2 = vec_adds(stage2, stage1); /* 160 + ((p0 - q0) + (p1 - q1) >> 2 + 1) >> 1 */ \ 533 vec160 = vec_ld(0, &A0v); \ 534 deltaneg = vec_subs(vec160, stage2); /* -d */ \ 535 delta = vec_subs(stage2, vec160); /* d */ \ 536 deltaneg = vec_min(tc0masked, deltaneg); \ 537 delta = vec_min(tc0masked, delta); \ 538 p0 = vec_subs(p0, deltaneg); \ 539 q0 = vec_subs(q0, delta); \ 540 p0 = vec_adds(p0, delta); \ 541 q0 = vec_adds(q0, deltaneg); \ 542} 543 544#define h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0) { \ 545 DECLARE_ALIGNED(16, unsigned char, temp)[16]; \ 546 register vec_u8 alphavec; \ 547 register vec_u8 betavec; \ 548 register vec_u8 mask; \ 549 register vec_u8 p1mask; \ 550 register vec_u8 q1mask; \ 551 register vector signed char tc0vec; \ 552 register vec_u8 finaltc0; \ 553 register vec_u8 tc0masked; \ 554 register vec_u8 newp1; \ 555 register vec_u8 newq1; \ 556 \ 557 temp[0] = alpha; \ 558 temp[1] = beta; \ 559 alphavec = vec_ld(0, temp); \ 560 betavec = vec_splat(alphavec, 0x1); \ 561 alphavec = vec_splat(alphavec, 0x0); \ 562 mask = h264_deblock_mask(p0, p1, q0, q1, alphavec, betavec); /*if in block */ \ 563 \ 564 AV_COPY32(temp, tc0); \ 565 tc0vec = vec_ld(0, (signed char*)temp); \ 566 tc0vec = vec_mergeh(tc0vec, tc0vec); \ 567 tc0vec = vec_mergeh(tc0vec, tc0vec); \ 568 mask = vec_and(mask, vec_cmpgt(tc0vec, vec_splat_s8(-1))); /* if tc0[i] >= 0 */ \ 569 finaltc0 = vec_and((vec_u8)tc0vec, mask); /* tc = tc0 */ \ 570 \ 571 p1mask = diff_lt_altivec(p2, p0, betavec); \ 572 p1mask = vec_and(p1mask, mask); /* if ( |p2 - p0| < beta) */ \ 573 tc0masked = vec_and(p1mask, (vec_u8)tc0vec); \ 574 finaltc0 = vec_sub(finaltc0, p1mask); /* tc++ */ \ 575 newp1 = h264_deblock_q1(p0, p1, p2, q0, tc0masked); \ 576 /*end if*/ \ 577 \ 578 q1mask = diff_lt_altivec(q2, q0, betavec); \ 579 q1mask = vec_and(q1mask, mask); /* if ( |q2 - q0| < beta ) */\ 580 tc0masked = vec_and(q1mask, (vec_u8)tc0vec); \ 581 finaltc0 = vec_sub(finaltc0, q1mask); /* tc++ */ \ 582 newq1 = h264_deblock_q1(p0, q1, q2, q0, tc0masked); \ 583 /*end if*/ \ 584 \ 585 h264_deblock_p0_q0(p0, p1, q0, q1, finaltc0); \ 586 p1 = newp1; \ 587 q1 = newq1; \ 588} 589 590static void h264_v_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) { 591 592 if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) >= 0) { 593 register vec_u8 p2 = vec_ld(-3*stride, pix); 594 register vec_u8 p1 = vec_ld(-2*stride, pix); 595 register vec_u8 p0 = vec_ld(-1*stride, pix); 596 register vec_u8 q0 = vec_ld(0, pix); 597 register vec_u8 q1 = vec_ld(stride, pix); 598 register vec_u8 q2 = vec_ld(2*stride, pix); 599 h264_loop_filter_luma_altivec(p2, p1, p0, q0, q1, q2, alpha, beta, tc0); 600 vec_st(p1, -2*stride, pix); 601 vec_st(p0, -1*stride, pix); 602 vec_st(q0, 0, pix); 603 vec_st(q1, stride, pix); 604 } 605} 606 607static void h264_h_loop_filter_luma_altivec(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0) { 608 609 register vec_u8 line0, line1, line2, line3, line4, line5; 610 if ((tc0[0] & tc0[1] & tc0[2] & tc0[3]) < 0) 611 return; 612 readAndTranspose16x6(pix-3, stride, line0, line1, line2, line3, line4, line5); 613 h264_loop_filter_luma_altivec(line0, line1, line2, line3, line4, line5, alpha, beta, tc0); 614 transpose4x16(line1, line2, line3, line4); 615 write16x4(pix-2, stride, line1, line2, line3, line4); 616} 617 618static av_always_inline 619void weight_h264_W_altivec(uint8_t *block, int stride, int height, 620 int log2_denom, int weight, int offset, int w) 621{ 622 int y, aligned; 623 vec_u8 vblock; 624 vec_s16 vtemp, vweight, voffset, v0, v1; 625 vec_u16 vlog2_denom; 626 DECLARE_ALIGNED(16, int32_t, temp)[4]; 627 LOAD_ZERO; 628 629 offset <<= log2_denom; 630 if(log2_denom) offset += 1<<(log2_denom-1); 631 temp[0] = log2_denom; 632 temp[1] = weight; 633 temp[2] = offset; 634 635 vtemp = (vec_s16)vec_ld(0, temp); 636 vlog2_denom = (vec_u16)vec_splat(vtemp, 1); 637 vweight = vec_splat(vtemp, 3); 638 voffset = vec_splat(vtemp, 5); 639 aligned = !((unsigned long)block & 0xf); 640 641 for (y = 0; y < height; y++) { 642 vblock = vec_ld(0, block); 643 644 v0 = (vec_s16)vec_mergeh(zero_u8v, vblock); 645 v1 = (vec_s16)vec_mergel(zero_u8v, vblock); 646 647 if (w == 16 || aligned) { 648 v0 = vec_mladd(v0, vweight, zero_s16v); 649 v0 = vec_adds(v0, voffset); 650 v0 = vec_sra(v0, vlog2_denom); 651 } 652 if (w == 16 || !aligned) { 653 v1 = vec_mladd(v1, vweight, zero_s16v); 654 v1 = vec_adds(v1, voffset); 655 v1 = vec_sra(v1, vlog2_denom); 656 } 657 vblock = vec_packsu(v0, v1); 658 vec_st(vblock, 0, block); 659 660 block += stride; 661 } 662} 663 664static av_always_inline 665void biweight_h264_W_altivec(uint8_t *dst, uint8_t *src, int stride, int height, 666 int log2_denom, int weightd, int weights, int offset, int w) 667{ 668 int y, dst_aligned, src_aligned; 669 vec_u8 vsrc, vdst; 670 vec_s16 vtemp, vweights, vweightd, voffset, v0, v1, v2, v3; 671 vec_u16 vlog2_denom; 672 DECLARE_ALIGNED(16, int32_t, temp)[4]; 673 LOAD_ZERO; 674 675 offset = ((offset + 1) | 1) << log2_denom; 676 temp[0] = log2_denom+1; 677 temp[1] = weights; 678 temp[2] = weightd; 679 temp[3] = offset; 680 681 vtemp = (vec_s16)vec_ld(0, temp); 682 vlog2_denom = (vec_u16)vec_splat(vtemp, 1); 683 vweights = vec_splat(vtemp, 3); 684 vweightd = vec_splat(vtemp, 5); 685 voffset = vec_splat(vtemp, 7); 686 dst_aligned = !((unsigned long)dst & 0xf); 687 src_aligned = !((unsigned long)src & 0xf); 688 689 for (y = 0; y < height; y++) { 690 vdst = vec_ld(0, dst); 691 vsrc = vec_ld(0, src); 692 693 v0 = (vec_s16)vec_mergeh(zero_u8v, vdst); 694 v1 = (vec_s16)vec_mergel(zero_u8v, vdst); 695 v2 = (vec_s16)vec_mergeh(zero_u8v, vsrc); 696 v3 = (vec_s16)vec_mergel(zero_u8v, vsrc); 697 698 if (w == 8) { 699 if (src_aligned) 700 v3 = v2; 701 else 702 v2 = v3; 703 } 704 705 if (w == 16 || dst_aligned) { 706 v0 = vec_mladd(v0, vweightd, zero_s16v); 707 v2 = vec_mladd(v2, vweights, zero_s16v); 708 709 v0 = vec_adds(v0, voffset); 710 v0 = vec_adds(v0, v2); 711 v0 = vec_sra(v0, vlog2_denom); 712 } 713 if (w == 16 || !dst_aligned) { 714 v1 = vec_mladd(v1, vweightd, zero_s16v); 715 v3 = vec_mladd(v3, vweights, zero_s16v); 716 717 v1 = vec_adds(v1, voffset); 718 v1 = vec_adds(v1, v3); 719 v1 = vec_sra(v1, vlog2_denom); 720 } 721 vdst = vec_packsu(v0, v1); 722 vec_st(vdst, 0, dst); 723 724 dst += stride; 725 src += stride; 726 } 727} 728 729#define H264_WEIGHT(W) \ 730static void weight_h264_pixels ## W ## _altivec(uint8_t *block, int stride, int height, \ 731 int log2_denom, int weight, int offset) \ 732{ \ 733 weight_h264_W_altivec(block, stride, height, log2_denom, weight, offset, W); \ 734}\ 735static void biweight_h264_pixels ## W ## _altivec(uint8_t *dst, uint8_t *src, int stride, int height, \ 736 int log2_denom, int weightd, int weights, int offset) \ 737{ \ 738 biweight_h264_W_altivec(dst, src, stride, height, log2_denom, weightd, weights, offset, W); \ 739} 740 741H264_WEIGHT(16) 742H264_WEIGHT( 8) 743#endif /* HAVE_ALTIVEC */ 744 745av_cold void ff_h264dsp_init_ppc(H264DSPContext *c, const int bit_depth, 746 const int chroma_format_idc) 747{ 748#if HAVE_ALTIVEC 749 if (!PPC_ALTIVEC(av_get_cpu_flags())) 750 return; 751 752 if (bit_depth == 8) { 753 c->h264_idct_add = h264_idct_add_altivec; 754 if (chroma_format_idc <= 1) 755 c->h264_idct_add8 = h264_idct_add8_altivec; 756 c->h264_idct_add16 = h264_idct_add16_altivec; 757 c->h264_idct_add16intra = h264_idct_add16intra_altivec; 758 c->h264_idct_dc_add= h264_idct_dc_add_altivec; 759 c->h264_idct8_dc_add = h264_idct8_dc_add_altivec; 760 c->h264_idct8_add = h264_idct8_add_altivec; 761 c->h264_idct8_add4 = h264_idct8_add4_altivec; 762 c->h264_v_loop_filter_luma= h264_v_loop_filter_luma_altivec; 763 c->h264_h_loop_filter_luma= h264_h_loop_filter_luma_altivec; 764 765 c->weight_h264_pixels_tab[0] = weight_h264_pixels16_altivec; 766 c->weight_h264_pixels_tab[1] = weight_h264_pixels8_altivec; 767 c->biweight_h264_pixels_tab[0] = biweight_h264_pixels16_altivec; 768 c->biweight_h264_pixels_tab[1] = biweight_h264_pixels8_altivec; 769 } 770#endif /* HAVE_ALTIVEC */ 771} 772