| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at> | ||
| 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 | #ifndef AVCODEC_X86_CABAC_H | ||
| 22 | #define AVCODEC_X86_CABAC_H | ||
| 23 | |||
| 24 | #include <stddef.h> | ||
| 25 | |||
| 26 | #include "libavcodec/cabac.h" | ||
| 27 | #include "libavutil/attributes.h" | ||
| 28 | #include "libavutil/macros.h" | ||
| 29 | #include "libavutil/x86/asm.h" | ||
| 30 | #include "config.h" | ||
| 31 | |||
| 32 | #if defined(__INTEL_COMPILER) && defined(_MSC_VER) | ||
| 33 | # define BROKEN_COMPILER 1 | ||
| 34 | #else | ||
| 35 | # define BROKEN_COMPILER 0 | ||
| 36 | #endif | ||
| 37 | |||
| 38 | #if HAVE_INLINE_ASM | ||
| 39 | |||
| 40 | #ifndef UNCHECKED_BITSTREAM_READER | ||
| 41 | #define UNCHECKED_BITSTREAM_READER !CONFIG_SAFE_BITSTREAM_READER | ||
| 42 | #endif | ||
| 43 | |||
| 44 | #if UNCHECKED_BITSTREAM_READER | ||
| 45 | #define END_CHECK(end) "" | ||
| 46 | #else | ||
| 47 | #define END_CHECK(end) \ | ||
| 48 | "cmp "end" , %%"FF_REG_c" \n\t"\ | ||
| 49 | "jge 1f \n\t" | ||
| 50 | #endif | ||
| 51 | |||
| 52 | #ifdef BROKEN_RELOCATIONS | ||
| 53 | #define TABLES_ARG , "r"(tables) | ||
| 54 | |||
| 55 | #if HAVE_FAST_CMOV | ||
| 56 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ | ||
| 57 | "cmp "low" , "tmp" \n\t"\ | ||
| 58 | "cmova %%ecx , "range" \n\t"\ | ||
| 59 | "sbb %%rcx , %%rcx \n\t"\ | ||
| 60 | "and %%ecx , "tmp" \n\t"\ | ||
| 61 | "xor %%rcx , "retq" \n\t"\ | ||
| 62 | "sub "tmp" , "low" \n\t" | ||
| 63 | #else /* HAVE_FAST_CMOV */ | ||
| 64 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ | ||
| 65 | /* P4 Prescott has crappy cmov,sbb,64-bit shift so avoid them */ \ | ||
| 66 | "sub "low" , "tmp" \n\t"\ | ||
| 67 | "sar $31 , "tmp" \n\t"\ | ||
| 68 | "sub %%ecx , "range" \n\t"\ | ||
| 69 | "and "tmp" , "range" \n\t"\ | ||
| 70 | "add %%ecx , "range" \n\t"\ | ||
| 71 | "shl $17 , %%ecx \n\t"\ | ||
| 72 | "and "tmp" , %%ecx \n\t"\ | ||
| 73 | "sub %%ecx , "low" \n\t"\ | ||
| 74 | "xor "tmp" , "ret" \n\t"\ | ||
| 75 | "movslq "ret" , "retq" \n\t" | ||
| 76 | #endif /* HAVE_FAST_CMOV */ | ||
| 77 | |||
| 78 | #ifdef __LZCNT__ | ||
| 79 | /* renormalisation shift = lzcnt(range) - 23, refill shift = tzcnt(low) - 16 */ | ||
| 80 | #define BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 81 | "lzcnt "range" , %%ecx \n\t"\ | ||
| 82 | "sub $23 , %%ecx \n\t" | ||
| 83 | #define BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 84 | "movzwl (%%"FF_REG_c") , "tmp" \n\t"\ | ||
| 85 | "tzcnt "low" , %%ecx \n\t"\ | ||
| 86 | "bswap "tmp" \n\t"\ | ||
| 87 | "shr $15 , "tmp" \n\t"\ | ||
| 88 | "sub $16 , %%ecx \n\t"\ | ||
| 89 | "sub $0xFFFF , "tmp" \n\t"\ | ||
| 90 | "shl %%cl , "tmp" \n\t"\ | ||
| 91 | "add "tmp" , "low" \n\t" | ||
| 92 | #else /* __LZCNT__ */ | ||
| 93 | #define BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 94 | "movzbl "norm_off"("tables", "rangeq"), %%ecx \n\t" | ||
| 95 | #define BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 96 | "movzwl (%%"FF_REG_c") , "tmp" \n\t"\ | ||
| 97 | "lea -1("low") , %%ecx \n\t"\ | ||
| 98 | "xor "low" , %%ecx \n\t"\ | ||
| 99 | "shr $15 , %%ecx \n\t"\ | ||
| 100 | "bswap "tmp" \n\t"\ | ||
| 101 | "shr $15 , "tmp" \n\t"\ | ||
| 102 | "movzbl "norm_off"("tables", %%rcx), %%ecx \n\t"\ | ||
| 103 | "sub $0xFFFF , "tmp" \n\t"\ | ||
| 104 | "neg %%ecx \n\t"\ | ||
| 105 | "add $7 , %%ecx \n\t"\ | ||
| 106 | "shl %%cl , "tmp" \n\t"\ | ||
| 107 | "add "tmp" , "low" \n\t" | ||
| 108 | #endif /* __LZCNT__ */ | ||
| 109 | |||
| 110 | #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \ | ||
| 111 | "movzbl "statep" , "ret" \n\t"\ | ||
| 112 | "mov "range" , "tmp" \n\t"\ | ||
| 113 | "and $0xC0 , "range" \n\t"\ | ||
| 114 | "lea ("ret", "range", 2), %%ecx \n\t"\ | ||
| 115 | "movzbl "lps_off"("tables", %%rcx), "range" \n\t"\ | ||
| 116 | "sub "range" , "tmp" \n\t"\ | ||
| 117 | "mov "tmp" , %%ecx \n\t"\ | ||
| 118 | "shl $17 , "tmp" \n\t"\ | ||
| 119 | BRANCHLESS_GET_CABAC_UPDATE(ret, retq, low, range, tmp) \ | ||
| 120 | BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 121 | "shl %%cl , "range" \n\t"\ | ||
| 122 | "movzbl "mlps_off"+128("tables", "retq"), "tmp" \n\t"\ | ||
| 123 | "shl %%cl , "low" \n\t"\ | ||
| 124 | "mov "tmpbyte" , "statep" \n\t"\ | ||
| 125 | "test "lowword" , "lowword" \n\t"\ | ||
| 126 | "jnz 2f \n\t"\ | ||
| 127 | "mov "byte" , %%"FF_REG_c" \n\t"\ | ||
| 128 | END_CHECK(end)\ | ||
| 129 | "add"FF_OPSIZE" $2 , "byte" \n\t"\ | ||
| 130 | "1: \n\t"\ | ||
| 131 | BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 132 | "2: \n\t" | ||
| 133 | |||
| 134 | #else /* BROKEN_RELOCATIONS */ | ||
| 135 | #define TABLES_ARG NAMED_CONSTRAINTS_ARRAY_ADD(ff_h264_cabac_tables) | ||
| 136 | #define RIP_ARG | ||
| 137 | |||
| 138 | #if HAVE_FAST_CMOV | ||
| 139 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\ | ||
| 140 | "mov "tmp" , %%ecx \n\t"\ | ||
| 141 | "shl $17 , "tmp" \n\t"\ | ||
| 142 | "cmp "low" , "tmp" \n\t"\ | ||
| 143 | "cmova %%ecx , "range" \n\t"\ | ||
| 144 | "sbb %%ecx , %%ecx \n\t"\ | ||
| 145 | "and %%ecx , "tmp" \n\t"\ | ||
| 146 | "xor %%ecx , "ret" \n\t"\ | ||
| 147 | "sub "tmp" , "low" \n\t" | ||
| 148 | #else /* HAVE_FAST_CMOV */ | ||
| 149 | #define BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp)\ | ||
| 150 | "mov "tmp" , %%ecx \n\t"\ | ||
| 151 | "shl $17 , "tmp" \n\t"\ | ||
| 152 | "sub "low" , "tmp" \n\t"\ | ||
| 153 | "sar $31 , "tmp" \n\t" /*lps_mask*/\ | ||
| 154 | "sub %%ecx , "range" \n\t" /*RangeLPS - range*/\ | ||
| 155 | "and "tmp" , "range" \n\t" /*(RangeLPS - range)&lps_mask*/\ | ||
| 156 | "add %%ecx , "range" \n\t" /*new range*/\ | ||
| 157 | "shl $17 , %%ecx \n\t"\ | ||
| 158 | "and "tmp" , %%ecx \n\t"\ | ||
| 159 | "sub %%ecx , "low" \n\t"\ | ||
| 160 | "xor "tmp" , "ret" \n\t" | ||
| 161 | #endif /* HAVE_FAST_CMOV */ | ||
| 162 | |||
| 163 | #ifdef __LZCNT__ | ||
| 164 | #define BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 165 | "lzcnt "range" , %%ecx \n\t"\ | ||
| 166 | "sub $23 , %%ecx \n\t" | ||
| 167 | #define BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 168 | "movzwl (%%"FF_REG_c") , "tmp" \n\t"\ | ||
| 169 | "tzcnt "low" , %%ecx \n\t"\ | ||
| 170 | "bswap "tmp" \n\t"\ | ||
| 171 | "shr $15 , "tmp" \n\t"\ | ||
| 172 | "sub $16 , %%ecx \n\t"\ | ||
| 173 | "sub $0xFFFF , "tmp" \n\t"\ | ||
| 174 | "shl %%cl , "tmp" \n\t"\ | ||
| 175 | "add "tmp" , "low" \n\t" | ||
| 176 | #else /* __LZCNT__ */ | ||
| 177 | #define BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 178 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"("range"), %%ecx \n\t" | ||
| 179 | #define BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 180 | "movzwl (%%"FF_REG_c") , "tmp" \n\t"\ | ||
| 181 | "lea -1("low") , %%ecx \n\t"\ | ||
| 182 | "xor "low" , %%ecx \n\t"\ | ||
| 183 | "shr $15 , %%ecx \n\t"\ | ||
| 184 | "bswap "tmp" \n\t"\ | ||
| 185 | "shr $15 , "tmp" \n\t"\ | ||
| 186 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"norm_off"(%%ecx), %%ecx \n\t"\ | ||
| 187 | "sub $0xFFFF , "tmp" \n\t"\ | ||
| 188 | "neg %%ecx \n\t"\ | ||
| 189 | "add $7 , %%ecx \n\t"\ | ||
| 190 | "shl %%cl , "tmp" \n\t"\ | ||
| 191 | "add "tmp" , "low" \n\t" | ||
| 192 | #endif /* __LZCNT__ */ | ||
| 193 | |||
| 194 | #define BRANCHLESS_GET_CABAC(ret, retq, statep, low, lowword, range, rangeq, tmp, tmpbyte, byte, end, norm_off, lps_off, mlps_off, tables) \ | ||
| 195 | "movzbl "statep" , "ret" \n\t"\ | ||
| 196 | "mov "range" , "tmp" \n\t"\ | ||
| 197 | "and $0xC0 , "range" \n\t"\ | ||
| 198 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"lps_off"("ret", "range", 2), "range" \n\t"\ | ||
| 199 | "sub "range" , "tmp" \n\t"\ | ||
| 200 | BRANCHLESS_GET_CABAC_UPDATE(ret, low, range, tmp) \ | ||
| 201 | BRANCHLESS_GET_CABAC_RENORM(range, rangeq, norm_off, tables) \ | ||
| 202 | "shl %%cl , "range" \n\t"\ | ||
| 203 | "movzbl "MANGLE(ff_h264_cabac_tables)"+"mlps_off"+128("ret"), "tmp" \n\t"\ | ||
| 204 | "shl %%cl , "low" \n\t"\ | ||
| 205 | "mov "tmpbyte" , "statep" \n\t"\ | ||
| 206 | "test "lowword" , "lowword" \n\t"\ | ||
| 207 | " jnz 2f \n\t"\ | ||
| 208 | "mov "byte" , %%"FF_REG_c" \n\t"\ | ||
| 209 | END_CHECK(end)\ | ||
| 210 | "add"FF_OPSIZE" $2 , "byte" \n\t"\ | ||
| 211 | "1: \n\t"\ | ||
| 212 | BRANCHLESS_GET_CABAC_REFILL(low, tmp, norm_off, tables) \ | ||
| 213 | "2: \n\t" | ||
| 214 | |||
| 215 | #endif /* BROKEN_RELOCATIONS */ | ||
| 216 | |||
| 217 | #if HAVE_X86_7REGS && !BROKEN_COMPILER | ||
| 218 | #define get_cabac_inline get_cabac_inline_x86 | ||
| 219 | static | ||
| 220 | #if ARCH_X86_32 | ||
| 221 | av_noinline | ||
| 222 | #else | ||
| 223 | av_always_inline | ||
| 224 | #endif | ||
| 225 | 1190408538 | int get_cabac_inline_x86(CABACContext *c, uint8_t *const state) | |
| 226 | { | ||
| 227 | int bit, tmp; | ||
| 228 | #ifdef BROKEN_RELOCATIONS | ||
| 229 | void *tables; | ||
| 230 | |||
| 231 | 1190408538 | __asm__ volatile( | |
| 232 | "lea "MANGLE(ff_h264_cabac_tables)", %0 \n\t" | ||
| 233 | : "=&r"(tables) | ||
| 234 | : NAMED_CONSTRAINTS_ARRAY(ff_h264_cabac_tables) | ||
| 235 | ); | ||
| 236 | #endif | ||
| 237 | |||
| 238 | 1190408538 | __asm__ volatile( | |
| 239 | BRANCHLESS_GET_CABAC("%0", "%q0", "(%4)", "%1", "%w1", | ||
| 240 | "%2", "%q2", "%3", "%b3", | ||
| 241 | "%c6(%5)", "%c7(%5)", | ||
| 242 | AV_STRINGIFY(H264_NORM_SHIFT_OFFSET), | ||
| 243 | AV_STRINGIFY(H264_LPS_RANGE_OFFSET), | ||
| 244 | AV_STRINGIFY(H264_MLPS_STATE_OFFSET), | ||
| 245 | "%8") | ||
| 246 | : "=&r"(bit), "=&r"(c->low), "=&r"(c->range), "=&q"(tmp) | ||
| 247 | : "r"(state), "r"(c), | ||
| 248 | "i"(offsetof(CABACContext, bytestream)), | ||
| 249 | "i"(offsetof(CABACContext, bytestream_end)) | ||
| 250 | TABLES_ARG | ||
| 251 | 1190408538 | ,"1"(c->low), "2"(c->range) | |
| 252 | : "%"FF_REG_c, "memory" | ||
| 253 | ); | ||
| 254 | 1190408538 | return bit & 1; | |
| 255 | } | ||
| 256 | #endif /* HAVE_X86_7REGS && !BROKEN_COMPILER */ | ||
| 257 | |||
| 258 | #if !BROKEN_COMPILER | ||
| 259 | #define get_cabac_bypass_sign get_cabac_bypass_sign_x86 | ||
| 260 | 182211427 | static av_always_inline int get_cabac_bypass_sign_x86(CABACContext *c, int val) | |
| 261 | { | ||
| 262 | x86_reg tmp; | ||
| 263 | 182211427 | __asm__ volatile( | |
| 264 | "movl %c6(%2), %k1 \n\t" | ||
| 265 | "movl %c3(%2), %%eax \n\t" | ||
| 266 | "shl $17, %k1 \n\t" | ||
| 267 | "add %%eax, %%eax \n\t" | ||
| 268 | "sub %k1, %%eax \n\t" | ||
| 269 | "cdq \n\t" | ||
| 270 | "and %%edx, %k1 \n\t" | ||
| 271 | "add %k1, %%eax \n\t" | ||
| 272 | "xor %%edx, %%ecx \n\t" | ||
| 273 | "sub %%edx, %%ecx \n\t" | ||
| 274 | "test %%ax, %%ax \n\t" | ||
| 275 | "jnz 1f \n\t" | ||
| 276 | "mov %c4(%2), %1 \n\t" | ||
| 277 | "subl $0xFFFF, %%eax \n\t" | ||
| 278 | "movzwl (%1), %%edx \n\t" | ||
| 279 | "bswap %%edx \n\t" | ||
| 280 | "shrl $15, %%edx \n\t" | ||
| 281 | #if UNCHECKED_BITSTREAM_READER | ||
| 282 | "add $2, %1 \n\t" | ||
| 283 | "addl %%edx, %%eax \n\t" | ||
| 284 | "mov %1, %c4(%2) \n\t" | ||
| 285 | #else | ||
| 286 | "addl %%edx, %%eax \n\t" | ||
| 287 | "cmp %c5(%2), %1 \n\t" | ||
| 288 | "jge 1f \n\t" | ||
| 289 | "add"FF_OPSIZE" $2, %c4(%2) \n\t" | ||
| 290 | #endif | ||
| 291 | "1: \n\t" | ||
| 292 | "movl %%eax, %c3(%2) \n\t" | ||
| 293 | |||
| 294 | : "+c"(val), "=&r"(tmp) | ||
| 295 | : "r"(c), | ||
| 296 | "i"(offsetof(CABACContext, low)), | ||
| 297 | "i"(offsetof(CABACContext, bytestream)), | ||
| 298 | "i"(offsetof(CABACContext, bytestream_end)), | ||
| 299 | "i"(offsetof(CABACContext, range)) | ||
| 300 | : "%eax", "%edx", "memory" | ||
| 301 | ); | ||
| 302 | 182211427 | return val; | |
| 303 | } | ||
| 304 | |||
| 305 | #define get_cabac_bypass get_cabac_bypass_x86 | ||
| 306 | 165119810 | static av_always_inline int get_cabac_bypass_x86(CABACContext *c) | |
| 307 | { | ||
| 308 | x86_reg tmp; | ||
| 309 | int res; | ||
| 310 | 165119810 | __asm__ volatile( | |
| 311 | "movl %c6(%2), %k1 \n\t" | ||
| 312 | "movl %c3(%2), %%eax \n\t" | ||
| 313 | "shl $17, %k1 \n\t" | ||
| 314 | "add %%eax, %%eax \n\t" | ||
| 315 | "sub %k1, %%eax \n\t" | ||
| 316 | "cdq \n\t" | ||
| 317 | "and %%edx, %k1 \n\t" | ||
| 318 | "add %k1, %%eax \n\t" | ||
| 319 | "inc %%edx \n\t" | ||
| 320 | "test %%ax, %%ax \n\t" | ||
| 321 | "jnz 1f \n\t" | ||
| 322 | "mov %c4(%2), %1 \n\t" | ||
| 323 | "subl $0xFFFF, %%eax \n\t" | ||
| 324 | "movzwl (%1), %%ecx \n\t" | ||
| 325 | "bswap %%ecx \n\t" | ||
| 326 | "shrl $15, %%ecx \n\t" | ||
| 327 | "addl %%ecx, %%eax \n\t" | ||
| 328 | "cmp %c5(%2), %1 \n\t" | ||
| 329 | "jge 1f \n\t" | ||
| 330 | "add"FF_OPSIZE" $2, %c4(%2) \n\t" | ||
| 331 | "1: \n\t" | ||
| 332 | "movl %%eax, %c3(%2) \n\t" | ||
| 333 | |||
| 334 | : "=&d"(res), "=&r"(tmp) | ||
| 335 | : "r"(c), | ||
| 336 | "i"(offsetof(CABACContext, low)), | ||
| 337 | "i"(offsetof(CABACContext, bytestream)), | ||
| 338 | "i"(offsetof(CABACContext, bytestream_end)), | ||
| 339 | "i"(offsetof(CABACContext, range)) | ||
| 340 | : "%eax", "%ecx", "memory" | ||
| 341 | ); | ||
| 342 | 165119810 | return res; | |
| 343 | } | ||
| 344 | #endif /* !BROKEN_COMPILER */ | ||
| 345 | |||
| 346 | #endif /* HAVE_INLINE_ASM */ | ||
| 347 | #endif /* AVCODEC_X86_CABAC_H */ | ||
| 348 |