libavcodec/i386/idct_sse2_xvid.c

   1 /*
   2  * XVID MPEG-4 VIDEO CODEC
   3  * - SSE2 inverse discrete cosine transform -
   4  *
   5  * Copyright(C) 2003 Pascal Massimino <skal@planet-d.net>
   6  *
   7  * Conversion to gcc syntax with modifications
   8  * by Alexander Strange <astrange@ithinksw.com>
   9  *
  10  * Originally from dct/x86_asm/fdct_sse2_skal.asm in Xvid.
  11  *
  12  * This file is part of FFmpeg.
  13  *
  14  * Vertical pass is an implementation of the scheme:
  15  *  Loeffler C., Ligtenberg A., and Moschytz C.S.:
  16  *  Practical Fast 1D DCT Algorithm with Eleven Multiplications,
  17  *  Proc. ICASSP 1989, 988-991.
  18  *
  19  * Horizontal pass is a double 4x4 vector/matrix multiplication,
  20  * (see also Intel's Application Note 922:
  21  *  http://developer.intel.com/vtune/cbts/strmsimd/922down.htm
  22  *  Copyright (C) 1999 Intel Corporation)
  23  *
  24  * More details at http://skal.planet-d.net/coding/dct.html
  25  *
  26  * FFmpeg is free software; you can redistribute it and/or
  27  * modify it under the terms of the GNU Lesser General Public
  28  * License as published by the Free Software Foundation; either
  29  * version 2.1 of the License, or (at your option) any later version.
  30  *
  31  * FFmpeg is distributed in the hope that it will be useful,
  32  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  33  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  34  * Lesser General Public License for more details.
  35  *
  36  * You should have received a copy of the GNU Lesser General Public License
  37  * along with FFmpeg; if not, write to the Free Software Foundation,
  38  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  39  */
  40
  41 #include "libavcodec/dsputil.h"
  42 #include "libavcodec/i386/idct_xvid.h"
  43
  44 /*!
  45  * @file idct_sse2_xvid.c
  46  * @brief SSE2 idct compatible with xvidmmx
  47  */
  48
  49 #define X8(x)     x,x,x,x,x,x,x,x
  50
  51 #define ROW_SHIFT 11
  52 #define COL_SHIFT 6
  53
  54 DECLARE_ASM_CONST(16, int16_t, tan1[]) = {X8(13036)}; // tan( pi/16)
  55 DECLARE_ASM_CONST(16, int16_t, tan2[]) = {X8(27146)}; // tan(2pi/16) = sqrt(2)-1
  56 DECLARE_ASM_CONST(16, int16_t, tan3[]) = {X8(43790)}; // tan(3pi/16)-1
  57 DECLARE_ASM_CONST(16, int16_t, sqrt2[])= {X8(23170)}; // 0.5/sqrt(2)
  58 DECLARE_ASM_CONST(8,  uint8_t, m127[]) = {X8(127)};
  59
  60 DECLARE_ASM_CONST(16, int16_t, iTab1[]) = {
  61  0x4000, 0x539f, 0xc000, 0xac61, 0x4000, 0xdd5d, 0x4000, 0xdd5d,
  62  0x4000, 0x22a3, 0x4000, 0x22a3, 0xc000, 0x539f, 0x4000, 0xac61,
  63  0x3249, 0x11a8, 0x4b42, 0xee58, 0x11a8, 0x4b42, 0x11a8, 0xcdb7,
  64  0x58c5, 0x4b42, 0xa73b, 0xcdb7, 0x3249, 0xa73b, 0x4b42, 0xa73b
  65 };
  66
  67 DECLARE_ASM_CONST(16, int16_t, iTab2[]) = {
  68  0x58c5, 0x73fc, 0xa73b, 0x8c04, 0x58c5, 0xcff5, 0x58c5, 0xcff5,
  69  0x58c5, 0x300b, 0x58c5, 0x300b, 0xa73b, 0x73fc, 0x58c5, 0x8c04,
  70  0x45bf, 0x187e, 0x6862, 0xe782, 0x187e, 0x6862, 0x187e, 0xba41,
  71  0x7b21, 0x6862, 0x84df, 0xba41, 0x45bf, 0x84df, 0x6862, 0x84df
  72 };
  73
  74 DECLARE_ASM_CONST(16, int16_t, iTab3[]) = {
  75  0x539f, 0x6d41, 0xac61, 0x92bf, 0x539f, 0xd2bf, 0x539f, 0xd2bf,
  76  0x539f, 0x2d41, 0x539f, 0x2d41, 0xac61, 0x6d41, 0x539f, 0x92bf,
  77  0x41b3, 0x1712, 0x6254, 0xe8ee, 0x1712, 0x6254, 0x1712, 0xbe4d,
  78  0x73fc, 0x6254, 0x8c04, 0xbe4d, 0x41b3, 0x8c04, 0x6254, 0x8c04
  79 };
  80
  81 DECLARE_ASM_CONST(16, int16_t, iTab4[]) = {
  82  0x4b42, 0x6254, 0xb4be, 0x9dac, 0x4b42, 0xd746, 0x4b42, 0xd746,
  83  0x4b42, 0x28ba, 0x4b42, 0x28ba, 0xb4be, 0x6254, 0x4b42, 0x9dac,
  84  0x3b21, 0x14c3, 0x587e, 0xeb3d, 0x14c3, 0x587e, 0x14c3, 0xc4df,
  85  0x6862, 0x587e, 0x979e, 0xc4df, 0x3b21, 0x979e, 0x587e, 0x979e
  86 };
  87
  88 DECLARE_ASM_CONST(16, int32_t, walkenIdctRounders[]) = {
  89  65536, 65536, 65536, 65536,
  90   3597,  3597,  3597,  3597,
  91   2260,  2260,  2260,  2260,
  92   1203,  1203,  1203,  1203,
  93    120,   120,   120,   120,
  94    512,   512,   512,   512
  95 };
  96
  97 // Temporary storage before the column pass
  98 #define ROW1 "%%xmm6"
  99 #define ROW3 "%%xmm4"
 100 #define ROW5 "%%xmm5"
 101 #define ROW7 "%%xmm7"
 102
 103 #define CLEAR_ODD(r) "pxor  "r","r" \n\t"
 104 #define PUT_ODD(dst) "pshufhw  $0x1B, %%xmm2, "dst"   \n\t"
 105
 106 #ifdef ARCH_X86_64
 107
 108 # define ROW0 "%%xmm8"
 109 # define REG0 ROW0
 110 # define ROW2 "%%xmm9"
 111 # define REG2 ROW2
 112 # define ROW4 "%%xmm10"
 113 # define REG4 ROW4
 114 # define ROW6 "%%xmm11"
 115 # define REG6 ROW6
 116 # define CLEAR_EVEN(r) CLEAR_ODD(r)
 117 # define PUT_EVEN(dst) PUT_ODD(dst)
 118 # define XMMS "%%xmm12"
 119 # define MOV_32_ONLY "#"
 120 # define SREG2 REG2
 121 # define TAN3 "%%xmm13"
 122 # define TAN1 "%%xmm14"
 123
 124 #else
 125
 126 # define ROW0 "(%0)"
 127 # define REG0 "%%xmm4"
 128 # define ROW2 "2*16(%0)"
 129 # define REG2 "%%xmm4"
 130 # define ROW4 "4*16(%0)"
 131 # define REG4 "%%xmm6"
 132 # define ROW6 "6*16(%0)"
 133 # define REG6 "%%xmm6"
 134 # define CLEAR_EVEN(r)
 135 # define PUT_EVEN(dst) \
 136     "pshufhw  $0x1B, %%xmm2, %%xmm2   \n\t" \
 137     "movdqa          %%xmm2, "dst"    \n\t"
 138 # define XMMS "%%xmm2"
 139 # define MOV_32_ONLY "movdqa "
 140 # define SREG2 "%%xmm7"
 141 # define TAN3 "%%xmm0"
 142 # define TAN1 "%%xmm2"
 143
 144 #endif
 145
 146 #define ROUND(x) "paddd   "MANGLE(x)
 147
 148 #define JZ(reg, to)                         \
 149     "testl     "reg","reg"            \n\t" \
 150     "jz        "to"                   \n\t"
 151
 152 #define JNZ(reg, to)                        \
 153     "testl     "reg","reg"            \n\t" \
 154     "jnz       "to"                   \n\t"
 155
 156 #define TEST_ONE_ROW(src, reg, clear)       \
 157     clear                                   \
 158     "movq     "src", %%mm1            \n\t" \
 159     "por    8+"src", %%mm1            \n\t" \
 160     "paddusb  %%mm0, %%mm1            \n\t" \
 161     "pmovmskb %%mm1, "reg"            \n\t"
 162
 163 #define TEST_TWO_ROWS(row1, row2, reg1, reg2, clear1, clear2) \
 164     clear1                                  \
 165     clear2                                  \
 166     "movq     "row1", %%mm1           \n\t" \
 167     "por    8+"row1", %%mm1           \n\t" \
 168     "movq     "row2", %%mm2           \n\t" \
 169     "por    8+"row2", %%mm2           \n\t" \
 170     "paddusb   %%mm0, %%mm1           \n\t" \
 171     "paddusb   %%mm0, %%mm2           \n\t" \
 172     "pmovmskb  %%mm1, "reg1"          \n\t" \
 173     "pmovmskb  %%mm2, "reg2"          \n\t"
 174
 175 ///IDCT pass on rows.
 176 #define iMTX_MULT(src, table, rounder, put) \
 177     "movdqa        "src", %%xmm3      \n\t" \
 178     "movdqa       %%xmm3, %%xmm0      \n\t" \
 179     "pshufd   $0x11, %%xmm3, %%xmm1   \n\t" /* 4602 */ \
 180     "punpcklqdq   %%xmm0, %%xmm0      \n\t" /* 0246 */ \
 181     "pmaddwd     "table", %%xmm0      \n\t" \
 182     "pmaddwd  16+"table", %%xmm1      \n\t" \
 183     "pshufd   $0xBB, %%xmm3, %%xmm2   \n\t" /* 5713 */ \
 184     "punpckhqdq   %%xmm3, %%xmm3      \n\t" /* 1357 */ \
 185     "pmaddwd  32+"table", %%xmm2      \n\t" \
 186     "pmaddwd  48+"table", %%xmm3      \n\t" \
 187     "paddd        %%xmm1, %%xmm0      \n\t" \
 188     "paddd        %%xmm3, %%xmm2      \n\t" \
 189     rounder",     %%xmm0              \n\t" \
 190     "movdqa       %%xmm2, %%xmm3      \n\t" \
 191     "paddd        %%xmm0, %%xmm2      \n\t" \
 192     "psubd        %%xmm3, %%xmm0      \n\t" \
 193     "psrad           $11, %%xmm2      \n\t" \
 194     "psrad           $11, %%xmm0      \n\t" \
 195     "packssdw     %%xmm0, %%xmm2      \n\t" \
 196     put                                     \
 197     "1:                               \n\t"
 198
 199 #define iLLM_HEAD                           \
 200     "movdqa   "MANGLE(tan3)", "TAN3"  \n\t" \
 201     "movdqa   "MANGLE(tan1)", "TAN1"  \n\t" \
 202
 203 ///IDCT pass on columns.
 204 #define iLLM_PASS(dct)                      \
 205     "movdqa   "TAN3", %%xmm1          \n\t" \
 206     "movdqa   "TAN1", %%xmm3          \n\t" \
 207     "pmulhw   %%xmm4, "TAN3"          \n\t" \
 208     "pmulhw   %%xmm5, %%xmm1          \n\t" \
 209     "paddsw   %%xmm4, "TAN3"          \n\t" \
 210     "paddsw   %%xmm5, %%xmm1          \n\t" \
 211     "psubsw   %%xmm5, "TAN3"          \n\t" \
 212     "paddsw   %%xmm4, %%xmm1          \n\t" \
 213     "pmulhw   %%xmm7, %%xmm3          \n\t" \
 214     "pmulhw   %%xmm6, "TAN1"          \n\t" \
 215     "paddsw   %%xmm6, %%xmm3          \n\t" \
 216     "psubsw   %%xmm7, "TAN1"          \n\t" \
 217     "movdqa   %%xmm3, %%xmm7          \n\t" \
 218     "movdqa   "TAN1", %%xmm6          \n\t" \
 219     "psubsw   %%xmm1, %%xmm3          \n\t" \
 220     "psubsw   "TAN3", "TAN1"          \n\t" \
 221     "paddsw   %%xmm7, %%xmm1          \n\t" \
 222     "paddsw   %%xmm6, "TAN3"          \n\t" \
 223     "movdqa   %%xmm3, %%xmm6          \n\t" \
 224     "psubsw   "TAN3", %%xmm3          \n\t" \
 225     "paddsw   %%xmm6, "TAN3"          \n\t" \
 226     "movdqa   "MANGLE(sqrt2)", %%xmm4 \n\t" \
 227     "pmulhw   %%xmm4, %%xmm3          \n\t" \
 228     "pmulhw   %%xmm4, "TAN3"          \n\t" \
 229     "paddsw   "TAN3", "TAN3"          \n\t" \
 230     "paddsw   %%xmm3, %%xmm3          \n\t" \
 231     "movdqa   "MANGLE(tan2)", %%xmm7  \n\t" \
 232     MOV_32_ONLY ROW2", "REG2"         \n\t" \
 233     MOV_32_ONLY ROW6", "REG6"         \n\t" \
 234     "movdqa   %%xmm7, %%xmm5          \n\t" \
 235     "pmulhw   "REG6", %%xmm7          \n\t" \
 236     "pmulhw   "REG2", %%xmm5          \n\t" \
 237     "paddsw   "REG2", %%xmm7          \n\t" \
 238     "psubsw   "REG6", %%xmm5          \n\t" \
 239     MOV_32_ONLY ROW0", "REG0"         \n\t" \
 240     MOV_32_ONLY ROW4", "REG4"         \n\t" \
 241     MOV_32_ONLY"  "TAN1", (%0)        \n\t" \
 242     "movdqa   "REG0", "XMMS"          \n\t" \
 243     "psubsw   "REG4", "REG0"          \n\t" \
 244     "paddsw   "XMMS", "REG4"          \n\t" \
 245     "movdqa   "REG4", "XMMS"          \n\t" \
 246     "psubsw   %%xmm7, "REG4"          \n\t" \
 247     "paddsw   "XMMS", %%xmm7          \n\t" \
 248     "movdqa   "REG0", "XMMS"          \n\t" \
 249     "psubsw   %%xmm5, "REG0"          \n\t" \
 250     "paddsw   "XMMS", %%xmm5          \n\t" \
 251     "movdqa   %%xmm5, "XMMS"          \n\t" \
 252     "psubsw   "TAN3", %%xmm5          \n\t" \
 253     "paddsw   "XMMS", "TAN3"          \n\t" \
 254     "movdqa   "REG0", "XMMS"          \n\t" \
 255     "psubsw   %%xmm3, "REG0"          \n\t" \
 256     "paddsw   "XMMS", %%xmm3          \n\t" \
 257     MOV_32_ONLY"  (%0), "TAN1"        \n\t" \
 258     "psraw        $6, %%xmm5          \n\t" \
 259     "psraw        $6, "REG0"          \n\t" \
 260     "psraw        $6, "TAN3"          \n\t" \
 261     "psraw        $6, %%xmm3          \n\t" \
 262     "movdqa   "TAN3", 1*16("dct")     \n\t" \
 263     "movdqa   %%xmm3, 2*16("dct")     \n\t" \
 264     "movdqa   "REG0", 5*16("dct")     \n\t" \
 265     "movdqa   %%xmm5, 6*16("dct")     \n\t" \
 266     "movdqa   %%xmm7, %%xmm0          \n\t" \
 267     "movdqa   "REG4", %%xmm4          \n\t" \
 268     "psubsw   %%xmm1, %%xmm7          \n\t" \
 269     "psubsw   "TAN1", "REG4"          \n\t" \
 270     "paddsw   %%xmm0, %%xmm1          \n\t" \
 271     "paddsw   %%xmm4, "TAN1"          \n\t" \
 272     "psraw        $6, %%xmm1          \n\t" \
 273     "psraw        $6, %%xmm7          \n\t" \
 274     "psraw        $6, "TAN1"          \n\t" \
 275     "psraw        $6, "REG4"          \n\t" \
 276     "movdqa   %%xmm1, ("dct")         \n\t" \
 277     "movdqa   "TAN1", 3*16("dct")     \n\t" \
 278     "movdqa   "REG4", 4*16("dct")     \n\t" \
 279     "movdqa   %%xmm7, 7*16("dct")     \n\t"
 280
 281 ///IDCT pass on columns, assuming rows 4-7 are zero.
 282 #define iLLM_PASS_SPARSE(dct)               \
 283     "pmulhw   %%xmm4, "TAN3"          \n\t" \
 284     "paddsw   %%xmm4, "TAN3"          \n\t" \
 285     "movdqa   %%xmm6, %%xmm3          \n\t" \
 286     "pmulhw   %%xmm6, "TAN1"          \n\t" \
 287     "movdqa   %%xmm4, %%xmm1          \n\t" \
 288     "psubsw   %%xmm1, %%xmm3          \n\t" \
 289     "paddsw   %%xmm6, %%xmm1          \n\t" \
 290     "movdqa   "TAN1", %%xmm6          \n\t" \
 291     "psubsw   "TAN3", "TAN1"          \n\t" \
 292     "paddsw   %%xmm6, "TAN3"          \n\t" \
 293     "movdqa   %%xmm3, %%xmm6          \n\t" \
 294     "psubsw   "TAN3", %%xmm3          \n\t" \
 295     "paddsw   %%xmm6, "TAN3"          \n\t" \
 296     "movdqa   "MANGLE(sqrt2)", %%xmm4 \n\t" \
 297     "pmulhw   %%xmm4, %%xmm3          \n\t" \
 298     "pmulhw   %%xmm4, "TAN3"          \n\t" \
 299     "paddsw   "TAN3", "TAN3"          \n\t" \
 300     "paddsw   %%xmm3, %%xmm3          \n\t" \
 301     "movdqa   "MANGLE(tan2)", %%xmm5  \n\t" \
 302     MOV_32_ONLY ROW2", "SREG2"        \n\t" \
 303     "pmulhw   "SREG2", %%xmm5         \n\t" \
 304     MOV_32_ONLY ROW0", "REG0"         \n\t" \
 305     "movdqa   "REG0", %%xmm6          \n\t" \
 306     "psubsw   "SREG2", %%xmm6         \n\t" \
 307     "paddsw   "REG0", "SREG2"         \n\t" \
 308     MOV_32_ONLY"  "TAN1", (%0)        \n\t" \
 309     "movdqa   "REG0", "XMMS"          \n\t" \
 310     "psubsw   %%xmm5, "REG0"          \n\t" \
 311     "paddsw   "XMMS", %%xmm5          \n\t" \
 312     "movdqa   %%xmm5, "XMMS"          \n\t" \
 313     "psubsw   "TAN3", %%xmm5          \n\t" \
 314     "paddsw   "XMMS", "TAN3"          \n\t" \
 315     "movdqa   "REG0", "XMMS"          \n\t" \
 316     "psubsw   %%xmm3, "REG0"          \n\t" \
 317     "paddsw   "XMMS", %%xmm3          \n\t" \
 318     MOV_32_ONLY"  (%0), "TAN1"        \n\t" \
 319     "psraw        $6, %%xmm5          \n\t" \
 320     "psraw        $6, "REG0"          \n\t" \
 321     "psraw        $6, "TAN3"          \n\t" \
 322     "psraw        $6, %%xmm3          \n\t" \
 323     "movdqa   "TAN3", 1*16("dct")     \n\t" \
 324     "movdqa   %%xmm3, 2*16("dct")     \n\t" \
 325     "movdqa   "REG0", 5*16("dct")     \n\t" \
 326     "movdqa   %%xmm5, 6*16("dct")     \n\t" \
 327     "movdqa   "SREG2", %%xmm0         \n\t" \
 328     "movdqa   %%xmm6, %%xmm4          \n\t" \
 329     "psubsw   %%xmm1, "SREG2"         \n\t" \
 330     "psubsw   "TAN1", %%xmm6          \n\t" \
 331     "paddsw   %%xmm0, %%xmm1          \n\t" \
 332     "paddsw   %%xmm4, "TAN1"          \n\t" \
 333     "psraw        $6, %%xmm1          \n\t" \
 334     "psraw        $6, "SREG2"         \n\t" \
 335     "psraw        $6, "TAN1"          \n\t" \
 336     "psraw        $6, %%xmm6          \n\t" \
 337     "movdqa   %%xmm1, ("dct")         \n\t" \
 338     "movdqa   "TAN1", 3*16("dct")     \n\t" \
 339     "movdqa   %%xmm6, 4*16("dct")     \n\t" \
 340     "movdqa   "SREG2", 7*16("dct")    \n\t"
 341
 342 inline void ff_idct_xvid_sse2(short *block)
 343 {
 344     __asm__ volatile(
 345     "movq     "MANGLE(m127)", %%mm0                              \n\t"
 346     iMTX_MULT("(%0)",     MANGLE(iTab1), ROUND(walkenIdctRounders),      PUT_EVEN(ROW0))
 347     iMTX_MULT("1*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders+1*16), PUT_ODD(ROW1))
 348     iMTX_MULT("2*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders+2*16), PUT_EVEN(ROW2))
 349
 350     TEST_TWO_ROWS("3*16(%0)", "4*16(%0)", "%%eax", "%%ecx", CLEAR_ODD(ROW3), CLEAR_EVEN(ROW4))
 351     JZ("%%eax", "1f")
 352     iMTX_MULT("3*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders+3*16), PUT_ODD(ROW3))
 353
 354     TEST_TWO_ROWS("5*16(%0)", "6*16(%0)", "%%eax", "%%edx", CLEAR_ODD(ROW5), CLEAR_EVEN(ROW6))
 355     TEST_ONE_ROW("7*16(%0)", "%%esi", CLEAR_ODD(ROW7))
 356     iLLM_HEAD
 357     ASMALIGN(4)
 358     JNZ("%%ecx", "2f")
 359     JNZ("%%eax", "3f")
 360     JNZ("%%edx", "4f")
 361     JNZ("%%esi", "5f")
 362     iLLM_PASS_SPARSE("%0")
 363     "jmp 6f                                                      \n\t"
 364     "2:                                                          \n\t"
 365     iMTX_MULT("4*16(%0)", MANGLE(iTab1), "#", PUT_EVEN(ROW4))
 366     "3:                                                          \n\t"
 367     iMTX_MULT("5*16(%0)", MANGLE(iTab4), ROUND(walkenIdctRounders+4*16), PUT_ODD(ROW5))
 368     JZ("%%edx", "1f")
 369     "4:                                                          \n\t"
 370     iMTX_MULT("6*16(%0)", MANGLE(iTab3), ROUND(walkenIdctRounders+5*16), PUT_EVEN(ROW6))
 371     JZ("%%esi", "1f")
 372     "5:                                                          \n\t"
 373     iMTX_MULT("7*16(%0)", MANGLE(iTab2), ROUND(walkenIdctRounders+5*16), PUT_ODD(ROW7))
 374 #ifndef ARCH_X86_64
 375     iLLM_HEAD
 376 #endif
 377     iLLM_PASS("%0")
 378     "6:                                                          \n\t"
 379     : "+r"(block)
 380     :
 381     : "%eax", "%ecx", "%edx", "%esi", "memory");
 382 }
 383
 384 void ff_idct_xvid_sse2_put(uint8_t *dest, int line_size, short *block)
 385 {
 386     ff_idct_xvid_sse2(block);
 387     put_pixels_clamped_mmx(block, dest, line_size);
 388 }
 389
 390 void ff_idct_xvid_sse2_add(uint8_t *dest, int line_size, short *block)
 391 {
 392     ff_idct_xvid_sse2(block);
 393     add_pixels_clamped_mmx(block, dest, line_size);
 394 }