Parent Directory
| 
Revision Log
Revision 1793 - (view) (download)
| 1 : | edgomez | 1382 | ;/**************************************************************************** |
| 2 : | ; * | ||
| 3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
| 4 : | ; * - MMX and XMM forward discrete cosine transform - | ||
| 5 : | ; * | ||
| 6 : | ; * Copyright(C) 2001 Peter Ross <pross@xvid.org> | ||
| 7 : | ; * 2002 Jaan Kalda | ||
| 8 : | ; * | ||
| 9 : | ; * This program is free software; you can redistribute it and/or modify it | ||
| 10 : | ; * 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 02111-1307 USA | ||
| 22 : | ; * | ||
| 23 : | Isibaar | 1793 | ; * $Id: idct_3dne.asm,v 1.8 2008-11-11 20:46:24 Isibaar Exp $ |
| 24 : | edgomez | 1382 | ; * |
| 25 : | ; ***************************************************************************/ | ||
| 26 : | |||
| 27 : | ; **************************************************************************** | ||
| 28 : | ; | ||
| 29 : | edgomez | 851 | ; Originally provided by Intel at AP-922 |
| 30 : | ; http://developer.intel.com/vtune/cbts/strmsimd/922down.htm | ||
| 31 : | ; (See more app notes at http://developer.intel.com/vtune/cbts/strmsimd/appnotes.htm) | ||
| 32 : | ; but in a limited edition. | ||
| 33 : | ; New macro implements a column part for precise iDCT | ||
| 34 : | edgomez | 1382 | ; The routine precision now satisfies IEEE standard 1180-1990. |
| 35 : | edgomez | 851 | ; |
| 36 : | edgomez | 1382 | ; Copyright(C) 2000-2001 Peter Gubanov <peter@elecard.net.ru> |
| 37 : | ; Rounding trick Copyright(C) 2000 Michel Lespinasse <walken@zoy.org> | ||
| 38 : | edgomez | 851 | ; |
| 39 : | ; http://www.elecard.com/peter/idct.html | ||
| 40 : | ; http://www.linuxvideo.org/mpeg2dec/ | ||
| 41 : | ; | ||
| 42 : | edgomez | 1382 | ; ***************************************************************************/ |
| 43 : | edgomez | 851 | ; |
| 44 : | ; These examples contain code fragments for first stage iDCT 8x8 | ||
| 45 : | ; (for rows) and first stage DCT 8x8 (for columns) | ||
| 46 : | ; | ||
| 47 : | |||
| 48 : | edgomez | 1382 | ; ***************************************************************************/ |
| 49 : | ; this 3dne function is compatible with iSSE, but is optimized specifically for | ||
| 50 : | ; K7 pipelines (ca 5% gain), for implementation details see the idct_mmx.asm | ||
| 51 : | ; file | ||
| 52 : | edgomez | 851 | ; |
| 53 : | edgomez | 1382 | ; ---------------------------------------------------------------------------- |
| 54 : | ; Athlon optimizations contributed by Jaan Kalda | ||
| 55 : | ;----------------------------------------------------------------------------- | ||
| 56 : | edgomez | 851 | |
| 57 : | edgomez | 1382 | BITS 32 |
| 58 : | edgomez | 851 | |
| 59 : | edgomez | 1382 | ;============================================================================= |
| 60 : | ; Macros and other preprocessor constants | ||
| 61 : | ;============================================================================= | ||
| 62 : | |||
| 63 : | %macro cglobal 1 | ||
| 64 : | edgomez | 851 | %ifdef PREFIX |
| 65 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 66 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
| 67 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 68 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
| 69 : | edgomez | 1535 | %else |
| 70 : | global _%1 | ||
| 71 : | %define %1 _%1 | ||
| 72 : | Isibaar | 1793 | %define ENDFUNC |
| 73 : | edgomez | 1535 | %endif |
| 74 : | edgomez | 851 | %else |
| 75 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 76 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
| 77 : | Isibaar | 1793 | %define ENDFUNC .endfunc |
| 78 : | edgomez | 1535 | %else |
| 79 : | global %1 | ||
| 80 : | Isibaar | 1793 | %define ENDFUNC |
| 81 : | edgomez | 1535 | %endif |
| 82 : | edgomez | 851 | %endif |
| 83 : | %endmacro | ||
| 84 : | |||
| 85 : | edgomez | 1382 | %define BITS_INV_ACC 5 ; 4 or 5 for IEEE |
| 86 : | edgomez | 851 | %define SHIFT_INV_ROW 16 - BITS_INV_ACC |
| 87 : | %define SHIFT_INV_COL 1 + BITS_INV_ACC | ||
| 88 : | %define RND_INV_ROW 1024 * (6 - BITS_INV_ACC) ; 1 << (SHIFT_INV_ROW-1) | ||
| 89 : | %define RND_INV_COL 16 * (BITS_INV_ACC - 3) ; 1 << (SHIFT_INV_COL-1) | ||
| 90 : | edgomez | 1382 | %define RND_INV_CORR RND_INV_COL - 1 ; correction -1.0 and round |
| 91 : | edgomez | 851 | |
| 92 : | edgomez | 1382 | %define BITS_FRW_ACC 3 ; 2 or 3 for accuracy |
| 93 : | edgomez | 851 | %define SHIFT_FRW_COL BITS_FRW_ACC |
| 94 : | %define SHIFT_FRW_ROW BITS_FRW_ACC + 17 | ||
| 95 : | edgomez | 1382 | %define RND_FRW_ROW 262144*(BITS_FRW_ACC - 1) ; 1 << (SHIFT_FRW_ROW-1) |
| 96 : | edgomez | 851 | |
| 97 : | edgomez | 1382 | ;============================================================================= |
| 98 : | ; Local Data (Read Only) | ||
| 99 : | ;============================================================================= | ||
| 100 : | |||
| 101 : | edgomez | 851 | %ifdef FORMAT_COFF |
| 102 : | edgomez | 1519 | SECTION .rodata |
| 103 : | edgomez | 851 | %else |
| 104 : | edgomez | 1519 | SECTION .rodata align=16 |
| 105 : | edgomez | 851 | %endif |
| 106 : | |||
| 107 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 108 : | ; Various memory constants (trigonometric values or rounding values) | ||
| 109 : | ;----------------------------------------------------------------------------- | ||
| 110 : | edgomez | 851 | |
| 111 : | edgomez | 1382 | ALIGN 16 |
| 112 : | one_corr: | ||
| 113 : | dw 1, 1, 1, 1 | ||
| 114 : | round_inv_row: | ||
| 115 : | dd RND_INV_ROW, RND_INV_ROW | ||
| 116 : | round_inv_col: | ||
| 117 : | dw RND_INV_COL, RND_INV_COL, RND_INV_COL, RND_INV_COL | ||
| 118 : | round_inv_corr: | ||
| 119 : | dw RND_INV_CORR, RND_INV_CORR, RND_INV_CORR, RND_INV_CORR | ||
| 120 : | round_frw_row: | ||
| 121 : | dd RND_FRW_ROW, RND_FRW_ROW | ||
| 122 : | tg_1_16: | ||
| 123 : | dw 13036, 13036, 13036, 13036 ; tg * (2<<16) + 0.5 | ||
| 124 : | tg_2_16: | ||
| 125 : | dw 27146, 27146, 27146, 27146 ; tg * (2<<16) + 0.5 | ||
| 126 : | tg_3_16: | ||
| 127 : | dw -21746, -21746, -21746, -21746 ; tg * (2<<16) + 0.5 | ||
| 128 : | cos_4_16: | ||
| 129 : | dw -19195, -19195, -19195, -19195 ; cos * (2<<16) + 0.5 | ||
| 130 : | ocos_4_16: | ||
| 131 : | dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 | ||
| 132 : | otg_3_16: | ||
| 133 : | dw 21895, 21895, 21895, 21895 ; tg * (2<<16) + 0.5 | ||
| 134 : | edgomez | 851 | |
| 135 : | %if SHIFT_INV_ROW == 12 ; assume SHIFT_INV_ROW == 12 | ||
| 136 : | edgomez | 1382 | rounder_0: |
| 137 : | dd 65536, 65536 | ||
| 138 : | rounder_4: | ||
| 139 : | dd 0, 0 | ||
| 140 : | rounder_1: | ||
| 141 : | dd 7195, 7195 | ||
| 142 : | rounder_7 | ||
| 143 : | dd 1024, 1024 | ||
| 144 : | rounder_2: | ||
| 145 : | dd 4520, 4520 | ||
| 146 : | rounder_6: | ||
| 147 : | dd 1024, 1024 | ||
| 148 : | rounder_3: | ||
| 149 : | dd 2407, 2407 | ||
| 150 : | rounder_5: | ||
| 151 : | dd 240, 240 | ||
| 152 : | edgomez | 851 | |
| 153 : | %elif SHIFT_INV_ROW == 11 ; assume SHIFT_INV_ROW == 11 | ||
| 154 : | edgomez | 1382 | rounder_0: |
| 155 : | dd 65536, 65536 | ||
| 156 : | rounder_4: | ||
| 157 : | dd 0, 0 | ||
| 158 : | rounder_1: | ||
| 159 : | dd 3597, 3597 | ||
| 160 : | rounder_7: | ||
| 161 : | dd 512, 512 | ||
| 162 : | rounder_2: | ||
| 163 : | dd 2260, 2260 | ||
| 164 : | rounder_6: | ||
| 165 : | dd 512, 512 | ||
| 166 : | rounder_3: | ||
| 167 : | dd 1203, 1203 | ||
| 168 : | rounder_5: | ||
| 169 : | dd 120, 120 | ||
| 170 : | edgomez | 851 | %else |
| 171 : | |||
| 172 : | edgomez | 1382 | %error invalid SHIFT_INV_ROW |
| 173 : | edgomez | 851 | |
| 174 : | %endif | ||
| 175 : | |||
| 176 : | ;----------------------------------------------------------------------------- | ||
| 177 : | edgomez | 1382 | ; Tables for xmm processors |
| 178 : | edgomez | 851 | ;----------------------------------------------------------------------------- |
| 179 : | |||
| 180 : | ; %3 for rows 0,4 - constants are multiplied by cos_4_16 | ||
| 181 : | edgomez | 1382 | tab_i_04_xmm: |
| 182 : | dw 16384, 21407, 16384, 8867 ; movq-> w05 w04 w01 w00 | ||
| 183 : | dw 16384, 8867, -16384, -21407 ; w07 w06 w03 w02 | ||
| 184 : | dw 16384, -8867, 16384, -21407 ; w13 w12 w09 w08 | ||
| 185 : | dw -16384, 21407, 16384, -8867 ; w15 w14 w11 w10 | ||
| 186 : | dw 22725, 19266, 19266, -4520 ; w21 w20 w17 w16 | ||
| 187 : | dw 12873, 4520, -22725, -12873 ; w23 w22 w19 w18 | ||
| 188 : | dw 12873, -22725, 4520, -12873 ; w29 w28 w25 w24 | ||
| 189 : | dw 4520, 19266, 19266, -22725 ; w31 w30 w27 w26 | ||
| 190 : | edgomez | 851 | |
| 191 : | ; %3 for rows 1,7 - constants are multiplied by cos_1_16 | ||
| 192 : | edgomez | 1382 | tab_i_17_xmm: |
| 193 : | dw 22725, 29692, 22725, 12299 ; movq-> w05 w04 w01 w00 | ||
| 194 : | dw 22725, 12299, -22725, -29692 ; w07 w06 w03 w02 | ||
| 195 : | dw 22725, -12299, 22725, -29692 ; w13 w12 w09 w08 | ||
| 196 : | dw -22725, 29692, 22725, -12299 ; w15 w14 w11 w10 | ||
| 197 : | dw 31521, 26722, 26722, -6270 ; w21 w20 w17 w16 | ||
| 198 : | dw 17855, 6270, -31521, -17855 ; w23 w22 w19 w18 | ||
| 199 : | dw 17855, -31521, 6270, -17855 ; w29 w28 w25 w24 | ||
| 200 : | dw 6270, 26722, 26722, -31521 ; w31 w30 w27 w26 | ||
| 201 : | edgomez | 851 | |
| 202 : | ; %3 for rows 2,6 - constants are multiplied by cos_2_16 | ||
| 203 : | edgomez | 1382 | tab_i_26_xmm: |
| 204 : | dw 21407, 27969, 21407, 11585 ; movq-> w05 w04 w01 w00 | ||
| 205 : | dw 21407, 11585, -21407, -27969 ; w07 w06 w03 w02 | ||
| 206 : | dw 21407, -11585, 21407, -27969 ; w13 w12 w09 w08 | ||
| 207 : | dw -21407, 27969, 21407, -11585 ; w15 w14 w11 w10 | ||
| 208 : | dw 29692, 25172, 25172, -5906 ; w21 w20 w17 w16 | ||
| 209 : | dw 16819, 5906, -29692, -16819 ; w23 w22 w19 w18 | ||
| 210 : | dw 16819, -29692, 5906, -16819 ; w29 w28 w25 w24 | ||
| 211 : | dw 5906, 25172, 25172, -29692 ; w31 w30 w27 w26 | ||
| 212 : | edgomez | 851 | |
| 213 : | ; %3 for rows 3,5 - constants are multiplied by cos_3_16 | ||
| 214 : | edgomez | 1382 | tab_i_35_xmm: |
| 215 : | dw 19266, 25172, 19266, 10426 ; movq-> w05 w04 w01 w00 | ||
| 216 : | dw 19266, 10426, -19266, -25172 ; w07 w06 w03 w02 | ||
| 217 : | dw 19266, -10426, 19266, -25172 ; w13 w12 w09 w08 | ||
| 218 : | dw -19266, 25172, 19266, -10426 ; w15 w14 w11 w10 | ||
| 219 : | dw 26722, 22654, 22654, -5315 ; w21 w20 w17 w16 | ||
| 220 : | dw 15137, 5315, -26722, -15137 ; w23 w22 w19 w18 | ||
| 221 : | dw 15137, -26722, 5315, -15137 ; w29 w28 w25 w24 | ||
| 222 : | dw 5315, 22654, 22654, -26722 ; w31 w30 w27 w26 | ||
| 223 : | edgomez | 851 | |
| 224 : | ;============================================================================= | ||
| 225 : | edgomez | 1382 | ; Code |
| 226 : | edgomez | 851 | ;============================================================================= |
| 227 : | |||
| 228 : | edgomez | 1382 | SECTION .text |
| 229 : | |||
| 230 : | cglobal idct_3dne | ||
| 231 : | |||
| 232 : | edgomez | 851 | ;----------------------------------------------------------------------------- |
| 233 : | edgomez | 1382 | ; void idct_3dne(uint16_t block[64]); |
| 234 : | edgomez | 851 | ;----------------------------------------------------------------------------- |
| 235 : | |||
| 236 : | edgomez | 1382 | ALIGN 16 |
| 237 : | idct_3dne: | ||
| 238 : | mov eax, [esp+4] | ||
| 239 : | edgomez | 851 | |
| 240 : | edgomez | 1382 | ; DCT_8_INV_ROW_1_s [eax+64], [eax+64], tab_i_04_sse, rounder_4 ;rounder_4=0 |
| 241 : | pshufw mm0, [eax+64],10001000b ; x2 x0 x2 x0 | ||
| 242 : | movq mm3, [tab_i_04_xmm] ; 3 ; w05 w04 w01 w00 | ||
| 243 : | pshufw mm1, [eax+64+8],10001000b ; x6 x4 x6 x4 | ||
| 244 : | movq mm4, [tab_i_04_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
| 245 : | pshufw mm2, [eax+64],11011101b ; x3 x1 x3 x1 | ||
| 246 : | pshufw mm5, [eax+64+8],11011101b ; x7 x5 x7 x5 | ||
| 247 : | movq mm6, [tab_i_04_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
| 248 : | pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
| 249 : | movq mm7, [tab_i_04_xmm+40] ; 7 ; w23 w22 w19 w18 ; | ||
| 250 : | pmaddwd mm0, [tab_i_04_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 | ||
| 251 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
| 252 : | pmaddwd mm1, [tab_i_04_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 | ||
| 253 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
| 254 : | pmaddwd mm2, [tab_i_04_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 | ||
| 255 : | pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
| 256 : | pmaddwd mm5, [tab_i_04_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 | ||
| 257 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
| 258 : | paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
| 259 : | pshufw mm1, [eax+80+8],10001000b ; x6 x4 x6 x4 | ||
| 260 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
| 261 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
| 262 : | paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
| 263 : | pshufw mm5, [eax+80],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle | ||
| 264 : | movq mm7, mm0 ; 7 ; a3 a2 | ||
| 265 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
| 266 : | paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
| 267 : | movq mm3, [tab_i_35_xmm] ; 3 ; w05 w04 w01 w00 | ||
| 268 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
| 269 : | paddd mm0, mm2 ; 0 free a3+b3 a2+b2 | ||
| 270 : | pshufw mm2, [eax+80],11011101b; x3 x1 x3 x1 | ||
| 271 : | pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
| 272 : | pmaddwd mm5, [tab_i_35_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 | ||
| 273 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
| 274 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
| 275 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
| 276 : | psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
| 277 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
| 278 : | packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 | ||
| 279 : | pshufw mm0, [eax+80+8],11011101b ; x7 x5 x7 x5 | ||
| 280 : | movq [eax+64], mm6 ; 3 ; save y3 y2 y1 y0 stall2 | ||
| 281 : | edgomez | 851 | |
| 282 : | edgomez | 1382 | ; DCT_8_INV_ROW_1_s [eax+80], [eax+80], tab_i_35_xmm, rounder_5 |
| 283 : | movq mm4, [tab_i_35_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
| 284 : | movq mm6, [tab_i_35_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
| 285 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 | ||
| 286 : | paddd mm3, [rounder_5] ; +rounder stall 6 | ||
| 287 : | paddd mm5, [rounder_5] ; +rounder | ||
| 288 : | movq [eax+64+8], mm7 ; 7 ; save y7 y6 y5 y4 | ||
| 289 : | movq mm7, [tab_i_35_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
| 290 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
| 291 : | pmaddwd mm1, [tab_i_35_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 | ||
| 292 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
| 293 : | pmaddwd mm2, [tab_i_35_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 | ||
| 294 : | pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
| 295 : | pmaddwd mm0, [tab_i_35_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 | ||
| 296 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
| 297 : | paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
| 298 : | pshufw mm1, [eax+96+8],10001000b ; x6 x4 x6 x4 | ||
| 299 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
| 300 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
| 301 : | paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
| 302 : | pshufw mm0, [eax+96],10001000b ; x2 x0 x2 x0 | ||
| 303 : | movq mm7, mm5 ; 7 ; a3 a2 | ||
| 304 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 | ||
| 305 : | paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
| 306 : | movq mm3, [tab_i_26_xmm] ; 3 ; w05 w04 w01 w00 | ||
| 307 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
| 308 : | paddd mm5, mm2 ; 0 free a3+b3 a2+b2 | ||
| 309 : | pshufw mm2, [eax+96],11011101b; x3 x1 x3 x1 | ||
| 310 : | pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
| 311 : | pmaddwd mm0, [tab_i_26_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 | ||
| 312 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
| 313 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
| 314 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
| 315 : | psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
| 316 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
| 317 : | packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 | ||
| 318 : | pshufw mm5, [eax+96+8],11011101b ; x7 x5 x7 x5 | ||
| 319 : | movq [eax+80], mm6 ; 3 ; save y3 y2 y1 y0 | ||
| 320 : | edgomez | 851 | |
| 321 : | edgomez | 1382 | ; DCT_8_INV_ROW_1_s [eax+96], [eax+96], tab_i_26_xmm, rounder_6 |
| 322 : | movq mm4, [tab_i_26_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
| 323 : | movq mm6, [tab_i_26_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
| 324 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 | ||
| 325 : | paddd mm3, [rounder_6] ; +rounder | ||
| 326 : | paddd mm0, [rounder_6] ; +rounder | ||
| 327 : | movq [eax+80+8], mm7 ; 7 ; save y7 y6 | ||
| 328 : | movq mm7, [tab_i_26_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
| 329 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
| 330 : | pmaddwd mm1, [tab_i_26_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 | ||
| 331 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
| 332 : | pmaddwd mm2, [tab_i_26_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 | ||
| 333 : | pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
| 334 : | pmaddwd mm5, [tab_i_26_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 | ||
| 335 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
| 336 : | paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
| 337 : | pshufw mm1, [eax+112+8],10001000b ; x6 x4 x6 x4 | ||
| 338 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
| 339 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
| 340 : | paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
| 341 : | pshufw mm5, [eax+112],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle | ||
| 342 : | movq mm7, mm0 ; 7 ; a3 a2 | ||
| 343 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
| 344 : | paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
| 345 : | movq mm3, [tab_i_17_xmm] ; 3 ; w05 w04 w01 w00 | ||
| 346 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
| 347 : | paddd mm0, mm2 ; 0 free a3+b3 a2+b2 | ||
| 348 : | pshufw mm2, [eax+112],11011101b; x3 x1 x3 x1 | ||
| 349 : | pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
| 350 : | pmaddwd mm5, [tab_i_17_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 | ||
| 351 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
| 352 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
| 353 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
| 354 : | psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
| 355 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
| 356 : | packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 | ||
| 357 : | pshufw mm0, [eax+112+8],11011101b ; x7 x5 x7 x5 | ||
| 358 : | movq [eax+96], mm6 ; 3 ; save y3 y2 y1 y0 stall2 | ||
| 359 : | edgomez | 851 | |
| 360 : | edgomez | 1382 | ; DCT_8_INV_ROW_1_s [eax+112], [eax+112], tab_i_17_xmm, rounder_7 |
| 361 : | movq mm4, [tab_i_17_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
| 362 : | movq mm6, [tab_i_17_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
| 363 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 | ||
| 364 : | paddd mm3, [rounder_7] ; +rounder stall 6 | ||
| 365 : | paddd mm5, [rounder_7] ; +rounder | ||
| 366 : | movq [eax+96+8], mm7 ; 7 ; save y7 y6 y5 y4 | ||
| 367 : | movq mm7, [tab_i_17_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
| 368 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
| 369 : | pmaddwd mm1, [tab_i_17_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 | ||
| 370 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
| 371 : | pmaddwd mm2, [tab_i_17_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 | ||
| 372 : | pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
| 373 : | pmaddwd mm0, [tab_i_17_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 | ||
| 374 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
| 375 : | paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
| 376 : | pshufw mm1, [eax+0+8],10001000b; x6 x4 x6 x4 | ||
| 377 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
| 378 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
| 379 : | paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
| 380 : | pshufw mm0, [eax+0],10001000b ; x2 x0 x2 x0 | ||
| 381 : | movq mm7, mm5 ; 7 ; a3 a2 | ||
| 382 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 | ||
| 383 : | paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
| 384 : | movq mm3, [tab_i_04_xmm] ; 3 ; w05 w04 w01 w00 | ||
| 385 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
| 386 : | paddd mm5, mm2 ; 0 free a3+b3 a2+b2 | ||
| 387 : | pshufw mm2, [eax+0],11011101b ; x3 x1 x3 x1 | ||
| 388 : | pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
| 389 : | pmaddwd mm0, [tab_i_04_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 | ||
| 390 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
| 391 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
| 392 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
| 393 : | psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
| 394 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
| 395 : | packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 | ||
| 396 : | pshufw mm5, [eax+0+8],11011101b; x7 x5 x7 x5 | ||
| 397 : | movq [eax+112], mm6 ; 3 ; save y3 y2 y1 y0 | ||
| 398 : | edgomez | 851 | |
| 399 : | edgomez | 1382 | ; DCT_8_INV_ROW_1_s [eax+0], 0, tab_i_04_xmm, rounder_0 |
| 400 : | movq mm4, [tab_i_04_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
| 401 : | movq mm6, [tab_i_04_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
| 402 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 | ||
| 403 : | paddd mm3, [rounder_0] ; +rounder | ||
| 404 : | paddd mm0, [rounder_0] ; +rounder | ||
| 405 : | movq [eax+112+8], mm7 ; 7 ; save y7 y6 | ||
| 406 : | movq mm7, [tab_i_04_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
| 407 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
| 408 : | pmaddwd mm1, [tab_i_04_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 | ||
| 409 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
| 410 : | pmaddwd mm2, [tab_i_04_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 | ||
| 411 : | pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
| 412 : | pmaddwd mm5, [tab_i_04_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 | ||
| 413 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
| 414 : | paddd mm0, mm1 ; 1 | ||
| 415 : | pshufw mm1, [eax+16+8],10001000b ; x6 x4 x6 x4 | ||
| 416 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
| 417 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
| 418 : | paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
| 419 : | pshufw mm5, [eax+16],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle | ||
| 420 : | movq mm7, mm0 ; 7 ; a3 a2 | ||
| 421 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
| 422 : | paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
| 423 : | movq mm3, [tab_i_17_xmm] ; 3 ; w05 w04 w01 w00 | ||
| 424 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
| 425 : | paddd mm0, mm2 ; 0 free a3+b3 a2+b2 | ||
| 426 : | pshufw mm2, [eax+16],11011101b; x3 x1 x3 x1 | ||
| 427 : | pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
| 428 : | pmaddwd mm5, [tab_i_17_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 | ||
| 429 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
| 430 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
| 431 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
| 432 : | psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
| 433 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
| 434 : | packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 | ||
| 435 : | pshufw mm0, [eax+16+8],11011101b ; x7 x5 x7 x5 | ||
| 436 : | movq [eax+0], mm6 ; 3 ; save y3 y2 y1 y0 stall2 | ||
| 437 : | edgomez | 851 | |
| 438 : | edgomez | 1382 | ; DCT_8_INV_ROW_1_s [eax+16], 16, tab_i_17_xmm, rounder_1 |
| 439 : | movq mm4, [tab_i_17_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
| 440 : | movq mm6, [tab_i_17_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
| 441 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 | ||
| 442 : | paddd mm3, [rounder_1] ; +rounder stall 6 | ||
| 443 : | paddd mm5, [rounder_1] ; +rounder | ||
| 444 : | movq [eax+0+8], mm7 ; 7 ; save y7 y6 y5 y4 | ||
| 445 : | movq mm7, [tab_i_17_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
| 446 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
| 447 : | pmaddwd mm1, [tab_i_17_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 | ||
| 448 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
| 449 : | pmaddwd mm2, [tab_i_17_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 | ||
| 450 : | pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
| 451 : | pmaddwd mm0, [tab_i_17_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 | ||
| 452 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
| 453 : | paddd mm5, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
| 454 : | pshufw mm1, [eax+32+8],10001000b ; x6 x4 x6 x4 | ||
| 455 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
| 456 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
| 457 : | paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
| 458 : | pshufw mm0, [eax+32],10001000b; x2 x0 x2 x0 | ||
| 459 : | movq mm7, mm5 ; 7 ; a3 a2 | ||
| 460 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 stall 5 | ||
| 461 : | paddd mm6, mm3 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
| 462 : | movq mm3, [tab_i_26_xmm] ; 3 ; w05 w04 w01 w00 | ||
| 463 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
| 464 : | paddd mm5, mm2 ; 0 free a3+b3 a2+b2 | ||
| 465 : | pshufw mm2, [eax+32],11011101b; x3 x1 x3 x1 | ||
| 466 : | pmaddwd mm3, mm0 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
| 467 : | pmaddwd mm0, [tab_i_26_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 | ||
| 468 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
| 469 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
| 470 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
| 471 : | psrad mm5, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
| 472 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
| 473 : | packssdw mm6, mm5 ; 0 free ; y3 y2 y1 y0 | ||
| 474 : | pshufw mm5, [eax+32+8],11011101b ; x7 x5 x7 x5 | ||
| 475 : | movq [eax+16], mm6 ; 3 ; save y3 y2 y1 y0 | ||
| 476 : | edgomez | 851 | |
| 477 : | edgomez | 1382 | ; DCT_8_INV_ROW_1_s [eax+32], 32, tab_i_26_xmm, rounder_2 |
| 478 : | movq mm4, [tab_i_26_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
| 479 : | movq mm6, [tab_i_26_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
| 480 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 STALL 6 | ||
| 481 : | paddd mm3, [rounder_2] ; +rounder | ||
| 482 : | paddd mm0, [rounder_2] ; +rounder | ||
| 483 : | movq [eax+16+8], mm7 ; 7 ; save y7 y6 | ||
| 484 : | movq mm7, [tab_i_26_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
| 485 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
| 486 : | pmaddwd mm1, [tab_i_26_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 | ||
| 487 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
| 488 : | pmaddwd mm2, [tab_i_26_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 | ||
| 489 : | pmaddwd mm7, mm5 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
| 490 : | pmaddwd mm5, [tab_i_26_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 | ||
| 491 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
| 492 : | paddd mm0, mm1 ; 1 free ; a3=sum(even3) a2=sum(even2) | ||
| 493 : | pshufw mm1, [eax+48+8],10001000b ; x6 x4 x6 x4 | ||
| 494 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
| 495 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
| 496 : | paddd mm2, mm5 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
| 497 : | pshufw mm5, [eax+48],10001000b; x2 x0 x2 x0 mm5 & mm0 exchanged for next cycle | ||
| 498 : | movq mm7, mm0 ; 7 ; a3 a2 | ||
| 499 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
| 500 : | paddd mm6, mm3 ; mm6 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
| 501 : | movq mm3, [tab_i_35_xmm] ; 3 ; w05 w04 w01 w00 | ||
| 502 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
| 503 : | paddd mm0, mm2 ; 0 free a3+b3 a2+b2 | ||
| 504 : | pshufw mm2, [eax+48],11011101b; x3 x1 x3 x1 | ||
| 505 : | pmaddwd mm3, mm5 ; x2*w05+x0*w04 x2*w01+x0*w00 | ||
| 506 : | pmaddwd mm5, [tab_i_35_xmm+16]; x2*w13+x0*w12 x2*w09+x0*w08 | ||
| 507 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
| 508 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
| 509 : | psrad mm6, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
| 510 : | psrad mm0, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
| 511 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
| 512 : | packssdw mm6, mm0 ; 0 free ; y3 y2 y1 y0 | ||
| 513 : | pshufw mm0, [eax+48+8],11011101b ; x7 x5 x7 x5 | ||
| 514 : | movq [eax+32], mm6 ; 3 ; save y3 y2 y1 y0 stall2 | ||
| 515 : | edgomez | 851 | |
| 516 : | edgomez | 1382 | ; DCT_8_INV_ROW_1_s [eax+48], [eax+48], tab_i_35_xmm, rounder_3 |
| 517 : | movq mm4, [tab_i_35_xmm+8] ; 4 ; w07 w06 w03 w02 | ||
| 518 : | movq mm6, [tab_i_35_xmm+32] ; 6 ; w21 w20 w17 w16 | ||
| 519 : | pshufw mm7, mm7, 10110001b ; y7 y6 y5 y4 | ||
| 520 : | paddd mm3, [rounder_3] ; +rounder stall 6 | ||
| 521 : | paddd mm5, [rounder_3] ; +rounder | ||
| 522 : | movq [eax+32+8], mm7 ; 7 ; save y7 y6 y5 y4 | ||
| 523 : | movq mm7, [tab_i_35_xmm+40] ; 7 ; w23 w22 w19 w18 | ||
| 524 : | pmaddwd mm4, mm1 ; x6*w07+x4*w06 x6*w03+x4*w02 | ||
| 525 : | pmaddwd mm1, [tab_i_35_xmm+24]; x6*w15+x4*w14 x6*w11+x4*w10 | ||
| 526 : | pmaddwd mm6, mm2 ; x3*w21+x1*w20 x3*w17+x1*w16 | ||
| 527 : | pmaddwd mm2, [tab_i_35_xmm+48]; x3*w29+x1*w28 x3*w25+x1*w24 | ||
| 528 : | pmaddwd mm7, mm0 ; 7 ; x7*w23+x5*w22 x7*w19+x5*w18 ; w23 w22 w19 w18 | ||
| 529 : | pmaddwd mm0, [tab_i_35_xmm+56]; x7*w31+x5*w30 x7*w27+x5*w26 | ||
| 530 : | paddd mm3, mm4 ; 4 free ; a1=sum(even1) a0=sum(even0) | ||
| 531 : | paddd mm5, mm1 ; mm1 free ; a3=sum(even3) a2=sum(even2) | ||
| 532 : | movq mm1, [tg_3_16] | ||
| 533 : | movq mm4, mm3 ; 4 ; a1 a0 | ||
| 534 : | paddd mm6, mm7 ; 7 free ; b1=sum(odd1) b0=sum(odd0) | ||
| 535 : | paddd mm2, mm0 ; 5 free ; b3=sum(odd3) b2=sum(odd2) | ||
| 536 : | movq mm0, [tg_3_16] | ||
| 537 : | movq mm7, mm5 ; 7 ; a3 a2 | ||
| 538 : | psubd mm4, mm6 ; 6 free ; a1-b1 a0-b0 | ||
| 539 : | paddd mm3, mm6 ; mm3 = mm3+mm6+mm5+mm4; a1+b1 a0+b0 | ||
| 540 : | psubd mm7, mm2 ; ; a3-b3 a2-b2 | ||
| 541 : | paddd mm2, mm5 ; 0 free a3+b3 a2+b2 | ||
| 542 : | movq mm5, [eax+16*5] | ||
| 543 : | psrad mm4, SHIFT_INV_ROW ; y6=a1-b1 y7=a0-b0 | ||
| 544 : | psrad mm7, SHIFT_INV_ROW ; y4=a3-b3 y5=a2-b2 | ||
| 545 : | psrad mm3, SHIFT_INV_ROW ; y1=a1+b1 y0=a0+b0 | ||
| 546 : | psrad mm2, SHIFT_INV_ROW ; y3=a3+b3 y2=a2+b2 | ||
| 547 : | movq mm6, [eax+16*1] | ||
| 548 : | packssdw mm7, mm4 ; 4 ; y6 y7 y4 y5 | ||
| 549 : | movq mm4, [tg_1_16] | ||
| 550 : | packssdw mm3, mm2 ; 0 free ; y3 y2 y1 y0 | ||
| 551 : | pshufw mm2, mm7, 10110001b ; y7 y6 y5 y4 | ||
| 552 : | edgomez | 851 | |
| 553 : | edgomez | 1382 | ; DCT_8_INV_COL_4 [eax+0],[eax+0] |
| 554 : | ; movq mm3,mmword ptr [eax+16*3] | ||
| 555 : | movq mm7, [eax+16*7] | ||
| 556 : | pmulhw mm0, mm3 ; x3*(tg_3_16-1) | ||
| 557 : | pmulhw mm1, mm5 ; x5*(tg_3_16-1) | ||
| 558 : | movq [eax+48+8], mm2 ; 7 ; save y7 y6 y5 y4 | ||
| 559 : | movq mm2, mm4 ; tg_1_16 | ||
| 560 : | pmulhw mm4, mm7 ; x7*tg_1_16 | ||
| 561 : | paddsw mm0, mm3 ; x3*tg_3_16 | ||
| 562 : | pmulhw mm2, mm6 ; x1*tg_1_16 | ||
| 563 : | paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) | ||
| 564 : | psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 | ||
| 565 : | movq [eax+48], mm3 ; 3 ; save y3 y2 y1 y0 | ||
| 566 : | movq mm3, [ocos_4_16] | ||
| 567 : | paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 | ||
| 568 : | paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 | ||
| 569 : | psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 | ||
| 570 : | movq mm5, mm4 ; tp17 | ||
| 571 : | movq mm6, mm2 ; tm17 | ||
| 572 : | paddsw mm5, mm1 ; tp17+tp35 = b0 | ||
| 573 : | psubsw mm6, mm0 ; tm17-tm35 = b3 | ||
| 574 : | psubsw mm4, mm1 ; tp17-tp35 = t1 | ||
| 575 : | paddsw mm2, mm0 ; tm17+tm35 = t2 | ||
| 576 : | movq mm7, [tg_2_16] | ||
| 577 : | movq mm1, mm4 ; t1 | ||
| 578 : | movq [eax+3*16], mm5 ; save b0 | ||
| 579 : | paddsw mm1, mm2 ; t1+t2 | ||
| 580 : | movq [eax+5*16], mm6 ; save b3 | ||
| 581 : | psubsw mm4, mm2 ; t1-t2 | ||
| 582 : | movq mm5, [eax+2*16] | ||
| 583 : | movq mm0, mm7 ; tg_2_16 | ||
| 584 : | movq mm6, [eax+6*16] | ||
| 585 : | pmulhw mm0, mm5 ; x2*tg_2_16 | ||
| 586 : | pmulhw mm7, mm6 ; x6*tg_2_16 | ||
| 587 : | edgomez | 851 | ; slot |
| 588 : | edgomez | 1382 | pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
| 589 : | edgomez | 851 | ; slot |
| 590 : | edgomez | 1382 | movq mm2, [eax+0*16] |
| 591 : | pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 | ||
| 592 : | psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 | ||
| 593 : | movq mm3, [eax+0*16] ; x0 | ||
| 594 : | movq mm6, [eax+4*16] | ||
| 595 : | paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 | ||
| 596 : | paddsw mm2, mm6 ; x0+x4 = tp04 | ||
| 597 : | psubsw mm3, mm6 ; x0-x4 = tm04 | ||
| 598 : | movq mm5, mm2 ; tp04 | ||
| 599 : | movq mm6, mm3 ; tm04 | ||
| 600 : | psubsw mm2, mm7 ; tp04-tp26 = a3 | ||
| 601 : | paddsw mm3, mm0 ; tm04+tm26 = a1 | ||
| 602 : | paddsw mm1, mm1 ; b1 | ||
| 603 : | paddsw mm4, mm4 ; b2 | ||
| 604 : | paddsw mm5, mm7 ; tp04+tp26 = a0 | ||
| 605 : | psubsw mm6, mm0 ; tm04-tm26 = a2 | ||
| 606 : | movq mm7, mm3 ; a1 | ||
| 607 : | movq mm0, mm6 ; a2 | ||
| 608 : | paddsw mm3, mm1 ; a1+b1 | ||
| 609 : | paddsw mm6, mm4 ; a2+b2 | ||
| 610 : | psraw mm3, SHIFT_INV_COL ; dst1 | ||
| 611 : | psubsw mm7, mm1 ; a1-b1 | ||
| 612 : | psraw mm6, SHIFT_INV_COL ; dst2 | ||
| 613 : | psubsw mm0, mm4 ; a2-b2 | ||
| 614 : | movq mm1, [eax+3*16] ; load b0 | ||
| 615 : | psraw mm7, SHIFT_INV_COL ; dst6 | ||
| 616 : | movq mm4, mm5 ; a0 | ||
| 617 : | psraw mm0, SHIFT_INV_COL ; dst5 | ||
| 618 : | movq [eax+1*16], mm3 | ||
| 619 : | paddsw mm5, mm1 ; a0+b0 | ||
| 620 : | movq [eax+2*16], mm6 | ||
| 621 : | psubsw mm4, mm1 ; a0-b0 | ||
| 622 : | movq mm3, [eax+5*16] ; load b3 | ||
| 623 : | psraw mm5, SHIFT_INV_COL ; dst0 | ||
| 624 : | movq mm6, mm2 ; a3 | ||
| 625 : | psraw mm4, SHIFT_INV_COL ; dst7 | ||
| 626 : | movq [eax+5*16], mm0 | ||
| 627 : | movq mm0, [tg_3_16] | ||
| 628 : | paddsw mm2, mm3 ; a3+b3 | ||
| 629 : | movq [eax+6*16], mm7 | ||
| 630 : | psubsw mm6, mm3 ; a3-b3 | ||
| 631 : | movq mm3, [eax+8+16*3] | ||
| 632 : | movq [eax+0*16], mm5 | ||
| 633 : | psraw mm2, SHIFT_INV_COL ; dst3 | ||
| 634 : | movq [eax+7*16], mm4 | ||
| 635 : | edgomez | 851 | |
| 636 : | edgomez | 1382 | ; DCT_8_INV_COL_4 [eax+8],[eax+8] |
| 637 : | movq mm1, mm0 ; tg_3_16 | ||
| 638 : | movq mm5, [eax+8+16*5] | ||
| 639 : | psraw mm6, SHIFT_INV_COL ; dst4 | ||
| 640 : | pmulhw mm0, mm3 ; x3*(tg_3_16-1) | ||
| 641 : | movq mm4, [tg_1_16] | ||
| 642 : | pmulhw mm1, mm5 ; x5*(tg_3_16-1) | ||
| 643 : | movq mm7, [eax+8+16*7] | ||
| 644 : | movq [eax+3*16], mm2 | ||
| 645 : | movq mm2, mm4 ; tg_1_16 | ||
| 646 : | movq [eax+4*16], mm6 | ||
| 647 : | movq mm6, [eax+8+16*1] | ||
| 648 : | pmulhw mm4, mm7 ; x7*tg_1_16 | ||
| 649 : | paddsw mm0, mm3 ; x3*tg_3_16 | ||
| 650 : | pmulhw mm2, mm6 ; x1*tg_1_16 | ||
| 651 : | paddsw mm1, mm3 ; x3+x5*(tg_3_16-1) | ||
| 652 : | psubsw mm0, mm5 ; x3*tg_3_16-x5 = tm35 | ||
| 653 : | movq mm3, [ocos_4_16] | ||
| 654 : | paddsw mm1, mm5 ; x3+x5*tg_3_16 = tp35 | ||
| 655 : | paddsw mm4, mm6 ; x1+tg_1_16*x7 = tp17 | ||
| 656 : | psubsw mm2, mm7 ; x1*tg_1_16-x7 = tm17 | ||
| 657 : | movq mm5, mm4 ; tp17 | ||
| 658 : | movq mm6, mm2 ; tm17 | ||
| 659 : | paddsw mm5, mm1 ; tp17+tp35 = b0 | ||
| 660 : | psubsw mm4, mm1 ; tp17-tp35 = t1 | ||
| 661 : | paddsw mm2, mm0 ; tm17+tm35 = t2 | ||
| 662 : | movq mm7, [tg_2_16] | ||
| 663 : | movq mm1, mm4 ; t1 | ||
| 664 : | psubsw mm6, mm0 ; tm17-tm35 = b3 | ||
| 665 : | movq [eax+8+3*16], mm5 ; save b0 | ||
| 666 : | movq [eax+8+5*16], mm6 ; save b3 | ||
| 667 : | psubsw mm4, mm2 ; t1-t2 | ||
| 668 : | movq mm5, [eax+8+2*16] | ||
| 669 : | movq mm0, mm7 ; tg_2_16 | ||
| 670 : | movq mm6, [eax+8+6*16] | ||
| 671 : | paddsw mm1, mm2 ; t1+t2 | ||
| 672 : | pmulhw mm0, mm5 ; x2*tg_2_16 | ||
| 673 : | pmulhw mm7, mm6 ; x6*tg_2_16 | ||
| 674 : | movq mm2, [eax+8+0*16] | ||
| 675 : | pmulhw mm4, mm3 ; ocos_4_16*(t1-t2) = b2/2 | ||
| 676 : | psubsw mm0, mm6 ; t2*tg_2_16-x6 = tm26 | ||
| 677 : | edgomez | 851 | ; slot |
| 678 : | edgomez | 1382 | pmulhw mm1, mm3 ; ocos_4_16*(t1+t2) = b1/2 |
| 679 : | edgomez | 851 | ; slot |
| 680 : | edgomez | 1382 | movq mm3, [eax+8+0*16] ; x0 |
| 681 : | movq mm6, [eax+8+4*16] | ||
| 682 : | paddsw mm7, mm5 ; x2+x6*tg_2_16 = tp26 | ||
| 683 : | paddsw mm2, mm6 ; x0+x4 = tp04 | ||
| 684 : | psubsw mm3, mm6 ; x0-x4 = tm04 | ||
| 685 : | movq mm5, mm2 ; tp04 | ||
| 686 : | movq mm6, mm3 ; tm04 | ||
| 687 : | psubsw mm2, mm7 ; tp04-tp26 = a3 | ||
| 688 : | paddsw mm3, mm0 ; tm04+tm26 = a1 | ||
| 689 : | paddsw mm1, mm1 ; b1 | ||
| 690 : | paddsw mm4, mm4 ; b2 | ||
| 691 : | paddsw mm5, mm7 ; tp04+tp26 = a0 | ||
| 692 : | psubsw mm6, mm0 ; tm04-tm26 = a2 | ||
| 693 : | movq mm7, mm3 ; a1 | ||
| 694 : | movq mm0, mm6 ; a2 | ||
| 695 : | paddsw mm3, mm1 ; a1+b1 | ||
| 696 : | paddsw mm6, mm4 ; a2+b2 | ||
| 697 : | psraw mm3, SHIFT_INV_COL ; dst1 | ||
| 698 : | psubsw mm7, mm1 ; a1-b1 | ||
| 699 : | psraw mm6, SHIFT_INV_COL ; dst2 | ||
| 700 : | psubsw mm0, mm4 ; a2-b2 | ||
| 701 : | movq mm1, [eax+8+3*16] ; load b0 | ||
| 702 : | psraw mm7, SHIFT_INV_COL ; dst6 | ||
| 703 : | movq mm4, mm5 ; a0 | ||
| 704 : | psraw mm0, SHIFT_INV_COL ; dst5 | ||
| 705 : | movq [eax+8+1*16], mm3 | ||
| 706 : | paddsw mm5, mm1 ; a0+b0 | ||
| 707 : | movq [eax+8+2*16], mm6 | ||
| 708 : | psubsw mm4, mm1 ; a0-b0 | ||
| 709 : | movq mm3, [eax+8+5*16] ; load b3 | ||
| 710 : | psraw mm5, SHIFT_INV_COL ; dst0 | ||
| 711 : | movq mm6, mm2 ; a3 | ||
| 712 : | psraw mm4, SHIFT_INV_COL ; dst7 | ||
| 713 : | movq [eax+8+5*16], mm0 | ||
| 714 : | paddsw mm2, mm3 ; a3+b3 | ||
| 715 : | movq [eax+8+6*16], mm7 | ||
| 716 : | psubsw mm6, mm3 ; a3-b3 | ||
| 717 : | movq [eax+8+0*16], mm5 | ||
| 718 : | psraw mm2, SHIFT_INV_COL ; dst3 | ||
| 719 : | movq [eax+8+7*16], mm4 | ||
| 720 : | psraw mm6, SHIFT_INV_COL ; dst4 | ||
| 721 : | movq [eax+8+3*16], mm2 | ||
| 722 : | movq [eax+8+4*16], mm6 | ||
| 723 : | edgomez | 851 | |
| 724 : | edgomez | 1382 | ret |
| 725 : | Isibaar | 1793 | ENDFUNC |
| 726 : | edgomez | 1540 | |
| 727 : | Isibaar | 1790 | |
| 728 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 729 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 730 : | %endif | ||
| 731 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |