Parent Directory
| 
Revision Log
Revision 649 -
(view)
(download)
Original Path: trunk/xvidcore/src/dct/x86_asm/fdct_mmx.asm
| 1 : | chl | 432 | ;/***************************************************************************** |
| 2 : | ; * | ||
| 3 : | ; * XVID MPEG-4 VIDEO CODEC | ||
| 4 : | chl | 460 | ; * mmx version - fast discrete cosine transformation |
| 5 : | chl | 432 | ; * |
| 6 : | chl | 460 | ; * Initial version provided by Intel at AppNote AP-922 |
| 7 : | chl | 432 | ; * Copyright (C) 1999 Intel Corporation, |
| 8 : | chl | 460 | ; * |
| 9 : | ; * completed and corrected | ||
| 10 : | chl | 432 | ; * Copyright (C) 2000 - Royce Shih-Wea Liao <liaor@iname.com>, |
| 11 : | chl | 460 | ; * |
| 12 : | chl | 432 | ; * ported to NASM and some minor changes |
| 13 : | ; * Copyright (C) 2001 - Michael Militzer <isibaar@xvid.org> | ||
| 14 : | chl | 460 | ; * |
| 15 : | edgomez | 649 | ; * This file is part of XviD, a free MPEG-4 video encoder/decoder |
| 16 : | chl | 432 | ; * |
| 17 : | edgomez | 649 | ; * XviD is free software; you can redistribute it and/or modify it |
| 18 : | ; * under the terms of the GNU General Public License as published by | ||
| 19 : | chl | 432 | ; * the Free Software Foundation; either version 2 of the License, or |
| 20 : | ; * (at your option) any later version. | ||
| 21 : | ; * | ||
| 22 : | ; * This program is distributed in the hope that it will be useful, | ||
| 23 : | ; * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
| 24 : | ; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | ||
| 25 : | ; * GNU General Public License for more details. | ||
| 26 : | ; * | ||
| 27 : | ; * You should have received a copy of the GNU General Public License | ||
| 28 : | ; * along with this program; if not, write to the Free Software | ||
| 29 : | ; * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | ||
| 30 : | ; * | ||
| 31 : | edgomez | 649 | ; * Under section 8 of the GNU General Public License, the copyright |
| 32 : | ; * holders of XVID explicitly forbid distribution in the following | ||
| 33 : | ; * countries: | ||
| 34 : | ; * | ||
| 35 : | ; * - Japan | ||
| 36 : | ; * - United States of America | ||
| 37 : | ; * | ||
| 38 : | ; * Linking XviD statically or dynamically with other modules is making a | ||
| 39 : | ; * combined work based on XviD. Thus, the terms and conditions of the | ||
| 40 : | ; * GNU General Public License cover the whole combination. | ||
| 41 : | ; * | ||
| 42 : | ; * As a special exception, the copyright holders of XviD give you | ||
| 43 : | ; * permission to link XviD with independent modules that communicate with | ||
| 44 : | ; * XviD solely through the VFW1.1 and DShow interfaces, regardless of the | ||
| 45 : | ; * license terms of these independent modules, and to copy and distribute | ||
| 46 : | ; * the resulting combined work under terms of your choice, provided that | ||
| 47 : | ; * every copy of the combined work is accompanied by a complete copy of | ||
| 48 : | ; * the source code of XviD (the version of XviD used to produce the | ||
| 49 : | ; * combined work), being distributed under the terms of the GNU General | ||
| 50 : | ; * Public License plus this exception. An independent module is a module | ||
| 51 : | ; * which is not derived from or based on XviD. | ||
| 52 : | ; * | ||
| 53 : | ; * Note that people who make modified versions of XviD are not obligated | ||
| 54 : | ; * to grant this special exception for their modified versions; it is | ||
| 55 : | ; * their choice whether to do so. The GNU General Public License gives | ||
| 56 : | ; * permission to release a modified version without this exception; this | ||
| 57 : | ; * exception also makes it possible to release a modified version which | ||
| 58 : | ; * carries forward this exception. | ||
| 59 : | ; * | ||
| 60 : | ; * $Id: fdct_mmx.asm,v 1.6 2002-11-16 23:51:58 edgomez Exp $ | ||
| 61 : | ; * | ||
| 62 : | chl | 460 | ; *************************************************************************/ |
| 63 : | Isibaar | 154 | |
| 64 : | BITS 32 | ||
| 65 : | |||
| 66 : | %macro cglobal 1 | ||
| 67 : | %ifdef PREFIX | ||
| 68 : | global _%1 | ||
| 69 : | %define %1 _%1 | ||
| 70 : | %else | ||
| 71 : | global %1 | ||
| 72 : | %endif | ||
| 73 : | %endmacro | ||
| 74 : | |||
| 75 : | %define INP eax | ||
| 76 : | %define TABLE ebx | ||
| 77 : | %define TABLEF ebx | ||
| 78 : | %define OUT ecx | ||
| 79 : | %define round_frw_row edx | ||
| 80 : | |||
| 81 : | %define INP_1 eax + 16 | ||
| 82 : | %define INP_2 eax + 32 | ||
| 83 : | %define INP_3 eax + 48 | ||
| 84 : | %define INP_4 eax + 64 | ||
| 85 : | %define INP_5 eax + 80 | ||
| 86 : | %define INP_6 eax + 96 | ||
| 87 : | %define INP_7 eax + 112 | ||
| 88 : | |||
| 89 : | %define OUT_1 ecx + 16 | ||
| 90 : | %define OUT_2 ecx + 32 | ||
| 91 : | %define OUT_3 ecx + 48 | ||
| 92 : | %define OUT_4 ecx + 64 | ||
| 93 : | %define OUT_5 ecx + 80 | ||
| 94 : | %define OUT_6 ecx + 96 | ||
| 95 : | %define OUT_7 ecx + 112 | ||
| 96 : | %define OUT_8 ecx + 128 | ||
| 97 : | |||
| 98 : | %define TABLE_1 ebx + 64 | ||
| 99 : | %define TABLE_2 ebx + 128 | ||
| 100 : | %define TABLE_3 ebx + 192 | ||
| 101 : | %define TABLE_4 ebx + 256 | ||
| 102 : | %define TABLE_5 ebx + 320 | ||
| 103 : | %define TABLE_6 ebx + 384 | ||
| 104 : | %define TABLE_7 ebx + 448 | ||
| 105 : | |||
| 106 : | %define x0 INP + 0*16 | ||
| 107 : | %define x1 INP + 1*16 | ||
| 108 : | %define x2 INP + 2*16 | ||
| 109 : | %define x3 INP + 3*16 | ||
| 110 : | %define x4 INP + 4*16 | ||
| 111 : | %define x5 INP + 5*16 | ||
| 112 : | %define x6 INP + 6*16 | ||
| 113 : | %define x7 INP + 7*16 | ||
| 114 : | %define y0 OUT + 0*16 | ||
| 115 : | %define y1 OUT + 1*16 | ||
| 116 : | %define y2 OUT + 2*16 | ||
| 117 : | %define y3 OUT + 3*16 | ||
| 118 : | %define y4 OUT + 4*16 | ||
| 119 : | %define y5 OUT + 5*16 | ||
| 120 : | %define y6 OUT + 6*16 | ||
| 121 : | %define y7 OUT + 7*16 | ||
| 122 : | |||
| 123 : | %define tg_1_16 (TABLEF + 0) | ||
| 124 : | %define tg_2_16 (TABLEF + 8) | ||
| 125 : | %define tg_3_16 (TABLEF + 16) | ||
| 126 : | %define cos_4_16 (TABLEF + 24) | ||
| 127 : | %define ocos_4_16 (TABLEF + 32) | ||
| 128 : | |||
| 129 : | ;---added for sse2 | ||
| 130 : | %define tg_1_16_2 (TABLEF + 0) | ||
| 131 : | %define tg_2_16_2 (TABLEF + 16) | ||
| 132 : | %define tg_3_16_2 (TABLEF + 32) | ||
| 133 : | %define cos_4_16_2 (TABLEF + 48) | ||
| 134 : | %define ocos_4_16_2 (TABLEF + 64) | ||
| 135 : | ;--- | ||
| 136 : | |||
| 137 : | %ifdef FORMAT_COFF | ||
| 138 : | SECTION .data data | ||
| 139 : | %else | ||
| 140 : | SECTION .data data align=64 | ||
| 141 : | %endif | ||
| 142 : | |||
| 143 : | |||
| 144 : | ALIGN 16 | ||
| 145 : | |||
| 146 : | suxen_drol | 207 | BITS_FRW_ACC equ 3 ; 2 or 3 for accuracy |
| 147 : | SHIFT_FRW_COL equ BITS_FRW_ACC | ||
| 148 : | SHIFT_FRW_ROW equ (BITS_FRW_ACC + 17) | ||
| 149 : | Isibaar | 154 | RND_FRW_ROW equ (1 << (SHIFT_FRW_ROW-1)) |
| 150 : | suxen_drol | 207 | |
| 151 : | SHIFT_FRW_ROW_CLIP2 equ (4) | ||
| 152 : | Isibaar | 154 | SHIFT_FRW_ROW_CLIP1 equ (SHIFT_FRW_ROW - SHIFT_FRW_ROW_CLIP2) |
| 153 : | |||
| 154 : | PASS1_BITS equ (2) | ||
| 155 : | CONST_BITS equ (13) | ||
| 156 : | |||
| 157 : | one_corr dw 1, 1, 1, 1 | ||
| 158 : | |||
| 159 : | ;---added for sse2 | ||
| 160 : | align 16 | ||
| 161 : | one_corr_2 dw 1, 1, 1, 1, | ||
| 162 : | dw 1, 1, 1, 1 | ||
| 163 : | ;--- | ||
| 164 : | |||
| 165 : | r_frw_row dd RND_FRW_ROW, RND_FRW_ROW | ||
| 166 : | |||
| 167 : | |||
| 168 : | tg_all_16 dw 13036, 13036, 13036, 13036, ; tg * (2<<16) + 0.5 | ||
| 169 : | dw 27146, 27146, 27146, 27146, ; tg * (2<<16) + 0.5 | ||
| 170 : | dw -21746, -21746, -21746, -21746, ; tg * (2<<16) + 0.5 | ||
| 171 : | dw -19195, -19195, -19195, -19195, ; cos * (2<<16) + 0.5 | ||
| 172 : | dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 | ||
| 173 : | |||
| 174 : | ;---added for sse2 | ||
| 175 : | align 16 | ||
| 176 : | tg_all_16_2 dw 13036, 13036, 13036, 13036, ; tg * (2<<16) + 0.5 | ||
| 177 : | dw 13036, 13036, 13036, 13036, | ||
| 178 : | dw 27146, 27146, 27146, 27146, ; tg * (2<<16) + 0.5 | ||
| 179 : | dw 27146, 27146, 27146, 27146, | ||
| 180 : | dw -21746, -21746, -21746, -21746, ; tg * (2<<16) + 0.5 | ||
| 181 : | dw -21746, -21746, -21746, -21746, | ||
| 182 : | dw -19195, -19195, -19195, -19195, ; cos * (2<<16) + 0.5 | ||
| 183 : | dw -19195, -19195, -19195, -19195, | ||
| 184 : | dw 23170, 23170, 23170, 23170 ; cos * (2<<15) + 0.5 | ||
| 185 : | dw 23170, 23170, 23170, 23170 | ||
| 186 : | ;--- | ||
| 187 : | |||
| 188 : | tab_frw_01234567 | ||
| 189 : | ; row0 | ||
| 190 : | dw 16384, 16384, 21407, -8867, ; w09 w01 w08 w00 | ||
| 191 : | dw 16384, 16384, 8867, -21407, ; w13 w05 w12 w04 | ||
| 192 : | dw 16384, -16384, 8867, 21407, ; w11 w03 w10 w02 | ||
| 193 : | dw -16384, 16384, -21407, -8867, ; w15 w07 w14 w06 | ||
| 194 : | dw 22725, 12873, 19266, -22725, ; w22 w20 w18 w16 | ||
| 195 : | dw 19266, 4520, -4520, -12873, ; w23 w21 w19 w17 | ||
| 196 : | dw 12873, 4520, 4520, 19266, ; w30 w28 w26 w24 | ||
| 197 : | dw -22725, 19266, -12873, -22725, ; w31 w29 w27 w25 | ||
| 198 : | |||
| 199 : | ; row1 | ||
| 200 : | dw 22725, 22725, 29692, -12299, ; w09 w01 w08 w00 | ||
| 201 : | dw 22725, 22725, 12299, -29692, ; w13 w05 w12 w04 | ||
| 202 : | dw 22725, -22725, 12299, 29692, ; w11 w03 w10 w02 | ||
| 203 : | dw -22725, 22725, -29692, -12299, ; w15 w07 w14 w06 | ||
| 204 : | dw 31521, 17855, 26722, -31521, ; w22 w20 w18 w16 | ||
| 205 : | dw 26722, 6270, -6270, -17855, ; w23 w21 w19 w17 | ||
| 206 : | dw 17855, 6270, 6270, 26722, ; w30 w28 w26 w24 | ||
| 207 : | dw -31521, 26722, -17855, -31521, ; w31 w29 w27 w25 | ||
| 208 : | |||
| 209 : | ; row2 | ||
| 210 : | dw 21407, 21407, 27969, -11585, ; w09 w01 w08 w00 | ||
| 211 : | dw 21407, 21407, 11585, -27969, ; w13 w05 w12 w04 | ||
| 212 : | dw 21407, -21407, 11585, 27969, ; w11 w03 w10 w02 | ||
| 213 : | dw -21407, 21407, -27969, -11585, ; w15 w07 w14 w06 | ||
| 214 : | dw 29692, 16819, 25172, -29692, ; w22 w20 w18 w16 | ||
| 215 : | dw 25172, 5906, -5906, -16819, ; w23 w21 w19 w17 | ||
| 216 : | dw 16819, 5906, 5906, 25172, ; w30 w28 w26 w24 | ||
| 217 : | dw -29692, 25172, -16819, -29692, ; w31 w29 w27 w25 | ||
| 218 : | |||
| 219 : | ; row3 | ||
| 220 : | dw 19266, 19266, 25172, -10426, ; w09 w01 w08 w00 | ||
| 221 : | dw 19266, 19266, 10426, -25172, ; w13 w05 w12 w04 | ||
| 222 : | dw 19266, -19266, 10426, 25172, ; w11 w03 w10 w02 | ||
| 223 : | dw -19266, 19266, -25172, -10426, ; w15 w07 w14 w06 | ||
| 224 : | dw 26722, 15137, 22654, -26722, ; w22 w20 w18 w16 | ||
| 225 : | dw 22654, 5315, -5315, -15137, ; w23 w21 w19 w17 | ||
| 226 : | dw 15137, 5315, 5315, 22654, ; w30 w28 w26 w24 | ||
| 227 : | dw -26722, 22654, -15137, -26722, ; w31 w29 w27 w25 | ||
| 228 : | |||
| 229 : | ; row4 | ||
| 230 : | dw 16384, 16384, 21407, -8867, ; w09 w01 w08 w00 | ||
| 231 : | dw 16384, 16384, 8867, -21407, ; w13 w05 w12 w04 | ||
| 232 : | dw 16384, -16384, 8867, 21407, ; w11 w03 w10 w02 | ||
| 233 : | dw -16384, 16384, -21407, -8867, ; w15 w07 w14 w06 | ||
| 234 : | dw 22725, 12873, 19266, -22725, ; w22 w20 w18 w16 | ||
| 235 : | dw 19266, 4520, -4520, -12873, ; w23 w21 w19 w17 | ||
| 236 : | dw 12873, 4520, 4520, 19266, ; w30 w28 w26 w24 | ||
| 237 : | dw -22725, 19266, -12873, -22725, ; w31 w29 w27 w25 | ||
| 238 : | |||
| 239 : | ; row5 | ||
| 240 : | dw 19266, 19266, 25172, -10426, ; w09 w01 w08 w00 | ||
| 241 : | dw 19266, 19266, 10426, -25172, ; w13 w05 w12 w04 | ||
| 242 : | dw 19266, -19266, 10426, 25172, ; w11 w03 w10 w02 | ||
| 243 : | dw -19266, 19266, -25172, -10426, ; w15 w07 w14 w06 | ||
| 244 : | dw 26722, 15137, 22654, -26722, ; w22 w20 w18 w16 | ||
| 245 : | dw 22654, 5315, -5315, -15137, ; w23 w21 w19 w17 | ||
| 246 : | dw 15137, 5315, 5315, 22654, ; w30 w28 w26 w24 | ||
| 247 : | dw -26722, 22654, -15137, -26722, ; w31 w29 w27 w25 | ||
| 248 : | |||
| 249 : | ; row6 | ||
| 250 : | dw 21407, 21407, 27969, -11585, ; w09 w01 w08 w00 | ||
| 251 : | dw 21407, 21407, 11585, -27969, ; w13 w05 w12 w04 | ||
| 252 : | dw 21407, -21407, 11585, 27969, ; w11 w03 w10 w02 | ||
| 253 : | dw -21407, 21407, -27969, -11585, ; w15 w07 w14 w06 | ||
| 254 : | dw 29692, 16819, 25172, -29692, ; w22 w20 w18 w16 | ||
| 255 : | dw 25172, 5906, -5906, -16819, ; w23 w21 w19 w17 | ||
| 256 : | dw 16819, 5906, 5906, 25172, ; w30 w28 w26 w24 | ||
| 257 : | dw -29692, 25172, -16819, -29692, ; w31 w29 w27 w25 | ||
| 258 : | |||
| 259 : | ; row7 | ||
| 260 : | dw 22725, 22725, 29692, -12299, ; w09 w01 w08 w00 | ||
| 261 : | dw 22725, 22725, 12299, -29692, ; w13 w05 w12 w04 | ||
| 262 : | dw 22725, -22725, 12299, 29692, ; w11 w03 w10 w02 | ||
| 263 : | dw -22725, 22725, -29692, -12299, ; w15 w07 w14 w06 | ||
| 264 : | dw 31521, 17855, 26722, -31521, ; w22 w20 w18 w16 | ||
| 265 : | dw 26722, 6270, -6270, -17855, ; w23 w21 w19 w17 | ||
| 266 : | dw 17855, 6270, 6270, 26722, ; w30 w28 w26 w24 | ||
| 267 : | dw -31521, 26722, -17855, -31521 ; w31 w29 w27 w25 | ||
| 268 : | |||
| 269 : | align 128 | ||
| 270 : | |||
| 271 : | FIX_1 dw 10703, 4433, 10703, 4433 | ||
| 272 : | dw 10703, 4433, 10703, 4433 | ||
| 273 : | |||
| 274 : | FIX_2 dw 4433, -10704, 4433, -10704 | ||
| 275 : | dw 4433, -10704, 4433, -10704 | ||
| 276 : | |||
| 277 : | rounder_5 dw 2+16,2+16,2+16,2+16,2+16,2+16,2+16,2+16 | ||
| 278 : | rounder_11 dd 1024,1024,1024,1024 | ||
| 279 : | rounder_18 dd 16384+131072,16384+131072,16384+131072,16384+131072 | ||
| 280 : | |||
| 281 : | align 128 | ||
| 282 : | |||
| 283 : | FIX_3 dw 6437, 2260, 6437, 2260 | ||
| 284 : | dw 6437, 2260, 6437, 2260 | ||
| 285 : | dw 11363, 9633, 11363, 9633 | ||
| 286 : | dw 11363, 9633, 11363, 9633 | ||
| 287 : | |||
| 288 : | FIX_4 dw -11362, -6436, -11362, -6436 | ||
| 289 : | dw -11362, -6436, -11362, -6436 | ||
| 290 : | dw 9633, -2259, 9633, -2259 | ||
| 291 : | dw 9633, -2259, 9633, -2259 | ||
| 292 : | |||
| 293 : | FIX_5 dw 2261, 9633, 2261, 9633 | ||
| 294 : | dw 2261, 9633, 2261, 9633 | ||
| 295 : | dw 6437, -11362, 6437, -11362 | ||
| 296 : | dw 6437, -11362, 6437, -11362 | ||
| 297 : | |||
| 298 : | FIX_6 dw 9633, -11363, 9633, -11363 | ||
| 299 : | dw 9633, -11363, 9633, -11363 | ||
| 300 : | dw 2260, -6436, 2260, -6436 | ||
| 301 : | dw 2260, -6436, 2260, -6436 | ||
| 302 : | |||
| 303 : | align 128 | ||
| 304 : | buffer dw 0,0,0,0,0,0,0,0 | ||
| 305 : | dw 0,0,0,0,0,0,0,0 | ||
| 306 : | dw 0,0,0,0,0,0,0,0 | ||
| 307 : | dw 0,0,0,0,0,0,0,0 | ||
| 308 : | dw 0,0,0,0,0,0,0,0 | ||
| 309 : | dw 0,0,0,0,0,0,0,0 | ||
| 310 : | dw 0,0,0,0,0,0,0,0 | ||
| 311 : | dw 0,0,0,0,0,0,0,0 | ||
| 312 : | |||
| 313 : | SECTION .text | ||
| 314 : | |||
| 315 : | ALIGN 16 | ||
| 316 : | |||
| 317 : | |||
| 318 : | cglobal fdct_mmx | ||
| 319 : | ;;void fdct_mmx(short *block); | ||
| 320 : | fdct_mmx: | ||
| 321 : | |||
| 322 : | push ebx | ||
| 323 : | |||
| 324 : | mov INP, dword [esp + 8] ; block | ||
| 325 : | |||
| 326 : | mov TABLEF, tg_all_16 | ||
| 327 : | mov OUT, INP | ||
| 328 : | |||
| 329 : | movq mm0, [x1] ; 0 ; x1 | ||
| 330 : | |||
| 331 : | movq mm1, [x6] ; 1 ; x6 | ||
| 332 : | movq mm2, mm0 ; 2 ; x1 | ||
| 333 : | |||
| 334 : | movq mm3, [x2] ; 3 ; x2 | ||
| 335 : | paddsw mm0, mm1 ; t1 = x[1] + x[6] | ||
| 336 : | |||
| 337 : | movq mm4, [x5] ; 4 ; x5 | ||
| 338 : | psllw mm0, SHIFT_FRW_COL ; t1 | ||
| 339 : | |||
| 340 : | movq mm5, [x0] ; 5 ; x0 | ||
| 341 : | paddsw mm4, mm3 ; t2 = x[2] + x[5] | ||
| 342 : | |||
| 343 : | paddsw mm5, [x7] ; t0 = x[0] + x[7] | ||
| 344 : | psllw mm4, SHIFT_FRW_COL ; t2 | ||
| 345 : | |||
| 346 : | movq mm6, mm0 ; 6 ; t1 | ||
| 347 : | psubsw mm2, mm1 ; 1 ; t6 = x[1] - x[6] | ||
| 348 : | |||
| 349 : | movq mm1, [tg_2_16] ; 1 ; tg_2_16 | ||
| 350 : | psubsw mm0, mm4 ; tm12 = t1 - t2 | ||
| 351 : | |||
| 352 : | movq mm7, [x3] ; 7 ; x3 | ||
| 353 : | pmulhw mm1, mm0 ; tm12*tg_2_16 | ||
| 354 : | |||
| 355 : | paddsw mm7, [x4] ; t3 = x[3] + x[4] | ||
| 356 : | psllw mm5, SHIFT_FRW_COL ; t0 | ||
| 357 : | |||
| 358 : | paddsw mm6, mm4 ; 4 ; tp12 = t1 + t2 | ||
| 359 : | psllw mm7, SHIFT_FRW_COL ; t3 | ||
| 360 : | |||
| 361 : | movq mm4, mm5 ; 4 ; t0 | ||
| 362 : | psubsw mm5, mm7 ; tm03 = t0 - t3 | ||
| 363 : | |||
| 364 : | paddsw mm1, mm5 ; y2 = tm03 + tm12*tg_2_16 | ||
| 365 : | paddsw mm4, mm7 ; 7 ; tp03 = t0 + t3 | ||
| 366 : | |||
| 367 : | por mm1, qword [one_corr] ; correction y2 +0.5 | ||
| 368 : | psllw mm2, SHIFT_FRW_COL+1 ; t6 | ||
| 369 : | |||
| 370 : | pmulhw mm5, [tg_2_16] ; tm03*tg_2_16 | ||
| 371 : | movq mm7, mm4 ; 7 ; tp03 | ||
| 372 : | |||
| 373 : | psubsw mm3, [x5] ; t5 = x[2] - x[5] | ||
| 374 : | psubsw mm4, mm6 ; y4 = tp03 - tp12 | ||
| 375 : | |||
| 376 : | movq [y2], mm1 ; 1 ; save y2 | ||
| 377 : | paddsw mm7, mm6 ; 6 ; y0 = tp03 + tp12 | ||
| 378 : | |||
| 379 : | movq mm1, [x3] ; 1 ; x3 | ||
| 380 : | psllw mm3, SHIFT_FRW_COL+1 ; t5 | ||
| 381 : | |||
| 382 : | psubsw mm1, [x4] ; t4 = x[3] - x[4] | ||
| 383 : | movq mm6, mm2 ; 6 ; t6 | ||
| 384 : | |||
| 385 : | movq [y4], mm4 ; 4 ; save y4 | ||
| 386 : | paddsw mm2, mm3 ; t6 + t5 | ||
| 387 : | |||
| 388 : | pmulhw mm2, [ocos_4_16] ; tp65 = (t6 + t5)*cos_4_16 | ||
| 389 : | psubsw mm6, mm3 ; 3 ; t6 - t5 | ||
| 390 : | |||
| 391 : | pmulhw mm6, [ocos_4_16] ; tm65 = (t6 - t5)*cos_4_16 | ||
| 392 : | psubsw mm5, mm0 ; 0 ; y6 = tm03*tg_2_16 - tm12 | ||
| 393 : | |||
| 394 : | por mm5, qword [one_corr] ; correction y6 +0.5 | ||
| 395 : | psllw mm1, SHIFT_FRW_COL ; t4 | ||
| 396 : | |||
| 397 : | por mm2, qword [one_corr] ; correction tp65 +0.5 | ||
| 398 : | movq mm4, mm1 ; 4 ; t4 | ||
| 399 : | |||
| 400 : | movq mm3, [x0] ; 3 ; x0 | ||
| 401 : | paddsw mm1, mm6 ; tp465 = t4 + tm65 | ||
| 402 : | |||
| 403 : | psubsw mm3, [x7] ; t7 = x[0] - x[7] | ||
| 404 : | psubsw mm4, mm6 ; 6 ; tm465 = t4 - tm65 | ||
| 405 : | |||
| 406 : | movq mm0, [tg_1_16] ; 0 ; tg_1_16 | ||
| 407 : | psllw mm3, SHIFT_FRW_COL ; t7 | ||
| 408 : | |||
| 409 : | movq mm6, [tg_3_16] ; 6 ; tg_3_16 | ||
| 410 : | pmulhw mm0, mm1 ; tp465*tg_1_16 | ||
| 411 : | |||
| 412 : | movq [y0], mm7 ; 7 ; save y0 | ||
| 413 : | pmulhw mm6, mm4 ; tm465*tg_3_16 | ||
| 414 : | |||
| 415 : | movq [y6], mm5 ; 5 ; save y6 | ||
| 416 : | movq mm7, mm3 ; 7 ; t7 | ||
| 417 : | |||
| 418 : | movq mm5, [tg_3_16] ; 5 ; tg_3_16 | ||
| 419 : | psubsw mm7, mm2 ; tm765 = t7 - tp65 | ||
| 420 : | |||
| 421 : | paddsw mm3, mm2 ; 2 ; tp765 = t7 + tp65 | ||
| 422 : | pmulhw mm5, mm7 ; tm765*tg_3_16 | ||
| 423 : | |||
| 424 : | paddsw mm0, mm3 ; y1 = tp765 + tp465*tg_1_16 | ||
| 425 : | paddsw mm6, mm4 ; tm465*tg_3_16 | ||
| 426 : | |||
| 427 : | pmulhw mm3, [tg_1_16] ; tp765*tg_1_16 | ||
| 428 : | |||
| 429 : | por mm0, qword [one_corr] ; correction y1 +0.5 | ||
| 430 : | paddsw mm5, mm7 ; tm765*tg_3_16 | ||
| 431 : | |||
| 432 : | psubsw mm7, mm6 ; 6 ; y3 = tm765 - tm465*tg_3_16 | ||
| 433 : | add INP, 0x08 | ||
| 434 : | |||
| 435 : | movq [y1], mm0 ; 0 ; save y1 | ||
| 436 : | paddsw mm5, mm4 ; 4 ; y5 = tm765*tg_3_16 + tm465 | ||
| 437 : | |||
| 438 : | movq [y3], mm7 ; 7 ; save y3 | ||
| 439 : | psubsw mm3, mm1 ; 1 ; y7 = tp765*tg_1_16 - tp465 | ||
| 440 : | |||
| 441 : | movq [y5], mm5 ; 5 ; save y5 | ||
| 442 : | |||
| 443 : | movq mm0, [x1] ; 0 ; x1 | ||
| 444 : | |||
| 445 : | movq [y7], mm3 ; 3 ; save y7 (columns 0-4) | ||
| 446 : | |||
| 447 : | movq mm1, [x6] ; 1 ; x6 | ||
| 448 : | movq mm2, mm0 ; 2 ; x1 | ||
| 449 : | |||
| 450 : | movq mm3, [x2] ; 3 ; x2 | ||
| 451 : | paddsw mm0, mm1 ; t1 = x[1] + x[6] | ||
| 452 : | |||
| 453 : | movq mm4, [x5] ; 4 ; x5 | ||
| 454 : | psllw mm0, SHIFT_FRW_COL ; t1 | ||
| 455 : | |||
| 456 : | movq mm5, [x0] ; 5 ; x0 | ||
| 457 : | paddsw mm4, mm3 ; t2 = x[2] + x[5] | ||
| 458 : | |||
| 459 : | paddsw mm5, [x7] ; t0 = x[0] + x[7] | ||
| 460 : | psllw mm4, SHIFT_FRW_COL ; t2 | ||
| 461 : | |||
| 462 : | movq mm6, mm0 ; 6 ; t1 | ||
| 463 : | psubsw mm2, mm1 ; 1 ; t6 = x[1] - x[6] | ||
| 464 : | |||
| 465 : | movq mm1, [tg_2_16] ; 1 ; tg_2_16 | ||
| 466 : | psubsw mm0, mm4 ; tm12 = t1 - t2 | ||
| 467 : | |||
| 468 : | movq mm7, [x3] ; 7 ; x3 | ||
| 469 : | pmulhw mm1, mm0 ; tm12*tg_2_16 | ||
| 470 : | |||
| 471 : | paddsw mm7, [x4] ; t3 = x[3] + x[4] | ||
| 472 : | psllw mm5, SHIFT_FRW_COL ; t0 | ||
| 473 : | |||
| 474 : | paddsw mm6, mm4 ; 4 ; tp12 = t1 + t2 | ||
| 475 : | psllw mm7, SHIFT_FRW_COL ; t3 | ||
| 476 : | |||
| 477 : | movq mm4, mm5 ; 4 ; t0 | ||
| 478 : | psubsw mm5, mm7 ; tm03 = t0 - t3 | ||
| 479 : | |||
| 480 : | paddsw mm1, mm5 ; y2 = tm03 + tm12*tg_2_16 | ||
| 481 : | paddsw mm4, mm7 ; 7 ; tp03 = t0 + t3 | ||
| 482 : | |||
| 483 : | por mm1, qword [one_corr] ; correction y2 +0.5 | ||
| 484 : | psllw mm2, SHIFT_FRW_COL+1 ; t6 | ||
| 485 : | |||
| 486 : | pmulhw mm5, [tg_2_16] ; tm03*tg_2_16 | ||
| 487 : | movq mm7, mm4 ; 7 ; tp03 | ||
| 488 : | |||
| 489 : | psubsw mm3, [x5] ; t5 = x[2] - x[5] | ||
| 490 : | psubsw mm4, mm6 ; y4 = tp03 - tp12 | ||
| 491 : | |||
| 492 : | movq [y2+8], mm1 ; 1 ; save y2 | ||
| 493 : | paddsw mm7, mm6 ; 6 ; y0 = tp03 + tp12 | ||
| 494 : | |||
| 495 : | movq mm1, [x3] ; 1 ; x3 | ||
| 496 : | psllw mm3, SHIFT_FRW_COL+1 ; t5 | ||
| 497 : | |||
| 498 : | psubsw mm1, [x4] ; t4 = x[3] - x[4] | ||
| 499 : | movq mm6, mm2 ; 6 ; t6 | ||
| 500 : | |||
| 501 : | movq [y4+8], mm4 ; 4 ; save y4 | ||
| 502 : | paddsw mm2, mm3 ; t6 + t5 | ||
| 503 : | |||
| 504 : | pmulhw mm2, [ocos_4_16] ; tp65 = (t6 + t5)*cos_4_16 | ||
| 505 : | psubsw mm6, mm3 ; 3 ; t6 - t5 | ||
| 506 : | |||
| 507 : | pmulhw mm6, [ocos_4_16] ; tm65 = (t6 - t5)*cos_4_16 | ||
| 508 : | psubsw mm5, mm0 ; 0 ; y6 = tm03*tg_2_16 - tm12 | ||
| 509 : | |||
| 510 : | por mm5, qword [one_corr] ; correction y6 +0.5 | ||
| 511 : | psllw mm1, SHIFT_FRW_COL ; t4 | ||
| 512 : | |||
| 513 : | por mm2, qword [one_corr] ; correction tp65 +0.5 | ||
| 514 : | movq mm4, mm1 ; 4 ; t4 | ||
| 515 : | |||
| 516 : | movq mm3, [x0] ; 3 ; x0 | ||
| 517 : | paddsw mm1, mm6 ; tp465 = t4 + tm65 | ||
| 518 : | |||
| 519 : | psubsw mm3, [x7] ; t7 = x[0] - x[7] | ||
| 520 : | psubsw mm4, mm6 ; 6 ; tm465 = t4 - tm65 | ||
| 521 : | |||
| 522 : | movq mm0, [tg_1_16] ; 0 ; tg_1_16 | ||
| 523 : | psllw mm3, SHIFT_FRW_COL ; t7 | ||
| 524 : | |||
| 525 : | movq mm6, [tg_3_16] ; 6 ; tg_3_16 | ||
| 526 : | pmulhw mm0, mm1 ; tp465*tg_1_16 | ||
| 527 : | |||
| 528 : | movq [y0+8], mm7 ; 7 ; save y0 | ||
| 529 : | pmulhw mm6, mm4 ; tm465*tg_3_16 | ||
| 530 : | |||
| 531 : | movq [y6+8], mm5 ; 5 ; save y6 | ||
| 532 : | movq mm7, mm3 ; 7 ; t7 | ||
| 533 : | |||
| 534 : | movq mm5, [tg_3_16] ; 5 ; tg_3_16 | ||
| 535 : | psubsw mm7, mm2 ; tm765 = t7 - tp65 | ||
| 536 : | |||
| 537 : | paddsw mm3, mm2 ; 2 ; tp765 = t7 + tp65 | ||
| 538 : | pmulhw mm5, mm7 ; tm765*tg_3_16 | ||
| 539 : | |||
| 540 : | paddsw mm0, mm3 ; y1 = tp765 + tp465*tg_1_16 | ||
| 541 : | paddsw mm6, mm4 ; tm465*tg_3_16 | ||
| 542 : | |||
| 543 : | pmulhw mm3, [tg_1_16] ; tp765*tg_1_16 | ||
| 544 : | |||
| 545 : | por mm0, qword [one_corr] ; correction y1 +0.5 | ||
| 546 : | paddsw mm5, mm7 ; tm765*tg_3_16 | ||
| 547 : | |||
| 548 : | psubsw mm7, mm6 ; 6 ; y3 = tm765 - tm465*tg_3_16 | ||
| 549 : | |||
| 550 : | movq [y1+8], mm0 ; 0 ; save y1 | ||
| 551 : | paddsw mm5, mm4 ; 4 ; y5 = tm765*tg_3_16 + tm465 | ||
| 552 : | |||
| 553 : | movq [y3+8], mm7 ; 7 ; save y3 | ||
| 554 : | psubsw mm3, mm1 ; 1 ; y7 = tp765*tg_1_16 - tp465 | ||
| 555 : | |||
| 556 : | movq [y5+8], mm5 ; 5 ; save y5 | ||
| 557 : | |||
| 558 : | movq [y7+8], mm3 ; 3 ; save y7 | ||
| 559 : | |||
| 560 : | |||
| 561 : | ;--------------------- columns | ||
| 562 : | |||
| 563 : | mov INP, [esp + 8] ; row 0 | ||
| 564 : | |||
| 565 : | mov TABLEF, tab_frw_01234567 ; row 0 | ||
| 566 : | mov OUT, INP | ||
| 567 : | |||
| 568 : | mov round_frw_row, r_frw_row | ||
| 569 : | |||
| 570 : | movd mm5, [INP+12] ; mm5 = 7 6 | ||
| 571 : | |||
| 572 : | punpcklwd mm5, [INP+8] | ||
| 573 : | |||
| 574 : | movq mm2, mm5 ; mm2 = 5 7 4 6 | ||
| 575 : | psrlq mm5, 32 ; mm5 = _ _ 5 7 | ||
| 576 : | |||
| 577 : | movq mm0, [INP] ; mm0 = 3 2 1 0 | ||
| 578 : | punpcklwd mm5, mm2 ; mm5 = 4 5 6 7 | ||
| 579 : | |||
| 580 : | movq mm1, mm0 ; mm1 = 3 2 1 0 | ||
| 581 : | paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0) | ||
| 582 : | |||
| 583 : | psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4) | ||
| 584 : | movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ] | ||
| 585 : | |||
| 586 : | punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ] | ||
| 587 : | |||
| 588 : | punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ] | ||
| 589 : | movq mm1, mm2 ; mm1 | ||
| 590 : | |||
| 591 : | movq mm2, mm0 ; 2 ; x3 x2 x1 x0 | ||
| 592 : | |||
| 593 : | movq mm3, [TABLE] ; 3 ; w06 w04 w02 w00 | ||
| 594 : | punpcklwd mm0, mm1 ; x5 x1 x4 x0 | ||
| 595 : | |||
| 596 : | movq mm5, mm0 ; 5 ; x5 x1 x4 x0 | ||
| 597 : | punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ] | ||
| 598 : | |||
| 599 : | movq mm4, [TABLE+8] ; 4 ; w07 w05 w03 w01 | ||
| 600 : | punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 | ||
| 601 : | |||
| 602 : | pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 | ||
| 603 : | movq mm6, mm2 ; 6 ; x7 x3 x6 x2 | ||
| 604 : | |||
| 605 : | movq mm1, [TABLE+32] ; 1 ; w22 w20 w18 w16 | ||
| 606 : | punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ] | ||
| 607 : | |||
| 608 : | pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 | ||
| 609 : | punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ] | ||
| 610 : | |||
| 611 : | pmaddwd mm0, [TABLE+16] ; x4*w14+x0*w12 x4*w10+x0*w08 | ||
| 612 : | punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ] | ||
| 613 : | |||
| 614 : | movq mm7, [TABLE+40] ; 7 ; w23 w21 w19 w17 | ||
| 615 : | pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 | ||
| 616 : | |||
| 617 : | paddd mm3, [round_frw_row] ; +rounder (y2,y0) | ||
| 618 : | pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 | ||
| 619 : | |||
| 620 : | pmaddwd mm2, [TABLE+24] ; x6*w15+x2*w13 x6*w11+x2*w09 | ||
| 621 : | paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) | ||
| 622 : | |||
| 623 : | pmaddwd mm5, [TABLE+48] ; x5*w30+x1*w28 x5*w26+x1*w24 | ||
| 624 : | |||
| 625 : | pmaddwd mm6, [TABLE+56] ; x7*w31+x3*w29 x7*w27+x3*w25 | ||
| 626 : | paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) | ||
| 627 : | |||
| 628 : | paddd mm0, [round_frw_row] ; +rounder (y6,y4) | ||
| 629 : | psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0) | ||
| 630 : | |||
| 631 : | paddd mm1, [round_frw_row] ; +rounder (y3,y1) | ||
| 632 : | paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) | ||
| 633 : | |||
| 634 : | paddd mm5, [round_frw_row] ; +rounder (y7,y5) | ||
| 635 : | psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0 | ||
| 636 : | |||
| 637 : | paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) | ||
| 638 : | psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2 | ||
| 639 : | |||
| 640 : | psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2 | ||
| 641 : | |||
| 642 : | packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767} | ||
| 643 : | |||
| 644 : | packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767} | ||
| 645 : | movq mm6, mm3 ; mm0 = y6 y4 y2 y0 | ||
| 646 : | |||
| 647 : | punpcklwd mm3, mm1 ; y3 y2 y1 y0 | ||
| 648 : | |||
| 649 : | punpckhwd mm6, mm1 ; y7 y6 y5 y4 | ||
| 650 : | |||
| 651 : | psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047} | ||
| 652 : | |||
| 653 : | psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047} | ||
| 654 : | |||
| 655 : | movq [OUT_1-16], mm3 ; 1 ; save y3 y2 y1 y0 | ||
| 656 : | |||
| 657 : | movq [OUT_1-8], mm6 ; 7 ; save y7 y6 y5 y4 | ||
| 658 : | |||
| 659 : | movd mm5, [INP_1+12] ; mm5 = 7 6 | ||
| 660 : | |||
| 661 : | punpcklwd mm5, [INP_1+8] | ||
| 662 : | |||
| 663 : | movq mm2, mm5 ; mm2 = 5 7 4 6 | ||
| 664 : | psrlq mm5, 32 ; mm5 = _ _ 5 7 | ||
| 665 : | |||
| 666 : | movq mm0, [INP_1] ; mm0 = 3 2 1 0 | ||
| 667 : | punpcklwd mm5, mm2 ; mm5 = 4 5 6 7 | ||
| 668 : | |||
| 669 : | movq mm1, mm0 ; mm1 = 3 2 1 0 | ||
| 670 : | paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0) | ||
| 671 : | |||
| 672 : | psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4) | ||
| 673 : | movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ] | ||
| 674 : | |||
| 675 : | punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ] | ||
| 676 : | |||
| 677 : | punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ] | ||
| 678 : | movq mm1, mm2 ; mm1 | ||
| 679 : | |||
| 680 : | movq mm2, mm0 ; 2 ; x3 x2 x1 x0 | ||
| 681 : | |||
| 682 : | movq mm3, [TABLE_1] ; 3 ; w06 w04 w02 w00 | ||
| 683 : | punpcklwd mm0, mm1 ; x5 x1 x4 x0 | ||
| 684 : | |||
| 685 : | movq mm5, mm0 ; 5 ; x5 x1 x4 x0 | ||
| 686 : | punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ] | ||
| 687 : | |||
| 688 : | movq mm4, [TABLE_1+8] ; 4 ; w07 w05 w03 w01 | ||
| 689 : | punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 | ||
| 690 : | |||
| 691 : | pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 | ||
| 692 : | movq mm6, mm2 ; 6 ; x7 x3 x6 x2 | ||
| 693 : | |||
| 694 : | movq mm1, [TABLE_1+32] ; 1 ; w22 w20 w18 w16 | ||
| 695 : | punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ] | ||
| 696 : | |||
| 697 : | pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 | ||
| 698 : | punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ] | ||
| 699 : | |||
| 700 : | pmaddwd mm0, [TABLE_1+16] ; x4*w14+x0*w12 x4*w10+x0*w08 | ||
| 701 : | punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ] | ||
| 702 : | |||
| 703 : | movq mm7, [TABLE_1+40] ; 7 ; w23 w21 w19 w17 | ||
| 704 : | pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 | ||
| 705 : | |||
| 706 : | paddd mm3, [round_frw_row] ; +rounder (y2,y0) | ||
| 707 : | pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 | ||
| 708 : | |||
| 709 : | pmaddwd mm2, [TABLE_1+24] ; x6*w15+x2*w13 x6*w11+x2*w09 | ||
| 710 : | paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) | ||
| 711 : | |||
| 712 : | pmaddwd mm5, [TABLE_1+48] ; x5*w30+x1*w28 x5*w26+x1*w24 | ||
| 713 : | |||
| 714 : | pmaddwd mm6, [TABLE_1+56] ; x7*w31+x3*w29 x7*w27+x3*w25 | ||
| 715 : | paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) | ||
| 716 : | |||
| 717 : | paddd mm0, [round_frw_row] ; +rounder (y6,y4) | ||
| 718 : | psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0) | ||
| 719 : | |||
| 720 : | paddd mm1, [round_frw_row] ; +rounder (y3,y1) | ||
| 721 : | paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) | ||
| 722 : | |||
| 723 : | paddd mm5, [round_frw_row] ; +rounder (y7,y5) | ||
| 724 : | psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0 | ||
| 725 : | |||
| 726 : | paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) | ||
| 727 : | psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2 | ||
| 728 : | |||
| 729 : | psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2 | ||
| 730 : | |||
| 731 : | packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767} | ||
| 732 : | |||
| 733 : | packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767} | ||
| 734 : | movq mm6, mm3 ; mm0 = y6 y4 y2 y0 | ||
| 735 : | |||
| 736 : | punpcklwd mm3, mm1 ; y3 y2 y1 y0 | ||
| 737 : | |||
| 738 : | punpckhwd mm6, mm1 ; y7 y6 y5 y4 | ||
| 739 : | |||
| 740 : | psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047} | ||
| 741 : | |||
| 742 : | psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047} | ||
| 743 : | |||
| 744 : | movq [OUT_2-16], mm3 ; 1 ; save y3 y2 y1 y0 | ||
| 745 : | |||
| 746 : | movq [OUT_2-8], mm6 ; 7 ; save y7 y6 y5 y4 | ||
| 747 : | |||
| 748 : | movd mm5, [INP_2+12] ; mm5 = 7 6 | ||
| 749 : | |||
| 750 : | punpcklwd mm5, [INP_2+8] | ||
| 751 : | |||
| 752 : | movq mm2, mm5 ; mm2 = 5 7 4 6 | ||
| 753 : | psrlq mm5, 32 ; mm5 = _ _ 5 7 | ||
| 754 : | |||
| 755 : | movq mm0, [INP_2] ; mm0 = 3 2 1 0 | ||
| 756 : | punpcklwd mm5, mm2 ; mm5 = 4 5 6 7 | ||
| 757 : | |||
| 758 : | movq mm1, mm0 ; mm1 = 3 2 1 0 | ||
| 759 : | paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0) | ||
| 760 : | |||
| 761 : | psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4) | ||
| 762 : | movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ] | ||
| 763 : | |||
| 764 : | punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ] | ||
| 765 : | |||
| 766 : | punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ] | ||
| 767 : | movq mm1, mm2 ; mm1 | ||
| 768 : | |||
| 769 : | movq mm2, mm0 ; 2 ; x3 x2 x1 x0 | ||
| 770 : | |||
| 771 : | movq mm3, [TABLE_2] ; 3 ; w06 w04 w02 w00 | ||
| 772 : | punpcklwd mm0, mm1 ; x5 x1 x4 x0 | ||
| 773 : | |||
| 774 : | movq mm5, mm0 ; 5 ; x5 x1 x4 x0 | ||
| 775 : | punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ] | ||
| 776 : | |||
| 777 : | movq mm4, [TABLE_2+8] ; 4 ; w07 w05 w03 w01 | ||
| 778 : | punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 | ||
| 779 : | |||
| 780 : | pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 | ||
| 781 : | movq mm6, mm2 ; 6 ; x7 x3 x6 x2 | ||
| 782 : | |||
| 783 : | movq mm1, [TABLE_2+32] ; 1 ; w22 w20 w18 w16 | ||
| 784 : | punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ] | ||
| 785 : | |||
| 786 : | pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 | ||
| 787 : | punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ] | ||
| 788 : | |||
| 789 : | pmaddwd mm0, [TABLE_2+16] ; x4*w14+x0*w12 x4*w10+x0*w08 | ||
| 790 : | punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ] | ||
| 791 : | |||
| 792 : | movq mm7, [TABLE_2+40] ; 7 ; w23 w21 w19 w17 | ||
| 793 : | pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 | ||
| 794 : | |||
| 795 : | paddd mm3, [round_frw_row] ; +rounder (y2,y0) | ||
| 796 : | pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 | ||
| 797 : | |||
| 798 : | pmaddwd mm2, [TABLE_2+24] ; x6*w15+x2*w13 x6*w11+x2*w09 | ||
| 799 : | paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) | ||
| 800 : | |||
| 801 : | pmaddwd mm5, [TABLE_2+48] ; x5*w30+x1*w28 x5*w26+x1*w24 | ||
| 802 : | |||
| 803 : | pmaddwd mm6, [TABLE_2+56] ; x7*w31+x3*w29 x7*w27+x3*w25 | ||
| 804 : | paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) | ||
| 805 : | |||
| 806 : | paddd mm0, [round_frw_row] ; +rounder (y6,y4) | ||
| 807 : | psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0) | ||
| 808 : | |||
| 809 : | paddd mm1, [round_frw_row] ; +rounder (y3,y1) | ||
| 810 : | paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) | ||
| 811 : | |||
| 812 : | paddd mm5, [round_frw_row] ; +rounder (y7,y5) | ||
| 813 : | psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0 | ||
| 814 : | |||
| 815 : | paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) | ||
| 816 : | psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2 | ||
| 817 : | |||
| 818 : | psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2 | ||
| 819 : | |||
| 820 : | packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767} | ||
| 821 : | |||
| 822 : | packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767} | ||
| 823 : | movq mm6, mm3 ; mm0 = y6 y4 y2 y0 | ||
| 824 : | |||
| 825 : | punpcklwd mm3, mm1 ; y3 y2 y1 y0 | ||
| 826 : | |||
| 827 : | punpckhwd mm6, mm1 ; y7 y6 y5 y4 | ||
| 828 : | |||
| 829 : | psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047} | ||
| 830 : | |||
| 831 : | psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047} | ||
| 832 : | |||
| 833 : | movq [OUT_3-16], mm3 ; 1 ; save y3 y2 y1 y0 | ||
| 834 : | |||
| 835 : | movq [OUT_3-8], mm6 ; 7 ; save y7 y6 y5 y4 | ||
| 836 : | |||
| 837 : | movd mm5, [INP_3+12] ; mm5 = 7 6 | ||
| 838 : | |||
| 839 : | punpcklwd mm5, [INP_3+8] | ||
| 840 : | |||
| 841 : | movq mm2, mm5 ; mm2 = 5 7 4 6 | ||
| 842 : | psrlq mm5, 32 ; mm5 = _ _ 5 7 | ||
| 843 : | |||
| 844 : | movq mm0, [INP_3] ; mm0 = 3 2 1 0 | ||
| 845 : | punpcklwd mm5, mm2 ; mm5 = 4 5 6 7 | ||
| 846 : | |||
| 847 : | movq mm1, mm0 ; mm1 = 3 2 1 0 | ||
| 848 : | paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0) | ||
| 849 : | |||
| 850 : | psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4) | ||
| 851 : | movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ] | ||
| 852 : | |||
| 853 : | punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ] | ||
| 854 : | |||
| 855 : | punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ] | ||
| 856 : | movq mm1, mm2 ; mm1 | ||
| 857 : | |||
| 858 : | movq mm2, mm0 ; 2 ; x3 x2 x1 x0 | ||
| 859 : | |||
| 860 : | movq mm3, [TABLE_3] ; 3 ; w06 w04 w02 w00 | ||
| 861 : | punpcklwd mm0, mm1 ; x5 x1 x4 x0 | ||
| 862 : | |||
| 863 : | movq mm5, mm0 ; 5 ; x5 x1 x4 x0 | ||
| 864 : | punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ] | ||
| 865 : | |||
| 866 : | movq mm4, [TABLE_3+8] ; 4 ; w07 w05 w03 w01 | ||
| 867 : | punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 | ||
| 868 : | |||
| 869 : | pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 | ||
| 870 : | movq mm6, mm2 ; 6 ; x7 x3 x6 x2 | ||
| 871 : | |||
| 872 : | movq mm1, [TABLE_3+32] ; 1 ; w22 w20 w18 w16 | ||
| 873 : | punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ] | ||
| 874 : | |||
| 875 : | pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 | ||
| 876 : | punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ] | ||
| 877 : | |||
| 878 : | pmaddwd mm0, [TABLE_3+16] ; x4*w14+x0*w12 x4*w10+x0*w08 | ||
| 879 : | punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ] | ||
| 880 : | |||
| 881 : | movq mm7, [TABLE_3+40] ; 7 ; w23 w21 w19 w17 | ||
| 882 : | pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 | ||
| 883 : | |||
| 884 : | paddd mm3, [round_frw_row] ; +rounder (y2,y0) | ||
| 885 : | pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 | ||
| 886 : | |||
| 887 : | pmaddwd mm2, [TABLE_3+24] ; x6*w15+x2*w13 x6*w11+x2*w09 | ||
| 888 : | paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) | ||
| 889 : | |||
| 890 : | pmaddwd mm5, [TABLE_3+48] ; x5*w30+x1*w28 x5*w26+x1*w24 | ||
| 891 : | |||
| 892 : | pmaddwd mm6, [TABLE_3+56] ; x7*w31+x3*w29 x7*w27+x3*w25 | ||
| 893 : | paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) | ||
| 894 : | |||
| 895 : | paddd mm0, [round_frw_row] ; +rounder (y6,y4) | ||
| 896 : | psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0) | ||
| 897 : | |||
| 898 : | paddd mm1, [round_frw_row] ; +rounder (y3,y1) | ||
| 899 : | paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) | ||
| 900 : | |||
| 901 : | paddd mm5, [round_frw_row] ; +rounder (y7,y5) | ||
| 902 : | psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0 | ||
| 903 : | |||
| 904 : | paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) | ||
| 905 : | psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2 | ||
| 906 : | |||
| 907 : | psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2 | ||
| 908 : | |||
| 909 : | packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767} | ||
| 910 : | |||
| 911 : | packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767} | ||
| 912 : | movq mm6, mm3 ; mm0 = y6 y4 y2 y0 | ||
| 913 : | |||
| 914 : | punpcklwd mm3, mm1 ; y3 y2 y1 y0 | ||
| 915 : | |||
| 916 : | punpckhwd mm6, mm1 ; y7 y6 y5 y4 | ||
| 917 : | |||
| 918 : | psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047} | ||
| 919 : | |||
| 920 : | psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047} | ||
| 921 : | |||
| 922 : | movq [OUT_4-16], mm3 ; 1 ; save y3 y2 y1 y0 | ||
| 923 : | |||
| 924 : | movq [OUT_4-8], mm6 ; 7 ; save y7 y6 y5 y4 | ||
| 925 : | |||
| 926 : | movd mm5, [INP_4+12] ; mm5 = 7 6 | ||
| 927 : | |||
| 928 : | punpcklwd mm5, [INP_4+8] | ||
| 929 : | |||
| 930 : | movq mm2, mm5 ; mm2 = 5 7 4 6 | ||
| 931 : | psrlq mm5, 32 ; mm5 = _ _ 5 7 | ||
| 932 : | |||
| 933 : | movq mm0, [INP_4] ; mm0 = 3 2 1 0 | ||
| 934 : | punpcklwd mm5, mm2 ; mm5 = 4 5 6 7 | ||
| 935 : | |||
| 936 : | movq mm1, mm0 ; mm1 = 3 2 1 0 | ||
| 937 : | paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0) | ||
| 938 : | |||
| 939 : | psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4) | ||
| 940 : | movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ] | ||
| 941 : | |||
| 942 : | punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ] | ||
| 943 : | |||
| 944 : | punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ] | ||
| 945 : | movq mm1, mm2 ; mm1 | ||
| 946 : | |||
| 947 : | movq mm2, mm0 ; 2 ; x3 x2 x1 x0 | ||
| 948 : | |||
| 949 : | movq mm3, [TABLE_4] ; 3 ; w06 w04 w02 w00 | ||
| 950 : | punpcklwd mm0, mm1 ; x5 x1 x4 x0 | ||
| 951 : | |||
| 952 : | movq mm5, mm0 ; 5 ; x5 x1 x4 x0 | ||
| 953 : | punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ] | ||
| 954 : | |||
| 955 : | movq mm4, [TABLE_4+8] ; 4 ; w07 w05 w03 w01 | ||
| 956 : | punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 | ||
| 957 : | |||
| 958 : | pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 | ||
| 959 : | movq mm6, mm2 ; 6 ; x7 x3 x6 x2 | ||
| 960 : | |||
| 961 : | movq mm1, [TABLE_4+32] ; 1 ; w22 w20 w18 w16 | ||
| 962 : | punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ] | ||
| 963 : | |||
| 964 : | pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 | ||
| 965 : | punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ] | ||
| 966 : | |||
| 967 : | pmaddwd mm0, [TABLE_4+16] ; x4*w14+x0*w12 x4*w10+x0*w08 | ||
| 968 : | punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ] | ||
| 969 : | |||
| 970 : | movq mm7, [TABLE_4+40] ; 7 ; w23 w21 w19 w17 | ||
| 971 : | pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 | ||
| 972 : | |||
| 973 : | paddd mm3, [round_frw_row] ; +rounder (y2,y0) | ||
| 974 : | pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 | ||
| 975 : | |||
| 976 : | pmaddwd mm2, [TABLE_4+24] ; x6*w15+x2*w13 x6*w11+x2*w09 | ||
| 977 : | paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) | ||
| 978 : | |||
| 979 : | pmaddwd mm5, [TABLE_4+48] ; x5*w30+x1*w28 x5*w26+x1*w24 | ||
| 980 : | |||
| 981 : | pmaddwd mm6, [TABLE_4+56] ; x7*w31+x3*w29 x7*w27+x3*w25 | ||
| 982 : | paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) | ||
| 983 : | |||
| 984 : | paddd mm0, [round_frw_row] ; +rounder (y6,y4) | ||
| 985 : | psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0) | ||
| 986 : | |||
| 987 : | paddd mm1, [round_frw_row] ; +rounder (y3,y1) | ||
| 988 : | paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) | ||
| 989 : | |||
| 990 : | paddd mm5, [round_frw_row] ; +rounder (y7,y5) | ||
| 991 : | psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0 | ||
| 992 : | |||
| 993 : | paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) | ||
| 994 : | psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2 | ||
| 995 : | |||
| 996 : | psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2 | ||
| 997 : | |||
| 998 : | packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767} | ||
| 999 : | |||
| 1000 : | packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767} | ||
| 1001 : | movq mm6, mm3 ; mm0 = y6 y4 y2 y0 | ||
| 1002 : | |||
| 1003 : | punpcklwd mm3, mm1 ; y3 y2 y1 y0 | ||
| 1004 : | |||
| 1005 : | punpckhwd mm6, mm1 ; y7 y6 y5 y4 | ||
| 1006 : | |||
| 1007 : | psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047} | ||
| 1008 : | |||
| 1009 : | psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047} | ||
| 1010 : | |||
| 1011 : | movq [OUT_5-16], mm3 ; 1 ; save y3 y2 y1 y0 | ||
| 1012 : | |||
| 1013 : | movq [OUT_5-8], mm6 ; 7 ; save y7 y6 y5 y4 | ||
| 1014 : | |||
| 1015 : | movd mm5, [INP_5+12] ; mm5 = 7 6 | ||
| 1016 : | |||
| 1017 : | punpcklwd mm5, [INP_5+8] | ||
| 1018 : | |||
| 1019 : | movq mm2, mm5 ; mm2 = 5 7 4 6 | ||
| 1020 : | psrlq mm5, 32 ; mm5 = _ _ 5 7 | ||
| 1021 : | |||
| 1022 : | movq mm0, [INP_5] ; mm0 = 3 2 1 0 | ||
| 1023 : | punpcklwd mm5, mm2 ; mm5 = 4 5 6 7 | ||
| 1024 : | |||
| 1025 : | movq mm1, mm0 ; mm1 = 3 2 1 0 | ||
| 1026 : | paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0) | ||
| 1027 : | |||
| 1028 : | psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4) | ||
| 1029 : | movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ] | ||
| 1030 : | |||
| 1031 : | punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ] | ||
| 1032 : | |||
| 1033 : | punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ] | ||
| 1034 : | movq mm1, mm2 ; mm1 | ||
| 1035 : | |||
| 1036 : | movq mm2, mm0 ; 2 ; x3 x2 x1 x0 | ||
| 1037 : | |||
| 1038 : | movq mm3, [TABLE_5] ; 3 ; w06 w04 w02 w00 | ||
| 1039 : | punpcklwd mm0, mm1 ; x5 x1 x4 x0 | ||
| 1040 : | |||
| 1041 : | movq mm5, mm0 ; 5 ; x5 x1 x4 x0 | ||
| 1042 : | punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ] | ||
| 1043 : | |||
| 1044 : | movq mm4, [TABLE_5+8] ; 4 ; w07 w05 w03 w01 | ||
| 1045 : | punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 | ||
| 1046 : | |||
| 1047 : | pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 | ||
| 1048 : | movq mm6, mm2 ; 6 ; x7 x3 x6 x2 | ||
| 1049 : | |||
| 1050 : | movq mm1, [TABLE_5+32] ; 1 ; w22 w20 w18 w16 | ||
| 1051 : | punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ] | ||
| 1052 : | |||
| 1053 : | pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 | ||
| 1054 : | punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ] | ||
| 1055 : | |||
| 1056 : | pmaddwd mm0, [TABLE_5+16] ; x4*w14+x0*w12 x4*w10+x0*w08 | ||
| 1057 : | punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ] | ||
| 1058 : | |||
| 1059 : | movq mm7, [TABLE_5+40] ; 7 ; w23 w21 w19 w17 | ||
| 1060 : | pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 | ||
| 1061 : | |||
| 1062 : | paddd mm3, [round_frw_row] ; +rounder (y2,y0) | ||
| 1063 : | pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 | ||
| 1064 : | |||
| 1065 : | pmaddwd mm2, [TABLE_5+24] ; x6*w15+x2*w13 x6*w11+x2*w09 | ||
| 1066 : | paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) | ||
| 1067 : | |||
| 1068 : | pmaddwd mm5, [TABLE_5+48] ; x5*w30+x1*w28 x5*w26+x1*w24 | ||
| 1069 : | |||
| 1070 : | pmaddwd mm6, [TABLE_5+56] ; x7*w31+x3*w29 x7*w27+x3*w25 | ||
| 1071 : | paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) | ||
| 1072 : | |||
| 1073 : | paddd mm0, [round_frw_row] ; +rounder (y6,y4) | ||
| 1074 : | psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0) | ||
| 1075 : | |||
| 1076 : | paddd mm1, [round_frw_row] ; +rounder (y3,y1) | ||
| 1077 : | paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) | ||
| 1078 : | |||
| 1079 : | paddd mm5, [round_frw_row] ; +rounder (y7,y5) | ||
| 1080 : | psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0 | ||
| 1081 : | |||
| 1082 : | paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) | ||
| 1083 : | psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2 | ||
| 1084 : | |||
| 1085 : | psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2 | ||
| 1086 : | |||
| 1087 : | packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767} | ||
| 1088 : | |||
| 1089 : | packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767} | ||
| 1090 : | movq mm6, mm3 ; mm0 = y6 y4 y2 y0 | ||
| 1091 : | |||
| 1092 : | punpcklwd mm3, mm1 ; y3 y2 y1 y0 | ||
| 1093 : | |||
| 1094 : | punpckhwd mm6, mm1 ; y7 y6 y5 y4 | ||
| 1095 : | |||
| 1096 : | psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047} | ||
| 1097 : | |||
| 1098 : | psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047} | ||
| 1099 : | |||
| 1100 : | movq [OUT_6-16], mm3 ; 1 ; save y3 y2 y1 y0 | ||
| 1101 : | |||
| 1102 : | movq [OUT_6-8], mm6 ; 7 ; save y7 y6 y5 y4 | ||
| 1103 : | |||
| 1104 : | movd mm5, [INP_6+12] ; mm5 = 7 6 | ||
| 1105 : | |||
| 1106 : | punpcklwd mm5, [INP_6+8] | ||
| 1107 : | |||
| 1108 : | movq mm2, mm5 ; mm2 = 5 7 4 6 | ||
| 1109 : | psrlq mm5, 32 ; mm5 = _ _ 5 7 | ||
| 1110 : | |||
| 1111 : | movq mm0, [INP_6] ; mm0 = 3 2 1 0 | ||
| 1112 : | punpcklwd mm5, mm2 ; mm5 = 4 5 6 7 | ||
| 1113 : | |||
| 1114 : | movq mm1, mm0 ; mm1 = 3 2 1 0 | ||
| 1115 : | paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0) | ||
| 1116 : | |||
| 1117 : | psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4) | ||
| 1118 : | movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ] | ||
| 1119 : | |||
| 1120 : | punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ] | ||
| 1121 : | |||
| 1122 : | punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ] | ||
| 1123 : | movq mm1, mm2 ; mm1 | ||
| 1124 : | |||
| 1125 : | movq mm2, mm0 ; 2 ; x3 x2 x1 x0 | ||
| 1126 : | |||
| 1127 : | movq mm3, [TABLE_6] ; 3 ; w06 w04 w02 w00 | ||
| 1128 : | punpcklwd mm0, mm1 ; x5 x1 x4 x0 | ||
| 1129 : | |||
| 1130 : | movq mm5, mm0 ; 5 ; x5 x1 x4 x0 | ||
| 1131 : | punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ] | ||
| 1132 : | |||
| 1133 : | movq mm4, [TABLE_6+8] ; 4 ; w07 w05 w03 w01 | ||
| 1134 : | punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 | ||
| 1135 : | |||
| 1136 : | pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 | ||
| 1137 : | movq mm6, mm2 ; 6 ; x7 x3 x6 x2 | ||
| 1138 : | |||
| 1139 : | movq mm1, [TABLE_6+32] ; 1 ; w22 w20 w18 w16 | ||
| 1140 : | punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ] | ||
| 1141 : | |||
| 1142 : | pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 | ||
| 1143 : | punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ] | ||
| 1144 : | |||
| 1145 : | pmaddwd mm0, [TABLE_6+16] ; x4*w14+x0*w12 x4*w10+x0*w08 | ||
| 1146 : | punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ] | ||
| 1147 : | |||
| 1148 : | movq mm7, [TABLE_6+40] ; 7 ; w23 w21 w19 w17 | ||
| 1149 : | pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 | ||
| 1150 : | |||
| 1151 : | paddd mm3, [round_frw_row] ; +rounder (y2,y0) | ||
| 1152 : | pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 | ||
| 1153 : | |||
| 1154 : | pmaddwd mm2, [TABLE_6+24] ; x6*w15+x2*w13 x6*w11+x2*w09 | ||
| 1155 : | paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) | ||
| 1156 : | |||
| 1157 : | pmaddwd mm5, [TABLE_6+48] ; x5*w30+x1*w28 x5*w26+x1*w24 | ||
| 1158 : | |||
| 1159 : | pmaddwd mm6, [TABLE_6+56] ; x7*w31+x3*w29 x7*w27+x3*w25 | ||
| 1160 : | paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) | ||
| 1161 : | |||
| 1162 : | paddd mm0, [round_frw_row] ; +rounder (y6,y4) | ||
| 1163 : | psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0) | ||
| 1164 : | |||
| 1165 : | paddd mm1, [round_frw_row] ; +rounder (y3,y1) | ||
| 1166 : | paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) | ||
| 1167 : | |||
| 1168 : | paddd mm5, [round_frw_row] ; +rounder (y7,y5) | ||
| 1169 : | psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0 | ||
| 1170 : | |||
| 1171 : | paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) | ||
| 1172 : | psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2 | ||
| 1173 : | |||
| 1174 : | psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2 | ||
| 1175 : | |||
| 1176 : | packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767} | ||
| 1177 : | |||
| 1178 : | packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767} | ||
| 1179 : | movq mm6, mm3 ; mm0 = y6 y4 y2 y0 | ||
| 1180 : | |||
| 1181 : | punpcklwd mm3, mm1 ; y3 y2 y1 y0 | ||
| 1182 : | |||
| 1183 : | punpckhwd mm6, mm1 ; y7 y6 y5 y4 | ||
| 1184 : | |||
| 1185 : | psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047} | ||
| 1186 : | |||
| 1187 : | psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047} | ||
| 1188 : | |||
| 1189 : | movq [OUT_7-16], mm3 ; 1 ; save y3 y2 y1 y0 | ||
| 1190 : | |||
| 1191 : | movq [OUT_7-8], mm6 ; 7 ; save y7 y6 y5 y4 | ||
| 1192 : | |||
| 1193 : | movd mm5, [INP_7+12] ; mm5 = 7 6 | ||
| 1194 : | |||
| 1195 : | punpcklwd mm5, [INP_7+8] | ||
| 1196 : | |||
| 1197 : | movq mm2, mm5 ; mm2 = 5 7 4 6 | ||
| 1198 : | psrlq mm5, 32 ; mm5 = _ _ 5 7 | ||
| 1199 : | |||
| 1200 : | movq mm0, [INP_7] ; mm0 = 3 2 1 0 | ||
| 1201 : | punpcklwd mm5, mm2 ; mm5 = 4 5 6 7 | ||
| 1202 : | |||
| 1203 : | movq mm1, mm0 ; mm1 = 3 2 1 0 | ||
| 1204 : | paddsw mm0, mm5 ; mm0 = [3+4, 2+5, 1+6, 0+7] (xt3, xt2, xt1, xt0) | ||
| 1205 : | |||
| 1206 : | psubsw mm1, mm5 ; mm1 = [3-4, 2-5, 1-6, 0-7] (xt7, xt6, xt5, xt4) | ||
| 1207 : | movq mm2, mm0 ; mm2 = [ xt3 xt2 xt1 xt0 ] | ||
| 1208 : | |||
| 1209 : | punpcklwd mm0, mm1 ; mm0 = [ xt5 xt1 xt4 xt0 ] | ||
| 1210 : | |||
| 1211 : | punpckhwd mm2, mm1 ; mm2 = [ xt7 xt3 xt6 xt2 ] | ||
| 1212 : | movq mm1, mm2 ; mm1 | ||
| 1213 : | |||
| 1214 : | movq mm2, mm0 ; 2 ; x3 x2 x1 x0 | ||
| 1215 : | |||
| 1216 : | movq mm3, [TABLE_7] ; 3 ; w06 w04 w02 w00 | ||
| 1217 : | punpcklwd mm0, mm1 ; x5 x1 x4 x0 | ||
| 1218 : | |||
| 1219 : | movq mm5, mm0 ; 5 ; x5 x1 x4 x0 | ||
| 1220 : | punpckldq mm0, mm0 ; x4 x0 x4 x0 [ xt2 xt0 xt2 xt0 ] | ||
| 1221 : | |||
| 1222 : | movq mm4, [TABLE_7+8] ; 4 ; w07 w05 w03 w01 | ||
| 1223 : | punpckhwd mm2, mm1 ; 1 ; x7 x3 x6 x2 | ||
| 1224 : | |||
| 1225 : | pmaddwd mm3, mm0 ; x4*w06+x0*w04 x4*w02+x0*w00 | ||
| 1226 : | movq mm6, mm2 ; 6 ; x7 x3 x6 x2 | ||
| 1227 : | |||
| 1228 : | movq mm1, [TABLE_7+32] ; 1 ; w22 w20 w18 w16 | ||
| 1229 : | punpckldq mm2, mm2 ; x6 x2 x6 x2 [ xt3 xt1 xt3 xt1 ] | ||
| 1230 : | |||
| 1231 : | pmaddwd mm4, mm2 ; x6*w07+x2*w05 x6*w03+x2*w01 | ||
| 1232 : | punpckhdq mm5, mm5 ; x5 x1 x5 x1 [ xt6 xt4 xt6 xt4 ] | ||
| 1233 : | |||
| 1234 : | pmaddwd mm0, [TABLE_7+16] ; x4*w14+x0*w12 x4*w10+x0*w08 | ||
| 1235 : | punpckhdq mm6, mm6 ; x7 x3 x7 x3 [ xt7 xt5 xt7 xt5 ] | ||
| 1236 : | |||
| 1237 : | movq mm7, [TABLE_7+40] ; 7 ; w23 w21 w19 w17 | ||
| 1238 : | pmaddwd mm1, mm5 ; x5*w22+x1*w20 x5*w18+x1*w16 | ||
| 1239 : | |||
| 1240 : | paddd mm3, [round_frw_row] ; +rounder (y2,y0) | ||
| 1241 : | pmaddwd mm7, mm6 ; x7*w23+x3*w21 x7*w19+x3*w17 | ||
| 1242 : | |||
| 1243 : | pmaddwd mm2, [TABLE_7+24] ; x6*w15+x2*w13 x6*w11+x2*w09 | ||
| 1244 : | paddd mm3, mm4 ; 4 ; a1=sum(even1) a0=sum(even0) | ||
| 1245 : | |||
| 1246 : | pmaddwd mm5, [TABLE_7+48] ; x5*w30+x1*w28 x5*w26+x1*w24 | ||
| 1247 : | |||
| 1248 : | pmaddwd mm6, [TABLE_7+56] ; x7*w31+x3*w29 x7*w27+x3*w25 | ||
| 1249 : | paddd mm1, mm7 ; 7 ; b1=sum(odd1) b0=sum(odd0) | ||
| 1250 : | |||
| 1251 : | paddd mm0, [round_frw_row] ; +rounder (y6,y4) | ||
| 1252 : | psrad mm3, SHIFT_FRW_ROW_CLIP1 ; (y2, y0) | ||
| 1253 : | |||
| 1254 : | paddd mm1, [round_frw_row] ; +rounder (y3,y1) | ||
| 1255 : | paddd mm0, mm2 ; 2 ; a3=sum(even3) a2=sum(even2) | ||
| 1256 : | |||
| 1257 : | paddd mm5, [round_frw_row] ; +rounder (y7,y5) | ||
| 1258 : | psrad mm1, SHIFT_FRW_ROW_CLIP1 ; y1=a1+b1 y0=a0+b0 | ||
| 1259 : | |||
| 1260 : | paddd mm5, mm6 ; 6 ; b3=sum(odd3) b2=sum(odd2) | ||
| 1261 : | psrad mm0, SHIFT_FRW_ROW_CLIP1 ; y3=a3+b3 y2=a2+b2 | ||
| 1262 : | |||
| 1263 : | psrad mm5, SHIFT_FRW_ROW_CLIP1 ; y4=a3-b3 y5=a2-b2 | ||
| 1264 : | |||
| 1265 : | packssdw mm3, mm0 ; 0 ; y6 y4 y2 y0, saturate {-32768,+32767} | ||
| 1266 : | |||
| 1267 : | packssdw mm1, mm5 ; 3 ; y7 y5 y3 y1, saturate {-32768,+32767} | ||
| 1268 : | movq mm6, mm3 ; mm0 = y6 y4 y2 y0 | ||
| 1269 : | |||
| 1270 : | punpcklwd mm3, mm1 ; y3 y2 y1 y0 | ||
| 1271 : | |||
| 1272 : | punpckhwd mm6, mm1 ; y7 y6 y5 y4 | ||
| 1273 : | |||
| 1274 : | psraw mm3, SHIFT_FRW_ROW_CLIP2 ; descale [y3 y2 y1 y0] to {-2048,+2047} | ||
| 1275 : | |||
| 1276 : | psraw mm6, SHIFT_FRW_ROW_CLIP2 ; descale [y7 y6 y5 y4] to {-2048,+2047} | ||
| 1277 : | |||
| 1278 : | movq [OUT_8-16], mm3 ; 1 ; save y3 y2 y1 y0 | ||
| 1279 : | |||
| 1280 : | movq [OUT_8-8], mm6 ; 7 ; save y7 y6 y5 y4 | ||
| 1281 : | |||
| 1282 : | pop ebx | ||
| 1283 : | emms | ||
| 1284 : | |||
| 1285 : | ret | ||
| 1286 : | |||
| 1287 : | |||
| 1288 : | ;******************************************************************* | ||
| 1289 : | ; This SSE2 code of the FDCT algoruthm coded by | ||
| 1290 : | ; Dmitry Rozhdestvensky and checked by Vladimir G. Ivanov | ||
| 1291 : | ;******************************************************************* | ||
| 1292 : | |||
| 1293 : | %macro transpose8x8sse2 3 ; source,dest,{0=first part, 1=second part} | ||
| 1294 : | ;I found this smart transposing algo at www.x86.org | ||
| 1295 : | |||
| 1296 : | movdqa xmm0,[%1+0*16] ; 07 06 05 04 03 02 01 00 | ||
| 1297 : | movdqa xmm6,[%1+2*16] ; 27 26 25 24 23 22 21 20 | ||
| 1298 : | movdqa xmm4,[%1+4*16] ; 47 46 45 44 43 42 41 40 | ||
| 1299 : | movdqa xmm7,[%1+6*16] ; 67 66 65 64 63 62 61 60 | ||
| 1300 : | |||
| 1301 : | %if %3=0 | ||
| 1302 : | punpcklwd xmm0,[%1+1*16] ; 13 03 12 02 11 01 10 00 | ||
| 1303 : | movdqa xmm2,xmm0 | ||
| 1304 : | punpcklwd xmm6,[%1+3*16] ; 33 23 32 22 31 21 30 20 | ||
| 1305 : | punpcklwd xmm4,[%1+5*16] ; 53 43 52 42 51 41 50 40 | ||
| 1306 : | movdqa xmm5,xmm4 | ||
| 1307 : | punpcklwd xmm7,[%1+7*16] ; 73 63 72 62 71 61 70 60 | ||
| 1308 : | %else | ||
| 1309 : | punpckhwd xmm0,[%1+1*16] ; | ||
| 1310 : | movdqa xmm2,xmm0 | ||
| 1311 : | punpckhwd xmm6,[%1+3*16] ; | ||
| 1312 : | punpckhwd xmm4,[%1+5*16] ; | ||
| 1313 : | movdqa xmm5,xmm4 | ||
| 1314 : | punpckhwd xmm7,[%1+7*16] ; | ||
| 1315 : | %endif | ||
| 1316 : | |||
| 1317 : | punpckldq xmm0,xmm6 ; 31 21 11 01 30 20 10 00 | ||
| 1318 : | movdqa xmm1,xmm0 | ||
| 1319 : | punpckldq xmm4,xmm7 ; 71 61 51 41 70 60 50 40 | ||
| 1320 : | punpckhdq xmm2,xmm6 ; 33 23 13 03 32 22 12 02 | ||
| 1321 : | movdqa xmm3,xmm2 | ||
| 1322 : | punpckhdq xmm5,xmm7 ; 73 63 53 43 72 62 52 42 | ||
| 1323 : | |||
| 1324 : | punpcklqdq xmm0,xmm4 ; 70 60 50 40 30 20 10 00 | ||
| 1325 : | punpcklqdq xmm2,xmm5 ; 72 62 52 42 32 22 21 02 | ||
| 1326 : | punpckhqdq xmm1,xmm4 ; 71 61 51 41 31 21 11 01 | ||
| 1327 : | punpckhqdq xmm3,xmm5 ; 73 63 53 43 33 23 13 03 | ||
| 1328 : | |||
| 1329 : | movdqa [%2+(0+%3*4)*16],xmm0 | ||
| 1330 : | movdqa [%2+(1+%3*4)*16],xmm1 | ||
| 1331 : | movdqa [%2+(2+%3*4)*16],xmm2 | ||
| 1332 : | movdqa [%2+(3+%3*4)*16],xmm3 | ||
| 1333 : | |||
| 1334 : | %endmacro | ||
| 1335 : | |||
| 1336 : | |||
| 1337 : | %macro makeoddpart 3 ; output, table, shift | ||
| 1338 : | |||
| 1339 : | |||
| 1340 : | movdqa xmm1,[%2+0 ] ;45l | ||
| 1341 : | movdqa xmm5,[%2+16] ;67l | ||
| 1342 : | movdqa xmm3,xmm1 ;45h | ||
| 1343 : | movdqa xmm7,xmm5 ;67h | ||
| 1344 : | |||
| 1345 : | pmaddwd xmm1,xmm2 ;[%2+0 ] | ||
| 1346 : | pmaddwd xmm5,xmm0 ;[%2+16] | ||
| 1347 : | paddd xmm1,xmm5 | ||
| 1348 : | %if %3=11 | ||
| 1349 : | movdqa xmm5,[rounder_11] | ||
| 1350 : | %else | ||
| 1351 : | movdqa xmm5,[rounder_18] | ||
| 1352 : | %endif | ||
| 1353 : | |||
| 1354 : | pmaddwd xmm3,xmm6 ;[%2+0 ] | ||
| 1355 : | pmaddwd xmm7,xmm4 ;[%2+16] | ||
| 1356 : | paddd xmm3,xmm7 | ||
| 1357 : | |||
| 1358 : | paddd xmm1,xmm5 ;rounder | ||
| 1359 : | paddd xmm3,xmm5 ;rounder | ||
| 1360 : | psrad xmm1,%3 | ||
| 1361 : | psrad xmm3,%3 | ||
| 1362 : | |||
| 1363 : | packssdw xmm1,xmm3 | ||
| 1364 : | movdqa [%1],xmm1 | ||
| 1365 : | |||
| 1366 : | %endmacro | ||
| 1367 : | |||
| 1368 : | %macro makeoddpartlast 3 ; output, table, shift | ||
| 1369 : | |||
| 1370 : | pmaddwd xmm2,[%2+0 ] | ||
| 1371 : | pmaddwd xmm0,[%2+16] | ||
| 1372 : | paddd xmm2,xmm0 | ||
| 1373 : | |||
| 1374 : | pmaddwd xmm6,[%2+0 ] | ||
| 1375 : | pmaddwd xmm4,[%2+16] | ||
| 1376 : | paddd xmm6,xmm4 | ||
| 1377 : | |||
| 1378 : | paddd xmm2,xmm5 ;rounder | ||
| 1379 : | paddd xmm6,xmm5 ;rounder | ||
| 1380 : | psrad xmm2,%3 | ||
| 1381 : | psrad xmm6,%3 | ||
| 1382 : | |||
| 1383 : | packssdw xmm2,xmm6 | ||
| 1384 : | movdqa [%1],xmm2 | ||
| 1385 : | |||
| 1386 : | %endmacro | ||
| 1387 : | |||
| 1388 : | |||
| 1389 : | %macro FDCT_1D 4 ; INP,OUTP,{0=first pass, 1=second pass}, shift={11,18} | ||
| 1390 : | |||
| 1391 : | ;movdqa xmm0,[%1+16*0] ;We do not load 0-3 values here for they | ||
| 1392 : | ;movdqa xmm1,[%1+16*1] ;stayed from transposition | ||
| 1393 : | ;movdqa xmm2,[%1+16*2] | ||
| 1394 : | ;movdqa xmm3,[%1+16*3] | ||
| 1395 : | %if %3<>0 | ||
| 1396 : | movdqa xmm7,[rounder_5] | ||
| 1397 : | %endif | ||
| 1398 : | |||
| 1399 : | paddsw xmm0,[%1+16*7] ;tmp0 | ||
| 1400 : | movdqa xmm4,xmm0 | ||
| 1401 : | paddsw xmm1,[%1+16*6] ;tmp1 | ||
| 1402 : | movdqa xmm5,xmm1 | ||
| 1403 : | paddsw xmm2,[%1+16*5] ;tmp2 | ||
| 1404 : | paddsw xmm3,[%1+16*4] ;tmp3 | ||
| 1405 : | |||
| 1406 : | paddsw xmm0,xmm3 ;tmp10 | ||
| 1407 : | %if %3<>0 ; In the second pass we must round and shift before | ||
| 1408 : | ; the tmp10+tmp11 and tmp10-tmp11 calculation | ||
| 1409 : | ; or the overflow will happen. | ||
| 1410 : | paddsw xmm0,xmm7 ;[rounder_5] | ||
| 1411 : | psraw xmm0,PASS1_BITS+3 | ||
| 1412 : | %endif | ||
| 1413 : | movdqa xmm6,xmm0 ;tmp10 | ||
| 1414 : | paddsw xmm1,xmm2 ;tmp11 | ||
| 1415 : | psubsw xmm4,xmm3 ;tmp13 | ||
| 1416 : | psubsw xmm5,xmm2 ;tmp12 | ||
| 1417 : | |||
| 1418 : | %if %3=0 | ||
| 1419 : | paddsw xmm0,xmm1 | ||
| 1420 : | psubsw xmm6,xmm1 | ||
| 1421 : | |||
| 1422 : | psllw xmm0,PASS1_BITS | ||
| 1423 : | psllw xmm6,PASS1_BITS | ||
| 1424 : | %else | ||
| 1425 : | paddsw xmm1,xmm7 ;[rounder_5] | ||
| 1426 : | psraw xmm1,PASS1_BITS+3 | ||
| 1427 : | |||
| 1428 : | paddsw xmm0,xmm1 | ||
| 1429 : | psubsw xmm6,xmm1 | ||
| 1430 : | %endif | ||
| 1431 : | |||
| 1432 : | movdqa xmm1,xmm4 | ||
| 1433 : | movdqa xmm2,xmm4 | ||
| 1434 : | |||
| 1435 : | movdqa [%2+16*0],xmm0 | ||
| 1436 : | movdqa [%2+16*4],xmm6 | ||
| 1437 : | |||
| 1438 : | movdqa xmm7,[FIX_1] | ||
| 1439 : | |||
| 1440 : | punpckhwd xmm1,xmm5 ; 12 13 12 13 12 13 12 13 high part | ||
| 1441 : | movdqa xmm6,xmm1 ;high | ||
| 1442 : | punpcklwd xmm2,xmm5 ; 12 13 12 13 12 13 12 13 low part | ||
| 1443 : | movdqa xmm0,xmm2 ;low | ||
| 1444 : | |||
| 1445 : | movdqa xmm4,[FIX_2] | ||
| 1446 : | |||
| 1447 : | %if %4=11 | ||
| 1448 : | movdqa xmm5,[rounder_11] | ||
| 1449 : | %else | ||
| 1450 : | movdqa xmm5,[rounder_18] | ||
| 1451 : | %endif | ||
| 1452 : | |||
| 1453 : | pmaddwd xmm2,xmm7 ;[FIX_1] | ||
| 1454 : | pmaddwd xmm1,xmm7 ;[FIX_1] | ||
| 1455 : | pmaddwd xmm0,xmm4 ;[FIX_2] | ||
| 1456 : | pmaddwd xmm6,xmm4 ;[FIX_2] | ||
| 1457 : | |||
| 1458 : | paddd xmm2,xmm5 ;rounder | ||
| 1459 : | paddd xmm1,xmm5 ;rounder | ||
| 1460 : | psrad xmm2,%4 | ||
| 1461 : | psrad xmm1,%4 | ||
| 1462 : | |||
| 1463 : | packssdw xmm2,xmm1 | ||
| 1464 : | movdqa [%2+16*2],xmm2 | ||
| 1465 : | |||
| 1466 : | paddd xmm0,xmm5 ;rounder | ||
| 1467 : | paddd xmm6,xmm5 ;rounder | ||
| 1468 : | psrad xmm0,%4 | ||
| 1469 : | psrad xmm6,%4 | ||
| 1470 : | |||
| 1471 : | packssdw xmm0,xmm6 | ||
| 1472 : | movdqa [%2+16*6],xmm0 | ||
| 1473 : | |||
| 1474 : | movdqa xmm0,[%1+16*0] | ||
| 1475 : | movdqa xmm1,[%1+16*1] | ||
| 1476 : | movdqa xmm2,[%1+16*2] | ||
| 1477 : | movdqa xmm3,[%1+16*3] | ||
| 1478 : | |||
| 1479 : | psubsw xmm0,[%1+16*7] ;tmp7 | ||
| 1480 : | movdqa xmm4,xmm0 | ||
| 1481 : | psubsw xmm1,[%1+16*6] ;tmp6 | ||
| 1482 : | psubsw xmm2,[%1+16*5] ;tmp5 | ||
| 1483 : | movdqa xmm6,xmm2 | ||
| 1484 : | psubsw xmm3,[%1+16*4] ;tmp4 | ||
| 1485 : | |||
| 1486 : | punpckhwd xmm4,xmm1 ; 6 7 6 7 6 7 6 7 high part | ||
| 1487 : | punpcklwd xmm0,xmm1 ; 6 7 6 7 6 7 6 7 low part | ||
| 1488 : | punpckhwd xmm6,xmm3 ; 4 5 4 5 4 5 4 5 high part | ||
| 1489 : | punpcklwd xmm2,xmm3 ; 4 5 4 5 4 5 4 5 low part | ||
| 1490 : | |||
| 1491 : | makeoddpart %2+16*1,FIX_3,%4 | ||
| 1492 : | makeoddpart %2+16*3,FIX_4,%4 | ||
| 1493 : | makeoddpart %2+16*5,FIX_5,%4 | ||
| 1494 : | makeoddpartlast %2+16*7,FIX_6,%4 | ||
| 1495 : | |||
| 1496 : | %endmacro | ||
| 1497 : | |||
| 1498 : | cglobal fdct_sse2 | ||
| 1499 : | ;;void f dct_sse2(short *block); | ||
| 1500 : | fdct_sse2: | ||
| 1501 : | |||
| 1502 : | push eax | ||
| 1503 : | push ebx | ||
| 1504 : | |||
| 1505 : | mov eax,[esp+4+2*4] | ||
| 1506 : | mov ebx,buffer | ||
| 1507 : | |||
| 1508 : | prefetchnta [FIX_1] | ||
| 1509 : | prefetchnta [FIX_3] | ||
| 1510 : | prefetchnta [FIX_5] | ||
| 1511 : | |||
| 1512 : | transpose8x8sse2 eax,ebx,1 ; First we transpose last 4 lines | ||
| 1513 : | transpose8x8sse2 eax,ebx,0 ; Then the firts 4 lines | ||
| 1514 : | |||
| 1515 : | ;processing columns (became rows after transposition) | ||
| 1516 : | |||
| 1517 : | FDCT_1D ebx,eax,0,CONST_BITS - PASS1_BITS | ||
| 1518 : | |||
| 1519 : | transpose8x8sse2 eax,ebx,1 | ||
| 1520 : | transpose8x8sse2 eax,ebx,0 | ||
| 1521 : | |||
| 1522 : | ;now processing rows | ||
| 1523 : | |||
| 1524 : | FDCT_1D ebx,eax,1,CONST_BITS + PASS1_BITS + 3 | ||
| 1525 : | |||
| 1526 : | pop ebx | ||
| 1527 : | pop eax | ||
| 1528 : | |||
| 1529 : | ret |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |