Parent Directory
| 
Revision Log
Revision 1790 -
(view)
(download)
Original Path: trunk/xvidcore/src/image/x86_asm/colorspace_rgb_mmx.asm
| 1 : | edgomez | 1382 | ;/***************************************************************************** |
| 2 : | edgomez | 851 | ; * |
| 3 : | edgomez | 1382 | ; * XVID MPEG-4 VIDEO CODEC |
| 4 : | ; * - RGB colorspace conversions - | ||
| 5 : | edgomez | 851 | ; * |
| 6 : | edgomez | 1382 | ; * Copyright(C) 2002-2003 Michael Militzer <isibaar@xvid.org> |
| 7 : | ; * 2002-2003 Peter Ross <pross@xvid.org> | ||
| 8 : | edgomez | 851 | ; * |
| 9 : | edgomez | 1382 | ; * This program is free software ; you can redistribute it and/or modify |
| 10 : | ; * it 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 : | edgomez | 851 | ; * |
| 14 : | edgomez | 1382 | ; * 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 : | edgomez | 851 | ; * |
| 19 : | edgomez | 1382 | ; * 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 : | edgomez | 851 | ; * |
| 23 : | edgomez | 1382 | ; ****************************************************************************/ |
| 24 : | edgomez | 851 | |
| 25 : | edgomez | 1382 | BITS 32 |
| 26 : | edgomez | 851 | |
| 27 : | edgomez | 1382 | %macro cglobal 1 |
| 28 : | edgomez | 851 | %ifdef PREFIX |
| 29 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 30 : | edgomez | 1540 | global _%1:function %1.endfunc-%1 |
| 31 : | %define %1 _%1:function %1.endfunc-%1 | ||
| 32 : | edgomez | 1535 | %else |
| 33 : | global _%1 | ||
| 34 : | %define %1 _%1 | ||
| 35 : | %endif | ||
| 36 : | edgomez | 851 | %else |
| 37 : | edgomez | 1535 | %ifdef MARK_FUNCS |
| 38 : | edgomez | 1540 | global %1:function %1.endfunc-%1 |
| 39 : | edgomez | 1535 | %else |
| 40 : | global %1 | ||
| 41 : | %endif | ||
| 42 : | edgomez | 851 | %endif |
| 43 : | %endmacro | ||
| 44 : | |||
| 45 : | edgomez | 1382 | ;============================================================================= |
| 46 : | ; Some constants | ||
| 47 : | ;============================================================================= | ||
| 48 : | edgomez | 851 | |
| 49 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 50 : | edgomez | 851 | ; RGB->YV12 yuv constants |
| 51 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 52 : | |||
| 53 : | edgomez | 851 | %define Y_R 0.257 |
| 54 : | %define Y_G 0.504 | ||
| 55 : | %define Y_B 0.098 | ||
| 56 : | %define Y_ADD 16 | ||
| 57 : | |||
| 58 : | %define U_R 0.148 | ||
| 59 : | %define U_G 0.291 | ||
| 60 : | %define U_B 0.439 | ||
| 61 : | %define U_ADD 128 | ||
| 62 : | |||
| 63 : | %define V_R 0.439 | ||
| 64 : | %define V_G 0.368 | ||
| 65 : | %define V_B 0.071 | ||
| 66 : | %define V_ADD 128 | ||
| 67 : | |||
| 68 : | edgomez | 1382 | ; Scaling used during conversion |
| 69 : | %define SCALEBITS 6 | ||
| 70 : | edgomez | 851 | |
| 71 : | edgomez | 1382 | ;============================================================================= |
| 72 : | ; Read only data | ||
| 73 : | ;============================================================================= | ||
| 74 : | edgomez | 851 | |
| 75 : | edgomez | 1382 | %ifdef FORMAT_COFF |
| 76 : | edgomez | 1519 | SECTION .rodata |
| 77 : | edgomez | 1382 | %else |
| 78 : | edgomez | 1519 | SECTION .rodata align=16 |
| 79 : | edgomez | 1382 | %endif |
| 80 : | ALIGN 16 | ||
| 81 : | edgomez | 851 | |
| 82 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 83 : | ; RGB->YV12 multiplication matrices | ||
| 84 : | ;----------------------------------------------------------------------------- | ||
| 85 : | ; FIX(Y_B) FIX(Y_G) FIX(Y_R) Ignored | ||
| 86 : | edgomez | 851 | |
| 87 : | chl | 1763 | bgr_y_mul: dw 25, 129, 66, 0 |
| 88 : | bgr_u_mul: dw 112, -74, -38, 0 | ||
| 89 : | bgr_v_mul: dw -18, -94, 112, 0 | ||
| 90 : | edgomez | 851 | |
| 91 : | chl | 1763 | ;----------------------------------------------------------------------------- |
| 92 : | ; BGR->YV12 multiplication matrices | ||
| 93 : | ;----------------------------------------------------------------------------- | ||
| 94 : | ; FIX(Y_R) FIX(Y_G) FIX(Y_B) Ignored | ||
| 95 : | edgomez | 851 | |
| 96 : | chl | 1763 | rgb_y_mul: dw 66, 129, 25, 0 |
| 97 : | rgb_u_mul: dw -38, -74, 112, 0 | ||
| 98 : | rgb_v_mul: dw 112, -94, -18, 0 | ||
| 99 : | |||
| 100 : | edgomez | 1382 | ;----------------------------------------------------------------------------- |
| 101 : | ; YV12->RGB data | ||
| 102 : | ;----------------------------------------------------------------------------- | ||
| 103 : | edgomez | 851 | |
| 104 : | edgomez | 1382 | Y_SUB: dw 16, 16, 16, 16 |
| 105 : | U_SUB: dw 128, 128, 128, 128 | ||
| 106 : | V_SUB: dw 128, 128, 128, 128 | ||
| 107 : | edgomez | 851 | |
| 108 : | edgomez | 1382 | Y_MUL: dw 74, 74, 74, 74 |
| 109 : | edgomez | 851 | |
| 110 : | edgomez | 1382 | UG_MUL: dw 25, 25, 25, 25 |
| 111 : | VG_MUL: dw 52, 52, 52, 52 | ||
| 112 : | edgomez | 851 | |
| 113 : | edgomez | 1382 | UB_MUL: dw 129, 129, 129, 129 |
| 114 : | VR_MUL: dw 102, 102, 102, 102 | ||
| 115 : | edgomez | 851 | |
| 116 : | edgomez | 1382 | BRIGHT: db 128, 128, 128, 128, 128, 128, 128, 128 |
| 117 : | |||
| 118 : | ;============================================================================= | ||
| 119 : | ; Helper macros used with the colorspace_mmx.inc file | ||
| 120 : | ;============================================================================= | ||
| 121 : | |||
| 122 : | edgomez | 851 | ;------------------------------------------------------------------------------ |
| 123 : | ; BGR_TO_YV12( BYTES ) | ||
| 124 : | ; | ||
| 125 : | ; BYTES 3=bgr(24bit), 4=bgra(32-bit) | ||
| 126 : | ; | ||
| 127 : | ; bytes=3/4, pixels = 2, vpixels=2 | ||
| 128 : | ;------------------------------------------------------------------------------ | ||
| 129 : | edgomez | 1382 | |
| 130 : | edgomez | 851 | %macro BGR_TO_YV12_INIT 2 |
| 131 : | chl | 1763 | movq mm7, [bgr_y_mul] |
| 132 : | edgomez | 851 | %endmacro |
| 133 : | |||
| 134 : | |||
| 135 : | %macro BGR_TO_YV12 2 | ||
| 136 : | edgomez | 1382 | ; y_out |
| 137 : | pxor mm4, mm4 | ||
| 138 : | pxor mm5, mm5 | ||
| 139 : | movd mm0, [edi] ; x_ptr[0...] | ||
| 140 : | movd mm2, [edi+edx] ; x_ptr[x_stride...] | ||
| 141 : | punpcklbw mm0, mm4 ; [ |b |g |r ] | ||
| 142 : | punpcklbw mm2, mm5 ; [ |b |g |r ] | ||
| 143 : | movq mm6, mm0 ; = [ |b4|g4|r4] | ||
| 144 : | paddw mm6, mm2 ; +[ |b4|g4|r4] | ||
| 145 : | pmaddwd mm0, mm7 ; *= Y_MUL | ||
| 146 : | pmaddwd mm2, mm7 ; *= Y_MUL | ||
| 147 : | movq mm4, mm0 ; [r] | ||
| 148 : | movq mm5, mm2 ; [r] | ||
| 149 : | psrlq mm4, 32 ; +[g] | ||
| 150 : | psrlq mm5, 32 ; +[g] | ||
| 151 : | paddd mm0, mm4 ; +[b] | ||
| 152 : | paddd mm2, mm5 ; +[b] | ||
| 153 : | edgomez | 851 | |
| 154 : | edgomez | 1382 | pxor mm4, mm4 |
| 155 : | pxor mm5, mm5 | ||
| 156 : | movd mm1, [edi+%1] ; src[%1...] | ||
| 157 : | movd mm3, [edi+edx+%1] ; src[x_stride+%1...] | ||
| 158 : | punpcklbw mm1, mm4 ; [ |b |g |r ] | ||
| 159 : | punpcklbw mm3, mm5 ; [ |b |g |r ] | ||
| 160 : | paddw mm6, mm1 ; +[ |b4|g4|r4] | ||
| 161 : | paddw mm6, mm3 ; +[ |b4|g4|r4] | ||
| 162 : | pmaddwd mm1, mm7 ; *= Y_MUL | ||
| 163 : | pmaddwd mm3, mm7 ; *= Y_MUL | ||
| 164 : | movq mm4, mm1 ; [r] | ||
| 165 : | movq mm5, mm3 ; [r] | ||
| 166 : | psrlq mm4, 32 ; +[g] | ||
| 167 : | psrlq mm5, 32 ; +[g] | ||
| 168 : | paddd mm1, mm4 ; +[b] | ||
| 169 : | paddd mm3, mm5 ; +[b] | ||
| 170 : | edgomez | 851 | |
| 171 : | edgomez | 1382 | push edx |
| 172 : | edgomez | 851 | |
| 173 : | edgomez | 1382 | movd edx, mm0 |
| 174 : | shr edx, 8 | ||
| 175 : | add edx, Y_ADD | ||
| 176 : | mov [esi], dl ; y_ptr[0] | ||
| 177 : | edgomez | 851 | |
| 178 : | edgomez | 1382 | movd edx, mm1 |
| 179 : | shr edx, 8 | ||
| 180 : | add edx, Y_ADD | ||
| 181 : | mov [esi + 1], dl ; y_ptr[1] | ||
| 182 : | edgomez | 851 | |
| 183 : | edgomez | 1382 | movd edx, mm2 |
| 184 : | shr edx, 8 | ||
| 185 : | add edx, Y_ADD | ||
| 186 : | mov [esi + eax + 0], dl ; y_ptr[y_stride + 0] | ||
| 187 : | edgomez | 851 | |
| 188 : | edgomez | 1382 | movd edx, mm3 |
| 189 : | shr edx, 8 | ||
| 190 : | add edx, Y_ADD | ||
| 191 : | mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] | ||
| 192 : | edgomez | 851 | |
| 193 : | edgomez | 1382 | ; u_ptr, v_ptr |
| 194 : | movq mm0, mm6 ; = [ |b4|g4|r4] | ||
| 195 : | chl | 1763 | pmaddwd mm6, [bgr_v_mul] ; *= V_MUL |
| 196 : | pmaddwd mm0, [bgr_u_mul] ; *= U_MUL | ||
| 197 : | edgomez | 1382 | movq mm1, mm0 |
| 198 : | movq mm2, mm6 | ||
| 199 : | psrlq mm1, 32 | ||
| 200 : | psrlq mm2, 32 | ||
| 201 : | paddd mm0, mm1 | ||
| 202 : | paddd mm2, mm6 | ||
| 203 : | edgomez | 851 | |
| 204 : | edgomez | 1382 | movd edx, mm0 |
| 205 : | shr edx, 10 | ||
| 206 : | add edx, U_ADD | ||
| 207 : | mov [ebx], dl | ||
| 208 : | edgomez | 851 | |
| 209 : | edgomez | 1382 | movd edx, mm2 |
| 210 : | shr edx, 10 | ||
| 211 : | add edx, V_ADD | ||
| 212 : | mov [ecx], dl | ||
| 213 : | edgomez | 851 | |
| 214 : | edgomez | 1382 | pop edx |
| 215 : | edgomez | 851 | %endmacro |
| 216 : | |||
| 217 : | ;------------------------------------------------------------------------------ | ||
| 218 : | chl | 1763 | ; RGB_TO_YV12( BYTES ) |
| 219 : | ; | ||
| 220 : | ; BYTES 3=rgb(24bit), 4=rgba(32-bit) | ||
| 221 : | ; | ||
| 222 : | ; bytes=3/4, pixels = 2, vpixels=2 | ||
| 223 : | ;------------------------------------------------------------------------------ | ||
| 224 : | |||
| 225 : | %macro RGB_TO_YV12_INIT 2 | ||
| 226 : | movq mm7, [rgb_y_mul] | ||
| 227 : | %endmacro | ||
| 228 : | |||
| 229 : | |||
| 230 : | %macro RGB_TO_YV12 2 | ||
| 231 : | ; y_out | ||
| 232 : | pxor mm4, mm4 | ||
| 233 : | pxor mm5, mm5 | ||
| 234 : | movd mm0, [edi] ; x_ptr[0...] | ||
| 235 : | movd mm2, [edi+edx] ; x_ptr[x_stride...] | ||
| 236 : | punpcklbw mm0, mm4 ; [ |b |g |r ] | ||
| 237 : | punpcklbw mm2, mm5 ; [ |b |g |r ] | ||
| 238 : | movq mm6, mm0 ; = [ |b4|g4|r4] | ||
| 239 : | paddw mm6, mm2 ; +[ |b4|g4|r4] | ||
| 240 : | pmaddwd mm0, mm7 ; *= Y_MUL | ||
| 241 : | pmaddwd mm2, mm7 ; *= Y_MUL | ||
| 242 : | movq mm4, mm0 ; [r] | ||
| 243 : | movq mm5, mm2 ; [r] | ||
| 244 : | psrlq mm4, 32 ; +[g] | ||
| 245 : | psrlq mm5, 32 ; +[g] | ||
| 246 : | paddd mm0, mm4 ; +[b] | ||
| 247 : | paddd mm2, mm5 ; +[b] | ||
| 248 : | |||
| 249 : | pxor mm4, mm4 | ||
| 250 : | pxor mm5, mm5 | ||
| 251 : | movd mm1, [edi+%1] ; src[%1...] | ||
| 252 : | movd mm3, [edi+edx+%1] ; src[x_stride+%1...] | ||
| 253 : | punpcklbw mm1, mm4 ; [ |b |g |r ] | ||
| 254 : | punpcklbw mm3, mm5 ; [ |b |g |r ] | ||
| 255 : | paddw mm6, mm1 ; +[ |b4|g4|r4] | ||
| 256 : | paddw mm6, mm3 ; +[ |b4|g4|r4] | ||
| 257 : | pmaddwd mm1, mm7 ; *= Y_MUL | ||
| 258 : | pmaddwd mm3, mm7 ; *= Y_MUL | ||
| 259 : | movq mm4, mm1 ; [r] | ||
| 260 : | movq mm5, mm3 ; [r] | ||
| 261 : | psrlq mm4, 32 ; +[g] | ||
| 262 : | psrlq mm5, 32 ; +[g] | ||
| 263 : | paddd mm1, mm4 ; +[b] | ||
| 264 : | paddd mm3, mm5 ; +[b] | ||
| 265 : | |||
| 266 : | push edx | ||
| 267 : | |||
| 268 : | movd edx, mm0 | ||
| 269 : | shr edx, 8 | ||
| 270 : | add edx, Y_ADD | ||
| 271 : | mov [esi], dl ; y_ptr[0] | ||
| 272 : | |||
| 273 : | movd edx, mm1 | ||
| 274 : | shr edx, 8 | ||
| 275 : | add edx, Y_ADD | ||
| 276 : | mov [esi + 1], dl ; y_ptr[1] | ||
| 277 : | |||
| 278 : | movd edx, mm2 | ||
| 279 : | shr edx, 8 | ||
| 280 : | add edx, Y_ADD | ||
| 281 : | mov [esi + eax + 0], dl ; y_ptr[y_stride + 0] | ||
| 282 : | |||
| 283 : | movd edx, mm3 | ||
| 284 : | shr edx, 8 | ||
| 285 : | add edx, Y_ADD | ||
| 286 : | mov [esi + eax + 1], dl ; y_ptr[y_stride + 1] | ||
| 287 : | |||
| 288 : | ; u_ptr, v_ptr | ||
| 289 : | movq mm0, mm6 ; = [ |b4|g4|r4] | ||
| 290 : | pmaddwd mm6, [rgb_v_mul] ; *= V_MUL | ||
| 291 : | pmaddwd mm0, [rgb_u_mul] ; *= U_MUL | ||
| 292 : | movq mm1, mm0 | ||
| 293 : | movq mm2, mm6 | ||
| 294 : | psrlq mm1, 32 | ||
| 295 : | psrlq mm2, 32 | ||
| 296 : | paddd mm0, mm1 | ||
| 297 : | paddd mm2, mm6 | ||
| 298 : | |||
| 299 : | movd edx, mm0 | ||
| 300 : | shr edx, 10 | ||
| 301 : | add edx, U_ADD | ||
| 302 : | mov [ebx], dl | ||
| 303 : | |||
| 304 : | movd edx, mm2 | ||
| 305 : | shr edx, 10 | ||
| 306 : | add edx, V_ADD | ||
| 307 : | mov [ecx], dl | ||
| 308 : | |||
| 309 : | pop edx | ||
| 310 : | %endmacro | ||
| 311 : | |||
| 312 : | ;------------------------------------------------------------------------------ | ||
| 313 : | edgomez | 851 | ; YV12_TO_BGR( BYTES ) |
| 314 : | ; | ||
| 315 : | ; BYTES 3=bgr(24-bit), 4=bgra(32-bit) | ||
| 316 : | ; | ||
| 317 : | ; bytes=3/4, pixels = 8, vpixels=2 | ||
| 318 : | ;------------------------------------------------------------------------------ | ||
| 319 : | edgomez | 1382 | |
| 320 : | edgomez | 851 | %macro YV12_TO_BGR_INIT 2 |
| 321 : | edgomez | 1382 | pxor mm7, mm7 ; clear mm7 |
| 322 : | edgomez | 851 | %endmacro |
| 323 : | |||
| 324 : | %macro YV12_TO_BGR 2 | ||
| 325 : | %define TEMP_Y1 esp | ||
| 326 : | %define TEMP_Y2 esp + 8 | ||
| 327 : | %define TEMP_G1 esp + 16 | ||
| 328 : | %define TEMP_G2 esp + 24 | ||
| 329 : | %define TEMP_B1 esp + 32 | ||
| 330 : | %define TEMP_B2 esp + 40 | ||
| 331 : | |||
| 332 : | edgomez | 1382 | movd mm2, [ebx] ; u_ptr[0] |
| 333 : | movd mm3, [ecx] ; v_ptr[0] | ||
| 334 : | punpcklbw mm2, mm7 ; u3u2u1u0 -> mm2 | ||
| 335 : | punpcklbw mm3, mm7 ; v3v2v1v0 -> mm3 | ||
| 336 : | psubsw mm2, [U_SUB] ; U - 128 | ||
| 337 : | psubsw mm3, [V_SUB] ; V - 128 | ||
| 338 : | movq mm4, mm2 | ||
| 339 : | movq mm5, mm3 | ||
| 340 : | pmullw mm2, [UG_MUL] | ||
| 341 : | pmullw mm3, [VG_MUL] | ||
| 342 : | movq mm6, mm2 ; u3u2u1u0 -> mm6 | ||
| 343 : | punpckhwd mm2, mm2 ; u3u3u2u2 -> mm2 | ||
| 344 : | punpcklwd mm6, mm6 ; u1u1u0u0 -> mm6 | ||
| 345 : | pmullw mm4, [UB_MUL] ; B_ADD -> mm4 | ||
| 346 : | movq mm0, mm3 | ||
| 347 : | punpckhwd mm3, mm3 ; v3v3v2v2 -> mm2 | ||
| 348 : | punpcklwd mm0, mm0 ; v1v1v0v0 -> mm6 | ||
| 349 : | paddsw mm2, mm3 | ||
| 350 : | paddsw mm6, mm0 | ||
| 351 : | pmullw mm5, [VR_MUL] ; R_ADD -> mm5 | ||
| 352 : | movq mm0, [esi] ; y7y6y5y4y3y2y1y0 -> mm0 | ||
| 353 : | movq mm1, mm0 | ||
| 354 : | punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 | ||
| 355 : | punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 | ||
| 356 : | psubsw mm0, [Y_SUB] ; Y - Y_SUB | ||
| 357 : | psubsw mm1, [Y_SUB] ; Y - Y_SUB | ||
| 358 : | pmullw mm1, [Y_MUL] | ||
| 359 : | pmullw mm0, [Y_MUL] | ||
| 360 : | movq [TEMP_Y2], mm1 ; y7y6y5y4 -> mm3 | ||
| 361 : | movq [TEMP_Y1], mm0 ; y3y2y1y0 -> mm7 | ||
| 362 : | psubsw mm1, mm2 ; g7g6g5g4 -> mm1 | ||
| 363 : | psubsw mm0, mm6 ; g3g2g1g0 -> mm0 | ||
| 364 : | psraw mm1, SCALEBITS | ||
| 365 : | psraw mm0, SCALEBITS | ||
| 366 : | packuswb mm0, mm1 ;g7g6g5g4g3g2g1g0 -> mm0 | ||
| 367 : | movq [TEMP_G1], mm0 | ||
| 368 : | movq mm0, [esi+eax] ; y7y6y5y4y3y2y1y0 -> mm0 | ||
| 369 : | movq mm1, mm0 | ||
| 370 : | punpckhbw mm1, mm7 ; y7y6y5y4 -> mm1 | ||
| 371 : | punpcklbw mm0, mm7 ; y3y2y1y0 -> mm0 | ||
| 372 : | psubsw mm0, [Y_SUB] ; Y - Y_SUB | ||
| 373 : | psubsw mm1, [Y_SUB] ; Y - Y_SUB | ||
| 374 : | pmullw mm1, [Y_MUL] | ||
| 375 : | pmullw mm0, [Y_MUL] | ||
| 376 : | movq mm3, mm1 | ||
| 377 : | psubsw mm1, mm2 ; g7g6g5g4 -> mm1 | ||
| 378 : | movq mm2, mm0 | ||
| 379 : | psubsw mm0, mm6 ; g3g2g1g0 -> mm0 | ||
| 380 : | psraw mm1, SCALEBITS | ||
| 381 : | psraw mm0, SCALEBITS | ||
| 382 : | packuswb mm0, mm1 ; g7g6g5g4g3g2g1g0 -> mm0 | ||
| 383 : | movq [TEMP_G2], mm0 | ||
| 384 : | movq mm0, mm4 | ||
| 385 : | punpckhwd mm4, mm4 ; u3u3u2u2 -> mm2 | ||
| 386 : | punpcklwd mm0, mm0 ; u1u1u0u0 -> mm6 | ||
| 387 : | movq mm1, mm3 ; y7y6y5y4 -> mm1 | ||
| 388 : | paddsw mm3, mm4 ; b7b6b5b4 -> mm3 | ||
| 389 : | movq mm7, mm2 ; y3y2y1y0 -> mm7 | ||
| 390 : | paddsw mm2, mm0 ; b3b2b1b0 -> mm2 | ||
| 391 : | psraw mm3, SCALEBITS | ||
| 392 : | psraw mm2, SCALEBITS | ||
| 393 : | packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 | ||
| 394 : | movq [TEMP_B2], mm2 | ||
| 395 : | movq mm3, [TEMP_Y2] | ||
| 396 : | movq mm2, [TEMP_Y1] | ||
| 397 : | movq mm6, mm3 ; TEMP_Y2 -> mm6 | ||
| 398 : | paddsw mm3, mm4 ; b7b6b5b4 -> mm3 | ||
| 399 : | movq mm4, mm2 ; TEMP_Y1 -> mm4 | ||
| 400 : | paddsw mm2, mm0 ; b3b2b1b0 -> mm2 | ||
| 401 : | psraw mm3, SCALEBITS | ||
| 402 : | psraw mm2, SCALEBITS | ||
| 403 : | packuswb mm2, mm3 ; b7b6b5b4b3b2b1b0 -> mm2 | ||
| 404 : | movq [TEMP_B1], mm2 | ||
| 405 : | movq mm0, mm5 | ||
| 406 : | punpckhwd mm5, mm5 ; v3v3v2v2 -> mm5 | ||
| 407 : | punpcklwd mm0, mm0 ; v1v1v0v0 -> mm0 | ||
| 408 : | paddsw mm1, mm5 ; r7r6r5r4 -> mm1 | ||
| 409 : | paddsw mm7, mm0 ; r3r2r1r0 -> mm7 | ||
| 410 : | psraw mm1, SCALEBITS | ||
| 411 : | psraw mm7, SCALEBITS | ||
| 412 : | packuswb mm7, mm1 ; r7r6r5r4r3r2r1r0 -> mm7 (TEMP_R2) | ||
| 413 : | paddsw mm6, mm5 ; r7r6r5r4 -> mm6 | ||
| 414 : | paddsw mm4, mm0 ; r3r2r1r0 -> mm4 | ||
| 415 : | psraw mm6, SCALEBITS | ||
| 416 : | psraw mm4, SCALEBITS | ||
| 417 : | packuswb mm4, mm6 ; r7r6r5r4r3r2r1r0 -> mm4 (TEMP_R1) | ||
| 418 : | movq mm0, [TEMP_B1] | ||
| 419 : | movq mm1, [TEMP_G1] | ||
| 420 : | movq mm6, mm7 | ||
| 421 : | movq mm2, mm0 | ||
| 422 : | punpcklbw mm2, mm4 ; r3b3r2b2r1b1r0b0 -> mm2 | ||
| 423 : | punpckhbw mm0, mm4 ; r7b7r6b6r5b5r4b4 -> mm0 | ||
| 424 : | pxor mm7, mm7 | ||
| 425 : | movq mm3, mm1 | ||
| 426 : | punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 | ||
| 427 : | punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 | ||
| 428 : | movq mm4, mm2 | ||
| 429 : | punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 | ||
| 430 : | punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 | ||
| 431 : | movq mm5, mm0 | ||
| 432 : | punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 | ||
| 433 : | punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 | ||
| 434 : | %if %1 == 3 ; BGR (24-bit) | ||
| 435 : | movd [edi], mm2 | ||
| 436 : | psrlq mm2, 32 | ||
| 437 : | movd [edi + 3], mm2 | ||
| 438 : | movd [edi + 6], mm4 | ||
| 439 : | psrlq mm4, 32 | ||
| 440 : | movd [edi + 9], mm4 | ||
| 441 : | movd [edi + 12], mm0 | ||
| 442 : | psrlq mm0, 32 | ||
| 443 : | movd [edi + 15], mm0 | ||
| 444 : | movq mm2, mm5 | ||
| 445 : | psrlq mm0, 8 ; 000000r5g5 -> mm0 | ||
| 446 : | psllq mm2, 32 ; 0r6g6b60000 -> mm2 | ||
| 447 : | psrlq mm5, 32 ; 00000r7g7b7 -> mm5 | ||
| 448 : | psrlq mm2, 16 ; 000r6g6b600 -> mm2 | ||
| 449 : | por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 | ||
| 450 : | psllq mm5, 40 ; r7g7b700000 -> mm5 | ||
| 451 : | por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 | ||
| 452 : | movq [edi + 16], mm5 | ||
| 453 : | movq mm0, [TEMP_B2] | ||
| 454 : | movq mm1, [TEMP_G2] | ||
| 455 : | movq mm2, mm0 | ||
| 456 : | punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 | ||
| 457 : | punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 | ||
| 458 : | movq mm3, mm1 | ||
| 459 : | punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 | ||
| 460 : | punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 | ||
| 461 : | movq mm4, mm2 | ||
| 462 : | punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 | ||
| 463 : | punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 | ||
| 464 : | movq mm5, mm0 | ||
| 465 : | punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 | ||
| 466 : | punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 | ||
| 467 : | movd [edi+edx], mm2 | ||
| 468 : | psrlq mm2, 32 | ||
| 469 : | movd [edi+edx + 3], mm2 | ||
| 470 : | movd [edi+edx + 6], mm4 | ||
| 471 : | psrlq mm4, 32 | ||
| 472 : | movd [edi+edx + 9], mm4 | ||
| 473 : | movd [edi+edx + 12], mm0 | ||
| 474 : | psrlq mm0, 32 | ||
| 475 : | movd [edi+edx + 15], mm0 | ||
| 476 : | movq mm2, mm5 | ||
| 477 : | psrlq mm0, 8 ; 000000r5g5 -> mm0 | ||
| 478 : | psllq mm2, 32 ; 0r6g6b60000 -> mm2 | ||
| 479 : | psrlq mm5, 32 ; 00000r7g7b7 -> mm5 | ||
| 480 : | psrlq mm2, 16 ; 000r6g6b600 -> mm2 | ||
| 481 : | por mm0, mm2 ; 000r6g6b6r5g5 -> mm0 | ||
| 482 : | psllq mm5, 40 ; r7g7b700000 -> mm5 | ||
| 483 : | por mm5, mm0 ; r7g7b7r6g6b6r5g5 -> mm5 | ||
| 484 : | movq [edi + edx + 16], mm5 | ||
| 485 : | edgomez | 851 | |
| 486 : | edgomez | 1382 | %else ; BGRA (32-bit) |
| 487 : | movq [edi], mm2 | ||
| 488 : | movq [edi + 8], mm4 | ||
| 489 : | movq [edi + 16], mm0 | ||
| 490 : | movq [edi + 24], mm5 | ||
| 491 : | movq mm0, [TEMP_B2] | ||
| 492 : | movq mm1, [TEMP_G2] | ||
| 493 : | movq mm2, mm0 | ||
| 494 : | punpcklbw mm2, mm6 ; r3b3r2b2r1b1r0b0 -> mm2 | ||
| 495 : | punpckhbw mm0, mm6 ; r7b7r6b6r5b5r4b4 -> mm0 | ||
| 496 : | movq mm3, mm1 | ||
| 497 : | punpcklbw mm1, mm7 ; 0g30g20g10g0 -> mm1 | ||
| 498 : | punpckhbw mm3, mm7 ; 0g70g60g50g4 -> mm3 | ||
| 499 : | movq mm4, mm2 | ||
| 500 : | punpcklbw mm2, mm1 ; 0r1g1b10r0g0b0 -> mm2 | ||
| 501 : | punpckhbw mm4, mm1 ; 0r3g3b30r2g2b2 -> mm4 | ||
| 502 : | movq mm5, mm0 | ||
| 503 : | punpcklbw mm0, mm3 ; 0r5g5b50r4g4b4 -> mm0 | ||
| 504 : | punpckhbw mm5, mm3 ; 0r7g7b70r6g6b6 -> mm5 | ||
| 505 : | movq [edi + edx], mm2 | ||
| 506 : | movq [edi + edx + 8], mm4 | ||
| 507 : | movq [edi + edx + 16], mm0 | ||
| 508 : | movq [edi + edx + 24], mm5 | ||
| 509 : | edgomez | 851 | %endif |
| 510 : | |||
| 511 : | %undef TEMP_Y1 | ||
| 512 : | %undef TEMP_Y2 | ||
| 513 : | %undef TEMP_G1 | ||
| 514 : | %undef TEMP_G2 | ||
| 515 : | %undef TEMP_B1 | ||
| 516 : | %undef TEMP_B2 | ||
| 517 : | %endmacro | ||
| 518 : | |||
| 519 : | edgomez | 1382 | ;============================================================================= |
| 520 : | ; Code | ||
| 521 : | ;============================================================================= | ||
| 522 : | edgomez | 851 | |
| 523 : | edgomez | 1382 | SECTION .text |
| 524 : | edgomez | 851 | |
| 525 : | edgomez | 1382 | %include "colorspace_mmx.inc" |
| 526 : | edgomez | 851 | |
| 527 : | ; input | ||
| 528 : | MAKE_COLORSPACE bgr_to_yv12_mmx,0, 3,2,2, BGR_TO_YV12, 3, -1 | ||
| 529 : | MAKE_COLORSPACE bgra_to_yv12_mmx,0, 4,2,2, BGR_TO_YV12, 4, -1 | ||
| 530 : | chl | 1763 | MAKE_COLORSPACE rgb_to_yv12_mmx,0, 3,2,2, RGB_TO_YV12, 3, -1 |
| 531 : | MAKE_COLORSPACE rgba_to_yv12_mmx,0, 4,2,2, RGB_TO_YV12, 4, -1 | ||
| 532 : | edgomez | 851 | |
| 533 : | ; output | ||
| 534 : | MAKE_COLORSPACE yv12_to_bgr_mmx,48, 3,8,2, YV12_TO_BGR, 3, -1 | ||
| 535 : | MAKE_COLORSPACE yv12_to_bgra_mmx,48, 4,8,2, YV12_TO_BGR, 4, -1 | ||
| 536 : | |||
| 537 : | Isibaar | 1790 | |
| 538 : | %ifidn __OUTPUT_FORMAT__,elf | ||
| 539 : | section ".note.GNU-stack" noalloc noexec nowrite progbits | ||
| 540 : | %endif | ||
| 541 : |
| No admin address has been configured | ViewVC Help |
| Powered by ViewVC 1.0.4 |