davem69 | 775355a | 2001-06-05 23:54:00 +0000 | [diff] [blame^] | 1 | /* $Id: xform.S,v 1.2 2001/06/05 23:54:01 davem69 Exp $ */ |
Brian Paul | 7943b34 | 2001-05-23 14:27:03 +0000 | [diff] [blame] | 2 | |
| 3 | /* TODO |
| 4 | * |
| 5 | * 1) It would be nice if load/store double could be used |
| 6 | * at least for the matrix parts. I think for the matrices |
| 7 | * it is safe, but for the vertices it probably is not due to |
| 8 | * things like glInterleavedArrays etc. |
| 9 | * |
| 10 | * UPDATE: Trying this now in sparc_matrix.h -DaveM_990624 |
| 11 | * |
| 12 | * 2) One extremely slick trick would be if we could enclose |
| 13 | * groups of xform calls on the same vertices such that |
| 14 | * we just load the matrix into f16-->f31 before the calls |
| 15 | * and then we would not have to do them here. This may be |
| 16 | * tricky and not much of a gain though. |
| 17 | */ |
| 18 | |
davem69 | 775355a | 2001-06-05 23:54:00 +0000 | [diff] [blame^] | 19 | #include "sparc_matrix.h" |
Brian Paul | 7943b34 | 2001-05-23 14:27:03 +0000 | [diff] [blame] | 20 | |
| 21 | .text |
| 22 | .align 64 |
| 23 | |
| 24 | __set_v4f_1: |
| 25 | ld [%o0 + V4F_FLAGS], %g2 |
| 26 | mov 1, %g1 |
| 27 | st %g1, [%o0 + V4F_SIZE] |
| 28 | or %g2, VEC_SIZE_1, %g2 |
| 29 | retl |
| 30 | st %g2, [%o0 + V4F_FLAGS] |
| 31 | __set_v4f_2: |
| 32 | ld [%o0 + V4F_FLAGS], %g2 |
| 33 | mov 2, %g1 |
| 34 | st %g1, [%o0 + V4F_SIZE] |
| 35 | or %g2, VEC_SIZE_2, %g2 |
| 36 | retl |
| 37 | st %g2, [%o0 + V4F_FLAGS] |
| 38 | __set_v4f_3: |
| 39 | ld [%o0 + V4F_FLAGS], %g2 |
| 40 | mov 3, %g1 |
| 41 | st %g1, [%o0 + V4F_SIZE] |
| 42 | or %g2, VEC_SIZE_3, %g2 |
| 43 | retl |
| 44 | st %g2, [%o0 + V4F_FLAGS] |
| 45 | __set_v4f_4: |
| 46 | ld [%o0 + V4F_FLAGS], %g2 |
| 47 | mov 4, %g1 |
| 48 | st %g1, [%o0 + V4F_SIZE] |
| 49 | or %g2, VEC_SIZE_4, %g2 |
| 50 | retl |
| 51 | st %g2, [%o0 + V4F_FLAGS] |
| 52 | |
Brian Paul | 7943b34 | 2001-05-23 14:27:03 +0000 | [diff] [blame] | 53 | /* First the raw versions. */ |
| 54 | |
| 55 | .globl _mesa_sparc_transform_points1_general |
| 56 | _mesa_sparc_transform_points1_general: |
| 57 | ld [%o2 + V4F_STRIDE], %o5 |
| 58 | LDPTR [%o2 + V4F_START], %g1 |
| 59 | LDPTR [%o0 + V4F_START], %g2 |
| 60 | ld [%o2 + V4F_COUNT], %g3 |
| 61 | |
| 62 | LDMATRIX_0_1_2_3_12_13_14_15(%o1) |
| 63 | |
| 64 | cmp %g3, 1 |
| 65 | st %g3, [%o0 + V4F_COUNT] |
| 66 | bl 3f |
| 67 | clr %o1 |
| 68 | |
| 69 | be 2f |
| 70 | andn %g3, 1, %o2 |
| 71 | |
| 72 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 73 | add %g1, %o5, %g1 ! IEU0 |
| 74 | ld [%g1 + 0x00], %f8 ! LSU Group |
| 75 | add %o1, 2, %o1 ! IEU0 |
| 76 | add %g1, %o5, %g1 ! IEU1 |
| 77 | fmuls %f0, M0, %f1 ! FGM Group 1-cycle stall on %f0 |
| 78 | fmuls %f0, M1, %f2 ! FGM Group |
| 79 | fmuls %f0, M2, %f3 ! FGM Group |
| 80 | fmuls %f0, M3, %f4 ! FGM Group |
| 81 | fmuls %f8, M0, %f9 ! FGM Group f1 available |
| 82 | fadds %f1, M12, %f1 ! FGA |
| 83 | st %f1, [%g2 + 0x00] ! LSU |
| 84 | fmuls %f8, M1, %f10 ! FGM Group f2 available |
| 85 | fadds %f2, M13, %f2 ! FGA |
| 86 | st %f2, [%g2 + 0x04] ! LSU |
| 87 | fmuls %f8, M2, %f11 ! FGM Group f3 available |
| 88 | fadds %f3, M14, %f3 ! FGA |
| 89 | st %f3, [%g2 + 0x08] ! LSU |
| 90 | fmuls %f8, M3, %f12 ! FGM Group f4 available |
| 91 | fadds %f4, M15, %f4 ! FGA |
| 92 | st %f4, [%g2 + 0x0c] ! LSU |
| 93 | fadds %f9, M12, %f9 ! FGA Group f9 available |
| 94 | st %f9, [%g2 + 0x10] ! LSU |
| 95 | fadds %f10, M13, %f10 ! FGA Group f10 available |
| 96 | st %f10, [%g2 + 0x14] ! LSU |
| 97 | fadds %f11, M14, %f11 ! FGA Group f11 available |
| 98 | st %f11, [%g2 + 0x18] ! LSU |
| 99 | fadds %f12, M15, %f12 ! FGA Group f12 available |
| 100 | st %f12, [%g2 + 0x1c] ! LSU |
| 101 | cmp %o1, %o2 ! IEU1 |
| 102 | bne 1b ! CTI |
| 103 | add %g2, 0x20, %g2 ! IEU0 Group |
| 104 | |
| 105 | cmp %o1, %g3 |
| 106 | be 3f |
| 107 | nop |
| 108 | |
| 109 | 2: ld [%g1 + 0x00], %f0 ! LSU Group |
| 110 | fmuls %f0, M0, %f1 ! FGM Group 1-cycle stall on %f0 |
| 111 | fmuls %f0, M1, %f2 ! FGM Group |
| 112 | fmuls %f0, M2, %f3 ! FGM Group |
| 113 | fmuls %f0, M3, %f4 ! FGM Group |
| 114 | fadds %f1, M12, %f1 ! FGA Group |
| 115 | st %f1, [%g2 + 0x00] ! LSU |
| 116 | fadds %f2, M13, %f2 ! FGA Group |
| 117 | st %f2, [%g2 + 0x04] ! LSU |
| 118 | fadds %f3, M14, %f3 ! FGA Group |
| 119 | st %f3, [%g2 + 0x08] ! LSU |
| 120 | fadds %f4, M15, %f4 ! FGA Group |
| 121 | st %f4, [%g2 + 0x0c] ! LSU |
| 122 | |
| 123 | 3: |
| 124 | ba __set_v4f_4 |
| 125 | nop |
| 126 | |
| 127 | .globl _mesa_sparc_transform_points1_identity |
| 128 | _mesa_sparc_transform_points1_identity: |
| 129 | cmp %o0, %o2 |
| 130 | be 4f |
| 131 | ld [%o2 + V4F_STRIDE], %o5 |
| 132 | LDPTR [%o2 + V4F_START], %g1 |
| 133 | LDPTR [%o0 + V4F_START], %g2 |
| 134 | ld [%o2 + V4F_COUNT], %g3 |
| 135 | |
| 136 | cmp %g3, 1 |
| 137 | st %g3, [%o0 + V4F_COUNT] |
| 138 | bl 3f |
| 139 | clr %o1 |
| 140 | |
| 141 | be 2f |
| 142 | andn %g3, 1, %o2 |
| 143 | |
| 144 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 145 | add %g1, %o5, %g1 ! IEU0 |
| 146 | ld [%g1 + 0x00], %f1 ! LSU Group |
| 147 | add %o1, 2, %o1 ! IEU0 |
| 148 | add %g1, %o5, %g1 ! IEU1 |
| 149 | st %f0, [%g2 + 0x00] ! LSU Group |
| 150 | cmp %o1, %o2 ! IEU1 |
| 151 | st %f1, [%g2 + 0x10] ! LSU Group |
| 152 | bne 1b ! CTI |
| 153 | add %g2, 0x20, %g2 ! IEU0 |
| 154 | |
| 155 | cmp %o1, %g3 |
| 156 | be 3f |
| 157 | nop |
| 158 | |
| 159 | 2: ld [%g1 + 0x00], %f0 |
| 160 | addx %g0, %g0, %g0 |
| 161 | st %f0, [%g2 + 0x00] |
| 162 | |
| 163 | 3: |
| 164 | ba __set_v4f_1 |
| 165 | nop |
| 166 | |
| 167 | 4: retl |
| 168 | nop |
| 169 | |
| 170 | .globl _mesa_sparc_transform_points1_2d |
| 171 | _mesa_sparc_transform_points1_2d: |
| 172 | ld [%o2 + V4F_STRIDE], %o5 |
| 173 | LDPTR [%o2 + V4F_START], %g1 |
| 174 | LDPTR [%o0 + V4F_START], %g2 |
| 175 | ld [%o2 + V4F_COUNT], %g3 |
| 176 | |
| 177 | LDMATRIX_0_1_12_13(%o1) |
| 178 | |
| 179 | cmp %g3, 1 |
| 180 | st %g3, [%o0 + V4F_COUNT] |
| 181 | bl 3f |
| 182 | clr %o1 |
| 183 | |
| 184 | be 2f |
| 185 | andn %g3, 1, %o2 |
| 186 | |
| 187 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 188 | add %g1, %o5, %g1 ! IEU0 |
| 189 | ld [%g1 + 0x00], %f8 ! LSU Group |
| 190 | add %o1, 2, %o1 ! IEU0 |
| 191 | add %g1, %o5, %g1 ! IEU1 |
| 192 | fmuls %f0, M0, %f1 ! FGM Group |
| 193 | fmuls %f0, M1, %f2 ! FGM Group |
| 194 | fmuls %f8, M0, %f9 ! FGM Group |
| 195 | fmuls %f8, M1, %f10 ! FGM Group |
| 196 | fadds %f1, M12, %f3 ! FGA Group f1 available |
| 197 | st %f3, [%g2 + 0x00] ! LSU |
| 198 | fadds %f2, M13, %f4 ! FGA Group f2 available |
| 199 | st %f4, [%g2 + 0x04] ! LSU |
| 200 | fadds %f9, M12, %f11 ! FGA Group f9 available |
| 201 | st %f11, [%g2 + 0x10] ! LSU |
| 202 | fadds %f10, M13, %f12 ! FGA Group f10 available |
| 203 | st %f12, [%g2 + 0x14] ! LSU |
| 204 | cmp %o1, %o2 ! IEU1 |
| 205 | bne 1b ! CTI |
| 206 | add %g2, 0x20, %g2 ! IEU0 Group |
| 207 | |
| 208 | cmp %o1, %g3 |
| 209 | be 3f |
| 210 | nop |
| 211 | |
| 212 | 2: ld [%g1 + 0x00], %f0 |
| 213 | fmuls %f0, M0, %f1 |
| 214 | fmuls %f0, M1, %f2 |
| 215 | fadds %f1, M12, %f3 |
| 216 | st %f3, [%g2 + 0x00] |
| 217 | fadds %f2, M13, %f4 |
| 218 | st %f4, [%g2 + 0x04] |
| 219 | |
| 220 | 3: |
| 221 | ba __set_v4f_2 |
| 222 | nop |
| 223 | |
| 224 | .globl _mesa_sparc_transform_points1_2d_no_rot |
| 225 | _mesa_sparc_transform_points1_2d_no_rot: |
| 226 | ld [%o2 + V4F_STRIDE], %o5 |
| 227 | LDPTR [%o2 + V4F_START], %g1 |
| 228 | LDPTR [%o0 + V4F_START], %g2 |
| 229 | ld [%o2 + V4F_COUNT], %g3 |
| 230 | |
| 231 | LDMATRIX_0_12_13(%o1) |
| 232 | |
| 233 | cmp %g3, 1 |
| 234 | st %g3, [%o0 + V4F_COUNT] |
| 235 | bl 3f |
| 236 | clr %o1 |
| 237 | |
| 238 | be 2f |
| 239 | andn %g3, 1, %o2 |
| 240 | |
| 241 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 242 | add %g1, %o5, %g1 ! IEU0 |
| 243 | ld [%g1 + 0x00], %f4 ! LSU Group |
| 244 | add %o1, 2, %o1 ! IEU0 |
| 245 | add %g1, %o5, %g1 ! IEU1 |
| 246 | fmuls %f0, M0, %f1 ! FGM Group |
| 247 | fmuls %f4, M0, %f5 ! FGM Group |
| 248 | fadds %f1, M12, %f3 ! FGA Group, 2 cycle stall, f1 available |
| 249 | st %f3, [%g2 + 0x00] ! LSU |
| 250 | st M13, [%g2 + 0x04] ! LSU Group, f5 available |
| 251 | fadds %f5, M12, %f6 ! FGA |
| 252 | st %f6, [%g2 + 0x10] ! LSU Group |
| 253 | st M13, [%g2 + 0x14] ! LSU Group |
| 254 | cmp %o1, %o2 ! IEU1 |
| 255 | bne 1b ! CTI |
| 256 | add %g2, 0x20, %g2 ! IEU0 Group |
| 257 | |
| 258 | cmp %o1, %g3 |
| 259 | be 3f |
| 260 | nop |
| 261 | |
| 262 | 2: ld [%g1 + 0x00], %f0 |
| 263 | fmuls %f0, M0, %f1 |
| 264 | fadds %f1, M12, %f3 |
| 265 | st %f3, [%g2 + 0x00] |
| 266 | st M13, [%g2 + 0x04] |
| 267 | |
| 268 | 3: |
| 269 | ba __set_v4f_2 |
| 270 | nop |
| 271 | |
| 272 | .globl _mesa_sparc_transform_points1_3d |
| 273 | _mesa_sparc_transform_points1_3d: |
| 274 | ld [%o2 + V4F_STRIDE], %o5 |
| 275 | LDPTR [%o2 + V4F_START], %g1 |
| 276 | LDPTR [%o0 + V4F_START], %g2 |
| 277 | ld [%o2 + V4F_COUNT], %g3 |
| 278 | |
| 279 | LDMATRIX_0_1_2_12_13_14(%o1) |
| 280 | |
| 281 | cmp %g3, 1 |
| 282 | st %g3, [%o0 + V4F_COUNT] |
| 283 | bl 3f |
| 284 | clr %o1 |
| 285 | |
| 286 | be 2f |
| 287 | andn %g3, 1, %o2 |
| 288 | |
| 289 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 290 | add %g1, %o5, %g1 ! IEU0 |
| 291 | ld [%g1 + 0x00], %f4 ! LSU Group |
| 292 | add %o1, 2, %o1 ! IEU0 |
| 293 | add %g1, %o5, %g1 ! IEU1 |
| 294 | fmuls %f0, M0, %f1 ! FGM Group |
| 295 | fmuls %f0, M1, %f2 ! FGM Group |
| 296 | fmuls %f0, M2, %f3 ! FGM Group |
| 297 | fmuls %f4, M0, %f5 ! FGM Group |
| 298 | fadds %f1, M12, %f1 ! FGA Group, f1 available |
| 299 | st %f1, [%g2 + 0x00] ! LSU |
| 300 | fmuls %f4, M1, %f6 ! FGM |
| 301 | fadds %f2, M13, %f2 ! FGA Group, f2 available |
| 302 | st %f2, [%g2 + 0x04] ! LSU |
| 303 | fmuls %f4, M2, %f7 ! FGM |
| 304 | fadds %f3, M14, %f3 ! FGA Group, f3 available |
| 305 | st %f3, [%g2 + 0x08] ! LSU |
| 306 | fadds %f5, M12, %f5 ! FGA Group, f5 available |
| 307 | st %f5, [%g2 + 0x10] ! LSU |
| 308 | fadds %f6, M13, %f6 ! FGA Group, f6 available |
| 309 | st %f6, [%g2 + 0x14] ! LSU |
| 310 | fadds %f7, M14, %f7 ! FGA Group, f7 available |
| 311 | st %f7, [%g2 + 0x18] ! LSU |
| 312 | cmp %o1, %o2 ! IEU1 |
| 313 | bne 1b ! CTI |
| 314 | add %g2, 0x20, %g2 ! IEU0 Group |
| 315 | |
| 316 | cmp %o1, %g3 |
| 317 | be 3f |
| 318 | nop |
| 319 | |
| 320 | 2: ld [%g1 + 0x00], %f0 |
| 321 | fmuls %f0, M0, %f1 |
| 322 | fmuls %f0, M1, %f2 |
| 323 | fmuls %f0, M2, %f3 |
| 324 | fadds %f1, M12, %f1 |
| 325 | st %f1, [%g2 + 0x00] |
| 326 | fadds %f2, M13, %f2 |
| 327 | st %f2, [%g2 + 0x04] |
| 328 | fadds %f3, M14, %f3 |
| 329 | st %f3, [%g2 + 0x08] |
| 330 | |
| 331 | 3: |
| 332 | ba __set_v4f_3 |
| 333 | nop |
| 334 | |
| 335 | .globl _mesa_sparc_transform_points1_3d_no_rot |
| 336 | _mesa_sparc_transform_points1_3d_no_rot: |
| 337 | ld [%o2 + V4F_STRIDE], %o5 |
| 338 | LDPTR [%o2 + V4F_START], %g1 |
| 339 | LDPTR [%o0 + V4F_START], %g2 |
| 340 | ld [%o2 + V4F_COUNT], %g3 |
| 341 | |
| 342 | LDMATRIX_0_12_13_14(%o1) |
| 343 | |
| 344 | cmp %g3, 1 |
| 345 | st %g3, [%o0 + V4F_COUNT] |
| 346 | bl 3f |
| 347 | clr %o1 |
| 348 | |
| 349 | be 2f |
| 350 | andn %g3, 1, %o2 |
| 351 | |
| 352 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 353 | add %g1, %o5, %g1 ! IEU0 |
| 354 | ld [%g1 + 0x00], %f2 ! LSU Group |
| 355 | add %o1, 2, %o1 ! IEU0 |
| 356 | add %g1, %o5, %g1 ! IEU1 |
| 357 | fmuls %f0, M0, %f1 ! FGM Group |
| 358 | fmuls %f2, M0, %f3 ! FGM Group |
| 359 | fadds %f1, M12, %f1 ! FGA Group, 2 cycle stall, f1 available |
| 360 | st %f1, [%g2 + 0x00] ! LSU |
| 361 | fadds %f3, M12, %f3 ! FGA Group, f3 available |
| 362 | st M13, [%g2 + 0x04] ! LSU |
| 363 | st M14, [%g2 + 0x08] ! LSU Group |
| 364 | st %f3, [%g2 + 0x10] ! LSU Group |
| 365 | st M13, [%g2 + 0x14] ! LSU Group |
| 366 | st M14, [%g2 + 0x18] ! LSU Group |
| 367 | cmp %o1, %o2 ! IEU1 |
| 368 | bne 1b ! CTI |
| 369 | add %g2, 0x20, %g2 ! IEU0 Group |
| 370 | |
| 371 | cmp %o1, %g3 |
| 372 | be 3f |
| 373 | nop |
| 374 | |
| 375 | 2: ld [%g1 + 0x00], %f0 |
| 376 | fmuls %f0, M0, %f1 |
| 377 | fadds %f1, M12, %f1 |
| 378 | st %f1, [%g2 + 0x00] |
| 379 | st M13, [%g2 + 0x04] |
| 380 | st M14, [%g2 + 0x08] |
| 381 | |
| 382 | 3: |
| 383 | ba __set_v4f_3 |
| 384 | nop |
| 385 | |
| 386 | .globl _mesa_sparc_transform_points1_perspective |
| 387 | _mesa_sparc_transform_points1_perspective: |
| 388 | ld [%o2 + V4F_STRIDE], %o5 |
| 389 | LDPTR [%o2 + V4F_START], %g1 |
| 390 | LDPTR [%o0 + V4F_START], %g2 |
| 391 | ld [%o2 + V4F_COUNT], %g3 |
| 392 | |
| 393 | LDMATRIX_0_14(%o1) |
| 394 | |
| 395 | cmp %g3, 1 |
| 396 | st %g3, [%o0 + V4F_COUNT] |
| 397 | bl 3f |
| 398 | clr %o1 |
| 399 | |
| 400 | be 2f |
| 401 | andn %g3, 1, %o2 |
| 402 | |
| 403 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 404 | add %g1, %o5, %g1 ! IEU0 |
| 405 | ld [%g1 + 0x00], %f2 ! LSU Group |
| 406 | add %o1, 2, %o1 ! IEU0 |
| 407 | add %g1, %o5, %g1 ! IEU1 |
| 408 | fmuls %f0, M0, %f1 ! FGM Group |
| 409 | st %f1, [%g2 + 0x00] ! LSU |
| 410 | fmuls %f2, M0, %f3 ! FGM Group |
| 411 | st %g0, [%g2 + 0x04] ! LSU |
| 412 | st M14, [%g2 + 0x08] ! LSU Group |
| 413 | st %g0, [%g2 + 0x0c] ! LSU Group |
| 414 | st %f3, [%g2 + 0x10] ! LSU Group |
| 415 | st %g0, [%g2 + 0x14] ! LSU Group |
| 416 | st M14, [%g2 + 0x18] ! LSU Group |
| 417 | st %g0, [%g2 + 0x1c] ! LSU Group |
| 418 | cmp %o1, %o2 ! IEU1 |
| 419 | bne 1b ! CTI |
| 420 | add %g2, 0x20, %g2 ! IEU0 Group |
| 421 | |
| 422 | cmp %o1, %g3 |
| 423 | be 3f |
| 424 | nop |
| 425 | |
| 426 | 2: ld [%g1 + 0x00], %f0 |
| 427 | fmuls %f0, M0, %f1 |
| 428 | st %f1, [%g2 + 0x00] |
| 429 | st %g0, [%g2 + 0x04] |
| 430 | st M14, [%g2 + 0x08] |
| 431 | st %g0, [%g2 + 0x0c] |
| 432 | |
| 433 | 3: |
| 434 | ba __set_v4f_4 |
| 435 | nop |
| 436 | |
| 437 | .globl _mesa_sparc_transform_points2_general |
| 438 | _mesa_sparc_transform_points2_general: |
| 439 | ld [%o2 + V4F_STRIDE], %o5 |
| 440 | LDPTR [%o2 + V4F_START], %g1 |
| 441 | LDPTR [%o0 + V4F_START], %g2 |
| 442 | ld [%o2 + V4F_COUNT], %g3 |
| 443 | |
| 444 | LDMATRIX_0_1_2_3_4_5_6_7_12_13_14_15(%o1) |
| 445 | |
| 446 | cmp %g3, 0 |
| 447 | st %g3, [%o0 + V4F_COUNT] |
| 448 | be 2f |
| 449 | clr %o1 |
| 450 | |
| 451 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 452 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 453 | add %o1, 1, %o1 ! IEU0 |
| 454 | add %g1, %o5, %g1 ! IEU1 |
| 455 | fmuls %f0, M0, %f2 ! FGM Group |
| 456 | fmuls %f0, M1, %f3 ! FGM Group |
| 457 | fmuls %f0, M2, %f4 ! FGM Group |
| 458 | fmuls %f0, M3, %f5 ! FGM Group |
| 459 | fadds %f2, M12, %f2 ! FGA Group f2 available |
| 460 | fmuls %f1, M4, %f6 ! FGM |
| 461 | fadds %f3, M13, %f3 ! FGA Group f3 available |
| 462 | fmuls %f1, M5, %f7 ! FGM |
| 463 | fadds %f4, M14, %f4 ! FGA Group f4 available |
| 464 | fmuls %f1, M6, %f8 ! FGM |
| 465 | fadds %f5, M15, %f5 ! FGA Group f5 available |
| 466 | fmuls %f1, M7, %f9 ! FGM |
| 467 | fadds %f2, %f6, %f2 ! FGA Group f6 available |
| 468 | st %f2, [%g2 + 0x00] ! LSU |
| 469 | fadds %f3, %f7, %f3 ! FGA Group f7 available |
| 470 | st %f3, [%g2 + 0x04] ! LSU |
| 471 | fadds %f4, %f8, %f4 ! FGA Group f8 available |
| 472 | st %f4, [%g2 + 0x08] ! LSU |
| 473 | fadds %f5, %f9, %f5 ! FGA Group f9 available |
| 474 | st %f5, [%g2 + 0x0c] ! LSU |
| 475 | cmp %o1, %g3 ! IEU1 |
| 476 | bne 1b ! CTI |
| 477 | add %g2, 0x10, %g2 ! IEU0 Group |
| 478 | 2: |
| 479 | ba __set_v4f_4 |
| 480 | nop |
| 481 | |
| 482 | .globl _mesa_sparc_transform_points2_identity |
| 483 | _mesa_sparc_transform_points2_identity: |
| 484 | cmp %o2, %o0 |
| 485 | be 3f |
| 486 | ld [%o2 + V4F_STRIDE], %o5 |
| 487 | LDPTR [%o2 + V4F_START], %g1 |
| 488 | LDPTR [%o0 + V4F_START], %g2 |
| 489 | ld [%o2 + V4F_COUNT], %g3 |
| 490 | |
| 491 | cmp %g3, 0 |
| 492 | st %g3, [%o0 + V4F_COUNT] |
| 493 | be 2f |
| 494 | clr %o1 |
| 495 | |
| 496 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 497 | add %o1, 1, %o1 ! IEU0 |
| 498 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 499 | add %g1, %o5, %g1 ! IEU0 |
| 500 | cmp %o1, %g3 ! IEU1 |
| 501 | st %f0, [%g2 + 0x00] ! LSU Group |
| 502 | st %f1, [%g2 + 0x04] ! LSU Group |
| 503 | bne 1b ! CTI |
| 504 | add %g2, 0x10, %g2 ! IEU0 |
| 505 | 2: |
| 506 | ba __set_v4f_2 |
| 507 | nop |
| 508 | |
| 509 | 3: retl |
| 510 | nop |
| 511 | |
| 512 | .globl _mesa_sparc_transform_points2_2d |
| 513 | _mesa_sparc_transform_points2_2d: |
| 514 | ld [%o2 + V4F_STRIDE], %o5 |
| 515 | LDPTR [%o2 + V4F_START], %g1 |
| 516 | LDPTR [%o0 + V4F_START], %g2 |
| 517 | ld [%o2 + V4F_COUNT], %g3 |
| 518 | |
| 519 | LDMATRIX_0_1_4_5_12_13(%o1) |
| 520 | |
| 521 | cmp %g3, 1 |
| 522 | st %g3, [%o0 + V4F_COUNT] |
| 523 | bl 3f |
| 524 | clr %o1 |
| 525 | |
| 526 | be 2f |
| 527 | andn %g3, 1, %o2 |
| 528 | |
| 529 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 530 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 531 | add %o1, 2, %o1 ! IEU0 |
| 532 | add %g1, %o5, %g1 ! IEU1 |
| 533 | fmuls %f0, M0, %f2 ! FGM |
| 534 | ld [%g1 + 0x00], %f8 ! LSU Group |
| 535 | fmuls %f0, M1, %f3 ! FGM |
| 536 | ld [%g1 + 0x04], %f9 ! LSU Group |
| 537 | fmuls %f1, M4, %f6 ! FGM |
| 538 | fmuls %f1, M5, %f7 ! FGM Group |
| 539 | add %g1, %o5, %g1 ! IEU0 |
| 540 | fmuls %f8, M0, %f10 ! FGM Group f2 available |
| 541 | fadds %f2, M12, %f2 ! FGA |
| 542 | fmuls %f8, M1, %f11 ! FGM Group f3 available |
| 543 | fadds %f3, M13, %f3 ! FGA |
| 544 | fmuls %f9, M4, %f12 ! FGM Group |
| 545 | fmuls %f9, M5, %f13 ! FGM Group |
| 546 | fadds %f10, M12, %f10 ! FGA Group f2, f10 available |
| 547 | fadds %f2, %f6, %f2 ! FGA Group f3, f11 available |
| 548 | st %f2, [%g2 + 0x00] ! LSU |
| 549 | fadds %f11, M13, %f11 ! FGA Group f12 available |
| 550 | fadds %f3, %f7, %f3 ! FGA Group f13 available |
| 551 | st %f3, [%g2 + 0x04] ! LSU |
| 552 | fadds %f10, %f12, %f10 ! FGA Group f10 available |
| 553 | st %f10, [%g2 + 0x10] ! LSU |
| 554 | fadds %f11, %f13, %f11 ! FGA Group f11 available |
| 555 | st %f11, [%g2 + 0x14] ! LSU |
| 556 | cmp %o1, %o2 ! IEU1 |
| 557 | bne 1b ! CTI |
| 558 | add %g2, 0x20, %g2 ! IEU0 Group |
| 559 | |
| 560 | cmp %o1, %g3 |
| 561 | be 3f |
| 562 | nop |
| 563 | |
| 564 | 2: ld [%g1 + 0x00], %f0 ! LSU Group |
| 565 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 566 | fmuls %f0, M0, %f2 ! FGM Group |
| 567 | fmuls %f0, M1, %f3 ! FGM Group |
| 568 | fmuls %f1, M4, %f6 ! FGM Group |
| 569 | fmuls %f1, M5, %f7 ! FGM Group |
| 570 | fadds %f2, M12, %f2 ! FGA Group f2 available |
| 571 | fadds %f3, M13, %f3 ! FGA Group f3 available |
| 572 | fadds %f2, %f6, %f2 ! FGA Group 2 cycle stall, f2 available |
| 573 | st %f2, [%g2 + 0x00] ! LSU |
| 574 | fadds %f3, %f7, %f3 ! FGA Group f3 available |
| 575 | st %f3, [%g2 + 0x04] ! LSU |
| 576 | |
| 577 | 3: |
| 578 | ba __set_v4f_2 |
| 579 | nop |
| 580 | |
| 581 | .globl _mesa_sparc_transform_points2_2d_no_rot |
| 582 | _mesa_sparc_transform_points2_2d_no_rot: |
| 583 | ld [%o2 + V4F_STRIDE], %o5 |
| 584 | LDPTR [%o2 + V4F_START], %g1 |
| 585 | LDPTR [%o0 + V4F_START], %g2 |
| 586 | ld [%o2 + V4F_COUNT], %g3 |
| 587 | |
| 588 | LDMATRIX_0_5_12_13(%o1) |
| 589 | |
| 590 | cmp %g3, 1 |
| 591 | st %g3, [%o0 + V4F_COUNT] |
| 592 | bl 3f |
| 593 | clr %o1 |
| 594 | |
| 595 | be 2f |
| 596 | andn %g3, 1, %o2 |
| 597 | |
| 598 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 599 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 600 | add %o1, 2, %o1 ! IEU0 |
| 601 | add %g1, %o5, %g1 ! IEU1 |
| 602 | ld [%g1 + 0x00], %f4 ! LSU Group |
| 603 | fmuls %f0, M0, %f2 ! FGM |
| 604 | ld [%g1 + 0x04], %f5 ! LSU Group |
| 605 | fmuls %f1, M5, %f3 ! FGM |
| 606 | fmuls %f4, M0, %f6 ! FGM Group |
| 607 | add %g1, %o5, %g1 ! IEU0 |
| 608 | fmuls %f5, M5, %f7 ! FGM Group |
| 609 | fadds %f2, M12, %f2 ! FGA Group f2 available |
| 610 | st %f2, [%g2 + 0x00] ! LSU |
| 611 | fadds %f3, M13, %f3 ! FGA Group f3 available |
| 612 | st %f3, [%g2 + 0x04] ! LSU |
| 613 | fadds %f6, M12, %f6 ! FGA Group f6 available |
| 614 | st %f6, [%g2 + 0x10] ! LSU |
| 615 | fadds %f7, M13, %f7 ! FGA Group f7 available |
| 616 | st %f7, [%g2 + 0x14] ! LSU |
| 617 | cmp %o1, %o2 ! IEU1 |
| 618 | bne 1b ! CTI |
| 619 | add %g2, 0x20, %g2 ! IEU0 Group |
| 620 | |
| 621 | cmp %o1, %g3 |
| 622 | be 3f |
| 623 | nop |
| 624 | |
| 625 | 2: ld [%g1 + 0x00], %f0 ! LSU Group |
| 626 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 627 | fmuls %f0, M0, %f2 ! FGM Group |
| 628 | fmuls %f1, M5, %f3 ! FGM Group |
| 629 | fadds %f2, M12, %f2 ! FGA Group, 2 cycle stall, f2 available |
| 630 | st %f2, [%g2 + 0x00] ! LSU |
| 631 | fadds %f3, M13, %f3 ! FGA Group f3 available |
| 632 | st %f3, [%g2 + 0x04] ! LSU |
| 633 | |
| 634 | 3: |
| 635 | ba __set_v4f_2 |
| 636 | nop |
| 637 | |
| 638 | /* orig: 12 cycles */ |
| 639 | .globl _mesa_sparc_transform_points2_3d |
| 640 | _mesa_sparc_transform_points2_3d: |
| 641 | ld [%o2 + V4F_STRIDE], %o5 |
| 642 | ld [%o2 + V4F_START], %g1 |
| 643 | ld [%o0 + V4F_START], %g2 |
| 644 | ld [%o2 + V4F_COUNT], %g3 |
| 645 | |
| 646 | LDMATRIX_0_1_2_3_4_5_6_12_13_14(%o1) |
| 647 | |
| 648 | cmp %g3, 1 |
| 649 | st %g3, [%o0 + V4F_COUNT] |
| 650 | bl 3f |
| 651 | clr %o1 |
| 652 | |
| 653 | be 2f |
| 654 | andn %g3, 1, %o2 |
| 655 | |
| 656 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 657 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 658 | add %o1, 2, %o1 ! IEU0 |
| 659 | add %g1, %o5, %g1 ! IEU1 |
| 660 | ld [%g1 + 0x00], %f9 ! LSU Group |
| 661 | fmuls %f0, M0, %f2 ! FGM |
| 662 | ld [%g1 + 0x04], %f10 ! LSU Group |
| 663 | fmuls %f0, M1, %f3 ! FGM |
| 664 | fmuls %f0, M2, %f4 ! FGM Group |
| 665 | add %g1, %o5, %g1 ! IEU0 |
| 666 | fmuls %f1, M4, %f6 ! FGM Group |
| 667 | fmuls %f1, M5, %f7 ! FGM Group f2 available |
| 668 | fadds %f2, M12, %f2 ! FGA |
| 669 | fmuls %f1, M6, %f8 ! FGM Group f3 available |
| 670 | fadds %f3, M13, %f3 ! FGA |
| 671 | fmuls %f9, M0, %f11 ! FGM Group f4 available |
| 672 | fadds %f4, M14, %f4 ! FGA |
| 673 | fmuls %f9, M1, %f12 ! FGM Group f6 available |
| 674 | fmuls %f9, M2, %f13 ! FGM Group f2, f7 available |
| 675 | fadds %f2, %f6, %f2 ! FGA |
| 676 | st %f2, [%g2 + 0x00] ! LSU |
| 677 | fmuls %f10, M4, %f14 ! FGM Group f3, f8 available |
| 678 | fadds %f3, %f7, %f3 ! FGA |
| 679 | st %f3, [%g2 + 0x04] ! LSU |
| 680 | fmuls %f10, M5, %f15 ! FGM Group f4, f11 available |
| 681 | fadds %f11, M12, %f11 ! FGA |
| 682 | fmuls %f10, M6, %f0 ! FGM Group f12 available |
| 683 | fadds %f12, M13, %f12 ! FGA |
| 684 | fadds %f13, M14, %f13 ! FGA Group f13 available |
| 685 | fadds %f4, %f8, %f4 ! FGA Group f14 available |
| 686 | st %f4, [%g2 + 0x08] ! LSU |
| 687 | fadds %f11, %f14, %f11 ! FGA Group f15, f11 available |
| 688 | st %f11, [%g2 + 0x10] ! LSU |
| 689 | fadds %f12, %f15, %f12 ! FGA Group f0, f12 available |
| 690 | st %f12, [%g2 + 0x14] ! LSU |
| 691 | fadds %f13, %f0, %f13 ! FGA Group f13 available |
| 692 | st %f13, [%g2 + 0x18] ! LSU |
| 693 | |
| 694 | cmp %o1, %o2 ! IEU1 |
| 695 | bne 1b ! CTI |
| 696 | add %g2, 0x20, %g2 ! IEU0 Group |
| 697 | |
| 698 | cmp %o1, %g3 |
| 699 | be 3f |
| 700 | nop |
| 701 | |
| 702 | 2: ld [%g1 + 0x00], %f0 ! LSU Group |
| 703 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 704 | fmuls %f0, M0, %f2 ! FGM Group |
| 705 | fmuls %f0, M1, %f3 ! FGM Group |
| 706 | fmuls %f0, M2, %f4 ! FGM Group |
| 707 | fmuls %f1, M4, %f6 ! FGM Group |
| 708 | fmuls %f1, M5, %f7 ! FGM Group f2 available |
| 709 | fadds %f2, M12, %f2 ! FGA |
| 710 | fmuls %f1, M6, %f8 ! FGM Group f3 available |
| 711 | fadds %f3, M13, %f3 ! FGA |
| 712 | fadds %f4, M14, %f4 ! FGA Group f4 available |
| 713 | fadds %f2, %f6, %f2 ! FGA Group stall, f2, f6, f7 available |
| 714 | st %f2, [%g2 + 0x00] ! LSU |
| 715 | fadds %f3, %f7, %f3 ! FGA Group f3, f8 available |
| 716 | st %f3, [%g2 + 0x04] ! LSU |
| 717 | fadds %f4, %f8, %f4 ! FGA Group f4 available |
| 718 | st %f4, [%g2 + 0x08] ! LSU |
| 719 | |
| 720 | 3: |
| 721 | ba __set_v4f_3 |
| 722 | nop |
| 723 | |
| 724 | .globl _mesa_sparc_transform_points2_3d_no_rot |
| 725 | _mesa_sparc_transform_points2_3d_no_rot: |
| 726 | ld [%o2 + V4F_STRIDE], %o5 |
| 727 | LDPTR [%o2 + V4F_START], %g1 |
| 728 | LDPTR [%o0 + V4F_START], %g2 |
| 729 | ld [%o2 + V4F_COUNT], %g3 |
| 730 | |
| 731 | LDMATRIX_0_5_12_13_14(%o1) |
| 732 | |
| 733 | cmp %g3, 1 |
| 734 | st %g3, [%o0 + V4F_COUNT] |
| 735 | bl 3f |
| 736 | clr %o3 |
| 737 | |
| 738 | be 2f |
| 739 | andn %g3, 1, %o2 |
| 740 | |
| 741 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 742 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 743 | add %o3, 2, %o3 ! IEU0 |
| 744 | add %g1, %o5, %g1 ! IEU1 |
| 745 | ld [%g1 + 0x00], %f4 ! LSU Group |
| 746 | fmuls %f0, M0, %f2 ! FGM |
| 747 | ld [%g1 + 0x04], %f5 ! LSU Group |
| 748 | fmuls %f1, M5, %f3 ! FGM |
| 749 | fmuls %f4, M0, %f6 ! FGM Group |
| 750 | add %g1, %o5, %g1 ! IEU0 |
| 751 | fmuls %f5, M5, %f7 ! FGM Group |
| 752 | fadds %f2, M12, %f2 ! FGA Group f2 available |
| 753 | st %f2, [%g2 + 0x00] ! LSU |
| 754 | fadds %f3, M13, %f3 ! FGA Group f3 available |
| 755 | st %f3, [%g2 + 0x04] ! LSU |
| 756 | fadds %f6, M12, %f6 ! FGA Group f6 available |
| 757 | st M14, [%g2 + 0x08] ! LSU |
| 758 | fadds %f7, M13, %f7 ! FGA Group f7 available |
| 759 | st %f6, [%g2 + 0x10] ! LSU |
| 760 | st %f7, [%g2 + 0x14] ! LSU Group |
| 761 | st M14, [%g2 + 0x18] ! LSU Group |
| 762 | cmp %o3, %o2 ! IEU1 |
| 763 | bne 1b ! CTI |
| 764 | add %g2, 0x20, %g2 ! IEU0 Group |
| 765 | |
| 766 | cmp %o3, %g3 |
| 767 | be 3f |
| 768 | nop |
| 769 | |
| 770 | 2: ld [%g1 + 0x00], %f0 ! LSU Group |
| 771 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 772 | fmuls %f0, M0, %f2 ! FGM Group |
| 773 | fmuls %f1, M5, %f3 ! FGM Group |
| 774 | fadds %f2, M12, %f2 ! FGA Group, 2 cycle stall, f2 available |
| 775 | st %f2, [%g2 + 0x00] ! LSU |
| 776 | fadds %f3, M13, %f3 ! FGA Group f3 available |
| 777 | st %f3, [%g2 + 0x04] ! LSU |
| 778 | st M14, [%g2 + 0x08] ! LSU Group |
| 779 | |
| 780 | 3: ld [%o1 + (14 * 0x4)], %g3 |
| 781 | cmp %g3, 0 |
| 782 | bne __set_v4f_3 |
| 783 | nop |
| 784 | ba __set_v4f_2 |
| 785 | nop |
| 786 | |
| 787 | .globl _mesa_sparc_transform_points2_perspective |
| 788 | _mesa_sparc_transform_points2_perspective: |
| 789 | ld [%o2 + V4F_STRIDE], %o5 |
| 790 | LDPTR [%o2 + V4F_START], %g1 |
| 791 | LDPTR [%o0 + V4F_START], %g2 |
| 792 | ld [%o2 + V4F_COUNT], %g3 |
| 793 | |
| 794 | LDMATRIX_0_5_14(%o1) |
| 795 | |
| 796 | cmp %g3, 0 |
| 797 | st %g3, [%o0 + V4F_COUNT] |
| 798 | be 2f |
| 799 | clr %o1 |
| 800 | |
| 801 | 1: ld [%g1 + 0x00], %f0 |
| 802 | ld [%g1 + 0x04], %f1 |
| 803 | add %o1, 1, %o1 |
| 804 | add %g1, %o5, %g1 |
| 805 | fmuls %f0, M0, %f2 |
| 806 | st %f2, [%g2 + 0x00] |
| 807 | fmuls %f1, M5, %f3 |
| 808 | st %f3, [%g2 + 0x04] |
| 809 | st M14, [%g2 + 0x08] |
| 810 | st %g0, [%g2 + 0x0c] |
| 811 | cmp %o1, %g3 |
| 812 | bne 1b |
| 813 | add %g2, 0x10, %g2 |
| 814 | 2: |
| 815 | ba __set_v4f_4 |
| 816 | nop |
| 817 | |
| 818 | .globl _mesa_sparc_transform_points3_general |
| 819 | _mesa_sparc_transform_points3_general: |
| 820 | ld [%o2 + V4F_STRIDE], %o5 |
| 821 | LDPTR [%o2 + V4F_START], %g1 |
| 822 | LDPTR [%o0 + V4F_START], %g2 |
| 823 | ld [%o2 + V4F_COUNT], %g3 |
| 824 | |
| 825 | LDMATRIX_0_1_2_3_4_5_6_7_8_9_10_11_12_13_14_15(%o1) |
| 826 | |
| 827 | cmp %g3, 0 |
| 828 | st %g3, [%o0 + V4F_COUNT] |
| 829 | be 2f |
| 830 | clr %o1 |
| 831 | |
| 832 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 833 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 834 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 835 | add %o1, 1, %o1 ! IEU0 |
| 836 | add %g1, %o5, %g1 ! IEU1 |
| 837 | fmuls %f0, M0, %f3 ! FGM |
| 838 | fmuls %f1, M4, %f7 ! FGM Group |
| 839 | fmuls %f0, M1, %f4 ! FGM Group |
| 840 | fmuls %f1, M5, %f8 ! FGM Group |
| 841 | fmuls %f0, M2, %f5 ! FGM Group f3 available |
| 842 | fmuls %f1, M6, %f9 ! FGM Group f7 available |
| 843 | fadds %f3, %f7, %f3 ! FGA |
| 844 | fmuls %f0, M3, %f6 ! FGM Group f4 available |
| 845 | fmuls %f1, M7, %f10 ! FGM Group f8 available |
| 846 | fadds %f4, %f8, %f4 ! FGA |
| 847 | fmuls %f2, M8, %f7 ! FGM Group f5 available |
| 848 | fmuls %f2, M9, %f8 ! FGM Group f9,f3 available |
| 849 | fadds %f5, %f9, %f5 ! FGA |
| 850 | fmuls %f2, M10, %f9 ! FGM Group f6 available |
| 851 | fadds %f6, %f10, %f6 ! FGA Group f10,f4 available |
| 852 | fmuls %f2, M11, %f10 ! FGM |
| 853 | fadds %f3, M12, %f3 ! FGA Group f7 available |
| 854 | fadds %f4, M13, %f4 ! FGA Group f8,f5 available |
| 855 | fadds %f5, M14, %f5 ! FGA Group f9 available |
| 856 | fadds %f6, M15, %f6 ! FGA Group f10,f6 available |
| 857 | fadds %f3, %f7, %f3 ! FGA Group f3 available |
| 858 | st %f3, [%g2 + 0x00] ! LSU |
| 859 | fadds %f4, %f8, %f4 ! FGA Group f4 available |
| 860 | st %f4, [%g2 + 0x04] ! LSU |
| 861 | fadds %f5, %f9, %f5 ! FGA Group f5 available |
| 862 | st %f5, [%g2 + 0x08] ! LSU |
| 863 | fadds %f6, %f10, %f6 ! FGA Group f6 available |
| 864 | st %f6, [%g2 + 0x0c] ! LSU |
| 865 | cmp %o1, %g3 ! IEU1 |
| 866 | bne 1b ! CTI |
| 867 | add %g2, 0x10, %g2 ! IEU0 Group |
| 868 | 2: |
| 869 | ba __set_v4f_4 |
| 870 | nop |
| 871 | |
| 872 | .globl _mesa_sparc_transform_points3_identity |
| 873 | _mesa_sparc_transform_points3_identity: |
| 874 | ld [%o2 + V4F_STRIDE], %o5 |
| 875 | LDPTR [%o2 + V4F_START], %g1 |
| 876 | LDPTR [%o0 + V4F_START], %g2 |
| 877 | ld [%o2 + V4F_COUNT], %g3 |
| 878 | |
| 879 | cmp %g3, 0 |
| 880 | st %g3, [%o0 + V4F_COUNT] |
| 881 | be 2f |
| 882 | clr %o1 |
| 883 | |
| 884 | 1: ld [%g1 + 0x00], %f0 |
| 885 | ld [%g1 + 0x04], %f1 |
| 886 | ld [%g1 + 0x08], %f2 |
| 887 | add %o1, 1, %o1 |
| 888 | add %g1, %o5, %g1 |
| 889 | cmp %o1, %g3 |
| 890 | st %f0, [%g2 + 0x00] |
| 891 | st %f1, [%g2 + 0x04] |
| 892 | st %f2, [%g2 + 0x08] |
| 893 | bne 1b |
| 894 | add %g2, 0x10, %g2 |
| 895 | 2: |
| 896 | ba __set_v4f_3 |
| 897 | nop |
| 898 | |
| 899 | .globl _mesa_sparc_transform_points3_2d |
| 900 | _mesa_sparc_transform_points3_2d: |
| 901 | ld [%o2 + V4F_STRIDE], %o5 |
| 902 | LDPTR [%o2 + V4F_START], %g1 |
| 903 | LDPTR [%o0 + V4F_START], %g2 |
| 904 | ld [%o2 + V4F_COUNT], %g3 |
| 905 | |
| 906 | LDMATRIX_0_1_4_5_12_13(%o1) |
| 907 | |
| 908 | cmp %g3, 0 |
| 909 | st %g3, [%o0 + V4F_COUNT] |
| 910 | be 2f |
| 911 | clr %o1 |
| 912 | |
| 913 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 914 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 915 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 916 | add %o1, 1, %o1 ! IEU0 |
| 917 | add %g1, %o5, %g1 ! IEU1 |
| 918 | fmuls %f0, M0, %f3 ! FGM |
| 919 | fmuls %f0, M1, %f4 ! FGM Group |
| 920 | fmuls %f1, M4, %f6 ! FGM Group |
| 921 | fmuls %f1, M5, %f7 ! FGM Group |
| 922 | fadds %f3, M12, %f3 ! FGA Group f3 available |
| 923 | fadds %f4, M13, %f4 ! FGA Group f4 available |
| 924 | fadds %f3, %f6, %f3 ! FGA Group f6 available |
| 925 | st %f3, [%g2 + 0x00] ! LSU |
| 926 | fadds %f4, %f7, %f4 ! FGA Group f7 available |
| 927 | st %f4, [%g2 + 0x04] ! LSU |
| 928 | st %f2, [%g2 + 0x08] ! LSU Group |
| 929 | cmp %o1, %g3 ! IEU1 |
| 930 | bne 1b ! CTI |
| 931 | add %g2, 0x10, %g2 ! IEU0 Group |
| 932 | 2: |
| 933 | ba __set_v4f_3 |
| 934 | nop |
| 935 | |
| 936 | .globl _mesa_sparc_transform_points3_2d_no_rot |
| 937 | _mesa_sparc_transform_points3_2d_no_rot: |
| 938 | ld [%o2 + V4F_STRIDE], %o5 |
| 939 | LDPTR [%o2 + V4F_START], %g1 |
| 940 | LDPTR [%o0 + V4F_START], %g2 |
| 941 | ld [%o2 + V4F_COUNT], %g3 |
| 942 | |
| 943 | LDMATRIX_0_5_12_13(%o1) |
| 944 | |
| 945 | cmp %g3, 0 |
| 946 | st %g3, [%o0 + V4F_COUNT] |
| 947 | be 2f |
| 948 | clr %o1 |
| 949 | |
| 950 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 951 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 952 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 953 | add %o1, 1, %o1 ! IEU0 |
| 954 | add %g1, %o5, %g1 ! IEU1 |
| 955 | fmuls %f0, M0, %f3 ! FGM |
| 956 | fmuls %f1, M5, %f4 ! FGM Group |
| 957 | st %f2, [%g2 + 0x08] ! LSU |
| 958 | fadds %f3, M12, %f3 ! FGA Group |
| 959 | st %f3, [%g2 + 0x00] ! LSU |
| 960 | fadds %f4, M13, %f4 ! FGA Group |
| 961 | st %f4, [%g2 + 0x04] ! LSU |
| 962 | cmp %o1, %g3 ! IEU1 |
| 963 | bne 1b ! CTI |
| 964 | add %g2, 0x10, %g2 ! IEU0 Group |
| 965 | 2: |
| 966 | ba __set_v4f_3 |
| 967 | nop |
| 968 | |
| 969 | .globl _mesa_sparc_transform_points3_3d |
| 970 | _mesa_sparc_transform_points3_3d: |
| 971 | ld [%o2 + V4F_STRIDE], %o5 |
| 972 | LDPTR [%o2 + V4F_START], %g1 |
| 973 | LDPTR [%o0 + V4F_START], %g2 |
| 974 | ld [%o2 + V4F_COUNT], %g3 |
| 975 | |
| 976 | LDMATRIX_0_1_2_4_5_6_8_9_10_12_13_14(%o1) |
| 977 | |
| 978 | cmp %g3, 0 |
| 979 | st %g3, [%o0 + V4F_COUNT] |
| 980 | be 2f |
| 981 | clr %o1 |
| 982 | |
| 983 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 984 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 985 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 986 | add %o1, 1, %o1 ! IEU0 |
| 987 | add %g1, %o5, %g1 ! IEU1 |
| 988 | fmuls %f0, M0, %f3 ! FGM |
| 989 | fmuls %f1, M4, %f6 ! FGM Group |
| 990 | fmuls %f0, M1, %f4 ! FGM Group |
| 991 | fmuls %f1, M5, %f7 ! FGM Group |
| 992 | fmuls %f0, M2, %f5 ! FGM Group f3 available |
| 993 | fmuls %f1, M6, %f8 ! FGM Group f6 available |
| 994 | fadds %f3, %f6, %f3 ! FGA |
| 995 | fmuls %f2, M8, %f9 ! FGM Group f4 available |
| 996 | fmuls %f2, M9, %f10 ! FGM Group f7 available |
| 997 | fadds %f4, %f7, %f4 ! FGA |
| 998 | fmuls %f2, M10, %f11 ! FGM Group f5 available |
| 999 | fadds %f5, %f8, %f5 ! FGA Group f8, f3 available |
| 1000 | fadds %f3, %f9, %f3 ! FGA Group f9 available |
| 1001 | fadds %f4, %f10, %f4 ! FGA Group f10, f4 available |
| 1002 | fadds %f5, %f11, %f5 ! FGA Group stall, f11, f5 available |
| 1003 | fadds %f3, M12, %f3 ! FGA Group f3 available |
| 1004 | st %f3, [%g2 + 0x00] ! LSU |
| 1005 | fadds %f4, M13, %f4 ! FGA Group f4 available |
| 1006 | st %f4, [%g2 + 0x04] ! LSU |
| 1007 | fadds %f5, M14, %f5 ! FGA Group f5 available |
| 1008 | st %f5, [%g2 + 0x08] ! LSU |
| 1009 | cmp %o1, %g3 ! IEU1 |
| 1010 | bne 1b ! CTI |
| 1011 | add %g2, 0x10, %g2 ! IEU0 Group |
| 1012 | 2: |
| 1013 | ba __set_v4f_3 |
| 1014 | nop |
| 1015 | |
| 1016 | .globl _mesa_sparc_transform_points3_3d_no_rot |
| 1017 | _mesa_sparc_transform_points3_3d_no_rot: |
| 1018 | ld [%o2 + V4F_STRIDE], %o5 |
| 1019 | LDPTR [%o2 + V4F_START], %g1 |
| 1020 | LDPTR [%o0 + V4F_START], %g2 |
| 1021 | ld [%o2 + V4F_COUNT], %g3 |
| 1022 | |
| 1023 | LDMATRIX_0_5_10_12_13_14(%o1) |
| 1024 | |
| 1025 | cmp %g3, 0 |
| 1026 | st %g3, [%o0 + V4F_COUNT] |
| 1027 | be 2f |
| 1028 | clr %o1 |
| 1029 | |
| 1030 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 1031 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 1032 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 1033 | add %o1, 1, %o1 ! IEU0 |
| 1034 | add %g1, %o5, %g1 ! IEU1 |
| 1035 | cmp %o1, %g3 ! IEU1 Group |
| 1036 | fmuls %f0, M0, %f3 ! FGM |
| 1037 | fmuls %f1, M5, %f4 ! FGM Group |
| 1038 | fmuls %f2, M10, %f5 ! FGM Group |
| 1039 | fadds %f3, M12, %f3 ! FGA Group, stall, f3 available |
| 1040 | st %f3, [%g2 + 0x00] ! LSU |
| 1041 | fadds %f4, M13, %f4 ! FGA Group, f4 available |
| 1042 | st %f4, [%g2 + 0x04] ! LSU |
| 1043 | fadds %f5, M14, %f5 ! FGA Group, f5 available |
| 1044 | st %f5, [%g2 + 0x08] ! LEU |
| 1045 | bne 1b ! CTI |
| 1046 | add %g2, 0x10, %g2 ! IEU0 Group |
| 1047 | 2: |
| 1048 | ba __set_v4f_3 |
| 1049 | nop |
| 1050 | |
| 1051 | .globl _mesa_sparc_transform_points3_perspective |
| 1052 | _mesa_sparc_transform_points3_perspective: |
| 1053 | ld [%o2 + V4F_STRIDE], %o5 |
| 1054 | LDPTR [%o2 + V4F_START], %g1 |
| 1055 | LDPTR [%o0 + V4F_START], %g2 |
| 1056 | ld [%o2 + V4F_COUNT], %g3 |
| 1057 | |
| 1058 | LDMATRIX_0_5_8_9_10_14(%o1) |
| 1059 | |
| 1060 | cmp %g3, 0 |
| 1061 | st %g3, [%o0 + V4F_COUNT] |
| 1062 | be 2f |
| 1063 | clr %o1 |
| 1064 | |
| 1065 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 1066 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 1067 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 1068 | add %o1, 1, %o1 ! IEU0 |
| 1069 | add %g1, %o5, %g1 ! IEU1 |
| 1070 | fmuls %f0, M0, %f3 ! FGM |
| 1071 | fmuls %f2, M8, %f6 ! FGM Group |
| 1072 | fmuls %f1, M5, %f4 ! FGM Group |
| 1073 | fmuls %f2, M9, %f7 ! FGM Group |
| 1074 | fmuls %f2, M10, %f5 ! FGM Group f3 available |
| 1075 | fadds %f3, %f6, %f3 ! FGA Group f6 available |
| 1076 | st %f3, [%g2 + 0x00] ! LSU |
| 1077 | fadds %f4, %f7, %f4 ! FGA Group stall, f4, f7 available |
| 1078 | st %f4, [%g2 + 0x04] ! LSU |
| 1079 | fadds %f5, M14, %f5 ! FGA Group |
| 1080 | st %f5, [%g2 + 0x08] ! LSU |
| 1081 | fnegs %f2, %f6 ! FGA Group |
| 1082 | st %f6, [%g2 + 0x0c] ! LSU |
| 1083 | cmp %o1, %g3 ! IEU1 |
| 1084 | bne 1b ! CTI |
| 1085 | add %g2, 0x10, %g2 ! IEU0 Group |
| 1086 | 2: |
| 1087 | ba __set_v4f_4 |
| 1088 | nop |
| 1089 | |
| 1090 | .globl _mesa_sparc_transform_points4_general |
| 1091 | _mesa_sparc_transform_points4_general: |
| 1092 | ld [%o2 + V4F_STRIDE], %o5 |
| 1093 | LDPTR [%o2 + V4F_START], %g1 |
| 1094 | LDPTR [%o0 + V4F_START], %g2 |
| 1095 | ld [%o2 + V4F_COUNT], %g3 |
| 1096 | |
| 1097 | LDMATRIX_0_1_2_3_4_5_6_7_8_9_10_11_12_13_14_15(%o1) |
| 1098 | |
| 1099 | cmp %g3, 0 |
| 1100 | st %g3, [%o0 + V4F_COUNT] |
| 1101 | be 2f |
| 1102 | clr %o1 |
| 1103 | |
| 1104 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 1105 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 1106 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 1107 | ld [%g1 + 0x0c], %f3 ! LSU Group |
| 1108 | add %o1, 1, %o1 ! IEU0 |
| 1109 | add %g1, %o5, %g1 ! IEU1 |
| 1110 | fmuls %f0, M0, %f4 ! FGM Group |
| 1111 | fmuls %f1, M4, %f8 ! FGM Group |
| 1112 | fmuls %f0, M1, %f5 ! FGM Group |
| 1113 | fmuls %f1, M5, %f9 ! FGM Group |
| 1114 | fmuls %f0, M2, %f6 ! FGM Group f4 available |
| 1115 | fmuls %f1, M6, %f10 ! FGM Group f8 available |
| 1116 | fadds %f4, %f8, %f4 ! FGA |
| 1117 | fmuls %f0, M3, %f7 ! FGM Group f5 available |
| 1118 | fmuls %f1, M7, %f11 ! FGM Group f9 available |
| 1119 | fadds %f5, %f9, %f5 ! FGA |
| 1120 | fmuls %f2, M8, %f12 ! FGM Group f6 available |
| 1121 | fmuls %f2, M9, %f13 ! FGM Group f10, f4 available |
| 1122 | fadds %f6, %f10, %f6 ! FGA |
| 1123 | fmuls %f2, M10, %f14 ! FGM Group f7 available |
| 1124 | fmuls %f2, M11, %f15 ! FGM Group f11, f5 available |
| 1125 | fadds %f7, %f11, %f7 ! FGA |
| 1126 | fmuls %f3, M12, %f8 ! FGM Group f12 available |
| 1127 | fadds %f4, %f12, %f4 ! FGA |
| 1128 | fmuls %f3, M13, %f9 ! FGM Group f13, f6 available |
| 1129 | fadds %f5, %f13, %f5 ! FGA |
| 1130 | fmuls %f3, M14, %f10 ! FGM Group f14 available |
| 1131 | fadds %f6, %f14, %f6 ! FGA |
| 1132 | fmuls %f3, M15, %f11 ! FGM Group f15, f7 available |
| 1133 | fadds %f7, %f15, %f7 ! FGA |
| 1134 | fadds %f4, %f8, %f4 ! FGA Group f8, f4 available |
| 1135 | st %f4, [%g2 + 0x00] ! LSU |
| 1136 | fadds %f5, %f9, %f5 ! FGA Group f9, f5 available |
| 1137 | st %f5, [%g2 + 0x04] ! LSU |
| 1138 | fadds %f6, %f10, %f6 ! FGA Group f10, f6 available |
| 1139 | st %f6, [%g2 + 0x08] ! LSU |
| 1140 | fadds %f7, %f11, %f7 ! FGA Group f11, f7 available |
| 1141 | st %f7, [%g2 + 0x0c] ! LSU |
| 1142 | cmp %o1, %g3 ! IEU1 |
| 1143 | bne 1b ! CTI |
| 1144 | add %g2, 0x10, %g2 ! IEU0 Group |
| 1145 | 2: |
| 1146 | ba __set_v4f_4 |
| 1147 | nop |
| 1148 | |
| 1149 | .globl _mesa_sparc_transform_points4_identity |
| 1150 | _mesa_sparc_transform_points4_identity: |
| 1151 | ld [%o2 + V4F_STRIDE], %o5 |
| 1152 | LDPTR [%o2 + V4F_START], %g1 |
| 1153 | LDPTR [%o0 + V4F_START], %g2 |
| 1154 | ld [%o2 + V4F_COUNT], %g3 |
| 1155 | |
| 1156 | cmp %g3, 0 |
| 1157 | st %g3, [%o0 + V4F_COUNT] |
| 1158 | be 2f |
| 1159 | clr %o1 |
| 1160 | |
| 1161 | 1: ld [%g1 + 0x00], %f0 |
| 1162 | ld [%g1 + 0x04], %f1 |
| 1163 | ld [%g1 + 0x08], %f2 |
| 1164 | add %o1, 1, %o1 |
| 1165 | ld [%g1 + 0x0c], %f3 |
| 1166 | add %g1, %o5, %g1 |
| 1167 | st %f0, [%g2 + 0x00] |
| 1168 | st %f1, [%g2 + 0x04] |
| 1169 | st %f2, [%g2 + 0x08] |
| 1170 | cmp %o1, %g3 |
| 1171 | st %f3, [%g2 + 0x0c] |
| 1172 | bne 1b |
| 1173 | add %g2, 0x10, %g2 |
| 1174 | 2: |
| 1175 | ba __set_v4f_4 |
| 1176 | nop |
| 1177 | |
| 1178 | .globl _mesa_sparc_transform_points4_2d |
| 1179 | _mesa_sparc_transform_points4_2d: |
| 1180 | ld [%o2 + V4F_STRIDE], %o5 |
| 1181 | LDPTR [%o2 + V4F_START], %g1 |
| 1182 | LDPTR [%o0 + V4F_START], %g2 |
| 1183 | ld [%o2 + V4F_COUNT], %g3 |
| 1184 | |
| 1185 | LDMATRIX_0_1_4_5_12_13(%o1) |
| 1186 | |
| 1187 | cmp %g3, 0 |
| 1188 | st %g3, [%o0 + V4F_COUNT] |
| 1189 | be 2f |
| 1190 | clr %o1 |
| 1191 | |
| 1192 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 1193 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 1194 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 1195 | ld [%g1 + 0x0c], %f3 ! LSU Group |
| 1196 | add %o1, 1, %o1 ! IEU0 |
| 1197 | add %g1, %o5, %g1 ! IEU1 |
| 1198 | fmuls %f0, M0, %f4 ! FGM |
| 1199 | fmuls %f1, M4, %f8 ! FGM Group |
| 1200 | fmuls %f0, M1, %f5 ! FGM Group |
| 1201 | fmuls %f1, M5, %f9 ! FGM Group f4 available |
| 1202 | fmuls %f3, M12, %f12 ! FGM Group |
| 1203 | fmuls %f3, M13, %f13 ! FGM Group f8 available |
| 1204 | fadds %f4, %f8, %f4 ! FGA |
| 1205 | fadds %f5, %f9, %f5 ! FGA Group stall, f5, f9 available |
| 1206 | fadds %f4, %f12, %f4 ! FGA Group 2 cycle stall, f4, f12, f13 avail |
| 1207 | st %f4, [%g2 + 0x00] ! LSU |
| 1208 | fadds %f5, %f13, %f5 ! FGA Group f5 available |
| 1209 | st %f5, [%g2 + 0x04] ! LSU |
| 1210 | st %f2, [%g2 + 0x08] ! LSU Group |
| 1211 | st %f3, [%g2 + 0x0c] ! LSU Group |
| 1212 | cmp %o1, %g3 ! IEU1 |
| 1213 | bne 1b ! CTI |
| 1214 | add %g2, 0x10, %g2 ! IEU0 Group |
| 1215 | 2: |
| 1216 | ba __set_v4f_4 |
| 1217 | nop |
| 1218 | |
| 1219 | .globl _mesa_sparc_transform_points4_2d_no_rot |
| 1220 | _mesa_sparc_transform_points4_2d_no_rot: |
| 1221 | ld [%o2 + V4F_STRIDE], %o5 |
| 1222 | LDPTR [%o2 + V4F_START], %g1 |
| 1223 | LDPTR [%o0 + V4F_START], %g2 |
| 1224 | ld [%o2 + V4F_COUNT], %g3 |
| 1225 | |
| 1226 | LDMATRIX_0_1_4_5_12_13(%o1) |
| 1227 | |
| 1228 | cmp %g3, 0 |
| 1229 | st %g3, [%o0 + V4F_COUNT] |
| 1230 | be 2f |
| 1231 | clr %o1 |
| 1232 | |
| 1233 | 1: ld [%g1 + 0x00], %f0 |
| 1234 | ld [%g1 + 0x04], %f1 |
| 1235 | ld [%g1 + 0x08], %f2 |
| 1236 | ld [%g1 + 0x0c], %f3 |
| 1237 | add %o1, 1, %o1 |
| 1238 | add %g1, %o5, %g1 |
| 1239 | fmuls %f0, M0, %f4 |
| 1240 | fmuls %f3, M12, %f8 |
| 1241 | fmuls %f1, M5, %f5 |
| 1242 | fmuls %f3, M13, %f9 |
| 1243 | fadds %f4, %f8, %f4 |
| 1244 | st %f4, [%g2 + 0x00] |
| 1245 | fadds %f5, %f9, %f5 |
| 1246 | st %f5, [%g2 + 0x04] |
| 1247 | st %f2, [%g2 + 0x08] |
| 1248 | st %f3, [%g2 + 0x0c] |
| 1249 | cmp %o1, %g3 |
| 1250 | bne 1b |
| 1251 | add %g2, 0x10, %g2 |
| 1252 | 2: |
| 1253 | ba __set_v4f_4 |
| 1254 | nop |
| 1255 | |
| 1256 | .globl _mesa_sparc_transform_points4_3d |
| 1257 | _mesa_sparc_transform_points4_3d: |
| 1258 | ld [%o2 + V4F_STRIDE], %o5 |
| 1259 | LDPTR [%o2 + V4F_START], %g1 |
| 1260 | LDPTR [%o0 + V4F_START], %g2 |
| 1261 | ld [%o2 + V4F_COUNT], %g3 |
| 1262 | |
| 1263 | LDMATRIX_0_1_2_4_5_6_8_9_10_12_13_14(%o1) |
| 1264 | |
| 1265 | cmp %g3, 0 |
| 1266 | st %g3, [%o0 + V4F_COUNT] |
| 1267 | be 2f |
| 1268 | clr %o1 |
| 1269 | |
| 1270 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 1271 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 1272 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 1273 | ld [%g1 + 0x0c], %f3 ! LSU Group |
| 1274 | add %o1, 1, %o1 ! IEU0 |
| 1275 | add %g1, %o5, %g1 ! IEU1 |
| 1276 | fmuls %f0, M0, %f4 ! FGM |
| 1277 | fmuls %f1, M4, %f7 ! FGM Group |
| 1278 | fmuls %f0, M1, %f5 ! FGM Group |
| 1279 | fmuls %f1, M5, %f8 ! FGM Group |
| 1280 | fmuls %f0, M2, %f6 ! FGM Group f4 available |
| 1281 | fmuls %f1, M6, %f9 ! FGM Group f7 available |
| 1282 | fadds %f4, %f7, %f4 ! FGA |
| 1283 | fmuls %f2, M8, %f10 ! FGM Group f5 available |
| 1284 | fmuls %f2, M9, %f11 ! FGM Group f8 available |
| 1285 | fadds %f5, %f8, %f5 ! FGA |
| 1286 | fmuls %f2, M10, %f12 ! FGM Group f6 available |
| 1287 | fmuls %f3, M12, %f13 ! FGM Group f9, f4 available |
| 1288 | fadds %f6, %f9, %f6 ! FGA |
| 1289 | fmuls %f3, M13, %f14 ! FGM Group f10 available |
| 1290 | fadds %f4, %f10, %f4 ! FGA |
| 1291 | fmuls %f3, M14, %f15 ! FGM Group f11, f5 available |
| 1292 | fadds %f5, %f11, %f5 ! FGA |
| 1293 | fadds %f6, %f12, %f6 ! FGA Group stall, f12, f13, f6 available |
| 1294 | fadds %f4, %f13, %f4 ! FGA Group f14, f4 available |
| 1295 | st %f4, [%g2 + 0x00] ! LSU |
| 1296 | fadds %f5, %f14, %f5 ! FGA Group f15, f5 available |
| 1297 | st %f5, [%g2 + 0x04] ! LSU |
| 1298 | fadds %f6, %f15, %f6 ! FGA Group f6 available |
| 1299 | st %f6, [%g2 + 0x08] ! LSU |
| 1300 | st %f3, [%g2 + 0x0c] ! LSU Group |
| 1301 | cmp %o1, %g3 ! IEU1 |
| 1302 | bne 1b ! CTI |
| 1303 | add %g2, 0x10, %g2 ! IEU0 Group |
| 1304 | 2: |
| 1305 | ba __set_v4f_4 |
| 1306 | nop |
| 1307 | |
| 1308 | .globl _mesa_sparc_transform_points4_3d_no_rot |
| 1309 | _mesa_sparc_transform_points4_3d_no_rot: |
| 1310 | ld [%o2 + V4F_STRIDE], %o5 |
| 1311 | LDPTR [%o2 + V4F_START], %g1 |
| 1312 | LDPTR [%o0 + V4F_START], %g2 |
| 1313 | ld [%o2 + V4F_COUNT], %g3 |
| 1314 | |
| 1315 | LDMATRIX_0_5_10_12_13_14(%o1) |
| 1316 | |
| 1317 | cmp %g3, 0 |
| 1318 | st %g3, [%o0 + V4F_COUNT] |
| 1319 | be 2f |
| 1320 | clr %o1 |
| 1321 | |
| 1322 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 1323 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 1324 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 1325 | ld [%g1 + 0x0c], %f3 ! LSU Group |
| 1326 | add %o1, 1, %o1 ! IEU0 |
| 1327 | add %g1, %o5, %g1 ! IEU1 |
| 1328 | fmuls %f0, M0, %f4 ! FGM |
| 1329 | fmuls %f3, M12, %f7 ! FGM Group |
| 1330 | fmuls %f1, M5, %f5 ! FGM Group |
| 1331 | fmuls %f3, M13, %f8 ! FGM Group |
| 1332 | fmuls %f2, M10, %f6 ! FGM Group f4 available |
| 1333 | fmuls %f3, M14, %f9 ! FGM Group f7 available |
| 1334 | fadds %f4, %f7, %f4 ! FGA |
| 1335 | st %f4, [%g2 + 0x00] ! LSU |
| 1336 | fadds %f5, %f8, %f5 ! FGA Group stall, f5, f8 available |
| 1337 | st %f5, [%g2 + 0x04] ! LSU |
| 1338 | fadds %f6, %f9, %f6 ! FGA Group stall, f6, f9 available |
| 1339 | st %f6, [%g2 + 0x08] ! LSU |
| 1340 | st %f3, [%g2 + 0x0c] ! LSU Group |
| 1341 | cmp %o1, %g3 ! IEU1 |
| 1342 | bne 1b ! CTI |
| 1343 | add %g2, 0x10, %g2 ! IEU0 Group |
| 1344 | 2: |
| 1345 | ba __set_v4f_4 |
| 1346 | nop |
| 1347 | |
| 1348 | .globl _mesa_sparc_transform_points4_perspective |
| 1349 | _mesa_sparc_transform_points4_perspective: |
| 1350 | ld [%o2 + V4F_STRIDE], %o5 |
| 1351 | LDPTR [%o2 + V4F_START], %g1 |
| 1352 | LDPTR [%o0 + V4F_START], %g2 |
| 1353 | ld [%o2 + V4F_COUNT], %g3 |
| 1354 | |
| 1355 | LDMATRIX_0_5_8_9_10_14(%o1) |
| 1356 | |
| 1357 | cmp %g3, 0 |
| 1358 | st %g3, [%o0 + V4F_COUNT] |
| 1359 | be 2f |
| 1360 | clr %o1 |
| 1361 | |
| 1362 | 1: ld [%g1 + 0x00], %f0 ! LSU Group |
| 1363 | ld [%g1 + 0x04], %f1 ! LSU Group |
| 1364 | ld [%g1 + 0x08], %f2 ! LSU Group |
| 1365 | ld [%g1 + 0x0c], %f3 ! LSU Group |
| 1366 | add %o1, 1, %o1 ! IEU0 |
| 1367 | add %g1, %o5, %g1 ! IEU1 |
| 1368 | fmuls %f0, M0, %f4 ! FGM |
| 1369 | fmuls %f2, M8, %f7 ! FGM Group |
| 1370 | fmuls %f1, M5, %f5 ! FGM Group |
| 1371 | fmuls %f2, M9, %f8 ! FGM Group |
| 1372 | fmuls %f2, M10, %f6 ! FGM Group f4 available |
| 1373 | fmuls %f3, M14, %f9 ! FGM Group f7 available |
| 1374 | fadds %f4, %f7, %f4 ! FGA |
| 1375 | st %f4, [%g2 + 0x00] ! LSU |
| 1376 | fadds %f5, %f8, %f5 ! FGA Group stall, f5, f8 available |
| 1377 | st %f5, [%g2 + 0x04] ! LSU |
| 1378 | fadds %f6, %f9, %f6 ! FGA Group stall, f6, f9 available |
| 1379 | st %f6, [%g2 + 0x08] ! LSU |
| 1380 | fnegs %f2, %f7 ! FGA Group |
| 1381 | st %f7, [%g2 + 0x0c] ! LSU |
| 1382 | cmp %o1, %g3 ! IEU1 |
| 1383 | bne 1b ! CTI |
| 1384 | add %g2, 0x10, %g2 ! IEU0 Group |
| 1385 | 2: |
| 1386 | ba __set_v4f_4 |
| 1387 | nop |