1 2/* 3 * Mesa 3-D graphics library 4 * 5 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved. 6 * 7 * Permission is hereby granted, free of charge, to any person obtaining a 8 * copy of this software and associated documentation files (the "Software"), 9 * to deal in the Software without restriction, including without limitation 10 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 11 * and/or sell copies of the Software, and to permit persons to whom the 12 * Software is furnished to do so, subject to the following conditions: 13 * 14 * The above copyright notice and this permission notice shall be included 15 * in all copies or substantial portions of the Software. 16 * 17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR 21 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 22 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 23 * OTHER DEALINGS IN THE SOFTWARE. 24 */ 25 26/** TODO: 27 * - insert PREFETCH instructions to avoid cache-misses ! 28 * - some more optimizations are possible... 29 * - for 40-50% more performance in the SSE-functions, the 30 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned ! 31 */ 32 33#ifdef USE_SSE_ASM 34#include "assyntax.h" 35#define MATH_ASM_PTR_SIZE 4 36#include "math/m_vector_asm.h" 37#include "xform_args.h" 38 39 SEG_TEXT 40 41#define S(i) REGOFF(i * 4, ESI) 42#define D(i) REGOFF(i * 4, EDI) 43#define M(i) REGOFF(i * 4, EDX) 44 45 46ALIGNTEXT4 47GLOBL GLNAME(_mesa_sse_transform_points2_general) 48HIDDEN (_mesa_sse_transform_points2_general) 49GLNAME( _mesa_sse_transform_points2_general ): 50 _CET_ENDBR 51#define FRAME_OFFSET 8 52 PUSH_L ( ESI ) 53 PUSH_L ( EDI ) 54 55 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 56 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 57 58 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 59 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 60 61 TEST_L( ECX, ECX ) 62 JZ( LLBL(K_GTP2GR_finish) ) /* count was zero; go to finish */ 63 64 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 65 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 66 67 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 68 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 69 70 SHL_L( CONST(4), ECX ) /* count *= 16 */ 71 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 72 73 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 74 ADD_L( EDI, ECX ) /* count += dest ptr */ 75 76ALIGNTEXT32 77 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */ 78 MOVAPS( M(4), XMM1 ) /* m7 | m6 | m5 | m4 */ 79 MOVAPS( M(12), XMM2 ) /* m15 | m14 | m13 | m12 */ 80 81ALIGNTEXT32 82LLBL(K_GTP2GR_top): 83 MOVSS( S(0), XMM3 ) /* ox */ 84 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox | ox */ 85 MULPS( XMM0, XMM3 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */ 86 MOVSS( S(1), XMM4 ) /* oy */ 87 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy | oy */ 88 MULPS( XMM1, XMM4 ) /* oy*m7 | oy*m6 | oy*m5 | oy*m4 */ 89 90 ADDPS( XMM4, XMM3 ) 91 ADDPS( XMM2, XMM3 ) 92 MOVAPS( XMM3, D(0) ) 93 94LLBL(K_GTP2GR_skip): 95 ADD_L ( CONST(16), EDI ) 96 ADD_L ( EAX, ESI ) 97 CMP_L ( ECX, EDI ) 98 JNE ( LLBL(K_GTP2GR_top) ) 99 100LLBL(K_GTP2GR_finish): 101 POP_L ( EDI ) 102 POP_L ( ESI ) 103 RET 104#undef FRAME_OFFSET 105 106 107ALIGNTEXT4 108GLOBL GLNAME(_mesa_sse_transform_points2_identity) 109HIDDEN(_mesa_sse_transform_points2_identity) 110GLNAME( _mesa_sse_transform_points2_identity ): 111 _CET_ENDBR 112#define FRAME_OFFSET 8 113 PUSH_L ( ESI ) 114 PUSH_L ( EDI ) 115 116 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 117 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 118 119 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 120 121 TEST_L( ECX, ECX) 122 JZ( LLBL(K_GTP2IR_finish) ) /* count was zero; go to finish */ 123 124 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 125 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 126 127 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 128 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 129 130 SHL_L( CONST(4), ECX ) /* count *= 16 */ 131 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 132 133 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 134 ADD_L( EDI, ECX ) /* count += dest ptr */ 135 136 CMP_L( ESI, EDI ) 137 JE( LLBL(K_GTP2IR_finish) ) 138 139 140ALIGNTEXT32 141LLBL(K_GTP2IR_top): 142 MOV_L ( S(0), EDX ) 143 MOV_L ( EDX, D(0) ) 144 MOV_L ( S(1), EDX ) 145 MOV_L ( EDX, D(1) ) 146 147LLBL(K_GTP2IR_skip): 148 ADD_L ( CONST(16), EDI ) 149 ADD_L ( EAX, ESI ) 150 CMP_L ( ECX, EDI ) 151 JNE ( LLBL(K_GTP2IR_top) ) 152 153LLBL(K_GTP2IR_finish): 154 POP_L ( EDI ) 155 POP_L ( ESI ) 156 RET 157#undef FRAME_OFFSET 158 159 160ALIGNTEXT4 161GLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot) 162HIDDEN(_mesa_sse_transform_points2_3d_no_rot) 163GLNAME(_mesa_sse_transform_points2_3d_no_rot): 164 _CET_ENDBR 165#define FRAME_OFFSET 8 166 PUSH_L( ESI ) 167 PUSH_L( EDI ) 168 169 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 170 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 171 172 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 173 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 174 175 TEST_L( ECX, ECX) 176 JZ( LLBL(K_GTP23DNRR_finish) ) /* count was zero; go to finish */ 177 178 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 179 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 180 181 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 182 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 183 184 SHL_L( CONST(4), ECX ) /* count *= 16 */ 185 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 186 187 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 188 ADD_L( EDI, ECX ) /* count += dest ptr */ 189 190 XORPS( XMM0, XMM0 ) /* clean the working register */ 191 192ALIGNTEXT32 193 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 194 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 195 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 196 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */ 197 MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */ 198 199ALIGNTEXT32 200LLBL(K_GTP23DNRR_top): 201 MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */ 202 MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */ 203 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */ 204 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */ 205 206 MOVSS ( XMM3, D(2) ) /* -> D(2) */ 207 208LLBL(K_GTP23DNRR_skip): 209 ADD_L ( CONST(16), EDI ) 210 ADD_L ( EAX, ESI ) 211 CMP_L ( ECX, EDI ) 212 JNE ( LLBL(K_GTP23DNRR_top) ) 213 214LLBL(K_GTP23DNRR_finish): 215 POP_L ( EDI ) 216 POP_L ( ESI ) 217 RET 218#undef FRAME_OFFSET 219 220 221ALIGNTEXT4 222GLOBL GLNAME(_mesa_sse_transform_points2_perspective) 223HIDDEN(_mesa_sse_transform_points2_perspective) 224GLNAME(_mesa_sse_transform_points2_perspective): 225 _CET_ENDBR 226#define FRAME_OFFSET 8 227 PUSH_L ( ESI ) 228 PUSH_L ( EDI ) 229 230 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 231 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 232 233 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 234 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 235 236 TEST_L( ECX, ECX) 237 JZ( LLBL(K_GTP23PR_finish) ) /* count was zero; go to finish */ 238 239 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 240 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 241 242 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 243 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 244 245 SHL_L( CONST(4), ECX ) /* count *= 16 */ 246 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 247 248 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 249 ADD_L( EDI, ECX ) /* count += dest ptr */ 250 251ALIGNTEXT32 252 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 253 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 254 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 255 MOVSS ( M(14), XMM3 ) /* m14 */ 256 XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */ 257 258ALIGNTEXT32 259LLBL(K_GTP23PR_top): 260 MOVLPS( S(0), XMM4 ) /* oy | ox */ 261 MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */ 262 MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */ 263 MOVSS( XMM3, D(2) ) /* ->D(2) */ 264 MOVSS( XMM0, D(3) ) /* ->D(3) */ 265 266LLBL(K_GTP23PR_skip): 267 ADD_L( CONST(16), EDI ) 268 ADD_L( EAX, ESI ) 269 CMP_L( ECX, EDI ) 270 JNE( LLBL(K_GTP23PR_top) ) 271 272LLBL(K_GTP23PR_finish): 273 POP_L ( EDI ) 274 POP_L ( ESI ) 275 RET 276#undef FRAME_OFFSET 277 278 279 280ALIGNTEXT4 281GLOBL GLNAME(_mesa_sse_transform_points2_2d) 282HIDDEN(_mesa_sse_transform_points2_2d) 283GLNAME(_mesa_sse_transform_points2_2d): 284 _CET_ENDBR 285#define FRAME_OFFSET 8 286 PUSH_L( ESI ) 287 PUSH_L( EDI ) 288 289 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 290 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 291 292 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 293 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 294 295 TEST_L( ECX, ECX) 296 JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */ 297 298 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 299 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 300 301 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 302 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 303 304 SHL_L( CONST(4), ECX ) /* count *= 16 */ 305 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 306 307 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 308 ADD_L( EDI, ECX ) /* count += dest ptr */ 309 310ALIGNTEXT32 311 MOVLPS( M(0), XMM0 ) /* m1 | m0 */ 312 MOVLPS( M(4), XMM1 ) /* m5 | m4 */ 313 MOVLPS( M(12), XMM2 ) /* m13 | m12 */ 314 315ALIGNTEXT32 316LLBL(K_GTP23P2DR_top): 317 MOVSS( S(0), XMM3 ) /* ox */ 318 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */ 319 MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */ 320 321 MOVSS( S(1), XMM4 ) /* oy */ 322 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */ 323 MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */ 324 325 ADDPS( XMM4, XMM3 ) 326 ADDPS( XMM2, XMM3 ) 327 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */ 328 329LLBL(K_GTP23P2DR_skip): 330 ADD_L ( CONST(16), EDI ) 331 ADD_L ( EAX, ESI ) 332 CMP_L ( ECX, EDI ) 333 JNE ( LLBL(K_GTP23P2DR_top) ) 334 335LLBL(K_GTP23P2DR_finish): 336 POP_L ( EDI ) 337 POP_L ( ESI ) 338 RET 339#undef FRAME_OFFSET 340 341 342 343ALIGNTEXT4 344GLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot) 345HIDDEN(_mesa_sse_transform_points2_2d_no_rot) 346GLNAME(_mesa_sse_transform_points2_2d_no_rot): 347 _CET_ENDBR 348#define FRAME_OFFSET 8 349 PUSH_L( ESI ) 350 PUSH_L( EDI ) 351 352 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 353 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 354 355 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 356 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 357 358 TEST_L( ECX, ECX) 359 JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */ 360 361 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 362 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 363 364 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 365 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 366 367 SHL_L( CONST(4), ECX ) /* count *= 16 */ 368 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 369 370 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 371 ADD_L( EDI, ECX ) /* count += dest ptr */ 372 373ALIGNTEXT32 374 MOVSS ( M(0), XMM1 ) /* m0 */ 375 MOVSS ( M(5), XMM2 ) /* m5 */ 376 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */ 377 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */ 378 379ALIGNTEXT32 380LLBL(K_GTP23P2DNRR_top): 381 MOVLPS( S(0), XMM0 ) /* oy | ox */ 382 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 383 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */ 384 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 385 386LLBL(K_GTP23P2DNRR_skip): 387 ADD_L( CONST(16), EDI ) 388 ADD_L( EAX, ESI ) 389 CMP_L( ECX, EDI ) 390 JNE( LLBL(K_GTP23P2DNRR_top) ) 391 392LLBL(K_GTP23P2DNRR_finish): 393 POP_L( EDI ) 394 POP_L( ESI ) 395 RET 396#undef FRAME_OFFSET 397 398 399 400ALIGNTEXT4 401GLOBL GLNAME(_mesa_sse_transform_points2_3d) 402HIDDEN(_mesa_sse_transform_points2_3d) 403GLNAME(_mesa_sse_transform_points2_3d): 404 _CET_ENDBR 405#define FRAME_OFFSET 8 406 PUSH_L( ESI ) 407 PUSH_L( EDI ) 408 409 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 410 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 411 412 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 413 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 414 415 TEST_L( ECX, ECX) 416 JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */ 417 418 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 419 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 420 421 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 422 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 423 424 SHL_L( CONST(4), ECX ) /* count *= 16 */ 425 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 426 427 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 428 ADD_L( EDI, ECX ) /* count += dest ptr */ 429 430ALIGNTEXT32 431 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */ 432 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */ 433 MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */ 434 435ALIGNTEXT32 436LLBL(K_GTP23P3DR_top): 437 MOVSS( S(0), XMM3 ) /* ox */ 438 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */ 439 MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */ 440 441 MOVSS( S(1), XMM4 ) /* oy */ 442 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */ 443 MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */ 444 445 ADDPS( XMM4, XMM3 ) 446 ADDPS( XMM2, XMM3 ) 447 448 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */ 449 UNPCKHPS( XMM3, XMM3 ) 450 MOVSS( XMM3, D(2) ) /* ->D(2) */ 451 452LLBL(K_GTP23P3DR_skip): 453 ADD_L( CONST(16), EDI ) 454 ADD_L( EAX, ESI ) 455 CMP_L( ECX, EDI ) 456 JNE( LLBL(K_GTP23P3DR_top) ) 457 458LLBL(K_GTP23P3DR_finish): 459 POP_L( EDI ) 460 POP_L( ESI ) 461 RET 462#undef FRAME_OFFSET 463#endif 464 465#if defined (__ELF__) && defined (__linux__) 466 .section .note.GNU-stack,"",%progbits 467#endif 468