sse_xform2.S revision 3474e9de924d92a941b4ea33ecc694f5fad2651f
1/* $Id: sse_xform2.S,v 1.4 2006/04/17 18:58:24 krh Exp $ */ 2 3/* 4 * Mesa 3-D graphics library 5 * Version: 3.5 6 * 7 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved. 8 * 9 * Permission is hereby granted, free of charge, to any person obtaining a 10 * copy of this software and associated documentation files (the "Software"), 11 * to deal in the Software without restriction, including without limitation 12 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 13 * and/or sell copies of the Software, and to permit persons to whom the 14 * Software is furnished to do so, subject to the following conditions: 15 * 16 * The above copyright notice and this permission notice shall be included 17 * in all copies or substantial portions of the Software. 18 * 19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS 20 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 22 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN 23 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 24 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. 25 */ 26 27/** TODO: 28 * - insert PREFETCH instructions to avoid cache-misses ! 29 * - some more optimizations are possible... 30 * - for 40-50% more performance in the SSE-functions, the 31 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned ! 32 */ 33 34#ifdef USE_SSE_ASM 35#include "assyntax.h" 36#include "matypes.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 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 ): 111 112#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): 164 165#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 191ALIGNTEXT32 192 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 193 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 194 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 195 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */ 196 MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */ 197 198ALIGNTEXT32 199LLBL(K_GTP23DNRR_top): 200 MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */ 201 MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */ 202 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */ 203 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */ 204 205 MOVSS ( XMM3, D(2) ) /* -> D(2) */ 206 207LLBL(K_GTP23DNRR_skip): 208 ADD_L ( CONST(16), EDI ) 209 ADD_L ( EAX, ESI ) 210 CMP_L ( ECX, EDI ) 211 JNE ( LLBL(K_GTP23DNRR_top) ) 212 213LLBL(K_GTP23DNRR_finish): 214 POP_L ( EDI ) 215 POP_L ( ESI ) 216 RET 217#undef FRAME_OFFSET 218 219 220ALIGNTEXT4 221GLOBL GLNAME(_mesa_sse_transform_points2_perspective) 222HIDDEN(_mesa_sse_transform_points2_perspective) 223GLNAME(_mesa_sse_transform_points2_perspective): 224 225#define FRAME_OFFSET 8 226 PUSH_L ( ESI ) 227 PUSH_L ( EDI ) 228 229 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 230 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 231 232 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 233 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 234 235 TEST_L( ECX, ECX) 236 JZ( LLBL(K_GTP23PR_finish) ) /* count was zero; go to finish */ 237 238 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 239 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 240 241 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 242 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 243 244 SHL_L( CONST(4), ECX ) /* count *= 16 */ 245 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 246 247 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 248 ADD_L( EDI, ECX ) /* count += dest ptr */ 249 250ALIGNTEXT32 251 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */ 252 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */ 253 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */ 254 MOVSS ( M(14), XMM3 ) /* m14 */ 255 XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */ 256 257ALIGNTEXT32 258LLBL(K_GTP23PR_top): 259 MOVLPS( S(0), XMM4 ) /* oy | ox */ 260 MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */ 261 MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */ 262 MOVSS( XMM3, D(2) ) /* ->D(2) */ 263 MOVSS( XMM0, D(3) ) /* ->D(3) */ 264 265LLBL(K_GTP23PR_skip): 266 ADD_L( CONST(16), EDI ) 267 ADD_L( EAX, ESI ) 268 CMP_L( ECX, EDI ) 269 JNE( LLBL(K_GTP23PR_top) ) 270 271LLBL(K_GTP23PR_finish): 272 POP_L ( EDI ) 273 POP_L ( ESI ) 274 RET 275#undef FRAME_OFFSET 276 277 278 279ALIGNTEXT4 280GLOBL GLNAME(_mesa_sse_transform_points2_2d) 281HIDDEN(_mesa_sse_transform_points2_2d) 282GLNAME(_mesa_sse_transform_points2_2d): 283 284#define FRAME_OFFSET 8 285 PUSH_L( ESI ) 286 PUSH_L( EDI ) 287 288 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 289 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 290 291 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 292 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 293 294 TEST_L( ECX, ECX) 295 JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */ 296 297 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 298 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 299 300 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 301 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 302 303 SHL_L( CONST(4), ECX ) /* count *= 16 */ 304 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 305 306 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 307 ADD_L( EDI, ECX ) /* count += dest ptr */ 308 309ALIGNTEXT32 310 MOVLPS( M(0), XMM0 ) /* m1 | m0 */ 311 MOVLPS( M(4), XMM1 ) /* m5 | m4 */ 312 MOVLPS( M(12), XMM2 ) /* m13 | m12 */ 313 314ALIGNTEXT32 315LLBL(K_GTP23P2DR_top): 316 MOVSS( S(0), XMM3 ) /* ox */ 317 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */ 318 MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */ 319 320 MOVSS( S(1), XMM4 ) /* oy */ 321 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */ 322 MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */ 323 324 ADDPS( XMM4, XMM3 ) 325 ADDPS( XMM2, XMM3 ) 326 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */ 327 328LLBL(K_GTP23P2DR_skip): 329 ADD_L ( CONST(16), EDI ) 330 ADD_L ( EAX, ESI ) 331 CMP_L ( ECX, EDI ) 332 JNE ( LLBL(K_GTP23P2DR_top) ) 333 334LLBL(K_GTP23P2DR_finish): 335 POP_L ( EDI ) 336 POP_L ( ESI ) 337 RET 338#undef FRAME_OFFSET 339 340 341 342ALIGNTEXT4 343GLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot) 344HIDDEN(_mesa_sse_transform_points2_2d_no_rot) 345GLNAME(_mesa_sse_transform_points2_2d_no_rot): 346 347#define FRAME_OFFSET 8 348 PUSH_L( ESI ) 349 PUSH_L( EDI ) 350 351 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 352 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 353 354 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 355 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 356 357 TEST_L( ECX, ECX) 358 JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */ 359 360 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 361 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 362 363 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 364 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 365 366 SHL_L( CONST(4), ECX ) /* count *= 16 */ 367 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 368 369 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 370 ADD_L( EDI, ECX ) /* count += dest ptr */ 371 372ALIGNTEXT32 373 MOVSS ( M(0), XMM1 ) /* m0 */ 374 MOVSS ( M(5), XMM2 ) /* m5 */ 375 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */ 376 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */ 377 378ALIGNTEXT32 379LLBL(K_GTP23P2DNRR_top): 380 MOVLPS( S(0), XMM0 ) /* oy | ox */ 381 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */ 382 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */ 383 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */ 384 385LLBL(K_GTP23P2DNRR_skip): 386 ADD_L( CONST(16), EDI ) 387 ADD_L( EAX, ESI ) 388 CMP_L( ECX, EDI ) 389 JNE( LLBL(K_GTP23P2DNRR_top) ) 390 391LLBL(K_GTP23P2DNRR_finish): 392 POP_L( EDI ) 393 POP_L( ESI ) 394 RET 395#undef FRAME_OFFSET 396 397 398 399ALIGNTEXT4 400GLOBL GLNAME(_mesa_sse_transform_points2_3d) 401HIDDEN(_mesa_sse_transform_points2_3d) 402GLNAME(_mesa_sse_transform_points2_3d): 403 404#define FRAME_OFFSET 8 405 PUSH_L( ESI ) 406 PUSH_L( EDI ) 407 408 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */ 409 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */ 410 411 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */ 412 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */ 413 414 TEST_L( ECX, ECX) 415 JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */ 416 417 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */ 418 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */ 419 420 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */ 421 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */ 422 423 SHL_L( CONST(4), ECX ) /* count *= 16 */ 424 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */ 425 426 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */ 427 ADD_L( EDI, ECX ) /* count += dest ptr */ 428 429ALIGNTEXT32 430 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */ 431 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */ 432 MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */ 433 434ALIGNTEXT32 435LLBL(K_GTP23P3DR_top): 436 MOVSS( S(0), XMM3 ) /* ox */ 437 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */ 438 MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */ 439 440 MOVSS( S(1), XMM4 ) /* oy */ 441 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */ 442 MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */ 443 444 ADDPS( XMM4, XMM3 ) 445 ADDPS( XMM2, XMM3 ) 446 447 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */ 448 UNPCKHPS( XMM3, XMM3 ) 449 MOVSS( XMM3, D(2) ) /* ->D(2) */ 450 451LLBL(K_GTP23P3DR_skip): 452 ADD_L( CONST(16), EDI ) 453 ADD_L( EAX, ESI ) 454 CMP_L( ECX, EDI ) 455 JNE( LLBL(K_GTP23P3DR_top) ) 456 457LLBL(K_GTP23P3DR_finish): 458 POP_L( EDI ) 459 POP_L( ESI ) 460 RET 461#undef FRAME_OFFSET 462#endif 463 464#if defined (__ELF__) && defined (__linux__) 465 .section .note.GNU-stack,"",%progbits 466#endif 467