3 * Mesa 3-D graphics library
5 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
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:
14 * The above copyright notice and this permission notice shall be included
15 * in all copies or substantial portions of the Software.
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.
29 #include "xform_args.h"
33 #define FRAME_OFFSET 4
37 GLOBL GLNAME( _mesa_3dnow_transform_points4_general )
38 HIDDEN(_mesa_3dnow_transform_points4_general)
39 GLNAME( _mesa_3dnow_transform_points4_general ):
43 MOV_L ( ARG_DEST, ECX )
44 MOV_L ( ARG_MATRIX, ESI )
45 MOV_L ( ARG_SOURCE, EAX )
46 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
47 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
48 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
49 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
53 MOV_L ( REGOFF(V4F_START, ECX), EDX )
55 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
56 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
57 MOV_L ( REGOFF(V4F_START, EAX), EAX )
60 JZ ( LLBL( G3TPGR_2 ) )
62 PREFETCHW ( REGIND(EDX) )
67 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
69 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
70 MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */
72 ADD_L ( EDI, EAX ) /* next vertex */
73 PREFETCH ( REGIND(EAX) )
75 MOVQ ( MM0, MM2 ) /* x1 | x0 */
76 MOVQ ( MM4, MM6 ) /* x3 | x2 */
78 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
79 PUNPCKHDQ ( MM2, MM2 ) /* x1 | x1 */
81 MOVQ ( MM0, MM1 ) /* x0 | x0 */
82 ADD_L ( CONST(16), EDX ) /* next r */
84 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
85 MOVQ ( MM2, MM3 ) /* x1 | x1 */
87 PFMUL ( REGOFF(8, ECX), MM1 ) /* x0*m3 | x0*m2 */
88 PUNPCKLDQ ( MM4, MM4 ) /* x2 | x2 */
90 PFMUL ( REGOFF(16, ECX), MM2 ) /* x1*m5 | x1*m4 */
91 MOVQ ( MM4, MM5 ) /* x2 | x2 */
93 PFMUL ( REGOFF(24, ECX), MM3 ) /* x1*m7 | x1*m6 */
94 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
96 PFMUL ( REGOFF(32, ECX), MM4 ) /* x2*m9 | x2*m8 */
97 MOVQ ( MM6, MM7 ) /* x3 | x3 */
99 PFMUL ( REGOFF(40, ECX), MM5 ) /* x2*m11 | x2*m10 */
102 PFMUL ( REGOFF(48, ECX), MM6 ) /* x3*m13 | x3*m12 */
105 PFMUL ( REGOFF(56, ECX), MM7 ) /* x3*m15 | x3*m14 */
112 MOVQ ( MM6, REGOFF(-16, EDX) )
114 MOVQ ( MM7, REGOFF(-8, EDX) )
116 DEC_L ( ESI ) /* decrement vertex counter */
117 JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */
130 GLOBL GLNAME( _mesa_3dnow_transform_points4_perspective )
131 HIDDEN(_mesa_3dnow_transform_points4_perspective)
132 GLNAME( _mesa_3dnow_transform_points4_perspective ):
136 MOV_L ( ARG_DEST, ECX )
137 MOV_L ( ARG_MATRIX, ESI )
138 MOV_L ( ARG_SOURCE, EAX )
139 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
140 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
141 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
142 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
146 MOV_L ( REGOFF(V4F_START, ECX), EDX )
148 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
149 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
150 MOV_L ( REGOFF(V4F_START, EAX), EAX )
153 JZ ( LLBL( G3TPPR_2 ) )
155 PREFETCH ( REGIND(EAX) )
156 PREFETCHW ( REGIND(EDX) )
158 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
159 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
161 MOVD ( REGOFF(40, ECX), MM1 ) /* | m22 */
162 PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m32 | m22 */
164 MOVQ ( REGOFF(32, ECX), MM2 ) /* m21 | m20 */
165 PXOR ( MM7, MM7 ) /* 0 | 0 */
170 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
172 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
173 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
174 MOVD ( REGOFF(8, EAX), MM3 ) /* | x2 */
176 ADD_L ( EDI, EAX ) /* next vertex */
177 PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */
179 MOVQ ( MM5, MM6 ) /* x3 | x2 */
180 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
182 PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */
183 ADD_L ( CONST(16), EDX ) /* next r */
185 PFMUL ( MM2, MM5 ) /* x2*m21 | x2*m20 */
186 PFSUBR ( MM7, MM3 ) /* | -x2 */
188 PFMUL ( MM1, MM6 ) /* x3*m32 | x2*m22 */
189 PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */
191 PFACC ( MM3, MM6 ) /* -x2 | x2*m22+x3*m32 */
192 MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */
194 MOVQ ( MM6, REGOFF(-8, EDX) ) /* write r2, r3 */
195 DEC_L ( ESI ) /* decrement vertex counter */
197 JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */
210 GLOBL GLNAME( _mesa_3dnow_transform_points4_3d )
211 HIDDEN(_mesa_3dnow_transform_points4_3d)
212 GLNAME( _mesa_3dnow_transform_points4_3d ):
216 MOV_L ( ARG_DEST, ECX )
217 MOV_L ( ARG_MATRIX, ESI )
218 MOV_L ( ARG_SOURCE, EAX )
219 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
220 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
221 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
222 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
226 MOV_L ( REGOFF(V4F_START, ECX), EDX )
228 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
229 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
230 MOV_L ( REGOFF(V4F_START, EAX), EAX )
233 JZ ( LLBL( G3TP3R_2 ) )
235 MOVD ( REGOFF(8, ECX), MM6 ) /* | m2 */
236 PUNPCKLDQ ( REGOFF(24, ECX), MM6 ) /* m6 | m2 */
238 MOVD ( REGOFF(40, ECX), MM7 ) /* | m10 */
239 PUNPCKLDQ ( REGOFF(56, ECX), MM7 ) /* m14 | m10 */
244 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
245 PREFETCH ( REGOFF(32, EAX) ) /* hopefully array is tightly packed */
247 MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */
248 MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */
250 MOVQ ( MM2, MM0 ) /* x1 | x0 */
251 MOVQ ( MM3, MM4 ) /* x3 | x2 */
253 MOVQ ( MM0, MM1 ) /* x1 | x0 */
254 MOVQ ( MM4, MM5 ) /* x3 | x2 */
256 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
257 PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */
259 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
260 PUNPCKLDQ ( MM3, MM3 ) /* x2 | x2 */
262 PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */
263 PUNPCKHDQ ( MM4, MM4 ) /* x3 | x3 */
265 PFMUL ( MM6, MM2 ) /* x1*m6 | x0*m2 */
266 PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */
268 PFMUL ( REGOFF(32, ECX), MM3 ) /* x2*m9 | x2*m8 */
269 ADD_L ( CONST(16), EDX ) /* next r */
271 PFMUL ( REGOFF(48, ECX), MM4 ) /* x3*m13 | x3*m12 */
272 PFADD ( MM1, MM3 ) /* x0*m1+..+x2*m9 | x0*m0+...+x2*m8 */
274 PFMUL ( MM7, MM5 ) /* x3*m14 | x2*m10 */
275 PFADD ( MM3, MM4 ) /* r1 | r0 */
277 PFACC ( MM2, MM5 ) /* x0*m2+x1*m6 | x2*m10+x3*m14 */
278 MOVD ( REGOFF(12, EAX), MM0 ) /* | x3 */
280 ADD_L ( EDI, EAX ) /* next vertex */
281 PFACC ( MM0, MM5 ) /* r3 | r2 */
283 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
284 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
286 DEC_L ( ESI ) /* decrement vertex counter */
287 JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */
300 GLOBL GLNAME( _mesa_3dnow_transform_points4_3d_no_rot )
301 HIDDEN(_mesa_3dnow_transform_points4_3d_no_rot)
302 GLNAME( _mesa_3dnow_transform_points4_3d_no_rot ):
305 MOV_L ( ARG_DEST, ECX )
306 MOV_L ( ARG_MATRIX, ESI )
307 MOV_L ( ARG_SOURCE, EAX )
308 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
309 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
310 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
311 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
315 MOV_L ( REGOFF(V4F_START, ECX), EDX )
317 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
318 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
319 MOV_L ( REGOFF(V4F_START, EAX), EAX )
322 JZ ( LLBL( G3TP3NRR_2 ) )
324 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
325 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
327 MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
328 PUNPCKLDQ ( REGOFF(56, ECX), MM2 ) /* m32 | m22 */
330 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
335 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
337 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
338 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
339 MOVD ( REGOFF(12, EAX), MM7 ) /* | x3 */
341 ADD_L ( EDI, EAX ) /* next vertex */
342 PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */
344 MOVQ ( MM5, MM6 ) /* x3 | x2 */
345 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
347 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
348 PFMUL ( MM2, MM5 ) /* x3*m32 | x2*m22 */
350 PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */
351 PFACC ( MM7, MM5 ) /* x3 | x2*m22+x3*m32 */
353 PFADD ( MM6, MM4 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */
354 ADD_L ( CONST(16), EDX ) /* next r */
356 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
357 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
359 DEC_L ( ESI ) /* decrement vertex counter */
360 JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
373 GLOBL GLNAME( _mesa_3dnow_transform_points4_2d )
374 HIDDEN(_mesa_3dnow_transform_points4_2d)
375 GLNAME( _mesa_3dnow_transform_points4_2d ):
379 MOV_L ( ARG_DEST, ECX )
380 MOV_L ( ARG_MATRIX, ESI )
381 MOV_L ( ARG_SOURCE, EAX )
382 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
383 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
384 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
385 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
389 MOV_L ( REGOFF(V4F_START, ECX), EDX )
391 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
392 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
393 MOV_L ( REGOFF(V4F_START, EAX), EAX )
396 JZ ( LLBL( G3TP2R_2 ) )
398 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
399 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
401 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
402 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
404 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
409 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
411 MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */
412 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
414 ADD_L ( EDI, EAX ) /* next vertex */
415 PREFETCH ( REGIND(EAX) )
417 MOVQ ( MM3, MM4 ) /* x1 | x0 */
418 MOVQ ( MM5, MM6 ) /* x3 | x2 */
420 PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */
421 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
423 PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */
424 ADD_L ( CONST(16), EDX ) /* next r */
426 PFACC ( MM4, MM3 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */
427 PFMUL ( MM2, MM6 ) /* x3*m31 | x3*m30 */
429 PFADD ( MM6, MM3 ) /* r1 | r0 */
430 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
432 MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
434 DEC_L ( ESI ) /* decrement vertex counter */
435 JNZ ( LLBL( G3TP2R_1 ) ) /* cnt > 0 ? -> process next vertex */
448 GLOBL GLNAME( _mesa_3dnow_transform_points4_2d_no_rot )
449 HIDDEN(_mesa_3dnow_transform_points4_2d_no_rot)
450 GLNAME( _mesa_3dnow_transform_points4_2d_no_rot ):
454 MOV_L ( ARG_DEST, ECX )
455 MOV_L ( ARG_MATRIX, ESI )
456 MOV_L ( ARG_SOURCE, EAX )
457 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
458 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
459 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
460 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
464 MOV_L ( REGOFF(V4F_START, ECX), EDX )
466 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
467 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
468 MOV_L ( REGOFF(V4F_START, EAX), EAX )
471 JZ ( LLBL( G3TP2NRR_3 ) )
473 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
474 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
476 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
481 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
483 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
484 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
486 ADD_L ( EDI, EAX ) /* next vertex */
487 PREFETCH ( REGIND(EAX) )
489 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
490 MOVQ ( MM5, MM6 ) /* x3 | x2 */
492 ADD_L ( CONST(16), EDX ) /* next r */
493 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
495 PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */
496 PFADD ( MM4, MM6 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */
498 MOVQ ( MM6, REGOFF(-16, EDX) ) /* write r0, r1 */
499 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
501 DEC_L ( ESI ) /* decrement vertex counter */
503 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
516 GLOBL GLNAME( _mesa_3dnow_transform_points4_identity )
517 HIDDEN(_mesa_3dnow_transform_points4_identity)
518 GLNAME( _mesa_3dnow_transform_points4_identity ):
522 MOV_L ( ARG_DEST, ECX )
523 MOV_L ( ARG_MATRIX, ESI )
524 MOV_L ( ARG_SOURCE, EAX )
525 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
526 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
527 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
528 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
532 MOV_L ( REGOFF(V4F_START, ECX), EDX )
534 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
535 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
536 MOV_L ( REGOFF(V4F_START, EAX), EAX )
539 JZ ( LLBL( G3TPIR_2 ) )
544 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
546 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
547 MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */
549 ADD_L ( EDI, EAX ) /* next vertex */
550 PREFETCH ( REGIND(EAX) )
552 ADD_L ( CONST(16), EDX ) /* next r */
553 MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */
555 MOVQ ( MM1, REGOFF(-8, EDX) ) /* r3 | r2 */
557 DEC_L ( ESI ) /* decrement vertex counter */
558 JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */
568 #if defined (__ELF__) && defined (__linux__)
569 .section .note.GNU-stack,"",%progbits