1 /* $Id: 3dnow_xform4.S,v 1.1 2001/03/29 06:46:16 gareth Exp $ */
4 * Mesa 3-D graphics library
7 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
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:
16 * The above copyright notice and this permission notice shall be included
17 * in all copies or substantial portions of the Software.
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.
28 #include "xform_args.h"
32 #define FRAME_OFFSET 4
36 GLOBL GLNAME( _mesa_3dnow_transform_points4_general )
37 GLNAME( _mesa_3dnow_transform_points4_general ):
41 MOV_L ( ARG_DEST, ECX )
42 MOV_L ( ARG_MATRIX, ESI )
43 MOV_L ( ARG_SOURCE, EAX )
44 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
45 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
46 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
47 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
51 MOV_L ( REGOFF(V4F_START, ECX), EDX )
53 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
54 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
55 MOV_L ( REGOFF(V4F_START, EAX), EAX )
58 JZ ( LLBL( G3TPGR_2 ) )
60 PREFETCHW ( REGIND(EDX) )
62 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
63 MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */
65 ADD_L ( EDI, EAX ) /* next vertex */
70 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
71 PREFETCH ( REGIND(EAX) )
73 MOVQ ( MM0, MM2 ) /* x1 | x0 */
74 MOVQ ( MM4, MM6 ) /* x3 | x2 */
76 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
77 PUNPCKHDQ ( MM2, MM2 ) /* x1 | x1 */
79 MOVQ ( MM0, MM1 ) /* x0 | x0 */
80 ADD_L ( CONST(16), EDX ) /* next r */
82 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
83 MOVQ ( MM2, MM3 ) /* x1 | x1 */
85 PFMUL ( REGOFF(8, ECX), MM1 ) /* x0*m3 | x0*m2 */
86 PUNPCKLDQ ( MM4, MM4 ) /* x2 | x2 */
88 PFMUL ( REGOFF(16, ECX), MM2 ) /* x1*m5 | x1*m4 */
89 MOVQ ( MM4, MM5 ) /* x2 | x2 */
91 PFMUL ( REGOFF(24, ECX), MM3 ) /* x1*m7 | x1*m6 */
92 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
94 PFMUL ( REGOFF(32, ECX), MM4 ) /* x2*m9 | x2*m8 */
95 MOVQ ( MM6, MM7 ) /* x3 | x3 */
97 PFMUL ( REGOFF(40, ECX), MM5 ) /* x2*m11 | x2*m10 */
100 PFMUL ( REGOFF(48, ECX), MM6 ) /* x3*m13 | x3*m12 */
103 PFMUL ( REGOFF(56, ECX), MM7 ) /* x3*m15 | x3*m14 */
110 MOVQ ( MM6, REGOFF(-16, EDX) )
112 MOVQ ( MM7, REGOFF(-8, EDX) )
113 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
115 MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */
116 ADD_L ( EDI, EAX ) /* next vertex */
118 DEC_L ( ESI ) /* decrement vertex counter */
119 JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */
132 GLOBL GLNAME( _mesa_3dnow_transform_points4_perspective )
133 GLNAME( _mesa_3dnow_transform_points4_perspective ):
137 MOV_L ( ARG_DEST, ECX )
138 MOV_L ( ARG_MATRIX, ESI )
139 MOV_L ( ARG_SOURCE, EAX )
140 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
141 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
142 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
143 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
147 MOV_L ( REGOFF(V4F_START, ECX), EDX )
149 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
150 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
151 MOV_L ( REGOFF(V4F_START, EAX), EAX )
154 JZ ( LLBL( G3TPPR_2 ) )
156 PREFETCH ( REGIND(EAX) )
157 PREFETCHW ( REGIND(EDX) )
159 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
160 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
162 MOVD ( REGOFF(40, ECX), MM1 ) /* | m22 */
163 PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m32 | m22 */
165 MOVQ ( REGOFF(32, ECX), MM2 ) /* m21 | m20 */
166 PXOR ( MM7, MM7 ) /* 0 | 0 */
168 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
169 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
170 MOVD ( REGOFF(8, EAX), MM3 ) /* | x2 */
172 ADD_L ( EDI, EAX ) /* next vertex */
177 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
178 PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */
180 MOVQ ( MM5, MM6 ) /* x3 | x2 */
181 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
183 PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */
184 ADD_L ( CONST(16), EDX ) /* next r */
186 PFMUL ( MM2, MM5 ) /* x2*m21 | x2*m20 */
187 PFSUBR ( MM7, MM3 ) /* | -x2 */
189 PFMUL ( MM1, MM6 ) /* x3*m32 | x2*m22 */
190 PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */
192 PFACC ( MM3, MM6 ) /* -x2 | x2*m22+x3*m32 */
193 MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */
195 MOVQ ( MM6, REGOFF(-8, EDX) ) /* write r2, r3 */
196 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
198 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
199 MOVD ( REGOFF(8, EAX), MM3 ) /* | x2 */
201 ADD_L ( EDI, EAX ) /* next vertex */
202 DEC_L ( ESI ) /* decrement vertex counter */
204 JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */
217 GLOBL GLNAME( _mesa_3dnow_transform_points4_3d )
218 GLNAME( _mesa_3dnow_transform_points4_3d ):
222 MOV_L ( ARG_DEST, ECX )
223 MOV_L ( ARG_MATRIX, ESI )
224 MOV_L ( ARG_SOURCE, EAX )
225 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
226 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
227 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
228 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
232 MOV_L ( REGOFF(V4F_START, ECX), EDX )
234 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
235 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
236 MOV_L ( REGOFF(V4F_START, EAX), EAX )
239 JZ ( LLBL( G3TP3R_2 ) )
241 MOVD ( REGOFF(8, ECX), MM6 ) /* | m2 */
242 PUNPCKLDQ ( REGOFF(24, ECX), MM6 ) /* m6 | m2 */
244 MOVD ( REGOFF(40, ECX), MM7 ) /* | m10 */
245 PUNPCKLDQ ( REGOFF(56, ECX), MM7 ) /* m14 | m10 */
247 MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */
248 MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */
253 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
254 PREFETCH ( REGOFF(32, EAX) ) /* hopefully array is tightly packed */
256 MOVQ ( MM2, MM0 ) /* x1 | x0 */
257 MOVQ ( MM3, MM4 ) /* x3 | x2 */
259 MOVQ ( MM0, MM1 ) /* x1 | x0 */
260 MOVQ ( MM4, MM5 ) /* x3 | x2 */
262 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
263 PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */
265 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
266 PUNPCKLDQ ( MM3, MM3 ) /* x2 | x2 */
268 PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */
269 PUNPCKHDQ ( MM4, MM4 ) /* x3 | x3 */
271 PFMUL ( MM6, MM2 ) /* x1*m6 | x0*m2 */
272 PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */
274 PFMUL ( REGOFF(32, ECX), MM3 ) /* x2*m9 | x2*m8 */
275 ADD_L ( CONST(16), EDX ) /* next r */
277 PFMUL ( REGOFF(48, ECX), MM4 ) /* x3*m13 | x3*m12 */
278 PFADD ( MM1, MM3 ) /* x0*m1+..+x2*m9 | x0*m0+...+x2*m8 */
280 PFMUL ( MM7, MM5 ) /* x3*m14 | x2*m10 */
281 PFADD ( MM3, MM4 ) /* r1 | r0 */
283 PFACC ( MM2, MM5 ) /* x0*m2+x1*m6 | x2*m10+x3*m14 */
284 MOVD ( REGOFF(12, EAX), MM0 ) /* | x3 */
286 ADD_L ( EDI, EAX ) /* next vertex */
287 PFACC ( MM0, MM5 ) /* r3 | r2 */
289 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
290 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
292 MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */
293 MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */
295 DEC_L ( ESI ) /* decrement vertex counter */
296 JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */
309 GLOBL GLNAME( _mesa_3dnow_transform_points4_3d_no_rot )
310 GLNAME( _mesa_3dnow_transform_points4_3d_no_rot ):
313 MOV_L ( ARG_DEST, ECX )
314 MOV_L ( ARG_MATRIX, ESI )
315 MOV_L ( ARG_SOURCE, EAX )
316 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
317 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
318 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
319 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
323 MOV_L ( REGOFF(V4F_START, ECX), EDX )
325 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
326 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
327 MOV_L ( REGOFF(V4F_START, EAX), EAX )
330 JZ ( LLBL( G3TP3NRR_2 ) )
332 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
333 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
335 MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
336 PUNPCKLDQ ( REGOFF(56, ECX), MM2 ) /* m32 | m22 */
338 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
340 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
341 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
342 MOVD ( REGOFF(12, EAX), MM7 ) /* | x3 */
344 ADD_L ( EDI, EAX ) /* next vertex */
349 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
350 PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */
352 MOVQ ( MM5, MM6 ) /* x3 | x2 */
353 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
355 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
356 PFMUL ( MM2, MM5 ) /* x3*m32 | x2*m22 */
358 PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */
359 PFACC ( MM7, MM5 ) /* x3 | x2*m22+x3*m32 */
361 PFADD ( MM6, MM4 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */
362 ADD_L ( CONST(16), EDX ) /* next r */
364 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
365 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
367 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
368 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
370 MOVD ( REGOFF(12, EAX), MM7 ) /* | x3 */
371 ADD_L ( EDI, EAX ) /* next vertex */
373 DEC_L ( ESI ) /* decrement vertex counter */
374 JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
387 GLOBL GLNAME( _mesa_3dnow_transform_points4_2d )
388 GLNAME( _mesa_3dnow_transform_points4_2d ):
392 MOV_L ( ARG_DEST, ECX )
393 MOV_L ( ARG_MATRIX, ESI )
394 MOV_L ( ARG_SOURCE, EAX )
395 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
396 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
397 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
398 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
402 MOV_L ( REGOFF(V4F_START, ECX), EDX )
404 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
405 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
406 MOV_L ( REGOFF(V4F_START, EAX), EAX )
409 JZ ( LLBL( G3TP2R_2 ) )
411 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
412 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
414 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
415 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
417 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
419 MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */
420 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
422 ADD_L ( EDI, EAX ) /* next vertex */
427 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
428 PREFETCH ( REGIND(EAX) )
430 MOVQ ( MM3, MM4 ) /* x1 | x0 */
431 MOVQ ( MM5, MM6 ) /* x3 | x2 */
433 PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */
434 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
436 PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */
437 ADD_L ( CONST(16), EDX ) /* next r */
439 PFACC ( MM4, MM3 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */
440 PFMUL ( MM2, MM6 ) /* x3*m31 | x3*m30 */
442 PFADD ( MM6, MM3 ) /* r1 | r0 */
443 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
445 MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
446 MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */
448 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
449 ADD_L ( EDI, EAX ) /* next vertex */
451 DEC_L ( ESI ) /* decrement vertex counter */
452 JNZ ( LLBL( G3TP2R_1 ) ) /* cnt > 0 ? -> process next vertex */
465 GLOBL GLNAME( _mesa_3dnow_transform_points4_2d_no_rot )
466 GLNAME( _mesa_3dnow_transform_points4_2d_no_rot ):
470 MOV_L ( ARG_DEST, ECX )
471 MOV_L ( ARG_MATRIX, ESI )
472 MOV_L ( ARG_SOURCE, EAX )
473 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
474 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
475 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
476 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
480 MOV_L ( REGOFF(V4F_START, ECX), EDX )
482 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
483 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
484 MOV_L ( REGOFF(V4F_START, EAX), EAX )
487 JZ ( LLBL( G3TP2NRR_3 ) )
489 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
490 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
492 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
494 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
495 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
497 ADD_L ( EDI, EAX ) /* next vertex */
502 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
503 PREFETCH ( REGIND(EAX) )
505 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
506 MOVQ ( MM5, MM6 ) /* x3 | x2 */
508 ADD_L ( CONST(16), EDX ) /* next r */
509 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
511 PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */
512 PFADD ( MM4, MM6 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */
514 MOVQ ( MM6, REGOFF(-16, EDX) ) /* write r0, r1 */
515 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
517 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
518 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
520 ADD_L ( EDI, EAX ) /* next vertex */
521 DEC_L ( ESI ) /* decrement vertex counter */
523 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
536 GLOBL GLNAME( _mesa_3dnow_transform_points4_identity )
537 GLNAME( _mesa_3dnow_transform_points4_identity ):
541 MOV_L ( ARG_DEST, ECX )
542 MOV_L ( ARG_MATRIX, ESI )
543 MOV_L ( ARG_SOURCE, EAX )
544 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
545 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
546 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
547 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
551 MOV_L ( REGOFF(V4F_START, ECX), EDX )
553 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
554 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
555 MOV_L ( REGOFF(V4F_START, EAX), EAX )
558 JZ ( LLBL( G3TPIR_2 ) )
560 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
561 MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */
563 ADD_L ( EDI, EAX ) /* next vertex */
568 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
569 PREFETCH ( REGIND(EAX) )
571 ADD_L ( CONST(16), EDX ) /* next r */
572 MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */
574 MOVQ ( MM1, REGOFF(-8, EDX) ) /* r3 | r2 */
575 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
577 MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */
578 ADD_L ( EDI, EAX ) /* next vertex */
580 DEC_L ( ESI ) /* decrement vertex counter */
581 JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */