1 /* $Id: 3dnow_xform3.S,v 1.2 2001/10/22 01:21:16 brianp 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_points3_general )
37 GLNAME( _mesa_3dnow_transform_points3_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) )
65 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
67 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
68 MOVD ( REGOFF(8, EAX), MM2 ) /* | x2 */
70 ADD_L ( EDI, EAX ) /* next vertex */
71 PREFETCH ( REGIND(EAX) )
73 MOVQ ( MM0, MM1 ) /* x1 | x0 */
74 PUNPCKLDQ ( MM2, MM2 ) /* x2 | x2 */
76 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
77 MOVQ ( MM2, MM5 ) /* x2 | x2 */
79 PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */
80 PFMUL ( REGOFF(32, ECX), MM2 ) /* x2*m9 | x2*m8 */
82 MOVQ ( MM0, MM3 ) /* x0 | x0 */
83 PFMUL ( REGOFF(40, ECX), MM5 ) /* x2*m11 | x2*m10 */
85 MOVQ ( MM1, MM4 ) /* x1 | x1 */
86 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
88 PFADD ( REGOFF(48, ECX), MM2 ) /* x2*m9+m13 | x2*m8+m12 */
89 PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */
91 PFADD ( REGOFF(56, ECX), MM5 ) /* x2*m11+m15 | x2*m10+m14 */
92 PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */
94 PFMUL ( REGOFF(8, ECX), MM3 ) /* x0*m3 | x0*m2 */
95 PFADD ( MM1, MM2 ) /* r1 | r0 */
97 PFMUL ( REGOFF(24, ECX), MM4 ) /* x1*m7 | x1*m6 */
98 ADD_L ( CONST(16), EDX ) /* next output vertex */
100 PFADD ( MM3, MM4 ) /* x0*m3+x1*m7 | x0*m2+x1*m6 */
101 MOVQ ( MM2, REGOFF(-16, EDX) ) /* write r0, r1 */
103 PFADD ( MM4, MM5 ) /* r3 | r2 */
104 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
106 DEC_L ( ESI ) /* decrement vertex counter */
107 JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */
120 GLOBL GLNAME( _mesa_3dnow_transform_points3_perspective )
121 GLNAME( _mesa_3dnow_transform_points3_perspective ):
125 MOV_L ( ARG_DEST, ECX )
126 MOV_L ( ARG_MATRIX, ESI )
127 MOV_L ( ARG_SOURCE, EAX )
128 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
129 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
130 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
131 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
135 MOV_L ( REGOFF(V4F_START, ECX), EDX )
137 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
138 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
139 MOV_L ( REGOFF(V4F_START, EAX), EAX )
142 JZ ( LLBL( G3TPPR_2 ) )
144 PREFETCH ( REGIND(EAX) )
145 PREFETCHW ( REGIND(EDX) )
147 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
148 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
150 MOVQ ( REGOFF(32, ECX), MM1 ) /* m21 | m20 */
151 MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
153 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
158 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
160 MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
161 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
163 ADD_L ( EDI, EAX ) /* next vertex */
164 PREFETCH ( REGIND(EAX) )
166 PXOR ( MM7, MM7 ) /* 0 | 0 */
167 MOVQ ( MM5, MM6 ) /* | x2 */
169 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
170 PFSUB ( MM5, MM7 ) /* | -x2 */
172 PFMUL ( MM2, MM6 ) /* | x2*m22 */
173 PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */
175 ADD_L ( CONST(16), EDX ) /* next r */
176 PFMUL ( MM1, MM5 ) /* x2*m21 | x2*m20 */
178 PFADD ( MM3, MM6 ) /* | x2*m22+m32 */
179 PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */
181 MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */
182 MOVD ( MM6, REGOFF(-8, EDX) ) /* write r2 */
184 MOVD ( MM7, REGOFF(-4, EDX) ) /* write r3 */
186 DEC_L ( ESI ) /* decrement vertex counter */
187 JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */
200 GLOBL GLNAME( _mesa_3dnow_transform_points3_3d )
201 GLNAME( _mesa_3dnow_transform_points3_3d ):
205 MOV_L ( ARG_DEST, ECX )
206 MOV_L ( ARG_MATRIX, ESI )
207 MOV_L ( ARG_SOURCE, EAX )
208 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
209 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
210 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
211 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
215 MOV_L ( REGOFF(V4F_START, ECX), EDX )
217 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
218 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
219 MOV_L ( REGOFF(V4F_START, EAX), EAX )
222 JZ ( LLBL( G3TP3R_2 ) )
224 PREFETCH ( REGIND(EAX) )
225 PREFETCH ( REGIND(EDX) )
227 MOVD ( REGOFF(8, ECX), MM7 ) /* | m2 */
228 PUNPCKLDQ ( REGOFF(24, ECX), MM7 ) /* m6 | m2 */
234 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
236 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
237 MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */
239 ADD_L ( EDI, EAX ) /* next vertex */
240 PREFETCH ( REGIND(EAX) )
242 MOVQ ( MM0, MM2 ) /* x1 | x0 */
243 ADD_L ( CONST(16), EDX ) /* next r */
245 PUNPCKLDQ ( MM2, MM2 ) /* x0 | x0 */
246 MOVQ ( MM0, MM3 ) /* x1 | x0 */
248 PFMUL ( REGIND(ECX), MM2 ) /* x0*m1 | x0*m0 */
249 PUNPCKHDQ ( MM3, MM3 ) /* x1 | x1 */
251 MOVQ ( MM1, MM4 ) /* | x2 */
252 PFMUL ( REGOFF(16, ECX), MM3 ) /* x1*m5 | x1*m4 */
254 PUNPCKLDQ ( MM4, MM4 ) /* x2 | x2 */
255 PFADD ( MM2, MM3 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */
257 PFMUL ( REGOFF(32, ECX), MM4 ) /* x2*m9 | x2*m8 */
258 PFADD ( REGOFF(48, ECX), MM3 ) /* x0*m1+...+m11 | x0*m0+x1*m4+m12 */
260 PFMUL ( MM7, MM0 ) /* x1*m6 | x0*m2 */
261 PFADD ( MM4, MM3 ) /* r1 | r0 */
263 PFMUL ( REGOFF(40, ECX), MM1 ) /* | x2*m10 */
264 PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m14 | x2*m10 */
268 MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
269 PFACC ( MM1, MM1 ) /* | r2 */
271 MOVD ( MM1, REGOFF(-8, EDX) ) /* write r2 */
273 DEC_L ( ESI ) /* decrement vertex counter */
274 JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */
287 GLOBL GLNAME( _mesa_3dnow_transform_points3_3d_no_rot )
288 GLNAME( _mesa_3dnow_transform_points3_3d_no_rot ):
292 MOV_L ( ARG_DEST, ECX )
293 MOV_L ( ARG_MATRIX, ESI )
294 MOV_L ( ARG_SOURCE, EAX )
295 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
296 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
297 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
298 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
302 MOV_L ( REGOFF(V4F_START, ECX), EDX )
304 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
305 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
306 MOV_L ( REGOFF(V4F_START, EAX), EAX )
309 JZ ( LLBL( G3TP3NRR_2 ) )
311 PREFETCH ( REGIND(EAX) )
312 PREFETCHW ( REGIND(EDX) )
314 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
315 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
317 MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
318 PUNPCKLDQ ( MM2, MM2 ) /* m22 | m22 */
320 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
321 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
323 PUNPCKLDQ ( MM3, MM3 ) /* m32 | m32 */
329 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
331 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
332 MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
334 ADD_L ( EDI, EAX ) /* next vertex */
335 PREFETCHW ( REGIND(EAX) )
337 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
339 PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
340 PFMUL ( MM2, MM5 ) /* | x2*m22 */
342 PFADD ( MM3, MM5 ) /* | x2*m22+m32 */
343 MOVQ ( MM4, REGIND(EDX) ) /* write r0, r1 */
345 ADD_L ( CONST(16), EDX ) /* next r */
346 DEC_L ( ESI ) /* decrement vertex counter */
348 MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 */
349 JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
362 GLOBL GLNAME( _mesa_3dnow_transform_points3_2d )
363 GLNAME( _mesa_3dnow_transform_points3_2d ):
367 MOV_L ( ARG_DEST, ECX )
368 MOV_L ( ARG_MATRIX, ESI )
369 MOV_L ( ARG_SOURCE, EAX )
370 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
371 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
372 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
373 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
377 MOV_L ( REGOFF(V4F_START, ECX), EDX )
379 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
380 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
381 MOV_L ( REGOFF(V4F_START, EAX), EAX )
384 JZ ( LLBL( G3TP2R_3) )
386 PREFETCH ( REGIND(EAX) )
387 PREFETCHW ( REGIND(EDX) )
389 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
390 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
392 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
393 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
395 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
400 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
402 MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */
403 MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
405 ADD_L ( EDI, EAX ) /* next vertex */
406 PREFETCH ( REGIND(EAX) )
408 MOVQ ( MM3, MM4 ) /* x1 | x0 */
409 PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */
411 ADD_L ( CONST(16), EDX ) /* next r */
412 PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */
414 PFACC ( MM4, MM3 ) /* x0*m00+x1*m10 | x0*m01+x1*m11 */
415 MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */
417 PFADD ( MM2, MM3 ) /* x0*...*m10+m30 | x0*...*m11+m31 */
418 MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
420 DEC_L ( ESI ) /* decrement vertex counter */
421 JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
434 GLOBL GLNAME( _mesa_3dnow_transform_points3_2d_no_rot )
435 GLNAME( _mesa_3dnow_transform_points3_2d_no_rot ):
439 MOV_L ( ARG_DEST, ECX )
440 MOV_L ( ARG_MATRIX, ESI )
441 MOV_L ( ARG_SOURCE, EAX )
442 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
443 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
444 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
445 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
449 MOV_L ( REGOFF(V4F_START, ECX), EDX )
451 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
452 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
453 MOV_L ( REGOFF(V4F_START, EAX), EAX )
456 JZ ( LLBL( G3TP2NRR_2 ) )
458 PREFETCH ( REGIND(EAX) )
459 PREFETCHW ( REGIND(EDX) )
461 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
462 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
464 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
470 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
472 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
473 MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
475 ADD_L ( EDI, EAX ) /* next vertex */
476 PREFETCH ( REGIND(EAX) )
478 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
479 ADD_L ( CONST(16), EDX ) /* next r */
481 PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
483 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
484 MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */
486 DEC_L ( ESI ) /* decrement vertex counter */
487 JNZ ( LLBL( G3TP2NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
500 GLOBL GLNAME( _mesa_3dnow_transform_points3_identity )
501 GLNAME( _mesa_3dnow_transform_points3_identity ):
505 MOV_L ( ARG_DEST, ECX )
506 MOV_L ( ARG_MATRIX, ESI )
507 MOV_L ( ARG_SOURCE, EAX )
508 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
509 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
510 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
511 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
515 MOV_L ( REGOFF(V4F_START, ECX), EDX )
517 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
518 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
519 MOV_L ( REGOFF(V4F_START, EAX), EAX )
522 JZ ( LLBL( G3TPIR_2 ) )
524 PREFETCHW ( REGIND(EDX) )
529 PREFETCHW ( REGOFF(32, EDX) )
531 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
532 MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */
534 ADD_L ( EDI, EAX ) /* next vertex */
535 ADD_L ( CONST(16), EDX ) /* next r */
537 DEC_L ( ESI ) /* decrement vertex counter */
538 MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */
540 MOVD ( MM1, REGOFF(-8, EDX) ) /* | r2 */
541 JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */