1 /* $Id: 3dnow_xform3.S,v 1.3 2004/04/26 10:10:25 alanh 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.
29 #include "xform_args.h"
33 #define FRAME_OFFSET 4
37 GLOBL GLNAME( _mesa_3dnow_transform_points3_general )
38 GLNAME( _mesa_3dnow_transform_points3_general ):
42 MOV_L ( ARG_DEST, ECX )
43 MOV_L ( ARG_MATRIX, ESI )
44 MOV_L ( ARG_SOURCE, EAX )
45 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
46 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
47 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
48 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
52 MOV_L ( REGOFF(V4F_START, ECX), EDX )
54 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
55 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
56 MOV_L ( REGOFF(V4F_START, EAX), EAX )
59 JZ ( LLBL( G3TPGR_2 ) )
61 PREFETCHW ( REGIND(EDX) )
66 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
68 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
69 MOVD ( REGOFF(8, EAX), MM2 ) /* | x2 */
71 ADD_L ( EDI, EAX ) /* next vertex */
72 PREFETCH ( REGIND(EAX) )
74 MOVQ ( MM0, MM1 ) /* x1 | x0 */
75 PUNPCKLDQ ( MM2, MM2 ) /* x2 | x2 */
77 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
78 MOVQ ( MM2, MM5 ) /* x2 | x2 */
80 PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */
81 PFMUL ( REGOFF(32, ECX), MM2 ) /* x2*m9 | x2*m8 */
83 MOVQ ( MM0, MM3 ) /* x0 | x0 */
84 PFMUL ( REGOFF(40, ECX), MM5 ) /* x2*m11 | x2*m10 */
86 MOVQ ( MM1, MM4 ) /* x1 | x1 */
87 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
89 PFADD ( REGOFF(48, ECX), MM2 ) /* x2*m9+m13 | x2*m8+m12 */
90 PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */
92 PFADD ( REGOFF(56, ECX), MM5 ) /* x2*m11+m15 | x2*m10+m14 */
93 PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */
95 PFMUL ( REGOFF(8, ECX), MM3 ) /* x0*m3 | x0*m2 */
96 PFADD ( MM1, MM2 ) /* r1 | r0 */
98 PFMUL ( REGOFF(24, ECX), MM4 ) /* x1*m7 | x1*m6 */
99 ADD_L ( CONST(16), EDX ) /* next output vertex */
101 PFADD ( MM3, MM4 ) /* x0*m3+x1*m7 | x0*m2+x1*m6 */
102 MOVQ ( MM2, REGOFF(-16, EDX) ) /* write r0, r1 */
104 PFADD ( MM4, MM5 ) /* r3 | r2 */
105 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
107 DEC_L ( ESI ) /* decrement vertex counter */
108 JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */
121 GLOBL GLNAME( _mesa_3dnow_transform_points3_perspective )
122 GLNAME( _mesa_3dnow_transform_points3_perspective ):
126 MOV_L ( ARG_DEST, ECX )
127 MOV_L ( ARG_MATRIX, ESI )
128 MOV_L ( ARG_SOURCE, EAX )
129 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
130 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
131 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
132 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
136 MOV_L ( REGOFF(V4F_START, ECX), EDX )
138 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
139 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
140 MOV_L ( REGOFF(V4F_START, EAX), EAX )
143 JZ ( LLBL( G3TPPR_2 ) )
145 PREFETCH ( REGIND(EAX) )
146 PREFETCHW ( REGIND(EDX) )
148 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
149 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
151 MOVQ ( REGOFF(32, ECX), MM1 ) /* m21 | m20 */
152 MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
154 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
159 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
161 MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
162 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
164 ADD_L ( EDI, EAX ) /* next vertex */
165 PREFETCH ( REGIND(EAX) )
167 PXOR ( MM7, MM7 ) /* 0 | 0 */
168 MOVQ ( MM5, MM6 ) /* | x2 */
170 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
171 PFSUB ( MM5, MM7 ) /* | -x2 */
173 PFMUL ( MM2, MM6 ) /* | x2*m22 */
174 PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */
176 ADD_L ( CONST(16), EDX ) /* next r */
177 PFMUL ( MM1, MM5 ) /* x2*m21 | x2*m20 */
179 PFADD ( MM3, MM6 ) /* | x2*m22+m32 */
180 PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */
182 MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */
183 MOVD ( MM6, REGOFF(-8, EDX) ) /* write r2 */
185 MOVD ( MM7, REGOFF(-4, EDX) ) /* write r3 */
187 DEC_L ( ESI ) /* decrement vertex counter */
188 JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */
201 GLOBL GLNAME( _mesa_3dnow_transform_points3_3d )
202 GLNAME( _mesa_3dnow_transform_points3_3d ):
206 MOV_L ( ARG_DEST, ECX )
207 MOV_L ( ARG_MATRIX, ESI )
208 MOV_L ( ARG_SOURCE, EAX )
209 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
210 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
211 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
212 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
216 MOV_L ( REGOFF(V4F_START, ECX), EDX )
218 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
219 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
220 MOV_L ( REGOFF(V4F_START, EAX), EAX )
223 JZ ( LLBL( G3TP3R_2 ) )
225 PREFETCH ( REGIND(EAX) )
226 PREFETCH ( REGIND(EDX) )
228 MOVD ( REGOFF(8, ECX), MM7 ) /* | m2 */
229 PUNPCKLDQ ( REGOFF(24, ECX), MM7 ) /* m6 | m2 */
235 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
237 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
238 MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */
240 ADD_L ( EDI, EAX ) /* next vertex */
241 PREFETCH ( REGIND(EAX) )
243 MOVQ ( MM0, MM2 ) /* x1 | x0 */
244 ADD_L ( CONST(16), EDX ) /* next r */
246 PUNPCKLDQ ( MM2, MM2 ) /* x0 | x0 */
247 MOVQ ( MM0, MM3 ) /* x1 | x0 */
249 PFMUL ( REGIND(ECX), MM2 ) /* x0*m1 | x0*m0 */
250 PUNPCKHDQ ( MM3, MM3 ) /* x1 | x1 */
252 MOVQ ( MM1, MM4 ) /* | x2 */
253 PFMUL ( REGOFF(16, ECX), MM3 ) /* x1*m5 | x1*m4 */
255 PUNPCKLDQ ( MM4, MM4 ) /* x2 | x2 */
256 PFADD ( MM2, MM3 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */
258 PFMUL ( REGOFF(32, ECX), MM4 ) /* x2*m9 | x2*m8 */
259 PFADD ( REGOFF(48, ECX), MM3 ) /* x0*m1+...+m11 | x0*m0+x1*m4+m12 */
261 PFMUL ( MM7, MM0 ) /* x1*m6 | x0*m2 */
262 PFADD ( MM4, MM3 ) /* r1 | r0 */
264 PFMUL ( REGOFF(40, ECX), MM1 ) /* | x2*m10 */
265 PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m14 | x2*m10 */
269 MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
270 PFACC ( MM1, MM1 ) /* | r2 */
272 MOVD ( MM1, REGOFF(-8, EDX) ) /* write r2 */
274 DEC_L ( ESI ) /* decrement vertex counter */
275 JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */
288 GLOBL GLNAME( _mesa_3dnow_transform_points3_3d_no_rot )
289 GLNAME( _mesa_3dnow_transform_points3_3d_no_rot ):
293 MOV_L ( ARG_DEST, ECX )
294 MOV_L ( ARG_MATRIX, ESI )
295 MOV_L ( ARG_SOURCE, EAX )
296 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
297 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
298 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
299 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
303 MOV_L ( REGOFF(V4F_START, ECX), EDX )
305 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
306 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
307 MOV_L ( REGOFF(V4F_START, EAX), EAX )
310 JZ ( LLBL( G3TP3NRR_2 ) )
312 PREFETCH ( REGIND(EAX) )
313 PREFETCHW ( REGIND(EDX) )
315 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
316 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
318 MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
319 PUNPCKLDQ ( MM2, MM2 ) /* m22 | m22 */
321 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
322 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
324 PUNPCKLDQ ( MM3, MM3 ) /* m32 | m32 */
330 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
332 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
333 MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
335 ADD_L ( EDI, EAX ) /* next vertex */
336 PREFETCHW ( REGIND(EAX) )
338 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
340 PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
341 PFMUL ( MM2, MM5 ) /* | x2*m22 */
343 PFADD ( MM3, MM5 ) /* | x2*m22+m32 */
344 MOVQ ( MM4, REGIND(EDX) ) /* write r0, r1 */
346 ADD_L ( CONST(16), EDX ) /* next r */
347 DEC_L ( ESI ) /* decrement vertex counter */
349 MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 */
350 JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
363 GLOBL GLNAME( _mesa_3dnow_transform_points3_2d )
364 GLNAME( _mesa_3dnow_transform_points3_2d ):
368 MOV_L ( ARG_DEST, ECX )
369 MOV_L ( ARG_MATRIX, ESI )
370 MOV_L ( ARG_SOURCE, EAX )
371 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
372 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
373 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
374 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
378 MOV_L ( REGOFF(V4F_START, ECX), EDX )
380 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
381 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
382 MOV_L ( REGOFF(V4F_START, EAX), EAX )
385 JZ ( LLBL( G3TP2R_3) )
387 PREFETCH ( REGIND(EAX) )
388 PREFETCHW ( REGIND(EDX) )
390 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
391 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
393 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
394 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
396 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
401 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
403 MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */
404 MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
406 ADD_L ( EDI, EAX ) /* next vertex */
407 PREFETCH ( REGIND(EAX) )
409 MOVQ ( MM3, MM4 ) /* x1 | x0 */
410 PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */
412 ADD_L ( CONST(16), EDX ) /* next r */
413 PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */
415 PFACC ( MM4, MM3 ) /* x0*m00+x1*m10 | x0*m01+x1*m11 */
416 MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */
418 PFADD ( MM2, MM3 ) /* x0*...*m10+m30 | x0*...*m11+m31 */
419 MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
421 DEC_L ( ESI ) /* decrement vertex counter */
422 JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
435 GLOBL GLNAME( _mesa_3dnow_transform_points3_2d_no_rot )
436 GLNAME( _mesa_3dnow_transform_points3_2d_no_rot ):
440 MOV_L ( ARG_DEST, ECX )
441 MOV_L ( ARG_MATRIX, ESI )
442 MOV_L ( ARG_SOURCE, EAX )
443 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
444 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
445 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
446 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
450 MOV_L ( REGOFF(V4F_START, ECX), EDX )
452 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
453 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
454 MOV_L ( REGOFF(V4F_START, EAX), EAX )
457 JZ ( LLBL( G3TP2NRR_2 ) )
459 PREFETCH ( REGIND(EAX) )
460 PREFETCHW ( REGIND(EDX) )
462 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
463 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
465 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
471 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
473 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
474 MOVD ( REGOFF(8, EAX), MM5 ) /* | x2 */
476 ADD_L ( EDI, EAX ) /* next vertex */
477 PREFETCH ( REGIND(EAX) )
479 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
480 ADD_L ( CONST(16), EDX ) /* next r */
482 PFADD ( MM1, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
484 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
485 MOVD ( MM5, REGOFF(-8, EDX) ) /* write r2 (=x2) */
487 DEC_L ( ESI ) /* decrement vertex counter */
488 JNZ ( LLBL( G3TP2NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
501 GLOBL GLNAME( _mesa_3dnow_transform_points3_identity )
502 GLNAME( _mesa_3dnow_transform_points3_identity ):
506 MOV_L ( ARG_DEST, ECX )
507 MOV_L ( ARG_MATRIX, ESI )
508 MOV_L ( ARG_SOURCE, EAX )
509 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
510 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
511 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
512 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
516 MOV_L ( REGOFF(V4F_START, ECX), EDX )
518 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
519 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
520 MOV_L ( REGOFF(V4F_START, EAX), EAX )
523 JZ ( LLBL( G3TPIR_2 ) )
525 PREFETCHW ( REGIND(EDX) )
530 PREFETCHW ( REGOFF(32, EDX) )
532 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
533 MOVD ( REGOFF(8, EAX), MM1 ) /* | x2 */
535 ADD_L ( EDI, EAX ) /* next vertex */
536 ADD_L ( CONST(16), EDX ) /* next r */
538 DEC_L ( ESI ) /* decrement vertex counter */
539 MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */
541 MOVD ( MM1, REGOFF(-8, EDX) ) /* | r2 */
542 JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */