1 /* $Id: 3dnow_xform4.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_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) )
65 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
67 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
68 MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */
70 ADD_L ( EDI, EAX ) /* next vertex */
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) )
114 DEC_L ( ESI ) /* decrement vertex counter */
115 JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */
128 GLOBL GLNAME( _mesa_3dnow_transform_points4_perspective )
129 GLNAME( _mesa_3dnow_transform_points4_perspective ):
133 MOV_L ( ARG_DEST, ECX )
134 MOV_L ( ARG_MATRIX, ESI )
135 MOV_L ( ARG_SOURCE, EAX )
136 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
137 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
138 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
139 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
143 MOV_L ( REGOFF(V4F_START, ECX), EDX )
145 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
146 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
147 MOV_L ( REGOFF(V4F_START, EAX), EAX )
150 JZ ( LLBL( G3TPPR_2 ) )
152 PREFETCH ( REGIND(EAX) )
153 PREFETCHW ( REGIND(EDX) )
155 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
156 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
158 MOVD ( REGOFF(40, ECX), MM1 ) /* | m22 */
159 PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m32 | m22 */
161 MOVQ ( REGOFF(32, ECX), MM2 ) /* m21 | m20 */
162 PXOR ( MM7, MM7 ) /* 0 | 0 */
167 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
169 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
170 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
171 MOVD ( REGOFF(8, EAX), MM3 ) /* | x2 */
173 ADD_L ( EDI, EAX ) /* next vertex */
174 PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */
176 MOVQ ( MM5, MM6 ) /* x3 | x2 */
177 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
179 PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */
180 ADD_L ( CONST(16), EDX ) /* next r */
182 PFMUL ( MM2, MM5 ) /* x2*m21 | x2*m20 */
183 PFSUBR ( MM7, MM3 ) /* | -x2 */
185 PFMUL ( MM1, MM6 ) /* x3*m32 | x2*m22 */
186 PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */
188 PFACC ( MM3, MM6 ) /* -x2 | x2*m22+x3*m32 */
189 MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */
191 MOVQ ( MM6, REGOFF(-8, EDX) ) /* write r2, r3 */
192 DEC_L ( ESI ) /* decrement vertex counter */
194 JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */
207 GLOBL GLNAME( _mesa_3dnow_transform_points4_3d )
208 GLNAME( _mesa_3dnow_transform_points4_3d ):
212 MOV_L ( ARG_DEST, ECX )
213 MOV_L ( ARG_MATRIX, ESI )
214 MOV_L ( ARG_SOURCE, EAX )
215 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
216 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
217 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
218 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
222 MOV_L ( REGOFF(V4F_START, ECX), EDX )
224 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
225 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
226 MOV_L ( REGOFF(V4F_START, EAX), EAX )
229 JZ ( LLBL( G3TP3R_2 ) )
231 MOVD ( REGOFF(8, ECX), MM6 ) /* | m2 */
232 PUNPCKLDQ ( REGOFF(24, ECX), MM6 ) /* m6 | m2 */
234 MOVD ( REGOFF(40, ECX), MM7 ) /* | m10 */
235 PUNPCKLDQ ( REGOFF(56, ECX), MM7 ) /* m14 | m10 */
240 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
241 PREFETCH ( REGOFF(32, EAX) ) /* hopefully array is tightly packed */
243 MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */
244 MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */
246 MOVQ ( MM2, MM0 ) /* x1 | x0 */
247 MOVQ ( MM3, MM4 ) /* x3 | x2 */
249 MOVQ ( MM0, MM1 ) /* x1 | x0 */
250 MOVQ ( MM4, MM5 ) /* x3 | x2 */
252 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
253 PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */
255 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
256 PUNPCKLDQ ( MM3, MM3 ) /* x2 | x2 */
258 PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */
259 PUNPCKHDQ ( MM4, MM4 ) /* x3 | x3 */
261 PFMUL ( MM6, MM2 ) /* x1*m6 | x0*m2 */
262 PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */
264 PFMUL ( REGOFF(32, ECX), MM3 ) /* x2*m9 | x2*m8 */
265 ADD_L ( CONST(16), EDX ) /* next r */
267 PFMUL ( REGOFF(48, ECX), MM4 ) /* x3*m13 | x3*m12 */
268 PFADD ( MM1, MM3 ) /* x0*m1+..+x2*m9 | x0*m0+...+x2*m8 */
270 PFMUL ( MM7, MM5 ) /* x3*m14 | x2*m10 */
271 PFADD ( MM3, MM4 ) /* r1 | r0 */
273 PFACC ( MM2, MM5 ) /* x0*m2+x1*m6 | x2*m10+x3*m14 */
274 MOVD ( REGOFF(12, EAX), MM0 ) /* | x3 */
276 ADD_L ( EDI, EAX ) /* next vertex */
277 PFACC ( MM0, MM5 ) /* r3 | r2 */
279 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
280 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
282 DEC_L ( ESI ) /* decrement vertex counter */
283 JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */
296 GLOBL GLNAME( _mesa_3dnow_transform_points4_3d_no_rot )
297 GLNAME( _mesa_3dnow_transform_points4_3d_no_rot ):
300 MOV_L ( ARG_DEST, ECX )
301 MOV_L ( ARG_MATRIX, ESI )
302 MOV_L ( ARG_SOURCE, EAX )
303 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
304 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
305 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
306 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
310 MOV_L ( REGOFF(V4F_START, ECX), EDX )
312 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
313 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
314 MOV_L ( REGOFF(V4F_START, EAX), EAX )
317 JZ ( LLBL( G3TP3NRR_2 ) )
319 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
320 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
322 MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
323 PUNPCKLDQ ( REGOFF(56, ECX), MM2 ) /* m32 | m22 */
325 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
330 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
332 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
333 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
334 MOVD ( REGOFF(12, EAX), MM7 ) /* | x3 */
336 ADD_L ( EDI, EAX ) /* next vertex */
337 PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */
339 MOVQ ( MM5, MM6 ) /* x3 | x2 */
340 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
342 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
343 PFMUL ( MM2, MM5 ) /* x3*m32 | x2*m22 */
345 PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */
346 PFACC ( MM7, MM5 ) /* x3 | x2*m22+x3*m32 */
348 PFADD ( MM6, MM4 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */
349 ADD_L ( CONST(16), EDX ) /* next r */
351 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
352 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
354 DEC_L ( ESI ) /* decrement vertex counter */
355 JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
368 GLOBL GLNAME( _mesa_3dnow_transform_points4_2d )
369 GLNAME( _mesa_3dnow_transform_points4_2d ):
373 MOV_L ( ARG_DEST, ECX )
374 MOV_L ( ARG_MATRIX, ESI )
375 MOV_L ( ARG_SOURCE, EAX )
376 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
377 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
378 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
379 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
383 MOV_L ( REGOFF(V4F_START, ECX), EDX )
385 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
386 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
387 MOV_L ( REGOFF(V4F_START, EAX), EAX )
390 JZ ( LLBL( G3TP2R_2 ) )
392 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
393 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
395 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
396 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
398 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
403 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
405 MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */
406 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
408 ADD_L ( EDI, EAX ) /* next vertex */
409 PREFETCH ( REGIND(EAX) )
411 MOVQ ( MM3, MM4 ) /* x1 | x0 */
412 MOVQ ( MM5, MM6 ) /* x3 | x2 */
414 PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */
415 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
417 PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */
418 ADD_L ( CONST(16), EDX ) /* next r */
420 PFACC ( MM4, MM3 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */
421 PFMUL ( MM2, MM6 ) /* x3*m31 | x3*m30 */
423 PFADD ( MM6, MM3 ) /* r1 | r0 */
424 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
426 MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
428 DEC_L ( ESI ) /* decrement vertex counter */
429 JNZ ( LLBL( G3TP2R_1 ) ) /* cnt > 0 ? -> process next vertex */
442 GLOBL GLNAME( _mesa_3dnow_transform_points4_2d_no_rot )
443 GLNAME( _mesa_3dnow_transform_points4_2d_no_rot ):
447 MOV_L ( ARG_DEST, ECX )
448 MOV_L ( ARG_MATRIX, ESI )
449 MOV_L ( ARG_SOURCE, EAX )
450 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
451 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
452 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
453 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
457 MOV_L ( REGOFF(V4F_START, ECX), EDX )
459 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
460 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
461 MOV_L ( REGOFF(V4F_START, EAX), EAX )
464 JZ ( LLBL( G3TP2NRR_3 ) )
466 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
467 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
469 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
474 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
476 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
477 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
479 ADD_L ( EDI, EAX ) /* next vertex */
480 PREFETCH ( REGIND(EAX) )
482 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
483 MOVQ ( MM5, MM6 ) /* x3 | x2 */
485 ADD_L ( CONST(16), EDX ) /* next r */
486 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
488 PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */
489 PFADD ( MM4, MM6 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */
491 MOVQ ( MM6, REGOFF(-16, EDX) ) /* write r0, r1 */
492 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
494 DEC_L ( ESI ) /* decrement vertex counter */
496 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
509 GLOBL GLNAME( _mesa_3dnow_transform_points4_identity )
510 GLNAME( _mesa_3dnow_transform_points4_identity ):
514 MOV_L ( ARG_DEST, ECX )
515 MOV_L ( ARG_MATRIX, ESI )
516 MOV_L ( ARG_SOURCE, EAX )
517 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
518 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
519 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
520 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
524 MOV_L ( REGOFF(V4F_START, ECX), EDX )
526 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
527 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
528 MOV_L ( REGOFF(V4F_START, EAX), EAX )
531 JZ ( LLBL( G3TPIR_2 ) )
536 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
538 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
539 MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */
541 ADD_L ( EDI, EAX ) /* next vertex */
542 PREFETCH ( REGIND(EAX) )
544 ADD_L ( CONST(16), EDX ) /* next r */
545 MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */
547 MOVQ ( MM1, REGOFF(-8, EDX) ) /* r3 | r2 */
549 DEC_L ( ESI ) /* decrement vertex counter */
550 JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */