move mesa32 over to new dir
[reactos.git] / reactos / lib / mesa32 / src / x86 / 3dnow_xform2.S
1 /* $Id: 3dnow_xform2.S,v 1.3 2005/01/04 14:33:47 brianp Exp $ */
2
3 /*
4 * Mesa 3-D graphics library
5 * Version: 3.5
6 *
7 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
8 *
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:
15 *
16 * The above copyright notice and this permission notice shall be included
17 * in all copies or substantial portions of the Software.
18 *
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.
25 */
26
27 #ifdef USE_3DNOW_ASM
28 #include "matypes.h"
29 #include "xform_args.h"
30
31 SEG_TEXT
32
33 #define FRAME_OFFSET 4
34
35
36 ALIGNTEXT16
37 GLOBL GLNAME( _mesa_3dnow_transform_points2_general )
38 HIDDEN(_mesa_3dnow_transform_points2_general)
39 GLNAME( _mesa_3dnow_transform_points2_general ):
40
41 PUSH_L ( ESI )
42
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) )
50
51 PUSH_L ( EDI )
52
53 MOV_L ( REGOFF(V4F_START, ECX), EDX )
54 MOV_L ( ESI, ECX )
55 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
56 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
57 MOV_L ( REGOFF(V4F_START, EAX), EAX )
58
59 TEST_L ( ESI, ESI )
60 JZ ( LLBL( G3TPGR_3 ) )
61
62 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
63 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
64
65 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
66 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
67
68 MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */
69 PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */
70
71 MOVD ( REGOFF(12, ECX), MM3 ) /* | m03 */
72 PUNPCKLDQ ( REGOFF(28, ECX), MM3 ) /* m13 | m03 */
73
74 MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */
75 MOVQ ( REGOFF(56, ECX), MM5 ) /* m33 | m32 */
76
77 ALIGNTEXT16
78 LLBL( G3TPGR_2 ):
79
80 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
81 MOVQ ( MM6, MM7 ) /* x1 | x0 */
82
83 PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */
84 PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */
85
86 PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */
87 PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */
88
89 MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */
90 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
91
92 MOVQ ( MM6, MM7 ) /* x1 | x0 */
93 PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */
94
95 PFMUL ( MM3, MM7 ) /* x1*m13 | x0*m03 */
96 ADD_L ( EDI, EAX ) /* next vertex */
97
98 PFACC ( MM7, MM6 ) /* x0*m03+x1*m13 | x0*x02+x1*m12 */
99 PFADD ( MM5, MM6 ) /* x0*...*m13+m33 | x0*...*m12+m32 */
100
101 MOVQ ( MM6, REGOFF(8, EDX) ) /* write r3, r2 */
102 ADD_L ( CONST(16), EDX ) /* next r */
103
104 DEC_L ( ESI ) /* decrement vertex counter */
105 JNZ ( LLBL( G3TPGR_2 ) ) /* cnt > 0 ? -> process next vertex */
106
107 LLBL( G3TPGR_3 ):
108
109 FEMMS
110 POP_L ( EDI )
111 POP_L ( ESI )
112 RET
113
114
115
116
117 ALIGNTEXT16
118 GLOBL GLNAME( _mesa_3dnow_transform_points2_perspective )
119 HIDDEN(_mesa_3dnow_transform_points2_perspective)
120 GLNAME( _mesa_3dnow_transform_points2_perspective ):
121
122 PUSH_L ( ESI )
123
124 MOV_L ( ARG_DEST, ECX )
125 MOV_L ( ARG_MATRIX, ESI )
126 MOV_L ( ARG_SOURCE, EAX )
127 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
128 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
129 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
130 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
131
132 PUSH_L ( EDI )
133
134 MOV_L ( REGOFF(V4F_START, ECX), EDX )
135 MOV_L ( ESI, ECX )
136 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
137 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
138 MOV_L ( REGOFF(V4F_START, EAX), EAX )
139
140 TEST_L ( ESI, ESI )
141 JZ ( LLBL( G3TPPR_3 ) )
142
143 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
144 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
145
146 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
147
148 ALIGNTEXT16
149 LLBL( G3TPPR_2 ):
150
151 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
152 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
153
154 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
155 MOVQ ( MM3, REGOFF(8, EDX) ) /* write r2 (=m32), r3 (=0) */
156
157 ADD_L ( EDI, EAX ) /* next vertex */
158 ADD_L ( CONST(16), EDX ) /* next r */
159
160 DEC_L ( ESI ) /* decrement vertex counter */
161 JNZ ( LLBL( G3TPPR_2 ) ) /* cnt > 0 ? -> process next vertex */
162
163 LLBL( G3TPPR_3 ):
164
165 FEMMS
166 POP_L ( EDI )
167 POP_L ( ESI )
168 RET
169
170
171
172
173 ALIGNTEXT16
174 GLOBL GLNAME( _mesa_3dnow_transform_points2_3d )
175 HIDDEN(_mesa_3dnow_transform_points2_3d)
176 GLNAME( _mesa_3dnow_transform_points2_3d ):
177
178 PUSH_L ( ESI )
179
180 MOV_L ( ARG_DEST, ECX )
181 MOV_L ( ARG_MATRIX, ESI )
182 MOV_L ( ARG_SOURCE, EAX )
183 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
184 OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) )
185 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
186 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
187
188 PUSH_L ( EDI )
189
190 MOV_L ( REGOFF(V4F_START, ECX), EDX )
191 MOV_L ( ESI, ECX )
192 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
193 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
194 MOV_L ( REGOFF(V4F_START, EAX), EAX )
195
196 TEST_L ( ESI, ESI )
197 JZ ( LLBL( G3TP3R_3 ) )
198
199 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
200 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
201
202 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
203 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
204
205 MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */
206 PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */
207
208 MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */
209 MOVD ( REGOFF(56, ECX), MM5 ) /* | m32 */
210
211 ALIGNTEXT16
212 LLBL( G3TP3R_2 ):
213
214 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
215 MOVQ ( MM6, MM7 ) /* x1 | x0 */
216
217 PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */
218 PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */
219
220 PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */
221 PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */
222
223 MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */
224 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
225
226 MOVQ ( MM6, MM7 ) /* x1 | x0 */
227 PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */
228
229 PFACC ( MM7, MM6 ) /* ***trash*** | x0*x02+x1*m12 */
230 PFADD ( MM5, MM6 ) /* ***trash*** | x0*...*m12+m32 */
231
232 MOVD ( MM6, REGOFF(8, EDX) ) /* write r2 */
233 ADD_L ( EDI, EAX ) /* next vertex */
234
235 ADD_L ( CONST(16), EDX ) /* next r */
236 DEC_L ( ESI ) /* decrement vertex counter */
237
238 JNZ ( LLBL( G3TP3R_2 ) ) /* cnt > 0 ? -> process next vertex */
239
240 LLBL( G3TP3R_3 ):
241
242 FEMMS
243 POP_L ( EDI )
244 POP_L ( ESI )
245 RET
246
247
248
249
250 ALIGNTEXT16
251 GLOBL GLNAME( _mesa_3dnow_transform_points2_3d_no_rot )
252 HIDDEN(_mesa_3dnow_transform_points2_3d_no_rot)
253 GLNAME( _mesa_3dnow_transform_points2_3d_no_rot ):
254
255 PUSH_L ( ESI )
256
257 MOV_L ( ARG_DEST, ECX )
258 MOV_L ( ARG_MATRIX, ESI )
259 MOV_L ( ARG_SOURCE, EAX )
260 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
261 OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) )
262 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
263 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
264
265 PUSH_L ( EDI )
266
267 MOV_L ( REGOFF(V4F_START, ECX), EDX )
268 MOV_L ( ESI, ECX )
269 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
270 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
271 MOV_L ( REGOFF(V4F_START, EAX), EAX )
272
273 TEST_L ( ESI, ESI )
274 JZ ( LLBL( G3TP3NRR_3 ) )
275
276 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
277 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
278
279 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
280 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
281
282 ALIGNTEXT16
283 LLBL( G3TP3NRR_2 ):
284
285 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
286 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
287
288 PFADD ( MM2, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
289 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
290
291 MOVD ( MM3, REGOFF(8, EDX) ) /* write r2 */
292 ADD_L ( EDI, EAX ) /* next vertex */
293
294 ADD_L ( CONST(16), EDX ) /* next r */
295 DEC_L ( ESI ) /* decrement vertex counter */
296
297 JNZ ( LLBL( G3TP3NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
298
299 LLBL( G3TP3NRR_3 ):
300
301 FEMMS
302 POP_L ( EDI )
303 POP_L ( ESI )
304 RET
305
306
307
308
309 ALIGNTEXT16
310 GLOBL GLNAME( _mesa_3dnow_transform_points2_2d )
311 HIDDEN(_mesa_3dnow_transform_points2_2d)
312 GLNAME( _mesa_3dnow_transform_points2_2d ):
313
314 PUSH_L ( ESI )
315
316 MOV_L ( ARG_DEST, ECX )
317 MOV_L ( ARG_MATRIX, ESI )
318 MOV_L ( ARG_SOURCE, EAX )
319 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
320 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
321 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
322 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
323
324 PUSH_L ( EDI )
325
326 MOV_L ( REGOFF(V4F_START, ECX), EDX )
327 MOV_L ( ESI, ECX )
328 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
329 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
330 MOV_L ( REGOFF(V4F_START, EAX), EAX )
331
332 TEST_L ( ESI, ESI )
333 JZ ( LLBL( G3TP2R_3 ) )
334
335 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
336 MOVQ ( REGOFF(16, ECX), MM1 ) /* m11 | m10 */
337
338 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
339
340 ALIGNTEXT16
341 LLBL( G3TP2R_2 ):
342
343 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
344 MOVD ( REGOFF(4, EAX), MM5 ) /* | x1 */
345
346 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
347 ADD_L ( EDI, EAX ) /* next vertex */
348
349 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
350 PUNPCKLDQ ( MM5, MM5 ) /* x1 | x1 */
351
352 PFMUL ( MM1, MM5 ) /* x1*m11 | x1*m10 */
353 PFADD ( MM2, MM4 ) /* x...x1*m11+31 | x0*..*m10+m30 */
354
355 PFADD ( MM5, MM4 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */
356 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
357
358 ADD_L ( CONST(16), EDX ) /* next r */
359 DEC_L ( ESI ) /* decrement vertex counter */
360
361 JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
362
363 LLBL( G3TP2R_3 ):
364
365 FEMMS
366 POP_L ( EDI )
367 POP_L ( ESI )
368 RET
369
370
371
372
373 ALIGNTEXT16
374 GLOBL GLNAME( _mesa_3dnow_transform_points2_2d_no_rot )
375 HIDDEN(_mesa_3dnow_transform_points2_2d_no_rot)
376 GLNAME( _mesa_3dnow_transform_points2_2d_no_rot ):
377
378 PUSH_L ( ESI )
379
380 MOV_L ( ARG_DEST, ECX )
381 MOV_L ( ARG_MATRIX, ESI )
382 MOV_L ( ARG_SOURCE, EAX )
383 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
384 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
385 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
386 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
387
388 PUSH_L ( EDI )
389
390 MOV_L ( REGOFF(V4F_START, ECX), EDX )
391 MOV_L ( ESI, ECX )
392 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
393 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
394 MOV_L ( REGOFF(V4F_START, EAX), EAX )
395
396 TEST_L ( ESI, ESI )
397 JZ ( LLBL( G3TP2NRR_3 ) )
398
399 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
400 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
401
402 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
403
404 ALIGNTEXT16
405 LLBL( G3TP2NRR_2 ):
406
407 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
408 ADD_L ( EDI, EAX ) /* next vertex */
409
410 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
411 PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */
412
413 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
414 ADD_L ( CONST(16), EDX ) /* next r */
415
416 DEC_L ( ESI ) /* decrement vertex counter */
417 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
418
419 LLBL( G3TP2NRR_3 ):
420
421 FEMMS
422 POP_L ( EDI )
423 POP_L ( ESI )
424 RET
425
426
427
428
429 ALIGNTEXT16
430 GLOBL GLNAME( _mesa_3dnow_transform_points2_identity )
431 HIDDEN(_mesa_3dnow_transform_points2_identity)
432 GLNAME( _mesa_3dnow_transform_points2_identity ):
433
434 PUSH_L ( ESI )
435
436 MOV_L ( ARG_DEST, ECX )
437 MOV_L ( ARG_MATRIX, ESI )
438 MOV_L ( ARG_SOURCE, EAX )
439 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
440 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
441 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
442 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
443
444 PUSH_L ( EDI )
445
446 MOV_L ( REGOFF(V4F_START, ECX), EDX )
447 MOV_L ( ESI, ECX )
448 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
449 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
450 MOV_L ( REGOFF(V4F_START, EAX), EAX )
451
452 TEST_L ( ESI, ESI )
453 JZ ( LLBL( G3TPIR_3 ) )
454
455 ALIGNTEXT16
456 LLBL( G3TPIR_3 ):
457
458 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
459 ADD_L ( EDI, EAX ) /* next vertex */
460
461 MOVQ ( MM0, REGIND(EDX) ) /* r1 | r0 */
462 ADD_L ( CONST(16), EDX ) /* next r */
463
464 DEC_L ( ESI ) /* decrement vertex counter */
465 JNZ ( LLBL( G3TPIR_3 ) ) /* cnt > 0 ? -> process next vertex */
466
467 LLBL( G3TPIR_4 ):
468
469 FEMMS
470 POP_L ( EDI )
471 POP_L ( ESI )
472 RET
473 #endif