cb3a8531df7696fbaf00bbcecb47eaea9221c35b
[mesa.git] / src / mesa / x86 / 3dnow_xform2.S
1 /* $Id: 3dnow_xform2.S,v 1.1 2001/03/29 06:46:16 gareth 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 #include "matypes.h"
28 #include "xform_args.h"
29
30 SEG_TEXT
31
32 #define FRAME_OFFSET 4
33
34
35 ALIGNTEXT16
36 GLOBL GLNAME( _mesa_3dnow_transform_points2_general )
37 GLNAME( _mesa_3dnow_transform_points2_general ):
38
39 PUSH_L ( ESI )
40
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) )
48
49 PUSH_L ( EDI )
50
51 MOV_L ( REGOFF(V4F_START, ECX), EDX )
52 MOV_L ( ESI, ECX )
53 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
54 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
55 MOV_L ( REGOFF(V4F_START, EAX), EAX )
56
57 TEST_L ( ESI, ESI )
58 JZ ( LLBL( G3TPGR_3 ) )
59
60 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
61 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
62
63 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
64 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
65
66 MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */
67 PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */
68
69 MOVD ( REGOFF(12, ECX), MM3 ) /* | m03 */
70 PUNPCKLDQ ( REGOFF(28, ECX), MM3 ) /* m13 | m03 */
71
72 MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */
73 MOVQ ( REGOFF(56, ECX), MM5 ) /* m33 | m32 */
74
75 ALIGNTEXT16
76 LLBL( G3TPGR_2 ):
77
78 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
79 MOVQ ( MM6, MM7 ) /* x1 | x0 */
80
81 PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */
82 PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */
83
84 PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */
85 PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */
86
87 MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */
88 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
89
90 MOVQ ( MM6, MM7 ) /* x1 | x0 */
91 PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */
92
93 PFMUL ( MM3, MM7 ) /* x1*m13 | x0*m03 */
94 ADD_L ( EDI, EAX ) /* next vertex */
95
96 PFACC ( MM7, MM6 ) /* x0*m03+x1*m13 | x0*x02+x1*m12 */
97 PFADD ( MM5, MM6 ) /* x0*...*m13+m33 | x0*...*m12+m32 */
98
99 MOVQ ( MM6, REGOFF(8, EDX) ) /* write r3, r2 */
100 ADD_L ( CONST(16), EDX ) /* next r */
101
102 DEC_L ( ESI ) /* decrement vertex counter */
103 JNZ ( LLBL( G3TPGR_2 ) ) /* cnt > 0 ? -> process next vertex */
104
105 LLBL( G3TPGR_3 ):
106
107 FEMMS
108 POP_L ( EDI )
109 POP_L ( ESI )
110 RET
111
112
113
114
115 ALIGNTEXT16
116 GLOBL GLNAME( _mesa_3dnow_transform_points2_perspective )
117 GLNAME( _mesa_3dnow_transform_points2_perspective ):
118
119 PUSH_L ( ESI )
120
121 MOV_L ( ARG_DEST, ECX )
122 MOV_L ( ARG_MATRIX, ESI )
123 MOV_L ( ARG_SOURCE, EAX )
124 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
125 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
126 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
127 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
128
129 PUSH_L ( EDI )
130
131 MOV_L ( REGOFF(V4F_START, ECX), EDX )
132 MOV_L ( ESI, ECX )
133 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
134 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
135 MOV_L ( REGOFF(V4F_START, EAX), EAX )
136
137 TEST_L ( ESI, ESI )
138 JZ ( LLBL( G3TPPR_3 ) )
139
140 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
141 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
142
143 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
144
145 ALIGNTEXT16
146 LLBL( G3TPPR_2 ):
147
148 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
149 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
150
151 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
152 MOVQ ( MM3, REGOFF(8, EDX) ) /* write r2 (=m32), r3 (=0) */
153
154 ADD_L ( EDI, EAX ) /* next vertex */
155 ADD_L ( CONST(16), EDX ) /* next r */
156
157 DEC_L ( ESI ) /* decrement vertex counter */
158 JNZ ( LLBL( G3TPPR_2 ) ) /* cnt > 0 ? -> process next vertex */
159
160 LLBL( G3TPPR_3 ):
161
162 FEMMS
163 POP_L ( EDI )
164 POP_L ( ESI )
165 RET
166
167
168
169
170 ALIGNTEXT16
171 GLOBL GLNAME( _mesa_3dnow_transform_points2_3d )
172 GLNAME( _mesa_3dnow_transform_points2_3d ):
173
174 PUSH_L ( ESI )
175
176 MOV_L ( ARG_DEST, ECX )
177 MOV_L ( ARG_MATRIX, ESI )
178 MOV_L ( ARG_SOURCE, EAX )
179 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
180 OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) )
181 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
182 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
183
184 PUSH_L ( EDI )
185
186 MOV_L ( REGOFF(V4F_START, ECX), EDX )
187 MOV_L ( ESI, ECX )
188 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
189 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
190 MOV_L ( REGOFF(V4F_START, EAX), EAX )
191
192 TEST_L ( ESI, ESI )
193 JZ ( LLBL( G3TP3R_3 ) )
194
195 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
196 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
197
198 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
199 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
200
201 MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */
202 PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */
203
204 MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */
205 MOVD ( REGOFF(56, ECX), MM5 ) /* | m32 */
206
207 ALIGNTEXT16
208 LLBL( G3TP3R_2 ):
209
210 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
211 MOVQ ( MM6, MM7 ) /* x1 | x0 */
212
213 PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */
214 PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */
215
216 PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */
217 PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */
218
219 MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */
220 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
221
222 MOVQ ( MM6, MM7 ) /* x1 | x0 */
223 PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */
224
225 PFACC ( MM7, MM6 ) /* ***trash*** | x0*x02+x1*m12 */
226 PFADD ( MM5, MM6 ) /* ***trash*** | x0*...*m12+m32 */
227
228 MOVD ( MM6, REGOFF(8, EDX) ) /* write r2 */
229 ADD_L ( EDI, EAX ) /* next vertex */
230
231 ADD_L ( CONST(16), EDX ) /* next r */
232 DEC_L ( ESI ) /* decrement vertex counter */
233
234 JNZ ( LLBL( G3TP3R_2 ) ) /* cnt > 0 ? -> process next vertex */
235
236 LLBL( G3TP3R_3 ):
237
238 FEMMS
239 POP_L ( EDI )
240 POP_L ( ESI )
241 RET
242
243
244
245
246 ALIGNTEXT16
247 GLOBL GLNAME( _mesa_3dnow_transform_points2_3d_no_rot )
248 GLNAME( _mesa_3dnow_transform_points2_3d_no_rot ):
249
250 PUSH_L ( ESI )
251
252 MOV_L ( ARG_DEST, ECX )
253 MOV_L ( ARG_MATRIX, ESI )
254 MOV_L ( ARG_SOURCE, EAX )
255 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
256 OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) )
257 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
258 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
259
260 PUSH_L ( EDI )
261
262 MOV_L ( REGOFF(V4F_START, ECX), EDX )
263 MOV_L ( ESI, ECX )
264 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
265 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
266 MOV_L ( REGOFF(V4F_START, EAX), EAX )
267
268 TEST_L ( ESI, ESI )
269 JZ ( LLBL( G3TP3NRR_3 ) )
270
271 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
272 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
273
274 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
275 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
276
277 ALIGNTEXT16
278 LLBL( G3TP3NRR_2 ):
279
280 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
281 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
282
283 PFADD ( MM2, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
284 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
285
286 MOVD ( MM3, REGOFF(8, EDX) ) /* write r2 */
287 ADD_L ( EDI, EAX ) /* next vertex */
288
289 ADD_L ( CONST(16), EDX ) /* next r */
290 DEC_L ( ESI ) /* decrement vertex counter */
291
292 JNZ ( LLBL( G3TP3NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
293
294 LLBL( G3TP3NRR_3 ):
295
296 FEMMS
297 POP_L ( EDI )
298 POP_L ( ESI )
299 RET
300
301
302
303
304 ALIGNTEXT16
305 GLOBL GLNAME( _mesa_3dnow_transform_points2_2d )
306 GLNAME( _mesa_3dnow_transform_points2_2d ):
307
308 PUSH_L ( ESI )
309
310 MOV_L ( ARG_DEST, ECX )
311 MOV_L ( ARG_MATRIX, ESI )
312 MOV_L ( ARG_SOURCE, EAX )
313 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
314 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
315 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
316 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
317
318 PUSH_L ( EDI )
319
320 MOV_L ( REGOFF(V4F_START, ECX), EDX )
321 MOV_L ( ESI, ECX )
322 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
323 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
324 MOV_L ( REGOFF(V4F_START, EAX), EAX )
325
326 TEST_L ( ESI, ESI )
327 JZ ( LLBL( G3TP2R_3 ) )
328
329 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
330 MOVQ ( REGOFF(16, ECX), MM1 ) /* m11 | m10 */
331
332 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
333
334 ALIGNTEXT16
335 LLBL( G3TP2R_2 ):
336
337 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
338 MOVD ( REGOFF(4, EAX), MM5 ) /* | x1 */
339
340 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
341 ADD_L ( EDI, EAX ) /* next vertex */
342
343 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
344 PUNPCKLDQ ( MM5, MM5 ) /* x1 | x1 */
345
346 PFMUL ( MM1, MM5 ) /* x1*m11 | x1*m10 */
347 PFADD ( MM2, MM4 ) /* x...x1*m11+31 | x0*..*m10+m30 */
348
349 PFADD ( MM5, MM4 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */
350 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
351
352 ADD_L ( CONST(16), EDX ) /* next r */
353 DEC_L ( ESI ) /* decrement vertex counter */
354
355 JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
356
357 LLBL( G3TP2R_3 ):
358
359 FEMMS
360 POP_L ( EDI )
361 POP_L ( ESI )
362 RET
363
364
365
366
367 ALIGNTEXT16
368 GLOBL GLNAME( _mesa_3dnow_transform_points2_2d_no_rot )
369 GLNAME( _mesa_3dnow_transform_points2_2d_no_rot ):
370
371 PUSH_L ( ESI )
372
373 MOV_L ( ARG_DEST, ECX )
374 MOV_L ( ARG_MATRIX, ESI )
375 MOV_L ( ARG_SOURCE, EAX )
376 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
377 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
378 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
379 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
380
381 PUSH_L ( EDI )
382
383 MOV_L ( REGOFF(V4F_START, ECX), EDX )
384 MOV_L ( ESI, ECX )
385 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
386 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
387 MOV_L ( REGOFF(V4F_START, EAX), EAX )
388
389 TEST_L ( ESI, ESI )
390 JZ ( LLBL( G3TP2NRR_3 ) )
391
392 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
393 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
394
395 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
396
397 ALIGNTEXT16
398 LLBL( G3TP2NRR_2 ):
399
400 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
401 ADD_L ( EDI, EAX ) /* next vertex */
402
403 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
404 PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */
405
406 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
407 ADD_L ( CONST(16), EDX ) /* next r */
408
409 DEC_L ( ESI ) /* decrement vertex counter */
410 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
411
412 LLBL( G3TP2NRR_3 ):
413
414 FEMMS
415 POP_L ( EDI )
416 POP_L ( ESI )
417 RET
418
419
420
421
422 ALIGNTEXT16
423 GLOBL GLNAME( _mesa_3dnow_transform_points2_identity )
424 GLNAME( _mesa_3dnow_transform_points2_identity ):
425
426 PUSH_L ( ESI )
427
428 MOV_L ( ARG_DEST, ECX )
429 MOV_L ( ARG_MATRIX, ESI )
430 MOV_L ( ARG_SOURCE, EAX )
431 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
432 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
433 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
434 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
435
436 PUSH_L ( EDI )
437
438 MOV_L ( REGOFF(V4F_START, ECX), EDX )
439 MOV_L ( ESI, ECX )
440 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
441 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
442 MOV_L ( REGOFF(V4F_START, EAX), EAX )
443
444 TEST_L ( ESI, ESI )
445 JZ ( LLBL( G3TPIR_3 ) )
446
447 ALIGNTEXT16
448 LLBL( G3TPIR_3 ):
449
450 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
451 ADD_L ( EDI, EAX ) /* next vertex */
452
453 MOVQ ( MM0, REGIND(EDX) ) /* r1 | r0 */
454 ADD_L ( CONST(16), EDX ) /* next r */
455
456 DEC_L ( ESI ) /* decrement vertex counter */
457 JNZ ( LLBL( G3TPIR_3 ) ) /* cnt > 0 ? -> process next vertex */
458
459 LLBL( G3TPIR_4 ):
460
461 FEMMS
462 POP_L ( EDI )
463 POP_L ( ESI )
464 RET