Massively cleaned up the code that calculates front/back/depth buffer
[mesa.git] / src / mesa / x86 / 3dnow_xform2.S
1 /* $Id: 3dnow_xform2.S,v 1.2 2004/04/26 10:10:25 alanh 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 GLNAME( _mesa_3dnow_transform_points2_general ):
39
40 PUSH_L ( ESI )
41
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) )
49
50 PUSH_L ( EDI )
51
52 MOV_L ( REGOFF(V4F_START, ECX), EDX )
53 MOV_L ( ESI, ECX )
54 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
55 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
56 MOV_L ( REGOFF(V4F_START, EAX), EAX )
57
58 TEST_L ( ESI, ESI )
59 JZ ( LLBL( G3TPGR_3 ) )
60
61 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
62 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
63
64 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
65 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
66
67 MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */
68 PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */
69
70 MOVD ( REGOFF(12, ECX), MM3 ) /* | m03 */
71 PUNPCKLDQ ( REGOFF(28, ECX), MM3 ) /* m13 | m03 */
72
73 MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */
74 MOVQ ( REGOFF(56, ECX), MM5 ) /* m33 | m32 */
75
76 ALIGNTEXT16
77 LLBL( G3TPGR_2 ):
78
79 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
80 MOVQ ( MM6, MM7 ) /* x1 | x0 */
81
82 PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */
83 PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */
84
85 PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */
86 PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */
87
88 MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */
89 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
90
91 MOVQ ( MM6, MM7 ) /* x1 | x0 */
92 PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */
93
94 PFMUL ( MM3, MM7 ) /* x1*m13 | x0*m03 */
95 ADD_L ( EDI, EAX ) /* next vertex */
96
97 PFACC ( MM7, MM6 ) /* x0*m03+x1*m13 | x0*x02+x1*m12 */
98 PFADD ( MM5, MM6 ) /* x0*...*m13+m33 | x0*...*m12+m32 */
99
100 MOVQ ( MM6, REGOFF(8, EDX) ) /* write r3, r2 */
101 ADD_L ( CONST(16), EDX ) /* next r */
102
103 DEC_L ( ESI ) /* decrement vertex counter */
104 JNZ ( LLBL( G3TPGR_2 ) ) /* cnt > 0 ? -> process next vertex */
105
106 LLBL( G3TPGR_3 ):
107
108 FEMMS
109 POP_L ( EDI )
110 POP_L ( ESI )
111 RET
112
113
114
115
116 ALIGNTEXT16
117 GLOBL GLNAME( _mesa_3dnow_transform_points2_perspective )
118 GLNAME( _mesa_3dnow_transform_points2_perspective ):
119
120 PUSH_L ( ESI )
121
122 MOV_L ( ARG_DEST, ECX )
123 MOV_L ( ARG_MATRIX, ESI )
124 MOV_L ( ARG_SOURCE, EAX )
125 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
126 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
127 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
128 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
129
130 PUSH_L ( EDI )
131
132 MOV_L ( REGOFF(V4F_START, ECX), EDX )
133 MOV_L ( ESI, ECX )
134 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
135 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
136 MOV_L ( REGOFF(V4F_START, EAX), EAX )
137
138 TEST_L ( ESI, ESI )
139 JZ ( LLBL( G3TPPR_3 ) )
140
141 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
142 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
143
144 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
145
146 ALIGNTEXT16
147 LLBL( G3TPPR_2 ):
148
149 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
150 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
151
152 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
153 MOVQ ( MM3, REGOFF(8, EDX) ) /* write r2 (=m32), r3 (=0) */
154
155 ADD_L ( EDI, EAX ) /* next vertex */
156 ADD_L ( CONST(16), EDX ) /* next r */
157
158 DEC_L ( ESI ) /* decrement vertex counter */
159 JNZ ( LLBL( G3TPPR_2 ) ) /* cnt > 0 ? -> process next vertex */
160
161 LLBL( G3TPPR_3 ):
162
163 FEMMS
164 POP_L ( EDI )
165 POP_L ( ESI )
166 RET
167
168
169
170
171 ALIGNTEXT16
172 GLOBL GLNAME( _mesa_3dnow_transform_points2_3d )
173 GLNAME( _mesa_3dnow_transform_points2_3d ):
174
175 PUSH_L ( ESI )
176
177 MOV_L ( ARG_DEST, ECX )
178 MOV_L ( ARG_MATRIX, ESI )
179 MOV_L ( ARG_SOURCE, EAX )
180 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
181 OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) )
182 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
183 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
184
185 PUSH_L ( EDI )
186
187 MOV_L ( REGOFF(V4F_START, ECX), EDX )
188 MOV_L ( ESI, ECX )
189 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
190 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
191 MOV_L ( REGOFF(V4F_START, EAX), EAX )
192
193 TEST_L ( ESI, ESI )
194 JZ ( LLBL( G3TP3R_3 ) )
195
196 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
197 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
198
199 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
200 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
201
202 MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */
203 PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */
204
205 MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */
206 MOVD ( REGOFF(56, ECX), MM5 ) /* | m32 */
207
208 ALIGNTEXT16
209 LLBL( G3TP3R_2 ):
210
211 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
212 MOVQ ( MM6, MM7 ) /* x1 | x0 */
213
214 PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */
215 PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */
216
217 PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */
218 PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */
219
220 MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */
221 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
222
223 MOVQ ( MM6, MM7 ) /* x1 | x0 */
224 PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */
225
226 PFACC ( MM7, MM6 ) /* ***trash*** | x0*x02+x1*m12 */
227 PFADD ( MM5, MM6 ) /* ***trash*** | x0*...*m12+m32 */
228
229 MOVD ( MM6, REGOFF(8, EDX) ) /* write r2 */
230 ADD_L ( EDI, EAX ) /* next vertex */
231
232 ADD_L ( CONST(16), EDX ) /* next r */
233 DEC_L ( ESI ) /* decrement vertex counter */
234
235 JNZ ( LLBL( G3TP3R_2 ) ) /* cnt > 0 ? -> process next vertex */
236
237 LLBL( G3TP3R_3 ):
238
239 FEMMS
240 POP_L ( EDI )
241 POP_L ( ESI )
242 RET
243
244
245
246
247 ALIGNTEXT16
248 GLOBL GLNAME( _mesa_3dnow_transform_points2_3d_no_rot )
249 GLNAME( _mesa_3dnow_transform_points2_3d_no_rot ):
250
251 PUSH_L ( ESI )
252
253 MOV_L ( ARG_DEST, ECX )
254 MOV_L ( ARG_MATRIX, ESI )
255 MOV_L ( ARG_SOURCE, EAX )
256 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
257 OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) )
258 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
259 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
260
261 PUSH_L ( EDI )
262
263 MOV_L ( REGOFF(V4F_START, ECX), EDX )
264 MOV_L ( ESI, ECX )
265 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
266 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
267 MOV_L ( REGOFF(V4F_START, EAX), EAX )
268
269 TEST_L ( ESI, ESI )
270 JZ ( LLBL( G3TP3NRR_3 ) )
271
272 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
273 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
274
275 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
276 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
277
278 ALIGNTEXT16
279 LLBL( G3TP3NRR_2 ):
280
281 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
282 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
283
284 PFADD ( MM2, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
285 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
286
287 MOVD ( MM3, REGOFF(8, EDX) ) /* write r2 */
288 ADD_L ( EDI, EAX ) /* next vertex */
289
290 ADD_L ( CONST(16), EDX ) /* next r */
291 DEC_L ( ESI ) /* decrement vertex counter */
292
293 JNZ ( LLBL( G3TP3NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
294
295 LLBL( G3TP3NRR_3 ):
296
297 FEMMS
298 POP_L ( EDI )
299 POP_L ( ESI )
300 RET
301
302
303
304
305 ALIGNTEXT16
306 GLOBL GLNAME( _mesa_3dnow_transform_points2_2d )
307 GLNAME( _mesa_3dnow_transform_points2_2d ):
308
309 PUSH_L ( ESI )
310
311 MOV_L ( ARG_DEST, ECX )
312 MOV_L ( ARG_MATRIX, ESI )
313 MOV_L ( ARG_SOURCE, EAX )
314 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
315 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
316 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
317 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
318
319 PUSH_L ( EDI )
320
321 MOV_L ( REGOFF(V4F_START, ECX), EDX )
322 MOV_L ( ESI, ECX )
323 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
324 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
325 MOV_L ( REGOFF(V4F_START, EAX), EAX )
326
327 TEST_L ( ESI, ESI )
328 JZ ( LLBL( G3TP2R_3 ) )
329
330 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
331 MOVQ ( REGOFF(16, ECX), MM1 ) /* m11 | m10 */
332
333 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
334
335 ALIGNTEXT16
336 LLBL( G3TP2R_2 ):
337
338 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
339 MOVD ( REGOFF(4, EAX), MM5 ) /* | x1 */
340
341 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
342 ADD_L ( EDI, EAX ) /* next vertex */
343
344 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
345 PUNPCKLDQ ( MM5, MM5 ) /* x1 | x1 */
346
347 PFMUL ( MM1, MM5 ) /* x1*m11 | x1*m10 */
348 PFADD ( MM2, MM4 ) /* x...x1*m11+31 | x0*..*m10+m30 */
349
350 PFADD ( MM5, MM4 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */
351 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
352
353 ADD_L ( CONST(16), EDX ) /* next r */
354 DEC_L ( ESI ) /* decrement vertex counter */
355
356 JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
357
358 LLBL( G3TP2R_3 ):
359
360 FEMMS
361 POP_L ( EDI )
362 POP_L ( ESI )
363 RET
364
365
366
367
368 ALIGNTEXT16
369 GLOBL GLNAME( _mesa_3dnow_transform_points2_2d_no_rot )
370 GLNAME( _mesa_3dnow_transform_points2_2d_no_rot ):
371
372 PUSH_L ( ESI )
373
374 MOV_L ( ARG_DEST, ECX )
375 MOV_L ( ARG_MATRIX, ESI )
376 MOV_L ( ARG_SOURCE, EAX )
377 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
378 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
379 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
380 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
381
382 PUSH_L ( EDI )
383
384 MOV_L ( REGOFF(V4F_START, ECX), EDX )
385 MOV_L ( ESI, ECX )
386 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
387 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
388 MOV_L ( REGOFF(V4F_START, EAX), EAX )
389
390 TEST_L ( ESI, ESI )
391 JZ ( LLBL( G3TP2NRR_3 ) )
392
393 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
394 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
395
396 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
397
398 ALIGNTEXT16
399 LLBL( G3TP2NRR_2 ):
400
401 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
402 ADD_L ( EDI, EAX ) /* next vertex */
403
404 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
405 PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */
406
407 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
408 ADD_L ( CONST(16), EDX ) /* next r */
409
410 DEC_L ( ESI ) /* decrement vertex counter */
411 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
412
413 LLBL( G3TP2NRR_3 ):
414
415 FEMMS
416 POP_L ( EDI )
417 POP_L ( ESI )
418 RET
419
420
421
422
423 ALIGNTEXT16
424 GLOBL GLNAME( _mesa_3dnow_transform_points2_identity )
425 GLNAME( _mesa_3dnow_transform_points2_identity ):
426
427 PUSH_L ( ESI )
428
429 MOV_L ( ARG_DEST, ECX )
430 MOV_L ( ARG_MATRIX, ESI )
431 MOV_L ( ARG_SOURCE, EAX )
432 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
433 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
434 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
435 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
436
437 PUSH_L ( EDI )
438
439 MOV_L ( REGOFF(V4F_START, ECX), EDX )
440 MOV_L ( ESI, ECX )
441 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
442 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
443 MOV_L ( REGOFF(V4F_START, EAX), EAX )
444
445 TEST_L ( ESI, ESI )
446 JZ ( LLBL( G3TPIR_3 ) )
447
448 ALIGNTEXT16
449 LLBL( G3TPIR_3 ):
450
451 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
452 ADD_L ( EDI, EAX ) /* next vertex */
453
454 MOVQ ( MM0, REGIND(EDX) ) /* r1 | r0 */
455 ADD_L ( CONST(16), EDX ) /* next r */
456
457 DEC_L ( ESI ) /* decrement vertex counter */
458 JNZ ( LLBL( G3TPIR_3 ) ) /* cnt > 0 ? -> process next vertex */
459
460 LLBL( G3TPIR_4 ):
461
462 FEMMS
463 POP_L ( EDI )
464 POP_L ( ESI )
465 RET
466 #endif