API dispath updates
[mesa.git] / src / mesa / x86 / 3dnow_xform1.S
1 /* $Id: 3dnow_xform1.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_points1_general )
37 GLNAME( _mesa_3dnow_transform_points1_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(4, 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 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
61 MOVQ ( REGOFF(8, ECX), MM1 ) /* m03 | m02 */
62
63 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
64 MOVQ ( REGOFF(56, ECX), MM3 ) /* m33 | m32 */
65
66 ALIGNTEXT16
67 LLBL( G3TPGR_2 ):
68
69 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
70 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
71
72 MOVQ ( MM4, MM5 ) /* x0 | x0 */
73 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
74
75 PFMUL ( MM1, MM5 ) /* x0*m03 | x0*m02 */
76 PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */
77
78 PFADD ( MM3, MM5 ) /* x0*m03+m33 | x0*m02+m32 */
79 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
80
81 MOVQ ( MM5, REGOFF(8, EDX) ) /* write r3, r2 */
82 ADD_L ( EDI, EAX ) /* next vertex */
83
84 ADD_L ( CONST(16), EDX ) /* next r */
85 DEC_L ( ESI ) /* decrement vertex counter */
86
87 JNZ ( LLBL( G3TPGR_2 ) ) /* cnt > 0 ? -> process next vertex */
88
89 LLBL( G3TPGR_3 ):
90
91 FEMMS
92 POP_L ( EDI )
93 POP_L ( ESI )
94 RET
95
96
97
98
99 ALIGNTEXT16
100 GLOBL GLNAME( _mesa_3dnow_transform_points1_identity )
101 GLNAME( _mesa_3dnow_transform_points1_identity ):
102
103 PUSH_L ( ESI )
104
105 MOV_L ( ARG_DEST, ECX )
106 MOV_L ( ARG_MATRIX, ESI )
107 MOV_L ( ARG_SOURCE, EAX )
108 MOV_L ( CONST(1), REGOFF(V4F_SIZE, ECX) )
109 OR_B ( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, ECX) )
110 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
111 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
112
113 PUSH_L ( EDI )
114
115 MOV_L ( REGOFF(4, ECX), EDX )
116 MOV_L ( ESI, ECX )
117 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
118 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
119 MOV_L ( REGOFF(V4F_START, EAX), EAX )
120
121 TEST_L ( ESI, ESI )
122 JZ ( LLBL( G3TPIR_4) )
123
124 ALIGNTEXT16
125 LLBL( G3TPIR_3 ):
126
127 MOVD ( REGIND(EAX), MM0 ) /* | x0 */
128 ADD_L ( EDI, EAX ) /* next vertex */
129
130 MOVD ( MM0, REGIND(EDX) ) /* | r0 */
131 ADD_L ( CONST(16), EDX ) /* next r */
132
133 DEC_L ( ESI ) /* decrement vertex counter */
134 JNZ ( LLBL( G3TPIR_3 ) ) /* cnt > 0 ? -> process next vertex */
135
136 LLBL( G3TPIR_4 ):
137
138 FEMMS
139 POP_L ( EDI )
140 POP_L ( ESI )
141 RET
142
143
144
145
146 ALIGNTEXT16
147 GLOBL GLNAME( _mesa_3dnow_transform_points1_3d_no_rot )
148 GLNAME( _mesa_3dnow_transform_points1_3d_no_rot ):
149
150 PUSH_L ( ESI )
151
152 MOV_L ( ARG_DEST, ECX )
153 MOV_L ( ARG_MATRIX, ESI )
154 MOV_L ( ARG_SOURCE, EAX )
155 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
156 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
157 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
158 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
159
160 PUSH_L ( EDI )
161
162 MOV_L ( REGOFF(4, ECX), EDX )
163 MOV_L ( ESI, ECX )
164 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
165 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
166 MOV_L ( REGOFF(V4F_START, EAX), EAX )
167
168 TEST_L ( ESI, ESI )
169 JZ ( LLBL( G3TP3NRR_3 ) )
170
171 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
172 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
173
174 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
175
176 ALIGNTEXT16
177 LLBL( G3TP3NRR_2 ):
178
179 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
180 PFMUL ( MM0, MM4 ) /* | x0*m00 */
181
182 PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */
183 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
184
185 MOVD ( MM3, REGOFF(8, EDX) ) /* write r2 */
186 ADD_L ( EDI, EAX ) /* next vertex */
187
188 ADD_L ( CONST(16), EDX ) /* next r */
189 DEC_L ( ESI ) /* decrement vertex counter */
190
191 JNZ ( LLBL( G3TP3NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
192
193 LLBL( G3TP3NRR_3 ):
194
195 FEMMS
196 POP_L ( EDI )
197 POP_L ( ESI )
198 RET
199
200
201
202
203 ALIGNTEXT16
204 GLOBL GLNAME( _mesa_3dnow_transform_points1_perspective )
205 GLNAME( _mesa_3dnow_transform_points1_perspective ):
206
207 PUSH_L ( ESI )
208
209 MOV_L ( ARG_DEST, ECX )
210 MOV_L ( ARG_MATRIX, ESI )
211 MOV_L ( ARG_SOURCE, EAX )
212 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
213 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
214 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
215 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
216
217 PUSH_L ( EDI )
218
219 MOV_L ( REGOFF(4, ECX), EDX )
220 MOV_L ( ESI, ECX )
221 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
222 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
223 MOV_L ( REGOFF(V4F_START, EAX), EAX )
224
225 TEST_L ( ESI, ESI )
226 JZ ( LLBL( G3TPPR_3 ) )
227
228 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
229 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
230
231 ALIGNTEXT16
232 LLBL( G3TPPR_2 ):
233
234 MOVD ( REGIND(EAX), MM4 ) /* 0 | x0 */
235 PFMUL ( MM0, MM4 ) /* 0 | x0*m00 */
236
237 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
238 MOVQ ( MM3, REGOFF(8, EDX) ) /* write r2 (=m32), r3 (=0) */
239
240 ADD_L ( EDI, EAX ) /* next vertex */
241 ADD_L ( CONST(16), EDX ) /* next r */
242
243 DEC_L ( ESI ) /* decrement vertex counter */
244 JNZ ( LLBL( G3TPPR_2 ) ) /* cnt > 0 ? -> process next vertex */
245
246 LLBL( G3TPPR_3 ):
247
248 FEMMS
249 POP_L ( EDI )
250 POP_L ( ESI )
251 RET
252
253
254
255
256 ALIGNTEXT16
257 GLOBL GLNAME( _mesa_3dnow_transform_points1_2d )
258 GLNAME( _mesa_3dnow_transform_points1_2d ):
259
260 PUSH_L ( ESI )
261
262 MOV_L ( ARG_DEST, ECX )
263 MOV_L ( ARG_MATRIX, ESI )
264 MOV_L ( ARG_SOURCE, EAX )
265 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
266 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
267 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
268 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
269
270 PUSH_L ( EDI )
271
272 MOV_L ( REGOFF(4, ECX), EDX )
273 MOV_L ( ESI, ECX )
274 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
275 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
276 MOV_L ( REGOFF(V4F_START, EAX), EAX )
277
278 TEST_L ( ESI, ESI )
279 JZ ( LLBL( G3TP2R_3 ) )
280
281 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
282 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
283
284 ALIGNTEXT16
285 LLBL( G3TP2R_2 ):
286
287 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
288 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
289
290 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
291 PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */
292
293 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
294 ADD_L ( EDI, EAX ) /* next vertex */
295
296 ADD_L ( CONST(16), EDX ) /* next r */
297 DEC_L ( ESI ) /* decrement vertex counter */
298
299 JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
300
301 LLBL( G3TP2R_3 ):
302
303 FEMMS
304 POP_L ( EDI )
305 POP_L ( ESI )
306 RET
307
308
309
310
311 ALIGNTEXT16
312 GLOBL GLNAME( _mesa_3dnow_transform_points1_2d_no_rot )
313 GLNAME( _mesa_3dnow_transform_points1_2d_no_rot ):
314
315 PUSH_L ( ESI )
316
317 MOV_L ( ARG_DEST, ECX )
318 MOV_L ( ARG_MATRIX, ESI )
319 MOV_L ( ARG_SOURCE, EAX )
320 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
321 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
322 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
323 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
324
325 PUSH_L ( EDI )
326
327 MOV_L ( REGOFF(4, ECX), EDX )
328 MOV_L ( ESI, ECX )
329 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
330 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
331 MOV_L ( REGOFF(V4F_START, EAX), EAX )
332
333 TEST_L ( ESI, ESI )
334 JZ ( LLBL( G3TP2NRR_3 ) )
335
336 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
337 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
338
339 ALIGNTEXT16
340 LLBL( G3TP2NRR_2 ):
341
342 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
343 ADD_L ( EDI, EAX ) /* next vertex */
344
345 PFMUL ( MM0, MM4 ) /* | x0*m00 */
346 PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */
347
348 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
349 ADD_L ( CONST(16), EDX ) /* next r */
350
351 DEC_L ( ESI ) /* decrement vertex counter */
352 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
353
354 LLBL( G3TP2NRR_3 ):
355
356 FEMMS
357 POP_L ( EDI )
358 POP_L ( ESI )
359 RET
360
361
362
363
364 ALIGNTEXT16
365 GLOBL GLNAME( _mesa_3dnow_transform_points1_3d )
366 GLNAME( _mesa_3dnow_transform_points1_3d ):
367
368 PUSH_L ( ESI )
369
370 MOV_L ( ARG_DEST, ECX )
371 MOV_L ( ARG_MATRIX, ESI )
372 MOV_L ( ARG_SOURCE, EAX )
373 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
374 OR_B ( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, ECX) )
375 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
376 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
377
378 PUSH_L ( EDI )
379
380 MOV_L ( REGOFF(4, ECX), EDX )
381 MOV_L ( ESI, ECX )
382 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
383 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
384 MOV_L ( REGOFF(V4F_START, EAX), EAX )
385
386 TEST_L ( ESI, ESI )
387 JZ ( LLBL( G3TP3R_3 ) )
388
389 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
390 MOVD ( REGOFF(8, ECX), MM1 ) /* | m02 */
391
392 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
393 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
394
395 ALIGNTEXT16
396 LLBL( G3TP3R_2 ):
397
398 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
399 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
400
401 MOVQ ( MM4, MM5 ) /* | x0 */
402 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
403
404 PFMUL ( MM1, MM5 ) /* | x0*m02 */
405 PFADD ( MM2, MM4 ) /* x0*m01+m31 | x0*m00+m30 */
406
407 PFADD ( MM3, MM5 ) /* | x0*m02+m32 */
408 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
409
410 MOVD ( MM5, REGOFF(8, EDX) ) /* write r2 */
411 ADD_L ( EDI, EAX ) /* next vertex */
412
413 ADD_L ( CONST(16), EDX ) /* next r */
414 DEC_L ( ESI ) /* decrement vertex counter */
415
416 JNZ ( LLBL( G3TP3R_2 ) ) /* cnt > 0 ? -> process next vertex */
417
418 LLBL( G3TP3R_3 ):
419
420 FEMMS
421 POP_L ( EDI )
422 POP_L ( ESI )
423 RET