Remove CVS keywords.
[mesa.git] / src / mesa / x86 / sse_xform2.S
1
2 /*
3 * Mesa 3-D graphics library
4 * Version: 3.5
5 *
6 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
7 *
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
14 *
15 * The above copyright notice and this permission notice shall be included
16 * in all copies or substantial portions of the Software.
17 *
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
22 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
23 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
24 */
25
26 /** TODO:
27 * - insert PREFETCH instructions to avoid cache-misses !
28 * - some more optimizations are possible...
29 * - for 40-50% more performance in the SSE-functions, the
30 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
31 */
32
33 #ifdef USE_SSE_ASM
34 #include "assyntax.h"
35 #include "matypes.h"
36 #include "xform_args.h"
37
38 SEG_TEXT
39
40 #define S(i) REGOFF(i * 4, ESI)
41 #define D(i) REGOFF(i * 4, EDI)
42 #define M(i) REGOFF(i * 4, EDX)
43
44
45 ALIGNTEXT4
46 GLOBL GLNAME(_mesa_sse_transform_points2_general)
47 HIDDEN (_mesa_sse_transform_points2_general)
48 GLNAME( _mesa_sse_transform_points2_general ):
49
50 #define FRAME_OFFSET 8
51 PUSH_L ( ESI )
52 PUSH_L ( EDI )
53
54 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
55 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
56
57 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
58 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
59
60 TEST_L( ECX, ECX )
61 JZ( LLBL(K_GTP2GR_finish) ) /* count was zero; go to finish */
62
63 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
64 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
65
66 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
67 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
68
69 SHL_L( CONST(4), ECX ) /* count *= 16 */
70 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
71
72 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
73 ADD_L( EDI, ECX ) /* count += dest ptr */
74
75 ALIGNTEXT32
76 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
77 MOVAPS( M(4), XMM1 ) /* m7 | m6 | m5 | m4 */
78 MOVAPS( M(12), XMM2 ) /* m15 | m14 | m13 | m12 */
79
80 ALIGNTEXT32
81 LLBL(K_GTP2GR_top):
82 MOVSS( S(0), XMM3 ) /* ox */
83 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox | ox */
84 MULPS( XMM0, XMM3 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
85 MOVSS( S(1), XMM4 ) /* oy */
86 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy | oy */
87 MULPS( XMM1, XMM4 ) /* oy*m7 | oy*m6 | oy*m5 | oy*m4 */
88
89 ADDPS( XMM4, XMM3 )
90 ADDPS( XMM2, XMM3 )
91 MOVAPS( XMM3, D(0) )
92
93 LLBL(K_GTP2GR_skip):
94 ADD_L ( CONST(16), EDI )
95 ADD_L ( EAX, ESI )
96 CMP_L ( ECX, EDI )
97 JNE ( LLBL(K_GTP2GR_top) )
98
99 LLBL(K_GTP2GR_finish):
100 POP_L ( EDI )
101 POP_L ( ESI )
102 RET
103 #undef FRAME_OFFSET
104
105
106 ALIGNTEXT4
107 GLOBL GLNAME(_mesa_sse_transform_points2_identity)
108 HIDDEN(_mesa_sse_transform_points2_identity)
109 GLNAME( _mesa_sse_transform_points2_identity ):
110
111 #define FRAME_OFFSET 8
112 PUSH_L ( ESI )
113 PUSH_L ( EDI )
114
115 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
116 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
117
118 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
119
120 TEST_L( ECX, ECX)
121 JZ( LLBL(K_GTP2IR_finish) ) /* count was zero; go to finish */
122
123 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
124 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
125
126 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
127 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
128
129 SHL_L( CONST(4), ECX ) /* count *= 16 */
130 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
131
132 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
133 ADD_L( EDI, ECX ) /* count += dest ptr */
134
135 CMP_L( ESI, EDI )
136 JE( LLBL(K_GTP2IR_finish) )
137
138
139 ALIGNTEXT32
140 LLBL(K_GTP2IR_top):
141 MOV_L ( S(0), EDX )
142 MOV_L ( EDX, D(0) )
143 MOV_L ( S(1), EDX )
144 MOV_L ( EDX, D(1) )
145
146 LLBL(K_GTP2IR_skip):
147 ADD_L ( CONST(16), EDI )
148 ADD_L ( EAX, ESI )
149 CMP_L ( ECX, EDI )
150 JNE ( LLBL(K_GTP2IR_top) )
151
152 LLBL(K_GTP2IR_finish):
153 POP_L ( EDI )
154 POP_L ( ESI )
155 RET
156 #undef FRAME_OFFSET
157
158
159 ALIGNTEXT4
160 GLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot)
161 HIDDEN(_mesa_sse_transform_points2_3d_no_rot)
162 GLNAME(_mesa_sse_transform_points2_3d_no_rot):
163
164 #define FRAME_OFFSET 8
165 PUSH_L( ESI )
166 PUSH_L( EDI )
167
168 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
169 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
170
171 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
172 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
173
174 TEST_L( ECX, ECX)
175 JZ( LLBL(K_GTP23DNRR_finish) ) /* count was zero; go to finish */
176
177 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
178 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
179
180 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
181 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
182
183 SHL_L( CONST(4), ECX ) /* count *= 16 */
184 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
185
186 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
187 ADD_L( EDI, ECX ) /* count += dest ptr */
188
189
190 ALIGNTEXT32
191 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
192 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
193 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
194 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
195 MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */
196
197 ALIGNTEXT32
198 LLBL(K_GTP23DNRR_top):
199 MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */
200 MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */
201 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
202 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
203
204 MOVSS ( XMM3, D(2) ) /* -> D(2) */
205
206 LLBL(K_GTP23DNRR_skip):
207 ADD_L ( CONST(16), EDI )
208 ADD_L ( EAX, ESI )
209 CMP_L ( ECX, EDI )
210 JNE ( LLBL(K_GTP23DNRR_top) )
211
212 LLBL(K_GTP23DNRR_finish):
213 POP_L ( EDI )
214 POP_L ( ESI )
215 RET
216 #undef FRAME_OFFSET
217
218
219 ALIGNTEXT4
220 GLOBL GLNAME(_mesa_sse_transform_points2_perspective)
221 HIDDEN(_mesa_sse_transform_points2_perspective)
222 GLNAME(_mesa_sse_transform_points2_perspective):
223
224 #define FRAME_OFFSET 8
225 PUSH_L ( ESI )
226 PUSH_L ( EDI )
227
228 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
229 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
230
231 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
232 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
233
234 TEST_L( ECX, ECX)
235 JZ( LLBL(K_GTP23PR_finish) ) /* count was zero; go to finish */
236
237 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
238 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
239
240 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
241 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
242
243 SHL_L( CONST(4), ECX ) /* count *= 16 */
244 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
245
246 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
247 ADD_L( EDI, ECX ) /* count += dest ptr */
248
249 ALIGNTEXT32
250 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
251 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
252 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
253 MOVSS ( M(14), XMM3 ) /* m14 */
254 XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
255
256 ALIGNTEXT32
257 LLBL(K_GTP23PR_top):
258 MOVLPS( S(0), XMM4 ) /* oy | ox */
259 MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */
260 MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */
261 MOVSS( XMM3, D(2) ) /* ->D(2) */
262 MOVSS( XMM0, D(3) ) /* ->D(3) */
263
264 LLBL(K_GTP23PR_skip):
265 ADD_L( CONST(16), EDI )
266 ADD_L( EAX, ESI )
267 CMP_L( ECX, EDI )
268 JNE( LLBL(K_GTP23PR_top) )
269
270 LLBL(K_GTP23PR_finish):
271 POP_L ( EDI )
272 POP_L ( ESI )
273 RET
274 #undef FRAME_OFFSET
275
276
277
278 ALIGNTEXT4
279 GLOBL GLNAME(_mesa_sse_transform_points2_2d)
280 HIDDEN(_mesa_sse_transform_points2_2d)
281 GLNAME(_mesa_sse_transform_points2_2d):
282
283 #define FRAME_OFFSET 8
284 PUSH_L( ESI )
285 PUSH_L( EDI )
286
287 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
288 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
289
290 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
291 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
292
293 TEST_L( ECX, ECX)
294 JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */
295
296 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
297 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
298
299 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
300 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
301
302 SHL_L( CONST(4), ECX ) /* count *= 16 */
303 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
304
305 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
306 ADD_L( EDI, ECX ) /* count += dest ptr */
307
308 ALIGNTEXT32
309 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
310 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
311 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
312
313 ALIGNTEXT32
314 LLBL(K_GTP23P2DR_top):
315 MOVSS( S(0), XMM3 ) /* ox */
316 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
317 MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
318
319 MOVSS( S(1), XMM4 ) /* oy */
320 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
321 MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
322
323 ADDPS( XMM4, XMM3 )
324 ADDPS( XMM2, XMM3 )
325 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
326
327 LLBL(K_GTP23P2DR_skip):
328 ADD_L ( CONST(16), EDI )
329 ADD_L ( EAX, ESI )
330 CMP_L ( ECX, EDI )
331 JNE ( LLBL(K_GTP23P2DR_top) )
332
333 LLBL(K_GTP23P2DR_finish):
334 POP_L ( EDI )
335 POP_L ( ESI )
336 RET
337 #undef FRAME_OFFSET
338
339
340
341 ALIGNTEXT4
342 GLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot)
343 HIDDEN(_mesa_sse_transform_points2_2d_no_rot)
344 GLNAME(_mesa_sse_transform_points2_2d_no_rot):
345
346 #define FRAME_OFFSET 8
347 PUSH_L( ESI )
348 PUSH_L( EDI )
349
350 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
351 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
352
353 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
354 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
355
356 TEST_L( ECX, ECX)
357 JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */
358
359 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
360 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
361
362 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
363 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
364
365 SHL_L( CONST(4), ECX ) /* count *= 16 */
366 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
367
368 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
369 ADD_L( EDI, ECX ) /* count += dest ptr */
370
371 ALIGNTEXT32
372 MOVSS ( M(0), XMM1 ) /* m0 */
373 MOVSS ( M(5), XMM2 ) /* m5 */
374 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
375 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
376
377 ALIGNTEXT32
378 LLBL(K_GTP23P2DNRR_top):
379 MOVLPS( S(0), XMM0 ) /* oy | ox */
380 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
381 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
382 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
383
384 LLBL(K_GTP23P2DNRR_skip):
385 ADD_L( CONST(16), EDI )
386 ADD_L( EAX, ESI )
387 CMP_L( ECX, EDI )
388 JNE( LLBL(K_GTP23P2DNRR_top) )
389
390 LLBL(K_GTP23P2DNRR_finish):
391 POP_L( EDI )
392 POP_L( ESI )
393 RET
394 #undef FRAME_OFFSET
395
396
397
398 ALIGNTEXT4
399 GLOBL GLNAME(_mesa_sse_transform_points2_3d)
400 HIDDEN(_mesa_sse_transform_points2_3d)
401 GLNAME(_mesa_sse_transform_points2_3d):
402
403 #define FRAME_OFFSET 8
404 PUSH_L( ESI )
405 PUSH_L( EDI )
406
407 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
408 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
409
410 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
411 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
412
413 TEST_L( ECX, ECX)
414 JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */
415
416 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
417 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
418
419 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
420 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
421
422 SHL_L( CONST(4), ECX ) /* count *= 16 */
423 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
424
425 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
426 ADD_L( EDI, ECX ) /* count += dest ptr */
427
428 ALIGNTEXT32
429 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
430 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
431 MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */
432
433 ALIGNTEXT32
434 LLBL(K_GTP23P3DR_top):
435 MOVSS( S(0), XMM3 ) /* ox */
436 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */
437 MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */
438
439 MOVSS( S(1), XMM4 ) /* oy */
440 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */
441 MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */
442
443 ADDPS( XMM4, XMM3 )
444 ADDPS( XMM2, XMM3 )
445
446 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
447 UNPCKHPS( XMM3, XMM3 )
448 MOVSS( XMM3, D(2) ) /* ->D(2) */
449
450 LLBL(K_GTP23P3DR_skip):
451 ADD_L( CONST(16), EDI )
452 ADD_L( EAX, ESI )
453 CMP_L( ECX, EDI )
454 JNE( LLBL(K_GTP23P3DR_top) )
455
456 LLBL(K_GTP23P3DR_finish):
457 POP_L( EDI )
458 POP_L( ESI )
459 RET
460 #undef FRAME_OFFSET
461 #endif
462
463 #if defined (__ELF__) && defined (__linux__)
464 .section .note.GNU-stack,"",%progbits
465 #endif