Merge remote branch 'upstream/gallium-0.1' into nouveau-gallium-0.1
[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 "matypes.h"
35 #include "xform_args.h"
36
37 SEG_TEXT
38
39 #define S(i) REGOFF(i * 4, ESI)
40 #define D(i) REGOFF(i * 4, EDI)
41 #define M(i) REGOFF(i * 4, EDX)
42
43
44 ALIGNTEXT4
45 GLOBL GLNAME(_mesa_sse_transform_points2_general)
46 HIDDEN (_mesa_sse_transform_points2_general)
47 GLNAME( _mesa_sse_transform_points2_general ):
48
49 #define FRAME_OFFSET 8
50 PUSH_L ( ESI )
51 PUSH_L ( EDI )
52
53 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
54 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
55
56 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
57 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
58
59 TEST_L( ECX, ECX )
60 JZ( LLBL(K_GTP2GR_finish) ) /* count was zero; go to finish */
61
62 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
63 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
64
65 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
66 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
67
68 SHL_L( CONST(4), ECX ) /* count *= 16 */
69 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
70
71 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
72 ADD_L( EDI, ECX ) /* count += dest ptr */
73
74 ALIGNTEXT32
75 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
76 MOVAPS( M(4), XMM1 ) /* m7 | m6 | m5 | m4 */
77 MOVAPS( M(12), XMM2 ) /* m15 | m14 | m13 | m12 */
78
79 ALIGNTEXT32
80 LLBL(K_GTP2GR_top):
81 MOVSS( S(0), XMM3 ) /* ox */
82 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox | ox */
83 MULPS( XMM0, XMM3 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
84 MOVSS( S(1), XMM4 ) /* oy */
85 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy | oy */
86 MULPS( XMM1, XMM4 ) /* oy*m7 | oy*m6 | oy*m5 | oy*m4 */
87
88 ADDPS( XMM4, XMM3 )
89 ADDPS( XMM2, XMM3 )
90 MOVAPS( XMM3, D(0) )
91
92 LLBL(K_GTP2GR_skip):
93 ADD_L ( CONST(16), EDI )
94 ADD_L ( EAX, ESI )
95 CMP_L ( ECX, EDI )
96 JNE ( LLBL(K_GTP2GR_top) )
97
98 LLBL(K_GTP2GR_finish):
99 POP_L ( EDI )
100 POP_L ( ESI )
101 RET
102 #undef FRAME_OFFSET
103
104
105 ALIGNTEXT4
106 GLOBL GLNAME(_mesa_sse_transform_points2_identity)
107 HIDDEN(_mesa_sse_transform_points2_identity)
108 GLNAME( _mesa_sse_transform_points2_identity ):
109
110 #define FRAME_OFFSET 8
111 PUSH_L ( ESI )
112 PUSH_L ( EDI )
113
114 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
115 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
116
117 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
118
119 TEST_L( ECX, ECX)
120 JZ( LLBL(K_GTP2IR_finish) ) /* count was zero; go to finish */
121
122 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
123 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
124
125 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
126 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
127
128 SHL_L( CONST(4), ECX ) /* count *= 16 */
129 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
130
131 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
132 ADD_L( EDI, ECX ) /* count += dest ptr */
133
134 CMP_L( ESI, EDI )
135 JE( LLBL(K_GTP2IR_finish) )
136
137
138 ALIGNTEXT32
139 LLBL(K_GTP2IR_top):
140 MOV_L ( S(0), EDX )
141 MOV_L ( EDX, D(0) )
142 MOV_L ( S(1), EDX )
143 MOV_L ( EDX, D(1) )
144
145 LLBL(K_GTP2IR_skip):
146 ADD_L ( CONST(16), EDI )
147 ADD_L ( EAX, ESI )
148 CMP_L ( ECX, EDI )
149 JNE ( LLBL(K_GTP2IR_top) )
150
151 LLBL(K_GTP2IR_finish):
152 POP_L ( EDI )
153 POP_L ( ESI )
154 RET
155 #undef FRAME_OFFSET
156
157
158 ALIGNTEXT4
159 GLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot)
160 HIDDEN(_mesa_sse_transform_points2_3d_no_rot)
161 GLNAME(_mesa_sse_transform_points2_3d_no_rot):
162
163 #define FRAME_OFFSET 8
164 PUSH_L( ESI )
165 PUSH_L( EDI )
166
167 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
168 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
169
170 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
171 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
172
173 TEST_L( ECX, ECX)
174 JZ( LLBL(K_GTP23DNRR_finish) ) /* count was zero; go to finish */
175
176 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
177 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
178
179 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
180 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
181
182 SHL_L( CONST(4), ECX ) /* count *= 16 */
183 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
184
185 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
186 ADD_L( EDI, ECX ) /* count += dest ptr */
187
188
189 ALIGNTEXT32
190 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
191 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
192 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
193 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
194 MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */
195
196 ALIGNTEXT32
197 LLBL(K_GTP23DNRR_top):
198 MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */
199 MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */
200 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
201 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
202
203 MOVSS ( XMM3, D(2) ) /* -> D(2) */
204
205 LLBL(K_GTP23DNRR_skip):
206 ADD_L ( CONST(16), EDI )
207 ADD_L ( EAX, ESI )
208 CMP_L ( ECX, EDI )
209 JNE ( LLBL(K_GTP23DNRR_top) )
210
211 LLBL(K_GTP23DNRR_finish):
212 POP_L ( EDI )
213 POP_L ( ESI )
214 RET
215 #undef FRAME_OFFSET
216
217
218 ALIGNTEXT4
219 GLOBL GLNAME(_mesa_sse_transform_points2_perspective)
220 HIDDEN(_mesa_sse_transform_points2_perspective)
221 GLNAME(_mesa_sse_transform_points2_perspective):
222
223 #define FRAME_OFFSET 8
224 PUSH_L ( ESI )
225 PUSH_L ( EDI )
226
227 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
228 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
229
230 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
231 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
232
233 TEST_L( ECX, ECX)
234 JZ( LLBL(K_GTP23PR_finish) ) /* count was zero; go to finish */
235
236 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
237 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
238
239 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
240 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
241
242 SHL_L( CONST(4), ECX ) /* count *= 16 */
243 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
244
245 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
246 ADD_L( EDI, ECX ) /* count += dest ptr */
247
248 ALIGNTEXT32
249 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
250 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
251 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
252 MOVSS ( M(14), XMM3 ) /* m14 */
253 XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
254
255 ALIGNTEXT32
256 LLBL(K_GTP23PR_top):
257 MOVLPS( S(0), XMM4 ) /* oy | ox */
258 MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */
259 MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */
260 MOVSS( XMM3, D(2) ) /* ->D(2) */
261 MOVSS( XMM0, D(3) ) /* ->D(3) */
262
263 LLBL(K_GTP23PR_skip):
264 ADD_L( CONST(16), EDI )
265 ADD_L( EAX, ESI )
266 CMP_L( ECX, EDI )
267 JNE( LLBL(K_GTP23PR_top) )
268
269 LLBL(K_GTP23PR_finish):
270 POP_L ( EDI )
271 POP_L ( ESI )
272 RET
273 #undef FRAME_OFFSET
274
275
276
277 ALIGNTEXT4
278 GLOBL GLNAME(_mesa_sse_transform_points2_2d)
279 HIDDEN(_mesa_sse_transform_points2_2d)
280 GLNAME(_mesa_sse_transform_points2_2d):
281
282 #define FRAME_OFFSET 8
283 PUSH_L( ESI )
284 PUSH_L( EDI )
285
286 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
287 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
288
289 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
290 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
291
292 TEST_L( ECX, ECX)
293 JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */
294
295 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
296 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
297
298 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
299 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
300
301 SHL_L( CONST(4), ECX ) /* count *= 16 */
302 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
303
304 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
305 ADD_L( EDI, ECX ) /* count += dest ptr */
306
307 ALIGNTEXT32
308 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
309 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
310 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
311
312 ALIGNTEXT32
313 LLBL(K_GTP23P2DR_top):
314 MOVSS( S(0), XMM3 ) /* ox */
315 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
316 MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
317
318 MOVSS( S(1), XMM4 ) /* oy */
319 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
320 MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
321
322 ADDPS( XMM4, XMM3 )
323 ADDPS( XMM2, XMM3 )
324 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
325
326 LLBL(K_GTP23P2DR_skip):
327 ADD_L ( CONST(16), EDI )
328 ADD_L ( EAX, ESI )
329 CMP_L ( ECX, EDI )
330 JNE ( LLBL(K_GTP23P2DR_top) )
331
332 LLBL(K_GTP23P2DR_finish):
333 POP_L ( EDI )
334 POP_L ( ESI )
335 RET
336 #undef FRAME_OFFSET
337
338
339
340 ALIGNTEXT4
341 GLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot)
342 HIDDEN(_mesa_sse_transform_points2_2d_no_rot)
343 GLNAME(_mesa_sse_transform_points2_2d_no_rot):
344
345 #define FRAME_OFFSET 8
346 PUSH_L( ESI )
347 PUSH_L( EDI )
348
349 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
350 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
351
352 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
353 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
354
355 TEST_L( ECX, ECX)
356 JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */
357
358 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
359 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
360
361 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
362 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
363
364 SHL_L( CONST(4), ECX ) /* count *= 16 */
365 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
366
367 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
368 ADD_L( EDI, ECX ) /* count += dest ptr */
369
370 ALIGNTEXT32
371 MOVSS ( M(0), XMM1 ) /* m0 */
372 MOVSS ( M(5), XMM2 ) /* m5 */
373 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
374 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
375
376 ALIGNTEXT32
377 LLBL(K_GTP23P2DNRR_top):
378 MOVLPS( S(0), XMM0 ) /* oy | ox */
379 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
380 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
381 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
382
383 LLBL(K_GTP23P2DNRR_skip):
384 ADD_L( CONST(16), EDI )
385 ADD_L( EAX, ESI )
386 CMP_L( ECX, EDI )
387 JNE( LLBL(K_GTP23P2DNRR_top) )
388
389 LLBL(K_GTP23P2DNRR_finish):
390 POP_L( EDI )
391 POP_L( ESI )
392 RET
393 #undef FRAME_OFFSET
394
395
396
397 ALIGNTEXT4
398 GLOBL GLNAME(_mesa_sse_transform_points2_3d)
399 HIDDEN(_mesa_sse_transform_points2_3d)
400 GLNAME(_mesa_sse_transform_points2_3d):
401
402 #define FRAME_OFFSET 8
403 PUSH_L( ESI )
404 PUSH_L( EDI )
405
406 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
407 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
408
409 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
410 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
411
412 TEST_L( ECX, ECX)
413 JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */
414
415 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
416 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
417
418 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
419 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
420
421 SHL_L( CONST(4), ECX ) /* count *= 16 */
422 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
423
424 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
425 ADD_L( EDI, ECX ) /* count += dest ptr */
426
427 ALIGNTEXT32
428 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
429 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
430 MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */
431
432 ALIGNTEXT32
433 LLBL(K_GTP23P3DR_top):
434 MOVSS( S(0), XMM3 ) /* ox */
435 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */
436 MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */
437
438 MOVSS( S(1), XMM4 ) /* oy */
439 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */
440 MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */
441
442 ADDPS( XMM4, XMM3 )
443 ADDPS( XMM2, XMM3 )
444
445 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
446 UNPCKHPS( XMM3, XMM3 )
447 MOVSS( XMM3, D(2) ) /* ->D(2) */
448
449 LLBL(K_GTP23P3DR_skip):
450 ADD_L( CONST(16), EDI )
451 ADD_L( EAX, ESI )
452 CMP_L( ECX, EDI )
453 JNE( LLBL(K_GTP23P3DR_top) )
454
455 LLBL(K_GTP23P3DR_finish):
456 POP_L( EDI )
457 POP_L( ESI )
458 RET
459 #undef FRAME_OFFSET
460 #endif
461
462 #if defined (__ELF__) && defined (__linux__)
463 .section .note.GNU-stack,"",%progbits
464 #endif