49f9889f2e8445c9ac76891a15ce7d8981162af9
[mesa.git] / src / mesa / x86 / sse_xform2.S
1 /* $Id: sse_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 /** TODO:
28 * - insert PREFETCH instructions to avoid cache-misses !
29 * - some more optimizations are possible...
30 * - for 40-50% more performance in the SSE-functions, the
31 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
32 */
33
34 #ifdef USE_SSE_ASM
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 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 GLNAME( _mesa_sse_transform_points2_identity ):
108
109 #define FRAME_OFFSET 8
110 PUSH_L ( ESI )
111 PUSH_L ( EDI )
112
113 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
114 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
115
116 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
117
118 TEST_L( ECX, ECX)
119 JZ( LLBL(K_GTP2IR_finish) ) /* count was zero; go to finish */
120
121 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
122 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
123
124 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
125 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
126
127 SHL_L( CONST(4), ECX ) /* count *= 16 */
128 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
129
130 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
131 ADD_L( EDI, ECX ) /* count += dest ptr */
132
133 CMP_L( ESI, EDI )
134 JE( LLBL(K_GTP2IR_finish) )
135
136
137 ALIGNTEXT32
138 LLBL(K_GTP2IR_top):
139 MOV_L ( S(0), EDX )
140 MOV_L ( EDX, D(0) )
141 MOV_L ( S(1), EDX )
142 MOV_L ( EDX, D(1) )
143
144 LLBL(K_GTP2IR_skip):
145 ADD_L ( CONST(16), EDI )
146 ADD_L ( EAX, ESI )
147 CMP_L ( ECX, EDI )
148 JNE ( LLBL(K_GTP2IR_top) )
149
150 LLBL(K_GTP2IR_finish):
151 POP_L ( EDI )
152 POP_L ( ESI )
153 RET
154 #undef FRAME_OFFSET
155
156
157 ALIGNTEXT4
158 GLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot)
159 GLNAME(_mesa_sse_transform_points2_3d_no_rot):
160
161 #define FRAME_OFFSET 8
162 PUSH_L( ESI )
163 PUSH_L( EDI )
164
165 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
166 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
167
168 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
169 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
170
171 TEST_L( ECX, ECX)
172 JZ( LLBL(K_GTP23DNRR_finish) ) /* count was zero; go to finish */
173
174 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
175 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
176
177 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
178 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
179
180 SHL_L( CONST(4), ECX ) /* count *= 16 */
181 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
182
183 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
184 ADD_L( EDI, ECX ) /* count += dest ptr */
185
186
187 ALIGNTEXT32
188 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
189 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
190 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
191 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
192 MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */
193
194 ALIGNTEXT32
195 LLBL(K_GTP23DNRR_top):
196 MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */
197 MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */
198 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
199 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
200
201 MOVSS ( XMM3, D(2) ) /* -> D(2) */
202
203 LLBL(K_GTP23DNRR_skip):
204 ADD_L ( CONST(16), EDI )
205 ADD_L ( EAX, ESI )
206 CMP_L ( ECX, EDI )
207 JNE ( LLBL(K_GTP23DNRR_top) )
208
209 LLBL(K_GTP23DNRR_finish):
210 POP_L ( EDI )
211 POP_L ( ESI )
212 RET
213 #undef FRAME_OFFSET
214
215
216 ALIGNTEXT4
217 GLOBL GLNAME(_mesa_sse_transform_points2_perspective)
218 GLNAME(_mesa_sse_transform_points2_perspective):
219
220 #define FRAME_OFFSET 8
221 PUSH_L ( ESI )
222 PUSH_L ( EDI )
223
224 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
225 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
226
227 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
228 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
229
230 TEST_L( ECX, ECX)
231 JZ( LLBL(K_GTP23PR_finish) ) /* count was zero; go to finish */
232
233 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
234 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
235
236 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
237 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
238
239 SHL_L( CONST(4), ECX ) /* count *= 16 */
240 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
241
242 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
243 ADD_L( EDI, ECX ) /* count += dest ptr */
244
245 ALIGNTEXT32
246 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
247 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
248 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
249 MOVSS ( M(14), XMM3 ) /* m14 */
250 XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
251
252 ALIGNTEXT32
253 LLBL(K_GTP23PR_top):
254 MOVLPS( S(0), XMM4 ) /* oy | ox */
255 MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */
256 MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */
257 MOVSS( XMM3, D(2) ) /* ->D(2) */
258 MOVSS( XMM0, D(3) ) /* ->D(3) */
259
260 LLBL(K_GTP23PR_skip):
261 ADD_L( CONST(16), EDI )
262 ADD_L( EAX, ESI )
263 CMP_L( ECX, EDI )
264 JNE( LLBL(K_GTP23PR_top) )
265
266 LLBL(K_GTP23PR_finish):
267 POP_L ( EDI )
268 POP_L ( ESI )
269 RET
270 #undef FRAME_OFFSET
271
272
273
274 ALIGNTEXT4
275 GLOBL GLNAME(_mesa_sse_transform_points2_2d)
276 GLNAME(_mesa_sse_transform_points2_2d):
277
278 #define FRAME_OFFSET 8
279 PUSH_L( ESI )
280 PUSH_L( EDI )
281
282 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
283 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
284
285 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
286 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
287
288 TEST_L( ECX, ECX)
289 JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */
290
291 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
292 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
293
294 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
295 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
296
297 SHL_L( CONST(4), ECX ) /* count *= 16 */
298 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
299
300 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
301 ADD_L( EDI, ECX ) /* count += dest ptr */
302
303 ALIGNTEXT32
304 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
305 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
306 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
307
308 ALIGNTEXT32
309 LLBL(K_GTP23P2DR_top):
310 MOVSS( S(0), XMM3 ) /* ox */
311 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
312 MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
313
314 MOVSS( S(1), XMM4 ) /* oy */
315 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
316 MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
317
318 ADDPS( XMM4, XMM3 )
319 ADDPS( XMM2, XMM3 )
320 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
321
322 LLBL(K_GTP23P2DR_skip):
323 ADD_L ( CONST(16), EDI )
324 ADD_L ( EAX, ESI )
325 CMP_L ( ECX, EDI )
326 JNE ( LLBL(K_GTP23P2DR_top) )
327
328 LLBL(K_GTP23P2DR_finish):
329 POP_L ( EDI )
330 POP_L ( ESI )
331 RET
332 #undef FRAME_OFFSET
333
334
335
336 ALIGNTEXT4
337 GLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot)
338 GLNAME(_mesa_sse_transform_points2_2d_no_rot):
339
340 #define FRAME_OFFSET 8
341 PUSH_L( ESI )
342 PUSH_L( EDI )
343
344 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
345 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
346
347 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
348 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
349
350 TEST_L( ECX, ECX)
351 JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */
352
353 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
354 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
355
356 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
357 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
358
359 SHL_L( CONST(4), ECX ) /* count *= 16 */
360 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
361
362 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
363 ADD_L( EDI, ECX ) /* count += dest ptr */
364
365 ALIGNTEXT32
366 MOVSS ( M(0), XMM1 ) /* m0 */
367 MOVSS ( M(5), XMM2 ) /* m5 */
368 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
369 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
370
371 ALIGNTEXT32
372 LLBL(K_GTP23P2DNRR_top):
373 MOVLPS( S(0), XMM0 ) /* oy | ox */
374 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
375 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
376 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
377
378 LLBL(K_GTP23P2DNRR_skip):
379 ADD_L( CONST(16), EDI )
380 ADD_L( EAX, ESI )
381 CMP_L( ECX, EDI )
382 JNE( LLBL(K_GTP23P2DNRR_top) )
383
384 LLBL(K_GTP23P2DNRR_finish):
385 POP_L( EDI )
386 POP_L( ESI )
387 RET
388 #undef FRAME_OFFSET
389
390
391
392 ALIGNTEXT4
393 GLOBL GLNAME(_mesa_sse_transform_points2_3d)
394 GLNAME(_mesa_sse_transform_points2_3d):
395
396 #define FRAME_OFFSET 8
397 PUSH_L( ESI )
398 PUSH_L( EDI )
399
400 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
401 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
402
403 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
404 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
405
406 TEST_L( ECX, ECX)
407 JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */
408
409 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
410 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
411
412 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
413 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
414
415 SHL_L( CONST(4), ECX ) /* count *= 16 */
416 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
417
418 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
419 ADD_L( EDI, ECX ) /* count += dest ptr */
420
421 ALIGNTEXT32
422 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
423 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
424 MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */
425
426 ALIGNTEXT32
427 LLBL(K_GTP23P3DR_top):
428 MOVSS( S(0), XMM3 ) /* ox */
429 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */
430 MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */
431
432 MOVSS( S(1), XMM4 ) /* oy */
433 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */
434 MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */
435
436 ADDPS( XMM4, XMM3 )
437 ADDPS( XMM2, XMM3 )
438
439 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
440 UNPCKHPS( XMM3, XMM3 )
441 MOVSS( XMM3, D(2) ) /* ->D(2) */
442
443 LLBL(K_GTP23P3DR_skip):
444 ADD_L( CONST(16), EDI )
445 ADD_L( EAX, ESI )
446 CMP_L( ECX, EDI )
447 JNE( LLBL(K_GTP23P3DR_top) )
448
449 LLBL(K_GTP23P3DR_finish):
450 POP_L( EDI )
451 POP_L( ESI )
452 RET
453 #undef FRAME_OFFSET
454 #endif