Merge branch 'gallium-vertex-linear' into gallium-0.1
[mesa.git] / src / mesa / x86 / sse_xform1.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_points1_general)
46 HIDDEN( _mesa_sse_transform_points1_general )
47 GLNAME( _mesa_sse_transform_points1_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 CMP_L( CONST(0), ECX ) /* count == 0 ? */
60 JE( LLBL(K_GTP1GR_finish) ) /* yes -> nothing to do. */
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
75 ALIGNTEXT32
76 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
77 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
78
79 ALIGNTEXT32
80 LLBL(K_GTP1GR_top):
81 MOVSS( S(0), XMM2 ) /* ox */
82 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
83 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
84 ADDPS( XMM1, XMM2 ) /* + | + | + | + */
85 MOVUPS( XMM2, D(0) )
86
87 LLBL(K_GTP1GR_skip):
88 ADD_L ( CONST(16), EDI )
89 ADD_L ( EAX, ESI )
90 CMP_L ( ECX, EDI )
91 JNE ( LLBL(K_GTP1GR_top) )
92
93 LLBL(K_GTP1GR_finish):
94 POP_L ( EDI )
95 POP_L ( ESI )
96 RET
97 #undef FRAME_OFFSET
98
99
100
101 ALIGNTEXT4
102 GLOBL GLNAME(_mesa_sse_transform_points1_identity)
103 HIDDEN(_mesa_sse_transform_points1_identity)
104 GLNAME( _mesa_sse_transform_points1_identity ):
105
106 #define FRAME_OFFSET 8
107 PUSH_L ( ESI )
108 PUSH_L ( EDI )
109
110 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
111 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
112
113 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
114
115 TEST_L( ECX, ECX)
116 JZ( LLBL(K_GTP1IR_finish) ) /* count was zero; go to finish */
117
118 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
119 OR_L( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
120
121 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
122 MOV_L( CONST(1), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
123
124 SHL_L( CONST(4), ECX ) /* count *= 16 */
125 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
126
127 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
128 ADD_L( EDI, ECX ) /* count += dest ptr */
129
130 CMP_L( ESI, EDI )
131 JE( LLBL(K_GTP1IR_finish) )
132
133
134 ALIGNTEXT32
135 LLBL(K_GTP1IR_top):
136 MOV_L( S(0), EDX )
137 MOV_L( EDX, D(0) )
138
139 LLBL(K_GTP1IR_skip):
140 ADD_L ( CONST(16), EDI )
141 ADD_L ( EAX, ESI )
142 CMP_L ( ECX, EDI )
143 JNE ( LLBL(K_GTP1IR_top) )
144
145 LLBL(K_GTP1IR_finish):
146 POP_L ( EDI )
147 POP_L ( ESI )
148 RET
149 #undef FRAME_OFFSET
150
151
152
153 ALIGNTEXT4
154 GLOBL GLNAME(_mesa_sse_transform_points1_3d_no_rot)
155 HIDDEN(_mesa_sse_transform_points1_3d_no_rot)
156 GLNAME(_mesa_sse_transform_points1_3d_no_rot):
157
158 #define FRAME_OFFSET 8
159 PUSH_L( ESI )
160 PUSH_L( EDI )
161
162 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
163 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
164
165 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
166 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
167
168 TEST_L( ECX, ECX)
169 JZ( LLBL(K_GTP13DNRR_finish) ) /* count was zero; go to finish */
170
171 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
172 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
173
174 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
175 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
176
177 SHL_L( CONST(4), ECX ) /* count *= 16 */
178 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
179
180 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
181 ADD_L( EDI, ECX ) /* count += dest ptr */
182
183
184 ALIGNTEXT32
185 MOVSS( M(0), XMM0 ) /* m0 */
186 MOVSS( M(12), XMM1 ) /* m12 */
187 MOVSS( M(13), XMM2 ) /* m13 */
188 MOVSS( M(14), XMM3 ) /* m14 */
189
190 ALIGNTEXT32
191 LLBL(K_GTP13DNRR_top):
192 MOVSS( S(0), XMM4 ) /* ox */
193 MULSS( XMM0, XMM4 ) /* ox*m0 */
194 ADDSS( XMM1, XMM4 ) /* ox*m0+m12 */
195 MOVSS( XMM4, D(0) )
196
197 MOVSS( XMM2, D(1) )
198 MOVSS( XMM3, D(2) )
199
200 LLBL(K_GTP13DNRR_skip):
201 ADD_L ( CONST(16), EDI )
202 ADD_L ( EAX, ESI )
203 CMP_L ( ECX, EDI )
204 JNE ( LLBL(K_GTP13DNRR_top) )
205
206 LLBL(K_GTP13DNRR_finish):
207 POP_L ( EDI )
208 POP_L ( ESI )
209 RET
210 #undef FRAME_OFFSET
211
212
213
214 ALIGNTEXT4
215 GLOBL GLNAME(_mesa_sse_transform_points1_perspective)
216 HIDDEN(_mesa_sse_transform_points1_perspective)
217 GLNAME(_mesa_sse_transform_points1_perspective):
218
219 #define FRAME_OFFSET 8
220 PUSH_L ( ESI )
221 PUSH_L ( EDI )
222
223 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
224 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
225
226 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
227 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
228
229 TEST_L( ECX, ECX)
230 JZ( LLBL(K_GTP13PR_finish) ) /* count was zero; go to finish */
231
232 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
233 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
234
235 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
236 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
237
238 SHL_L( CONST(4), ECX ) /* count *= 16 */
239 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
240
241 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
242 ADD_L( EDI, ECX ) /* count += dest ptr */
243
244
245 ALIGNTEXT32
246 XORPS( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
247 MOVSS( M(0), XMM1 ) /* m0 */
248 MOVSS( M(14), XMM2 ) /* m14 */
249
250 ALIGNTEXT32
251 LLBL(K_GTP13PR_top):
252 MOVSS( S(0), XMM3 ) /* ox */
253 MULSS( XMM1, XMM3 ) /* ox*m0 */
254 MOVSS( XMM3, D(0) ) /* ox*m0->D(0) */
255 MOVSS( XMM2, D(2) ) /* m14->D(2) */
256
257 MOVSS( XMM0, D(1) )
258 MOVSS( XMM0, D(3) )
259
260 LLBL(K_GTP13PR_skip):
261 ADD_L( CONST(16), EDI )
262 ADD_L( EAX, ESI )
263 CMP_L( ECX, EDI )
264 JNE( LLBL(K_GTP13PR_top) )
265
266 LLBL(K_GTP13PR_finish):
267 POP_L ( EDI )
268 POP_L ( ESI )
269 RET
270 #undef FRAME_OFFSET
271
272
273 ALIGNTEXT4
274 GLOBL GLNAME(_mesa_sse_transform_points1_2d)
275 HIDDEN(_mesa_sse_transform_points1_2d)
276 GLNAME(_mesa_sse_transform_points1_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_GTP13P2DR_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(12), XMM1 ) /* m13 | m12 */
306
307 ALIGNTEXT32
308 LLBL(K_GTP13P2DR_top):
309 MOVSS( S(0), XMM2 ) /* ox */
310 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
311 MULPS( XMM0, XMM2 ) /* - | - | ox*m1 | ox*m0 */
312 ADDPS( XMM1, XMM2 ) /* - | - | ox*m1+m13 | ox*m0+m12 */
313 MOVLPS( XMM2, D(0) )
314
315 LLBL(K_GTP13P2DR_skip):
316 ADD_L ( CONST(16), EDI )
317 ADD_L ( EAX, ESI )
318 CMP_L ( ECX, EDI )
319 JNE ( LLBL(K_GTP13P2DR_top) )
320
321 LLBL(K_GTP13P2DR_finish):
322 POP_L ( EDI )
323 POP_L ( ESI )
324 RET
325 #undef FRAME_OFFSET
326
327
328 ALIGNTEXT4
329 GLOBL GLNAME(_mesa_sse_transform_points1_2d_no_rot)
330 HIDDEN(_mesa_sse_transform_points1_2d_no_rot)
331 GLNAME(_mesa_sse_transform_points1_2d_no_rot):
332
333 #define FRAME_OFFSET 8
334 PUSH_L( ESI )
335 PUSH_L( EDI )
336
337 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
338 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
339
340 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
341 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
342
343 TEST_L( ECX, ECX)
344 JZ( LLBL(K_GTP13P2DNRR_finish) ) /* count was zero; go to finish */
345
346 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
347 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
348
349 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
350 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
351
352 SHL_L( CONST(4), ECX ) /* count *= 16 */
353 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
354
355 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
356 ADD_L( EDI, ECX ) /* count += dest ptr */
357
358 ALIGNTEXT32
359 MOVSS( M(0), XMM0 ) /* m0 */
360 MOVSS( M(12), XMM1 ) /* m12 */
361 MOVSS( M(13), XMM2 ) /* m13 */
362
363 ALIGNTEXT32
364 LLBL(K_GTP13P2DNRR_top):
365 MOVSS( S(0), XMM3 ) /* ox */
366 MULSS( XMM0, XMM3 ) /* ox*m0 */
367 ADDSS( XMM1, XMM3 ) /* ox*m0+m12 */
368 MOVSS( XMM3, D(0) )
369 MOVSS( XMM2, D(1) )
370
371 LLBL(K_GTP13P2DNRR_skip):
372 ADD_L( CONST(16), EDI )
373 ADD_L( EAX, ESI )
374 CMP_L( ECX, EDI )
375 JNE( LLBL(K_GTP13P2DNRR_top) )
376
377 LLBL(K_GTP13P2DNRR_finish):
378 POP_L( EDI )
379 POP_L( ESI )
380 RET
381 #undef FRAME_OFFSET
382
383
384
385 ALIGNTEXT4
386 GLOBL GLNAME(_mesa_sse_transform_points1_3d)
387 HIDDEN(_mesa_sse_transform_points1_3d)
388 GLNAME(_mesa_sse_transform_points1_3d):
389
390 #define FRAME_OFFSET 8
391 PUSH_L( ESI )
392 PUSH_L( EDI )
393
394 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
395 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
396
397 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
398 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
399
400 TEST_L( ECX, ECX)
401 JZ( LLBL(K_GTP13P3DR_finish) ) /* count was zero; go to finish */
402
403 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
404 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
405
406 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
407 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
408
409 SHL_L( CONST(4), ECX ) /* count *= 16 */
410 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
411
412 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
413 ADD_L( EDI, ECX ) /* count += dest ptr */
414
415
416 ALIGNTEXT32
417 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
418 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
419
420 ALIGNTEXT32
421 LLBL(K_GTP13P3DR_top):
422 MOVSS( S(0), XMM2 ) /* ox */
423 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
424 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
425 ADDPS( XMM1, XMM2 ) /* +m15 | +m14 | +m13 | +m12 */
426 MOVLPS( XMM2, D(0) ) /* - | - | ->D(1)| ->D(0)*/
427 UNPCKHPS( XMM2, XMM2 ) /* ox*m3+m15 | ox*m3+m15 | ox*m2+m14 | ox*m2+m14 */
428 MOVSS( XMM2, D(2) )
429
430 LLBL(K_GTP13P3DR_skip):
431 ADD_L( CONST(16), EDI )
432 ADD_L( EAX, ESI )
433 CMP_L( ECX, EDI )
434 JNE( LLBL(K_GTP13P3DR_top) )
435
436 LLBL(K_GTP13P3DR_finish):
437 POP_L( EDI )
438 POP_L( ESI )
439 RET
440 #undef FRAME_OFFSET
441 #endif
442
443 #if defined (__ELF__) && defined (__linux__)
444 .section .note.GNU-stack,"",%progbits
445 #endif