Consolidation of asm code in 3.5
[mesa.git] / src / mesa / x86 / sse_xform2.S
1 /* $Id: sse_xform2.S,v 1.1 2001/03/29 06:46:27 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 /** 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 #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 GLNAME( _mesa_sse_transform_points2_general ):
47
48 #define FRAME_OFFSET 8
49 PUSH_L ( ESI )
50 PUSH_L ( EDI )
51
52 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
53 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
54
55 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
56 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
57
58 TEST_L( ECX, ECX )
59 JZ( LLBL(K_GTP2GR_finish) ) /* count was zero; go to finish */
60
61 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
62 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
63
64 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
65 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
66
67 SHL_L( CONST(4), ECX ) /* count *= 16 */
68 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
69
70 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
71 ADD_L( EDI, ECX ) /* count += dest ptr */
72
73 ALIGNTEXT32
74 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
75 MOVAPS( M(4), XMM1 ) /* m7 | m6 | m5 | m4 */
76 MOVAPS( M(12), XMM2 ) /* m15 | m14 | m13 | m12 */
77
78 ALIGNTEXT32
79 LLBL(K_GTP2GR_top):
80 MOVSS( S(0), XMM3 ) /* ox */
81 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox | ox */
82 MULPS( XMM0, XMM3 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
83 MOVSS( S(1), XMM4 ) /* oy */
84 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy | oy */
85 MULPS( XMM1, XMM4 ) /* oy*m7 | oy*m6 | oy*m5 | oy*m4 */
86
87 ADDPS( XMM4, XMM3 )
88 ADDPS( XMM2, XMM3 )
89 MOVAPS( XMM3, D(0) )
90
91 LLBL(K_GTP2GR_skip):
92 ADD_L ( CONST(16), EDI )
93 ADD_L ( EAX, ESI )
94 CMP_L ( ECX, EDI )
95 JNE ( LLBL(K_GTP2GR_top) )
96
97 LLBL(K_GTP2GR_finish):
98 POP_L ( EDI )
99 POP_L ( ESI )
100 RET
101 #undef FRAME_OFFSET
102
103
104 ALIGNTEXT4
105 GLOBL GLNAME(_mesa_sse_transform_points2_identity)
106 GLNAME( _mesa_sse_transform_points2_identity ):
107
108 #define FRAME_OFFSET 8
109 PUSH_L ( ESI )
110 PUSH_L ( EDI )
111
112 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
113 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
114
115 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
116
117 TEST_L( ECX, ECX)
118 JZ( LLBL(K_GTP2IR_finish) ) /* count was zero; go to finish */
119
120 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
121 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
122
123 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
124 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
125
126 SHL_L( CONST(4), ECX ) /* count *= 16 */
127 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
128
129 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
130 ADD_L( EDI, ECX ) /* count += dest ptr */
131
132 CMP_L( ESI, EDI )
133 JE( LLBL(K_GTP2IR_finish) )
134
135
136 ALIGNTEXT32
137 LLBL(K_GTP2IR_top):
138 MOV_L ( S(0), EDX )
139 MOV_L ( EDX, D(0) )
140 MOV_L ( S(1), EDX )
141 MOV_L ( EDX, D(1) )
142
143 LLBL(K_GTP2IR_skip):
144 ADD_L ( CONST(16), EDI )
145 ADD_L ( EAX, ESI )
146 CMP_L ( ECX, EDI )
147 JNE ( LLBL(K_GTP2IR_top) )
148
149 LLBL(K_GTP2IR_finish):
150 POP_L ( EDI )
151 POP_L ( ESI )
152 RET
153 #undef FRAME_OFFSET
154
155
156 ALIGNTEXT4
157 GLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot)
158 GLNAME(_mesa_sse_transform_points2_3d_no_rot):
159
160 #define FRAME_OFFSET 8
161 PUSH_L( ESI )
162 PUSH_L( EDI )
163
164 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
165 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
166
167 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
168 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
169
170 TEST_L( ECX, ECX)
171 JZ( LLBL(K_GTP23DNRR_finish) ) /* count was zero; go to finish */
172
173 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
174 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
175
176 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
177 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
178
179 SHL_L( CONST(4), ECX ) /* count *= 16 */
180 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
181
182 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
183 ADD_L( EDI, ECX ) /* count += dest ptr */
184
185
186 ALIGNTEXT32
187 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
188 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
189 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
190 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
191 MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */
192
193 ALIGNTEXT32
194 LLBL(K_GTP23DNRR_top):
195 MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */
196 MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */
197 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
198 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
199
200 MOVSS ( XMM3, D(2) ) /* -> D(2) */
201
202 LLBL(K_GTP23DNRR_skip):
203 ADD_L ( CONST(16), EDI )
204 ADD_L ( EAX, ESI )
205 CMP_L ( ECX, EDI )
206 JNE ( LLBL(K_GTP23DNRR_top) )
207
208 LLBL(K_GTP23DNRR_finish):
209 POP_L ( EDI )
210 POP_L ( ESI )
211 RET
212 #undef FRAME_OFFSET
213
214
215 ALIGNTEXT4
216 GLOBL GLNAME(_mesa_sse_transform_points2_perspective)
217 GLNAME(_mesa_sse_transform_points2_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_GTP23PR_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 ALIGNTEXT32
245 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
246 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
247 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
248 MOVSS ( M(14), XMM3 ) /* m14 */
249 XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
250
251 ALIGNTEXT32
252 LLBL(K_GTP23PR_top):
253 MOVLPS( S(0), XMM4 ) /* oy | ox */
254 MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */
255 MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */
256 MOVSS( XMM3, D(2) ) /* ->D(2) */
257 MOVSS( XMM0, D(3) ) /* ->D(3) */
258
259 LLBL(K_GTP23PR_skip):
260 ADD_L( CONST(16), EDI )
261 ADD_L( EAX, ESI )
262 CMP_L( ECX, EDI )
263 JNE( LLBL(K_GTP23PR_top) )
264
265 LLBL(K_GTP23PR_finish):
266 POP_L ( EDI )
267 POP_L ( ESI )
268 RET
269 #undef FRAME_OFFSET
270
271
272
273 ALIGNTEXT4
274 GLOBL GLNAME(_mesa_sse_transform_points2_2d)
275 GLNAME(_mesa_sse_transform_points2_2d):
276
277 #define FRAME_OFFSET 8
278 PUSH_L( ESI )
279 PUSH_L( EDI )
280
281 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
282 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
283
284 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
285 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
286
287 TEST_L( ECX, ECX)
288 JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */
289
290 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
291 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
292
293 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
294 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
295
296 SHL_L( CONST(4), ECX ) /* count *= 16 */
297 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
298
299 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
300 ADD_L( EDI, ECX ) /* count += dest ptr */
301
302 ALIGNTEXT32
303 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
304 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
305 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
306
307 ALIGNTEXT32
308 LLBL(K_GTP23P2DR_top):
309 MOVSS( S(0), XMM3 ) /* ox */
310 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
311 MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
312
313 MOVSS( S(1), XMM4 ) /* oy */
314 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
315 MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
316
317 ADDPS( XMM4, XMM3 )
318 ADDPS( XMM2, XMM3 )
319 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
320
321 LLBL(K_GTP23P2DR_skip):
322 ADD_L ( CONST(16), EDI )
323 ADD_L ( EAX, ESI )
324 CMP_L ( ECX, EDI )
325 JNE ( LLBL(K_GTP23P2DR_top) )
326
327 LLBL(K_GTP23P2DR_finish):
328 POP_L ( EDI )
329 POP_L ( ESI )
330 RET
331 #undef FRAME_OFFSET
332
333
334
335 ALIGNTEXT4
336 GLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot)
337 GLNAME(_mesa_sse_transform_points2_2d_no_rot):
338
339 #define FRAME_OFFSET 8
340 PUSH_L( ESI )
341 PUSH_L( EDI )
342
343 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
344 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
345
346 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
347 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
348
349 TEST_L( ECX, ECX)
350 JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */
351
352 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
353 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
354
355 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
356 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
357
358 SHL_L( CONST(4), ECX ) /* count *= 16 */
359 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
360
361 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
362 ADD_L( EDI, ECX ) /* count += dest ptr */
363
364 ALIGNTEXT32
365 MOVSS ( M(0), XMM1 ) /* m0 */
366 MOVSS ( M(5), XMM2 ) /* m5 */
367 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
368 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
369
370 ALIGNTEXT32
371 LLBL(K_GTP23P2DNRR_top):
372 MOVLPS( S(0), XMM0 ) /* oy | ox */
373 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
374 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
375 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
376
377 LLBL(K_GTP23P2DNRR_skip):
378 ADD_L( CONST(16), EDI )
379 ADD_L( EAX, ESI )
380 CMP_L( ECX, EDI )
381 JNE( LLBL(K_GTP23P2DNRR_top) )
382
383 LLBL(K_GTP23P2DNRR_finish):
384 POP_L( EDI )
385 POP_L( ESI )
386 RET
387 #undef FRAME_OFFSET
388
389
390
391 ALIGNTEXT4
392 GLOBL GLNAME(_mesa_sse_transform_points2_3d)
393 GLNAME(_mesa_sse_transform_points2_3d):
394
395 #define FRAME_OFFSET 8
396 PUSH_L( ESI )
397 PUSH_L( EDI )
398
399 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
400 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
401
402 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
403 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
404
405 TEST_L( ECX, ECX)
406 JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */
407
408 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
409 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
410
411 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
412 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
413
414 SHL_L( CONST(4), ECX ) /* count *= 16 */
415 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
416
417 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
418 ADD_L( EDI, ECX ) /* count += dest ptr */
419
420 ALIGNTEXT32
421 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
422 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
423 MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */
424
425 ALIGNTEXT32
426 LLBL(K_GTP23P3DR_top):
427 MOVSS( S(0), XMM3 ) /* ox */
428 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */
429 MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */
430
431 MOVSS( S(1), XMM4 ) /* oy */
432 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */
433 MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */
434
435 ADDPS( XMM4, XMM3 )
436 ADDPS( XMM2, XMM3 )
437
438 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
439 UNPCKHPS( XMM3, XMM3 )
440 MOVSS( XMM3, D(2) ) /* ->D(2) */
441
442 LLBL(K_GTP23P3DR_skip):
443 ADD_L( CONST(16), EDI )
444 ADD_L( EAX, ESI )
445 CMP_L( ECX, EDI )
446 JNE( LLBL(K_GTP23P3DR_top) )
447
448 LLBL(K_GTP23P3DR_finish):
449 POP_L( EDI )
450 POP_L( ESI )
451 RET
452 #undef FRAME_OFFSET