90768ba1707d33ec28223f77c8d0fce9a287e86d
[mesa.git] / src / mesa / x86 / sse_xform1.S
1 /* $Id: sse_xform1.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_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 GLNAME( _mesa_sse_transform_points1_identity ):
104
105 #define FRAME_OFFSET 8
106 PUSH_L ( ESI )
107 PUSH_L ( EDI )
108
109 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
110 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
111
112 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
113
114 TEST_L( ECX, ECX)
115 JZ( LLBL(K_GTP1IR_finish) ) /* count was zero; go to finish */
116
117 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
118 OR_L( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
119
120 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
121 MOV_L( CONST(1), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
122
123 SHL_L( CONST(4), ECX ) /* count *= 16 */
124 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
125
126 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
127 ADD_L( EDI, ECX ) /* count += dest ptr */
128
129 CMP_L( ESI, EDI )
130 JE( LLBL(K_GTP1IR_finish) )
131
132
133 ALIGNTEXT32
134 LLBL(K_GTP1IR_top):
135 MOV_L( S(0), EDX )
136 MOV_L( EDX, D(0) )
137
138 LLBL(K_GTP1IR_skip):
139 ADD_L ( CONST(16), EDI )
140 ADD_L ( EAX, ESI )
141 CMP_L ( ECX, EDI )
142 JNE ( LLBL(K_GTP1IR_top) )
143
144 LLBL(K_GTP1IR_finish):
145 POP_L ( EDI )
146 POP_L ( ESI )
147 RET
148 #undef FRAME_OFFSET
149
150
151
152 ALIGNTEXT4
153 GLOBL GLNAME(_mesa_sse_transform_points1_3d_no_rot)
154 GLNAME(_mesa_sse_transform_points1_3d_no_rot):
155
156 #define FRAME_OFFSET 8
157 PUSH_L( ESI )
158 PUSH_L( EDI )
159
160 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
161 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
162
163 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
164 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
165
166 TEST_L( ECX, ECX)
167 JZ( LLBL(K_GTP13DNRR_finish) ) /* count was zero; go to finish */
168
169 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
170 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
171
172 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
173 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
174
175 SHL_L( CONST(4), ECX ) /* count *= 16 */
176 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
177
178 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
179 ADD_L( EDI, ECX ) /* count += dest ptr */
180
181
182 ALIGNTEXT32
183 MOVSS( M(0), XMM0 ) /* m0 */
184 MOVSS( M(12), XMM1 ) /* m12 */
185 MOVSS( M(13), XMM2 ) /* m13 */
186 MOVSS( M(14), XMM3 ) /* m14 */
187
188 ALIGNTEXT32
189 LLBL(K_GTP13DNRR_top):
190 MOVSS( S(0), XMM4 ) /* ox */
191 MULSS( XMM0, XMM4 ) /* ox*m0 */
192 ADDSS( XMM1, XMM4 ) /* ox*m0+m12 */
193 MOVSS( XMM4, D(0) )
194
195 MOVSS( XMM2, D(1) )
196 MOVSS( XMM3, D(2) )
197
198 LLBL(K_GTP13DNRR_skip):
199 ADD_L ( CONST(16), EDI )
200 ADD_L ( EAX, ESI )
201 CMP_L ( ECX, EDI )
202 JNE ( LLBL(K_GTP13DNRR_top) )
203
204 LLBL(K_GTP13DNRR_finish):
205 POP_L ( EDI )
206 POP_L ( ESI )
207 RET
208 #undef FRAME_OFFSET
209
210
211
212 ALIGNTEXT4
213 GLOBL GLNAME(_mesa_sse_transform_points1_perspective)
214 GLNAME(_mesa_sse_transform_points1_perspective):
215
216 #define FRAME_OFFSET 8
217 PUSH_L ( ESI )
218 PUSH_L ( EDI )
219
220 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
221 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
222
223 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
224 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
225
226 TEST_L( ECX, ECX)
227 JZ( LLBL(K_GTP13PR_finish) ) /* count was zero; go to finish */
228
229 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
230 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
231
232 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
233 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
234
235 SHL_L( CONST(4), ECX ) /* count *= 16 */
236 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
237
238 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
239 ADD_L( EDI, ECX ) /* count += dest ptr */
240
241
242 ALIGNTEXT32
243 XORPS( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
244 MOVSS( M(0), XMM1 ) /* m0 */
245 MOVSS( M(14), XMM2 ) /* m14 */
246
247 ALIGNTEXT32
248 LLBL(K_GTP13PR_top):
249 MOVSS( S(0), XMM3 ) /* ox */
250 MULSS( XMM1, XMM3 ) /* ox*m0 */
251 MOVSS( XMM3, D(0) ) /* ox*m0->D(0) */
252 MOVSS( XMM2, D(2) ) /* m14->D(2) */
253
254 MOVSS( XMM0, D(1) )
255 MOVSS( XMM0, D(3) )
256
257 LLBL(K_GTP13PR_skip):
258 ADD_L( CONST(16), EDI )
259 ADD_L( EAX, ESI )
260 CMP_L( ECX, EDI )
261 JNE( LLBL(K_GTP13PR_top) )
262
263 LLBL(K_GTP13PR_finish):
264 POP_L ( EDI )
265 POP_L ( ESI )
266 RET
267 #undef FRAME_OFFSET
268
269
270 ALIGNTEXT4
271 GLOBL GLNAME(_mesa_sse_transform_points1_2d)
272 GLNAME(_mesa_sse_transform_points1_2d):
273
274 #define FRAME_OFFSET 8
275 PUSH_L( ESI )
276 PUSH_L( EDI )
277
278 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
279 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
280
281 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
282 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
283
284 TEST_L( ECX, ECX)
285 JZ( LLBL(K_GTP13P2DR_finish) ) /* count was zero; go to finish */
286
287 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
288 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
289
290 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
291 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
292
293 SHL_L( CONST(4), ECX ) /* count *= 16 */
294 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
295
296 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
297 ADD_L( EDI, ECX ) /* count += dest ptr */
298
299 ALIGNTEXT32
300 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
301 MOVLPS( M(12), XMM1 ) /* m13 | m12 */
302
303 ALIGNTEXT32
304 LLBL(K_GTP13P2DR_top):
305 MOVSS( S(0), XMM2 ) /* ox */
306 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
307 MULPS( XMM0, XMM2 ) /* - | - | ox*m1 | ox*m0 */
308 ADDPS( XMM1, XMM2 ) /* - | - | ox*m1+m13 | ox*m0+m12 */
309 MOVLPS( XMM2, D(0) )
310
311 LLBL(K_GTP13P2DR_skip):
312 ADD_L ( CONST(16), EDI )
313 ADD_L ( EAX, ESI )
314 CMP_L ( ECX, EDI )
315 JNE ( LLBL(K_GTP13P2DR_top) )
316
317 LLBL(K_GTP13P2DR_finish):
318 POP_L ( EDI )
319 POP_L ( ESI )
320 RET
321 #undef FRAME_OFFSET
322
323
324 ALIGNTEXT4
325 GLOBL GLNAME(_mesa_sse_transform_points1_2d_no_rot)
326 GLNAME(_mesa_sse_transform_points1_2d_no_rot):
327
328 #define FRAME_OFFSET 8
329 PUSH_L( ESI )
330 PUSH_L( EDI )
331
332 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
333 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
334
335 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
336 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
337
338 TEST_L( ECX, ECX)
339 JZ( LLBL(K_GTP13P2DNRR_finish) ) /* count was zero; go to finish */
340
341 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
342 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
343
344 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
345 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
346
347 SHL_L( CONST(4), ECX ) /* count *= 16 */
348 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
349
350 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
351 ADD_L( EDI, ECX ) /* count += dest ptr */
352
353 ALIGNTEXT32
354 MOVSS( M(0), XMM0 ) /* m0 */
355 MOVSS( M(12), XMM1 ) /* m12 */
356 MOVSS( M(13), XMM2 ) /* m13 */
357
358 ALIGNTEXT32
359 LLBL(K_GTP13P2DNRR_top):
360 MOVSS( S(0), XMM3 ) /* ox */
361 MULSS( XMM0, XMM3 ) /* ox*m0 */
362 ADDSS( XMM1, XMM3 ) /* ox*m0+m12 */
363 MOVSS( XMM3, D(0) )
364 MOVSS( XMM2, D(1) )
365
366 LLBL(K_GTP13P2DNRR_skip):
367 ADD_L( CONST(16), EDI )
368 ADD_L( EAX, ESI )
369 CMP_L( ECX, EDI )
370 JNE( LLBL(K_GTP13P2DNRR_top) )
371
372 LLBL(K_GTP13P2DNRR_finish):
373 POP_L( EDI )
374 POP_L( ESI )
375 RET
376 #undef FRAME_OFFSET
377
378
379
380 ALIGNTEXT4
381 GLOBL GLNAME(_mesa_sse_transform_points1_3d)
382 GLNAME(_mesa_sse_transform_points1_3d):
383
384 #define FRAME_OFFSET 8
385 PUSH_L( ESI )
386 PUSH_L( EDI )
387
388 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
389 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
390
391 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
392 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
393
394 TEST_L( ECX, ECX)
395 JZ( LLBL(K_GTP13P3DR_finish) ) /* count was zero; go to finish */
396
397 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
398 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
399
400 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
401 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
402
403 SHL_L( CONST(4), ECX ) /* count *= 16 */
404 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
405
406 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
407 ADD_L( EDI, ECX ) /* count += dest ptr */
408
409
410 ALIGNTEXT32
411 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
412 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
413
414 ALIGNTEXT32
415 LLBL(K_GTP13P3DR_top):
416 MOVSS( S(0), XMM2 ) /* ox */
417 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
418 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
419 ADDPS( XMM1, XMM2 ) /* +m15 | +m14 | +m13 | +m12 */
420 MOVLPS( XMM2, D(0) ) /* - | - | ->D(1)| ->D(0)*/
421 UNPCKHPS( XMM2, XMM2 ) /* ox*m3+m15 | ox*m3+m15 | ox*m2+m14 | ox*m2+m14 */
422 MOVSS( XMM2, D(2) )
423
424 LLBL(K_GTP13P3DR_skip):
425 ADD_L( CONST(16), EDI )
426 ADD_L( EAX, ESI )
427 CMP_L( ECX, EDI )
428 JNE( LLBL(K_GTP13P3DR_top) )
429
430 LLBL(K_GTP13P3DR_finish):
431 POP_L( EDI )
432 POP_L( ESI )
433 RET
434 #undef FRAME_OFFSET
435 #endif