3 * Mesa 3-D graphics library
5 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice shall be included
15 * in all copies or substantial portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
21 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
22 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
23 * OTHER DEALINGS IN THE SOFTWARE.
28 #define MATH_ASM_PTR_SIZE 4
29 #include "math/m_vector_asm.h"
30 #include "xform_args.h"
34 #define FRAME_OFFSET 4
38 GLOBL GLNAME( _mesa_3dnow_transform_points4_general )
39 HIDDEN(_mesa_3dnow_transform_points4_general)
40 GLNAME( _mesa_3dnow_transform_points4_general ):
44 MOV_L ( ARG_DEST, ECX )
45 MOV_L ( ARG_MATRIX, ESI )
46 MOV_L ( ARG_SOURCE, EAX )
47 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
48 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
49 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
50 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
54 MOV_L ( REGOFF(V4F_START, ECX), EDX )
56 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
57 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
58 MOV_L ( REGOFF(V4F_START, EAX), EAX )
61 JZ ( LLBL( G3TPGR_2 ) )
63 PREFETCHW ( REGIND(EDX) )
68 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
70 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
71 MOVQ ( REGOFF(8, EAX), MM4 ) /* x3 | x2 */
73 ADD_L ( EDI, EAX ) /* next vertex */
74 PREFETCH ( REGIND(EAX) )
76 MOVQ ( MM0, MM2 ) /* x1 | x0 */
77 MOVQ ( MM4, MM6 ) /* x3 | x2 */
79 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
80 PUNPCKHDQ ( MM2, MM2 ) /* x1 | x1 */
82 MOVQ ( MM0, MM1 ) /* x0 | x0 */
83 ADD_L ( CONST(16), EDX ) /* next r */
85 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
86 MOVQ ( MM2, MM3 ) /* x1 | x1 */
88 PFMUL ( REGOFF(8, ECX), MM1 ) /* x0*m3 | x0*m2 */
89 PUNPCKLDQ ( MM4, MM4 ) /* x2 | x2 */
91 PFMUL ( REGOFF(16, ECX), MM2 ) /* x1*m5 | x1*m4 */
92 MOVQ ( MM4, MM5 ) /* x2 | x2 */
94 PFMUL ( REGOFF(24, ECX), MM3 ) /* x1*m7 | x1*m6 */
95 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
97 PFMUL ( REGOFF(32, ECX), MM4 ) /* x2*m9 | x2*m8 */
98 MOVQ ( MM6, MM7 ) /* x3 | x3 */
100 PFMUL ( REGOFF(40, ECX), MM5 ) /* x2*m11 | x2*m10 */
103 PFMUL ( REGOFF(48, ECX), MM6 ) /* x3*m13 | x3*m12 */
106 PFMUL ( REGOFF(56, ECX), MM7 ) /* x3*m15 | x3*m14 */
113 MOVQ ( MM6, REGOFF(-16, EDX) )
115 MOVQ ( MM7, REGOFF(-8, EDX) )
117 DEC_L ( ESI ) /* decrement vertex counter */
118 JNZ ( LLBL( G3TPGR_1 ) ) /* cnt > 0 ? -> process next vertex */
131 GLOBL GLNAME( _mesa_3dnow_transform_points4_perspective )
132 HIDDEN(_mesa_3dnow_transform_points4_perspective)
133 GLNAME( _mesa_3dnow_transform_points4_perspective ):
137 MOV_L ( ARG_DEST, ECX )
138 MOV_L ( ARG_MATRIX, ESI )
139 MOV_L ( ARG_SOURCE, EAX )
140 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
141 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
142 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
143 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
147 MOV_L ( REGOFF(V4F_START, ECX), EDX )
149 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
150 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
151 MOV_L ( REGOFF(V4F_START, EAX), EAX )
154 JZ ( LLBL( G3TPPR_2 ) )
156 PREFETCH ( REGIND(EAX) )
157 PREFETCHW ( REGIND(EDX) )
159 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
160 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
162 MOVD ( REGOFF(40, ECX), MM1 ) /* | m22 */
163 PUNPCKLDQ ( REGOFF(56, ECX), MM1 ) /* m32 | m22 */
165 MOVQ ( REGOFF(32, ECX), MM2 ) /* m21 | m20 */
166 PXOR ( MM7, MM7 ) /* 0 | 0 */
171 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
173 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
174 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
175 MOVD ( REGOFF(8, EAX), MM3 ) /* | x2 */
177 ADD_L ( EDI, EAX ) /* next vertex */
178 PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */
180 MOVQ ( MM5, MM6 ) /* x3 | x2 */
181 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
183 PUNPCKLDQ ( MM5, MM5 ) /* x2 | x2 */
184 ADD_L ( CONST(16), EDX ) /* next r */
186 PFMUL ( MM2, MM5 ) /* x2*m21 | x2*m20 */
187 PFSUBR ( MM7, MM3 ) /* | -x2 */
189 PFMUL ( MM1, MM6 ) /* x3*m32 | x2*m22 */
190 PFADD ( MM4, MM5 ) /* x1*m11+x2*m21 | x0*m00+x2*m20 */
192 PFACC ( MM3, MM6 ) /* -x2 | x2*m22+x3*m32 */
193 MOVQ ( MM5, REGOFF(-16, EDX) ) /* write r0, r1 */
195 MOVQ ( MM6, REGOFF(-8, EDX) ) /* write r2, r3 */
196 DEC_L ( ESI ) /* decrement vertex counter */
198 JNZ ( LLBL( G3TPPR_1 ) ) /* cnt > 0 ? -> process next vertex */
211 GLOBL GLNAME( _mesa_3dnow_transform_points4_3d )
212 HIDDEN(_mesa_3dnow_transform_points4_3d)
213 GLNAME( _mesa_3dnow_transform_points4_3d ):
217 MOV_L ( ARG_DEST, ECX )
218 MOV_L ( ARG_MATRIX, ESI )
219 MOV_L ( ARG_SOURCE, EAX )
220 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
221 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
222 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
223 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
227 MOV_L ( REGOFF(V4F_START, ECX), EDX )
229 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
230 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
231 MOV_L ( REGOFF(V4F_START, EAX), EAX )
234 JZ ( LLBL( G3TP3R_2 ) )
236 MOVD ( REGOFF(8, ECX), MM6 ) /* | m2 */
237 PUNPCKLDQ ( REGOFF(24, ECX), MM6 ) /* m6 | m2 */
239 MOVD ( REGOFF(40, ECX), MM7 ) /* | m10 */
240 PUNPCKLDQ ( REGOFF(56, ECX), MM7 ) /* m14 | m10 */
245 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
246 PREFETCH ( REGOFF(32, EAX) ) /* hopefully array is tightly packed */
248 MOVQ ( REGIND(EAX), MM2 ) /* x1 | x0 */
249 MOVQ ( REGOFF(8, EAX), MM3 ) /* x3 | x2 */
251 MOVQ ( MM2, MM0 ) /* x1 | x0 */
252 MOVQ ( MM3, MM4 ) /* x3 | x2 */
254 MOVQ ( MM0, MM1 ) /* x1 | x0 */
255 MOVQ ( MM4, MM5 ) /* x3 | x2 */
257 PUNPCKLDQ ( MM0, MM0 ) /* x0 | x0 */
258 PUNPCKHDQ ( MM1, MM1 ) /* x1 | x1 */
260 PFMUL ( REGIND(ECX), MM0 ) /* x0*m1 | x0*m0 */
261 PUNPCKLDQ ( MM3, MM3 ) /* x2 | x2 */
263 PFMUL ( REGOFF(16, ECX), MM1 ) /* x1*m5 | x1*m4 */
264 PUNPCKHDQ ( MM4, MM4 ) /* x3 | x3 */
266 PFMUL ( MM6, MM2 ) /* x1*m6 | x0*m2 */
267 PFADD ( MM0, MM1 ) /* x0*m1+x1*m5 | x0*m0+x1*m4 */
269 PFMUL ( REGOFF(32, ECX), MM3 ) /* x2*m9 | x2*m8 */
270 ADD_L ( CONST(16), EDX ) /* next r */
272 PFMUL ( REGOFF(48, ECX), MM4 ) /* x3*m13 | x3*m12 */
273 PFADD ( MM1, MM3 ) /* x0*m1+..+x2*m9 | x0*m0+...+x2*m8 */
275 PFMUL ( MM7, MM5 ) /* x3*m14 | x2*m10 */
276 PFADD ( MM3, MM4 ) /* r1 | r0 */
278 PFACC ( MM2, MM5 ) /* x0*m2+x1*m6 | x2*m10+x3*m14 */
279 MOVD ( REGOFF(12, EAX), MM0 ) /* | x3 */
281 ADD_L ( EDI, EAX ) /* next vertex */
282 PFACC ( MM0, MM5 ) /* r3 | r2 */
284 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
285 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
287 DEC_L ( ESI ) /* decrement vertex counter */
288 JNZ ( LLBL( G3TP3R_1 ) ) /* cnt > 0 ? -> process next vertex */
301 GLOBL GLNAME( _mesa_3dnow_transform_points4_3d_no_rot )
302 HIDDEN(_mesa_3dnow_transform_points4_3d_no_rot)
303 GLNAME( _mesa_3dnow_transform_points4_3d_no_rot ):
306 MOV_L ( ARG_DEST, ECX )
307 MOV_L ( ARG_MATRIX, ESI )
308 MOV_L ( ARG_SOURCE, EAX )
309 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
310 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
311 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
312 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
316 MOV_L ( REGOFF(V4F_START, ECX), EDX )
318 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
319 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
320 MOV_L ( REGOFF(V4F_START, EAX), EAX )
323 JZ ( LLBL( G3TP3NRR_2 ) )
325 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
326 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
328 MOVD ( REGOFF(40, ECX), MM2 ) /* | m22 */
329 PUNPCKLDQ ( REGOFF(56, ECX), MM2 ) /* m32 | m22 */
331 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
336 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
338 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
339 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
340 MOVD ( REGOFF(12, EAX), MM7 ) /* | x3 */
342 ADD_L ( EDI, EAX ) /* next vertex */
343 PREFETCH ( REGOFF(32, EAX) ) /* hopefully stride is zero */
345 MOVQ ( MM5, MM6 ) /* x3 | x2 */
346 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
348 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
349 PFMUL ( MM2, MM5 ) /* x3*m32 | x2*m22 */
351 PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */
352 PFACC ( MM7, MM5 ) /* x3 | x2*m22+x3*m32 */
354 PFADD ( MM6, MM4 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */
355 ADD_L ( CONST(16), EDX ) /* next r */
357 MOVQ ( MM4, REGOFF(-16, EDX) ) /* write r0, r1 */
358 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
360 DEC_L ( ESI ) /* decrement vertex counter */
361 JNZ ( LLBL( G3TP3NRR_1 ) ) /* cnt > 0 ? -> process next vertex */
374 GLOBL GLNAME( _mesa_3dnow_transform_points4_2d )
375 HIDDEN(_mesa_3dnow_transform_points4_2d)
376 GLNAME( _mesa_3dnow_transform_points4_2d ):
380 MOV_L ( ARG_DEST, ECX )
381 MOV_L ( ARG_MATRIX, ESI )
382 MOV_L ( ARG_SOURCE, EAX )
383 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
384 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
385 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
386 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
390 MOV_L ( REGOFF(V4F_START, ECX), EDX )
392 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
393 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
394 MOV_L ( REGOFF(V4F_START, EAX), EAX )
397 JZ ( LLBL( G3TP2R_2 ) )
399 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
400 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
402 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
403 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
405 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
410 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
412 MOVQ ( REGIND(EAX), MM3 ) /* x1 | x0 */
413 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
415 ADD_L ( EDI, EAX ) /* next vertex */
416 PREFETCH ( REGIND(EAX) )
418 MOVQ ( MM3, MM4 ) /* x1 | x0 */
419 MOVQ ( MM5, MM6 ) /* x3 | x2 */
421 PFMUL ( MM1, MM4 ) /* x1*m11 | x0*m01 */
422 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
424 PFMUL ( MM0, MM3 ) /* x1*m10 | x0*m00 */
425 ADD_L ( CONST(16), EDX ) /* next r */
427 PFACC ( MM4, MM3 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */
428 PFMUL ( MM2, MM6 ) /* x3*m31 | x3*m30 */
430 PFADD ( MM6, MM3 ) /* r1 | r0 */
431 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
433 MOVQ ( MM3, REGOFF(-16, EDX) ) /* write r0, r1 */
435 DEC_L ( ESI ) /* decrement vertex counter */
436 JNZ ( LLBL( G3TP2R_1 ) ) /* cnt > 0 ? -> process next vertex */
449 GLOBL GLNAME( _mesa_3dnow_transform_points4_2d_no_rot )
450 HIDDEN(_mesa_3dnow_transform_points4_2d_no_rot)
451 GLNAME( _mesa_3dnow_transform_points4_2d_no_rot ):
455 MOV_L ( ARG_DEST, ECX )
456 MOV_L ( ARG_MATRIX, ESI )
457 MOV_L ( ARG_SOURCE, EAX )
458 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
459 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
460 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
461 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
465 MOV_L ( REGOFF(V4F_START, ECX), EDX )
467 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
468 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
469 MOV_L ( REGOFF(V4F_START, EAX), EAX )
472 JZ ( LLBL( G3TP2NRR_3 ) )
474 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
475 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
477 MOVQ ( REGOFF(48, ECX), MM1 ) /* m31 | m30 */
482 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
484 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
485 MOVQ ( REGOFF(8, EAX), MM5 ) /* x3 | x2 */
487 ADD_L ( EDI, EAX ) /* next vertex */
488 PREFETCH ( REGIND(EAX) )
490 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
491 MOVQ ( MM5, MM6 ) /* x3 | x2 */
493 ADD_L ( CONST(16), EDX ) /* next r */
494 PUNPCKHDQ ( MM6, MM6 ) /* x3 | x3 */
496 PFMUL ( MM1, MM6 ) /* x3*m31 | x3*m30 */
497 PFADD ( MM4, MM6 ) /* x1*m11+x3*m31 | x0*m00+x3*m30 */
499 MOVQ ( MM6, REGOFF(-16, EDX) ) /* write r0, r1 */
500 MOVQ ( MM5, REGOFF(-8, EDX) ) /* write r2, r3 */
502 DEC_L ( ESI ) /* decrement vertex counter */
504 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
517 GLOBL GLNAME( _mesa_3dnow_transform_points4_identity )
518 HIDDEN(_mesa_3dnow_transform_points4_identity)
519 GLNAME( _mesa_3dnow_transform_points4_identity ):
523 MOV_L ( ARG_DEST, ECX )
524 MOV_L ( ARG_MATRIX, ESI )
525 MOV_L ( ARG_SOURCE, EAX )
526 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
527 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
528 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
529 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
533 MOV_L ( REGOFF(V4F_START, ECX), EDX )
535 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
536 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
537 MOV_L ( REGOFF(V4F_START, EAX), EAX )
540 JZ ( LLBL( G3TPIR_2 ) )
545 PREFETCHW ( REGOFF(32, EDX) ) /* prefetch 2 vertices ahead */
547 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
548 MOVQ ( REGOFF(8, EAX), MM1 ) /* x3 | x2 */
550 ADD_L ( EDI, EAX ) /* next vertex */
551 PREFETCH ( REGIND(EAX) )
553 ADD_L ( CONST(16), EDX ) /* next r */
554 MOVQ ( MM0, REGOFF(-16, EDX) ) /* r1 | r0 */
556 MOVQ ( MM1, REGOFF(-8, EDX) ) /* r3 | r2 */
558 DEC_L ( ESI ) /* decrement vertex counter */
559 JNZ ( LLBL( G3TPIR_1 ) ) /* cnt > 0 ? -> process next vertex */
569 #if defined (__ELF__) && defined (__linux__)
570 .section .note.GNU-stack,"",%progbits