1 /* $Id: 3dnow_xform2.S,v 1.1 2001/03/29 06:46:16 gareth Exp $ */
4 * Mesa 3-D graphics library
7 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
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:
16 * The above copyright notice and this permission notice shall be included
17 * in all copies or substantial portions of the Software.
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.
28 #include "xform_args.h"
32 #define FRAME_OFFSET 4
36 GLOBL GLNAME( _mesa_3dnow_transform_points2_general )
37 GLNAME( _mesa_3dnow_transform_points2_general ):
41 MOV_L ( ARG_DEST, ECX )
42 MOV_L ( ARG_MATRIX, ESI )
43 MOV_L ( ARG_SOURCE, EAX )
44 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
45 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
46 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
47 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
51 MOV_L ( REGOFF(V4F_START, ECX), EDX )
53 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
54 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
55 MOV_L ( REGOFF(V4F_START, EAX), EAX )
58 JZ ( LLBL( G3TPGR_3 ) )
60 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
61 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
63 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
64 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
66 MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */
67 PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */
69 MOVD ( REGOFF(12, ECX), MM3 ) /* | m03 */
70 PUNPCKLDQ ( REGOFF(28, ECX), MM3 ) /* m13 | m03 */
72 MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */
73 MOVQ ( REGOFF(56, ECX), MM5 ) /* m33 | m32 */
78 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
79 MOVQ ( MM6, MM7 ) /* x1 | x0 */
81 PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */
82 PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */
84 PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */
85 PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */
87 MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */
88 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
90 MOVQ ( MM6, MM7 ) /* x1 | x0 */
91 PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */
93 PFMUL ( MM3, MM7 ) /* x1*m13 | x0*m03 */
94 ADD_L ( EDI, EAX ) /* next vertex */
96 PFACC ( MM7, MM6 ) /* x0*m03+x1*m13 | x0*x02+x1*m12 */
97 PFADD ( MM5, MM6 ) /* x0*...*m13+m33 | x0*...*m12+m32 */
99 MOVQ ( MM6, REGOFF(8, EDX) ) /* write r3, r2 */
100 ADD_L ( CONST(16), EDX ) /* next r */
102 DEC_L ( ESI ) /* decrement vertex counter */
103 JNZ ( LLBL( G3TPGR_2 ) ) /* cnt > 0 ? -> process next vertex */
116 GLOBL GLNAME( _mesa_3dnow_transform_points2_perspective )
117 GLNAME( _mesa_3dnow_transform_points2_perspective ):
121 MOV_L ( ARG_DEST, ECX )
122 MOV_L ( ARG_MATRIX, ESI )
123 MOV_L ( ARG_SOURCE, EAX )
124 MOV_L ( CONST(4), REGOFF(V4F_SIZE, ECX) )
125 OR_B ( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, ECX) )
126 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
127 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
131 MOV_L ( REGOFF(V4F_START, ECX), EDX )
133 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
134 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
135 MOV_L ( REGOFF(V4F_START, EAX), EAX )
138 JZ ( LLBL( G3TPPR_3 ) )
140 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
141 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
143 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
148 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
149 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
151 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
152 MOVQ ( MM3, REGOFF(8, EDX) ) /* write r2 (=m32), r3 (=0) */
154 ADD_L ( EDI, EAX ) /* next vertex */
155 ADD_L ( CONST(16), EDX ) /* next r */
157 DEC_L ( ESI ) /* decrement vertex counter */
158 JNZ ( LLBL( G3TPPR_2 ) ) /* cnt > 0 ? -> process next vertex */
171 GLOBL GLNAME( _mesa_3dnow_transform_points2_3d )
172 GLNAME( _mesa_3dnow_transform_points2_3d ):
176 MOV_L ( ARG_DEST, ECX )
177 MOV_L ( ARG_MATRIX, ESI )
178 MOV_L ( ARG_SOURCE, EAX )
179 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
180 OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) )
181 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
182 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
186 MOV_L ( REGOFF(V4F_START, ECX), EDX )
188 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
189 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
190 MOV_L ( REGOFF(V4F_START, EAX), EAX )
193 JZ ( LLBL( G3TP3R_3 ) )
195 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
196 PUNPCKLDQ ( REGOFF(16, ECX), MM0 ) /* m10 | m00 */
198 MOVD ( REGOFF(4, ECX), MM1 ) /* | m01 */
199 PUNPCKLDQ ( REGOFF(20, ECX), MM1 ) /* m11 | m01 */
201 MOVD ( REGOFF(8, ECX), MM2 ) /* | m02 */
202 PUNPCKLDQ ( REGOFF(24, ECX), MM2 ) /* m12 | m02 */
204 MOVQ ( REGOFF(48, ECX), MM4 ) /* m31 | m30 */
205 MOVD ( REGOFF(56, ECX), MM5 ) /* | m32 */
210 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
211 MOVQ ( MM6, MM7 ) /* x1 | x0 */
213 PFMUL ( MM0, MM6 ) /* x1*m10 | x0*m00 */
214 PFMUL ( MM1, MM7 ) /* x1*m11 | x0*m01 */
216 PFACC ( MM7, MM6 ) /* x0*m01+x1*m11 | x0*x00+x1*m10 */
217 PFADD ( MM4, MM6 ) /* x0*...*m11+m31 | x0*...*m10+m30 */
219 MOVQ ( MM6, REGIND(EDX) ) /* write r1, r0 */
220 MOVQ ( REGIND(EAX), MM6 ) /* x1 | x0 */
222 MOVQ ( MM6, MM7 ) /* x1 | x0 */
223 PFMUL ( MM2, MM6 ) /* x1*m12 | x0*m02 */
225 PFACC ( MM7, MM6 ) /* ***trash*** | x0*x02+x1*m12 */
226 PFADD ( MM5, MM6 ) /* ***trash*** | x0*...*m12+m32 */
228 MOVD ( MM6, REGOFF(8, EDX) ) /* write r2 */
229 ADD_L ( EDI, EAX ) /* next vertex */
231 ADD_L ( CONST(16), EDX ) /* next r */
232 DEC_L ( ESI ) /* decrement vertex counter */
234 JNZ ( LLBL( G3TP3R_2 ) ) /* cnt > 0 ? -> process next vertex */
247 GLOBL GLNAME( _mesa_3dnow_transform_points2_3d_no_rot )
248 GLNAME( _mesa_3dnow_transform_points2_3d_no_rot ):
252 MOV_L ( ARG_DEST, ECX )
253 MOV_L ( ARG_MATRIX, ESI )
254 MOV_L ( ARG_SOURCE, EAX )
255 MOV_L ( CONST(3), REGOFF(V4F_SIZE, ECX) )
256 OR_B ( CONST(VEC_SIZE_3 ), REGOFF(V4F_FLAGS, ECX) )
257 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
258 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
262 MOV_L ( REGOFF(V4F_START, ECX), EDX )
264 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
265 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
266 MOV_L ( REGOFF(V4F_START, EAX), EAX )
269 JZ ( LLBL( G3TP3NRR_3 ) )
271 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
272 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
274 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
275 MOVD ( REGOFF(56, ECX), MM3 ) /* | m32 */
280 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
281 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
283 PFADD ( MM2, MM4 ) /* x1*m11+m31 | x0*m00+m30 */
284 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
286 MOVD ( MM3, REGOFF(8, EDX) ) /* write r2 */
287 ADD_L ( EDI, EAX ) /* next vertex */
289 ADD_L ( CONST(16), EDX ) /* next r */
290 DEC_L ( ESI ) /* decrement vertex counter */
292 JNZ ( LLBL( G3TP3NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
305 GLOBL GLNAME( _mesa_3dnow_transform_points2_2d )
306 GLNAME( _mesa_3dnow_transform_points2_2d ):
310 MOV_L ( ARG_DEST, ECX )
311 MOV_L ( ARG_MATRIX, ESI )
312 MOV_L ( ARG_SOURCE, EAX )
313 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
314 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
315 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
316 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
320 MOV_L ( REGOFF(V4F_START, ECX), EDX )
322 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
323 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
324 MOV_L ( REGOFF(V4F_START, EAX), EAX )
327 JZ ( LLBL( G3TP2R_3 ) )
329 MOVQ ( REGIND(ECX), MM0 ) /* m01 | m00 */
330 MOVQ ( REGOFF(16, ECX), MM1 ) /* m11 | m10 */
332 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
337 MOVD ( REGIND(EAX), MM4 ) /* | x0 */
338 MOVD ( REGOFF(4, EAX), MM5 ) /* | x1 */
340 PUNPCKLDQ ( MM4, MM4 ) /* x0 | x0 */
341 ADD_L ( EDI, EAX ) /* next vertex */
343 PFMUL ( MM0, MM4 ) /* x0*m01 | x0*m00 */
344 PUNPCKLDQ ( MM5, MM5 ) /* x1 | x1 */
346 PFMUL ( MM1, MM5 ) /* x1*m11 | x1*m10 */
347 PFADD ( MM2, MM4 ) /* x...x1*m11+31 | x0*..*m10+m30 */
349 PFADD ( MM5, MM4 ) /* x0*m01+x1*m11 | x0*m00+x1*m10 */
350 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
352 ADD_L ( CONST(16), EDX ) /* next r */
353 DEC_L ( ESI ) /* decrement vertex counter */
355 JNZ ( LLBL( G3TP2R_2 ) ) /* cnt > 0 ? -> process next vertex */
368 GLOBL GLNAME( _mesa_3dnow_transform_points2_2d_no_rot )
369 GLNAME( _mesa_3dnow_transform_points2_2d_no_rot ):
373 MOV_L ( ARG_DEST, ECX )
374 MOV_L ( ARG_MATRIX, ESI )
375 MOV_L ( ARG_SOURCE, EAX )
376 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
377 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
378 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
379 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
383 MOV_L ( REGOFF(V4F_START, ECX), EDX )
385 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
386 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
387 MOV_L ( REGOFF(V4F_START, EAX), EAX )
390 JZ ( LLBL( G3TP2NRR_3 ) )
392 MOVD ( REGIND(ECX), MM0 ) /* | m00 */
393 PUNPCKLDQ ( REGOFF(20, ECX), MM0 ) /* m11 | m00 */
395 MOVQ ( REGOFF(48, ECX), MM2 ) /* m31 | m30 */
400 MOVQ ( REGIND(EAX), MM4 ) /* x1 | x0 */
401 ADD_L ( EDI, EAX ) /* next vertex */
403 PFMUL ( MM0, MM4 ) /* x1*m11 | x0*m00 */
404 PFADD ( MM2, MM4 ) /* m31 | x0*m00+m30 */
406 MOVQ ( MM4, REGIND(EDX) ) /* write r1, r0 */
407 ADD_L ( CONST(16), EDX ) /* next r */
409 DEC_L ( ESI ) /* decrement vertex counter */
410 JNZ ( LLBL( G3TP2NRR_2 ) ) /* cnt > 0 ? -> process next vertex */
423 GLOBL GLNAME( _mesa_3dnow_transform_points2_identity )
424 GLNAME( _mesa_3dnow_transform_points2_identity ):
428 MOV_L ( ARG_DEST, ECX )
429 MOV_L ( ARG_MATRIX, ESI )
430 MOV_L ( ARG_SOURCE, EAX )
431 MOV_L ( CONST(2), REGOFF(V4F_SIZE, ECX) )
432 OR_B ( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, ECX) )
433 MOV_L ( REGOFF(V4F_COUNT, EAX), EDX )
434 MOV_L ( EDX, REGOFF(V4F_COUNT, ECX) )
438 MOV_L ( REGOFF(V4F_START, ECX), EDX )
440 MOV_L ( REGOFF(V4F_COUNT, EAX), ESI )
441 MOV_L ( REGOFF(V4F_STRIDE, EAX), EDI )
442 MOV_L ( REGOFF(V4F_START, EAX), EAX )
445 JZ ( LLBL( G3TPIR_3 ) )
450 MOVQ ( REGIND(EAX), MM0 ) /* x1 | x0 */
451 ADD_L ( EDI, EAX ) /* next vertex */
453 MOVQ ( MM0, REGIND(EDX) ) /* r1 | r0 */
454 ADD_L ( CONST(16), EDX ) /* next r */
456 DEC_L ( ESI ) /* decrement vertex counter */
457 JNZ ( LLBL( G3TPIR_3 ) ) /* cnt > 0 ? -> process next vertex */