1 /* $Id: sse_xform3.S,v 1.2 2004/04/26 10:10:25 alanh 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 * - 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 !
36 #include "xform_args.h"
40 #define S(i) REGOFF(i * 4, ESI)
41 #define D(i) REGOFF(i * 4, EDI)
42 #define M(i) REGOFF(i * 4, EDX)
46 GLOBL GLNAME(_mesa_sse_transform_points3_general)
47 GLNAME( _mesa_sse_transform_points3_general ):
49 #define FRAME_OFFSET 8
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 */
56 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
57 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
59 CMP_L ( CONST(0), ECX ) /* count == 0 ? */
60 JE ( LLBL(K_GTPGR_finish) ) /* yes -> nothing to do. */
62 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
63 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
65 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
66 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
68 SHL_L( CONST(4), ECX ) /* count *= 16 */
69 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
71 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
72 ADD_L( EDI, ECX ) /* count += dest ptr */
76 MOVAPS ( REGOFF(0, EDX), XMM0 ) /* m0 | m1 | m2 | m3 */
77 MOVAPS ( REGOFF(16, EDX), XMM1 ) /* m4 | m5 | m6 | m7 */
78 MOVAPS ( REGOFF(32, EDX), XMM2 ) /* m8 | m9 | m10 | m11 */
79 MOVAPS ( REGOFF(48, EDX), XMM3 ) /* m12 | m13 | m14 | m15 */
84 MOVSS ( REGOFF(0, ESI), XMM4 ) /* | | | ox */
85 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox | ox */
86 MOVSS ( REGOFF(4, ESI), XMM5 ) /* | | | oy */
87 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy | oy */
88 MOVSS ( REGOFF(8, ESI), XMM6 ) /* | | | oz */
89 SHUFPS ( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz | oz */
91 MULPS ( XMM0, XMM4 ) /* m3*ox | m2*ox | m1*ox | m0*ox */
92 MULPS ( XMM1, XMM5 ) /* m7*oy | m6*oy | m5*oy | m4*oy */
93 MULPS ( XMM2, XMM6 ) /* m11*oz | m10*oz | m9*oz | m8*oz */
99 MOVAPS ( XMM4, REGOFF(0, EDI) )
102 ADD_L ( CONST(16), EDI )
105 JNE ( LLBL(K_GTPGR_top) )
107 LLBL(K_GTPGR_finish):
115 GLOBL GLNAME(_mesa_sse_transform_points3_identity)
116 GLNAME( _mesa_sse_transform_points3_identity ):
118 #define FRAME_OFFSET 8
122 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
123 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
125 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
128 JZ( LLBL(K_GTPIR_finish) ) /* count was zero; go to finish */
130 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
131 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
133 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
134 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
136 SHL_L( CONST(4), ECX ) /* count *= 16 */
137 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
139 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
140 ADD_L( EDI, ECX ) /* count += dest ptr */
143 JE( LLBL(K_GTPIR_finish) )
148 MOVLPS ( S(0), XMM0 )
149 MOVLPS ( XMM0, D(0) )
154 ADD_L ( CONST(16), EDI )
157 JNE ( LLBL(K_GTPIR_top) )
159 LLBL(K_GTPIR_finish):
169 GLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot)
170 GLNAME(_mesa_sse_transform_points3_3d_no_rot):
172 #define FRAME_OFFSET 8
176 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
177 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
180 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
181 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
184 JZ( LLBL(K_GTP3DNRR_finish) ) /* count was zero; go to finish */
186 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
187 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
189 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
190 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
192 SHL_L( CONST(4), ECX ) /* count *= 16 */
193 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
195 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
196 ADD_L( EDI, ECX ) /* count += dest ptr */
200 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
201 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
202 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
203 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
204 MOVSS ( M(10), XMM3 ) /* - | - | - | m10 */
205 MOVSS ( M(14), XMM4 ) /* - | - | - | m14 */
208 LLBL(K_GTP3DNRR_top):
210 MOVLPS ( S(0), XMM0 ) /* - | - | s1 | s0 */
211 MULPS ( XMM1, XMM0 ) /* - | - | s1*m5 | s0*m0 */
212 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
213 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
215 MOVSS ( S(2), XMM0 ) /* sz */
216 MULSS ( XMM3, XMM0 ) /* sz*m10 */
217 ADDSS ( XMM4, XMM0 ) /* +m14 */
218 MOVSS ( XMM0, D(2) ) /* -> D(2) */
220 LLBL(K_GTP3DNRR_skip):
221 ADD_L ( CONST(16), EDI )
224 JNE ( LLBL(K_GTP3DNRR_top) )
226 LLBL(K_GTP3DNRR_finish):
235 GLOBL GLNAME(_mesa_sse_transform_points3_perspective)
236 GLNAME(_mesa_sse_transform_points3_perspective):
238 #define FRAME_OFFSET 8
242 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
243 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
245 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
246 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
249 JZ( LLBL(K_GTP3PR_finish) ) /* count was zero; go to finish */
251 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
252 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
254 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
255 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
257 SHL_L( CONST(4), ECX ) /* count *= 16 */
258 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
260 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
261 ADD_L( EDI, ECX ) /* count += dest ptr */
264 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
265 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
266 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
267 MOVLPS ( M(8), XMM2 ) /* - | - | m9 | m8 */
268 MOVSS ( M(10), XMM3 ) /* m10 */
269 MOVSS ( M(14), XMM4 ) /* m14 */
270 XORPS ( XMM6, XMM6 ) /* 0 */
274 MOVLPS ( S(0), XMM0 ) /* oy | ox */
275 MULPS ( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
276 MOVSS ( S(2), XMM5 ) /* oz */
277 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oz | oz */
278 MULPS ( XMM2, XMM5 ) /* oz*m9 | oz*m8 */
279 ADDPS ( XMM5, XMM0 ) /* +oy*m5 | +ox*m0 */
280 MOVLPS ( XMM0, D(0) ) /* ->D(1) | ->D(0) */
282 MOVSS ( S(2), XMM0 ) /* oz */
283 MULSS ( XMM3, XMM0 ) /* oz*m10 */
284 ADDSS ( XMM4, XMM0 ) /* +m14 */
285 MOVSS ( XMM0, D(2) ) /* ->D(2) */
287 MOVSS ( S(2), XMM0 ) /* oz */
288 MOVSS ( XMM6, XMM5 ) /* 0 */
289 SUBPS ( XMM0, XMM5 ) /* -oz */
290 MOVSS ( XMM5, D(3) ) /* ->D(3) */
293 ADD_L( CONST(16), EDI )
296 JNE( LLBL(K_GTP3PR_top) )
298 LLBL(K_GTP3PR_finish):
307 GLOBL GLNAME(_mesa_sse_transform_points3_2d)
308 GLNAME(_mesa_sse_transform_points3_2d):
310 #define FRAME_OFFSET 8
314 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
315 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
317 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
318 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
321 JZ( LLBL(K_GTP3P2DR_finish) ) /* count was zero; go to finish */
323 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
324 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
326 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
327 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
329 SHL_L( CONST(4), ECX ) /* count *= 16 */
330 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
332 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
333 ADD_L( EDI, ECX ) /* count += dest ptr */
336 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
337 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
338 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
341 LLBL(K_GTP3P2DR_top):
342 MOVSS ( S(0), XMM3 ) /* ox */
343 SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
344 MULPS ( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
345 MOVSS ( S(1), XMM4 ) /* oy */
346 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
347 MULPS ( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
351 MOVLPS ( XMM3, D(0) )
356 LLBL(K_GTP3P2DR_skip):
357 ADD_L ( CONST(16), EDI )
360 JNE ( LLBL(K_GTP3P2DR_top) )
362 LLBL(K_GTP3P2DR_finish):
371 GLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot)
372 GLNAME(_mesa_sse_transform_points3_2d_no_rot):
374 #define FRAME_OFFSET 8
378 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
379 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
381 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
382 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
385 JZ( LLBL(K_GTP3P2DNRR_finish) ) /* count was zero; go to finish */
387 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
388 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
390 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
391 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
393 SHL_L( CONST(4), ECX ) /* count *= 16 */
394 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
396 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
397 ADD_L( EDI, ECX ) /* count += dest ptr */
400 MOVSS ( M(0), XMM1 ) /* m0 */
401 MOVSS ( M(5), XMM2 ) /* m5 */
402 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
403 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
406 LLBL(K_GTP3P2DNRR_top):
407 MOVLPS( S(0), XMM0 ) /* oy | ox */
408 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
409 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
410 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
415 LLBL(K_GTP3P2DNRR_skip):
416 ADD_L( CONST(16), EDI )
419 JNE( LLBL(K_GTP3P2DNRR_top) )
421 LLBL(K_GTP3P2DNRR_finish):
431 GLOBL GLNAME(_mesa_sse_transform_points3_3d)
432 GLNAME(_mesa_sse_transform_points3_3d):
434 #define FRAME_OFFSET 8
438 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
439 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
442 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
443 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
446 JZ( LLBL(K_GTP3P3DR_finish) ) /* count was zero; go to finish */
448 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
449 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
451 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
452 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
454 SHL_L( CONST(4), ECX ) /* count *= 16 */
455 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
457 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
458 ADD_L( EDI, ECX ) /* count += dest ptr */
462 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
463 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
464 MOVAPS( M(8), XMM2 ) /* m10 | m9 | m8 */
465 MOVAPS( M(12), XMM3 ) /* m14 | m13 | m12 */
468 LLBL(K_GTP3P3DR_top):
470 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox */
471 MULPS( XMM0, XMM4 ) /* ox*m2 | ox*m1 | ox*m0 */
474 SHUFPS( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy */
475 MULPS( XMM1, XMM5 ) /* oy*m6 | oy*m5 | oy*m4 */
478 SHUFPS( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz */
479 MULPS( XMM2, XMM6 ) /* oz*m10 | oz*m9 | oz*m8 */
481 ADDPS( XMM5, XMM4 ) /* + | + | + */
482 ADDPS( XMM6, XMM4 ) /* + | + | + */
483 ADDPS( XMM3, XMM4 ) /* + | + | + */
485 MOVLPS( XMM4, D(0) ) /* => D(1) | => D(0) */
486 UNPCKHPS( XMM4, XMM4 )
489 LLBL(K_GTP3P3DR_skip):
490 ADD_L( CONST(16), EDI )
493 JNE( LLBL(K_GTP3P3DR_top) )
495 LLBL(K_GTP3P3DR_finish):