3 * Mesa 3-D graphics library
6 * Copyright (C) 1999-2001 Brian Paul All Rights Reserved.
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice shall be included
16 * in all copies or substantial portions of the Software.
18 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
19 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
21 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
22 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
23 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
27 * - insert PREFETCH instructions to avoid cache-misses !
28 * - some more optimizations are possible...
29 * - for 40-50% more performance in the SSE-functions, the
30 * data (trans-matrix, src_vert, dst_vert) needs to be 16byte aligned !
35 #include "xform_args.h"
39 #define S(i) REGOFF(i * 4, ESI)
40 #define D(i) REGOFF(i * 4, EDI)
41 #define M(i) REGOFF(i * 4, EDX)
45 GLOBL GLNAME(_mesa_sse_transform_points3_general)
46 HIDDEN(_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 HIDDEN(_mesa_sse_transform_points3_identity)
117 GLNAME( _mesa_sse_transform_points3_identity ):
119 #define FRAME_OFFSET 8
123 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
124 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
126 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
129 JZ( LLBL(K_GTPIR_finish) ) /* count was zero; go to finish */
131 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
132 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
134 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
135 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
137 SHL_L( CONST(4), ECX ) /* count *= 16 */
138 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
140 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
141 ADD_L( EDI, ECX ) /* count += dest ptr */
144 JE( LLBL(K_GTPIR_finish) )
149 MOVLPS ( S(0), XMM0 )
150 MOVLPS ( XMM0, D(0) )
155 ADD_L ( CONST(16), EDI )
158 JNE ( LLBL(K_GTPIR_top) )
160 LLBL(K_GTPIR_finish):
170 GLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot)
171 HIDDEN(_mesa_sse_transform_points3_3d_no_rot)
172 GLNAME(_mesa_sse_transform_points3_3d_no_rot):
174 #define FRAME_OFFSET 8
178 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
179 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
182 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
183 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
186 JZ( LLBL(K_GTP3DNRR_finish) ) /* count was zero; go to finish */
188 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
189 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
191 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
192 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
194 SHL_L( CONST(4), ECX ) /* count *= 16 */
195 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
197 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
198 ADD_L( EDI, ECX ) /* count += dest ptr */
202 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
203 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
204 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
205 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
206 MOVSS ( M(10), XMM3 ) /* - | - | - | m10 */
207 MOVSS ( M(14), XMM4 ) /* - | - | - | m14 */
210 LLBL(K_GTP3DNRR_top):
212 MOVLPS ( S(0), XMM0 ) /* - | - | s1 | s0 */
213 MULPS ( XMM1, XMM0 ) /* - | - | s1*m5 | s0*m0 */
214 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
215 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
217 MOVSS ( S(2), XMM0 ) /* sz */
218 MULSS ( XMM3, XMM0 ) /* sz*m10 */
219 ADDSS ( XMM4, XMM0 ) /* +m14 */
220 MOVSS ( XMM0, D(2) ) /* -> D(2) */
222 LLBL(K_GTP3DNRR_skip):
223 ADD_L ( CONST(16), EDI )
226 JNE ( LLBL(K_GTP3DNRR_top) )
228 LLBL(K_GTP3DNRR_finish):
237 GLOBL GLNAME(_mesa_sse_transform_points3_perspective)
238 HIDDEN(_mesa_sse_transform_points3_perspective)
239 GLNAME(_mesa_sse_transform_points3_perspective):
241 #define FRAME_OFFSET 8
245 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
246 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
248 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
249 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
252 JZ( LLBL(K_GTP3PR_finish) ) /* count was zero; go to finish */
254 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
255 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
257 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
258 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
260 SHL_L( CONST(4), ECX ) /* count *= 16 */
261 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
263 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
264 ADD_L( EDI, ECX ) /* count += dest ptr */
267 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
268 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
269 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
270 MOVLPS ( M(8), XMM2 ) /* - | - | m9 | m8 */
271 MOVSS ( M(10), XMM3 ) /* m10 */
272 MOVSS ( M(14), XMM4 ) /* m14 */
273 XORPS ( XMM6, XMM6 ) /* 0 */
277 MOVLPS ( S(0), XMM0 ) /* oy | ox */
278 MULPS ( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
279 MOVSS ( S(2), XMM5 ) /* oz */
280 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oz | oz */
281 MULPS ( XMM2, XMM5 ) /* oz*m9 | oz*m8 */
282 ADDPS ( XMM5, XMM0 ) /* +oy*m5 | +ox*m0 */
283 MOVLPS ( XMM0, D(0) ) /* ->D(1) | ->D(0) */
285 MOVSS ( S(2), XMM0 ) /* oz */
286 MULSS ( XMM3, XMM0 ) /* oz*m10 */
287 ADDSS ( XMM4, XMM0 ) /* +m14 */
288 MOVSS ( XMM0, D(2) ) /* ->D(2) */
290 MOVSS ( S(2), XMM0 ) /* oz */
291 MOVSS ( XMM6, XMM5 ) /* 0 */
292 SUBPS ( XMM0, XMM5 ) /* -oz */
293 MOVSS ( XMM5, D(3) ) /* ->D(3) */
296 ADD_L( CONST(16), EDI )
299 JNE( LLBL(K_GTP3PR_top) )
301 LLBL(K_GTP3PR_finish):
310 GLOBL GLNAME(_mesa_sse_transform_points3_2d)
311 HIDDEN(_mesa_sse_transform_points3_2d)
312 GLNAME(_mesa_sse_transform_points3_2d):
314 #define FRAME_OFFSET 8
318 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
319 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
321 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
322 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
325 JZ( LLBL(K_GTP3P2DR_finish) ) /* count was zero; go to finish */
327 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
328 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
330 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
331 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
333 SHL_L( CONST(4), ECX ) /* count *= 16 */
334 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
336 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
337 ADD_L( EDI, ECX ) /* count += dest ptr */
340 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
341 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
342 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
345 LLBL(K_GTP3P2DR_top):
346 MOVSS ( S(0), XMM3 ) /* ox */
347 SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
348 MULPS ( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
349 MOVSS ( S(1), XMM4 ) /* oy */
350 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
351 MULPS ( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
355 MOVLPS ( XMM3, D(0) )
360 LLBL(K_GTP3P2DR_skip):
361 ADD_L ( CONST(16), EDI )
364 JNE ( LLBL(K_GTP3P2DR_top) )
366 LLBL(K_GTP3P2DR_finish):
375 GLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot)
376 HIDDEN(_mesa_sse_transform_points3_2d_no_rot)
377 GLNAME(_mesa_sse_transform_points3_2d_no_rot):
379 #define FRAME_OFFSET 8
383 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
384 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
386 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
387 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
390 JZ( LLBL(K_GTP3P2DNRR_finish) ) /* count was zero; go to finish */
392 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
393 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
395 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
396 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
398 SHL_L( CONST(4), ECX ) /* count *= 16 */
399 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
401 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
402 ADD_L( EDI, ECX ) /* count += dest ptr */
405 MOVSS ( M(0), XMM1 ) /* m0 */
406 MOVSS ( M(5), XMM2 ) /* m5 */
407 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
408 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
411 LLBL(K_GTP3P2DNRR_top):
412 MOVLPS( S(0), XMM0 ) /* oy | ox */
413 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
414 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
415 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
420 LLBL(K_GTP3P2DNRR_skip):
421 ADD_L( CONST(16), EDI )
424 JNE( LLBL(K_GTP3P2DNRR_top) )
426 LLBL(K_GTP3P2DNRR_finish):
436 GLOBL GLNAME(_mesa_sse_transform_points3_3d)
437 HIDDEN(_mesa_sse_transform_points3_3d)
438 GLNAME(_mesa_sse_transform_points3_3d):
440 #define FRAME_OFFSET 8
444 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
445 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
448 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
449 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
452 JZ( LLBL(K_GTP3P3DR_finish) ) /* count was zero; go to finish */
454 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
455 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
457 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
458 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
460 SHL_L( CONST(4), ECX ) /* count *= 16 */
461 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
463 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
464 ADD_L( EDI, ECX ) /* count += dest ptr */
468 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
469 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
470 MOVAPS( M(8), XMM2 ) /* m10 | m9 | m8 */
471 MOVAPS( M(12), XMM3 ) /* m14 | m13 | m12 */
474 LLBL(K_GTP3P3DR_top):
476 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox */
477 MULPS( XMM0, XMM4 ) /* ox*m2 | ox*m1 | ox*m0 */
480 SHUFPS( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy */
481 MULPS( XMM1, XMM5 ) /* oy*m6 | oy*m5 | oy*m4 */
484 SHUFPS( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz */
485 MULPS( XMM2, XMM6 ) /* oz*m10 | oz*m9 | oz*m8 */
487 ADDPS( XMM5, XMM4 ) /* + | + | + */
488 ADDPS( XMM6, XMM4 ) /* + | + | + */
489 ADDPS( XMM3, XMM4 ) /* + | + | + */
491 MOVLPS( XMM4, D(0) ) /* => D(1) | => D(0) */
492 UNPCKHPS( XMM4, XMM4 )
495 LLBL(K_GTP3P3DR_skip):
496 ADD_L( CONST(16), EDI )
499 JNE( LLBL(K_GTP3P3DR_top) )
501 LLBL(K_GTP3P3DR_finish):
508 #if defined (__ELF__) && defined (__linux__)
509 .section .note.GNU-stack,"",%progbits