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.
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 #define MATH_ASM_PTR_SIZE 4
36 #include "math/m_vector_asm.h"
37 #include "xform_args.h"
41 #define S(i) REGOFF(i * 4, ESI)
42 #define D(i) REGOFF(i * 4, EDI)
43 #define M(i) REGOFF(i * 4, EDX)
47 GLOBL GLNAME(_mesa_sse_transform_points2_general)
48 HIDDEN (_mesa_sse_transform_points2_general)
49 GLNAME( _mesa_sse_transform_points2_general ):
51 #define FRAME_OFFSET 8
55 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
56 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
58 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
59 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
62 JZ( LLBL(K_GTP2GR_finish) ) /* count was zero; go to finish */
64 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
65 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
67 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
68 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
70 SHL_L( CONST(4), ECX ) /* count *= 16 */
71 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
73 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
74 ADD_L( EDI, ECX ) /* count += dest ptr */
77 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
78 MOVAPS( M(4), XMM1 ) /* m7 | m6 | m5 | m4 */
79 MOVAPS( M(12), XMM2 ) /* m15 | m14 | m13 | m12 */
83 MOVSS( S(0), XMM3 ) /* ox */
84 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox | ox */
85 MULPS( XMM0, XMM3 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
86 MOVSS( S(1), XMM4 ) /* oy */
87 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy | oy */
88 MULPS( XMM1, XMM4 ) /* oy*m7 | oy*m6 | oy*m5 | oy*m4 */
95 ADD_L ( CONST(16), EDI )
98 JNE ( LLBL(K_GTP2GR_top) )
100 LLBL(K_GTP2GR_finish):
108 GLOBL GLNAME(_mesa_sse_transform_points2_identity)
109 HIDDEN(_mesa_sse_transform_points2_identity)
110 GLNAME( _mesa_sse_transform_points2_identity ):
112 #define FRAME_OFFSET 8
116 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
117 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
119 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
122 JZ( LLBL(K_GTP2IR_finish) ) /* count was zero; go to finish */
124 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
125 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
127 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
128 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
130 SHL_L( CONST(4), ECX ) /* count *= 16 */
131 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
133 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
134 ADD_L( EDI, ECX ) /* count += dest ptr */
137 JE( LLBL(K_GTP2IR_finish) )
148 ADD_L ( CONST(16), EDI )
151 JNE ( LLBL(K_GTP2IR_top) )
153 LLBL(K_GTP2IR_finish):
161 GLOBL GLNAME(_mesa_sse_transform_points2_3d_no_rot)
162 HIDDEN(_mesa_sse_transform_points2_3d_no_rot)
163 GLNAME(_mesa_sse_transform_points2_3d_no_rot):
165 #define FRAME_OFFSET 8
169 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
170 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
172 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
173 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
176 JZ( LLBL(K_GTP23DNRR_finish) ) /* count was zero; go to finish */
178 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
179 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
181 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
182 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
184 SHL_L( CONST(4), ECX ) /* count *= 16 */
185 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
187 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
188 ADD_L( EDI, ECX ) /* count += dest ptr */
190 XORPS( XMM0, XMM0 ) /* clean the working register */
193 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
194 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
195 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
196 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
197 MOVSS ( M(14), XMM3 ) /* - | - | - | m14 */
200 LLBL(K_GTP23DNRR_top):
201 MOVLPS ( S(0), XMM0 ) /* - | - | oy | ox */
202 MULPS ( XMM1, XMM0 ) /* - | - | oy*m5 | ox*m0 */
203 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
204 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
206 MOVSS ( XMM3, D(2) ) /* -> D(2) */
208 LLBL(K_GTP23DNRR_skip):
209 ADD_L ( CONST(16), EDI )
212 JNE ( LLBL(K_GTP23DNRR_top) )
214 LLBL(K_GTP23DNRR_finish):
222 GLOBL GLNAME(_mesa_sse_transform_points2_perspective)
223 HIDDEN(_mesa_sse_transform_points2_perspective)
224 GLNAME(_mesa_sse_transform_points2_perspective):
226 #define FRAME_OFFSET 8
230 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
231 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
233 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
234 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
237 JZ( LLBL(K_GTP23PR_finish) ) /* count was zero; go to finish */
239 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
240 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
242 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
243 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
245 SHL_L( CONST(4), ECX ) /* count *= 16 */
246 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
248 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
249 ADD_L( EDI, ECX ) /* count += dest ptr */
252 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
253 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
254 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
255 MOVSS ( M(14), XMM3 ) /* m14 */
256 XORPS ( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
260 MOVLPS( S(0), XMM4 ) /* oy | ox */
261 MULPS( XMM1, XMM4 ) /* oy*m5 | ox*m0 */
262 MOVLPS( XMM4, D(0) ) /* ->D(1) | ->D(0) */
263 MOVSS( XMM3, D(2) ) /* ->D(2) */
264 MOVSS( XMM0, D(3) ) /* ->D(3) */
266 LLBL(K_GTP23PR_skip):
267 ADD_L( CONST(16), EDI )
270 JNE( LLBL(K_GTP23PR_top) )
272 LLBL(K_GTP23PR_finish):
281 GLOBL GLNAME(_mesa_sse_transform_points2_2d)
282 HIDDEN(_mesa_sse_transform_points2_2d)
283 GLNAME(_mesa_sse_transform_points2_2d):
285 #define FRAME_OFFSET 8
289 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
290 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
292 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
293 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
296 JZ( LLBL(K_GTP23P2DR_finish) ) /* count was zero; go to finish */
298 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
299 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
301 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
302 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
304 SHL_L( CONST(4), ECX ) /* count *= 16 */
305 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
307 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
308 ADD_L( EDI, ECX ) /* count += dest ptr */
311 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
312 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
313 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
316 LLBL(K_GTP23P2DR_top):
317 MOVSS( S(0), XMM3 ) /* ox */
318 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
319 MULPS( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
321 MOVSS( S(1), XMM4 ) /* oy */
322 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
323 MULPS( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
327 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
329 LLBL(K_GTP23P2DR_skip):
330 ADD_L ( CONST(16), EDI )
333 JNE ( LLBL(K_GTP23P2DR_top) )
335 LLBL(K_GTP23P2DR_finish):
344 GLOBL GLNAME(_mesa_sse_transform_points2_2d_no_rot)
345 HIDDEN(_mesa_sse_transform_points2_2d_no_rot)
346 GLNAME(_mesa_sse_transform_points2_2d_no_rot):
348 #define FRAME_OFFSET 8
352 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
353 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
355 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
356 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
359 JZ( LLBL(K_GTP23P2DNRR_finish) ) /* count was zero; go to finish */
361 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
362 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
364 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
365 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
367 SHL_L( CONST(4), ECX ) /* count *= 16 */
368 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
370 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
371 ADD_L( EDI, ECX ) /* count += dest ptr */
374 MOVSS ( M(0), XMM1 ) /* m0 */
375 MOVSS ( M(5), XMM2 ) /* m5 */
376 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
377 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
380 LLBL(K_GTP23P2DNRR_top):
381 MOVLPS( S(0), XMM0 ) /* oy | ox */
382 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
383 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
384 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
386 LLBL(K_GTP23P2DNRR_skip):
387 ADD_L( CONST(16), EDI )
390 JNE( LLBL(K_GTP23P2DNRR_top) )
392 LLBL(K_GTP23P2DNRR_finish):
401 GLOBL GLNAME(_mesa_sse_transform_points2_3d)
402 HIDDEN(_mesa_sse_transform_points2_3d)
403 GLNAME(_mesa_sse_transform_points2_3d):
405 #define FRAME_OFFSET 8
409 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
410 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
412 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
413 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
416 JZ( LLBL(K_GTP23P3DR_finish) ) /* count was zero; go to finish */
418 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
419 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
421 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
422 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
424 SHL_L( CONST(4), ECX ) /* count *= 16 */
425 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
427 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
428 ADD_L( EDI, ECX ) /* count += dest ptr */
431 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
432 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
433 MOVAPS( M(12), XMM2 ) /* m14 | m13 | m12 */
436 LLBL(K_GTP23P3DR_top):
437 MOVSS( S(0), XMM3 ) /* ox */
438 SHUFPS( CONST(0x0), XMM3, XMM3 ) /* ox | ox | ox */
439 MULPS( XMM0, XMM3 ) /* ox*m2 | ox*m1 | ox*m0 */
441 MOVSS( S(1), XMM4 ) /* oy */
442 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* oy | oy | oy */
443 MULPS( XMM1, XMM4 ) /* oy*m6 | oy*m5 | oy*m4 */
448 MOVLPS( XMM3, D(0) ) /* ->D(1) | ->D(0) */
449 UNPCKHPS( XMM3, XMM3 )
450 MOVSS( XMM3, D(2) ) /* ->D(2) */
452 LLBL(K_GTP23P3DR_skip):
453 ADD_L( CONST(16), EDI )
456 JNE( LLBL(K_GTP23P3DR_top) )
458 LLBL(K_GTP23P3DR_finish):
465 #if defined (__ELF__) && defined (__linux__)
466 .section .note.GNU-stack,"",%progbits