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_points1_general)
46 HIDDEN( _mesa_sse_transform_points1_general )
47 GLNAME( _mesa_sse_transform_points1_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_GTP1GR_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( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
77 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
81 MOVSS( S(0), XMM2 ) /* ox */
82 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
83 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
84 ADDPS( XMM1, XMM2 ) /* + | + | + | + */
88 ADD_L ( CONST(16), EDI )
91 JNE ( LLBL(K_GTP1GR_top) )
93 LLBL(K_GTP1GR_finish):
102 GLOBL GLNAME(_mesa_sse_transform_points1_identity)
103 HIDDEN(_mesa_sse_transform_points1_identity)
104 GLNAME( _mesa_sse_transform_points1_identity ):
106 #define FRAME_OFFSET 8
110 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
111 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
113 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
116 JZ( LLBL(K_GTP1IR_finish) ) /* count was zero; go to finish */
118 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
119 OR_L( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
121 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
122 MOV_L( CONST(1), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
124 SHL_L( CONST(4), ECX ) /* count *= 16 */
125 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
127 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
128 ADD_L( EDI, ECX ) /* count += dest ptr */
131 JE( LLBL(K_GTP1IR_finish) )
140 ADD_L ( CONST(16), EDI )
143 JNE ( LLBL(K_GTP1IR_top) )
145 LLBL(K_GTP1IR_finish):
154 GLOBL GLNAME(_mesa_sse_transform_points1_3d_no_rot)
155 HIDDEN(_mesa_sse_transform_points1_3d_no_rot)
156 GLNAME(_mesa_sse_transform_points1_3d_no_rot):
158 #define FRAME_OFFSET 8
162 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
163 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
165 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
166 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
169 JZ( LLBL(K_GTP13DNRR_finish) ) /* count was zero; go to finish */
171 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
172 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
174 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
175 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
177 SHL_L( CONST(4), ECX ) /* count *= 16 */
178 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
180 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
181 ADD_L( EDI, ECX ) /* count += dest ptr */
185 MOVSS( M(0), XMM0 ) /* m0 */
186 MOVSS( M(12), XMM1 ) /* m12 */
187 MOVSS( M(13), XMM2 ) /* m13 */
188 MOVSS( M(14), XMM3 ) /* m14 */
191 LLBL(K_GTP13DNRR_top):
192 MOVSS( S(0), XMM4 ) /* ox */
193 MULSS( XMM0, XMM4 ) /* ox*m0 */
194 ADDSS( XMM1, XMM4 ) /* ox*m0+m12 */
200 LLBL(K_GTP13DNRR_skip):
201 ADD_L ( CONST(16), EDI )
204 JNE ( LLBL(K_GTP13DNRR_top) )
206 LLBL(K_GTP13DNRR_finish):
215 GLOBL GLNAME(_mesa_sse_transform_points1_perspective)
216 HIDDEN(_mesa_sse_transform_points1_perspective)
217 GLNAME(_mesa_sse_transform_points1_perspective):
219 #define FRAME_OFFSET 8
223 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
224 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
226 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
227 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
230 JZ( LLBL(K_GTP13PR_finish) ) /* count was zero; go to finish */
232 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
233 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
235 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
236 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
238 SHL_L( CONST(4), ECX ) /* count *= 16 */
239 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
241 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
242 ADD_L( EDI, ECX ) /* count += dest ptr */
246 XORPS( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
247 MOVSS( M(0), XMM1 ) /* m0 */
248 MOVSS( M(14), XMM2 ) /* m14 */
252 MOVSS( S(0), XMM3 ) /* ox */
253 MULSS( XMM1, XMM3 ) /* ox*m0 */
254 MOVSS( XMM3, D(0) ) /* ox*m0->D(0) */
255 MOVSS( XMM2, D(2) ) /* m14->D(2) */
260 LLBL(K_GTP13PR_skip):
261 ADD_L( CONST(16), EDI )
264 JNE( LLBL(K_GTP13PR_top) )
266 LLBL(K_GTP13PR_finish):
274 GLOBL GLNAME(_mesa_sse_transform_points1_2d)
275 HIDDEN(_mesa_sse_transform_points1_2d)
276 GLNAME(_mesa_sse_transform_points1_2d):
278 #define FRAME_OFFSET 8
282 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
283 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
285 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
286 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
289 JZ( LLBL(K_GTP13P2DR_finish) ) /* count was zero; go to finish */
291 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
292 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
294 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
295 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
297 SHL_L( CONST(4), ECX ) /* count *= 16 */
298 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
300 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
301 ADD_L( EDI, ECX ) /* count += dest ptr */
304 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
305 MOVLPS( M(12), XMM1 ) /* m13 | m12 */
308 LLBL(K_GTP13P2DR_top):
309 MOVSS( S(0), XMM2 ) /* ox */
310 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
311 MULPS( XMM0, XMM2 ) /* - | - | ox*m1 | ox*m0 */
312 ADDPS( XMM1, XMM2 ) /* - | - | ox*m1+m13 | ox*m0+m12 */
315 LLBL(K_GTP13P2DR_skip):
316 ADD_L ( CONST(16), EDI )
319 JNE ( LLBL(K_GTP13P2DR_top) )
321 LLBL(K_GTP13P2DR_finish):
329 GLOBL GLNAME(_mesa_sse_transform_points1_2d_no_rot)
330 HIDDEN(_mesa_sse_transform_points1_2d_no_rot)
331 GLNAME(_mesa_sse_transform_points1_2d_no_rot):
333 #define FRAME_OFFSET 8
337 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
338 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
340 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
341 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
344 JZ( LLBL(K_GTP13P2DNRR_finish) ) /* count was zero; go to finish */
346 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
347 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
349 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
350 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
352 SHL_L( CONST(4), ECX ) /* count *= 16 */
353 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
355 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
356 ADD_L( EDI, ECX ) /* count += dest ptr */
359 MOVSS( M(0), XMM0 ) /* m0 */
360 MOVSS( M(12), XMM1 ) /* m12 */
361 MOVSS( M(13), XMM2 ) /* m13 */
364 LLBL(K_GTP13P2DNRR_top):
365 MOVSS( S(0), XMM3 ) /* ox */
366 MULSS( XMM0, XMM3 ) /* ox*m0 */
367 ADDSS( XMM1, XMM3 ) /* ox*m0+m12 */
371 LLBL(K_GTP13P2DNRR_skip):
372 ADD_L( CONST(16), EDI )
375 JNE( LLBL(K_GTP13P2DNRR_top) )
377 LLBL(K_GTP13P2DNRR_finish):
386 GLOBL GLNAME(_mesa_sse_transform_points1_3d)
387 HIDDEN(_mesa_sse_transform_points1_3d)
388 GLNAME(_mesa_sse_transform_points1_3d):
390 #define FRAME_OFFSET 8
394 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
395 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
397 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
398 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
401 JZ( LLBL(K_GTP13P3DR_finish) ) /* count was zero; go to finish */
403 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
404 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
406 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
407 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
409 SHL_L( CONST(4), ECX ) /* count *= 16 */
410 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
412 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
413 ADD_L( EDI, ECX ) /* count += dest ptr */
417 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
418 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
421 LLBL(K_GTP13P3DR_top):
422 MOVSS( S(0), XMM2 ) /* ox */
423 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
424 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
425 ADDPS( XMM1, XMM2 ) /* +m15 | +m14 | +m13 | +m12 */
426 MOVLPS( XMM2, D(0) ) /* - | - | ->D(1)| ->D(0)*/
427 UNPCKHPS( XMM2, XMM2 ) /* ox*m3+m15 | ox*m3+m15 | ox*m2+m14 | ox*m2+m14 */
430 LLBL(K_GTP13P3DR_skip):
431 ADD_L( CONST(16), EDI )
434 JNE( LLBL(K_GTP13P3DR_top) )
436 LLBL(K_GTP13P3DR_finish):
443 #if defined (__ELF__) && defined (__linux__)
444 .section .note.GNU-stack,"",%progbits