1 /* $Id: sse_xform1.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_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 GLNAME( _mesa_sse_transform_points1_identity ):
105 #define FRAME_OFFSET 8
109 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
110 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
112 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
115 JZ( LLBL(K_GTP1IR_finish) ) /* count was zero; go to finish */
117 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
118 OR_L( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
120 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
121 MOV_L( CONST(1), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
123 SHL_L( CONST(4), ECX ) /* count *= 16 */
124 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
126 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
127 ADD_L( EDI, ECX ) /* count += dest ptr */
130 JE( LLBL(K_GTP1IR_finish) )
139 ADD_L ( CONST(16), EDI )
142 JNE ( LLBL(K_GTP1IR_top) )
144 LLBL(K_GTP1IR_finish):
153 GLOBL GLNAME(_mesa_sse_transform_points1_3d_no_rot)
154 GLNAME(_mesa_sse_transform_points1_3d_no_rot):
156 #define FRAME_OFFSET 8
160 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
161 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
163 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
164 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
167 JZ( LLBL(K_GTP13DNRR_finish) ) /* count was zero; go to finish */
169 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
170 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
172 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
173 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
175 SHL_L( CONST(4), ECX ) /* count *= 16 */
176 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
178 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
179 ADD_L( EDI, ECX ) /* count += dest ptr */
183 MOVSS( M(0), XMM0 ) /* m0 */
184 MOVSS( M(12), XMM1 ) /* m12 */
185 MOVSS( M(13), XMM2 ) /* m13 */
186 MOVSS( M(14), XMM3 ) /* m14 */
189 LLBL(K_GTP13DNRR_top):
190 MOVSS( S(0), XMM4 ) /* ox */
191 MULSS( XMM0, XMM4 ) /* ox*m0 */
192 ADDSS( XMM1, XMM4 ) /* ox*m0+m12 */
198 LLBL(K_GTP13DNRR_skip):
199 ADD_L ( CONST(16), EDI )
202 JNE ( LLBL(K_GTP13DNRR_top) )
204 LLBL(K_GTP13DNRR_finish):
213 GLOBL GLNAME(_mesa_sse_transform_points1_perspective)
214 GLNAME(_mesa_sse_transform_points1_perspective):
216 #define FRAME_OFFSET 8
220 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
221 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
223 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
224 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
227 JZ( LLBL(K_GTP13PR_finish) ) /* count was zero; go to finish */
229 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
230 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
232 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
233 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
235 SHL_L( CONST(4), ECX ) /* count *= 16 */
236 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
238 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
239 ADD_L( EDI, ECX ) /* count += dest ptr */
243 XORPS( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
244 MOVSS( M(0), XMM1 ) /* m0 */
245 MOVSS( M(14), XMM2 ) /* m14 */
249 MOVSS( S(0), XMM3 ) /* ox */
250 MULSS( XMM1, XMM3 ) /* ox*m0 */
251 MOVSS( XMM3, D(0) ) /* ox*m0->D(0) */
252 MOVSS( XMM2, D(2) ) /* m14->D(2) */
257 LLBL(K_GTP13PR_skip):
258 ADD_L( CONST(16), EDI )
261 JNE( LLBL(K_GTP13PR_top) )
263 LLBL(K_GTP13PR_finish):
271 GLOBL GLNAME(_mesa_sse_transform_points1_2d)
272 GLNAME(_mesa_sse_transform_points1_2d):
274 #define FRAME_OFFSET 8
278 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
279 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
281 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
282 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
285 JZ( LLBL(K_GTP13P2DR_finish) ) /* count was zero; go to finish */
287 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
288 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
290 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
291 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
293 SHL_L( CONST(4), ECX ) /* count *= 16 */
294 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
296 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
297 ADD_L( EDI, ECX ) /* count += dest ptr */
300 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
301 MOVLPS( M(12), XMM1 ) /* m13 | m12 */
304 LLBL(K_GTP13P2DR_top):
305 MOVSS( S(0), XMM2 ) /* ox */
306 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
307 MULPS( XMM0, XMM2 ) /* - | - | ox*m1 | ox*m0 */
308 ADDPS( XMM1, XMM2 ) /* - | - | ox*m1+m13 | ox*m0+m12 */
311 LLBL(K_GTP13P2DR_skip):
312 ADD_L ( CONST(16), EDI )
315 JNE ( LLBL(K_GTP13P2DR_top) )
317 LLBL(K_GTP13P2DR_finish):
325 GLOBL GLNAME(_mesa_sse_transform_points1_2d_no_rot)
326 GLNAME(_mesa_sse_transform_points1_2d_no_rot):
328 #define FRAME_OFFSET 8
332 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
333 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
335 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
336 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
339 JZ( LLBL(K_GTP13P2DNRR_finish) ) /* count was zero; go to finish */
341 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
342 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
344 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
345 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
347 SHL_L( CONST(4), ECX ) /* count *= 16 */
348 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
350 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
351 ADD_L( EDI, ECX ) /* count += dest ptr */
354 MOVSS( M(0), XMM0 ) /* m0 */
355 MOVSS( M(12), XMM1 ) /* m12 */
356 MOVSS( M(13), XMM2 ) /* m13 */
359 LLBL(K_GTP13P2DNRR_top):
360 MOVSS( S(0), XMM3 ) /* ox */
361 MULSS( XMM0, XMM3 ) /* ox*m0 */
362 ADDSS( XMM1, XMM3 ) /* ox*m0+m12 */
366 LLBL(K_GTP13P2DNRR_skip):
367 ADD_L( CONST(16), EDI )
370 JNE( LLBL(K_GTP13P2DNRR_top) )
372 LLBL(K_GTP13P2DNRR_finish):
381 GLOBL GLNAME(_mesa_sse_transform_points1_3d)
382 GLNAME(_mesa_sse_transform_points1_3d):
384 #define FRAME_OFFSET 8
388 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
389 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
391 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
392 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
395 JZ( LLBL(K_GTP13P3DR_finish) ) /* count was zero; go to finish */
397 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
398 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
400 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
401 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
403 SHL_L( CONST(4), ECX ) /* count *= 16 */
404 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
406 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
407 ADD_L( EDI, ECX ) /* count += dest ptr */
411 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
412 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
415 LLBL(K_GTP13P3DR_top):
416 MOVSS( S(0), XMM2 ) /* ox */
417 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
418 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
419 ADDPS( XMM1, XMM2 ) /* +m15 | +m14 | +m13 | +m12 */
420 MOVLPS( XMM2, D(0) ) /* - | - | ->D(1)| ->D(0)*/
421 UNPCKHPS( XMM2, XMM2 ) /* ox*m3+m15 | ox*m3+m15 | ox*m2+m14 | ox*m2+m14 */
424 LLBL(K_GTP13P3DR_skip):
425 ADD_L( CONST(16), EDI )
428 JNE( LLBL(K_GTP13P3DR_top) )
430 LLBL(K_GTP13P3DR_finish):