1 /* $Id: sse_xform3.S,v 1.1 2001/03/29 06:46:27 gareth 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 !
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 GLNAME( _mesa_sse_transform_points3_general ):
48 #define FRAME_OFFSET 8
52 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
53 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
55 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
56 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
58 CMP_L ( CONST(0), ECX ) /* count == 0 ? */
59 JE ( LLBL(K_GTPGR_finish) ) /* yes -> nothing to do. */
61 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
62 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
64 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
65 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
67 SHL_L( CONST(4), ECX ) /* count *= 16 */
68 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
70 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
71 ADD_L( EDI, ECX ) /* count += dest ptr */
75 MOVAPS ( REGOFF(0, EDX), XMM0 ) /* m0 | m1 | m2 | m3 */
76 MOVAPS ( REGOFF(16, EDX), XMM1 ) /* m4 | m5 | m6 | m7 */
77 MOVAPS ( REGOFF(32, EDX), XMM2 ) /* m8 | m9 | m10 | m11 */
78 MOVAPS ( REGOFF(48, EDX), XMM3 ) /* m12 | m13 | m14 | m15 */
83 MOVSS ( REGOFF(0, ESI), XMM4 ) /* | | | ox */
84 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox | ox */
85 MOVSS ( REGOFF(4, ESI), XMM5 ) /* | | | oy */
86 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy | oy */
87 MOVSS ( REGOFF(8, ESI), XMM6 ) /* | | | oz */
88 SHUFPS ( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz | oz */
90 MULPS ( XMM0, XMM4 ) /* m3*ox | m2*ox | m1*ox | m0*ox */
91 MULPS ( XMM1, XMM5 ) /* m7*oy | m6*oy | m5*oy | m4*oy */
92 MULPS ( XMM2, XMM6 ) /* m11*oz | m10*oz | m9*oz | m8*oz */
98 MOVAPS ( XMM4, REGOFF(0, EDI) )
101 ADD_L ( CONST(16), EDI )
104 JNE ( LLBL(K_GTPGR_top) )
106 LLBL(K_GTPGR_finish):
114 GLOBL GLNAME(_mesa_sse_transform_points3_identity)
115 GLNAME( _mesa_sse_transform_points3_identity ):
117 #define FRAME_OFFSET 8
121 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
122 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
124 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
127 JZ( LLBL(K_GTPIR_finish) ) /* count was zero; go to finish */
129 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
130 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
132 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
133 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
135 SHL_L( CONST(4), ECX ) /* count *= 16 */
136 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
138 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
139 ADD_L( EDI, ECX ) /* count += dest ptr */
142 JE( LLBL(K_GTPIR_finish) )
147 MOVLPS ( S(0), XMM0 )
148 MOVLPS ( XMM0, D(0) )
153 ADD_L ( CONST(16), EDI )
156 JNE ( LLBL(K_GTPIR_top) )
158 LLBL(K_GTPIR_finish):
168 GLOBL GLNAME(_mesa_sse_transform_points3_3d_no_rot)
169 GLNAME(_mesa_sse_transform_points3_3d_no_rot):
171 #define FRAME_OFFSET 8
175 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
176 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
179 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
180 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
183 JZ( LLBL(K_GTP3DNRR_finish) ) /* count was zero; go to finish */
185 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
186 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
188 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
189 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
191 SHL_L( CONST(4), ECX ) /* count *= 16 */
192 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
194 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
195 ADD_L( EDI, ECX ) /* count += dest ptr */
199 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
200 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
201 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
202 MOVLPS ( M(12), XMM2 ) /* - | - | m13 | m12 */
203 MOVSS ( M(10), XMM3 ) /* - | - | - | m10 */
204 MOVSS ( M(14), XMM4 ) /* - | - | - | m14 */
207 LLBL(K_GTP3DNRR_top):
209 MOVLPS ( S(0), XMM0 ) /* - | - | s1 | s0 */
210 MULPS ( XMM1, XMM0 ) /* - | - | s1*m5 | s0*m0 */
211 ADDPS ( XMM2, XMM0 ) /* - | - | +m13 | +m12 */
212 MOVLPS ( XMM0, D(0) ) /* -> D(1) | -> D(0) */
214 MOVSS ( S(2), XMM0 ) /* sz */
215 MULSS ( XMM3, XMM0 ) /* sz*m10 */
216 ADDSS ( XMM4, XMM0 ) /* +m14 */
217 MOVSS ( XMM0, D(2) ) /* -> D(2) */
219 LLBL(K_GTP3DNRR_skip):
220 ADD_L ( CONST(16), EDI )
223 JNE ( LLBL(K_GTP3DNRR_top) )
225 LLBL(K_GTP3DNRR_finish):
234 GLOBL GLNAME(_mesa_sse_transform_points3_perspective)
235 GLNAME(_mesa_sse_transform_points3_perspective):
237 #define FRAME_OFFSET 8
241 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
242 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
244 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
245 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
248 JZ( LLBL(K_GTP3PR_finish) ) /* count was zero; go to finish */
250 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
251 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
253 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
254 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
256 SHL_L( CONST(4), ECX ) /* count *= 16 */
257 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
259 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
260 ADD_L( EDI, ECX ) /* count += dest ptr */
263 MOVSS ( M(0), XMM1 ) /* - | - | - | m0 */
264 MOVSS ( M(5), XMM2 ) /* - | - | - | m5 */
265 UNPCKLPS ( XMM2, XMM1 ) /* - | - | m5 | m0 */
266 MOVLPS ( M(8), XMM2 ) /* - | - | m9 | m8 */
267 MOVSS ( M(10), XMM3 ) /* m10 */
268 MOVSS ( M(14), XMM4 ) /* m14 */
269 XORPS ( XMM6, XMM6 ) /* 0 */
273 MOVLPS ( S(0), XMM0 ) /* oy | ox */
274 MULPS ( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
275 MOVSS ( S(2), XMM5 ) /* oz */
276 SHUFPS ( CONST(0x0), XMM5, XMM5 ) /* oz | oz */
277 MULPS ( XMM2, XMM5 ) /* oz*m9 | oz*m8 */
278 ADDPS ( XMM5, XMM0 ) /* +oy*m5 | +ox*m0 */
279 MOVLPS ( XMM0, D(0) ) /* ->D(1) | ->D(0) */
281 MOVSS ( S(2), XMM0 ) /* oz */
282 MULSS ( XMM3, XMM0 ) /* oz*m10 */
283 ADDSS ( XMM4, XMM0 ) /* +m14 */
284 MOVSS ( XMM0, D(2) ) /* ->D(2) */
286 MOVSS ( S(2), XMM0 ) /* oz */
287 MOVSS ( XMM6, XMM5 ) /* 0 */
288 SUBPS ( XMM0, XMM5 ) /* -oz */
289 MOVSS ( XMM5, D(3) ) /* ->D(3) */
292 ADD_L( CONST(16), EDI )
295 JNE( LLBL(K_GTP3PR_top) )
297 LLBL(K_GTP3PR_finish):
306 GLOBL GLNAME(_mesa_sse_transform_points3_2d)
307 GLNAME(_mesa_sse_transform_points3_2d):
309 #define FRAME_OFFSET 8
313 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
314 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
316 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
317 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
320 JZ( LLBL(K_GTP3P2DR_finish) ) /* count was zero; go to finish */
322 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
323 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
325 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
326 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
328 SHL_L( CONST(4), ECX ) /* count *= 16 */
329 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
331 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
332 ADD_L( EDI, ECX ) /* count += dest ptr */
335 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
336 MOVLPS( M(4), XMM1 ) /* m5 | m4 */
337 MOVLPS( M(12), XMM2 ) /* m13 | m12 */
340 LLBL(K_GTP3P2DR_top):
341 MOVSS ( S(0), XMM3 ) /* ox */
342 SHUFPS ( CONST(0x0), XMM3, XMM3 ) /* ox | ox */
343 MULPS ( XMM0, XMM3 ) /* ox*m1 | ox*m0 */
344 MOVSS ( S(1), XMM4 ) /* oy */
345 SHUFPS ( CONST(0x0), XMM4, XMM4 ) /* oy | oy */
346 MULPS ( XMM1, XMM4 ) /* oy*m5 | oy*m4 */
350 MOVLPS ( XMM3, D(0) )
355 LLBL(K_GTP3P2DR_skip):
356 ADD_L ( CONST(16), EDI )
359 JNE ( LLBL(K_GTP3P2DR_top) )
361 LLBL(K_GTP3P2DR_finish):
370 GLOBL GLNAME(_mesa_sse_transform_points3_2d_no_rot)
371 GLNAME(_mesa_sse_transform_points3_2d_no_rot):
373 #define FRAME_OFFSET 8
377 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
378 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
380 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
381 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
384 JZ( LLBL(K_GTP3P2DNRR_finish) ) /* count was zero; go to finish */
386 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
387 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
389 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
390 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
392 SHL_L( CONST(4), ECX ) /* count *= 16 */
393 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
395 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
396 ADD_L( EDI, ECX ) /* count += dest ptr */
399 MOVSS ( M(0), XMM1 ) /* m0 */
400 MOVSS ( M(5), XMM2 ) /* m5 */
401 UNPCKLPS ( XMM2, XMM1 ) /* m5 | m0 */
402 MOVLPS ( M(12), XMM2 ) /* m13 | m12 */
405 LLBL(K_GTP3P2DNRR_top):
406 MOVLPS( S(0), XMM0 ) /* oy | ox */
407 MULPS( XMM1, XMM0 ) /* oy*m5 | ox*m0 */
408 ADDPS( XMM2, XMM0 ) /* +m13 | +m12 */
409 MOVLPS( XMM0, D(0) ) /* ->D(1) | ->D(0) */
414 LLBL(K_GTP3P2DNRR_skip):
415 ADD_L( CONST(16), EDI )
418 JNE( LLBL(K_GTP3P2DNRR_top) )
420 LLBL(K_GTP3P2DNRR_finish):
430 GLOBL GLNAME(_mesa_sse_transform_points3_3d)
431 GLNAME(_mesa_sse_transform_points3_3d):
433 #define FRAME_OFFSET 8
437 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
438 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
441 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
442 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
445 JZ( LLBL(K_GTP3P3DR_finish) ) /* count was zero; go to finish */
447 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
448 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
450 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
451 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
453 SHL_L( CONST(4), ECX ) /* count *= 16 */
454 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
456 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
457 ADD_L( EDI, ECX ) /* count += dest ptr */
461 MOVAPS( M(0), XMM0 ) /* m2 | m1 | m0 */
462 MOVAPS( M(4), XMM1 ) /* m6 | m5 | m4 */
463 MOVAPS( M(8), XMM2 ) /* m10 | m9 | m8 */
464 MOVAPS( M(12), XMM3 ) /* m14 | m13 | m12 */
467 LLBL(K_GTP3P3DR_top):
469 SHUFPS( CONST(0x0), XMM4, XMM4 ) /* ox | ox | ox */
470 MULPS( XMM0, XMM4 ) /* ox*m2 | ox*m1 | ox*m0 */
473 SHUFPS( CONST(0x0), XMM5, XMM5 ) /* oy | oy | oy */
474 MULPS( XMM1, XMM5 ) /* oy*m6 | oy*m5 | oy*m4 */
477 SHUFPS( CONST(0x0), XMM6, XMM6 ) /* oz | oz | oz */
478 MULPS( XMM2, XMM6 ) /* oz*m10 | oz*m9 | oz*m8 */
480 ADDPS( XMM5, XMM4 ) /* + | + | + */
481 ADDPS( XMM6, XMM4 ) /* + | + | + */
482 ADDPS( XMM3, XMM4 ) /* + | + | + */
484 MOVLPS( XMM4, D(0) ) /* => D(1) | => D(0) */
485 UNPCKHPS( XMM4, XMM4 )
488 LLBL(K_GTP3P3DR_skip):
489 ADD_L( CONST(16), EDI )
492 JNE( LLBL(K_GTP3P3DR_top) )
494 LLBL(K_GTP3P3DR_finish):