1 /* $Id: sse_xform1.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_points1_general)
46 GLNAME( _mesa_sse_transform_points1_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_GTP1GR_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( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
76 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
80 MOVSS( S(0), XMM2 ) /* ox */
81 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
82 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
83 ADDPS( XMM1, XMM2 ) /* + | + | + | + */
87 ADD_L ( CONST(16), EDI )
90 JNE ( LLBL(K_GTP1GR_top) )
92 LLBL(K_GTP1GR_finish):
101 GLOBL GLNAME(_mesa_sse_transform_points1_identity)
102 GLNAME( _mesa_sse_transform_points1_identity ):
104 #define FRAME_OFFSET 8
108 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
109 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
111 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
114 JZ( LLBL(K_GTP1IR_finish) ) /* count was zero; go to finish */
116 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
117 OR_L( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
119 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
120 MOV_L( CONST(1), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
122 SHL_L( CONST(4), ECX ) /* count *= 16 */
123 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
125 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
126 ADD_L( EDI, ECX ) /* count += dest ptr */
129 JE( LLBL(K_GTP1IR_finish) )
138 ADD_L ( CONST(16), EDI )
141 JNE ( LLBL(K_GTP1IR_top) )
143 LLBL(K_GTP1IR_finish):
152 GLOBL GLNAME(_mesa_sse_transform_points1_3d_no_rot)
153 GLNAME(_mesa_sse_transform_points1_3d_no_rot):
155 #define FRAME_OFFSET 8
159 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
160 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
162 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
163 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
166 JZ( LLBL(K_GTP13DNRR_finish) ) /* count was zero; go to finish */
168 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
169 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
171 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
172 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
174 SHL_L( CONST(4), ECX ) /* count *= 16 */
175 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
177 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
178 ADD_L( EDI, ECX ) /* count += dest ptr */
182 MOVSS( M(0), XMM0 ) /* m0 */
183 MOVSS( M(12), XMM1 ) /* m12 */
184 MOVSS( M(13), XMM2 ) /* m13 */
185 MOVSS( M(14), XMM3 ) /* m14 */
188 LLBL(K_GTP13DNRR_top):
189 MOVSS( S(0), XMM4 ) /* ox */
190 MULSS( XMM0, XMM4 ) /* ox*m0 */
191 ADDSS( XMM1, XMM4 ) /* ox*m0+m12 */
197 LLBL(K_GTP13DNRR_skip):
198 ADD_L ( CONST(16), EDI )
201 JNE ( LLBL(K_GTP13DNRR_top) )
203 LLBL(K_GTP13DNRR_finish):
212 GLOBL GLNAME(_mesa_sse_transform_points1_perspective)
213 GLNAME(_mesa_sse_transform_points1_perspective):
215 #define FRAME_OFFSET 8
219 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
220 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
222 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
223 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
226 JZ( LLBL(K_GTP13PR_finish) ) /* count was zero; go to finish */
228 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
229 OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
231 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
232 MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
234 SHL_L( CONST(4), ECX ) /* count *= 16 */
235 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
237 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
238 ADD_L( EDI, ECX ) /* count += dest ptr */
242 XORPS( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
243 MOVSS( M(0), XMM1 ) /* m0 */
244 MOVSS( M(14), XMM2 ) /* m14 */
248 MOVSS( S(0), XMM3 ) /* ox */
249 MULSS( XMM1, XMM3 ) /* ox*m0 */
250 MOVSS( XMM3, D(0) ) /* ox*m0->D(0) */
251 MOVSS( XMM2, D(2) ) /* m14->D(2) */
256 LLBL(K_GTP13PR_skip):
257 ADD_L( CONST(16), EDI )
260 JNE( LLBL(K_GTP13PR_top) )
262 LLBL(K_GTP13PR_finish):
270 GLOBL GLNAME(_mesa_sse_transform_points1_2d)
271 GLNAME(_mesa_sse_transform_points1_2d):
273 #define FRAME_OFFSET 8
277 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
278 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
280 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
281 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
284 JZ( LLBL(K_GTP13P2DR_finish) ) /* count was zero; go to finish */
286 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
287 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
289 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
290 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
292 SHL_L( CONST(4), ECX ) /* count *= 16 */
293 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
295 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
296 ADD_L( EDI, ECX ) /* count += dest ptr */
299 MOVLPS( M(0), XMM0 ) /* m1 | m0 */
300 MOVLPS( M(12), XMM1 ) /* m13 | m12 */
303 LLBL(K_GTP13P2DR_top):
304 MOVSS( S(0), XMM2 ) /* ox */
305 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
306 MULPS( XMM0, XMM2 ) /* - | - | ox*m1 | ox*m0 */
307 ADDPS( XMM1, XMM2 ) /* - | - | ox*m1+m13 | ox*m0+m12 */
310 LLBL(K_GTP13P2DR_skip):
311 ADD_L ( CONST(16), EDI )
314 JNE ( LLBL(K_GTP13P2DR_top) )
316 LLBL(K_GTP13P2DR_finish):
324 GLOBL GLNAME(_mesa_sse_transform_points1_2d_no_rot)
325 GLNAME(_mesa_sse_transform_points1_2d_no_rot):
327 #define FRAME_OFFSET 8
331 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
332 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
334 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
335 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
338 JZ( LLBL(K_GTP13P2DNRR_finish) ) /* count was zero; go to finish */
340 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
341 OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
343 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
344 MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
346 SHL_L( CONST(4), ECX ) /* count *= 16 */
347 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
349 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
350 ADD_L( EDI, ECX ) /* count += dest ptr */
353 MOVSS( M(0), XMM0 ) /* m0 */
354 MOVSS( M(12), XMM1 ) /* m12 */
355 MOVSS( M(13), XMM2 ) /* m13 */
358 LLBL(K_GTP13P2DNRR_top):
359 MOVSS( S(0), XMM3 ) /* ox */
360 MULSS( XMM0, XMM3 ) /* ox*m0 */
361 ADDSS( XMM1, XMM3 ) /* ox*m0+m12 */
365 LLBL(K_GTP13P2DNRR_skip):
366 ADD_L( CONST(16), EDI )
369 JNE( LLBL(K_GTP13P2DNRR_top) )
371 LLBL(K_GTP13P2DNRR_finish):
380 GLOBL GLNAME(_mesa_sse_transform_points1_3d)
381 GLNAME(_mesa_sse_transform_points1_3d):
383 #define FRAME_OFFSET 8
387 MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
388 MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
390 MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
391 MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
394 JZ( LLBL(K_GTP13P3DR_finish) ) /* count was zero; go to finish */
396 MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
397 OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
399 MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
400 MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
402 SHL_L( CONST(4), ECX ) /* count *= 16 */
403 MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
405 MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
406 ADD_L( EDI, ECX ) /* count += dest ptr */
410 MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
411 MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
414 LLBL(K_GTP13P3DR_top):
415 MOVSS( S(0), XMM2 ) /* ox */
416 SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
417 MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
418 ADDPS( XMM1, XMM2 ) /* +m15 | +m14 | +m13 | +m12 */
419 MOVLPS( XMM2, D(0) ) /* - | - | ->D(1)| ->D(0)*/
420 UNPCKHPS( XMM2, XMM2 ) /* ox*m3+m15 | ox*m3+m15 | ox*m2+m14 | ox*m2+m14 */
423 LLBL(K_GTP13P3DR_skip):
424 ADD_L( CONST(16), EDI )
427 JNE( LLBL(K_GTP13P3DR_top) )
429 LLBL(K_GTP13P3DR_finish):