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 !
34
#include "xform_args.h"
38
#define S(i) REGOFF(i * 4, ESI)
39
#define D(i) REGOFF(i * 4, EDI)
40
#define M(i) REGOFF(i * 4, EDX)
44
GLOBL GLNAME(_mesa_sse_transform_points1_general)
45
GLNAME( _mesa_sse_transform_points1_general ):
47
#define FRAME_OFFSET 8
51
MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
52
MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
54
MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
55
MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
57
CMP_L( CONST(0), ECX ) /* count == 0 ? */
58
JE( LLBL(K_GTP1GR_finish) ) /* yes -> nothing to do. */
60
MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
61
OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
63
MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
64
MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
66
SHL_L( CONST(4), ECX ) /* count *= 16 */
67
MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
69
MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
70
ADD_L( EDI, ECX ) /* count += dest ptr */
74
MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
75
MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
79
MOVSS( S(0), XMM2 ) /* ox */
80
SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
81
MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
82
ADDPS( XMM1, XMM2 ) /* + | + | + | + */
86
ADD_L ( CONST(16), EDI )
89
JNE ( LLBL(K_GTP1GR_top) )
91
LLBL(K_GTP1GR_finish):
100
GLOBL GLNAME(_mesa_sse_transform_points1_identity)
101
GLNAME( _mesa_sse_transform_points1_identity ):
103
#define FRAME_OFFSET 8
107
MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
108
MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
110
MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
113
JZ( LLBL(K_GTP1IR_finish) ) /* count was zero; go to finish */
115
MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
116
OR_L( CONST(VEC_SIZE_1), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
118
MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
119
MOV_L( CONST(1), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
121
SHL_L( CONST(4), ECX ) /* count *= 16 */
122
MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
124
MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
125
ADD_L( EDI, ECX ) /* count += dest ptr */
128
JE( LLBL(K_GTP1IR_finish) )
137
ADD_L ( CONST(16), EDI )
140
JNE ( LLBL(K_GTP1IR_top) )
142
LLBL(K_GTP1IR_finish):
151
GLOBL GLNAME(_mesa_sse_transform_points1_3d_no_rot)
152
GLNAME(_mesa_sse_transform_points1_3d_no_rot):
154
#define FRAME_OFFSET 8
158
MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
159
MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
161
MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
162
MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
165
JZ( LLBL(K_GTP13DNRR_finish) ) /* count was zero; go to finish */
167
MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
168
OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
170
MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
171
MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
173
SHL_L( CONST(4), ECX ) /* count *= 16 */
174
MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
176
MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
177
ADD_L( EDI, ECX ) /* count += dest ptr */
181
MOVSS( M(0), XMM0 ) /* m0 */
182
MOVSS( M(12), XMM1 ) /* m12 */
183
MOVSS( M(13), XMM2 ) /* m13 */
184
MOVSS( M(14), XMM3 ) /* m14 */
187
LLBL(K_GTP13DNRR_top):
188
MOVSS( S(0), XMM4 ) /* ox */
189
MULSS( XMM0, XMM4 ) /* ox*m0 */
190
ADDSS( XMM1, XMM4 ) /* ox*m0+m12 */
196
LLBL(K_GTP13DNRR_skip):
197
ADD_L ( CONST(16), EDI )
200
JNE ( LLBL(K_GTP13DNRR_top) )
202
LLBL(K_GTP13DNRR_finish):
211
GLOBL GLNAME(_mesa_sse_transform_points1_perspective)
212
GLNAME(_mesa_sse_transform_points1_perspective):
214
#define FRAME_OFFSET 8
218
MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
219
MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
221
MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
222
MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
225
JZ( LLBL(K_GTP13PR_finish) ) /* count was zero; go to finish */
227
MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
228
OR_L( CONST(VEC_SIZE_4), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
230
MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
231
MOV_L( CONST(4), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
233
SHL_L( CONST(4), ECX ) /* count *= 16 */
234
MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
236
MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
237
ADD_L( EDI, ECX ) /* count += dest ptr */
241
XORPS( XMM0, XMM0 ) /* 0 | 0 | 0 | 0 */
242
MOVSS( M(0), XMM1 ) /* m0 */
243
MOVSS( M(14), XMM2 ) /* m14 */
247
MOVSS( S(0), XMM3 ) /* ox */
248
MULSS( XMM1, XMM3 ) /* ox*m0 */
249
MOVSS( XMM3, D(0) ) /* ox*m0->D(0) */
250
MOVSS( XMM2, D(2) ) /* m14->D(2) */
255
LLBL(K_GTP13PR_skip):
256
ADD_L( CONST(16), EDI )
259
JNE( LLBL(K_GTP13PR_top) )
261
LLBL(K_GTP13PR_finish):
269
GLOBL GLNAME(_mesa_sse_transform_points1_2d)
270
GLNAME(_mesa_sse_transform_points1_2d):
272
#define FRAME_OFFSET 8
276
MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
277
MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
279
MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
280
MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
283
JZ( LLBL(K_GTP13P2DR_finish) ) /* count was zero; go to finish */
285
MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
286
OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
288
MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
289
MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
291
SHL_L( CONST(4), ECX ) /* count *= 16 */
292
MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
294
MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
295
ADD_L( EDI, ECX ) /* count += dest ptr */
298
MOVLPS( M(0), XMM0 ) /* m1 | m0 */
299
MOVLPS( M(12), XMM1 ) /* m13 | m12 */
302
LLBL(K_GTP13P2DR_top):
303
MOVSS( S(0), XMM2 ) /* ox */
304
SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
305
MULPS( XMM0, XMM2 ) /* - | - | ox*m1 | ox*m0 */
306
ADDPS( XMM1, XMM2 ) /* - | - | ox*m1+m13 | ox*m0+m12 */
309
LLBL(K_GTP13P2DR_skip):
310
ADD_L ( CONST(16), EDI )
313
JNE ( LLBL(K_GTP13P2DR_top) )
315
LLBL(K_GTP13P2DR_finish):
323
GLOBL GLNAME(_mesa_sse_transform_points1_2d_no_rot)
324
GLNAME(_mesa_sse_transform_points1_2d_no_rot):
326
#define FRAME_OFFSET 8
330
MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
331
MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
333
MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
334
MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
337
JZ( LLBL(K_GTP13P2DNRR_finish) ) /* count was zero; go to finish */
339
MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
340
OR_L( CONST(VEC_SIZE_2), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
342
MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
343
MOV_L( CONST(2), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
345
SHL_L( CONST(4), ECX ) /* count *= 16 */
346
MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
348
MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
349
ADD_L( EDI, ECX ) /* count += dest ptr */
352
MOVSS( M(0), XMM0 ) /* m0 */
353
MOVSS( M(12), XMM1 ) /* m12 */
354
MOVSS( M(13), XMM2 ) /* m13 */
357
LLBL(K_GTP13P2DNRR_top):
358
MOVSS( S(0), XMM3 ) /* ox */
359
MULSS( XMM0, XMM3 ) /* ox*m0 */
360
ADDSS( XMM1, XMM3 ) /* ox*m0+m12 */
364
LLBL(K_GTP13P2DNRR_skip):
365
ADD_L( CONST(16), EDI )
368
JNE( LLBL(K_GTP13P2DNRR_top) )
370
LLBL(K_GTP13P2DNRR_finish):
379
GLOBL GLNAME(_mesa_sse_transform_points1_3d)
380
GLNAME(_mesa_sse_transform_points1_3d):
382
#define FRAME_OFFSET 8
386
MOV_L( REGOFF(OFFSET_SOURCE+8, ESP), ESI ) /* ptr to source GLvector4f */
387
MOV_L( REGOFF(OFFSET_DEST+8, ESP), EDI ) /* ptr to dest GLvector4f */
389
MOV_L( ARG_MATRIX, EDX ) /* ptr to matrix */
390
MOV_L( REGOFF(V4F_COUNT, ESI), ECX ) /* source count */
393
JZ( LLBL(K_GTP13P3DR_finish) ) /* count was zero; go to finish */
395
MOV_L( REGOFF(V4F_STRIDE, ESI), EAX ) /* stride */
396
OR_L( CONST(VEC_SIZE_3), REGOFF(V4F_FLAGS, EDI) ) /* set dest flags */
398
MOV_L( ECX, REGOFF(V4F_COUNT, EDI) ) /* set dest count */
399
MOV_L( CONST(3), REGOFF(V4F_SIZE, EDI) ) /* set dest size */
401
SHL_L( CONST(4), ECX ) /* count *= 16 */
402
MOV_L( REGOFF(V4F_START, ESI), ESI ) /* ptr to first source vertex */
404
MOV_L( REGOFF(V4F_START, EDI), EDI ) /* ptr to first dest vertex */
405
ADD_L( EDI, ECX ) /* count += dest ptr */
409
MOVAPS( M(0), XMM0 ) /* m3 | m2 | m1 | m0 */
410
MOVAPS( M(12), XMM1 ) /* m15 | m14 | m13 | m12 */
413
LLBL(K_GTP13P3DR_top):
414
MOVSS( S(0), XMM2 ) /* ox */
415
SHUFPS( CONST(0x0), XMM2, XMM2 ) /* ox | ox | ox | ox */
416
MULPS( XMM0, XMM2 ) /* ox*m3 | ox*m2 | ox*m1 | ox*m0 */
417
ADDPS( XMM1, XMM2 ) /* +m15 | +m14 | +m13 | +m12 */
418
MOVLPS( XMM2, D(0) ) /* - | - | ->D(1)| ->D(0)*/
419
UNPCKHPS( XMM2, XMM2 ) /* ox*m3+m15 | ox*m3+m15 | ox*m2+m14 | ox*m2+m14 */
422
LLBL(K_GTP13P3DR_skip):
423
ADD_L( CONST(16), EDI )
426
JNE( LLBL(K_GTP13P3DR_top) )
428
LLBL(K_GTP13P3DR_finish):