2 * Mesa 3-D graphics library
5 * Copyright (C) 1999-2004 Brian Paul All Rights Reserved.
7 * Permission is hereby granted, free of charge, to any person obtaining a
8 * copy of this software and associated documentation files (the "Software"),
9 * to deal in the Software without restriction, including without limitation
10 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
11 * and/or sell copies of the Software, and to permit persons to whom the
12 * Software is furnished to do so, subject to the following conditions:
14 * The above copyright notice and this permission notice shall be included
15 * in all copies or substantial portions of the Software.
17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
18 * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
20 * BRIAN PAUL BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN
21 * AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
22 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 * \file t_arb_program.c
27 * Compile vertex programs to an intermediate representation.
28 * Execute vertex programs over a buffer of vertices.
29 * \author Keith Whitwell, Brian Paul
37 #include "arbprogparse.h"
39 #include "math/m_matrix.h"
40 #include "math/m_translate.h"
41 #include "t_context.h"
42 #include "t_pipeline.h"
43 #include "t_vp_build.h"
44 #include "t_vb_arbprogram.h"
48 /*--------------------------------------------------------------------------- */
53 void (*print
)( union instruction
, const struct opcode_info
* );
58 union instruction
*csr
;
62 #define ARB_VP_MACHINE(stage) ((struct arb_vp_machine *)(stage->privatePtr))
67 * Set x to positive or negative infinity.
69 * XXX: FIXME - type punning.
71 #if defined(USE_IEEE) || defined(_WIN32)
72 #define SET_POS_INFINITY(x) ( *((GLuint *) (void *)&x) = 0x7F800000 )
73 #define SET_NEG_INFINITY(x) ( *((GLuint *) (void *)&x) = 0xFF800000 )
75 #define SET_POS_INFINITY(x) x = __MAXFLOAT
76 #define SET_NEG_INFINITY(x) x = -__MAXFLOAT
77 #define IS_INF_OR_NAN(t) ((t) == __MAXFLOAT)
79 #define SET_POS_INFINITY(x) x = (GLfloat) HUGE_VAL
80 #define SET_NEG_INFINITY(x) x = (GLfloat) -HUGE_VAL
83 #define FREXPF(a,b) frexpf(a,b)
85 #define PUFF(x) ((x)[1] = (x)[2] = (x)[3] = (x)[0])
87 /* FIXME: more type punning (despite use of fi_type...)
89 #define SET_FLOAT_BITS(x, bits) ((fi_type *) (void *) &(x))->i = bits
92 static GLfloat
RoughApproxLog2(GLfloat t
)
97 static GLfloat
RoughApproxPow2(GLfloat t
)
100 /* This isn't nearly accurate enough - it discards all of t's
105 fi
.i
= (fi
.i
<< 23) + 0x3f800000;
108 return (GLfloat
) _mesa_pow(2.0, t
);
112 static GLfloat
RoughApproxPower(GLfloat x
, GLfloat y
)
115 return RoughApproxPow2(y
* RoughApproxLog2(x
));
117 return (GLfloat
) _mesa_pow(x
, y
);
126 * Perform a reduced swizzle:
128 static void do_RSW( struct arb_vp_machine
*m
, union instruction op
)
130 GLfloat
*result
= m
->File
[0][op
.rsw
.dst
];
131 const GLfloat
*arg0
= m
->File
[op
.rsw
.file0
][op
.rsw
.idx0
];
132 GLuint swz
= op
.rsw
.swz
;
133 GLuint neg
= op
.rsw
.neg
;
135 result
[0] = arg0
[GET_RSW(swz
, 0)];
136 result
[1] = arg0
[GET_RSW(swz
, 1)];
137 result
[2] = arg0
[GET_RSW(swz
, 2)];
138 result
[3] = arg0
[GET_RSW(swz
, 3)];
141 if (neg
& 0x1) result
[0] = -result
[0];
142 if (neg
& 0x2) result
[1] = -result
[1];
143 if (neg
& 0x4) result
[2] = -result
[2];
144 if (neg
& 0x8) result
[3] = -result
[3];
148 /* Used to implement write masking. To make things easier for the sse
149 * generator I've gone back to a 1 argument version of this function
150 * (dst.msk = arg), rather than the semantically cleaner (dst = SEL
153 * That means this is the only instruction which doesn't write a full
154 * 4 dwords out. This would make such a program harder to analyse,
155 * but it looks like analysis is going to take place on a higher level
158 static void do_MSK( struct arb_vp_machine
*m
, union instruction op
)
160 GLfloat
*dst
= m
->File
[0][op
.msk
.dst
];
161 const GLfloat
*arg
= m
->File
[op
.msk
.file
][op
.msk
.idx
];
163 if (op
.msk
.mask
& 0x1) dst
[0] = arg
[0];
164 if (op
.msk
.mask
& 0x2) dst
[1] = arg
[1];
165 if (op
.msk
.mask
& 0x4) dst
[2] = arg
[2];
166 if (op
.msk
.mask
& 0x8) dst
[3] = arg
[3];
170 static void do_PRT( struct arb_vp_machine
*m
, union instruction op
)
172 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
174 _mesa_printf("%d: %f %f %f %f\n", m
->vtx_nr
,
175 arg0
[0], arg0
[1], arg0
[2], arg0
[3]);
180 * The traditional ALU and texturing instructions. All operate on
181 * internal registers and ignore write masks and swizzling issues.
184 static void do_ABS( struct arb_vp_machine
*m
, union instruction op
)
186 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
187 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
189 result
[0] = (arg0
[0] < 0.0) ? -arg0
[0] : arg0
[0];
190 result
[1] = (arg0
[1] < 0.0) ? -arg0
[1] : arg0
[1];
191 result
[2] = (arg0
[2] < 0.0) ? -arg0
[2] : arg0
[2];
192 result
[3] = (arg0
[3] < 0.0) ? -arg0
[3] : arg0
[3];
195 static void do_ADD( struct arb_vp_machine
*m
, union instruction op
)
197 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
198 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
199 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
201 result
[0] = arg0
[0] + arg1
[0];
202 result
[1] = arg0
[1] + arg1
[1];
203 result
[2] = arg0
[2] + arg1
[2];
204 result
[3] = arg0
[3] + arg1
[3];
208 static void do_DP3( struct arb_vp_machine
*m
, union instruction op
)
210 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
211 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
212 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
214 result
[0] = (arg0
[0] * arg1
[0] +
223 static void do_DP4( struct arb_vp_machine
*m
, union instruction op
)
225 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
226 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
227 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
229 result
[0] = (arg0
[0] * arg1
[0] +
237 static void do_DPH( struct arb_vp_machine
*m
, union instruction op
)
239 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
240 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
241 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
243 result
[0] = (arg0
[0] * arg1
[0] +
251 static void do_DST( struct arb_vp_machine
*m
, union instruction op
)
253 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
254 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
255 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
258 result
[1] = arg0
[1] * arg1
[1];
264 static void do_EX2( struct arb_vp_machine
*m
, union instruction op
)
266 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
267 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
269 result
[0] = (GLfloat
)RoughApproxPow2(arg0
[0]);
273 static void do_EXP( struct arb_vp_machine
*m
, union instruction op
)
275 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
276 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
277 GLfloat tmp
= arg0
[0];
278 GLfloat flr_tmp
= FLOORF(tmp
);
280 /* KW: nvvertexec has an optimized version of this which is pretty
281 * hard to understand/validate, but avoids the RoughApproxPow2.
283 result
[0] = (GLfloat
) (1 << (int)flr_tmp
);
284 result
[1] = tmp
- flr_tmp
;
285 result
[2] = RoughApproxPow2(tmp
);
289 static void do_FLR( struct arb_vp_machine
*m
, union instruction op
)
291 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
292 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
294 result
[0] = FLOORF(arg0
[0]);
295 result
[1] = FLOORF(arg0
[1]);
296 result
[2] = FLOORF(arg0
[2]);
297 result
[3] = FLOORF(arg0
[3]);
300 static void do_FRC( struct arb_vp_machine
*m
, union instruction op
)
302 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
303 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
305 result
[0] = arg0
[0] - FLOORF(arg0
[0]);
306 result
[1] = arg0
[1] - FLOORF(arg0
[1]);
307 result
[2] = arg0
[2] - FLOORF(arg0
[2]);
308 result
[3] = arg0
[3] - FLOORF(arg0
[3]);
311 static void do_LG2( struct arb_vp_machine
*m
, union instruction op
)
313 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
314 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
316 result
[0] = RoughApproxLog2(arg0
[0]);
322 static void do_LIT( struct arb_vp_machine
*m
, union instruction op
)
324 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
325 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
327 const GLfloat epsilon
= 1.0F
/ 256.0F
; /* per NV spec */
330 tmp
[0] = MAX2(arg0
[0], 0.0F
);
331 tmp
[1] = MAX2(arg0
[1], 0.0F
);
332 tmp
[3] = CLAMP(arg0
[3], -(128.0F
- epsilon
), (128.0F
- epsilon
));
336 result
[2] = (tmp
[0] > 0.0) ? RoughApproxPower(tmp
[1], tmp
[3]) : 0.0F
;
341 static void do_LOG( struct arb_vp_machine
*m
, union instruction op
)
343 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
344 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
345 GLfloat tmp
= FABSF(arg0
[0]);
347 GLfloat mantissa
= FREXPF(tmp
, &exponent
);
349 result
[0] = (GLfloat
) (exponent
- 1);
350 result
[1] = 2.0 * mantissa
; /* map [.5, 1) -> [1, 2) */
351 result
[2] = result
[0] + LOG2(result
[1]);
355 static void do_MAX( struct arb_vp_machine
*m
, union instruction op
)
357 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
358 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
359 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
361 result
[0] = (arg0
[0] > arg1
[0]) ? arg0
[0] : arg1
[0];
362 result
[1] = (arg0
[1] > arg1
[1]) ? arg0
[1] : arg1
[1];
363 result
[2] = (arg0
[2] > arg1
[2]) ? arg0
[2] : arg1
[2];
364 result
[3] = (arg0
[3] > arg1
[3]) ? arg0
[3] : arg1
[3];
368 static void do_MIN( struct arb_vp_machine
*m
, union instruction op
)
370 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
371 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
372 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
374 result
[0] = (arg0
[0] < arg1
[0]) ? arg0
[0] : arg1
[0];
375 result
[1] = (arg0
[1] < arg1
[1]) ? arg0
[1] : arg1
[1];
376 result
[2] = (arg0
[2] < arg1
[2]) ? arg0
[2] : arg1
[2];
377 result
[3] = (arg0
[3] < arg1
[3]) ? arg0
[3] : arg1
[3];
380 static void do_MOV( struct arb_vp_machine
*m
, union instruction op
)
382 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
383 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
391 static void do_MUL( struct arb_vp_machine
*m
, union instruction op
)
393 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
394 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
395 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
397 result
[0] = arg0
[0] * arg1
[0];
398 result
[1] = arg0
[1] * arg1
[1];
399 result
[2] = arg0
[2] * arg1
[2];
400 result
[3] = arg0
[3] * arg1
[3];
404 static void do_POW( struct arb_vp_machine
*m
, union instruction op
)
406 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
407 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
408 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
410 result
[0] = (GLfloat
)RoughApproxPower(arg0
[0], arg1
[0]);
414 static void do_REL( struct arb_vp_machine
*m
, union instruction op
)
416 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
417 GLuint idx
= (op
.alu
.idx0
+ (GLint
)m
->File
[0][REG_ADDR
][0]) & (MAX_NV_VERTEX_PROGRAM_PARAMS
-1);
418 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][idx
];
426 static void do_RCP( struct arb_vp_machine
*m
, union instruction op
)
428 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
429 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
431 result
[0] = 1.0F
/ arg0
[0];
435 static void do_RSQ( struct arb_vp_machine
*m
, union instruction op
)
437 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
438 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
440 result
[0] = INV_SQRTF(FABSF(arg0
[0]));
445 static void do_SGE( struct arb_vp_machine
*m
, union instruction op
)
447 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
448 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
449 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
451 result
[0] = (arg0
[0] >= arg1
[0]) ? 1.0F
: 0.0F
;
452 result
[1] = (arg0
[1] >= arg1
[1]) ? 1.0F
: 0.0F
;
453 result
[2] = (arg0
[2] >= arg1
[2]) ? 1.0F
: 0.0F
;
454 result
[3] = (arg0
[3] >= arg1
[3]) ? 1.0F
: 0.0F
;
458 static void do_SLT( struct arb_vp_machine
*m
, union instruction op
)
460 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
461 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
462 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
464 result
[0] = (arg0
[0] < arg1
[0]) ? 1.0F
: 0.0F
;
465 result
[1] = (arg0
[1] < arg1
[1]) ? 1.0F
: 0.0F
;
466 result
[2] = (arg0
[2] < arg1
[2]) ? 1.0F
: 0.0F
;
467 result
[3] = (arg0
[3] < arg1
[3]) ? 1.0F
: 0.0F
;
470 static void do_SUB( struct arb_vp_machine
*m
, union instruction op
)
472 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
473 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
474 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
476 result
[0] = arg0
[0] - arg1
[0];
477 result
[1] = arg0
[1] - arg1
[1];
478 result
[2] = arg0
[2] - arg1
[2];
479 result
[3] = arg0
[3] - arg1
[3];
483 static void do_XPD( struct arb_vp_machine
*m
, union instruction op
)
485 GLfloat
*result
= m
->File
[0][op
.alu
.dst
];
486 const GLfloat
*arg0
= m
->File
[op
.alu
.file0
][op
.alu
.idx0
];
487 const GLfloat
*arg1
= m
->File
[op
.alu
.file1
][op
.alu
.idx1
];
489 result
[0] = arg0
[1] * arg1
[2] - arg0
[2] * arg1
[1];
490 result
[1] = arg0
[2] * arg1
[0] - arg0
[0] * arg1
[2];
491 result
[2] = arg0
[0] * arg1
[1] - arg0
[1] * arg1
[0];
494 static void do_NOP( struct arb_vp_machine
*m
, union instruction op
)
498 /* Some useful debugging functions:
500 static void print_mask( GLuint mask
)
503 if (mask
&0x1) _mesa_printf("x");
504 if (mask
&0x2) _mesa_printf("y");
505 if (mask
&0x4) _mesa_printf("z");
506 if (mask
&0x8) _mesa_printf("w");
509 static void print_reg( GLuint file
, GLuint reg
)
511 static const char *reg_file
[] = {
521 else if (reg
>= REG_ARG0
&& reg
<= REG_ARG1
)
522 _mesa_printf("ARG%d", reg
- REG_ARG0
);
523 else if (reg
>= REG_TMP0
&& reg
<= REG_TMP11
)
524 _mesa_printf("TMP%d", reg
- REG_TMP0
);
525 else if (reg
>= REG_IN0
&& reg
<= REG_IN31
)
526 _mesa_printf("IN%d", reg
- REG_IN0
);
527 else if (reg
>= REG_OUT0
&& reg
<= REG_OUT14
)
528 _mesa_printf("OUT%d", reg
- REG_OUT0
);
529 else if (reg
== REG_ADDR
)
530 _mesa_printf("ADDR");
531 else if (reg
== REG_ID
)
534 _mesa_printf("REG%d", reg
);
537 _mesa_printf("%s:%d", reg_file
[file
], reg
);
541 static void print_RSW( union instruction op
, const struct opcode_info
*info
)
543 GLuint swz
= op
.rsw
.swz
;
544 GLuint neg
= op
.rsw
.neg
;
547 _mesa_printf("%s ", info
->string
);
548 print_reg(0, op
.rsw
.dst
);
550 print_reg(op
.rsw
.file0
, op
.rsw
.idx0
);
552 for (i
= 0; i
< 4; i
++, swz
>>= 2) {
553 const char *cswz
= "xyzw";
556 _mesa_printf("%c", cswz
[swz
&0x3]);
562 static void print_ALU( union instruction op
, const struct opcode_info
*info
)
564 _mesa_printf("%s ", info
->string
);
565 print_reg(0, op
.alu
.dst
);
567 print_reg(op
.alu
.file0
, op
.alu
.idx0
);
568 if (info
->nr_args
> 1) {
570 print_reg(op
.alu
.file1
, op
.alu
.idx1
);
575 static void print_MSK( union instruction op
, const struct opcode_info
*info
)
577 _mesa_printf("%s ", info
->string
);
578 print_reg(0, op
.msk
.dst
);
579 print_mask(op
.msk
.mask
);
581 print_reg(op
.msk
.file
, op
.msk
.idx
);
586 static void print_NOP( union instruction op
, const struct opcode_info
*info
)
594 static const struct opcode_info opcode_info
[] =
596 { 1, "ABS", print_ALU
},
597 { 2, "ADD", print_ALU
},
598 { 1, "ARL", print_NOP
},
599 { 2, "DP3", print_ALU
},
600 { 2, "DP4", print_ALU
},
601 { 2, "DPH", print_ALU
},
602 { 2, "DST", print_ALU
},
603 { 0, "END", print_NOP
},
604 { 1, "EX2", print_ALU
},
605 { 1, "EXP", print_ALU
},
606 { 1, "FLR", print_ALU
},
607 { 1, "FRC", print_ALU
},
608 { 1, "LG2", print_ALU
},
609 { 1, "LIT", print_ALU
},
610 { 1, "LOG", print_ALU
},
611 { 3, "MAD", print_NOP
},
612 { 2, "MAX", print_ALU
},
613 { 2, "MIN", print_ALU
},
614 { 1, "MOV", print_ALU
},
615 { 2, "MUL", print_ALU
},
616 { 2, "POW", print_ALU
},
617 { 1, "PRT", print_ALU
}, /* PRINT */
618 { 1, "RCC", print_NOP
},
619 { 1, "RCP", print_ALU
},
620 { 1, "RSQ", print_ALU
},
621 { 2, "SGE", print_ALU
},
622 { 2, "SLT", print_ALU
},
623 { 2, "SUB", print_ALU
},
624 { 1, "SWZ", print_NOP
},
625 { 2, "XPD", print_ALU
},
626 { 1, "RSW", print_RSW
},
627 { 2, "MSK", print_MSK
},
628 { 1, "REL", print_ALU
},
631 void _tnl_disassem_vba_insn( union instruction op
)
633 const struct opcode_info
*info
= &opcode_info
[op
.alu
.opcode
];
634 info
->print( op
, info
);
638 static void (* const opcode_func
[])(struct arb_vp_machine
*, union instruction
) =
675 static union instruction
*cvp_next_instruction( struct compilation
*cp
)
677 union instruction
*op
= cp
->csr
++;
682 static struct reg
cvp_make_reg( GLuint file
, GLuint idx
)
690 static struct reg
cvp_emit_rel( struct compilation
*cp
,
694 union instruction
*op
= cvp_next_instruction(cp
);
695 op
->alu
.opcode
= REL
;
696 op
->alu
.file0
= reg
.file
;
697 op
->alu
.idx0
= reg
.idx
;
698 op
->alu
.dst
= tmpreg
.idx
;
703 static struct reg
cvp_load_reg( struct compilation
*cp
,
709 struct reg tmpreg
= cvp_make_reg(FILE_REG
, tmpidx
);
713 case PROGRAM_TEMPORARY
:
714 return cvp_make_reg(FILE_REG
, REG_TMP0
+ index
);
717 return cvp_make_reg(FILE_REG
, REG_IN0
+ index
);
720 return cvp_make_reg(FILE_REG
, REG_OUT0
+ index
);
722 /* These two aren't populated by the parser?
724 case PROGRAM_LOCAL_PARAM
:
725 reg
= cvp_make_reg(FILE_LOCAL_PARAM
, index
);
727 return cvp_emit_rel(cp
, reg
, tmpreg
);
731 case PROGRAM_ENV_PARAM
:
732 reg
= cvp_make_reg(FILE_ENV_PARAM
, index
);
734 return cvp_emit_rel(cp
, reg
, tmpreg
);
738 case PROGRAM_STATE_VAR
:
739 reg
= cvp_make_reg(FILE_STATE_PARAM
, index
);
741 return cvp_emit_rel(cp
, reg
, tmpreg
);
747 case PROGRAM_WRITE_ONLY
:
748 case PROGRAM_ADDRESS
:
751 return tmpreg
; /* can't happen */
755 static struct reg
cvp_emit_arg( struct compilation
*cp
,
756 const struct vp_src_register
*src
,
759 struct reg reg
= cvp_load_reg( cp
, src
->File
, src
->Index
, src
->RelAddr
, arg
);
760 union instruction rsw
, noop
;
762 /* Emit any necessary swizzling.
765 rsw
.rsw
.neg
= src
->Negate
? WRITEMASK_XYZW
: 0;
766 rsw
.rsw
.swz
= ((GET_SWZ(src
->Swizzle
, 0) << 0) |
767 (GET_SWZ(src
->Swizzle
, 1) << 2) |
768 (GET_SWZ(src
->Swizzle
, 2) << 4) |
769 (GET_SWZ(src
->Swizzle
, 3) << 6));
773 noop
.rsw
.swz
= RSW_NOOP
;
775 if (rsw
.dword
!= noop
.dword
) {
776 union instruction
*op
= cvp_next_instruction(cp
);
777 struct reg rsw_reg
= cvp_make_reg(FILE_REG
, REG_ARG0
+ arg
);
778 op
->dword
= rsw
.dword
;
779 op
->rsw
.opcode
= RSW
;
780 op
->rsw
.file0
= reg
.file
;
781 op
->rsw
.idx0
= reg
.idx
;
782 op
->rsw
.dst
= rsw_reg
.idx
;
789 static GLuint
cvp_choose_result( struct compilation
*cp
,
790 const struct vp_dst_register
*dst
,
791 union instruction
*fixup
)
793 GLuint mask
= dst
->WriteMask
;
797 case PROGRAM_TEMPORARY
:
798 idx
= REG_TMP0
+ dst
->Index
;
801 idx
= REG_OUT0
+ dst
->Index
;
805 return REG_RES
; /* can't happen */
808 /* Optimization: When writing (with a writemask) to an undefined
809 * value for the first time, the writemask may be ignored.
811 if (mask
!= WRITEMASK_XYZW
&& (cp
->reg_active
& (1 << idx
))) {
812 fixup
->msk
.opcode
= MSK
;
813 fixup
->msk
.dst
= idx
;
814 fixup
->msk
.file
= FILE_REG
;
815 fixup
->msk
.idx
= REG_RES
;
816 fixup
->msk
.mask
= mask
;
817 cp
->reg_active
|= 1 << idx
;
822 cp
->reg_active
|= 1 << idx
;
827 static struct reg
cvp_emit_rsw( struct compilation
*cp
,
836 if (swz
!= RSW_NOOP
|| neg
!= 0) {
837 union instruction
*op
= cvp_next_instruction(cp
);
838 op
->rsw
.opcode
= RSW
;
840 op
->rsw
.file0
= src
.file
;
841 op
->rsw
.idx0
= src
.idx
;
845 retval
.file
= FILE_REG
;
850 /* Oops. Degenerate case:
852 union instruction
*op
= cvp_next_instruction(cp
);
853 op
->alu
.opcode
= VP_OPCODE_MOV
;
855 op
->alu
.file0
= src
.file
;
856 op
->alu
.idx0
= src
.idx
;
858 retval
.file
= FILE_REG
;
868 static void cvp_emit_inst( struct compilation
*cp
,
869 const struct vp_instruction
*inst
)
871 const struct opcode_info
*info
= &opcode_info
[inst
->Opcode
];
872 union instruction
*op
;
873 union instruction fixup
;
877 assert(sizeof(*op
) == sizeof(GLuint
));
879 /* Need to handle SWZ, ARL specially.
881 switch (inst
->Opcode
) {
882 /* Split into mul and add:
885 result
= cvp_choose_result( cp
, &inst
->DstReg
, &fixup
);
886 for (i
= 0; i
< 3; i
++)
887 reg
[i
] = cvp_emit_arg( cp
, &inst
->SrcReg
[i
], REG_ARG0
+i
);
889 op
= cvp_next_instruction(cp
);
890 op
->alu
.opcode
= VP_OPCODE_MUL
;
891 op
->alu
.file0
= reg
[0].file
;
892 op
->alu
.idx0
= reg
[0].idx
;
893 op
->alu
.file1
= reg
[1].file
;
894 op
->alu
.idx1
= reg
[1].idx
;
895 op
->alu
.dst
= REG_ARG0
;
897 op
= cvp_next_instruction(cp
);
898 op
->alu
.opcode
= VP_OPCODE_ADD
;
899 op
->alu
.file0
= FILE_REG
;
900 op
->alu
.idx0
= REG_ARG0
;
901 op
->alu
.file1
= reg
[2].file
;
902 op
->alu
.idx1
= reg
[2].idx
;
903 op
->alu
.dst
= result
;
907 reg
[0] = cvp_emit_arg( cp
, &inst
->SrcReg
[0], REG_ARG0
);
909 op
= cvp_next_instruction(cp
);
910 op
->alu
.opcode
= VP_OPCODE_FLR
;
911 op
->alu
.dst
= REG_ADDR
;
912 op
->alu
.file0
= reg
[0].file
;
913 op
->alu
.idx0
= reg
[0].idx
;
916 case VP_OPCODE_SWZ
: {
917 GLuint swz0
= 0, swz1
= 0;
918 GLuint neg0
= 0, neg1
= 0;
921 /* Translate 3-bit-per-element swizzle into two 2-bit swizzles,
922 * one from the source register the other from a constant
925 for (i
= 0; i
< 4; i
++) {
926 GLuint swzelt
= GET_SWZ(inst
->SrcReg
[0].Swizzle
, i
);
927 if (swzelt
>= SWIZZLE_ZERO
) {
928 neg0
|= inst
->SrcReg
[0].Negate
& (1<<i
);
929 if (swzelt
== SWIZZLE_ONE
)
930 swz0
|= SWIZZLE_W
<< (i
*2);
931 else if (i
< SWIZZLE_W
)
936 neg1
|= inst
->SrcReg
[0].Negate
& (1<<i
);
937 swz1
|= swzelt
<< (i
*2);
941 result
= cvp_choose_result( cp
, &inst
->DstReg
, &fixup
);
942 reg
[0].file
= FILE_REG
;
944 reg
[1] = cvp_emit_arg( cp
, &inst
->SrcReg
[0], REG_ARG0
);
946 if (mask
== WRITEMASK_XYZW
) {
947 cvp_emit_rsw(cp
, result
, reg
[0], neg0
, swz0
, GL_TRUE
);
950 else if (mask
== 0) {
951 cvp_emit_rsw(cp
, result
, reg
[1], neg1
, swz1
, GL_TRUE
);
954 cvp_emit_rsw(cp
, result
, reg
[0], neg0
, swz0
, GL_TRUE
);
955 reg
[1] = cvp_emit_rsw(cp
, REG_ARG0
, reg
[1], neg1
, swz1
, GL_FALSE
);
957 op
= cvp_next_instruction(cp
);
958 op
->msk
.opcode
= MSK
;
959 op
->msk
.dst
= result
;
960 op
->msk
.file
= reg
[1].file
;
961 op
->msk
.idx
= reg
[1].idx
;
965 if (result
== REG_RES
) {
966 op
= cvp_next_instruction(cp
);
967 op
->dword
= fixup
.dword
;
971 case VP_OPCODE_PRINT
:
976 result
= cvp_choose_result( cp
, &inst
->DstReg
, &fixup
);
977 for (i
= 0; i
< info
->nr_args
; i
++)
978 reg
[i
] = cvp_emit_arg( cp
, &inst
->SrcReg
[i
], REG_ARG0
+ i
);
980 op
= cvp_next_instruction(cp
);
981 op
->alu
.opcode
= inst
->Opcode
;
982 op
->alu
.file0
= reg
[0].file
;
983 op
->alu
.idx0
= reg
[0].idx
;
984 op
->alu
.file1
= reg
[1].file
;
985 op
->alu
.idx1
= reg
[1].idx
;
986 op
->alu
.dst
= result
;
988 if (result
== REG_RES
) {
989 op
= cvp_next_instruction(cp
);
990 op
->dword
= fixup
.dword
;
996 static void free_tnl_data( struct vertex_program
*program
)
998 struct tnl_compiled_program
*p
= program
->TnlData
;
999 if (p
->compiled_func
) free((void *)p
->compiled_func
);
1001 program
->TnlData
= NULL
;
1004 static void compile_vertex_program( struct vertex_program
*program
,
1005 GLboolean try_codegen
)
1007 struct compilation cp
;
1008 struct tnl_compiled_program
*p
= CALLOC_STRUCT(tnl_compiled_program
);
1011 _mesa_printf("%s\n", __FUNCTION__
);
1013 if (program
->TnlData
)
1014 free_tnl_data( program
);
1016 program
->TnlData
= p
;
1018 /* Initialize cp. Note that ctx and VB aren't used in compilation
1019 * so we don't have to worry about statechanges:
1021 memset(&cp
, 0, sizeof(cp
));
1022 cp
.csr
= p
->instructions
;
1024 /* Compile instructions:
1026 for (i
= 0; i
< program
->Base
.NumInstructions
; i
++) {
1027 cvp_emit_inst(&cp
, &program
->Instructions
[i
]);
1032 p
->nr_instructions
= cp
.csr
- p
->instructions
;
1034 /* Print/disassemble:
1037 for (i
= 0; i
< p
->nr_instructions
; i
++) {
1038 _tnl_disassem_vba_insn(p
->instructions
[i
]);
1040 _mesa_printf("\n\n");
1045 _tnl_sse_codegen_vertex_program(p
);
1053 /* ----------------------------------------------------------------------
1056 static void userclip( GLcontext
*ctx
,
1059 GLubyte
*clipormask
,
1060 GLubyte
*clipandmask
)
1064 for (p
= 0; p
< ctx
->Const
.MaxClipPlanes
; p
++) {
1065 if (ctx
->Transform
.ClipPlanesEnabled
& (1 << p
)) {
1067 const GLfloat a
= ctx
->Transform
._ClipUserPlane
[p
][0];
1068 const GLfloat b
= ctx
->Transform
._ClipUserPlane
[p
][1];
1069 const GLfloat c
= ctx
->Transform
._ClipUserPlane
[p
][2];
1070 const GLfloat d
= ctx
->Transform
._ClipUserPlane
[p
][3];
1071 GLfloat
*coord
= (GLfloat
*)clip
->data
;
1072 GLuint stride
= clip
->stride
;
1073 GLuint count
= clip
->count
;
1075 for (nr
= 0, i
= 0 ; i
< count
; i
++) {
1076 GLfloat dp
= (coord
[0] * a
+
1083 clipmask
[i
] |= CLIP_USER_BIT
;
1086 STRIDE_F(coord
, stride
);
1090 *clipormask
|= CLIP_USER_BIT
;
1092 *clipandmask
|= CLIP_USER_BIT
;
1101 static GLboolean
do_ndc_cliptest( struct arb_vp_machine
*m
)
1103 GLcontext
*ctx
= m
->ctx
;
1104 TNLcontext
*tnl
= TNL_CONTEXT(ctx
);
1105 struct vertex_buffer
*VB
= m
->VB
;
1107 /* Cliptest and perspective divide. Clip functions must clear
1111 m
->andmask
= CLIP_ALL_BITS
;
1113 if (tnl
->NeedNdcCoords
) {
1115 _mesa_clip_tab
[VB
->ClipPtr
->size
]( VB
->ClipPtr
,
1123 _mesa_clip_np_tab
[VB
->ClipPtr
->size
]( VB
->ClipPtr
,
1131 /* All vertices are outside the frustum */
1135 /* Test userclip planes. This contributes to VB->ClipMask.
1137 if (ctx
->Transform
.ClipPlanesEnabled
&& !ctx
->VertexProgram
._Enabled
) {
1149 VB
->ClipAndMask
= m
->andmask
;
1150 VB
->ClipOrMask
= m
->ormask
;
1151 VB
->ClipMask
= m
->clipmask
;
1157 static INLINE
void call_func( struct tnl_compiled_program
*p
,
1158 struct arb_vp_machine
*m
)
1160 p
->compiled_func(m
);
1164 * Execute the given vertex program.
1166 * TODO: Integrate the t_vertex.c code here, to build machine vertices
1167 * directly at this point.
1169 * TODO: Eliminate the VB struct entirely and just use
1170 * struct arb_vertex_machine.
1173 run_arb_vertex_program(GLcontext
*ctx
, struct tnl_pipeline_stage
*stage
)
1175 struct vertex_program
*program
= (ctx
->VertexProgram
._Enabled
?
1176 ctx
->VertexProgram
.Current
:
1178 struct vertex_buffer
*VB
= &TNL_CONTEXT(ctx
)->vb
;
1179 struct arb_vp_machine
*m
= ARB_VP_MACHINE(stage
);
1180 struct tnl_compiled_program
*p
;
1181 GLuint i
, j
, outputs
;
1183 if (!program
|| program
->IsNVProgram
)
1186 if (program
->Parameters
) {
1187 _mesa_load_state_parameters(ctx
, program
->Parameters
);
1190 p
= (struct tnl_compiled_program
*)program
->TnlData
;
1193 /* Initialize regs where necessary:
1195 ASSIGN_4V(m
->File
[0][REG_ID
], 0, 0, 0, 1);
1197 m
->nr_inputs
= m
->nr_outputs
= 0;
1199 for (i
= 0; i
< _TNL_ATTRIB_MAX
; i
++) {
1200 if (program
->InputsRead
& (1<<i
)) {
1201 GLuint j
= m
->nr_inputs
++;
1202 m
->input
[j
].idx
= i
;
1203 m
->input
[j
].data
= (GLfloat
*)m
->VB
->AttribPtr
[i
]->data
;
1204 m
->input
[j
].stride
= m
->VB
->AttribPtr
[i
]->stride
;
1205 m
->input
[j
].size
= m
->VB
->AttribPtr
[i
]->size
;
1206 ASSIGN_4V(m
->File
[0][REG_IN0
+ i
], 0, 0, 0, 1);
1210 for (i
= 0; i
< 15; i
++) {
1211 if (program
->OutputsWritten
& (1<<i
)) {
1212 GLuint j
= m
->nr_outputs
++;
1213 m
->output
[j
].idx
= i
;
1214 m
->output
[j
].data
= (GLfloat
*)m
->attribs
[i
].data
;
1219 /* Run the actual program:
1221 for (m
->vtx_nr
= 0; m
->vtx_nr
< VB
->Count
; m
->vtx_nr
++) {
1222 for (j
= 0; j
< m
->nr_inputs
; j
++) {
1223 GLuint idx
= REG_IN0
+ m
->input
[j
].idx
;
1224 switch (m
->input
[j
].size
) {
1225 case 4: m
->File
[0][idx
][3] = m
->input
[j
].data
[3];
1226 case 3: m
->File
[0][idx
][2] = m
->input
[j
].data
[2];
1227 case 2: m
->File
[0][idx
][1] = m
->input
[j
].data
[1];
1228 case 1: m
->File
[0][idx
][0] = m
->input
[j
].data
[0];
1231 STRIDE_F(m
->input
[j
].data
, m
->input
[j
].stride
);
1234 if (p
->compiled_func
) {
1238 for (j
= 0; j
< p
->nr_instructions
; j
++) {
1239 union instruction inst
= p
->instructions
[j
];
1240 opcode_func
[inst
.alu
.opcode
]( m
, inst
);
1244 for (j
= 0; j
< m
->nr_outputs
; j
++) {
1245 GLuint idx
= REG_OUT0
+ m
->output
[j
].idx
;
1246 m
->output
[j
].data
[0] = m
->File
[0][idx
][0];
1247 m
->output
[j
].data
[1] = m
->File
[0][idx
][1];
1248 m
->output
[j
].data
[2] = m
->File
[0][idx
][2];
1249 m
->output
[j
].data
[3] = m
->File
[0][idx
][3];
1250 m
->output
[j
].data
+= 4;
1254 /* Setup the VB pointers so that the next pipeline stages get
1255 * their data from the right place (the program output arrays).
1257 * TODO: 1) Have tnl use these RESULT values for outputs rather
1258 * than trying to shoe-horn inputs and outputs into one set of
1261 * TODO: 2) Integrate t_vertex.c so that we just go straight ahead
1262 * and build machine vertices here.
1264 VB
->ClipPtr
= &m
->attribs
[VERT_RESULT_HPOS
];
1265 VB
->ClipPtr
->count
= VB
->Count
;
1267 outputs
= program
->OutputsWritten
;
1269 if (outputs
& (1<<VERT_RESULT_COL0
)) {
1270 VB
->ColorPtr
[0] = &m
->attribs
[VERT_RESULT_COL0
];
1271 VB
->AttribPtr
[VERT_ATTRIB_COLOR0
] = VB
->ColorPtr
[0];
1274 if (outputs
& (1<<VERT_RESULT_BFC0
)) {
1275 VB
->ColorPtr
[1] = &m
->attribs
[VERT_RESULT_BFC0
];
1278 if (outputs
& (1<<VERT_RESULT_COL1
)) {
1279 VB
->SecondaryColorPtr
[0] = &m
->attribs
[VERT_RESULT_COL1
];
1280 VB
->AttribPtr
[VERT_ATTRIB_COLOR1
] = VB
->SecondaryColorPtr
[0];
1283 if (outputs
& (1<<VERT_RESULT_BFC1
)) {
1284 VB
->SecondaryColorPtr
[1] = &m
->attribs
[VERT_RESULT_BFC1
];
1287 if (outputs
& (1<<VERT_RESULT_FOGC
)) {
1288 VB
->FogCoordPtr
= &m
->attribs
[VERT_RESULT_FOGC
];
1289 VB
->AttribPtr
[VERT_ATTRIB_FOG
] = VB
->FogCoordPtr
;
1292 if (outputs
& (1<<VERT_RESULT_PSIZ
)) {
1293 VB
->PointSizePtr
= &m
->attribs
[VERT_RESULT_PSIZ
];
1294 VB
->AttribPtr
[_TNL_ATTRIB_POINTSIZE
] = &m
->attribs
[VERT_RESULT_PSIZ
];
1297 for (i
= 0; i
< ctx
->Const
.MaxTextureUnits
; i
++) {
1298 if (outputs
& (1<<(VERT_RESULT_TEX0
+i
))) {
1299 VB
->TexCoordPtr
[i
] = &m
->attribs
[VERT_RESULT_TEX0
+ i
];
1300 VB
->AttribPtr
[VERT_ATTRIB_TEX0
+i
] = VB
->TexCoordPtr
[i
];
1305 for (i
= 0; i
< VB
->Count
; i
++) {
1306 printf("Out %d: %f %f %f %f %f %f %f %f\n", i
,
1307 VEC_ELT(VB
->ClipPtr
, GLfloat
, i
)[0],
1308 VEC_ELT(VB
->ClipPtr
, GLfloat
, i
)[1],
1309 VEC_ELT(VB
->ClipPtr
, GLfloat
, i
)[2],
1310 VEC_ELT(VB
->ClipPtr
, GLfloat
, i
)[3],
1311 VEC_ELT(VB
->TexCoordPtr
[0], GLfloat
, i
)[0],
1312 VEC_ELT(VB
->TexCoordPtr
[0], GLfloat
, i
)[1],
1313 VEC_ELT(VB
->TexCoordPtr
[0], GLfloat
, i
)[2],
1314 VEC_ELT(VB
->TexCoordPtr
[0], GLfloat
, i
)[3]);
1318 /* Perform NDC and cliptest operations:
1320 return do_ndc_cliptest(m
);
1325 validate_vertex_program( GLcontext
*ctx
, struct tnl_pipeline_stage
*stage
)
1327 struct arb_vp_machine
*m
= ARB_VP_MACHINE(stage
);
1328 struct vertex_program
*program
=
1329 (ctx
->VertexProgram
._Enabled
? ctx
->VertexProgram
.Current
: 0);
1331 if (!program
&& ctx
->_MaintainTnlProgram
) {
1332 program
= ctx
->_TnlProgram
;
1336 if (!program
->TnlData
)
1337 compile_vertex_program( program
, m
->try_codegen
);
1339 /* Grab the state GL state and put into registers:
1341 m
->File
[FILE_LOCAL_PARAM
] = program
->Base
.LocalParams
;
1342 m
->File
[FILE_ENV_PARAM
] = ctx
->VertexProgram
.Parameters
;
1343 m
->File
[FILE_STATE_PARAM
] = program
->Parameters
->ParameterValues
;
1354 * Called the first time stage->run is called. In effect, don't
1355 * allocate data until the first time the stage is run.
1357 static GLboolean
init_vertex_program( GLcontext
*ctx
,
1358 struct tnl_pipeline_stage
*stage
)
1360 TNLcontext
*tnl
= TNL_CONTEXT(ctx
);
1361 struct vertex_buffer
*VB
= &(tnl
->vb
);
1362 struct arb_vp_machine
*m
;
1363 const GLuint size
= VB
->Size
;
1366 stage
->privatePtr
= MALLOC(sizeof(*m
));
1367 m
= ARB_VP_MACHINE(stage
);
1371 /* arb_vertex_machine struct should subsume the VB:
1376 m
->File
[0] = ALIGN_MALLOC(REG_MAX
* sizeof(GLfloat
) * 4, 16);
1378 if (_mesa_getenv("MESA_EXPERIMENTAL"))
1381 /* Allocate arrays of vertex output values */
1382 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
1383 _mesa_vector4f_alloc( &m
->attribs
[i
], 0, size
, 32 );
1384 m
->attribs
[i
].size
= 4;
1387 /* a few other misc allocations */
1388 _mesa_vector4f_alloc( &m
->ndcCoords
, 0, size
, 32 );
1389 m
->clipmask
= (GLubyte
*) ALIGN_MALLOC(sizeof(GLubyte
)*size
, 32 );
1391 if (ctx
->_MaintainTnlProgram
)
1392 _mesa_allow_light_in_model( ctx
, GL_FALSE
);
1401 * Destructor for this pipeline stage.
1403 static void dtr( struct tnl_pipeline_stage
*stage
)
1405 struct arb_vp_machine
*m
= ARB_VP_MACHINE(stage
);
1410 /* free the vertex program result arrays */
1411 for (i
= 0; i
< VERT_RESULT_MAX
; i
++)
1412 _mesa_vector4f_free( &m
->attribs
[i
] );
1414 /* free misc arrays */
1415 _mesa_vector4f_free( &m
->ndcCoords
);
1416 ALIGN_FREE( m
->clipmask
);
1417 ALIGN_FREE( m
->File
[0] );
1420 stage
->privatePtr
= NULL
;
1425 * Public description of this pipeline stage.
1427 const struct tnl_pipeline_stage _tnl_arb_vertex_program_stage
=
1430 NULL
, /* private_data */
1431 init_vertex_program
, /* create */
1433 validate_vertex_program
, /* validate */
1434 run_arb_vertex_program
/* run */