2 Copyright (C) Intel Corp. 2006. All Rights Reserved.
3 Intel funded Tungsten Graphics (http://www.tungstengraphics.com) to
4 develop this 3D driver.
6 Permission is hereby granted, free of charge, to any person obtaining
7 a copy of this software and associated documentation files (the
8 "Software"), to deal in the Software without restriction, including
9 without limitation the rights to use, copy, modify, merge, publish,
10 distribute, sublicense, and/or sell copies of the Software, and to
11 permit persons to whom the Software is furnished to do so, subject to
12 the following conditions:
14 The above copyright notice and this permission notice (including the
15 next paragraph) shall be included in all copies or substantial
16 portions of the Software.
18 THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
19 EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
20 MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
21 IN NO EVENT SHALL THE COPYRIGHT OWNER(S) AND/OR ITS SUPPLIERS BE
22 LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
23 OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
24 WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
26 **********************************************************************/
29 * Keith Whitwell <keith@tungstengraphics.com>
32 #include "pipe/p_shader_tokens.h"
34 #include "util/u_memory.h"
35 #include "util/u_math.h"
37 #include "tgsi/tgsi_ureg.h"
38 #include "tgsi/tgsi_ureg_parse.h"
39 #include "tgsi/tgsi_dump.h"
40 #include "tgsi/tgsi_info.h"
42 #include "brw_context.h"
44 #include "brw_debug.h"
48 static struct brw_reg
get_tmp( struct brw_vs_compile
*c
)
50 struct brw_reg tmp
= brw_vec8_grf(c
->last_tmp
, 0);
52 if (++c
->last_tmp
> c
->prog_data
.total_grf
)
53 c
->prog_data
.total_grf
= c
->last_tmp
;
58 static void release_tmp( struct brw_vs_compile
*c
, struct brw_reg tmp
)
60 if (tmp
.nr
== c
->last_tmp
-1)
64 static void release_tmps( struct brw_vs_compile
*c
)
66 c
->last_tmp
= c
->first_tmp
;
71 * Preallocate GRF register before code emit.
72 * Do things as simply as possible. Allocate and populate all regs
75 static void brw_vs_alloc_regs( struct brw_vs_compile
*c
)
77 GLuint i
, reg
= 0, mrf
;
78 int attributes_in_vue
;
80 /* Determine whether to use a real constant buffer or use a block
81 * of GRF registers for constants. The later is faster but only
82 * works if everything fits in the GRF.
83 * XXX this heuristic/check may need some fine tuning...
85 if (c
->vp
->info
.file_max
[TGSI_FILE_CONSTANT
] +
86 c
->vp
->info
.file_max
[TGSI_FILE_TEMPORARY
] + 21 > BRW_MAX_GRF
)
87 c
->vp
->use_const_buffer
= GL_TRUE
;
89 c
->vp
->use_const_buffer
= GL_FALSE
;
91 /*printf("use_const_buffer = %d\n", c->vp->use_const_buffer);*/
93 /* r0 -- reserved as usual
95 c
->r0
= brw_vec8_grf(reg
, 0);
98 /* User clip planes from curbe:
100 if (c
->key
.nr_userclip
) {
101 for (i
= 0; i
< c
->key
.nr_userclip
; i
++) {
102 c
->userplane
[i
] = stride( brw_vec4_grf(reg
+3+i
/2, (i
%2) * 4), 0, 4, 1);
105 /* Deal with curbe alignment:
107 reg
+= ((6 + c
->key
.nr_userclip
+ 3) / 4) * 2;
110 /* Vertex program parameters from curbe:
112 if (c
->vp
->use_const_buffer
) {
113 /* get constants from a real constant buffer */
114 c
->prog_data
.curb_read_length
= 0;
115 c
->prog_data
.nr_params
= 4; /* XXX 0 causes a bug elsewhere... */
118 /* use a section of the GRF for constants */
119 GLuint nr_params
= c
->vp
->info
.file_max
[TGSI_FILE_CONSTANT
] + 1;
120 for (i
= 0; i
< nr_params
; i
++) {
121 c
->regs
[TGSI_FILE_CONSTANT
][i
] = stride( brw_vec4_grf(reg
+i
/2, (i
%2) * 4), 0, 4, 1);
123 reg
+= (nr_params
+ 1) / 2;
124 c
->prog_data
.curb_read_length
= reg
- 1;
125 c
->prog_data
.nr_params
= nr_params
* 4;
128 /* Allocate input regs:
130 c
->nr_inputs
= c
->vp
->info
.num_inputs
;
131 for (i
= 0; i
< c
->nr_inputs
; i
++) {
132 c
->regs
[TGSI_FILE_INPUT
][i
] = brw_vec8_grf(reg
, 0);
136 /* If there are no inputs, we'll still be reading one attribute's worth
137 * because it's required -- see urb_read_length setting.
139 if (c
->nr_inputs
== 0)
142 /* Allocate outputs. The non-position outputs go straight into message regs.
145 c
->first_output
= reg
;
146 c
->first_overflow_output
= 0;
148 if (c
->chipset
.is_igdng
)
153 /* XXX: need to access vertex output semantics here:
155 c
->nr_outputs
= c
->prog_data
.nr_outputs
;
156 for (i
= 0; i
< c
->prog_data
.nr_outputs
; i
++) {
157 assert(i
< Elements(c
->regs
[TGSI_FILE_OUTPUT
]));
159 /* XXX: Hardwire position to zero:
162 c
->regs
[TGSI_FILE_OUTPUT
][i
] = brw_vec8_grf(reg
, 0);
165 /* XXX: disable psiz:
168 c
->regs
[TGSI_FILE_OUTPUT
][i
] = brw_vec8_grf(reg
, 0);
170 mrf
++; /* just a placeholder? XXX fix later stages & remove this */
173 c
->regs
[TGSI_FILE_OUTPUT
][i
] = brw_message_reg(mrf
);
177 /* too many vertex results to fit in MRF, use GRF for overflow */
178 if (!c
->first_overflow_output
)
179 c
->first_overflow_output
= i
;
180 c
->regs
[TGSI_FILE_OUTPUT
][i
] = brw_vec8_grf(reg
, 0);
185 /* Allocate program temporaries:
188 for (i
= 0; i
< c
->vp
->info
.file_max
[TGSI_FILE_TEMPORARY
]+1; i
++) {
189 c
->regs
[TGSI_FILE_TEMPORARY
][i
] = brw_vec8_grf(reg
, 0);
193 /* Address reg(s). Don't try to use the internal address reg until
196 for (i
= 0; i
< c
->vp
->info
.file_max
[TGSI_FILE_ADDRESS
]+1; i
++) {
197 c
->regs
[TGSI_FILE_ADDRESS
][i
] = brw_reg(BRW_GENERAL_REGISTER_FILE
,
201 BRW_VERTICAL_STRIDE_8
,
203 BRW_HORIZONTAL_STRIDE_1
,
209 if (c
->vp
->use_const_buffer
) {
210 for (i
= 0; i
< 3; i
++) {
211 c
->current_const
[i
].index
= -1;
212 c
->current_const
[i
].reg
= brw_vec8_grf(reg
, 0);
217 for (i
= 0; i
< 128; i
++) {
218 if (c
->output_regs
[i
].used_in_src
) {
219 c
->output_regs
[i
].reg
= brw_vec8_grf(reg
, 0);
224 c
->stack
= brw_uw16_reg(BRW_GENERAL_REGISTER_FILE
, reg
, 0);
227 /* Some opcodes need an internal temporary:
230 c
->last_tmp
= reg
; /* for allocation purposes */
232 /* Each input reg holds data from two vertices. The
233 * urb_read_length is the number of registers read from *each*
234 * vertex urb, so is half the amount:
236 c
->prog_data
.urb_read_length
= (c
->nr_inputs
+ 1) / 2;
238 /* Setting this field to 0 leads to undefined behavior according to the
239 * the VS_STATE docs. Our VUEs will always have at least one attribute
240 * sitting in them, even if it's padding.
242 if (c
->prog_data
.urb_read_length
== 0)
243 c
->prog_data
.urb_read_length
= 1;
245 /* The VS VUEs are shared by VF (outputting our inputs) and VS, so size
246 * them to fit the biggest thing they need to.
248 attributes_in_vue
= MAX2(c
->nr_outputs
, c
->nr_inputs
);
250 if (c
->chipset
.is_igdng
)
251 c
->prog_data
.urb_entry_size
= (attributes_in_vue
+ 6 + 3) / 4;
253 c
->prog_data
.urb_entry_size
= (attributes_in_vue
+ 2 + 3) / 4;
255 c
->prog_data
.total_grf
= reg
;
257 if (BRW_DEBUG
& DEBUG_VS
) {
258 debug_printf("%s NumAddrRegs %d\n", __FUNCTION__
,
259 c
->vp
->info
.file_max
[TGSI_FILE_ADDRESS
]+1);
260 debug_printf("%s NumTemps %d\n", __FUNCTION__
,
261 c
->vp
->info
.file_max
[TGSI_FILE_TEMPORARY
]+1);
262 debug_printf("%s reg = %d\n", __FUNCTION__
, reg
);
268 * If an instruction uses a temp reg both as a src and the dest, we
269 * sometimes need to allocate an intermediate temporary.
271 static void unalias1( struct brw_vs_compile
*c
,
274 void (*func
)( struct brw_vs_compile
*,
278 if (dst
.file
== arg0
.file
&& dst
.nr
== arg0
.nr
) {
279 struct brw_compile
*p
= &c
->func
;
280 struct brw_reg tmp
= brw_writemask(get_tmp(c
), dst
.dw1
.bits
.writemask
);
282 brw_MOV(p
, dst
, tmp
);
292 * Checkes if 2-operand instruction needs an intermediate temporary.
294 static void unalias2( struct brw_vs_compile
*c
,
298 void (*func
)( struct brw_vs_compile
*,
303 if ((dst
.file
== arg0
.file
&& dst
.nr
== arg0
.nr
) ||
304 (dst
.file
== arg1
.file
&& dst
.nr
== arg1
.nr
)) {
305 struct brw_compile
*p
= &c
->func
;
306 struct brw_reg tmp
= brw_writemask(get_tmp(c
), dst
.dw1
.bits
.writemask
);
307 func(c
, tmp
, arg0
, arg1
);
308 brw_MOV(p
, dst
, tmp
);
312 func(c
, dst
, arg0
, arg1
);
318 * Checkes if 3-operand instruction needs an intermediate temporary.
320 static void unalias3( struct brw_vs_compile
*c
,
325 void (*func
)( struct brw_vs_compile
*,
331 if ((dst
.file
== arg0
.file
&& dst
.nr
== arg0
.nr
) ||
332 (dst
.file
== arg1
.file
&& dst
.nr
== arg1
.nr
) ||
333 (dst
.file
== arg2
.file
&& dst
.nr
== arg2
.nr
)) {
334 struct brw_compile
*p
= &c
->func
;
335 struct brw_reg tmp
= brw_writemask(get_tmp(c
), dst
.dw1
.bits
.writemask
);
336 func(c
, tmp
, arg0
, arg1
, arg2
);
337 brw_MOV(p
, dst
, tmp
);
341 func(c
, dst
, arg0
, arg1
, arg2
);
345 static void emit_sop( struct brw_compile
*p
,
351 brw_MOV(p
, dst
, brw_imm_f(0.0f
));
352 brw_CMP(p
, brw_null_reg(), cond
, arg0
, arg1
);
353 brw_MOV(p
, dst
, brw_imm_f(1.0f
));
354 brw_set_predicate_control_flag_value(p
, 0xff);
357 static void emit_seq( struct brw_compile
*p
,
360 struct brw_reg arg1
)
362 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_EQ
);
365 static void emit_sne( struct brw_compile
*p
,
368 struct brw_reg arg1
)
370 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_NEQ
);
372 static void emit_slt( struct brw_compile
*p
,
375 struct brw_reg arg1
)
377 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_L
);
380 static void emit_sle( struct brw_compile
*p
,
383 struct brw_reg arg1
)
385 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_LE
);
388 static void emit_sgt( struct brw_compile
*p
,
391 struct brw_reg arg1
)
393 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_G
);
396 static void emit_sge( struct brw_compile
*p
,
399 struct brw_reg arg1
)
401 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_GE
);
404 static void emit_max( struct brw_compile
*p
,
407 struct brw_reg arg1
)
409 brw_CMP(p
, brw_null_reg(), BRW_CONDITIONAL_L
, arg0
, arg1
);
410 brw_SEL(p
, dst
, arg1
, arg0
);
411 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
414 static void emit_min( struct brw_compile
*p
,
417 struct brw_reg arg1
)
419 brw_CMP(p
, brw_null_reg(), BRW_CONDITIONAL_L
, arg0
, arg1
);
420 brw_SEL(p
, dst
, arg0
, arg1
);
421 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
425 static void emit_math1( struct brw_vs_compile
*c
,
431 /* There are various odd behaviours with SEND on the simulator. In
432 * addition there are documented issues with the fact that the GEN4
433 * processor doesn't do dependency control properly on SEND
434 * results. So, on balance, this kludge to get around failures
435 * with writemasked math results looks like it might be necessary
436 * whether that turns out to be a simulator bug or not:
438 struct brw_compile
*p
= &c
->func
;
439 struct brw_reg tmp
= dst
;
440 GLboolean need_tmp
= (dst
.dw1
.bits
.writemask
!= 0xf ||
441 dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
449 BRW_MATH_SATURATE_NONE
,
452 BRW_MATH_DATA_SCALAR
,
456 brw_MOV(p
, dst
, tmp
);
462 static void emit_math2( struct brw_vs_compile
*c
,
469 struct brw_compile
*p
= &c
->func
;
470 struct brw_reg tmp
= dst
;
471 GLboolean need_tmp
= (dst
.dw1
.bits
.writemask
!= 0xf ||
472 dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
477 brw_MOV(p
, brw_message_reg(3), arg1
);
482 BRW_MATH_SATURATE_NONE
,
485 BRW_MATH_DATA_SCALAR
,
489 brw_MOV(p
, dst
, tmp
);
495 static void emit_exp_noalias( struct brw_vs_compile
*c
,
497 struct brw_reg arg0
)
499 struct brw_compile
*p
= &c
->func
;
502 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_X
) {
503 struct brw_reg tmp
= get_tmp(c
);
504 struct brw_reg tmp_d
= retype(tmp
, BRW_REGISTER_TYPE_D
);
506 /* tmp_d = floor(arg0.x) */
507 brw_RNDD(p
, tmp_d
, brw_swizzle1(arg0
, 0));
509 /* result[0] = 2.0 ^ tmp */
511 /* Adjust exponent for floating point:
514 brw_ADD(p
, brw_writemask(tmp_d
, BRW_WRITEMASK_X
), tmp_d
, brw_imm_d(127));
516 /* Install exponent and sign.
517 * Excess drops off the edge:
519 brw_SHL(p
, brw_writemask(retype(dst
, BRW_REGISTER_TYPE_D
), BRW_WRITEMASK_X
),
520 tmp_d
, brw_imm_d(23));
525 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_Y
) {
526 /* result[1] = arg0.x - floor(arg0.x) */
527 brw_FRC(p
, brw_writemask(dst
, BRW_WRITEMASK_Y
), brw_swizzle1(arg0
, 0));
530 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_Z
) {
531 /* As with the LOG instruction, we might be better off just
532 * doing a taylor expansion here, seeing as we have to do all
535 * If mathbox partial precision is too low, consider also:
536 * result[3] = result[0] * EXP(result[1])
539 BRW_MATH_FUNCTION_EXP
,
540 brw_writemask(dst
, BRW_WRITEMASK_Z
),
541 brw_swizzle1(arg0
, 0),
542 BRW_MATH_PRECISION_FULL
);
545 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_W
) {
546 /* result[3] = 1.0; */
547 brw_MOV(p
, brw_writemask(dst
, BRW_WRITEMASK_W
), brw_imm_f(1));
552 static void emit_log_noalias( struct brw_vs_compile
*c
,
554 struct brw_reg arg0
)
556 struct brw_compile
*p
= &c
->func
;
557 struct brw_reg tmp
= dst
;
558 struct brw_reg tmp_ud
= retype(tmp
, BRW_REGISTER_TYPE_UD
);
559 struct brw_reg arg0_ud
= retype(arg0
, BRW_REGISTER_TYPE_UD
);
560 GLboolean need_tmp
= (dst
.dw1
.bits
.writemask
!= 0xf ||
561 dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
565 tmp_ud
= retype(tmp
, BRW_REGISTER_TYPE_UD
);
568 /* Perform mant = frexpf(fabsf(x), &exp), adjust exp and mnt
571 * These almost look likey they could be joined up, but not really
574 * result[0].f = (x.i & ((1<<31)-1) >> 23) - 127
575 * result[1].i = (x.i & ((1<<23)-1) + (127<<23)
577 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_XZ
) {
579 brw_writemask(tmp_ud
, BRW_WRITEMASK_X
),
580 brw_swizzle1(arg0_ud
, 0),
581 brw_imm_ud((1U<<31)-1));
584 brw_writemask(tmp_ud
, BRW_WRITEMASK_X
),
589 brw_writemask(tmp
, BRW_WRITEMASK_X
),
590 retype(tmp_ud
, BRW_REGISTER_TYPE_D
), /* does it matter? */
594 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_YZ
) {
596 brw_writemask(tmp_ud
, BRW_WRITEMASK_Y
),
597 brw_swizzle1(arg0_ud
, 0),
598 brw_imm_ud((1<<23)-1));
601 brw_writemask(tmp_ud
, BRW_WRITEMASK_Y
),
603 brw_imm_ud(127<<23));
606 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_Z
) {
607 /* result[2] = result[0] + LOG2(result[1]); */
609 /* Why bother? The above is just a hint how to do this with a
610 * taylor series. Maybe we *should* use a taylor series as by
611 * the time all the above has been done it's almost certainly
612 * quicker than calling the mathbox, even with low precision.
615 * - result[0] + mathbox.LOG2(result[1])
616 * - mathbox.LOG2(arg0.x)
617 * - result[0] + inline_taylor_approx(result[1])
620 BRW_MATH_FUNCTION_LOG
,
621 brw_writemask(tmp
, BRW_WRITEMASK_Z
),
622 brw_swizzle1(tmp
, 1),
623 BRW_MATH_PRECISION_FULL
);
626 brw_writemask(tmp
, BRW_WRITEMASK_Z
),
627 brw_swizzle1(tmp
, 2),
628 brw_swizzle1(tmp
, 0));
631 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_W
) {
632 /* result[3] = 1.0; */
633 brw_MOV(p
, brw_writemask(tmp
, BRW_WRITEMASK_W
), brw_imm_f(1));
637 brw_MOV(p
, dst
, tmp
);
643 /* Need to unalias - consider swizzles: r0 = DST r0.xxxx r1
645 static void emit_dst_noalias( struct brw_vs_compile
*c
,
650 struct brw_compile
*p
= &c
->func
;
652 /* There must be a better way to do this:
654 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_X
)
655 brw_MOV(p
, brw_writemask(dst
, BRW_WRITEMASK_X
), brw_imm_f(1.0));
656 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_Y
)
657 brw_MUL(p
, brw_writemask(dst
, BRW_WRITEMASK_Y
), arg0
, arg1
);
658 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_Z
)
659 brw_MOV(p
, brw_writemask(dst
, BRW_WRITEMASK_Z
), arg0
);
660 if (dst
.dw1
.bits
.writemask
& BRW_WRITEMASK_W
)
661 brw_MOV(p
, brw_writemask(dst
, BRW_WRITEMASK_W
), arg1
);
665 static void emit_xpd( struct brw_compile
*p
,
670 brw_MUL(p
, brw_null_reg(), brw_swizzle(t
, 1,2,0,3), brw_swizzle(u
,2,0,1,3));
671 brw_MAC(p
, dst
, negate(brw_swizzle(t
, 2,0,1,3)), brw_swizzle(u
,1,2,0,3));
675 static void emit_lit_noalias( struct brw_vs_compile
*c
,
677 struct brw_reg arg0
)
679 struct brw_compile
*p
= &c
->func
;
680 struct brw_instruction
*if_insn
;
681 struct brw_reg tmp
= dst
;
682 GLboolean need_tmp
= (dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
687 brw_MOV(p
, brw_writemask(dst
, BRW_WRITEMASK_YZ
), brw_imm_f(0));
688 brw_MOV(p
, brw_writemask(dst
, BRW_WRITEMASK_XW
), brw_imm_f(1));
690 /* Need to use BRW_EXECUTE_8 and also do an 8-wide compare in order
691 * to get all channels active inside the IF. In the clipping code
692 * we run with NoMask, so it's not an option and we can use
693 * BRW_EXECUTE_1 for all comparisions.
695 brw_CMP(p
, brw_null_reg(), BRW_CONDITIONAL_G
, brw_swizzle1(arg0
,0), brw_imm_f(0));
696 if_insn
= brw_IF(p
, BRW_EXECUTE_8
);
698 brw_MOV(p
, brw_writemask(dst
, BRW_WRITEMASK_Y
), brw_swizzle1(arg0
,0));
700 brw_CMP(p
, brw_null_reg(), BRW_CONDITIONAL_G
, brw_swizzle1(arg0
,1), brw_imm_f(0));
701 brw_MOV(p
, brw_writemask(tmp
, BRW_WRITEMASK_Z
), brw_swizzle1(arg0
,1));
702 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
705 BRW_MATH_FUNCTION_POW
,
706 brw_writemask(dst
, BRW_WRITEMASK_Z
),
707 brw_swizzle1(tmp
, 2),
708 brw_swizzle1(arg0
, 3),
709 BRW_MATH_PRECISION_PARTIAL
);
712 brw_ENDIF(p
, if_insn
);
717 static void emit_lrp_noalias(struct brw_vs_compile
*c
,
723 struct brw_compile
*p
= &c
->func
;
725 brw_ADD(p
, dst
, negate(arg0
), brw_imm_f(1.0));
726 brw_MUL(p
, brw_null_reg(), dst
, arg2
);
727 brw_MAC(p
, dst
, arg0
, arg1
);
730 /** 3 or 4-component vector normalization */
731 static void emit_nrm( struct brw_vs_compile
*c
,
736 struct brw_compile
*p
= &c
->func
;
737 struct brw_reg tmp
= get_tmp(c
);
739 /* tmp = dot(arg0, arg0) */
741 brw_DP3(p
, tmp
, arg0
, arg0
);
743 brw_DP4(p
, tmp
, arg0
, arg0
);
745 /* tmp = 1 / sqrt(tmp) */
746 emit_math1(c
, BRW_MATH_FUNCTION_RSQ
, tmp
, tmp
, BRW_MATH_PRECISION_FULL
);
748 /* dst = arg0 * tmp */
749 brw_MUL(p
, dst
, arg0
, tmp
);
755 static struct brw_reg
756 get_constant(struct brw_vs_compile
*c
,
757 const struct ureg_instruction
*inst
,
760 const struct ureg_src src
= inst
->src
[argIndex
];
761 struct brw_compile
*p
= &c
->func
;
762 struct brw_reg const_reg
;
763 struct brw_reg const2_reg
;
764 const GLboolean relAddr
= src
.Indirect
;
766 assert(argIndex
< 3);
768 if (c
->current_const
[argIndex
].index
!= src
.Index
|| relAddr
) {
769 struct brw_reg addrReg
= c
->regs
[TGSI_FILE_ADDRESS
][0];
771 c
->current_const
[argIndex
].index
= src
.Index
;
774 printf(" fetch const[%d] for arg %d into reg %d\n",
775 src
.Index
, argIndex
, c
->current_const
[argIndex
].reg
.nr
);
777 /* need to fetch the constant now */
779 c
->current_const
[argIndex
].reg
,/* writeback dest */
781 relAddr
, /* relative indexing? */
782 addrReg
, /* address register */
783 16 * src
.Index
, /* byte offset */
784 SURF_INDEX_VERT_CONST_BUFFER
/* binding table index */
789 const2_reg
= get_tmp(c
);
791 /* use upper half of address reg for second read */
792 addrReg
= stride(addrReg
, 0, 4, 0);
796 const2_reg
, /* writeback dest */
798 relAddr
, /* relative indexing? */
799 addrReg
, /* address register */
800 16 * src
.Index
, /* byte offset */
801 SURF_INDEX_VERT_CONST_BUFFER
806 const_reg
= c
->current_const
[argIndex
].reg
;
809 /* merge the two Owords into the constant register */
810 /* const_reg[7..4] = const2_reg[7..4] */
812 suboffset(stride(const_reg
, 0, 4, 1), 4),
813 suboffset(stride(const2_reg
, 0, 4, 1), 4));
814 release_tmp(c
, const2_reg
);
817 /* replicate lower four floats into upper half (to get XYZWXYZW) */
818 const_reg
= stride(const_reg
, 0, 4, 0);
827 /* TODO: relative addressing!
829 static struct brw_reg
get_reg( struct brw_vs_compile
*c
,
830 enum tgsi_file_type file
,
834 case TGSI_FILE_TEMPORARY
:
835 case TGSI_FILE_INPUT
:
836 case TGSI_FILE_OUTPUT
:
837 case TGSI_FILE_CONSTANT
:
838 assert(c
->regs
[file
][index
].nr
!= 0);
839 return c
->regs
[file
][index
];
841 case TGSI_FILE_ADDRESS
:
843 return c
->regs
[file
][index
];
845 case TGSI_FILE_NULL
: /* undef values */
846 return brw_null_reg();
850 return brw_null_reg();
856 * Indirect addressing: get reg[[arg] + offset].
858 static struct brw_reg
deref( struct brw_vs_compile
*c
,
862 struct brw_compile
*p
= &c
->func
;
863 struct brw_reg tmp
= vec4(get_tmp(c
));
864 struct brw_reg addr_reg
= c
->regs
[TGSI_FILE_ADDRESS
][0];
865 struct brw_reg vp_address
= retype(vec1(addr_reg
), BRW_REGISTER_TYPE_UW
);
866 GLuint byte_offset
= arg
.nr
* 32 + arg
.subnr
+ offset
* 16;
867 struct brw_reg indirect
= brw_vec4_indirect(0,0);
870 brw_push_insn_state(p
);
871 brw_set_access_mode(p
, BRW_ALIGN_1
);
873 /* This is pretty clunky - load the address register twice and
874 * fetch each 4-dword value in turn. There must be a way to do
875 * this in a single pass, but I couldn't get it to work.
877 brw_ADD(p
, brw_address_reg(0), vp_address
, brw_imm_d(byte_offset
));
878 brw_MOV(p
, tmp
, indirect
);
880 brw_ADD(p
, brw_address_reg(0), suboffset(vp_address
, 8), brw_imm_d(byte_offset
));
881 brw_MOV(p
, suboffset(tmp
, 4), indirect
);
883 brw_pop_insn_state(p
);
886 /* NOTE: tmp not released */
892 * Get brw reg corresponding to the instruction's [argIndex] src reg.
893 * TODO: relative addressing!
895 static struct brw_reg
896 get_src_reg( struct brw_vs_compile
*c
,
897 const struct ureg_instruction
*inst
,
900 const GLuint file
= inst
->src
[argIndex
].File
;
901 const GLint index
= inst
->src
[argIndex
].Index
;
902 const GLboolean relAddr
= inst
->src
[argIndex
].Indirect
;
905 case TGSI_FILE_TEMPORARY
:
906 case TGSI_FILE_INPUT
:
907 case TGSI_FILE_OUTPUT
:
909 return deref(c
, c
->regs
[file
][0], index
);
912 assert(c
->regs
[file
][index
].nr
!= 0);
913 return c
->regs
[file
][index
];
916 case TGSI_FILE_CONSTANT
:
917 if (c
->vp
->use_const_buffer
) {
918 return get_constant(c
, inst
, argIndex
);
921 return deref(c
, c
->regs
[TGSI_FILE_CONSTANT
][0], index
);
924 assert(c
->regs
[TGSI_FILE_CONSTANT
][index
].nr
!= 0);
925 return c
->regs
[TGSI_FILE_CONSTANT
][index
];
927 case TGSI_FILE_ADDRESS
:
929 return c
->regs
[file
][index
];
932 /* this is a normal case since we loop over all three src args */
933 return brw_null_reg();
937 return brw_null_reg();
942 static void emit_arl( struct brw_vs_compile
*c
,
944 struct brw_reg arg0
)
946 struct brw_compile
*p
= &c
->func
;
947 struct brw_reg tmp
= dst
;
948 GLboolean need_tmp
= (dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
953 brw_RNDD(p
, tmp
, arg0
); /* tmp = round(arg0) */
954 brw_MUL(p
, dst
, tmp
, brw_imm_d(16)); /* dst = tmp * 16 */
962 * Return the brw reg for the given instruction's src argument.
964 static struct brw_reg
get_arg( struct brw_vs_compile
*c
,
965 const struct ureg_instruction
*inst
,
968 const struct ureg_src src
= inst
->src
[argIndex
];
971 if (src
.File
== TGSI_FILE_NULL
)
972 return brw_null_reg();
974 reg
= get_src_reg(c
, inst
, argIndex
);
976 /* Convert 3-bit swizzle to 2-bit.
978 reg
.dw1
.bits
.swizzle
= BRW_SWIZZLE4(src
.SwizzleX
,
983 /* Note this is ok for non-swizzle instructions:
985 reg
.negate
= src
.Negate
? 1 : 0;
992 * Get brw register for the given program dest register.
994 static struct brw_reg
get_dst( struct brw_vs_compile
*c
,
995 struct ureg_dst dst
)
1000 case TGSI_FILE_TEMPORARY
:
1001 case TGSI_FILE_OUTPUT
:
1002 assert(c
->regs
[dst
.File
][dst
.Index
].nr
!= 0);
1003 reg
= c
->regs
[dst
.File
][dst
.Index
];
1005 case TGSI_FILE_ADDRESS
:
1006 assert(dst
.Index
== 0);
1007 reg
= c
->regs
[dst
.File
][dst
.Index
];
1009 case TGSI_FILE_NULL
:
1010 /* we may hit this for OPCODE_END, OPCODE_KIL, etc */
1011 reg
= brw_null_reg();
1015 reg
= brw_null_reg();
1018 reg
.dw1
.bits
.writemask
= dst
.WriteMask
;
1027 * Post-vertex-program processing. Send the results to the URB.
1029 static void emit_vertex_write( struct brw_vs_compile
*c
)
1031 struct brw_compile
*p
= &c
->func
;
1032 struct brw_reg m0
= brw_message_reg(0);
1033 struct brw_reg pos
= c
->regs
[TGSI_FILE_OUTPUT
][VERT_RESULT_HPOS
];
1036 GLuint len_vertext_header
= 2;
1038 if (c
->key
.copy_edgeflag
) {
1041 get_reg(c
, TGSI_FILE_OUTPUT
, 0),
1042 get_reg(c
, TGSI_FILE_INPUT
, 0));
1045 /* Build ndc coords */
1047 /* ndc = 1.0 / pos.w */
1048 emit_math1(c
, BRW_MATH_FUNCTION_INV
, ndc
, brw_swizzle1(pos
, 3), BRW_MATH_PRECISION_FULL
);
1049 /* ndc.xyz = pos * ndc */
1050 brw_MUL(p
, brw_writemask(ndc
, BRW_WRITEMASK_XYZ
), pos
, ndc
);
1052 /* Update the header for point size, user clipping flags, and -ve rhw
1055 if (c
->prog_data
.writes_psiz
||
1056 c
->key
.nr_userclip
||
1059 struct brw_reg header1
= retype(get_tmp(c
), BRW_REGISTER_TYPE_UD
);
1062 brw_MOV(p
, header1
, brw_imm_ud(0));
1064 brw_set_access_mode(p
, BRW_ALIGN_16
);
1066 if (c
->prog_data
.writes_psiz
) {
1067 struct brw_reg psiz
= c
->regs
[TGSI_FILE_OUTPUT
][VERT_RESULT_PSIZ
];
1068 brw_MUL(p
, brw_writemask(header1
, BRW_WRITEMASK_W
), brw_swizzle1(psiz
, 0), brw_imm_f(1<<11));
1069 brw_AND(p
, brw_writemask(header1
, BRW_WRITEMASK_W
), header1
, brw_imm_ud(0x7ff<<8));
1072 for (i
= 0; i
< c
->key
.nr_userclip
; i
++) {
1073 brw_set_conditionalmod(p
, BRW_CONDITIONAL_L
);
1074 brw_DP4(p
, brw_null_reg(), pos
, c
->userplane
[i
]);
1075 brw_OR(p
, brw_writemask(header1
, BRW_WRITEMASK_W
), header1
, brw_imm_ud(1<<i
));
1076 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1079 /* i965 clipping workaround:
1080 * 1) Test for -ve rhw
1082 * set ndc = (0,0,0,0)
1085 * Later, clipping will detect ucp[6] and ensure the primitive is
1086 * clipped against all fixed planes.
1088 if (c
->chipset
.is_965
) {
1090 vec8(brw_null_reg()),
1092 brw_swizzle1(ndc
, 3),
1095 brw_OR(p
, brw_writemask(header1
, BRW_WRITEMASK_W
), header1
, brw_imm_ud(1<<6));
1096 brw_MOV(p
, ndc
, brw_imm_f(0));
1097 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1100 brw_set_access_mode(p
, BRW_ALIGN_1
); /* why? */
1101 brw_MOV(p
, retype(brw_message_reg(1), BRW_REGISTER_TYPE_UD
), header1
);
1102 brw_set_access_mode(p
, BRW_ALIGN_16
);
1104 release_tmp(c
, header1
);
1107 brw_MOV(p
, retype(brw_message_reg(1), BRW_REGISTER_TYPE_UD
), brw_imm_ud(0));
1110 /* Emit the (interleaved) headers for the two vertices - an 8-reg
1111 * of zeros followed by two sets of NDC coordinates:
1113 brw_set_access_mode(p
, BRW_ALIGN_1
);
1114 brw_MOV(p
, offset(m0
, 2), ndc
);
1116 if (c
->chipset
.is_igdng
) {
1117 /* There are 20 DWs (D0-D19) in VUE vertex header on IGDNG */
1118 brw_MOV(p
, offset(m0
, 3), pos
); /* a portion of vertex header */
1119 /* m4, m5 contain the distances from vertex to the user clip planeXXX.
1120 * Seems it is useless for us.
1121 * m6 is used for aligning, so that the remainder of vertex element is
1124 brw_MOV(p
, offset(m0
, 7), pos
); /* the remainder of vertex element */
1125 len_vertext_header
= 6;
1127 brw_MOV(p
, offset(m0
, 3), pos
);
1128 len_vertext_header
= 2;
1131 eot
= (c
->first_overflow_output
== 0);
1134 brw_null_reg(), /* dest */
1135 0, /* starting mrf reg nr */
1139 MIN2(c
->nr_outputs
+ 1 + len_vertext_header
, (BRW_MAX_MRF
-1)), /* msg len */
1140 0, /* response len */
1142 eot
, /* writes complete */
1143 0, /* urb destination offset */
1144 BRW_URB_SWIZZLE_INTERLEAVE
);
1146 if (c
->first_overflow_output
> 0) {
1147 /* Not all of the vertex outputs/results fit into the MRF.
1148 * Move the overflowed attributes from the GRF to the MRF and
1149 * issue another brw_urb_WRITE().
1151 /* XXX I'm not 100% sure about which MRF regs to use here. Starting
1155 for (i
= c
->first_overflow_output
; i
< c
->prog_data
.nr_outputs
; i
++) {
1156 /* move from GRF to MRF */
1157 brw_MOV(p
, brw_message_reg(4+mrf
), c
->regs
[TGSI_FILE_OUTPUT
][i
]);
1162 brw_null_reg(), /* dest */
1163 4, /* starting mrf reg nr */
1167 mrf
+1, /* msg len */
1168 0, /* response len */
1170 1, /* writes complete */
1171 BRW_MAX_MRF
-1, /* urb destination offset */
1172 BRW_URB_SWIZZLE_INTERLEAVE
);
1178 * Called after code generation to resolve subroutine calls and the
1180 * \param end_inst points to brw code for END instruction
1181 * \param last_inst points to last instruction emitted before vertex write
1184 post_vs_emit( struct brw_vs_compile
*c
,
1185 struct brw_instruction
*end_inst
,
1186 struct brw_instruction
*last_inst
)
1190 brw_resolve_cals(&c
->func
);
1192 /* patch up the END code to jump past subroutines, etc */
1193 offset
= last_inst
- end_inst
;
1195 brw_set_src1(end_inst
, brw_imm_d(offset
* 16));
1197 end_inst
->header
.opcode
= BRW_OPCODE_NOP
;
1202 get_predicate(const struct ureg_instruction
*inst
)
1204 /* XXX: disabling for now
1207 if (inst
->dst
.CondMask
== COND_TR
)
1208 return BRW_PREDICATE_NONE
;
1210 /* All of GLSL only produces predicates for COND_NE and one channel per
1211 * vector. Fail badly if someone starts doing something else, as it might
1212 * mean infinite looping or something.
1214 * We'd like to support all the condition codes, but our hardware doesn't
1215 * quite match the Mesa IR, which is modeled after the NV extensions. For
1216 * those, the instruction may update the condition codes or not, then any
1217 * later instruction may use one of those condition codes. For gen4, the
1218 * instruction may update the flags register based on one of the condition
1219 * codes output by the instruction, and then further instructions may
1220 * predicate on that. We can probably support this, but it won't
1221 * necessarily be easy.
1223 /* assert(inst->dst.CondMask == COND_NE); */
1225 switch (inst
->dst
.CondSwizzle
) {
1227 return BRW_PREDICATE_ALIGN16_REPLICATE_X
;
1229 return BRW_PREDICATE_ALIGN16_REPLICATE_Y
;
1231 return BRW_PREDICATE_ALIGN16_REPLICATE_Z
;
1233 return BRW_PREDICATE_ALIGN16_REPLICATE_W
;
1235 debug_printf("Unexpected predicate: 0x%08x\n",
1236 inst
->dst
.CondMask
);
1237 return BRW_PREDICATE_NORMAL
;
1240 return BRW_PREDICATE_NORMAL
;
1244 static void emit_insn(struct brw_vs_compile
*c
,
1245 const struct ureg_instruction
*inst
)
1247 struct brw_compile
*p
= &c
->func
;
1248 struct brw_reg args
[3], dst
;
1252 printf("%d: ", insn
);
1253 _mesa_print_instruction(inst
);
1256 /* Get argument regs.
1258 for (i
= 0; i
< 3; i
++) {
1259 args
[i
] = get_arg(c
, inst
, i
);
1262 /* Get dest regs. Note that it is possible for a reg to be both
1263 * dst and arg, given the static allocation of registers. So
1264 * care needs to be taken emitting multi-operation instructions.
1266 dst
= get_dst(c
, inst
->dst
);
1268 if (inst
->dst
.Saturate
) {
1269 debug_printf("Unsupported saturate in vertex shader");
1272 switch (inst
->opcode
) {
1273 case TGSI_OPCODE_ABS
:
1274 brw_MOV(p
, dst
, brw_abs(args
[0]));
1276 case TGSI_OPCODE_ADD
:
1277 brw_ADD(p
, dst
, args
[0], args
[1]);
1279 case TGSI_OPCODE_COS
:
1280 emit_math1(c
, BRW_MATH_FUNCTION_COS
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1282 case TGSI_OPCODE_DP3
:
1283 brw_DP3(p
, dst
, args
[0], args
[1]);
1285 case TGSI_OPCODE_DP4
:
1286 brw_DP4(p
, dst
, args
[0], args
[1]);
1288 case TGSI_OPCODE_DPH
:
1289 brw_DPH(p
, dst
, args
[0], args
[1]);
1291 case TGSI_OPCODE_NRM
:
1292 emit_nrm(c
, dst
, args
[0], 3);
1294 case TGSI_OPCODE_NRM4
:
1295 emit_nrm(c
, dst
, args
[0], 4);
1297 case TGSI_OPCODE_DST
:
1298 unalias2(c
, dst
, args
[0], args
[1], emit_dst_noalias
);
1300 case TGSI_OPCODE_EXP
:
1301 unalias1(c
, dst
, args
[0], emit_exp_noalias
);
1303 case TGSI_OPCODE_EX2
:
1304 emit_math1(c
, BRW_MATH_FUNCTION_EXP
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1306 case TGSI_OPCODE_ARL
:
1307 emit_arl(c
, dst
, args
[0]);
1309 case TGSI_OPCODE_FLR
:
1310 brw_RNDD(p
, dst
, args
[0]);
1312 case TGSI_OPCODE_FRC
:
1313 brw_FRC(p
, dst
, args
[0]);
1315 case TGSI_OPCODE_LOG
:
1316 unalias1(c
, dst
, args
[0], emit_log_noalias
);
1318 case TGSI_OPCODE_LG2
:
1319 emit_math1(c
, BRW_MATH_FUNCTION_LOG
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1321 case TGSI_OPCODE_LIT
:
1322 unalias1(c
, dst
, args
[0], emit_lit_noalias
);
1324 case TGSI_OPCODE_LRP
:
1325 unalias3(c
, dst
, args
[0], args
[1], args
[2], emit_lrp_noalias
);
1327 case TGSI_OPCODE_MAD
:
1328 brw_MOV(p
, brw_acc_reg(), args
[2]);
1329 brw_MAC(p
, dst
, args
[0], args
[1]);
1331 case TGSI_OPCODE_MAX
:
1332 emit_max(p
, dst
, args
[0], args
[1]);
1334 case TGSI_OPCODE_MIN
:
1335 emit_min(p
, dst
, args
[0], args
[1]);
1337 case TGSI_OPCODE_MOV
:
1338 brw_MOV(p
, dst
, args
[0]);
1340 case TGSI_OPCODE_MUL
:
1341 brw_MUL(p
, dst
, args
[0], args
[1]);
1343 case TGSI_OPCODE_POW
:
1344 emit_math2(c
, BRW_MATH_FUNCTION_POW
, dst
, args
[0], args
[1], BRW_MATH_PRECISION_FULL
);
1346 case TGSI_OPCODE_RCP
:
1347 emit_math1(c
, BRW_MATH_FUNCTION_INV
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1349 case TGSI_OPCODE_RSQ
:
1350 emit_math1(c
, BRW_MATH_FUNCTION_RSQ
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1352 case TGSI_OPCODE_SEQ
:
1353 emit_seq(p
, dst
, args
[0], args
[1]);
1355 case TGSI_OPCODE_SIN
:
1356 emit_math1(c
, BRW_MATH_FUNCTION_SIN
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1358 case TGSI_OPCODE_SNE
:
1359 emit_sne(p
, dst
, args
[0], args
[1]);
1361 case TGSI_OPCODE_SGE
:
1362 emit_sge(p
, dst
, args
[0], args
[1]);
1364 case TGSI_OPCODE_SGT
:
1365 emit_sgt(p
, dst
, args
[0], args
[1]);
1367 case TGSI_OPCODE_SLT
:
1368 emit_slt(p
, dst
, args
[0], args
[1]);
1370 case TGSI_OPCODE_SLE
:
1371 emit_sle(p
, dst
, args
[0], args
[1]);
1373 case TGSI_OPCODE_SUB
:
1374 brw_ADD(p
, dst
, args
[0], negate(args
[1]));
1376 case TGSI_OPCODE_TRUNC
:
1377 /* round toward zero */
1378 brw_RNDZ(p
, dst
, args
[0]);
1380 case TGSI_OPCODE_XPD
:
1381 emit_xpd(p
, dst
, args
[0], args
[1]);
1383 case TGSI_OPCODE_IF
:
1384 assert(c
->if_depth
< MAX_IF_DEPTH
);
1385 c
->if_inst
[c
->if_depth
] = brw_IF(p
, BRW_EXECUTE_8
);
1386 /* Note that brw_IF smashes the predicate_control field. */
1387 c
->if_inst
[c
->if_depth
]->header
.predicate_control
= get_predicate(inst
);
1390 case TGSI_OPCODE_ELSE
:
1391 c
->if_inst
[c
->if_depth
-1] = brw_ELSE(p
, c
->if_inst
[c
->if_depth
-1]);
1393 case TGSI_OPCODE_ENDIF
:
1394 assert(c
->if_depth
> 0);
1395 brw_ENDIF(p
, c
->if_inst
[--c
->if_depth
]);
1397 case TGSI_OPCODE_BGNLOOP
:
1398 c
->loop_inst
[c
->loop_depth
++] = brw_DO(p
, BRW_EXECUTE_8
);
1400 case TGSI_OPCODE_BRK
:
1401 brw_set_predicate_control(p
, get_predicate(inst
));
1403 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1405 case TGSI_OPCODE_CONT
:
1406 brw_set_predicate_control(p
, get_predicate(inst
));
1408 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1410 case TGSI_OPCODE_ENDLOOP
:
1412 struct brw_instruction
*inst0
, *inst1
;
1417 if (c
->chipset
.is_igdng
)
1420 inst0
= inst1
= brw_WHILE(p
, c
->loop_inst
[c
->loop_depth
]);
1421 /* patch all the BREAK/CONT instructions from last BEGINLOOP */
1422 while (inst0
> c
->loop_inst
[c
->loop_depth
]) {
1424 if (inst0
->header
.opcode
== TGSI_OPCODE_BRK
) {
1425 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
+ 1);
1426 inst0
->bits3
.if_else
.pop_count
= 0;
1428 else if (inst0
->header
.opcode
== TGSI_OPCODE_CONT
) {
1429 inst0
->bits3
.if_else
.jump_count
= br
* (inst1
- inst0
);
1430 inst0
->bits3
.if_else
.pop_count
= 0;
1435 case TGSI_OPCODE_BRA
:
1436 brw_set_predicate_control(p
, get_predicate(inst
));
1437 brw_ADD(p
, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16));
1438 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1440 case TGSI_OPCODE_CAL
:
1441 brw_set_access_mode(p
, BRW_ALIGN_1
);
1442 brw_ADD(p
, deref_1d(c
->stack_index
, 0), brw_ip_reg(), brw_imm_d(3*16));
1443 brw_set_access_mode(p
, BRW_ALIGN_16
);
1444 brw_ADD(p
, get_addr_reg(c
->stack_index
),
1445 get_addr_reg(c
->stack_index
), brw_imm_d(4));
1446 brw_save_call(p
, inst
->label
, p
->nr_insn
);
1447 brw_ADD(p
, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16));
1449 case TGSI_OPCODE_RET
:
1450 brw_ADD(p
, get_addr_reg(c
->stack_index
),
1451 get_addr_reg(c
->stack_index
), brw_imm_d(-4));
1452 brw_set_access_mode(p
, BRW_ALIGN_1
);
1453 brw_MOV(p
, brw_ip_reg(), deref_1d(c
->stack_index
, 0));
1454 brw_set_access_mode(p
, BRW_ALIGN_16
);
1456 case TGSI_OPCODE_END
:
1457 c
->end_offset
= p
->nr_insn
;
1458 /* this instruction will get patched later to jump past subroutine
1461 brw_ADD(p
, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16));
1463 case TGSI_OPCODE_BGNSUB
:
1464 brw_save_label(p
, p
->nr_insn
, p
->nr_insn
);
1466 case TGSI_OPCODE_ENDSUB
:
1470 debug_printf("Unsupported opcode %i (%s) in vertex shader",
1472 tgsi_get_opcode_name(inst
->opcode
));
1475 /* Set the predication update on the last instruction of the native
1476 * instruction sequence.
1478 * This would be problematic if it was set on a math instruction,
1479 * but that shouldn't be the case with the current GLSL compiler.
1484 if (inst
->CondUpdate
) {
1485 struct brw_instruction
*hw_insn
= &p
->store
[p
->nr_insn
- 1];
1487 assert(hw_insn
->header
.destreg__conditionalmod
== 0);
1488 hw_insn
->header
.destreg__conditionalmod
= BRW_CONDITIONAL_NZ
;
1496 /* Emit the vertex program instructions here.
1498 void brw_vs_emit(struct brw_vs_compile
*c
)
1500 struct brw_compile
*p
= &c
->func
;
1501 struct brw_instruction
*end_inst
, *last_inst
;
1502 struct ureg_parse_context parse
;
1503 struct ureg_declaration
*decl
;
1504 struct ureg_declaration
*imm
;
1505 struct ureg_declaration
*insn
;
1507 if (BRW_DEBUG
& DEBUG_VS
)
1508 tgsi_dump(c
->vp
->tokens
, 0);
1510 c
->stack_index
= brw_indirect(0, 0);
1512 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1513 brw_set_access_mode(p
, BRW_ALIGN_16
);
1515 /* Static register allocation
1517 brw_vs_alloc_regs(c
);
1518 brw_MOV(p
, get_addr_reg(c
->stack_index
), brw_address(c
->stack
));
1520 while (ureg_next_decl(&parse
, &decl
)) {
1523 while (ureg_next_immediate(&parse
, &imm
)) {
1526 while (ureg_next_instruction(&parse
, &insn
)) {
1529 end_inst
= &p
->store
[end_offset
];
1530 last_inst
= &p
->store
[p
->nr_insn
];
1532 /* The END instruction will be patched to jump to this code */
1533 emit_vertex_write(c
);
1535 post_vs_emit(c
, end_inst
, last_inst
);
1537 if (BRW_DEBUG
& DEBUG_VS
) {
1540 debug_printf("vs-native:\n");
1541 for (i
= 0; i
< p
->nr_insn
; i
++)
1542 brw_disasm(stderr
, &p
->store
[i
]);