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>
33 #include "main/macros.h"
34 #include "shader/program.h"
35 #include "shader/prog_parameter.h"
36 #include "shader/prog_print.h"
37 #include "brw_context.h"
41 static struct brw_reg
get_tmp( struct brw_vs_compile
*c
)
43 struct brw_reg tmp
= brw_vec8_grf(c
->last_tmp
, 0);
45 if (++c
->last_tmp
> c
->prog_data
.total_grf
)
46 c
->prog_data
.total_grf
= c
->last_tmp
;
51 static void release_tmp( struct brw_vs_compile
*c
, struct brw_reg tmp
)
53 if (tmp
.nr
== c
->last_tmp
-1)
57 static void release_tmps( struct brw_vs_compile
*c
)
59 c
->last_tmp
= c
->first_tmp
;
64 * Preallocate GRF register before code emit.
65 * Do things as simply as possible. Allocate and populate all regs
68 static void brw_vs_alloc_regs( struct brw_vs_compile
*c
)
70 GLuint i
, reg
= 0, mrf
;
74 if (c
->vp
->program
.Base
.Parameters
->NumParameters
>= 6)
75 c
->use_const_buffer
= 1;
78 c
->use_const_buffer
= GL_FALSE
;
79 /*printf("use_const_buffer = %d\n", c->use_const_buffer);*/
81 /* r0 -- reserved as usual
83 c
->r0
= brw_vec8_grf(reg
, 0);
86 /* User clip planes from curbe:
88 if (c
->key
.nr_userclip
) {
89 for (i
= 0; i
< c
->key
.nr_userclip
; i
++) {
90 c
->userplane
[i
] = stride( brw_vec4_grf(reg
+3+i
/2, (i
%2) * 4), 0, 4, 1);
93 /* Deal with curbe alignment:
95 reg
+= ((6 + c
->key
.nr_userclip
+ 3) / 4) * 2;
98 /* Vertex program parameters from curbe:
100 if (c
->use_const_buffer
) {
101 /* get constants from a real constant buffer */
102 c
->prog_data
.curb_read_length
= 0;
105 /* use a section of the GRF for constants */
106 nr_params
= c
->vp
->program
.Base
.Parameters
->NumParameters
;
107 for (i
= 0; i
< nr_params
; i
++) {
108 c
->regs
[PROGRAM_STATE_VAR
][i
] = stride( brw_vec4_grf(reg
+i
/2, (i
%2) * 4), 0, 4, 1);
110 reg
+= (nr_params
+ 1) / 2;
111 c
->prog_data
.curb_read_length
= reg
- 1;
114 /* Allocate input regs:
117 for (i
= 0; i
< VERT_ATTRIB_MAX
; i
++) {
118 if (c
->prog_data
.inputs_read
& (1 << i
)) {
120 c
->regs
[PROGRAM_INPUT
][i
] = brw_vec8_grf(reg
, 0);
125 /* Allocate outputs: TODO: could organize the non-position outputs
126 * to go straight into message regs.
129 c
->first_output
= reg
;
131 for (i
= 0; i
< VERT_RESULT_MAX
; i
++) {
132 if (c
->prog_data
.outputs_written
& (1 << i
)) {
134 if (i
== VERT_RESULT_HPOS
) {
135 c
->regs
[PROGRAM_OUTPUT
][i
] = brw_vec8_grf(reg
, 0);
138 else if (i
== VERT_RESULT_PSIZ
) {
139 c
->regs
[PROGRAM_OUTPUT
][i
] = brw_vec8_grf(reg
, 0);
141 mrf
++; /* just a placeholder? XXX fix later stages & remove this */
144 c
->regs
[PROGRAM_OUTPUT
][i
] = brw_message_reg(mrf
);
150 /* Allocate program temporaries:
152 for (i
= 0; i
< c
->vp
->program
.Base
.NumTemporaries
; i
++) {
153 c
->regs
[PROGRAM_TEMPORARY
][i
] = brw_vec8_grf(reg
, 0);
157 /* Address reg(s). Don't try to use the internal address reg until
160 for (i
= 0; i
< c
->vp
->program
.Base
.NumAddressRegs
; i
++) {
161 c
->regs
[PROGRAM_ADDRESS
][i
] = brw_reg(BRW_GENERAL_REGISTER_FILE
,
165 BRW_VERTICAL_STRIDE_8
,
167 BRW_HORIZONTAL_STRIDE_1
,
173 for (i
= 0; i
< 128; i
++) {
174 if (c
->output_regs
[i
].used_in_src
) {
175 c
->output_regs
[i
].reg
= brw_vec8_grf(reg
, 0);
180 c
->stack
= brw_uw16_reg(BRW_GENERAL_REGISTER_FILE
, reg
, 0);
183 /* Some opcodes need an internal temporary:
186 c
->last_tmp
= reg
; /* for allocation purposes */
188 /* Each input reg holds data from two vertices. The
189 * urb_read_length is the number of registers read from *each*
190 * vertex urb, so is half the amount:
192 c
->prog_data
.urb_read_length
= (c
->nr_inputs
+ 1) / 2;
194 c
->prog_data
.urb_entry_size
= (c
->nr_outputs
+ 2 + 3) / 4;
195 c
->prog_data
.total_grf
= reg
;
197 if (c
->use_const_buffer
) {
198 for (i
= 0; i
< 3; i
++) {
199 c
->current_const
[i
].index
= -1;
200 c
->current_const
[i
].reg
= get_tmp(c
);
204 if (INTEL_DEBUG
& DEBUG_VS
) {
205 _mesa_printf("%s NumAddrRegs %d\n", __FUNCTION__
, c
->vp
->program
.Base
.NumAddressRegs
);
206 _mesa_printf("%s NumTemps %d\n", __FUNCTION__
, c
->vp
->program
.Base
.NumTemporaries
);
207 _mesa_printf("%s reg = %d\n", __FUNCTION__
, reg
);
213 * If an instruction uses a temp reg both as a src and the dest, we
214 * sometimes need to allocate an intermediate temporary.
216 static void unalias1( struct brw_vs_compile
*c
,
219 void (*func
)( struct brw_vs_compile
*,
223 if (dst
.file
== arg0
.file
&& dst
.nr
== arg0
.nr
) {
224 struct brw_compile
*p
= &c
->func
;
225 struct brw_reg tmp
= brw_writemask(get_tmp(c
), dst
.dw1
.bits
.writemask
);
227 brw_MOV(p
, dst
, tmp
);
237 * Checkes if 2-operand instruction needs an intermediate temporary.
239 static void unalias2( struct brw_vs_compile
*c
,
243 void (*func
)( struct brw_vs_compile
*,
248 if ((dst
.file
== arg0
.file
&& dst
.nr
== arg0
.nr
) ||
249 (dst
.file
== arg1
.file
&& dst
.nr
== arg1
.nr
)) {
250 struct brw_compile
*p
= &c
->func
;
251 struct brw_reg tmp
= brw_writemask(get_tmp(c
), dst
.dw1
.bits
.writemask
);
252 func(c
, tmp
, arg0
, arg1
);
253 brw_MOV(p
, dst
, tmp
);
257 func(c
, dst
, arg0
, arg1
);
263 * Checkes if 3-operand instruction needs an intermediate temporary.
265 static void unalias3( struct brw_vs_compile
*c
,
270 void (*func
)( struct brw_vs_compile
*,
276 if ((dst
.file
== arg0
.file
&& dst
.nr
== arg0
.nr
) ||
277 (dst
.file
== arg1
.file
&& dst
.nr
== arg1
.nr
) ||
278 (dst
.file
== arg2
.file
&& dst
.nr
== arg2
.nr
)) {
279 struct brw_compile
*p
= &c
->func
;
280 struct brw_reg tmp
= brw_writemask(get_tmp(c
), dst
.dw1
.bits
.writemask
);
281 func(c
, tmp
, arg0
, arg1
, arg2
);
282 brw_MOV(p
, dst
, tmp
);
286 func(c
, dst
, arg0
, arg1
, arg2
);
290 static void emit_sop( struct brw_compile
*p
,
296 brw_MOV(p
, dst
, brw_imm_f(0.0f
));
297 brw_CMP(p
, brw_null_reg(), cond
, arg0
, arg1
);
298 brw_MOV(p
, dst
, brw_imm_f(1.0f
));
299 brw_set_predicate_control_flag_value(p
, 0xff);
302 static void emit_seq( struct brw_compile
*p
,
305 struct brw_reg arg1
)
307 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_EQ
);
310 static void emit_sne( struct brw_compile
*p
,
313 struct brw_reg arg1
)
315 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_NEQ
);
317 static void emit_slt( struct brw_compile
*p
,
320 struct brw_reg arg1
)
322 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_L
);
325 static void emit_sle( struct brw_compile
*p
,
328 struct brw_reg arg1
)
330 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_LE
);
333 static void emit_sgt( struct brw_compile
*p
,
336 struct brw_reg arg1
)
338 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_G
);
341 static void emit_sge( struct brw_compile
*p
,
344 struct brw_reg arg1
)
346 emit_sop(p
, dst
, arg0
, arg1
, BRW_CONDITIONAL_GE
);
349 static void emit_max( struct brw_compile
*p
,
352 struct brw_reg arg1
)
354 brw_CMP(p
, brw_null_reg(), BRW_CONDITIONAL_L
, arg0
, arg1
);
355 brw_SEL(p
, dst
, arg1
, arg0
);
356 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
359 static void emit_min( struct brw_compile
*p
,
362 struct brw_reg arg1
)
364 brw_CMP(p
, brw_null_reg(), BRW_CONDITIONAL_L
, arg0
, arg1
);
365 brw_SEL(p
, dst
, arg0
, arg1
);
366 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
370 static void emit_math1( struct brw_vs_compile
*c
,
376 /* There are various odd behaviours with SEND on the simulator. In
377 * addition there are documented issues with the fact that the GEN4
378 * processor doesn't do dependency control properly on SEND
379 * results. So, on balance, this kludge to get around failures
380 * with writemasked math results looks like it might be necessary
381 * whether that turns out to be a simulator bug or not:
383 struct brw_compile
*p
= &c
->func
;
384 struct brw_reg tmp
= dst
;
385 GLboolean need_tmp
= (dst
.dw1
.bits
.writemask
!= 0xf ||
386 dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
394 BRW_MATH_SATURATE_NONE
,
397 BRW_MATH_DATA_SCALAR
,
401 brw_MOV(p
, dst
, tmp
);
407 static void emit_math2( struct brw_vs_compile
*c
,
414 struct brw_compile
*p
= &c
->func
;
415 struct brw_reg tmp
= dst
;
416 GLboolean need_tmp
= (dst
.dw1
.bits
.writemask
!= 0xf ||
417 dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
422 brw_MOV(p
, brw_message_reg(3), arg1
);
427 BRW_MATH_SATURATE_NONE
,
430 BRW_MATH_DATA_SCALAR
,
434 brw_MOV(p
, dst
, tmp
);
440 static void emit_exp_noalias( struct brw_vs_compile
*c
,
442 struct brw_reg arg0
)
444 struct brw_compile
*p
= &c
->func
;
447 if (dst
.dw1
.bits
.writemask
& WRITEMASK_X
) {
448 struct brw_reg tmp
= get_tmp(c
);
449 struct brw_reg tmp_d
= retype(tmp
, BRW_REGISTER_TYPE_D
);
451 /* tmp_d = floor(arg0.x) */
452 brw_RNDD(p
, tmp_d
, brw_swizzle1(arg0
, 0));
454 /* result[0] = 2.0 ^ tmp */
456 /* Adjust exponent for floating point:
459 brw_ADD(p
, brw_writemask(tmp_d
, WRITEMASK_X
), tmp_d
, brw_imm_d(127));
461 /* Install exponent and sign.
462 * Excess drops off the edge:
464 brw_SHL(p
, brw_writemask(retype(dst
, BRW_REGISTER_TYPE_D
), WRITEMASK_X
),
465 tmp_d
, brw_imm_d(23));
470 if (dst
.dw1
.bits
.writemask
& WRITEMASK_Y
) {
471 /* result[1] = arg0.x - floor(arg0.x) */
472 brw_FRC(p
, brw_writemask(dst
, WRITEMASK_Y
), brw_swizzle1(arg0
, 0));
475 if (dst
.dw1
.bits
.writemask
& WRITEMASK_Z
) {
476 /* As with the LOG instruction, we might be better off just
477 * doing a taylor expansion here, seeing as we have to do all
480 * If mathbox partial precision is too low, consider also:
481 * result[3] = result[0] * EXP(result[1])
484 BRW_MATH_FUNCTION_EXP
,
485 brw_writemask(dst
, WRITEMASK_Z
),
486 brw_swizzle1(arg0
, 0),
487 BRW_MATH_PRECISION_FULL
);
490 if (dst
.dw1
.bits
.writemask
& WRITEMASK_W
) {
491 /* result[3] = 1.0; */
492 brw_MOV(p
, brw_writemask(dst
, WRITEMASK_W
), brw_imm_f(1));
497 static void emit_log_noalias( struct brw_vs_compile
*c
,
499 struct brw_reg arg0
)
501 struct brw_compile
*p
= &c
->func
;
502 struct brw_reg tmp
= dst
;
503 struct brw_reg tmp_ud
= retype(tmp
, BRW_REGISTER_TYPE_UD
);
504 struct brw_reg arg0_ud
= retype(arg0
, BRW_REGISTER_TYPE_UD
);
505 GLboolean need_tmp
= (dst
.dw1
.bits
.writemask
!= 0xf ||
506 dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
510 tmp_ud
= retype(tmp
, BRW_REGISTER_TYPE_UD
);
513 /* Perform mant = frexpf(fabsf(x), &exp), adjust exp and mnt
516 * These almost look likey they could be joined up, but not really
519 * result[0].f = (x.i & ((1<<31)-1) >> 23) - 127
520 * result[1].i = (x.i & ((1<<23)-1) + (127<<23)
522 if (dst
.dw1
.bits
.writemask
& WRITEMASK_XZ
) {
524 brw_writemask(tmp_ud
, WRITEMASK_X
),
525 brw_swizzle1(arg0_ud
, 0),
526 brw_imm_ud((1U<<31)-1));
529 brw_writemask(tmp_ud
, WRITEMASK_X
),
534 brw_writemask(tmp
, WRITEMASK_X
),
535 retype(tmp_ud
, BRW_REGISTER_TYPE_D
), /* does it matter? */
539 if (dst
.dw1
.bits
.writemask
& WRITEMASK_YZ
) {
541 brw_writemask(tmp_ud
, WRITEMASK_Y
),
542 brw_swizzle1(arg0_ud
, 0),
543 brw_imm_ud((1<<23)-1));
546 brw_writemask(tmp_ud
, WRITEMASK_Y
),
548 brw_imm_ud(127<<23));
551 if (dst
.dw1
.bits
.writemask
& WRITEMASK_Z
) {
552 /* result[2] = result[0] + LOG2(result[1]); */
554 /* Why bother? The above is just a hint how to do this with a
555 * taylor series. Maybe we *should* use a taylor series as by
556 * the time all the above has been done it's almost certainly
557 * quicker than calling the mathbox, even with low precision.
560 * - result[0] + mathbox.LOG2(result[1])
561 * - mathbox.LOG2(arg0.x)
562 * - result[0] + inline_taylor_approx(result[1])
565 BRW_MATH_FUNCTION_LOG
,
566 brw_writemask(tmp
, WRITEMASK_Z
),
567 brw_swizzle1(tmp
, 1),
568 BRW_MATH_PRECISION_FULL
);
571 brw_writemask(tmp
, WRITEMASK_Z
),
572 brw_swizzle1(tmp
, 2),
573 brw_swizzle1(tmp
, 0));
576 if (dst
.dw1
.bits
.writemask
& WRITEMASK_W
) {
577 /* result[3] = 1.0; */
578 brw_MOV(p
, brw_writemask(tmp
, WRITEMASK_W
), brw_imm_f(1));
582 brw_MOV(p
, dst
, tmp
);
588 /* Need to unalias - consider swizzles: r0 = DST r0.xxxx r1
590 static void emit_dst_noalias( struct brw_vs_compile
*c
,
595 struct brw_compile
*p
= &c
->func
;
597 /* There must be a better way to do this:
599 if (dst
.dw1
.bits
.writemask
& WRITEMASK_X
)
600 brw_MOV(p
, brw_writemask(dst
, WRITEMASK_X
), brw_imm_f(1.0));
601 if (dst
.dw1
.bits
.writemask
& WRITEMASK_Y
)
602 brw_MUL(p
, brw_writemask(dst
, WRITEMASK_Y
), arg0
, arg1
);
603 if (dst
.dw1
.bits
.writemask
& WRITEMASK_Z
)
604 brw_MOV(p
, brw_writemask(dst
, WRITEMASK_Z
), arg0
);
605 if (dst
.dw1
.bits
.writemask
& WRITEMASK_W
)
606 brw_MOV(p
, brw_writemask(dst
, WRITEMASK_W
), arg1
);
610 static void emit_xpd( struct brw_compile
*p
,
615 brw_MUL(p
, brw_null_reg(), brw_swizzle(t
, 1,2,0,3), brw_swizzle(u
,2,0,1,3));
616 brw_MAC(p
, dst
, negate(brw_swizzle(t
, 2,0,1,3)), brw_swizzle(u
,1,2,0,3));
620 static void emit_lit_noalias( struct brw_vs_compile
*c
,
622 struct brw_reg arg0
)
624 struct brw_compile
*p
= &c
->func
;
625 struct brw_instruction
*if_insn
;
626 struct brw_reg tmp
= dst
;
627 GLboolean need_tmp
= (dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
632 brw_MOV(p
, brw_writemask(dst
, WRITEMASK_YZ
), brw_imm_f(0));
633 brw_MOV(p
, brw_writemask(dst
, WRITEMASK_XW
), brw_imm_f(1));
635 /* Need to use BRW_EXECUTE_8 and also do an 8-wide compare in order
636 * to get all channels active inside the IF. In the clipping code
637 * we run with NoMask, so it's not an option and we can use
638 * BRW_EXECUTE_1 for all comparisions.
640 brw_CMP(p
, brw_null_reg(), BRW_CONDITIONAL_G
, brw_swizzle1(arg0
,0), brw_imm_f(0));
641 if_insn
= brw_IF(p
, BRW_EXECUTE_8
);
643 brw_MOV(p
, brw_writemask(dst
, WRITEMASK_Y
), brw_swizzle1(arg0
,0));
645 brw_CMP(p
, brw_null_reg(), BRW_CONDITIONAL_G
, brw_swizzle1(arg0
,1), brw_imm_f(0));
646 brw_MOV(p
, brw_writemask(tmp
, WRITEMASK_Z
), brw_swizzle1(arg0
,1));
647 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
650 BRW_MATH_FUNCTION_POW
,
651 brw_writemask(dst
, WRITEMASK_Z
),
652 brw_swizzle1(tmp
, 2),
653 brw_swizzle1(arg0
, 3),
654 BRW_MATH_PRECISION_PARTIAL
);
657 brw_ENDIF(p
, if_insn
);
660 static void emit_lrp_noalias(struct brw_vs_compile
*c
,
666 struct brw_compile
*p
= &c
->func
;
668 brw_ADD(p
, dst
, negate(arg0
), brw_imm_f(1.0));
669 brw_MUL(p
, brw_null_reg(), dst
, arg2
);
670 brw_MAC(p
, dst
, arg0
, arg1
);
673 /** 3 or 4-component vector normalization */
674 static void emit_nrm( struct brw_vs_compile
*c
,
679 struct brw_compile
*p
= &c
->func
;
680 struct brw_reg tmp
= get_tmp(c
);
682 /* tmp = dot(arg0, arg0) */
684 brw_DP3(p
, tmp
, arg0
, arg0
);
686 brw_DP4(p
, tmp
, arg0
, arg0
);
688 /* tmp = 1 / sqrt(tmp) */
689 emit_math1(c
, BRW_MATH_FUNCTION_RSQ
, tmp
, tmp
, BRW_MATH_PRECISION_FULL
);
691 /* dst = arg0 * tmp */
692 brw_MUL(p
, dst
, arg0
, tmp
);
698 static struct brw_reg
699 get_constant(struct brw_vs_compile
*c
,
700 const struct prog_instruction
*inst
,
703 const struct prog_src_register
*src
= &inst
->SrcReg
[argIndex
];
704 struct brw_compile
*p
= &c
->func
;
705 struct brw_reg const_reg
;
707 if (c
->current_const
[argIndex
].index
!= src
->Index
) {
708 struct brw_reg src_reg
= get_tmp(c
);
709 struct brw_reg t
= get_tmp(c
);
711 c
->current_const
[argIndex
].index
= src
->Index
;
713 brw_MOV(p
, t
, brw_vec8_grf(0, 0));/*SAVE*/
716 printf(" fetch const[%d] for arg %d into reg %d\n",
717 src
->Index
, argIndex
, c
->current_const
[argIndex
].reg
.nr
);
720 /* need to fetch the constant now */
722 c
->current_const
[argIndex
].reg
, /* writeback dest */
723 src_reg
, /* src reg */
725 src
->RelAddr
, /* relative indexing? */
726 16 * src
->Index
, /* byte offset */
727 SURF_INDEX_VERT_CONST_BUFFER
/* binding table index */
730 brw_MOV(p
, brw_vec8_grf(0, 0), t
);/*RESTORE*/
731 release_tmp(c
, src_reg
);
735 /* replicate lower four floats into upper four floats (to get XYZWXYZW) */
736 const_reg
= c
->current_const
[argIndex
].reg
;
737 const_reg
= stride(const_reg
, 0, 4, 0);
745 /* TODO: relative addressing!
747 static struct brw_reg
get_reg( struct brw_vs_compile
*c
,
748 gl_register_file file
,
752 case PROGRAM_TEMPORARY
:
755 assert(c
->regs
[file
][index
].nr
!= 0);
756 return c
->regs
[file
][index
];
757 case PROGRAM_STATE_VAR
:
758 case PROGRAM_CONSTANT
:
759 case PROGRAM_UNIFORM
:
760 assert(c
->regs
[PROGRAM_STATE_VAR
][index
].nr
!= 0);
761 return c
->regs
[PROGRAM_STATE_VAR
][index
];
762 case PROGRAM_ADDRESS
:
764 return c
->regs
[file
][index
];
766 case PROGRAM_UNDEFINED
: /* undef values */
767 return brw_null_reg();
769 case PROGRAM_LOCAL_PARAM
:
770 case PROGRAM_ENV_PARAM
:
771 case PROGRAM_WRITE_ONLY
:
774 return brw_null_reg();
780 * Get brw reg corresponding to the instruction's [argIndex] src reg.
781 * TODO: relative addressing!
783 static struct brw_reg
784 get_src_reg( struct brw_vs_compile
*c
,
785 const struct prog_instruction
*inst
,
788 const GLuint file
= inst
->SrcReg
[argIndex
].File
;
789 const GLint index
= inst
->SrcReg
[argIndex
].Index
;
792 case PROGRAM_TEMPORARY
:
795 assert(c
->regs
[file
][index
].nr
!= 0);
796 return c
->regs
[file
][index
];
797 case PROGRAM_STATE_VAR
:
798 case PROGRAM_CONSTANT
:
799 case PROGRAM_UNIFORM
:
800 if (c
->use_const_buffer
) {
801 return get_constant(c
, inst
, argIndex
);
804 assert(c
->regs
[PROGRAM_STATE_VAR
][index
].nr
!= 0);
805 return c
->regs
[PROGRAM_STATE_VAR
][index
];
807 case PROGRAM_ADDRESS
:
809 return c
->regs
[file
][index
];
811 case PROGRAM_UNDEFINED
:
812 /* this is a normal case since we loop over all three src args */
813 return brw_null_reg();
815 case PROGRAM_LOCAL_PARAM
:
816 case PROGRAM_ENV_PARAM
:
817 case PROGRAM_WRITE_ONLY
:
820 return brw_null_reg();
826 * Indirect addressing: get reg[[arg] + offset].
828 static struct brw_reg
deref( struct brw_vs_compile
*c
,
832 struct brw_compile
*p
= &c
->func
;
833 struct brw_reg tmp
= vec4(get_tmp(c
));
834 struct brw_reg addr_reg
= c
->regs
[PROGRAM_ADDRESS
][0];
835 struct brw_reg vp_address
= retype(vec1(addr_reg
), BRW_REGISTER_TYPE_UW
);
836 GLuint byte_offset
= arg
.nr
* 32 + arg
.subnr
+ offset
* 16;
837 struct brw_reg indirect
= brw_vec4_indirect(0,0);
840 brw_push_insn_state(p
);
841 brw_set_access_mode(p
, BRW_ALIGN_1
);
843 /* This is pretty clunky - load the address register twice and
844 * fetch each 4-dword value in turn. There must be a way to do
845 * this in a single pass, but I couldn't get it to work.
847 brw_ADD(p
, brw_address_reg(0), vp_address
, brw_imm_d(byte_offset
));
848 brw_MOV(p
, tmp
, indirect
);
850 brw_ADD(p
, brw_address_reg(0), suboffset(vp_address
, 8), brw_imm_d(byte_offset
));
851 brw_MOV(p
, suboffset(tmp
, 4), indirect
);
853 brw_pop_insn_state(p
);
860 static void emit_arl( struct brw_vs_compile
*c
,
862 struct brw_reg arg0
)
864 struct brw_compile
*p
= &c
->func
;
865 struct brw_reg tmp
= dst
;
866 GLboolean need_tmp
= (dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
871 brw_RNDD(p
, tmp
, arg0
);
872 brw_MUL(p
, dst
, tmp
, brw_imm_d(16));
880 * Return the brw reg for the given instruction's src argument.
881 * Will return mangled results for SWZ op. The emit_swz() function
882 * ignores this result and recalculates taking extended swizzles into
885 static struct brw_reg
get_arg( struct brw_vs_compile
*c
,
886 const struct prog_instruction
*inst
,
889 const struct prog_src_register
*src
= &inst
->SrcReg
[argIndex
];
892 if (src
->File
== PROGRAM_UNDEFINED
)
893 return brw_null_reg();
897 reg
= deref(c
, c
->regs
[PROGRAM_STATE_VAR
][0], src
->Index
);
900 reg
= get_src_reg(c
, inst
, argIndex
);
903 /* Convert 3-bit swizzle to 2-bit.
905 reg
.dw1
.bits
.swizzle
= BRW_SWIZZLE4(GET_SWZ(src
->Swizzle
, 0),
906 GET_SWZ(src
->Swizzle
, 1),
907 GET_SWZ(src
->Swizzle
, 2),
908 GET_SWZ(src
->Swizzle
, 3));
910 /* Note this is ok for non-swizzle instructions:
912 reg
.negate
= src
->NegateBase
? 1 : 0;
919 * Get brw register for the given program dest register.
921 static struct brw_reg
get_dst( struct brw_vs_compile
*c
,
922 struct prog_dst_register dst
)
927 case PROGRAM_TEMPORARY
:
929 assert(c
->regs
[dst
.File
][dst
.Index
].nr
!= 0);
930 reg
= c
->regs
[dst
.File
][dst
.Index
];
932 case PROGRAM_UNDEFINED
:
933 /* we may hit this for OPCODE_END, OPCODE_KIL, etc */
934 reg
= brw_null_reg();
938 reg
= brw_null_reg();
941 reg
.dw1
.bits
.writemask
= dst
.WriteMask
;
947 static void emit_swz( struct brw_vs_compile
*c
,
949 const struct prog_instruction
*inst
)
951 const GLuint argIndex
= 0;
952 const struct prog_src_register src
= inst
->SrcReg
[argIndex
];
953 struct brw_compile
*p
= &c
->func
;
954 GLuint zeros_mask
= 0;
955 GLuint ones_mask
= 0;
958 GLboolean need_tmp
= (src
.NegateBase
&&
959 dst
.file
!= BRW_GENERAL_REGISTER_FILE
);
960 struct brw_reg tmp
= dst
;
966 for (i
= 0; i
< 4; i
++) {
967 if (dst
.dw1
.bits
.writemask
& (1<<i
)) {
968 GLubyte s
= GET_SWZ(src
.Swizzle
, i
);
987 /* Do src first, in case dst aliases src:
993 arg0
= deref(c
, c
->regs
[PROGRAM_STATE_VAR
][0], src
.Index
);
995 arg0
= get_src_reg(c
, inst
, argIndex
);
997 arg0
= brw_swizzle(arg0
,
998 src_swz
[0], src_swz
[1],
999 src_swz
[2], src_swz
[3]);
1001 brw_MOV(p
, brw_writemask(tmp
, src_mask
), arg0
);
1005 brw_MOV(p
, brw_writemask(tmp
, zeros_mask
), brw_imm_f(0));
1008 brw_MOV(p
, brw_writemask(tmp
, ones_mask
), brw_imm_f(1));
1011 brw_MOV(p
, brw_writemask(tmp
, src
.NegateBase
), negate(tmp
));
1014 brw_MOV(p
, dst
, tmp
);
1015 release_tmp(c
, tmp
);
1021 * Post-vertex-program processing. Send the results to the URB.
1023 static void emit_vertex_write( struct brw_vs_compile
*c
)
1025 struct brw_compile
*p
= &c
->func
;
1026 struct brw_reg m0
= brw_message_reg(0);
1027 struct brw_reg pos
= c
->regs
[PROGRAM_OUTPUT
][VERT_RESULT_HPOS
];
1030 if (c
->key
.copy_edgeflag
) {
1032 get_reg(c
, PROGRAM_OUTPUT
, VERT_RESULT_EDGE
),
1033 get_reg(c
, PROGRAM_INPUT
, VERT_ATTRIB_EDGEFLAG
));
1036 /* Build ndc coords */
1038 emit_math1(c
, BRW_MATH_FUNCTION_INV
, ndc
, brw_swizzle1(pos
, 3), BRW_MATH_PRECISION_FULL
);
1039 brw_MUL(p
, brw_writemask(ndc
, WRITEMASK_XYZ
), pos
, ndc
);
1041 /* Update the header for point size, user clipping flags, and -ve rhw
1044 if ((c
->prog_data
.outputs_written
& (1<<VERT_RESULT_PSIZ
)) ||
1045 c
->key
.nr_userclip
|| !BRW_IS_G4X(p
->brw
))
1047 struct brw_reg header1
= retype(get_tmp(c
), BRW_REGISTER_TYPE_UD
);
1050 brw_MOV(p
, header1
, brw_imm_ud(0));
1052 brw_set_access_mode(p
, BRW_ALIGN_16
);
1054 if (c
->prog_data
.outputs_written
& (1<<VERT_RESULT_PSIZ
)) {
1055 struct brw_reg psiz
= c
->regs
[PROGRAM_OUTPUT
][VERT_RESULT_PSIZ
];
1056 brw_MUL(p
, brw_writemask(header1
, WRITEMASK_W
), brw_swizzle1(psiz
, 0), brw_imm_f(1<<11));
1057 brw_AND(p
, brw_writemask(header1
, WRITEMASK_W
), header1
, brw_imm_ud(0x7ff<<8));
1060 for (i
= 0; i
< c
->key
.nr_userclip
; i
++) {
1061 brw_set_conditionalmod(p
, BRW_CONDITIONAL_L
);
1062 brw_DP4(p
, brw_null_reg(), pos
, c
->userplane
[i
]);
1063 brw_OR(p
, brw_writemask(header1
, WRITEMASK_W
), header1
, brw_imm_ud(1<<i
));
1064 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1067 /* i965 clipping workaround:
1068 * 1) Test for -ve rhw
1070 * set ndc = (0,0,0,0)
1073 * Later, clipping will detect ucp[6] and ensure the primitive is
1074 * clipped against all fixed planes.
1076 if (!BRW_IS_G4X(p
->brw
)) {
1078 vec8(brw_null_reg()),
1080 brw_swizzle1(ndc
, 3),
1083 brw_OR(p
, brw_writemask(header1
, WRITEMASK_W
), header1
, brw_imm_ud(1<<6));
1084 brw_MOV(p
, ndc
, brw_imm_f(0));
1085 brw_set_predicate_control(p
, BRW_PREDICATE_NONE
);
1088 brw_set_access_mode(p
, BRW_ALIGN_1
); /* why? */
1089 brw_MOV(p
, retype(brw_message_reg(1), BRW_REGISTER_TYPE_UD
), header1
);
1090 brw_set_access_mode(p
, BRW_ALIGN_16
);
1092 release_tmp(c
, header1
);
1095 brw_MOV(p
, retype(brw_message_reg(1), BRW_REGISTER_TYPE_UD
), brw_imm_ud(0));
1098 /* Emit the (interleaved) headers for the two vertices - an 8-reg
1099 * of zeros followed by two sets of NDC coordinates:
1101 brw_set_access_mode(p
, BRW_ALIGN_1
);
1102 brw_MOV(p
, offset(m0
, 2), ndc
);
1103 brw_MOV(p
, offset(m0
, 3), pos
);
1106 brw_null_reg(), /* dest */
1107 0, /* starting mrf reg nr */
1111 c
->nr_outputs
+ 3, /* msg len */
1112 0, /* response len */
1114 1, /* writes complete */
1115 0, /* urb destination offset */
1116 BRW_URB_SWIZZLE_INTERLEAVE
);
1121 * Called after code generation to resolve subroutine calls and the
1123 * \param end_inst points to brw code for END instruction
1124 * \param last_inst points to last instruction emitted before vertex write
1127 post_vs_emit( struct brw_vs_compile
*c
,
1128 struct brw_instruction
*end_inst
,
1129 struct brw_instruction
*last_inst
)
1133 brw_resolve_cals(&c
->func
);
1135 /* patch up the END code to jump past subroutines, etc */
1136 offset
= last_inst
- end_inst
;
1137 brw_set_src1(end_inst
, brw_imm_d(offset
* 16));
1141 /* Emit the vertex program instructions here.
1143 void brw_vs_emit(struct brw_vs_compile
*c
)
1146 struct brw_compile
*p
= &c
->func
;
1147 GLuint nr_insns
= c
->vp
->program
.Base
.NumInstructions
;
1148 GLuint insn
, if_insn
= 0;
1149 GLuint end_offset
= 0;
1150 struct brw_instruction
*end_inst
, *last_inst
;
1151 struct brw_instruction
*if_inst
[MAX_IFSN
];
1152 struct brw_indirect stack_index
= brw_indirect(0, 0);
1157 if (INTEL_DEBUG
& DEBUG_VS
) {
1158 _mesa_printf("vs-emit:\n");
1159 _mesa_print_program(&c
->vp
->program
.Base
);
1163 brw_set_compression_control(p
, BRW_COMPRESSION_NONE
);
1164 brw_set_access_mode(p
, BRW_ALIGN_16
);
1166 /* Message registers can't be read, so copy the output into GRF register
1167 if they are used in source registers */
1168 for (insn
= 0; insn
< nr_insns
; insn
++) {
1170 struct prog_instruction
*inst
= &c
->vp
->program
.Base
.Instructions
[insn
];
1171 for (i
= 0; i
< 3; i
++) {
1172 struct prog_src_register
*src
= &inst
->SrcReg
[i
];
1173 GLuint index
= src
->Index
;
1174 GLuint file
= src
->File
;
1175 if (file
== PROGRAM_OUTPUT
&& index
!= VERT_RESULT_HPOS
)
1176 c
->output_regs
[index
].used_in_src
= GL_TRUE
;
1180 /* Static register allocation
1182 brw_vs_alloc_regs(c
);
1183 brw_MOV(p
, get_addr_reg(stack_index
), brw_address(c
->stack
));
1185 for (insn
= 0; insn
< nr_insns
; insn
++) {
1187 struct prog_instruction
*inst
= &c
->vp
->program
.Base
.Instructions
[insn
];
1188 struct brw_reg args
[3], dst
;
1191 /* Get argument regs. SWZ is special and does this itself.
1193 if (inst
->Opcode
!= OPCODE_SWZ
)
1194 for (i
= 0; i
< 3; i
++) {
1195 struct prog_src_register
*src
= &inst
->SrcReg
[i
];
1198 if (file
== PROGRAM_OUTPUT
&& c
->output_regs
[index
].used_in_src
)
1199 args
[i
] = c
->output_regs
[index
].reg
;
1201 args
[i
] = get_arg(c
, inst
, i
);
1204 /* Get dest regs. Note that it is possible for a reg to be both
1205 * dst and arg, given the static allocation of registers. So
1206 * care needs to be taken emitting multi-operation instructions.
1208 index
= inst
->DstReg
.Index
;
1209 file
= inst
->DstReg
.File
;
1210 if (file
== PROGRAM_OUTPUT
&& c
->output_regs
[index
].used_in_src
)
1211 dst
= c
->output_regs
[index
].reg
;
1213 dst
= get_dst(c
, inst
->DstReg
);
1215 if (inst
->SaturateMode
!= SATURATE_OFF
) {
1216 _mesa_problem(NULL
, "Unsupported saturate %d in vertex shader",
1217 inst
->SaturateMode
);
1220 switch (inst
->Opcode
) {
1222 brw_MOV(p
, dst
, brw_abs(args
[0]));
1225 brw_ADD(p
, dst
, args
[0], args
[1]);
1228 emit_math1(c
, BRW_MATH_FUNCTION_COS
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1231 brw_DP3(p
, dst
, args
[0], args
[1]);
1234 brw_DP4(p
, dst
, args
[0], args
[1]);
1237 brw_DPH(p
, dst
, args
[0], args
[1]);
1240 emit_nrm(c
, dst
, args
[0], 3);
1243 emit_nrm(c
, dst
, args
[0], 4);
1246 unalias2(c
, dst
, args
[0], args
[1], emit_dst_noalias
);
1249 unalias1(c
, dst
, args
[0], emit_exp_noalias
);
1252 emit_math1(c
, BRW_MATH_FUNCTION_EXP
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1255 emit_arl(c
, dst
, args
[0]);
1258 brw_RNDD(p
, dst
, args
[0]);
1261 brw_FRC(p
, dst
, args
[0]);
1264 unalias1(c
, dst
, args
[0], emit_log_noalias
);
1267 emit_math1(c
, BRW_MATH_FUNCTION_LOG
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1270 unalias1(c
, dst
, args
[0], emit_lit_noalias
);
1273 unalias3(c
, dst
, args
[0], args
[1], args
[2], emit_lrp_noalias
);
1276 brw_MOV(p
, brw_acc_reg(), args
[2]);
1277 brw_MAC(p
, dst
, args
[0], args
[1]);
1280 emit_max(p
, dst
, args
[0], args
[1]);
1283 emit_min(p
, dst
, args
[0], args
[1]);
1286 brw_MOV(p
, dst
, args
[0]);
1289 brw_MUL(p
, dst
, args
[0], args
[1]);
1292 emit_math2(c
, BRW_MATH_FUNCTION_POW
, dst
, args
[0], args
[1], BRW_MATH_PRECISION_FULL
);
1295 emit_math1(c
, BRW_MATH_FUNCTION_INV
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1298 emit_math1(c
, BRW_MATH_FUNCTION_RSQ
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1302 emit_seq(p
, dst
, args
[0], args
[1]);
1305 emit_math1(c
, BRW_MATH_FUNCTION_SIN
, dst
, args
[0], BRW_MATH_PRECISION_FULL
);
1308 emit_sne(p
, dst
, args
[0], args
[1]);
1311 emit_sge(p
, dst
, args
[0], args
[1]);
1314 emit_sgt(p
, dst
, args
[0], args
[1]);
1317 emit_slt(p
, dst
, args
[0], args
[1]);
1320 emit_sle(p
, dst
, args
[0], args
[1]);
1323 brw_ADD(p
, dst
, args
[0], negate(args
[1]));
1326 /* The args[0] value can't be used here as it won't have
1327 * correctly encoded the full swizzle:
1329 emit_swz(c
, dst
, inst
);
1332 /* round toward zero */
1333 brw_RNDZ(p
, dst
, args
[0]);
1336 emit_xpd(p
, dst
, args
[0], args
[1]);
1339 assert(if_insn
< MAX_IFSN
);
1340 if_inst
[if_insn
++] = brw_IF(p
, BRW_EXECUTE_8
);
1343 if_inst
[if_insn
-1] = brw_ELSE(p
, if_inst
[if_insn
-1]);
1346 assert(if_insn
> 0);
1347 brw_ENDIF(p
, if_inst
[--if_insn
]);
1350 brw_set_predicate_control(p
, BRW_PREDICATE_NORMAL
);
1351 brw_ADD(p
, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16));
1352 brw_set_predicate_control_flag_value(p
, 0xff);
1355 brw_set_access_mode(p
, BRW_ALIGN_1
);
1356 brw_ADD(p
, deref_1d(stack_index
, 0), brw_ip_reg(), brw_imm_d(3*16));
1357 brw_set_access_mode(p
, BRW_ALIGN_16
);
1358 brw_ADD(p
, get_addr_reg(stack_index
),
1359 get_addr_reg(stack_index
), brw_imm_d(4));
1360 brw_save_call(p
, inst
->Comment
, p
->nr_insn
);
1361 brw_ADD(p
, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16));
1364 brw_ADD(p
, get_addr_reg(stack_index
),
1365 get_addr_reg(stack_index
), brw_imm_d(-4));
1366 brw_set_access_mode(p
, BRW_ALIGN_1
);
1367 brw_MOV(p
, brw_ip_reg(), deref_1d(stack_index
, 0));
1368 brw_set_access_mode(p
, BRW_ALIGN_16
);
1371 end_offset
= p
->nr_insn
;
1372 /* this instruction will get patched later to jump past subroutine
1375 brw_ADD(p
, brw_ip_reg(), brw_ip_reg(), brw_imm_d(1*16));
1381 brw_save_label(p
, inst
->Comment
, p
->nr_insn
);
1387 _mesa_problem(NULL
, "Unsupported opcode %i (%s) in vertex shader",
1388 inst
->Opcode
, inst
->Opcode
< MAX_OPCODE
?
1389 _mesa_opcode_string(inst
->Opcode
) :
1393 if ((inst
->DstReg
.File
== PROGRAM_OUTPUT
)
1394 && (inst
->DstReg
.Index
!= VERT_RESULT_HPOS
)
1395 && c
->output_regs
[inst
->DstReg
.Index
].used_in_src
) {
1396 brw_MOV(p
, get_dst(c
, inst
->DstReg
), dst
);
1399 /* Result color clamping.
1401 * When destination register is an output register and
1402 * it's primary/secondary front/back color, we have to clamp
1403 * the result to [0,1]. This is done by enabling the
1404 * saturation bit for the last instruction.
1406 * We don't use brw_set_saturate() as it modifies
1407 * p->current->header.saturate, which affects all the subsequent
1408 * instructions. Instead, we directly modify the header
1409 * of the last (already stored) instruction.
1411 if (inst
->DstReg
.File
== PROGRAM_OUTPUT
) {
1412 if ((inst
->DstReg
.Index
== VERT_RESULT_COL0
)
1413 || (inst
->DstReg
.Index
== VERT_RESULT_COL1
)
1414 || (inst
->DstReg
.Index
== VERT_RESULT_BFC0
)
1415 || (inst
->DstReg
.Index
== VERT_RESULT_BFC1
)) {
1416 p
->store
[p
->nr_insn
-1].header
.saturate
= 1;
1423 end_inst
= &p
->store
[end_offset
];
1424 last_inst
= &p
->store
[p
->nr_insn
];
1426 /* The END instruction will be patched to jump to this code */
1427 emit_vertex_write(c
);
1429 post_vs_emit(c
, end_inst
, last_inst
);