2 * Copyright 2008 Ben Skeggs
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY,
18 * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF
19 * OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
23 #include "pipe/p_context.h"
24 #include "pipe/p_defines.h"
25 #include "pipe/p_state.h"
26 #include "pipe/p_inlines.h"
28 #include "pipe/p_shader_tokens.h"
29 #include "tgsi/tgsi_parse.h"
30 #include "tgsi/tgsi_util.h"
32 #include "nv50_context.h"
34 #define NV50_SU_MAX_TEMP 127
35 #define NV50_SU_MAX_ADDR 4
36 //#define NV50_PROGRAM_DUMP
38 /* $a5 and $a6 always seem to be 0, and using $a7 gives you noise */
40 /* ARL - gallium craps itself on progs/vp/arl.txt
42 * MSB - Like MAD, but MUL+SUB
43 * - Fuck it off, introduce a way to negate args for ops that
46 * Look into inlining IMMD for ops other than MOV (make it general?)
47 * - Maybe even relax restrictions a bit, can't do P_RESULT + P_IMMD,
48 * but can emit to P_TEMP first - then MOV later. NVIDIA does this
50 * In ops such as ADD it's possible to construct a bad opcode in the !is_long()
51 * case, if the emit_src() causes the inst to suddenly become long.
53 * Verify half-insns work where expected - and force disable them where they
54 * don't work - MUL has it forcibly disabled atm as it fixes POW..
56 * FUCK! watch dst==src vectors, can overwrite components that are needed.
57 * ie. SUB R0, R0.yzxw, R0
59 * Things to check with renouveau:
60 * FP attr/result assignment - how?
62 * - 0x16bc maps vp output onto fp hpos
63 * - 0x16c0 maps vp output onto fp col0
67 * 0x16bc->0x16e8 --> some binding between vp/fp regs
68 * 0x16b8 --> VP output count
70 * 0x1298 --> "MOV rcol.x, fcol.y" "MOV depr, fcol.y" = 0x00000005
71 * "MOV rcol.x, fcol.y" = 0x00000004
72 * 0x19a8 --> as above but 0x00000100 and 0x00000000
73 * - 0x00100000 used when KIL used
74 * 0x196c --> as above but 0x00000011 and 0x00000000
76 * 0x1988 --> 0xXXNNNNNN
77 * - XX == FP high something
93 int rhw
; /* result hw for FP outputs, or interpolant index */
94 int acc
; /* instruction where this reg is last read (first insn == 1) */
97 #define NV50_MOD_NEG 1
98 #define NV50_MOD_ABS 2
99 #define NV50_MOD_SAT 4
101 /* arbitrary limits */
102 #define MAX_IF_DEPTH 4
103 #define MAX_LOOP_DEPTH 4
106 struct nv50_program
*p
;
109 struct nv50_reg
*r_temp
[NV50_SU_MAX_TEMP
];
110 struct nv50_reg r_addr
[NV50_SU_MAX_ADDR
];
113 struct nv50_reg
*temp
;
115 struct nv50_reg
*attr
;
117 struct nv50_reg
*result
;
119 struct nv50_reg
*param
;
121 struct nv50_reg
*immd
;
124 struct nv50_reg
**addr
;
127 struct nv50_reg
*temp_temp
[16];
128 unsigned temp_temp_nr
;
130 /* broadcast and destination replacement regs */
131 struct nv50_reg
*r_brdc
;
132 struct nv50_reg
*r_dst
[4];
134 struct nv50_reg reg_instances
[16];
135 unsigned reg_instance_nr
;
137 unsigned interp_mode
[32];
138 /* perspective interpolation registers */
139 struct nv50_reg
*iv_p
;
140 struct nv50_reg
*iv_c
;
142 struct nv50_program_exec
*if_cond
;
143 struct nv50_program_exec
*if_insn
[MAX_IF_DEPTH
];
144 struct nv50_program_exec
*br_join
[MAX_IF_DEPTH
];
145 struct nv50_program_exec
*br_loop
[MAX_LOOP_DEPTH
]; /* for BRK branch */
146 int if_lvl
, loop_lvl
;
147 unsigned loop_pos
[MAX_LOOP_DEPTH
];
149 /* current instruction and total number of insns */
156 static INLINE
struct nv50_reg
*
157 reg_instance(struct nv50_pc
*pc
, struct nv50_reg
*reg
)
159 struct nv50_reg
*dup
= NULL
;
161 assert(pc
->reg_instance_nr
< 16);
162 dup
= &pc
->reg_instances
[pc
->reg_instance_nr
++];
170 ctor_reg(struct nv50_reg
*reg
, unsigned type
, int index
, int hw
)
180 static INLINE
unsigned
181 popcnt4(uint32_t val
)
183 static const unsigned cnt
[16]
184 = { 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4 };
185 return cnt
[val
& 0xf];
189 terminate_mbb(struct nv50_pc
*pc
)
193 /* remove records of temporary address register values */
194 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
)
195 if (pc
->r_addr
[i
].index
< 0)
196 pc
->r_addr
[i
].rhw
= -1;
200 alloc_reg(struct nv50_pc
*pc
, struct nv50_reg
*reg
)
204 if (reg
->type
== P_RESULT
) {
205 if (pc
->p
->cfg
.high_result
< (reg
->hw
+ 1))
206 pc
->p
->cfg
.high_result
= reg
->hw
+ 1;
209 if (reg
->type
!= P_TEMP
)
213 /*XXX: do this here too to catch FP temp-as-attr usage..
214 * not clean, but works */
215 if (pc
->p
->cfg
.high_temp
< (reg
->hw
+ 1))
216 pc
->p
->cfg
.high_temp
= reg
->hw
+ 1;
220 if (reg
->rhw
!= -1) {
221 /* try to allocate temporary with index rhw first */
222 if (!(pc
->r_temp
[reg
->rhw
])) {
223 pc
->r_temp
[reg
->rhw
] = reg
;
225 if (pc
->p
->cfg
.high_temp
< (reg
->rhw
+ 1))
226 pc
->p
->cfg
.high_temp
= reg
->rhw
+ 1;
229 /* make sure we don't get things like $r0 needs to go
230 * in $r1 and $r1 in $r0
232 i
= pc
->result_nr
* 4;
235 for (; i
< NV50_SU_MAX_TEMP
; i
++) {
236 if (!(pc
->r_temp
[i
])) {
239 if (pc
->p
->cfg
.high_temp
< (i
+ 1))
240 pc
->p
->cfg
.high_temp
= i
+ 1;
248 /* XXX: For shaders that aren't executed linearly (e.g. shaders that
249 * contain loops), we need to assign all hw regs to TGSI TEMPs early,
250 * lest we risk temp_temps overwriting regs alloc'd "later".
252 static struct nv50_reg
*
253 alloc_temp(struct nv50_pc
*pc
, struct nv50_reg
*dst
)
258 if (dst
&& dst
->type
== P_TEMP
&& dst
->hw
== -1)
261 for (i
= 0; i
< NV50_SU_MAX_TEMP
; i
++) {
262 if (!pc
->r_temp
[i
]) {
263 r
= MALLOC_STRUCT(nv50_reg
);
264 ctor_reg(r
, P_TEMP
, -1, i
);
274 /* Assign the hw of the discarded temporary register src
275 * to the tgsi register dst and free src.
278 assimilate_temp(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
280 assert(src
->index
== -1 && src
->hw
!= -1);
283 pc
->r_temp
[dst
->hw
] = NULL
;
284 pc
->r_temp
[src
->hw
] = dst
;
290 /* release the hardware resource held by r */
292 release_hw(struct nv50_pc
*pc
, struct nv50_reg
*r
)
294 assert(r
->type
== P_TEMP
);
298 assert(pc
->r_temp
[r
->hw
] == r
);
299 pc
->r_temp
[r
->hw
] = NULL
;
307 free_temp(struct nv50_pc
*pc
, struct nv50_reg
*r
)
309 if (r
->index
== -1) {
312 FREE(pc
->r_temp
[hw
]);
313 pc
->r_temp
[hw
] = NULL
;
318 alloc_temp4(struct nv50_pc
*pc
, struct nv50_reg
*dst
[4], int idx
)
322 if ((idx
+ 4) >= NV50_SU_MAX_TEMP
)
325 if (pc
->r_temp
[idx
] || pc
->r_temp
[idx
+ 1] ||
326 pc
->r_temp
[idx
+ 2] || pc
->r_temp
[idx
+ 3])
327 return alloc_temp4(pc
, dst
, idx
+ 4);
329 for (i
= 0; i
< 4; i
++) {
330 dst
[i
] = MALLOC_STRUCT(nv50_reg
);
331 ctor_reg(dst
[i
], P_TEMP
, -1, idx
+ i
);
332 pc
->r_temp
[idx
+ i
] = dst
[i
];
339 free_temp4(struct nv50_pc
*pc
, struct nv50_reg
*reg
[4])
343 for (i
= 0; i
< 4; i
++)
344 free_temp(pc
, reg
[i
]);
347 static struct nv50_reg
*
348 temp_temp(struct nv50_pc
*pc
)
350 if (pc
->temp_temp_nr
>= 16)
353 pc
->temp_temp
[pc
->temp_temp_nr
] = alloc_temp(pc
, NULL
);
354 return pc
->temp_temp
[pc
->temp_temp_nr
++];
358 kill_temp_temp(struct nv50_pc
*pc
)
362 for (i
= 0; i
< pc
->temp_temp_nr
; i
++)
363 free_temp(pc
, pc
->temp_temp
[i
]);
364 pc
->temp_temp_nr
= 0;
368 ctor_immd(struct nv50_pc
*pc
, float x
, float y
, float z
, float w
)
370 pc
->immd_buf
= REALLOC(pc
->immd_buf
, (pc
->immd_nr
* 4 * sizeof(float)),
371 (pc
->immd_nr
+ 1) * 4 * sizeof(float));
372 pc
->immd_buf
[(pc
->immd_nr
* 4) + 0] = x
;
373 pc
->immd_buf
[(pc
->immd_nr
* 4) + 1] = y
;
374 pc
->immd_buf
[(pc
->immd_nr
* 4) + 2] = z
;
375 pc
->immd_buf
[(pc
->immd_nr
* 4) + 3] = w
;
377 return pc
->immd_nr
++;
381 ctor_immd_4u32(struct nv50_pc
*pc
, uint32_t x
, uint32_t y
, uint32_t z
, uint32_t w
)
383 pc
->immd_buf
= REALLOC(pc
->immd_buf
, (pc
->immd_nr
* 4 * sizeof(uint32_t)),
384 (pc
->immd_nr
+ 1) * 4 * sizeof(uint32_t));
385 pc
->immd_buf
[(pc
->immd_nr
* 4) + 0] = x
;
386 pc
->immd_buf
[(pc
->immd_nr
* 4) + 1] = y
;
387 pc
->immd_buf
[(pc
->immd_nr
* 4) + 2] = z
;
388 pc
->immd_buf
[(pc
->immd_nr
* 4) + 3] = w
;
390 return pc
->immd_nr
++;
393 static struct nv50_reg
*
394 alloc_immd(struct nv50_pc
*pc
, float f
)
396 struct nv50_reg
*r
= MALLOC_STRUCT(nv50_reg
);
399 for (hw
= 0; hw
< pc
->immd_nr
* 4; hw
++)
400 if (pc
->immd_buf
[hw
] == f
)
403 if (hw
== pc
->immd_nr
* 4)
404 hw
= ctor_immd(pc
, f
, -f
, 0.5 * f
, 0) * 4;
406 ctor_reg(r
, P_IMMD
, -1, hw
);
410 static struct nv50_program_exec
*
411 exec(struct nv50_pc
*pc
)
413 struct nv50_program_exec
*e
= CALLOC_STRUCT(nv50_program_exec
);
420 emit(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
422 struct nv50_program
*p
= pc
->p
;
425 p
->exec_tail
->next
= e
;
429 p
->exec_size
+= (e
->inst
[0] & 1) ? 2 : 1;
432 static INLINE
void set_long(struct nv50_pc
*, struct nv50_program_exec
*);
435 is_long(struct nv50_program_exec
*e
)
443 is_immd(struct nv50_program_exec
*e
)
445 if (is_long(e
) && (e
->inst
[1] & 3) == 3)
451 set_pred(struct nv50_pc
*pc
, unsigned pred
, unsigned idx
,
452 struct nv50_program_exec
*e
)
455 e
->inst
[1] &= ~((0x1f << 7) | (0x3 << 12));
456 e
->inst
[1] |= (pred
<< 7) | (idx
<< 12);
460 set_pred_wr(struct nv50_pc
*pc
, unsigned on
, unsigned idx
,
461 struct nv50_program_exec
*e
)
464 e
->inst
[1] &= ~((0x3 << 4) | (1 << 6));
465 e
->inst
[1] |= (idx
<< 4) | (on
<< 6);
469 set_long(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
475 set_pred(pc
, 0xf, 0, e
);
476 set_pred_wr(pc
, 0, 0, e
);
480 set_dst(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_program_exec
*e
)
482 if (dst
->type
== P_RESULT
) {
484 e
->inst
[1] |= 0x00000008;
490 e
->inst
[0] |= (dst
->hw
<< 2);
494 set_immd(struct nv50_pc
*pc
, struct nv50_reg
*imm
, struct nv50_program_exec
*e
)
497 float f
= pc
->immd_buf
[imm
->hw
];
499 if (imm
->mod
& NV50_MOD_ABS
)
501 val
= fui((imm
->mod
& NV50_MOD_NEG
) ? -f
: f
);
504 /*XXX: can't be predicated - bits overlap.. catch cases where both
505 * are required and avoid them. */
506 set_pred(pc
, 0, 0, e
);
507 set_pred_wr(pc
, 0, 0, e
);
509 e
->inst
[1] |= 0x00000002 | 0x00000001;
510 e
->inst
[0] |= (val
& 0x3f) << 16;
511 e
->inst
[1] |= (val
>> 6) << 2;
515 set_addr(struct nv50_program_exec
*e
, struct nv50_reg
*a
)
517 assert(!(e
->inst
[0] & 0x0c000000));
518 assert(!(e
->inst
[1] & 0x00000004));
520 e
->inst
[0] |= (a
->hw
& 3) << 26;
521 e
->inst
[1] |= (a
->hw
>> 2) << 2;
525 emit_add_addr_imm(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
526 struct nv50_reg
*src0
, uint16_t src1_val
)
528 struct nv50_program_exec
*e
= exec(pc
);
530 e
->inst
[0] = 0xd0000000 | (src1_val
<< 9);
531 e
->inst
[1] = 0x20000000;
533 e
->inst
[0] |= dst
->hw
<< 2;
534 if (src0
) /* otherwise will add to $a0, which is always 0 */
540 static struct nv50_reg
*
541 alloc_addr(struct nv50_pc
*pc
, struct nv50_reg
*ref
)
544 struct nv50_reg
*a_tgsi
= NULL
, *a
= NULL
;
547 /* allocate for TGSI address reg */
548 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
) {
549 if (pc
->r_addr
[i
].index
>= 0)
551 if (pc
->r_addr
[i
].rhw
>= 0 &&
552 pc
->r_addr
[i
].acc
== pc
->insn_cur
)
555 pc
->r_addr
[i
].rhw
= -1;
556 pc
->r_addr
[i
].index
= i
;
557 return &pc
->r_addr
[i
];
563 /* Allocate and set an address reg so we can access 'ref'.
565 * If and r_addr has index < 0, it is not reserved for TGSI,
566 * and index will be the negative of the TGSI addr index the
567 * value in rhw is relative to, or -256 if rhw is an offset
568 * from 0. If rhw < 0, the reg has not been initialized.
570 for (i
= NV50_SU_MAX_ADDR
- 1; i
>= 0; --i
) {
571 if (pc
->r_addr
[i
].index
>= 0) /* occupied for TGSI */
573 if (pc
->r_addr
[i
].rhw
< 0) { /* unused */
577 if (!a
&& pc
->r_addr
[i
].acc
!= pc
->insn_cur
)
580 if (ref
->hw
- pc
->r_addr
[i
].rhw
>= 128)
583 if ((ref
->acc
>= 0 && pc
->r_addr
[i
].index
== -256) ||
584 (ref
->acc
< 0 && -pc
->r_addr
[i
].index
== ref
->index
)) {
585 pc
->r_addr
[i
].acc
= pc
->insn_cur
;
586 return &pc
->r_addr
[i
];
592 a_tgsi
= pc
->addr
[ref
->index
];
594 emit_add_addr_imm(pc
, a
, a_tgsi
, (ref
->hw
& ~0x7f) * 4);
596 a
->rhw
= ref
->hw
& ~0x7f;
597 a
->acc
= pc
->insn_cur
;
598 a
->index
= a_tgsi
? -ref
->index
: -256;
602 #define INTERP_LINEAR 0
603 #define INTERP_FLAT 1
604 #define INTERP_PERSPECTIVE 2
605 #define INTERP_CENTROID 4
607 /* interpolant index has been stored in dst->rhw */
609 emit_interp(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*iv
,
612 assert(dst
->rhw
!= -1);
613 struct nv50_program_exec
*e
= exec(pc
);
615 e
->inst
[0] |= 0x80000000;
617 e
->inst
[0] |= (dst
->rhw
<< 16);
619 if (mode
& INTERP_FLAT
) {
620 e
->inst
[0] |= (1 << 8);
622 if (mode
& INTERP_PERSPECTIVE
) {
623 e
->inst
[0] |= (1 << 25);
625 e
->inst
[0] |= (iv
->hw
<< 9);
628 if (mode
& INTERP_CENTROID
)
629 e
->inst
[0] |= (1 << 24);
636 set_data(struct nv50_pc
*pc
, struct nv50_reg
*src
, unsigned m
, unsigned s
,
637 struct nv50_program_exec
*e
)
641 e
->param
.index
= src
->hw
& 127;
643 e
->param
.mask
= m
<< (s
% 32);
646 set_addr(e
, alloc_addr(pc
, src
));
649 assert(src
->type
== P_CONST
);
650 set_addr(e
, pc
->addr
[src
->index
]);
653 e
->inst
[1] |= (((src
->type
== P_IMMD
) ? 0 : 1) << 22);
657 emit_mov(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
659 struct nv50_program_exec
*e
= exec(pc
);
661 e
->inst
[0] = 0x10000000;
667 if (!is_long(e
) && src
->type
== P_IMMD
) {
668 set_immd(pc
, src
, e
);
669 /*XXX: 32-bit, but steals part of "half" reg space - need to
670 * catch and handle this case if/when we do half-regs
673 if (src
->type
== P_IMMD
|| src
->type
== P_CONST
) {
675 set_data(pc
, src
, 0x7f, 9, e
);
676 e
->inst
[1] |= 0x20000000; /* mov from c[] */
678 if (src
->type
== P_ATTR
) {
680 e
->inst
[1] |= 0x00200000;
686 e
->inst
[0] |= (src
->hw
<< 9);
689 if (is_long(e
) && !is_immd(e
)) {
690 e
->inst
[1] |= 0x04000000; /* 32-bit */
691 e
->inst
[1] |= 0x0000c000; /* 32-bit c[] load / lane mask 0:1 */
692 if (!(e
->inst
[1] & 0x20000000))
693 e
->inst
[1] |= 0x00030000; /* lane mask 2:3 */
695 e
->inst
[0] |= 0x00008000;
701 emit_mov_immdval(struct nv50_pc
*pc
, struct nv50_reg
*dst
, float f
)
703 struct nv50_reg
*imm
= alloc_immd(pc
, f
);
704 emit_mov(pc
, dst
, imm
);
709 emit_nop(struct nv50_pc
*pc
)
711 struct nv50_program_exec
*e
= exec(pc
);
713 e
->inst
[0] = 0xf0000000;
715 e
->inst
[1] = 0xe0000000;
720 check_swap_src_0_1(struct nv50_pc
*pc
,
721 struct nv50_reg
**s0
, struct nv50_reg
**s1
)
723 struct nv50_reg
*src0
= *s0
, *src1
= *s1
;
725 if (src0
->type
== P_CONST
) {
726 if (src1
->type
!= P_CONST
) {
732 if (src1
->type
== P_ATTR
) {
733 if (src0
->type
!= P_ATTR
) {
744 set_src_0_restricted(struct nv50_pc
*pc
, struct nv50_reg
*src
,
745 struct nv50_program_exec
*e
)
747 struct nv50_reg
*temp
;
749 if (src
->type
!= P_TEMP
) {
750 temp
= temp_temp(pc
);
751 emit_mov(pc
, temp
, src
);
758 e
->inst
[0] |= (src
->hw
<< 9);
762 set_src_0(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
764 if (src
->type
== P_ATTR
) {
766 e
->inst
[1] |= 0x00200000;
768 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
769 struct nv50_reg
*temp
= temp_temp(pc
);
771 emit_mov(pc
, temp
, src
);
778 e
->inst
[0] |= (src
->hw
<< 9);
782 set_src_1(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
784 if (src
->type
== P_ATTR
) {
785 struct nv50_reg
*temp
= temp_temp(pc
);
787 emit_mov(pc
, temp
, src
);
790 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
791 assert(!(e
->inst
[0] & 0x00800000));
792 if (e
->inst
[0] & 0x01000000) {
793 struct nv50_reg
*temp
= temp_temp(pc
);
795 emit_mov(pc
, temp
, src
);
798 set_data(pc
, src
, 0x7f, 16, e
);
799 e
->inst
[0] |= 0x00800000;
806 e
->inst
[0] |= ((src
->hw
& 127) << 16);
810 set_src_2(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
814 if (src
->type
== P_ATTR
) {
815 struct nv50_reg
*temp
= temp_temp(pc
);
817 emit_mov(pc
, temp
, src
);
820 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
821 assert(!(e
->inst
[0] & 0x01000000));
822 if (e
->inst
[0] & 0x00800000) {
823 struct nv50_reg
*temp
= temp_temp(pc
);
825 emit_mov(pc
, temp
, src
);
828 set_data(pc
, src
, 0x7f, 32+14, e
);
829 e
->inst
[0] |= 0x01000000;
834 e
->inst
[1] |= ((src
->hw
& 127) << 14);
838 emit_mov_from_pred(struct nv50_pc
*pc
, struct nv50_reg
*dst
, int pred
)
840 struct nv50_program_exec
*e
= exec(pc
);
842 assert(dst
->type
== P_TEMP
);
843 e
->inst
[1] = 0x20000000 | (pred
<< 12);
851 emit_mov_to_pred(struct nv50_pc
*pc
, int pred
, struct nv50_reg
*src
)
853 struct nv50_program_exec
*e
= exec(pc
);
855 e
->inst
[0] = 0x000001fc;
856 e
->inst
[1] = 0xa0000008;
858 set_pred_wr(pc
, 1, pred
, e
);
859 set_src_0_restricted(pc
, src
, e
);
865 emit_mul(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
866 struct nv50_reg
*src1
)
868 struct nv50_program_exec
*e
= exec(pc
);
870 e
->inst
[0] |= 0xc0000000;
875 check_swap_src_0_1(pc
, &src0
, &src1
);
877 set_src_0(pc
, src0
, e
);
878 if (src1
->type
== P_IMMD
&& !is_long(e
)) {
879 if (src0
->mod
& NV50_MOD_NEG
)
880 e
->inst
[0] |= 0x00008000;
881 set_immd(pc
, src1
, e
);
883 set_src_1(pc
, src1
, e
);
884 if ((src0
->mod
^ src1
->mod
) & NV50_MOD_NEG
) {
886 e
->inst
[1] |= 0x08000000;
888 e
->inst
[0] |= 0x00008000;
896 emit_add(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
897 struct nv50_reg
*src0
, struct nv50_reg
*src1
)
899 struct nv50_program_exec
*e
= exec(pc
);
901 e
->inst
[0] = 0xb0000000;
904 check_swap_src_0_1(pc
, &src0
, &src1
);
906 if (!pc
->allow32
|| (src0
->mod
| src1
->mod
) || src1
->hw
> 63) {
908 e
->inst
[1] |= ((src0
->mod
& NV50_MOD_NEG
) << 26) |
909 ((src1
->mod
& NV50_MOD_NEG
) << 27);
913 set_src_0(pc
, src0
, e
);
914 if (src1
->type
== P_CONST
|| src1
->type
== P_ATTR
|| is_long(e
))
915 set_src_2(pc
, src1
, e
);
917 if (src1
->type
== P_IMMD
)
918 set_immd(pc
, src1
, e
);
920 set_src_1(pc
, src1
, e
);
926 emit_arl(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
,
929 struct nv50_program_exec
*e
= exec(pc
);
932 e
->inst
[1] |= 0xc0000000;
934 e
->inst
[0] |= dst
->hw
<< 2;
935 e
->inst
[0] |= s
<< 16; /* shift left */
936 set_src_0_restricted(pc
, src
, e
);
942 emit_minmax(struct nv50_pc
*pc
, unsigned sub
, struct nv50_reg
*dst
,
943 struct nv50_reg
*src0
, struct nv50_reg
*src1
)
945 struct nv50_program_exec
*e
= exec(pc
);
948 e
->inst
[0] |= 0xb0000000;
949 e
->inst
[1] |= (sub
<< 29);
951 check_swap_src_0_1(pc
, &src0
, &src1
);
953 set_src_0(pc
, src0
, e
);
954 set_src_1(pc
, src1
, e
);
956 if (src0
->mod
& NV50_MOD_ABS
)
957 e
->inst
[1] |= 0x00100000;
958 if (src1
->mod
& NV50_MOD_ABS
)
959 e
->inst
[1] |= 0x00080000;
965 emit_sub(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
966 struct nv50_reg
*src1
)
968 src1
->mod
^= NV50_MOD_NEG
;
969 emit_add(pc
, dst
, src0
, src1
);
970 src1
->mod
^= NV50_MOD_NEG
;
974 emit_bitop2(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
975 struct nv50_reg
*src1
, unsigned op
)
977 struct nv50_program_exec
*e
= exec(pc
);
979 e
->inst
[0] = 0xd0000000;
982 check_swap_src_0_1(pc
, &src0
, &src1
);
984 set_src_0(pc
, src0
, e
);
986 if (op
!= TGSI_OPCODE_AND
&& op
!= TGSI_OPCODE_OR
&&
987 op
!= TGSI_OPCODE_XOR
)
988 assert(!"invalid bit op");
990 if (src1
->type
== P_IMMD
&& src0
->type
== P_TEMP
&& pc
->allow32
) {
991 set_immd(pc
, src1
, e
);
992 if (op
== TGSI_OPCODE_OR
)
993 e
->inst
[0] |= 0x0100;
995 if (op
== TGSI_OPCODE_XOR
)
996 e
->inst
[0] |= 0x8000;
998 set_src_1(pc
, src1
, e
);
999 e
->inst
[1] |= 0x04000000; /* 32 bit */
1000 if (op
== TGSI_OPCODE_OR
)
1001 e
->inst
[1] |= 0x4000;
1003 if (op
== TGSI_OPCODE_XOR
)
1004 e
->inst
[1] |= 0x8000;
1011 emit_mad(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
1012 struct nv50_reg
*src1
, struct nv50_reg
*src2
)
1014 struct nv50_program_exec
*e
= exec(pc
);
1016 e
->inst
[0] |= 0xe0000000;
1018 check_swap_src_0_1(pc
, &src0
, &src1
);
1019 set_dst(pc
, dst
, e
);
1020 set_src_0(pc
, src0
, e
);
1021 set_src_1(pc
, src1
, e
);
1022 set_src_2(pc
, src2
, e
);
1024 if ((src0
->mod
^ src1
->mod
) & NV50_MOD_NEG
)
1025 e
->inst
[1] |= 0x04000000;
1026 if (src2
->mod
& NV50_MOD_NEG
)
1027 e
->inst
[1] |= 0x08000000;
1033 emit_msb(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
1034 struct nv50_reg
*src1
, struct nv50_reg
*src2
)
1036 src2
->mod
^= NV50_MOD_NEG
;
1037 emit_mad(pc
, dst
, src0
, src1
, src2
);
1038 src2
->mod
^= NV50_MOD_NEG
;
1042 emit_flop(struct nv50_pc
*pc
, unsigned sub
,
1043 struct nv50_reg
*dst
, struct nv50_reg
*src
)
1045 struct nv50_program_exec
*e
= exec(pc
);
1047 e
->inst
[0] |= 0x90000000;
1050 e
->inst
[1] |= (sub
<< 29);
1053 set_dst(pc
, dst
, e
);
1055 if (sub
== 0 || sub
== 2)
1056 set_src_0_restricted(pc
, src
, e
);
1058 set_src_0(pc
, src
, e
);
1064 emit_preex2(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1066 struct nv50_program_exec
*e
= exec(pc
);
1068 e
->inst
[0] |= 0xb0000000;
1070 set_dst(pc
, dst
, e
);
1071 set_src_0(pc
, src
, e
);
1073 e
->inst
[1] |= (6 << 29) | 0x00004000;
1079 emit_precossin(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1081 struct nv50_program_exec
*e
= exec(pc
);
1083 e
->inst
[0] |= 0xb0000000;
1085 set_dst(pc
, dst
, e
);
1086 set_src_0(pc
, src
, e
);
1088 e
->inst
[1] |= (6 << 29);
1093 #define CVTOP_RN 0x01
1094 #define CVTOP_FLOOR 0x03
1095 #define CVTOP_CEIL 0x05
1096 #define CVTOP_TRUNC 0x07
1097 #define CVTOP_SAT 0x08
1098 #define CVTOP_ABS 0x10
1100 /* 0x04 == 32 bit dst */
1101 /* 0x40 == dst is float */
1102 /* 0x80 == src is float */
1103 #define CVT_F32_F32 0xc4
1104 #define CVT_F32_S32 0x44
1105 #define CVT_S32_F32 0x8c
1106 #define CVT_S32_S32 0x0c
1107 #define CVT_NEG 0x20
1111 emit_cvt(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
,
1112 int wp
, unsigned cvn
, unsigned fmt
)
1114 struct nv50_program_exec
*e
;
1119 e
->inst
[0] |= 0xa0000000;
1120 e
->inst
[1] |= 0x00004000; /* 32 bit src */
1121 e
->inst
[1] |= (cvn
<< 16);
1122 e
->inst
[1] |= (fmt
<< 24);
1123 set_src_0(pc
, src
, e
);
1126 set_pred_wr(pc
, 1, wp
, e
);
1129 set_dst(pc
, dst
, e
);
1131 e
->inst
[0] |= 0x000001fc;
1132 e
->inst
[1] |= 0x00000008;
1138 /* nv50 Condition codes:
1145 * 0x7 = set condition code ? (used before bra.lt/le/gt/ge)
1146 * 0x8 = unordered bit (allows NaN)
1149 emit_set(struct nv50_pc
*pc
, unsigned ccode
, struct nv50_reg
*dst
, int wp
,
1150 struct nv50_reg
*src0
, struct nv50_reg
*src1
)
1152 static const unsigned cc_swapped
[8] = { 0, 4, 2, 6, 1, 5, 3, 7 };
1154 struct nv50_program_exec
*e
= exec(pc
);
1155 struct nv50_reg
*rdst
;
1158 if (check_swap_src_0_1(pc
, &src0
, &src1
))
1159 ccode
= cc_swapped
[ccode
& 7] | (ccode
& 8);
1162 if (dst
&& dst
->type
!= P_TEMP
)
1163 dst
= alloc_temp(pc
, NULL
);
1167 e
->inst
[0] |= 0xb0000000;
1168 e
->inst
[1] |= 0x60000000 | (ccode
<< 14);
1170 /* XXX: decuda will disasm as .u16 and use .lo/.hi regs, but
1171 * that doesn't seem to match what the hw actually does
1172 e->inst[1] |= 0x04000000; << breaks things, u32 by default ?
1176 set_pred_wr(pc
, 1, wp
, e
);
1178 set_dst(pc
, dst
, e
);
1180 e
->inst
[0] |= 0x000001fc;
1181 e
->inst
[1] |= 0x00000008;
1184 set_src_0(pc
, src0
, e
);
1185 set_src_1(pc
, src1
, e
);
1188 pc
->if_cond
= pc
->p
->exec_tail
; /* record for OPCODE_IF */
1190 /* cvt.f32.u32/s32 (?) if we didn't only write the predicate */
1192 emit_cvt(pc
, rdst
, dst
, -1, CVTOP_ABS
| CVTOP_RN
, CVT_F32_S32
);
1193 if (rdst
&& rdst
!= dst
)
1197 static INLINE
unsigned
1198 map_tgsi_setop_cc(unsigned op
)
1201 case TGSI_OPCODE_SLT
: return 0x1;
1202 case TGSI_OPCODE_SGE
: return 0x6;
1203 case TGSI_OPCODE_SEQ
: return 0x2;
1204 case TGSI_OPCODE_SGT
: return 0x4;
1205 case TGSI_OPCODE_SLE
: return 0x3;
1206 case TGSI_OPCODE_SNE
: return 0xd;
1214 emit_flr(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1216 emit_cvt(pc
, dst
, src
, -1, CVTOP_FLOOR
, CVT_F32_F32
| CVT_RI
);
1220 emit_pow(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1221 struct nv50_reg
*v
, struct nv50_reg
*e
)
1223 struct nv50_reg
*temp
= alloc_temp(pc
, NULL
);
1225 emit_flop(pc
, 3, temp
, v
);
1226 emit_mul(pc
, temp
, temp
, e
);
1227 emit_preex2(pc
, temp
, temp
);
1228 emit_flop(pc
, 6, dst
, temp
);
1230 free_temp(pc
, temp
);
1234 emit_abs(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1236 emit_cvt(pc
, dst
, src
, -1, CVTOP_ABS
, CVT_F32_F32
);
1240 emit_sat(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1242 emit_cvt(pc
, dst
, src
, -1, CVTOP_SAT
, CVT_F32_F32
);
1246 emit_lit(struct nv50_pc
*pc
, struct nv50_reg
**dst
, unsigned mask
,
1247 struct nv50_reg
**src
)
1249 struct nv50_reg
*one
= alloc_immd(pc
, 1.0);
1250 struct nv50_reg
*zero
= alloc_immd(pc
, 0.0);
1251 struct nv50_reg
*neg128
= alloc_immd(pc
, -127.999999);
1252 struct nv50_reg
*pos128
= alloc_immd(pc
, 127.999999);
1253 struct nv50_reg
*tmp
[4];
1254 boolean allow32
= pc
->allow32
;
1256 pc
->allow32
= FALSE
;
1258 if (mask
& (3 << 1)) {
1259 tmp
[0] = alloc_temp(pc
, NULL
);
1260 emit_minmax(pc
, 4, tmp
[0], src
[0], zero
);
1263 if (mask
& (1 << 2)) {
1264 set_pred_wr(pc
, 1, 0, pc
->p
->exec_tail
);
1266 tmp
[1] = temp_temp(pc
);
1267 emit_minmax(pc
, 4, tmp
[1], src
[1], zero
);
1269 tmp
[3] = temp_temp(pc
);
1270 emit_minmax(pc
, 4, tmp
[3], src
[3], neg128
);
1271 emit_minmax(pc
, 5, tmp
[3], tmp
[3], pos128
);
1273 emit_pow(pc
, dst
[2], tmp
[1], tmp
[3]);
1274 emit_mov(pc
, dst
[2], zero
);
1275 set_pred(pc
, 3, 0, pc
->p
->exec_tail
);
1278 if (mask
& (1 << 1))
1279 assimilate_temp(pc
, dst
[1], tmp
[0]);
1281 if (mask
& (1 << 2))
1282 free_temp(pc
, tmp
[0]);
1284 pc
->allow32
= allow32
;
1286 /* do this last, in case src[i,j] == dst[0,3] */
1287 if (mask
& (1 << 0))
1288 emit_mov(pc
, dst
[0], one
);
1290 if (mask
& (1 << 3))
1291 emit_mov(pc
, dst
[3], one
);
1300 emit_neg(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1302 emit_cvt(pc
, dst
, src
, -1, CVTOP_RN
, CVT_F32_F32
| CVT_NEG
);
1306 emit_kil(struct nv50_pc
*pc
, struct nv50_reg
*src
)
1308 struct nv50_program_exec
*e
;
1309 const int r_pred
= 1;
1310 unsigned cvn
= CVT_F32_F32
;
1312 if (src
->mod
& NV50_MOD_NEG
)
1314 /* write predicate reg */
1315 emit_cvt(pc
, NULL
, src
, r_pred
, CVTOP_RN
, cvn
);
1317 /* conditional discard */
1319 e
->inst
[0] = 0x00000002;
1321 set_pred(pc
, 0x1 /* LT */, r_pred
, e
);
1325 static struct nv50_program_exec
*
1326 emit_branch(struct nv50_pc
*pc
, int pred
, unsigned cc
,
1327 struct nv50_program_exec
**join
)
1329 struct nv50_program_exec
*e
= exec(pc
);
1333 e
->inst
[0] |= 0xa0000002;
1340 e
->inst
[0] |= 0x10000002;
1342 set_pred(pc
, cc
, pred
, e
);
1344 return pc
->p
->exec_tail
;
1350 #define QOP_MOV_SRC1 3
1352 /* For a quad of threads / top left, top right, bottom left, bottom right
1353 * pixels, do a different operation, and take src0 from a specific thread.
1356 emit_quadop(struct nv50_pc
*pc
, struct nv50_reg
*dst
, int wp
, int lane_src0
,
1357 struct nv50_reg
*src0
, struct nv50_reg
*src1
, ubyte qop
)
1359 struct nv50_program_exec
*e
= exec(pc
);
1361 e
->inst
[0] = 0xc0000000;
1362 e
->inst
[1] = 0x80000000;
1364 e
->inst
[0] |= lane_src0
<< 16;
1365 set_src_0(pc
, src0
, e
);
1366 set_src_2(pc
, src1
, e
);
1369 set_pred_wr(pc
, 1, wp
, e
);
1372 set_dst(pc
, dst
, e
);
1374 e
->inst
[0] |= 0x000001fc;
1375 e
->inst
[1] |= 0x00000008;
1378 e
->inst
[0] |= (qop
& 3) << 20;
1379 e
->inst
[1] |= (qop
>> 2) << 22;
1385 load_cube_tex_coords(struct nv50_pc
*pc
, struct nv50_reg
*t
[4],
1386 struct nv50_reg
**src
, unsigned arg
, boolean proj
)
1388 int mod
[3] = { src
[0]->mod
, src
[1]->mod
, src
[2]->mod
};
1390 src
[0]->mod
|= NV50_MOD_ABS
;
1391 src
[1]->mod
|= NV50_MOD_ABS
;
1392 src
[2]->mod
|= NV50_MOD_ABS
;
1394 emit_minmax(pc
, 4, t
[2], src
[0], src
[1]);
1395 emit_minmax(pc
, 4, t
[2], src
[2], t
[2]);
1397 src
[0]->mod
= mod
[0];
1398 src
[1]->mod
= mod
[1];
1399 src
[2]->mod
= mod
[2];
1401 if (proj
&& 0 /* looks more correct without this */)
1402 emit_mul(pc
, t
[2], t
[2], src
[3]);
1404 if (arg
== 4) /* there is no textureProj(samplerCubeShadow) */
1405 emit_mov(pc
, t
[3], src
[3]);
1407 emit_flop(pc
, 0, t
[2], t
[2]);
1409 emit_mul(pc
, t
[0], src
[0], t
[2]);
1410 emit_mul(pc
, t
[1], src
[1], t
[2]);
1411 emit_mul(pc
, t
[2], src
[2], t
[2]);
1415 load_proj_tex_coords(struct nv50_pc
*pc
, struct nv50_reg
*t
[4],
1416 struct nv50_reg
**src
, unsigned dim
, unsigned arg
)
1420 if (src
[0]->type
== P_TEMP
&& src
[0]->rhw
!= -1) {
1421 mode
= pc
->interp_mode
[src
[0]->index
] | INTERP_PERSPECTIVE
;
1423 t
[3]->rhw
= src
[3]->rhw
;
1424 emit_interp(pc
, t
[3], NULL
, (mode
& INTERP_CENTROID
));
1425 emit_flop(pc
, 0, t
[3], t
[3]);
1427 for (c
= 0; c
< dim
; ++c
) {
1428 t
[c
]->rhw
= src
[c
]->rhw
;
1429 emit_interp(pc
, t
[c
], t
[3], mode
);
1431 if (arg
!= dim
) { /* depth reference value */
1432 t
[dim
]->rhw
= src
[2]->rhw
;
1433 emit_interp(pc
, t
[dim
], t
[3], mode
);
1436 /* XXX: for some reason the blob sometimes uses MAD
1437 * (mad f32 $rX $rY $rZ neg $r63)
1439 emit_flop(pc
, 0, t
[3], src
[3]);
1440 for (c
= 0; c
< dim
; ++c
)
1441 emit_mul(pc
, t
[c
], src
[c
], t
[3]);
1442 if (arg
!= dim
) /* depth reference value */
1443 emit_mul(pc
, t
[dim
], src
[2], t
[3]);
1448 get_tex_dim(unsigned type
, unsigned *dim
, unsigned *arg
)
1451 case TGSI_TEXTURE_1D
:
1454 case TGSI_TEXTURE_SHADOW1D
:
1458 case TGSI_TEXTURE_UNKNOWN
:
1459 case TGSI_TEXTURE_2D
:
1460 case TGSI_TEXTURE_RECT
:
1463 case TGSI_TEXTURE_SHADOW2D
:
1464 case TGSI_TEXTURE_SHADOWRECT
:
1468 case TGSI_TEXTURE_3D
:
1469 case TGSI_TEXTURE_CUBE
:
1478 /* We shouldn't execute TEXLOD if any of the pixels in a quad have
1479 * different LOD values, so branch off groups of equal LOD.
1482 emit_texlod_sequence(struct nv50_pc
*pc
, struct nv50_reg
*tlod
,
1483 struct nv50_reg
*src
, struct nv50_program_exec
*tex
)
1485 struct nv50_program_exec
*join_at
;
1486 unsigned i
, target
= pc
->p
->exec_size
+ 7 * 2;
1488 /* Subtract lod of each pixel from lod of top left pixel, jump
1489 * texlod insn if result is 0, then repeat for 2 other pixels.
1491 emit_quadop(pc
, NULL
, 0, 0, tlod
, tlod
, 0x55);
1492 emit_branch(pc
, 0, 2, &join_at
)->param
.index
= target
;
1494 for (i
= 1; i
< 4; ++i
) {
1495 emit_quadop(pc
, NULL
, 0, i
, tlod
, tlod
, 0x55);
1496 emit_branch(pc
, 0, 2, NULL
)->param
.index
= target
;
1499 emit_mov(pc
, tlod
, src
); /* target */
1500 emit(pc
, tex
); /* texlod */
1502 join_at
->param
.index
= target
+ 2 * 2;
1504 pc
->p
->exec_tail
->inst
[1] |= 2; /* join _after_ tex */
1508 emit_texbias_sequence(struct nv50_pc
*pc
, struct nv50_reg
*t
[4], unsigned arg
,
1509 struct nv50_program_exec
*tex
)
1511 struct nv50_program_exec
*e
;
1512 struct nv50_reg imm_1248
, *t123
[4][4], *r_bits
= alloc_temp(pc
, NULL
);
1514 unsigned n
, c
, i
, cc
[4] = { 0x0a, 0x13, 0x11, 0x10 };
1516 pc
->allow32
= FALSE
;
1517 ctor_reg(&imm_1248
, P_IMMD
, -1, ctor_immd_4u32(pc
, 1, 2, 4, 8) * 4);
1519 /* Subtract bias value of thread i from bias values of each thread,
1520 * store result in r_pred, and set bit i in r_bits if result was 0.
1523 for (i
= 0; i
< 4; ++i
, ++imm_1248
.hw
) {
1524 emit_quadop(pc
, NULL
, r_pred
, i
, t
[arg
], t
[arg
], 0x55);
1525 emit_mov(pc
, r_bits
, &imm_1248
);
1526 set_pred(pc
, 2, r_pred
, pc
->p
->exec_tail
);
1528 emit_mov_to_pred(pc
, r_pred
, r_bits
);
1530 /* The lanes of a quad are now grouped by the bit in r_pred they have
1531 * set. Put the input values for TEX into a new register set for each
1532 * group and execute TEX only for a specific group.
1533 * We cannot use the same register set for each group because we need
1534 * the derivatives, which are implicitly calculated, to be correct.
1536 for (i
= 1; i
< 4; ++i
) {
1537 alloc_temp4(pc
, t123
[i
], 0);
1539 for (c
= 0; c
<= arg
; ++c
)
1540 emit_mov(pc
, t123
[i
][c
], t
[c
]);
1542 *(e
= exec(pc
)) = *(tex
);
1543 e
->inst
[0] &= ~0x01fc;
1544 set_dst(pc
, t123
[i
][0], e
);
1545 set_pred(pc
, cc
[i
], r_pred
, e
);
1548 /* finally TEX on the original regs (where we kept the input) */
1549 set_pred(pc
, cc
[0], r_pred
, tex
);
1552 /* put the 3 * n other results into regs for lane 0 */
1553 n
= popcnt4(((e
->inst
[0] >> 25) & 0x3) | ((e
->inst
[1] >> 12) & 0xc));
1554 for (i
= 1; i
< 4; ++i
) {
1555 for (c
= 0; c
< n
; ++c
) {
1556 emit_mov(pc
, t
[c
], t123
[i
][c
]);
1557 set_pred(pc
, cc
[i
], r_pred
, pc
->p
->exec_tail
);
1559 free_temp4(pc
, t123
[i
]);
1563 free_temp(pc
, r_bits
);
1567 emit_tex(struct nv50_pc
*pc
, struct nv50_reg
**dst
, unsigned mask
,
1568 struct nv50_reg
**src
, unsigned unit
, unsigned type
,
1569 boolean proj
, int bias_lod
)
1571 struct nv50_reg
*t
[4];
1572 struct nv50_program_exec
*e
;
1573 unsigned c
, dim
, arg
;
1575 /* t[i] must be within a single 128 bit super-reg */
1576 alloc_temp4(pc
, t
, 0);
1579 e
->inst
[0] = 0xf0000000;
1581 set_dst(pc
, t
[0], e
);
1583 /* TIC and TSC binding indices (TSC is ignored as TSC_LINKED = TRUE): */
1584 e
->inst
[0] |= (unit
<< 9) /* | (unit << 17) */;
1586 /* live flag (don't set if TEX results affect input to another TEX): */
1587 /* e->inst[0] |= 0x00000004; */
1589 get_tex_dim(type
, &dim
, &arg
);
1591 if (type
== TGSI_TEXTURE_CUBE
) {
1592 e
->inst
[0] |= 0x08000000;
1593 load_cube_tex_coords(pc
, t
, src
, arg
, proj
);
1596 load_proj_tex_coords(pc
, t
, src
, dim
, arg
);
1598 for (c
= 0; c
< dim
; c
++)
1599 emit_mov(pc
, t
[c
], src
[c
]);
1600 if (arg
!= dim
) /* depth reference value (always src.z here) */
1601 emit_mov(pc
, t
[dim
], src
[2]);
1604 e
->inst
[0] |= (mask
& 0x3) << 25;
1605 e
->inst
[1] |= (mask
& 0xc) << 12;
1608 e
->inst
[0] |= (arg
- 1) << 22;
1612 e
->inst
[0] |= arg
<< 22;
1613 e
->inst
[1] |= 0x20000000; /* texbias */
1614 emit_mov(pc
, t
[arg
], src
[3]);
1615 emit_texbias_sequence(pc
, t
, arg
, e
);
1617 e
->inst
[0] |= arg
<< 22;
1618 e
->inst
[1] |= 0x40000000; /* texlod */
1619 emit_mov(pc
, t
[arg
], src
[3]);
1620 emit_texlod_sequence(pc
, t
[arg
], src
[3], e
);
1625 if (mask
& 1) emit_mov(pc
, dst
[0], t
[c
++]);
1626 if (mask
& 2) emit_mov(pc
, dst
[1], t
[c
++]);
1627 if (mask
& 4) emit_mov(pc
, dst
[2], t
[c
++]);
1628 if (mask
& 8) emit_mov(pc
, dst
[3], t
[c
]);
1632 /* XXX: if p.e. MUL is used directly after TEX, it would still use
1633 * the texture coordinates, not the fetched values: latency ? */
1635 for (c
= 0; c
< 4; c
++) {
1636 if (mask
& (1 << c
))
1637 assimilate_temp(pc
, dst
[c
], t
[c
]);
1639 free_temp(pc
, t
[c
]);
1645 emit_ddx(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1647 struct nv50_program_exec
*e
= exec(pc
);
1649 assert(src
->type
== P_TEMP
);
1651 e
->inst
[0] = 0xc0140000;
1652 e
->inst
[1] = 0x89800000;
1654 set_dst(pc
, dst
, e
);
1655 set_src_0(pc
, src
, e
);
1656 set_src_2(pc
, src
, e
);
1662 emit_ddy(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1664 struct nv50_reg
*r
= src
;
1665 struct nv50_program_exec
*e
= exec(pc
);
1667 assert(src
->type
== P_TEMP
);
1669 if (!(src
->mod
& NV50_MOD_NEG
)) { /* ! double negation */
1670 r
= alloc_temp(pc
, NULL
);
1671 emit_neg(pc
, r
, src
);
1674 e
->inst
[0] = 0xc0150000;
1675 e
->inst
[1] = 0x8a400000;
1677 set_dst(pc
, dst
, e
);
1678 set_src_0(pc
, r
, e
);
1679 set_src_2(pc
, r
, e
);
1688 convert_to_long(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
1690 unsigned q
= 0, m
= ~0;
1692 assert(!is_long(e
));
1694 switch (e
->inst
[0] >> 28) {
1701 /* INTERP (move centroid, perspective and flat bits) */
1703 q
= (e
->inst
[0] & (3 << 24)) >> (24 - 16);
1704 q
|= (e
->inst
[0] & (1 << 8)) << (18 - 8);
1712 q
= ((e
->inst
[0] & (~m
)) >> 2);
1717 q
= ((e
->inst
[0] & (~m
)) << 12);
1720 /* MAD (if src2 == dst) */
1721 q
= ((e
->inst
[0] & 0x1fc) << 12);
1735 /* Some operations support an optional negation flag. */
1737 negate_supported(const struct tgsi_full_instruction
*insn
, int i
)
1739 switch (insn
->Instruction
.Opcode
) {
1740 case TGSI_OPCODE_DDY
:
1741 case TGSI_OPCODE_DP3
:
1742 case TGSI_OPCODE_DP4
:
1743 case TGSI_OPCODE_MUL
:
1744 case TGSI_OPCODE_KIL
:
1745 case TGSI_OPCODE_ADD
:
1746 case TGSI_OPCODE_SUB
:
1747 case TGSI_OPCODE_MAD
:
1749 case TGSI_OPCODE_POW
:
1758 /* Return a read mask for source registers deduced from opcode & write mask. */
1760 nv50_tgsi_src_mask(const struct tgsi_full_instruction
*insn
, int c
)
1762 unsigned x
, mask
= insn
->Dst
[0].Register
.WriteMask
;
1764 switch (insn
->Instruction
.Opcode
) {
1765 case TGSI_OPCODE_COS
:
1766 case TGSI_OPCODE_SIN
:
1767 return (mask
& 0x8) | ((mask
& 0x7) ? 0x1 : 0x0);
1768 case TGSI_OPCODE_DP3
:
1770 case TGSI_OPCODE_DP4
:
1771 case TGSI_OPCODE_DPH
:
1772 case TGSI_OPCODE_KIL
: /* WriteMask ignored */
1774 case TGSI_OPCODE_DST
:
1775 return mask
& (c
? 0xa : 0x6);
1776 case TGSI_OPCODE_EX2
:
1777 case TGSI_OPCODE_LG2
:
1778 case TGSI_OPCODE_POW
:
1779 case TGSI_OPCODE_RCP
:
1780 case TGSI_OPCODE_RSQ
:
1781 case TGSI_OPCODE_SCS
:
1783 case TGSI_OPCODE_IF
:
1785 case TGSI_OPCODE_LIT
:
1787 case TGSI_OPCODE_TEX
:
1788 case TGSI_OPCODE_TXB
:
1789 case TGSI_OPCODE_TXL
:
1790 case TGSI_OPCODE_TXP
:
1792 const struct tgsi_instruction_texture
*tex
;
1794 assert(insn
->Instruction
.Texture
);
1795 tex
= &insn
->Texture
;
1798 if (insn
->Instruction
.Opcode
!= TGSI_OPCODE_TEX
&&
1799 insn
->Instruction
.Opcode
!= TGSI_OPCODE_TXD
)
1800 mask
|= 0x8; /* bias, lod or proj */
1802 switch (tex
->Texture
) {
1803 case TGSI_TEXTURE_1D
:
1806 case TGSI_TEXTURE_SHADOW1D
:
1809 case TGSI_TEXTURE_2D
:
1817 case TGSI_OPCODE_XPD
:
1819 if (mask
& 1) x
|= 0x6;
1820 if (mask
& 2) x
|= 0x5;
1821 if (mask
& 4) x
|= 0x3;
1830 static struct nv50_reg
*
1831 tgsi_dst(struct nv50_pc
*pc
, int c
, const struct tgsi_full_dst_register
*dst
)
1833 switch (dst
->Register
.File
) {
1834 case TGSI_FILE_TEMPORARY
:
1835 return &pc
->temp
[dst
->Register
.Index
* 4 + c
];
1836 case TGSI_FILE_OUTPUT
:
1837 return &pc
->result
[dst
->Register
.Index
* 4 + c
];
1838 case TGSI_FILE_ADDRESS
:
1840 struct nv50_reg
*r
= pc
->addr
[dst
->Register
.Index
* 4 + c
];
1842 r
= alloc_addr(pc
, NULL
);
1843 pc
->addr
[dst
->Register
.Index
* 4 + c
] = r
;
1848 case TGSI_FILE_NULL
:
1857 static struct nv50_reg
*
1858 tgsi_src(struct nv50_pc
*pc
, int chan
, const struct tgsi_full_src_register
*src
,
1861 struct nv50_reg
*r
= NULL
;
1862 struct nv50_reg
*temp
;
1863 unsigned sgn
, c
, swz
;
1865 if (src
->Register
.File
!= TGSI_FILE_CONSTANT
)
1866 assert(!src
->Register
.Indirect
);
1868 sgn
= tgsi_util_get_full_src_register_sign_mode(src
, chan
);
1870 c
= tgsi_util_get_full_src_register_swizzle(src
, chan
);
1872 case TGSI_SWIZZLE_X
:
1873 case TGSI_SWIZZLE_Y
:
1874 case TGSI_SWIZZLE_Z
:
1875 case TGSI_SWIZZLE_W
:
1876 switch (src
->Register
.File
) {
1877 case TGSI_FILE_INPUT
:
1878 r
= &pc
->attr
[src
->Register
.Index
* 4 + c
];
1880 case TGSI_FILE_TEMPORARY
:
1881 r
= &pc
->temp
[src
->Register
.Index
* 4 + c
];
1883 case TGSI_FILE_CONSTANT
:
1884 if (!src
->Register
.Indirect
) {
1885 r
= &pc
->param
[src
->Register
.Index
* 4 + c
];
1888 /* Indicate indirection by setting r->acc < 0 and
1889 * use the index field to select the address reg.
1891 r
= MALLOC_STRUCT(nv50_reg
);
1892 swz
= tgsi_util_get_src_register_swizzle(
1894 ctor_reg(r
, P_CONST
,
1895 src
->Indirect
.Index
* 4 + swz
,
1896 src
->Register
.Index
* 4 + c
);
1899 case TGSI_FILE_IMMEDIATE
:
1900 r
= &pc
->immd
[src
->Register
.Index
* 4 + c
];
1902 case TGSI_FILE_SAMPLER
:
1904 case TGSI_FILE_ADDRESS
:
1905 r
= pc
->addr
[src
->Register
.Index
* 4 + c
];
1919 case TGSI_UTIL_SIGN_KEEP
:
1921 case TGSI_UTIL_SIGN_CLEAR
:
1922 temp
= temp_temp(pc
);
1923 emit_abs(pc
, temp
, r
);
1926 case TGSI_UTIL_SIGN_TOGGLE
:
1928 r
->mod
= NV50_MOD_NEG
;
1930 temp
= temp_temp(pc
);
1931 emit_neg(pc
, temp
, r
);
1935 case TGSI_UTIL_SIGN_SET
:
1936 temp
= temp_temp(pc
);
1937 emit_cvt(pc
, temp
, r
, -1, CVTOP_ABS
, CVT_F32_F32
| CVT_NEG
);
1948 /* return TRUE for ops that produce only a single result */
1950 is_scalar_op(unsigned op
)
1953 case TGSI_OPCODE_COS
:
1954 case TGSI_OPCODE_DP2
:
1955 case TGSI_OPCODE_DP3
:
1956 case TGSI_OPCODE_DP4
:
1957 case TGSI_OPCODE_DPH
:
1958 case TGSI_OPCODE_EX2
:
1959 case TGSI_OPCODE_LG2
:
1960 case TGSI_OPCODE_POW
:
1961 case TGSI_OPCODE_RCP
:
1962 case TGSI_OPCODE_RSQ
:
1963 case TGSI_OPCODE_SIN
:
1965 case TGSI_OPCODE_KIL:
1966 case TGSI_OPCODE_LIT:
1967 case TGSI_OPCODE_SCS:
1975 /* Returns a bitmask indicating which dst components depend
1976 * on source s, component c (reverse of nv50_tgsi_src_mask).
1979 nv50_tgsi_dst_revdep(unsigned op
, int s
, int c
)
1981 if (is_scalar_op(op
))
1985 case TGSI_OPCODE_DST
:
1986 return (1 << c
) & (s
? 0xa : 0x6);
1987 case TGSI_OPCODE_XPD
:
1997 case TGSI_OPCODE_LIT
:
1998 case TGSI_OPCODE_SCS
:
1999 case TGSI_OPCODE_TEX
:
2000 case TGSI_OPCODE_TXB
:
2001 case TGSI_OPCODE_TXL
:
2002 case TGSI_OPCODE_TXP
:
2003 /* these take care of dangerous swizzles themselves */
2005 case TGSI_OPCODE_IF
:
2006 case TGSI_OPCODE_KIL
:
2007 /* don't call this function for these ops */
2011 /* linear vector instruction */
2016 static INLINE boolean
2017 has_pred(struct nv50_program_exec
*e
, unsigned cc
)
2019 if (!is_long(e
) || is_immd(e
))
2021 return ((e
->inst
[1] & 0x780) == (cc
<< 7));
2024 /* on ENDIF see if we can do "@p0.neu single_op" instead of:
2031 nv50_kill_branch(struct nv50_pc
*pc
)
2033 int lvl
= pc
->if_lvl
;
2035 if (pc
->if_insn
[lvl
]->next
!= pc
->p
->exec_tail
)
2038 /* if ccode == 'true', the BRA is from an ELSE and the predicate
2039 * reg may no longer be valid, since we currently always use $p0
2041 if (has_pred(pc
->if_insn
[lvl
], 0xf))
2043 assert(pc
->if_insn
[lvl
] && pc
->br_join
[lvl
]);
2045 /* We'll use the exec allocated for JOIN_AT (as we can't easily
2046 * update prev's next); if exec_tail is BRK, update the pointer.
2048 if (pc
->loop_lvl
&& pc
->br_loop
[pc
->loop_lvl
- 1] == pc
->p
->exec_tail
)
2049 pc
->br_loop
[pc
->loop_lvl
- 1] = pc
->br_join
[lvl
];
2051 pc
->p
->exec_size
-= 4; /* remove JOIN_AT and BRA */
2053 *pc
->br_join
[lvl
] = *pc
->p
->exec_tail
;
2055 FREE(pc
->if_insn
[lvl
]);
2056 FREE(pc
->p
->exec_tail
);
2058 pc
->p
->exec_tail
= pc
->br_join
[lvl
];
2059 pc
->p
->exec_tail
->next
= NULL
;
2060 set_pred(pc
, 0xd, 0, pc
->p
->exec_tail
);
2066 nv50_program_tx_insn(struct nv50_pc
*pc
,
2067 const struct tgsi_full_instruction
*inst
)
2069 struct nv50_reg
*rdst
[4], *dst
[4], *brdc
, *src
[3][4], *temp
;
2070 unsigned mask
, sat
, unit
;
2073 mask
= inst
->Dst
[0].Register
.WriteMask
;
2074 sat
= inst
->Instruction
.Saturate
== TGSI_SAT_ZERO_ONE
;
2076 memset(src
, 0, sizeof(src
));
2078 for (c
= 0; c
< 4; c
++) {
2079 if ((mask
& (1 << c
)) && !pc
->r_dst
[c
])
2080 dst
[c
] = tgsi_dst(pc
, c
, &inst
->Dst
[0]);
2082 dst
[c
] = pc
->r_dst
[c
];
2086 for (i
= 0; i
< inst
->Instruction
.NumSrcRegs
; i
++) {
2087 const struct tgsi_full_src_register
*fs
= &inst
->Src
[i
];
2091 src_mask
= nv50_tgsi_src_mask(inst
, i
);
2092 neg_supp
= negate_supported(inst
, i
);
2094 if (fs
->Register
.File
== TGSI_FILE_SAMPLER
)
2095 unit
= fs
->Register
.Index
;
2097 for (c
= 0; c
< 4; c
++)
2098 if (src_mask
& (1 << c
))
2099 src
[i
][c
] = reg_instance(pc
,
2100 tgsi_src(pc
, c
, fs
, neg_supp
));
2103 brdc
= temp
= pc
->r_brdc
;
2104 if (brdc
&& brdc
->type
!= P_TEMP
) {
2105 temp
= temp_temp(pc
);
2110 for (c
= 0; c
< 4; c
++) {
2111 if (!(mask
& (1 << c
)) || dst
[c
]->type
== P_TEMP
)
2113 /* rdst[c] = dst[c]; */ /* done above */
2114 dst
[c
] = temp_temp(pc
);
2118 assert(brdc
|| !is_scalar_op(inst
->Instruction
.Opcode
));
2120 switch (inst
->Instruction
.Opcode
) {
2121 case TGSI_OPCODE_ABS
:
2122 for (c
= 0; c
< 4; c
++) {
2123 if (!(mask
& (1 << c
)))
2125 emit_abs(pc
, dst
[c
], src
[0][c
]);
2128 case TGSI_OPCODE_ADD
:
2129 for (c
= 0; c
< 4; c
++) {
2130 if (!(mask
& (1 << c
)))
2132 emit_add(pc
, dst
[c
], src
[0][c
], src
[1][c
]);
2135 case TGSI_OPCODE_AND
:
2136 case TGSI_OPCODE_XOR
:
2137 case TGSI_OPCODE_OR
:
2138 for (c
= 0; c
< 4; c
++) {
2139 if (!(mask
& (1 << c
)))
2141 emit_bitop2(pc
, dst
[c
], src
[0][c
], src
[1][c
],
2142 inst
->Instruction
.Opcode
);
2145 case TGSI_OPCODE_ARL
:
2147 temp
= temp_temp(pc
);
2148 emit_cvt(pc
, temp
, src
[0][0], -1, CVTOP_FLOOR
, CVT_S32_F32
);
2149 emit_arl(pc
, dst
[0], temp
, 4);
2151 case TGSI_OPCODE_BGNLOOP
:
2152 pc
->loop_pos
[pc
->loop_lvl
++] = pc
->p
->exec_size
;
2155 case TGSI_OPCODE_BRK
:
2156 emit_branch(pc
, -1, 0, NULL
);
2157 assert(pc
->loop_lvl
> 0);
2158 pc
->br_loop
[pc
->loop_lvl
- 1] = pc
->p
->exec_tail
;
2160 case TGSI_OPCODE_CEIL
:
2161 for (c
= 0; c
< 4; c
++) {
2162 if (!(mask
& (1 << c
)))
2164 emit_cvt(pc
, dst
[c
], src
[0][c
], -1,
2165 CVTOP_CEIL
, CVT_F32_F32
| CVT_RI
);
2168 case TGSI_OPCODE_CMP
:
2169 pc
->allow32
= FALSE
;
2170 for (c
= 0; c
< 4; c
++) {
2171 if (!(mask
& (1 << c
)))
2173 emit_cvt(pc
, NULL
, src
[0][c
], 1, CVTOP_RN
, CVT_F32_F32
);
2174 emit_mov(pc
, dst
[c
], src
[1][c
]);
2175 set_pred(pc
, 0x1, 1, pc
->p
->exec_tail
); /* @SF */
2176 emit_mov(pc
, dst
[c
], src
[2][c
]);
2177 set_pred(pc
, 0x6, 1, pc
->p
->exec_tail
); /* @NSF */
2180 case TGSI_OPCODE_COS
:
2182 emit_precossin(pc
, temp
, src
[0][3]);
2183 emit_flop(pc
, 5, dst
[3], temp
);
2187 temp
= brdc
= temp_temp(pc
);
2189 emit_precossin(pc
, temp
, src
[0][0]);
2190 emit_flop(pc
, 5, brdc
, temp
);
2192 case TGSI_OPCODE_DDX
:
2193 for (c
= 0; c
< 4; c
++) {
2194 if (!(mask
& (1 << c
)))
2196 emit_ddx(pc
, dst
[c
], src
[0][c
]);
2199 case TGSI_OPCODE_DDY
:
2200 for (c
= 0; c
< 4; c
++) {
2201 if (!(mask
& (1 << c
)))
2203 emit_ddy(pc
, dst
[c
], src
[0][c
]);
2206 case TGSI_OPCODE_DP3
:
2207 emit_mul(pc
, temp
, src
[0][0], src
[1][0]);
2208 emit_mad(pc
, temp
, src
[0][1], src
[1][1], temp
);
2209 emit_mad(pc
, brdc
, src
[0][2], src
[1][2], temp
);
2211 case TGSI_OPCODE_DP4
:
2212 emit_mul(pc
, temp
, src
[0][0], src
[1][0]);
2213 emit_mad(pc
, temp
, src
[0][1], src
[1][1], temp
);
2214 emit_mad(pc
, temp
, src
[0][2], src
[1][2], temp
);
2215 emit_mad(pc
, brdc
, src
[0][3], src
[1][3], temp
);
2217 case TGSI_OPCODE_DPH
:
2218 emit_mul(pc
, temp
, src
[0][0], src
[1][0]);
2219 emit_mad(pc
, temp
, src
[0][1], src
[1][1], temp
);
2220 emit_mad(pc
, temp
, src
[0][2], src
[1][2], temp
);
2221 emit_add(pc
, brdc
, src
[1][3], temp
);
2223 case TGSI_OPCODE_DST
:
2224 if (mask
& (1 << 1))
2225 emit_mul(pc
, dst
[1], src
[0][1], src
[1][1]);
2226 if (mask
& (1 << 2))
2227 emit_mov(pc
, dst
[2], src
[0][2]);
2228 if (mask
& (1 << 3))
2229 emit_mov(pc
, dst
[3], src
[1][3]);
2230 if (mask
& (1 << 0))
2231 emit_mov_immdval(pc
, dst
[0], 1.0f
);
2233 case TGSI_OPCODE_ELSE
:
2234 emit_branch(pc
, -1, 0, NULL
);
2235 pc
->if_insn
[--pc
->if_lvl
]->param
.index
= pc
->p
->exec_size
;
2236 pc
->if_insn
[pc
->if_lvl
++] = pc
->p
->exec_tail
;
2239 case TGSI_OPCODE_ENDIF
:
2240 pc
->if_insn
[--pc
->if_lvl
]->param
.index
= pc
->p
->exec_size
;
2242 /* try to replace branch over 1 insn with a predicated insn */
2243 if (nv50_kill_branch(pc
) == TRUE
)
2246 if (pc
->br_join
[pc
->if_lvl
]) {
2247 pc
->br_join
[pc
->if_lvl
]->param
.index
= pc
->p
->exec_size
;
2248 pc
->br_join
[pc
->if_lvl
] = NULL
;
2251 /* emit a NOP as join point, we could set it on the next
2252 * one, but would have to make sure it is long and !immd
2255 pc
->p
->exec_tail
->inst
[1] |= 2;
2257 case TGSI_OPCODE_ENDLOOP
:
2258 emit_branch(pc
, -1, 0, NULL
);
2259 pc
->p
->exec_tail
->param
.index
= pc
->loop_pos
[--pc
->loop_lvl
];
2260 pc
->br_loop
[pc
->loop_lvl
]->param
.index
= pc
->p
->exec_size
;
2263 case TGSI_OPCODE_EX2
:
2264 emit_preex2(pc
, temp
, src
[0][0]);
2265 emit_flop(pc
, 6, brdc
, temp
);
2267 case TGSI_OPCODE_FLR
:
2268 for (c
= 0; c
< 4; c
++) {
2269 if (!(mask
& (1 << c
)))
2271 emit_flr(pc
, dst
[c
], src
[0][c
]);
2274 case TGSI_OPCODE_FRC
:
2275 temp
= temp_temp(pc
);
2276 for (c
= 0; c
< 4; c
++) {
2277 if (!(mask
& (1 << c
)))
2279 emit_flr(pc
, temp
, src
[0][c
]);
2280 emit_sub(pc
, dst
[c
], src
[0][c
], temp
);
2283 case TGSI_OPCODE_IF
:
2284 /* emitting a join_at may not be necessary */
2285 assert(pc
->if_lvl
< MAX_IF_DEPTH
);
2286 /* set_pred_wr(pc, 1, 0, pc->if_cond); */
2287 emit_cvt(pc
, NULL
, src
[0][0], 0, CVTOP_ABS
| CVTOP_RN
,
2289 emit_branch(pc
, 0, 2, &pc
->br_join
[pc
->if_lvl
]);
2290 pc
->if_insn
[pc
->if_lvl
++] = pc
->p
->exec_tail
;
2293 case TGSI_OPCODE_KIL
:
2294 emit_kil(pc
, src
[0][0]);
2295 emit_kil(pc
, src
[0][1]);
2296 emit_kil(pc
, src
[0][2]);
2297 emit_kil(pc
, src
[0][3]);
2299 case TGSI_OPCODE_LIT
:
2300 emit_lit(pc
, &dst
[0], mask
, &src
[0][0]);
2302 case TGSI_OPCODE_LG2
:
2303 emit_flop(pc
, 3, brdc
, src
[0][0]);
2305 case TGSI_OPCODE_LRP
:
2306 temp
= temp_temp(pc
);
2307 for (c
= 0; c
< 4; c
++) {
2308 if (!(mask
& (1 << c
)))
2310 emit_sub(pc
, temp
, src
[1][c
], src
[2][c
]);
2311 emit_mad(pc
, dst
[c
], temp
, src
[0][c
], src
[2][c
]);
2314 case TGSI_OPCODE_MAD
:
2315 for (c
= 0; c
< 4; c
++) {
2316 if (!(mask
& (1 << c
)))
2318 emit_mad(pc
, dst
[c
], src
[0][c
], src
[1][c
], src
[2][c
]);
2321 case TGSI_OPCODE_MAX
:
2322 for (c
= 0; c
< 4; c
++) {
2323 if (!(mask
& (1 << c
)))
2325 emit_minmax(pc
, 4, dst
[c
], src
[0][c
], src
[1][c
]);
2328 case TGSI_OPCODE_MIN
:
2329 for (c
= 0; c
< 4; c
++) {
2330 if (!(mask
& (1 << c
)))
2332 emit_minmax(pc
, 5, dst
[c
], src
[0][c
], src
[1][c
]);
2335 case TGSI_OPCODE_MOV
:
2336 for (c
= 0; c
< 4; c
++) {
2337 if (!(mask
& (1 << c
)))
2339 emit_mov(pc
, dst
[c
], src
[0][c
]);
2342 case TGSI_OPCODE_MUL
:
2343 for (c
= 0; c
< 4; c
++) {
2344 if (!(mask
& (1 << c
)))
2346 emit_mul(pc
, dst
[c
], src
[0][c
], src
[1][c
]);
2349 case TGSI_OPCODE_POW
:
2350 emit_pow(pc
, brdc
, src
[0][0], src
[1][0]);
2352 case TGSI_OPCODE_RCP
:
2353 emit_flop(pc
, 0, brdc
, src
[0][0]);
2355 case TGSI_OPCODE_RSQ
:
2356 emit_flop(pc
, 2, brdc
, src
[0][0]);
2358 case TGSI_OPCODE_SCS
:
2359 temp
= temp_temp(pc
);
2361 emit_precossin(pc
, temp
, src
[0][0]);
2362 if (mask
& (1 << 0))
2363 emit_flop(pc
, 5, dst
[0], temp
);
2364 if (mask
& (1 << 1))
2365 emit_flop(pc
, 4, dst
[1], temp
);
2366 if (mask
& (1 << 2))
2367 emit_mov_immdval(pc
, dst
[2], 0.0);
2368 if (mask
& (1 << 3))
2369 emit_mov_immdval(pc
, dst
[3], 1.0);
2371 case TGSI_OPCODE_SIN
:
2373 emit_precossin(pc
, temp
, src
[0][3]);
2374 emit_flop(pc
, 4, dst
[3], temp
);
2378 temp
= brdc
= temp_temp(pc
);
2380 emit_precossin(pc
, temp
, src
[0][0]);
2381 emit_flop(pc
, 4, brdc
, temp
);
2383 case TGSI_OPCODE_SLT
:
2384 case TGSI_OPCODE_SGE
:
2385 case TGSI_OPCODE_SEQ
:
2386 case TGSI_OPCODE_SGT
:
2387 case TGSI_OPCODE_SLE
:
2388 case TGSI_OPCODE_SNE
:
2389 i
= map_tgsi_setop_cc(inst
->Instruction
.Opcode
);
2390 for (c
= 0; c
< 4; c
++) {
2391 if (!(mask
& (1 << c
)))
2393 emit_set(pc
, i
, dst
[c
], -1, src
[0][c
], src
[1][c
]);
2396 case TGSI_OPCODE_SUB
:
2397 for (c
= 0; c
< 4; c
++) {
2398 if (!(mask
& (1 << c
)))
2400 emit_sub(pc
, dst
[c
], src
[0][c
], src
[1][c
]);
2403 case TGSI_OPCODE_TEX
:
2404 emit_tex(pc
, dst
, mask
, src
[0], unit
,
2405 inst
->Texture
.Texture
, FALSE
, 0);
2407 case TGSI_OPCODE_TXB
:
2408 emit_tex(pc
, dst
, mask
, src
[0], unit
,
2409 inst
->Texture
.Texture
, FALSE
, -1);
2411 case TGSI_OPCODE_TXL
:
2412 emit_tex(pc
, dst
, mask
, src
[0], unit
,
2413 inst
->Texture
.Texture
, FALSE
, 1);
2415 case TGSI_OPCODE_TXP
:
2416 emit_tex(pc
, dst
, mask
, src
[0], unit
,
2417 inst
->Texture
.Texture
, TRUE
, 0);
2419 case TGSI_OPCODE_TRUNC
:
2420 for (c
= 0; c
< 4; c
++) {
2421 if (!(mask
& (1 << c
)))
2423 emit_cvt(pc
, dst
[c
], src
[0][c
], -1,
2424 CVTOP_TRUNC
, CVT_F32_F32
| CVT_RI
);
2427 case TGSI_OPCODE_XPD
:
2428 temp
= temp_temp(pc
);
2429 if (mask
& (1 << 0)) {
2430 emit_mul(pc
, temp
, src
[0][2], src
[1][1]);
2431 emit_msb(pc
, dst
[0], src
[0][1], src
[1][2], temp
);
2433 if (mask
& (1 << 1)) {
2434 emit_mul(pc
, temp
, src
[0][0], src
[1][2]);
2435 emit_msb(pc
, dst
[1], src
[0][2], src
[1][0], temp
);
2437 if (mask
& (1 << 2)) {
2438 emit_mul(pc
, temp
, src
[0][1], src
[1][0]);
2439 emit_msb(pc
, dst
[2], src
[0][0], src
[1][1], temp
);
2441 if (mask
& (1 << 3))
2442 emit_mov_immdval(pc
, dst
[3], 1.0);
2444 case TGSI_OPCODE_END
:
2447 NOUVEAU_ERR("invalid opcode %d\n", inst
->Instruction
.Opcode
);
2453 emit_sat(pc
, brdc
, brdc
);
2454 for (c
= 0; c
< 4; c
++)
2455 if ((mask
& (1 << c
)) && dst
[c
] != brdc
)
2456 emit_mov(pc
, dst
[c
], brdc
);
2459 for (c
= 0; c
< 4; c
++) {
2460 if (!(mask
& (1 << c
)))
2462 /* In this case we saturate later, and dst[c] won't
2463 * be another temp_temp (and thus lost), since rdst
2464 * already is TEMP (see above). */
2465 if (rdst
[c
]->type
== P_TEMP
&& rdst
[c
]->index
< 0)
2467 emit_sat(pc
, rdst
[c
], dst
[c
]);
2471 for (i
= 0; i
< inst
->Instruction
.NumSrcRegs
; i
++) {
2472 for (c
= 0; c
< 4; c
++) {
2475 if (src
[i
][c
]->acc
< 0 && src
[i
][c
]->type
== P_CONST
)
2476 FREE(src
[i
][c
]); /* indirect constant */
2481 pc
->reg_instance_nr
= 0;
2487 prep_inspect_insn(struct nv50_pc
*pc
, const struct tgsi_full_instruction
*insn
)
2489 struct nv50_reg
*reg
= NULL
;
2490 const struct tgsi_full_src_register
*src
;
2491 const struct tgsi_dst_register
*dst
;
2492 unsigned i
, c
, k
, mask
;
2494 dst
= &insn
->Dst
[0].Register
;
2495 mask
= dst
->WriteMask
;
2497 if (dst
->File
== TGSI_FILE_TEMPORARY
)
2500 if (dst
->File
== TGSI_FILE_OUTPUT
)
2504 for (c
= 0; c
< 4; c
++) {
2505 if (!(mask
& (1 << c
)))
2507 reg
[dst
->Index
* 4 + c
].acc
= pc
->insn_nr
;
2511 for (i
= 0; i
< insn
->Instruction
.NumSrcRegs
; i
++) {
2512 src
= &insn
->Src
[i
];
2514 if (src
->Register
.File
== TGSI_FILE_TEMPORARY
)
2517 if (src
->Register
.File
== TGSI_FILE_INPUT
)
2522 mask
= nv50_tgsi_src_mask(insn
, i
);
2524 for (c
= 0; c
< 4; c
++) {
2525 if (!(mask
& (1 << c
)))
2527 k
= tgsi_util_get_full_src_register_swizzle(src
, c
);
2529 reg
[src
->Register
.Index
* 4 + k
].acc
= pc
->insn_nr
;
2534 /* Returns a bitmask indicating which dst components need to be
2535 * written to temporaries first to avoid 'corrupting' sources.
2537 * m[i] (out) indicate component to write in the i-th position
2538 * rdep[c] (in) bitmasks of dst[i] that require dst[c] as source
2541 nv50_revdep_reorder(unsigned m
[4], unsigned rdep
[4])
2543 unsigned i
, c
, x
, unsafe
;
2545 for (c
= 0; c
< 4; c
++)
2548 /* Swap as long as a dst component written earlier is depended on
2549 * by one written later, but the next one isn't depended on by it.
2551 for (c
= 0; c
< 3; c
++) {
2552 if (rdep
[m
[c
+ 1]] & (1 << m
[c
]))
2553 continue; /* if next one is depended on by us */
2554 for (i
= c
+ 1; i
< 4; i
++)
2555 /* if we are depended on by a later one */
2556 if (rdep
[m
[c
]] & (1 << m
[i
]))
2569 /* mark dependencies that could not be resolved by reordering */
2570 for (i
= 0; i
< 3; ++i
)
2571 for (c
= i
+ 1; c
< 4; ++c
)
2572 if (rdep
[m
[i
]] & (1 << m
[c
]))
2575 /* NOTE: $unsafe is with respect to order, not component */
2579 /* Select a suitable dst register for broadcasting scalar results,
2580 * or return NULL if we have to allocate an extra TEMP.
2582 * If e.g. only 1 component is written, we may also emit the final
2583 * result to a write-only register.
2585 static struct nv50_reg
*
2586 tgsi_broadcast_dst(struct nv50_pc
*pc
,
2587 const struct tgsi_full_dst_register
*fd
, unsigned mask
)
2589 if (fd
->Register
.File
== TGSI_FILE_TEMPORARY
) {
2590 int c
= ffs(~mask
& fd
->Register
.WriteMask
);
2592 return tgsi_dst(pc
, c
- 1, fd
);
2594 int c
= ffs(fd
->Register
.WriteMask
) - 1;
2595 if ((1 << c
) == fd
->Register
.WriteMask
)
2596 return tgsi_dst(pc
, c
, fd
);
2602 /* Scan source swizzles and return a bitmask indicating dst regs that
2603 * also occur among the src regs, and fill rdep for nv50_revdep_reoder.
2606 nv50_tgsi_scan_swizzle(const struct tgsi_full_instruction
*insn
,
2609 const struct tgsi_full_dst_register
*fd
= &insn
->Dst
[0];
2610 const struct tgsi_full_src_register
*fs
;
2611 unsigned i
, deqs
= 0;
2613 for (i
= 0; i
< 4; ++i
)
2616 for (i
= 0; i
< insn
->Instruction
.NumSrcRegs
; i
++) {
2617 unsigned chn
, mask
= nv50_tgsi_src_mask(insn
, i
);
2618 boolean neg_supp
= negate_supported(insn
, i
);
2621 if (fs
->Register
.File
!= fd
->Register
.File
||
2622 fs
->Register
.Index
!= fd
->Register
.Index
)
2625 for (chn
= 0; chn
< 4; ++chn
) {
2628 if (!(mask
& (1 << chn
))) /* src is not read */
2630 c
= tgsi_util_get_full_src_register_swizzle(fs
, chn
);
2631 s
= tgsi_util_get_full_src_register_sign_mode(fs
, chn
);
2633 if (!(fd
->Register
.WriteMask
& (1 << c
)))
2636 /* no danger if src is copied to TEMP first */
2637 if ((s
!= TGSI_UTIL_SIGN_KEEP
) &&
2638 (s
!= TGSI_UTIL_SIGN_TOGGLE
|| !neg_supp
))
2641 rdep
[c
] |= nv50_tgsi_dst_revdep(
2642 insn
->Instruction
.Opcode
, i
, chn
);
2651 nv50_tgsi_insn(struct nv50_pc
*pc
, const union tgsi_full_token
*tok
)
2653 struct tgsi_full_instruction insn
= tok
->FullInstruction
;
2654 const struct tgsi_full_dst_register
*fd
;
2655 unsigned i
, deqs
, rdep
[4], m
[4];
2657 fd
= &tok
->FullInstruction
.Dst
[0];
2658 deqs
= nv50_tgsi_scan_swizzle(&insn
, rdep
);
2660 if (is_scalar_op(insn
.Instruction
.Opcode
)) {
2661 pc
->r_brdc
= tgsi_broadcast_dst(pc
, fd
, deqs
);
2663 pc
->r_brdc
= temp_temp(pc
);
2664 return nv50_program_tx_insn(pc
, &insn
);
2669 return nv50_program_tx_insn(pc
, &insn
);
2671 deqs
= nv50_revdep_reorder(m
, rdep
);
2673 for (i
= 0; i
< 4; ++i
) {
2674 assert(pc
->r_dst
[m
[i
]] == NULL
);
2676 insn
.Dst
[0].Register
.WriteMask
=
2677 fd
->Register
.WriteMask
& (1 << m
[i
]);
2679 if (!insn
.Dst
[0].Register
.WriteMask
)
2682 if (deqs
& (1 << i
))
2683 pc
->r_dst
[m
[i
]] = alloc_temp(pc
, NULL
);
2685 if (!nv50_program_tx_insn(pc
, &insn
))
2689 for (i
= 0; i
< 4; i
++) {
2690 struct nv50_reg
*reg
= pc
->r_dst
[i
];
2693 pc
->r_dst
[i
] = NULL
;
2695 if (insn
.Instruction
.Saturate
== TGSI_SAT_ZERO_ONE
)
2696 emit_sat(pc
, tgsi_dst(pc
, i
, fd
), reg
);
2698 emit_mov(pc
, tgsi_dst(pc
, i
, fd
), reg
);
2706 load_interpolant(struct nv50_pc
*pc
, struct nv50_reg
*reg
)
2708 struct nv50_reg
*iv
, **ppiv
;
2709 unsigned mode
= pc
->interp_mode
[reg
->index
];
2711 ppiv
= (mode
& INTERP_CENTROID
) ? &pc
->iv_c
: &pc
->iv_p
;
2714 if ((mode
& INTERP_PERSPECTIVE
) && !iv
) {
2715 iv
= *ppiv
= alloc_temp(pc
, NULL
);
2716 iv
->rhw
= popcnt4(pc
->p
->cfg
.regs
[1] >> 24) - 1;
2718 emit_interp(pc
, iv
, NULL
, mode
& INTERP_CENTROID
);
2719 emit_flop(pc
, 0, iv
, iv
);
2721 /* XXX: when loading interpolants dynamically, move these
2722 * to the program head, or make sure it can't be skipped.
2726 emit_interp(pc
, reg
, iv
, mode
);
2729 /* The face input is always at v[255] (varying space), with a
2730 * value of 0 for back-facing, and 0xffffffff for front-facing.
2733 load_frontfacing(struct nv50_pc
*pc
, struct nv50_reg
*a
)
2735 struct nv50_reg
*one
= alloc_immd(pc
, 1.0f
);
2737 assert(a
->rhw
== -1);
2738 alloc_reg(pc
, a
); /* do this before rhw is set */
2740 load_interpolant(pc
, a
);
2741 emit_bitop2(pc
, a
, a
, one
, TGSI_OPCODE_AND
);
2747 nv50_program_tx_prep(struct nv50_pc
*pc
)
2749 struct tgsi_parse_context tp
;
2750 struct nv50_program
*p
= pc
->p
;
2751 boolean ret
= FALSE
;
2752 unsigned i
, c
, flat_nr
= 0;
2754 tgsi_parse_init(&tp
, pc
->p
->pipe
.tokens
);
2755 while (!tgsi_parse_end_of_tokens(&tp
)) {
2756 const union tgsi_full_token
*tok
= &tp
.FullToken
;
2758 tgsi_parse_token(&tp
);
2759 switch (tok
->Token
.Type
) {
2760 case TGSI_TOKEN_TYPE_IMMEDIATE
:
2762 const struct tgsi_full_immediate
*imm
=
2763 &tp
.FullToken
.FullImmediate
;
2765 ctor_immd(pc
, imm
->u
[0].Float
,
2771 case TGSI_TOKEN_TYPE_DECLARATION
:
2773 const struct tgsi_full_declaration
*d
;
2774 unsigned si
, last
, first
, mode
;
2776 d
= &tp
.FullToken
.FullDeclaration
;
2777 first
= d
->Range
.First
;
2778 last
= d
->Range
.Last
;
2780 switch (d
->Declaration
.File
) {
2781 case TGSI_FILE_TEMPORARY
:
2783 case TGSI_FILE_OUTPUT
:
2784 if (!d
->Declaration
.Semantic
||
2785 p
->type
== PIPE_SHADER_FRAGMENT
)
2788 si
= d
->Semantic
.Index
;
2789 switch (d
->Semantic
.Name
) {
2790 case TGSI_SEMANTIC_BCOLOR
:
2791 p
->cfg
.two_side
[si
].hw
= first
;
2792 if (p
->cfg
.io_nr
> first
)
2793 p
->cfg
.io_nr
= first
;
2795 case TGSI_SEMANTIC_PSIZE
:
2796 p
->cfg
.psiz
= first
;
2797 if (p
->cfg
.io_nr
> first
)
2798 p
->cfg
.io_nr
= first
;
2801 case TGSI_SEMANTIC_CLIP_DISTANCE:
2802 p->cfg.clpd = MIN2(p->cfg.clpd, first);
2809 case TGSI_FILE_INPUT
:
2811 if (p
->type
!= PIPE_SHADER_FRAGMENT
)
2814 switch (d
->Declaration
.Interpolate
) {
2815 case TGSI_INTERPOLATE_CONSTANT
:
2819 case TGSI_INTERPOLATE_PERSPECTIVE
:
2820 mode
= INTERP_PERSPECTIVE
;
2821 p
->cfg
.regs
[1] |= 0x08 << 24;
2824 mode
= INTERP_LINEAR
;
2827 if (d
->Declaration
.Centroid
)
2828 mode
|= INTERP_CENTROID
;
2831 for (i
= first
; i
<= last
; i
++)
2832 pc
->interp_mode
[i
] = mode
;
2835 case TGSI_FILE_ADDRESS
:
2836 case TGSI_FILE_CONSTANT
:
2837 case TGSI_FILE_SAMPLER
:
2840 NOUVEAU_ERR("bad decl file %d\n",
2841 d
->Declaration
.File
);
2846 case TGSI_TOKEN_TYPE_INSTRUCTION
:
2848 prep_inspect_insn(pc
, &tok
->FullInstruction
);
2855 if (p
->type
== PIPE_SHADER_VERTEX
) {
2858 for (i
= 0; i
< pc
->attr_nr
* 4; ++i
) {
2859 if (pc
->attr
[i
].acc
) {
2860 pc
->attr
[i
].hw
= rid
++;
2861 p
->cfg
.attr
[i
/ 32] |= 1 << (i
% 32);
2865 for (i
= 0, rid
= 0; i
< pc
->result_nr
; ++i
) {
2866 p
->cfg
.io
[i
].hw
= rid
;
2867 p
->cfg
.io
[i
].id
= i
;
2869 for (c
= 0; c
< 4; ++c
) {
2871 if (!pc
->result
[n
].acc
)
2873 pc
->result
[n
].hw
= rid
++;
2874 p
->cfg
.io
[i
].mask
|= 1 << c
;
2878 for (c
= 0; c
< 2; ++c
)
2879 if (p
->cfg
.two_side
[c
].hw
< 0x40)
2880 p
->cfg
.two_side
[c
] = p
->cfg
.io
[
2881 p
->cfg
.two_side
[c
].hw
];
2883 if (p
->cfg
.psiz
< 0x40)
2884 p
->cfg
.psiz
= p
->cfg
.io
[p
->cfg
.psiz
].hw
;
2886 if (p
->type
== PIPE_SHADER_FRAGMENT
) {
2888 unsigned n
= 0, m
= pc
->attr_nr
- flat_nr
;
2892 int base
= (TGSI_SEMANTIC_POSITION
==
2893 p
->info
.input_semantic_name
[0]) ? 0 : 1;
2895 /* non-flat interpolants have to be mapped to
2896 * the lower hardware IDs, so sort them:
2898 for (i
= 0; i
< pc
->attr_nr
; i
++) {
2899 if (pc
->interp_mode
[i
] == INTERP_FLAT
)
2900 p
->cfg
.io
[m
++].id
= i
;
2902 if (!(pc
->interp_mode
[i
] & INTERP_PERSPECTIVE
))
2903 p
->cfg
.io
[n
].linear
= TRUE
;
2904 p
->cfg
.io
[n
++].id
= i
;
2908 if (!base
) /* set w-coordinate mask from perspective interp */
2909 p
->cfg
.io
[0].mask
|= p
->cfg
.regs
[1] >> 24;
2911 aid
= popcnt4( /* if fcrd isn't contained in cfg.io */
2912 base
? (p
->cfg
.regs
[1] >> 24) : p
->cfg
.io
[0].mask
);
2914 for (n
= 0; n
< pc
->attr_nr
; ++n
) {
2915 p
->cfg
.io
[n
].hw
= rid
= aid
;
2916 i
= p
->cfg
.io
[n
].id
;
2918 if (p
->info
.input_semantic_name
[n
] ==
2919 TGSI_SEMANTIC_FACE
) {
2920 load_frontfacing(pc
, &pc
->attr
[i
* 4]);
2924 for (c
= 0; c
< 4; ++c
) {
2925 if (!pc
->attr
[i
* 4 + c
].acc
)
2927 pc
->attr
[i
* 4 + c
].rhw
= rid
++;
2928 p
->cfg
.io
[n
].mask
|= 1 << c
;
2930 load_interpolant(pc
, &pc
->attr
[i
* 4 + c
]);
2932 aid
+= popcnt4(p
->cfg
.io
[n
].mask
);
2936 p
->cfg
.regs
[1] |= p
->cfg
.io
[0].mask
<< 24;
2938 m
= popcnt4(p
->cfg
.regs
[1] >> 24);
2940 /* set count of non-position inputs and of non-flat
2941 * non-position inputs for FP_INTERPOLANT_CTRL
2943 p
->cfg
.regs
[1] |= aid
- m
;
2946 i
= p
->cfg
.io
[pc
->attr_nr
- flat_nr
].hw
;
2947 p
->cfg
.regs
[1] |= (i
- m
) << 16;
2949 p
->cfg
.regs
[1] |= p
->cfg
.regs
[1] << 16;
2951 /* mark color semantic for light-twoside */
2953 for (i
= 0; i
< pc
->attr_nr
; i
++) {
2956 sn
= p
->info
.input_semantic_name
[p
->cfg
.io
[i
].id
];
2957 si
= p
->info
.input_semantic_index
[p
->cfg
.io
[i
].id
];
2959 if (sn
== TGSI_SEMANTIC_COLOR
) {
2960 p
->cfg
.two_side
[si
] = p
->cfg
.io
[i
];
2962 /* increase colour count */
2963 p
->cfg
.regs
[0] += popcnt4(
2964 p
->cfg
.two_side
[si
].mask
) << 16;
2966 n
= MIN2(n
, p
->cfg
.io
[i
].hw
- m
);
2970 p
->cfg
.regs
[0] += n
;
2972 /* Initialize FP results:
2973 * FragDepth is always first TGSI and last hw output
2975 i
= p
->info
.writes_z
? 4 : 0;
2976 for (rid
= 0; i
< pc
->result_nr
* 4; i
++)
2977 pc
->result
[i
].rhw
= rid
++;
2978 if (p
->info
.writes_z
)
2979 pc
->result
[2].rhw
= rid
;
2981 p
->cfg
.high_result
= rid
;
2983 /* separate/different colour results for MRTs ? */
2984 if (pc
->result_nr
- (p
->info
.writes_z
? 1 : 0) > 1)
2985 p
->cfg
.regs
[2] |= 1;
2991 pc
->immd
= MALLOC(pc
->immd_nr
* 4 * sizeof(struct nv50_reg
));
2995 for (i
= 0; i
< pc
->immd_nr
; i
++) {
2996 for (c
= 0; c
< 4; c
++, rid
++)
2997 ctor_reg(&pc
->immd
[rid
], P_IMMD
, i
, rid
);
3004 free_temp(pc
, pc
->iv_p
);
3006 free_temp(pc
, pc
->iv_c
);
3008 tgsi_parse_free(&tp
);
3013 free_nv50_pc(struct nv50_pc
*pc
)
3030 ctor_nv50_pc(struct nv50_pc
*pc
, struct nv50_program
*p
)
3033 unsigned rtype
[2] = { P_ATTR
, P_RESULT
};
3036 pc
->temp_nr
= p
->info
.file_max
[TGSI_FILE_TEMPORARY
] + 1;
3037 pc
->attr_nr
= p
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
3038 pc
->result_nr
= p
->info
.file_max
[TGSI_FILE_OUTPUT
] + 1;
3039 pc
->param_nr
= p
->info
.file_max
[TGSI_FILE_CONSTANT
] + 1;
3040 pc
->addr_nr
= p
->info
.file_max
[TGSI_FILE_ADDRESS
] + 1;
3041 assert(pc
->addr_nr
<= 2);
3043 p
->cfg
.high_temp
= 4;
3045 p
->cfg
.two_side
[0].hw
= 0x40;
3046 p
->cfg
.two_side
[1].hw
= 0x40;
3049 case PIPE_SHADER_VERTEX
:
3052 p
->cfg
.io_nr
= pc
->result_nr
;
3054 case PIPE_SHADER_FRAGMENT
:
3055 rtype
[0] = rtype
[1] = P_TEMP
;
3057 p
->cfg
.regs
[0] = 0x01000004;
3058 p
->cfg
.io_nr
= pc
->attr_nr
;
3060 if (p
->info
.writes_z
) {
3061 p
->cfg
.regs
[2] |= 0x00000100;
3062 p
->cfg
.regs
[3] |= 0x00000011;
3064 if (p
->info
.uses_kill
)
3065 p
->cfg
.regs
[2] |= 0x00100000;
3070 pc
->temp
= MALLOC(pc
->temp_nr
* 4 * sizeof(struct nv50_reg
));
3074 for (i
= 0; i
< pc
->temp_nr
* 4; ++i
)
3075 ctor_reg(&pc
->temp
[i
], P_TEMP
, i
/ 4, -1);
3079 pc
->attr
= MALLOC(pc
->attr_nr
* 4 * sizeof(struct nv50_reg
));
3083 for (i
= 0; i
< pc
->attr_nr
* 4; ++i
)
3084 ctor_reg(&pc
->attr
[i
], rtype
[0], i
/ 4, -1);
3087 if (pc
->result_nr
) {
3088 unsigned nr
= pc
->result_nr
* 4;
3090 pc
->result
= MALLOC(nr
* sizeof(struct nv50_reg
));
3094 for (i
= 0; i
< nr
; ++i
)
3095 ctor_reg(&pc
->result
[i
], rtype
[1], i
/ 4, -1);
3101 pc
->param
= MALLOC(pc
->param_nr
* 4 * sizeof(struct nv50_reg
));
3105 for (i
= 0; i
< pc
->param_nr
; ++i
)
3106 for (c
= 0; c
< 4; ++c
, ++rid
)
3107 ctor_reg(&pc
->param
[rid
], P_CONST
, i
, rid
);
3111 pc
->addr
= CALLOC(pc
->addr_nr
* 4, sizeof(struct nv50_reg
*));
3115 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
)
3116 ctor_reg(&pc
->r_addr
[i
], P_ADDR
, -256, i
+ 1);
3122 nv50_fp_move_results(struct nv50_pc
*pc
)
3124 struct nv50_reg reg
;
3127 ctor_reg(®
, P_TEMP
, -1, -1);
3129 for (i
= 0; i
< pc
->result_nr
* 4; ++i
) {
3130 if (pc
->result
[i
].rhw
< 0 || pc
->result
[i
].hw
< 0)
3132 if (pc
->result
[i
].rhw
!= pc
->result
[i
].hw
) {
3133 reg
.hw
= pc
->result
[i
].rhw
;
3134 emit_mov(pc
, ®
, &pc
->result
[i
]);
3140 nv50_program_fixup_insns(struct nv50_pc
*pc
)
3142 struct nv50_program_exec
*e
, **bra_list
;
3145 bra_list
= CALLOC(pc
->p
->exec_size
, sizeof(struct nv50_program_exec
*));
3147 /* Collect branch instructions, we need to adjust their offsets
3148 * when converting 32 bit instructions to 64 bit ones
3150 for (n
= 0, e
= pc
->p
->exec_head
; e
; e
= e
->next
)
3151 if (e
->param
.index
>= 0 && !e
->param
.mask
)
3154 /* last instruction must be long so it can have the exit bit set */
3155 if (!is_long(pc
->p
->exec_tail
))
3156 convert_to_long(pc
, pc
->p
->exec_tail
);
3158 pc
->p
->exec_tail
->inst
[1] |= 1;
3160 /* !immd on exit insn simultaneously means !join */
3161 assert(!is_immd(pc
->p
->exec_head
));
3162 assert(!is_immd(pc
->p
->exec_tail
));
3164 /* Make sure we don't have any single 32 bit instructions. */
3165 for (e
= pc
->p
->exec_head
, pos
= 0; e
; e
= e
->next
) {
3166 pos
+= is_long(e
) ? 2 : 1;
3168 if ((pos
& 1) && (!e
->next
|| is_long(e
->next
))) {
3169 for (i
= 0; i
< n
; ++i
)
3170 if (bra_list
[i
]->param
.index
>= pos
)
3171 bra_list
[i
]->param
.index
+= 1;
3172 convert_to_long(pc
, e
);
3181 nv50_program_tx(struct nv50_program
*p
)
3183 struct tgsi_parse_context parse
;
3187 pc
= CALLOC_STRUCT(nv50_pc
);
3191 ret
= ctor_nv50_pc(pc
, p
);
3195 ret
= nv50_program_tx_prep(pc
);
3199 tgsi_parse_init(&parse
, pc
->p
->pipe
.tokens
);
3200 while (!tgsi_parse_end_of_tokens(&parse
)) {
3201 const union tgsi_full_token
*tok
= &parse
.FullToken
;
3203 /* don't allow half insn/immd on first and last instruction */
3205 if (pc
->insn_cur
== 0 || pc
->insn_cur
+ 2 == pc
->insn_nr
)
3206 pc
->allow32
= FALSE
;
3208 tgsi_parse_token(&parse
);
3210 switch (tok
->Token
.Type
) {
3211 case TGSI_TOKEN_TYPE_INSTRUCTION
:
3213 ret
= nv50_tgsi_insn(pc
, tok
);
3222 if (pc
->p
->type
== PIPE_SHADER_FRAGMENT
)
3223 nv50_fp_move_results(pc
);
3225 nv50_program_fixup_insns(pc
);
3227 p
->param_nr
= pc
->param_nr
* 4;
3228 p
->immd_nr
= pc
->immd_nr
* 4;
3229 p
->immd
= pc
->immd_buf
;
3232 tgsi_parse_free(&parse
);
3240 nv50_program_validate(struct nv50_context
*nv50
, struct nv50_program
*p
)
3242 if (nv50_program_tx(p
) == FALSE
)
3244 p
->translated
= TRUE
;
3248 nv50_program_upload_data(struct nv50_context
*nv50
, float *map
,
3249 unsigned start
, unsigned count
, unsigned cbuf
)
3251 struct nouveau_channel
*chan
= nv50
->screen
->base
.channel
;
3252 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
3255 unsigned nr
= count
> 2047 ? 2047 : count
;
3257 BEGIN_RING(chan
, tesla
, NV50TCL_CB_ADDR
, 1);
3258 OUT_RING (chan
, (cbuf
<< 0) | (start
<< 8));
3259 BEGIN_RING(chan
, tesla
, NV50TCL_CB_DATA(0) | 0x40000000, nr
);
3260 OUT_RINGp (chan
, map
, nr
);
3269 nv50_program_validate_data(struct nv50_context
*nv50
, struct nv50_program
*p
)
3271 struct pipe_screen
*pscreen
= nv50
->pipe
.screen
;
3273 if (!p
->data
[0] && p
->immd_nr
) {
3274 struct nouveau_resource
*heap
= nv50
->screen
->immd_heap
[0];
3276 if (nouveau_resource_alloc(heap
, p
->immd_nr
, p
, &p
->data
[0])) {
3277 while (heap
->next
&& heap
->size
< p
->immd_nr
) {
3278 struct nv50_program
*evict
= heap
->next
->priv
;
3279 nouveau_resource_free(&evict
->data
[0]);
3282 if (nouveau_resource_alloc(heap
, p
->immd_nr
, p
,
3287 /* immediates only need to be uploaded again when freed */
3288 nv50_program_upload_data(nv50
, p
->immd
, p
->data
[0]->start
,
3289 p
->immd_nr
, NV50_CB_PMISC
);
3292 assert(p
->param_nr
<= 512);
3296 float *map
= pipe_buffer_map(pscreen
, nv50
->constbuf
[p
->type
],
3297 PIPE_BUFFER_USAGE_CPU_READ
);
3299 if (p
->type
== PIPE_SHADER_VERTEX
)
3304 nv50_program_upload_data(nv50
, map
, 0, p
->param_nr
, cb
);
3305 pipe_buffer_unmap(pscreen
, nv50
->constbuf
[p
->type
]);
3310 nv50_program_validate_code(struct nv50_context
*nv50
, struct nv50_program
*p
)
3312 struct nouveau_channel
*chan
= nv50
->screen
->base
.channel
;
3313 struct nv50_program_exec
*e
;
3315 boolean upload
= FALSE
;
3318 nouveau_bo_new(chan
->device
, NOUVEAU_BO_VRAM
, 0x100,
3319 p
->exec_size
* 4, &p
->bo
);
3323 if (p
->data
[0] && p
->data
[0]->start
!= p
->data_start
[0])
3329 up
= MALLOC(p
->exec_size
* 4);
3331 for (i
= 0, e
= p
->exec_head
; e
; e
= e
->next
) {
3332 unsigned ei
, ci
, bs
;
3334 if (e
->param
.index
>= 0 && e
->param
.mask
) {
3335 bs
= (e
->inst
[1] >> 22) & 0x07;
3337 ei
= e
->param
.shift
>> 5;
3338 ci
= e
->param
.index
;
3340 ci
+= p
->data
[bs
]->start
;
3342 e
->inst
[ei
] &= ~e
->param
.mask
;
3343 e
->inst
[ei
] |= (ci
<< e
->param
.shift
);
3345 if (e
->param
.index
>= 0) {
3346 /* zero mask means param is a jump/branch offset */
3347 assert(!(e
->param
.index
& 1));
3348 /* seem to be 8 byte steps */
3349 ei
= (e
->param
.index
>> 1) + 0 /* START_ID */;
3351 e
->inst
[0] &= 0xf0000fff;
3352 e
->inst
[0] |= ei
<< 12;
3355 up
[i
++] = e
->inst
[0];
3357 up
[i
++] = e
->inst
[1];
3359 assert(i
== p
->exec_size
);
3362 p
->data_start
[0] = p
->data
[0]->start
;
3364 #ifdef NV50_PROGRAM_DUMP
3365 NOUVEAU_ERR("-------\n");
3366 for (e
= p
->exec_head
; e
; e
= e
->next
) {
3367 NOUVEAU_ERR("0x%08x\n", e
->inst
[0]);
3369 NOUVEAU_ERR("0x%08x\n", e
->inst
[1]);
3372 nv50_upload_sifc(nv50
, p
->bo
, 0, NOUVEAU_BO_VRAM
,
3373 NV50_2D_DST_FORMAT_R8_UNORM
, 65536, 1, 262144,
3374 up
, NV50_2D_SIFC_FORMAT_R8_UNORM
, 0,
3375 0, 0, p
->exec_size
* 4, 1, 1);
3381 nv50_vertprog_validate(struct nv50_context
*nv50
)
3383 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
3384 struct nv50_program
*p
= nv50
->vertprog
;
3385 struct nouveau_stateobj
*so
;
3387 if (!p
->translated
) {
3388 nv50_program_validate(nv50
, p
);
3393 nv50_program_validate_data(nv50
, p
);
3394 nv50_program_validate_code(nv50
, p
);
3397 so_method(so
, tesla
, NV50TCL_VP_ADDRESS_HIGH
, 2);
3398 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
3399 NOUVEAU_BO_HIGH
, 0, 0);
3400 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
3401 NOUVEAU_BO_LOW
, 0, 0);
3402 so_method(so
, tesla
, NV50TCL_VP_ATTR_EN_0
, 2);
3403 so_data (so
, p
->cfg
.attr
[0]);
3404 so_data (so
, p
->cfg
.attr
[1]);
3405 so_method(so
, tesla
, NV50TCL_VP_REG_ALLOC_RESULT
, 1);
3406 so_data (so
, p
->cfg
.high_result
);
3407 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP_SIZE
, 2);
3408 so_data (so
, p
->cfg
.high_result
); //8);
3409 so_data (so
, p
->cfg
.high_temp
);
3410 so_method(so
, tesla
, NV50TCL_VP_START_ID
, 1);
3411 so_data (so
, 0); /* program start offset */
3412 so_ref(so
, &nv50
->state
.vertprog
);
3417 nv50_fragprog_validate(struct nv50_context
*nv50
)
3419 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
3420 struct nv50_program
*p
= nv50
->fragprog
;
3421 struct nouveau_stateobj
*so
;
3423 if (!p
->translated
) {
3424 nv50_program_validate(nv50
, p
);
3429 nv50_program_validate_data(nv50
, p
);
3430 nv50_program_validate_code(nv50
, p
);
3433 so_method(so
, tesla
, NV50TCL_FP_ADDRESS_HIGH
, 2);
3434 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
3435 NOUVEAU_BO_HIGH
, 0, 0);
3436 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
3437 NOUVEAU_BO_LOW
, 0, 0);
3438 so_method(so
, tesla
, NV50TCL_FP_REG_ALLOC_TEMP
, 1);
3439 so_data (so
, p
->cfg
.high_temp
);
3440 so_method(so
, tesla
, NV50TCL_FP_RESULT_COUNT
, 1);
3441 so_data (so
, p
->cfg
.high_result
);
3442 so_method(so
, tesla
, NV50TCL_FP_CTRL_UNK19A8
, 1);
3443 so_data (so
, p
->cfg
.regs
[2]);
3444 so_method(so
, tesla
, NV50TCL_FP_CTRL_UNK196C
, 1);
3445 so_data (so
, p
->cfg
.regs
[3]);
3446 so_method(so
, tesla
, NV50TCL_FP_START_ID
, 1);
3447 so_data (so
, 0); /* program start offset */
3448 so_ref(so
, &nv50
->state
.fragprog
);
3453 nv50_pntc_replace(struct nv50_context
*nv50
, uint32_t pntc
[8], unsigned base
)
3455 struct nv50_program
*fp
= nv50
->fragprog
;
3456 struct nv50_program
*vp
= nv50
->vertprog
;
3457 unsigned i
, c
, m
= base
;
3459 /* XXX: this might not work correctly in all cases yet - we'll
3460 * just assume that an FP generic input that is not written in
3461 * the VP is PointCoord.
3463 memset(pntc
, 0, 8 * sizeof(uint32_t));
3465 for (i
= 0; i
< fp
->cfg
.io_nr
; i
++) {
3467 uint8_t j
, k
= fp
->cfg
.io
[i
].id
;
3468 unsigned n
= popcnt4(fp
->cfg
.io
[i
].mask
);
3470 if (fp
->info
.input_semantic_name
[k
] != TGSI_SEMANTIC_GENERIC
) {
3475 for (j
= 0; j
< vp
->info
.num_outputs
; ++j
) {
3476 sn
= vp
->info
.output_semantic_name
[j
];
3477 si
= vp
->info
.output_semantic_index
[j
];
3479 if (sn
== fp
->info
.input_semantic_name
[k
] &&
3480 si
== fp
->info
.input_semantic_index
[k
])
3484 if (j
< vp
->info
.num_outputs
) {
3486 nv50
->rasterizer
->pipe
.sprite_coord_mode
[si
];
3488 if (mode
== PIPE_SPRITE_COORD_NONE
) {
3494 /* this is either PointCoord or replaced by sprite coords */
3495 for (c
= 0; c
< 4; c
++) {
3496 if (!(fp
->cfg
.io
[i
].mask
& (1 << c
)))
3498 pntc
[m
/ 8] |= (c
+ 1) << ((m
% 8) * 4);
3505 nv50_sreg4_map(uint32_t *p_map
, int mid
, uint32_t lin
[4],
3506 struct nv50_sreg4
*fpi
, struct nv50_sreg4
*vpo
)
3509 uint8_t mv
= vpo
->mask
, mf
= fpi
->mask
, oid
= vpo
->hw
;
3510 uint8_t *map
= (uint8_t *)p_map
;
3512 for (c
= 0; c
< 4; ++c
) {
3514 if (fpi
->linear
== TRUE
)
3515 lin
[mid
/ 32] |= 1 << (mid
% 32);
3516 map
[mid
++] = (mv
& 1) ? oid
: ((c
== 3) ? 0x41 : 0x40);
3528 nv50_linkage_validate(struct nv50_context
*nv50
)
3530 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
3531 struct nv50_program
*vp
= nv50
->vertprog
;
3532 struct nv50_program
*fp
= nv50
->fragprog
;
3533 struct nouveau_stateobj
*so
;
3534 struct nv50_sreg4 dummy
, *vpo
;
3536 uint32_t map
[16], lin
[4], reg
[5], pcrd
[8];
3538 memset(map
, 0, sizeof(map
));
3539 memset(lin
, 0, sizeof(lin
));
3541 reg
[1] = 0x00000004; /* low and high clip distance map ids */
3542 reg
[2] = 0x00000000; /* layer index map id (disabled, GP only) */
3543 reg
[3] = 0x00000000; /* point size map id & enable */
3544 reg
[0] = fp
->cfg
.regs
[0]; /* colour semantic reg */
3545 reg
[4] = fp
->cfg
.regs
[1]; /* interpolant info */
3547 dummy
.linear
= FALSE
;
3548 dummy
.mask
= 0xf; /* map all components of HPOS */
3549 m
= nv50_sreg4_map(map
, m
, lin
, &dummy
, &vp
->cfg
.io
[0]);
3553 if (vp
->cfg
.clpd
< 0x40) {
3554 for (c
= 0; c
< vp
->cfg
.clpd_nr
; ++c
)
3555 map
[m
++] = vp
->cfg
.clpd
+ c
;
3559 reg
[0] |= m
<< 8; /* adjust BFC0 id */
3561 /* if light_twoside is active, it seems FFC0_ID == BFC0_ID is bad */
3562 if (nv50
->rasterizer
->pipe
.light_twoside
) {
3563 vpo
= &vp
->cfg
.two_side
[0];
3565 m
= nv50_sreg4_map(map
, m
, lin
, &fp
->cfg
.two_side
[0], &vpo
[0]);
3566 m
= nv50_sreg4_map(map
, m
, lin
, &fp
->cfg
.two_side
[1], &vpo
[1]);
3569 reg
[0] += m
- 4; /* adjust FFC0 id */
3570 reg
[4] |= m
<< 8; /* set mid where 'normal' FP inputs start */
3572 for (i
= 0; i
< fp
->cfg
.io_nr
; i
++) {
3573 ubyte sn
= fp
->info
.input_semantic_name
[fp
->cfg
.io
[i
].id
];
3574 ubyte si
= fp
->info
.input_semantic_index
[fp
->cfg
.io
[i
].id
];
3576 /* position must be mapped first */
3577 assert(i
== 0 || sn
!= TGSI_SEMANTIC_POSITION
);
3579 /* maybe even remove these from cfg.io */
3580 if (sn
== TGSI_SEMANTIC_POSITION
|| sn
== TGSI_SEMANTIC_FACE
)
3583 /* VP outputs and vp->cfg.io are in the same order */
3584 for (n
= 0; n
< vp
->info
.num_outputs
; ++n
) {
3585 if (vp
->info
.output_semantic_name
[n
] == sn
&&
3586 vp
->info
.output_semantic_index
[n
] == si
)
3589 vpo
= (n
< vp
->info
.num_outputs
) ? &vp
->cfg
.io
[n
] : &dummy
;
3591 m
= nv50_sreg4_map(map
, m
, lin
, &fp
->cfg
.io
[i
], vpo
);
3594 if (nv50
->rasterizer
->pipe
.point_size_per_vertex
) {
3595 map
[m
/ 4] |= vp
->cfg
.psiz
<< ((m
% 4) * 8);
3596 reg
[3] = (m
++ << 4) | 1;
3599 /* now fill the stateobj */
3603 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP_SIZE
, 1);
3605 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP(0), n
);
3606 so_datap (so
, map
, n
);
3608 so_method(so
, tesla
, NV50TCL_MAP_SEMANTIC_0
, 4);
3609 so_datap (so
, reg
, 4);
3611 so_method(so
, tesla
, NV50TCL_FP_INTERPOLANT_CTRL
, 1);
3612 so_data (so
, reg
[4]);
3614 so_method(so
, tesla
, 0x1540, 4);
3615 so_datap (so
, lin
, 4);
3617 if (nv50
->rasterizer
->pipe
.point_sprite
) {
3618 nv50_pntc_replace(nv50
, pcrd
, (reg
[4] >> 8) & 0xff);
3620 so_method(so
, tesla
, NV50TCL_POINT_COORD_REPLACE_MAP(0), 8);
3621 so_datap (so
, pcrd
, 8);
3624 so_ref(so
, &nv50
->state
.programs
);
3629 nv50_program_destroy(struct nv50_context
*nv50
, struct nv50_program
*p
)
3631 while (p
->exec_head
) {
3632 struct nv50_program_exec
*e
= p
->exec_head
;
3634 p
->exec_head
= e
->next
;
3637 p
->exec_tail
= NULL
;
3640 nouveau_bo_ref(NULL
, &p
->bo
);
3642 nouveau_resource_free(&p
->data
[0]);