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
)
161 assert(pc
->reg_instance_nr
< 16);
162 ri
= &pc
->reg_instances
[pc
->reg_instance_nr
++];
171 ctor_reg(struct nv50_reg
*reg
, unsigned type
, int index
, int hw
)
181 static INLINE
unsigned
182 popcnt4(uint32_t val
)
184 static const unsigned cnt
[16]
185 = { 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4 };
186 return cnt
[val
& 0xf];
190 terminate_mbb(struct nv50_pc
*pc
)
194 /* remove records of temporary address register values */
195 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
)
196 if (pc
->r_addr
[i
].index
< 0)
197 pc
->r_addr
[i
].rhw
= -1;
201 alloc_reg(struct nv50_pc
*pc
, struct nv50_reg
*reg
)
205 if (reg
->type
== P_RESULT
) {
206 if (pc
->p
->cfg
.high_result
< (reg
->hw
+ 1))
207 pc
->p
->cfg
.high_result
= reg
->hw
+ 1;
210 if (reg
->type
!= P_TEMP
)
214 /*XXX: do this here too to catch FP temp-as-attr usage..
215 * not clean, but works */
216 if (pc
->p
->cfg
.high_temp
< (reg
->hw
+ 1))
217 pc
->p
->cfg
.high_temp
= reg
->hw
+ 1;
221 if (reg
->rhw
!= -1) {
222 /* try to allocate temporary with index rhw first */
223 if (!(pc
->r_temp
[reg
->rhw
])) {
224 pc
->r_temp
[reg
->rhw
] = reg
;
226 if (pc
->p
->cfg
.high_temp
< (reg
->rhw
+ 1))
227 pc
->p
->cfg
.high_temp
= reg
->rhw
+ 1;
230 /* make sure we don't get things like $r0 needs to go
231 * in $r1 and $r1 in $r0
233 i
= pc
->result_nr
* 4;
236 for (; i
< NV50_SU_MAX_TEMP
; i
++) {
237 if (!(pc
->r_temp
[i
])) {
240 if (pc
->p
->cfg
.high_temp
< (i
+ 1))
241 pc
->p
->cfg
.high_temp
= i
+ 1;
249 /* XXX: For shaders that aren't executed linearly (e.g. shaders that
250 * contain loops), we need to assign all hw regs to TGSI TEMPs early,
251 * lest we risk temp_temps overwriting regs alloc'd "later".
253 static struct nv50_reg
*
254 alloc_temp(struct nv50_pc
*pc
, struct nv50_reg
*dst
)
259 if (dst
&& dst
->type
== P_TEMP
&& dst
->hw
== -1)
262 for (i
= 0; i
< NV50_SU_MAX_TEMP
; i
++) {
263 if (!pc
->r_temp
[i
]) {
264 r
= MALLOC_STRUCT(nv50_reg
);
265 ctor_reg(r
, P_TEMP
, -1, i
);
275 /* Assign the hw of the discarded temporary register src
276 * to the tgsi register dst and free src.
279 assimilate_temp(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
281 assert(src
->index
== -1 && src
->hw
!= -1);
284 pc
->r_temp
[dst
->hw
] = NULL
;
285 pc
->r_temp
[src
->hw
] = dst
;
291 /* release the hardware resource held by r */
293 release_hw(struct nv50_pc
*pc
, struct nv50_reg
*r
)
295 assert(r
->type
== P_TEMP
);
299 assert(pc
->r_temp
[r
->hw
] == r
);
300 pc
->r_temp
[r
->hw
] = NULL
;
308 free_temp(struct nv50_pc
*pc
, struct nv50_reg
*r
)
310 if (r
->index
== -1) {
313 FREE(pc
->r_temp
[hw
]);
314 pc
->r_temp
[hw
] = NULL
;
319 alloc_temp4(struct nv50_pc
*pc
, struct nv50_reg
*dst
[4], int idx
)
323 if ((idx
+ 4) >= NV50_SU_MAX_TEMP
)
326 if (pc
->r_temp
[idx
] || pc
->r_temp
[idx
+ 1] ||
327 pc
->r_temp
[idx
+ 2] || pc
->r_temp
[idx
+ 3])
328 return alloc_temp4(pc
, dst
, idx
+ 4);
330 for (i
= 0; i
< 4; i
++) {
331 dst
[i
] = MALLOC_STRUCT(nv50_reg
);
332 ctor_reg(dst
[i
], P_TEMP
, -1, idx
+ i
);
333 pc
->r_temp
[idx
+ i
] = dst
[i
];
340 free_temp4(struct nv50_pc
*pc
, struct nv50_reg
*reg
[4])
344 for (i
= 0; i
< 4; i
++)
345 free_temp(pc
, reg
[i
]);
348 static struct nv50_reg
*
349 temp_temp(struct nv50_pc
*pc
)
351 if (pc
->temp_temp_nr
>= 16)
354 pc
->temp_temp
[pc
->temp_temp_nr
] = alloc_temp(pc
, NULL
);
355 return pc
->temp_temp
[pc
->temp_temp_nr
++];
359 kill_temp_temp(struct nv50_pc
*pc
)
363 for (i
= 0; i
< pc
->temp_temp_nr
; i
++)
364 free_temp(pc
, pc
->temp_temp
[i
]);
365 pc
->temp_temp_nr
= 0;
369 ctor_immd_4u32(struct nv50_pc
*pc
,
370 uint32_t x
, uint32_t y
, uint32_t z
, uint32_t w
)
372 unsigned size
= pc
->immd_nr
* 4 * sizeof(uint32_t);
374 pc
->immd_buf
= REALLOC(pc
->immd_buf
, size
, size
+ 4 * sizeof(uint32_t));
376 pc
->immd_buf
[(pc
->immd_nr
* 4) + 0] = x
;
377 pc
->immd_buf
[(pc
->immd_nr
* 4) + 1] = y
;
378 pc
->immd_buf
[(pc
->immd_nr
* 4) + 2] = z
;
379 pc
->immd_buf
[(pc
->immd_nr
* 4) + 3] = w
;
381 return pc
->immd_nr
++;
385 ctor_immd_4f32(struct nv50_pc
*pc
, float x
, float y
, float z
, float w
)
387 return ctor_immd_4u32(pc
, fui(x
), fui(y
), fui(z
), fui(w
));
390 static struct nv50_reg
*
391 alloc_immd(struct nv50_pc
*pc
, float f
)
393 struct nv50_reg
*r
= MALLOC_STRUCT(nv50_reg
);
396 for (hw
= 0; hw
< pc
->immd_nr
* 4; hw
++)
397 if (pc
->immd_buf
[hw
] == fui(f
))
400 if (hw
== pc
->immd_nr
* 4)
401 hw
= ctor_immd_4f32(pc
, f
, -f
, 0.5 * f
, 0) * 4;
403 ctor_reg(r
, P_IMMD
, -1, hw
);
407 static struct nv50_program_exec
*
408 exec(struct nv50_pc
*pc
)
410 struct nv50_program_exec
*e
= CALLOC_STRUCT(nv50_program_exec
);
417 emit(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
419 struct nv50_program
*p
= pc
->p
;
422 p
->exec_tail
->next
= e
;
426 p
->exec_size
+= (e
->inst
[0] & 1) ? 2 : 1;
429 static INLINE
void set_long(struct nv50_pc
*, struct nv50_program_exec
*);
432 is_long(struct nv50_program_exec
*e
)
440 is_immd(struct nv50_program_exec
*e
)
442 if (is_long(e
) && (e
->inst
[1] & 3) == 3)
448 set_pred(struct nv50_pc
*pc
, unsigned pred
, unsigned idx
,
449 struct nv50_program_exec
*e
)
452 e
->inst
[1] &= ~((0x1f << 7) | (0x3 << 12));
453 e
->inst
[1] |= (pred
<< 7) | (idx
<< 12);
457 set_pred_wr(struct nv50_pc
*pc
, unsigned on
, unsigned idx
,
458 struct nv50_program_exec
*e
)
461 e
->inst
[1] &= ~((0x3 << 4) | (1 << 6));
462 e
->inst
[1] |= (idx
<< 4) | (on
<< 6);
466 set_long(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
472 set_pred(pc
, 0xf, 0, e
);
473 set_pred_wr(pc
, 0, 0, e
);
477 set_dst(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_program_exec
*e
)
479 if (dst
->type
== P_RESULT
) {
481 e
->inst
[1] |= 0x00000008;
487 e
->inst
[0] |= (dst
->hw
<< 2);
491 set_immd(struct nv50_pc
*pc
, struct nv50_reg
*imm
, struct nv50_program_exec
*e
)
497 u
.ui
= pc
->immd_buf
[imm
->hw
];
499 u
.f
= (imm
->mod
& NV50_MOD_ABS
) ? fabsf(u
.f
) : u
.f
;
500 u
.f
= (imm
->mod
& NV50_MOD_NEG
) ? -u
.f
: u
.f
;
503 /* XXX: can't be predicated - bits overlap; cases where both
504 * are required should be avoided by using pc->allow32 */
505 set_pred(pc
, 0, 0, e
);
506 set_pred_wr(pc
, 0, 0, e
);
508 e
->inst
[1] |= 0x00000002 | 0x00000001;
509 e
->inst
[0] |= (u
.ui
& 0x3f) << 16;
510 e
->inst
[1] |= (u
.ui
>> 6) << 2;
514 set_addr(struct nv50_program_exec
*e
, struct nv50_reg
*a
)
516 assert(!(e
->inst
[0] & 0x0c000000));
517 assert(!(e
->inst
[1] & 0x00000004));
519 e
->inst
[0] |= (a
->hw
& 3) << 26;
520 e
->inst
[1] |= (a
->hw
>> 2) << 2;
524 emit_add_addr_imm(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
525 struct nv50_reg
*src0
, uint16_t src1_val
)
527 struct nv50_program_exec
*e
= exec(pc
);
529 e
->inst
[0] = 0xd0000000 | (src1_val
<< 9);
530 e
->inst
[1] = 0x20000000;
532 e
->inst
[0] |= dst
->hw
<< 2;
533 if (src0
) /* otherwise will add to $a0, which is always 0 */
539 static struct nv50_reg
*
540 alloc_addr(struct nv50_pc
*pc
, struct nv50_reg
*ref
)
543 struct nv50_reg
*a_tgsi
= NULL
, *a
= NULL
;
546 /* allocate for TGSI address reg */
547 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
) {
548 if (pc
->r_addr
[i
].index
>= 0)
550 if (pc
->r_addr
[i
].rhw
>= 0 &&
551 pc
->r_addr
[i
].acc
== pc
->insn_cur
)
554 pc
->r_addr
[i
].rhw
= -1;
555 pc
->r_addr
[i
].index
= i
;
556 return &pc
->r_addr
[i
];
562 /* Allocate and set an address reg so we can access 'ref'.
564 * If and r_addr has index < 0, it is not reserved for TGSI,
565 * and index will be the negative of the TGSI addr index the
566 * value in rhw is relative to, or -256 if rhw is an offset
567 * from 0. If rhw < 0, the reg has not been initialized.
569 for (i
= NV50_SU_MAX_ADDR
- 1; i
>= 0; --i
) {
570 if (pc
->r_addr
[i
].index
>= 0) /* occupied for TGSI */
572 if (pc
->r_addr
[i
].rhw
< 0) { /* unused */
576 if (!a
&& pc
->r_addr
[i
].acc
!= pc
->insn_cur
)
579 if (ref
->hw
- pc
->r_addr
[i
].rhw
>= 128)
582 if ((ref
->acc
>= 0 && pc
->r_addr
[i
].index
== -256) ||
583 (ref
->acc
< 0 && -pc
->r_addr
[i
].index
== ref
->index
)) {
584 pc
->r_addr
[i
].acc
= pc
->insn_cur
;
585 return &pc
->r_addr
[i
];
591 a_tgsi
= pc
->addr
[ref
->index
];
593 emit_add_addr_imm(pc
, a
, a_tgsi
, (ref
->hw
& ~0x7f) * 4);
595 a
->rhw
= ref
->hw
& ~0x7f;
596 a
->acc
= pc
->insn_cur
;
597 a
->index
= a_tgsi
? -ref
->index
: -256;
601 #define INTERP_LINEAR 0
602 #define INTERP_FLAT 1
603 #define INTERP_PERSPECTIVE 2
604 #define INTERP_CENTROID 4
606 /* interpolant index has been stored in dst->rhw */
608 emit_interp(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*iv
,
611 assert(dst
->rhw
!= -1);
612 struct nv50_program_exec
*e
= exec(pc
);
614 e
->inst
[0] |= 0x80000000;
616 e
->inst
[0] |= (dst
->rhw
<< 16);
618 if (mode
& INTERP_FLAT
) {
619 e
->inst
[0] |= (1 << 8);
621 if (mode
& INTERP_PERSPECTIVE
) {
622 e
->inst
[0] |= (1 << 25);
624 e
->inst
[0] |= (iv
->hw
<< 9);
627 if (mode
& INTERP_CENTROID
)
628 e
->inst
[0] |= (1 << 24);
635 set_data(struct nv50_pc
*pc
, struct nv50_reg
*src
, unsigned m
, unsigned s
,
636 struct nv50_program_exec
*e
)
640 e
->param
.index
= src
->hw
& 127;
642 e
->param
.mask
= m
<< (s
% 32);
645 set_addr(e
, alloc_addr(pc
, src
));
648 assert(src
->type
== P_CONST
);
649 set_addr(e
, pc
->addr
[src
->index
]);
652 e
->inst
[1] |= (((src
->type
== P_IMMD
) ? 0 : 1) << 22);
656 emit_mov(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
658 struct nv50_program_exec
*e
= exec(pc
);
660 e
->inst
[0] = 0x10000000;
666 if (!is_long(e
) && src
->type
== P_IMMD
) {
667 set_immd(pc
, src
, e
);
668 /*XXX: 32-bit, but steals part of "half" reg space - need to
669 * catch and handle this case if/when we do half-regs
672 if (src
->type
== P_IMMD
|| src
->type
== P_CONST
) {
674 set_data(pc
, src
, 0x7f, 9, e
);
675 e
->inst
[1] |= 0x20000000; /* mov from c[] */
677 if (src
->type
== P_ATTR
) {
679 e
->inst
[1] |= 0x00200000;
685 e
->inst
[0] |= (src
->hw
<< 9);
688 if (is_long(e
) && !is_immd(e
)) {
689 e
->inst
[1] |= 0x04000000; /* 32-bit */
690 e
->inst
[1] |= 0x0000c000; /* 32-bit c[] load / lane mask 0:1 */
691 if (!(e
->inst
[1] & 0x20000000))
692 e
->inst
[1] |= 0x00030000; /* lane mask 2:3 */
694 e
->inst
[0] |= 0x00008000;
700 emit_mov_immdval(struct nv50_pc
*pc
, struct nv50_reg
*dst
, float f
)
702 struct nv50_reg
*imm
= alloc_immd(pc
, f
);
703 emit_mov(pc
, dst
, imm
);
708 emit_nop(struct nv50_pc
*pc
)
710 struct nv50_program_exec
*e
= exec(pc
);
712 e
->inst
[0] = 0xf0000000;
714 e
->inst
[1] = 0xe0000000;
719 check_swap_src_0_1(struct nv50_pc
*pc
,
720 struct nv50_reg
**s0
, struct nv50_reg
**s1
)
722 struct nv50_reg
*src0
= *s0
, *src1
= *s1
;
724 if (src0
->type
== P_CONST
) {
725 if (src1
->type
!= P_CONST
) {
731 if (src1
->type
== P_ATTR
) {
732 if (src0
->type
!= P_ATTR
) {
743 set_src_0_restricted(struct nv50_pc
*pc
, struct nv50_reg
*src
,
744 struct nv50_program_exec
*e
)
746 struct nv50_reg
*temp
;
748 if (src
->type
!= P_TEMP
) {
749 temp
= temp_temp(pc
);
750 emit_mov(pc
, temp
, src
);
757 e
->inst
[0] |= (src
->hw
<< 9);
761 set_src_0(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
763 if (src
->type
== P_ATTR
) {
765 e
->inst
[1] |= 0x00200000;
767 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
768 struct nv50_reg
*temp
= temp_temp(pc
);
770 emit_mov(pc
, temp
, src
);
777 e
->inst
[0] |= (src
->hw
<< 9);
781 set_src_1(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
783 if (src
->type
== P_ATTR
) {
784 struct nv50_reg
*temp
= temp_temp(pc
);
786 emit_mov(pc
, temp
, src
);
789 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
790 assert(!(e
->inst
[0] & 0x00800000));
791 if (e
->inst
[0] & 0x01000000) {
792 struct nv50_reg
*temp
= temp_temp(pc
);
794 emit_mov(pc
, temp
, src
);
797 set_data(pc
, src
, 0x7f, 16, e
);
798 e
->inst
[0] |= 0x00800000;
805 e
->inst
[0] |= ((src
->hw
& 127) << 16);
809 set_src_2(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
813 if (src
->type
== P_ATTR
) {
814 struct nv50_reg
*temp
= temp_temp(pc
);
816 emit_mov(pc
, temp
, src
);
819 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
820 assert(!(e
->inst
[0] & 0x01000000));
821 if (e
->inst
[0] & 0x00800000) {
822 struct nv50_reg
*temp
= temp_temp(pc
);
824 emit_mov(pc
, temp
, src
);
827 set_data(pc
, src
, 0x7f, 32+14, e
);
828 e
->inst
[0] |= 0x01000000;
833 e
->inst
[1] |= ((src
->hw
& 127) << 14);
837 emit_mov_from_pred(struct nv50_pc
*pc
, struct nv50_reg
*dst
, int pred
)
839 struct nv50_program_exec
*e
= exec(pc
);
841 assert(dst
->type
== P_TEMP
);
842 e
->inst
[1] = 0x20000000 | (pred
<< 12);
850 emit_mov_to_pred(struct nv50_pc
*pc
, int pred
, struct nv50_reg
*src
)
852 struct nv50_program_exec
*e
= exec(pc
);
854 e
->inst
[0] = 0x000001fc;
855 e
->inst
[1] = 0xa0000008;
857 set_pred_wr(pc
, 1, pred
, e
);
858 set_src_0_restricted(pc
, src
, e
);
864 emit_mul(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
865 struct nv50_reg
*src1
)
867 struct nv50_program_exec
*e
= exec(pc
);
869 e
->inst
[0] |= 0xc0000000;
874 check_swap_src_0_1(pc
, &src0
, &src1
);
876 set_src_0(pc
, src0
, e
);
877 if (src1
->type
== P_IMMD
&& !is_long(e
)) {
878 if (src0
->mod
& NV50_MOD_NEG
)
879 e
->inst
[0] |= 0x00008000;
880 set_immd(pc
, src1
, e
);
882 set_src_1(pc
, src1
, e
);
883 if ((src0
->mod
^ src1
->mod
) & NV50_MOD_NEG
) {
885 e
->inst
[1] |= 0x08000000;
887 e
->inst
[0] |= 0x00008000;
895 emit_add(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
896 struct nv50_reg
*src0
, struct nv50_reg
*src1
)
898 struct nv50_program_exec
*e
= exec(pc
);
900 e
->inst
[0] = 0xb0000000;
903 check_swap_src_0_1(pc
, &src0
, &src1
);
905 if (!pc
->allow32
|| (src0
->mod
| src1
->mod
) || src1
->hw
> 63) {
907 e
->inst
[1] |= ((src0
->mod
& NV50_MOD_NEG
) << 26) |
908 ((src1
->mod
& NV50_MOD_NEG
) << 27);
912 set_src_0(pc
, src0
, e
);
913 if (src1
->type
== P_CONST
|| src1
->type
== P_ATTR
|| is_long(e
))
914 set_src_2(pc
, src1
, e
);
916 if (src1
->type
== P_IMMD
)
917 set_immd(pc
, src1
, e
);
919 set_src_1(pc
, src1
, e
);
925 emit_arl(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
,
928 struct nv50_program_exec
*e
= exec(pc
);
931 e
->inst
[1] |= 0xc0000000;
933 e
->inst
[0] |= dst
->hw
<< 2;
934 e
->inst
[0] |= s
<< 16; /* shift left */
935 set_src_0_restricted(pc
, src
, e
);
941 emit_minmax(struct nv50_pc
*pc
, unsigned sub
, struct nv50_reg
*dst
,
942 struct nv50_reg
*src0
, struct nv50_reg
*src1
)
944 struct nv50_program_exec
*e
= exec(pc
);
947 e
->inst
[0] |= 0xb0000000;
948 e
->inst
[1] |= (sub
<< 29);
950 check_swap_src_0_1(pc
, &src0
, &src1
);
952 set_src_0(pc
, src0
, e
);
953 set_src_1(pc
, src1
, e
);
955 if (src0
->mod
& NV50_MOD_ABS
)
956 e
->inst
[1] |= 0x00100000;
957 if (src1
->mod
& NV50_MOD_ABS
)
958 e
->inst
[1] |= 0x00080000;
964 emit_sub(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
965 struct nv50_reg
*src1
)
967 src1
->mod
^= NV50_MOD_NEG
;
968 emit_add(pc
, dst
, src0
, src1
);
969 src1
->mod
^= NV50_MOD_NEG
;
973 emit_bitop2(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
974 struct nv50_reg
*src1
, unsigned op
)
976 struct nv50_program_exec
*e
= exec(pc
);
978 e
->inst
[0] = 0xd0000000;
981 check_swap_src_0_1(pc
, &src0
, &src1
);
983 set_src_0(pc
, src0
, e
);
985 if (op
!= TGSI_OPCODE_AND
&& op
!= TGSI_OPCODE_OR
&&
986 op
!= TGSI_OPCODE_XOR
)
987 assert(!"invalid bit op");
989 if (src1
->type
== P_IMMD
&& src0
->type
== P_TEMP
&& pc
->allow32
) {
990 set_immd(pc
, src1
, e
);
991 if (op
== TGSI_OPCODE_OR
)
992 e
->inst
[0] |= 0x0100;
994 if (op
== TGSI_OPCODE_XOR
)
995 e
->inst
[0] |= 0x8000;
997 set_src_1(pc
, src1
, e
);
998 e
->inst
[1] |= 0x04000000; /* 32 bit */
999 if (op
== TGSI_OPCODE_OR
)
1000 e
->inst
[1] |= 0x4000;
1002 if (op
== TGSI_OPCODE_XOR
)
1003 e
->inst
[1] |= 0x8000;
1010 emit_mad(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
1011 struct nv50_reg
*src1
, struct nv50_reg
*src2
)
1013 struct nv50_program_exec
*e
= exec(pc
);
1015 e
->inst
[0] |= 0xe0000000;
1017 check_swap_src_0_1(pc
, &src0
, &src1
);
1018 set_dst(pc
, dst
, e
);
1019 set_src_0(pc
, src0
, e
);
1020 set_src_1(pc
, src1
, e
);
1021 set_src_2(pc
, src2
, e
);
1023 if ((src0
->mod
^ src1
->mod
) & NV50_MOD_NEG
)
1024 e
->inst
[1] |= 0x04000000;
1025 if (src2
->mod
& NV50_MOD_NEG
)
1026 e
->inst
[1] |= 0x08000000;
1032 emit_msb(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
1033 struct nv50_reg
*src1
, struct nv50_reg
*src2
)
1035 src2
->mod
^= NV50_MOD_NEG
;
1036 emit_mad(pc
, dst
, src0
, src1
, src2
);
1037 src2
->mod
^= NV50_MOD_NEG
;
1041 emit_flop(struct nv50_pc
*pc
, unsigned sub
,
1042 struct nv50_reg
*dst
, struct nv50_reg
*src
)
1044 struct nv50_program_exec
*e
= exec(pc
);
1046 e
->inst
[0] |= 0x90000000;
1049 e
->inst
[1] |= (sub
<< 29);
1052 set_dst(pc
, dst
, e
);
1054 if (sub
== 0 || sub
== 2)
1055 set_src_0_restricted(pc
, src
, e
);
1057 set_src_0(pc
, src
, e
);
1063 emit_preex2(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1065 struct nv50_program_exec
*e
= exec(pc
);
1067 e
->inst
[0] |= 0xb0000000;
1069 set_dst(pc
, dst
, e
);
1070 set_src_0(pc
, src
, e
);
1072 e
->inst
[1] |= (6 << 29) | 0x00004000;
1078 emit_precossin(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1080 struct nv50_program_exec
*e
= exec(pc
);
1082 e
->inst
[0] |= 0xb0000000;
1084 set_dst(pc
, dst
, e
);
1085 set_src_0(pc
, src
, e
);
1087 e
->inst
[1] |= (6 << 29);
1092 #define CVTOP_RN 0x01
1093 #define CVTOP_FLOOR 0x03
1094 #define CVTOP_CEIL 0x05
1095 #define CVTOP_TRUNC 0x07
1096 #define CVTOP_SAT 0x08
1097 #define CVTOP_ABS 0x10
1099 /* 0x04 == 32 bit dst */
1100 /* 0x40 == dst is float */
1101 /* 0x80 == src is float */
1102 #define CVT_F32_F32 0xc4
1103 #define CVT_F32_S32 0x44
1104 #define CVT_S32_F32 0x8c
1105 #define CVT_S32_S32 0x0c
1106 #define CVT_NEG 0x20
1110 emit_cvt(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
,
1111 int wp
, unsigned cvn
, unsigned fmt
)
1113 struct nv50_program_exec
*e
;
1118 e
->inst
[0] |= 0xa0000000;
1119 e
->inst
[1] |= 0x00004000; /* 32 bit src */
1120 e
->inst
[1] |= (cvn
<< 16);
1121 e
->inst
[1] |= (fmt
<< 24);
1122 set_src_0(pc
, src
, e
);
1125 set_pred_wr(pc
, 1, wp
, e
);
1128 set_dst(pc
, dst
, e
);
1130 e
->inst
[0] |= 0x000001fc;
1131 e
->inst
[1] |= 0x00000008;
1137 /* nv50 Condition codes:
1144 * 0x7 = set condition code ? (used before bra.lt/le/gt/ge)
1145 * 0x8 = unordered bit (allows NaN)
1148 emit_set(struct nv50_pc
*pc
, unsigned ccode
, struct nv50_reg
*dst
, int wp
,
1149 struct nv50_reg
*src0
, struct nv50_reg
*src1
)
1151 static const unsigned cc_swapped
[8] = { 0, 4, 2, 6, 1, 5, 3, 7 };
1153 struct nv50_program_exec
*e
= exec(pc
);
1154 struct nv50_reg
*rdst
;
1157 if (check_swap_src_0_1(pc
, &src0
, &src1
))
1158 ccode
= cc_swapped
[ccode
& 7] | (ccode
& 8);
1161 if (dst
&& dst
->type
!= P_TEMP
)
1162 dst
= alloc_temp(pc
, NULL
);
1166 e
->inst
[0] |= 0xb0000000;
1167 e
->inst
[1] |= 0x60000000 | (ccode
<< 14);
1169 /* XXX: decuda will disasm as .u16 and use .lo/.hi regs, but
1170 * that doesn't seem to match what the hw actually does
1171 e->inst[1] |= 0x04000000; << breaks things, u32 by default ?
1175 set_pred_wr(pc
, 1, wp
, e
);
1177 set_dst(pc
, dst
, e
);
1179 e
->inst
[0] |= 0x000001fc;
1180 e
->inst
[1] |= 0x00000008;
1183 set_src_0(pc
, src0
, e
);
1184 set_src_1(pc
, src1
, e
);
1187 pc
->if_cond
= pc
->p
->exec_tail
; /* record for OPCODE_IF */
1189 /* cvt.f32.u32/s32 (?) if we didn't only write the predicate */
1191 emit_cvt(pc
, rdst
, dst
, -1, CVTOP_ABS
| CVTOP_RN
, CVT_F32_S32
);
1192 if (rdst
&& rdst
!= dst
)
1196 static INLINE
unsigned
1197 map_tgsi_setop_cc(unsigned op
)
1200 case TGSI_OPCODE_SLT
: return 0x1;
1201 case TGSI_OPCODE_SGE
: return 0x6;
1202 case TGSI_OPCODE_SEQ
: return 0x2;
1203 case TGSI_OPCODE_SGT
: return 0x4;
1204 case TGSI_OPCODE_SLE
: return 0x3;
1205 case TGSI_OPCODE_SNE
: return 0xd;
1213 emit_flr(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1215 emit_cvt(pc
, dst
, src
, -1, CVTOP_FLOOR
, CVT_F32_F32
| CVT_RI
);
1219 emit_pow(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1220 struct nv50_reg
*v
, struct nv50_reg
*e
)
1222 struct nv50_reg
*temp
= alloc_temp(pc
, NULL
);
1224 emit_flop(pc
, 3, temp
, v
);
1225 emit_mul(pc
, temp
, temp
, e
);
1226 emit_preex2(pc
, temp
, temp
);
1227 emit_flop(pc
, 6, dst
, temp
);
1229 free_temp(pc
, temp
);
1233 emit_abs(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1235 emit_cvt(pc
, dst
, src
, -1, CVTOP_ABS
, CVT_F32_F32
);
1239 emit_sat(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1241 emit_cvt(pc
, dst
, src
, -1, CVTOP_SAT
, CVT_F32_F32
);
1245 emit_lit(struct nv50_pc
*pc
, struct nv50_reg
**dst
, unsigned mask
,
1246 struct nv50_reg
**src
)
1248 struct nv50_reg
*one
= alloc_immd(pc
, 1.0);
1249 struct nv50_reg
*zero
= alloc_immd(pc
, 0.0);
1250 struct nv50_reg
*neg128
= alloc_immd(pc
, -127.999999);
1251 struct nv50_reg
*pos128
= alloc_immd(pc
, 127.999999);
1252 struct nv50_reg
*tmp
[4];
1253 boolean allow32
= pc
->allow32
;
1255 pc
->allow32
= FALSE
;
1257 if (mask
& (3 << 1)) {
1258 tmp
[0] = alloc_temp(pc
, NULL
);
1259 emit_minmax(pc
, 4, tmp
[0], src
[0], zero
);
1262 if (mask
& (1 << 2)) {
1263 set_pred_wr(pc
, 1, 0, pc
->p
->exec_tail
);
1265 tmp
[1] = temp_temp(pc
);
1266 emit_minmax(pc
, 4, tmp
[1], src
[1], zero
);
1268 tmp
[3] = temp_temp(pc
);
1269 emit_minmax(pc
, 4, tmp
[3], src
[3], neg128
);
1270 emit_minmax(pc
, 5, tmp
[3], tmp
[3], pos128
);
1272 emit_pow(pc
, dst
[2], tmp
[1], tmp
[3]);
1273 emit_mov(pc
, dst
[2], zero
);
1274 set_pred(pc
, 3, 0, pc
->p
->exec_tail
);
1277 if (mask
& (1 << 1))
1278 assimilate_temp(pc
, dst
[1], tmp
[0]);
1280 if (mask
& (1 << 2))
1281 free_temp(pc
, tmp
[0]);
1283 pc
->allow32
= allow32
;
1285 /* do this last, in case src[i,j] == dst[0,3] */
1286 if (mask
& (1 << 0))
1287 emit_mov(pc
, dst
[0], one
);
1289 if (mask
& (1 << 3))
1290 emit_mov(pc
, dst
[3], one
);
1299 emit_neg(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1301 emit_cvt(pc
, dst
, src
, -1, CVTOP_RN
, CVT_F32_F32
| CVT_NEG
);
1305 emit_kil(struct nv50_pc
*pc
, struct nv50_reg
*src
)
1307 struct nv50_program_exec
*e
;
1308 const int r_pred
= 1;
1309 unsigned cvn
= CVT_F32_F32
;
1311 if (src
->mod
& NV50_MOD_NEG
)
1313 /* write predicate reg */
1314 emit_cvt(pc
, NULL
, src
, r_pred
, CVTOP_RN
, cvn
);
1316 /* conditional discard */
1318 e
->inst
[0] = 0x00000002;
1320 set_pred(pc
, 0x1 /* LT */, r_pred
, e
);
1324 static struct nv50_program_exec
*
1325 emit_branch(struct nv50_pc
*pc
, int pred
, unsigned cc
,
1326 struct nv50_program_exec
**join
)
1328 struct nv50_program_exec
*e
= exec(pc
);
1332 e
->inst
[0] |= 0xa0000002;
1339 e
->inst
[0] |= 0x10000002;
1341 set_pred(pc
, cc
, pred
, e
);
1343 return pc
->p
->exec_tail
;
1349 #define QOP_MOV_SRC1 3
1351 /* For a quad of threads / top left, top right, bottom left, bottom right
1352 * pixels, do a different operation, and take src0 from a specific thread.
1355 emit_quadop(struct nv50_pc
*pc
, struct nv50_reg
*dst
, int wp
, int lane_src0
,
1356 struct nv50_reg
*src0
, struct nv50_reg
*src1
, ubyte qop
)
1358 struct nv50_program_exec
*e
= exec(pc
);
1360 e
->inst
[0] = 0xc0000000;
1361 e
->inst
[1] = 0x80000000;
1363 e
->inst
[0] |= lane_src0
<< 16;
1364 set_src_0(pc
, src0
, e
);
1365 set_src_2(pc
, src1
, e
);
1368 set_pred_wr(pc
, 1, wp
, e
);
1371 set_dst(pc
, dst
, e
);
1373 e
->inst
[0] |= 0x000001fc;
1374 e
->inst
[1] |= 0x00000008;
1377 e
->inst
[0] |= (qop
& 3) << 20;
1378 e
->inst
[1] |= (qop
>> 2) << 22;
1384 load_cube_tex_coords(struct nv50_pc
*pc
, struct nv50_reg
*t
[4],
1385 struct nv50_reg
**src
, unsigned arg
, boolean proj
)
1387 int mod
[3] = { src
[0]->mod
, src
[1]->mod
, src
[2]->mod
};
1389 src
[0]->mod
|= NV50_MOD_ABS
;
1390 src
[1]->mod
|= NV50_MOD_ABS
;
1391 src
[2]->mod
|= NV50_MOD_ABS
;
1393 emit_minmax(pc
, 4, t
[2], src
[0], src
[1]);
1394 emit_minmax(pc
, 4, t
[2], src
[2], t
[2]);
1396 src
[0]->mod
= mod
[0];
1397 src
[1]->mod
= mod
[1];
1398 src
[2]->mod
= mod
[2];
1400 if (proj
&& 0 /* looks more correct without this */)
1401 emit_mul(pc
, t
[2], t
[2], src
[3]);
1403 if (arg
== 4) /* there is no textureProj(samplerCubeShadow) */
1404 emit_mov(pc
, t
[3], src
[3]);
1406 emit_flop(pc
, 0, t
[2], t
[2]);
1408 emit_mul(pc
, t
[0], src
[0], t
[2]);
1409 emit_mul(pc
, t
[1], src
[1], t
[2]);
1410 emit_mul(pc
, t
[2], src
[2], t
[2]);
1414 load_proj_tex_coords(struct nv50_pc
*pc
, struct nv50_reg
*t
[4],
1415 struct nv50_reg
**src
, unsigned dim
, unsigned arg
)
1419 if (src
[0]->type
== P_TEMP
&& src
[0]->rhw
!= -1) {
1420 mode
= pc
->interp_mode
[src
[0]->index
] | INTERP_PERSPECTIVE
;
1422 t
[3]->rhw
= src
[3]->rhw
;
1423 emit_interp(pc
, t
[3], NULL
, (mode
& INTERP_CENTROID
));
1424 emit_flop(pc
, 0, t
[3], t
[3]);
1426 for (c
= 0; c
< dim
; ++c
) {
1427 t
[c
]->rhw
= src
[c
]->rhw
;
1428 emit_interp(pc
, t
[c
], t
[3], mode
);
1430 if (arg
!= dim
) { /* depth reference value */
1431 t
[dim
]->rhw
= src
[2]->rhw
;
1432 emit_interp(pc
, t
[dim
], t
[3], mode
);
1435 /* XXX: for some reason the blob sometimes uses MAD
1436 * (mad f32 $rX $rY $rZ neg $r63)
1438 emit_flop(pc
, 0, t
[3], src
[3]);
1439 for (c
= 0; c
< dim
; ++c
)
1440 emit_mul(pc
, t
[c
], src
[c
], t
[3]);
1441 if (arg
!= dim
) /* depth reference value */
1442 emit_mul(pc
, t
[dim
], src
[2], t
[3]);
1447 get_tex_dim(unsigned type
, unsigned *dim
, unsigned *arg
)
1450 case TGSI_TEXTURE_1D
:
1453 case TGSI_TEXTURE_SHADOW1D
:
1457 case TGSI_TEXTURE_UNKNOWN
:
1458 case TGSI_TEXTURE_2D
:
1459 case TGSI_TEXTURE_RECT
:
1462 case TGSI_TEXTURE_SHADOW2D
:
1463 case TGSI_TEXTURE_SHADOWRECT
:
1467 case TGSI_TEXTURE_3D
:
1468 case TGSI_TEXTURE_CUBE
:
1477 /* We shouldn't execute TEXLOD if any of the pixels in a quad have
1478 * different LOD values, so branch off groups of equal LOD.
1481 emit_texlod_sequence(struct nv50_pc
*pc
, struct nv50_reg
*tlod
,
1482 struct nv50_reg
*src
, struct nv50_program_exec
*tex
)
1484 struct nv50_program_exec
*join_at
;
1485 unsigned i
, target
= pc
->p
->exec_size
+ 7 * 2;
1487 /* Subtract lod of each pixel from lod of top left pixel, jump
1488 * texlod insn if result is 0, then repeat for 2 other pixels.
1490 emit_quadop(pc
, NULL
, 0, 0, tlod
, tlod
, 0x55);
1491 emit_branch(pc
, 0, 2, &join_at
)->param
.index
= target
;
1493 for (i
= 1; i
< 4; ++i
) {
1494 emit_quadop(pc
, NULL
, 0, i
, tlod
, tlod
, 0x55);
1495 emit_branch(pc
, 0, 2, NULL
)->param
.index
= target
;
1498 emit_mov(pc
, tlod
, src
); /* target */
1499 emit(pc
, tex
); /* texlod */
1501 join_at
->param
.index
= target
+ 2 * 2;
1503 pc
->p
->exec_tail
->inst
[1] |= 2; /* join _after_ tex */
1507 emit_texbias_sequence(struct nv50_pc
*pc
, struct nv50_reg
*t
[4], unsigned arg
,
1508 struct nv50_program_exec
*tex
)
1510 struct nv50_program_exec
*e
;
1511 struct nv50_reg imm_1248
, *t123
[4][4], *r_bits
= alloc_temp(pc
, NULL
);
1513 unsigned n
, c
, i
, cc
[4] = { 0x0a, 0x13, 0x11, 0x10 };
1515 pc
->allow32
= FALSE
;
1516 ctor_reg(&imm_1248
, P_IMMD
, -1, ctor_immd_4u32(pc
, 1, 2, 4, 8) * 4);
1518 /* Subtract bias value of thread i from bias values of each thread,
1519 * store result in r_pred, and set bit i in r_bits if result was 0.
1522 for (i
= 0; i
< 4; ++i
, ++imm_1248
.hw
) {
1523 emit_quadop(pc
, NULL
, r_pred
, i
, t
[arg
], t
[arg
], 0x55);
1524 emit_mov(pc
, r_bits
, &imm_1248
);
1525 set_pred(pc
, 2, r_pred
, pc
->p
->exec_tail
);
1527 emit_mov_to_pred(pc
, r_pred
, r_bits
);
1529 /* The lanes of a quad are now grouped by the bit in r_pred they have
1530 * set. Put the input values for TEX into a new register set for each
1531 * group and execute TEX only for a specific group.
1532 * We cannot use the same register set for each group because we need
1533 * the derivatives, which are implicitly calculated, to be correct.
1535 for (i
= 1; i
< 4; ++i
) {
1536 alloc_temp4(pc
, t123
[i
], 0);
1538 for (c
= 0; c
<= arg
; ++c
)
1539 emit_mov(pc
, t123
[i
][c
], t
[c
]);
1541 *(e
= exec(pc
)) = *(tex
);
1542 e
->inst
[0] &= ~0x01fc;
1543 set_dst(pc
, t123
[i
][0], e
);
1544 set_pred(pc
, cc
[i
], r_pred
, e
);
1547 /* finally TEX on the original regs (where we kept the input) */
1548 set_pred(pc
, cc
[0], r_pred
, tex
);
1551 /* put the 3 * n other results into regs for lane 0 */
1552 n
= popcnt4(((e
->inst
[0] >> 25) & 0x3) | ((e
->inst
[1] >> 12) & 0xc));
1553 for (i
= 1; i
< 4; ++i
) {
1554 for (c
= 0; c
< n
; ++c
) {
1555 emit_mov(pc
, t
[c
], t123
[i
][c
]);
1556 set_pred(pc
, cc
[i
], r_pred
, pc
->p
->exec_tail
);
1558 free_temp4(pc
, t123
[i
]);
1562 free_temp(pc
, r_bits
);
1566 emit_tex(struct nv50_pc
*pc
, struct nv50_reg
**dst
, unsigned mask
,
1567 struct nv50_reg
**src
, unsigned unit
, unsigned type
,
1568 boolean proj
, int bias_lod
)
1570 struct nv50_reg
*t
[4];
1571 struct nv50_program_exec
*e
;
1572 unsigned c
, dim
, arg
;
1574 /* t[i] must be within a single 128 bit super-reg */
1575 alloc_temp4(pc
, t
, 0);
1578 e
->inst
[0] = 0xf0000000;
1580 set_dst(pc
, t
[0], e
);
1582 /* TIC and TSC binding indices (TSC is ignored as TSC_LINKED = TRUE): */
1583 e
->inst
[0] |= (unit
<< 9) /* | (unit << 17) */;
1585 /* live flag (don't set if TEX results affect input to another TEX): */
1586 /* e->inst[0] |= 0x00000004; */
1588 get_tex_dim(type
, &dim
, &arg
);
1590 if (type
== TGSI_TEXTURE_CUBE
) {
1591 e
->inst
[0] |= 0x08000000;
1592 load_cube_tex_coords(pc
, t
, src
, arg
, proj
);
1595 load_proj_tex_coords(pc
, t
, src
, dim
, arg
);
1597 for (c
= 0; c
< dim
; c
++)
1598 emit_mov(pc
, t
[c
], src
[c
]);
1599 if (arg
!= dim
) /* depth reference value (always src.z here) */
1600 emit_mov(pc
, t
[dim
], src
[2]);
1603 e
->inst
[0] |= (mask
& 0x3) << 25;
1604 e
->inst
[1] |= (mask
& 0xc) << 12;
1607 e
->inst
[0] |= (arg
- 1) << 22;
1611 e
->inst
[0] |= arg
<< 22;
1612 e
->inst
[1] |= 0x20000000; /* texbias */
1613 emit_mov(pc
, t
[arg
], src
[3]);
1614 emit_texbias_sequence(pc
, t
, arg
, e
);
1616 e
->inst
[0] |= arg
<< 22;
1617 e
->inst
[1] |= 0x40000000; /* texlod */
1618 emit_mov(pc
, t
[arg
], src
[3]);
1619 emit_texlod_sequence(pc
, t
[arg
], src
[3], e
);
1624 if (mask
& 1) emit_mov(pc
, dst
[0], t
[c
++]);
1625 if (mask
& 2) emit_mov(pc
, dst
[1], t
[c
++]);
1626 if (mask
& 4) emit_mov(pc
, dst
[2], t
[c
++]);
1627 if (mask
& 8) emit_mov(pc
, dst
[3], t
[c
]);
1631 /* XXX: if p.e. MUL is used directly after TEX, it would still use
1632 * the texture coordinates, not the fetched values: latency ? */
1634 for (c
= 0; c
< 4; c
++) {
1635 if (mask
& (1 << c
))
1636 assimilate_temp(pc
, dst
[c
], t
[c
]);
1638 free_temp(pc
, t
[c
]);
1644 emit_ddx(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1646 struct nv50_program_exec
*e
= exec(pc
);
1648 assert(src
->type
== P_TEMP
);
1650 e
->inst
[0] = 0xc0140000;
1651 e
->inst
[1] = 0x89800000;
1653 set_dst(pc
, dst
, e
);
1654 set_src_0(pc
, src
, e
);
1655 set_src_2(pc
, src
, e
);
1661 emit_ddy(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1663 struct nv50_reg
*r
= src
;
1664 struct nv50_program_exec
*e
= exec(pc
);
1666 assert(src
->type
== P_TEMP
);
1668 if (!(src
->mod
& NV50_MOD_NEG
)) { /* ! double negation */
1669 r
= alloc_temp(pc
, NULL
);
1670 emit_neg(pc
, r
, src
);
1673 e
->inst
[0] = 0xc0150000;
1674 e
->inst
[1] = 0x8a400000;
1676 set_dst(pc
, dst
, e
);
1677 set_src_0(pc
, r
, e
);
1678 set_src_2(pc
, r
, e
);
1687 convert_to_long(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
1689 unsigned q
= 0, m
= ~0;
1691 assert(!is_long(e
));
1693 switch (e
->inst
[0] >> 28) {
1700 /* INTERP (move centroid, perspective and flat bits) */
1702 q
= (e
->inst
[0] & (3 << 24)) >> (24 - 16);
1703 q
|= (e
->inst
[0] & (1 << 8)) << (18 - 8);
1711 q
= ((e
->inst
[0] & (~m
)) >> 2);
1716 q
= ((e
->inst
[0] & (~m
)) << 12);
1719 /* MAD (if src2 == dst) */
1720 q
= ((e
->inst
[0] & 0x1fc) << 12);
1734 /* Some operations support an optional negation flag. */
1736 negate_supported(const struct tgsi_full_instruction
*insn
, int i
)
1738 switch (insn
->Instruction
.Opcode
) {
1739 case TGSI_OPCODE_DDY
:
1740 case TGSI_OPCODE_DP3
:
1741 case TGSI_OPCODE_DP4
:
1742 case TGSI_OPCODE_MUL
:
1743 case TGSI_OPCODE_KIL
:
1744 case TGSI_OPCODE_ADD
:
1745 case TGSI_OPCODE_SUB
:
1746 case TGSI_OPCODE_MAD
:
1748 case TGSI_OPCODE_POW
:
1757 /* Return a read mask for source registers deduced from opcode & write mask. */
1759 nv50_tgsi_src_mask(const struct tgsi_full_instruction
*insn
, int c
)
1761 unsigned x
, mask
= insn
->Dst
[0].Register
.WriteMask
;
1763 switch (insn
->Instruction
.Opcode
) {
1764 case TGSI_OPCODE_COS
:
1765 case TGSI_OPCODE_SIN
:
1766 return (mask
& 0x8) | ((mask
& 0x7) ? 0x1 : 0x0);
1767 case TGSI_OPCODE_DP3
:
1769 case TGSI_OPCODE_DP4
:
1770 case TGSI_OPCODE_DPH
:
1771 case TGSI_OPCODE_KIL
: /* WriteMask ignored */
1773 case TGSI_OPCODE_DST
:
1774 return mask
& (c
? 0xa : 0x6);
1775 case TGSI_OPCODE_EX2
:
1776 case TGSI_OPCODE_LG2
:
1777 case TGSI_OPCODE_POW
:
1778 case TGSI_OPCODE_RCP
:
1779 case TGSI_OPCODE_RSQ
:
1780 case TGSI_OPCODE_SCS
:
1782 case TGSI_OPCODE_IF
:
1784 case TGSI_OPCODE_LIT
:
1786 case TGSI_OPCODE_TEX
:
1787 case TGSI_OPCODE_TXB
:
1788 case TGSI_OPCODE_TXL
:
1789 case TGSI_OPCODE_TXP
:
1791 const struct tgsi_instruction_texture
*tex
;
1793 assert(insn
->Instruction
.Texture
);
1794 tex
= &insn
->Texture
;
1797 if (insn
->Instruction
.Opcode
!= TGSI_OPCODE_TEX
&&
1798 insn
->Instruction
.Opcode
!= TGSI_OPCODE_TXD
)
1799 mask
|= 0x8; /* bias, lod or proj */
1801 switch (tex
->Texture
) {
1802 case TGSI_TEXTURE_1D
:
1805 case TGSI_TEXTURE_SHADOW1D
:
1808 case TGSI_TEXTURE_2D
:
1816 case TGSI_OPCODE_XPD
:
1818 if (mask
& 1) x
|= 0x6;
1819 if (mask
& 2) x
|= 0x5;
1820 if (mask
& 4) x
|= 0x3;
1829 static struct nv50_reg
*
1830 tgsi_dst(struct nv50_pc
*pc
, int c
, const struct tgsi_full_dst_register
*dst
)
1832 switch (dst
->Register
.File
) {
1833 case TGSI_FILE_TEMPORARY
:
1834 return &pc
->temp
[dst
->Register
.Index
* 4 + c
];
1835 case TGSI_FILE_OUTPUT
:
1836 return &pc
->result
[dst
->Register
.Index
* 4 + c
];
1837 case TGSI_FILE_ADDRESS
:
1839 struct nv50_reg
*r
= pc
->addr
[dst
->Register
.Index
* 4 + c
];
1841 r
= alloc_addr(pc
, NULL
);
1842 pc
->addr
[dst
->Register
.Index
* 4 + c
] = r
;
1847 case TGSI_FILE_NULL
:
1856 static struct nv50_reg
*
1857 tgsi_src(struct nv50_pc
*pc
, int chan
, const struct tgsi_full_src_register
*src
,
1860 struct nv50_reg
*r
= NULL
;
1861 struct nv50_reg
*temp
;
1862 unsigned sgn
, c
, swz
;
1864 if (src
->Register
.File
!= TGSI_FILE_CONSTANT
)
1865 assert(!src
->Register
.Indirect
);
1867 sgn
= tgsi_util_get_full_src_register_sign_mode(src
, chan
);
1869 c
= tgsi_util_get_full_src_register_swizzle(src
, chan
);
1871 case TGSI_SWIZZLE_X
:
1872 case TGSI_SWIZZLE_Y
:
1873 case TGSI_SWIZZLE_Z
:
1874 case TGSI_SWIZZLE_W
:
1875 switch (src
->Register
.File
) {
1876 case TGSI_FILE_INPUT
:
1877 r
= &pc
->attr
[src
->Register
.Index
* 4 + c
];
1879 case TGSI_FILE_TEMPORARY
:
1880 r
= &pc
->temp
[src
->Register
.Index
* 4 + c
];
1882 case TGSI_FILE_CONSTANT
:
1883 if (!src
->Register
.Indirect
) {
1884 r
= &pc
->param
[src
->Register
.Index
* 4 + c
];
1887 /* Indicate indirection by setting r->acc < 0 and
1888 * use the index field to select the address reg.
1890 r
= reg_instance(pc
, NULL
);
1891 swz
= tgsi_util_get_src_register_swizzle(
1893 ctor_reg(r
, P_CONST
,
1894 src
->Indirect
.Index
* 4 + swz
,
1895 src
->Register
.Index
* 4 + c
);
1898 case TGSI_FILE_IMMEDIATE
:
1899 r
= &pc
->immd
[src
->Register
.Index
* 4 + c
];
1901 case TGSI_FILE_SAMPLER
:
1903 case TGSI_FILE_ADDRESS
:
1904 r
= pc
->addr
[src
->Register
.Index
* 4 + c
];
1918 case TGSI_UTIL_SIGN_KEEP
:
1920 case TGSI_UTIL_SIGN_CLEAR
:
1921 temp
= temp_temp(pc
);
1922 emit_abs(pc
, temp
, r
);
1925 case TGSI_UTIL_SIGN_TOGGLE
:
1927 r
->mod
= NV50_MOD_NEG
;
1929 temp
= temp_temp(pc
);
1930 emit_neg(pc
, temp
, r
);
1934 case TGSI_UTIL_SIGN_SET
:
1935 temp
= temp_temp(pc
);
1936 emit_cvt(pc
, temp
, r
, -1, CVTOP_ABS
, CVT_F32_F32
| CVT_NEG
);
1944 if (r
&& r
->acc
>= 0 && r
!= temp
)
1945 return reg_instance(pc
, r
);
1949 /* return TRUE for ops that produce only a single result */
1951 is_scalar_op(unsigned op
)
1954 case TGSI_OPCODE_COS
:
1955 case TGSI_OPCODE_DP2
:
1956 case TGSI_OPCODE_DP3
:
1957 case TGSI_OPCODE_DP4
:
1958 case TGSI_OPCODE_DPH
:
1959 case TGSI_OPCODE_EX2
:
1960 case TGSI_OPCODE_LG2
:
1961 case TGSI_OPCODE_POW
:
1962 case TGSI_OPCODE_RCP
:
1963 case TGSI_OPCODE_RSQ
:
1964 case TGSI_OPCODE_SIN
:
1966 case TGSI_OPCODE_KIL:
1967 case TGSI_OPCODE_LIT:
1968 case TGSI_OPCODE_SCS:
1976 /* Returns a bitmask indicating which dst components depend
1977 * on source s, component c (reverse of nv50_tgsi_src_mask).
1980 nv50_tgsi_dst_revdep(unsigned op
, int s
, int c
)
1982 if (is_scalar_op(op
))
1986 case TGSI_OPCODE_DST
:
1987 return (1 << c
) & (s
? 0xa : 0x6);
1988 case TGSI_OPCODE_XPD
:
1998 case TGSI_OPCODE_LIT
:
1999 case TGSI_OPCODE_SCS
:
2000 case TGSI_OPCODE_TEX
:
2001 case TGSI_OPCODE_TXB
:
2002 case TGSI_OPCODE_TXL
:
2003 case TGSI_OPCODE_TXP
:
2004 /* these take care of dangerous swizzles themselves */
2006 case TGSI_OPCODE_IF
:
2007 case TGSI_OPCODE_KIL
:
2008 /* don't call this function for these ops */
2012 /* linear vector instruction */
2017 static INLINE boolean
2018 has_pred(struct nv50_program_exec
*e
, unsigned cc
)
2020 if (!is_long(e
) || is_immd(e
))
2022 return ((e
->inst
[1] & 0x780) == (cc
<< 7));
2025 /* on ENDIF see if we can do "@p0.neu single_op" instead of:
2032 nv50_kill_branch(struct nv50_pc
*pc
)
2034 int lvl
= pc
->if_lvl
;
2036 if (pc
->if_insn
[lvl
]->next
!= pc
->p
->exec_tail
)
2039 /* if ccode == 'true', the BRA is from an ELSE and the predicate
2040 * reg may no longer be valid, since we currently always use $p0
2042 if (has_pred(pc
->if_insn
[lvl
], 0xf))
2044 assert(pc
->if_insn
[lvl
] && pc
->br_join
[lvl
]);
2046 /* We'll use the exec allocated for JOIN_AT (as we can't easily
2047 * update prev's next); if exec_tail is BRK, update the pointer.
2049 if (pc
->loop_lvl
&& pc
->br_loop
[pc
->loop_lvl
- 1] == pc
->p
->exec_tail
)
2050 pc
->br_loop
[pc
->loop_lvl
- 1] = pc
->br_join
[lvl
];
2052 pc
->p
->exec_size
-= 4; /* remove JOIN_AT and BRA */
2054 *pc
->br_join
[lvl
] = *pc
->p
->exec_tail
;
2056 FREE(pc
->if_insn
[lvl
]);
2057 FREE(pc
->p
->exec_tail
);
2059 pc
->p
->exec_tail
= pc
->br_join
[lvl
];
2060 pc
->p
->exec_tail
->next
= NULL
;
2061 set_pred(pc
, 0xd, 0, pc
->p
->exec_tail
);
2067 nv50_program_tx_insn(struct nv50_pc
*pc
,
2068 const struct tgsi_full_instruction
*inst
)
2070 struct nv50_reg
*rdst
[4], *dst
[4], *brdc
, *src
[3][4], *temp
;
2071 unsigned mask
, sat
, unit
;
2074 mask
= inst
->Dst
[0].Register
.WriteMask
;
2075 sat
= inst
->Instruction
.Saturate
== TGSI_SAT_ZERO_ONE
;
2077 memset(src
, 0, sizeof(src
));
2079 for (c
= 0; c
< 4; c
++) {
2080 if ((mask
& (1 << c
)) && !pc
->r_dst
[c
])
2081 dst
[c
] = tgsi_dst(pc
, c
, &inst
->Dst
[0]);
2083 dst
[c
] = pc
->r_dst
[c
];
2087 for (i
= 0; i
< inst
->Instruction
.NumSrcRegs
; i
++) {
2088 const struct tgsi_full_src_register
*fs
= &inst
->Src
[i
];
2092 src_mask
= nv50_tgsi_src_mask(inst
, i
);
2093 neg_supp
= negate_supported(inst
, i
);
2095 if (fs
->Register
.File
== TGSI_FILE_SAMPLER
)
2096 unit
= fs
->Register
.Index
;
2098 for (c
= 0; c
< 4; c
++)
2099 if (src_mask
& (1 << c
))
2100 src
[i
][c
] = 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
]);
2472 pc
->reg_instance_nr
= 0;
2478 prep_inspect_insn(struct nv50_pc
*pc
, const struct tgsi_full_instruction
*insn
)
2480 struct nv50_reg
*reg
= NULL
;
2481 const struct tgsi_full_src_register
*src
;
2482 const struct tgsi_dst_register
*dst
;
2483 unsigned i
, c
, k
, mask
;
2485 dst
= &insn
->Dst
[0].Register
;
2486 mask
= dst
->WriteMask
;
2488 if (dst
->File
== TGSI_FILE_TEMPORARY
)
2491 if (dst
->File
== TGSI_FILE_OUTPUT
)
2495 for (c
= 0; c
< 4; c
++) {
2496 if (!(mask
& (1 << c
)))
2498 reg
[dst
->Index
* 4 + c
].acc
= pc
->insn_nr
;
2502 for (i
= 0; i
< insn
->Instruction
.NumSrcRegs
; i
++) {
2503 src
= &insn
->Src
[i
];
2505 if (src
->Register
.File
== TGSI_FILE_TEMPORARY
)
2508 if (src
->Register
.File
== TGSI_FILE_INPUT
)
2513 mask
= nv50_tgsi_src_mask(insn
, i
);
2515 for (c
= 0; c
< 4; c
++) {
2516 if (!(mask
& (1 << c
)))
2518 k
= tgsi_util_get_full_src_register_swizzle(src
, c
);
2520 reg
[src
->Register
.Index
* 4 + k
].acc
= pc
->insn_nr
;
2525 /* Returns a bitmask indicating which dst components need to be
2526 * written to temporaries first to avoid 'corrupting' sources.
2528 * m[i] (out) indicate component to write in the i-th position
2529 * rdep[c] (in) bitmasks of dst[i] that require dst[c] as source
2532 nv50_revdep_reorder(unsigned m
[4], unsigned rdep
[4])
2534 unsigned i
, c
, x
, unsafe
;
2536 for (c
= 0; c
< 4; c
++)
2539 /* Swap as long as a dst component written earlier is depended on
2540 * by one written later, but the next one isn't depended on by it.
2542 for (c
= 0; c
< 3; c
++) {
2543 if (rdep
[m
[c
+ 1]] & (1 << m
[c
]))
2544 continue; /* if next one is depended on by us */
2545 for (i
= c
+ 1; i
< 4; i
++)
2546 /* if we are depended on by a later one */
2547 if (rdep
[m
[c
]] & (1 << m
[i
]))
2560 /* mark dependencies that could not be resolved by reordering */
2561 for (i
= 0; i
< 3; ++i
)
2562 for (c
= i
+ 1; c
< 4; ++c
)
2563 if (rdep
[m
[i
]] & (1 << m
[c
]))
2566 /* NOTE: $unsafe is with respect to order, not component */
2570 /* Select a suitable dst register for broadcasting scalar results,
2571 * or return NULL if we have to allocate an extra TEMP.
2573 * If e.g. only 1 component is written, we may also emit the final
2574 * result to a write-only register.
2576 static struct nv50_reg
*
2577 tgsi_broadcast_dst(struct nv50_pc
*pc
,
2578 const struct tgsi_full_dst_register
*fd
, unsigned mask
)
2580 if (fd
->Register
.File
== TGSI_FILE_TEMPORARY
) {
2581 int c
= ffs(~mask
& fd
->Register
.WriteMask
);
2583 return tgsi_dst(pc
, c
- 1, fd
);
2585 int c
= ffs(fd
->Register
.WriteMask
) - 1;
2586 if ((1 << c
) == fd
->Register
.WriteMask
)
2587 return tgsi_dst(pc
, c
, fd
);
2593 /* Scan source swizzles and return a bitmask indicating dst regs that
2594 * also occur among the src regs, and fill rdep for nv50_revdep_reoder.
2597 nv50_tgsi_scan_swizzle(const struct tgsi_full_instruction
*insn
,
2600 const struct tgsi_full_dst_register
*fd
= &insn
->Dst
[0];
2601 const struct tgsi_full_src_register
*fs
;
2602 unsigned i
, deqs
= 0;
2604 for (i
= 0; i
< 4; ++i
)
2607 for (i
= 0; i
< insn
->Instruction
.NumSrcRegs
; i
++) {
2608 unsigned chn
, mask
= nv50_tgsi_src_mask(insn
, i
);
2609 boolean neg_supp
= negate_supported(insn
, i
);
2612 if (fs
->Register
.File
!= fd
->Register
.File
||
2613 fs
->Register
.Index
!= fd
->Register
.Index
)
2616 for (chn
= 0; chn
< 4; ++chn
) {
2619 if (!(mask
& (1 << chn
))) /* src is not read */
2621 c
= tgsi_util_get_full_src_register_swizzle(fs
, chn
);
2622 s
= tgsi_util_get_full_src_register_sign_mode(fs
, chn
);
2624 if (!(fd
->Register
.WriteMask
& (1 << c
)))
2627 /* no danger if src is copied to TEMP first */
2628 if ((s
!= TGSI_UTIL_SIGN_KEEP
) &&
2629 (s
!= TGSI_UTIL_SIGN_TOGGLE
|| !neg_supp
))
2632 rdep
[c
] |= nv50_tgsi_dst_revdep(
2633 insn
->Instruction
.Opcode
, i
, chn
);
2642 nv50_tgsi_insn(struct nv50_pc
*pc
, const union tgsi_full_token
*tok
)
2644 struct tgsi_full_instruction insn
= tok
->FullInstruction
;
2645 const struct tgsi_full_dst_register
*fd
;
2646 unsigned i
, deqs
, rdep
[4], m
[4];
2648 fd
= &tok
->FullInstruction
.Dst
[0];
2649 deqs
= nv50_tgsi_scan_swizzle(&insn
, rdep
);
2651 if (is_scalar_op(insn
.Instruction
.Opcode
)) {
2652 pc
->r_brdc
= tgsi_broadcast_dst(pc
, fd
, deqs
);
2654 pc
->r_brdc
= temp_temp(pc
);
2655 return nv50_program_tx_insn(pc
, &insn
);
2660 return nv50_program_tx_insn(pc
, &insn
);
2662 deqs
= nv50_revdep_reorder(m
, rdep
);
2664 for (i
= 0; i
< 4; ++i
) {
2665 assert(pc
->r_dst
[m
[i
]] == NULL
);
2667 insn
.Dst
[0].Register
.WriteMask
=
2668 fd
->Register
.WriteMask
& (1 << m
[i
]);
2670 if (!insn
.Dst
[0].Register
.WriteMask
)
2673 if (deqs
& (1 << i
))
2674 pc
->r_dst
[m
[i
]] = alloc_temp(pc
, NULL
);
2676 if (!nv50_program_tx_insn(pc
, &insn
))
2680 for (i
= 0; i
< 4; i
++) {
2681 struct nv50_reg
*reg
= pc
->r_dst
[i
];
2684 pc
->r_dst
[i
] = NULL
;
2686 if (insn
.Instruction
.Saturate
== TGSI_SAT_ZERO_ONE
)
2687 emit_sat(pc
, tgsi_dst(pc
, i
, fd
), reg
);
2689 emit_mov(pc
, tgsi_dst(pc
, i
, fd
), reg
);
2697 load_interpolant(struct nv50_pc
*pc
, struct nv50_reg
*reg
)
2699 struct nv50_reg
*iv
, **ppiv
;
2700 unsigned mode
= pc
->interp_mode
[reg
->index
];
2702 ppiv
= (mode
& INTERP_CENTROID
) ? &pc
->iv_c
: &pc
->iv_p
;
2705 if ((mode
& INTERP_PERSPECTIVE
) && !iv
) {
2706 iv
= *ppiv
= alloc_temp(pc
, NULL
);
2707 iv
->rhw
= popcnt4(pc
->p
->cfg
.regs
[1] >> 24) - 1;
2709 emit_interp(pc
, iv
, NULL
, mode
& INTERP_CENTROID
);
2710 emit_flop(pc
, 0, iv
, iv
);
2712 /* XXX: when loading interpolants dynamically, move these
2713 * to the program head, or make sure it can't be skipped.
2717 emit_interp(pc
, reg
, iv
, mode
);
2720 /* The face input is always at v[255] (varying space), with a
2721 * value of 0 for back-facing, and 0xffffffff for front-facing.
2724 load_frontfacing(struct nv50_pc
*pc
, struct nv50_reg
*a
)
2726 struct nv50_reg
*one
= alloc_immd(pc
, 1.0f
);
2728 assert(a
->rhw
== -1);
2729 alloc_reg(pc
, a
); /* do this before rhw is set */
2731 load_interpolant(pc
, a
);
2732 emit_bitop2(pc
, a
, a
, one
, TGSI_OPCODE_AND
);
2738 nv50_program_tx_prep(struct nv50_pc
*pc
)
2740 struct tgsi_parse_context tp
;
2741 struct nv50_program
*p
= pc
->p
;
2742 boolean ret
= FALSE
;
2743 unsigned i
, c
, flat_nr
= 0;
2745 tgsi_parse_init(&tp
, pc
->p
->pipe
.tokens
);
2746 while (!tgsi_parse_end_of_tokens(&tp
)) {
2747 const union tgsi_full_token
*tok
= &tp
.FullToken
;
2749 tgsi_parse_token(&tp
);
2750 switch (tok
->Token
.Type
) {
2751 case TGSI_TOKEN_TYPE_IMMEDIATE
:
2753 const struct tgsi_full_immediate
*imm
=
2754 &tp
.FullToken
.FullImmediate
;
2756 ctor_immd_4f32(pc
, imm
->u
[0].Float
,
2762 case TGSI_TOKEN_TYPE_DECLARATION
:
2764 const struct tgsi_full_declaration
*d
;
2765 unsigned si
, last
, first
, mode
;
2767 d
= &tp
.FullToken
.FullDeclaration
;
2768 first
= d
->Range
.First
;
2769 last
= d
->Range
.Last
;
2771 switch (d
->Declaration
.File
) {
2772 case TGSI_FILE_TEMPORARY
:
2774 case TGSI_FILE_OUTPUT
:
2775 if (!d
->Declaration
.Semantic
||
2776 p
->type
== PIPE_SHADER_FRAGMENT
)
2779 si
= d
->Semantic
.Index
;
2780 switch (d
->Semantic
.Name
) {
2781 case TGSI_SEMANTIC_BCOLOR
:
2782 p
->cfg
.two_side
[si
].hw
= first
;
2783 if (p
->cfg
.io_nr
> first
)
2784 p
->cfg
.io_nr
= first
;
2786 case TGSI_SEMANTIC_PSIZE
:
2787 p
->cfg
.psiz
= first
;
2788 if (p
->cfg
.io_nr
> first
)
2789 p
->cfg
.io_nr
= first
;
2792 case TGSI_SEMANTIC_CLIP_DISTANCE:
2793 p->cfg.clpd = MIN2(p->cfg.clpd, first);
2800 case TGSI_FILE_INPUT
:
2802 if (p
->type
!= PIPE_SHADER_FRAGMENT
)
2805 switch (d
->Declaration
.Interpolate
) {
2806 case TGSI_INTERPOLATE_CONSTANT
:
2810 case TGSI_INTERPOLATE_PERSPECTIVE
:
2811 mode
= INTERP_PERSPECTIVE
;
2812 p
->cfg
.regs
[1] |= 0x08 << 24;
2815 mode
= INTERP_LINEAR
;
2818 if (d
->Declaration
.Centroid
)
2819 mode
|= INTERP_CENTROID
;
2822 for (i
= first
; i
<= last
; i
++)
2823 pc
->interp_mode
[i
] = mode
;
2826 case TGSI_FILE_ADDRESS
:
2827 case TGSI_FILE_CONSTANT
:
2828 case TGSI_FILE_SAMPLER
:
2831 NOUVEAU_ERR("bad decl file %d\n",
2832 d
->Declaration
.File
);
2837 case TGSI_TOKEN_TYPE_INSTRUCTION
:
2839 prep_inspect_insn(pc
, &tok
->FullInstruction
);
2846 if (p
->type
== PIPE_SHADER_VERTEX
) {
2849 for (i
= 0; i
< pc
->attr_nr
* 4; ++i
) {
2850 if (pc
->attr
[i
].acc
) {
2851 pc
->attr
[i
].hw
= rid
++;
2852 p
->cfg
.attr
[i
/ 32] |= 1 << (i
% 32);
2856 for (i
= 0, rid
= 0; i
< pc
->result_nr
; ++i
) {
2857 p
->cfg
.io
[i
].hw
= rid
;
2858 p
->cfg
.io
[i
].id
= i
;
2860 for (c
= 0; c
< 4; ++c
) {
2862 if (!pc
->result
[n
].acc
)
2864 pc
->result
[n
].hw
= rid
++;
2865 p
->cfg
.io
[i
].mask
|= 1 << c
;
2869 for (c
= 0; c
< 2; ++c
)
2870 if (p
->cfg
.two_side
[c
].hw
< 0x40)
2871 p
->cfg
.two_side
[c
] = p
->cfg
.io
[
2872 p
->cfg
.two_side
[c
].hw
];
2874 if (p
->cfg
.psiz
< 0x40)
2875 p
->cfg
.psiz
= p
->cfg
.io
[p
->cfg
.psiz
].hw
;
2877 if (p
->type
== PIPE_SHADER_FRAGMENT
) {
2879 unsigned n
= 0, m
= pc
->attr_nr
- flat_nr
;
2883 int base
= (TGSI_SEMANTIC_POSITION
==
2884 p
->info
.input_semantic_name
[0]) ? 0 : 1;
2886 /* non-flat interpolants have to be mapped to
2887 * the lower hardware IDs, so sort them:
2889 for (i
= 0; i
< pc
->attr_nr
; i
++) {
2890 if (pc
->interp_mode
[i
] == INTERP_FLAT
)
2891 p
->cfg
.io
[m
++].id
= i
;
2893 if (!(pc
->interp_mode
[i
] & INTERP_PERSPECTIVE
))
2894 p
->cfg
.io
[n
].linear
= TRUE
;
2895 p
->cfg
.io
[n
++].id
= i
;
2899 if (!base
) /* set w-coordinate mask from perspective interp */
2900 p
->cfg
.io
[0].mask
|= p
->cfg
.regs
[1] >> 24;
2902 aid
= popcnt4( /* if fcrd isn't contained in cfg.io */
2903 base
? (p
->cfg
.regs
[1] >> 24) : p
->cfg
.io
[0].mask
);
2905 for (n
= 0; n
< pc
->attr_nr
; ++n
) {
2906 p
->cfg
.io
[n
].hw
= rid
= aid
;
2907 i
= p
->cfg
.io
[n
].id
;
2909 if (p
->info
.input_semantic_name
[n
] ==
2910 TGSI_SEMANTIC_FACE
) {
2911 load_frontfacing(pc
, &pc
->attr
[i
* 4]);
2915 for (c
= 0; c
< 4; ++c
) {
2916 if (!pc
->attr
[i
* 4 + c
].acc
)
2918 pc
->attr
[i
* 4 + c
].rhw
= rid
++;
2919 p
->cfg
.io
[n
].mask
|= 1 << c
;
2921 load_interpolant(pc
, &pc
->attr
[i
* 4 + c
]);
2923 aid
+= popcnt4(p
->cfg
.io
[n
].mask
);
2927 p
->cfg
.regs
[1] |= p
->cfg
.io
[0].mask
<< 24;
2929 m
= popcnt4(p
->cfg
.regs
[1] >> 24);
2931 /* set count of non-position inputs and of non-flat
2932 * non-position inputs for FP_INTERPOLANT_CTRL
2934 p
->cfg
.regs
[1] |= aid
- m
;
2937 i
= p
->cfg
.io
[pc
->attr_nr
- flat_nr
].hw
;
2938 p
->cfg
.regs
[1] |= (i
- m
) << 16;
2940 p
->cfg
.regs
[1] |= p
->cfg
.regs
[1] << 16;
2942 /* mark color semantic for light-twoside */
2944 for (i
= 0; i
< pc
->attr_nr
; i
++) {
2947 sn
= p
->info
.input_semantic_name
[p
->cfg
.io
[i
].id
];
2948 si
= p
->info
.input_semantic_index
[p
->cfg
.io
[i
].id
];
2950 if (sn
== TGSI_SEMANTIC_COLOR
) {
2951 p
->cfg
.two_side
[si
] = p
->cfg
.io
[i
];
2953 /* increase colour count */
2954 p
->cfg
.regs
[0] += popcnt4(
2955 p
->cfg
.two_side
[si
].mask
) << 16;
2957 n
= MIN2(n
, p
->cfg
.io
[i
].hw
- m
);
2961 p
->cfg
.regs
[0] += n
;
2963 /* Initialize FP results:
2964 * FragDepth is always first TGSI and last hw output
2966 i
= p
->info
.writes_z
? 4 : 0;
2967 for (rid
= 0; i
< pc
->result_nr
* 4; i
++)
2968 pc
->result
[i
].rhw
= rid
++;
2969 if (p
->info
.writes_z
)
2970 pc
->result
[2].rhw
= rid
;
2972 p
->cfg
.high_result
= rid
;
2974 /* separate/different colour results for MRTs ? */
2975 if (pc
->result_nr
- (p
->info
.writes_z
? 1 : 0) > 1)
2976 p
->cfg
.regs
[2] |= 1;
2982 pc
->immd
= MALLOC(pc
->immd_nr
* 4 * sizeof(struct nv50_reg
));
2986 for (i
= 0; i
< pc
->immd_nr
; i
++) {
2987 for (c
= 0; c
< 4; c
++, rid
++)
2988 ctor_reg(&pc
->immd
[rid
], P_IMMD
, i
, rid
);
2995 free_temp(pc
, pc
->iv_p
);
2997 free_temp(pc
, pc
->iv_c
);
2999 tgsi_parse_free(&tp
);
3004 free_nv50_pc(struct nv50_pc
*pc
)
3021 ctor_nv50_pc(struct nv50_pc
*pc
, struct nv50_program
*p
)
3024 unsigned rtype
[2] = { P_ATTR
, P_RESULT
};
3027 pc
->temp_nr
= p
->info
.file_max
[TGSI_FILE_TEMPORARY
] + 1;
3028 pc
->attr_nr
= p
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
3029 pc
->result_nr
= p
->info
.file_max
[TGSI_FILE_OUTPUT
] + 1;
3030 pc
->param_nr
= p
->info
.file_max
[TGSI_FILE_CONSTANT
] + 1;
3031 pc
->addr_nr
= p
->info
.file_max
[TGSI_FILE_ADDRESS
] + 1;
3032 assert(pc
->addr_nr
<= 2);
3034 p
->cfg
.high_temp
= 4;
3036 p
->cfg
.two_side
[0].hw
= 0x40;
3037 p
->cfg
.two_side
[1].hw
= 0x40;
3040 case PIPE_SHADER_VERTEX
:
3043 p
->cfg
.io_nr
= pc
->result_nr
;
3045 case PIPE_SHADER_FRAGMENT
:
3046 rtype
[0] = rtype
[1] = P_TEMP
;
3048 p
->cfg
.regs
[0] = 0x01000004;
3049 p
->cfg
.io_nr
= pc
->attr_nr
;
3051 if (p
->info
.writes_z
) {
3052 p
->cfg
.regs
[2] |= 0x00000100;
3053 p
->cfg
.regs
[3] |= 0x00000011;
3055 if (p
->info
.uses_kill
)
3056 p
->cfg
.regs
[2] |= 0x00100000;
3061 pc
->temp
= MALLOC(pc
->temp_nr
* 4 * sizeof(struct nv50_reg
));
3065 for (i
= 0; i
< pc
->temp_nr
* 4; ++i
)
3066 ctor_reg(&pc
->temp
[i
], P_TEMP
, i
/ 4, -1);
3070 pc
->attr
= MALLOC(pc
->attr_nr
* 4 * sizeof(struct nv50_reg
));
3074 for (i
= 0; i
< pc
->attr_nr
* 4; ++i
)
3075 ctor_reg(&pc
->attr
[i
], rtype
[0], i
/ 4, -1);
3078 if (pc
->result_nr
) {
3079 unsigned nr
= pc
->result_nr
* 4;
3081 pc
->result
= MALLOC(nr
* sizeof(struct nv50_reg
));
3085 for (i
= 0; i
< nr
; ++i
)
3086 ctor_reg(&pc
->result
[i
], rtype
[1], i
/ 4, -1);
3092 pc
->param
= MALLOC(pc
->param_nr
* 4 * sizeof(struct nv50_reg
));
3096 for (i
= 0; i
< pc
->param_nr
; ++i
)
3097 for (c
= 0; c
< 4; ++c
, ++rid
)
3098 ctor_reg(&pc
->param
[rid
], P_CONST
, i
, rid
);
3102 pc
->addr
= CALLOC(pc
->addr_nr
* 4, sizeof(struct nv50_reg
*));
3106 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
)
3107 ctor_reg(&pc
->r_addr
[i
], P_ADDR
, -256, i
+ 1);
3113 nv50_fp_move_results(struct nv50_pc
*pc
)
3115 struct nv50_reg reg
;
3118 ctor_reg(®
, P_TEMP
, -1, -1);
3120 for (i
= 0; i
< pc
->result_nr
* 4; ++i
) {
3121 if (pc
->result
[i
].rhw
< 0 || pc
->result
[i
].hw
< 0)
3123 if (pc
->result
[i
].rhw
!= pc
->result
[i
].hw
) {
3124 reg
.hw
= pc
->result
[i
].rhw
;
3125 emit_mov(pc
, ®
, &pc
->result
[i
]);
3131 nv50_program_fixup_insns(struct nv50_pc
*pc
)
3133 struct nv50_program_exec
*e
, **bra_list
;
3136 bra_list
= CALLOC(pc
->p
->exec_size
, sizeof(struct nv50_program_exec
*));
3138 /* Collect branch instructions, we need to adjust their offsets
3139 * when converting 32 bit instructions to 64 bit ones
3141 for (n
= 0, e
= pc
->p
->exec_head
; e
; e
= e
->next
)
3142 if (e
->param
.index
>= 0 && !e
->param
.mask
)
3145 /* last instruction must be long so it can have the exit bit set */
3146 if (!is_long(pc
->p
->exec_tail
))
3147 convert_to_long(pc
, pc
->p
->exec_tail
);
3149 pc
->p
->exec_tail
->inst
[1] |= 1;
3151 /* !immd on exit insn simultaneously means !join */
3152 assert(!is_immd(pc
->p
->exec_head
));
3153 assert(!is_immd(pc
->p
->exec_tail
));
3155 /* Make sure we don't have any single 32 bit instructions. */
3156 for (e
= pc
->p
->exec_head
, pos
= 0; e
; e
= e
->next
) {
3157 pos
+= is_long(e
) ? 2 : 1;
3159 if ((pos
& 1) && (!e
->next
|| is_long(e
->next
))) {
3160 for (i
= 0; i
< n
; ++i
)
3161 if (bra_list
[i
]->param
.index
>= pos
)
3162 bra_list
[i
]->param
.index
+= 1;
3163 convert_to_long(pc
, e
);
3172 nv50_program_tx(struct nv50_program
*p
)
3174 struct tgsi_parse_context parse
;
3178 pc
= CALLOC_STRUCT(nv50_pc
);
3182 ret
= ctor_nv50_pc(pc
, p
);
3186 ret
= nv50_program_tx_prep(pc
);
3190 tgsi_parse_init(&parse
, pc
->p
->pipe
.tokens
);
3191 while (!tgsi_parse_end_of_tokens(&parse
)) {
3192 const union tgsi_full_token
*tok
= &parse
.FullToken
;
3194 /* don't allow half insn/immd on first and last instruction */
3196 if (pc
->insn_cur
== 0 || pc
->insn_cur
+ 2 == pc
->insn_nr
)
3197 pc
->allow32
= FALSE
;
3199 tgsi_parse_token(&parse
);
3201 switch (tok
->Token
.Type
) {
3202 case TGSI_TOKEN_TYPE_INSTRUCTION
:
3204 ret
= nv50_tgsi_insn(pc
, tok
);
3213 if (pc
->p
->type
== PIPE_SHADER_FRAGMENT
)
3214 nv50_fp_move_results(pc
);
3216 nv50_program_fixup_insns(pc
);
3218 p
->param_nr
= pc
->param_nr
* 4;
3219 p
->immd_nr
= pc
->immd_nr
* 4;
3220 p
->immd
= pc
->immd_buf
;
3223 tgsi_parse_free(&parse
);
3231 nv50_program_validate(struct nv50_context
*nv50
, struct nv50_program
*p
)
3233 if (nv50_program_tx(p
) == FALSE
)
3235 p
->translated
= TRUE
;
3239 nv50_program_upload_data(struct nv50_context
*nv50
, uint32_t *map
,
3240 unsigned start
, unsigned count
, unsigned cbuf
)
3242 struct nouveau_channel
*chan
= nv50
->screen
->base
.channel
;
3243 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
3246 unsigned nr
= count
> 2047 ? 2047 : count
;
3248 BEGIN_RING(chan
, tesla
, NV50TCL_CB_ADDR
, 1);
3249 OUT_RING (chan
, (cbuf
<< 0) | (start
<< 8));
3250 BEGIN_RING(chan
, tesla
, NV50TCL_CB_DATA(0) | 0x40000000, nr
);
3251 OUT_RINGp (chan
, map
, nr
);
3260 nv50_program_validate_data(struct nv50_context
*nv50
, struct nv50_program
*p
)
3262 struct pipe_screen
*pscreen
= nv50
->pipe
.screen
;
3264 if (!p
->data
[0] && p
->immd_nr
) {
3265 struct nouveau_resource
*heap
= nv50
->screen
->immd_heap
[0];
3267 if (nouveau_resource_alloc(heap
, p
->immd_nr
, p
, &p
->data
[0])) {
3268 while (heap
->next
&& heap
->size
< p
->immd_nr
) {
3269 struct nv50_program
*evict
= heap
->next
->priv
;
3270 nouveau_resource_free(&evict
->data
[0]);
3273 if (nouveau_resource_alloc(heap
, p
->immd_nr
, p
,
3278 /* immediates only need to be uploaded again when freed */
3279 nv50_program_upload_data(nv50
, p
->immd
, p
->data
[0]->start
,
3280 p
->immd_nr
, NV50_CB_PMISC
);
3283 assert(p
->param_nr
<= 512);
3287 uint32_t *map
= pipe_buffer_map(pscreen
, nv50
->constbuf
[p
->type
],
3288 PIPE_BUFFER_USAGE_CPU_READ
);
3290 if (p
->type
== PIPE_SHADER_VERTEX
)
3295 nv50_program_upload_data(nv50
, map
, 0, p
->param_nr
, cb
);
3296 pipe_buffer_unmap(pscreen
, nv50
->constbuf
[p
->type
]);
3301 nv50_program_validate_code(struct nv50_context
*nv50
, struct nv50_program
*p
)
3303 struct nouveau_channel
*chan
= nv50
->screen
->base
.channel
;
3304 struct nv50_program_exec
*e
;
3306 boolean upload
= FALSE
;
3309 nouveau_bo_new(chan
->device
, NOUVEAU_BO_VRAM
, 0x100,
3310 p
->exec_size
* 4, &p
->bo
);
3314 if (p
->data
[0] && p
->data
[0]->start
!= p
->data_start
[0])
3320 up
= MALLOC(p
->exec_size
* 4);
3322 for (i
= 0, e
= p
->exec_head
; e
; e
= e
->next
) {
3323 unsigned ei
, ci
, bs
;
3325 if (e
->param
.index
>= 0 && e
->param
.mask
) {
3326 bs
= (e
->inst
[1] >> 22) & 0x07;
3328 ei
= e
->param
.shift
>> 5;
3329 ci
= e
->param
.index
;
3331 ci
+= p
->data
[bs
]->start
;
3333 e
->inst
[ei
] &= ~e
->param
.mask
;
3334 e
->inst
[ei
] |= (ci
<< e
->param
.shift
);
3336 if (e
->param
.index
>= 0) {
3337 /* zero mask means param is a jump/branch offset */
3338 assert(!(e
->param
.index
& 1));
3339 /* seem to be 8 byte steps */
3340 ei
= (e
->param
.index
>> 1) + 0 /* START_ID */;
3342 e
->inst
[0] &= 0xf0000fff;
3343 e
->inst
[0] |= ei
<< 12;
3346 up
[i
++] = e
->inst
[0];
3348 up
[i
++] = e
->inst
[1];
3350 assert(i
== p
->exec_size
);
3353 p
->data_start
[0] = p
->data
[0]->start
;
3355 #ifdef NV50_PROGRAM_DUMP
3356 NOUVEAU_ERR("-------\n");
3357 for (e
= p
->exec_head
; e
; e
= e
->next
) {
3358 NOUVEAU_ERR("0x%08x\n", e
->inst
[0]);
3360 NOUVEAU_ERR("0x%08x\n", e
->inst
[1]);
3363 nv50_upload_sifc(nv50
, p
->bo
, 0, NOUVEAU_BO_VRAM
,
3364 NV50_2D_DST_FORMAT_R8_UNORM
, 65536, 1, 262144,
3365 up
, NV50_2D_SIFC_FORMAT_R8_UNORM
, 0,
3366 0, 0, p
->exec_size
* 4, 1, 1);
3372 nv50_vertprog_validate(struct nv50_context
*nv50
)
3374 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
3375 struct nv50_program
*p
= nv50
->vertprog
;
3376 struct nouveau_stateobj
*so
;
3378 if (!p
->translated
) {
3379 nv50_program_validate(nv50
, p
);
3384 nv50_program_validate_data(nv50
, p
);
3385 nv50_program_validate_code(nv50
, p
);
3388 so_method(so
, tesla
, NV50TCL_VP_ADDRESS_HIGH
, 2);
3389 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
3390 NOUVEAU_BO_HIGH
, 0, 0);
3391 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
3392 NOUVEAU_BO_LOW
, 0, 0);
3393 so_method(so
, tesla
, NV50TCL_VP_ATTR_EN_0
, 2);
3394 so_data (so
, p
->cfg
.attr
[0]);
3395 so_data (so
, p
->cfg
.attr
[1]);
3396 so_method(so
, tesla
, NV50TCL_VP_REG_ALLOC_RESULT
, 1);
3397 so_data (so
, p
->cfg
.high_result
);
3398 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP_SIZE
, 2);
3399 so_data (so
, p
->cfg
.high_result
); //8);
3400 so_data (so
, p
->cfg
.high_temp
);
3401 so_method(so
, tesla
, NV50TCL_VP_START_ID
, 1);
3402 so_data (so
, 0); /* program start offset */
3403 so_ref(so
, &nv50
->state
.vertprog
);
3408 nv50_fragprog_validate(struct nv50_context
*nv50
)
3410 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
3411 struct nv50_program
*p
= nv50
->fragprog
;
3412 struct nouveau_stateobj
*so
;
3414 if (!p
->translated
) {
3415 nv50_program_validate(nv50
, p
);
3420 nv50_program_validate_data(nv50
, p
);
3421 nv50_program_validate_code(nv50
, p
);
3424 so_method(so
, tesla
, NV50TCL_FP_ADDRESS_HIGH
, 2);
3425 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
3426 NOUVEAU_BO_HIGH
, 0, 0);
3427 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
3428 NOUVEAU_BO_LOW
, 0, 0);
3429 so_method(so
, tesla
, NV50TCL_FP_REG_ALLOC_TEMP
, 1);
3430 so_data (so
, p
->cfg
.high_temp
);
3431 so_method(so
, tesla
, NV50TCL_FP_RESULT_COUNT
, 1);
3432 so_data (so
, p
->cfg
.high_result
);
3433 so_method(so
, tesla
, NV50TCL_FP_CTRL_UNK19A8
, 1);
3434 so_data (so
, p
->cfg
.regs
[2]);
3435 so_method(so
, tesla
, NV50TCL_FP_CTRL_UNK196C
, 1);
3436 so_data (so
, p
->cfg
.regs
[3]);
3437 so_method(so
, tesla
, NV50TCL_FP_START_ID
, 1);
3438 so_data (so
, 0); /* program start offset */
3439 so_ref(so
, &nv50
->state
.fragprog
);
3444 nv50_pntc_replace(struct nv50_context
*nv50
, uint32_t pntc
[8], unsigned base
)
3446 struct nv50_program
*fp
= nv50
->fragprog
;
3447 struct nv50_program
*vp
= nv50
->vertprog
;
3448 unsigned i
, c
, m
= base
;
3450 /* XXX: this might not work correctly in all cases yet - we'll
3451 * just assume that an FP generic input that is not written in
3452 * the VP is PointCoord.
3454 memset(pntc
, 0, 8 * sizeof(uint32_t));
3456 for (i
= 0; i
< fp
->cfg
.io_nr
; i
++) {
3458 uint8_t j
, k
= fp
->cfg
.io
[i
].id
;
3459 unsigned n
= popcnt4(fp
->cfg
.io
[i
].mask
);
3461 if (fp
->info
.input_semantic_name
[k
] != TGSI_SEMANTIC_GENERIC
) {
3466 for (j
= 0; j
< vp
->info
.num_outputs
; ++j
) {
3467 sn
= vp
->info
.output_semantic_name
[j
];
3468 si
= vp
->info
.output_semantic_index
[j
];
3470 if (sn
== fp
->info
.input_semantic_name
[k
] &&
3471 si
== fp
->info
.input_semantic_index
[k
])
3475 if (j
< vp
->info
.num_outputs
) {
3477 nv50
->rasterizer
->pipe
.sprite_coord_mode
[si
];
3479 if (mode
== PIPE_SPRITE_COORD_NONE
) {
3485 /* this is either PointCoord or replaced by sprite coords */
3486 for (c
= 0; c
< 4; c
++) {
3487 if (!(fp
->cfg
.io
[i
].mask
& (1 << c
)))
3489 pntc
[m
/ 8] |= (c
+ 1) << ((m
% 8) * 4);
3496 nv50_sreg4_map(uint32_t *p_map
, int mid
, uint32_t lin
[4],
3497 struct nv50_sreg4
*fpi
, struct nv50_sreg4
*vpo
)
3500 uint8_t mv
= vpo
->mask
, mf
= fpi
->mask
, oid
= vpo
->hw
;
3501 uint8_t *map
= (uint8_t *)p_map
;
3503 for (c
= 0; c
< 4; ++c
) {
3505 if (fpi
->linear
== TRUE
)
3506 lin
[mid
/ 32] |= 1 << (mid
% 32);
3507 map
[mid
++] = (mv
& 1) ? oid
: ((c
== 3) ? 0x41 : 0x40);
3519 nv50_linkage_validate(struct nv50_context
*nv50
)
3521 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
3522 struct nv50_program
*vp
= nv50
->vertprog
;
3523 struct nv50_program
*fp
= nv50
->fragprog
;
3524 struct nouveau_stateobj
*so
;
3525 struct nv50_sreg4 dummy
, *vpo
;
3527 uint32_t map
[16], lin
[4], reg
[5], pcrd
[8];
3529 memset(map
, 0, sizeof(map
));
3530 memset(lin
, 0, sizeof(lin
));
3532 reg
[1] = 0x00000004; /* low and high clip distance map ids */
3533 reg
[2] = 0x00000000; /* layer index map id (disabled, GP only) */
3534 reg
[3] = 0x00000000; /* point size map id & enable */
3535 reg
[0] = fp
->cfg
.regs
[0]; /* colour semantic reg */
3536 reg
[4] = fp
->cfg
.regs
[1]; /* interpolant info */
3538 dummy
.linear
= FALSE
;
3539 dummy
.mask
= 0xf; /* map all components of HPOS */
3540 m
= nv50_sreg4_map(map
, m
, lin
, &dummy
, &vp
->cfg
.io
[0]);
3544 if (vp
->cfg
.clpd
< 0x40) {
3545 for (c
= 0; c
< vp
->cfg
.clpd_nr
; ++c
)
3546 map
[m
++] = vp
->cfg
.clpd
+ c
;
3550 reg
[0] |= m
<< 8; /* adjust BFC0 id */
3552 /* if light_twoside is active, it seems FFC0_ID == BFC0_ID is bad */
3553 if (nv50
->rasterizer
->pipe
.light_twoside
) {
3554 vpo
= &vp
->cfg
.two_side
[0];
3556 m
= nv50_sreg4_map(map
, m
, lin
, &fp
->cfg
.two_side
[0], &vpo
[0]);
3557 m
= nv50_sreg4_map(map
, m
, lin
, &fp
->cfg
.two_side
[1], &vpo
[1]);
3560 reg
[0] += m
- 4; /* adjust FFC0 id */
3561 reg
[4] |= m
<< 8; /* set mid where 'normal' FP inputs start */
3563 for (i
= 0; i
< fp
->cfg
.io_nr
; i
++) {
3564 ubyte sn
= fp
->info
.input_semantic_name
[fp
->cfg
.io
[i
].id
];
3565 ubyte si
= fp
->info
.input_semantic_index
[fp
->cfg
.io
[i
].id
];
3567 /* position must be mapped first */
3568 assert(i
== 0 || sn
!= TGSI_SEMANTIC_POSITION
);
3570 /* maybe even remove these from cfg.io */
3571 if (sn
== TGSI_SEMANTIC_POSITION
|| sn
== TGSI_SEMANTIC_FACE
)
3574 /* VP outputs and vp->cfg.io are in the same order */
3575 for (n
= 0; n
< vp
->info
.num_outputs
; ++n
) {
3576 if (vp
->info
.output_semantic_name
[n
] == sn
&&
3577 vp
->info
.output_semantic_index
[n
] == si
)
3580 vpo
= (n
< vp
->info
.num_outputs
) ? &vp
->cfg
.io
[n
] : &dummy
;
3582 m
= nv50_sreg4_map(map
, m
, lin
, &fp
->cfg
.io
[i
], vpo
);
3585 if (nv50
->rasterizer
->pipe
.point_size_per_vertex
) {
3586 map
[m
/ 4] |= vp
->cfg
.psiz
<< ((m
% 4) * 8);
3587 reg
[3] = (m
++ << 4) | 1;
3590 /* now fill the stateobj */
3594 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP_SIZE
, 1);
3596 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP(0), n
);
3597 so_datap (so
, map
, n
);
3599 so_method(so
, tesla
, NV50TCL_MAP_SEMANTIC_0
, 4);
3600 so_datap (so
, reg
, 4);
3602 so_method(so
, tesla
, NV50TCL_FP_INTERPOLANT_CTRL
, 1);
3603 so_data (so
, reg
[4]);
3605 so_method(so
, tesla
, 0x1540, 4);
3606 so_datap (so
, lin
, 4);
3608 if (nv50
->rasterizer
->pipe
.point_sprite
) {
3609 nv50_pntc_replace(nv50
, pcrd
, (reg
[4] >> 8) & 0xff);
3611 so_method(so
, tesla
, NV50TCL_POINT_COORD_REPLACE_MAP(0), 8);
3612 so_datap (so
, pcrd
, 8);
3615 so_ref(so
, &nv50
->state
.programs
);
3620 nv50_program_destroy(struct nv50_context
*nv50
, struct nv50_program
*p
)
3622 while (p
->exec_head
) {
3623 struct nv50_program_exec
*e
= p
->exec_head
;
3625 p
->exec_head
= e
->next
;
3628 p
->exec_tail
= NULL
;
3631 nouveau_bo_ref(NULL
, &p
->bo
);
3633 nouveau_resource_free(&p
->data
[0]);