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) */
96 int vtx
; /* vertex index, for GP inputs (TGSI Dimension.Index) */
97 int indirect
[2]; /* index into pc->addr, or -1 */
100 #define NV50_MOD_NEG 1
101 #define NV50_MOD_ABS 2
102 #define NV50_MOD_NEG_ABS (NV50_MOD_NEG | NV50_MOD_ABS)
103 #define NV50_MOD_SAT 4
104 #define NV50_MOD_I32 8
106 /* NV50_MOD_I32 is used to indicate integer mode for neg/abs */
108 /* STACK: Conditionals and loops have to use the (per warp) stack.
109 * Stack entries consist of an entry type (divergent path, join at),
110 * a mask indicating the active threads of the warp, and an address.
111 * MPs can store 12 stack entries internally, if we need more (and
112 * we probably do), we have to create a stack buffer in VRAM.
114 /* impose low limits for now */
115 #define NV50_MAX_COND_NESTING 4
116 #define NV50_MAX_LOOP_NESTING 3
118 #define JOIN_ON(e) e; pc->p->exec_tail->inst[1] |= 2
121 struct nv50_program
*p
;
124 struct nv50_reg
*r_temp
[NV50_SU_MAX_TEMP
];
125 struct nv50_reg r_addr
[NV50_SU_MAX_ADDR
];
128 struct nv50_reg
*temp
;
130 struct nv50_reg
*attr
;
132 struct nv50_reg
*result
;
134 struct nv50_reg
*param
;
136 struct nv50_reg
*immd
;
139 struct nv50_reg
**addr
;
141 struct nv50_reg
*sysval
;
144 struct nv50_reg
*temp_temp
[16];
145 struct nv50_program_exec
*temp_temp_exec
[16];
146 unsigned temp_temp_nr
;
148 /* broadcast and destination replacement regs */
149 struct nv50_reg
*r_brdc
;
150 struct nv50_reg
*r_dst
[4];
152 struct nv50_reg reg_instances
[16];
153 unsigned reg_instance_nr
;
155 unsigned interp_mode
[32];
156 /* perspective interpolation registers */
157 struct nv50_reg
*iv_p
;
158 struct nv50_reg
*iv_c
;
160 struct nv50_program_exec
*if_insn
[NV50_MAX_COND_NESTING
];
161 struct nv50_program_exec
*if_join
[NV50_MAX_COND_NESTING
];
162 struct nv50_program_exec
*loop_brka
[NV50_MAX_LOOP_NESTING
];
163 int if_lvl
, loop_lvl
;
164 unsigned loop_pos
[NV50_MAX_LOOP_NESTING
];
166 unsigned *insn_pos
; /* actual program offset of each TGSI insn */
167 boolean in_subroutine
;
169 /* current instruction and total number of insns */
175 uint8_t edgeflag_out
;
178 static struct nv50_reg
*get_address_reg(struct nv50_pc
*, struct nv50_reg
*);
181 ctor_reg(struct nv50_reg
*reg
, unsigned type
, int index
, int hw
)
190 reg
->indirect
[0] = reg
->indirect
[1] = -1;
193 static INLINE
unsigned
194 popcnt4(uint32_t val
)
196 static const unsigned cnt
[16]
197 = { 0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4 };
198 return cnt
[val
& 0xf];
202 terminate_mbb(struct nv50_pc
*pc
)
206 /* remove records of temporary address register values */
207 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
)
208 if (pc
->r_addr
[i
].index
< 0)
209 pc
->r_addr
[i
].acc
= 0;
213 alloc_reg(struct nv50_pc
*pc
, struct nv50_reg
*reg
)
217 if (reg
->type
== P_RESULT
) {
218 if (pc
->p
->cfg
.high_result
< (reg
->hw
+ 1))
219 pc
->p
->cfg
.high_result
= reg
->hw
+ 1;
222 if (reg
->type
!= P_TEMP
)
226 /*XXX: do this here too to catch FP temp-as-attr usage..
227 * not clean, but works */
228 if (pc
->p
->cfg
.high_temp
< (reg
->hw
+ 1))
229 pc
->p
->cfg
.high_temp
= reg
->hw
+ 1;
233 if (reg
->rhw
!= -1) {
234 /* try to allocate temporary with index rhw first */
235 if (!(pc
->r_temp
[reg
->rhw
])) {
236 pc
->r_temp
[reg
->rhw
] = reg
;
238 if (pc
->p
->cfg
.high_temp
< (reg
->rhw
+ 1))
239 pc
->p
->cfg
.high_temp
= reg
->rhw
+ 1;
242 /* make sure we don't get things like $r0 needs to go
243 * in $r1 and $r1 in $r0
245 i
= pc
->result_nr
* 4;
248 for (; i
< NV50_SU_MAX_TEMP
; i
++) {
249 if (!(pc
->r_temp
[i
])) {
252 if (pc
->p
->cfg
.high_temp
< (i
+ 1))
253 pc
->p
->cfg
.high_temp
= i
+ 1;
258 NOUVEAU_ERR("out of registers\n");
262 static INLINE
struct nv50_reg
*
263 reg_instance(struct nv50_pc
*pc
, struct nv50_reg
*reg
)
267 assert(pc
->reg_instance_nr
< 16);
268 ri
= &pc
->reg_instances
[pc
->reg_instance_nr
++];
272 reg
->indirect
[0] = reg
->indirect
[1] = -1;
278 /* XXX: For shaders that aren't executed linearly (e.g. shaders that
279 * contain loops), we need to assign all hw regs to TGSI TEMPs early,
280 * lest we risk temp_temps overwriting regs alloc'd "later".
282 static struct nv50_reg
*
283 alloc_temp(struct nv50_pc
*pc
, struct nv50_reg
*dst
)
288 if (dst
&& dst
->type
== P_TEMP
&& dst
->hw
== -1)
291 for (i
= 0; i
< NV50_SU_MAX_TEMP
; i
++) {
292 if (!pc
->r_temp
[i
]) {
293 r
= MALLOC_STRUCT(nv50_reg
);
294 ctor_reg(r
, P_TEMP
, -1, i
);
300 NOUVEAU_ERR("out of registers\n");
305 /* release the hardware resource held by r */
307 release_hw(struct nv50_pc
*pc
, struct nv50_reg
*r
)
309 assert(r
->type
== P_TEMP
);
313 assert(pc
->r_temp
[r
->hw
] == r
);
314 pc
->r_temp
[r
->hw
] = NULL
;
322 free_temp(struct nv50_pc
*pc
, struct nv50_reg
*r
)
324 if (r
->index
== -1) {
327 FREE(pc
->r_temp
[hw
]);
328 pc
->r_temp
[hw
] = NULL
;
333 alloc_temp4(struct nv50_pc
*pc
, struct nv50_reg
*dst
[4], int idx
)
337 if ((idx
+ 4) >= NV50_SU_MAX_TEMP
)
340 if (pc
->r_temp
[idx
] || pc
->r_temp
[idx
+ 1] ||
341 pc
->r_temp
[idx
+ 2] || pc
->r_temp
[idx
+ 3])
342 return alloc_temp4(pc
, dst
, idx
+ 4);
344 for (i
= 0; i
< 4; i
++) {
345 dst
[i
] = MALLOC_STRUCT(nv50_reg
);
346 ctor_reg(dst
[i
], P_TEMP
, -1, idx
+ i
);
347 pc
->r_temp
[idx
+ i
] = dst
[i
];
354 free_temp4(struct nv50_pc
*pc
, struct nv50_reg
*reg
[4])
358 for (i
= 0; i
< 4; i
++)
359 free_temp(pc
, reg
[i
]);
362 static struct nv50_reg
*
363 temp_temp(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
365 if (pc
->temp_temp_nr
>= 16)
368 pc
->temp_temp
[pc
->temp_temp_nr
] = alloc_temp(pc
, NULL
);
369 pc
->temp_temp_exec
[pc
->temp_temp_nr
] = e
;
370 return pc
->temp_temp
[pc
->temp_temp_nr
++];
373 /* This *must* be called for all nv50_program_exec that have been
374 * given as argument to temp_temp, or the temps will be leaked !
377 kill_temp_temp(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
381 for (i
= 0; i
< pc
->temp_temp_nr
; i
++)
382 if (pc
->temp_temp_exec
[i
] == e
)
383 free_temp(pc
, pc
->temp_temp
[i
]);
385 pc
->temp_temp_nr
= 0;
389 ctor_immd_4u32(struct nv50_pc
*pc
,
390 uint32_t x
, uint32_t y
, uint32_t z
, uint32_t w
)
392 unsigned size
= pc
->immd_nr
* 4 * sizeof(uint32_t);
394 pc
->immd_buf
= REALLOC(pc
->immd_buf
, size
, size
+ 4 * sizeof(uint32_t));
396 pc
->immd_buf
[(pc
->immd_nr
* 4) + 0] = x
;
397 pc
->immd_buf
[(pc
->immd_nr
* 4) + 1] = y
;
398 pc
->immd_buf
[(pc
->immd_nr
* 4) + 2] = z
;
399 pc
->immd_buf
[(pc
->immd_nr
* 4) + 3] = w
;
401 return pc
->immd_nr
++;
405 ctor_immd_4f32(struct nv50_pc
*pc
, float x
, float y
, float z
, float w
)
407 return ctor_immd_4u32(pc
, fui(x
), fui(y
), fui(z
), fui(w
));
410 static struct nv50_reg
*
411 alloc_immd(struct nv50_pc
*pc
, float f
)
413 struct nv50_reg
*r
= MALLOC_STRUCT(nv50_reg
);
416 for (hw
= 0; hw
< pc
->immd_nr
* 4; hw
++)
417 if (pc
->immd_buf
[hw
] == fui(f
))
420 if (hw
== pc
->immd_nr
* 4)
421 hw
= ctor_immd_4f32(pc
, f
, -f
, 0.5 * f
, 0) * 4;
423 ctor_reg(r
, P_IMMD
, -1, hw
);
427 static struct nv50_program_exec
*
428 exec(struct nv50_pc
*pc
)
430 struct nv50_program_exec
*e
= CALLOC_STRUCT(nv50_program_exec
);
437 emit(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
439 struct nv50_program
*p
= pc
->p
;
442 p
->exec_tail
->next
= e
;
446 p
->exec_size
+= (e
->inst
[0] & 1) ? 2 : 1;
448 kill_temp_temp(pc
, e
);
451 static INLINE
void set_long(struct nv50_pc
*, struct nv50_program_exec
*);
454 is_long(struct nv50_program_exec
*e
)
462 is_immd(struct nv50_program_exec
*e
)
464 if (is_long(e
) && (e
->inst
[1] & 3) == 3)
470 is_join(struct nv50_program_exec
*e
)
472 if (is_long(e
) && (e
->inst
[1] & 3) == 2)
478 set_pred(struct nv50_pc
*pc
, unsigned pred
, unsigned idx
,
479 struct nv50_program_exec
*e
)
483 e
->inst
[1] &= ~((0x1f << 7) | (0x3 << 12));
484 e
->inst
[1] |= (pred
<< 7) | (idx
<< 12);
488 set_pred_wr(struct nv50_pc
*pc
, unsigned on
, unsigned idx
,
489 struct nv50_program_exec
*e
)
492 e
->inst
[1] &= ~((0x3 << 4) | (1 << 6));
493 e
->inst
[1] |= (idx
<< 4) | (on
<< 6);
497 set_long(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
503 set_pred(pc
, 0xf, 0, e
);
504 set_pred_wr(pc
, 0, 0, e
);
508 set_dst(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_program_exec
*e
)
510 if (dst
->type
== P_RESULT
) {
512 e
->inst
[1] |= 0x00000008;
518 e
->inst
[0] |= (dst
->hw
<< 2);
522 set_immd(struct nv50_pc
*pc
, struct nv50_reg
*imm
, struct nv50_program_exec
*e
)
525 /* XXX: can't be predicated - bits overlap; cases where both
526 * are required should be avoided by using pc->allow32 */
527 set_pred(pc
, 0, 0, e
);
528 set_pred_wr(pc
, 0, 0, e
);
530 e
->inst
[1] |= 0x00000002 | 0x00000001;
531 e
->inst
[0] |= (pc
->immd_buf
[imm
->hw
] & 0x3f) << 16;
532 e
->inst
[1] |= (pc
->immd_buf
[imm
->hw
] >> 6) << 2;
536 set_addr(struct nv50_program_exec
*e
, struct nv50_reg
*a
)
538 assert(a
->type
== P_ADDR
);
540 assert(!(e
->inst
[0] & 0x0c000000));
541 assert(!(e
->inst
[1] & 0x00000004));
543 e
->inst
[0] |= (a
->hw
& 3) << 26;
544 e
->inst
[1] |= a
->hw
& 4;
548 emit_arl(struct nv50_pc
*, struct nv50_reg
*, struct nv50_reg
*, uint8_t);
551 emit_shl_imm(struct nv50_pc
*, struct nv50_reg
*, struct nv50_reg
*, int);
554 emit_mov_from_addr(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
555 struct nv50_reg
*src
)
557 struct nv50_program_exec
*e
= exec(pc
);
559 e
->inst
[1] = 0x40000000;
568 emit_add_addr_imm(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
569 struct nv50_reg
*src0
, uint16_t src1_val
)
571 struct nv50_program_exec
*e
= exec(pc
);
573 e
->inst
[0] = 0xd0000000 | (src1_val
<< 9);
574 e
->inst
[1] = 0x20000000;
576 e
->inst
[0] |= dst
->hw
<< 2;
577 if (src0
) /* otherwise will add to $a0, which is always 0 */
583 #define INTERP_LINEAR 0
584 #define INTERP_FLAT 1
585 #define INTERP_PERSPECTIVE 2
586 #define INTERP_CENTROID 4
588 /* interpolant index has been stored in dst->rhw */
590 emit_interp(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*iv
,
593 assert(dst
->rhw
!= -1);
594 struct nv50_program_exec
*e
= exec(pc
);
596 e
->inst
[0] |= 0x80000000;
598 e
->inst
[0] |= (dst
->rhw
<< 16);
600 if (mode
& INTERP_FLAT
) {
601 e
->inst
[0] |= (1 << 8);
603 if (mode
& INTERP_PERSPECTIVE
) {
604 e
->inst
[0] |= (1 << 25);
606 e
->inst
[0] |= (iv
->hw
<< 9);
609 if (mode
& INTERP_CENTROID
)
610 e
->inst
[0] |= (1 << 24);
617 set_data(struct nv50_pc
*pc
, struct nv50_reg
*src
, unsigned m
, unsigned s
,
618 struct nv50_program_exec
*e
)
622 e
->param
.index
= src
->hw
& 127;
624 e
->param
.mask
= m
<< (s
% 32);
626 if (src
->hw
< 0 || src
->hw
> 127) /* need (additional) address reg */
627 set_addr(e
, get_address_reg(pc
, src
));
630 assert(src
->type
== P_CONST
);
631 set_addr(e
, pc
->addr
[src
->indirect
[0]]);
634 e
->inst
[1] |= (((src
->type
== P_IMMD
) ? 0 : 1) << 22);
637 /* Never apply nv50_reg::mod in emit_mov, or carefully check the code !!! */
639 emit_mov(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
641 struct nv50_program_exec
*e
= exec(pc
);
643 e
->inst
[0] = 0x10000000;
649 if (!is_long(e
) && src
->type
== P_IMMD
) {
650 set_immd(pc
, src
, e
);
651 /*XXX: 32-bit, but steals part of "half" reg space - need to
652 * catch and handle this case if/when we do half-regs
655 if (src
->type
== P_IMMD
|| src
->type
== P_CONST
) {
657 set_data(pc
, src
, 0x7f, 9, e
);
658 e
->inst
[1] |= 0x20000000; /* mov from c[] */
660 if (src
->type
== P_ATTR
) {
662 e
->inst
[1] |= 0x00200000;
665 /* indirect (vertex base + c) load from p[] */
666 e
->inst
[0] |= 0x01800000;
667 set_addr(e
, get_address_reg(pc
, src
));
674 e
->inst
[0] |= (src
->hw
<< 9);
677 if (is_long(e
) && !is_immd(e
)) {
678 e
->inst
[1] |= 0x04000000; /* 32-bit */
679 e
->inst
[1] |= 0x0000c000; /* 32-bit c[] load / lane mask 0:1 */
680 if (!(e
->inst
[1] & 0x20000000))
681 e
->inst
[1] |= 0x00030000; /* lane mask 2:3 */
683 e
->inst
[0] |= 0x00008000;
689 emit_mov_immdval(struct nv50_pc
*pc
, struct nv50_reg
*dst
, float f
)
691 struct nv50_reg
*imm
= alloc_immd(pc
, f
);
692 emit_mov(pc
, dst
, imm
);
696 /* Assign the hw of the discarded temporary register src
697 * to the tgsi register dst and free src.
700 assimilate_temp(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
702 assert(src
->index
== -1 && src
->hw
!= -1);
704 if (pc
->if_lvl
|| pc
->loop_lvl
||
705 (dst
->type
!= P_TEMP
) ||
706 (src
->hw
< pc
->result_nr
* 4 &&
707 pc
->p
->type
== PIPE_SHADER_FRAGMENT
) ||
708 pc
->p
->info
.opcode_count
[TGSI_OPCODE_CAL
] ||
709 pc
->p
->info
.opcode_count
[TGSI_OPCODE_BRA
]) {
711 emit_mov(pc
, dst
, src
);
717 pc
->r_temp
[dst
->hw
] = NULL
;
718 pc
->r_temp
[src
->hw
] = dst
;
725 emit_nop(struct nv50_pc
*pc
)
727 struct nv50_program_exec
*e
= exec(pc
);
729 e
->inst
[0] = 0xf0000000;
731 e
->inst
[1] = 0xe0000000;
736 check_swap_src_0_1(struct nv50_pc
*pc
,
737 struct nv50_reg
**s0
, struct nv50_reg
**s1
)
739 struct nv50_reg
*src0
= *s0
, *src1
= *s1
;
741 if (src0
->type
== P_CONST
) {
742 if (src1
->type
!= P_CONST
) {
748 if (src1
->type
== P_ATTR
) {
749 if (src0
->type
!= P_ATTR
) {
760 set_src_0_restricted(struct nv50_pc
*pc
, struct nv50_reg
*src
,
761 struct nv50_program_exec
*e
)
763 struct nv50_reg
*temp
;
765 if (src
->type
!= P_TEMP
) {
766 temp
= temp_temp(pc
, e
);
767 emit_mov(pc
, temp
, src
);
774 e
->inst
[0] |= (src
->hw
<< 9);
778 set_src_0(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
780 if (src
->type
== P_ATTR
) {
782 e
->inst
[1] |= 0x00200000;
785 e
->inst
[0] |= 0x01800000; /* src from p[] */
786 set_addr(e
, get_address_reg(pc
, src
));
789 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
790 struct nv50_reg
*temp
= temp_temp(pc
, e
);
792 emit_mov(pc
, temp
, src
);
799 e
->inst
[0] |= (src
->hw
<< 9);
803 set_src_1(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
805 if (src
->type
== P_ATTR
) {
806 struct nv50_reg
*temp
= temp_temp(pc
, e
);
808 emit_mov(pc
, temp
, src
);
811 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
812 if (e
->inst
[0] & 0x01800000) {
813 struct nv50_reg
*temp
= temp_temp(pc
, e
);
815 emit_mov(pc
, temp
, src
);
818 assert(!(e
->inst
[0] & 0x00800000));
819 set_data(pc
, src
, 0x7f, 16, e
);
820 e
->inst
[0] |= 0x00800000;
827 e
->inst
[0] |= ((src
->hw
& 127) << 16);
831 set_src_2(struct nv50_pc
*pc
, struct nv50_reg
*src
, struct nv50_program_exec
*e
)
835 if (src
->type
== P_ATTR
) {
836 struct nv50_reg
*temp
= temp_temp(pc
, e
);
838 emit_mov(pc
, temp
, src
);
841 if (src
->type
== P_CONST
|| src
->type
== P_IMMD
) {
842 if (e
->inst
[0] & 0x01800000) {
843 struct nv50_reg
*temp
= temp_temp(pc
, e
);
845 emit_mov(pc
, temp
, src
);
848 assert(!(e
->inst
[0] & 0x01000000));
849 set_data(pc
, src
, 0x7f, 32+14, e
);
850 e
->inst
[0] |= 0x01000000;
855 e
->inst
[1] |= ((src
->hw
& 127) << 14);
859 set_half_src(struct nv50_pc
*pc
, struct nv50_reg
*src
, int lh
,
860 struct nv50_program_exec
*e
, int pos
)
862 struct nv50_reg
*r
= src
;
865 if (r
->type
!= P_TEMP
) {
866 r
= temp_temp(pc
, e
);
867 emit_mov(pc
, r
, src
);
870 if (r
->hw
> (NV50_SU_MAX_TEMP
/ 2)) {
871 NOUVEAU_ERR("out of low GPRs\n");
875 e
->inst
[pos
/ 32] |= ((src
->hw
* 2) + lh
) << (pos
% 32);
879 emit_mov_from_pred(struct nv50_pc
*pc
, struct nv50_reg
*dst
, int pred
)
881 struct nv50_program_exec
*e
= exec(pc
);
883 assert(dst
->type
== P_TEMP
);
884 e
->inst
[1] = 0x20000000 | (pred
<< 12);
892 emit_mov_to_pred(struct nv50_pc
*pc
, int pred
, struct nv50_reg
*src
)
894 struct nv50_program_exec
*e
= exec(pc
);
896 e
->inst
[0] = 0x000001fc;
897 e
->inst
[1] = 0xa0000008;
899 set_pred_wr(pc
, 1, pred
, e
);
900 set_src_0_restricted(pc
, src
, e
);
906 emit_mul(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
907 struct nv50_reg
*src1
)
909 struct nv50_program_exec
*e
= exec(pc
);
911 e
->inst
[0] |= 0xc0000000;
916 check_swap_src_0_1(pc
, &src0
, &src1
);
918 set_src_0(pc
, src0
, e
);
919 if (src1
->type
== P_IMMD
&& !is_long(e
)) {
920 if (src0
->mod
^ src1
->mod
)
921 e
->inst
[0] |= 0x00008000;
922 set_immd(pc
, src1
, e
);
924 set_src_1(pc
, src1
, e
);
925 if ((src0
->mod
^ src1
->mod
) & NV50_MOD_NEG
) {
927 e
->inst
[1] |= 0x08000000;
929 e
->inst
[0] |= 0x00008000;
937 emit_add(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
938 struct nv50_reg
*src0
, struct nv50_reg
*src1
)
940 struct nv50_program_exec
*e
= exec(pc
);
942 e
->inst
[0] = 0xb0000000;
945 check_swap_src_0_1(pc
, &src0
, &src1
);
947 if (!pc
->allow32
|| (src0
->mod
| src1
->mod
) || src1
->hw
> 63) {
949 e
->inst
[1] |= ((src0
->mod
& NV50_MOD_NEG
) << 26) |
950 ((src1
->mod
& NV50_MOD_NEG
) << 27);
954 set_src_0(pc
, src0
, e
);
955 if (src1
->type
== P_CONST
|| src1
->type
== P_ATTR
|| is_long(e
))
956 set_src_2(pc
, src1
, e
);
958 if (src1
->type
== P_IMMD
)
959 set_immd(pc
, src1
, e
);
961 set_src_1(pc
, src1
, e
);
967 emit_arl(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
,
970 struct nv50_program_exec
*e
= exec(pc
);
973 e
->inst
[1] |= 0xc0000000;
975 e
->inst
[0] |= dst
->hw
<< 2;
976 e
->inst
[0] |= s
<< 16; /* shift left */
977 set_src_0(pc
, src
, e
);
983 address_reg_suitable(struct nv50_reg
*a
, struct nv50_reg
*r
)
988 if (r
->vtx
!= a
->vtx
)
991 return (r
->indirect
[1] == a
->indirect
[1]);
993 if (r
->hw
< a
->rhw
|| (r
->hw
- a
->rhw
) >= 128)
997 return (a
->index
== r
->indirect
[0]);
998 return (a
->indirect
[0] == r
->indirect
[0]);
1002 load_vertex_base(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1003 struct nv50_reg
*a
, int shift
)
1005 struct nv50_reg mem
, *temp
;
1007 ctor_reg(&mem
, P_ATTR
, -1, dst
->vtx
);
1009 assert(dst
->type
== P_ADDR
);
1011 emit_arl(pc
, dst
, &mem
, 0);
1014 temp
= alloc_temp(pc
, NULL
);
1017 emit_mov_from_addr(pc
, temp
, a
);
1019 emit_shl_imm(pc
, temp
, temp
, shift
);
1020 emit_arl(pc
, dst
, temp
, MAX2(shift
, 0));
1022 emit_mov(pc
, temp
, &mem
);
1023 set_addr(pc
->p
->exec_tail
, dst
);
1025 emit_arl(pc
, dst
, temp
, 0);
1026 free_temp(pc
, temp
);
1029 /* case (ref == NULL): allocate address register for TGSI_FILE_ADDRESS
1030 * case (vtx >= 0, acc >= 0): load vertex base from a[vtx * 4] to $aX
1031 * case (vtx >= 0, acc < 0): load vertex base from s[$aY + vtx * 4] to $aX
1032 * case (vtx < 0, acc >= 0): memory address too high to encode
1033 * case (vtx < 0, acc < 0): get source register for TGSI_FILE_ADDRESS
1035 static struct nv50_reg
*
1036 get_address_reg(struct nv50_pc
*pc
, struct nv50_reg
*ref
)
1039 struct nv50_reg
*a_ref
, *a
= NULL
;
1041 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
) {
1042 if (pc
->r_addr
[i
].acc
== 0)
1043 a
= &pc
->r_addr
[i
]; /* an unused address reg */
1045 if (address_reg_suitable(&pc
->r_addr
[i
], ref
)) {
1046 pc
->r_addr
[i
].acc
= pc
->insn_cur
;
1047 return &pc
->r_addr
[i
];
1049 if (!a
&& pc
->r_addr
[i
].index
< 0 &&
1050 pc
->r_addr
[i
].acc
< pc
->insn_cur
)
1054 /* We'll be able to spill address regs when this
1055 * mess is replaced with a proper compiler ...
1057 NOUVEAU_ERR("out of address regs\n");
1062 /* initialize and reserve for this TGSI instruction */
1064 a
->index
= a
->indirect
[0] = a
->indirect
[1] = -1;
1065 a
->acc
= pc
->insn_cur
;
1073 /* now put in the correct value ... */
1075 if (ref
->vtx
>= 0) {
1076 a
->indirect
[1] = ref
->indirect
[1];
1078 /* For an indirect vertex index, we need to shift address right
1079 * by 2, the address register will contain vtx * 16, we need to
1080 * load from a[vtx * 4].
1082 load_vertex_base(pc
, a
, (ref
->acc
< 0) ?
1083 pc
->addr
[ref
->indirect
[1]] : NULL
, -2);
1085 assert(ref
->acc
< 0 || ref
->indirect
[0] < 0);
1087 a
->rhw
= ref
->hw
& ~0x7f;
1088 a
->indirect
[0] = ref
->indirect
[0];
1089 a_ref
= (ref
->acc
< 0) ? pc
->addr
[ref
->indirect
[0]] : NULL
;
1091 emit_add_addr_imm(pc
, a
, a_ref
, a
->rhw
* 4);
1096 #define NV50_MAX_F32 0x880
1097 #define NV50_MAX_S32 0x08c
1098 #define NV50_MAX_U32 0x084
1099 #define NV50_MIN_F32 0x8a0
1100 #define NV50_MIN_S32 0x0ac
1101 #define NV50_MIN_U32 0x0a4
1104 emit_minmax(struct nv50_pc
*pc
, unsigned sub
, struct nv50_reg
*dst
,
1105 struct nv50_reg
*src0
, struct nv50_reg
*src1
)
1107 struct nv50_program_exec
*e
= exec(pc
);
1110 e
->inst
[0] |= 0x30000000 | ((sub
& 0x800) << 20);
1111 e
->inst
[1] |= (sub
<< 24);
1113 check_swap_src_0_1(pc
, &src0
, &src1
);
1114 set_dst(pc
, dst
, e
);
1115 set_src_0(pc
, src0
, e
);
1116 set_src_1(pc
, src1
, e
);
1118 if (src0
->mod
& NV50_MOD_ABS
)
1119 e
->inst
[1] |= 0x00100000;
1120 if (src1
->mod
& NV50_MOD_ABS
)
1121 e
->inst
[1] |= 0x00080000;
1127 emit_sub(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
1128 struct nv50_reg
*src1
)
1130 src1
->mod
^= NV50_MOD_NEG
;
1131 emit_add(pc
, dst
, src0
, src1
);
1132 src1
->mod
^= NV50_MOD_NEG
;
1136 emit_bitop2(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
1137 struct nv50_reg
*src1
, unsigned op
)
1139 struct nv50_program_exec
*e
= exec(pc
);
1141 e
->inst
[0] = 0xd0000000;
1144 check_swap_src_0_1(pc
, &src0
, &src1
);
1145 set_dst(pc
, dst
, e
);
1146 set_src_0(pc
, src0
, e
);
1148 if (op
!= TGSI_OPCODE_AND
&& op
!= TGSI_OPCODE_OR
&&
1149 op
!= TGSI_OPCODE_XOR
)
1150 assert(!"invalid bit op");
1152 assert(!(src0
->mod
| src1
->mod
));
1154 if (src1
->type
== P_IMMD
&& src0
->type
== P_TEMP
&& pc
->allow32
) {
1155 set_immd(pc
, src1
, e
);
1156 if (op
== TGSI_OPCODE_OR
)
1157 e
->inst
[0] |= 0x0100;
1159 if (op
== TGSI_OPCODE_XOR
)
1160 e
->inst
[0] |= 0x8000;
1162 set_src_1(pc
, src1
, e
);
1163 e
->inst
[1] |= 0x04000000; /* 32 bit */
1164 if (op
== TGSI_OPCODE_OR
)
1165 e
->inst
[1] |= 0x4000;
1167 if (op
== TGSI_OPCODE_XOR
)
1168 e
->inst
[1] |= 0x8000;
1175 emit_not(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1177 struct nv50_program_exec
*e
= exec(pc
);
1179 e
->inst
[0] = 0xd0000000;
1180 e
->inst
[1] = 0x0402c000;
1182 set_dst(pc
, dst
, e
);
1183 set_src_1(pc
, src
, e
);
1189 emit_shift(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1190 struct nv50_reg
*src0
, struct nv50_reg
*src1
, unsigned dir
)
1192 struct nv50_program_exec
*e
= exec(pc
);
1194 e
->inst
[0] = 0x30000000;
1195 e
->inst
[1] = 0xc4000000;
1198 set_dst(pc
, dst
, e
);
1199 set_src_0(pc
, src0
, e
);
1201 if (src1
->type
== P_IMMD
) {
1202 e
->inst
[1] |= (1 << 20);
1203 e
->inst
[0] |= (pc
->immd_buf
[src1
->hw
] & 0x7f) << 16;
1205 set_src_1(pc
, src1
, e
);
1207 if (dir
!= TGSI_OPCODE_SHL
)
1208 e
->inst
[1] |= (1 << 29);
1210 if (dir
== TGSI_OPCODE_ISHR
)
1211 e
->inst
[1] |= (1 << 27);
1217 emit_shl_imm(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1218 struct nv50_reg
*src
, int s
)
1220 struct nv50_program_exec
*e
= exec(pc
);
1222 e
->inst
[0] = 0x30000000;
1223 e
->inst
[1] = 0xc4100000;
1225 e
->inst
[1] |= 1 << 29;
1228 e
->inst
[1] |= ((s
& 0x7f) << 16);
1231 set_dst(pc
, dst
, e
);
1232 set_src_0(pc
, src
, e
);
1238 emit_mad(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
1239 struct nv50_reg
*src1
, struct nv50_reg
*src2
)
1241 struct nv50_program_exec
*e
= exec(pc
);
1243 e
->inst
[0] |= 0xe0000000;
1245 check_swap_src_0_1(pc
, &src0
, &src1
);
1246 set_dst(pc
, dst
, e
);
1247 set_src_0(pc
, src0
, e
);
1248 set_src_1(pc
, src1
, e
);
1249 set_src_2(pc
, src2
, e
);
1251 if ((src0
->mod
^ src1
->mod
) & NV50_MOD_NEG
)
1252 e
->inst
[1] |= 0x04000000;
1253 if (src2
->mod
& NV50_MOD_NEG
)
1254 e
->inst
[1] |= 0x08000000;
1260 emit_msb(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src0
,
1261 struct nv50_reg
*src1
, struct nv50_reg
*src2
)
1263 src2
->mod
^= NV50_MOD_NEG
;
1264 emit_mad(pc
, dst
, src0
, src1
, src2
);
1265 src2
->mod
^= NV50_MOD_NEG
;
1268 #define NV50_FLOP_RCP 0
1269 #define NV50_FLOP_RSQ 2
1270 #define NV50_FLOP_LG2 3
1271 #define NV50_FLOP_SIN 4
1272 #define NV50_FLOP_COS 5
1273 #define NV50_FLOP_EX2 6
1275 /* rcp, rsqrt, lg2 support neg and abs */
1277 emit_flop(struct nv50_pc
*pc
, unsigned sub
,
1278 struct nv50_reg
*dst
, struct nv50_reg
*src
)
1280 struct nv50_program_exec
*e
= exec(pc
);
1282 e
->inst
[0] |= 0x90000000;
1283 if (sub
|| src
->mod
) {
1285 e
->inst
[1] |= (sub
<< 29);
1288 set_dst(pc
, dst
, e
);
1289 set_src_0_restricted(pc
, src
, e
);
1291 assert(!src
->mod
|| sub
< 4);
1293 if (src
->mod
& NV50_MOD_NEG
)
1294 e
->inst
[1] |= 0x04000000;
1295 if (src
->mod
& NV50_MOD_ABS
)
1296 e
->inst
[1] |= 0x00100000;
1302 emit_preex2(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1304 struct nv50_program_exec
*e
= exec(pc
);
1306 e
->inst
[0] |= 0xb0000000;
1308 set_dst(pc
, dst
, e
);
1309 set_src_0(pc
, src
, e
);
1311 e
->inst
[1] |= (6 << 29) | 0x00004000;
1313 if (src
->mod
& NV50_MOD_NEG
)
1314 e
->inst
[1] |= 0x04000000;
1315 if (src
->mod
& NV50_MOD_ABS
)
1316 e
->inst
[1] |= 0x00100000;
1322 emit_precossin(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1324 struct nv50_program_exec
*e
= exec(pc
);
1326 e
->inst
[0] |= 0xb0000000;
1328 set_dst(pc
, dst
, e
);
1329 set_src_0(pc
, src
, e
);
1331 e
->inst
[1] |= (6 << 29);
1333 if (src
->mod
& NV50_MOD_NEG
)
1334 e
->inst
[1] |= 0x04000000;
1335 if (src
->mod
& NV50_MOD_ABS
)
1336 e
->inst
[1] |= 0x00100000;
1341 #define CVT_RN (0x00 << 16)
1342 #define CVT_FLOOR (0x02 << 16)
1343 #define CVT_CEIL (0x04 << 16)
1344 #define CVT_TRUNC (0x06 << 16)
1345 #define CVT_SAT (0x08 << 16)
1346 #define CVT_ABS (0x10 << 16)
1348 #define CVT_X32_X32 0x04004000
1349 #define CVT_X32_S32 0x04014000
1350 #define CVT_F32_F32 ((0xc0 << 24) | CVT_X32_X32)
1351 #define CVT_S32_F32 ((0x88 << 24) | CVT_X32_X32)
1352 #define CVT_U32_F32 ((0x80 << 24) | CVT_X32_X32)
1353 #define CVT_F32_S32 ((0x40 << 24) | CVT_X32_S32)
1354 #define CVT_F32_U32 ((0x40 << 24) | CVT_X32_X32)
1355 #define CVT_S32_S32 ((0x08 << 24) | CVT_X32_S32)
1356 #define CVT_S32_U32 ((0x08 << 24) | CVT_X32_X32)
1357 #define CVT_U32_S32 ((0x00 << 24) | CVT_X32_S32)
1359 #define CVT_NEG 0x20000000
1360 #define CVT_RI 0x08000000
1363 emit_cvt(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
,
1364 int wp
, uint32_t cvn
)
1366 struct nv50_program_exec
*e
;
1370 if (src
->mod
& NV50_MOD_NEG
) cvn
|= CVT_NEG
;
1371 if (src
->mod
& NV50_MOD_ABS
) cvn
|= CVT_ABS
;
1373 e
->inst
[0] = 0xa0000000;
1376 set_src_0(pc
, src
, e
);
1379 set_pred_wr(pc
, 1, wp
, e
);
1382 set_dst(pc
, dst
, e
);
1384 e
->inst
[0] |= 0x000001fc;
1385 e
->inst
[1] |= 0x00000008;
1391 /* nv50 Condition codes:
1398 * 0x7 = set condition code ? (used before bra.lt/le/gt/ge)
1399 * 0x8 = unordered bit (allows NaN)
1401 * mode = 0x04 (u32), 0x0c (s32), 0x80 (f32)
1404 emit_set(struct nv50_pc
*pc
, unsigned ccode
, struct nv50_reg
*dst
, int wp
,
1405 struct nv50_reg
*src0
, struct nv50_reg
*src1
, uint8_t mode
)
1407 static const unsigned cc_swapped
[8] = { 0, 4, 2, 6, 1, 5, 3, 7 };
1409 struct nv50_program_exec
*e
= exec(pc
);
1410 struct nv50_reg
*rdst
;
1413 if (check_swap_src_0_1(pc
, &src0
, &src1
))
1414 ccode
= cc_swapped
[ccode
& 7] | (ccode
& 8);
1417 if (dst
&& dst
->type
!= P_TEMP
)
1418 dst
= alloc_temp(pc
, NULL
);
1421 e
->inst
[0] |= 0x30000000 | (mode
<< 24);
1422 e
->inst
[1] |= 0x60000000 | (ccode
<< 14);
1425 set_pred_wr(pc
, 1, wp
, e
);
1427 set_dst(pc
, dst
, e
);
1429 e
->inst
[0] |= 0x000001fc;
1430 e
->inst
[1] |= 0x00000008;
1433 set_src_0(pc
, src0
, e
);
1434 set_src_1(pc
, src1
, e
);
1438 if (rdst
&& mode
== 0x80) /* convert to float ? */
1439 emit_cvt(pc
, rdst
, dst
, -1, CVT_ABS
| CVT_F32_S32
);
1440 if (rdst
&& rdst
!= dst
)
1445 map_tgsi_setop_hw(unsigned op
, uint8_t *cc
, uint8_t *ty
)
1448 case TGSI_OPCODE_SLT
: *cc
= 0x1; *ty
= 0x80; break;
1449 case TGSI_OPCODE_SGE
: *cc
= 0x6; *ty
= 0x80; break;
1450 case TGSI_OPCODE_SEQ
: *cc
= 0x2; *ty
= 0x80; break;
1451 case TGSI_OPCODE_SGT
: *cc
= 0x4; *ty
= 0x80; break;
1452 case TGSI_OPCODE_SLE
: *cc
= 0x3; *ty
= 0x80; break;
1453 case TGSI_OPCODE_SNE
: *cc
= 0xd; *ty
= 0x80; break;
1455 case TGSI_OPCODE_ISLT
: *cc
= 0x1; *ty
= 0x0c; break;
1456 case TGSI_OPCODE_ISGE
: *cc
= 0x6; *ty
= 0x0c; break;
1457 case TGSI_OPCODE_USEQ
: *cc
= 0x2; *ty
= 0x04; break;
1458 case TGSI_OPCODE_USGE
: *cc
= 0x6; *ty
= 0x04; break;
1459 case TGSI_OPCODE_USLT
: *cc
= 0x1; *ty
= 0x04; break;
1460 case TGSI_OPCODE_USNE
: *cc
= 0x5; *ty
= 0x04; break;
1468 emit_add_b32(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1469 struct nv50_reg
*src0
, struct nv50_reg
*rsrc1
)
1471 struct nv50_program_exec
*e
= exec(pc
);
1472 struct nv50_reg
*src1
;
1474 e
->inst
[0] = 0x20000000;
1476 alloc_reg(pc
, rsrc1
);
1477 check_swap_src_0_1(pc
, &src0
, &rsrc1
);
1480 if (src0
->mod
& rsrc1
->mod
& NV50_MOD_NEG
) {
1481 src1
= temp_temp(pc
, e
);
1482 emit_cvt(pc
, src1
, rsrc1
, -1, CVT_S32_S32
);
1485 if (!pc
->allow32
|| src1
->hw
> 63 ||
1486 (src1
->type
!= P_TEMP
&& src1
->type
!= P_IMMD
))
1489 set_dst(pc
, dst
, e
);
1490 set_src_0(pc
, src0
, e
);
1493 e
->inst
[1] |= 1 << 26;
1494 set_src_2(pc
, src1
, e
);
1496 e
->inst
[0] |= 0x8000;
1497 if (src1
->type
== P_IMMD
)
1498 set_immd(pc
, src1
, e
);
1500 set_src_1(pc
, src1
, e
);
1503 if (src0
->mod
& NV50_MOD_NEG
)
1504 e
->inst
[0] |= 1 << 28;
1506 if (src1
->mod
& NV50_MOD_NEG
)
1507 e
->inst
[0] |= 1 << 22;
1513 emit_mad_u16(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1514 struct nv50_reg
*src0
, int lh_0
, struct nv50_reg
*src1
, int lh_1
,
1515 struct nv50_reg
*src2
)
1517 struct nv50_program_exec
*e
= exec(pc
);
1519 e
->inst
[0] = 0x60000000;
1522 set_dst(pc
, dst
, e
);
1524 set_half_src(pc
, src0
, lh_0
, e
, 9);
1525 set_half_src(pc
, src1
, lh_1
, e
, 16);
1526 alloc_reg(pc
, src2
);
1527 if (is_long(e
) || (src2
->type
!= P_TEMP
) || (src2
->hw
!= dst
->hw
))
1528 set_src_2(pc
, src2
, e
);
1534 emit_mul_u16(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1535 struct nv50_reg
*src0
, int lh_0
, struct nv50_reg
*src1
, int lh_1
)
1537 struct nv50_program_exec
*e
= exec(pc
);
1539 e
->inst
[0] = 0x40000000;
1541 set_dst(pc
, dst
, e
);
1543 set_half_src(pc
, src0
, lh_0
, e
, 9);
1544 set_half_src(pc
, src1
, lh_1
, e
, 16);
1550 emit_sad(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1551 struct nv50_reg
*src0
, struct nv50_reg
*src1
, struct nv50_reg
*src2
)
1553 struct nv50_program_exec
*e
= exec(pc
);
1555 e
->inst
[0] = 0x50000000;
1558 check_swap_src_0_1(pc
, &src0
, &src1
);
1559 set_dst(pc
, dst
, e
);
1560 set_src_0(pc
, src0
, e
);
1561 set_src_1(pc
, src1
, e
);
1562 alloc_reg(pc
, src2
);
1563 if (is_long(e
) || (src2
->type
!= dst
->type
) || (src2
->hw
!= dst
->hw
))
1564 set_src_2(pc
, src2
, e
);
1567 e
->inst
[1] |= 0x0c << 24;
1569 e
->inst
[0] |= 0x81 << 8;
1575 emit_flr(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1577 emit_cvt(pc
, dst
, src
, -1, CVT_FLOOR
| CVT_F32_F32
| CVT_RI
);
1581 emit_pow(struct nv50_pc
*pc
, struct nv50_reg
*dst
,
1582 struct nv50_reg
*v
, struct nv50_reg
*e
)
1584 struct nv50_reg
*temp
= alloc_temp(pc
, NULL
);
1586 emit_flop(pc
, NV50_FLOP_LG2
, temp
, v
);
1587 emit_mul(pc
, temp
, temp
, e
);
1588 emit_preex2(pc
, temp
, temp
);
1589 emit_flop(pc
, NV50_FLOP_EX2
, dst
, temp
);
1591 free_temp(pc
, temp
);
1595 emit_sat(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
1597 emit_cvt(pc
, dst
, src
, -1, CVT_SAT
| CVT_F32_F32
);
1601 emit_lit(struct nv50_pc
*pc
, struct nv50_reg
**dst
, unsigned mask
,
1602 struct nv50_reg
**src
)
1604 struct nv50_reg
*one
= alloc_immd(pc
, 1.0);
1605 struct nv50_reg
*zero
= alloc_immd(pc
, 0.0);
1606 struct nv50_reg
*neg128
= alloc_immd(pc
, -127.999999);
1607 struct nv50_reg
*pos128
= alloc_immd(pc
, 127.999999);
1608 struct nv50_reg
*tmp
[4];
1609 boolean allow32
= pc
->allow32
;
1611 pc
->allow32
= FALSE
;
1613 if (mask
& (3 << 1)) {
1614 tmp
[0] = alloc_temp(pc
, NULL
);
1615 emit_minmax(pc
, NV50_MAX_F32
, tmp
[0], src
[0], zero
);
1618 if (mask
& (1 << 2)) {
1619 set_pred_wr(pc
, 1, 0, pc
->p
->exec_tail
);
1621 tmp
[1] = temp_temp(pc
, NULL
);
1622 emit_minmax(pc
, NV50_MAX_F32
, tmp
[1], src
[1], zero
);
1624 tmp
[3] = temp_temp(pc
, NULL
);
1625 emit_minmax(pc
, NV50_MAX_F32
, tmp
[3], src
[3], neg128
);
1626 emit_minmax(pc
, NV50_MIN_F32
, tmp
[3], tmp
[3], pos128
);
1628 emit_pow(pc
, dst
[2], tmp
[1], tmp
[3]);
1629 emit_mov(pc
, dst
[2], zero
);
1630 set_pred(pc
, 3, 0, pc
->p
->exec_tail
);
1633 if (mask
& (1 << 1))
1634 assimilate_temp(pc
, dst
[1], tmp
[0]);
1636 if (mask
& (1 << 2))
1637 free_temp(pc
, tmp
[0]);
1639 pc
->allow32
= allow32
;
1641 /* do this last, in case src[i,j] == dst[0,3] */
1642 if (mask
& (1 << 0))
1643 emit_mov(pc
, dst
[0], one
);
1645 if (mask
& (1 << 3))
1646 emit_mov(pc
, dst
[3], one
);
1655 emit_kil(struct nv50_pc
*pc
, struct nv50_reg
*src
)
1657 struct nv50_program_exec
*e
;
1658 const int r_pred
= 1;
1661 e
->inst
[0] = 0x00000002; /* discard */
1662 set_long(pc
, e
); /* sets cond code to ALWAYS */
1665 set_pred(pc
, 0x1 /* cc = LT */, r_pred
, e
);
1666 /* write to predicate reg */
1667 emit_cvt(pc
, NULL
, src
, r_pred
, CVT_F32_F32
);
1673 static struct nv50_program_exec
*
1674 emit_control_flow(struct nv50_pc
*pc
, unsigned op
, int pred
, unsigned cc
)
1676 struct nv50_program_exec
*e
= exec(pc
);
1678 e
->inst
[0] = (op
<< 28) | 2;
1681 set_pred(pc
, cc
, pred
, e
);
1687 static INLINE
struct nv50_program_exec
*
1688 emit_breakaddr(struct nv50_pc
*pc
)
1690 return emit_control_flow(pc
, 0x4, -1, 0);
1694 emit_break(struct nv50_pc
*pc
, int pred
, unsigned cc
)
1696 emit_control_flow(pc
, 0x5, pred
, cc
);
1699 static INLINE
struct nv50_program_exec
*
1700 emit_joinat(struct nv50_pc
*pc
)
1702 return emit_control_flow(pc
, 0xa, -1, 0);
1705 static INLINE
struct nv50_program_exec
*
1706 emit_branch(struct nv50_pc
*pc
, int pred
, unsigned cc
)
1708 return emit_control_flow(pc
, 0x1, pred
, cc
);
1711 static INLINE
struct nv50_program_exec
*
1712 emit_call(struct nv50_pc
*pc
, int pred
, unsigned cc
)
1714 return emit_control_flow(pc
, 0x2, pred
, cc
);
1718 emit_ret(struct nv50_pc
*pc
, int pred
, unsigned cc
)
1720 emit_control_flow(pc
, 0x3, pred
, cc
);
1724 emit_prim_cmd(struct nv50_pc
*pc
, unsigned cmd
)
1726 struct nv50_program_exec
*e
= exec(pc
);
1728 e
->inst
[0] = 0xf0000000 | (cmd
<< 9);
1729 e
->inst
[1] = 0xc0000000;
1738 #define QOP_MOV_SRC1 3
1740 /* For a quad of threads / top left, top right, bottom left, bottom right
1741 * pixels, do a different operation, and take src0 from a specific thread.
1744 emit_quadop(struct nv50_pc
*pc
, struct nv50_reg
*dst
, int wp
, int lane_src0
,
1745 struct nv50_reg
*src0
, struct nv50_reg
*src1
, ubyte qop
)
1747 struct nv50_program_exec
*e
= exec(pc
);
1749 e
->inst
[0] = 0xc0000000;
1750 e
->inst
[1] = 0x80000000;
1752 e
->inst
[0] |= lane_src0
<< 16;
1753 set_src_0(pc
, src0
, e
);
1754 set_src_2(pc
, src1
, e
);
1757 set_pred_wr(pc
, 1, wp
, e
);
1760 set_dst(pc
, dst
, e
);
1762 e
->inst
[0] |= 0x000001fc;
1763 e
->inst
[1] |= 0x00000008;
1766 e
->inst
[0] |= (qop
& 3) << 20;
1767 e
->inst
[1] |= (qop
>> 2) << 22;
1773 load_cube_tex_coords(struct nv50_pc
*pc
, struct nv50_reg
*t
[4],
1774 struct nv50_reg
**src
, unsigned arg
, boolean proj
)
1776 int mod
[3] = { src
[0]->mod
, src
[1]->mod
, src
[2]->mod
};
1778 src
[0]->mod
|= NV50_MOD_ABS
;
1779 src
[1]->mod
|= NV50_MOD_ABS
;
1780 src
[2]->mod
|= NV50_MOD_ABS
;
1782 emit_minmax(pc
, NV50_MAX_F32
, t
[2], src
[0], src
[1]);
1783 emit_minmax(pc
, NV50_MAX_F32
, t
[2], src
[2], t
[2]);
1785 src
[0]->mod
= mod
[0];
1786 src
[1]->mod
= mod
[1];
1787 src
[2]->mod
= mod
[2];
1789 if (proj
&& 0 /* looks more correct without this */)
1790 emit_mul(pc
, t
[2], t
[2], src
[3]);
1792 if (arg
== 4) /* there is no textureProj(samplerCubeShadow) */
1793 emit_mov(pc
, t
[3], src
[3]);
1795 emit_flop(pc
, NV50_FLOP_RCP
, t
[2], t
[2]);
1797 emit_mul(pc
, t
[0], src
[0], t
[2]);
1798 emit_mul(pc
, t
[1], src
[1], t
[2]);
1799 emit_mul(pc
, t
[2], src
[2], t
[2]);
1803 load_proj_tex_coords(struct nv50_pc
*pc
, struct nv50_reg
*t
[4],
1804 struct nv50_reg
**src
, unsigned dim
, unsigned arg
)
1808 if (src
[0]->type
== P_TEMP
&& src
[0]->rhw
!= -1) {
1809 mode
= pc
->interp_mode
[src
[0]->index
] | INTERP_PERSPECTIVE
;
1811 t
[3]->rhw
= src
[3]->rhw
;
1812 emit_interp(pc
, t
[3], NULL
, (mode
& INTERP_CENTROID
));
1813 emit_flop(pc
, NV50_FLOP_RCP
, t
[3], t
[3]);
1815 for (c
= 0; c
< dim
; ++c
) {
1816 t
[c
]->rhw
= src
[c
]->rhw
;
1817 emit_interp(pc
, t
[c
], t
[3], mode
);
1819 if (arg
!= dim
) { /* depth reference value */
1820 t
[dim
]->rhw
= src
[2]->rhw
;
1821 emit_interp(pc
, t
[dim
], t
[3], mode
);
1824 /* XXX: for some reason the blob sometimes uses MAD
1825 * (mad f32 $rX $rY $rZ neg $r63)
1827 emit_flop(pc
, NV50_FLOP_RCP
, t
[3], src
[3]);
1828 for (c
= 0; c
< dim
; ++c
)
1829 emit_mul(pc
, t
[c
], src
[c
], t
[3]);
1830 if (arg
!= dim
) /* depth reference value */
1831 emit_mul(pc
, t
[dim
], src
[2], t
[3]);
1836 get_tex_dim(unsigned type
, unsigned *dim
, unsigned *arg
)
1839 case TGSI_TEXTURE_1D
:
1842 case TGSI_TEXTURE_SHADOW1D
:
1846 case TGSI_TEXTURE_UNKNOWN
:
1847 case TGSI_TEXTURE_2D
:
1848 case TGSI_TEXTURE_RECT
:
1851 case TGSI_TEXTURE_SHADOW2D
:
1852 case TGSI_TEXTURE_SHADOWRECT
:
1856 case TGSI_TEXTURE_3D
:
1857 case TGSI_TEXTURE_CUBE
:
1866 /* We shouldn't execute TEXLOD if any of the pixels in a quad have
1867 * different LOD values, so branch off groups of equal LOD.
1870 emit_texlod_sequence(struct nv50_pc
*pc
, struct nv50_reg
*tlod
,
1871 struct nv50_reg
*src
, struct nv50_program_exec
*tex
)
1873 struct nv50_program_exec
*join_at
;
1874 unsigned i
, target
= pc
->p
->exec_size
+ 9 * 2;
1876 if (pc
->p
->type
!= PIPE_SHADER_FRAGMENT
) {
1880 pc
->allow32
= FALSE
;
1882 /* Subtract lod of each pixel from lod of top left pixel, jump
1883 * texlod insn if result is 0, then repeat for 2 other pixels.
1885 join_at
= emit_joinat(pc
);
1886 emit_quadop(pc
, NULL
, 0, 0, tlod
, tlod
, 0x55);
1887 emit_branch(pc
, 0, 2)->param
.index
= target
;
1889 for (i
= 1; i
< 4; ++i
) {
1890 emit_quadop(pc
, NULL
, 0, i
, tlod
, tlod
, 0x55);
1891 emit_branch(pc
, 0, 2)->param
.index
= target
;
1894 emit_mov(pc
, tlod
, src
); /* target */
1895 emit(pc
, tex
); /* texlod */
1897 join_at
->param
.index
= target
+ 2 * 2;
1898 JOIN_ON(emit_nop(pc
)); /* join _after_ tex */
1902 emit_texbias_sequence(struct nv50_pc
*pc
, struct nv50_reg
*t
[4], unsigned arg
,
1903 struct nv50_program_exec
*tex
)
1905 struct nv50_program_exec
*e
;
1906 struct nv50_reg imm_1248
, *t123
[4][4], *r_bits
= alloc_temp(pc
, NULL
);
1908 unsigned n
, c
, i
, cc
[4] = { 0x0a, 0x13, 0x11, 0x10 };
1910 pc
->allow32
= FALSE
;
1911 ctor_reg(&imm_1248
, P_IMMD
, -1, ctor_immd_4u32(pc
, 1, 2, 4, 8) * 4);
1913 /* Subtract bias value of thread i from bias values of each thread,
1914 * store result in r_pred, and set bit i in r_bits if result was 0.
1917 for (i
= 0; i
< 4; ++i
, ++imm_1248
.hw
) {
1918 emit_quadop(pc
, NULL
, r_pred
, i
, t
[arg
], t
[arg
], 0x55);
1919 emit_mov(pc
, r_bits
, &imm_1248
);
1920 set_pred(pc
, 2, r_pred
, pc
->p
->exec_tail
);
1922 emit_mov_to_pred(pc
, r_pred
, r_bits
);
1924 /* The lanes of a quad are now grouped by the bit in r_pred they have
1925 * set. Put the input values for TEX into a new register set for each
1926 * group and execute TEX only for a specific group.
1927 * We cannot use the same register set for each group because we need
1928 * the derivatives, which are implicitly calculated, to be correct.
1930 for (i
= 1; i
< 4; ++i
) {
1931 alloc_temp4(pc
, t123
[i
], 0);
1933 for (c
= 0; c
<= arg
; ++c
)
1934 emit_mov(pc
, t123
[i
][c
], t
[c
]);
1936 *(e
= exec(pc
)) = *(tex
);
1937 e
->inst
[0] &= ~0x01fc;
1938 set_dst(pc
, t123
[i
][0], e
);
1939 set_pred(pc
, cc
[i
], r_pred
, e
);
1942 /* finally TEX on the original regs (where we kept the input) */
1943 set_pred(pc
, cc
[0], r_pred
, tex
);
1946 /* put the 3 * n other results into regs for lane 0 */
1947 n
= popcnt4(((e
->inst
[0] >> 25) & 0x3) | ((e
->inst
[1] >> 12) & 0xc));
1948 for (i
= 1; i
< 4; ++i
) {
1949 for (c
= 0; c
< n
; ++c
) {
1950 emit_mov(pc
, t
[c
], t123
[i
][c
]);
1951 set_pred(pc
, cc
[i
], r_pred
, pc
->p
->exec_tail
);
1953 free_temp4(pc
, t123
[i
]);
1957 free_temp(pc
, r_bits
);
1961 emit_tex(struct nv50_pc
*pc
, struct nv50_reg
**dst
, unsigned mask
,
1962 struct nv50_reg
**src
, unsigned unit
, unsigned type
,
1963 boolean proj
, int bias_lod
)
1965 struct nv50_reg
*t
[4];
1966 struct nv50_program_exec
*e
;
1967 unsigned c
, dim
, arg
;
1969 /* t[i] must be within a single 128 bit super-reg */
1970 alloc_temp4(pc
, t
, 0);
1973 e
->inst
[0] = 0xf0000000;
1975 set_dst(pc
, t
[0], e
);
1977 /* TIC and TSC binding indices (TSC is ignored as TSC_LINKED = TRUE): */
1978 e
->inst
[0] |= (unit
<< 9) /* | (unit << 17) */;
1980 /* live flag (don't set if TEX results affect input to another TEX): */
1981 /* e->inst[0] |= 0x00000004; */
1983 get_tex_dim(type
, &dim
, &arg
);
1985 if (type
== TGSI_TEXTURE_CUBE
) {
1986 e
->inst
[0] |= 0x08000000;
1987 load_cube_tex_coords(pc
, t
, src
, arg
, proj
);
1990 load_proj_tex_coords(pc
, t
, src
, dim
, arg
);
1992 for (c
= 0; c
< dim
; c
++)
1993 emit_mov(pc
, t
[c
], src
[c
]);
1994 if (arg
!= dim
) /* depth reference value (always src.z here) */
1995 emit_mov(pc
, t
[dim
], src
[2]);
1998 e
->inst
[0] |= (mask
& 0x3) << 25;
1999 e
->inst
[1] |= (mask
& 0xc) << 12;
2002 e
->inst
[0] |= (arg
- 1) << 22;
2006 assert(pc
->p
->type
== PIPE_SHADER_FRAGMENT
);
2007 e
->inst
[0] |= arg
<< 22;
2008 e
->inst
[1] |= 0x20000000; /* texbias */
2009 emit_mov(pc
, t
[arg
], src
[3]);
2010 emit_texbias_sequence(pc
, t
, arg
, e
);
2012 e
->inst
[0] |= arg
<< 22;
2013 e
->inst
[1] |= 0x40000000; /* texlod */
2014 emit_mov(pc
, t
[arg
], src
[3]);
2015 emit_texlod_sequence(pc
, t
[arg
], src
[3], e
);
2020 if (mask
& 1) emit_mov(pc
, dst
[0], t
[c
++]);
2021 if (mask
& 2) emit_mov(pc
, dst
[1], t
[c
++]);
2022 if (mask
& 4) emit_mov(pc
, dst
[2], t
[c
++]);
2023 if (mask
& 8) emit_mov(pc
, dst
[3], t
[c
]);
2027 /* XXX: if p.e. MUL is used directly after TEX, it would still use
2028 * the texture coordinates, not the fetched values: latency ? */
2030 for (c
= 0; c
< 4; c
++) {
2031 if (mask
& (1 << c
))
2032 assimilate_temp(pc
, dst
[c
], t
[c
]);
2034 free_temp(pc
, t
[c
]);
2040 emit_ddx(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
2042 struct nv50_program_exec
*e
= exec(pc
);
2044 assert(src
->type
== P_TEMP
);
2046 e
->inst
[0] = (src
->mod
& NV50_MOD_NEG
) ? 0xc0240000 : 0xc0140000;
2047 e
->inst
[1] = (src
->mod
& NV50_MOD_NEG
) ? 0x86400000 : 0x89800000;
2049 set_dst(pc
, dst
, e
);
2050 set_src_0(pc
, src
, e
);
2051 set_src_2(pc
, src
, e
);
2057 emit_ddy(struct nv50_pc
*pc
, struct nv50_reg
*dst
, struct nv50_reg
*src
)
2059 struct nv50_program_exec
*e
= exec(pc
);
2061 assert(src
->type
== P_TEMP
);
2063 e
->inst
[0] = (src
->mod
& NV50_MOD_NEG
) ? 0xc0250000 : 0xc0150000;
2064 e
->inst
[1] = (src
->mod
& NV50_MOD_NEG
) ? 0x85800000 : 0x8a400000;
2066 set_dst(pc
, dst
, e
);
2067 set_src_0(pc
, src
, e
);
2068 set_src_2(pc
, src
, e
);
2074 convert_to_long(struct nv50_pc
*pc
, struct nv50_program_exec
*e
)
2076 unsigned q
= 0, m
= ~0;
2078 assert(!is_long(e
));
2080 switch (e
->inst
[0] >> 28) {
2088 /* ADD, SUB, SUBR b32 */
2089 m
= ~(0x8000 | (127 << 16));
2090 q
= ((e
->inst
[0] & (~m
)) >> 2) | (1 << 26);
2095 q
= (0x0c << 24) | ((e
->inst
[0] & (0x7f << 2)) << 12);
2099 q
= (e
->inst
[0] & (0x7f << 2)) << 12;
2102 /* INTERP (move centroid, perspective and flat bits) */
2104 q
= (e
->inst
[0] & (3 << 24)) >> (24 - 16);
2105 q
|= (e
->inst
[0] & (1 << 8)) << (18 - 8);
2113 q
= ((e
->inst
[0] & (~m
)) >> 2);
2118 q
= ((e
->inst
[0] & (~m
)) << 12);
2121 /* MAD (if src2 == dst) */
2122 q
= ((e
->inst
[0] & 0x1fc) << 12);
2136 /* Some operations support an optional negation flag. */
2138 get_supported_mods(const struct tgsi_full_instruction
*insn
, int i
)
2140 switch (insn
->Instruction
.Opcode
) {
2141 case TGSI_OPCODE_ADD
:
2142 case TGSI_OPCODE_COS
:
2143 case TGSI_OPCODE_DDX
:
2144 case TGSI_OPCODE_DDY
:
2145 case TGSI_OPCODE_DP3
:
2146 case TGSI_OPCODE_DP4
:
2147 case TGSI_OPCODE_EX2
:
2148 case TGSI_OPCODE_KIL
:
2149 case TGSI_OPCODE_LG2
:
2150 case TGSI_OPCODE_MAD
:
2151 case TGSI_OPCODE_MUL
:
2152 case TGSI_OPCODE_POW
:
2153 case TGSI_OPCODE_RCP
:
2154 case TGSI_OPCODE_RSQ
: /* ignored, RSQ = rsqrt(abs(src.x)) */
2155 case TGSI_OPCODE_SCS
:
2156 case TGSI_OPCODE_SIN
:
2157 case TGSI_OPCODE_SUB
:
2158 return NV50_MOD_NEG
;
2159 case TGSI_OPCODE_MAX
:
2160 case TGSI_OPCODE_MIN
:
2161 case TGSI_OPCODE_INEG
: /* tgsi src sign toggle/set would be stupid */
2162 return NV50_MOD_ABS
;
2163 case TGSI_OPCODE_CEIL
:
2164 case TGSI_OPCODE_FLR
:
2165 case TGSI_OPCODE_TRUNC
:
2166 return NV50_MOD_NEG
| NV50_MOD_ABS
;
2167 case TGSI_OPCODE_F2I
:
2168 case TGSI_OPCODE_F2U
:
2169 case TGSI_OPCODE_I2F
:
2170 case TGSI_OPCODE_U2F
:
2171 return NV50_MOD_NEG
| NV50_MOD_ABS
| NV50_MOD_I32
;
2172 case TGSI_OPCODE_UADD
:
2173 return NV50_MOD_NEG
| NV50_MOD_I32
;
2174 case TGSI_OPCODE_SAD
:
2175 case TGSI_OPCODE_SHL
:
2176 case TGSI_OPCODE_IMAX
:
2177 case TGSI_OPCODE_IMIN
:
2178 case TGSI_OPCODE_ISHR
:
2179 case TGSI_OPCODE_NOT
:
2180 case TGSI_OPCODE_UMAD
:
2181 case TGSI_OPCODE_UMAX
:
2182 case TGSI_OPCODE_UMIN
:
2183 case TGSI_OPCODE_UMUL
:
2184 case TGSI_OPCODE_USHR
:
2185 return NV50_MOD_I32
;
2191 /* Return a read mask for source registers deduced from opcode & write mask. */
2193 nv50_tgsi_src_mask(const struct tgsi_full_instruction
*insn
, int c
)
2195 unsigned x
, mask
= insn
->Dst
[0].Register
.WriteMask
;
2197 switch (insn
->Instruction
.Opcode
) {
2198 case TGSI_OPCODE_COS
:
2199 case TGSI_OPCODE_SIN
:
2200 return (mask
& 0x8) | ((mask
& 0x7) ? 0x1 : 0x0);
2201 case TGSI_OPCODE_DP3
:
2203 case TGSI_OPCODE_DP4
:
2204 case TGSI_OPCODE_DPH
:
2205 case TGSI_OPCODE_KIL
: /* WriteMask ignored */
2207 case TGSI_OPCODE_DST
:
2208 return mask
& (c
? 0xa : 0x6);
2209 case TGSI_OPCODE_EX2
:
2210 case TGSI_OPCODE_EXP
:
2211 case TGSI_OPCODE_LG2
:
2212 case TGSI_OPCODE_LOG
:
2213 case TGSI_OPCODE_POW
:
2214 case TGSI_OPCODE_RCP
:
2215 case TGSI_OPCODE_RSQ
:
2216 case TGSI_OPCODE_SCS
:
2218 case TGSI_OPCODE_IF
:
2220 case TGSI_OPCODE_LIT
:
2222 case TGSI_OPCODE_TEX
:
2223 case TGSI_OPCODE_TXB
:
2224 case TGSI_OPCODE_TXL
:
2225 case TGSI_OPCODE_TXP
:
2227 const struct tgsi_instruction_texture
*tex
;
2229 assert(insn
->Instruction
.Texture
);
2230 tex
= &insn
->Texture
;
2233 if (insn
->Instruction
.Opcode
!= TGSI_OPCODE_TEX
&&
2234 insn
->Instruction
.Opcode
!= TGSI_OPCODE_TXD
)
2235 mask
|= 0x8; /* bias, lod or proj */
2237 switch (tex
->Texture
) {
2238 case TGSI_TEXTURE_1D
:
2241 case TGSI_TEXTURE_SHADOW1D
:
2244 case TGSI_TEXTURE_2D
:
2252 case TGSI_OPCODE_XPD
:
2254 if (mask
& 1) x
|= 0x6;
2255 if (mask
& 2) x
|= 0x5;
2256 if (mask
& 4) x
|= 0x3;
2265 static struct nv50_reg
*
2266 tgsi_dst(struct nv50_pc
*pc
, int c
, const struct tgsi_full_dst_register
*dst
)
2268 switch (dst
->Register
.File
) {
2269 case TGSI_FILE_TEMPORARY
:
2270 return &pc
->temp
[dst
->Register
.Index
* 4 + c
];
2271 case TGSI_FILE_OUTPUT
:
2272 return &pc
->result
[dst
->Register
.Index
* 4 + c
];
2273 case TGSI_FILE_ADDRESS
:
2275 struct nv50_reg
*r
= pc
->addr
[dst
->Register
.Index
* 4 + c
];
2277 r
= get_address_reg(pc
, NULL
);
2278 r
->index
= dst
->Register
.Index
* 4 + c
;
2279 pc
->addr
[r
->index
] = r
;
2284 case TGSI_FILE_NULL
:
2286 case TGSI_FILE_SYSTEM_VALUE
:
2287 assert(pc
->sysval
[dst
->Register
.Index
].type
== P_RESULT
);
2289 return &pc
->sysval
[dst
->Register
.Index
];
2297 static struct nv50_reg
*
2298 tgsi_src(struct nv50_pc
*pc
, int chan
, const struct tgsi_full_src_register
*src
,
2301 struct nv50_reg
*r
= NULL
;
2302 struct nv50_reg
*temp
= NULL
;
2303 unsigned sgn
, c
, swz
, cvn
;
2305 if (src
->Register
.File
!= TGSI_FILE_CONSTANT
)
2306 assert(!src
->Register
.Indirect
);
2308 sgn
= tgsi_util_get_full_src_register_sign_mode(src
, chan
);
2310 c
= tgsi_util_get_full_src_register_swizzle(src
, chan
);
2312 case TGSI_SWIZZLE_X
:
2313 case TGSI_SWIZZLE_Y
:
2314 case TGSI_SWIZZLE_Z
:
2315 case TGSI_SWIZZLE_W
:
2316 switch (src
->Register
.File
) {
2317 case TGSI_FILE_INPUT
:
2318 r
= &pc
->attr
[src
->Register
.Index
* 4 + c
];
2320 if (!src
->Dimension
.Dimension
)
2322 r
= reg_instance(pc
, r
);
2323 r
->vtx
= src
->Dimension
.Index
;
2325 if (!src
->Dimension
.Indirect
)
2327 swz
= tgsi_util_get_src_register_swizzle(
2328 &src
->DimIndirect
, 0);
2330 r
->indirect
[1] = src
->DimIndirect
.Index
* 4 + swz
;
2332 case TGSI_FILE_TEMPORARY
:
2333 r
= &pc
->temp
[src
->Register
.Index
* 4 + c
];
2335 case TGSI_FILE_CONSTANT
:
2336 if (!src
->Register
.Indirect
) {
2337 r
= &pc
->param
[src
->Register
.Index
* 4 + c
];
2340 /* Indicate indirection by setting r->acc < 0 and
2341 * use the index field to select the address reg.
2343 r
= reg_instance(pc
, NULL
);
2344 ctor_reg(r
, P_CONST
, -1, src
->Register
.Index
* 4 + c
);
2346 swz
= tgsi_util_get_src_register_swizzle(
2349 r
->indirect
[0] = src
->Indirect
.Index
* 4 + swz
;
2351 case TGSI_FILE_IMMEDIATE
:
2352 r
= &pc
->immd
[src
->Register
.Index
* 4 + c
];
2354 case TGSI_FILE_SAMPLER
:
2356 case TGSI_FILE_ADDRESS
:
2357 r
= pc
->addr
[src
->Register
.Index
* 4 + c
];
2360 case TGSI_FILE_SYSTEM_VALUE
:
2362 r
= &pc
->sysval
[src
->Register
.Index
];
2374 cvn
= (mod
& NV50_MOD_I32
) ? CVT_S32_S32
: CVT_F32_F32
;
2377 case TGSI_UTIL_SIGN_CLEAR
:
2378 r
->mod
= NV50_MOD_ABS
;
2380 case TGSI_UTIL_SIGN_SET
:
2381 r
->mod
= NV50_MOD_NEG_ABS
;
2383 case TGSI_UTIL_SIGN_TOGGLE
:
2384 r
->mod
= NV50_MOD_NEG
;
2387 assert(!r
->mod
&& sgn
== TGSI_UTIL_SIGN_KEEP
);
2391 if ((r
->mod
& mod
) != r
->mod
) {
2392 temp
= temp_temp(pc
, NULL
);
2393 emit_cvt(pc
, temp
, r
, -1, cvn
);
2397 r
->mod
|= mod
& NV50_MOD_I32
;
2400 if (r
->acc
>= 0 && r
->vtx
< 0 && r
!= temp
)
2401 return reg_instance(pc
, r
); /* will clear r->mod */
2405 /* return TRUE for ops that produce only a single result */
2407 is_scalar_op(unsigned op
)
2410 case TGSI_OPCODE_COS
:
2411 case TGSI_OPCODE_DP2
:
2412 case TGSI_OPCODE_DP3
:
2413 case TGSI_OPCODE_DP4
:
2414 case TGSI_OPCODE_DPH
:
2415 case TGSI_OPCODE_EX2
:
2416 case TGSI_OPCODE_LG2
:
2417 case TGSI_OPCODE_POW
:
2418 case TGSI_OPCODE_RCP
:
2419 case TGSI_OPCODE_RSQ
:
2420 case TGSI_OPCODE_SIN
:
2422 case TGSI_OPCODE_KIL:
2423 case TGSI_OPCODE_LIT:
2424 case TGSI_OPCODE_SCS:
2432 /* Returns a bitmask indicating which dst components depend
2433 * on source s, component c (reverse of nv50_tgsi_src_mask).
2436 nv50_tgsi_dst_revdep(unsigned op
, int s
, int c
)
2438 if (is_scalar_op(op
))
2442 case TGSI_OPCODE_DST
:
2443 return (1 << c
) & (s
? 0xa : 0x6);
2444 case TGSI_OPCODE_XPD
:
2454 case TGSI_OPCODE_EXP
:
2455 case TGSI_OPCODE_LOG
:
2456 case TGSI_OPCODE_LIT
:
2457 case TGSI_OPCODE_SCS
:
2458 case TGSI_OPCODE_TEX
:
2459 case TGSI_OPCODE_TXB
:
2460 case TGSI_OPCODE_TXL
:
2461 case TGSI_OPCODE_TXP
:
2462 /* these take care of dangerous swizzles themselves */
2464 case TGSI_OPCODE_IF
:
2465 case TGSI_OPCODE_KIL
:
2466 /* don't call this function for these ops */
2470 /* linear vector instruction */
2475 static INLINE boolean
2476 has_pred(struct nv50_program_exec
*e
, unsigned cc
)
2478 if (!is_long(e
) || is_immd(e
))
2480 return ((e
->inst
[1] & 0x780) == (cc
<< 7));
2483 /* on ENDIF see if we can do "@p0.neu single_op" instead of:
2490 nv50_kill_branch(struct nv50_pc
*pc
)
2492 int lvl
= pc
->if_lvl
;
2494 if (pc
->if_insn
[lvl
]->next
!= pc
->p
->exec_tail
)
2496 if (is_immd(pc
->p
->exec_tail
))
2499 /* if ccode == 'true', the BRA is from an ELSE and the predicate
2500 * reg may no longer be valid, since we currently always use $p0
2502 if (has_pred(pc
->if_insn
[lvl
], 0xf))
2504 assert(pc
->if_insn
[lvl
] && pc
->if_join
[lvl
]);
2506 /* We'll use the exec allocated for JOIN_AT (we can't easily
2507 * access nv50_program_exec's prev).
2509 pc
->p
->exec_size
-= 4; /* remove JOIN_AT and BRA */
2511 *pc
->if_join
[lvl
] = *pc
->p
->exec_tail
;
2513 FREE(pc
->if_insn
[lvl
]);
2514 FREE(pc
->p
->exec_tail
);
2516 pc
->p
->exec_tail
= pc
->if_join
[lvl
];
2517 pc
->p
->exec_tail
->next
= NULL
;
2518 set_pred(pc
, 0xd, 0, pc
->p
->exec_tail
);
2524 nv50_fp_move_results(struct nv50_pc
*pc
)
2526 struct nv50_reg reg
;
2529 ctor_reg(®
, P_TEMP
, -1, -1);
2531 for (i
= 0; i
< pc
->result_nr
* 4; ++i
) {
2532 if (pc
->result
[i
].rhw
< 0 || pc
->result
[i
].hw
< 0)
2534 if (pc
->result
[i
].rhw
!= pc
->result
[i
].hw
) {
2535 reg
.hw
= pc
->result
[i
].rhw
;
2536 emit_mov(pc
, ®
, &pc
->result
[i
]);
2542 nv50_program_tx_insn(struct nv50_pc
*pc
,
2543 const struct tgsi_full_instruction
*inst
)
2545 struct nv50_reg
*rdst
[4], *dst
[4], *brdc
, *src
[3][4], *temp
;
2546 unsigned mask
, sat
, unit
;
2549 mask
= inst
->Dst
[0].Register
.WriteMask
;
2550 sat
= inst
->Instruction
.Saturate
== TGSI_SAT_ZERO_ONE
;
2552 memset(src
, 0, sizeof(src
));
2554 for (c
= 0; c
< 4; c
++) {
2555 if ((mask
& (1 << c
)) && !pc
->r_dst
[c
])
2556 dst
[c
] = tgsi_dst(pc
, c
, &inst
->Dst
[0]);
2558 dst
[c
] = pc
->r_dst
[c
];
2562 for (i
= 0; i
< inst
->Instruction
.NumSrcRegs
; i
++) {
2563 const struct tgsi_full_src_register
*fs
= &inst
->Src
[i
];
2567 src_mask
= nv50_tgsi_src_mask(inst
, i
);
2568 mod_supp
= get_supported_mods(inst
, i
);
2570 if (fs
->Register
.File
== TGSI_FILE_SAMPLER
)
2571 unit
= fs
->Register
.Index
;
2573 for (c
= 0; c
< 4; c
++)
2574 if (src_mask
& (1 << c
))
2575 src
[i
][c
] = tgsi_src(pc
, c
, fs
, mod_supp
);
2578 brdc
= temp
= pc
->r_brdc
;
2579 if (brdc
&& brdc
->type
!= P_TEMP
) {
2580 temp
= temp_temp(pc
, NULL
);
2585 for (c
= 0; c
< 4; c
++) {
2586 if (!(mask
& (1 << c
)) || dst
[c
]->type
== P_TEMP
)
2588 /* rdst[c] = dst[c]; */ /* done above */
2589 dst
[c
] = temp_temp(pc
, NULL
);
2593 assert(brdc
|| !is_scalar_op(inst
->Instruction
.Opcode
));
2595 switch (inst
->Instruction
.Opcode
) {
2596 case TGSI_OPCODE_ABS
:
2597 for (c
= 0; c
< 4; c
++) {
2598 if (!(mask
& (1 << c
)))
2600 emit_cvt(pc
, dst
[c
], src
[0][c
], -1,
2601 CVT_ABS
| CVT_F32_F32
);
2604 case TGSI_OPCODE_ADD
:
2605 for (c
= 0; c
< 4; c
++) {
2606 if (!(mask
& (1 << c
)))
2608 emit_add(pc
, dst
[c
], src
[0][c
], src
[1][c
]);
2611 case TGSI_OPCODE_AND
:
2612 case TGSI_OPCODE_XOR
:
2613 case TGSI_OPCODE_OR
:
2614 for (c
= 0; c
< 4; c
++) {
2615 if (!(mask
& (1 << c
)))
2617 emit_bitop2(pc
, dst
[c
], src
[0][c
], src
[1][c
],
2618 inst
->Instruction
.Opcode
);
2621 case TGSI_OPCODE_ARL
:
2622 temp
= temp_temp(pc
, NULL
);
2623 for (c
= 0; c
< 4; c
++) {
2624 if (!(mask
& (1 << c
)))
2626 emit_cvt(pc
, temp
, src
[0][c
], -1,
2627 CVT_FLOOR
| CVT_S32_F32
);
2628 emit_arl(pc
, dst
[c
], temp
, 4);
2631 case TGSI_OPCODE_BGNLOOP
:
2632 pc
->loop_brka
[pc
->loop_lvl
] = emit_breakaddr(pc
);
2633 pc
->loop_pos
[pc
->loop_lvl
++] = pc
->p
->exec_size
;
2636 case TGSI_OPCODE_BGNSUB
:
2637 assert(!pc
->in_subroutine
);
2638 pc
->in_subroutine
= TRUE
;
2639 /* probably not necessary, but align to 8 byte boundary */
2640 if (!is_long(pc
->p
->exec_tail
))
2641 convert_to_long(pc
, pc
->p
->exec_tail
);
2643 case TGSI_OPCODE_BRK
:
2644 assert(pc
->loop_lvl
> 0);
2645 emit_break(pc
, -1, 0);
2647 case TGSI_OPCODE_CAL
:
2648 assert(inst
->Label
.Label
< pc
->insn_nr
);
2649 emit_call(pc
, -1, 0)->param
.index
= inst
->Label
.Label
;
2650 /* replaced by actual offset in nv50_program_fixup_insns */
2652 case TGSI_OPCODE_CEIL
:
2653 for (c
= 0; c
< 4; c
++) {
2654 if (!(mask
& (1 << c
)))
2656 emit_cvt(pc
, dst
[c
], src
[0][c
], -1,
2657 CVT_CEIL
| CVT_F32_F32
| CVT_RI
);
2660 case TGSI_OPCODE_CMP
:
2661 pc
->allow32
= FALSE
;
2662 for (c
= 0; c
< 4; c
++) {
2663 if (!(mask
& (1 << c
)))
2665 emit_cvt(pc
, NULL
, src
[0][c
], 1, CVT_F32_F32
);
2666 emit_mov(pc
, dst
[c
], src
[1][c
]);
2667 set_pred(pc
, 0x1, 1, pc
->p
->exec_tail
); /* @SF */
2668 emit_mov(pc
, dst
[c
], src
[2][c
]);
2669 set_pred(pc
, 0x6, 1, pc
->p
->exec_tail
); /* @NSF */
2672 case TGSI_OPCODE_CONT
:
2673 assert(pc
->loop_lvl
> 0);
2674 emit_branch(pc
, -1, 0)->param
.index
=
2675 pc
->loop_pos
[pc
->loop_lvl
- 1];
2677 case TGSI_OPCODE_COS
:
2679 emit_precossin(pc
, temp
, src
[0][3]);
2680 emit_flop(pc
, NV50_FLOP_COS
, dst
[3], temp
);
2684 temp
= brdc
= temp_temp(pc
, NULL
);
2686 emit_precossin(pc
, temp
, src
[0][0]);
2687 emit_flop(pc
, NV50_FLOP_COS
, brdc
, temp
);
2689 case TGSI_OPCODE_DDX
:
2690 for (c
= 0; c
< 4; c
++) {
2691 if (!(mask
& (1 << c
)))
2693 emit_ddx(pc
, dst
[c
], src
[0][c
]);
2696 case TGSI_OPCODE_DDY
:
2697 for (c
= 0; c
< 4; c
++) {
2698 if (!(mask
& (1 << c
)))
2700 emit_ddy(pc
, dst
[c
], src
[0][c
]);
2703 case TGSI_OPCODE_DP3
:
2704 emit_mul(pc
, temp
, src
[0][0], src
[1][0]);
2705 emit_mad(pc
, temp
, src
[0][1], src
[1][1], temp
);
2706 emit_mad(pc
, brdc
, src
[0][2], src
[1][2], temp
);
2708 case TGSI_OPCODE_DP4
:
2709 emit_mul(pc
, temp
, src
[0][0], src
[1][0]);
2710 emit_mad(pc
, temp
, src
[0][1], src
[1][1], temp
);
2711 emit_mad(pc
, temp
, src
[0][2], src
[1][2], temp
);
2712 emit_mad(pc
, brdc
, src
[0][3], src
[1][3], temp
);
2714 case TGSI_OPCODE_DPH
:
2715 emit_mul(pc
, temp
, src
[0][0], src
[1][0]);
2716 emit_mad(pc
, temp
, src
[0][1], src
[1][1], temp
);
2717 emit_mad(pc
, temp
, src
[0][2], src
[1][2], temp
);
2718 emit_add(pc
, brdc
, src
[1][3], temp
);
2720 case TGSI_OPCODE_DST
:
2721 if (mask
& (1 << 1))
2722 emit_mul(pc
, dst
[1], src
[0][1], src
[1][1]);
2723 if (mask
& (1 << 2))
2724 emit_mov(pc
, dst
[2], src
[0][2]);
2725 if (mask
& (1 << 3))
2726 emit_mov(pc
, dst
[3], src
[1][3]);
2727 if (mask
& (1 << 0))
2728 emit_mov_immdval(pc
, dst
[0], 1.0f
);
2730 case TGSI_OPCODE_ELSE
:
2731 emit_branch(pc
, -1, 0);
2732 pc
->if_insn
[--pc
->if_lvl
]->param
.index
= pc
->p
->exec_size
;
2733 pc
->if_insn
[pc
->if_lvl
++] = pc
->p
->exec_tail
;
2736 case TGSI_OPCODE_EMIT
:
2737 emit_prim_cmd(pc
, 1);
2739 case TGSI_OPCODE_ENDIF
:
2740 pc
->if_insn
[--pc
->if_lvl
]->param
.index
= pc
->p
->exec_size
;
2742 /* try to replace branch over 1 insn with a predicated insn */
2743 if (nv50_kill_branch(pc
) == TRUE
)
2746 if (pc
->if_join
[pc
->if_lvl
]) {
2747 pc
->if_join
[pc
->if_lvl
]->param
.index
= pc
->p
->exec_size
;
2748 pc
->if_join
[pc
->if_lvl
] = NULL
;
2751 /* emit a NOP as join point, we could set it on the next
2752 * one, but would have to make sure it is long and !immd
2754 JOIN_ON(emit_nop(pc
));
2756 case TGSI_OPCODE_ENDLOOP
:
2757 emit_branch(pc
, -1, 0)->param
.index
=
2758 pc
->loop_pos
[--pc
->loop_lvl
];
2759 pc
->loop_brka
[pc
->loop_lvl
]->param
.index
= pc
->p
->exec_size
;
2762 case TGSI_OPCODE_ENDPRIM
:
2763 emit_prim_cmd(pc
, 2);
2765 case TGSI_OPCODE_ENDSUB
:
2766 assert(pc
->in_subroutine
);
2768 pc
->in_subroutine
= FALSE
;
2770 case TGSI_OPCODE_EX2
:
2771 emit_preex2(pc
, temp
, src
[0][0]);
2772 emit_flop(pc
, NV50_FLOP_EX2
, brdc
, temp
);
2774 case TGSI_OPCODE_EXP
:
2776 struct nv50_reg
*t
[2];
2779 t
[0] = temp_temp(pc
, NULL
);
2780 t
[1] = temp_temp(pc
, NULL
);
2783 emit_mov(pc
, t
[0], src
[0][0]);
2785 emit_flr(pc
, t
[1], src
[0][0]);
2787 if (mask
& (1 << 1))
2788 emit_sub(pc
, dst
[1], t
[0], t
[1]);
2789 if (mask
& (1 << 0)) {
2790 emit_preex2(pc
, t
[1], t
[1]);
2791 emit_flop(pc
, NV50_FLOP_EX2
, dst
[0], t
[1]);
2793 if (mask
& (1 << 2)) {
2794 emit_preex2(pc
, t
[0], t
[0]);
2795 emit_flop(pc
, NV50_FLOP_EX2
, dst
[2], t
[0]);
2797 if (mask
& (1 << 3))
2798 emit_mov_immdval(pc
, dst
[3], 1.0f
);
2801 case TGSI_OPCODE_F2I
:
2802 for (c
= 0; c
< 4; c
++) {
2803 if (!(mask
& (1 << c
)))
2805 emit_cvt(pc
, dst
[c
], src
[0][c
], -1,
2806 CVT_TRUNC
| CVT_S32_F32
);
2809 case TGSI_OPCODE_F2U
:
2810 for (c
= 0; c
< 4; c
++) {
2811 if (!(mask
& (1 << c
)))
2813 emit_cvt(pc
, dst
[c
], src
[0][c
], -1,
2814 CVT_TRUNC
| CVT_U32_F32
);
2817 case TGSI_OPCODE_FLR
:
2818 for (c
= 0; c
< 4; c
++) {
2819 if (!(mask
& (1 << c
)))
2821 emit_flr(pc
, dst
[c
], src
[0][c
]);
2824 case TGSI_OPCODE_FRC
:
2825 temp
= temp_temp(pc
, NULL
);
2826 for (c
= 0; c
< 4; c
++) {
2827 if (!(mask
& (1 << c
)))
2829 emit_flr(pc
, temp
, src
[0][c
]);
2830 emit_sub(pc
, dst
[c
], src
[0][c
], temp
);
2833 case TGSI_OPCODE_I2F
:
2834 for (c
= 0; c
< 4; c
++) {
2835 if (!(mask
& (1 << c
)))
2837 emit_cvt(pc
, dst
[c
], src
[0][c
], -1, CVT_F32_S32
);
2840 case TGSI_OPCODE_IF
:
2841 assert(pc
->if_lvl
< NV50_MAX_COND_NESTING
);
2842 emit_cvt(pc
, NULL
, src
[0][0], 0, CVT_ABS
| CVT_F32_F32
);
2843 pc
->if_join
[pc
->if_lvl
] = emit_joinat(pc
);
2844 pc
->if_insn
[pc
->if_lvl
++] = emit_branch(pc
, 0, 2);;
2847 case TGSI_OPCODE_IMAX
:
2848 for (c
= 0; c
< 4; c
++) {
2849 if (!(mask
& (1 << c
)))
2851 emit_minmax(pc
, 0x08c, dst
[c
], src
[0][c
], src
[1][c
]);
2854 case TGSI_OPCODE_IMIN
:
2855 for (c
= 0; c
< 4; c
++) {
2856 if (!(mask
& (1 << c
)))
2858 emit_minmax(pc
, 0x0ac, dst
[c
], src
[0][c
], src
[1][c
]);
2861 case TGSI_OPCODE_INEG
:
2862 for (c
= 0; c
< 4; c
++) {
2863 if (!(mask
& (1 << c
)))
2865 emit_cvt(pc
, dst
[c
], src
[0][c
], -1,
2866 CVT_S32_S32
| CVT_NEG
);
2869 case TGSI_OPCODE_KIL
:
2870 assert(src
[0][0] && src
[0][1] && src
[0][2] && src
[0][3]);
2871 emit_kil(pc
, src
[0][0]);
2872 emit_kil(pc
, src
[0][1]);
2873 emit_kil(pc
, src
[0][2]);
2874 emit_kil(pc
, src
[0][3]);
2876 case TGSI_OPCODE_KILP
:
2879 case TGSI_OPCODE_LIT
:
2880 emit_lit(pc
, &dst
[0], mask
, &src
[0][0]);
2882 case TGSI_OPCODE_LG2
:
2883 emit_flop(pc
, NV50_FLOP_LG2
, brdc
, src
[0][0]);
2885 case TGSI_OPCODE_LOG
:
2887 struct nv50_reg
*t
[2];
2889 t
[0] = temp_temp(pc
, NULL
);
2890 if (mask
& (1 << 1))
2891 t
[1] = temp_temp(pc
, NULL
);
2895 emit_cvt(pc
, t
[0], src
[0][0], -1, CVT_ABS
| CVT_F32_F32
);
2896 emit_flop(pc
, NV50_FLOP_LG2
, t
[1], t
[0]);
2897 if (mask
& (1 << 2))
2898 emit_mov(pc
, dst
[2], t
[1]);
2899 emit_flr(pc
, t
[1], t
[1]);
2900 if (mask
& (1 << 0))
2901 emit_mov(pc
, dst
[0], t
[1]);
2902 if (mask
& (1 << 1)) {
2903 t
[1]->mod
= NV50_MOD_NEG
;
2904 emit_preex2(pc
, t
[1], t
[1]);
2906 emit_flop(pc
, NV50_FLOP_EX2
, t
[1], t
[1]);
2907 emit_mul(pc
, dst
[1], t
[0], t
[1]);
2909 if (mask
& (1 << 3))
2910 emit_mov_immdval(pc
, dst
[3], 1.0f
);
2913 case TGSI_OPCODE_LRP
:
2914 temp
= temp_temp(pc
, NULL
);
2915 for (c
= 0; c
< 4; c
++) {
2916 if (!(mask
& (1 << c
)))
2918 emit_sub(pc
, temp
, src
[1][c
], src
[2][c
]);
2919 emit_mad(pc
, dst
[c
], temp
, src
[0][c
], src
[2][c
]);
2922 case TGSI_OPCODE_MAD
:
2923 for (c
= 0; c
< 4; c
++) {
2924 if (!(mask
& (1 << c
)))
2926 emit_mad(pc
, dst
[c
], src
[0][c
], src
[1][c
], src
[2][c
]);
2929 case TGSI_OPCODE_MAX
:
2930 for (c
= 0; c
< 4; c
++) {
2931 if (!(mask
& (1 << c
)))
2933 emit_minmax(pc
, 0x880, dst
[c
], src
[0][c
], src
[1][c
]);
2936 case TGSI_OPCODE_MIN
:
2937 for (c
= 0; c
< 4; c
++) {
2938 if (!(mask
& (1 << c
)))
2940 emit_minmax(pc
, 0x8a0, dst
[c
], src
[0][c
], src
[1][c
]);
2943 case TGSI_OPCODE_MOV
:
2944 for (c
= 0; c
< 4; c
++) {
2945 if (!(mask
& (1 << c
)))
2947 emit_mov(pc
, dst
[c
], src
[0][c
]);
2950 case TGSI_OPCODE_MUL
:
2951 for (c
= 0; c
< 4; c
++) {
2952 if (!(mask
& (1 << c
)))
2954 emit_mul(pc
, dst
[c
], src
[0][c
], src
[1][c
]);
2957 case TGSI_OPCODE_NOT
:
2958 for (c
= 0; c
< 4; c
++) {
2959 if (!(mask
& (1 << c
)))
2961 emit_not(pc
, dst
[c
], src
[0][c
]);
2964 case TGSI_OPCODE_POW
:
2965 emit_pow(pc
, brdc
, src
[0][0], src
[1][0]);
2967 case TGSI_OPCODE_RCP
:
2968 if (!sat
&& popcnt4(mask
) == 1)
2969 brdc
= dst
[ffs(mask
) - 1];
2970 emit_flop(pc
, NV50_FLOP_RCP
, brdc
, src
[0][0]);
2972 case TGSI_OPCODE_RET
:
2973 if (pc
->p
->type
== PIPE_SHADER_FRAGMENT
&& !pc
->in_subroutine
)
2974 nv50_fp_move_results(pc
);
2975 emit_ret(pc
, -1, 0);
2977 case TGSI_OPCODE_RSQ
:
2978 if (!sat
&& popcnt4(mask
) == 1)
2979 brdc
= dst
[ffs(mask
) - 1];
2980 src
[0][0]->mod
|= NV50_MOD_ABS
;
2981 emit_flop(pc
, NV50_FLOP_RSQ
, brdc
, src
[0][0]);
2983 case TGSI_OPCODE_SAD
:
2984 for (c
= 0; c
< 4; c
++) {
2985 if (!(mask
& (1 << c
)))
2987 emit_sad(pc
, dst
[c
], src
[0][c
], src
[1][c
], src
[2][c
]);
2990 case TGSI_OPCODE_SCS
:
2991 temp
= temp_temp(pc
, NULL
);
2993 emit_precossin(pc
, temp
, src
[0][0]);
2994 if (mask
& (1 << 0))
2995 emit_flop(pc
, NV50_FLOP_COS
, dst
[0], temp
);
2996 if (mask
& (1 << 1))
2997 emit_flop(pc
, NV50_FLOP_SIN
, dst
[1], temp
);
2998 if (mask
& (1 << 2))
2999 emit_mov_immdval(pc
, dst
[2], 0.0);
3000 if (mask
& (1 << 3))
3001 emit_mov_immdval(pc
, dst
[3], 1.0);
3003 case TGSI_OPCODE_SHL
:
3004 case TGSI_OPCODE_ISHR
:
3005 case TGSI_OPCODE_USHR
:
3006 for (c
= 0; c
< 4; c
++) {
3007 if (!(mask
& (1 << c
)))
3009 emit_shift(pc
, dst
[c
], src
[0][c
], src
[1][c
],
3010 inst
->Instruction
.Opcode
);
3013 case TGSI_OPCODE_SIN
:
3015 emit_precossin(pc
, temp
, src
[0][3]);
3016 emit_flop(pc
, NV50_FLOP_SIN
, dst
[3], temp
);
3020 temp
= brdc
= temp_temp(pc
, NULL
);
3022 emit_precossin(pc
, temp
, src
[0][0]);
3023 emit_flop(pc
, NV50_FLOP_SIN
, brdc
, temp
);
3025 case TGSI_OPCODE_SLT
:
3026 case TGSI_OPCODE_SGE
:
3027 case TGSI_OPCODE_SEQ
:
3028 case TGSI_OPCODE_SGT
:
3029 case TGSI_OPCODE_SLE
:
3030 case TGSI_OPCODE_SNE
:
3031 case TGSI_OPCODE_ISLT
:
3032 case TGSI_OPCODE_ISGE
:
3033 case TGSI_OPCODE_USEQ
:
3034 case TGSI_OPCODE_USGE
:
3035 case TGSI_OPCODE_USLT
:
3036 case TGSI_OPCODE_USNE
:
3040 map_tgsi_setop_hw(inst
->Instruction
.Opcode
, &cc
, &ty
);
3042 for (c
= 0; c
< 4; c
++) {
3043 if (!(mask
& (1 << c
)))
3045 emit_set(pc
, cc
, dst
[c
], -1, src
[0][c
], src
[1][c
], ty
);
3049 case TGSI_OPCODE_SUB
:
3050 for (c
= 0; c
< 4; c
++) {
3051 if (!(mask
& (1 << c
)))
3053 emit_sub(pc
, dst
[c
], src
[0][c
], src
[1][c
]);
3056 case TGSI_OPCODE_TEX
:
3057 emit_tex(pc
, dst
, mask
, src
[0], unit
,
3058 inst
->Texture
.Texture
, FALSE
, 0);
3060 case TGSI_OPCODE_TXB
:
3061 emit_tex(pc
, dst
, mask
, src
[0], unit
,
3062 inst
->Texture
.Texture
, FALSE
, -1);
3064 case TGSI_OPCODE_TXL
:
3065 emit_tex(pc
, dst
, mask
, src
[0], unit
,
3066 inst
->Texture
.Texture
, FALSE
, 1);
3068 case TGSI_OPCODE_TXP
:
3069 emit_tex(pc
, dst
, mask
, src
[0], unit
,
3070 inst
->Texture
.Texture
, TRUE
, 0);
3072 case TGSI_OPCODE_TRUNC
:
3073 for (c
= 0; c
< 4; c
++) {
3074 if (!(mask
& (1 << c
)))
3076 emit_cvt(pc
, dst
[c
], src
[0][c
], -1,
3077 CVT_TRUNC
| CVT_F32_F32
| CVT_RI
);
3080 case TGSI_OPCODE_U2F
:
3081 for (c
= 0; c
< 4; c
++) {
3082 if (!(mask
& (1 << c
)))
3084 emit_cvt(pc
, dst
[c
], src
[0][c
], -1, CVT_F32_U32
);
3087 case TGSI_OPCODE_UADD
:
3088 for (c
= 0; c
< 4; c
++) {
3089 if (!(mask
& (1 << c
)))
3091 emit_add_b32(pc
, dst
[c
], src
[0][c
], src
[1][c
]);
3094 case TGSI_OPCODE_UMAX
:
3095 for (c
= 0; c
< 4; c
++) {
3096 if (!(mask
& (1 << c
)))
3098 emit_minmax(pc
, 0x084, dst
[c
], src
[0][c
], src
[1][c
]);
3101 case TGSI_OPCODE_UMIN
:
3102 for (c
= 0; c
< 4; c
++) {
3103 if (!(mask
& (1 << c
)))
3105 emit_minmax(pc
, 0x0a4, dst
[c
], src
[0][c
], src
[1][c
]);
3108 case TGSI_OPCODE_UMAD
:
3111 temp
= temp_temp(pc
, NULL
);
3112 for (c
= 0; c
< 4; c
++) {
3113 if (!(mask
& (1 << c
)))
3115 emit_mul_u16(pc
, temp
, src
[0][c
], 0, src
[1][c
], 1);
3116 emit_mad_u16(pc
, temp
, src
[0][c
], 1, src
[1][c
], 0,
3118 emit_shl_imm(pc
, temp
, temp
, 16);
3119 emit_mad_u16(pc
, temp
, src
[0][c
], 0, src
[1][c
], 0,
3121 emit_add_b32(pc
, dst
[c
], temp
, src
[2][c
]);
3125 case TGSI_OPCODE_UMUL
:
3128 temp
= temp_temp(pc
, NULL
);
3129 for (c
= 0; c
< 4; c
++) {
3130 if (!(mask
& (1 << c
)))
3132 emit_mul_u16(pc
, temp
, src
[0][c
], 0, src
[1][c
], 1);
3133 emit_mad_u16(pc
, temp
, src
[0][c
], 1, src
[1][c
], 0,
3135 emit_shl_imm(pc
, temp
, temp
, 16);
3136 emit_mad_u16(pc
, dst
[c
], src
[0][c
], 0, src
[1][c
], 0,
3141 case TGSI_OPCODE_XPD
:
3142 temp
= temp_temp(pc
, NULL
);
3143 if (mask
& (1 << 0)) {
3144 emit_mul(pc
, temp
, src
[0][2], src
[1][1]);
3145 emit_msb(pc
, dst
[0], src
[0][1], src
[1][2], temp
);
3147 if (mask
& (1 << 1)) {
3148 emit_mul(pc
, temp
, src
[0][0], src
[1][2]);
3149 emit_msb(pc
, dst
[1], src
[0][2], src
[1][0], temp
);
3151 if (mask
& (1 << 2)) {
3152 emit_mul(pc
, temp
, src
[0][1], src
[1][0]);
3153 emit_msb(pc
, dst
[2], src
[0][0], src
[1][1], temp
);
3155 if (mask
& (1 << 3))
3156 emit_mov_immdval(pc
, dst
[3], 1.0);
3158 case TGSI_OPCODE_END
:
3159 if (pc
->p
->type
== PIPE_SHADER_FRAGMENT
)
3160 nv50_fp_move_results(pc
);
3162 /* last insn must be long so it can have the exit bit set */
3163 if (!is_long(pc
->p
->exec_tail
))
3164 convert_to_long(pc
, pc
->p
->exec_tail
);
3166 if (is_immd(pc
->p
->exec_tail
) || is_join(pc
->p
->exec_tail
))
3169 pc
->p
->exec_tail
->inst
[1] |= 1; /* set exit bit */
3174 NOUVEAU_ERR("invalid opcode %d\n", inst
->Instruction
.Opcode
);
3180 emit_sat(pc
, brdc
, brdc
);
3181 for (c
= 0; c
< 4; c
++)
3182 if ((mask
& (1 << c
)) && dst
[c
] != brdc
)
3183 emit_mov(pc
, dst
[c
], brdc
);
3186 for (c
= 0; c
< 4; c
++) {
3187 if (!(mask
& (1 << c
)))
3189 /* In this case we saturate later, and dst[c] won't
3190 * be another temp_temp (and thus lost), since rdst
3191 * already is TEMP (see above). */
3192 if (rdst
[c
]->type
== P_TEMP
&& rdst
[c
]->index
< 0)
3194 emit_sat(pc
, rdst
[c
], dst
[c
]);
3198 kill_temp_temp(pc
, NULL
);
3199 pc
->reg_instance_nr
= 0;
3205 prep_inspect_insn(struct nv50_pc
*pc
, const struct tgsi_full_instruction
*insn
)
3207 struct nv50_reg
*r
, *reg
= NULL
;
3208 const struct tgsi_full_src_register
*src
;
3209 const struct tgsi_dst_register
*dst
;
3210 unsigned i
, c
, k
, mask
;
3212 dst
= &insn
->Dst
[0].Register
;
3213 mask
= dst
->WriteMask
;
3215 if (dst
->File
== TGSI_FILE_TEMPORARY
)
3218 if (dst
->File
== TGSI_FILE_OUTPUT
) {
3221 if (insn
->Instruction
.Opcode
== TGSI_OPCODE_MOV
&&
3222 dst
->Index
== pc
->edgeflag_out
&&
3223 insn
->Src
[0].Register
.File
== TGSI_FILE_INPUT
)
3224 pc
->p
->cfg
.edgeflag_in
= insn
->Src
[0].Register
.Index
;
3228 for (c
= 0; c
< 4; c
++) {
3229 if (!(mask
& (1 << c
)))
3231 reg
[dst
->Index
* 4 + c
].acc
= pc
->insn_nr
;
3235 for (i
= 0; i
< insn
->Instruction
.NumSrcRegs
; i
++) {
3236 src
= &insn
->Src
[i
];
3238 if (src
->Register
.File
== TGSI_FILE_TEMPORARY
)
3241 if (src
->Register
.File
== TGSI_FILE_INPUT
)
3246 mask
= nv50_tgsi_src_mask(insn
, i
);
3248 for (c
= 0; c
< 4; c
++) {
3249 if (!(mask
& (1 << c
)))
3251 k
= tgsi_util_get_full_src_register_swizzle(src
, c
);
3253 r
= ®
[src
->Register
.Index
* 4 + k
];
3255 /* If used before written, pre-allocate the reg,
3256 * lest we overwrite results from a subroutine.
3258 if (!r
->acc
&& r
->type
== P_TEMP
)
3261 r
->acc
= pc
->insn_nr
;
3266 /* Returns a bitmask indicating which dst components need to be
3267 * written to temporaries first to avoid 'corrupting' sources.
3269 * m[i] (out) indicate component to write in the i-th position
3270 * rdep[c] (in) bitmasks of dst[i] that require dst[c] as source
3273 nv50_revdep_reorder(unsigned m
[4], unsigned rdep
[4])
3275 unsigned i
, c
, x
, unsafe
;
3277 for (c
= 0; c
< 4; c
++)
3280 /* Swap as long as a dst component written earlier is depended on
3281 * by one written later, but the next one isn't depended on by it.
3283 for (c
= 0; c
< 3; c
++) {
3284 if (rdep
[m
[c
+ 1]] & (1 << m
[c
]))
3285 continue; /* if next one is depended on by us */
3286 for (i
= c
+ 1; i
< 4; i
++)
3287 /* if we are depended on by a later one */
3288 if (rdep
[m
[c
]] & (1 << m
[i
]))
3301 /* mark dependencies that could not be resolved by reordering */
3302 for (i
= 0; i
< 3; ++i
)
3303 for (c
= i
+ 1; c
< 4; ++c
)
3304 if (rdep
[m
[i
]] & (1 << m
[c
]))
3307 /* NOTE: $unsafe is with respect to order, not component */
3311 /* Select a suitable dst register for broadcasting scalar results,
3312 * or return NULL if we have to allocate an extra TEMP.
3314 * If e.g. only 1 component is written, we may also emit the final
3315 * result to a write-only register.
3317 static struct nv50_reg
*
3318 tgsi_broadcast_dst(struct nv50_pc
*pc
,
3319 const struct tgsi_full_dst_register
*fd
, unsigned mask
)
3321 if (fd
->Register
.File
== TGSI_FILE_TEMPORARY
) {
3322 int c
= ffs(~mask
& fd
->Register
.WriteMask
);
3324 return tgsi_dst(pc
, c
- 1, fd
);
3326 int c
= ffs(fd
->Register
.WriteMask
) - 1;
3327 if ((1 << c
) == fd
->Register
.WriteMask
)
3328 return tgsi_dst(pc
, c
, fd
);
3334 /* Scan source swizzles and return a bitmask indicating dst regs that
3335 * also occur among the src regs, and fill rdep for nv50_revdep_reoder.
3338 nv50_tgsi_scan_swizzle(const struct tgsi_full_instruction
*insn
,
3341 const struct tgsi_full_dst_register
*fd
= &insn
->Dst
[0];
3342 const struct tgsi_full_src_register
*fs
;
3343 unsigned i
, deqs
= 0;
3345 for (i
= 0; i
< 4; ++i
)
3348 for (i
= 0; i
< insn
->Instruction
.NumSrcRegs
; i
++) {
3349 unsigned chn
, mask
= nv50_tgsi_src_mask(insn
, i
);
3350 int ms
= get_supported_mods(insn
, i
);
3353 if (fs
->Register
.File
!= fd
->Register
.File
||
3354 fs
->Register
.Index
!= fd
->Register
.Index
)
3357 for (chn
= 0; chn
< 4; ++chn
) {
3360 if (!(mask
& (1 << chn
))) /* src is not read */
3362 c
= tgsi_util_get_full_src_register_swizzle(fs
, chn
);
3363 s
= tgsi_util_get_full_src_register_sign_mode(fs
, chn
);
3365 if (!(fd
->Register
.WriteMask
& (1 << c
)))
3368 if (s
== TGSI_UTIL_SIGN_TOGGLE
&& !(ms
& NV50_MOD_NEG
))
3370 if (s
== TGSI_UTIL_SIGN_CLEAR
&& !(ms
& NV50_MOD_ABS
))
3372 if ((s
== TGSI_UTIL_SIGN_SET
) && ((ms
& 3) != 3))
3375 rdep
[c
] |= nv50_tgsi_dst_revdep(
3376 insn
->Instruction
.Opcode
, i
, chn
);
3385 nv50_tgsi_insn(struct nv50_pc
*pc
, const union tgsi_full_token
*tok
)
3387 struct tgsi_full_instruction insn
= tok
->FullInstruction
;
3388 const struct tgsi_full_dst_register
*fd
;
3389 unsigned i
, deqs
, rdep
[4], m
[4];
3391 fd
= &tok
->FullInstruction
.Dst
[0];
3392 deqs
= nv50_tgsi_scan_swizzle(&insn
, rdep
);
3394 if (is_scalar_op(insn
.Instruction
.Opcode
)) {
3395 pc
->r_brdc
= tgsi_broadcast_dst(pc
, fd
, deqs
);
3397 pc
->r_brdc
= temp_temp(pc
, NULL
);
3398 return nv50_program_tx_insn(pc
, &insn
);
3402 if (!deqs
|| (!rdep
[0] && !rdep
[1] && !rdep
[2] && !rdep
[3]))
3403 return nv50_program_tx_insn(pc
, &insn
);
3405 deqs
= nv50_revdep_reorder(m
, rdep
);
3407 for (i
= 0; i
< 4; ++i
) {
3408 assert(pc
->r_dst
[m
[i
]] == NULL
);
3410 insn
.Dst
[0].Register
.WriteMask
=
3411 fd
->Register
.WriteMask
& (1 << m
[i
]);
3413 if (!insn
.Dst
[0].Register
.WriteMask
)
3416 if (deqs
& (1 << i
))
3417 pc
->r_dst
[m
[i
]] = alloc_temp(pc
, NULL
);
3419 if (!nv50_program_tx_insn(pc
, &insn
))
3423 for (i
= 0; i
< 4; i
++) {
3424 struct nv50_reg
*reg
= pc
->r_dst
[i
];
3427 pc
->r_dst
[i
] = NULL
;
3429 if (insn
.Instruction
.Saturate
== TGSI_SAT_ZERO_ONE
)
3430 emit_sat(pc
, tgsi_dst(pc
, i
, fd
), reg
);
3432 emit_mov(pc
, tgsi_dst(pc
, i
, fd
), reg
);
3440 load_interpolant(struct nv50_pc
*pc
, struct nv50_reg
*reg
)
3442 struct nv50_reg
*iv
, **ppiv
;
3443 unsigned mode
= pc
->interp_mode
[reg
->index
];
3445 ppiv
= (mode
& INTERP_CENTROID
) ? &pc
->iv_c
: &pc
->iv_p
;
3448 if ((mode
& INTERP_PERSPECTIVE
) && !iv
) {
3449 iv
= *ppiv
= alloc_temp(pc
, NULL
);
3450 iv
->rhw
= popcnt4(pc
->p
->cfg
.regs
[1] >> 24) - 1;
3452 emit_interp(pc
, iv
, NULL
, mode
& INTERP_CENTROID
);
3453 emit_flop(pc
, NV50_FLOP_RCP
, iv
, iv
);
3455 /* XXX: when loading interpolants dynamically, move these
3456 * to the program head, or make sure it can't be skipped.
3460 emit_interp(pc
, reg
, iv
, mode
);
3463 /* The face input is always at v[255] (varying space), with a
3464 * value of 0 for back-facing, and 0xffffffff for front-facing.
3467 load_frontfacing(struct nv50_pc
*pc
, struct nv50_reg
*sv
)
3469 struct nv50_reg
*temp
= alloc_temp(pc
, NULL
);
3473 emit_interp(pc
, temp
, NULL
, INTERP_FLAT
);
3475 emit_cvt(pc
, sv
, temp
, r_pred
, CVT_ABS
| CVT_F32_S32
);
3477 emit_not(pc
, temp
, temp
);
3478 set_pred(pc
, 0x2, r_pred
, pc
->p
->exec_tail
);
3479 emit_cvt(pc
, sv
, temp
, -1, CVT_F32_S32
);
3480 set_pred(pc
, 0x2, r_pred
, pc
->p
->exec_tail
);
3482 free_temp(pc
, temp
);
3486 copy_semantic_info(struct nv50_program
*p
)
3490 for (i
= 0; i
< p
->cfg
.in_nr
; ++i
) {
3491 id
= p
->cfg
.in
[i
].id
;
3492 p
->cfg
.in
[i
].sn
= p
->info
.input_semantic_name
[id
];
3493 p
->cfg
.in
[i
].si
= p
->info
.input_semantic_index
[id
];
3496 for (i
= 0; i
< p
->cfg
.out_nr
; ++i
) {
3497 id
= p
->cfg
.out
[i
].id
;
3498 p
->cfg
.out
[i
].sn
= p
->info
.output_semantic_name
[id
];
3499 p
->cfg
.out
[i
].si
= p
->info
.output_semantic_index
[id
];
3504 nv50_program_tx_prep(struct nv50_pc
*pc
)
3506 struct tgsi_parse_context tp
;
3507 struct nv50_program
*p
= pc
->p
;
3508 boolean ret
= FALSE
;
3509 unsigned i
, c
, instance_id
, vertex_id
, flat_nr
= 0;
3511 tgsi_parse_init(&tp
, pc
->p
->pipe
.tokens
);
3512 while (!tgsi_parse_end_of_tokens(&tp
)) {
3513 const union tgsi_full_token
*tok
= &tp
.FullToken
;
3515 tgsi_parse_token(&tp
);
3516 switch (tok
->Token
.Type
) {
3517 case TGSI_TOKEN_TYPE_IMMEDIATE
:
3519 const struct tgsi_full_immediate
*imm
=
3520 &tp
.FullToken
.FullImmediate
;
3522 ctor_immd_4f32(pc
, imm
->u
[0].Float
,
3528 case TGSI_TOKEN_TYPE_DECLARATION
:
3530 const struct tgsi_full_declaration
*d
;
3531 unsigned si
, last
, first
, mode
;
3533 d
= &tp
.FullToken
.FullDeclaration
;
3534 first
= d
->Range
.First
;
3535 last
= d
->Range
.Last
;
3537 switch (d
->Declaration
.File
) {
3538 case TGSI_FILE_TEMPORARY
:
3540 case TGSI_FILE_OUTPUT
:
3541 if (!d
->Declaration
.Semantic
||
3542 p
->type
== PIPE_SHADER_FRAGMENT
)
3545 si
= d
->Semantic
.Index
;
3546 switch (d
->Semantic
.Name
) {
3547 case TGSI_SEMANTIC_BCOLOR
:
3548 p
->cfg
.two_side
[si
].hw
= first
;
3549 if (p
->cfg
.out_nr
> first
)
3550 p
->cfg
.out_nr
= first
;
3552 case TGSI_SEMANTIC_PSIZE
:
3553 p
->cfg
.psiz
= first
;
3554 if (p
->cfg
.out_nr
> first
)
3555 p
->cfg
.out_nr
= first
;
3557 case TGSI_SEMANTIC_EDGEFLAG
:
3558 pc
->edgeflag_out
= first
;
3561 case TGSI_SEMANTIC_CLIP_DISTANCE:
3562 p->cfg.clpd = MIN2(p->cfg.clpd, first);
3569 case TGSI_FILE_INPUT
:
3571 if (p
->type
!= PIPE_SHADER_FRAGMENT
)
3574 switch (d
->Declaration
.Interpolate
) {
3575 case TGSI_INTERPOLATE_CONSTANT
:
3579 case TGSI_INTERPOLATE_PERSPECTIVE
:
3580 mode
= INTERP_PERSPECTIVE
;
3581 p
->cfg
.regs
[1] |= 0x08 << 24;
3584 mode
= INTERP_LINEAR
;
3587 if (d
->Declaration
.Centroid
)
3588 mode
|= INTERP_CENTROID
;
3591 for (i
= first
; i
<= last
; i
++)
3592 pc
->interp_mode
[i
] = mode
;
3595 case TGSI_FILE_SYSTEM_VALUE
:
3596 assert(d
->Declaration
.Semantic
);
3597 switch (d
->Semantic
.Name
) {
3598 case TGSI_SEMANTIC_FACE
:
3599 assert(p
->type
== PIPE_SHADER_FRAGMENT
);
3600 load_frontfacing(pc
,
3601 &pc
->sysval
[first
]);
3603 case TGSI_SEMANTIC_INSTANCEID
:
3604 assert(p
->type
== PIPE_SHADER_VERTEX
);
3605 instance_id
= first
;
3606 p
->cfg
.regs
[0] |= (1 << 4);
3608 case TGSI_SEMANTIC_PRIMID
:
3609 assert(p
->type
!= PIPE_SHADER_VERTEX
);
3610 p
->cfg
.prim_id
= first
;
3613 case TGSI_SEMANTIC_PRIMIDIN:
3614 assert(p->type == PIPE_SHADER_GEOMETRY);
3615 pc->sysval[first].hw = 6;
3616 p->cfg.regs[0] |= (1 << 8);
3618 case TGSI_SEMANTIC_VERTEXID:
3619 assert(p->type == PIPE_SHADER_VERTEX);
3621 p->cfg.regs[0] |= (1 << 12) | (1 << 0);
3626 case TGSI_FILE_ADDRESS
:
3627 case TGSI_FILE_CONSTANT
:
3628 case TGSI_FILE_SAMPLER
:
3631 NOUVEAU_ERR("bad decl file %d\n",
3632 d
->Declaration
.File
);
3637 case TGSI_TOKEN_TYPE_INSTRUCTION
:
3639 prep_inspect_insn(pc
, &tok
->FullInstruction
);
3646 if (p
->type
== PIPE_SHADER_VERTEX
|| p
->type
== PIPE_SHADER_GEOMETRY
) {
3649 if (p
->type
== PIPE_SHADER_GEOMETRY
) {
3650 for (i
= 0; i
< pc
->attr_nr
; ++i
) {
3651 p
->cfg
.in
[i
].hw
= rid
;
3652 p
->cfg
.in
[i
].id
= i
;
3654 for (c
= 0; c
< 4; ++c
) {
3656 if (!pc
->attr
[n
].acc
)
3658 pc
->attr
[n
].hw
= rid
++;
3659 p
->cfg
.in
[i
].mask
|= 1 << c
;
3663 for (i
= 0; i
< pc
->attr_nr
* 4; ++i
) {
3664 if (pc
->attr
[i
].acc
) {
3665 pc
->attr
[i
].hw
= rid
++;
3666 p
->cfg
.attr
[i
/ 32] |= 1 << (i
% 32);
3669 if (p
->cfg
.regs
[0] & (1 << 0))
3670 pc
->sysval
[vertex_id
].hw
= rid
++;
3671 if (p
->cfg
.regs
[0] & (1 << 4))
3672 pc
->sysval
[instance_id
].hw
= rid
++;
3675 for (i
= 0, rid
= 0; i
< pc
->result_nr
; ++i
) {
3676 p
->cfg
.out
[i
].hw
= rid
;
3677 p
->cfg
.out
[i
].id
= i
;
3679 for (c
= 0; c
< 4; ++c
) {
3681 if (!pc
->result
[n
].acc
)
3683 pc
->result
[n
].hw
= rid
++;
3684 p
->cfg
.out
[i
].mask
|= 1 << c
;
3687 if (p
->cfg
.prim_id
< 0x40) {
3688 /* GP has to write to PrimitiveID */
3689 ctor_reg(&pc
->sysval
[p
->cfg
.prim_id
],
3690 P_RESULT
, p
->cfg
.prim_id
, rid
);
3691 p
->cfg
.prim_id
= rid
++;
3694 for (c
= 0; c
< 2; ++c
)
3695 if (p
->cfg
.two_side
[c
].hw
< 0x40)
3696 p
->cfg
.two_side
[c
] = p
->cfg
.out
[
3697 p
->cfg
.two_side
[c
].hw
];
3699 if (p
->cfg
.psiz
< 0x40)
3700 p
->cfg
.psiz
= p
->cfg
.out
[p
->cfg
.psiz
].hw
;
3702 copy_semantic_info(p
);
3704 if (p
->type
== PIPE_SHADER_FRAGMENT
) {
3706 unsigned n
= 0, m
= pc
->attr_nr
- flat_nr
;
3710 base
= (TGSI_SEMANTIC_POSITION
==
3711 p
->info
.input_semantic_name
[0]) ? 0 : 1;
3713 /* non-flat interpolants have to be mapped to
3714 * the lower hardware IDs, so sort them:
3716 for (i
= 0; i
< pc
->attr_nr
; i
++) {
3717 if (pc
->interp_mode
[i
] == INTERP_FLAT
)
3718 p
->cfg
.in
[m
++].id
= i
;
3720 if (!(pc
->interp_mode
[i
] & INTERP_PERSPECTIVE
))
3721 p
->cfg
.in
[n
].linear
= TRUE
;
3722 p
->cfg
.in
[n
++].id
= i
;
3725 copy_semantic_info(p
);
3727 if (!base
) /* set w-coordinate mask from perspective interp */
3728 p
->cfg
.in
[0].mask
|= p
->cfg
.regs
[1] >> 24;
3730 aid
= popcnt4( /* if fcrd isn't contained in cfg.io */
3731 base
? (p
->cfg
.regs
[1] >> 24) : p
->cfg
.in
[0].mask
);
3733 for (n
= 0; n
< pc
->attr_nr
; ++n
) {
3734 p
->cfg
.in
[n
].hw
= rid
= aid
;
3735 i
= p
->cfg
.in
[n
].id
;
3737 if (p
->info
.input_semantic_name
[n
] ==
3738 TGSI_SEMANTIC_FACE
) {
3739 load_frontfacing(pc
, &pc
->attr
[i
* 4]);
3743 for (c
= 0; c
< 4; ++c
) {
3744 if (!pc
->attr
[i
* 4 + c
].acc
)
3746 pc
->attr
[i
* 4 + c
].rhw
= rid
++;
3747 p
->cfg
.in
[n
].mask
|= 1 << c
;
3749 load_interpolant(pc
, &pc
->attr
[i
* 4 + c
]);
3751 aid
+= popcnt4(p
->cfg
.in
[n
].mask
);
3755 p
->cfg
.regs
[1] |= p
->cfg
.in
[0].mask
<< 24;
3757 m
= popcnt4(p
->cfg
.regs
[1] >> 24);
3759 /* set count of non-position inputs and of non-flat
3760 * non-position inputs for FP_INTERPOLANT_CTRL
3762 p
->cfg
.regs
[1] |= aid
- m
;
3765 i
= p
->cfg
.in
[pc
->attr_nr
- flat_nr
].hw
;
3766 p
->cfg
.regs
[1] |= (i
- m
) << 16;
3768 p
->cfg
.regs
[1] |= p
->cfg
.regs
[1] << 16;
3770 /* mark color semantic for light-twoside */
3772 for (i
= 0; i
< p
->cfg
.in_nr
; i
++) {
3773 if (p
->cfg
.in
[i
].sn
== TGSI_SEMANTIC_COLOR
) {
3774 n
= MIN2(n
, p
->cfg
.in
[i
].hw
- m
);
3775 p
->cfg
.two_side
[p
->cfg
.in
[i
].si
] = p
->cfg
.in
[i
];
3777 p
->cfg
.regs
[0] += /* increase colour count */
3778 popcnt4(p
->cfg
.in
[i
].mask
) << 16;
3782 p
->cfg
.regs
[0] += n
;
3784 if (p
->cfg
.prim_id
< 0x40) {
3785 pc
->sysval
[p
->cfg
.prim_id
].rhw
= rid
++;
3786 emit_interp(pc
, &pc
->sysval
[p
->cfg
.prim_id
], NULL
,
3788 /* increase FP_INTERPOLANT_CTRL_COUNT */
3789 p
->cfg
.regs
[1] += 1;
3792 /* Initialize FP results:
3793 * FragDepth is always first TGSI and last hw output
3795 i
= p
->info
.writes_z
? 4 : 0;
3796 for (rid
= 0; i
< pc
->result_nr
* 4; i
++)
3797 pc
->result
[i
].rhw
= rid
++;
3798 if (p
->info
.writes_z
)
3799 pc
->result
[2].rhw
= rid
;
3801 p
->cfg
.high_result
= rid
;
3803 /* separate/different colour results for MRTs ? */
3804 if (pc
->result_nr
- (p
->info
.writes_z
? 1 : 0) > 1)
3805 p
->cfg
.regs
[2] |= 1;
3811 pc
->immd
= MALLOC(pc
->immd_nr
* 4 * sizeof(struct nv50_reg
));
3815 for (i
= 0; i
< pc
->immd_nr
; i
++) {
3816 for (c
= 0; c
< 4; c
++, rid
++)
3817 ctor_reg(&pc
->immd
[rid
], P_IMMD
, i
, rid
);
3824 free_temp(pc
, pc
->iv_p
);
3826 free_temp(pc
, pc
->iv_c
);
3828 tgsi_parse_free(&tp
);
3833 free_nv50_pc(struct nv50_pc
*pc
)
3853 static INLINE
uint32_t
3854 nv50_map_gs_output_prim(unsigned pprim
)
3857 case PIPE_PRIM_POINTS
:
3858 return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_POINTS
;
3859 case PIPE_PRIM_LINE_STRIP
:
3860 return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_LINE_STRIP
;
3861 case PIPE_PRIM_TRIANGLE_STRIP
:
3862 return NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE_TRIANGLE_STRIP
;
3864 NOUVEAU_ERR("invalid GS_OUTPUT_PRIMITIVE: %u\n", pprim
);
3871 ctor_nv50_pc(struct nv50_pc
*pc
, struct nv50_program
*p
)
3874 unsigned rtype
[2] = { P_ATTR
, P_RESULT
};
3877 pc
->temp_nr
= p
->info
.file_max
[TGSI_FILE_TEMPORARY
] + 1;
3878 pc
->attr_nr
= p
->info
.file_max
[TGSI_FILE_INPUT
] + 1;
3879 pc
->result_nr
= p
->info
.file_max
[TGSI_FILE_OUTPUT
] + 1;
3880 pc
->param_nr
= p
->info
.file_max
[TGSI_FILE_CONSTANT
] + 1;
3881 pc
->addr_nr
= p
->info
.file_max
[TGSI_FILE_ADDRESS
] + 1;
3882 assert(pc
->addr_nr
<= 2);
3883 pc
->sysval_nr
= p
->info
.file_max
[TGSI_FILE_SYSTEM_VALUE
] + 1;
3885 p
->cfg
.high_temp
= 4;
3887 p
->cfg
.two_side
[0].hw
= 0x40;
3888 p
->cfg
.two_side
[1].hw
= 0x40;
3889 p
->cfg
.prim_id
= 0x40;
3891 p
->cfg
.edgeflag_in
= pc
->edgeflag_out
= 0xff;
3893 for (i
= 0; i
< p
->info
.num_properties
; ++i
) {
3894 unsigned *data
= &p
->info
.properties
[i
].data
[0];
3896 switch (p
->info
.properties
[i
].name
) {
3897 case TGSI_PROPERTY_GS_OUTPUT_PRIM
:
3898 p
->cfg
.prim_type
= nv50_map_gs_output_prim(data
[0]);
3900 case TGSI_PROPERTY_GS_MAX_VERTICES
:
3901 p
->cfg
.vert_count
= data
[0];
3909 case PIPE_SHADER_VERTEX
:
3912 p
->cfg
.out_nr
= pc
->result_nr
;
3914 case PIPE_SHADER_GEOMETRY
:
3915 assert(p
->cfg
.prim_type
);
3916 assert(p
->cfg
.vert_count
);
3920 p
->cfg
.prim_id
= 0x80;
3921 p
->cfg
.out_nr
= pc
->result_nr
;
3922 p
->cfg
.in_nr
= pc
->attr_nr
;
3924 p
->cfg
.two_side
[0].hw
= 0x80;
3925 p
->cfg
.two_side
[1].hw
= 0x80;
3927 case PIPE_SHADER_FRAGMENT
:
3928 rtype
[0] = rtype
[1] = P_TEMP
;
3930 p
->cfg
.regs
[0] = 0x01000004;
3931 p
->cfg
.in_nr
= pc
->attr_nr
;
3933 if (p
->info
.writes_z
) {
3934 p
->cfg
.regs
[2] |= 0x00000100;
3935 p
->cfg
.regs
[3] |= 0x00000011;
3937 if (p
->info
.uses_kill
)
3938 p
->cfg
.regs
[2] |= 0x00100000;
3943 pc
->temp
= MALLOC(pc
->temp_nr
* 4 * sizeof(struct nv50_reg
));
3947 for (i
= 0; i
< pc
->temp_nr
* 4; ++i
)
3948 ctor_reg(&pc
->temp
[i
], P_TEMP
, i
/ 4, -1);
3952 pc
->attr
= MALLOC(pc
->attr_nr
* 4 * sizeof(struct nv50_reg
));
3956 for (i
= 0; i
< pc
->attr_nr
* 4; ++i
)
3957 ctor_reg(&pc
->attr
[i
], rtype
[0], i
/ 4, -1);
3960 if (pc
->result_nr
) {
3961 unsigned nr
= pc
->result_nr
* 4;
3963 pc
->result
= MALLOC(nr
* sizeof(struct nv50_reg
));
3967 for (i
= 0; i
< nr
; ++i
)
3968 ctor_reg(&pc
->result
[i
], rtype
[1], i
/ 4, -1);
3974 pc
->param
= MALLOC(pc
->param_nr
* 4 * sizeof(struct nv50_reg
));
3978 for (i
= 0; i
< pc
->param_nr
; ++i
)
3979 for (c
= 0; c
< 4; ++c
, ++rid
)
3980 ctor_reg(&pc
->param
[rid
], P_CONST
, i
, rid
);
3984 pc
->addr
= CALLOC(pc
->addr_nr
* 4, sizeof(struct nv50_reg
*));
3988 for (i
= 0; i
< NV50_SU_MAX_ADDR
; ++i
)
3989 ctor_reg(&pc
->r_addr
[i
], P_ADDR
, -1, i
+ 1);
3991 if (pc
->sysval_nr
) {
3992 pc
->sysval
= CALLOC(pc
->sysval_nr
, sizeof(struct nv50_reg
*));
3995 /* will only ever use SYSTEM_VALUE[i].x (hopefully) */
3996 for (i
= 0; i
< pc
->sysval_nr
; ++i
)
3997 ctor_reg(&pc
->sysval
[i
], rtype
[0], i
, -1);
4004 nv50_program_fixup_insns(struct nv50_pc
*pc
)
4006 struct nv50_program_exec
*e
, **bra_list
;
4009 bra_list
= CALLOC(pc
->p
->exec_size
, sizeof(struct nv50_program_exec
*));
4011 /* Collect branch instructions, we need to adjust their offsets
4012 * when converting 32 bit instructions to 64 bit ones
4014 for (n
= 0, e
= pc
->p
->exec_head
; e
; e
= e
->next
)
4015 if (e
->param
.index
>= 0 && !e
->param
.mask
)
4018 /* Make sure we don't have any single 32 bit instructions. */
4019 for (e
= pc
->p
->exec_head
, pos
= 0; e
; e
= e
->next
) {
4020 pos
+= is_long(e
) ? 2 : 1;
4022 if ((pos
& 1) && (!e
->next
|| is_long(e
->next
))) {
4023 for (i
= 0; i
< n
; ++i
)
4024 if (bra_list
[i
]->param
.index
>= pos
)
4025 bra_list
[i
]->param
.index
+= 1;
4026 for (i
= 0; i
< pc
->insn_nr
; ++i
)
4027 if (pc
->insn_pos
[i
] >= pos
)
4028 pc
->insn_pos
[i
] += 1;
4029 convert_to_long(pc
, e
);
4036 if (!pc
->p
->info
.opcode_count
[TGSI_OPCODE_CAL
])
4039 /* fill in CALL offsets */
4040 for (e
= pc
->p
->exec_head
; e
; e
= e
->next
) {
4041 if ((e
->inst
[0] & 2) && (e
->inst
[0] >> 28) == 0x2)
4042 e
->param
.index
= pc
->insn_pos
[e
->param
.index
];
4047 nv50_program_tx(struct nv50_program
*p
)
4049 struct tgsi_parse_context parse
;
4053 pc
= CALLOC_STRUCT(nv50_pc
);
4057 ret
= ctor_nv50_pc(pc
, p
);
4061 ret
= nv50_program_tx_prep(pc
);
4065 pc
->insn_pos
= MALLOC(pc
->insn_nr
* sizeof(unsigned));
4067 tgsi_parse_init(&parse
, pc
->p
->pipe
.tokens
);
4068 while (!tgsi_parse_end_of_tokens(&parse
)) {
4069 const union tgsi_full_token
*tok
= &parse
.FullToken
;
4071 /* previously allow32 was FALSE for first & last instruction */
4074 tgsi_parse_token(&parse
);
4076 switch (tok
->Token
.Type
) {
4077 case TGSI_TOKEN_TYPE_INSTRUCTION
:
4078 pc
->insn_pos
[pc
->insn_cur
] = pc
->p
->exec_size
;
4080 ret
= nv50_tgsi_insn(pc
, tok
);
4089 nv50_program_fixup_insns(pc
);
4091 p
->param_nr
= pc
->param_nr
* 4;
4092 p
->immd_nr
= pc
->immd_nr
* 4;
4093 p
->immd
= pc
->immd_buf
;
4096 tgsi_parse_free(&parse
);
4104 nv50_program_validate(struct nv50_context
*nv50
, struct nv50_program
*p
)
4106 if (nv50_program_tx(p
) == FALSE
)
4108 p
->translated
= TRUE
;
4112 nv50_program_upload_data(struct nv50_context
*nv50
, uint32_t *map
,
4113 unsigned start
, unsigned count
, unsigned cbuf
)
4115 struct nouveau_channel
*chan
= nv50
->screen
->base
.channel
;
4116 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
4119 unsigned nr
= count
> 2047 ? 2047 : count
;
4121 BEGIN_RING(chan
, tesla
, NV50TCL_CB_ADDR
, 1);
4122 OUT_RING (chan
, (cbuf
<< 0) | (start
<< 8));
4123 BEGIN_RING(chan
, tesla
, NV50TCL_CB_DATA(0) | 0x40000000, nr
);
4124 OUT_RINGp (chan
, map
, nr
);
4133 nv50_program_validate_data(struct nv50_context
*nv50
, struct nv50_program
*p
)
4135 struct pipe_screen
*pscreen
= nv50
->pipe
.screen
;
4137 if (!p
->data
[0] && p
->immd_nr
) {
4138 struct nouveau_resource
*heap
= nv50
->screen
->immd_heap
[0];
4140 if (nouveau_resource_alloc(heap
, p
->immd_nr
, p
, &p
->data
[0])) {
4141 while (heap
->next
&& heap
->size
< p
->immd_nr
) {
4142 struct nv50_program
*evict
= heap
->next
->priv
;
4143 nouveau_resource_free(&evict
->data
[0]);
4146 if (nouveau_resource_alloc(heap
, p
->immd_nr
, p
,
4151 /* immediates only need to be uploaded again when freed */
4152 nv50_program_upload_data(nv50
, p
->immd
, p
->data
[0]->start
,
4153 p
->immd_nr
, NV50_CB_PMISC
);
4156 assert(p
->param_nr
<= 512);
4160 uint32_t *map
= pipe_buffer_map(pscreen
,
4161 nv50
->constbuf
[p
->type
],
4162 PIPE_BUFFER_USAGE_CPU_READ
);
4164 case PIPE_SHADER_GEOMETRY
: cb
= NV50_CB_PGP
; break;
4165 case PIPE_SHADER_FRAGMENT
: cb
= NV50_CB_PFP
; break;
4168 assert(p
->type
== PIPE_SHADER_VERTEX
);
4172 nv50_program_upload_data(nv50
, map
, 0, p
->param_nr
, cb
);
4173 pipe_buffer_unmap(pscreen
, nv50
->constbuf
[p
->type
]);
4178 nv50_program_validate_code(struct nv50_context
*nv50
, struct nv50_program
*p
)
4180 struct nouveau_channel
*chan
= nv50
->screen
->base
.channel
;
4181 struct nv50_program_exec
*e
;
4183 boolean upload
= FALSE
;
4186 nouveau_bo_new(chan
->device
, NOUVEAU_BO_VRAM
, 0x100,
4187 p
->exec_size
* 4, &p
->bo
);
4191 if (p
->data
[0] && p
->data
[0]->start
!= p
->data_start
[0])
4197 up
= MALLOC(p
->exec_size
* 4);
4199 for (i
= 0, e
= p
->exec_head
; e
; e
= e
->next
) {
4200 unsigned ei
, ci
, bs
;
4202 if (e
->param
.index
>= 0 && e
->param
.mask
) {
4203 bs
= (e
->inst
[1] >> 22) & 0x07;
4205 ei
= e
->param
.shift
>> 5;
4206 ci
= e
->param
.index
;
4208 ci
+= p
->data
[bs
]->start
;
4210 e
->inst
[ei
] &= ~e
->param
.mask
;
4211 e
->inst
[ei
] |= (ci
<< e
->param
.shift
);
4213 if (e
->param
.index
>= 0) {
4214 /* zero mask means param is a jump/branch offset */
4215 assert(!(e
->param
.index
& 1));
4216 /* seem to be 8 byte steps */
4217 ei
= (e
->param
.index
>> 1) + 0 /* START_ID */;
4219 e
->inst
[0] &= 0xf0000fff;
4220 e
->inst
[0] |= ei
<< 12;
4223 up
[i
++] = e
->inst
[0];
4225 up
[i
++] = e
->inst
[1];
4227 assert(i
== p
->exec_size
);
4230 p
->data_start
[0] = p
->data
[0]->start
;
4232 #ifdef NV50_PROGRAM_DUMP
4233 NOUVEAU_ERR("-------\n");
4234 for (e
= p
->exec_head
; e
; e
= e
->next
) {
4235 NOUVEAU_ERR("0x%08x\n", e
->inst
[0]);
4237 NOUVEAU_ERR("0x%08x\n", e
->inst
[1]);
4240 nv50_upload_sifc(nv50
, p
->bo
, 0, NOUVEAU_BO_VRAM
,
4241 NV50_2D_DST_FORMAT_R8_UNORM
, 65536, 1, 262144,
4242 up
, NV50_2D_SIFC_FORMAT_R8_UNORM
, 0,
4243 0, 0, p
->exec_size
* 4, 1, 1);
4249 nv50_vertprog_validate(struct nv50_context
*nv50
)
4251 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
4252 struct nv50_program
*p
= nv50
->vertprog
;
4253 struct nouveau_stateobj
*so
;
4255 if (!p
->translated
) {
4256 nv50_program_validate(nv50
, p
);
4261 nv50_program_validate_data(nv50
, p
);
4262 nv50_program_validate_code(nv50
, p
);
4264 so
= so_new(5, 7, 2);
4265 so_method(so
, tesla
, NV50TCL_VP_ADDRESS_HIGH
, 2);
4266 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
4267 NOUVEAU_BO_HIGH
, 0, 0);
4268 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
4269 NOUVEAU_BO_LOW
, 0, 0);
4270 so_method(so
, tesla
, NV50TCL_VP_ATTR_EN_0
, 2);
4271 so_data (so
, p
->cfg
.attr
[0]);
4272 so_data (so
, p
->cfg
.attr
[1]);
4273 so_method(so
, tesla
, NV50TCL_VP_REG_ALLOC_RESULT
, 1);
4274 so_data (so
, p
->cfg
.high_result
);
4275 so_method(so
, tesla
, NV50TCL_VP_REG_ALLOC_TEMP
, 1);
4276 so_data (so
, p
->cfg
.high_temp
);
4277 so_method(so
, tesla
, NV50TCL_VP_START_ID
, 1);
4278 so_data (so
, 0); /* program start offset */
4279 so_ref(so
, &nv50
->state
.vertprog
);
4284 nv50_fragprog_validate(struct nv50_context
*nv50
)
4286 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
4287 struct nv50_program
*p
= nv50
->fragprog
;
4288 struct nouveau_stateobj
*so
;
4290 if (!p
->translated
) {
4291 nv50_program_validate(nv50
, p
);
4296 nv50_program_validate_data(nv50
, p
);
4297 nv50_program_validate_code(nv50
, p
);
4299 so
= so_new(6, 7, 2);
4300 so_method(so
, tesla
, NV50TCL_FP_ADDRESS_HIGH
, 2);
4301 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
4302 NOUVEAU_BO_HIGH
, 0, 0);
4303 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
4304 NOUVEAU_BO_LOW
, 0, 0);
4305 so_method(so
, tesla
, NV50TCL_FP_REG_ALLOC_TEMP
, 1);
4306 so_data (so
, p
->cfg
.high_temp
);
4307 so_method(so
, tesla
, NV50TCL_FP_RESULT_COUNT
, 1);
4308 so_data (so
, p
->cfg
.high_result
);
4309 so_method(so
, tesla
, NV50TCL_FP_CONTROL
, 1);
4310 so_data (so
, p
->cfg
.regs
[2]);
4311 so_method(so
, tesla
, NV50TCL_FP_CTRL_UNK196C
, 1);
4312 so_data (so
, p
->cfg
.regs
[3]);
4313 so_method(so
, tesla
, NV50TCL_FP_START_ID
, 1);
4314 so_data (so
, 0); /* program start offset */
4315 so_ref(so
, &nv50
->state
.fragprog
);
4320 nv50_geomprog_validate(struct nv50_context
*nv50
)
4322 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
4323 struct nv50_program
*p
= nv50
->geomprog
;
4324 struct nouveau_stateobj
*so
;
4326 if (!p
->translated
) {
4327 nv50_program_validate(nv50
, p
);
4332 nv50_program_validate_data(nv50
, p
);
4333 nv50_program_validate_code(nv50
, p
);
4335 so
= so_new(6, 7, 2);
4336 so_method(so
, tesla
, NV50TCL_GP_ADDRESS_HIGH
, 2);
4337 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
4338 NOUVEAU_BO_HIGH
, 0, 0);
4339 so_reloc (so
, p
->bo
, 0, NOUVEAU_BO_VRAM
| NOUVEAU_BO_RD
|
4340 NOUVEAU_BO_LOW
, 0, 0);
4341 so_method(so
, tesla
, NV50TCL_GP_REG_ALLOC_TEMP
, 1);
4342 so_data (so
, p
->cfg
.high_temp
);
4343 so_method(so
, tesla
, NV50TCL_GP_REG_ALLOC_RESULT
, 1);
4344 so_data (so
, p
->cfg
.high_result
);
4345 so_method(so
, tesla
, NV50TCL_GP_OUTPUT_PRIMITIVE_TYPE
, 1);
4346 so_data (so
, p
->cfg
.prim_type
);
4347 so_method(so
, tesla
, NV50TCL_GP_VERTEX_OUTPUT_COUNT
, 1);
4348 so_data (so
, p
->cfg
.vert_count
);
4349 so_method(so
, tesla
, NV50TCL_GP_START_ID
, 1);
4351 so_ref(so
, &nv50
->state
.geomprog
);
4356 nv50_pntc_replace(struct nv50_context
*nv50
, uint32_t pntc
[8], unsigned base
)
4358 struct nv50_program
*vp
;
4359 struct nv50_program
*fp
= nv50
->fragprog
;
4360 unsigned i
, c
, m
= base
;
4361 uint32_t origin
= 0x00000010;
4363 vp
= nv50
->geomprog
? nv50
->geomprog
: nv50
->vertprog
;
4365 /* XXX: this might not work correctly in all cases yet - we'll
4366 * just assume that an FP generic input that is not written in
4367 * the VP is PointCoord.
4369 memset(pntc
, 0, 8 * sizeof(uint32_t));
4371 for (i
= 0; i
< fp
->cfg
.in_nr
; i
++) {
4372 unsigned j
, n
= popcnt4(fp
->cfg
.in
[i
].mask
);
4374 if (fp
->cfg
.in
[i
].sn
!= TGSI_SEMANTIC_GENERIC
) {
4379 for (j
= 0; j
< vp
->cfg
.out_nr
; ++j
)
4380 if (vp
->cfg
.out
[j
].sn
== fp
->cfg
.in
[i
].sn
&&
4381 vp
->cfg
.out
[j
].si
== fp
->cfg
.in
[i
].si
)
4384 if (j
< vp
->cfg
.out_nr
) {
4385 ubyte mode
= nv50
->rasterizer
->pipe
.sprite_coord_mode
[
4388 if (mode
== PIPE_SPRITE_COORD_NONE
) {
4392 if (mode
== PIPE_SPRITE_COORD_LOWER_LEFT
)
4396 /* this is either PointCoord or replaced by sprite coords */
4397 for (c
= 0; c
< 4; c
++) {
4398 if (!(fp
->cfg
.in
[i
].mask
& (1 << c
)))
4400 pntc
[m
/ 8] |= (c
+ 1) << ((m
% 8) * 4);
4408 nv50_vec4_map(uint32_t *map32
, int mid
, uint8_t zval
, uint32_t lin
[4],
4409 struct nv50_sreg4
*fpi
, struct nv50_sreg4
*vpo
)
4412 uint8_t mv
= vpo
->mask
, mf
= fpi
->mask
, oid
= vpo
->hw
;
4413 uint8_t *map
= (uint8_t *)map32
;
4415 for (c
= 0; c
< 4; ++c
) {
4417 if (fpi
->linear
== TRUE
)
4418 lin
[mid
/ 32] |= 1 << (mid
% 32);
4422 map
[mid
] = (c
== 3) ? (zval
+ 1) : zval
;
4435 nv50_fp_linkage_validate(struct nv50_context
*nv50
)
4437 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
4438 struct nv50_program
*vp
= nv50
->vertprog
;
4439 struct nv50_program
*fp
= nv50
->fragprog
;
4440 struct nouveau_stateobj
*so
;
4441 struct nv50_sreg4 dummy
;
4443 uint32_t map
[16], lin
[4], reg
[6], pcrd
[8];
4444 uint8_t zval
= 0x40;
4446 if (nv50
->geomprog
) {
4447 vp
= nv50
->geomprog
;
4450 memset(map
, 0, sizeof(map
));
4451 memset(lin
, 0, sizeof(lin
));
4453 reg
[1] = 0x00000004; /* low and high clip distance map ids */
4454 reg
[2] = 0x00000000; /* layer index map id (disabled, GP only) */
4455 reg
[3] = 0x00000000; /* point size map id & enable */
4456 reg
[5] = 0x00000000; /* primitive ID map slot */
4457 reg
[0] = fp
->cfg
.regs
[0]; /* colour semantic reg */
4458 reg
[4] = fp
->cfg
.regs
[1]; /* interpolant info */
4460 dummy
.linear
= FALSE
;
4461 dummy
.mask
= 0xf; /* map all components of HPOS */
4462 m
= nv50_vec4_map(map
, m
, zval
, lin
, &dummy
, &vp
->cfg
.out
[0]);
4466 if (vp
->cfg
.clpd
< 0x40) {
4467 for (c
= 0; c
< vp
->cfg
.clpd_nr
; ++c
) {
4468 map
[m
/ 4] |= (vp
->cfg
.clpd
+ c
) << ((m
% 4) * 8);
4474 reg
[0] |= m
<< 8; /* adjust BFC0 id */
4476 /* if light_twoside is active, it seems FFC0_ID == BFC0_ID is bad */
4477 if (nv50
->rasterizer
->pipe
.light_twoside
) {
4478 struct nv50_sreg4
*vpo
= &vp
->cfg
.two_side
[0];
4479 struct nv50_sreg4
*fpi
= &fp
->cfg
.two_side
[0];
4481 m
= nv50_vec4_map(map
, m
, zval
, lin
, &fpi
[0], &vpo
[0]);
4482 m
= nv50_vec4_map(map
, m
, zval
, lin
, &fpi
[1], &vpo
[1]);
4485 reg
[0] += m
- 4; /* adjust FFC0 id */
4486 reg
[4] |= m
<< 8; /* set mid where 'normal' FP inputs start */
4488 for (i
= 0; i
< fp
->cfg
.in_nr
; i
++) {
4489 /* maybe even remove these from cfg.io */
4490 if (fp
->cfg
.in
[i
].sn
== TGSI_SEMANTIC_POSITION
||
4491 fp
->cfg
.in
[i
].sn
== TGSI_SEMANTIC_FACE
)
4494 for (n
= 0; n
< vp
->cfg
.out_nr
; ++n
)
4495 if (vp
->cfg
.out
[n
].sn
== fp
->cfg
.in
[i
].sn
&&
4496 vp
->cfg
.out
[n
].si
== fp
->cfg
.in
[i
].si
)
4499 m
= nv50_vec4_map(map
, m
, zval
, lin
, &fp
->cfg
.in
[i
],
4500 (n
< vp
->cfg
.out_nr
) ?
4501 &vp
->cfg
.out
[n
] : &dummy
);
4503 /* PrimitiveID either is replaced by the system value, or
4504 * written by the geometry shader into an output register
4506 if (fp
->cfg
.prim_id
< 0x40) {
4507 map
[m
/ 4] |= vp
->cfg
.prim_id
<< ((m
% 4) * 8);
4511 if (nv50
->rasterizer
->pipe
.point_size_per_vertex
) {
4512 map
[m
/ 4] |= vp
->cfg
.psiz
<< ((m
% 4) * 8);
4513 reg
[3] = (m
++ << 4) | 1;
4516 /* now fill the stateobj (at most 28 so_data) */
4517 so
= so_new(10, 54, 0);
4521 if (vp
->type
== PIPE_SHADER_GEOMETRY
) {
4522 so_method(so
, tesla
, NV50TCL_GP_RESULT_MAP_SIZE
, 1);
4524 so_method(so
, tesla
, NV50TCL_GP_RESULT_MAP(0), n
);
4525 so_datap (so
, map
, n
);
4527 so_method(so
, tesla
, NV50TCL_VP_GP_BUILTIN_ATTR_EN
, 1);
4528 so_data (so
, vp
->cfg
.regs
[0]);
4530 so_method(so
, tesla
, NV50TCL_MAP_SEMANTIC_4
, 1);
4531 so_data (so
, reg
[5]);
4533 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP_SIZE
, 1);
4535 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP(0), n
);
4536 so_datap (so
, map
, n
);
4539 so_method(so
, tesla
, NV50TCL_MAP_SEMANTIC_0
, 4);
4540 so_datap (so
, reg
, 4);
4542 so_method(so
, tesla
, NV50TCL_FP_INTERPOLANT_CTRL
, 1);
4543 so_data (so
, reg
[4]);
4545 so_method(so
, tesla
, NV50TCL_NOPERSPECTIVE_BITMAP(0), 4);
4546 so_datap (so
, lin
, 4);
4548 if (nv50
->rasterizer
->pipe
.point_sprite
) {
4549 so_method(so
, tesla
, NV50TCL_POINT_SPRITE_CTRL
, 1);
4551 nv50_pntc_replace(nv50
, pcrd
, (reg
[4] >> 8) & 0xff));
4553 so_method(so
, tesla
, NV50TCL_POINT_COORD_REPLACE_MAP(0), 8);
4554 so_datap (so
, pcrd
, 8);
4557 so_method(so
, tesla
, NV50TCL_GP_ENABLE
, 1);
4558 so_data (so
, (vp
->type
== PIPE_SHADER_GEOMETRY
) ? 1 : 0);
4560 so_ref(so
, &nv50
->state
.fp_linkage
);
4565 construct_vp_gp_mapping(uint32_t *map32
, int m
,
4566 struct nv50_program
*vp
, struct nv50_program
*gp
)
4568 uint8_t *map
= (uint8_t *)map32
;
4571 for (i
= 0; i
< gp
->cfg
.in_nr
; ++i
) {
4572 uint8_t oid
, mv
= 0, mg
= gp
->cfg
.in
[i
].mask
;
4574 for (j
= 0; j
< vp
->cfg
.out_nr
; ++j
) {
4575 if (vp
->cfg
.out
[j
].sn
== gp
->cfg
.in
[i
].sn
&&
4576 vp
->cfg
.out
[j
].si
== gp
->cfg
.in
[i
].si
) {
4577 mv
= vp
->cfg
.out
[j
].mask
;
4578 oid
= vp
->cfg
.out
[j
].hw
;
4583 for (c
= 0; c
< 4; ++c
, mv
>>= 1, mg
>>= 1) {
4588 map
[m
++] = (c
== 3) ? 0x41 : 0x40;
4596 nv50_gp_linkage_validate(struct nv50_context
*nv50
)
4598 struct nouveau_grobj
*tesla
= nv50
->screen
->tesla
;
4599 struct nouveau_stateobj
*so
;
4600 struct nv50_program
*vp
= nv50
->vertprog
;
4601 struct nv50_program
*gp
= nv50
->geomprog
;
4606 so_ref(NULL
, &nv50
->state
.gp_linkage
);
4609 memset(map
, 0, sizeof(map
));
4611 m
= construct_vp_gp_mapping(map
, m
, vp
, gp
);
4613 so
= so_new(3, 24 - 3, 0);
4615 so_method(so
, tesla
, NV50TCL_VP_GP_BUILTIN_ATTR_EN
, 1);
4616 so_data (so
, vp
->cfg
.regs
[0] | gp
->cfg
.regs
[0]);
4619 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP_SIZE
, 1);
4623 so_method(so
, tesla
, NV50TCL_VP_RESULT_MAP(0), m
);
4624 so_datap (so
, map
, m
);
4626 so_ref(so
, &nv50
->state
.gp_linkage
);
4631 nv50_program_destroy(struct nv50_context
*nv50
, struct nv50_program
*p
)
4633 while (p
->exec_head
) {
4634 struct nv50_program_exec
*e
= p
->exec_head
;
4636 p
->exec_head
= e
->next
;
4639 p
->exec_tail
= NULL
;
4642 nouveau_bo_ref(NULL
, &p
->bo
);
4644 nouveau_resource_free(&p
->data
[0]);