2 * Copyright © 2016 Broadcom
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 (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
25 #include "util/format/u_format.h"
26 #include "util/u_helpers.h"
27 #include "util/u_math.h"
28 #include "util/u_memory.h"
29 #include "util/ralloc.h"
30 #include "util/hash_table.h"
31 #include "compiler/nir/nir.h"
32 #include "compiler/nir/nir_builder.h"
33 #include "common/v3d_device_info.h"
34 #include "v3d_compiler.h"
36 /* We don't do any address packing. */
37 #define __gen_user_data void
38 #define __gen_address_type uint32_t
39 #define __gen_address_offset(reloc) (*reloc)
40 #define __gen_emit_reloc(cl, reloc)
41 #include "cle/v3d_packet_v41_pack.h"
43 #define GENERAL_TMU_LOOKUP_PER_QUAD (0 << 7)
44 #define GENERAL_TMU_LOOKUP_PER_PIXEL (1 << 7)
45 #define GENERAL_TMU_LOOKUP_TYPE_8BIT_I (0 << 0)
46 #define GENERAL_TMU_LOOKUP_TYPE_16BIT_I (1 << 0)
47 #define GENERAL_TMU_LOOKUP_TYPE_VEC2 (2 << 0)
48 #define GENERAL_TMU_LOOKUP_TYPE_VEC3 (3 << 0)
49 #define GENERAL_TMU_LOOKUP_TYPE_VEC4 (4 << 0)
50 #define GENERAL_TMU_LOOKUP_TYPE_8BIT_UI (5 << 0)
51 #define GENERAL_TMU_LOOKUP_TYPE_16BIT_UI (6 << 0)
52 #define GENERAL_TMU_LOOKUP_TYPE_32BIT_UI (7 << 0)
54 #define V3D_TSY_SET_QUORUM 0
55 #define V3D_TSY_INC_WAITERS 1
56 #define V3D_TSY_DEC_WAITERS 2
57 #define V3D_TSY_INC_QUORUM 3
58 #define V3D_TSY_DEC_QUORUM 4
59 #define V3D_TSY_FREE_ALL 5
60 #define V3D_TSY_RELEASE 6
61 #define V3D_TSY_ACQUIRE 7
62 #define V3D_TSY_WAIT 8
63 #define V3D_TSY_WAIT_INC 9
64 #define V3D_TSY_WAIT_CHECK 10
65 #define V3D_TSY_WAIT_INC_CHECK 11
66 #define V3D_TSY_WAIT_CV 12
67 #define V3D_TSY_INC_SEMAPHORE 13
68 #define V3D_TSY_DEC_SEMAPHORE 14
69 #define V3D_TSY_SET_QUORUM_FREE_ALL 15
72 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
75 resize_qreg_array(struct v3d_compile
*c
,
80 if (*size
>= decl_size
)
83 uint32_t old_size
= *size
;
84 *size
= MAX2(*size
* 2, decl_size
);
85 *regs
= reralloc(c
, *regs
, struct qreg
, *size
);
87 fprintf(stderr
, "Malloc failure\n");
91 for (uint32_t i
= old_size
; i
< *size
; i
++)
92 (*regs
)[i
] = c
->undef
;
96 vir_emit_thrsw(struct v3d_compile
*c
)
101 /* Always thread switch after each texture operation for now.
103 * We could do better by batching a bunch of texture fetches up and
104 * then doing one thread switch and collecting all their results
107 c
->last_thrsw
= vir_NOP(c
);
108 c
->last_thrsw
->qpu
.sig
.thrsw
= true;
109 c
->last_thrsw_at_top_level
= !c
->in_control_flow
;
111 /* We need to lock the scoreboard before any tlb acess happens. If this
112 * thread switch comes after we have emitted a tlb load, then it means
113 * that we can't lock on the last thread switch any more.
115 if (c
->emitted_tlb_load
)
116 c
->lock_scoreboard_on_first_thrsw
= true;
120 v3d_get_op_for_atomic_add(nir_intrinsic_instr
*instr
, unsigned src
)
122 if (nir_src_is_const(instr
->src
[src
])) {
123 int64_t add_val
= nir_src_as_int(instr
->src
[src
]);
125 return V3D_TMU_OP_WRITE_AND_READ_INC
;
126 else if (add_val
== -1)
127 return V3D_TMU_OP_WRITE_OR_READ_DEC
;
130 return V3D_TMU_OP_WRITE_ADD_READ_PREFETCH
;
134 v3d_general_tmu_op(nir_intrinsic_instr
*instr
)
136 switch (instr
->intrinsic
) {
137 case nir_intrinsic_load_ssbo
:
138 case nir_intrinsic_load_ubo
:
139 case nir_intrinsic_load_uniform
:
140 case nir_intrinsic_load_shared
:
141 case nir_intrinsic_load_scratch
:
142 case nir_intrinsic_store_ssbo
:
143 case nir_intrinsic_store_shared
:
144 case nir_intrinsic_store_scratch
:
145 return V3D_TMU_OP_REGULAR
;
146 case nir_intrinsic_ssbo_atomic_add
:
147 return v3d_get_op_for_atomic_add(instr
, 2);
148 case nir_intrinsic_shared_atomic_add
:
149 return v3d_get_op_for_atomic_add(instr
, 1);
150 case nir_intrinsic_ssbo_atomic_imin
:
151 case nir_intrinsic_shared_atomic_imin
:
152 return V3D_TMU_OP_WRITE_SMIN
;
153 case nir_intrinsic_ssbo_atomic_umin
:
154 case nir_intrinsic_shared_atomic_umin
:
155 return V3D_TMU_OP_WRITE_UMIN_FULL_L1_CLEAR
;
156 case nir_intrinsic_ssbo_atomic_imax
:
157 case nir_intrinsic_shared_atomic_imax
:
158 return V3D_TMU_OP_WRITE_SMAX
;
159 case nir_intrinsic_ssbo_atomic_umax
:
160 case nir_intrinsic_shared_atomic_umax
:
161 return V3D_TMU_OP_WRITE_UMAX
;
162 case nir_intrinsic_ssbo_atomic_and
:
163 case nir_intrinsic_shared_atomic_and
:
164 return V3D_TMU_OP_WRITE_AND_READ_INC
;
165 case nir_intrinsic_ssbo_atomic_or
:
166 case nir_intrinsic_shared_atomic_or
:
167 return V3D_TMU_OP_WRITE_OR_READ_DEC
;
168 case nir_intrinsic_ssbo_atomic_xor
:
169 case nir_intrinsic_shared_atomic_xor
:
170 return V3D_TMU_OP_WRITE_XOR_READ_NOT
;
171 case nir_intrinsic_ssbo_atomic_exchange
:
172 case nir_intrinsic_shared_atomic_exchange
:
173 return V3D_TMU_OP_WRITE_XCHG_READ_FLUSH
;
174 case nir_intrinsic_ssbo_atomic_comp_swap
:
175 case nir_intrinsic_shared_atomic_comp_swap
:
176 return V3D_TMU_OP_WRITE_CMPXCHG_READ_FLUSH
;
178 unreachable("unknown intrinsic op");
183 * Implements indirect uniform loads and SSBO accesses through the TMU general
184 * memory access interface.
187 ntq_emit_tmu_general(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
,
188 bool is_shared_or_scratch
)
190 uint32_t tmu_op
= v3d_general_tmu_op(instr
);
192 /* If we were able to replace atomic_add for an inc/dec, then we
193 * need/can to do things slightly different, like not loading the
194 * amount to add/sub, as that is implicit.
196 bool atomic_add_replaced
=
197 ((instr
->intrinsic
== nir_intrinsic_ssbo_atomic_add
||
198 instr
->intrinsic
== nir_intrinsic_shared_atomic_add
) &&
199 (tmu_op
== V3D_TMU_OP_WRITE_AND_READ_INC
||
200 tmu_op
== V3D_TMU_OP_WRITE_OR_READ_DEC
));
202 bool is_store
= (instr
->intrinsic
== nir_intrinsic_store_ssbo
||
203 instr
->intrinsic
== nir_intrinsic_store_scratch
||
204 instr
->intrinsic
== nir_intrinsic_store_shared
);
206 bool is_load
= (instr
->intrinsic
== nir_intrinsic_load_uniform
||
207 instr
->intrinsic
== nir_intrinsic_load_ubo
||
208 instr
->intrinsic
== nir_intrinsic_load_ssbo
||
209 instr
->intrinsic
== nir_intrinsic_load_scratch
||
210 instr
->intrinsic
== nir_intrinsic_load_shared
);
213 c
->tmu_dirty_rcl
= true;
215 bool has_index
= !is_shared_or_scratch
;
218 if (instr
->intrinsic
== nir_intrinsic_load_uniform
) {
220 } else if (instr
->intrinsic
== nir_intrinsic_load_ssbo
||
221 instr
->intrinsic
== nir_intrinsic_load_ubo
||
222 instr
->intrinsic
== nir_intrinsic_load_scratch
||
223 instr
->intrinsic
== nir_intrinsic_load_shared
||
224 atomic_add_replaced
) {
225 offset_src
= 0 + has_index
;
226 } else if (is_store
) {
227 offset_src
= 1 + has_index
;
229 offset_src
= 0 + has_index
;
232 bool dynamic_src
= !nir_src_is_const(instr
->src
[offset_src
]);
233 uint32_t const_offset
= 0;
235 const_offset
= nir_src_as_uint(instr
->src
[offset_src
]);
237 struct qreg base_offset
;
238 if (instr
->intrinsic
== nir_intrinsic_load_uniform
) {
239 const_offset
+= nir_intrinsic_base(instr
);
240 base_offset
= vir_uniform(c
, QUNIFORM_UBO_ADDR
,
241 v3d_unit_data_create(0, const_offset
));
243 } else if (instr
->intrinsic
== nir_intrinsic_load_ubo
) {
244 uint32_t index
= nir_src_as_uint(instr
->src
[0]) + 1;
245 /* Note that QUNIFORM_UBO_ADDR takes a UBO index shifted up by
246 * 1 (0 is gallium's constant buffer 0).
249 vir_uniform(c
, QUNIFORM_UBO_ADDR
,
250 v3d_unit_data_create(index
, const_offset
));
252 } else if (is_shared_or_scratch
) {
253 /* Shared and scratch variables have no buffer index, and all
254 * start from a common base that we set up at the start of
257 if (instr
->intrinsic
== nir_intrinsic_load_scratch
||
258 instr
->intrinsic
== nir_intrinsic_store_scratch
) {
259 base_offset
= c
->spill_base
;
261 base_offset
= c
->cs_shared_offset
;
262 const_offset
+= nir_intrinsic_base(instr
);
265 base_offset
= vir_uniform(c
, QUNIFORM_SSBO_OFFSET
,
266 nir_src_as_uint(instr
->src
[is_store
?
270 struct qreg tmud
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUD
);
271 unsigned writemask
= is_store
? nir_intrinsic_write_mask(instr
) : 0;
272 uint32_t base_const_offset
= const_offset
;
273 int first_component
= -1;
274 int last_component
= -1;
276 int tmu_writes
= 1; /* address */
279 /* Find the first set of consecutive components that
280 * are enabled in the writemask and emit the TMUD
281 * instructions for them.
283 first_component
= ffs(writemask
) - 1;
284 last_component
= first_component
;
285 while (writemask
& BITFIELD_BIT(last_component
+ 1))
288 assert(first_component
>= 0 &&
289 first_component
<= last_component
&&
290 last_component
< instr
->num_components
);
292 struct qreg tmud
= vir_reg(QFILE_MAGIC
,
294 for (int i
= first_component
; i
<= last_component
; i
++) {
296 ntq_get_src(c
, instr
->src
[0], i
);
297 vir_MOV_dest(c
, tmud
, data
);
301 /* Update the offset for the TMU write based on the
302 * the first component we are writing.
304 const_offset
= base_const_offset
+ first_component
* 4;
306 /* Clear these components from the writemask */
307 uint32_t written_mask
=
308 BITFIELD_RANGE(first_component
, tmu_writes
- 1);
309 writemask
&= ~written_mask
;
310 } else if (!is_load
&& !atomic_add_replaced
) {
312 ntq_get_src(c
, instr
->src
[1 + has_index
], 0);
313 vir_MOV_dest(c
, tmud
, data
);
315 if (tmu_op
== V3D_TMU_OP_WRITE_CMPXCHG_READ_FLUSH
) {
316 data
= ntq_get_src(c
, instr
->src
[2 + has_index
],
318 vir_MOV_dest(c
, tmud
, data
);
323 /* Make sure we won't exceed the 16-entry TMU fifo if each
324 * thread is storing at the same time.
326 while (tmu_writes
> 16 / c
->threads
)
329 /* The spec says that for atomics, the TYPE field is ignored,
330 * but that doesn't seem to be the case for CMPXCHG. Just use
331 * the number of tmud writes we did to decide the type (or
332 * choose "32bit" for atomic reads, which has been fine).
334 uint32_t num_components
;
335 if (is_load
|| atomic_add_replaced
) {
336 num_components
= instr
->num_components
;
338 assert(tmu_writes
> 1);
339 num_components
= tmu_writes
- 1;
342 uint32_t perquad
= is_load
343 ? GENERAL_TMU_LOOKUP_PER_QUAD
344 : GENERAL_TMU_LOOKUP_PER_PIXEL
;
345 uint32_t config
= (0xffffff00 |
348 if (num_components
== 1) {
349 config
|= GENERAL_TMU_LOOKUP_TYPE_32BIT_UI
;
351 config
|= GENERAL_TMU_LOOKUP_TYPE_VEC2
+
355 if (vir_in_nonuniform_control_flow(c
)) {
356 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
362 tmua
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUA
);
364 tmua
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUAU
);
368 struct qreg offset
= base_offset
;
369 if (const_offset
!= 0) {
370 offset
= vir_ADD(c
, offset
,
371 vir_uniform_ui(c
, const_offset
));
374 ntq_get_src(c
, instr
->src
[offset_src
], 0);
375 tmu
= vir_ADD_dest(c
, tmua
, offset
, data
);
377 if (const_offset
!= 0) {
378 tmu
= vir_ADD_dest(c
, tmua
, base_offset
,
379 vir_uniform_ui(c
, const_offset
));
381 tmu
= vir_MOV_dest(c
, tmua
, base_offset
);
387 vir_get_uniform_index(c
, QUNIFORM_CONSTANT
,
391 if (vir_in_nonuniform_control_flow(c
))
392 vir_set_cond(tmu
, V3D_QPU_COND_IFA
);
396 /* Read the result, or wait for the TMU op to complete. */
397 for (int i
= 0; i
< nir_intrinsic_dest_components(instr
); i
++) {
398 ntq_store_dest(c
, &instr
->dest
, i
,
399 vir_MOV(c
, vir_LDTMU(c
)));
402 if (nir_intrinsic_dest_components(instr
) == 0)
404 } while (is_store
&& writemask
!= 0);
408 ntq_init_ssa_def(struct v3d_compile
*c
, nir_ssa_def
*def
)
410 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
411 def
->num_components
);
412 _mesa_hash_table_insert(c
->def_ht
, def
, qregs
);
417 is_ld_signal(const struct v3d_qpu_sig
*sig
)
419 return (sig
->ldunif
||
431 * This function is responsible for getting VIR results into the associated
432 * storage for a NIR instruction.
434 * If it's a NIR SSA def, then we just set the associated hash table entry to
437 * If it's a NIR reg, then we need to update the existing qreg assigned to the
438 * NIR destination with the incoming value. To do that without introducing
439 * new MOVs, we require that the incoming qreg either be a uniform, or be
440 * SSA-defined by the previous VIR instruction in the block and rewritable by
441 * this function. That lets us sneak ahead and insert the SF flag beforehand
442 * (knowing that the previous instruction doesn't depend on flags) and rewrite
443 * its destination to be the NIR reg's destination
446 ntq_store_dest(struct v3d_compile
*c
, nir_dest
*dest
, int chan
,
449 struct qinst
*last_inst
= NULL
;
450 if (!list_is_empty(&c
->cur_block
->instructions
))
451 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
453 assert((result
.file
== QFILE_TEMP
&&
454 last_inst
&& last_inst
== c
->defs
[result
.index
]));
457 assert(chan
< dest
->ssa
.num_components
);
460 struct hash_entry
*entry
=
461 _mesa_hash_table_search(c
->def_ht
, &dest
->ssa
);
466 qregs
= ntq_init_ssa_def(c
, &dest
->ssa
);
468 qregs
[chan
] = result
;
470 nir_register
*reg
= dest
->reg
.reg
;
471 assert(dest
->reg
.base_offset
== 0);
472 assert(reg
->num_array_elems
== 0);
473 struct hash_entry
*entry
=
474 _mesa_hash_table_search(c
->def_ht
, reg
);
475 struct qreg
*qregs
= entry
->data
;
477 /* If the previous instruction can't be predicated for
478 * the store into the nir_register, then emit a MOV
481 if (vir_in_nonuniform_control_flow(c
) &&
482 is_ld_signal(&c
->defs
[last_inst
->dst
.index
]->qpu
.sig
)) {
483 result
= vir_MOV(c
, result
);
484 last_inst
= c
->defs
[result
.index
];
487 /* We know they're both temps, so just rewrite index. */
488 c
->defs
[last_inst
->dst
.index
] = NULL
;
489 last_inst
->dst
.index
= qregs
[chan
].index
;
491 /* If we're in control flow, then make this update of the reg
492 * conditional on the execution mask.
494 if (vir_in_nonuniform_control_flow(c
)) {
495 last_inst
->dst
.index
= qregs
[chan
].index
;
497 /* Set the flags to the current exec mask.
499 c
->cursor
= vir_before_inst(last_inst
);
500 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
502 c
->cursor
= vir_after_inst(last_inst
);
504 vir_set_cond(last_inst
, V3D_QPU_COND_IFA
);
510 ntq_get_src(struct v3d_compile
*c
, nir_src src
, int i
)
512 struct hash_entry
*entry
;
514 entry
= _mesa_hash_table_search(c
->def_ht
, src
.ssa
);
515 assert(i
< src
.ssa
->num_components
);
517 nir_register
*reg
= src
.reg
.reg
;
518 entry
= _mesa_hash_table_search(c
->def_ht
, reg
);
519 assert(reg
->num_array_elems
== 0);
520 assert(src
.reg
.base_offset
== 0);
521 assert(i
< reg
->num_components
);
524 struct qreg
*qregs
= entry
->data
;
529 ntq_get_alu_src(struct v3d_compile
*c
, nir_alu_instr
*instr
,
532 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
533 unsigned chan
= ffs(instr
->dest
.write_mask
) - 1;
534 struct qreg r
= ntq_get_src(c
, instr
->src
[src
].src
,
535 instr
->src
[src
].swizzle
[chan
]);
537 assert(!instr
->src
[src
].abs
);
538 assert(!instr
->src
[src
].negate
);
544 ntq_minify(struct v3d_compile
*c
, struct qreg size
, struct qreg level
)
546 return vir_MAX(c
, vir_SHR(c
, size
, level
), vir_uniform_ui(c
, 1));
550 ntq_emit_txs(struct v3d_compile
*c
, nir_tex_instr
*instr
)
552 unsigned unit
= instr
->texture_index
;
553 int lod_index
= nir_tex_instr_src_index(instr
, nir_tex_src_lod
);
554 int dest_size
= nir_tex_instr_dest_size(instr
);
556 struct qreg lod
= c
->undef
;
558 lod
= ntq_get_src(c
, instr
->src
[lod_index
].src
, 0);
560 for (int i
= 0; i
< dest_size
; i
++) {
562 enum quniform_contents contents
;
564 if (instr
->is_array
&& i
== dest_size
- 1)
565 contents
= QUNIFORM_TEXTURE_ARRAY_SIZE
;
567 contents
= QUNIFORM_TEXTURE_WIDTH
+ i
;
569 struct qreg size
= vir_uniform(c
, contents
, unit
);
571 switch (instr
->sampler_dim
) {
572 case GLSL_SAMPLER_DIM_1D
:
573 case GLSL_SAMPLER_DIM_2D
:
574 case GLSL_SAMPLER_DIM_MS
:
575 case GLSL_SAMPLER_DIM_3D
:
576 case GLSL_SAMPLER_DIM_CUBE
:
577 /* Don't minify the array size. */
578 if (!(instr
->is_array
&& i
== dest_size
- 1)) {
579 size
= ntq_minify(c
, size
, lod
);
583 case GLSL_SAMPLER_DIM_RECT
:
584 /* There's no LOD field for rects */
588 unreachable("Bad sampler type");
591 ntq_store_dest(c
, &instr
->dest
, i
, size
);
596 ntq_emit_tex(struct v3d_compile
*c
, nir_tex_instr
*instr
)
598 unsigned unit
= instr
->texture_index
;
600 /* Since each texture sampling op requires uploading uniforms to
601 * reference the texture, there's no HW support for texture size and
602 * you just upload uniforms containing the size.
605 case nir_texop_query_levels
:
606 ntq_store_dest(c
, &instr
->dest
, 0,
607 vir_uniform(c
, QUNIFORM_TEXTURE_LEVELS
, unit
));
610 ntq_emit_txs(c
, instr
);
616 if (c
->devinfo
->ver
>= 40)
617 v3d40_vir_emit_tex(c
, instr
);
619 v3d33_vir_emit_tex(c
, instr
);
623 ntq_fsincos(struct v3d_compile
*c
, struct qreg src
, bool is_cos
)
625 struct qreg input
= vir_FMUL(c
, src
, vir_uniform_f(c
, 1.0f
/ M_PI
));
627 input
= vir_FADD(c
, input
, vir_uniform_f(c
, 0.5));
629 struct qreg periods
= vir_FROUND(c
, input
);
630 struct qreg sin_output
= vir_SIN(c
, vir_FSUB(c
, input
, periods
));
631 return vir_XOR(c
, sin_output
, vir_SHL(c
,
632 vir_FTOIN(c
, periods
),
633 vir_uniform_ui(c
, -1)));
637 ntq_fsign(struct v3d_compile
*c
, struct qreg src
)
639 struct qreg t
= vir_get_temp(c
);
641 vir_MOV_dest(c
, t
, vir_uniform_f(c
, 0.0));
642 vir_set_pf(vir_FMOV_dest(c
, vir_nop_reg(), src
), V3D_QPU_PF_PUSHZ
);
643 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_f(c
, 1.0));
644 vir_set_pf(vir_FMOV_dest(c
, vir_nop_reg(), src
), V3D_QPU_PF_PUSHN
);
645 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_f(c
, -1.0));
646 return vir_MOV(c
, t
);
650 emit_fragcoord_input(struct v3d_compile
*c
, int attr
)
652 c
->inputs
[attr
* 4 + 0] = vir_FXCD(c
);
653 c
->inputs
[attr
* 4 + 1] = vir_FYCD(c
);
654 c
->inputs
[attr
* 4 + 2] = c
->payload_z
;
655 c
->inputs
[attr
* 4 + 3] = vir_RECIP(c
, c
->payload_w
);
659 emit_fragment_varying(struct v3d_compile
*c
, nir_variable
*var
,
660 uint8_t swizzle
, int array_index
)
662 struct qreg r3
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R3
);
663 struct qreg r5
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R5
);
666 if (c
->devinfo
->ver
>= 41) {
667 struct qinst
*ldvary
= vir_add_inst(V3D_QPU_A_NOP
, c
->undef
,
669 ldvary
->qpu
.sig
.ldvary
= true;
670 vary
= vir_emit_def(c
, ldvary
);
672 vir_NOP(c
)->qpu
.sig
.ldvary
= true;
676 /* For gl_PointCoord input or distance along a line, we'll be called
677 * with no nir_variable, and we don't count toward VPM size so we
678 * don't track an input slot.
681 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
684 int i
= c
->num_inputs
++;
686 v3d_slot_from_slot_and_component(var
->data
.location
+
687 array_index
, swizzle
);
689 switch (var
->data
.interpolation
) {
690 case INTERP_MODE_NONE
:
691 /* If a gl_FrontColor or gl_BackColor input has no interp
692 * qualifier, then if we're using glShadeModel(GL_FLAT) it
693 * needs to be flat shaded.
695 switch (var
->data
.location
+ array_index
) {
696 case VARYING_SLOT_COL0
:
697 case VARYING_SLOT_COL1
:
698 case VARYING_SLOT_BFC0
:
699 case VARYING_SLOT_BFC1
:
700 if (c
->fs_key
->shade_model_flat
) {
701 BITSET_SET(c
->flat_shade_flags
, i
);
702 vir_MOV_dest(c
, c
->undef
, vary
);
703 return vir_MOV(c
, r5
);
705 return vir_FADD(c
, vir_FMUL(c
, vary
,
712 case INTERP_MODE_SMOOTH
:
713 if (var
->data
.centroid
) {
714 BITSET_SET(c
->centroid_flags
, i
);
715 return vir_FADD(c
, vir_FMUL(c
, vary
,
716 c
->payload_w_centroid
), r5
);
718 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
720 case INTERP_MODE_NOPERSPECTIVE
:
721 BITSET_SET(c
->noperspective_flags
, i
);
722 return vir_FADD(c
, vir_MOV(c
, vary
), r5
);
723 case INTERP_MODE_FLAT
:
724 BITSET_SET(c
->flat_shade_flags
, i
);
725 vir_MOV_dest(c
, c
->undef
, vary
);
726 return vir_MOV(c
, r5
);
728 unreachable("Bad interp mode");
733 emit_fragment_input(struct v3d_compile
*c
, int attr
, nir_variable
*var
,
736 for (int i
= 0; i
< glsl_get_vector_elements(var
->type
); i
++) {
737 int chan
= var
->data
.location_frac
+ i
;
738 c
->inputs
[attr
* 4 + chan
] =
739 emit_fragment_varying(c
, var
, chan
, array_index
);
744 emit_compact_fragment_input(struct v3d_compile
*c
, int attr
, nir_variable
*var
,
747 /* Compact variables are scalar arrays where each set of 4 elements
748 * consumes a single location.
750 int loc_offset
= array_index
/ 4;
751 int chan
= var
->data
.location_frac
+ array_index
% 4;
752 c
->inputs
[(attr
+ loc_offset
) * 4 + chan
] =
753 emit_fragment_varying(c
, var
, chan
, loc_offset
);
757 add_output(struct v3d_compile
*c
,
758 uint32_t decl_offset
,
762 uint32_t old_array_size
= c
->outputs_array_size
;
763 resize_qreg_array(c
, &c
->outputs
, &c
->outputs_array_size
,
766 if (old_array_size
!= c
->outputs_array_size
) {
767 c
->output_slots
= reralloc(c
,
769 struct v3d_varying_slot
,
770 c
->outputs_array_size
);
773 c
->output_slots
[decl_offset
] =
774 v3d_slot_from_slot_and_component(slot
, swizzle
);
778 * If compare_instr is a valid comparison instruction, emits the
779 * compare_instr's comparison and returns the sel_instr's return value based
780 * on the compare_instr's result.
783 ntq_emit_comparison(struct v3d_compile
*c
,
784 nir_alu_instr
*compare_instr
,
785 enum v3d_qpu_cond
*out_cond
)
787 struct qreg src0
= ntq_get_alu_src(c
, compare_instr
, 0);
789 if (nir_op_infos
[compare_instr
->op
].num_inputs
> 1)
790 src1
= ntq_get_alu_src(c
, compare_instr
, 1);
791 bool cond_invert
= false;
792 struct qreg nop
= vir_nop_reg();
794 switch (compare_instr
->op
) {
797 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
800 vir_set_pf(vir_XOR_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
805 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
809 vir_set_pf(vir_XOR_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
815 vir_set_pf(vir_FCMP_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
818 vir_set_pf(vir_MIN_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
822 vir_set_pf(vir_SUB_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHC
);
828 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHN
);
831 vir_set_pf(vir_MIN_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
834 vir_set_pf(vir_SUB_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHC
);
838 vir_set_pf(vir_MOV_dest(c
, nop
, src0
), V3D_QPU_PF_PUSHZ
);
843 vir_set_pf(vir_FMOV_dest(c
, nop
, src0
), V3D_QPU_PF_PUSHZ
);
851 *out_cond
= cond_invert
? V3D_QPU_COND_IFNA
: V3D_QPU_COND_IFA
;
856 /* Finds an ALU instruction that generates our src value that could
857 * (potentially) be greedily emitted in the consuming instruction.
859 static struct nir_alu_instr
*
860 ntq_get_alu_parent(nir_src src
)
862 if (!src
.is_ssa
|| src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
864 nir_alu_instr
*instr
= nir_instr_as_alu(src
.ssa
->parent_instr
);
868 /* If the ALU instr's srcs are non-SSA, then we would have to avoid
869 * moving emission of the ALU instr down past another write of the
872 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
873 if (!instr
->src
[i
].src
.is_ssa
)
880 /* Turns a NIR bool into a condition code to predicate on. */
881 static enum v3d_qpu_cond
882 ntq_emit_bool_to_cond(struct v3d_compile
*c
, nir_src src
)
884 nir_alu_instr
*compare
= ntq_get_alu_parent(src
);
888 enum v3d_qpu_cond cond
;
889 if (ntq_emit_comparison(c
, compare
, &cond
))
893 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), ntq_get_src(c
, src
, 0)),
895 return V3D_QPU_COND_IFNA
;
899 ntq_emit_alu(struct v3d_compile
*c
, nir_alu_instr
*instr
)
901 /* This should always be lowered to ALU operations for V3D. */
902 assert(!instr
->dest
.saturate
);
904 /* Vectors are special in that they have non-scalarized writemasks,
905 * and just take the first swizzle channel for each argument in order
906 * into each writemask channel.
908 if (instr
->op
== nir_op_vec2
||
909 instr
->op
== nir_op_vec3
||
910 instr
->op
== nir_op_vec4
) {
912 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
913 srcs
[i
] = ntq_get_src(c
, instr
->src
[i
].src
,
914 instr
->src
[i
].swizzle
[0]);
915 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
916 ntq_store_dest(c
, &instr
->dest
.dest
, i
,
917 vir_MOV(c
, srcs
[i
]));
921 /* General case: We can just grab the one used channel per src. */
922 struct qreg src
[nir_op_infos
[instr
->op
].num_inputs
];
923 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
924 src
[i
] = ntq_get_alu_src(c
, instr
, i
);
931 result
= vir_MOV(c
, src
[0]);
935 result
= vir_XOR(c
, src
[0], vir_uniform_ui(c
, 1 << 31));
938 result
= vir_NEG(c
, src
[0]);
942 result
= vir_FMUL(c
, src
[0], src
[1]);
945 result
= vir_FADD(c
, src
[0], src
[1]);
948 result
= vir_FSUB(c
, src
[0], src
[1]);
951 result
= vir_FMIN(c
, src
[0], src
[1]);
954 result
= vir_FMAX(c
, src
[0], src
[1]);
958 nir_alu_instr
*src0_alu
= ntq_get_alu_parent(instr
->src
[0].src
);
959 if (src0_alu
&& src0_alu
->op
== nir_op_fround_even
) {
960 result
= vir_FTOIN(c
, ntq_get_alu_src(c
, src0_alu
, 0));
962 result
= vir_FTOIZ(c
, src
[0]);
968 result
= vir_FTOUZ(c
, src
[0]);
971 result
= vir_ITOF(c
, src
[0]);
974 result
= vir_UTOF(c
, src
[0]);
977 result
= vir_AND(c
, src
[0], vir_uniform_f(c
, 1.0));
980 result
= vir_AND(c
, src
[0], vir_uniform_ui(c
, 1));
984 result
= vir_ADD(c
, src
[0], src
[1]);
987 result
= vir_SHR(c
, src
[0], src
[1]);
990 result
= vir_SUB(c
, src
[0], src
[1]);
993 result
= vir_ASR(c
, src
[0], src
[1]);
996 result
= vir_SHL(c
, src
[0], src
[1]);
999 result
= vir_MIN(c
, src
[0], src
[1]);
1002 result
= vir_UMIN(c
, src
[0], src
[1]);
1005 result
= vir_MAX(c
, src
[0], src
[1]);
1008 result
= vir_UMAX(c
, src
[0], src
[1]);
1011 result
= vir_AND(c
, src
[0], src
[1]);
1014 result
= vir_OR(c
, src
[0], src
[1]);
1017 result
= vir_XOR(c
, src
[0], src
[1]);
1020 result
= vir_NOT(c
, src
[0]);
1023 case nir_op_ufind_msb
:
1024 result
= vir_SUB(c
, vir_uniform_ui(c
, 31), vir_CLZ(c
, src
[0]));
1028 result
= vir_UMUL(c
, src
[0], src
[1]);
1035 enum v3d_qpu_cond cond
;
1036 ASSERTED
bool ok
= ntq_emit_comparison(c
, instr
, &cond
);
1038 result
= vir_MOV(c
, vir_SEL(c
, cond
,
1039 vir_uniform_f(c
, 1.0),
1040 vir_uniform_f(c
, 0.0)));
1055 case nir_op_ult32
: {
1056 enum v3d_qpu_cond cond
;
1057 ASSERTED
bool ok
= ntq_emit_comparison(c
, instr
, &cond
);
1059 result
= vir_MOV(c
, vir_SEL(c
, cond
,
1060 vir_uniform_ui(c
, ~0),
1061 vir_uniform_ui(c
, 0)));
1065 case nir_op_b32csel
:
1068 ntq_emit_bool_to_cond(c
, instr
->src
[0].src
),
1073 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), src
[0]),
1075 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
1080 result
= vir_RECIP(c
, src
[0]);
1083 result
= vir_RSQRT(c
, src
[0]);
1086 result
= vir_EXP(c
, src
[0]);
1089 result
= vir_LOG(c
, src
[0]);
1093 result
= vir_FCEIL(c
, src
[0]);
1096 result
= vir_FFLOOR(c
, src
[0]);
1098 case nir_op_fround_even
:
1099 result
= vir_FROUND(c
, src
[0]);
1102 result
= vir_FTRUNC(c
, src
[0]);
1106 result
= ntq_fsincos(c
, src
[0], false);
1109 result
= ntq_fsincos(c
, src
[0], true);
1113 result
= ntq_fsign(c
, src
[0]);
1117 result
= vir_FMOV(c
, src
[0]);
1118 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_ABS
);
1123 result
= vir_MAX(c
, src
[0], vir_NEG(c
, src
[0]));
1127 case nir_op_fddx_coarse
:
1128 case nir_op_fddx_fine
:
1129 result
= vir_FDX(c
, src
[0]);
1133 case nir_op_fddy_coarse
:
1134 case nir_op_fddy_fine
:
1135 result
= vir_FDY(c
, src
[0]);
1138 case nir_op_uadd_carry
:
1139 vir_set_pf(vir_ADD_dest(c
, vir_nop_reg(), src
[0], src
[1]),
1141 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
1142 vir_uniform_ui(c
, ~0),
1143 vir_uniform_ui(c
, 0)));
1146 case nir_op_pack_half_2x16_split
:
1147 result
= vir_VFPACK(c
, src
[0], src
[1]);
1150 case nir_op_unpack_half_2x16_split_x
:
1151 result
= vir_FMOV(c
, src
[0]);
1152 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_L
);
1155 case nir_op_unpack_half_2x16_split_y
:
1156 result
= vir_FMOV(c
, src
[0]);
1157 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_H
);
1161 fprintf(stderr
, "unknown NIR ALU inst: ");
1162 nir_print_instr(&instr
->instr
, stderr
);
1163 fprintf(stderr
, "\n");
1167 /* We have a scalar result, so the instruction should only have a
1168 * single channel written to.
1170 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
1171 ntq_store_dest(c
, &instr
->dest
.dest
,
1172 ffs(instr
->dest
.write_mask
) - 1, result
);
1175 /* Each TLB read/write setup (a render target or depth buffer) takes an 8-bit
1176 * specifier. They come from a register that's preloaded with 0xffffffff
1177 * (0xff gets you normal vec4 f16 RT0 writes), and when one is neaded the low
1178 * 8 bits are shifted off the bottom and 0xff shifted in from the top.
1180 #define TLB_TYPE_F16_COLOR (3 << 6)
1181 #define TLB_TYPE_I32_COLOR (1 << 6)
1182 #define TLB_TYPE_F32_COLOR (0 << 6)
1183 #define TLB_RENDER_TARGET_SHIFT 3 /* Reversed! 7 = RT 0, 0 = RT 7. */
1184 #define TLB_SAMPLE_MODE_PER_SAMPLE (0 << 2)
1185 #define TLB_SAMPLE_MODE_PER_PIXEL (1 << 2)
1186 #define TLB_F16_SWAP_HI_LO (1 << 1)
1187 #define TLB_VEC_SIZE_4_F16 (1 << 0)
1188 #define TLB_VEC_SIZE_2_F16 (0 << 0)
1189 #define TLB_VEC_SIZE_MINUS_1_SHIFT 0
1191 /* Triggers Z/Stencil testing, used when the shader state's "FS modifies Z"
1194 #define TLB_TYPE_DEPTH ((2 << 6) | (0 << 4))
1195 #define TLB_DEPTH_TYPE_INVARIANT (0 << 2) /* Unmodified sideband input used */
1196 #define TLB_DEPTH_TYPE_PER_PIXEL (1 << 2) /* QPU result used */
1197 #define TLB_V42_DEPTH_TYPE_INVARIANT (0 << 3) /* Unmodified sideband input used */
1198 #define TLB_V42_DEPTH_TYPE_PER_PIXEL (1 << 3) /* QPU result used */
1200 /* Stencil is a single 32-bit write. */
1201 #define TLB_TYPE_STENCIL_ALPHA ((2 << 6) | (1 << 4))
1204 vir_emit_tlb_color_write(struct v3d_compile
*c
, unsigned rt
)
1206 if (!(c
->fs_key
->cbufs
& (1 << rt
)) || !c
->output_color_var
[rt
])
1209 struct qreg tlb_reg
= vir_magic_reg(V3D_QPU_WADDR_TLB
);
1210 struct qreg tlbu_reg
= vir_magic_reg(V3D_QPU_WADDR_TLBU
);
1212 nir_variable
*var
= c
->output_color_var
[rt
];
1213 int num_components
= glsl_get_vector_elements(var
->type
);
1214 uint32_t conf
= 0xffffff00;
1217 conf
|= c
->msaa_per_sample_output
? TLB_SAMPLE_MODE_PER_SAMPLE
:
1218 TLB_SAMPLE_MODE_PER_PIXEL
;
1219 conf
|= (7 - rt
) << TLB_RENDER_TARGET_SHIFT
;
1221 if (c
->fs_key
->swap_color_rb
& (1 << rt
))
1222 num_components
= MAX2(num_components
, 3);
1223 assert(num_components
!= 0);
1225 enum glsl_base_type type
= glsl_get_base_type(var
->type
);
1226 bool is_int_format
= type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_UINT
;
1227 bool is_32b_tlb_format
= is_int_format
||
1228 (c
->fs_key
->f32_color_rb
& (1 << rt
));
1230 if (is_int_format
) {
1231 /* The F32 vs I32 distinction was dropped in 4.2. */
1232 if (c
->devinfo
->ver
< 42)
1233 conf
|= TLB_TYPE_I32_COLOR
;
1235 conf
|= TLB_TYPE_F32_COLOR
;
1236 conf
|= ((num_components
- 1) << TLB_VEC_SIZE_MINUS_1_SHIFT
);
1238 if (c
->fs_key
->f32_color_rb
& (1 << rt
)) {
1239 conf
|= TLB_TYPE_F32_COLOR
;
1240 conf
|= ((num_components
- 1) <<
1241 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1243 conf
|= TLB_TYPE_F16_COLOR
;
1244 conf
|= TLB_F16_SWAP_HI_LO
;
1245 if (num_components
>= 3)
1246 conf
|= TLB_VEC_SIZE_4_F16
;
1248 conf
|= TLB_VEC_SIZE_2_F16
;
1252 int num_samples
= c
->msaa_per_sample_output
? V3D_MAX_SAMPLES
: 1;
1253 for (int i
= 0; i
< num_samples
; i
++) {
1254 struct qreg
*color
= c
->msaa_per_sample_output
?
1255 &c
->sample_colors
[(rt
* V3D_MAX_SAMPLES
+ i
) * 4] :
1256 &c
->outputs
[var
->data
.driver_location
* 4];
1258 struct qreg r
= color
[0];
1259 struct qreg g
= color
[1];
1260 struct qreg b
= color
[2];
1261 struct qreg a
= color
[3];
1263 if (c
->fs_key
->swap_color_rb
& (1 << rt
)) {
1268 if (c
->fs_key
->sample_alpha_to_one
)
1269 a
= vir_uniform_f(c
, 1.0);
1271 if (is_32b_tlb_format
) {
1273 inst
= vir_MOV_dest(c
, tlbu_reg
, r
);
1275 vir_get_uniform_index(c
,
1279 inst
= vir_MOV_dest(c
, tlb_reg
, r
);
1282 if (num_components
>= 2)
1283 vir_MOV_dest(c
, tlb_reg
, g
);
1284 if (num_components
>= 3)
1285 vir_MOV_dest(c
, tlb_reg
, b
);
1286 if (num_components
>= 4)
1287 vir_MOV_dest(c
, tlb_reg
, a
);
1289 inst
= vir_VFPACK_dest(c
, tlb_reg
, r
, g
);
1290 if (conf
!= ~0 && i
== 0) {
1291 inst
->dst
= tlbu_reg
;
1293 vir_get_uniform_index(c
,
1298 if (num_components
>= 3)
1299 inst
= vir_VFPACK_dest(c
, tlb_reg
, b
, a
);
1305 emit_frag_end(struct v3d_compile
*c
)
1308 if (c->output_sample_mask_index != -1) {
1309 vir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1313 bool has_any_tlb_color_write
= false;
1314 for (int rt
= 0; rt
< V3D_MAX_DRAW_BUFFERS
; rt
++) {
1315 if (c
->fs_key
->cbufs
& (1 << rt
) && c
->output_color_var
[rt
])
1316 has_any_tlb_color_write
= true;
1319 if (c
->fs_key
->sample_alpha_to_coverage
&& c
->output_color_var
[0]) {
1320 struct nir_variable
*var
= c
->output_color_var
[0];
1321 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
1323 vir_SETMSF_dest(c
, vir_nop_reg(),
1326 vir_FTOC(c
, color
[3])));
1329 struct qreg tlbu_reg
= vir_magic_reg(V3D_QPU_WADDR_TLBU
);
1330 if (c
->output_position_index
!= -1) {
1331 struct qinst
*inst
= vir_MOV_dest(c
, tlbu_reg
,
1332 c
->outputs
[c
->output_position_index
]);
1333 uint8_t tlb_specifier
= TLB_TYPE_DEPTH
;
1335 if (c
->devinfo
->ver
>= 42) {
1336 tlb_specifier
|= (TLB_V42_DEPTH_TYPE_PER_PIXEL
|
1337 TLB_SAMPLE_MODE_PER_PIXEL
);
1339 tlb_specifier
|= TLB_DEPTH_TYPE_PER_PIXEL
;
1341 inst
->uniform
= vir_get_uniform_index(c
, QUNIFORM_CONSTANT
,
1345 } else if (c
->s
->info
.fs
.uses_discard
||
1346 !c
->s
->info
.fs
.early_fragment_tests
||
1347 c
->fs_key
->sample_alpha_to_coverage
||
1348 !has_any_tlb_color_write
) {
1349 /* Emit passthrough Z if it needed to be delayed until shader
1350 * end due to potential discards.
1352 * Since (single-threaded) fragment shaders always need a TLB
1353 * write, emit passthrouh Z if we didn't have any color
1354 * buffers and flag us as potentially discarding, so that we
1355 * can use Z as the TLB write.
1357 c
->s
->info
.fs
.uses_discard
= true;
1359 struct qinst
*inst
= vir_MOV_dest(c
, tlbu_reg
,
1361 uint8_t tlb_specifier
= TLB_TYPE_DEPTH
;
1363 if (c
->devinfo
->ver
>= 42) {
1364 /* The spec says the PER_PIXEL flag is ignored for
1365 * invariant writes, but the simulator demands it.
1367 tlb_specifier
|= (TLB_V42_DEPTH_TYPE_INVARIANT
|
1368 TLB_SAMPLE_MODE_PER_PIXEL
);
1370 tlb_specifier
|= TLB_DEPTH_TYPE_INVARIANT
;
1373 inst
->uniform
= vir_get_uniform_index(c
,
1380 /* XXX: Performance improvement: Merge Z write and color writes TLB
1383 for (int rt
= 0; rt
< V3D_MAX_DRAW_BUFFERS
; rt
++)
1384 vir_emit_tlb_color_write(c
, rt
);
1388 vir_VPM_WRITE_indirect(struct v3d_compile
*c
,
1390 struct qreg vpm_index
)
1392 assert(c
->devinfo
->ver
>= 40);
1393 vir_STVPMV(c
, vpm_index
, val
);
1397 vir_VPM_WRITE(struct v3d_compile
*c
, struct qreg val
, uint32_t vpm_index
)
1399 if (c
->devinfo
->ver
>= 40) {
1400 vir_VPM_WRITE_indirect(c
, val
, vir_uniform_ui(c
, vpm_index
));
1402 /* XXX: v3d33_vir_vpm_write_setup(c); */
1403 vir_MOV_dest(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_VPM
), val
);
1408 emit_vert_end(struct v3d_compile
*c
)
1410 /* GFXH-1684: VPM writes need to be complete by the end of the shader.
1412 if (c
->devinfo
->ver
>= 40 && c
->devinfo
->ver
<= 42)
1417 emit_geom_end(struct v3d_compile
*c
)
1419 /* GFXH-1684: VPM writes need to be complete by the end of the shader.
1421 if (c
->devinfo
->ver
>= 40 && c
->devinfo
->ver
<= 42)
1426 v3d_optimize_nir(struct nir_shader
*s
)
1429 unsigned lower_flrp
=
1430 (s
->options
->lower_flrp16
? 16 : 0) |
1431 (s
->options
->lower_flrp32
? 32 : 0) |
1432 (s
->options
->lower_flrp64
? 64 : 0);
1437 NIR_PASS_V(s
, nir_lower_vars_to_ssa
);
1438 NIR_PASS(progress
, s
, nir_lower_alu_to_scalar
, NULL
, NULL
);
1439 NIR_PASS(progress
, s
, nir_lower_phis_to_scalar
);
1440 NIR_PASS(progress
, s
, nir_copy_prop
);
1441 NIR_PASS(progress
, s
, nir_opt_remove_phis
);
1442 NIR_PASS(progress
, s
, nir_opt_dce
);
1443 NIR_PASS(progress
, s
, nir_opt_dead_cf
);
1444 NIR_PASS(progress
, s
, nir_opt_cse
);
1445 NIR_PASS(progress
, s
, nir_opt_peephole_select
, 8, true, true);
1446 NIR_PASS(progress
, s
, nir_opt_algebraic
);
1447 NIR_PASS(progress
, s
, nir_opt_constant_folding
);
1449 if (lower_flrp
!= 0) {
1450 bool lower_flrp_progress
= false;
1452 NIR_PASS(lower_flrp_progress
, s
, nir_lower_flrp
,
1454 false /* always_precise */,
1455 s
->options
->lower_ffma
);
1456 if (lower_flrp_progress
) {
1457 NIR_PASS(progress
, s
, nir_opt_constant_folding
);
1461 /* Nothing should rematerialize any flrps, so we only
1462 * need to do this lowering once.
1467 NIR_PASS(progress
, s
, nir_opt_undef
);
1470 NIR_PASS(progress
, s
, nir_opt_move
, nir_move_load_ubo
);
1474 driver_location_compare(const void *in_a
, const void *in_b
)
1476 const nir_variable
*const *a
= in_a
;
1477 const nir_variable
*const *b
= in_b
;
1479 if ((*a
)->data
.driver_location
== (*b
)->data
.driver_location
)
1480 return (*a
)->data
.location_frac
- (*b
)->data
.location_frac
;
1482 return (*a
)->data
.driver_location
- (*b
)->data
.driver_location
;
1486 ntq_emit_vpm_read(struct v3d_compile
*c
,
1487 uint32_t *num_components_queued
,
1488 uint32_t *remaining
,
1491 struct qreg vpm
= vir_reg(QFILE_VPM
, vpm_index
);
1493 if (c
->devinfo
->ver
>= 40 ) {
1494 return vir_LDVPMV_IN(c
,
1496 (*num_components_queued
)++));
1499 if (*num_components_queued
!= 0) {
1500 (*num_components_queued
)--;
1501 return vir_MOV(c
, vpm
);
1504 uint32_t num_components
= MIN2(*remaining
, 32);
1506 v3d33_vir_vpm_read_setup(c
, num_components
);
1508 *num_components_queued
= num_components
- 1;
1509 *remaining
-= num_components
;
1511 return vir_MOV(c
, vpm
);
1515 ntq_setup_vs_inputs(struct v3d_compile
*c
)
1517 /* Figure out how many components of each vertex attribute the shader
1518 * uses. Each variable should have been split to individual
1519 * components and unused ones DCEed. The vertex fetcher will load
1520 * from the start of the attribute to the number of components we
1521 * declare we need in c->vattr_sizes[].
1523 nir_foreach_shader_in_variable(var
, c
->s
) {
1524 /* No VS attribute array support. */
1525 assert(MAX2(glsl_get_length(var
->type
), 1) == 1);
1527 unsigned loc
= var
->data
.driver_location
;
1528 int start_component
= var
->data
.location_frac
;
1529 int num_components
= glsl_get_components(var
->type
);
1531 c
->vattr_sizes
[loc
] = MAX2(c
->vattr_sizes
[loc
],
1532 start_component
+ num_components
);
1535 unsigned num_components
= 0;
1536 uint32_t vpm_components_queued
= 0;
1537 bool uses_iid
= c
->s
->info
.system_values_read
&
1538 (1ull << SYSTEM_VALUE_INSTANCE_ID
);
1539 bool uses_vid
= c
->s
->info
.system_values_read
&
1540 (1ull << SYSTEM_VALUE_VERTEX_ID
);
1541 num_components
+= uses_iid
;
1542 num_components
+= uses_vid
;
1544 for (int i
= 0; i
< ARRAY_SIZE(c
->vattr_sizes
); i
++)
1545 num_components
+= c
->vattr_sizes
[i
];
1548 c
->iid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1549 &num_components
, ~0);
1553 c
->vid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1554 &num_components
, ~0);
1557 /* The actual loads will happen directly in nir_intrinsic_load_input
1558 * on newer versions.
1560 if (c
->devinfo
->ver
>= 40)
1563 for (int loc
= 0; loc
< ARRAY_SIZE(c
->vattr_sizes
); loc
++) {
1564 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1567 for (int i
= 0; i
< c
->vattr_sizes
[loc
]; i
++) {
1568 c
->inputs
[loc
* 4 + i
] =
1569 ntq_emit_vpm_read(c
,
1570 &vpm_components_queued
,
1577 if (c
->devinfo
->ver
>= 40) {
1578 assert(vpm_components_queued
== num_components
);
1580 assert(vpm_components_queued
== 0);
1581 assert(num_components
== 0);
1586 program_reads_point_coord(struct v3d_compile
*c
)
1588 nir_foreach_shader_in_variable(var
, c
->s
) {
1589 if (util_varying_is_point_coord(var
->data
.location
,
1590 c
->fs_key
->point_sprite_mask
)) {
1599 get_sorted_input_variables(struct v3d_compile
*c
,
1600 unsigned *num_entries
,
1601 nir_variable
***vars
)
1604 nir_foreach_shader_in_variable(var
, c
->s
)
1607 *vars
= ralloc_array(c
, nir_variable
*, *num_entries
);
1610 nir_foreach_shader_in_variable(var
, c
->s
)
1613 /* Sort the variables so that we emit the input setup in
1614 * driver_location order. This is required for VPM reads, whose data
1615 * is fetched into the VPM in driver_location (TGSI register index)
1618 qsort(*vars
, *num_entries
, sizeof(**vars
), driver_location_compare
);
1622 ntq_setup_gs_inputs(struct v3d_compile
*c
)
1624 nir_variable
**vars
;
1625 unsigned num_entries
;
1626 get_sorted_input_variables(c
, &num_entries
, &vars
);
1628 for (unsigned i
= 0; i
< num_entries
; i
++) {
1629 nir_variable
*var
= vars
[i
];
1631 /* All GS inputs are arrays with as many entries as vertices
1632 * in the input primitive, but here we only care about the
1633 * per-vertex input type.
1635 const struct glsl_type
*type
= glsl_without_array(var
->type
);
1636 unsigned array_len
= MAX2(glsl_get_length(type
), 1);
1637 unsigned loc
= var
->data
.driver_location
;
1639 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1640 (loc
+ array_len
) * 4);
1642 for (unsigned j
= 0; j
< array_len
; j
++) {
1643 unsigned num_elements
= glsl_get_vector_elements(type
);
1644 for (unsigned k
= 0; k
< num_elements
; k
++) {
1645 unsigned chan
= var
->data
.location_frac
+ k
;
1646 unsigned input_idx
= c
->num_inputs
++;
1647 struct v3d_varying_slot slot
=
1648 v3d_slot_from_slot_and_component(var
->data
.location
+ j
, chan
);
1649 c
->input_slots
[input_idx
] = slot
;
1657 ntq_setup_fs_inputs(struct v3d_compile
*c
)
1659 nir_variable
**vars
;
1660 unsigned num_entries
;
1661 get_sorted_input_variables(c
, &num_entries
, &vars
);
1663 for (unsigned i
= 0; i
< num_entries
; i
++) {
1664 nir_variable
*var
= vars
[i
];
1665 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1666 unsigned loc
= var
->data
.driver_location
;
1668 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1669 (loc
+ array_len
) * 4);
1671 if (var
->data
.location
== VARYING_SLOT_POS
) {
1672 emit_fragcoord_input(c
, loc
);
1673 } else if (util_varying_is_point_coord(var
->data
.location
,
1674 c
->fs_key
->point_sprite_mask
)) {
1675 c
->inputs
[loc
* 4 + 0] = c
->point_x
;
1676 c
->inputs
[loc
* 4 + 1] = c
->point_y
;
1677 } else if (var
->data
.compact
) {
1678 for (int j
= 0; j
< array_len
; j
++)
1679 emit_compact_fragment_input(c
, loc
, var
, j
);
1681 for (int j
= 0; j
< array_len
; j
++)
1682 emit_fragment_input(c
, loc
+ j
, var
, j
);
1688 ntq_setup_outputs(struct v3d_compile
*c
)
1690 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
)
1693 nir_foreach_shader_out_variable(var
, c
->s
) {
1694 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1695 unsigned loc
= var
->data
.driver_location
* 4;
1697 assert(array_len
== 1);
1700 for (int i
= 0; i
< 4 - var
->data
.location_frac
; i
++) {
1701 add_output(c
, loc
+ var
->data
.location_frac
+ i
,
1703 var
->data
.location_frac
+ i
);
1706 switch (var
->data
.location
) {
1707 case FRAG_RESULT_COLOR
:
1708 c
->output_color_var
[0] = var
;
1709 c
->output_color_var
[1] = var
;
1710 c
->output_color_var
[2] = var
;
1711 c
->output_color_var
[3] = var
;
1713 case FRAG_RESULT_DATA0
:
1714 case FRAG_RESULT_DATA1
:
1715 case FRAG_RESULT_DATA2
:
1716 case FRAG_RESULT_DATA3
:
1717 c
->output_color_var
[var
->data
.location
-
1718 FRAG_RESULT_DATA0
] = var
;
1720 case FRAG_RESULT_DEPTH
:
1721 c
->output_position_index
= loc
;
1723 case FRAG_RESULT_SAMPLE_MASK
:
1724 c
->output_sample_mask_index
= loc
;
1731 * Sets up the mapping from nir_register to struct qreg *.
1733 * Each nir_register gets a struct qreg per 32-bit component being stored.
1736 ntq_setup_registers(struct v3d_compile
*c
, struct exec_list
*list
)
1738 foreach_list_typed(nir_register
, nir_reg
, node
, list
) {
1739 unsigned array_len
= MAX2(nir_reg
->num_array_elems
, 1);
1740 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
1742 nir_reg
->num_components
);
1744 _mesa_hash_table_insert(c
->def_ht
, nir_reg
, qregs
);
1746 for (int i
= 0; i
< array_len
* nir_reg
->num_components
; i
++)
1747 qregs
[i
] = vir_get_temp(c
);
1752 ntq_emit_load_const(struct v3d_compile
*c
, nir_load_const_instr
*instr
)
1754 /* XXX perf: Experiment with using immediate loads to avoid having
1755 * these end up in the uniform stream. Watch out for breaking the
1756 * small immediates optimization in the process!
1758 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1759 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1760 qregs
[i
] = vir_uniform_ui(c
, instr
->value
[i
].u32
);
1762 _mesa_hash_table_insert(c
->def_ht
, &instr
->def
, qregs
);
1766 ntq_emit_ssa_undef(struct v3d_compile
*c
, nir_ssa_undef_instr
*instr
)
1768 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1770 /* VIR needs there to be *some* value, so pick 0 (same as for
1771 * ntq_setup_registers().
1773 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1774 qregs
[i
] = vir_uniform_ui(c
, 0);
1778 ntq_emit_image_size(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1780 unsigned image_index
= nir_src_as_uint(instr
->src
[0]);
1781 bool is_array
= nir_intrinsic_image_array(instr
);
1783 assert(nir_src_as_uint(instr
->src
[1]) == 0);
1785 ntq_store_dest(c
, &instr
->dest
, 0,
1786 vir_uniform(c
, QUNIFORM_IMAGE_WIDTH
, image_index
));
1787 if (instr
->num_components
> 1) {
1788 ntq_store_dest(c
, &instr
->dest
, 1,
1790 instr
->num_components
== 2 && is_array
?
1791 QUNIFORM_IMAGE_ARRAY_SIZE
:
1792 QUNIFORM_IMAGE_HEIGHT
,
1795 if (instr
->num_components
> 2) {
1796 ntq_store_dest(c
, &instr
->dest
, 2,
1799 QUNIFORM_IMAGE_ARRAY_SIZE
:
1800 QUNIFORM_IMAGE_DEPTH
,
1806 vir_emit_tlb_color_read(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1808 assert(c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
);
1810 int rt
= nir_src_as_uint(instr
->src
[0]);
1811 assert(rt
< V3D_MAX_DRAW_BUFFERS
);
1813 int sample_index
= nir_intrinsic_base(instr
) ;
1814 assert(sample_index
< V3D_MAX_SAMPLES
);
1816 int component
= nir_intrinsic_component(instr
);
1817 assert(component
< 4);
1819 /* We need to emit our TLB reads after we have acquired the scoreboard
1820 * lock, or the GPU will hang. Usually, we do our scoreboard locking on
1821 * the last thread switch to improve parallelism, however, that is only
1822 * guaranteed to happen before the tlb color writes.
1824 * To fix that, we make sure we always emit a thread switch before the
1825 * first tlb color read. If that happens to be the last thread switch
1826 * we emit, then everything is fine, but otherwsie, if any code after
1827 * this point needs to emit additional thread switches, then we will
1828 * switch the strategy to locking the scoreboard on the first thread
1829 * switch instead -- see vir_emit_thrsw().
1831 if (!c
->emitted_tlb_load
) {
1832 if (!c
->last_thrsw_at_top_level
) {
1833 assert(c
->devinfo
->ver
>= 41);
1837 c
->emitted_tlb_load
= true;
1840 struct qreg
*color_reads_for_sample
=
1841 &c
->color_reads
[(rt
* V3D_MAX_SAMPLES
+ sample_index
) * 4];
1843 if (color_reads_for_sample
[component
].file
== QFILE_NULL
) {
1844 enum pipe_format rt_format
= c
->fs_key
->color_fmt
[rt
].format
;
1845 int num_components
=
1846 util_format_get_nr_components(rt_format
);
1848 const bool swap_rb
= c
->fs_key
->swap_color_rb
& (1 << rt
);
1850 num_components
= MAX2(num_components
, 3);
1852 nir_variable
*var
= c
->output_color_var
[rt
];
1853 enum glsl_base_type type
= glsl_get_base_type(var
->type
);
1855 bool is_int_format
= type
== GLSL_TYPE_INT
||
1856 type
== GLSL_TYPE_UINT
;
1858 bool is_32b_tlb_format
= is_int_format
||
1859 (c
->fs_key
->f32_color_rb
& (1 << rt
));
1861 int num_samples
= c
->fs_key
->msaa
? V3D_MAX_SAMPLES
: 1;
1863 uint32_t conf
= 0xffffff00;
1864 conf
|= c
->fs_key
->msaa
? TLB_SAMPLE_MODE_PER_SAMPLE
:
1865 TLB_SAMPLE_MODE_PER_PIXEL
;
1866 conf
|= (7 - rt
) << TLB_RENDER_TARGET_SHIFT
;
1868 if (is_32b_tlb_format
) {
1869 /* The F32 vs I32 distinction was dropped in 4.2. */
1870 conf
|= (c
->devinfo
->ver
< 42 && is_int_format
) ?
1871 TLB_TYPE_I32_COLOR
: TLB_TYPE_F32_COLOR
;
1873 conf
|= ((num_components
- 1) <<
1874 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1876 conf
|= TLB_TYPE_F16_COLOR
;
1877 conf
|= TLB_F16_SWAP_HI_LO
;
1879 if (num_components
>= 3)
1880 conf
|= TLB_VEC_SIZE_4_F16
;
1882 conf
|= TLB_VEC_SIZE_2_F16
;
1886 for (int i
= 0; i
< num_samples
; i
++) {
1887 struct qreg r
, g
, b
, a
;
1888 if (is_32b_tlb_format
) {
1889 r
= conf
!= 0xffffffff && i
== 0?
1890 vir_TLBU_COLOR_READ(c
, conf
) :
1891 vir_TLB_COLOR_READ(c
);
1892 if (num_components
>= 2)
1893 g
= vir_TLB_COLOR_READ(c
);
1894 if (num_components
>= 3)
1895 b
= vir_TLB_COLOR_READ(c
);
1896 if (num_components
>= 4)
1897 a
= vir_TLB_COLOR_READ(c
);
1899 struct qreg rg
= conf
!= 0xffffffff && i
== 0 ?
1900 vir_TLBU_COLOR_READ(c
, conf
) :
1901 vir_TLB_COLOR_READ(c
);
1902 r
= vir_FMOV(c
, rg
);
1903 vir_set_unpack(c
->defs
[r
.index
], 0,
1905 g
= vir_FMOV(c
, rg
);
1906 vir_set_unpack(c
->defs
[g
.index
], 0,
1909 if (num_components
> 2) {
1910 struct qreg ba
= vir_TLB_COLOR_READ(c
);
1911 b
= vir_FMOV(c
, ba
);
1912 vir_set_unpack(c
->defs
[b
.index
], 0,
1914 a
= vir_FMOV(c
, ba
);
1915 vir_set_unpack(c
->defs
[a
.index
], 0,
1920 struct qreg
*color_reads
=
1921 &c
->color_reads
[(rt
* V3D_MAX_SAMPLES
+ i
) * 4];
1923 color_reads
[0] = swap_rb
? b
: r
;
1924 if (num_components
>= 2)
1926 if (num_components
>= 3)
1927 color_reads
[2] = swap_rb
? r
: b
;
1928 if (num_components
>= 4)
1933 assert(color_reads_for_sample
[component
].file
!= QFILE_NULL
);
1934 ntq_store_dest(c
, &instr
->dest
, 0,
1935 vir_MOV(c
, color_reads_for_sample
[component
]));
1939 ntq_emit_load_uniform(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1941 if (nir_src_is_const(instr
->src
[0])) {
1942 int offset
= (nir_intrinsic_base(instr
) +
1943 nir_src_as_uint(instr
->src
[0]));
1944 assert(offset
% 4 == 0);
1945 /* We need dwords */
1946 offset
= offset
/ 4;
1947 for (int i
= 0; i
< instr
->num_components
; i
++) {
1948 ntq_store_dest(c
, &instr
->dest
, i
,
1949 vir_uniform(c
, QUNIFORM_UNIFORM
,
1953 ntq_emit_tmu_general(c
, instr
, false);
1958 ntq_emit_load_input(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1960 /* XXX: Use ldvpmv (uniform offset) or ldvpmd (non-uniform offset)
1961 * and enable PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR.
1964 nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[0]);
1966 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
&& c
->devinfo
->ver
>= 40) {
1967 /* Emit the LDVPM directly now, rather than at the top
1968 * of the shader like we did for V3D 3.x (which needs
1969 * vpmsetup when not just taking the next offset).
1971 * Note that delaying like this may introduce stalls,
1972 * as LDVPMV takes a minimum of 1 instruction but may
1973 * be slower if the VPM unit is busy with another QPU.
1976 if (c
->s
->info
.system_values_read
&
1977 (1ull << SYSTEM_VALUE_INSTANCE_ID
)) {
1980 if (c
->s
->info
.system_values_read
&
1981 (1ull << SYSTEM_VALUE_VERTEX_ID
)) {
1984 for (int i
= 0; i
< offset
; i
++)
1985 index
+= c
->vattr_sizes
[i
];
1986 index
+= nir_intrinsic_component(instr
);
1987 for (int i
= 0; i
< instr
->num_components
; i
++) {
1988 struct qreg vpm_offset
= vir_uniform_ui(c
, index
++);
1989 ntq_store_dest(c
, &instr
->dest
, i
,
1990 vir_LDVPMV_IN(c
, vpm_offset
));
1993 for (int i
= 0; i
< instr
->num_components
; i
++) {
1994 int comp
= nir_intrinsic_component(instr
) + i
;
1995 ntq_store_dest(c
, &instr
->dest
, i
,
1996 vir_MOV(c
, c
->inputs
[offset
* 4 + comp
]));
2002 ntq_emit_per_sample_color_write(struct v3d_compile
*c
,
2003 nir_intrinsic_instr
*instr
)
2005 assert(instr
->intrinsic
== nir_intrinsic_store_tlb_sample_color_v3d
);
2007 unsigned rt
= nir_src_as_uint(instr
->src
[1]);
2008 assert(rt
< V3D_MAX_DRAW_BUFFERS
);
2010 unsigned sample_idx
= nir_intrinsic_base(instr
);
2011 assert(sample_idx
< V3D_MAX_SAMPLES
);
2013 unsigned offset
= (rt
* V3D_MAX_SAMPLES
+ sample_idx
) * 4;
2014 for (int i
= 0; i
< instr
->num_components
; i
++) {
2015 c
->sample_colors
[offset
+ i
] =
2016 vir_MOV(c
, ntq_get_src(c
, instr
->src
[0], i
));
2021 ntq_emit_color_write(struct v3d_compile
*c
,
2022 nir_intrinsic_instr
*instr
)
2024 unsigned offset
= (nir_intrinsic_base(instr
) +
2025 nir_src_as_uint(instr
->src
[1])) * 4 +
2026 nir_intrinsic_component(instr
);
2027 for (int i
= 0; i
< instr
->num_components
; i
++) {
2028 c
->outputs
[offset
+ i
] =
2029 vir_MOV(c
, ntq_get_src(c
, instr
->src
[0], i
));
2034 emit_store_output_gs(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
2036 assert(instr
->num_components
== 1);
2038 struct qreg offset
= ntq_get_src(c
, instr
->src
[1], 0);
2040 uint32_t base_offset
= nir_intrinsic_base(instr
);
2043 offset
= vir_ADD(c
, vir_uniform_ui(c
, base_offset
), offset
);
2045 /* Usually, for VS or FS, we only emit outputs once at program end so
2046 * our VPM writes are never in non-uniform control flow, but this
2047 * is not true for GS, where we are emitting multiple vertices.
2049 if (vir_in_nonuniform_control_flow(c
)) {
2050 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2054 struct qreg val
= ntq_get_src(c
, instr
->src
[0], 0);
2056 /* The offset isn’t necessarily dynamically uniform for a geometry
2057 * shader. This can happen if the shader sometimes doesn’t emit one of
2058 * the vertices. In that case subsequent vertices will be written to
2059 * different offsets in the VPM and we need to use the scatter write
2060 * instruction to have a different offset for each lane.
2062 if (nir_src_is_dynamically_uniform(instr
->src
[1]))
2063 vir_VPM_WRITE_indirect(c
, val
, offset
);
2065 vir_STVPMD(c
, offset
, val
);
2067 if (vir_in_nonuniform_control_flow(c
)) {
2068 struct qinst
*last_inst
=
2069 (struct qinst
*)c
->cur_block
->instructions
.prev
;
2070 vir_set_cond(last_inst
, V3D_QPU_COND_IFA
);
2075 ntq_emit_store_output(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
2077 /* XXX perf: Use stvpmv with uniform non-constant offsets and
2078 * stvpmd with non-uniform offsets and enable
2079 * PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR.
2081 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2082 ntq_emit_color_write(c
, instr
);
2083 } else if (c
->s
->info
.stage
== MESA_SHADER_GEOMETRY
) {
2084 emit_store_output_gs(c
, instr
);
2086 assert(c
->s
->info
.stage
== MESA_SHADER_VERTEX
);
2087 assert(instr
->num_components
== 1);
2090 ntq_get_src(c
, instr
->src
[0], 0),
2091 nir_intrinsic_base(instr
));
2096 ntq_emit_intrinsic(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
2098 switch (instr
->intrinsic
) {
2099 case nir_intrinsic_load_uniform
:
2100 ntq_emit_load_uniform(c
, instr
);
2103 case nir_intrinsic_load_ubo
:
2104 ntq_emit_tmu_general(c
, instr
, false);
2107 case nir_intrinsic_ssbo_atomic_add
:
2108 case nir_intrinsic_ssbo_atomic_imin
:
2109 case nir_intrinsic_ssbo_atomic_umin
:
2110 case nir_intrinsic_ssbo_atomic_imax
:
2111 case nir_intrinsic_ssbo_atomic_umax
:
2112 case nir_intrinsic_ssbo_atomic_and
:
2113 case nir_intrinsic_ssbo_atomic_or
:
2114 case nir_intrinsic_ssbo_atomic_xor
:
2115 case nir_intrinsic_ssbo_atomic_exchange
:
2116 case nir_intrinsic_ssbo_atomic_comp_swap
:
2117 case nir_intrinsic_load_ssbo
:
2118 case nir_intrinsic_store_ssbo
:
2119 ntq_emit_tmu_general(c
, instr
, false);
2122 case nir_intrinsic_shared_atomic_add
:
2123 case nir_intrinsic_shared_atomic_imin
:
2124 case nir_intrinsic_shared_atomic_umin
:
2125 case nir_intrinsic_shared_atomic_imax
:
2126 case nir_intrinsic_shared_atomic_umax
:
2127 case nir_intrinsic_shared_atomic_and
:
2128 case nir_intrinsic_shared_atomic_or
:
2129 case nir_intrinsic_shared_atomic_xor
:
2130 case nir_intrinsic_shared_atomic_exchange
:
2131 case nir_intrinsic_shared_atomic_comp_swap
:
2132 case nir_intrinsic_load_shared
:
2133 case nir_intrinsic_store_shared
:
2134 case nir_intrinsic_load_scratch
:
2135 case nir_intrinsic_store_scratch
:
2136 ntq_emit_tmu_general(c
, instr
, true);
2139 case nir_intrinsic_image_load
:
2140 case nir_intrinsic_image_store
:
2141 case nir_intrinsic_image_atomic_add
:
2142 case nir_intrinsic_image_atomic_imin
:
2143 case nir_intrinsic_image_atomic_umin
:
2144 case nir_intrinsic_image_atomic_imax
:
2145 case nir_intrinsic_image_atomic_umax
:
2146 case nir_intrinsic_image_atomic_and
:
2147 case nir_intrinsic_image_atomic_or
:
2148 case nir_intrinsic_image_atomic_xor
:
2149 case nir_intrinsic_image_atomic_exchange
:
2150 case nir_intrinsic_image_atomic_comp_swap
:
2151 v3d40_vir_emit_image_load_store(c
, instr
);
2154 case nir_intrinsic_get_buffer_size
:
2155 ntq_store_dest(c
, &instr
->dest
, 0,
2156 vir_uniform(c
, QUNIFORM_GET_BUFFER_SIZE
,
2157 nir_src_as_uint(instr
->src
[0])));
2160 case nir_intrinsic_load_user_clip_plane
:
2161 for (int i
= 0; i
< nir_intrinsic_dest_components(instr
); i
++) {
2162 ntq_store_dest(c
, &instr
->dest
, i
,
2163 vir_uniform(c
, QUNIFORM_USER_CLIP_PLANE
,
2164 nir_intrinsic_ucp_id(instr
) *
2169 case nir_intrinsic_load_viewport_x_scale
:
2170 ntq_store_dest(c
, &instr
->dest
, 0,
2171 vir_uniform(c
, QUNIFORM_VIEWPORT_X_SCALE
, 0));
2174 case nir_intrinsic_load_viewport_y_scale
:
2175 ntq_store_dest(c
, &instr
->dest
, 0,
2176 vir_uniform(c
, QUNIFORM_VIEWPORT_Y_SCALE
, 0));
2179 case nir_intrinsic_load_viewport_z_scale
:
2180 ntq_store_dest(c
, &instr
->dest
, 0,
2181 vir_uniform(c
, QUNIFORM_VIEWPORT_Z_SCALE
, 0));
2184 case nir_intrinsic_load_viewport_z_offset
:
2185 ntq_store_dest(c
, &instr
->dest
, 0,
2186 vir_uniform(c
, QUNIFORM_VIEWPORT_Z_OFFSET
, 0));
2189 case nir_intrinsic_load_alpha_ref_float
:
2190 ntq_store_dest(c
, &instr
->dest
, 0,
2191 vir_uniform(c
, QUNIFORM_ALPHA_REF
, 0));
2194 case nir_intrinsic_load_line_coord
:
2195 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->line_x
));
2198 case nir_intrinsic_load_line_width
:
2199 ntq_store_dest(c
, &instr
->dest
, 0,
2200 vir_uniform(c
, QUNIFORM_LINE_WIDTH
, 0));
2203 case nir_intrinsic_load_aa_line_width
:
2204 ntq_store_dest(c
, &instr
->dest
, 0,
2205 vir_uniform(c
, QUNIFORM_AA_LINE_WIDTH
, 0));
2208 case nir_intrinsic_load_sample_mask_in
:
2209 ntq_store_dest(c
, &instr
->dest
, 0, vir_MSF(c
));
2212 case nir_intrinsic_load_helper_invocation
:
2213 vir_set_pf(vir_MSF_dest(c
, vir_nop_reg()), V3D_QPU_PF_PUSHZ
);
2214 ntq_store_dest(c
, &instr
->dest
, 0,
2215 vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
2216 vir_uniform_ui(c
, ~0),
2217 vir_uniform_ui(c
, 0))));
2220 case nir_intrinsic_load_front_face
:
2221 /* The register contains 0 (front) or 1 (back), and we need to
2222 * turn it into a NIR bool where true means front.
2224 ntq_store_dest(c
, &instr
->dest
, 0,
2226 vir_uniform_ui(c
, -1),
2230 case nir_intrinsic_load_instance_id
:
2231 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->iid
));
2234 case nir_intrinsic_load_vertex_id
:
2235 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->vid
));
2238 case nir_intrinsic_load_tlb_color_v3d
:
2239 vir_emit_tlb_color_read(c
, instr
);
2242 case nir_intrinsic_load_input
:
2243 ntq_emit_load_input(c
, instr
);
2246 case nir_intrinsic_store_tlb_sample_color_v3d
:
2247 ntq_emit_per_sample_color_write(c
, instr
);
2250 case nir_intrinsic_store_output
:
2251 ntq_emit_store_output(c
, instr
);
2254 case nir_intrinsic_image_size
:
2255 ntq_emit_image_size(c
, instr
);
2258 case nir_intrinsic_discard
:
2259 if (vir_in_nonuniform_control_flow(c
)) {
2260 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2262 vir_set_cond(vir_SETMSF_dest(c
, vir_nop_reg(),
2263 vir_uniform_ui(c
, 0)),
2266 vir_SETMSF_dest(c
, vir_nop_reg(),
2267 vir_uniform_ui(c
, 0));
2271 case nir_intrinsic_discard_if
: {
2272 enum v3d_qpu_cond cond
= ntq_emit_bool_to_cond(c
, instr
->src
[0]);
2274 if (vir_in_nonuniform_control_flow(c
)) {
2275 struct qinst
*exec_flag
= vir_MOV_dest(c
, vir_nop_reg(),
2277 if (cond
== V3D_QPU_COND_IFA
) {
2278 vir_set_uf(exec_flag
, V3D_QPU_UF_ANDZ
);
2280 vir_set_uf(exec_flag
, V3D_QPU_UF_NORNZ
);
2281 cond
= V3D_QPU_COND_IFA
;
2285 vir_set_cond(vir_SETMSF_dest(c
, vir_nop_reg(),
2286 vir_uniform_ui(c
, 0)), cond
);
2291 case nir_intrinsic_memory_barrier
:
2292 case nir_intrinsic_memory_barrier_buffer
:
2293 case nir_intrinsic_memory_barrier_image
:
2294 case nir_intrinsic_memory_barrier_shared
:
2295 case nir_intrinsic_memory_barrier_tcs_patch
:
2296 case nir_intrinsic_group_memory_barrier
:
2297 /* We don't do any instruction scheduling of these NIR
2298 * instructions between each other, so we just need to make
2299 * sure that the TMU operations before the barrier are flushed
2300 * before the ones after the barrier. That is currently
2301 * handled by having a THRSW in each of them and a LDTMU
2302 * series or a TMUWT after.
2306 case nir_intrinsic_control_barrier
:
2307 /* Emit a TSY op to get all invocations in the workgroup
2308 * (actually supergroup) to block until the last invocation
2309 * reaches the TSY op.
2311 if (c
->devinfo
->ver
>= 42) {
2312 vir_BARRIERID_dest(c
, vir_reg(QFILE_MAGIC
,
2313 V3D_QPU_WADDR_SYNCB
));
2315 struct qinst
*sync
=
2316 vir_BARRIERID_dest(c
,
2317 vir_reg(QFILE_MAGIC
,
2318 V3D_QPU_WADDR_SYNCU
));
2320 vir_get_uniform_index(c
, QUNIFORM_CONSTANT
,
2322 V3D_TSY_WAIT_INC_CHECK
);
2326 /* The blocking of a TSY op only happens at the next thread
2327 * switch. No texturing may be outstanding at the time of a
2328 * TSY blocking operation.
2333 case nir_intrinsic_load_num_work_groups
:
2334 for (int i
= 0; i
< 3; i
++) {
2335 ntq_store_dest(c
, &instr
->dest
, i
,
2336 vir_uniform(c
, QUNIFORM_NUM_WORK_GROUPS
,
2341 case nir_intrinsic_load_local_invocation_index
:
2342 ntq_store_dest(c
, &instr
->dest
, 0,
2343 vir_SHR(c
, c
->cs_payload
[1],
2344 vir_uniform_ui(c
, 32 - c
->local_invocation_index_bits
)));
2347 case nir_intrinsic_load_work_group_id
:
2348 ntq_store_dest(c
, &instr
->dest
, 0,
2349 vir_AND(c
, c
->cs_payload
[0],
2350 vir_uniform_ui(c
, 0xffff)));
2351 ntq_store_dest(c
, &instr
->dest
, 1,
2352 vir_SHR(c
, c
->cs_payload
[0],
2353 vir_uniform_ui(c
, 16)));
2354 ntq_store_dest(c
, &instr
->dest
, 2,
2355 vir_AND(c
, c
->cs_payload
[1],
2356 vir_uniform_ui(c
, 0xffff)));
2359 case nir_intrinsic_load_subgroup_id
:
2360 ntq_store_dest(c
, &instr
->dest
, 0, vir_EIDX(c
));
2363 case nir_intrinsic_load_per_vertex_input
: {
2364 /* col: vertex index, row = varying index */
2365 struct qreg col
= ntq_get_src(c
, instr
->src
[0], 0);
2366 uint32_t row_idx
= nir_intrinsic_base(instr
) * 4 +
2367 nir_intrinsic_component(instr
);
2368 for (int i
= 0; i
< instr
->num_components
; i
++) {
2369 struct qreg row
= vir_uniform_ui(c
, row_idx
++);
2370 ntq_store_dest(c
, &instr
->dest
, i
,
2371 vir_LDVPMG_IN(c
, row
, col
));
2376 case nir_intrinsic_emit_vertex
:
2377 case nir_intrinsic_end_primitive
:
2378 unreachable("Should have been lowered in v3d_nir_lower_io");
2381 case nir_intrinsic_load_primitive_id
: {
2382 /* gl_PrimitiveIdIn is written by the GBG in the first word of
2383 * VPM output header. According to docs, we should read this
2384 * using ldvpm(v,d)_in (See Table 71).
2386 ntq_store_dest(c
, &instr
->dest
, 0,
2387 vir_LDVPMV_IN(c
, vir_uniform_ui(c
, 0)));
2391 case nir_intrinsic_load_invocation_id
:
2392 ntq_store_dest(c
, &instr
->dest
, 0, vir_IID(c
));
2395 case nir_intrinsic_load_fb_layers_v3d
:
2396 ntq_store_dest(c
, &instr
->dest
, 0,
2397 vir_uniform(c
, QUNIFORM_FB_LAYERS
, 0));
2400 case nir_intrinsic_load_sample_id
:
2401 ntq_store_dest(c
, &instr
->dest
, 0, vir_SAMPID(c
));
2405 fprintf(stderr
, "Unknown intrinsic: ");
2406 nir_print_instr(&instr
->instr
, stderr
);
2407 fprintf(stderr
, "\n");
2412 /* Clears (activates) the execute flags for any channels whose jump target
2413 * matches this block.
2415 * XXX perf: Could we be using flpush/flpop somehow for our execution channel
2418 * XXX perf: For uniform control flow, we should be able to skip c->execute
2419 * handling entirely.
2422 ntq_activate_execute_for_block(struct v3d_compile
*c
)
2424 vir_set_pf(vir_XOR_dest(c
, vir_nop_reg(),
2425 c
->execute
, vir_uniform_ui(c
, c
->cur_block
->index
)),
2428 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
2432 ntq_emit_uniform_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
2434 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
2435 bool empty_else_block
=
2436 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
2437 exec_list_is_empty(&nir_else_block
->instr_list
));
2439 struct qblock
*then_block
= vir_new_block(c
);
2440 struct qblock
*after_block
= vir_new_block(c
);
2441 struct qblock
*else_block
;
2442 if (empty_else_block
)
2443 else_block
= after_block
;
2445 else_block
= vir_new_block(c
);
2447 /* Set up the flags for the IF condition (taking the THEN branch). */
2448 enum v3d_qpu_cond cond
= ntq_emit_bool_to_cond(c
, if_stmt
->condition
);
2451 vir_BRANCH(c
, cond
== V3D_QPU_COND_IFA
?
2452 V3D_QPU_BRANCH_COND_ALLNA
:
2453 V3D_QPU_BRANCH_COND_ALLA
);
2454 vir_link_blocks(c
->cur_block
, else_block
);
2455 vir_link_blocks(c
->cur_block
, then_block
);
2457 /* Process the THEN block. */
2458 vir_set_emit_block(c
, then_block
);
2459 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
2461 if (!empty_else_block
) {
2462 /* At the end of the THEN block, jump to ENDIF */
2463 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALWAYS
);
2464 vir_link_blocks(c
->cur_block
, after_block
);
2466 /* Emit the else block. */
2467 vir_set_emit_block(c
, else_block
);
2468 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
2471 vir_link_blocks(c
->cur_block
, after_block
);
2473 vir_set_emit_block(c
, after_block
);
2477 ntq_emit_nonuniform_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
2479 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
2480 bool empty_else_block
=
2481 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
2482 exec_list_is_empty(&nir_else_block
->instr_list
));
2484 struct qblock
*then_block
= vir_new_block(c
);
2485 struct qblock
*after_block
= vir_new_block(c
);
2486 struct qblock
*else_block
;
2487 if (empty_else_block
)
2488 else_block
= after_block
;
2490 else_block
= vir_new_block(c
);
2492 bool was_uniform_control_flow
= false;
2493 if (!vir_in_nonuniform_control_flow(c
)) {
2494 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
2495 was_uniform_control_flow
= true;
2498 /* Set up the flags for the IF condition (taking the THEN branch). */
2499 enum v3d_qpu_cond cond
= ntq_emit_bool_to_cond(c
, if_stmt
->condition
);
2501 /* Update the flags+cond to mean "Taking the ELSE branch (!cond) and
2502 * was previously active (execute Z) for updating the exec flags.
2504 if (was_uniform_control_flow
) {
2505 cond
= v3d_qpu_cond_invert(cond
);
2507 struct qinst
*inst
= vir_MOV_dest(c
, vir_nop_reg(), c
->execute
);
2508 if (cond
== V3D_QPU_COND_IFA
) {
2509 vir_set_uf(inst
, V3D_QPU_UF_NORNZ
);
2511 vir_set_uf(inst
, V3D_QPU_UF_ANDZ
);
2512 cond
= V3D_QPU_COND_IFA
;
2516 vir_MOV_cond(c
, cond
,
2518 vir_uniform_ui(c
, else_block
->index
));
2520 /* Jump to ELSE if nothing is active for THEN, otherwise fall
2523 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
), V3D_QPU_PF_PUSHZ
);
2524 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLNA
);
2525 vir_link_blocks(c
->cur_block
, else_block
);
2526 vir_link_blocks(c
->cur_block
, then_block
);
2528 /* Process the THEN block. */
2529 vir_set_emit_block(c
, then_block
);
2530 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
2532 if (!empty_else_block
) {
2533 /* Handle the end of the THEN block. First, all currently
2534 * active channels update their execute flags to point to
2537 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2539 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2540 vir_uniform_ui(c
, after_block
->index
));
2542 /* If everything points at ENDIF, then jump there immediately. */
2543 vir_set_pf(vir_XOR_dest(c
, vir_nop_reg(),
2545 vir_uniform_ui(c
, after_block
->index
)),
2547 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLA
);
2548 vir_link_blocks(c
->cur_block
, after_block
);
2549 vir_link_blocks(c
->cur_block
, else_block
);
2551 vir_set_emit_block(c
, else_block
);
2552 ntq_activate_execute_for_block(c
);
2553 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
2556 vir_link_blocks(c
->cur_block
, after_block
);
2558 vir_set_emit_block(c
, after_block
);
2559 if (was_uniform_control_flow
)
2560 c
->execute
= c
->undef
;
2562 ntq_activate_execute_for_block(c
);
2566 ntq_emit_if(struct v3d_compile
*c
, nir_if
*nif
)
2568 bool was_in_control_flow
= c
->in_control_flow
;
2569 c
->in_control_flow
= true;
2570 if (!vir_in_nonuniform_control_flow(c
) &&
2571 nir_src_is_dynamically_uniform(nif
->condition
)) {
2572 ntq_emit_uniform_if(c
, nif
);
2574 ntq_emit_nonuniform_if(c
, nif
);
2576 c
->in_control_flow
= was_in_control_flow
;
2580 ntq_emit_jump(struct v3d_compile
*c
, nir_jump_instr
*jump
)
2582 switch (jump
->type
) {
2583 case nir_jump_break
:
2584 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2586 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2587 vir_uniform_ui(c
, c
->loop_break_block
->index
));
2590 case nir_jump_continue
:
2591 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2593 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2594 vir_uniform_ui(c
, c
->loop_cont_block
->index
));
2597 case nir_jump_return
:
2598 unreachable("All returns shouold be lowered\n");
2602 case nir_jump_goto_if
:
2603 unreachable("not supported\n");
2609 ntq_emit_instr(struct v3d_compile
*c
, nir_instr
*instr
)
2611 switch (instr
->type
) {
2612 case nir_instr_type_alu
:
2613 ntq_emit_alu(c
, nir_instr_as_alu(instr
));
2616 case nir_instr_type_intrinsic
:
2617 ntq_emit_intrinsic(c
, nir_instr_as_intrinsic(instr
));
2620 case nir_instr_type_load_const
:
2621 ntq_emit_load_const(c
, nir_instr_as_load_const(instr
));
2624 case nir_instr_type_ssa_undef
:
2625 ntq_emit_ssa_undef(c
, nir_instr_as_ssa_undef(instr
));
2628 case nir_instr_type_tex
:
2629 ntq_emit_tex(c
, nir_instr_as_tex(instr
));
2632 case nir_instr_type_jump
:
2633 ntq_emit_jump(c
, nir_instr_as_jump(instr
));
2637 fprintf(stderr
, "Unknown NIR instr type: ");
2638 nir_print_instr(instr
, stderr
);
2639 fprintf(stderr
, "\n");
2645 ntq_emit_block(struct v3d_compile
*c
, nir_block
*block
)
2647 nir_foreach_instr(instr
, block
) {
2648 ntq_emit_instr(c
, instr
);
2652 static void ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
2655 ntq_emit_loop(struct v3d_compile
*c
, nir_loop
*loop
)
2657 bool was_in_control_flow
= c
->in_control_flow
;
2658 c
->in_control_flow
= true;
2660 bool was_uniform_control_flow
= false;
2661 if (!vir_in_nonuniform_control_flow(c
)) {
2662 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
2663 was_uniform_control_flow
= true;
2666 struct qblock
*save_loop_cont_block
= c
->loop_cont_block
;
2667 struct qblock
*save_loop_break_block
= c
->loop_break_block
;
2669 c
->loop_cont_block
= vir_new_block(c
);
2670 c
->loop_break_block
= vir_new_block(c
);
2672 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
2673 vir_set_emit_block(c
, c
->loop_cont_block
);
2674 ntq_activate_execute_for_block(c
);
2676 ntq_emit_cf_list(c
, &loop
->body
);
2678 /* Re-enable any previous continues now, so our ANYA check below
2681 * XXX: Use the .ORZ flags update, instead.
2683 vir_set_pf(vir_XOR_dest(c
,
2686 vir_uniform_ui(c
, c
->loop_cont_block
->index
)),
2688 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
2690 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
), V3D_QPU_PF_PUSHZ
);
2692 struct qinst
*branch
= vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ANYA
);
2693 /* Pixels that were not dispatched or have been discarded should not
2694 * contribute to looping again.
2696 branch
->qpu
.branch
.msfign
= V3D_QPU_MSFIGN_P
;
2697 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
2698 vir_link_blocks(c
->cur_block
, c
->loop_break_block
);
2700 vir_set_emit_block(c
, c
->loop_break_block
);
2701 if (was_uniform_control_flow
)
2702 c
->execute
= c
->undef
;
2704 ntq_activate_execute_for_block(c
);
2706 c
->loop_break_block
= save_loop_break_block
;
2707 c
->loop_cont_block
= save_loop_cont_block
;
2711 c
->in_control_flow
= was_in_control_flow
;
2715 ntq_emit_function(struct v3d_compile
*c
, nir_function_impl
*func
)
2717 fprintf(stderr
, "FUNCTIONS not handled.\n");
2722 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
)
2724 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2725 switch (node
->type
) {
2726 case nir_cf_node_block
:
2727 ntq_emit_block(c
, nir_cf_node_as_block(node
));
2730 case nir_cf_node_if
:
2731 ntq_emit_if(c
, nir_cf_node_as_if(node
));
2734 case nir_cf_node_loop
:
2735 ntq_emit_loop(c
, nir_cf_node_as_loop(node
));
2738 case nir_cf_node_function
:
2739 ntq_emit_function(c
, nir_cf_node_as_function(node
));
2743 fprintf(stderr
, "Unknown NIR node type\n");
2750 ntq_emit_impl(struct v3d_compile
*c
, nir_function_impl
*impl
)
2752 ntq_setup_registers(c
, &impl
->registers
);
2753 ntq_emit_cf_list(c
, &impl
->body
);
2757 nir_to_vir(struct v3d_compile
*c
)
2759 switch (c
->s
->info
.stage
) {
2760 case MESA_SHADER_FRAGMENT
:
2761 c
->payload_w
= vir_MOV(c
, vir_reg(QFILE_REG
, 0));
2762 c
->payload_w_centroid
= vir_MOV(c
, vir_reg(QFILE_REG
, 1));
2763 c
->payload_z
= vir_MOV(c
, vir_reg(QFILE_REG
, 2));
2765 /* V3D 4.x can disable implicit point coordinate varyings if
2766 * they are not used.
2768 if (c
->fs_key
->is_points
&&
2769 (c
->devinfo
->ver
< 40 || program_reads_point_coord(c
))) {
2770 c
->point_x
= emit_fragment_varying(c
, NULL
, 0, 0);
2771 c
->point_y
= emit_fragment_varying(c
, NULL
, 0, 0);
2772 c
->uses_implicit_point_line_varyings
= true;
2773 } else if (c
->fs_key
->is_lines
&&
2774 (c
->devinfo
->ver
< 40 ||
2775 (c
->s
->info
.system_values_read
&
2776 BITFIELD64_BIT(SYSTEM_VALUE_LINE_COORD
)))) {
2777 c
->line_x
= emit_fragment_varying(c
, NULL
, 0, 0);
2778 c
->uses_implicit_point_line_varyings
= true;
2781 case MESA_SHADER_COMPUTE
:
2782 /* Set up the TSO for barriers, assuming we do some. */
2783 if (c
->devinfo
->ver
< 42) {
2784 vir_BARRIERID_dest(c
, vir_reg(QFILE_MAGIC
,
2785 V3D_QPU_WADDR_SYNC
));
2788 c
->cs_payload
[0] = vir_MOV(c
, vir_reg(QFILE_REG
, 0));
2789 c
->cs_payload
[1] = vir_MOV(c
, vir_reg(QFILE_REG
, 2));
2791 /* Set up the division between gl_LocalInvocationIndex and
2792 * wg_in_mem in the payload reg.
2794 int wg_size
= (c
->s
->info
.cs
.local_size
[0] *
2795 c
->s
->info
.cs
.local_size
[1] *
2796 c
->s
->info
.cs
.local_size
[2]);
2797 c
->local_invocation_index_bits
=
2798 ffs(util_next_power_of_two(MAX2(wg_size
, 64))) - 1;
2799 assert(c
->local_invocation_index_bits
<= 8);
2801 if (c
->s
->info
.cs
.shared_size
) {
2802 struct qreg wg_in_mem
= vir_SHR(c
, c
->cs_payload
[1],
2803 vir_uniform_ui(c
, 16));
2804 if (c
->s
->info
.cs
.local_size
[0] != 1 ||
2805 c
->s
->info
.cs
.local_size
[1] != 1 ||
2806 c
->s
->info
.cs
.local_size
[2] != 1) {
2808 c
->local_invocation_index_bits
);
2809 int wg_mask
= (1 << wg_bits
) - 1;
2810 wg_in_mem
= vir_AND(c
, wg_in_mem
,
2811 vir_uniform_ui(c
, wg_mask
));
2813 struct qreg shared_per_wg
=
2814 vir_uniform_ui(c
, c
->s
->info
.cs
.shared_size
);
2816 c
->cs_shared_offset
=
2818 vir_uniform(c
, QUNIFORM_SHARED_OFFSET
,0),
2819 vir_UMUL(c
, wg_in_mem
, shared_per_wg
));
2826 if (c
->s
->scratch_size
) {
2827 v3d_setup_spill_base(c
);
2828 c
->spill_size
+= V3D_CHANNELS
* c
->s
->scratch_size
;
2831 switch (c
->s
->info
.stage
) {
2832 case MESA_SHADER_VERTEX
:
2833 ntq_setup_vs_inputs(c
);
2835 case MESA_SHADER_GEOMETRY
:
2836 ntq_setup_gs_inputs(c
);
2838 case MESA_SHADER_FRAGMENT
:
2839 ntq_setup_fs_inputs(c
);
2841 case MESA_SHADER_COMPUTE
:
2844 unreachable("unsupported shader stage");
2847 ntq_setup_outputs(c
);
2849 /* Find the main function and emit the body. */
2850 nir_foreach_function(function
, c
->s
) {
2851 assert(strcmp(function
->name
, "main") == 0);
2852 assert(function
->impl
);
2853 ntq_emit_impl(c
, function
->impl
);
2857 const nir_shader_compiler_options v3d_nir_options
= {
2858 .lower_all_io_to_temps
= true,
2859 .lower_extract_byte
= true,
2860 .lower_extract_word
= true,
2861 .lower_bitfield_insert_to_shifts
= true,
2862 .lower_bitfield_extract_to_shifts
= true,
2863 .lower_bitfield_reverse
= true,
2864 .lower_bit_count
= true,
2865 .lower_cs_local_id_from_index
= true,
2866 .lower_ffract
= true,
2868 .lower_pack_unorm_2x16
= true,
2869 .lower_pack_snorm_2x16
= true,
2870 .lower_pack_unorm_4x8
= true,
2871 .lower_pack_snorm_4x8
= true,
2872 .lower_unpack_unorm_4x8
= true,
2873 .lower_unpack_snorm_4x8
= true,
2874 .lower_pack_half_2x16
= true,
2875 .lower_unpack_half_2x16
= true,
2877 .lower_find_lsb
= true,
2879 .lower_flrp32
= true,
2882 .lower_fsqrt
= true,
2883 .lower_ifind_msb
= true,
2884 .lower_isign
= true,
2885 .lower_ldexp
= true,
2886 .lower_mul_high
= true,
2887 .lower_wpos_pntc
= true,
2888 .lower_rotate
= true,
2889 .lower_to_scalar
= true,
2893 * When demoting a shader down to single-threaded, removes the THRSW
2894 * instructions (one will still be inserted at v3d_vir_to_qpu() for the
2898 vir_remove_thrsw(struct v3d_compile
*c
)
2900 vir_for_each_block(block
, c
) {
2901 vir_for_each_inst_safe(inst
, block
) {
2902 if (inst
->qpu
.sig
.thrsw
)
2903 vir_remove_instruction(c
, inst
);
2907 c
->last_thrsw
= NULL
;
2911 vir_emit_last_thrsw(struct v3d_compile
*c
)
2913 /* On V3D before 4.1, we need a TMU op to be outstanding when thread
2914 * switching, so disable threads if we didn't do any TMU ops (each of
2915 * which would have emitted a THRSW).
2917 if (!c
->last_thrsw_at_top_level
&& c
->devinfo
->ver
< 41) {
2920 vir_remove_thrsw(c
);
2924 /* If we're threaded and the last THRSW was in conditional code, then
2925 * we need to emit another one so that we can flag it as the last
2928 if (c
->last_thrsw
&& !c
->last_thrsw_at_top_level
) {
2929 assert(c
->devinfo
->ver
>= 41);
2933 /* If we're threaded, then we need to mark the last THRSW instruction
2934 * so we can emit a pair of them at QPU emit time.
2936 * For V3D 4.x, we can spawn the non-fragment shaders already in the
2937 * post-last-THRSW state, so we can skip this.
2939 if (!c
->last_thrsw
&& c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2940 assert(c
->devinfo
->ver
>= 41);
2945 c
->last_thrsw
->is_last_thrsw
= true;
2948 /* There's a flag in the shader for "center W is needed for reasons other than
2949 * non-centroid varyings", so we just walk the program after VIR optimization
2950 * to see if it's used. It should be harmless to set even if we only use
2951 * center W for varyings.
2954 vir_check_payload_w(struct v3d_compile
*c
)
2956 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
)
2959 vir_for_each_inst_inorder(inst
, c
) {
2960 for (int i
= 0; i
< vir_get_nsrc(inst
); i
++) {
2961 if (inst
->src
[i
].file
== QFILE_REG
&&
2962 inst
->src
[i
].index
== 0) {
2963 c
->uses_center_w
= true;
2972 v3d_nir_to_vir(struct v3d_compile
*c
)
2974 if (V3D_DEBUG
& (V3D_DEBUG_NIR
|
2975 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2976 fprintf(stderr
, "%s prog %d/%d NIR:\n",
2977 vir_get_stage_name(c
),
2978 c
->program_id
, c
->variant_id
);
2979 nir_print_shader(c
->s
, stderr
);
2984 /* Emit the last THRSW before STVPM and TLB writes. */
2985 vir_emit_last_thrsw(c
);
2987 switch (c
->s
->info
.stage
) {
2988 case MESA_SHADER_FRAGMENT
:
2991 case MESA_SHADER_GEOMETRY
:
2994 case MESA_SHADER_VERTEX
:
2997 case MESA_SHADER_COMPUTE
:
3000 unreachable("bad stage");
3003 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
3004 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
3005 fprintf(stderr
, "%s prog %d/%d pre-opt VIR:\n",
3006 vir_get_stage_name(c
),
3007 c
->program_id
, c
->variant_id
);
3009 fprintf(stderr
, "\n");
3014 vir_check_payload_w(c
);
3016 /* XXX perf: On VC4, we do a VIR-level instruction scheduling here.
3017 * We used that on that platform to pipeline TMU writes and reduce the
3018 * number of thread switches, as well as try (mostly successfully) to
3019 * reduce maximum register pressure to allow more threads. We should
3020 * do something of that sort for V3D -- either instruction scheduling
3021 * here, or delay the the THRSW and LDTMUs from our texture
3022 * instructions until the results are needed.
3025 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
3026 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
3027 fprintf(stderr
, "%s prog %d/%d VIR:\n",
3028 vir_get_stage_name(c
),
3029 c
->program_id
, c
->variant_id
);
3031 fprintf(stderr
, "\n");
3034 /* Attempt to allocate registers for the temporaries. If we fail,
3035 * reduce thread count and try again.
3037 int min_threads
= (c
->devinfo
->ver
>= 41) ? 2 : 1;
3038 struct qpu_reg
*temp_registers
;
3041 temp_registers
= v3d_register_allocate(c
, &spilled
);
3048 if (c
->threads
== min_threads
) {
3049 if (c
->fallback_scheduler
) {
3051 "Failed to register allocate at %d "
3056 c
->compilation_result
=
3057 V3D_COMPILATION_FAILED_REGISTER_ALLOCATION
;
3063 if (c
->threads
== 1)
3064 vir_remove_thrsw(c
);
3068 (V3D_DEBUG
& (V3D_DEBUG_VIR
|
3069 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
)))) {
3070 fprintf(stderr
, "%s prog %d/%d spilled VIR:\n",
3071 vir_get_stage_name(c
),
3072 c
->program_id
, c
->variant_id
);
3074 fprintf(stderr
, "\n");
3077 v3d_vir_to_qpu(c
, temp_registers
);