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 config
= (0xffffff00 |
344 GENERAL_TMU_LOOKUP_PER_PIXEL
);
345 if (num_components
== 1) {
346 config
|= GENERAL_TMU_LOOKUP_TYPE_32BIT_UI
;
348 config
|= GENERAL_TMU_LOOKUP_TYPE_VEC2
+
352 if (vir_in_nonuniform_control_flow(c
)) {
353 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
359 tmua
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUA
);
361 tmua
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_TMUAU
);
365 struct qreg offset
= base_offset
;
366 if (const_offset
!= 0) {
367 offset
= vir_ADD(c
, offset
,
368 vir_uniform_ui(c
, const_offset
));
371 ntq_get_src(c
, instr
->src
[offset_src
], 0);
372 tmu
= vir_ADD_dest(c
, tmua
, offset
, data
);
374 if (const_offset
!= 0) {
375 tmu
= vir_ADD_dest(c
, tmua
, base_offset
,
376 vir_uniform_ui(c
, const_offset
));
378 tmu
= vir_MOV_dest(c
, tmua
, base_offset
);
384 vir_get_uniform_index(c
, QUNIFORM_CONSTANT
,
388 if (vir_in_nonuniform_control_flow(c
))
389 vir_set_cond(tmu
, V3D_QPU_COND_IFA
);
393 /* Read the result, or wait for the TMU op to complete. */
394 for (int i
= 0; i
< nir_intrinsic_dest_components(instr
); i
++) {
395 ntq_store_dest(c
, &instr
->dest
, i
,
396 vir_MOV(c
, vir_LDTMU(c
)));
399 if (nir_intrinsic_dest_components(instr
) == 0)
401 } while (is_store
&& writemask
!= 0);
405 ntq_init_ssa_def(struct v3d_compile
*c
, nir_ssa_def
*def
)
407 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
408 def
->num_components
);
409 _mesa_hash_table_insert(c
->def_ht
, def
, qregs
);
414 is_ld_signal(const struct v3d_qpu_sig
*sig
)
416 return (sig
->ldunif
||
428 * This function is responsible for getting VIR results into the associated
429 * storage for a NIR instruction.
431 * If it's a NIR SSA def, then we just set the associated hash table entry to
434 * If it's a NIR reg, then we need to update the existing qreg assigned to the
435 * NIR destination with the incoming value. To do that without introducing
436 * new MOVs, we require that the incoming qreg either be a uniform, or be
437 * SSA-defined by the previous VIR instruction in the block and rewritable by
438 * this function. That lets us sneak ahead and insert the SF flag beforehand
439 * (knowing that the previous instruction doesn't depend on flags) and rewrite
440 * its destination to be the NIR reg's destination
443 ntq_store_dest(struct v3d_compile
*c
, nir_dest
*dest
, int chan
,
446 struct qinst
*last_inst
= NULL
;
447 if (!list_is_empty(&c
->cur_block
->instructions
))
448 last_inst
= (struct qinst
*)c
->cur_block
->instructions
.prev
;
450 assert((result
.file
== QFILE_TEMP
&&
451 last_inst
&& last_inst
== c
->defs
[result
.index
]));
454 assert(chan
< dest
->ssa
.num_components
);
457 struct hash_entry
*entry
=
458 _mesa_hash_table_search(c
->def_ht
, &dest
->ssa
);
463 qregs
= ntq_init_ssa_def(c
, &dest
->ssa
);
465 qregs
[chan
] = result
;
467 nir_register
*reg
= dest
->reg
.reg
;
468 assert(dest
->reg
.base_offset
== 0);
469 assert(reg
->num_array_elems
== 0);
470 struct hash_entry
*entry
=
471 _mesa_hash_table_search(c
->def_ht
, reg
);
472 struct qreg
*qregs
= entry
->data
;
474 /* If the previous instruction can't be predicated for
475 * the store into the nir_register, then emit a MOV
478 if (vir_in_nonuniform_control_flow(c
) &&
479 is_ld_signal(&c
->defs
[last_inst
->dst
.index
]->qpu
.sig
)) {
480 result
= vir_MOV(c
, result
);
481 last_inst
= c
->defs
[result
.index
];
484 /* We know they're both temps, so just rewrite index. */
485 c
->defs
[last_inst
->dst
.index
] = NULL
;
486 last_inst
->dst
.index
= qregs
[chan
].index
;
488 /* If we're in control flow, then make this update of the reg
489 * conditional on the execution mask.
491 if (vir_in_nonuniform_control_flow(c
)) {
492 last_inst
->dst
.index
= qregs
[chan
].index
;
494 /* Set the flags to the current exec mask.
496 c
->cursor
= vir_before_inst(last_inst
);
497 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
499 c
->cursor
= vir_after_inst(last_inst
);
501 vir_set_cond(last_inst
, V3D_QPU_COND_IFA
);
507 ntq_get_src(struct v3d_compile
*c
, nir_src src
, int i
)
509 struct hash_entry
*entry
;
511 entry
= _mesa_hash_table_search(c
->def_ht
, src
.ssa
);
512 assert(i
< src
.ssa
->num_components
);
514 nir_register
*reg
= src
.reg
.reg
;
515 entry
= _mesa_hash_table_search(c
->def_ht
, reg
);
516 assert(reg
->num_array_elems
== 0);
517 assert(src
.reg
.base_offset
== 0);
518 assert(i
< reg
->num_components
);
521 struct qreg
*qregs
= entry
->data
;
526 ntq_get_alu_src(struct v3d_compile
*c
, nir_alu_instr
*instr
,
529 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
530 unsigned chan
= ffs(instr
->dest
.write_mask
) - 1;
531 struct qreg r
= ntq_get_src(c
, instr
->src
[src
].src
,
532 instr
->src
[src
].swizzle
[chan
]);
534 assert(!instr
->src
[src
].abs
);
535 assert(!instr
->src
[src
].negate
);
541 ntq_minify(struct v3d_compile
*c
, struct qreg size
, struct qreg level
)
543 return vir_MAX(c
, vir_SHR(c
, size
, level
), vir_uniform_ui(c
, 1));
547 ntq_emit_txs(struct v3d_compile
*c
, nir_tex_instr
*instr
)
549 unsigned unit
= instr
->texture_index
;
550 int lod_index
= nir_tex_instr_src_index(instr
, nir_tex_src_lod
);
551 int dest_size
= nir_tex_instr_dest_size(instr
);
553 struct qreg lod
= c
->undef
;
555 lod
= ntq_get_src(c
, instr
->src
[lod_index
].src
, 0);
557 for (int i
= 0; i
< dest_size
; i
++) {
559 enum quniform_contents contents
;
561 if (instr
->is_array
&& i
== dest_size
- 1)
562 contents
= QUNIFORM_TEXTURE_ARRAY_SIZE
;
564 contents
= QUNIFORM_TEXTURE_WIDTH
+ i
;
566 struct qreg size
= vir_uniform(c
, contents
, unit
);
568 switch (instr
->sampler_dim
) {
569 case GLSL_SAMPLER_DIM_1D
:
570 case GLSL_SAMPLER_DIM_2D
:
571 case GLSL_SAMPLER_DIM_MS
:
572 case GLSL_SAMPLER_DIM_3D
:
573 case GLSL_SAMPLER_DIM_CUBE
:
574 /* Don't minify the array size. */
575 if (!(instr
->is_array
&& i
== dest_size
- 1)) {
576 size
= ntq_minify(c
, size
, lod
);
580 case GLSL_SAMPLER_DIM_RECT
:
581 /* There's no LOD field for rects */
585 unreachable("Bad sampler type");
588 ntq_store_dest(c
, &instr
->dest
, i
, size
);
593 ntq_emit_tex(struct v3d_compile
*c
, nir_tex_instr
*instr
)
595 unsigned unit
= instr
->texture_index
;
597 /* Since each texture sampling op requires uploading uniforms to
598 * reference the texture, there's no HW support for texture size and
599 * you just upload uniforms containing the size.
602 case nir_texop_query_levels
:
603 ntq_store_dest(c
, &instr
->dest
, 0,
604 vir_uniform(c
, QUNIFORM_TEXTURE_LEVELS
, unit
));
607 ntq_emit_txs(c
, instr
);
613 if (c
->devinfo
->ver
>= 40)
614 v3d40_vir_emit_tex(c
, instr
);
616 v3d33_vir_emit_tex(c
, instr
);
620 ntq_fsincos(struct v3d_compile
*c
, struct qreg src
, bool is_cos
)
622 struct qreg input
= vir_FMUL(c
, src
, vir_uniform_f(c
, 1.0f
/ M_PI
));
624 input
= vir_FADD(c
, input
, vir_uniform_f(c
, 0.5));
626 struct qreg periods
= vir_FROUND(c
, input
);
627 struct qreg sin_output
= vir_SIN(c
, vir_FSUB(c
, input
, periods
));
628 return vir_XOR(c
, sin_output
, vir_SHL(c
,
629 vir_FTOIN(c
, periods
),
630 vir_uniform_ui(c
, -1)));
634 ntq_fsign(struct v3d_compile
*c
, struct qreg src
)
636 struct qreg t
= vir_get_temp(c
);
638 vir_MOV_dest(c
, t
, vir_uniform_f(c
, 0.0));
639 vir_set_pf(vir_FMOV_dest(c
, vir_nop_reg(), src
), V3D_QPU_PF_PUSHZ
);
640 vir_MOV_cond(c
, V3D_QPU_COND_IFNA
, t
, vir_uniform_f(c
, 1.0));
641 vir_set_pf(vir_FMOV_dest(c
, vir_nop_reg(), src
), V3D_QPU_PF_PUSHN
);
642 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, t
, vir_uniform_f(c
, -1.0));
643 return vir_MOV(c
, t
);
647 emit_fragcoord_input(struct v3d_compile
*c
, int attr
)
649 c
->inputs
[attr
* 4 + 0] = vir_FXCD(c
);
650 c
->inputs
[attr
* 4 + 1] = vir_FYCD(c
);
651 c
->inputs
[attr
* 4 + 2] = c
->payload_z
;
652 c
->inputs
[attr
* 4 + 3] = vir_RECIP(c
, c
->payload_w
);
656 emit_fragment_varying(struct v3d_compile
*c
, nir_variable
*var
,
657 uint8_t swizzle
, int array_index
)
659 struct qreg r3
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R3
);
660 struct qreg r5
= vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_R5
);
663 if (c
->devinfo
->ver
>= 41) {
664 struct qinst
*ldvary
= vir_add_inst(V3D_QPU_A_NOP
, c
->undef
,
666 ldvary
->qpu
.sig
.ldvary
= true;
667 vary
= vir_emit_def(c
, ldvary
);
669 vir_NOP(c
)->qpu
.sig
.ldvary
= true;
673 /* For gl_PointCoord input or distance along a line, we'll be called
674 * with no nir_variable, and we don't count toward VPM size so we
675 * don't track an input slot.
678 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
681 int i
= c
->num_inputs
++;
683 v3d_slot_from_slot_and_component(var
->data
.location
+
684 array_index
, swizzle
);
686 switch (var
->data
.interpolation
) {
687 case INTERP_MODE_NONE
:
688 /* If a gl_FrontColor or gl_BackColor input has no interp
689 * qualifier, then if we're using glShadeModel(GL_FLAT) it
690 * needs to be flat shaded.
692 switch (var
->data
.location
+ array_index
) {
693 case VARYING_SLOT_COL0
:
694 case VARYING_SLOT_COL1
:
695 case VARYING_SLOT_BFC0
:
696 case VARYING_SLOT_BFC1
:
697 if (c
->fs_key
->shade_model_flat
) {
698 BITSET_SET(c
->flat_shade_flags
, i
);
699 vir_MOV_dest(c
, c
->undef
, vary
);
700 return vir_MOV(c
, r5
);
702 return vir_FADD(c
, vir_FMUL(c
, vary
,
709 case INTERP_MODE_SMOOTH
:
710 if (var
->data
.centroid
) {
711 BITSET_SET(c
->centroid_flags
, i
);
712 return vir_FADD(c
, vir_FMUL(c
, vary
,
713 c
->payload_w_centroid
), r5
);
715 return vir_FADD(c
, vir_FMUL(c
, vary
, c
->payload_w
), r5
);
717 case INTERP_MODE_NOPERSPECTIVE
:
718 BITSET_SET(c
->noperspective_flags
, i
);
719 return vir_FADD(c
, vir_MOV(c
, vary
), r5
);
720 case INTERP_MODE_FLAT
:
721 BITSET_SET(c
->flat_shade_flags
, i
);
722 vir_MOV_dest(c
, c
->undef
, vary
);
723 return vir_MOV(c
, r5
);
725 unreachable("Bad interp mode");
730 emit_fragment_input(struct v3d_compile
*c
, int attr
, nir_variable
*var
,
733 for (int i
= 0; i
< glsl_get_vector_elements(var
->type
); i
++) {
734 int chan
= var
->data
.location_frac
+ i
;
735 c
->inputs
[attr
* 4 + chan
] =
736 emit_fragment_varying(c
, var
, chan
, array_index
);
741 add_output(struct v3d_compile
*c
,
742 uint32_t decl_offset
,
746 uint32_t old_array_size
= c
->outputs_array_size
;
747 resize_qreg_array(c
, &c
->outputs
, &c
->outputs_array_size
,
750 if (old_array_size
!= c
->outputs_array_size
) {
751 c
->output_slots
= reralloc(c
,
753 struct v3d_varying_slot
,
754 c
->outputs_array_size
);
757 c
->output_slots
[decl_offset
] =
758 v3d_slot_from_slot_and_component(slot
, swizzle
);
762 * If compare_instr is a valid comparison instruction, emits the
763 * compare_instr's comparison and returns the sel_instr's return value based
764 * on the compare_instr's result.
767 ntq_emit_comparison(struct v3d_compile
*c
,
768 nir_alu_instr
*compare_instr
,
769 enum v3d_qpu_cond
*out_cond
)
771 struct qreg src0
= ntq_get_alu_src(c
, compare_instr
, 0);
773 if (nir_op_infos
[compare_instr
->op
].num_inputs
> 1)
774 src1
= ntq_get_alu_src(c
, compare_instr
, 1);
775 bool cond_invert
= false;
776 struct qreg nop
= vir_nop_reg();
778 switch (compare_instr
->op
) {
781 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
784 vir_set_pf(vir_XOR_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
789 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
793 vir_set_pf(vir_XOR_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHZ
);
799 vir_set_pf(vir_FCMP_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
802 vir_set_pf(vir_MIN_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
806 vir_set_pf(vir_SUB_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHC
);
812 vir_set_pf(vir_FCMP_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHN
);
815 vir_set_pf(vir_MIN_dest(c
, nop
, src1
, src0
), V3D_QPU_PF_PUSHC
);
818 vir_set_pf(vir_SUB_dest(c
, nop
, src0
, src1
), V3D_QPU_PF_PUSHC
);
822 vir_set_pf(vir_MOV_dest(c
, nop
, src0
), V3D_QPU_PF_PUSHZ
);
827 vir_set_pf(vir_FMOV_dest(c
, nop
, src0
), V3D_QPU_PF_PUSHZ
);
835 *out_cond
= cond_invert
? V3D_QPU_COND_IFNA
: V3D_QPU_COND_IFA
;
840 /* Finds an ALU instruction that generates our src value that could
841 * (potentially) be greedily emitted in the consuming instruction.
843 static struct nir_alu_instr
*
844 ntq_get_alu_parent(nir_src src
)
846 if (!src
.is_ssa
|| src
.ssa
->parent_instr
->type
!= nir_instr_type_alu
)
848 nir_alu_instr
*instr
= nir_instr_as_alu(src
.ssa
->parent_instr
);
852 /* If the ALU instr's srcs are non-SSA, then we would have to avoid
853 * moving emission of the ALU instr down past another write of the
856 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
857 if (!instr
->src
[i
].src
.is_ssa
)
864 /* Turns a NIR bool into a condition code to predicate on. */
865 static enum v3d_qpu_cond
866 ntq_emit_bool_to_cond(struct v3d_compile
*c
, nir_src src
)
868 nir_alu_instr
*compare
= ntq_get_alu_parent(src
);
872 enum v3d_qpu_cond cond
;
873 if (ntq_emit_comparison(c
, compare
, &cond
))
877 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), ntq_get_src(c
, src
, 0)),
879 return V3D_QPU_COND_IFNA
;
883 ntq_emit_alu(struct v3d_compile
*c
, nir_alu_instr
*instr
)
885 /* This should always be lowered to ALU operations for V3D. */
886 assert(!instr
->dest
.saturate
);
888 /* Vectors are special in that they have non-scalarized writemasks,
889 * and just take the first swizzle channel for each argument in order
890 * into each writemask channel.
892 if (instr
->op
== nir_op_vec2
||
893 instr
->op
== nir_op_vec3
||
894 instr
->op
== nir_op_vec4
) {
896 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
897 srcs
[i
] = ntq_get_src(c
, instr
->src
[i
].src
,
898 instr
->src
[i
].swizzle
[0]);
899 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++)
900 ntq_store_dest(c
, &instr
->dest
.dest
, i
,
901 vir_MOV(c
, srcs
[i
]));
905 /* General case: We can just grab the one used channel per src. */
906 struct qreg src
[nir_op_infos
[instr
->op
].num_inputs
];
907 for (int i
= 0; i
< nir_op_infos
[instr
->op
].num_inputs
; i
++) {
908 src
[i
] = ntq_get_alu_src(c
, instr
, i
);
915 result
= vir_MOV(c
, src
[0]);
919 result
= vir_XOR(c
, src
[0], vir_uniform_ui(c
, 1 << 31));
922 result
= vir_NEG(c
, src
[0]);
926 result
= vir_FMUL(c
, src
[0], src
[1]);
929 result
= vir_FADD(c
, src
[0], src
[1]);
932 result
= vir_FSUB(c
, src
[0], src
[1]);
935 result
= vir_FMIN(c
, src
[0], src
[1]);
938 result
= vir_FMAX(c
, src
[0], src
[1]);
942 nir_alu_instr
*src0_alu
= ntq_get_alu_parent(instr
->src
[0].src
);
943 if (src0_alu
&& src0_alu
->op
== nir_op_fround_even
) {
944 result
= vir_FTOIN(c
, ntq_get_alu_src(c
, src0_alu
, 0));
946 result
= vir_FTOIZ(c
, src
[0]);
952 result
= vir_FTOUZ(c
, src
[0]);
955 result
= vir_ITOF(c
, src
[0]);
958 result
= vir_UTOF(c
, src
[0]);
961 result
= vir_AND(c
, src
[0], vir_uniform_f(c
, 1.0));
964 result
= vir_AND(c
, src
[0], vir_uniform_ui(c
, 1));
968 result
= vir_ADD(c
, src
[0], src
[1]);
971 result
= vir_SHR(c
, src
[0], src
[1]);
974 result
= vir_SUB(c
, src
[0], src
[1]);
977 result
= vir_ASR(c
, src
[0], src
[1]);
980 result
= vir_SHL(c
, src
[0], src
[1]);
983 result
= vir_MIN(c
, src
[0], src
[1]);
986 result
= vir_UMIN(c
, src
[0], src
[1]);
989 result
= vir_MAX(c
, src
[0], src
[1]);
992 result
= vir_UMAX(c
, src
[0], src
[1]);
995 result
= vir_AND(c
, src
[0], src
[1]);
998 result
= vir_OR(c
, src
[0], src
[1]);
1001 result
= vir_XOR(c
, src
[0], src
[1]);
1004 result
= vir_NOT(c
, src
[0]);
1007 case nir_op_ufind_msb
:
1008 result
= vir_SUB(c
, vir_uniform_ui(c
, 31), vir_CLZ(c
, src
[0]));
1012 result
= vir_UMUL(c
, src
[0], src
[1]);
1019 enum v3d_qpu_cond cond
;
1020 ASSERTED
bool ok
= ntq_emit_comparison(c
, instr
, &cond
);
1022 result
= vir_MOV(c
, vir_SEL(c
, cond
,
1023 vir_uniform_f(c
, 1.0),
1024 vir_uniform_f(c
, 0.0)));
1039 case nir_op_ult32
: {
1040 enum v3d_qpu_cond cond
;
1041 ASSERTED
bool ok
= ntq_emit_comparison(c
, instr
, &cond
);
1043 result
= vir_MOV(c
, vir_SEL(c
, cond
,
1044 vir_uniform_ui(c
, ~0),
1045 vir_uniform_ui(c
, 0)));
1049 case nir_op_b32csel
:
1052 ntq_emit_bool_to_cond(c
, instr
->src
[0].src
),
1057 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), src
[0]),
1059 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFNA
,
1064 result
= vir_RECIP(c
, src
[0]);
1067 result
= vir_RSQRT(c
, src
[0]);
1070 result
= vir_EXP(c
, src
[0]);
1073 result
= vir_LOG(c
, src
[0]);
1077 result
= vir_FCEIL(c
, src
[0]);
1080 result
= vir_FFLOOR(c
, src
[0]);
1082 case nir_op_fround_even
:
1083 result
= vir_FROUND(c
, src
[0]);
1086 result
= vir_FTRUNC(c
, src
[0]);
1090 result
= ntq_fsincos(c
, src
[0], false);
1093 result
= ntq_fsincos(c
, src
[0], true);
1097 result
= ntq_fsign(c
, src
[0]);
1101 result
= vir_FMOV(c
, src
[0]);
1102 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_ABS
);
1107 result
= vir_MAX(c
, src
[0], vir_NEG(c
, src
[0]));
1111 case nir_op_fddx_coarse
:
1112 case nir_op_fddx_fine
:
1113 result
= vir_FDX(c
, src
[0]);
1117 case nir_op_fddy_coarse
:
1118 case nir_op_fddy_fine
:
1119 result
= vir_FDY(c
, src
[0]);
1122 case nir_op_uadd_carry
:
1123 vir_set_pf(vir_ADD_dest(c
, vir_nop_reg(), src
[0], src
[1]),
1125 result
= vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
1126 vir_uniform_ui(c
, ~0),
1127 vir_uniform_ui(c
, 0)));
1130 case nir_op_pack_half_2x16_split
:
1131 result
= vir_VFPACK(c
, src
[0], src
[1]);
1134 case nir_op_unpack_half_2x16_split_x
:
1135 result
= vir_FMOV(c
, src
[0]);
1136 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_L
);
1139 case nir_op_unpack_half_2x16_split_y
:
1140 result
= vir_FMOV(c
, src
[0]);
1141 vir_set_unpack(c
->defs
[result
.index
], 0, V3D_QPU_UNPACK_H
);
1145 fprintf(stderr
, "unknown NIR ALU inst: ");
1146 nir_print_instr(&instr
->instr
, stderr
);
1147 fprintf(stderr
, "\n");
1151 /* We have a scalar result, so the instruction should only have a
1152 * single channel written to.
1154 assert(util_is_power_of_two_or_zero(instr
->dest
.write_mask
));
1155 ntq_store_dest(c
, &instr
->dest
.dest
,
1156 ffs(instr
->dest
.write_mask
) - 1, result
);
1159 /* Each TLB read/write setup (a render target or depth buffer) takes an 8-bit
1160 * specifier. They come from a register that's preloaded with 0xffffffff
1161 * (0xff gets you normal vec4 f16 RT0 writes), and when one is neaded the low
1162 * 8 bits are shifted off the bottom and 0xff shifted in from the top.
1164 #define TLB_TYPE_F16_COLOR (3 << 6)
1165 #define TLB_TYPE_I32_COLOR (1 << 6)
1166 #define TLB_TYPE_F32_COLOR (0 << 6)
1167 #define TLB_RENDER_TARGET_SHIFT 3 /* Reversed! 7 = RT 0, 0 = RT 7. */
1168 #define TLB_SAMPLE_MODE_PER_SAMPLE (0 << 2)
1169 #define TLB_SAMPLE_MODE_PER_PIXEL (1 << 2)
1170 #define TLB_F16_SWAP_HI_LO (1 << 1)
1171 #define TLB_VEC_SIZE_4_F16 (1 << 0)
1172 #define TLB_VEC_SIZE_2_F16 (0 << 0)
1173 #define TLB_VEC_SIZE_MINUS_1_SHIFT 0
1175 /* Triggers Z/Stencil testing, used when the shader state's "FS modifies Z"
1178 #define TLB_TYPE_DEPTH ((2 << 6) | (0 << 4))
1179 #define TLB_DEPTH_TYPE_INVARIANT (0 << 2) /* Unmodified sideband input used */
1180 #define TLB_DEPTH_TYPE_PER_PIXEL (1 << 2) /* QPU result used */
1181 #define TLB_V42_DEPTH_TYPE_INVARIANT (0 << 3) /* Unmodified sideband input used */
1182 #define TLB_V42_DEPTH_TYPE_PER_PIXEL (1 << 3) /* QPU result used */
1184 /* Stencil is a single 32-bit write. */
1185 #define TLB_TYPE_STENCIL_ALPHA ((2 << 6) | (1 << 4))
1188 vir_emit_tlb_color_write(struct v3d_compile
*c
, unsigned rt
)
1190 if (!(c
->fs_key
->cbufs
& (1 << rt
)) || !c
->output_color_var
[rt
])
1193 struct qreg tlb_reg
= vir_magic_reg(V3D_QPU_WADDR_TLB
);
1194 struct qreg tlbu_reg
= vir_magic_reg(V3D_QPU_WADDR_TLBU
);
1196 nir_variable
*var
= c
->output_color_var
[rt
];
1197 int num_components
= glsl_get_vector_elements(var
->type
);
1198 uint32_t conf
= 0xffffff00;
1201 conf
|= c
->msaa_per_sample_output
? TLB_SAMPLE_MODE_PER_SAMPLE
:
1202 TLB_SAMPLE_MODE_PER_PIXEL
;
1203 conf
|= (7 - rt
) << TLB_RENDER_TARGET_SHIFT
;
1205 if (c
->fs_key
->swap_color_rb
& (1 << rt
))
1206 num_components
= MAX2(num_components
, 3);
1207 assert(num_components
!= 0);
1209 enum glsl_base_type type
= glsl_get_base_type(var
->type
);
1210 bool is_int_format
= type
== GLSL_TYPE_INT
|| type
== GLSL_TYPE_UINT
;
1211 bool is_32b_tlb_format
= is_int_format
||
1212 (c
->fs_key
->f32_color_rb
& (1 << rt
));
1214 if (is_int_format
) {
1215 /* The F32 vs I32 distinction was dropped in 4.2. */
1216 if (c
->devinfo
->ver
< 42)
1217 conf
|= TLB_TYPE_I32_COLOR
;
1219 conf
|= TLB_TYPE_F32_COLOR
;
1220 conf
|= ((num_components
- 1) << TLB_VEC_SIZE_MINUS_1_SHIFT
);
1222 if (c
->fs_key
->f32_color_rb
& (1 << rt
)) {
1223 conf
|= TLB_TYPE_F32_COLOR
;
1224 conf
|= ((num_components
- 1) <<
1225 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1227 conf
|= TLB_TYPE_F16_COLOR
;
1228 conf
|= TLB_F16_SWAP_HI_LO
;
1229 if (num_components
>= 3)
1230 conf
|= TLB_VEC_SIZE_4_F16
;
1232 conf
|= TLB_VEC_SIZE_2_F16
;
1236 int num_samples
= c
->msaa_per_sample_output
? V3D_MAX_SAMPLES
: 1;
1237 for (int i
= 0; i
< num_samples
; i
++) {
1238 struct qreg
*color
= c
->msaa_per_sample_output
?
1239 &c
->sample_colors
[(rt
* V3D_MAX_SAMPLES
+ i
) * 4] :
1240 &c
->outputs
[var
->data
.driver_location
* 4];
1242 struct qreg r
= color
[0];
1243 struct qreg g
= color
[1];
1244 struct qreg b
= color
[2];
1245 struct qreg a
= color
[3];
1247 if (c
->fs_key
->swap_color_rb
& (1 << rt
)) {
1252 if (c
->fs_key
->sample_alpha_to_one
)
1253 a
= vir_uniform_f(c
, 1.0);
1255 if (is_32b_tlb_format
) {
1257 inst
= vir_MOV_dest(c
, tlbu_reg
, r
);
1259 vir_get_uniform_index(c
,
1263 inst
= vir_MOV_dest(c
, tlb_reg
, r
);
1266 if (num_components
>= 2)
1267 vir_MOV_dest(c
, tlb_reg
, g
);
1268 if (num_components
>= 3)
1269 vir_MOV_dest(c
, tlb_reg
, b
);
1270 if (num_components
>= 4)
1271 vir_MOV_dest(c
, tlb_reg
, a
);
1273 inst
= vir_VFPACK_dest(c
, tlb_reg
, r
, g
);
1274 if (conf
!= ~0 && i
== 0) {
1275 inst
->dst
= tlbu_reg
;
1277 vir_get_uniform_index(c
,
1282 if (num_components
>= 3)
1283 inst
= vir_VFPACK_dest(c
, tlb_reg
, b
, a
);
1289 emit_frag_end(struct v3d_compile
*c
)
1292 if (c->output_sample_mask_index != -1) {
1293 vir_MS_MASK(c, c->outputs[c->output_sample_mask_index]);
1297 bool has_any_tlb_color_write
= false;
1298 for (int rt
= 0; rt
< V3D_MAX_DRAW_BUFFERS
; rt
++) {
1299 if (c
->fs_key
->cbufs
& (1 << rt
) && c
->output_color_var
[rt
])
1300 has_any_tlb_color_write
= true;
1303 if (c
->fs_key
->sample_alpha_to_coverage
&& c
->output_color_var
[0]) {
1304 struct nir_variable
*var
= c
->output_color_var
[0];
1305 struct qreg
*color
= &c
->outputs
[var
->data
.driver_location
* 4];
1307 vir_SETMSF_dest(c
, vir_nop_reg(),
1310 vir_FTOC(c
, color
[3])));
1313 struct qreg tlbu_reg
= vir_magic_reg(V3D_QPU_WADDR_TLBU
);
1314 if (c
->output_position_index
!= -1) {
1315 struct qinst
*inst
= vir_MOV_dest(c
, tlbu_reg
,
1316 c
->outputs
[c
->output_position_index
]);
1317 uint8_t tlb_specifier
= TLB_TYPE_DEPTH
;
1319 if (c
->devinfo
->ver
>= 42) {
1320 tlb_specifier
|= (TLB_V42_DEPTH_TYPE_PER_PIXEL
|
1321 TLB_SAMPLE_MODE_PER_PIXEL
);
1323 tlb_specifier
|= TLB_DEPTH_TYPE_PER_PIXEL
;
1325 inst
->uniform
= vir_get_uniform_index(c
, QUNIFORM_CONSTANT
,
1329 } else if (c
->s
->info
.fs
.uses_discard
||
1330 !c
->s
->info
.fs
.early_fragment_tests
||
1331 c
->fs_key
->sample_alpha_to_coverage
||
1332 !has_any_tlb_color_write
) {
1333 /* Emit passthrough Z if it needed to be delayed until shader
1334 * end due to potential discards.
1336 * Since (single-threaded) fragment shaders always need a TLB
1337 * write, emit passthrouh Z if we didn't have any color
1338 * buffers and flag us as potentially discarding, so that we
1339 * can use Z as the TLB write.
1341 c
->s
->info
.fs
.uses_discard
= true;
1343 struct qinst
*inst
= vir_MOV_dest(c
, tlbu_reg
,
1345 uint8_t tlb_specifier
= TLB_TYPE_DEPTH
;
1347 if (c
->devinfo
->ver
>= 42) {
1348 /* The spec says the PER_PIXEL flag is ignored for
1349 * invariant writes, but the simulator demands it.
1351 tlb_specifier
|= (TLB_V42_DEPTH_TYPE_INVARIANT
|
1352 TLB_SAMPLE_MODE_PER_PIXEL
);
1354 tlb_specifier
|= TLB_DEPTH_TYPE_INVARIANT
;
1357 inst
->uniform
= vir_get_uniform_index(c
,
1364 /* XXX: Performance improvement: Merge Z write and color writes TLB
1367 for (int rt
= 0; rt
< V3D_MAX_DRAW_BUFFERS
; rt
++)
1368 vir_emit_tlb_color_write(c
, rt
);
1372 vir_VPM_WRITE_indirect(struct v3d_compile
*c
,
1374 struct qreg vpm_index
)
1376 assert(c
->devinfo
->ver
>= 40);
1377 vir_STVPMV(c
, vpm_index
, val
);
1381 vir_VPM_WRITE(struct v3d_compile
*c
, struct qreg val
, uint32_t vpm_index
)
1383 if (c
->devinfo
->ver
>= 40) {
1384 vir_VPM_WRITE_indirect(c
, val
, vir_uniform_ui(c
, vpm_index
));
1386 /* XXX: v3d33_vir_vpm_write_setup(c); */
1387 vir_MOV_dest(c
, vir_reg(QFILE_MAGIC
, V3D_QPU_WADDR_VPM
), val
);
1392 emit_vert_end(struct v3d_compile
*c
)
1394 /* GFXH-1684: VPM writes need to be complete by the end of the shader.
1396 if (c
->devinfo
->ver
>= 40 && c
->devinfo
->ver
<= 42)
1401 emit_geom_end(struct v3d_compile
*c
)
1403 /* GFXH-1684: VPM writes need to be complete by the end of the shader.
1405 if (c
->devinfo
->ver
>= 40 && c
->devinfo
->ver
<= 42)
1410 v3d_optimize_nir(struct nir_shader
*s
)
1413 unsigned lower_flrp
=
1414 (s
->options
->lower_flrp16
? 16 : 0) |
1415 (s
->options
->lower_flrp32
? 32 : 0) |
1416 (s
->options
->lower_flrp64
? 64 : 0);
1421 NIR_PASS_V(s
, nir_lower_vars_to_ssa
);
1422 NIR_PASS(progress
, s
, nir_lower_alu_to_scalar
, NULL
, NULL
);
1423 NIR_PASS(progress
, s
, nir_lower_phis_to_scalar
);
1424 NIR_PASS(progress
, s
, nir_copy_prop
);
1425 NIR_PASS(progress
, s
, nir_opt_remove_phis
);
1426 NIR_PASS(progress
, s
, nir_opt_dce
);
1427 NIR_PASS(progress
, s
, nir_opt_dead_cf
);
1428 NIR_PASS(progress
, s
, nir_opt_cse
);
1429 NIR_PASS(progress
, s
, nir_opt_peephole_select
, 8, true, true);
1430 NIR_PASS(progress
, s
, nir_opt_algebraic
);
1431 NIR_PASS(progress
, s
, nir_opt_constant_folding
);
1433 if (lower_flrp
!= 0) {
1434 bool lower_flrp_progress
= false;
1436 NIR_PASS(lower_flrp_progress
, s
, nir_lower_flrp
,
1438 false /* always_precise */,
1439 s
->options
->lower_ffma
);
1440 if (lower_flrp_progress
) {
1441 NIR_PASS(progress
, s
, nir_opt_constant_folding
);
1445 /* Nothing should rematerialize any flrps, so we only
1446 * need to do this lowering once.
1451 NIR_PASS(progress
, s
, nir_opt_undef
);
1454 NIR_PASS(progress
, s
, nir_opt_move
, nir_move_load_ubo
);
1458 driver_location_compare(const void *in_a
, const void *in_b
)
1460 const nir_variable
*const *a
= in_a
;
1461 const nir_variable
*const *b
= in_b
;
1463 if ((*a
)->data
.driver_location
== (*b
)->data
.driver_location
)
1464 return (*a
)->data
.location_frac
- (*b
)->data
.location_frac
;
1466 return (*a
)->data
.driver_location
- (*b
)->data
.driver_location
;
1470 ntq_emit_vpm_read(struct v3d_compile
*c
,
1471 uint32_t *num_components_queued
,
1472 uint32_t *remaining
,
1475 struct qreg vpm
= vir_reg(QFILE_VPM
, vpm_index
);
1477 if (c
->devinfo
->ver
>= 40 ) {
1478 return vir_LDVPMV_IN(c
,
1480 (*num_components_queued
)++));
1483 if (*num_components_queued
!= 0) {
1484 (*num_components_queued
)--;
1485 return vir_MOV(c
, vpm
);
1488 uint32_t num_components
= MIN2(*remaining
, 32);
1490 v3d33_vir_vpm_read_setup(c
, num_components
);
1492 *num_components_queued
= num_components
- 1;
1493 *remaining
-= num_components
;
1495 return vir_MOV(c
, vpm
);
1499 ntq_setup_vs_inputs(struct v3d_compile
*c
)
1501 /* Figure out how many components of each vertex attribute the shader
1502 * uses. Each variable should have been split to individual
1503 * components and unused ones DCEed. The vertex fetcher will load
1504 * from the start of the attribute to the number of components we
1505 * declare we need in c->vattr_sizes[].
1507 nir_foreach_variable(var
, &c
->s
->inputs
) {
1508 /* No VS attribute array support. */
1509 assert(MAX2(glsl_get_length(var
->type
), 1) == 1);
1511 unsigned loc
= var
->data
.driver_location
;
1512 int start_component
= var
->data
.location_frac
;
1513 int num_components
= glsl_get_components(var
->type
);
1515 c
->vattr_sizes
[loc
] = MAX2(c
->vattr_sizes
[loc
],
1516 start_component
+ num_components
);
1519 unsigned num_components
= 0;
1520 uint32_t vpm_components_queued
= 0;
1521 bool uses_iid
= c
->s
->info
.system_values_read
&
1522 (1ull << SYSTEM_VALUE_INSTANCE_ID
);
1523 bool uses_vid
= c
->s
->info
.system_values_read
&
1524 (1ull << SYSTEM_VALUE_VERTEX_ID
);
1525 num_components
+= uses_iid
;
1526 num_components
+= uses_vid
;
1528 for (int i
= 0; i
< ARRAY_SIZE(c
->vattr_sizes
); i
++)
1529 num_components
+= c
->vattr_sizes
[i
];
1532 c
->iid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1533 &num_components
, ~0);
1537 c
->vid
= ntq_emit_vpm_read(c
, &vpm_components_queued
,
1538 &num_components
, ~0);
1541 /* The actual loads will happen directly in nir_intrinsic_load_input
1542 * on newer versions.
1544 if (c
->devinfo
->ver
>= 40)
1547 for (int loc
= 0; loc
< ARRAY_SIZE(c
->vattr_sizes
); loc
++) {
1548 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1551 for (int i
= 0; i
< c
->vattr_sizes
[loc
]; i
++) {
1552 c
->inputs
[loc
* 4 + i
] =
1553 ntq_emit_vpm_read(c
,
1554 &vpm_components_queued
,
1561 if (c
->devinfo
->ver
>= 40) {
1562 assert(vpm_components_queued
== num_components
);
1564 assert(vpm_components_queued
== 0);
1565 assert(num_components
== 0);
1570 program_reads_point_coord(struct v3d_compile
*c
)
1572 nir_foreach_variable(var
, &c
->s
->inputs
) {
1573 if (util_varying_is_point_coord(var
->data
.location
,
1574 c
->fs_key
->point_sprite_mask
)) {
1583 get_sorted_input_variables(struct v3d_compile
*c
,
1584 unsigned *num_entries
,
1585 nir_variable
***vars
)
1588 nir_foreach_variable(var
, &c
->s
->inputs
)
1591 *vars
= ralloc_array(c
, nir_variable
*, *num_entries
);
1594 nir_foreach_variable(var
, &c
->s
->inputs
)
1597 /* Sort the variables so that we emit the input setup in
1598 * driver_location order. This is required for VPM reads, whose data
1599 * is fetched into the VPM in driver_location (TGSI register index)
1602 qsort(*vars
, *num_entries
, sizeof(**vars
), driver_location_compare
);
1606 ntq_setup_gs_inputs(struct v3d_compile
*c
)
1608 nir_variable
**vars
;
1609 unsigned num_entries
;
1610 get_sorted_input_variables(c
, &num_entries
, &vars
);
1612 for (unsigned i
= 0; i
< num_entries
; i
++) {
1613 nir_variable
*var
= vars
[i
];
1615 /* All GS inputs are arrays with as many entries as vertices
1616 * in the input primitive, but here we only care about the
1617 * per-vertex input type.
1619 const struct glsl_type
*type
= glsl_without_array(var
->type
);
1620 unsigned array_len
= MAX2(glsl_get_length(type
), 1);
1621 unsigned loc
= var
->data
.driver_location
;
1623 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1624 (loc
+ array_len
) * 4);
1626 for (unsigned j
= 0; j
< array_len
; j
++) {
1627 unsigned num_elements
= glsl_get_vector_elements(type
);
1628 for (unsigned k
= 0; k
< num_elements
; k
++) {
1629 unsigned chan
= var
->data
.location_frac
+ k
;
1630 unsigned input_idx
= c
->num_inputs
++;
1631 struct v3d_varying_slot slot
=
1632 v3d_slot_from_slot_and_component(var
->data
.location
+ j
, chan
);
1633 c
->input_slots
[input_idx
] = slot
;
1641 ntq_setup_fs_inputs(struct v3d_compile
*c
)
1643 nir_variable
**vars
;
1644 unsigned num_entries
;
1645 get_sorted_input_variables(c
, &num_entries
, &vars
);
1647 for (unsigned i
= 0; i
< num_entries
; i
++) {
1648 nir_variable
*var
= vars
[i
];
1649 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1650 unsigned loc
= var
->data
.driver_location
;
1652 resize_qreg_array(c
, &c
->inputs
, &c
->inputs_array_size
,
1653 (loc
+ array_len
) * 4);
1655 if (var
->data
.location
== VARYING_SLOT_POS
) {
1656 emit_fragcoord_input(c
, loc
);
1657 } else if (util_varying_is_point_coord(var
->data
.location
,
1658 c
->fs_key
->point_sprite_mask
)) {
1659 c
->inputs
[loc
* 4 + 0] = c
->point_x
;
1660 c
->inputs
[loc
* 4 + 1] = c
->point_y
;
1662 for (int j
= 0; j
< array_len
; j
++)
1663 emit_fragment_input(c
, loc
+ j
, var
, j
);
1669 ntq_setup_outputs(struct v3d_compile
*c
)
1671 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
)
1674 nir_foreach_variable(var
, &c
->s
->outputs
) {
1675 unsigned array_len
= MAX2(glsl_get_length(var
->type
), 1);
1676 unsigned loc
= var
->data
.driver_location
* 4;
1678 assert(array_len
== 1);
1681 for (int i
= 0; i
< 4 - var
->data
.location_frac
; i
++) {
1682 add_output(c
, loc
+ var
->data
.location_frac
+ i
,
1684 var
->data
.location_frac
+ i
);
1687 switch (var
->data
.location
) {
1688 case FRAG_RESULT_COLOR
:
1689 c
->output_color_var
[0] = var
;
1690 c
->output_color_var
[1] = var
;
1691 c
->output_color_var
[2] = var
;
1692 c
->output_color_var
[3] = var
;
1694 case FRAG_RESULT_DATA0
:
1695 case FRAG_RESULT_DATA1
:
1696 case FRAG_RESULT_DATA2
:
1697 case FRAG_RESULT_DATA3
:
1698 c
->output_color_var
[var
->data
.location
-
1699 FRAG_RESULT_DATA0
] = var
;
1701 case FRAG_RESULT_DEPTH
:
1702 c
->output_position_index
= loc
;
1704 case FRAG_RESULT_SAMPLE_MASK
:
1705 c
->output_sample_mask_index
= loc
;
1712 * Sets up the mapping from nir_register to struct qreg *.
1714 * Each nir_register gets a struct qreg per 32-bit component being stored.
1717 ntq_setup_registers(struct v3d_compile
*c
, struct exec_list
*list
)
1719 foreach_list_typed(nir_register
, nir_reg
, node
, list
) {
1720 unsigned array_len
= MAX2(nir_reg
->num_array_elems
, 1);
1721 struct qreg
*qregs
= ralloc_array(c
->def_ht
, struct qreg
,
1723 nir_reg
->num_components
);
1725 _mesa_hash_table_insert(c
->def_ht
, nir_reg
, qregs
);
1727 for (int i
= 0; i
< array_len
* nir_reg
->num_components
; i
++)
1728 qregs
[i
] = vir_get_temp(c
);
1733 ntq_emit_load_const(struct v3d_compile
*c
, nir_load_const_instr
*instr
)
1735 /* XXX perf: Experiment with using immediate loads to avoid having
1736 * these end up in the uniform stream. Watch out for breaking the
1737 * small immediates optimization in the process!
1739 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1740 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1741 qregs
[i
] = vir_uniform_ui(c
, instr
->value
[i
].u32
);
1743 _mesa_hash_table_insert(c
->def_ht
, &instr
->def
, qregs
);
1747 ntq_emit_ssa_undef(struct v3d_compile
*c
, nir_ssa_undef_instr
*instr
)
1749 struct qreg
*qregs
= ntq_init_ssa_def(c
, &instr
->def
);
1751 /* VIR needs there to be *some* value, so pick 0 (same as for
1752 * ntq_setup_registers().
1754 for (int i
= 0; i
< instr
->def
.num_components
; i
++)
1755 qregs
[i
] = vir_uniform_ui(c
, 0);
1759 ntq_emit_image_size(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1761 unsigned image_index
= nir_src_as_uint(instr
->src
[0]);
1762 bool is_array
= nir_intrinsic_image_array(instr
);
1764 ntq_store_dest(c
, &instr
->dest
, 0,
1765 vir_uniform(c
, QUNIFORM_IMAGE_WIDTH
, image_index
));
1766 if (instr
->num_components
> 1) {
1767 ntq_store_dest(c
, &instr
->dest
, 1,
1769 instr
->num_components
== 2 && is_array
?
1770 QUNIFORM_IMAGE_ARRAY_SIZE
:
1771 QUNIFORM_IMAGE_HEIGHT
,
1774 if (instr
->num_components
> 2) {
1775 ntq_store_dest(c
, &instr
->dest
, 2,
1778 QUNIFORM_IMAGE_ARRAY_SIZE
:
1779 QUNIFORM_IMAGE_DEPTH
,
1785 vir_emit_tlb_color_read(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1787 assert(c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
);
1789 int rt
= nir_src_as_uint(instr
->src
[0]);
1790 assert(rt
< V3D_MAX_DRAW_BUFFERS
);
1792 int sample_index
= nir_intrinsic_base(instr
) ;
1793 assert(sample_index
< V3D_MAX_SAMPLES
);
1795 int component
= nir_intrinsic_component(instr
);
1796 assert(component
< 4);
1798 /* We need to emit our TLB reads after we have acquired the scoreboard
1799 * lock, or the GPU will hang. Usually, we do our scoreboard locking on
1800 * the last thread switch to improve parallelism, however, that is only
1801 * guaranteed to happen before the tlb color writes.
1803 * To fix that, we make sure we always emit a thread switch before the
1804 * first tlb color read. If that happens to be the last thread switch
1805 * we emit, then everything is fine, but otherwsie, if any code after
1806 * this point needs to emit additional thread switches, then we will
1807 * switch the strategy to locking the scoreboard on the first thread
1808 * switch instead -- see vir_emit_thrsw().
1810 if (!c
->emitted_tlb_load
) {
1811 if (!c
->last_thrsw_at_top_level
) {
1812 assert(c
->devinfo
->ver
>= 41);
1816 c
->emitted_tlb_load
= true;
1819 struct qreg
*color_reads_for_sample
=
1820 &c
->color_reads
[(rt
* V3D_MAX_SAMPLES
+ sample_index
) * 4];
1822 if (color_reads_for_sample
[component
].file
== QFILE_NULL
) {
1823 enum pipe_format rt_format
= c
->fs_key
->color_fmt
[rt
].format
;
1824 int num_components
=
1825 util_format_get_nr_components(rt_format
);
1827 const bool swap_rb
= c
->fs_key
->swap_color_rb
& (1 << rt
);
1829 num_components
= MAX2(num_components
, 3);
1831 nir_variable
*var
= c
->output_color_var
[rt
];
1832 enum glsl_base_type type
= glsl_get_base_type(var
->type
);
1834 bool is_int_format
= type
== GLSL_TYPE_INT
||
1835 type
== GLSL_TYPE_UINT
;
1837 bool is_32b_tlb_format
= is_int_format
||
1838 (c
->fs_key
->f32_color_rb
& (1 << rt
));
1840 int num_samples
= c
->fs_key
->msaa
? V3D_MAX_SAMPLES
: 1;
1842 uint32_t conf
= 0xffffff00;
1843 conf
|= c
->fs_key
->msaa
? TLB_SAMPLE_MODE_PER_SAMPLE
:
1844 TLB_SAMPLE_MODE_PER_PIXEL
;
1845 conf
|= (7 - rt
) << TLB_RENDER_TARGET_SHIFT
;
1847 if (is_32b_tlb_format
) {
1848 /* The F32 vs I32 distinction was dropped in 4.2. */
1849 conf
|= (c
->devinfo
->ver
< 42 && is_int_format
) ?
1850 TLB_TYPE_I32_COLOR
: TLB_TYPE_F32_COLOR
;
1852 conf
|= ((num_components
- 1) <<
1853 TLB_VEC_SIZE_MINUS_1_SHIFT
);
1855 conf
|= TLB_TYPE_F16_COLOR
;
1856 conf
|= TLB_F16_SWAP_HI_LO
;
1858 if (num_components
>= 3)
1859 conf
|= TLB_VEC_SIZE_4_F16
;
1861 conf
|= TLB_VEC_SIZE_2_F16
;
1865 for (int i
= 0; i
< num_samples
; i
++) {
1866 struct qreg r
, g
, b
, a
;
1867 if (is_32b_tlb_format
) {
1868 r
= conf
!= 0xffffffff && i
== 0?
1869 vir_TLBU_COLOR_READ(c
, conf
) :
1870 vir_TLB_COLOR_READ(c
);
1871 if (num_components
>= 2)
1872 g
= vir_TLB_COLOR_READ(c
);
1873 if (num_components
>= 3)
1874 b
= vir_TLB_COLOR_READ(c
);
1875 if (num_components
>= 4)
1876 a
= vir_TLB_COLOR_READ(c
);
1878 struct qreg rg
= conf
!= 0xffffffff && i
== 0 ?
1879 vir_TLBU_COLOR_READ(c
, conf
) :
1880 vir_TLB_COLOR_READ(c
);
1881 r
= vir_FMOV(c
, rg
);
1882 vir_set_unpack(c
->defs
[r
.index
], 0,
1884 g
= vir_FMOV(c
, rg
);
1885 vir_set_unpack(c
->defs
[g
.index
], 0,
1888 if (num_components
> 2) {
1889 struct qreg ba
= vir_TLB_COLOR_READ(c
);
1890 b
= vir_FMOV(c
, ba
);
1891 vir_set_unpack(c
->defs
[b
.index
], 0,
1893 a
= vir_FMOV(c
, ba
);
1894 vir_set_unpack(c
->defs
[a
.index
], 0,
1899 struct qreg
*color_reads
=
1900 &c
->color_reads
[(rt
* V3D_MAX_SAMPLES
+ i
) * 4];
1902 color_reads
[0] = swap_rb
? b
: r
;
1903 if (num_components
>= 2)
1905 if (num_components
>= 3)
1906 color_reads
[2] = swap_rb
? r
: b
;
1907 if (num_components
>= 4)
1912 assert(color_reads_for_sample
[component
].file
!= QFILE_NULL
);
1913 ntq_store_dest(c
, &instr
->dest
, 0,
1914 vir_MOV(c
, color_reads_for_sample
[component
]));
1918 ntq_emit_load_uniform(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1920 if (nir_src_is_const(instr
->src
[0])) {
1921 int offset
= (nir_intrinsic_base(instr
) +
1922 nir_src_as_uint(instr
->src
[0]));
1923 assert(offset
% 4 == 0);
1924 /* We need dwords */
1925 offset
= offset
/ 4;
1926 for (int i
= 0; i
< instr
->num_components
; i
++) {
1927 ntq_store_dest(c
, &instr
->dest
, i
,
1928 vir_uniform(c
, QUNIFORM_UNIFORM
,
1932 ntq_emit_tmu_general(c
, instr
, false);
1937 ntq_emit_load_input(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
1939 /* XXX: Use ldvpmv (uniform offset) or ldvpmd (non-uniform offset)
1940 * and enable PIPE_SHADER_CAP_INDIRECT_INPUT_ADDR.
1943 nir_intrinsic_base(instr
) + nir_src_as_uint(instr
->src
[0]);
1945 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
&& c
->devinfo
->ver
>= 40) {
1946 /* Emit the LDVPM directly now, rather than at the top
1947 * of the shader like we did for V3D 3.x (which needs
1948 * vpmsetup when not just taking the next offset).
1950 * Note that delaying like this may introduce stalls,
1951 * as LDVPMV takes a minimum of 1 instruction but may
1952 * be slower if the VPM unit is busy with another QPU.
1955 if (c
->s
->info
.system_values_read
&
1956 (1ull << SYSTEM_VALUE_INSTANCE_ID
)) {
1959 if (c
->s
->info
.system_values_read
&
1960 (1ull << SYSTEM_VALUE_VERTEX_ID
)) {
1963 for (int i
= 0; i
< offset
; i
++)
1964 index
+= c
->vattr_sizes
[i
];
1965 index
+= nir_intrinsic_component(instr
);
1966 for (int i
= 0; i
< instr
->num_components
; i
++) {
1967 struct qreg vpm_offset
= vir_uniform_ui(c
, index
++);
1968 ntq_store_dest(c
, &instr
->dest
, i
,
1969 vir_LDVPMV_IN(c
, vpm_offset
));
1972 for (int i
= 0; i
< instr
->num_components
; i
++) {
1973 int comp
= nir_intrinsic_component(instr
) + i
;
1974 ntq_store_dest(c
, &instr
->dest
, i
,
1975 vir_MOV(c
, c
->inputs
[offset
* 4 + comp
]));
1981 ntq_emit_per_sample_color_write(struct v3d_compile
*c
,
1982 nir_intrinsic_instr
*instr
)
1984 assert(instr
->intrinsic
== nir_intrinsic_store_tlb_sample_color_v3d
);
1986 unsigned rt
= nir_src_as_uint(instr
->src
[1]);
1987 assert(rt
< V3D_MAX_DRAW_BUFFERS
);
1989 unsigned sample_idx
= nir_intrinsic_base(instr
);
1990 assert(sample_idx
< V3D_MAX_SAMPLES
);
1992 unsigned offset
= (rt
* V3D_MAX_SAMPLES
+ sample_idx
) * 4;
1993 for (int i
= 0; i
< instr
->num_components
; i
++) {
1994 c
->sample_colors
[offset
+ i
] =
1995 vir_MOV(c
, ntq_get_src(c
, instr
->src
[0], i
));
2000 ntq_emit_color_write(struct v3d_compile
*c
,
2001 nir_intrinsic_instr
*instr
)
2003 unsigned offset
= (nir_intrinsic_base(instr
) +
2004 nir_src_as_uint(instr
->src
[1])) * 4 +
2005 nir_intrinsic_component(instr
);
2006 for (int i
= 0; i
< instr
->num_components
; i
++) {
2007 c
->outputs
[offset
+ i
] =
2008 vir_MOV(c
, ntq_get_src(c
, instr
->src
[0], i
));
2013 emit_store_output_gs(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
2015 assert(instr
->num_components
== 1);
2017 uint32_t base_offset
= nir_intrinsic_base(instr
);
2018 struct qreg src_offset
= ntq_get_src(c
, instr
->src
[1], 0);
2019 struct qreg offset
=
2020 vir_ADD(c
, vir_uniform_ui(c
, base_offset
), src_offset
);
2022 /* Usually, for VS or FS, we only emit outputs once at program end so
2023 * our VPM writes are never in non-uniform control flow, but this
2024 * is not true for GS, where we are emitting multiple vertices.
2026 if (vir_in_nonuniform_control_flow(c
)) {
2027 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2031 struct qreg val
= ntq_get_src(c
, instr
->src
[0], 0);
2033 /* The offset isn’t necessarily dynamically uniform for a geometry
2034 * shader. This can happen if the shader sometimes doesn’t emit one of
2035 * the vertices. In that case subsequent vertices will be written to
2036 * different offsets in the VPM and we need to use the scatter write
2037 * instruction to have a different offset for each lane.
2039 if (nir_src_is_dynamically_uniform(instr
->src
[1]))
2040 vir_VPM_WRITE_indirect(c
, val
, offset
);
2042 vir_STVPMD(c
, offset
, val
);
2044 if (vir_in_nonuniform_control_flow(c
)) {
2045 struct qinst
*last_inst
=
2046 (struct qinst
*)c
->cur_block
->instructions
.prev
;
2047 vir_set_cond(last_inst
, V3D_QPU_COND_IFA
);
2052 ntq_emit_store_output(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
2054 /* XXX perf: Use stvpmv with uniform non-constant offsets and
2055 * stvpmd with non-uniform offsets and enable
2056 * PIPE_SHADER_CAP_INDIRECT_OUTPUT_ADDR.
2058 if (c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2059 ntq_emit_color_write(c
, instr
);
2060 } else if (c
->s
->info
.stage
== MESA_SHADER_GEOMETRY
) {
2061 emit_store_output_gs(c
, instr
);
2063 assert(c
->s
->info
.stage
== MESA_SHADER_VERTEX
);
2064 assert(instr
->num_components
== 1);
2067 ntq_get_src(c
, instr
->src
[0], 0),
2068 nir_intrinsic_base(instr
));
2073 ntq_emit_intrinsic(struct v3d_compile
*c
, nir_intrinsic_instr
*instr
)
2075 switch (instr
->intrinsic
) {
2076 case nir_intrinsic_load_uniform
:
2077 ntq_emit_load_uniform(c
, instr
);
2080 case nir_intrinsic_load_ubo
:
2081 ntq_emit_tmu_general(c
, instr
, false);
2084 case nir_intrinsic_ssbo_atomic_add
:
2085 case nir_intrinsic_ssbo_atomic_imin
:
2086 case nir_intrinsic_ssbo_atomic_umin
:
2087 case nir_intrinsic_ssbo_atomic_imax
:
2088 case nir_intrinsic_ssbo_atomic_umax
:
2089 case nir_intrinsic_ssbo_atomic_and
:
2090 case nir_intrinsic_ssbo_atomic_or
:
2091 case nir_intrinsic_ssbo_atomic_xor
:
2092 case nir_intrinsic_ssbo_atomic_exchange
:
2093 case nir_intrinsic_ssbo_atomic_comp_swap
:
2094 case nir_intrinsic_load_ssbo
:
2095 case nir_intrinsic_store_ssbo
:
2096 ntq_emit_tmu_general(c
, instr
, false);
2099 case nir_intrinsic_shared_atomic_add
:
2100 case nir_intrinsic_shared_atomic_imin
:
2101 case nir_intrinsic_shared_atomic_umin
:
2102 case nir_intrinsic_shared_atomic_imax
:
2103 case nir_intrinsic_shared_atomic_umax
:
2104 case nir_intrinsic_shared_atomic_and
:
2105 case nir_intrinsic_shared_atomic_or
:
2106 case nir_intrinsic_shared_atomic_xor
:
2107 case nir_intrinsic_shared_atomic_exchange
:
2108 case nir_intrinsic_shared_atomic_comp_swap
:
2109 case nir_intrinsic_load_shared
:
2110 case nir_intrinsic_store_shared
:
2111 case nir_intrinsic_load_scratch
:
2112 case nir_intrinsic_store_scratch
:
2113 ntq_emit_tmu_general(c
, instr
, true);
2116 case nir_intrinsic_image_load
:
2117 case nir_intrinsic_image_store
:
2118 case nir_intrinsic_image_atomic_add
:
2119 case nir_intrinsic_image_atomic_imin
:
2120 case nir_intrinsic_image_atomic_umin
:
2121 case nir_intrinsic_image_atomic_imax
:
2122 case nir_intrinsic_image_atomic_umax
:
2123 case nir_intrinsic_image_atomic_and
:
2124 case nir_intrinsic_image_atomic_or
:
2125 case nir_intrinsic_image_atomic_xor
:
2126 case nir_intrinsic_image_atomic_exchange
:
2127 case nir_intrinsic_image_atomic_comp_swap
:
2128 v3d40_vir_emit_image_load_store(c
, instr
);
2131 case nir_intrinsic_get_buffer_size
:
2132 ntq_store_dest(c
, &instr
->dest
, 0,
2133 vir_uniform(c
, QUNIFORM_GET_BUFFER_SIZE
,
2134 nir_src_as_uint(instr
->src
[0])));
2137 case nir_intrinsic_load_user_clip_plane
:
2138 for (int i
= 0; i
< nir_intrinsic_dest_components(instr
); i
++) {
2139 ntq_store_dest(c
, &instr
->dest
, i
,
2140 vir_uniform(c
, QUNIFORM_USER_CLIP_PLANE
,
2141 nir_intrinsic_ucp_id(instr
) *
2146 case nir_intrinsic_load_viewport_x_scale
:
2147 ntq_store_dest(c
, &instr
->dest
, 0,
2148 vir_uniform(c
, QUNIFORM_VIEWPORT_X_SCALE
, 0));
2151 case nir_intrinsic_load_viewport_y_scale
:
2152 ntq_store_dest(c
, &instr
->dest
, 0,
2153 vir_uniform(c
, QUNIFORM_VIEWPORT_Y_SCALE
, 0));
2156 case nir_intrinsic_load_viewport_z_scale
:
2157 ntq_store_dest(c
, &instr
->dest
, 0,
2158 vir_uniform(c
, QUNIFORM_VIEWPORT_Z_SCALE
, 0));
2161 case nir_intrinsic_load_viewport_z_offset
:
2162 ntq_store_dest(c
, &instr
->dest
, 0,
2163 vir_uniform(c
, QUNIFORM_VIEWPORT_Z_OFFSET
, 0));
2166 case nir_intrinsic_load_alpha_ref_float
:
2167 ntq_store_dest(c
, &instr
->dest
, 0,
2168 vir_uniform(c
, QUNIFORM_ALPHA_REF
, 0));
2171 case nir_intrinsic_load_line_coord
:
2172 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->line_x
));
2175 case nir_intrinsic_load_line_width
:
2176 ntq_store_dest(c
, &instr
->dest
, 0,
2177 vir_uniform(c
, QUNIFORM_LINE_WIDTH
, 0));
2180 case nir_intrinsic_load_aa_line_width
:
2181 ntq_store_dest(c
, &instr
->dest
, 0,
2182 vir_uniform(c
, QUNIFORM_AA_LINE_WIDTH
, 0));
2185 case nir_intrinsic_load_sample_mask_in
:
2186 ntq_store_dest(c
, &instr
->dest
, 0, vir_MSF(c
));
2189 case nir_intrinsic_load_helper_invocation
:
2190 vir_set_pf(vir_MSF_dest(c
, vir_nop_reg()), V3D_QPU_PF_PUSHZ
);
2191 ntq_store_dest(c
, &instr
->dest
, 0,
2192 vir_MOV(c
, vir_SEL(c
, V3D_QPU_COND_IFA
,
2193 vir_uniform_ui(c
, ~0),
2194 vir_uniform_ui(c
, 0))));
2197 case nir_intrinsic_load_front_face
:
2198 /* The register contains 0 (front) or 1 (back), and we need to
2199 * turn it into a NIR bool where true means front.
2201 ntq_store_dest(c
, &instr
->dest
, 0,
2203 vir_uniform_ui(c
, -1),
2207 case nir_intrinsic_load_instance_id
:
2208 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->iid
));
2211 case nir_intrinsic_load_vertex_id
:
2212 ntq_store_dest(c
, &instr
->dest
, 0, vir_MOV(c
, c
->vid
));
2215 case nir_intrinsic_load_tlb_color_v3d
:
2216 vir_emit_tlb_color_read(c
, instr
);
2219 case nir_intrinsic_load_input
:
2220 ntq_emit_load_input(c
, instr
);
2223 case nir_intrinsic_store_tlb_sample_color_v3d
:
2224 ntq_emit_per_sample_color_write(c
, instr
);
2227 case nir_intrinsic_store_output
:
2228 ntq_emit_store_output(c
, instr
);
2231 case nir_intrinsic_image_size
:
2232 ntq_emit_image_size(c
, instr
);
2235 case nir_intrinsic_discard
:
2236 if (vir_in_nonuniform_control_flow(c
)) {
2237 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2239 vir_set_cond(vir_SETMSF_dest(c
, vir_nop_reg(),
2240 vir_uniform_ui(c
, 0)),
2243 vir_SETMSF_dest(c
, vir_nop_reg(),
2244 vir_uniform_ui(c
, 0));
2248 case nir_intrinsic_discard_if
: {
2249 enum v3d_qpu_cond cond
= ntq_emit_bool_to_cond(c
, instr
->src
[0]);
2251 if (vir_in_nonuniform_control_flow(c
)) {
2252 struct qinst
*exec_flag
= vir_MOV_dest(c
, vir_nop_reg(),
2254 if (cond
== V3D_QPU_COND_IFA
) {
2255 vir_set_uf(exec_flag
, V3D_QPU_UF_ANDZ
);
2257 vir_set_uf(exec_flag
, V3D_QPU_UF_NORNZ
);
2258 cond
= V3D_QPU_COND_IFA
;
2262 vir_set_cond(vir_SETMSF_dest(c
, vir_nop_reg(),
2263 vir_uniform_ui(c
, 0)), cond
);
2268 case nir_intrinsic_memory_barrier
:
2269 case nir_intrinsic_memory_barrier_buffer
:
2270 case nir_intrinsic_memory_barrier_image
:
2271 case nir_intrinsic_memory_barrier_shared
:
2272 case nir_intrinsic_memory_barrier_tcs_patch
:
2273 case nir_intrinsic_group_memory_barrier
:
2274 /* We don't do any instruction scheduling of these NIR
2275 * instructions between each other, so we just need to make
2276 * sure that the TMU operations before the barrier are flushed
2277 * before the ones after the barrier. That is currently
2278 * handled by having a THRSW in each of them and a LDTMU
2279 * series or a TMUWT after.
2283 case nir_intrinsic_control_barrier
:
2284 /* Emit a TSY op to get all invocations in the workgroup
2285 * (actually supergroup) to block until the last invocation
2286 * reaches the TSY op.
2288 if (c
->devinfo
->ver
>= 42) {
2289 vir_BARRIERID_dest(c
, vir_reg(QFILE_MAGIC
,
2290 V3D_QPU_WADDR_SYNCB
));
2292 struct qinst
*sync
=
2293 vir_BARRIERID_dest(c
,
2294 vir_reg(QFILE_MAGIC
,
2295 V3D_QPU_WADDR_SYNCU
));
2297 vir_get_uniform_index(c
, QUNIFORM_CONSTANT
,
2299 V3D_TSY_WAIT_INC_CHECK
);
2303 /* The blocking of a TSY op only happens at the next thread
2304 * switch. No texturing may be outstanding at the time of a
2305 * TSY blocking operation.
2310 case nir_intrinsic_load_num_work_groups
:
2311 for (int i
= 0; i
< 3; i
++) {
2312 ntq_store_dest(c
, &instr
->dest
, i
,
2313 vir_uniform(c
, QUNIFORM_NUM_WORK_GROUPS
,
2318 case nir_intrinsic_load_local_invocation_index
:
2319 ntq_store_dest(c
, &instr
->dest
, 0,
2320 vir_SHR(c
, c
->cs_payload
[1],
2321 vir_uniform_ui(c
, 32 - c
->local_invocation_index_bits
)));
2324 case nir_intrinsic_load_work_group_id
:
2325 ntq_store_dest(c
, &instr
->dest
, 0,
2326 vir_AND(c
, c
->cs_payload
[0],
2327 vir_uniform_ui(c
, 0xffff)));
2328 ntq_store_dest(c
, &instr
->dest
, 1,
2329 vir_SHR(c
, c
->cs_payload
[0],
2330 vir_uniform_ui(c
, 16)));
2331 ntq_store_dest(c
, &instr
->dest
, 2,
2332 vir_AND(c
, c
->cs_payload
[1],
2333 vir_uniform_ui(c
, 0xffff)));
2336 case nir_intrinsic_load_subgroup_id
:
2337 ntq_store_dest(c
, &instr
->dest
, 0, vir_EIDX(c
));
2340 case nir_intrinsic_load_per_vertex_input
: {
2341 /* col: vertex index, row = varying index */
2342 struct qreg col
= ntq_get_src(c
, instr
->src
[0], 0);
2343 uint32_t row_idx
= nir_intrinsic_base(instr
) * 4 +
2344 nir_intrinsic_component(instr
);
2345 for (int i
= 0; i
< instr
->num_components
; i
++) {
2346 struct qreg row
= vir_uniform_ui(c
, row_idx
++);
2347 ntq_store_dest(c
, &instr
->dest
, i
,
2348 vir_LDVPMG_IN(c
, row
, col
));
2353 case nir_intrinsic_emit_vertex
:
2354 case nir_intrinsic_end_primitive
:
2355 unreachable("Should have been lowered in v3d_nir_lower_io");
2358 case nir_intrinsic_load_primitive_id
: {
2359 /* gl_PrimitiveIdIn is written by the GBG in the first word of
2360 * VPM output header. According to docs, we should read this
2361 * using ldvpm(v,d)_in (See Table 71).
2363 ntq_store_dest(c
, &instr
->dest
, 0,
2364 vir_LDVPMV_IN(c
, vir_uniform_ui(c
, 0)));
2368 case nir_intrinsic_load_invocation_id
:
2369 ntq_store_dest(c
, &instr
->dest
, 0, vir_IID(c
));
2372 case nir_intrinsic_load_fb_layers_v3d
:
2373 ntq_store_dest(c
, &instr
->dest
, 0,
2374 vir_uniform(c
, QUNIFORM_FB_LAYERS
, 0));
2378 fprintf(stderr
, "Unknown intrinsic: ");
2379 nir_print_instr(&instr
->instr
, stderr
);
2380 fprintf(stderr
, "\n");
2385 /* Clears (activates) the execute flags for any channels whose jump target
2386 * matches this block.
2388 * XXX perf: Could we be using flpush/flpop somehow for our execution channel
2391 * XXX perf: For uniform control flow, we should be able to skip c->execute
2392 * handling entirely.
2395 ntq_activate_execute_for_block(struct v3d_compile
*c
)
2397 vir_set_pf(vir_XOR_dest(c
, vir_nop_reg(),
2398 c
->execute
, vir_uniform_ui(c
, c
->cur_block
->index
)),
2401 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
2405 ntq_emit_uniform_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
2407 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
2408 bool empty_else_block
=
2409 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
2410 exec_list_is_empty(&nir_else_block
->instr_list
));
2412 struct qblock
*then_block
= vir_new_block(c
);
2413 struct qblock
*after_block
= vir_new_block(c
);
2414 struct qblock
*else_block
;
2415 if (empty_else_block
)
2416 else_block
= after_block
;
2418 else_block
= vir_new_block(c
);
2420 /* Set up the flags for the IF condition (taking the THEN branch). */
2421 enum v3d_qpu_cond cond
= ntq_emit_bool_to_cond(c
, if_stmt
->condition
);
2424 vir_BRANCH(c
, cond
== V3D_QPU_COND_IFA
?
2425 V3D_QPU_BRANCH_COND_ALLNA
:
2426 V3D_QPU_BRANCH_COND_ALLA
);
2427 vir_link_blocks(c
->cur_block
, else_block
);
2428 vir_link_blocks(c
->cur_block
, then_block
);
2430 /* Process the THEN block. */
2431 vir_set_emit_block(c
, then_block
);
2432 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
2434 if (!empty_else_block
) {
2435 /* At the end of the THEN block, jump to ENDIF */
2436 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALWAYS
);
2437 vir_link_blocks(c
->cur_block
, after_block
);
2439 /* Emit the else block. */
2440 vir_set_emit_block(c
, else_block
);
2441 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
2444 vir_link_blocks(c
->cur_block
, after_block
);
2446 vir_set_emit_block(c
, after_block
);
2450 ntq_emit_nonuniform_if(struct v3d_compile
*c
, nir_if
*if_stmt
)
2452 nir_block
*nir_else_block
= nir_if_first_else_block(if_stmt
);
2453 bool empty_else_block
=
2454 (nir_else_block
== nir_if_last_else_block(if_stmt
) &&
2455 exec_list_is_empty(&nir_else_block
->instr_list
));
2457 struct qblock
*then_block
= vir_new_block(c
);
2458 struct qblock
*after_block
= vir_new_block(c
);
2459 struct qblock
*else_block
;
2460 if (empty_else_block
)
2461 else_block
= after_block
;
2463 else_block
= vir_new_block(c
);
2465 bool was_uniform_control_flow
= false;
2466 if (!vir_in_nonuniform_control_flow(c
)) {
2467 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
2468 was_uniform_control_flow
= true;
2471 /* Set up the flags for the IF condition (taking the THEN branch). */
2472 enum v3d_qpu_cond cond
= ntq_emit_bool_to_cond(c
, if_stmt
->condition
);
2474 /* Update the flags+cond to mean "Taking the ELSE branch (!cond) and
2475 * was previously active (execute Z) for updating the exec flags.
2477 if (was_uniform_control_flow
) {
2478 cond
= v3d_qpu_cond_invert(cond
);
2480 struct qinst
*inst
= vir_MOV_dest(c
, vir_nop_reg(), c
->execute
);
2481 if (cond
== V3D_QPU_COND_IFA
) {
2482 vir_set_uf(inst
, V3D_QPU_UF_NORNZ
);
2484 vir_set_uf(inst
, V3D_QPU_UF_ANDZ
);
2485 cond
= V3D_QPU_COND_IFA
;
2489 vir_MOV_cond(c
, cond
,
2491 vir_uniform_ui(c
, else_block
->index
));
2493 /* Jump to ELSE if nothing is active for THEN, otherwise fall
2496 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
), V3D_QPU_PF_PUSHZ
);
2497 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLNA
);
2498 vir_link_blocks(c
->cur_block
, else_block
);
2499 vir_link_blocks(c
->cur_block
, then_block
);
2501 /* Process the THEN block. */
2502 vir_set_emit_block(c
, then_block
);
2503 ntq_emit_cf_list(c
, &if_stmt
->then_list
);
2505 if (!empty_else_block
) {
2506 /* Handle the end of the THEN block. First, all currently
2507 * active channels update their execute flags to point to
2510 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2512 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2513 vir_uniform_ui(c
, after_block
->index
));
2515 /* If everything points at ENDIF, then jump there immediately. */
2516 vir_set_pf(vir_XOR_dest(c
, vir_nop_reg(),
2518 vir_uniform_ui(c
, after_block
->index
)),
2520 vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ALLA
);
2521 vir_link_blocks(c
->cur_block
, after_block
);
2522 vir_link_blocks(c
->cur_block
, else_block
);
2524 vir_set_emit_block(c
, else_block
);
2525 ntq_activate_execute_for_block(c
);
2526 ntq_emit_cf_list(c
, &if_stmt
->else_list
);
2529 vir_link_blocks(c
->cur_block
, after_block
);
2531 vir_set_emit_block(c
, after_block
);
2532 if (was_uniform_control_flow
)
2533 c
->execute
= c
->undef
;
2535 ntq_activate_execute_for_block(c
);
2539 ntq_emit_if(struct v3d_compile
*c
, nir_if
*nif
)
2541 bool was_in_control_flow
= c
->in_control_flow
;
2542 c
->in_control_flow
= true;
2543 if (!vir_in_nonuniform_control_flow(c
) &&
2544 nir_src_is_dynamically_uniform(nif
->condition
)) {
2545 ntq_emit_uniform_if(c
, nif
);
2547 ntq_emit_nonuniform_if(c
, nif
);
2549 c
->in_control_flow
= was_in_control_flow
;
2553 ntq_emit_jump(struct v3d_compile
*c
, nir_jump_instr
*jump
)
2555 switch (jump
->type
) {
2556 case nir_jump_break
:
2557 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2559 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2560 vir_uniform_ui(c
, c
->loop_break_block
->index
));
2563 case nir_jump_continue
:
2564 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
),
2566 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
,
2567 vir_uniform_ui(c
, c
->loop_cont_block
->index
));
2570 case nir_jump_return
:
2571 unreachable("All returns shouold be lowered\n");
2576 ntq_emit_instr(struct v3d_compile
*c
, nir_instr
*instr
)
2578 switch (instr
->type
) {
2579 case nir_instr_type_alu
:
2580 ntq_emit_alu(c
, nir_instr_as_alu(instr
));
2583 case nir_instr_type_intrinsic
:
2584 ntq_emit_intrinsic(c
, nir_instr_as_intrinsic(instr
));
2587 case nir_instr_type_load_const
:
2588 ntq_emit_load_const(c
, nir_instr_as_load_const(instr
));
2591 case nir_instr_type_ssa_undef
:
2592 ntq_emit_ssa_undef(c
, nir_instr_as_ssa_undef(instr
));
2595 case nir_instr_type_tex
:
2596 ntq_emit_tex(c
, nir_instr_as_tex(instr
));
2599 case nir_instr_type_jump
:
2600 ntq_emit_jump(c
, nir_instr_as_jump(instr
));
2604 fprintf(stderr
, "Unknown NIR instr type: ");
2605 nir_print_instr(instr
, stderr
);
2606 fprintf(stderr
, "\n");
2612 ntq_emit_block(struct v3d_compile
*c
, nir_block
*block
)
2614 nir_foreach_instr(instr
, block
) {
2615 ntq_emit_instr(c
, instr
);
2619 static void ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
);
2622 ntq_emit_loop(struct v3d_compile
*c
, nir_loop
*loop
)
2624 bool was_in_control_flow
= c
->in_control_flow
;
2625 c
->in_control_flow
= true;
2627 bool was_uniform_control_flow
= false;
2628 if (!vir_in_nonuniform_control_flow(c
)) {
2629 c
->execute
= vir_MOV(c
, vir_uniform_ui(c
, 0));
2630 was_uniform_control_flow
= true;
2633 struct qblock
*save_loop_cont_block
= c
->loop_cont_block
;
2634 struct qblock
*save_loop_break_block
= c
->loop_break_block
;
2636 c
->loop_cont_block
= vir_new_block(c
);
2637 c
->loop_break_block
= vir_new_block(c
);
2639 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
2640 vir_set_emit_block(c
, c
->loop_cont_block
);
2641 ntq_activate_execute_for_block(c
);
2643 ntq_emit_cf_list(c
, &loop
->body
);
2645 /* Re-enable any previous continues now, so our ANYA check below
2648 * XXX: Use the .ORZ flags update, instead.
2650 vir_set_pf(vir_XOR_dest(c
,
2653 vir_uniform_ui(c
, c
->loop_cont_block
->index
)),
2655 vir_MOV_cond(c
, V3D_QPU_COND_IFA
, c
->execute
, vir_uniform_ui(c
, 0));
2657 vir_set_pf(vir_MOV_dest(c
, vir_nop_reg(), c
->execute
), V3D_QPU_PF_PUSHZ
);
2659 struct qinst
*branch
= vir_BRANCH(c
, V3D_QPU_BRANCH_COND_ANYA
);
2660 /* Pixels that were not dispatched or have been discarded should not
2661 * contribute to looping again.
2663 branch
->qpu
.branch
.msfign
= V3D_QPU_MSFIGN_P
;
2664 vir_link_blocks(c
->cur_block
, c
->loop_cont_block
);
2665 vir_link_blocks(c
->cur_block
, c
->loop_break_block
);
2667 vir_set_emit_block(c
, c
->loop_break_block
);
2668 if (was_uniform_control_flow
)
2669 c
->execute
= c
->undef
;
2671 ntq_activate_execute_for_block(c
);
2673 c
->loop_break_block
= save_loop_break_block
;
2674 c
->loop_cont_block
= save_loop_cont_block
;
2678 c
->in_control_flow
= was_in_control_flow
;
2682 ntq_emit_function(struct v3d_compile
*c
, nir_function_impl
*func
)
2684 fprintf(stderr
, "FUNCTIONS not handled.\n");
2689 ntq_emit_cf_list(struct v3d_compile
*c
, struct exec_list
*list
)
2691 foreach_list_typed(nir_cf_node
, node
, node
, list
) {
2692 switch (node
->type
) {
2693 case nir_cf_node_block
:
2694 ntq_emit_block(c
, nir_cf_node_as_block(node
));
2697 case nir_cf_node_if
:
2698 ntq_emit_if(c
, nir_cf_node_as_if(node
));
2701 case nir_cf_node_loop
:
2702 ntq_emit_loop(c
, nir_cf_node_as_loop(node
));
2705 case nir_cf_node_function
:
2706 ntq_emit_function(c
, nir_cf_node_as_function(node
));
2710 fprintf(stderr
, "Unknown NIR node type\n");
2717 ntq_emit_impl(struct v3d_compile
*c
, nir_function_impl
*impl
)
2719 ntq_setup_registers(c
, &impl
->registers
);
2720 ntq_emit_cf_list(c
, &impl
->body
);
2724 nir_to_vir(struct v3d_compile
*c
)
2726 switch (c
->s
->info
.stage
) {
2727 case MESA_SHADER_FRAGMENT
:
2728 c
->payload_w
= vir_MOV(c
, vir_reg(QFILE_REG
, 0));
2729 c
->payload_w_centroid
= vir_MOV(c
, vir_reg(QFILE_REG
, 1));
2730 c
->payload_z
= vir_MOV(c
, vir_reg(QFILE_REG
, 2));
2732 /* V3D 4.x can disable implicit point coordinate varyings if
2733 * they are not used.
2735 if (c
->fs_key
->is_points
&&
2736 (c
->devinfo
->ver
< 40 || program_reads_point_coord(c
))) {
2737 c
->point_x
= emit_fragment_varying(c
, NULL
, 0, 0);
2738 c
->point_y
= emit_fragment_varying(c
, NULL
, 0, 0);
2739 c
->uses_implicit_point_line_varyings
= true;
2740 } else if (c
->fs_key
->is_lines
&&
2741 (c
->devinfo
->ver
< 40 ||
2742 (c
->s
->info
.system_values_read
&
2743 BITFIELD64_BIT(SYSTEM_VALUE_LINE_COORD
)))) {
2744 c
->line_x
= emit_fragment_varying(c
, NULL
, 0, 0);
2745 c
->uses_implicit_point_line_varyings
= true;
2748 case MESA_SHADER_COMPUTE
:
2749 /* Set up the TSO for barriers, assuming we do some. */
2750 if (c
->devinfo
->ver
< 42) {
2751 vir_BARRIERID_dest(c
, vir_reg(QFILE_MAGIC
,
2752 V3D_QPU_WADDR_SYNC
));
2755 c
->cs_payload
[0] = vir_MOV(c
, vir_reg(QFILE_REG
, 0));
2756 c
->cs_payload
[1] = vir_MOV(c
, vir_reg(QFILE_REG
, 2));
2758 /* Set up the division between gl_LocalInvocationIndex and
2759 * wg_in_mem in the payload reg.
2761 int wg_size
= (c
->s
->info
.cs
.local_size
[0] *
2762 c
->s
->info
.cs
.local_size
[1] *
2763 c
->s
->info
.cs
.local_size
[2]);
2764 c
->local_invocation_index_bits
=
2765 ffs(util_next_power_of_two(MAX2(wg_size
, 64))) - 1;
2766 assert(c
->local_invocation_index_bits
<= 8);
2768 if (c
->s
->info
.cs
.shared_size
) {
2769 struct qreg wg_in_mem
= vir_SHR(c
, c
->cs_payload
[1],
2770 vir_uniform_ui(c
, 16));
2771 if (c
->s
->info
.cs
.local_size
[0] != 1 ||
2772 c
->s
->info
.cs
.local_size
[1] != 1 ||
2773 c
->s
->info
.cs
.local_size
[2] != 1) {
2775 c
->local_invocation_index_bits
);
2776 int wg_mask
= (1 << wg_bits
) - 1;
2777 wg_in_mem
= vir_AND(c
, wg_in_mem
,
2778 vir_uniform_ui(c
, wg_mask
));
2780 struct qreg shared_per_wg
=
2781 vir_uniform_ui(c
, c
->s
->info
.cs
.shared_size
);
2783 c
->cs_shared_offset
=
2785 vir_uniform(c
, QUNIFORM_SHARED_OFFSET
,0),
2786 vir_UMUL(c
, wg_in_mem
, shared_per_wg
));
2793 if (c
->s
->scratch_size
) {
2794 v3d_setup_spill_base(c
);
2795 c
->spill_size
+= V3D_CHANNELS
* c
->s
->scratch_size
;
2798 switch (c
->s
->info
.stage
) {
2799 case MESA_SHADER_VERTEX
:
2800 ntq_setup_vs_inputs(c
);
2802 case MESA_SHADER_GEOMETRY
:
2803 ntq_setup_gs_inputs(c
);
2805 case MESA_SHADER_FRAGMENT
:
2806 ntq_setup_fs_inputs(c
);
2808 case MESA_SHADER_COMPUTE
:
2811 unreachable("unsupported shader stage");
2814 ntq_setup_outputs(c
);
2816 /* Find the main function and emit the body. */
2817 nir_foreach_function(function
, c
->s
) {
2818 assert(strcmp(function
->name
, "main") == 0);
2819 assert(function
->impl
);
2820 ntq_emit_impl(c
, function
->impl
);
2824 const nir_shader_compiler_options v3d_nir_options
= {
2825 .lower_all_io_to_temps
= true,
2826 .lower_extract_byte
= true,
2827 .lower_extract_word
= true,
2828 .lower_bitfield_insert_to_shifts
= true,
2829 .lower_bitfield_extract_to_shifts
= true,
2830 .lower_bitfield_reverse
= true,
2831 .lower_bit_count
= true,
2832 .lower_cs_local_id_from_index
= true,
2833 .lower_ffract
= true,
2835 .lower_pack_unorm_2x16
= true,
2836 .lower_pack_snorm_2x16
= true,
2837 .lower_pack_unorm_4x8
= true,
2838 .lower_pack_snorm_4x8
= true,
2839 .lower_unpack_unorm_4x8
= true,
2840 .lower_unpack_snorm_4x8
= true,
2841 .lower_pack_half_2x16
= true,
2842 .lower_unpack_half_2x16
= true,
2844 .lower_find_lsb
= true,
2846 .lower_flrp32
= true,
2849 .lower_fsqrt
= true,
2850 .lower_ifind_msb
= true,
2851 .lower_isign
= true,
2852 .lower_ldexp
= true,
2853 .lower_mul_high
= true,
2854 .lower_wpos_pntc
= true,
2855 .lower_rotate
= true,
2856 .lower_to_scalar
= true,
2860 * When demoting a shader down to single-threaded, removes the THRSW
2861 * instructions (one will still be inserted at v3d_vir_to_qpu() for the
2865 vir_remove_thrsw(struct v3d_compile
*c
)
2867 vir_for_each_block(block
, c
) {
2868 vir_for_each_inst_safe(inst
, block
) {
2869 if (inst
->qpu
.sig
.thrsw
)
2870 vir_remove_instruction(c
, inst
);
2874 c
->last_thrsw
= NULL
;
2878 vir_emit_last_thrsw(struct v3d_compile
*c
)
2880 /* On V3D before 4.1, we need a TMU op to be outstanding when thread
2881 * switching, so disable threads if we didn't do any TMU ops (each of
2882 * which would have emitted a THRSW).
2884 if (!c
->last_thrsw_at_top_level
&& c
->devinfo
->ver
< 41) {
2887 vir_remove_thrsw(c
);
2891 /* If we're threaded and the last THRSW was in conditional code, then
2892 * we need to emit another one so that we can flag it as the last
2895 if (c
->last_thrsw
&& !c
->last_thrsw_at_top_level
) {
2896 assert(c
->devinfo
->ver
>= 41);
2900 /* If we're threaded, then we need to mark the last THRSW instruction
2901 * so we can emit a pair of them at QPU emit time.
2903 * For V3D 4.x, we can spawn the non-fragment shaders already in the
2904 * post-last-THRSW state, so we can skip this.
2906 if (!c
->last_thrsw
&& c
->s
->info
.stage
== MESA_SHADER_FRAGMENT
) {
2907 assert(c
->devinfo
->ver
>= 41);
2912 c
->last_thrsw
->is_last_thrsw
= true;
2915 /* There's a flag in the shader for "center W is needed for reasons other than
2916 * non-centroid varyings", so we just walk the program after VIR optimization
2917 * to see if it's used. It should be harmless to set even if we only use
2918 * center W for varyings.
2921 vir_check_payload_w(struct v3d_compile
*c
)
2923 if (c
->s
->info
.stage
!= MESA_SHADER_FRAGMENT
)
2926 vir_for_each_inst_inorder(inst
, c
) {
2927 for (int i
= 0; i
< vir_get_nsrc(inst
); i
++) {
2928 if (inst
->src
[i
].file
== QFILE_REG
&&
2929 inst
->src
[i
].index
== 0) {
2930 c
->uses_center_w
= true;
2939 v3d_nir_to_vir(struct v3d_compile
*c
)
2941 if (V3D_DEBUG
& (V3D_DEBUG_NIR
|
2942 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2943 fprintf(stderr
, "%s prog %d/%d NIR:\n",
2944 vir_get_stage_name(c
),
2945 c
->program_id
, c
->variant_id
);
2946 nir_print_shader(c
->s
, stderr
);
2951 /* Emit the last THRSW before STVPM and TLB writes. */
2952 vir_emit_last_thrsw(c
);
2954 switch (c
->s
->info
.stage
) {
2955 case MESA_SHADER_FRAGMENT
:
2958 case MESA_SHADER_GEOMETRY
:
2961 case MESA_SHADER_VERTEX
:
2964 case MESA_SHADER_COMPUTE
:
2967 unreachable("bad stage");
2970 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
2971 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2972 fprintf(stderr
, "%s prog %d/%d pre-opt VIR:\n",
2973 vir_get_stage_name(c
),
2974 c
->program_id
, c
->variant_id
);
2976 fprintf(stderr
, "\n");
2981 vir_check_payload_w(c
);
2983 /* XXX perf: On VC4, we do a VIR-level instruction scheduling here.
2984 * We used that on that platform to pipeline TMU writes and reduce the
2985 * number of thread switches, as well as try (mostly successfully) to
2986 * reduce maximum register pressure to allow more threads. We should
2987 * do something of that sort for V3D -- either instruction scheduling
2988 * here, or delay the the THRSW and LDTMUs from our texture
2989 * instructions until the results are needed.
2992 if (V3D_DEBUG
& (V3D_DEBUG_VIR
|
2993 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
))) {
2994 fprintf(stderr
, "%s prog %d/%d VIR:\n",
2995 vir_get_stage_name(c
),
2996 c
->program_id
, c
->variant_id
);
2998 fprintf(stderr
, "\n");
3001 /* Attempt to allocate registers for the temporaries. If we fail,
3002 * reduce thread count and try again.
3004 int min_threads
= (c
->devinfo
->ver
>= 41) ? 2 : 1;
3005 struct qpu_reg
*temp_registers
;
3008 temp_registers
= v3d_register_allocate(c
, &spilled
);
3015 if (c
->threads
== min_threads
) {
3016 fprintf(stderr
, "Failed to register allocate at %d threads:\n",
3025 if (c
->threads
== 1)
3026 vir_remove_thrsw(c
);
3030 (V3D_DEBUG
& (V3D_DEBUG_VIR
|
3031 v3d_debug_flag_for_shader_stage(c
->s
->info
.stage
)))) {
3032 fprintf(stderr
, "%s prog %d/%d spilled VIR:\n",
3033 vir_get_stage_name(c
),
3034 c
->program_id
, c
->variant_id
);
3036 fprintf(stderr
, "\n");
3039 v3d_vir_to_qpu(c
, temp_registers
);