2 * Copyright © 2018 Valve Corporation
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
30 #include "vulkan/radv_shader.h"
37 * The general idea of this pass is:
38 * The CFG is traversed in reverse postorder (forward) and loops are processed
39 * several times until no progress is made.
40 * Per BB two wait_ctx is maintained: an in-context and out-context.
41 * The in-context is the joined out-contexts of the predecessors.
42 * The context contains a map: gpr -> wait_entry
43 * consisting of the information about the cnt values to be waited for.
44 * Note: After merge-nodes, it might occur that for the same register
45 * multiple cnt values are to be waited for.
47 * The values are updated according to the encountered instructions:
48 * - additional events increment the counter of waits of the same type
49 * - or erase gprs with counters higher than to be waited for.
52 // TODO: do a more clever insertion of wait_cnt (lgkm_cnt) when there is a load followed by a use of a previous load
54 /* Instructions of the same event will finish in-order except for smem
55 * and maybe flat. Instructions of different events may not finish in-order. */
56 enum wait_event
: uint16_t {
61 event_vmem_store
= 1 << 4, /* GFX10+ */
63 event_exp_pos
= 1 << 6,
64 event_exp_param
= 1 << 7,
65 event_exp_mrt_null
= 1 << 8,
66 event_gds_gpr_lock
= 1 << 9,
67 event_vmem_gpr_lock
= 1 << 10,
68 event_sendmsg
= 1 << 11,
71 enum counter_type
: uint8_t {
73 counter_lgkm
= 1 << 1,
78 static const uint16_t exp_events
= event_exp_pos
| event_exp_param
| event_exp_mrt_null
| event_gds_gpr_lock
| event_vmem_gpr_lock
;
79 static const uint16_t lgkm_events
= event_smem
| event_lds
| event_gds
| event_flat
| event_sendmsg
;
80 static const uint16_t vm_events
= event_vmem
| event_flat
;
81 static const uint16_t vs_events
= event_vmem_store
;
83 uint8_t get_counters_for_event(wait_event ev
)
93 case event_vmem_store
:
96 return counter_vm
| counter_lgkm
;
99 case event_exp_mrt_null
:
100 case event_gds_gpr_lock
:
101 case event_vmem_gpr_lock
:
109 static const uint8_t unset_counter
= 0xff;
117 vm(unset_counter
), exp(unset_counter
), lgkm(unset_counter
), vs(unset_counter
) {}
118 wait_imm(uint16_t vm_
, uint16_t exp_
, uint16_t lgkm_
, uint16_t vs_
) :
119 vm(vm_
), exp(exp_
), lgkm(lgkm_
), vs(vs_
) {}
121 wait_imm(enum chip_class chip
, uint16_t packed
) : vs(unset_counter
)
125 vm
|= (packed
>> 10) & 0x30;
127 exp
= (packed
>> 4) & 0x7;
129 lgkm
= (packed
>> 8) & 0xf;
131 lgkm
|= (packed
>> 8) & 0x30;
134 uint16_t pack(enum chip_class chip
) const
137 assert(exp
== unset_counter
|| exp
<= 0x7);
140 assert(lgkm
== unset_counter
|| lgkm
<= 0x3f);
141 assert(vm
== unset_counter
|| vm
<= 0x3f);
142 imm
= ((vm
& 0x30) << 10) | ((lgkm
& 0x3f) << 8) | ((exp
& 0x7) << 4) | (vm
& 0xf);
145 assert(lgkm
== unset_counter
|| lgkm
<= 0xf);
146 assert(vm
== unset_counter
|| vm
<= 0x3f);
147 imm
= ((vm
& 0x30) << 10) | ((lgkm
& 0xf) << 8) | ((exp
& 0x7) << 4) | (vm
& 0xf);
150 assert(lgkm
== unset_counter
|| lgkm
<= 0xf);
151 assert(vm
== unset_counter
|| vm
<= 0xf);
152 imm
= ((lgkm
& 0xf) << 8) | ((exp
& 0x7) << 4) | (vm
& 0xf);
155 if (chip
< GFX9
&& vm
== wait_imm::unset_counter
)
156 imm
|= 0xc000; /* should have no effect on pre-GFX9 and now we won't have to worry about the architecture when interpreting the immediate */
157 if (chip
< GFX10
&& lgkm
== wait_imm::unset_counter
)
158 imm
|= 0x3000; /* should have no effect on pre-GFX10 and now we won't have to worry about the architecture when interpreting the immediate */
162 bool combine(const wait_imm
& other
)
164 bool changed
= other
.vm
< vm
|| other
.exp
< exp
|| other
.lgkm
< lgkm
|| other
.vs
< vs
;
165 vm
= std::min(vm
, other
.vm
);
166 exp
= std::min(exp
, other
.exp
);
167 lgkm
= std::min(lgkm
, other
.lgkm
);
168 vs
= std::min(vs
, other
.vs
);
174 return vm
== unset_counter
&& exp
== unset_counter
&&
175 lgkm
== unset_counter
&& vs
== unset_counter
;
181 uint16_t events
; /* use wait_event notion */
182 uint8_t counters
; /* use counter_type notion */
186 wait_entry(wait_event event
, wait_imm imm
, bool logical
, bool wait_on_read
)
187 : imm(imm
), events(event
), counters(get_counters_for_event(event
)),
188 wait_on_read(wait_on_read
), logical(logical
) {}
190 bool join(const wait_entry
& other
)
192 bool changed
= (other
.events
& ~events
) ||
193 (other
.counters
& ~counters
) ||
194 (other
.wait_on_read
&& !wait_on_read
);
195 events
|= other
.events
;
196 counters
|= other
.counters
;
197 changed
|= imm
.combine(other
.imm
);
198 wait_on_read
= wait_on_read
|| other
.wait_on_read
;
199 assert(logical
== other
.logical
);
203 void remove_counter(counter_type counter
)
205 counters
&= ~counter
;
207 if (counter
== counter_lgkm
) {
208 imm
.lgkm
= wait_imm::unset_counter
;
209 events
&= ~(event_smem
| event_lds
| event_gds
| event_sendmsg
);
212 if (counter
== counter_vm
) {
213 imm
.vm
= wait_imm::unset_counter
;
214 events
&= ~event_vmem
;
217 if (counter
== counter_exp
) {
218 imm
.exp
= wait_imm::unset_counter
;
219 events
&= ~(event_exp_pos
| event_exp_param
| event_exp_mrt_null
| event_gds_gpr_lock
| event_vmem_gpr_lock
);
222 if (counter
== counter_vs
) {
223 imm
.vs
= wait_imm::unset_counter
;
224 events
&= ~event_vmem_store
;
227 if (!(counters
& counter_lgkm
) && !(counters
& counter_vm
))
228 events
&= ~event_flat
;
234 enum chip_class chip_class
;
236 uint16_t max_exp_cnt
;
237 uint16_t max_lgkm_cnt
;
239 uint16_t unordered_events
= event_smem
| event_flat
;
243 uint8_t lgkm_cnt
= 0;
245 bool pending_flat_lgkm
= false;
246 bool pending_flat_vm
= false;
247 bool pending_s_buffer_store
= false; /* GFX10 workaround */
249 wait_imm barrier_imm
[barrier_count
];
251 std::map
<PhysReg
,wait_entry
> gpr_map
;
254 wait_ctx(Program
*program_
)
256 chip_class(program_
->chip_class
),
257 max_vm_cnt(program_
->chip_class
>= GFX9
? 62 : 14),
259 max_lgkm_cnt(program_
->chip_class
>= GFX10
? 62 : 14),
260 max_vs_cnt(program_
->chip_class
>= GFX10
? 62 : 0),
261 unordered_events(event_smem
| (program_
->chip_class
< GFX10
? event_flat
: 0)) {}
263 bool join(const wait_ctx
* other
, bool logical
)
265 bool changed
= other
->exp_cnt
> exp_cnt
||
266 other
->vm_cnt
> vm_cnt
||
267 other
->lgkm_cnt
> lgkm_cnt
||
268 other
->vs_cnt
> vs_cnt
||
269 (other
->pending_flat_lgkm
&& !pending_flat_lgkm
) ||
270 (other
->pending_flat_vm
&& !pending_flat_vm
);
272 exp_cnt
= std::max(exp_cnt
, other
->exp_cnt
);
273 vm_cnt
= std::max(vm_cnt
, other
->vm_cnt
);
274 lgkm_cnt
= std::max(lgkm_cnt
, other
->lgkm_cnt
);
275 vs_cnt
= std::max(vs_cnt
, other
->vs_cnt
);
276 pending_flat_lgkm
|= other
->pending_flat_lgkm
;
277 pending_flat_vm
|= other
->pending_flat_vm
;
278 pending_s_buffer_store
|= other
->pending_s_buffer_store
;
280 for (std::pair
<PhysReg
,wait_entry
> entry
: other
->gpr_map
)
282 std::map
<PhysReg
,wait_entry
>::iterator it
= gpr_map
.find(entry
.first
);
283 if (entry
.second
.logical
!= logical
)
286 if (it
!= gpr_map
.end()) {
287 changed
|= it
->second
.join(entry
.second
);
289 gpr_map
.insert(entry
);
294 for (unsigned i
= 0; i
< barrier_count
; i
++)
295 changed
|= barrier_imm
[i
].combine(other
->barrier_imm
[i
]);
301 wait_imm
check_instr(Instruction
* instr
, wait_ctx
& ctx
)
305 for (const Operand op
: instr
->operands
) {
306 if (op
.isConstant() || op
.isUndefined())
309 /* check consecutively read gprs */
310 for (unsigned j
= 0; j
< op
.size(); j
++) {
311 PhysReg reg
{op
.physReg() + j
};
312 std::map
<PhysReg
,wait_entry
>::iterator it
= ctx
.gpr_map
.find(reg
);
313 if (it
== ctx
.gpr_map
.end() || !it
->second
.wait_on_read
)
316 wait
.combine(it
->second
.imm
);
320 for (const Definition
& def
: instr
->definitions
) {
321 /* check consecutively written gprs */
322 for (unsigned j
= 0; j
< def
.getTemp().size(); j
++)
324 PhysReg reg
{def
.physReg() + j
};
326 std::map
<PhysReg
,wait_entry
>::iterator it
= ctx
.gpr_map
.find(reg
);
327 if (it
== ctx
.gpr_map
.end())
330 /* Vector Memory reads and writes return in the order they were issued */
331 if (instr
->isVMEM() && ((it
->second
.events
& vm_events
) == event_vmem
)) {
332 it
->second
.remove_counter(counter_vm
);
333 if (!it
->second
.counters
)
334 it
= ctx
.gpr_map
.erase(it
);
338 /* LDS reads and writes return in the order they were issued. same for GDS */
339 if (instr
->format
== Format::DS
) {
340 bool gds
= static_cast<DS_instruction
*>(instr
)->gds
;
341 if ((it
->second
.events
& lgkm_events
) == (gds
? event_gds
: event_lds
)) {
342 it
->second
.remove_counter(counter_lgkm
);
343 if (!it
->second
.counters
)
344 it
= ctx
.gpr_map
.erase(it
);
349 wait
.combine(it
->second
.imm
);
356 wait_imm
parse_wait_instr(wait_ctx
& ctx
, Instruction
*instr
)
358 if (instr
->opcode
== aco_opcode::s_waitcnt_vscnt
&&
359 instr
->definitions
[0].physReg() == sgpr_null
) {
361 imm
.vs
= std::min
<uint8_t>(imm
.vs
, static_cast<SOPK_instruction
*>(instr
)->imm
);
363 } else if (instr
->opcode
== aco_opcode::s_waitcnt
) {
364 return wait_imm(ctx
.chip_class
, static_cast<SOPP_instruction
*>(instr
)->imm
);
369 wait_imm
kill(Instruction
* instr
, wait_ctx
& ctx
)
372 if (ctx
.exp_cnt
|| ctx
.vm_cnt
|| ctx
.lgkm_cnt
)
373 imm
.combine(check_instr(instr
, ctx
));
375 imm
.combine(parse_wait_instr(ctx
, instr
));
377 if (ctx
.chip_class
>= GFX10
) {
378 /* Seems to be required on GFX10 to achieve correct behaviour.
379 * It shouldn't cost anything anyways since we're about to do s_endpgm.
381 if (ctx
.lgkm_cnt
&& instr
->opcode
== aco_opcode::s_dcache_wb
)
384 /* GFX10: A store followed by a load at the same address causes a problem because
385 * the load doesn't load the correct values unless we wait for the store first.
386 * This is NOT mitigated by an s_nop.
388 * TODO: Refine this when we have proper alias analysis.
390 SMEM_instruction
*smem
= static_cast<SMEM_instruction
*>(instr
);
391 if (ctx
.pending_s_buffer_store
&&
392 !smem
->definitions
.empty() &&
393 !smem
->can_reorder
&& smem
->barrier
== barrier_buffer
) {
398 if (instr
->format
== Format::PSEUDO_BARRIER
) {
399 uint32_t workgroup_size
= UINT32_MAX
;
400 if (ctx
.program
->stage
& sw_cs
) {
401 unsigned* bsize
= ctx
.program
->info
->cs
.block_size
;
402 workgroup_size
= bsize
[0] * bsize
[1] * bsize
[2];
404 switch (instr
->opcode
) {
405 case aco_opcode::p_memory_barrier_common
:
406 imm
.combine(ctx
.barrier_imm
[ffs(barrier_atomic
) - 1]);
407 imm
.combine(ctx
.barrier_imm
[ffs(barrier_buffer
) - 1]);
408 imm
.combine(ctx
.barrier_imm
[ffs(barrier_image
) - 1]);
409 if (workgroup_size
> ctx
.program
->wave_size
)
410 imm
.combine(ctx
.barrier_imm
[ffs(barrier_shared
) - 1]);
412 case aco_opcode::p_memory_barrier_atomic
:
413 imm
.combine(ctx
.barrier_imm
[ffs(barrier_atomic
) - 1]);
415 /* see comment in aco_scheduler.cpp's can_move_instr() on why these barriers are merged */
416 case aco_opcode::p_memory_barrier_buffer
:
417 case aco_opcode::p_memory_barrier_image
:
418 imm
.combine(ctx
.barrier_imm
[ffs(barrier_buffer
) - 1]);
419 imm
.combine(ctx
.barrier_imm
[ffs(barrier_image
) - 1]);
421 case aco_opcode::p_memory_barrier_shared
:
422 if (workgroup_size
> ctx
.program
->wave_size
)
423 imm
.combine(ctx
.barrier_imm
[ffs(barrier_shared
) - 1]);
425 case aco_opcode::p_memory_barrier_gs_data
:
426 imm
.combine(ctx
.barrier_imm
[ffs(barrier_gs_data
) - 1]);
428 case aco_opcode::p_memory_barrier_gs_sendmsg
:
429 imm
.combine(ctx
.barrier_imm
[ffs(barrier_gs_sendmsg
) - 1]);
438 if (ctx
.pending_flat_vm
&& imm
.vm
!= wait_imm::unset_counter
)
440 if (ctx
.pending_flat_lgkm
&& imm
.lgkm
!= wait_imm::unset_counter
)
444 ctx
.exp_cnt
= std::min(ctx
.exp_cnt
, imm
.exp
);
445 ctx
.vm_cnt
= std::min(ctx
.vm_cnt
, imm
.vm
);
446 ctx
.lgkm_cnt
= std::min(ctx
.lgkm_cnt
, imm
.lgkm
);
447 ctx
.vs_cnt
= std::min(ctx
.vs_cnt
, imm
.vs
);
449 /* update barrier wait imms */
450 for (unsigned i
= 0; i
< barrier_count
; i
++) {
451 wait_imm
& bar
= ctx
.barrier_imm
[i
];
452 if (bar
.exp
!= wait_imm::unset_counter
&& imm
.exp
<= bar
.exp
)
453 bar
.exp
= wait_imm::unset_counter
;
454 if (bar
.vm
!= wait_imm::unset_counter
&& imm
.vm
<= bar
.vm
)
455 bar
.vm
= wait_imm::unset_counter
;
456 if (bar
.lgkm
!= wait_imm::unset_counter
&& imm
.lgkm
<= bar
.lgkm
)
457 bar
.lgkm
= wait_imm::unset_counter
;
458 if (bar
.vs
!= wait_imm::unset_counter
&& imm
.vs
<= bar
.vs
)
459 bar
.vs
= wait_imm::unset_counter
;
462 /* remove all gprs with higher counter from map */
463 std::map
<PhysReg
,wait_entry
>::iterator it
= ctx
.gpr_map
.begin();
464 while (it
!= ctx
.gpr_map
.end())
466 if (imm
.exp
!= wait_imm::unset_counter
&& imm
.exp
<= it
->second
.imm
.exp
)
467 it
->second
.remove_counter(counter_exp
);
468 if (imm
.vm
!= wait_imm::unset_counter
&& imm
.vm
<= it
->second
.imm
.vm
)
469 it
->second
.remove_counter(counter_vm
);
470 if (imm
.lgkm
!= wait_imm::unset_counter
&& imm
.lgkm
<= it
->second
.imm
.lgkm
)
471 it
->second
.remove_counter(counter_lgkm
);
472 if (imm
.lgkm
!= wait_imm::unset_counter
&& imm
.vs
<= it
->second
.imm
.vs
)
473 it
->second
.remove_counter(counter_vs
);
474 if (!it
->second
.counters
)
475 it
= ctx
.gpr_map
.erase(it
);
482 ctx
.pending_flat_vm
= false;
484 ctx
.pending_flat_lgkm
= false;
485 ctx
.pending_s_buffer_store
= false;
491 void update_barrier_imm(wait_ctx
& ctx
, uint8_t counters
, barrier_interaction barrier
)
493 unsigned barrier_index
= ffs(barrier
) - 1;
494 for (unsigned i
= 0; i
< barrier_count
; i
++) {
495 wait_imm
& bar
= ctx
.barrier_imm
[i
];
496 if (i
== barrier_index
) {
497 if (counters
& counter_lgkm
)
499 if (counters
& counter_vm
)
501 if (counters
& counter_exp
)
503 if (counters
& counter_vs
)
506 if (counters
& counter_lgkm
&& bar
.lgkm
!= wait_imm::unset_counter
&& bar
.lgkm
< ctx
.max_lgkm_cnt
)
508 if (counters
& counter_vm
&& bar
.vm
!= wait_imm::unset_counter
&& bar
.vm
< ctx
.max_vm_cnt
)
510 if (counters
& counter_exp
&& bar
.exp
!= wait_imm::unset_counter
&& bar
.exp
< ctx
.max_exp_cnt
)
512 if (counters
& counter_vs
&& bar
.vs
!= wait_imm::unset_counter
&& bar
.vs
< ctx
.max_vs_cnt
)
518 void update_counters(wait_ctx
& ctx
, wait_event event
, barrier_interaction barrier
=barrier_none
)
520 uint8_t counters
= get_counters_for_event(event
);
522 if (counters
& counter_lgkm
&& ctx
.lgkm_cnt
<= ctx
.max_lgkm_cnt
)
524 if (counters
& counter_vm
&& ctx
.vm_cnt
<= ctx
.max_vm_cnt
)
526 if (counters
& counter_exp
&& ctx
.exp_cnt
<= ctx
.max_exp_cnt
)
528 if (counters
& counter_vs
&& ctx
.vs_cnt
<= ctx
.max_vs_cnt
)
531 update_barrier_imm(ctx
, counters
, barrier
);
533 if (ctx
.unordered_events
& event
)
536 if (ctx
.pending_flat_lgkm
)
537 counters
&= ~counter_lgkm
;
538 if (ctx
.pending_flat_vm
)
539 counters
&= ~counter_vm
;
541 for (std::pair
<const PhysReg
,wait_entry
>& e
: ctx
.gpr_map
) {
542 wait_entry
& entry
= e
.second
;
544 if (entry
.events
& ctx
.unordered_events
)
547 assert(entry
.events
);
549 if ((counters
& counter_exp
) && (entry
.events
& exp_events
) == event
&& entry
.imm
.exp
< ctx
.max_exp_cnt
)
551 if ((counters
& counter_lgkm
) && (entry
.events
& lgkm_events
) == event
&& entry
.imm
.lgkm
< ctx
.max_lgkm_cnt
)
553 if ((counters
& counter_vm
) && (entry
.events
& vm_events
) == event
&& entry
.imm
.vm
< ctx
.max_vm_cnt
)
555 if ((counters
& counter_vs
) && (entry
.events
& vs_events
) == event
&& entry
.imm
.vs
< ctx
.max_vs_cnt
)
560 void update_counters_for_flat_load(wait_ctx
& ctx
, barrier_interaction barrier
=barrier_none
)
562 assert(ctx
.chip_class
< GFX10
);
564 if (ctx
.lgkm_cnt
<= ctx
.max_lgkm_cnt
)
566 if (ctx
.vm_cnt
<= ctx
.max_vm_cnt
)
569 update_barrier_imm(ctx
, counter_vm
| counter_lgkm
, barrier
);
571 for (std::pair
<PhysReg
,wait_entry
> e
: ctx
.gpr_map
)
573 if (e
.second
.counters
& counter_vm
)
575 if (e
.second
.counters
& counter_lgkm
)
576 e
.second
.imm
.lgkm
= 0;
578 ctx
.pending_flat_lgkm
= true;
579 ctx
.pending_flat_vm
= true;
582 void insert_wait_entry(wait_ctx
& ctx
, PhysReg reg
, RegClass rc
, wait_event event
, bool wait_on_read
)
584 uint16_t counters
= get_counters_for_event(event
);
586 if (counters
& counter_lgkm
)
588 if (counters
& counter_vm
)
590 if (counters
& counter_exp
)
592 if (counters
& counter_vs
)
595 wait_entry
new_entry(event
, imm
, !rc
.is_linear(), wait_on_read
);
597 for (unsigned i
= 0; i
< rc
.size(); i
++) {
598 auto it
= ctx
.gpr_map
.emplace(PhysReg
{reg
.reg
+i
}, new_entry
);
600 it
.first
->second
.join(new_entry
);
604 void insert_wait_entry(wait_ctx
& ctx
, Operand op
, wait_event event
)
606 if (!op
.isConstant() && !op
.isUndefined())
607 insert_wait_entry(ctx
, op
.physReg(), op
.regClass(), event
, false);
610 void insert_wait_entry(wait_ctx
& ctx
, Definition def
, wait_event event
)
612 insert_wait_entry(ctx
, def
.physReg(), def
.regClass(), event
, true);
615 void gen(Instruction
* instr
, wait_ctx
& ctx
)
617 switch (instr
->format
) {
619 Export_instruction
* exp_instr
= static_cast<Export_instruction
*>(instr
);
622 if (exp_instr
->dest
<= 9)
623 ev
= event_exp_mrt_null
;
624 else if (exp_instr
->dest
<= 15)
627 ev
= event_exp_param
;
628 update_counters(ctx
, ev
);
630 /* insert new entries for exported vgprs */
631 for (unsigned i
= 0; i
< 4; i
++)
633 if (exp_instr
->enabled_mask
& (1 << i
)) {
634 unsigned idx
= exp_instr
->compressed
? i
>> 1 : i
;
635 assert(idx
< exp_instr
->operands
.size());
636 insert_wait_entry(ctx
, exp_instr
->operands
[idx
], ev
);
640 insert_wait_entry(ctx
, exec
, s2
, ev
, false);
644 if (ctx
.chip_class
< GFX10
&& !instr
->definitions
.empty())
645 update_counters_for_flat_load(ctx
, barrier_buffer
);
647 update_counters(ctx
, event_flat
, barrier_buffer
);
649 if (!instr
->definitions
.empty())
650 insert_wait_entry(ctx
, instr
->definitions
[0], event_flat
);
654 SMEM_instruction
*smem
= static_cast<SMEM_instruction
*>(instr
);
655 update_counters(ctx
, event_smem
, static_cast<SMEM_instruction
*>(instr
)->barrier
);
657 if (!instr
->definitions
.empty())
658 insert_wait_entry(ctx
, instr
->definitions
[0], event_smem
);
659 else if (ctx
.chip_class
>= GFX10
&&
660 !smem
->can_reorder
&&
661 smem
->barrier
== barrier_buffer
)
662 ctx
.pending_s_buffer_store
= true;
667 bool gds
= static_cast<DS_instruction
*>(instr
)->gds
;
668 update_counters(ctx
, gds
? event_gds
: event_lds
, gds
? barrier_none
: barrier_shared
);
670 update_counters(ctx
, event_gds_gpr_lock
);
672 if (!instr
->definitions
.empty())
673 insert_wait_entry(ctx
, instr
->definitions
[0], gds
? event_gds
: event_lds
);
676 for (const Operand
& op
: instr
->operands
)
677 insert_wait_entry(ctx
, op
, event_gds_gpr_lock
);
678 insert_wait_entry(ctx
, exec
, s2
, event_gds_gpr_lock
, false);
685 case Format::GLOBAL
: {
686 wait_event ev
= !instr
->definitions
.empty() || ctx
.chip_class
< GFX10
? event_vmem
: event_vmem_store
;
687 update_counters(ctx
, ev
, get_barrier_interaction(instr
));
689 if (!instr
->definitions
.empty())
690 insert_wait_entry(ctx
, instr
->definitions
[0], ev
);
692 if (instr
->operands
.size() == 4 && ctx
.chip_class
== GFX6
) {
694 update_counters(ctx
, event_vmem_gpr_lock
);
695 insert_wait_entry(ctx
, instr
->operands
[3], event_vmem_gpr_lock
);
700 if (instr
->opcode
== aco_opcode::s_sendmsg
||
701 instr
->opcode
== aco_opcode::s_sendmsghalt
)
702 update_counters(ctx
, event_sendmsg
, get_barrier_interaction(instr
));
709 void emit_waitcnt(wait_ctx
& ctx
, std::vector
<aco_ptr
<Instruction
>>& instructions
, wait_imm imm
)
711 if (imm
.vs
!= wait_imm::unset_counter
) {
712 assert(ctx
.chip_class
>= GFX10
);
713 SOPK_instruction
* waitcnt_vs
= create_instruction
<SOPK_instruction
>(aco_opcode::s_waitcnt_vscnt
, Format::SOPK
, 0, 1);
714 waitcnt_vs
->definitions
[0] = Definition(sgpr_null
, s1
);
715 waitcnt_vs
->imm
= imm
.vs
;
716 instructions
.emplace_back(waitcnt_vs
);
717 imm
.vs
= wait_imm::unset_counter
;
720 SOPP_instruction
* waitcnt
= create_instruction
<SOPP_instruction
>(aco_opcode::s_waitcnt
, Format::SOPP
, 0, 0);
721 waitcnt
->imm
= imm
.pack(ctx
.chip_class
);
723 instructions
.emplace_back(waitcnt
);
727 void handle_block(Program
*program
, Block
& block
, wait_ctx
& ctx
)
729 std::vector
<aco_ptr
<Instruction
>> new_instructions
;
732 for (aco_ptr
<Instruction
>& instr
: block
.instructions
) {
733 bool is_wait
= !parse_wait_instr(ctx
, instr
.get()).empty();
735 queued_imm
.combine(kill(instr
.get(), ctx
));
737 gen(instr
.get(), ctx
);
739 if (instr
->format
!= Format::PSEUDO_BARRIER
&& !is_wait
) {
740 if (!queued_imm
.empty()) {
741 emit_waitcnt(ctx
, new_instructions
, queued_imm
);
742 queued_imm
= wait_imm();
744 new_instructions
.emplace_back(std::move(instr
));
748 if (!queued_imm
.empty())
749 emit_waitcnt(ctx
, new_instructions
, queued_imm
);
751 block
.instructions
.swap(new_instructions
);
754 } /* end namespace */
756 void insert_wait_states(Program
* program
)
759 std::vector
<bool> done(program
->blocks
.size());
760 wait_ctx in_ctx
[program
->blocks
.size()];
761 wait_ctx out_ctx
[program
->blocks
.size()];
762 for (unsigned i
= 0; i
< program
->blocks
.size(); i
++)
763 in_ctx
[i
] = wait_ctx(program
);
764 std::stack
<unsigned> loop_header_indices
;
765 unsigned loop_progress
= 0;
767 for (unsigned i
= 0; i
< program
->blocks
.size();) {
768 Block
& current
= program
->blocks
[i
++];
769 wait_ctx ctx
= in_ctx
[current
.index
];
771 if (current
.kind
& block_kind_loop_header
) {
772 loop_header_indices
.push(current
.index
);
773 } else if (current
.kind
& block_kind_loop_exit
) {
775 if (loop_progress
== loop_header_indices
.size()) {
776 i
= loop_header_indices
.top();
779 loop_header_indices
.pop();
780 loop_progress
= std::min
<unsigned>(loop_progress
, loop_header_indices
.size());
785 bool changed
= false;
786 for (unsigned b
: current
.linear_preds
)
787 changed
|= ctx
.join(&out_ctx
[b
], false);
788 for (unsigned b
: current
.logical_preds
)
789 changed
|= ctx
.join(&out_ctx
[b
], true);
791 in_ctx
[current
.index
] = ctx
;
793 if (done
[current
.index
] && !changed
)
796 if (current
.instructions
.empty()) {
797 out_ctx
[current
.index
] = ctx
;
801 loop_progress
= std::max
<unsigned>(loop_progress
, current
.loop_nest_depth
);
802 done
[current
.index
] = true;
804 handle_block(program
, current
, ctx
);
806 out_ctx
[current
.index
] = ctx
;