}
}
-static RegisterDemand getLiveChanges(aco_ptr<Instruction>& instr)
-{
- RegisterDemand changes;
- for (const Definition& def : instr->definitions) {
- if (!def.isTemp() || def.isKill())
- continue;
- changes += def.getTemp();
- }
-
- for (const Operand& op : instr->operands) {
- if (!op.isTemp() || !op.isFirstKill())
- continue;
- changes -= op.getTemp();
- }
-
- return changes;
-}
-
-static RegisterDemand getTempRegisters(aco_ptr<Instruction>& instr)
-{
- RegisterDemand temp_registers;
- for (const Definition& def : instr->definitions) {
- if (!def.isTemp() || !def.isKill())
- continue;
- temp_registers += def.getTemp();
- }
- return temp_registers;
-}
-
void MoveState::downwards_advance_helper()
{
source_idx--;
int dest_insert_idx = clause ? insert_idx_clause : insert_idx;
RegisterDemand register_pressure = clause ? total_demand_clause : total_demand;
- const RegisterDemand candidate_diff = getLiveChanges(instr);
- const RegisterDemand temp = getTempRegisters(instr);
+ const RegisterDemand candidate_diff = get_live_changes(instr);
+ const RegisterDemand temp = get_temp_registers(instr);
if (RegisterDemand(register_pressure - candidate_diff).exceeds(max_registers))
return move_fail_pressure;
- const RegisterDemand temp2 = getTempRegisters(block->instructions[dest_insert_idx - 1]);
+ const RegisterDemand temp2 = get_temp_registers(block->instructions[dest_insert_idx - 1]);
const RegisterDemand new_demand = register_demand[dest_insert_idx - 1] - temp2 + temp;
if (new_demand.exceeds(max_registers))
return move_fail_pressure;
}
/* check if register pressure is low enough: the diff is negative if register pressure is decreased */
- const RegisterDemand candidate_diff = getLiveChanges(instr);
- const RegisterDemand temp = getTempRegisters(instr);
+ const RegisterDemand candidate_diff = get_live_changes(instr);
+ const RegisterDemand temp = get_temp_registers(instr);
if (RegisterDemand(total_demand + candidate_diff).exceeds(max_registers))
return move_fail_pressure;
- const RegisterDemand temp2 = getTempRegisters(block->instructions[insert_idx - 1]);
+ const RegisterDemand temp2 = get_temp_registers(block->instructions[insert_idx - 1]);
const RegisterDemand new_demand = register_demand[insert_idx - 1] - temp2 + candidate_diff + temp;
if (new_demand.exceeds(max_registers))
return move_fail_pressure;
source_idx++;
}
-static bool is_spill_reload(aco_ptr<Instruction>& instr)
-{
- return instr->opcode == aco_opcode::p_spill || instr->opcode == aco_opcode::p_reload;
-}
-
-bool can_reorder(Instruction* candidate)
-{
- switch (candidate->format) {
- case Format::SMEM:
- return static_cast<SMEM_instruction*>(candidate)->can_reorder;
- case Format::MUBUF:
- return static_cast<MUBUF_instruction*>(candidate)->can_reorder;
- case Format::MIMG:
- return static_cast<MIMG_instruction*>(candidate)->can_reorder;
- case Format::MTBUF:
- return static_cast<MTBUF_instruction*>(candidate)->can_reorder;
- case Format::FLAT:
- case Format::GLOBAL:
- case Format::SCRATCH:
- return static_cast<FLAT_instruction*>(candidate)->can_reorder;
- default:
- return true;
- }
-}
-
-bool is_gs_or_done_sendmsg(Instruction *instr)
+bool is_gs_or_done_sendmsg(const Instruction *instr)
{
if (instr->opcode == aco_opcode::s_sendmsg) {
- uint16_t imm = static_cast<SOPP_instruction*>(instr)->imm;
+ uint16_t imm = static_cast<const SOPP_instruction*>(instr)->imm;
return (imm & sendmsg_id_mask) == _sendmsg_gs ||
(imm & sendmsg_id_mask) == _sendmsg_gs_done;
}
return false;
}
-bool is_done_sendmsg(Instruction *instr)
+bool is_done_sendmsg(const Instruction *instr)
{
if (instr->opcode == aco_opcode::s_sendmsg) {
- uint16_t imm = static_cast<SOPP_instruction*>(instr)->imm;
+ uint16_t imm = static_cast<const SOPP_instruction*>(instr)->imm;
return (imm & sendmsg_id_mask) == _sendmsg_gs_done;
}
return false;
}
-barrier_interaction get_barrier_interaction(Instruction* instr)
+memory_sync_info get_sync_info_with_hack(const Instruction* instr)
{
- switch (instr->format) {
- case Format::SMEM:
- return static_cast<SMEM_instruction*>(instr)->barrier;
- case Format::MUBUF:
- return static_cast<MUBUF_instruction*>(instr)->barrier;
- case Format::MIMG:
- return static_cast<MIMG_instruction*>(instr)->barrier;
- case Format::MTBUF:
- return static_cast<MTBUF_instruction*>(instr)->barrier;
- case Format::FLAT:
- case Format::GLOBAL:
- case Format::SCRATCH:
- return static_cast<FLAT_instruction*>(instr)->barrier;
- case Format::DS:
- return barrier_shared;
- case Format::SOPP:
- if (is_done_sendmsg(instr))
- return (barrier_interaction)(barrier_gs_data | barrier_gs_sendmsg);
- else if (is_gs_or_done_sendmsg(instr))
- return barrier_gs_sendmsg;
+ memory_sync_info sync = get_sync_info(instr);
+ if (instr->format == Format::SMEM && !instr->operands.empty() && instr->operands[0].bytes() == 16) {
+ // FIXME: currently, it doesn't seem beneficial to omit this due to how our scheduler works
+ sync.storage = (storage_class)(sync.storage | storage_buffer);
+ sync.semantics = (memory_semantics)(sync.semantics | semantic_private);
+ }
+ return sync;
+}
+
+struct memory_event_set {
+ bool has_control_barrier;
+
+ unsigned bar_acquire;
+ unsigned bar_release;
+ unsigned bar_classes;
+
+ unsigned access_acquire;
+ unsigned access_release;
+ unsigned access_relaxed;
+ unsigned access_atomic;
+};
+
+struct hazard_query {
+ bool contains_spill;
+ bool contains_sendmsg;
+ memory_event_set mem_events;
+ unsigned aliasing_storage; /* storage classes which are accessed (non-SMEM) */
+ unsigned aliasing_storage_smem; /* storage classes which are accessed (SMEM) */
+};
+
+void init_hazard_query(hazard_query *query) {
+ query->contains_spill = false;
+ query->contains_sendmsg = false;
+ memset(&query->mem_events, 0, sizeof(query->mem_events));
+ query->aliasing_storage = 0;
+ query->aliasing_storage_smem = 0;
+}
+
+void add_memory_event(memory_event_set *set, Instruction *instr, memory_sync_info *sync)
+{
+ set->has_control_barrier |= is_done_sendmsg(instr);
+ if (instr->opcode == aco_opcode::p_barrier) {
+ Pseudo_barrier_instruction *bar = static_cast<Pseudo_barrier_instruction*>(instr);
+ if (bar->sync.semantics & semantic_acquire)
+ set->bar_acquire |= bar->sync.storage;
+ if (bar->sync.semantics & semantic_release)
+ set->bar_release |= bar->sync.storage;
+ set->bar_classes |= bar->sync.storage;
+
+ set->has_control_barrier |= bar->exec_scope > scope_invocation;
+ }
+
+ if (!sync->storage)
+ return;
+
+ if (sync->semantics & semantic_acquire)
+ set->access_acquire |= sync->storage;
+ if (sync->semantics & semantic_release)
+ set->access_release |= sync->storage;
+
+ if (!(sync->semantics & semantic_private)) {
+ if (sync->semantics & semantic_atomic)
+ set->access_atomic |= sync->storage;
else
- return barrier_none;
- default:
- return barrier_none;
+ set->access_relaxed |= sync->storage;
}
}
-bool can_move_instr(aco_ptr<Instruction>& instr, Instruction* current, int moving_interaction)
+void add_to_hazard_query(hazard_query *query, Instruction *instr)
{
- /* don't move exports so that they stay closer together */
- if (instr->format == Format::EXP)
- return false;
+ if (instr->opcode == aco_opcode::p_spill || instr->opcode == aco_opcode::p_reload)
+ query->contains_spill = true;
+ query->contains_sendmsg |= instr->opcode == aco_opcode::s_sendmsg;
- /* don't move s_memtime/s_memrealtime */
- if (instr->opcode == aco_opcode::s_memtime || instr->opcode == aco_opcode::s_memrealtime)
- return false;
-
- /* handle barriers */
-
- /* TODO: instead of stopping, maybe try to move the barriers and any
- * instructions interacting with them instead? */
- if (instr->format != Format::PSEUDO_BARRIER) {
- if (instr->opcode == aco_opcode::s_barrier) {
- return can_reorder(current) && moving_interaction == barrier_none;
- } else if (is_gs_or_done_sendmsg(instr.get())) {
- int interaction = get_barrier_interaction(current);
- interaction |= moving_interaction;
- return !(interaction & get_barrier_interaction(instr.get()));
- } else {
- return true;
- }
+ memory_sync_info sync = get_sync_info_with_hack(instr);
+
+ add_memory_event(&query->mem_events, instr, &sync);
+
+ if (!(sync.semantics & semantic_can_reorder)) {
+ unsigned storage = sync.storage;
+ /* images and buffer/global memory can alias */ //TODO: more precisely, buffer images and buffer/global memory can alias
+ if (storage & (storage_buffer | storage_image))
+ storage |= storage_buffer | storage_image;
+ if (instr->format == Format::SMEM)
+ query->aliasing_storage_smem |= storage;
+ else
+ query->aliasing_storage |= storage;
+ }
+}
+
+enum HazardResult {
+ hazard_success,
+ hazard_fail_reorder_vmem_smem,
+ hazard_fail_reorder_ds,
+ hazard_fail_reorder_sendmsg,
+ hazard_fail_spill,
+ hazard_fail_export,
+ hazard_fail_barrier,
+ /* Must stop at these failures. The hazard query code doesn't consider them
+ * when added. */
+ hazard_fail_exec,
+ hazard_fail_unreorderable,
+};
+
+HazardResult perform_hazard_query(hazard_query *query, Instruction *instr, bool upwards)
+{
+ if (instr->opcode == aco_opcode::p_exit_early_if)
+ return hazard_fail_exec;
+ for (const Definition& def : instr->definitions) {
+ if (def.isFixed() && def.physReg() == exec)
+ return hazard_fail_exec;
}
- int interaction = get_barrier_interaction(current);
- interaction |= moving_interaction;
-
- switch (instr->opcode) {
- case aco_opcode::p_memory_barrier_atomic:
- return !(interaction & barrier_atomic);
- /* For now, buffer and image barriers are treated the same. this is because of
- * dEQP-VK.memory_model.message_passing.core11.u32.coherent.fence_fence.atomicwrite.device.payload_nonlocal.buffer.guard_nonlocal.image.comp
- * which seems to use an image load to determine if the result of a buffer load is valid. So the ordering of the two loads is important.
- * I /think/ we should probably eventually expand the meaning of a buffer barrier so that all buffer operations before it, must stay before it
- * and that both image and buffer operations after it, must stay after it. We should also do the same for image barriers.
- * Or perhaps the problem is that we don't have a combined barrier instruction for both buffers and images, but the CTS test expects us to?
- * Either way, this solution should work. */
- case aco_opcode::p_memory_barrier_buffer:
- case aco_opcode::p_memory_barrier_image:
- return !(interaction & (barrier_image | barrier_buffer));
- case aco_opcode::p_memory_barrier_shared:
- return !(interaction & barrier_shared);
- case aco_opcode::p_memory_barrier_common:
- return !(interaction & (barrier_image | barrier_buffer | barrier_shared | barrier_atomic));
- case aco_opcode::p_memory_barrier_gs_data:
- return !(interaction & barrier_gs_data);
- case aco_opcode::p_memory_barrier_gs_sendmsg:
- return !(interaction & barrier_gs_sendmsg);
- default:
- return false;
+ /* don't move exports so that they stay closer together */
+ if (instr->format == Format::EXP)
+ return hazard_fail_export;
+
+ /* don't move non-reorderable instructions */
+ if (instr->opcode == aco_opcode::s_memtime ||
+ instr->opcode == aco_opcode::s_memrealtime ||
+ instr->opcode == aco_opcode::s_setprio ||
+ instr->opcode == aco_opcode::s_getreg_b32)
+ return hazard_fail_unreorderable;
+
+ memory_event_set instr_set;
+ memset(&instr_set, 0, sizeof(instr_set));
+ memory_sync_info sync = get_sync_info_with_hack(instr);
+ add_memory_event(&instr_set, instr, &sync);
+
+ memory_event_set *first = &instr_set;
+ memory_event_set *second = &query->mem_events;
+ if (upwards)
+ std::swap(first, second);
+
+ /* everything after barrier(acquire) happens after the atomics/control_barriers before
+ * everything after load(acquire) happens after the load
+ */
+ if ((first->has_control_barrier || first->access_atomic) && second->bar_acquire)
+ return hazard_fail_barrier;
+ if (((first->access_acquire || first->bar_acquire) && second->bar_classes) ||
+ ((first->access_acquire | first->bar_acquire) & (second->access_relaxed | second->access_atomic)))
+ return hazard_fail_barrier;
+
+ /* everything before barrier(release) happens before the atomics/control_barriers after *
+ * everything before store(release) happens before the store
+ */
+ if (first->bar_release && (second->has_control_barrier || second->access_atomic))
+ return hazard_fail_barrier;
+ if ((first->bar_classes && (second->bar_release || second->access_release)) ||
+ ((first->access_relaxed | first->access_atomic) & (second->bar_release | second->access_release)))
+ return hazard_fail_barrier;
+
+ /* don't move memory barriers around other memory barriers */
+ if (first->bar_classes && second->bar_classes)
+ return hazard_fail_barrier;
+
+ /* Don't move memory accesses to before control barriers. I don't think
+ * this is necessary for the Vulkan memory model, but it might be for GLSL450. */
+ unsigned control_classes = storage_buffer | storage_atomic_counter | storage_image | storage_shared;
+ if (first->has_control_barrier && ((second->access_atomic | second->access_relaxed) & control_classes))
+ return hazard_fail_barrier;
+
+ /* don't move memory loads/stores past potentially aliasing loads/stores */
+ unsigned aliasing_storage = instr->format == Format::SMEM ?
+ query->aliasing_storage_smem :
+ query->aliasing_storage;
+ if ((sync.storage & aliasing_storage) && !(sync.semantics & semantic_can_reorder)) {
+ unsigned intersect = sync.storage & aliasing_storage;
+ if (intersect & storage_shared)
+ return hazard_fail_reorder_ds;
+ return hazard_fail_reorder_vmem_smem;
}
+
+ if ((instr->opcode == aco_opcode::p_spill || instr->opcode == aco_opcode::p_reload) &&
+ query->contains_spill)
+ return hazard_fail_spill;
+
+ if (instr->opcode == aco_opcode::s_sendmsg && query->contains_sendmsg)
+ return hazard_fail_reorder_sendmsg;
+
+ return hazard_success;
}
void schedule_SMEM(sched_ctx& ctx, Block* block,
int window_size = SMEM_WINDOW_SIZE;
int max_moves = SMEM_MAX_MOVES;
int16_t k = 0;
- bool can_reorder_cur = can_reorder(current);
/* don't move s_memtime/s_memrealtime */
if (current->opcode == aco_opcode::s_memtime || current->opcode == aco_opcode::s_memrealtime)
return;
/* first, check if we have instructions before current to move down */
- int moving_interaction = barrier_none;
- bool moving_spill = false;
+ hazard_query hq;
+ init_hazard_query(&hq);
+ add_to_hazard_query(&hq, current);
ctx.mv.downwards_init(idx, false, false);
assert(candidate_idx >= 0);
assert(candidate_idx == ctx.mv.source_idx);
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
- bool can_reorder_candidate = can_reorder(candidate.get());
/* break if we'd make the previous SMEM instruction stall */
bool can_stall_prev_smem = idx <= ctx.last_SMEM_dep_idx && candidate_idx < ctx.last_SMEM_dep_idx;
break;
/* break when encountering another MEM instruction, logical_start or barriers */
- if (!can_reorder_candidate && !can_reorder_cur)
- break;
if (candidate->opcode == aco_opcode::p_logical_start)
break;
- if (candidate->opcode == aco_opcode::p_exit_early_if)
- break;
- if (!can_move_instr(candidate, current, moving_interaction))
- break;
if (candidate->isVMEM())
break;
- /* if current depends on candidate, add additional dependencies and continue */
- bool writes_exec = false;
- for (const Definition& def : candidate->definitions) {
- if (def.isFixed() && def.physReg() == exec)
- writes_exec = true;
- }
- if (writes_exec)
- break;
-
bool can_move_down = true;
- if (moving_spill && is_spill_reload(candidate))
- can_move_down = false;
- if ((moving_interaction & barrier_shared) && candidate->format == Format::DS)
+
+ HazardResult haz = perform_hazard_query(&hq, candidate.get(), false);
+ if (haz == hazard_fail_reorder_ds || haz == hazard_fail_spill || haz == hazard_fail_reorder_sendmsg || haz == hazard_fail_barrier || haz == hazard_fail_export)
can_move_down = false;
- moving_interaction |= get_barrier_interaction(candidate.get());
- moving_spill |= is_spill_reload(candidate);
- if (!can_move_down) {
+ else if (haz != hazard_success)
+ break;
+
+ /* don't use LDS/GDS instructions to hide latency since it can
+ * significanly worsen LDS scheduling */
+ if (candidate->format == Format::DS || !can_move_down) {
+ add_to_hazard_query(&hq, candidate.get());
ctx.mv.downwards_skip();
- can_reorder_cur &= can_reorder_candidate;
continue;
}
MoveResult res = ctx.mv.downwards_move(false);
if (res == move_fail_ssa || res == move_fail_rar) {
+ add_to_hazard_query(&hq, candidate.get());
ctx.mv.downwards_skip();
- can_reorder_cur &= can_reorder_candidate;
continue;
} else if (res == move_fail_pressure) {
break;
}
/* find the first instruction depending on current or find another MEM */
- moving_interaction = barrier_none;
- moving_spill = false;
- can_reorder_cur = true;
-
ctx.mv.upwards_init(idx + 1, false);
bool found_dependency = false;
assert(candidate_idx == ctx.mv.source_idx);
assert(candidate_idx < (int) block->instructions.size());
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
- bool can_reorder_candidate = can_reorder(candidate.get());
if (candidate->opcode == aco_opcode::p_logical_end)
break;
- if (!can_move_instr(candidate, current, moving_interaction))
- break;
-
- const bool writes_exec = std::any_of(candidate->definitions.begin(), candidate->definitions.end(),
- [](const Definition& def) { return def.isFixed() && def.physReg() == exec;});
- if (writes_exec)
- break;
/* check if candidate depends on current */
bool is_dependency = !found_dependency && !ctx.mv.upwards_check_deps();
/* no need to steal from following VMEM instructions */
if (is_dependency && candidate->isVMEM())
break;
- if (moving_spill && is_spill_reload(candidate))
- is_dependency = true;
- if ((moving_interaction & barrier_shared) && candidate->format == Format::DS)
- is_dependency = true;
- moving_interaction |= get_barrier_interaction(candidate.get());
- moving_spill |= is_spill_reload(candidate);
+
+ if (found_dependency) {
+ HazardResult haz = perform_hazard_query(&hq, candidate.get(), true);
+ if (haz == hazard_fail_reorder_ds || haz == hazard_fail_spill ||
+ haz == hazard_fail_reorder_sendmsg || haz == hazard_fail_barrier ||
+ haz == hazard_fail_export)
+ is_dependency = true;
+ else if (haz != hazard_success)
+ break;
+ }
+
if (is_dependency) {
if (!found_dependency) {
ctx.mv.upwards_set_insert_idx(candidate_idx);
+ init_hazard_query(&hq);
found_dependency = true;
}
}
- if (!can_reorder_candidate && !can_reorder_cur)
- break;
-
- if (!found_dependency) {
- ctx.mv.upwards_skip();
- k++;
- continue;
- }
-
- if (is_dependency) {
+ if (is_dependency || !found_dependency) {
+ if (found_dependency)
+ add_to_hazard_query(&hq, candidate.get());
+ else
+ k++;
ctx.mv.upwards_skip();
- can_reorder_cur &= can_reorder_candidate;
continue;
}
/* no need to steal from following VMEM instructions */
if (res == move_fail_ssa && candidate->isVMEM())
break;
+ add_to_hazard_query(&hq, candidate.get());
ctx.mv.upwards_skip();
- can_reorder_cur &= can_reorder_candidate;
continue;
} else if (res == move_fail_pressure) {
break;
int max_moves = VMEM_MAX_MOVES;
int clause_max_grab_dist = VMEM_CLAUSE_MAX_GRAB_DIST;
int16_t k = 0;
- /* initially true as we don't pull other VMEM instructions
- * through the current instruction */
- bool can_reorder_vmem = true;
- bool can_reorder_smem = true;
/* first, check if we have instructions before current to move down */
- int moving_interaction = barrier_none;
- bool moving_spill = false;
+ hazard_query indep_hq;
+ hazard_query clause_hq;
+ init_hazard_query(&indep_hq);
+ init_hazard_query(&clause_hq);
+ add_to_hazard_query(&indep_hq, current);
ctx.mv.downwards_init(idx, true, true);
assert(candidate_idx == ctx.mv.source_idx);
assert(candidate_idx >= 0);
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
- bool can_reorder_candidate = can_reorder(candidate.get());
bool is_vmem = candidate->isVMEM() || candidate->isFlatOrGlobal();
/* break when encountering another VMEM instruction, logical_start or barriers */
- if (!can_reorder_smem && candidate->format == Format::SMEM && !can_reorder_candidate)
- break;
if (candidate->opcode == aco_opcode::p_logical_start)
break;
- if (candidate->opcode == aco_opcode::p_exit_early_if)
- break;
- if (!can_move_instr(candidate, current, moving_interaction))
- break;
/* break if we'd make the previous SMEM instruction stall */
bool can_stall_prev_smem = idx <= ctx.last_SMEM_dep_idx && candidate_idx < ctx.last_SMEM_dep_idx;
bool same_resource = true;
if (current->isVMEM())
same_resource = candidate->operands[0].tempId() == current->operands[0].tempId();
- bool can_reorder = can_reorder_vmem || can_reorder_candidate;
int grab_dist = ctx.mv.insert_idx_clause - candidate_idx;
/* We can't easily tell how much this will decrease the def-to-use
* distances, so just use how far it will be moved as a heuristic. */
- part_of_clause = can_reorder && same_resource && grab_dist < clause_max_grab_dist;
+ part_of_clause = same_resource && grab_dist < clause_max_grab_dist;
}
/* if current depends on candidate, add additional dependencies and continue */
bool can_move_down = !is_vmem || part_of_clause;
- bool writes_exec = false;
- for (const Definition& def : candidate->definitions) {
- if (def.isFixed() && def.physReg() == exec)
- writes_exec = true;
- }
- if (writes_exec)
- break;
- if (moving_spill && is_spill_reload(candidate))
+ HazardResult haz = perform_hazard_query(part_of_clause ? &clause_hq : &indep_hq, candidate.get(), false);
+ if (haz == hazard_fail_reorder_ds || haz == hazard_fail_spill ||
+ haz == hazard_fail_reorder_sendmsg || haz == hazard_fail_barrier ||
+ haz == hazard_fail_export)
can_move_down = false;
- if ((moving_interaction & barrier_shared) && candidate->format == Format::DS)
- can_move_down = false;
- moving_interaction |= get_barrier_interaction(candidate.get());
- moving_spill |= is_spill_reload(candidate);
+ else if (haz != hazard_success)
+ break;
+
if (!can_move_down) {
+ add_to_hazard_query(&indep_hq, candidate.get());
+ add_to_hazard_query(&clause_hq, candidate.get());
ctx.mv.downwards_skip();
- can_reorder_smem &= candidate->format != Format::SMEM || can_reorder_candidate;
- can_reorder_vmem &= !is_vmem || can_reorder_candidate;
continue;
}
+ Instruction *candidate_ptr = candidate.get();
MoveResult res = ctx.mv.downwards_move(part_of_clause);
if (res == move_fail_ssa || res == move_fail_rar) {
+ add_to_hazard_query(&indep_hq, candidate.get());
+ add_to_hazard_query(&clause_hq, candidate.get());
ctx.mv.downwards_skip();
- can_reorder_smem &= candidate->format != Format::SMEM || can_reorder_candidate;
- can_reorder_vmem &= !is_vmem || can_reorder_candidate;
continue;
} else if (res == move_fail_pressure) {
break;
}
+ if (part_of_clause)
+ add_to_hazard_query(&indep_hq, candidate_ptr);
k++;
if (candidate_idx < ctx.last_SMEM_dep_idx)
ctx.last_SMEM_stall++;
}
/* find the first instruction depending on current or find another VMEM */
- moving_interaction = barrier_none;
- moving_spill = false;
- // TODO: differentiate between loads and stores (load-load can always reorder)
- can_reorder_vmem = true;
- can_reorder_smem = true;
-
ctx.mv.upwards_init(idx + 1, true);
bool found_dependency = false;
assert(candidate_idx == ctx.mv.source_idx);
assert(candidate_idx < (int) block->instructions.size());
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
- bool can_reorder_candidate = can_reorder(candidate.get());
bool is_vmem = candidate->isVMEM() || candidate->isFlatOrGlobal();
if (candidate->opcode == aco_opcode::p_logical_end)
break;
- if (!can_move_instr(candidate, current, moving_interaction))
- break;
-
- const bool writes_exec = std::any_of(candidate->definitions.begin(), candidate->definitions.end(),
- [](const Definition& def) {return def.isFixed() && def.physReg() == exec; });
- if (writes_exec)
- break;
/* check if candidate depends on current */
bool is_dependency = false;
- if (candidate->format == Format::SMEM)
- is_dependency = !can_reorder_smem && !can_reorder_candidate;
- if (is_vmem)
- is_dependency = !can_reorder_vmem && !can_reorder_candidate;
+ if (found_dependency) {
+ HazardResult haz = perform_hazard_query(&indep_hq, candidate.get(), true);
+ if (haz == hazard_fail_reorder_ds || haz == hazard_fail_spill ||
+ haz == hazard_fail_reorder_vmem_smem || haz == hazard_fail_reorder_sendmsg ||
+ haz == hazard_fail_barrier || haz == hazard_fail_export)
+ is_dependency = true;
+ else if (haz != hazard_success)
+ break;
+ }
+
is_dependency |= !found_dependency && !ctx.mv.upwards_check_deps();
- if (moving_spill && is_spill_reload(candidate))
- is_dependency = true;
- if ((moving_interaction & barrier_shared) && candidate->format == Format::DS)
- is_dependency = true;
- moving_interaction |= get_barrier_interaction(candidate.get());
- moving_spill |= is_spill_reload(candidate);
if (is_dependency) {
- /* update flag whether we can reorder other memory instructions */
- can_reorder_smem &= candidate->format != Format::SMEM || can_reorder_candidate;
- can_reorder_vmem &= !is_vmem || can_reorder_candidate;
-
if (!found_dependency) {
ctx.mv.upwards_set_insert_idx(candidate_idx);
+ init_hazard_query(&indep_hq);
found_dependency = true;
}
-
} else if (is_vmem) {
/* don't move up dependencies of other VMEM instructions */
for (const Definition& def : candidate->definitions) {
}
if (is_dependency || !found_dependency) {
+ if (found_dependency)
+ add_to_hazard_query(&indep_hq, candidate.get());
ctx.mv.upwards_skip();
continue;
}
MoveResult res = ctx.mv.upwards_move();
if (res == move_fail_ssa || res == move_fail_rar) {
+ add_to_hazard_query(&indep_hq, candidate.get());
ctx.mv.upwards_skip();
- can_reorder_smem &= candidate->format != Format::SMEM || can_reorder_candidate;
- can_reorder_vmem &= !is_vmem || can_reorder_candidate;
continue;
} else if (res == move_fail_pressure) {
break;
ctx.mv.downwards_init(idx, true, false);
- /* first, check if we have instructions before current to move down */
- int moving_interaction = barrier_none;
- bool moving_spill = false;
+ hazard_query hq;
+ init_hazard_query(&hq);
+ add_to_hazard_query(&hq, current);
for (int candidate_idx = idx - 1; k < max_moves && candidate_idx > (int) idx - window_size; candidate_idx--) {
assert(candidate_idx >= 0);
aco_ptr<Instruction>& candidate = block->instructions[candidate_idx];
- /* break when encountering logical_start or barriers */
if (candidate->opcode == aco_opcode::p_logical_start)
break;
- if (candidate->opcode == aco_opcode::p_exit_early_if)
- break;
if (candidate->isVMEM() || candidate->format == Format::SMEM || candidate->isFlatOrGlobal())
break;
- if (!can_move_instr(candidate, current, moving_interaction))
- break;
- /* if current depends on candidate, add additional dependencies and continue */
- bool writes_exec = false;
- for (unsigned i = 0; i < candidate->definitions.size(); i++) {
- if (candidate->definitions[i].isFixed() && candidate->definitions[i].physReg() == exec)
- writes_exec = true;
- }
- if (writes_exec)
+ HazardResult haz = perform_hazard_query(&hq, candidate.get(), false);
+ if (haz == hazard_fail_exec || haz == hazard_fail_unreorderable)
break;
- bool can_move_down = true;
- if (moving_spill && is_spill_reload(candidate))
- can_move_down = false;
- if ((moving_interaction & barrier_shared) && candidate->format == Format::DS)
- can_move_down = false;
- moving_interaction |= get_barrier_interaction(candidate.get());
- moving_spill |= is_spill_reload(candidate);
- if (!can_move_down) {
+ if (haz != hazard_success) {
+ add_to_hazard_query(&hq, candidate.get());
ctx.mv.downwards_skip();
continue;
}
MoveResult res = ctx.mv.downwards_move(false);
if (res == move_fail_ssa || res == move_fail_rar) {
+ add_to_hazard_query(&hq, candidate.get());
ctx.mv.downwards_skip();
continue;
} else if (res == move_fail_pressure) {
}
}
- if ((program->stage & hw_vs) && block->index == program->blocks.size() - 1) {
+ if ((program->stage & (hw_vs | hw_ngg_gs)) && (block->kind & block_kind_export_end)) {
/* Try to move position exports as far up as possible, to reduce register
* usage and because ISA reference guides say so. */
for (unsigned idx = 0; idx < block->instructions.size(); idx++) {
void schedule_program(Program *program, live& live_vars)
{
+ /* don't use program->max_reg_demand because that is affected by max_waves_per_simd */
+ RegisterDemand demand;
+ for (Block& block : program->blocks)
+ demand.update(block.register_demand);
+
sched_ctx ctx;
ctx.mv.depends_on.resize(program->peekAllocationId());
ctx.mv.RAR_dependencies.resize(program->peekAllocationId());
* seem to hurt anything else. */
if (program->num_waves <= 5)
ctx.num_waves = program->num_waves;
- else if (program->max_reg_demand.vgpr >= 32)
+ else if (demand.vgpr >= 29)
ctx.num_waves = 5;
- else if (program->max_reg_demand.vgpr >= 28)
+ else if (demand.vgpr >= 25)
ctx.num_waves = 6;
- else if (program->max_reg_demand.vgpr >= 24)
- ctx.num_waves = 7;
else
- ctx.num_waves = 8;
+ ctx.num_waves = 7;
ctx.num_waves = std::max<uint16_t>(ctx.num_waves, program->min_waves);
+ ctx.num_waves = std::min<uint16_t>(ctx.num_waves, program->max_waves);
assert(ctx.num_waves > 0 && ctx.num_waves <= program->num_waves);
ctx.mv.max_registers = { int16_t(get_addr_vgpr_from_waves(program, ctx.num_waves) - 2),