RPCS3/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2024-11-22 02:32:36 +01:00
Code Issues Releases Wiki Activity

SPU: Optimize and enable SPU GETLLAR Polling detection by default

Browse Source 
* Make this setting guard all reservation waitings. (renamed)
* Revert atomic list usage: it's more expensive and not needed because it has a timeout and is not optimized for the rest of the waitables.
This commit is contained in:
Eladash 2022-08-17 14:28:29 +03:00 committed by Ivan
parent 33a4f05ffa
commit 6210a8491f
3 changed files with 89 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

128
rpcs3/Emu/Cell/SPUThread.cpp
View File

@ -3388,31 +3388,86 @@ bool spu_thread::process_mfc_cmd()
last_faddr = 0;
}
if (addr == raddr && !g_use_rtm && g_cfg.core.spu_getllar_polling_detection && rtime == vm::reservation_acquire(addr) && cmp_rdata(rdata, data))
{
// Spinning, might as well yield cpu resources
std::this_thread::yield();
// Reset perf
perf0.restart();
}
alignas(64) spu_rdata_t temp;
u64 ntime;
rsx::reservation_lock rsx_lock(addr, 128);
if (ch_events.load().events & SPU_EVENT_LR)
{
// There is no longer a need to concern about LR event if it has already been raised.
raddr = 0;
}
if (raddr)
{
// Save rdata from previous reservation
mov_rdata(temp, rdata);
if (raddr != addr)
{
// Last check for event before we replace the reservation with a new one
if (reservation_check(raddr, rdata))
{
set_events(SPU_EVENT_LR);
}
}
else
{
// Check if we can reuse our existing reservation
if (rtime == vm::reservation_acquire(addr) && cmp_rdata(rdata, data) && rtime == vm::reservation_acquire(addr) && cmp_rdata(rdata, data))
{
mov_rdata(_ref<spu_rdata_t>(ch_mfc_cmd.lsa & 0x3ff80), rdata);
ch_atomic_stat.set_value(MFC_GETLLAR_SUCCESS);
// Need to check twice for it to be accurate, the code is before and not not after this check for:
// 1. Reduce time between reservation accesses so TSX panelty would be lowered
// 2. Increase the chance of change detection: if GETLLAR has been called again new data is probably wanted
if (rtime == vm::reservation_acquire(addr) && cmp_rdata(rdata, data))
{
// Validation that it is indeed GETLLAR busy-waiting (large time window is intentional)
if ((g_cfg.core.spu_reservation_busy_waiting && !g_use_rtm) || last_getllar != pc || perf0.get() - last_gtsc >= 50'000)
{
if (g_cfg.core.mfc_debug)
{
auto& dump = mfc_history[mfc_dump_idx++ % spu_thread::max_mfc_dump_idx];
dump.cmd = ch_mfc_cmd;
dump.cmd.eah = pc;
std::memcpy(dump.data, rdata, 128);
}
last_getllar = pc;
last_gtsc = perf0.get();
if (g_cfg.core.spu_reservation_busy_waiting)
{
busy_wait();
}
return true;
}
// Spinning, might as well yield cpu resources
state += cpu_flag::wait;
vm::reservation_notifier(addr).wait(rtime, -128, atomic_wait_timeout{100'000});
// Reset perf
perf0.restart();
// Quick check if there were reservation changes
if (rtime == vm::reservation_acquire(addr))
{
// None at least in rtime
if (g_cfg.core.mfc_debug)
{
auto& dump = mfc_history[mfc_dump_idx++ % spu_thread::max_mfc_dump_idx];
dump.cmd = ch_mfc_cmd;
dump.cmd.eah = pc;
std::memcpy(dump.data, rdata, 128);
}
last_gtsc = perf0.get();
return true;
}
}
}
// We can't, LR needs to be set now
set_events(SPU_EVENT_LR);
static_cast<void>(test_stopped());
}
}
u64 ntime;
rsx::reservation_lock rsx_lock(addr, 128);
for (u64 i = 0; i != umax; [&]()
{
if (state & cpu_flag::pause)
@ -3492,26 +3547,11 @@ bool spu_thread::process_mfc_cmd()
break;
}
if (raddr && raddr != addr)
{
// Last check for event before we replace the reservation with a new one
if (reservation_check(raddr, temp))
{
set_events(SPU_EVENT_LR);
}
}
else if (raddr == addr)
{
// Lost previous reservation on polling
if (ntime != rtime || !cmp_rdata(rdata, temp))
{
set_events(SPU_EVENT_LR);
}
}
raddr = addr;
rtime = ntime;
mov_rdata(_ref<spu_rdata_t>(ch_mfc_cmd.lsa & 0x3ff80), rdata);
last_getllar = pc;
last_gtsc = perf0.get();
ch_atomic_stat.set_value(MFC_GETLLAR_SUCCESS);
@ -4205,11 +4245,15 @@ s64 spu_thread::get_ch_value(u32 ch)
if (raddr)
{
thread_ctrl::wait_on_custom<2>([&](atomic_wait::list<4>& list)
// Don't busy-wait with TSX - memory is sensitive
if (g_use_rtm || !g_cfg.core.spu_reservation_busy_waiting)
{
list.set<0>(state, old);
list.set<1>(vm::reservation_notifier(raddr), rtime, -128);
}, 100);
vm::reservation_notifier(raddr).wait(rtime, -128, atomic_wait_timeout{100'000});
}
else
{
busy_wait();
}
continue;
}

2
rpcs3/Emu/Cell/SPUThread.h
View File

@ -836,6 +836,8 @@ public:
u32 last_faddr = 0;
u64 last_fail = 0;
u64 last_succ = 0;
u64 last_gtsc = 0;
u32 last_getllar = umax; // LS address of last GETLLAR (if matches current GETLLAR we can let the thread rest)
std::vector<mfc_cmd_dump> mfc_history;
u64 mfc_dump_idx = 0;

2
rpcs3/Emu/system_config.h
View File

@ -31,7 +31,7 @@ struct cfg_root : cfg::node
cfg::_enum<thread_scheduler_mode> thread_scheduler{this, "Thread Scheduler Mode", thread_scheduler_mode::os};
cfg::_bool set_daz_and_ftz{ this, "Set DAZ and FTZ", false };
cfg::_enum<spu_decoder_type> spu_decoder{ this, "SPU Decoder", spu_decoder_type::llvm };
cfg::_bool spu_getllar_polling_detection{ this, "SPU GETLLAR polling detection", false, true };
cfg::_bool spu_reservation_busy_waiting{ this, "SPU Reservation Busy Waiting", false, true };
cfg::_bool spu_debug{ this, "SPU Debug" };
cfg::_bool mfc_debug{ this, "MFC Debug" };
cfg::_int<0, 6> preferred_spu_threads{ this, "Preferred SPU Threads", 0, true }; // Number of hardware threads dedicated to heavy simultaneous spu tasks