Core timing 2.0 #4913

B3n30 · 2019-09-01T16:56:14Z

This is a rewrite of Core::CoreTiming. The goal was to be able to emulate multiple cores on a single host thread

What Changed?

There is a timer for each core.
All timers are handled by a TimingManager (called Core::Timing)
Cores are handled consecutively by running a fixed amount of ticks or until the next event on that core
If an event is scheduled in the same core, it could execute less ticks
Events scheduled on another core will be scheduled as exact as possible. This means for cores that weren't already executed in the current cycle we can schedule them without a hassle. But for events scheduled on a core that already executed the current cycle and thus is ahead of time we will just schedule it from the time on that that core is already (it's a small delay).
All this event scheduling could lead to some cores being ahead of others. Thus in the next cycle of core execution, we try to bring all cores back to the one that is the most advanced

What is not in this PR?

All cores currently using the same type of scheduling threads. At a later stage we could think about having different scheduling types for cores (e.g. syscore is believed to use preemtive scheduling)
Layer2 cache
Adjusting CPU Clock Frequency
ApplicationCpuTimeLimit for syscore
The Register Debugger Widget will only show registers of core 0 for now. This needs a rework but can be done in a sperate PR
Unittests only test the case for one timer in TimingManager. So test for multiple timers would be nice

What I'm not sure of

Changes in gdbstub
Changes in cheats
Changes in RPC_Server

What does it fix? Why do we need this?

Games that broke due to removing the priority boost hack
Without that change a accurate new 3ds emulation is not possible

Sorry that it became such a big PR...

This change is

B3n30 · 2019-09-01T19:34:35Z

ci failure is unrelated

wwylele · 2019-09-02T01:27:08Z

src/core/hle/service/ldr_ro/ldr_ro.cpp

@@ -115,7 +115,7 @@ void RO::Initialize(Kernel::HLERequestContext& ctx) {
        return;
    }

-    CROHelper crs(crs_address, *process, system.Memory(), system.CPU());
+    CROHelper crs(crs_address, *process, system.Memory(), system.GetRunningCore());


This should invalidate the cache of all core instead of just the current one.

wwylele · 2019-09-02T01:28:28Z

src/core/rpc/rpc_server.cpp

-        Core::CPU().InvalidateCacheRange(address, data_size);
+
+        // Is current core correct here?
+        Core::GetRunningCore().InvalidateCacheRange(address, data_size);


ditto, this should invalidate cache of all cores

wwylele · 2019-09-02T01:33:12Z

src/core/hle/kernel/thread.cpp

@@ -33,13 +33,19 @@ void Thread::Acquire(Thread* thread) {
    ASSERT_MSG(!ShouldWait(thread), "object unavailable!");
 }

+u32 ThreadManager::next_thread_id = 1;


Please let this live in the kernel object instead of putting it as a global state.

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/core_timing.h

 };

+using SharedTimer = std::shared_ptr<Timing::Timer>;


TBH I feel this typedef is pretty useless and actually add to the difficulty when reading and understanding the code.

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/hle/kernel/kernel.cpp

+void KernelSystem::SetCPUs(std::vector<std::shared_ptr<ARM_Interface>> cpus) {
+    ASSERT(cpus.size() == thread_managers.size());
+    u32 i = 0;
+    for (auto cpu : cpus) {


Suggested change

for (auto cpu : cpus) {

for (auto& cpu : cpus) {

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/hle/kernel/kernel.cpp

+        current_process = process;
+        SetCurrentMemoryPageTable(&process->vm_manager.page_table);
+    } else {
+        stored_processes[core_id];


Eh.. what is this? I feel like this isn't doing anything

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/core.cpp

@@ -177,31 +229,50 @@ void System::Reschedule() {
    }

    reschedule_pending = false;
-    kernel->GetThreadManager().Reschedule();
+    for (auto core : cpu_cores) {


Suggested change

for (auto core : cpu_cores) {

for (const auto& core : cpu_cores) {

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/core_timing.h

-    // executing the first cycle of each slice to prepare the slice length and downcount for
-    // that slice.
-    bool is_global_timer_sane = true;
+    std::unordered_map<std::size_t, std::shared_ptr<Timer>> timers;


I feel this can just be a vector

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/core_timing.h

+        u64 idled_cycles;
+    };
+
+    Timing(std::size_t num_cores);


Suggested change

Timing(std::size_t num_cores);

explicit Timing(std::size_t num_cores);

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/core_timing.h

@@ -125,6 +125,8 @@ inline u64 cyclesToMs(s64 cycles) {

 namespace Core {

+class System;


Is this used?

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/core_timing.cpp

@@ -7,19 +7,27 @@
 #include <tuple>
 #include "common/assert.h"
 #include "common/logging/log.h"
+#include "core/core.h"


This seems unnecessary

zhaowenlan1779 · 2019-09-03T14:58:19Z

src/core/hle/kernel/thread.cpp

    thread->status = ThreadStatus::Dormant;
    thread->entry_point = entry_point;
    thread->stack_top = stack_top;
    thread->nominal_priority = thread->current_priority = priority;
-    thread->last_running_ticks = timing.GetTicks();
+    thread->last_running_ticks = timing.GetGlobalTicks();


Why is global ticks used here? Should this be the ticks of the target processer, or the current core instead?

I decided to go with GlobalTicks which is the ticks of the core thats advanced the most. The reason for this is because in any other case it would always create a thread in the past either for the current core or the target one. We could also take tha max of ticks from current core and target core but I think the overall max is fine too

robzombie91 · 2019-09-03T16:28:22Z

Tested my first game with this "SMT Devil Survivor Overclocked" The game refused to process the fmv that comes after the intro logo's, thus you cant go any farther into the game.

Edit: Log Output

FearlessTobi · 2019-09-03T16:34:39Z

Fixes Donkey Kong Country Returns 3D and Sonic Boom: Fire & Ice.

lioncash · 2019-09-03T17:29:24Z

src/core/core.h

@@ -248,6 +266,8 @@ class System {
        return registered_swkbd;
    }

+    bool initalized = false;


This should be private

B3n30 · 2019-09-03T18:53:02Z

After looking into the threads SMT Devil Survivor Overclocked creates, they are all created in core0, so multicore implementation can't fix it.
I wonder if there is still something wrong with our scheduler?

Edit: List of all threads created:

core/hle/kernel/svc.cpp:CreateThread:922: called entrypoint=0x002DA5D8 (thread-002DA5D8), arg=0x003EA880, stacktop=0x003EA880, threadpriority=0x00000018, processorid=0x00000000 : created handle=0x00020006
core/hle/kernel/svc.cpp:CreateThread:922: called entrypoint=0x002DA5D8 (thread-002DA5D8), arg=0x003DE890, stacktop=0x003DE890, threadpriority=0x0000002F, processorid=0x00000000 : created handle=0x0005000F
core/hle/kernel/svc.cpp:CreateThread:922: called entrypoint=0x002DA5D8 (thread-002DA5D8), arg=0x003E7650, stacktop=0x003E7650, threadpriority=0x0000001A, processorid=0x00000000 : created handle=0x000B0022
core/hle/kernel/svc.cpp:CreateThread:922: called entrypoint=0x002DA5D8 (thread-002DA5D8), arg=0x09000FC8, stacktop=0x09000FC8, threadpriority=0x00000022, processorid=0x00000000 : created handle=0x000D8028
core/hle/kernel/svc.cpp:CreateThread:922: called entrypoint=0x002DA5D8 (thread-002DA5D8), arg=0x003A9D38, stacktop=0x003A9D38, threadpriority=0x00000031, processorid=0x00000000 : created handle=0x000E0029
core/hle/kernel/svc.cpp:CreateThread:922: called entrypoint=0x002DA5D8 (thread-002DA5D8), arg=0x003ADE50, stacktop=0x003ADE50, threadpriority=0x00000032, processorid=0x00000000 : created handle=0x000E802A

… because Thread->status == WaitIPC

…ll broken)

B3n30 · 2020-02-09T17:46:19Z

Rebased and addressed comments. I hope I didn't miss anything

purpasmart96 · 2020-02-19T22:51:06Z

As long as it doesn't break anything. LGTM

B3n30 force-pushed the core-timing_2.0 branch from 4708172 to 0125969 Compare Sep 1, 2019

wwylele reviewed Sep 2, 2019

View changes

zhaowenlan1779 reviewed Sep 3, 2019

View changes

lioncash reviewed Sep 3, 2019

View changes

This comment has been minimized.

Sign in to view

jroweboy force-pushed the core-timing_2.0 branch from 0125969 to e71bfca Compare Jan 1, 2020

jroweboy added canary-merge and removed canary-merge labels Jan 1, 2020

B3n30 added 8 commits Feb 9, 2020

Core::Timing: Add multiple timer, one for each core

648f083

revert clang-format; work on tests for CoreTiming

44af48e

Kernel:: Add support for multiple cores, asserts in HandleSyncRequest…

37bfd82

… because Thread->status == WaitIPC

Add some TRACE_LOGs

78ad3d6

fix tests

1a101e2

make some adjustments to qt-debugger, cheats and gdbstub(probably sti…

ecd8bc9

…ll broken)

Make ARM_Interface::id private, rework ARM_Interface ctor

b921692

ReRename TimingManager to Timing for smaler diff

48cf6c9

B3n30 force-pushed the core-timing_2.0 branch 2 times, most recently from f33464e to 58b2abd Compare Feb 9, 2020

B3n30 force-pushed the core-timing_2.0 branch from 58b2abd to 88536bc Compare Feb 9, 2020

addressed review comments

bc8e481

B3n30 force-pushed the core-timing_2.0 branch from 88536bc to bc8e481 Compare Feb 9, 2020

B3n30 added canary-merge and removed canary-merge labels Feb 9, 2020

B3n30 merged commit 55ec703 into citra-emu:master Feb 21, 2020
3 checks passed

This was referenced Feb 21, 2020

Sonic Boom Shattered Crystals and Sonic Boom Fire & Ice regression #2654

Closed

Donkey Kong Country Returns 3D Boot on Black screen and freeze #2599

Closed

benderscruffy mentioned this pull request Feb 29, 2020

Games sending threads on Syscore starving. #2411

Closed

B3n30 mentioned this pull request Mar 13, 2020

Pokemon Ultra Sun Battle Lag #5128

Open

hamish-milne mentioned this pull request Mar 13, 2020

Emulate the system CPU core #435

Open

Dragios mentioned this pull request Mar 24, 2020

Phoenix Wright: Ace Attorney - Dual Destinies – Fail to boot if system titles update installed (Regression) #5141

Closed

Subv mentioned this pull request Apr 12, 2020

Theatrhythm Dragon Quest - Gameplay/audio desync issue #4527

Closed

LOuroboros mentioned this pull request Apr 14, 2020

Performance regression in Super Mario 3D Land #5209

Closed

Dragios mentioned this pull request Apr 23, 2020

Games that broke after removing priority boost hack #2554

Closed

Core timing 2.0 #4913

Core timing 2.0 #4913

B3n30 commented Sep 1, 2019 •

edited

B3n30 commented Sep 1, 2019

wwylele Sep 2, 2019 •

edited

wwylele Sep 2, 2019

wwylele Sep 2, 2019

zhaowenlan1779 Sep 3, 2019

zhaowenlan1779 Sep 3, 2019 •

edited

zhaowenlan1779 Sep 3, 2019

zhaowenlan1779 Sep 3, 2019

zhaowenlan1779 Sep 3, 2019

zhaowenlan1779 Sep 3, 2019

zhaowenlan1779 Sep 3, 2019

zhaowenlan1779 Sep 3, 2019

zhaowenlan1779 Sep 3, 2019

B3n30 Feb 9, 2020

robzombie91 commented Sep 3, 2019 •

edited

FearlessTobi commented Sep 3, 2019

lioncash Sep 3, 2019

B3n30 commented Sep 3, 2019 •

edited

This comment has been minimized.

B3n30 commented Feb 9, 2020

purpasmart96 commented Feb 19, 2020

	for (auto core : cpu_cores) {
	for (const auto& core : cpu_cores) {

	Timing(std::size_t num_cores);
	explicit Timing(std::size_t num_cores);

		@@ -125,6 +125,8 @@ inline u64 cyclesToMs(s64 cycles) {

		namespace Core {

		class System;

Core timing 2.0 #4913

Core timing 2.0 #4913

Conversation

B3n30 commented Sep 1, 2019 • edited

What Changed?

What is not in this PR?

What I'm not sure of

What does it fix? Why do we need this?

B3n30 commented Sep 1, 2019

wwylele Sep 2, 2019 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

zhaowenlan1779 Sep 3, 2019 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

robzombie91 commented Sep 3, 2019 • edited

FearlessTobi commented Sep 3, 2019

Choose a reason for hiding this comment

B3n30 commented Sep 3, 2019 • edited

This comment has been minimized.

B3n30 commented Feb 9, 2020

purpasmart96 commented Feb 19, 2020

B3n30 commented Sep 1, 2019 •

edited

wwylele Sep 2, 2019 •

edited

zhaowenlan1779 Sep 3, 2019 •

edited

robzombie91 commented Sep 3, 2019 •

edited

B3n30 commented Sep 3, 2019 •

edited