spu: Speed hacks - Do not starve PPU threads

optionally hint to the OS scheduler to give less attention to SPUs

ui: Add speed 'hacks' as configurable options

Utilities/Thread.cpp

@@ -1750,6 +1750,7 @@ void named_thread::start_thread(const std::shared_ptr<void>& _this)
 		try
 		{
 			LOG_TRACE(GENERAL, "Thread started");
+			on_spawn();
 			on_task();
 			LOG_TRACE(GENERAL, "Thread ended");
 		}

Utilities/Thread.h

@@ -258,6 +258,9 @@ protected:
 	// Thread finalization (called after on_task)
 	virtual void on_exit() {}
 
+	// Called once upon thread spawn within the thread's own context
+	virtual void on_spawn() {}
+
 public:
 	// ID initialization
 	virtual void on_init(const std::shared_ptr<void>& _this)

rpcs3/Emu/CPU/CPUThread.cpp

@@ -4,6 +4,11 @@
 #include "CPUThread.h"
 #include "Emu/IdManager.h"
 #include "Utilities/GDBDebugServer.h"
+#include "Utilities/Config.h"
+
+#ifdef _WIN32
+#include <Windows.h>
+#endif
 
 DECLARE(cpu_thread::g_threads_created){0};
 DECLARE(cpu_thread::g_threads_deleted){0};
@@ -178,3 +183,34 @@ std::string cpu_thread::dump() const
 {
 	return fmt::format("Type: %s\n" "State: %s\n", typeid(*this).name(), state.load());
 }
+
+void cpu_thread::set_native_priority(int priority)
+{
+#ifdef _WIN32
+	HANDLE _this_thread = GetCurrentThread();
+	INT native_priority = THREAD_PRIORITY_NORMAL;
+
+	switch (priority)
+	{
+	default:
+	case 0:
+		break;
+	case 1:
+		native_priority = THREAD_PRIORITY_ABOVE_NORMAL;
+		break;
+	case -1:
+		native_priority = THREAD_PRIORITY_BELOW_NORMAL;
+		break;
+	}
+
+	SetThreadPriority(_this_thread, native_priority);
+#endif // _WIN32
+}
+
+void cpu_thread::set_ideal_processor_core(int core)
+{
+#ifdef _WIN32
+	HANDLE _this_thread = GetCurrentThread();
+	SetThreadIdealProcessor(_this_thread, core);
+#endif
+}

rpcs3/Emu/CPU/CPUThread.h

@@ -65,6 +65,10 @@ public:
 
 	// Callback for cpu_flag::suspend
 	virtual void cpu_sleep() {}
+
+	//native scheduler tweaks
+	void set_native_priority(int priority);
+	void set_ideal_processor_core(int core);
 };
 
 inline cpu_thread* get_current_cpu_thread() noexcept

rpcs3/Emu/Cell/SPUThread.cpp

@@ -20,6 +20,8 @@
 
 #include <cmath>
 #include <cfenv>
+#include <atomic>
+#include <thread>
 
 #ifdef _MSC_VER
 bool operator ==(const u128& lhs, const u128& rhs)
@@ -49,10 +51,14 @@ cfg::map_entry<spu_decoder_type> g_cfg_spu_decoder(cfg::root.core, "SPU Decoder"
 });
 
 cfg::bool_entry g_cfg_spu_debug(cfg::root.core, "SPU Debug");
+cfg::bool_entry g_cfg_core_bind_spu_cores(cfg::root.core, "Bind SPU threads to secondary cores");
+cfg::bool_entry g_cfg_core_lower_spu_priority(cfg::root.core, "Lower SPU thread priority");
 
 const spu_decoder<spu_interpreter_precise> s_spu_interpreter_precise;
 const spu_decoder<spu_interpreter_fast> s_spu_interpreter_fast;
 
+std::atomic<u64> g_num_spu_threads = { 0ull };
+
 void spu_int_ctrl_t::set(u64 ints)
 {
 	// leave only enabled interrupts
@@ -134,6 +140,29 @@ spu_imm_table_t::spu_imm_table_t()
 	}
 }
 
+void SPUThread::on_spawn()
+{
+	if (g_cfg_core_bind_spu_cores)
+	{
+		//Get next secondary core number
+		auto core_count = std::thread::hardware_concurrency();
+		if (core_count > 0 && core_count <= 16)
+		{
+			auto half_count = core_count / 2;
+			auto assigned_secondary_core = ((g_num_spu_threads % half_count) * 2) + 1;
+
+			set_ideal_processor_core(assigned_secondary_core);
+		}
+	}
+
+	if (g_cfg_core_lower_spu_priority)
+	{
+		set_native_priority(-1);
+	}
+
+	g_num_spu_threads++;
+}
+
 void SPUThread::on_init(const std::shared_ptr<void>& _this)
 {
 	if (!offset)
@@ -208,7 +237,7 @@ extern thread_local std::string(*g_tls_log_prefix)();
 void SPUThread::cpu_task()
 {
 	std::fesetround(FE_TOWARDZERO);
-
+	
 	if (g_cfg_spu_decoder.get() == spu_decoder_type::asmjit)
 	{
 		if (!spu_db) spu_db = fxm::get_always<SPUDatabase>();

rpcs3/Emu/Cell/SPUThread.h

@@ -504,6 +504,7 @@ public:
 class SPUThread : public cpu_thread
 {
 public:
+	virtual void on_spawn() override;
 	virtual void on_init(const std::shared_ptr<void>&) override;
 	virtual std::string get_name() const override;
 	virtual std::string dump() const override;

rpcs3/Gui/SettingsDialog.cpp

@@ -226,7 +226,7 @@ SettingsDialog::SettingsDialog(wxWindow* parent, const std::string& path)
 
 	std::vector<std::unique_ptr<cfg_adapter>> pads;
 
-	static const u32 width  = 458;
+	static const u32 width  = 512;
 	static const u32 height = 400;
 
 	// Settings panels
@@ -310,6 +310,8 @@ SettingsDialog::SettingsDialog(wxWindow* parent, const std::string& path)
 	wxComboBox* cbox_sys_lang = new wxComboBox(p_system, wxID_ANY, wxEmptyString, wxDefaultPosition, wxDefaultSize, 0, NULL, wxCB_READONLY);
 
 	wxCheckBox* chbox_core_hook_stfunc = new wxCheckBox(p_core, wxID_ANY, "Hook static functions");
+	wxCheckBox* chbox_core_bind_spu_threads = new wxCheckBox(p_core, wxID_ANY, "Bind SPU threads to secondary cores");
+	wxCheckBox* chbox_core_lower_spu_priority = new wxCheckBox(p_core, wxID_ANY, "Lower SPU thread priority");
 	wxCheckBox* chbox_vfs_enable_host_root = new wxCheckBox(p_system, wxID_ANY, "Enable /host_root/");
 	wxCheckBox* chbox_gs_log_prog = new wxCheckBox(p_graphics, wxID_ANY, "Log Shader Programs");
 	wxCheckBox* chbox_gs_dump_depth = new wxCheckBox(p_graphics, wxID_ANY, "Write Depth Buffer");
@@ -391,6 +393,8 @@ SettingsDialog::SettingsDialog(wxWindow* parent, const std::string& path)
 	EnableModuleList(rbox_lib_loader->GetSelection());
 
 	pads.emplace_back(std::make_unique<checkbox_pad>(cfg_location{ "Core", "Hook static functions" }, chbox_core_hook_stfunc));
+	pads.emplace_back(std::make_unique<checkbox_pad>(cfg_location{ "Core", "Bind SPU threads to secondary cores" }, chbox_core_bind_spu_threads));
+	pads.emplace_back(std::make_unique<checkbox_pad>(cfg_location{ "Core", "Lower SPU thread priority" }, chbox_core_lower_spu_priority));
 	pads.emplace_back(std::make_unique<checkbox_pad>(cfg_location{ "VFS", "Enable /host_root/" }, chbox_vfs_enable_host_root));
 
 	pads.emplace_back(std::make_unique<combobox_pad>(cfg_location{ "Video", "Renderer" }, cbox_gs_render));
@@ -482,6 +486,8 @@ SettingsDialog::SettingsDialog(wxWindow* parent, const std::string& path)
 	s_subpanel_core1->Add(rbox_ppu_decoder, wxSizerFlags().Border(wxALL, 5).Expand());
 	s_subpanel_core1->Add(rbox_spu_decoder, wxSizerFlags().Border(wxALL, 5).Expand());
 	s_subpanel_core1->Add(chbox_core_hook_stfunc, wxSizerFlags().Border(wxALL, 5).Expand());
+	s_subpanel_core1->Add(chbox_core_bind_spu_threads, wxSizerFlags().Border(wxALL, 5).Expand());
+	s_subpanel_core1->Add(chbox_core_lower_spu_priority, wxSizerFlags().Border(wxALL, 5).Expand());
 	s_subpanel_core2->Add(rbox_lib_loader, wxSizerFlags().Border(wxALL, 5).Expand());
 	s_subpanel_core2->Add(s_round_core_lle, wxSizerFlags().Border(wxALL, 5).Expand());
 	s_subpanel_core->Add(s_subpanel_core1);