[core, gpu, threads] Multithreading refactor (#4254)

This insufferable work tries to cover some holes on previous threading implementation from yuzu's team, starting with Windows and Linux reordering of priorities (NICE), reworks previous Android's threading and cpu affinity with adpf, adjust emulated clocks/gpu for better "accuracy" with their work, bumps android minSDK for all flavors, legacy will now work with AP 29 to cover A10 - A12, standard will reach A13 as base and finally the optimized build will come with API 35, mostly targeted on devices with A15 support and newer, NDK and AGP wasn't upgraded yet. The performance cost efficiency have been improved based on device power configuration; preventing overheating if certain devices tended to fall into NICE0 (not allocated threads priority, all task ran with higher priority, 11 tasks running within the limited 2 - 7 threads available on the most common configuration 1x3x4 or 1x4x3).

Reviewed-on: https://git.eden-emu.dev/eden-emu/eden/pulls/4254

src/common/thread.cpp

@@ -1,5 +1,6 @@
 // SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
 // SPDX-License-Identifier: GPL-3.0-or-later
+
 // SPDX-FileCopyrightText: 2013 Dolphin Emulator Project
 // SPDX-FileCopyrightText: 2014 Citra Emulator Project
 // SPDX-License-Identifier: GPL-2.0-or-later
@@ -9,6 +10,7 @@
 #include <string>
 #include <thread>
 
+#include "common/adpf.h"
 #include "common/error.h"
 #include "common/logging.h"
 #include "common/assert.h"
@@ -39,6 +41,402 @@
 #include <unistd.h>
 #endif
 
+#ifdef __linux__
+#include <sys/resource.h>
+#include <algorithm>
+
+namespace {
+constexpr int NICE_AUDIO = -16;
+constexpr int NICE_URGENT_DISPLAY = -8;
+constexpr int NICE_DISPLAY = -4;
+constexpr int NICE_DEFAULT = 0;
+constexpr int NICE_BACKGROUND = 10;
+
+int LowestAllowedNice() {
+    static const int lowest = [] {
+        rlimit limit{};
+        if (getrlimit(RLIMIT_NICE, &limit) != 0) {
+            return 0;
+        }
+        if (limit.rlim_cur >= 40) {
+            return -20;
+        }
+        return 20 - static_cast<int>(limit.rlim_cur);
+    }();
+    return lowest;
+}
+
+int NiceValueForPriority(Common::ThreadPriority priority) {
+    const int wanted = [priority] {
+        switch (priority) {
+        case Common::ThreadPriority::Low: return NICE_BACKGROUND;
+        case Common::ThreadPriority::Normal: return NICE_DEFAULT;
+        case Common::ThreadPriority::High: return NICE_DISPLAY;
+        case Common::ThreadPriority::VeryHigh: return NICE_URGENT_DISPLAY;
+        case Common::ThreadPriority::Critical: return NICE_AUDIO;
+        default: return NICE_DEFAULT;
+        }
+    }();
+    return (std::max)(wanted, (std::min)(NICE_DEFAULT, LowestAllowedNice()));
+}
+} // Anonymous namespace
+#endif
+
+#ifdef __ANDROID__
+#include <sys/utsname.h>
+#include <atomic>
+#include <cerrno>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <fstream>
+#include <mutex>
+#include <vector>
+
+namespace {
+constexpr size_t ANDROID_MINIMUM_PERFORMANCE_CORES = 4;
+
+// A core counts as a performance core while it is within this much of the fastest one.
+constexpr s64 ANDROID_PERFORMANCE_CAPACITY_PERCENT = 50;
+
+constexpr std::chrono::nanoseconds ANDROID_POLICY_POLL_INTERVAL = std::chrono::milliseconds{500};
+
+enum class CoreGroup {
+    Unrestricted,
+    Performance,
+    Efficiency,
+};
+
+struct ThreadPolicy {
+    pid_t tid;
+    CoreGroup group;
+    s32 nice_value;
+    bool has_nice;
+};
+
+struct CoreInfo {
+    s64 weight;
+    u64 midr;
+    s32 cpu;
+};
+
+struct CpuTopologyState {
+    std::mutex topology_mutex;
+    cpu_set_t allowed{};
+    cpu_set_t performance{};
+    cpu_set_t efficiency{};
+    bool separated = false;
+    bool initialized = false;
+
+    pid_t canary_tid = 0;
+    cpu_set_t canary_mask{};
+    bool canary_valid = false;
+
+    std::atomic<s64> next_poll_ns{0};
+
+    std::mutex policy_mutex;
+    std::vector<ThreadPolicy> policies;
+};
+
+CpuTopologyState& State() {
+    static CpuTopologyState* const state = new CpuTopologyState();
+    return *state;
+}
+
+struct PolicyRegistration {
+    ~PolicyRegistration() {
+        const pid_t tid = gettid();
+        ::Common::ADPF::RemoveCurrentThread();
+        CpuTopologyState& state = State();
+        std::scoped_lock lock{state.policy_mutex};
+        std::erase_if(state.policies,
+                      [tid](const ThreadPolicy& policy) { return policy.tid == tid; });
+    }
+};
+
+thread_local PolicyRegistration t_policy_registration;
+
+s32 PossibleCpuCount() {
+    std::ifstream file("/sys/devices/system/cpu/possible");
+    std::string list;
+    if (file && std::getline(file, list) && !list.empty()) {
+        s64 highest = -1;
+        const char* cursor = list.c_str();
+        while (*cursor != '\0') {
+            char* end = nullptr;
+            const s64 value = std::strtol(cursor, &end, 10);
+            if (end == cursor) {
+                break;
+            }
+            highest = (std::max)(highest, value);
+            cursor = end;
+            while (*cursor == '-' || *cursor == ',') {
+                ++cursor;
+            }
+        }
+        if (highest >= 0) {
+            return static_cast<s32>((std::min<s64>)(highest + 1, CPU_SETSIZE));
+        }
+    }
+    const s64 configured = sysconf(_SC_NPROCESSORS_CONF);
+    if (configured > 0) {
+        return static_cast<s32>((std::min<s64>)(configured, CPU_SETSIZE));
+    }
+    return static_cast<s32>((std::min<s64>)(std::thread::hardware_concurrency(), CPU_SETSIZE));
+}
+
+s64 ReadCpuScalar(s32 cpu, const char* node) {
+    s64 value = 0;
+    std::ifstream file("/sys/devices/system/cpu/cpu" + std::to_string(cpu) + "/" + node);
+    if (!file || !(file >> value) || value <= 0) {
+        return 0;
+    }
+    return value;
+}
+
+u64 ReadCpuMidr(s32 cpu) {
+    u64 midr = 0;
+    std::ifstream file("/sys/devices/system/cpu/cpu" + std::to_string(cpu) +
+                       "/regs/identification/midr_el1");
+    if (!file || !(file >> std::hex >> midr)) {
+        return 0;
+    }
+    return midr;
+}
+
+std::vector<CoreInfo> CollectCores(const cpu_set_t& allowed, s32 total, const char* node,
+                                   bool require_all) {
+    std::vector<CoreInfo> cores;
+    for (s32 cpu = 0; cpu < total; ++cpu) {
+        if (!CPU_ISSET(cpu, &allowed)) {
+            continue;
+        }
+        const s64 weight = ReadCpuScalar(cpu, node);
+        if (weight <= 0) {
+            if (require_all) {
+                return {};
+            }
+            LOG_WARNING(Common, "Could not read {} for CPU {}, treating it as an efficiency core",
+                        node, cpu);
+            continue;
+        }
+        cores.push_back(CoreInfo{weight, ReadCpuMidr(cpu), cpu});
+    }
+    return cores;
+}
+
+bool WeightsAreUniform(const std::vector<CoreInfo>& cores) {
+    return std::all_of(cores.begin(), cores.end(),
+                       [&](const CoreInfo& core) { return core.weight == cores.front().weight; });
+}
+
+bool MidrsAreDistinct(const std::vector<CoreInfo>& cores) {
+    return std::none_of(cores.begin(), cores.end(),
+                        [](const CoreInfo& core) { return core.midr == 0; }) &&
+           std::any_of(cores.begin(), cores.end(),
+                       [&](const CoreInfo& core) { return core.midr != cores.front().midr; });
+}
+
+void ComputeTopologyLocked(CpuTopologyState& state) {
+    state.initialized = true;
+    state.separated = false;
+    CPU_ZERO(&state.allowed);
+    CPU_ZERO(&state.performance);
+    CPU_ZERO(&state.efficiency);
+
+    if (sched_getaffinity(getpid(), sizeof(state.allowed), &state.allowed) != 0) {
+        LOG_WARNING(Common, "Could not query process CPU affinity: {}",
+                    ::Common::GetLastErrorMsg());
+        return;
+    }
+
+    const s32 total = PossibleCpuCount();
+    auto cores = CollectCores(state.allowed, total, "cpu_capacity", true);
+    if (cores.empty() || (WeightsAreUniform(cores) && MidrsAreDistinct(cores))) {
+        auto by_frequency = CollectCores(state.allowed, total, "cpufreq/cpuinfo_max_freq", false);
+        if (!by_frequency.empty()) {
+            cores = std::move(by_frequency);
+        }
+    }
+    if (cores.empty()) {
+        LOG_WARNING(Common, "Could not determine CPU topology, thread placement is disabled");
+        return;
+    }
+
+    if (WeightsAreUniform(cores)) {
+        if (MidrsAreDistinct(cores)) {
+            LOG_WARNING(Common, "CPU clusters differ but rank identically, thread placement is "
+                                "disabled");
+        } else {
+            LOG_INFO(Common, "CPU cores are symmetric, thread placement is disabled");
+        }
+        return;
+    }
+
+    std::sort(cores.begin(), cores.end(), [](const CoreInfo& lhs, const CoreInfo& rhs) {
+        if (lhs.weight != rhs.weight) {
+            return lhs.weight > rhs.weight;
+        }
+        return lhs.cpu < rhs.cpu;
+    });
+
+    const s64 fastest = cores.front().weight;
+    size_t taken = 0;
+    for (const auto& core : cores) {
+        const bool fast_enough =
+            core.weight * 100 >= fastest * ANDROID_PERFORMANCE_CAPACITY_PERCENT;
+        if (!fast_enough && taken >= ANDROID_MINIMUM_PERFORMANCE_CORES) {
+            break;
+        }
+        CPU_SET(core.cpu, &state.performance);
+        ++taken;
+    }
+    if (taken == 0) {
+        return;
+    }
+
+    for (s32 cpu = 0; cpu < total; ++cpu) {
+        if (CPU_ISSET(cpu, &state.allowed) && !CPU_ISSET(cpu, &state.performance)) {
+            CPU_SET(cpu, &state.efficiency);
+        }
+    }
+
+    state.separated = CPU_COUNT(&state.efficiency) > 0;
+    LOG_INFO(Common, "CPU topology: {} performance cores, {} efficiency cores, separation {}",
+             CPU_COUNT(&state.performance), CPU_COUNT(&state.efficiency),
+             state.separated ? "enabled" : "unavailable");
+}
+
+void EnsureTopologyLocked(CpuTopologyState& state) {
+    if (!state.initialized) {
+        ComputeTopologyLocked(state);
+    }
+}
+
+void RefreshTopologyLocked(CpuTopologyState& state) {
+    if (!state.initialized) {
+        ComputeTopologyLocked(state);
+        return;
+    }
+    cpu_set_t current;
+    CPU_ZERO(&current);
+    if (sched_getaffinity(getpid(), sizeof(current), &current) != 0) {
+        return;
+    }
+    if (std::memcmp(&current, &state.allowed, sizeof(current)) != 0) {
+        ComputeTopologyLocked(state);
+    }
+}
+
+bool ApplyCoreGroupLocked(CpuTopologyState& state, pid_t tid, CoreGroup group,
+                          bool* gone = nullptr) {
+    const bool restrict_group = group != CoreGroup::Unrestricted && state.separated;
+    const cpu_set_t* mask = &state.allowed;
+    if (restrict_group) {
+        mask = group == CoreGroup::Performance ? &state.performance : &state.efficiency;
+    }
+    if (CPU_COUNT(mask) == 0) {
+        return false;
+    }
+    if (sched_setaffinity(tid, sizeof(*mask), mask) != 0) {
+        if (gone != nullptr && errno == ESRCH) {
+            *gone = true;
+            return false;
+        }
+        LOG_WARNING(Common, "Could not restrict thread {} to its core group: {}", tid,
+                    ::Common::GetLastErrorMsg());
+        return false;
+    }
+    return true;
+}
+
+bool KernelPreservesRequestedAffinity() {
+    utsname info{};
+    if (uname(&info) != 0) {
+        return false;
+    }
+    s32 major = 0;
+    s32 minor = 0;
+    if (std::sscanf(info.release, "%d.%d", &major, &minor) != 2) {
+        return false;
+    }
+    return major > 6 || (major == 6 && minor >= 2);
+}
+
+void SnapshotCanaryLocked(CpuTopologyState& state) {
+    state.canary_valid = false;
+    if (!state.separated) {
+        return;
+    }
+    for (const auto& policy : state.policies) {
+        if (policy.group == CoreGroup::Unrestricted) {
+            continue;
+        }
+        cpu_set_t mask;
+        CPU_ZERO(&mask);
+        if (sched_getaffinity(policy.tid, sizeof(mask), &mask) != 0) {
+            continue;
+        }
+        state.canary_tid = policy.tid;
+        state.canary_mask = mask;
+        state.canary_valid = true;
+        return;
+    }
+}
+
+bool DueForPoll(CpuTopologyState& state) {
+    const auto now = std::chrono::steady_clock::now().time_since_epoch();
+    const s64 now_ns = std::chrono::duration_cast<std::chrono::nanoseconds>(now).count();
+    s64 next = state.next_poll_ns.load(std::memory_order_relaxed);
+    if (now_ns < next) {
+        return false;
+    }
+    return state.next_poll_ns.compare_exchange_strong(
+        next, now_ns + ANDROID_POLICY_POLL_INTERVAL.count(), std::memory_order_relaxed);
+}
+
+ThreadPolicy& AcquirePolicyLocked(CpuTopologyState& state, pid_t tid) {
+    for (auto& policy : state.policies) {
+        if (policy.tid == tid) {
+            return policy;
+        }
+    }
+    return state.policies.emplace_back(ThreadPolicy{tid, CoreGroup::Unrestricted, 0, false});
+}
+
+void SetCurrentThreadCoreGroup(CoreGroup group) {
+    const pid_t tid = gettid();
+    (void)&t_policy_registration;
+
+    CoreGroup effective = group;
+    if (tid == getpid() && group != CoreGroup::Unrestricted) {
+        LOG_WARNING(Common, "Refusing to place the main thread: the CPU topology is read from it");
+        effective = CoreGroup::Unrestricted;
+    }
+
+    CpuTopologyState& state = State();
+    std::scoped_lock topology_lock{state.topology_mutex};
+    EnsureTopologyLocked(state);
+    ApplyCoreGroupLocked(state, tid, effective);
+
+    std::scoped_lock policy_lock{state.policy_mutex};
+    AcquirePolicyLocked(state, tid).group = effective;
+    if (!state.canary_valid) {
+        SnapshotCanaryLocked(state);
+    }
+}
+
+void RememberCurrentThreadNice(pid_t tid, s32 nice_value) {
+    (void)&t_policy_registration;
+    CpuTopologyState& state = State();
+    std::scoped_lock lock{state.policy_mutex};
+    ThreadPolicy& policy = AcquirePolicyLocked(state, tid);
+    policy.nice_value = nice_value;
+    policy.has_nice = true;
+}
+} // Anonymous namespace
+#endif
+
 #include "common/cpu_features.h"
 #ifdef ARCHITECTURE_x86_64
 #ifdef _MSC_VER
@@ -48,7 +446,6 @@
 #endif
 #include "common/x64/rdtsc.h"
 #endif
-#include "core/core_timing.h"
 
 namespace Common {
 
@@ -78,21 +475,31 @@ void SetCurrentThreadPriority(ThreadPriority new_priority) {
         }
     }();
     set_thread_priority(find_thread(NULL), priority);
-#else
-    pthread_t this_thread = pthread_self();
-    const auto scheduling_type = SCHED_OTHER;
-    s32 max_prio = sched_get_priority_max(scheduling_type);
-    s32 min_prio = sched_get_priority_min(scheduling_type);
-    u32 level = (std::max)(u32(new_priority) + 1, 4U);
-
-    struct sched_param params;
-    if (max_prio > min_prio) {
-        params.sched_priority = min_prio + ((max_prio - min_prio) * level) / 4;
-    } else {
-        params.sched_priority = min_prio - ((min_prio - max_prio) * level) / 4;
+#elif defined(__ANDROID__)
+    const int nice_value = NiceValueForPriority(new_priority);
+    const pid_t tid = gettid();
+    if (setpriority(PRIO_PROCESS, static_cast<id_t>(tid), nice_value) != 0) {
+        LOG_WARNING(Common, "Could not set thread nice value to {}: {}", nice_value,
+                    GetLastErrorMsg());
+        return;
+    }
+    RememberCurrentThreadNice(tid, nice_value);
+#elif defined(__linux__)
+    const int nice_value = NiceValueForPriority(new_priority);
+    if (setpriority(PRIO_PROCESS, 0, nice_value) != 0) {
+        LOG_DEBUG(Common, "Could not set thread nice value to {}: {}", nice_value,
+                  GetLastErrorMsg());
+    }
+#else
+    const s32 max_prio = sched_get_priority_max(SCHED_OTHER);
+    const s32 min_prio = sched_get_priority_min(SCHED_OTHER);
+    if (max_prio > min_prio) {
+        const u32 level = (std::min)(static_cast<u32>(new_priority), 4U);
+        sched_param params{};
+        params.sched_priority =
+            min_prio + static_cast<s32>(static_cast<u32>(max_prio - min_prio) * level) / 4;
+        pthread_setschedparam(pthread_self(), SCHED_OTHER, &params);
     }
-
-    pthread_setschedparam(this_thread, scheduling_type, &params);
 #endif
 }
 
@@ -132,29 +539,94 @@ void SetCurrentThreadName(const char* name) {
 #endif
 }
 
-void PinCurrentThreadToPerformanceCore(size_t core_id) {
-    ASSERT(core_id < 4);
-    // If we set a flag for a CPU that doesn't exist, the thread may not be allowed to
-    // run in ANY processor!
-    auto const total_cores = std::thread::hardware_concurrency();
-    if (core_id < total_cores) {
+void SetCurrentThreadToPerformanceCores() {
 #if defined(__ANDROID__)
-        cpu_set_t set;
-        CPU_ZERO(&set);
-        CPU_SET(core_id, &set);
-        sched_setaffinity(pthread_self(), sizeof(set), &set);
-#elif defined(__linux__) || defined(__FreeBSD__)
-        cpu_set_t set;
-        CPU_ZERO(&set);
-        CPU_SET(core_id, &set);
-        pthread_setaffinity_np(pthread_self(), sizeof(set), &set);
-#elif defined(_WIN32)
-        DWORD set = 1UL << core_id;
-        SetThreadAffinityMask(GetCurrentThread(), set);
-#else
-        // No pin functionality implemented
-#endif
+    if (ADPF::AddCurrentThread(ADPF::Session::Render)) {
+        SetCurrentThreadCoreGroup(CoreGroup::Unrestricted);
+        return;
     }
+    SetCurrentThreadCoreGroup(CoreGroup::Performance);
+#endif
+}
+
+void SetCurrentThreadToEfficiencyCores() {
+#if defined(__ANDROID__)
+    if (ADPF::AddCurrentThread(ADPF::Session::Background)) {
+        SetCurrentThreadCoreGroup(CoreGroup::Unrestricted);
+        return;
+    }
+    SetCurrentThreadCoreGroup(CoreGroup::Efficiency);
+#endif
+}
+
+void SetCurrentThreadToBackgroundWork() {
+#if defined(__ANDROID__)
+    ADPF::AddCurrentThread(ADPF::Session::Background);
+    SetCurrentThreadCoreGroup(CoreGroup::Unrestricted);
+#endif
+}
+
+void SetCurrentThreadToAllCores() {
+#if defined(__ANDROID__)
+    ADPF::RemoveCurrentThread();
+    SetCurrentThreadCoreGroup(CoreGroup::Unrestricted);
+#endif
+}
+
+void RefreshThreadPolicies() {
+#if defined(__ANDROID__)
+    CpuTopologyState& state = State();
+    std::scoped_lock topology_lock{state.topology_mutex};
+    RefreshTopologyLocked(state);
+
+    std::scoped_lock policy_lock{state.policy_mutex};
+    std::erase_if(state.policies, [&state](const ThreadPolicy& policy) {
+        bool gone = false;
+        if (policy.has_nice &&
+            setpriority(PRIO_PROCESS, static_cast<id_t>(policy.tid), policy.nice_value) != 0 &&
+            errno == ESRCH) {
+            gone = true;
+        }
+        if (!gone) {
+            ApplyCoreGroupLocked(state, policy.tid, policy.group, &gone);
+        }
+        return gone;
+    });
+    SnapshotCanaryLocked(state);
+#endif
+}
+
+void PollThreadPolicies() {
+#if defined(__ANDROID__)
+    static const bool needed = !KernelPreservesRequestedAffinity();
+    if (!needed) {
+        return;
+    }
+
+    CpuTopologyState& state = State();
+    if (!DueForPoll(state)) {
+        return;
+    }
+
+    pid_t tid;
+    cpu_set_t expected;
+    {
+        std::scoped_lock lock{state.topology_mutex};
+        if (!state.canary_valid) {
+            return;
+        }
+        tid = state.canary_tid;
+        expected = state.canary_mask;
+    }
+
+    cpu_set_t current;
+    CPU_ZERO(&current);
+    if (sched_getaffinity(tid, sizeof(current), &current) == 0 &&
+        std::memcmp(&current, &expected, sizeof(current)) == 0) {
+        return;
+    }
+    RefreshThreadPolicies();
+#endif
 }
 
 #ifdef ARCHITECTURE_x86_64