cfg: asio driver selector must list both 32-bit and 64-bit drivers (#734)

## Link to GitHub Issue or related Pull Request, if one exists
#730 

## Description of change
spicecfg is a 32-bit application. When it uses the ASIO SDK it only saw
the 32-bit drivers.

There are some ASIO drivers that have different names under 32-bit and
64-bit (Xonar AE is one of them)... so we have to manually scan the
registry and surface both the WOW32 and WOW64 nodes.

## Testing
<img width="329" height="309" alt="image"
src="https://github.com/user-attachments/assets/7abecf76-6835-4df5-8c2e-e7b425130f4c"
/>

Checked both 32-bit and 64-bit configurator, identical results.

src/spice2x/CMakeLists.txt

@@ -494,6 +494,8 @@ set(SOURCE_FILES ${SOURCE_FILES}
         # hooks
         hooks/audio/acm.cpp
         hooks/audio/audio.cpp
+        hooks/audio/asio_driver_scan.cpp
+        hooks/audio/asio_proxy.cpp
         hooks/audio/buffer.cpp
         hooks/audio/mme.cpp
         hooks/audio/util.cpp
@@ -502,9 +504,12 @@ set(SOURCE_FILES ${SOURCE_FILES}
         hooks/audio/backends/mmdevice/device.cpp
         hooks/audio/backends/mmdevice/device_collection.cpp
         hooks/audio/backends/mmdevice/device_enumerator.cpp
+        hooks/audio/backends/mmdevice/null_device.cpp
+        hooks/audio/backends/mmdevice/null_discard_backend.cpp
         hooks/audio/backends/wasapi/audio_client.cpp
         hooks/audio/backends/wasapi/audio_render_client.cpp
         hooks/audio/backends/wasapi/downmix.cpp
+        hooks/audio/backends/wasapi/resample.cpp
         hooks/audio/backends/wasapi/dummy_audio_client.cpp
         hooks/audio/backends/wasapi/dummy_audio_clock.cpp
         hooks/audio/backends/wasapi/dummy_audio_render_client.cpp

src/spice2x/games/gitadora/asio.cpp

@@ -21,6 +21,7 @@ namespace games::gitadora {
 
     static decltype(RegCloseKey) *RegCloseKey_orig = nullptr;
     static decltype(RegEnumKeyA) *RegEnumKeyA_orig = nullptr;
+    static decltype(RegOpenKeyA) *RegOpenKeyA_orig = nullptr;
     static decltype(RegOpenKeyExA) *RegOpenKeyExA_orig = nullptr;
     static decltype(RegQueryValueExA) *RegQueryValueExA_orig = nullptr;
 
@@ -30,7 +31,6 @@ namespace games::gitadora {
     static LONG WINAPI RegOpenKeyExA_hook(HKEY hKey, LPCSTR lpSubKey, DWORD ulOptions, REGSAM samDesired,
             PHKEY phkResult)
     {
-        // ASIO\XONAR redirect to ASIO\<configured>
         if (ASIO_DRIVER.has_value() &&
             lpSubKey != nullptr &&
             phkResult != nullptr &&
@@ -60,7 +60,10 @@ namespace games::gitadora {
             return result;
         }
 
-        // open of the ASIO root: hand back a sentinel
+        return RegOpenKeyExA_orig(hKey, lpSubKey, ulOptions, samDesired, phkResult);
+    }
+
+    static LONG WINAPI RegOpenKeyA_hook(HKEY hKey, LPCSTR lpSubKey, PHKEY phkResult) {
         if (ASIO_DRIVER.has_value() &&
             lpSubKey != nullptr &&
             phkResult != nullptr &&
@@ -68,10 +71,10 @@ namespace games::gitadora {
             _stricmp(lpSubKey, "software\\asio") == 0)
         {
             *phkResult = PARENT_ASIO_REG_HANDLE;
-            return RegOpenKeyExA_orig(hKey, lpSubKey, ulOptions, samDesired, &real_asio_reg_handle);
+            return RegOpenKeyA_orig(hKey, lpSubKey, &real_asio_reg_handle);
         }
 
-        return RegOpenKeyExA_orig(hKey, lpSubKey, ulOptions, samDesired, phkResult);
+        return RegOpenKeyA_orig(hKey, lpSubKey, phkResult);
     }
 
     static LONG WINAPI RegEnumKeyA_hook(HKEY hKey, DWORD dwIndex, LPSTR lpName, DWORD cchName) {
@@ -158,6 +161,8 @@ namespace games::gitadora {
                 "RegCloseKey", RegCloseKey_hook, avs::game::DLL_INSTANCE);
         RegEnumKeyA_orig = detour::iat_try(
                 "RegEnumKeyA", RegEnumKeyA_hook, avs::game::DLL_INSTANCE);
+        RegOpenKeyA_orig = detour::iat_try(
+                "RegOpenKeyA", RegOpenKeyA_hook, avs::game::DLL_INSTANCE);
         RegOpenKeyExA_orig = detour::iat_try(
                 "RegOpenKeyExA", RegOpenKeyExA_hook, avs::game::DLL_INSTANCE);
         RegQueryValueExA_orig = detour::iat_try(

src/spice2x/games/gitadora/gitadora.cpp

@@ -8,6 +8,7 @@
 #include <ksmedia.h>
 
 #include "cfg/configurator.h"
+#include "hooks/audio/audio.h"
 #include "hooks/audio/mme.h"
 #include "hooks/graphics/graphics.h"
 #include "misc/wintouchemu.h"
@@ -34,6 +35,7 @@ namespace games::gitadora {
     std::optional<socd::SocdAlgorithm> PICK_ALGO = socd::SocdAlgorithm::PreferRecent;
     std::optional<uint8_t> ARENA_WINDOW_COUNT = std::nullopt;
     std::optional<std::string> ASIO_DRIVER = std::nullopt;
+    bool ALLOW_REALTEK_AUDIO = false;
 
     /*
      * Prevent GitaDora from creating folders on F drive
@@ -593,6 +595,12 @@ namespace games::gitadora {
 #ifdef SPICE64
         // gitadora arena model 
         auto aio = libutils::try_library("libaio.dll");
+
+        // before we start patching and hooking things, detect invalid configuration
+        if (aio != nullptr && !is_arena_model()) {
+            log_fatal("gitadora", "arena model i/o (libaio.dll) detected, but <spec> is not an arena model - bad prop XML files?");
+        }
+
         if (aio != nullptr) {
             SETUPAPI_SETTINGS settings{};
             settings.class_guid[0] = 0x86E0D1E0;
@@ -623,7 +631,25 @@ namespace games::gitadora {
 
             // volume change prevention
             hooks::audio::mme::init(avs::game::DLL_INSTANCE);
-            
+
+            // fake Realtek audio injection
+            // if ASIO init succeeds, game tries to look for audio device with `Realtek` in friendly name
+            // if ASIO init fails, game opens default audio device
+            // therefore, it's safe to enable this hook by default regardless of ASIO preference
+            // (unless the user explicitly disables it, of course)
+            if (ALLOW_REALTEK_AUDIO) {
+                log_info(
+                    "gitadora",
+                    "fake Realtek audio injection disabled "
+                    "(user's real Realtek audio may be used after successful ASIO init)");
+            } else {
+                log_info(
+                    "gitadora",
+                    "fake Realtek audio injection enabled "
+                    "(create a fake Realtek audio device to prevent crashes after successful ASIO init)");
+                hooks::audio::INJECT_FAKE_REALTEK_AUDIO = true;
+            }
+
             // monitor/touch hooks (windowed or full screen)
             if (GRAPHICS_PREVENT_SECONDARY_WINDOWS) {
                 // enable touch hook for subscreen overlay

src/spice2x/games/gitadora/gitadora.h

@@ -20,6 +20,7 @@ namespace games::gitadora {
     extern std::optional<socd::SocdAlgorithm> PICK_ALGO;
     extern std::optional<uint8_t> ARENA_WINDOW_COUNT;
     extern std::optional<std::string> ASIO_DRIVER;
+    extern bool ALLOW_REALTEK_AUDIO;
 
     class GitaDoraGame : public games::Game {
     public:

src/spice2x/games/iidx/iidx.cpp

@@ -893,38 +893,10 @@ namespace games::iidx {
             }
         }
 
-        // patch iidx32+ for asio compatibility
-        // only do this if NOT wasapi (as opposed to checking if it's asio)
-        // the patch is only really needed for (some) non-XONAR devices but since people sometimes disguise
-        // other devices as a XONAR, don't check for the exact string (common ASIO workaround for INF)
-        if (avs::game::is_ext(2024090100, INT_MAX) &&
-            !(SOUND_OUTPUT_DEVICE_IN_EFFECT.has_value() &&
-              SOUND_OUTPUT_DEVICE_IN_EFFECT.value() == "wasapi")) {
-
-            // in iidx32 final:
-            // ff 50 08      call   QWORD PTR [rax+0x8]     ; ASIO instance AddRef
-            // 48 8b 4b 08   mov    rcx,QWORD PTR [rbx+0x8]
-            // 48 8b 01      mov    rax,QWORD PTR [rcx]
-            // ff 50 08      call   QWORD PTR [rax+0x8]     ; ASIO instance AddRef
-
-            intptr_t result = replace_pattern(
-                avs::game::DLL_INSTANCE,
-                "FF50????????????????FF50??4533C94533C0418D51",
-                "????????????????????909090??????????????????",
-                0, 0);
-
-            if (result == 0) {
-                log_warning(
-                    "iidx",
-                    "Failed to apply ASIO compatibility fix for iidx32+. "
-                    "Unless patches are applied, your ASIO device may hang or fail to work");
-            } else {
-                log_info(
-                    "iidx",
-                    "Successfully applied ASIO compatibility fix for iidx32+ using signature matching @ 0x{:x}.",
-                    result);
-            }
-        }
+        // note: the iidx32+ ASIO refcount bug (a duplicate AddRef on the ASIO instance with
+        // no matching Release, which leaks the driver and can hang non-XONAR devices) is now
+        // handled transparently by the WrappedAsio proxy (see hooks/audio/asio_proxy.cpp),
+        // so no game-DLL signature patch is needed here anymore
 
 #endif
 

src/spice2x/hooks/audio/asio_driver_scan.cpp

@@ -0,0 +1,74 @@
+#include "asio_driver_scan.h"
+
+#include <algorithm>
+
+#include <windows.h>
+
+#include "util/utils.h"
+
+namespace hooks::audio {
+
+    static constexpr char ASIO_REG_PATH[] = "software\\asio";
+    static constexpr char ASIO_REG_DESC[] = "description";
+
+    // enumerate a single registry view, appending to entries while merging
+    // duplicates discovered in another view. Drivers are matched by name (not
+    // CLSID): the game's ASIO loader selects drivers by name, and some vendors
+    // register the same CLSID under different 32-bit/64-bit names (e.g. "XONAR
+    // SOUND CARD" vs "XONAR SOUND CARD(64)"), which are distinct user choices.
+    static void scan_view(
+            REGSAM wow64_flag,
+            bool is_64bit,
+            std::vector<AsioDriverScanEntry> &entries) {
+
+        HKEY hkEnum = nullptr;
+        if (RegOpenKeyExA(HKEY_LOCAL_MACHINE, ASIO_REG_PATH, 0,
+                KEY_READ | wow64_flag, &hkEnum) != ERROR_SUCCESS) {
+            return;
+        }
+
+        char key_name[256];
+        for (DWORD index = 0;
+             RegEnumKeyA(hkEnum, index, key_name, sizeof(key_name)) == ERROR_SUCCESS;
+             index++) {
+
+            // read description (display name), fall back to the key name
+            char desc[256] = { 0 };
+            DWORD size = sizeof(desc);
+            std::string name = key_name;
+            if (RegGetValueA(hkEnum,
+                             key_name,
+                             ASIO_REG_DESC,
+                             RRF_RT_REG_SZ | wow64_flag,
+                             nullptr,
+                             desc,
+                             &size) == ERROR_SUCCESS && desc[0]) {
+                name = desc;
+            }
+
+            // merge with an existing entry from the other view (match by name)
+            const std::string name_lower = strtolower(name);
+            auto it = std::find_if(entries.begin(), entries.end(), [&](const auto &e) {
+                return strtolower(e.name) == name_lower;
+            });
+            if (it == entries.end()) {
+                entries.push_back({ name });
+                it = entries.end() - 1;
+            }
+            it->found_32bit |= !is_64bit;
+            it->found_64bit |= is_64bit;
+        }
+
+        RegCloseKey(hkEnum);
+    }
+
+    std::vector<AsioDriverScanEntry> scan_asio_drivers() {
+        std::vector<AsioDriverScanEntry> entries;
+
+        // 64-bit view first so it wins ordering when present in both
+        scan_view(KEY_WOW64_64KEY, true, entries);
+        scan_view(KEY_WOW64_32KEY, false, entries);
+
+        return entries;
+    }
+}

src/spice2x/hooks/audio/asio_driver_scan.h

@@ -0,0 +1,15 @@
+#pragma once
+
+#include <string>
+#include <vector>
+
+namespace hooks::audio {
+
+    struct AsioDriverScanEntry {
+        std::string name;
+        bool found_32bit = false;
+        bool found_64bit = false;
+    };
+
+    std::vector<AsioDriverScanEntry> scan_asio_drivers();
+}

src/spice2x/hooks/audio/asio_proxy.cpp

@@ -0,0 +1,813 @@
+#include "asio_proxy.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <mutex>
+#include <vector>
+
+#include "external/asio/asiolist.h"
+#include "hooks/audio/audio.h"
+#include "util/logging.h"
+#include "util/utils.h"
+
+namespace {
+
+    // readable name for an ASIO sample type (e.g. "ASIOSTInt32LSB"), falling back to the
+    // numeric value for unknown types
+    const char *asio_sample_type_name(AsioSampleType type) {
+        switch (type) {
+            case ASIOSTInt16MSB: return "ASIOSTInt16MSB";
+            case ASIOSTInt24MSB: return "ASIOSTInt24MSB";
+            case ASIOSTInt32MSB: return "ASIOSTInt32MSB";
+            case ASIOSTFloat32MSB: return "ASIOSTFloat32MSB";
+            case ASIOSTFloat64MSB: return "ASIOSTFloat64MSB";
+            case ASIOSTInt32MSB16: return "ASIOSTInt32MSB16";
+            case ASIOSTInt32MSB18: return "ASIOSTInt32MSB18";
+            case ASIOSTInt32MSB20: return "ASIOSTInt32MSB20";
+            case ASIOSTInt32MSB24: return "ASIOSTInt32MSB24";
+            case ASIOSTInt16LSB: return "ASIOSTInt16LSB";
+            case ASIOSTInt24LSB: return "ASIOSTInt24LSB";
+            case ASIOSTInt32LSB: return "ASIOSTInt32LSB";
+            case ASIOSTFloat32LSB: return "ASIOSTFloat32LSB";
+            case ASIOSTFloat64LSB: return "ASIOSTFloat64LSB";
+            case ASIOSTInt32LSB16: return "ASIOSTInt32LSB16";
+            case ASIOSTInt32LSB18: return "ASIOSTInt32LSB18";
+            case ASIOSTInt32LSB20: return "ASIOSTInt32LSB20";
+            case ASIOSTInt32LSB24: return "ASIOSTInt32LSB24";
+            default: return "unknown";
+        }
+    }
+
+    // duration in milliseconds of a buffer of the given frame count at a sample rate, or a
+    // negative sentinel when the frame count or sample rate is unusable
+    double frames_to_ms(long frames, AsioSampleRate sample_rate) {
+        if (frames < 0 || sample_rate <= 0.0) {
+            return -1.0;
+        }
+        return (frames * 1000.0) / sample_rate;
+    }
+
+    // scales one planar ASIO output buffer (frames samples of the given type) by gain in
+    // place, clamping integer formats so a boost saturates instead of wrapping. unsupported
+    // formats are left untouched. runs on the driver's realtime thread, so no allocation,
+    // locking or logging here
+    void apply_gain_planar(void *buffer, long frames, AsioSampleType type, float gain) {
+        if (buffer == nullptr || frames <= 0) {
+            return;
+        }
+        switch (type) {
+            case ASIOSTFloat32LSB: {
+                auto p = static_cast<float *>(buffer);
+                for (long i = 0; i < frames; i++) {
+                    p[i] = std::clamp(p[i] * gain, -1.0f, 1.0f);
+                }
+                break;
+            }
+            case ASIOSTFloat64LSB: {
+                auto p = static_cast<double *>(buffer);
+                for (long i = 0; i < frames; i++) {
+                    p[i] = std::clamp(p[i] * static_cast<double>(gain), -1.0, 1.0);
+                }
+                break;
+            }
+            case ASIOSTInt16LSB: {
+                auto p = static_cast<int16_t *>(buffer);
+                for (long i = 0; i < frames; i++) {
+                    p[i] = static_cast<int16_t>(
+                        std::clamp(std::lround(p[i] * gain), -32768L, 32767L));
+                }
+                break;
+            }
+            case ASIOSTInt24LSB: {
+                // packed 24-bit little-endian, 3 bytes per sample
+                auto bytes = static_cast<uint8_t *>(buffer);
+                for (long i = 0; i < frames; i++) {
+                    uint8_t *s = bytes + i * 3;
+                    int32_t v = s[0] | (s[1] << 8) | (s[2] << 16);
+                    if (v & 0x800000) {
+                        v |= ~0xFFFFFF; // sign extend
+                    }
+                    int64_t scaled = std::clamp<int64_t>(
+                        std::llround(static_cast<double>(v) * gain), -8388608, 8388607);
+                    s[0] = scaled & 0xFF;
+                    s[1] = (scaled >> 8) & 0xFF;
+                    s[2] = (scaled >> 16) & 0xFF;
+                }
+                break;
+            }
+            case ASIOSTInt32LSB: {
+                auto p = static_cast<int32_t *>(buffer);
+                for (long i = 0; i < frames; i++) {
+                    p[i] = static_cast<int32_t>(std::clamp<int64_t>(
+                        std::llround(static_cast<double>(p[i]) * gain), INT32_MIN, INT32_MAX));
+                }
+                break;
+            }
+            default:
+                // unsupported format (MSB, aligned 32-bit, DSD): leave untouched
+                break;
+        }
+    }
+
+    // live wrappers by CLSID. ASIO drivers are single-instance, so a host that creates a
+    // new instance without releasing the old one has leaked it; we track this to tear the
+    // stale one down
+    std::mutex g_wrappers_mutex;
+    std::vector<std::pair<CLSID, WrappedAsio *>> g_wrappers;
+
+    // register a wrapper as the live instance for its CLSID, returning any stale wrapper it
+    // replaces so the caller can tear it down outside the lock
+    WrappedAsio *register_wrapper(REFCLSID clsid, WrappedAsio *wrapper) {
+        std::lock_guard lock(g_wrappers_mutex);
+        for (auto &entry : g_wrappers) {
+            if (IsEqualCLSID(entry.first, clsid)) {
+                WrappedAsio *stale = entry.second;
+                entry.second = wrapper;
+                return stale;
+            }
+        }
+
+        g_wrappers.emplace_back(clsid, wrapper);
+        return nullptr;
+    }
+
+    void unregister_wrapper(WrappedAsio *wrapper) {
+        std::lock_guard lock(g_wrappers_mutex);
+        for (auto it = g_wrappers.begin(); it != g_wrappers.end(); ++it) {
+            if (it->second == wrapper) {
+                g_wrappers.erase(it);
+                return;
+            }
+        }
+    }
+
+    // ASIO drivers registered on this system (CLSID + registry name), scanned once
+    const std::vector<std::pair<CLSID, std::string>> &registered_asio_drivers() {
+        static const std::vector<std::pair<CLSID, std::string>> drivers = [] {
+            std::vector<std::pair<CLSID, std::string>> result;
+            AsioDriverList driver_list;
+            for (const auto &driver : driver_list.driver_list) {
+                result.emplace_back(driver.clsid, driver.name);
+                log_info(
+                    "audio::wrappedasio",
+                    "registered ASIO driver: name='{}', clsid={}",
+                    driver.name,
+                    guid2s(driver.clsid));
+            }
+
+            log_info("audio::wrappedasio", "discovered {} registered ASIO driver(s)", result.size());
+            return result;
+        }();
+
+        return drivers;
+    }
+
+    std::string registered_asio_name(REFCLSID clsid) {
+        for (const auto &driver : registered_asio_drivers()) {
+            if (IsEqualCLSID(driver.first, clsid)) {
+                return driver.second;
+            }
+        }
+
+        return guid2s(clsid);
+    }
+}
+
+namespace hooks::audio::asio {
+
+    bool is_asio_creation(REFCLSID rclsid, REFIID riid) {
+
+        // ASIO hosts request the driver using its own CLSID as the interface id
+        if (!IsEqualGUID(rclsid, riid)) {
+            return false;
+        }
+
+        for (const auto &driver : registered_asio_drivers()) {
+            if (IsEqualCLSID(driver.first, rclsid)) {
+                return true;
+            }
+        }
+
+        return false;
+    }
+}
+
+#pragma region IUnknown
+bool WrappedAsio::FORCE_TWO_CHANNELS = false;
+std::atomic<WrappedAsio *> WrappedAsio::volume_active_instance {nullptr};
+
+WrappedAsio::~WrappedAsio() {
+    unregister_wrapper(this);
+
+    this->detach_volume();
+
+    // our refcount is decoupled from the host's (see AddRef/Release), so pReal's count is
+    // exactly one here and this releases/unloads it deterministically
+    this->pReal->Release();
+
+    log_info("audio::wrappedasio", "destroying wrapped ASIO driver, clsid={}", guid2s(this->clsid));
+}
+
+HRESULT STDMETHODCALLTYPE WrappedAsio::QueryInterface(REFIID riid, void **ppv) {
+    if (ppv == nullptr) {
+        return E_POINTER;
+    }
+
+    // ASIO hosts query for the driver using its own CLSID as the IID
+    if (IsEqualIID(riid, IID_IUnknown) || IsEqualIID(riid, this->clsid)) {
+        this->AddRef();
+        *ppv = static_cast<IAsio *>(this);
+
+        return S_OK;
+    }
+
+    // the host is asking for some other interface; forward to the real driver. a failure
+    // here is a common reason a host discards a driver and retries
+    const HRESULT ret = this->pReal->QueryInterface(riid, ppv);
+    if (SUCCEEDED(ret)) {
+        log_info("audio::wrappedasio", "QueryInterface({}) -> forwarded to real driver", guid2s(riid));
+    } else {
+        log_info(
+            "audio::wrappedasio",
+            "QueryInterface({}) -> not supported by driver, hr={:#x}",
+            guid2s(riid),
+            static_cast<unsigned long>(ret));
+    }
+
+    return ret;
+}
+
+ULONG STDMETHODCALLTYPE WrappedAsio::AddRef() {
+    // decoupled from the real driver: we count host references on the wrapper and hold a
+    // single reference on pReal for our lifetime. this neutralizes a host bug (iidx32+)
+    // that takes a duplicate AddRef with no matching Release, which would leak the driver
+    return ++this->ref_count;
+}
+
+ULONG STDMETHODCALLTYPE WrappedAsio::Release() {
+    const ULONG refs = --this->ref_count;
+    if (refs == 0) {
+        delete this;
+    }
+
+    return refs;
+}
+#pragma endregion
+
+#pragma region IAsio
+AsioBool __thiscall WrappedAsio::init(void *sys_handle) {
+    const AsioBool result = this->pReal->init(sys_handle);
+    if (result == AsioTrue) {
+        log_info(
+            "audio::wrappedasio",
+            "init succeeded for '{}' (driver version {})",
+            this->driver_name,
+            this->pReal->get_driver_version());
+    } else {
+        char message[128] = {};
+        this->pReal->get_error_message(message);
+        log_warning("audio::wrappedasio", "init failed: {}", message);
+    }
+
+    return result;
+}
+
+void __thiscall WrappedAsio::get_driver_name(char *name) {
+    this->pReal->get_driver_name(name);
+}
+
+long __thiscall WrappedAsio::get_driver_version() {
+    return this->pReal->get_driver_version();
+}
+
+void __thiscall WrappedAsio::get_error_message(char *string) {
+    this->pReal->get_error_message(string);
+}
+
+AsioError __thiscall WrappedAsio::start() {
+    const AsioError result = this->pReal->start();
+    if (result == ASE_OK) {
+        log_info(
+            "audio::wrappedasio",
+            "start succeeded, ASIO stream is now running on '{}'",
+            this->driver_name);
+    } else {
+        log_warning("audio::wrappedasio", "start failed, err={}", static_cast<long>(result));
+    }
+
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::stop() {
+    const AsioError result = this->pReal->stop();
+    if (result == ASE_OK) {
+        log_info("audio::wrappedasio", "stop succeeded, ASIO stream on '{}' halted", this->driver_name);
+    } else {
+        log_warning("audio::wrappedasio", "stop failed, err={}", static_cast<long>(result));
+    }
+
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::get_channels(long *num_input_channels, long *num_output_channels) {
+    const AsioError result = this->pReal->get_channels(num_input_channels, num_output_channels);
+    if (result != ASE_OK) {
+        log_warning("audio::wrappedasio", "get_channels failed, err={}", static_cast<long>(result));
+        return result;
+    }
+
+    if (FORCE_TWO_CHANNELS
+        && num_output_channels != nullptr
+        && *num_output_channels < FORCED_OUTPUT_CHANNELS)
+    {
+        // the device has fewer outputs than the game hardcodes; report the count it
+        // expects so it proceeds to create_buffers, where we forward only the real
+        // front pair and discard the rest
+        log_info(
+            "audio::wrappedasio",
+            "reporting output channel count as {} (device has {}) for forced two-channel",
+            FORCED_OUTPUT_CHANNELS,
+            *num_output_channels);
+        *num_output_channels = FORCED_OUTPUT_CHANNELS;
+    }
+
+    log_info(
+        "audio::wrappedasio",
+        "get_channels -> in={}, out={}",
+        num_input_channels ? *num_input_channels : -1,
+        num_output_channels ? *num_output_channels : -1);
+
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::get_latencies(long *input_latency, long *output_latency) {
+    const AsioError result = this->pReal->get_latencies(input_latency, output_latency);
+    if (result == ASE_OK) {
+        // include millisecond equivalents alongside the frame counts for readability
+        AsioSampleRate sample_rate = 0.0;
+        this->pReal->get_sample_rate(&sample_rate);
+        const long in_frames = input_latency ? *input_latency : -1;
+        const long out_frames = output_latency ? *output_latency : -1;
+        log_info(
+            "audio::wrappedasio",
+            "get_latencies -> in={} frames ({:.2f} ms), out={} frames ({:.2f} ms)",
+            in_frames,
+            frames_to_ms(in_frames, sample_rate),
+            out_frames,
+            frames_to_ms(out_frames, sample_rate));
+    } else {
+        log_warning("audio::wrappedasio", "get_latencies failed, err={}", static_cast<long>(result));
+    }
+
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::get_buffer_size(
+    long *min_size,
+    long *max_size,
+    long *preferred_size,
+    long *granularity)
+{
+    const AsioError result = this->pReal->get_buffer_size(min_size, max_size, preferred_size, granularity);
+    if (result != ASE_OK) {
+        log_warning("audio::wrappedasio", "get_buffer_size failed, err={}", static_cast<long>(result));
+        return result;
+    }
+
+    // include millisecond equivalents alongside the frame counts for readability
+    AsioSampleRate sample_rate = 0.0;
+    this->pReal->get_sample_rate(&sample_rate);
+    const long min_frames = min_size ? *min_size : -1;
+    const long max_frames = max_size ? *max_size : -1;
+    const long preferred_frames = preferred_size ? *preferred_size : -1;
+    log_info(
+        "audio::wrappedasio",
+        "get_buffer_size -> min={} frames ({:.2f} ms), max={} frames ({:.2f} ms), "
+        "preferred={} frames ({:.2f} ms), granularity={}",
+        min_frames,
+        frames_to_ms(min_frames, sample_rate),
+        max_frames,
+        frames_to_ms(max_frames, sample_rate),
+        preferred_frames,
+        frames_to_ms(preferred_frames, sample_rate),
+        granularity ? *granularity : -1);
+
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::can_sample_rate(AsioSampleRate sample_rate) {
+    const AsioError result = this->pReal->can_sample_rate(sample_rate);
+    if (result == ASE_OK) {
+        log_misc("audio::wrappedasio", "can_sample_rate({} Hz) -> supported", sample_rate);
+    } else {
+        log_misc(
+            "audio::wrappedasio",
+            "can_sample_rate({} Hz) -> not supported, err={}",
+            sample_rate,
+            static_cast<long>(result));
+    }
+
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::get_sample_rate(AsioSampleRate *sample_rate) {
+    const AsioError result = this->pReal->get_sample_rate(sample_rate);
+    if (result == ASE_OK) {
+        log_misc("audio::wrappedasio", "get_sample_rate -> {} Hz", sample_rate ? *sample_rate : 0.0);
+    } else {
+        log_warning("audio::wrappedasio", "get_sample_rate failed, err={}", static_cast<long>(result));
+    }
+
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::set_sample_rate(AsioSampleRate sample_rate) {
+    const AsioError result = this->pReal->set_sample_rate(sample_rate);
+    if (result == ASE_OK) {
+        log_info("audio::wrappedasio", "set_sample_rate({} Hz) succeeded", sample_rate);
+    } else {
+        log_warning(
+            "audio::wrappedasio",
+            "set_sample_rate({} Hz) failed, err={}",
+            sample_rate,
+            static_cast<long>(result));
+    }
+
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::get_clock_sources(ASIOClockSource *clocks, long *num_sources) {
+    return this->pReal->get_clock_sources(clocks, num_sources);
+}
+
+AsioError __thiscall WrappedAsio::set_clock_source(long reference) {
+    return this->pReal->set_clock_source(reference);
+}
+
+AsioError __thiscall WrappedAsio::get_sample_position(ASIOSamples *s_pos, ASIOTimeStamp *t_stamp) {
+    return this->pReal->get_sample_position(s_pos, t_stamp);
+}
+
+AsioError __thiscall WrappedAsio::get_channel_info(AsioChannelInfo *info) {
+    // forced two-channel: the game probes all output channels it thinks exist, but the
+    // device only has the real front pair. fabricate a plausible entry for the channels
+    // beyond the device without touching the real driver - they are discarded in
+    // create_buffers anyway
+    long real_in = 0, real_out = 0;
+    if (FORCE_TWO_CHANNELS
+        && info != nullptr
+        && info->is_input == AsioFalse
+        && this->pReal->get_channels(&real_in, &real_out) == ASE_OK
+        && info->channel >= real_out)
+    {
+        const long channel = info->channel;
+        info->is_active = AsioTrue;
+        info->channel_group = 0;
+        info->type = ASIOSTInt32LSB;
+        snprintf(info->name, sizeof(info->name), "Fake ASIO OUT %ld", channel);
+        log_info(
+            "audio::wrappedasio",
+            "get_channel_info(channel={}, dir=output) -> fake channel, type={} ({})",
+            channel,
+            asio_sample_type_name(info->type),
+            static_cast<long>(info->type));
+
+        return ASE_OK;
+    }
+
+    const AsioError result = this->pReal->get_channel_info(info);
+    if (result == ASE_OK && info != nullptr) {
+        log_info(
+            "audio::wrappedasio",
+            "get_channel_info(channel={}, dir={}) -> active={}, group={}, type={} ({}), name='{}'",
+            info->channel,
+            info->is_input == AsioTrue ? "input" : "output",
+            info->is_active == AsioTrue,
+            info->channel_group,
+            asio_sample_type_name(info->type),
+            static_cast<long>(info->type),
+            info->name);
+    } else if (result != ASE_OK) {
+        log_warning("audio::wrappedasio", "get_channel_info failed, err={}", static_cast<long>(result));
+    }
+
+    return result;
+}
+
+AsioCallbacks *WrappedAsio::install_volume_callbacks(AsioCallbacks *game_callbacks) {
+    const float gain = hooks::audio::VOLUME_BOOST;
+
+    // start from a clean slate; a previous buffer set may have left state behind
+    this->volume_channels.clear();
+    this->volume_active = false;
+
+    // no boost configured (or no callbacks to wrap): pass the game's callbacks straight
+    // through and do zero realtime work, exactly as before
+    if (gain == 1.0f || game_callbacks == nullptr) {
+        return game_callbacks;
+    }
+
+    this->volume_active = true;
+    this->volume_gain = gain;
+    this->volume_game_callbacks = *game_callbacks;
+
+    // wrap only the buffer-switch callbacks, where the audio data lives and we apply the
+    // gain. the other two carry no data we touch, so forward the game's own pointers
+    // unchanged - the driver expects them non-null and the game already owns their context
+    this->volume_proxy_callbacks = {};
+    this->volume_proxy_callbacks.buffer_switch = &WrappedAsio::volume_buffer_switch;
+    this->volume_proxy_callbacks.sample_rate_did_change = game_callbacks->sample_rate_did_change;
+    this->volume_proxy_callbacks.asio_message = game_callbacks->asio_message;
+    this->volume_proxy_callbacks.buffer_switch_time_info =
+        game_callbacks->buffer_switch_time_info ? &WrappedAsio::volume_buffer_switch_time_info : nullptr;
+
+    return &this->volume_proxy_callbacks;
+}
+
+void WrappedAsio::record_volume_output_channel(const AsioBufferInfo &info) {
+    if (info.is_input != AsioFalse) {
+        return;
+    }
+
+    // ask the real driver for this channel's sample format so the realtime path knows how
+    // to scale it; fall back to a sentinel that apply_gain_planar leaves untouched
+    AsioChannelInfo ci {};
+    ci.channel = info.channel_num;
+    ci.is_input = AsioFalse;
+    AsioSampleType type = ASIOSTLastEntry;
+    if (this->pReal->get_channel_info(&ci) == ASE_OK) {
+        type = ci.type;
+    }
+
+    VolumeOutputChannel ch;
+    ch.buffers[0] = info.buffers[0];
+    ch.buffers[1] = info.buffers[1];
+    ch.type = type;
+    this->volume_channels.push_back(ch);
+}
+
+void WrappedAsio::publish_volume(long buffer_size) {
+    if (!this->volume_active) {
+        return;
+    }
+
+    // everything the realtime thread reads is now in place; make ourselves reachable
+    this->volume_buffer_size = buffer_size;
+    WrappedAsio::volume_active_instance.store(this, std::memory_order_release);
+    log_info(
+        "audio::wrappedasio",
+        "volume boost active: gain={}, scaling {} output channel(s)",
+        this->volume_gain,
+        this->volume_channels.size());
+}
+
+void WrappedAsio::detach_volume() {
+    // stop our realtime trampolines from reaching this wrapper, but only if we are the
+    // currently published instance
+    WrappedAsio *expected = this;
+    WrappedAsio::volume_active_instance.compare_exchange_strong(expected, nullptr);
+}
+
+void WrappedAsio::apply_output_volume(long double_buffer_index) {
+    if (double_buffer_index != 0 && double_buffer_index != 1) {
+        return;
+    }
+    const float gain = this->volume_gain;
+    const long frames = this->volume_buffer_size;
+    for (const VolumeOutputChannel &ch : this->volume_channels) {
+        apply_gain_planar(ch.buffers[double_buffer_index], frames, ch.type, gain);
+    }
+}
+
+void __cdecl WrappedAsio::volume_buffer_switch(long double_buffer_index, AsioBool direct_process) {
+    WrappedAsio *self = WrappedAsio::volume_active_instance.load(std::memory_order_acquire);
+    if (self == nullptr) {
+        return;
+    }
+
+    // let the game write its samples into the driver buffers first, then scale them before
+    // the driver plays this half on the next switch
+    if (self->volume_game_callbacks.buffer_switch != nullptr) {
+        self->volume_game_callbacks.buffer_switch(double_buffer_index, direct_process);
+    }
+    self->apply_output_volume(double_buffer_index);
+}
+
+AsioTime * __cdecl WrappedAsio::volume_buffer_switch_time_info(
+    AsioTime *params, long double_buffer_index, AsioBool direct_process)
+{
+    WrappedAsio *self = WrappedAsio::volume_active_instance.load(std::memory_order_acquire);
+    if (self == nullptr) {
+        return params;
+    }
+
+    AsioTime *ret = params;
+    if (self->volume_game_callbacks.buffer_switch_time_info != nullptr) {
+        ret = self->volume_game_callbacks.buffer_switch_time_info(
+            params, double_buffer_index, direct_process);
+    } else if (self->volume_game_callbacks.buffer_switch != nullptr) {
+        self->volume_game_callbacks.buffer_switch(double_buffer_index, direct_process);
+    }
+    self->apply_output_volume(double_buffer_index);
+    return ret;
+}
+
+AsioError __thiscall WrappedAsio::create_buffers(
+    AsioBufferInfo *buffer_infos,
+    long num_channels,
+    long buffer_size,
+    AsioCallbacks *callbacks)
+{
+    // swap in our buffer-switch trampolines if a volume boost is configured, so the real
+    // driver calls us and we scale its output after the game fills it (no-op otherwise)
+    AsioCallbacks *effective = this->install_volume_callbacks(callbacks);
+
+    if (FORCE_TWO_CHANNELS) {
+        return this->create_buffers_front_pair(buffer_infos, num_channels, buffer_size, effective);
+    }
+
+    const AsioError result = this->pReal->create_buffers(buffer_infos, num_channels, buffer_size, effective);
+    if (result == ASE_OK) {
+        log_info(
+            "audio::wrappedasio",
+            "create_buffers(channels={}, size={} frames) succeeded",
+            num_channels,
+            buffer_size);
+
+        // capture the device output channels we will scale, then publish ourselves to the
+        // realtime thread once everything is in place
+        if (this->volume_active) {
+            for (long i = 0; i < num_channels; i++) {
+                this->record_volume_output_channel(buffer_infos[i]);
+            }
+            this->publish_volume(buffer_size);
+        }
+    } else {
+        log_warning(
+            "audio::wrappedasio",
+            "create_buffers(channels={}, size={} frames) failed, err={}",
+            num_channels,
+            buffer_size,
+            static_cast<long>(result));
+    }
+
+    return result;
+}
+
+AsioError WrappedAsio::create_buffers_front_pair(
+    AsioBufferInfo *buffer_infos,
+    long num_channels,
+    long buffer_size,
+    AsioCallbacks *callbacks)
+{
+    // front-pair extraction (forced two-channel ASIO): the game asks for more output
+    // channels than the real device has (e.g. 8 vs 2). forward only the channels the
+    // device actually provides (channel 0/1 = front L/R) and hand the game throwaway
+    // buffers for the rest, so its front mix lands on the device and the surround
+    // channels are discarded. the game writes directly into the driver/dummy buffers
+    // from its own bufferSwitch; the only realtime work we do is the volume boost (if
+    // configured), which scales the forwarded device channels via our trampolines
+    long real_in = 0, real_out = 0;
+    const AsioError ch_result = this->pReal->get_channels(&real_in, &real_out);
+    if (ch_result != ASE_OK) {
+        log_warning(
+            "audio::wrappedasio",
+            "create_buffers: get_channels failed, err={}",
+            static_cast<long>(ch_result));
+        return ch_result;
+    }
+
+    // partition the requested channels: those the device can serve are forwarded, the rest
+    // are discarded. record source indices for both so we can patch the game's array after
+    std::vector<AsioBufferInfo> forwarded;
+    std::vector<long> forwarded_src;
+    std::vector<long> discarded_src;
+    forwarded.reserve(num_channels);
+    forwarded_src.reserve(num_channels);
+    discarded_src.reserve(num_channels);
+    for (long i = 0; i < num_channels; i++) {
+        const AsioBufferInfo &bi = buffer_infos[i];
+        const long limit = (bi.is_input == AsioTrue) ? real_in : real_out;
+        if (bi.channel_num < limit) {
+            forwarded.push_back(bi);
+            forwarded_src.push_back(i);
+        } else {
+            discarded_src.push_back(i);
+        }
+    }
+
+    const AsioError result = this->pReal->create_buffers(
+        forwarded.data(), static_cast<long>(forwarded.size()), buffer_size, callbacks);
+    if (result != ASE_OK) {
+        log_warning(
+            "audio::wrappedasio",
+            "create_buffers(forwarded={} of {}, size={} frames) failed, err={}",
+            forwarded.size(),
+            num_channels,
+            buffer_size,
+            static_cast<long>(result));
+        return result;
+    }
+
+    // copy the real driver buffer pointers back into the game's array
+    for (size_t k = 0; k < forwarded.size(); k++) {
+        AsioBufferInfo &dst = buffer_infos[forwarded_src[k]];
+        dst.buffers[0] = forwarded[k].buffers[0];
+        dst.buffers[1] = forwarded[k].buffers[1];
+
+        // only the forwarded channels reach the device, so those are the ones the volume
+        // boost scales (the discarded channels go to throwaway buffers below)
+        if (this->volume_active) {
+            this->record_volume_output_channel(dst);
+        }
+    }
+
+    this->publish_volume(buffer_size);
+
+    // hand throwaway double buffers to the discarded channels. sized generously at
+    // 8 bytes/sample (covers every ASIO sample type) so the game can never overrun them
+    // regardless of the negotiated format
+    this->dummy_buffers.clear();
+    this->dummy_buffers.reserve(discarded_src.size() * 2);
+    const size_t dummy_bytes = static_cast<size_t>(buffer_size) * 8;
+    for (const long i : discarded_src) {
+        for (void *&buffer : buffer_infos[i].buffers) {
+            auto buf = std::make_unique<uint8_t[]>(dummy_bytes);
+            std::memset(buf.get(), 0, dummy_bytes);
+            buffer = buf.get();
+            this->dummy_buffers.push_back(std::move(buf));
+        }
+    }
+
+    log_info(
+        "audio::wrappedasio",
+        "create_buffers: front-pair extraction - forwarded {} channel(s) to device, "
+        "discarded {} (requested {}, size={} frames)",
+        forwarded.size(),
+        discarded_src.size(),
+        num_channels,
+        buffer_size);
+
+    return ASE_OK;
+}
+
+AsioError __thiscall WrappedAsio::dispose_buffers() {
+    // stop our realtime trampolines from touching buffers the driver is about to free
+    this->detach_volume();
+
+    const AsioError result = this->pReal->dispose_buffers();
+    this->dummy_buffers.clear();
+    this->volume_channels.clear();
+    this->volume_active = false;
+    return result;
+}
+
+AsioError __thiscall WrappedAsio::control_panel() {
+    return this->pReal->control_panel();
+}
+
+AsioError __thiscall WrappedAsio::future(long selector, void *opt) {
+    return this->pReal->future(selector, opt);
+}
+
+AsioError __thiscall WrappedAsio::output_ready() {
+    return this->pReal->output_ready();
+}
+#pragma endregion
+
+namespace hooks::audio::asio {
+
+    IUnknown *wrap(REFCLSID clsid, void *real) {
+        log_info("audio::wrappedasio", "wrapping ASIO driver interface, clsid={}", guid2s(clsid));
+
+        auto *wrapper = new WrappedAsio(
+            reinterpret_cast<IAsio *>(real), clsid, registered_asio_name(clsid));
+
+        // if the host already had a live wrapper for this CLSID it leaked the previous
+        // instance (ASIO is single-instance); tear it down now so the real driver is
+        // released before the reinit. works around games (iidx32+) that ref twice but
+        // deref once before re-initializing.
+        //
+        // FlexASIO 1.9 and many DAC ASIO drivers can't handle this; FlexASIO 1.10 and
+        // Xonar AE can
+        if (WrappedAsio *stale = register_wrapper(clsid, wrapper)) {
+            log_info(
+                "audio::wrappedasio",
+                "host did not release previous instance for clsid={}, forcing teardown",
+                guid2s(clsid));
+
+            // the stale instance may still have a running stream on the driver's realtime
+            // thread. stop and dispose it before deleting so no in-flight buffer switch
+            // (e.g. our volume trampoline) touches buffers we are about to free. ASIO
+            // guarantees no further buffer_switch once stop() returns, and dispose_buffers
+            // detaches our trampolines, fully quiescing the realtime path before delete
+            stale->stop();
+            stale->dispose_buffers();
+            delete stale;
+        }
+
+        return static_cast<IAsio *>(wrapper);
+    }
+}

src/spice2x/hooks/audio/asio_proxy.h

@@ -0,0 +1,160 @@
+#pragma once
+
+#include <atomic>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include <windows.h>
+
+#include "external/asio/asio.h"
+#include "external/asio/iasiodrv.h"
+
+namespace hooks::audio::asio {
+
+    // returns true if a CoCreateInstance call is instantiating a registered ASIO driver.
+    // ASIO hosts pass the driver CLSID as both class id and interface id; we also validate
+    // it against the system's registered ASIO drivers to avoid false positives
+    bool is_asio_creation(REFCLSID rclsid, REFIID riid);
+
+    // wrap a real ASIO driver instance, taking ownership of the supplied reference, and
+    // return a proxy that forwards every call to it
+    IUnknown *wrap(REFCLSID clsid, void *real);
+}
+
+// transparent proxy around a real ASIO driver; a single place to intercept ASIO traffic
+struct WrappedAsio final : IAsio {
+    WrappedAsio(IAsio *real, REFCLSID clsid, std::string name)
+        : pReal(real), clsid(clsid), driver_name(std::move(name)) {
+    }
+
+    WrappedAsio(const WrappedAsio &) = delete;
+    WrappedAsio &operator=(const WrappedAsio &) = delete;
+
+    virtual ~WrappedAsio();
+
+    // when set, the proxy presents the game's expected multichannel layout to the host so
+    // it proceeds to create_buffers, then forwards only the device's real front pair and
+    // discards the rest (see create_buffers). set once at boot, before any wrapper exists,
+    // so it needs no synchronization
+    static bool FORCE_TWO_CHANNELS;
+
+    // some games hardcode a multichannel ASIO output and bail before create_buffers if
+    // get_channels reports fewer, so we report at least this many output channels when
+    // FORCE_TWO_CHANNELS is active
+    static constexpr long FORCED_OUTPUT_CHANNELS = 8;
+
+#pragma region IUnknown
+    HRESULT STDMETHODCALLTYPE QueryInterface(REFIID riid, void **ppv) override;
+    ULONG STDMETHODCALLTYPE AddRef() override;
+    ULONG STDMETHODCALLTYPE Release() override;
+#pragma endregion
+
+#pragma region IAsio
+    AsioBool __thiscall init(void *sys_handle) override;
+    void __thiscall get_driver_name(char *name) override;
+    long __thiscall get_driver_version() override;
+    void __thiscall get_error_message(char *string) override;
+    AsioError __thiscall start() override;
+    AsioError __thiscall stop() override;
+    AsioError __thiscall get_channels(long *num_input_channels, long *num_output_channels) override;
+    AsioError __thiscall get_latencies(long *input_latency, long *output_latency) override;
+    AsioError __thiscall get_buffer_size(
+        long *min_size,
+        long *max_size,
+        long *preferred_size,
+        long *granularity) override;
+    AsioError __thiscall can_sample_rate(AsioSampleRate sample_rate) override;
+    AsioError __thiscall get_sample_rate(AsioSampleRate *sample_rate) override;
+    AsioError __thiscall set_sample_rate(AsioSampleRate sample_rate) override;
+    AsioError __thiscall get_clock_sources(ASIOClockSource *clocks, long *num_sources) override;
+    AsioError __thiscall set_clock_source(long reference) override;
+    AsioError __thiscall get_sample_position(ASIOSamples *s_pos, ASIOTimeStamp *t_stamp) override;
+    AsioError __thiscall get_channel_info(AsioChannelInfo *info) override;
+    AsioError __thiscall create_buffers(
+        AsioBufferInfo *buffer_infos,
+        long num_channels,
+        long buffer_size,
+        AsioCallbacks *callbacks) override;
+    AsioError __thiscall dispose_buffers() override;
+    AsioError __thiscall control_panel() override;
+    AsioError __thiscall future(long selector, void *opt) override;
+    AsioError __thiscall output_ready() override;
+#pragma endregion
+
+private:
+    // create_buffers implementation used when FORCE_TWO_CHANNELS is active: forwards only
+    // the channels the real device has and hands the game throwaway buffers for the rest
+    AsioError create_buffers_front_pair(
+        AsioBufferInfo *buffer_infos,
+        long num_channels,
+        long buffer_size,
+        AsioCallbacks *callbacks);
+
+    // if hooks::audio::VOLUME_BOOST is set, saves the game's callbacks and returns a proxy
+    // callback set (our buffer-switch trampolines) to hand the real driver instead, so we
+    // can scale its output buffers after the game fills them. otherwise returns the game's
+    // callbacks unchanged. called at create_buffers time, before the stream starts
+    AsioCallbacks *install_volume_callbacks(AsioCallbacks *game_callbacks);
+
+    // records a device output channel whose buffers we scale by the volume boost. queries
+    // the real driver for the channel's sample format. called at create_buffers time
+    void record_volume_output_channel(const AsioBufferInfo &info);
+
+    // publishes the captured volume state to the realtime thread once the buffers exist,
+    // making our trampolines start scaling. called at the end of either create_buffers path
+    void publish_volume(long buffer_size);
+
+    // detaches this instance from the realtime trampolines so they stop touching its
+    // buffers. called from dispose_buffers and the destructor
+    void detach_volume();
+
+    // multiplies every recorded output channel's buffer for the given double-buffer index
+    // by the volume boost. runs on the driver's realtime thread from our buffer switch
+    void apply_output_volume(long double_buffer_index);
+
+    // realtime-thread trampolines for the buffer-switch callbacks, handed to the real
+    // driver in place of the game's; ASIO callbacks carry no user data, so they reach the
+    // active wrapper through volume_active_instance, call the game's original, then scale.
+    // the other two callbacks (sample_rate_did_change, asio_message) are forwarded as the
+    // game's own pointers, so they need no trampoline
+    static void __cdecl volume_buffer_switch(long double_buffer_index, AsioBool direct_process);
+    static AsioTime * __cdecl volume_buffer_switch_time_info(
+        AsioTime *params, long double_buffer_index, AsioBool direct_process);
+
+    // the single wrapper whose proxy callbacks are installed (ASIO is single-instance with
+    // one running stream); read by the static trampolines to reach the right wrapper
+    static std::atomic<WrappedAsio *> volume_active_instance;
+
+    IAsio *const pReal;
+    const CLSID clsid;
+
+    // registry name of the driver (not get_driver_name), used in our logs as a single
+    // unambiguous name; constant for our lifetime
+    std::string driver_name;
+
+    // our own reference count; we hold one reference on pReal and release it when this
+    // drops to zero
+    std::atomic<ULONG> ref_count {1};
+
+    // throwaway double buffers handed to the channels we discard when FORCE_TWO_CHANNELS
+    // is active (see create_buffers). owned for the lifetime of the buffer set and freed
+    // in dispose_buffers; only read by the game from its own bufferSwitch, never by us
+    std::vector<std::unique_ptr<uint8_t[]>> dummy_buffers;
+
+    // one device output channel scaled by the volume boost in our buffer switch
+    struct VolumeOutputChannel {
+        void *buffers[2];
+        AsioSampleType type;
+    };
+
+    // volume boost state, captured at create_buffers time and published to the realtime
+    // thread via volume_active_instance once fully built; untouched while the stream runs.
+    // volume_active gates whether we install our proxy callbacks at all
+    bool volume_active = false;
+    float volume_gain = 1.0f;
+    long volume_buffer_size = 0;
+    AsioCallbacks volume_game_callbacks {};
+    AsioCallbacks volume_proxy_callbacks {};
+    std::vector<VolumeOutputChannel> volume_channels;
+};

src/spice2x/hooks/audio/audio.cpp

@@ -14,6 +14,7 @@
 
 #include "audio_private.h"
 #include "acm.h"
+#include "asio_proxy.h"
 
 #ifdef _MSC_VER
 DEFINE_GUID(CLSID_MMDeviceEnumerator,
@@ -41,9 +42,12 @@ namespace hooks::audio {
     bool VOLUME_HOOK_ENABLED = true;
     std::optional<DownmixAlgorithm> DOWNMIX_ALGORITHM = std::nullopt;
     float VOLUME_BOOST = 1.0f;
+    std::optional<uint32_t> RESAMPLE_RATE = std::nullopt;
+    std::optional<uint32_t> EXCLUSIVE_BUFFER_MS = std::nullopt;
     bool USE_DUMMY = false;
     WAVEFORMATEXTENSIBLE FORMAT {};
     std::optional<Backend> BACKEND = std::nullopt;
+    bool INJECT_FAKE_REALTEK_AUDIO = false;
     std::optional<size_t> ASIO_DRIVER_ID = std::nullopt;
     std::string ASIO_DRIVER_NAME = "";
     bool ASIO_FORCE_UNLOAD_ON_STOP = false;
@@ -93,6 +97,10 @@ static HRESULT STDAPICALLTYPE CoCreateInstance_hook(
         // wrap object
         auto mmde = reinterpret_cast<IMMDeviceEnumerator **>(ppv);
         *mmde = new WrappedIMMDeviceEnumerator(*mmde);
+    } else if (ppv != nullptr && *ppv != nullptr && hooks::audio::asio::is_asio_creation(rclsid, riid)) {
+
+        // wrap every ASIO driver so calls pass through to the real driver
+        *ppv = hooks::audio::asio::wrap(rclsid, *ppv);
     }
 
     // return original result

src/spice2x/hooks/audio/audio.h

@@ -1,5 +1,6 @@
 #pragma once
 
+#include <cstdint>
 #include <optional>
 #include <string>
 
@@ -29,9 +30,21 @@ namespace hooks::audio {
     extern bool VOLUME_HOOK_ENABLED;
     extern std::optional<DownmixAlgorithm> DOWNMIX_ALGORITHM;
     extern float VOLUME_BOOST;
+
+    // target sample rate the hooked output is resampled to, if set
+    extern std::optional<uint32_t> RESAMPLE_RATE;
+
+    // minimum WASAPI exclusive buffer duration (milliseconds), if set. enlarges the device buffer
+    // to avoid underrun crackle on endpoints that cannot service a tiny buffer in time.
+    extern std::optional<uint32_t> EXCLUSIVE_BUFFER_MS;
     extern bool USE_DUMMY;
     extern WAVEFORMATEXTENSIBLE FORMAT;
     extern std::optional<Backend> BACKEND;
+
+    // when true, a synthetic "Realtek" render endpoint is injected into device enumeration that
+    // discards all audio. used by gitadora arena, whose device search crashes when no render
+    // endpoint reports a "Realtek" friendly name.
+    extern bool INJECT_FAKE_REALTEK_AUDIO;
     extern std::optional<size_t> ASIO_DRIVER_ID;
     extern std::string ASIO_DRIVER_NAME;
     extern bool ASIO_FORCE_UNLOAD_ON_STOP;

src/spice2x/hooks/audio/backends/mmdevice/device_collection.cpp

@@ -1,5 +1,6 @@
 #include "device_collection.h"
 #include "device.h"
+#include "null_device.h"
 #include "util/utils.h"
 #include "util/logging.h"
 
@@ -28,17 +29,48 @@ ULONG STDMETHODCALLTYPE WrappedIMMDeviceCollection::Release() {
     return refs;
 }
 
+bool WrappedIMMDeviceCollection::should_inject_fake_realtek() const {
+    return null_render_device_enabled()
+        && (data_flow == eRender || data_flow == eAll);
+}
+
 HRESULT STDMETHODCALLTYPE WrappedIMMDeviceCollection::GetCount(UINT *pcDevices) {
+    // when active, hide all real devices and present only the synthetic one
+    if (should_inject_fake_realtek()) {
+        if (pcDevices == nullptr) {
+            return E_POINTER;
+        }
+        *pcDevices = 1;
+        return S_OK;
+    }
+
     return pReal->GetCount(pcDevices);
 }
 
 HRESULT STDMETHODCALLTYPE WrappedIMMDeviceCollection::Item(UINT nDevice, IMMDevice **ppDevice) {
+    if (ppDevice == nullptr) {
+        return E_POINTER;
+    }
+
+    // when active, the only device in the collection is the synthetic fake Realtek
+    // render device; all real devices are hidden
+    if (should_inject_fake_realtek()) {
+        if (nDevice != 0) {
+            return E_INVALIDARG;
+        }
+        log_info("audio", "WrappedIMMDeviceCollection::Item[{}] -> synthetic fake Realtek render device", nDevice);
+        *ppDevice = new NullMMDevice();
+        return S_OK;
+    }
+
     log_info("audio", "WrappedIMMDeviceCollection::Item[{}]", nDevice);
 
     // call original
     const auto hr = pReal->Item(nDevice, ppDevice);
 
     // wrap interface
-    *ppDevice = new WrappedIMMDevice(*ppDevice);
+    if (SUCCEEDED(hr) && *ppDevice != nullptr) {
+        *ppDevice = new WrappedIMMDevice(*ppDevice);
+    }
     return hr;
 }

src/spice2x/hooks/audio/backends/mmdevice/device_collection.h

@@ -4,7 +4,8 @@
 #include <mmdeviceapi.h>
 
 struct WrappedIMMDeviceCollection : IMMDeviceCollection {
-    explicit WrappedIMMDeviceCollection(IMMDeviceCollection *orig) : pReal(orig) {
+    WrappedIMMDeviceCollection(IMMDeviceCollection *orig, EDataFlow dataFlow)
+        : pReal(orig), data_flow(dataFlow) {
     }
 
     WrappedIMMDeviceCollection(const WrappedIMMDeviceCollection &) = delete;
@@ -24,5 +25,9 @@ struct WrappedIMMDeviceCollection : IMMDeviceCollection {
 #pragma endregion
 
 private:
+    // whether the synthetic fake Realtek render device should be appended to this collection
+    bool should_inject_fake_realtek() const;
+
     IMMDeviceCollection *const pReal;
+    const EDataFlow data_flow;
 };

src/spice2x/hooks/audio/backends/mmdevice/device_enumerator.cpp

@@ -45,7 +45,7 @@ HRESULT STDMETHODCALLTYPE WrappedIMMDeviceEnumerator::EnumAudioEndpoints(
 {
     const auto hr = pReal->EnumAudioEndpoints(dataFlow, dwStateMask, ppDevices);
     if (SUCCEEDED(hr) && (ppDevices != nullptr) && (*ppDevices != nullptr)) {
-        *ppDevices = new WrappedIMMDeviceCollection(*ppDevices);
+        *ppDevices = new WrappedIMMDeviceCollection(*ppDevices, dataFlow);
     }
     return hr;
 }

src/spice2x/hooks/audio/backends/mmdevice/null_device.cpp

@@ -0,0 +1,195 @@
+#include "null_device.h"
+
+#include <atomic>
+#include <cstring>
+
+#include <audioclient.h>
+
+#include "hooks/audio/audio.h"
+#include "hooks/audio/audio_private.h"
+#include "hooks/audio/backends/wasapi/dummy_audio_client.h"
+#include "util/logging.h"
+#include "util/utils.h"
+
+#include "null_discard_backend.h"
+
+// friendly name reported by the synthetic device. must contain "Realtek" so the
+// gitadora arena device search matches it.
+static const wchar_t NULL_DEVICE_FRIENDLY_NAME[] = L"Realtek High Definition Audio";
+
+// arbitrary identifier reported by the synthetic device.
+static const wchar_t NULL_DEVICE_ID[] = L"{spice2x-null-render-device}";
+
+// PKEY_Device_FriendlyName, hardcoded to avoid pulling in functiondiscoverykeys_devpkey.h
+static const PROPERTYKEY PKEY_DEVICE_FRIENDLY_NAME_LOCAL = {
+    { 0xa45c254e, 0xdf1c, 0x4efd, { 0x80, 0x20, 0x67, 0xd1, 0x46, 0xa8, 0x50, 0xe0 } },
+    14
+};
+
+bool null_render_device_enabled() {
+    return hooks::audio::INJECT_FAKE_REALTEK_AUDIO;
+}
+
+// duplicate a wide string into CoTaskMem so the caller can free it with
+// CoTaskMemFree / PropVariantClear as the COM API contract requires.
+static LPWSTR co_task_wcsdup(const wchar_t *src) {
+    const size_t bytes = (wcslen(src) + 1) * sizeof(wchar_t);
+    auto *dst = static_cast<LPWSTR>(CoTaskMemAlloc(bytes));
+    if (dst != nullptr) {
+        memcpy(dst, src, bytes);
+    }
+    return dst;
+}
+
+namespace {
+
+    // minimal IPropertyStore that only answers PKEY_Device_FriendlyName.
+    struct NullPropertyStore : IPropertyStore {
+        std::atomic<ULONG> ref_cnt = 1;
+
+        virtual ~NullPropertyStore() = default;
+
+        HRESULT STDMETHODCALLTYPE QueryInterface(REFIID riid, void **ppvObj) override {
+            if (ppvObj == nullptr) {
+                return E_POINTER;
+            }
+            if (riid == __uuidof(IUnknown) || riid == __uuidof(IPropertyStore)) {
+                this->AddRef();
+                *ppvObj = this;
+                return S_OK;
+            }
+            *ppvObj = nullptr;
+            return E_NOINTERFACE;
+        }
+        ULONG STDMETHODCALLTYPE AddRef() override {
+            return ++this->ref_cnt;
+        }
+        ULONG STDMETHODCALLTYPE Release() override {
+            const ULONG refs = --this->ref_cnt;
+            if (refs == 0) {
+                delete this;
+            }
+            return refs;
+        }
+        HRESULT STDMETHODCALLTYPE GetCount(DWORD *cProps) override {
+            if (cProps == nullptr) {
+                return E_POINTER;
+            }
+            *cProps = 1;
+            return S_OK;
+        }
+        HRESULT STDMETHODCALLTYPE GetAt(DWORD iProp, PROPERTYKEY *pkey) override {
+            if (pkey == nullptr) {
+                return E_POINTER;
+            }
+            if (iProp != 0) {
+                return E_INVALIDARG;
+            }
+            *pkey = PKEY_DEVICE_FRIENDLY_NAME_LOCAL;
+            return S_OK;
+        }
+        HRESULT STDMETHODCALLTYPE GetValue(REFPROPERTYKEY key, PROPVARIANT *pv) override {
+            if (pv == nullptr) {
+                return E_POINTER;
+            }
+            PropVariantInit(pv);
+            if (key.fmtid == PKEY_DEVICE_FRIENDLY_NAME_LOCAL.fmtid
+                    && key.pid == PKEY_DEVICE_FRIENDLY_NAME_LOCAL.pid) {
+                pv->pwszVal = co_task_wcsdup(NULL_DEVICE_FRIENDLY_NAME);
+                if (pv->pwszVal == nullptr) {
+                    return E_OUTOFMEMORY;
+                }
+                pv->vt = VT_LPWSTR;
+            }
+            // unknown keys are returned as VT_EMPTY / S_OK
+            return S_OK;
+        }
+        HRESULT STDMETHODCALLTYPE SetValue(REFPROPERTYKEY, REFPROPVARIANT) override {
+            return STG_E_ACCESSDENIED;
+        }
+        HRESULT STDMETHODCALLTYPE Commit() override {
+            return S_OK;
+        }
+    };
+}
+
+#pragma region IUnknown
+HRESULT STDMETHODCALLTYPE NullMMDevice::QueryInterface(REFIID riid, void **ppvObj) {
+    if (ppvObj == nullptr) {
+        return E_POINTER;
+    }
+    if (riid == __uuidof(IUnknown) || riid == __uuidof(IMMDevice)) {
+        this->AddRef();
+        *ppvObj = this;
+        return S_OK;
+    }
+    *ppvObj = nullptr;
+    return E_NOINTERFACE;
+}
+ULONG STDMETHODCALLTYPE NullMMDevice::AddRef() {
+    return ++this->ref_cnt;
+}
+ULONG STDMETHODCALLTYPE NullMMDevice::Release() {
+    const ULONG refs = --this->ref_cnt;
+    if (refs == 0) {
+        delete this;
+    }
+    return refs;
+}
+#pragma endregion
+
+#pragma region IMMDevice
+HRESULT STDMETHODCALLTYPE NullMMDevice::Activate(
+    REFIID iid,
+    DWORD,
+    PROPVARIANT *,
+    void **ppInterface)
+{
+    if (ppInterface == nullptr) {
+        return E_POINTER;
+    }
+    *ppInterface = nullptr;
+
+    log_info("audio::null", "NullMMDevice::Activate {}", guid2s(iid));
+
+    if (iid == IID_IAudioClient) {
+
+        // release any previously persisted client
+        if (hooks::audio::CLIENT != nullptr) {
+            hooks::audio::CLIENT->Release();
+        }
+
+        auto *client = static_cast<IAudioClient *>(new DummyIAudioClient(new NullDiscardBackend()));
+        *ppInterface = client;
+
+        // persist the audio client
+        hooks::audio::CLIENT = client;
+        hooks::audio::CLIENT->AddRef();
+
+        return S_OK;
+    }
+
+    return E_NOINTERFACE;
+}
+HRESULT STDMETHODCALLTYPE NullMMDevice::OpenPropertyStore(DWORD, IPropertyStore **ppProperties) {
+    if (ppProperties == nullptr) {
+        return E_POINTER;
+    }
+    *ppProperties = new NullPropertyStore();
+    return S_OK;
+}
+HRESULT STDMETHODCALLTYPE NullMMDevice::GetId(LPWSTR *ppstrId) {
+    if (ppstrId == nullptr) {
+        return E_POINTER;
+    }
+    *ppstrId = co_task_wcsdup(NULL_DEVICE_ID);
+    return *ppstrId != nullptr ? S_OK : E_OUTOFMEMORY;
+}
+HRESULT STDMETHODCALLTYPE NullMMDevice::GetState(DWORD *pdwState) {
+    if (pdwState == nullptr) {
+        return E_POINTER;
+    }
+    *pdwState = DEVICE_STATE_ACTIVE;
+    return S_OK;
+}
+#pragma endregion

src/spice2x/hooks/audio/backends/mmdevice/null_device.h

@@ -0,0 +1,39 @@
+#pragma once
+
+#include <atomic>
+
+#include <mmdeviceapi.h>
+
+// returns true when a synthetic render endpoint should be injected into device
+// enumeration. games like gitadora arena search the render endpoint list for a
+// device whose friendly name contains "Realtek" and crash with a null pointer
+// dereference when no match exists. presenting a fake match that routes to the
+// null audio backend lets the search succeed while discarding the audio.
+bool null_render_device_enabled();
+
+// fake IMMDevice that reports a "Realtek" friendly name and activates straight
+// into the null audio backend, never touching real hardware.
+struct NullMMDevice : IMMDevice {
+    NullMMDevice() = default;
+
+    NullMMDevice(const NullMMDevice &) = delete;
+    NullMMDevice &operator=(const NullMMDevice &) = delete;
+
+    virtual ~NullMMDevice() = default;
+
+#pragma region IUnknown
+    HRESULT STDMETHODCALLTYPE QueryInterface(REFIID riid, void **ppvObj) override;
+    ULONG STDMETHODCALLTYPE AddRef() override;
+    ULONG STDMETHODCALLTYPE Release() override;
+#pragma endregion
+
+#pragma region IMMDevice
+    HRESULT STDMETHODCALLTYPE Activate(REFIID iid, DWORD dwClsCtx, PROPVARIANT *pActivationParams, void **ppInterface) override;
+    HRESULT STDMETHODCALLTYPE OpenPropertyStore(DWORD stgmAccess, IPropertyStore **ppProperties) override;
+    HRESULT STDMETHODCALLTYPE GetId(LPWSTR *ppstrId) override;
+    HRESULT STDMETHODCALLTYPE GetState(DWORD *pdwState) override;
+#pragma endregion
+
+private:
+    std::atomic<ULONG> ref_cnt = 1;
+};

src/spice2x/hooks/audio/backends/mmdevice/null_discard_backend.cpp

@@ -0,0 +1,141 @@
+#include "null_discard_backend.h"
+
+#include <algorithm>
+#include <chrono>
+
+#include "hooks/audio/util.h"
+#include "util/logging.h"
+#include "util/precise_timer.h"
+
+NullDiscardBackend::~NullDiscardBackend() {
+    this->running = false;
+    if (this->pacing_thread.joinable()) {
+        this->pacing_thread.join();
+    }
+}
+
+const WAVEFORMATEXTENSIBLE &NullDiscardBackend::format() const noexcept {
+    return this->format_;
+}
+
+HRESULT NullDiscardBackend::on_initialize(
+    AUDCLNT_SHAREMODE *,
+    DWORD *,
+    REFERENCE_TIME *hnsBufferDuration,
+    REFERENCE_TIME *,
+    const WAVEFORMATEX *pFormat,
+    LPCGUID)
+{
+    copy_wave_format(&this->format_, pFormat);
+
+    // honor the game's requested buffer duration, falling back to 10 ms
+    constexpr REFERENCE_TIME DEFAULT_REFTIME = 100000; // 10 ms in 100-ns units
+    this->period_reftime = (hnsBufferDuration && *hnsBufferDuration > 0)
+            ? *hnsBufferDuration
+            : DEFAULT_REFTIME;
+
+    this->buffer_frames = std::max<uint32_t>(1, static_cast<uint32_t>(
+            static_cast<double>(this->format_.Format.nSamplesPerSec)
+                    * this->period_reftime / 10000000.0 + 0.5));
+
+    log_info("audio::null", "initializing null render device with {} channels, {} Hz, {}-bit",
+             this->format_.Format.nChannels,
+             this->format_.Format.nSamplesPerSec,
+             this->format_.Format.wBitsPerSample);
+
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_get_buffer_size(uint32_t *buffer_frames) {
+    *buffer_frames = this->buffer_frames;
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_get_stream_latency(REFERENCE_TIME *latency) {
+    *latency = this->period_reftime;
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_get_current_padding(std::optional<uint32_t> &padding_frames) {
+    // discarded immediately, so the buffer always reads as fully drained
+    padding_frames = 0;
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_is_format_supported(
+    AUDCLNT_SHAREMODE *,
+    const WAVEFORMATEX *,
+    WAVEFORMATEX **ppClosestMatch)
+{
+    if (ppClosestMatch) {
+        *ppClosestMatch = nullptr;
+    }
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_get_mix_format(WAVEFORMATEX **) {
+    return E_NOTIMPL;
+}
+
+HRESULT NullDiscardBackend::on_get_device_period(
+    REFERENCE_TIME *default_device_period,
+    REFERENCE_TIME *minimum_device_period)
+{
+    if (default_device_period) {
+        *default_device_period = this->period_reftime;
+    }
+    if (minimum_device_period) {
+        *minimum_device_period = this->period_reftime;
+    }
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_start() {
+    if (!this->running.exchange(true)) {
+        this->pacing_thread = std::thread(&NullDiscardBackend::pace_loop, this);
+    }
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_stop() {
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_set_event_handle(HANDLE *event_handle) {
+    // keep the game's event so pace_loop() can wake it; there is no real device behind it
+    this->relay_handle = *event_handle;
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_get_buffer(uint32_t num_frames_requested, BYTE **ppData) {
+    const size_t buffer_size =
+            static_cast<size_t>(this->format_.Format.nBlockAlign) * num_frames_requested;
+    if (this->scratch.size() < buffer_size) {
+        this->scratch.resize(buffer_size);
+    }
+    *ppData = this->scratch.data();
+    return S_OK;
+}
+
+HRESULT NullDiscardBackend::on_release_buffer(uint32_t, DWORD) {
+    // discard the audio entirely
+    return S_OK;
+}
+
+void NullDiscardBackend::pace_loop() {
+    using namespace std::chrono;
+
+    timeutils::PreciseSleepTimer timer;
+
+    // audio is discarded, so timing precision and drift do not matter; just wake the
+    // game once per buffer period to keep its render thread from blocking on the event.
+    const auto period = duration_cast<steady_clock::duration>(
+            duration<double>(this->period_reftime / 10000000.0));
+
+    while (this->running.load()) {
+        if (this->relay_handle) {
+            SetEvent(this->relay_handle);
+        }
+        timer.sleep(period);
+    }
+}

src/spice2x/hooks/audio/backends/mmdevice/null_discard_backend.h

@@ -0,0 +1,54 @@
+#pragma once
+
+#include <atomic>
+#include <optional>
+#include <thread>
+#include <vector>
+
+#include <audioclient.h>
+
+#include "hooks/audio/implementations/backend.h"
+
+// discards all audio while pacing the game's event handle once per buffer period, so the game
+// keeps running normally with nothing output to any real device. routed through the shared
+// DummyIAudioClient, the same plumbing the asio backend uses.
+struct NullDiscardBackend final : AudioBackend {
+    ~NullDiscardBackend() final;
+
+    const WAVEFORMATEXTENSIBLE &format() const noexcept override;
+
+    HRESULT on_initialize(
+        AUDCLNT_SHAREMODE *,
+        DWORD *,
+        REFERENCE_TIME *hnsBufferDuration,
+        REFERENCE_TIME *,
+        const WAVEFORMATEX *pFormat,
+        LPCGUID) override;
+    HRESULT on_get_buffer_size(uint32_t *buffer_frames) override;
+    HRESULT on_get_stream_latency(REFERENCE_TIME *latency) override;
+    HRESULT on_get_current_padding(std::optional<uint32_t> &padding_frames) override;
+    HRESULT on_is_format_supported(
+        AUDCLNT_SHAREMODE *,
+        const WAVEFORMATEX *,
+        WAVEFORMATEX **ppClosestMatch) override;
+    HRESULT on_get_mix_format(WAVEFORMATEX **) override;
+    HRESULT on_get_device_period(
+        REFERENCE_TIME *default_device_period,
+        REFERENCE_TIME *minimum_device_period) override;
+    HRESULT on_start() override;
+    HRESULT on_stop() override;
+    HRESULT on_set_event_handle(HANDLE *event_handle) override;
+    HRESULT on_get_buffer(uint32_t num_frames_requested, BYTE **ppData) override;
+    HRESULT on_release_buffer(uint32_t, DWORD) override;
+
+private:
+    void pace_loop();
+
+    WAVEFORMATEXTENSIBLE format_ {};
+    uint32_t buffer_frames = 0;
+    REFERENCE_TIME period_reftime = 0;
+    HANDLE relay_handle = nullptr;
+    std::vector<BYTE> scratch;
+    std::thread pacing_thread;
+    std::atomic<bool> running = false;
+};

src/spice2x/hooks/audio/backends/wasapi/audio_client.cpp

@@ -181,13 +181,22 @@ HRESULT STDMETHODCALLTYPE WrappedIAudioClient::Initialize(
         games::gitadora::fix_audio_channel_mask(const_cast<WAVEFORMATEX *>(pFormat));
     }
 
+    // when resampling, open the real device at the target rate while the game keeps writing its
+    // native-rate audio into the scratch buffer. this runs on whatever device_format is now: the
+    // game's native format, or the stereo format produced above when downmix is also active, so
+    // the two stages chain as multi-channel -> stereo -> resampled stereo.
+    WAVEFORMATEXTENSIBLE resample_storage = {};
+    if (auto target_rate = hooks::audio::Resampler::resolve(device_format)) {
+        const uint32_t src_rate = device_format->nSamplesPerSec;
+        this->resample.setup(device_format, &resample_storage, *target_rate);
+        device_format = reinterpret_cast<const WAVEFORMATEX *>(&resample_storage);
+        log_info("audio::wasapi", "resample enabled: {} Hz -> {} Hz{}",
+                src_rate, *target_rate, this->downmix.enabled ? " (after downmix)" : "");
+    }
+
     // verbose output
     log_info("audio::wasapi", "IAudioClient::Initialize hook hit");
-    log_info("audio::wasapi", "... ShareMode         : {}", share_mode_str(ShareMode));
-    log_info("audio::wasapi", "... StreamFlags       : {}", stream_flags_str(StreamFlags));
-    log_info("audio::wasapi", "... hnsBufferDuration : {}", hnsBufferDuration);
-    log_info("audio::wasapi", "... hnsPeriodicity    : {}", hnsPeriodicity);
-    print_format(device_format);
+    print_format(ShareMode, StreamFlags, hnsBufferDuration, hnsPeriodicity, device_format);
 
     if (this->backend) {
         SAFE_CALL("AudioBackend", "on_initialize", this->backend->on_initialize(
@@ -199,22 +208,48 @@ HRESULT STDMETHODCALLTYPE WrappedIAudioClient::Initialize(
                 AudioSessionGuid));
 
         log_info("audio::wasapi", "AudioBackend::on_initialize call finished");
-        log_info("audio::wasapi", "... ShareMode         : {}", share_mode_str(ShareMode));
-        log_info("audio::wasapi", "... StreamFlags       : {}", stream_flags_str(StreamFlags));
-        log_info("audio::wasapi", "... hnsBufferDuration : {}", hnsBufferDuration);
-        log_info("audio::wasapi", "... hnsPeriodicity    : {}", hnsPeriodicity);
-        print_format(pFormat);
+        print_format(ShareMode, StreamFlags, hnsBufferDuration, hnsPeriodicity, pFormat);
     }
 
     // check for exclusive mode
     if (ShareMode == AUDCLNT_SHAREMODE_EXCLUSIVE) {
         this->exclusive_mode = true;
         this->frame_size = device_format->nChannels * (device_format->wBitsPerSample / 8);
+
+        // optionally enlarge the exclusive buffer. games request a very small buffer (e.g. 3 ms)
+        // which some endpoints (notably NVIDIA HDMI/DP display audio) cannot service in time,
+        // underrunning mid-period and crackling. a larger buffer gives the device slack. exclusive
+        // mode requires periodicity == buffer_duration, so raise both together; the initialize
+        // paths below handle any required buffer-size realignment.
+        if (hooks::audio::EXCLUSIVE_BUFFER_MS.has_value()) {
+            const REFERENCE_TIME min_duration =
+                    (REFERENCE_TIME) hooks::audio::EXCLUSIVE_BUFFER_MS.value() * 10000;
+            if (hnsBufferDuration < min_duration) {
+                log_info("audio::wasapi",
+                        "raising exclusive buffer from {} hns to {} hns ({} ms)",
+                        hnsBufferDuration, min_duration, hooks::audio::EXCLUSIVE_BUFFER_MS.value());
+                hnsBufferDuration = min_duration;
+                if (hnsPeriodicity != 0) {
+                    hnsPeriodicity = min_duration;
+                }
+            }
+        }
     }
 
-    // call next
+    // call next. the resampler owns the device interaction whenever it is active (including when
+    // chained after the downmix), otherwise the downmix does, otherwise the device is opened
+    // directly.
     HRESULT ret;
-    if (this->downmix.enabled) {
+    if (this->resample.enabled) {
+        ret = this->resample.initialize(
+                pReal,
+                ShareMode,
+                StreamFlags,
+                hnsBufferDuration,
+                hnsPeriodicity,
+                device_format,
+                AudioSessionGuid);
+    } else if (this->downmix.enabled) {
         ret = this->downmix.initialize(
                 pReal,
                 ShareMode,
@@ -224,7 +259,9 @@ HRESULT STDMETHODCALLTYPE WrappedIAudioClient::Initialize(
                 device_format,
                 AudioSessionGuid);
     } else {
-        ret = pReal->Initialize(
+        ret = initialize_with_alignment_retry(
+                pReal,
+                "audio::wasapi",
                 ShareMode,
                 StreamFlags,
                 hnsBufferDuration,
@@ -263,7 +300,15 @@ HRESULT STDMETHODCALLTYPE WrappedIAudioClient::GetBufferSize(UINT32 *pNumBufferF
         }
     }
 
-    CHECK_RESULT(pReal->GetBufferSize(pNumBufferFrames));
+    HRESULT ret = pReal->GetBufferSize(pNumBufferFrames);
+
+    // report the buffer size at the game's native rate; the real device buffer is at the
+    // resampled rate, so translate it back so the game paces its writes correctly.
+    if (SUCCEEDED(ret) && this->resample.enabled && pNumBufferFrames) {
+        *pNumBufferFrames = this->resample.frames_device_to_game(*pNumBufferFrames);
+    }
+
+    CHECK_RESULT(ret);
 }
 HRESULT STDMETHODCALLTYPE WrappedIAudioClient::GetStreamLatency(REFERENCE_TIME *phnsLatency) {
     static std::once_flag printed;
@@ -305,7 +350,15 @@ HRESULT STDMETHODCALLTYPE WrappedIAudioClient::GetCurrentPadding(UINT32 *pNumPad
         }
     }
 
-    CHECK_RESULT(pReal->GetCurrentPadding(pNumPaddingFrames));
+    HRESULT ret = pReal->GetCurrentPadding(pNumPaddingFrames);
+
+    // the device buffer is at the resampled rate; report padding at the game's native rate so the
+    // game's free-space calculation stays paced correctly.
+    if (SUCCEEDED(ret) && this->resample.enabled && pNumPaddingFrames) {
+        *pNumPaddingFrames = this->resample.padding_device_to_game(*pNumPaddingFrames);
+    }
+
+    CHECK_RESULT(ret);
 }
 HRESULT STDMETHODCALLTYPE WrappedIAudioClient::IsFormatSupported(
     AUDCLNT_SHAREMODE ShareMode,
@@ -323,16 +376,42 @@ HRESULT STDMETHODCALLTYPE WrappedIAudioClient::IsFormatSupported(
         fix_rec_format(const_cast<WAVEFORMATEX *>(pFormat));
     }
 
+    // log the format the game is asking about
+    log_info("audio::wasapi", "IAudioClient::IsFormatSupported hook hit");
+    print_format(ShareMode, pFormat);
+
     // when downmixing, the real device is opened as stereo, so check whether the equivalent
-    // stereo format is supported instead of the multi-channel one.
+    // stereo format is supported instead of the multi-channel one. when resampling is also active
+    // it chains onto that stereo format, so check the resampled stereo format.
     if (resolve_downmix(pFormat)) {
         WAVEFORMATEXTENSIBLE stereo_storage = {};
         hooks::audio::Downmix::make_stereo_format(pFormat, &stereo_storage);
-        const auto stereo_format = reinterpret_cast<const WAVEFORMATEX *>(&stereo_storage);
+        const WAVEFORMATEX *check_format = reinterpret_cast<const WAVEFORMATEX *>(&stereo_storage);
 
-        CHECK_RESULT(pReal->IsFormatSupported(ShareMode, stereo_format, ppClosestMatch));
+        WAVEFORMATEXTENSIBLE resample_storage = {};
+        if (auto target_rate = hooks::audio::Resampler::resolve(check_format)) {
+            hooks::audio::Resampler::make_device_format(check_format, &resample_storage, *target_rate);
+            check_format = reinterpret_cast<const WAVEFORMATEX *>(&resample_storage);
+        }
+
+        log_info("audio::wasapi", "... checking device format instead (after downmix/resample):");
+        print_format(check_format);
+
+        CHECK_RESULT(pReal->IsFormatSupported(ShareMode, check_format, ppClosestMatch));
     } else if (games::gitadora::is_arena_model()) {
         games::gitadora::fix_audio_channel_mask(const_cast<WAVEFORMATEX *>(pFormat));
+    } else if (auto target_rate = hooks::audio::Resampler::resolve(pFormat)) {
+
+        // when resampling, the real device is opened at the target rate, so check whether the
+        // equivalent format at that rate is supported instead of the game's native rate.
+        WAVEFORMATEXTENSIBLE resample_storage = {};
+        hooks::audio::Resampler::make_device_format(pFormat, &resample_storage, *target_rate);
+        const auto resample_format = reinterpret_cast<const WAVEFORMATEX *>(&resample_storage);
+
+        log_info("audio::wasapi", "... checking device format instead (after resample):");
+        print_format(resample_format);
+
+        CHECK_RESULT(pReal->IsFormatSupported(ShareMode, resample_format, ppClosestMatch));
     }
 
     if (this->backend) {

src/spice2x/hooks/audio/backends/wasapi/audio_client.h

@@ -9,9 +9,9 @@
 #include "util/logging.h"
 
 #include "downmix.h"
+#include "resample.h"
 
 #include "audio_render_client.h"
-
 // {1FBC8530-AF3E-4128-B418-115DE72F76B6}
 static const GUID IID_WrappedIAudioClient = {
     0x1fbc8530, 0xaf3e, 0x4128, { 0xb4, 0x18, 0x11, 0x5d, 0xe7, 0x2f, 0x76, 0xb6 }
@@ -103,4 +103,9 @@ struct WrappedIAudioClient : IAudioClient3 {
     // surround -> stereo downmix. the real device is opened as stereo while the game keeps
     // writing multi-channel audio into a scratch buffer that we downmix in the render client.
     hooks::audio::Downmix downmix;
+
+    // native-rate -> target-rate sample-rate conversion. the real device is opened at the target
+    // rate while the game keeps writing its native-rate audio into a scratch buffer that we
+    // resample in the render client.
+    hooks::audio::Resampler resample;
 };

src/spice2x/hooks/audio/backends/wasapi/audio_render_client.cpp

@@ -7,6 +7,7 @@
 
 #include "audio_client.h"
 #include "hooks/audio/audio.h"
+#include "util.h"
 #include "wasapi_private.h"
 
 const char CLASS_NAME[] = "WrappedIAudioRenderClient";
@@ -17,9 +18,7 @@ static void apply_gain(BYTE *buffer, UINT32 frames, const WAVEFORMATEXTENSIBLE &
     const WAVEFORMATEX &f = fmt.Format;
     const size_t samples = (size_t) frames * f.nChannels;
 
-    // KSDATAFORMAT_SUBTYPE_IEEE_FLOAT has Data1 == 3, _PCM has Data1 == 1
-    bool is_float = f.wFormatTag == WAVE_FORMAT_IEEE_FLOAT
-        || (f.wFormatTag == WAVE_FORMAT_EXTENSIBLE && fmt.SubFormat.Data1 == 0x00000003);
+    bool is_float = is_ieee_float(&f);
 
     if (is_float && f.wBitsPerSample == 32) {
         auto p = reinterpret_cast<float *>(buffer);
@@ -113,10 +112,23 @@ HRESULT STDMETHODCALLTYPE WrappedIAudioRenderClient::GetBuffer(UINT32 NumFramesR
         return S_OK;
     }
 
+    // downmix + resample chained: the game writes its multi-channel native-rate audio into the
+    // downmix scratch, which is downmixed to stereo and then resampled on release. size the
+    // resampler's (stereo) input scratch now and hand the game the multi-channel downmix scratch.
+    if (this->client->downmix.enabled && this->client->resample.enabled) {
+        BYTE *resample_scratch = nullptr;
+        this->client->resample.get_buffer(NumFramesRequested, &resample_scratch);
+        CHECK_RESULT(this->client->downmix.get_scratch(NumFramesRequested, ppData));
+
     // surround downmix: reserve the real (stereo) device buffer, but hand the game a
     // multi-channel scratch buffer that we downmix on release
-    if (this->client->downmix.enabled) {
+    } else if (this->client->downmix.enabled) {
         CHECK_RESULT(this->client->downmix.get_buffer(pReal, NumFramesRequested, ppData));
+
+    // resample: hand the game a native-rate scratch buffer that we convert on release. the real
+    // device buffer is acquired in ReleaseBuffer once the converted frame count is known.
+    } else if (this->client->resample.enabled) {
+        CHECK_RESULT(this->client->resample.get_buffer(NumFramesRequested, ppData));
     }
 
     // call original
@@ -143,6 +155,34 @@ HRESULT STDMETHODCALLTYPE WrappedIAudioRenderClient::ReleaseBuffer(UINT32 NumFra
         return S_OK;
     }
 
+    // downmix + resample chained: downmix the game's multi-channel scratch into the resampler's
+    // stereo input scratch, then let the resampler convert and push it to the device. a silent
+    // buffer skips the downmix and feeds silence straight through.
+    if (this->client->downmix.enabled && this->client->resample.enabled) {
+        if ((dwFlags & AUDCLNT_BUFFERFLAGS_SILENT) == 0) {
+            this->client->downmix.downmix_into(
+                    this->client->resample.input_data(), NumFramesWritten);
+        }
+        return this->client->resample.flush(
+                pReal,
+                this->client->pReal,
+                NumFramesWritten,
+                dwFlags,
+                hooks::audio::VOLUME_BOOST);
+    }
+
+    // resample: convert the game's native-rate scratch and push as many output frames as the
+    // device has room for, applying the volume boost to the converted output. handles acquiring
+    // and releasing the real device buffer itself.
+    if (this->client->resample.enabled) {
+        return this->client->resample.flush(
+                pReal,
+                this->client->pReal,
+                NumFramesWritten,
+                dwFlags,
+                hooks::audio::VOLUME_BOOST);
+    }
+
     // resolve the real device buffer for whichever path produced the audio
     BYTE *device_buffer;
     if (this->client->downmix.enabled) {

src/spice2x/hooks/audio/backends/wasapi/downmix.cpp

@@ -11,6 +11,8 @@
 
 #include "util/logging.h"
 
+#include "util.h"
+
 namespace hooks::audio {
 
     namespace {
@@ -45,69 +47,6 @@ namespace hooks::audio {
         }
     }
 
-    // read one sample at `p` as a normalized float in [-1, 1]
-    static inline float read_sample(const BYTE *p, int bytes, bool is_float) {
-        if (is_float) {
-            float v;
-            memcpy(&v, p, sizeof(float));
-            return v;
-        }
-        switch (bytes) {
-            case 2: {
-                int16_t v;
-                memcpy(&v, p, sizeof(v));
-                return v * (1.0f / 32768.0f);
-            }
-            case 3: {
-                int32_t v = p[0] | (p[1] << 8) | (p[2] << 16);
-                if (v & 0x800000) {
-                    v |= ~0xFFFFFF; // sign extend
-                }
-                return v * (1.0f / 8388608.0f);
-            }
-            case 4: {
-                int32_t v;
-                memcpy(&v, p, sizeof(v));
-                return (float) (v * (1.0 / 2147483648.0));
-            }
-            default:
-                return 0.0f;
-        }
-    }
-
-    // write the normalized float `value` to the sample at `p`, clamping to the format's range
-    static inline void write_sample(BYTE *p, int bytes, bool is_float, float value) {
-        if (is_float) {
-            float v = std::clamp(value, -1.0f, 1.0f);
-            memcpy(p, &v, sizeof(v));
-            return;
-        }
-        switch (bytes) {
-            case 2: {
-                int16_t v = (int16_t) std::clamp(
-                    (int) std::lround(value * 32768.0f), -32768, 32767);
-                memcpy(p, &v, sizeof(v));
-                break;
-            }
-            case 3: {
-                int32_t v = (int32_t) std::clamp<int64_t>(
-                    std::llround((double) value * 8388608.0), -8388608, 8388607);
-                p[0] = v & 0xFF;
-                p[1] = (v >> 8) & 0xFF;
-                p[2] = (v >> 16) & 0xFF;
-                break;
-            }
-            case 4: {
-                int32_t v = (int32_t) std::clamp<int64_t>(
-                    std::llround((double) value * 2147483648.0), INT32_MIN, INT32_MAX);
-                memcpy(p, &v, sizeof(v));
-                break;
-            }
-            default:
-                break;
-        }
-    }
-
     void Downmix::setup(const WAVEFORMATEX *game_format, WAVEFORMATEXTENSIBLE *stereo_out,
             DownmixAlgorithm algorithm) {
         this->enabled = true;
@@ -115,11 +54,7 @@ namespace hooks::audio {
         this->bytes_per_sample = game_format->wBitsPerSample / 8;
         this->game_frame_size = game_format->nChannels * this->bytes_per_sample;
 
-        // KSDATAFORMAT_SUBTYPE_IEEE_FLOAT has Data1 == 3 (matches apply_gain detection)
-        this->is_float = game_format->wFormatTag == WAVE_FORMAT_IEEE_FLOAT
-            || (game_format->wFormatTag == WAVE_FORMAT_EXTENSIBLE
-                && reinterpret_cast<const WAVEFORMATEXTENSIBLE *>(game_format)
-                    ->SubFormat.Data1 == 0x00000003);
+        this->is_float = is_ieee_float(game_format);
 
         // supported: 16/24/32-bit integer PCM and 32-bit float; anything else mixes to silence
         const bool supported = this->is_float
@@ -155,28 +90,10 @@ namespace hooks::audio {
             REFERENCE_TIME buffer_duration, REFERENCE_TIME periodicity,
             const WAVEFORMATEX *device_format, LPCGUID session_guid) {
 
-        HRESULT ret = real->Initialize(share_mode, stream_flags, buffer_duration, periodicity,
-                device_format, session_guid);
-
         // the smaller stereo buffer can end up unaligned for the device when the game sized the
-        // duration for its larger multi-channel format. recover by asking the device for the next
-        // aligned buffer size and re-initializing with a matching duration.
-        if (ret == AUDCLNT_E_BUFFER_SIZE_NOT_ALIGNED) {
-            UINT32 aligned_frames = 0;
-            if (SUCCEEDED(real->GetBufferSize(&aligned_frames)) && aligned_frames > 0) {
-                REFERENCE_TIME aligned_duration = (REFERENCE_TIME)
-                        (10000.0 * 1000 / device_format->nSamplesPerSec * aligned_frames + 0.5);
-
-                log_info("audio::downmix",
-                        "buffer not aligned, retrying with {} frames ({} hns)",
-                        aligned_frames, aligned_duration);
-
-                ret = real->Initialize(share_mode, stream_flags, aligned_duration,
-                        periodicity != 0 ? aligned_duration : 0, device_format, session_guid);
-            }
-        }
-
-        return ret;
+        // duration for its larger multi-channel format; the helper recovers from that.
+        return initialize_with_alignment_retry(real, "audio::downmix", share_mode, stream_flags,
+                buffer_duration, periodicity, device_format, session_guid);
     }
 
     void Downmix::add_channel(int channel, DWORD speaker, float gain) {
@@ -313,6 +230,21 @@ namespace hooks::audio {
         return S_OK;
     }
 
+    HRESULT Downmix::get_scratch(UINT32 frames, BYTE **ppData) {
+        const size_t needed = (size_t) frames * this->game_frame_size;
+        if (this->scratch.size() < needed) {
+            this->scratch.resize(needed);
+        }
+
+        *ppData = this->scratch.data();
+
+        return S_OK;
+    }
+
+    void Downmix::downmix_into(BYTE *dst, UINT32 frames) const {
+        this->process(dst, this->scratch.data(), frames);
+    }
+
     void Downmix::write_device_buffer(UINT32 frames, DWORD flags) {
         const int bps = this->bytes_per_sample;
         const int dst_stride = 2 * bps;

src/spice2x/hooks/audio/backends/wasapi/downmix.h

@@ -100,6 +100,14 @@ namespace hooks::audio {
         // grab the real stereo device buffer and hand the game the scratch buffer to write into.
         HRESULT get_buffer(IAudioRenderClient *real, UINT32 frames, BYTE **ppData);
 
+        // size the scratch and hand it to the game without acquiring a device buffer. used when a
+        // later stage (the resampler) owns the device interaction.
+        HRESULT get_scratch(UINT32 frames, BYTE **ppData);
+
+        // downmix the scratch the game wrote into the caller's stereo buffer, without touching the
+        // device. used to feed the resampler when the two stages are chained.
+        void downmix_into(BYTE *dst, UINT32 frames) const;
+
         // mix the scratch buffer into the stereo device buffer held since get_buffer. the caller
         // owns releasing the device buffer afterwards (see current_buffer / buffer_released).
         void write_device_buffer(UINT32 frames, DWORD flags);

src/spice2x/hooks/audio/backends/wasapi/dummy_audio_client.cpp

@@ -73,11 +73,7 @@ HRESULT STDMETHODCALLTYPE DummyIAudioClient::Initialize(
 
     // verbose output
     log_info("audio::wasapi", "IAudioClient::Initialize hook hit");
-    log_info("audio::wasapi", "... ShareMode         : {}", share_mode_str(ShareMode));
-    log_info("audio::wasapi", "... StreamFlags       : {}", stream_flags_str(StreamFlags));
-    log_info("audio::wasapi", "... hnsBufferDuration : {}", hnsBufferDuration);
-    log_info("audio::wasapi", "... hnsPeriodicity    : {}", hnsPeriodicity);
-    print_format(pFormat);
+    print_format(ShareMode, StreamFlags, hnsBufferDuration, hnsPeriodicity, pFormat);
 
     CHECK_RESULT(this->backend->on_initialize(
         &ShareMode,

src/spice2x/hooks/audio/backends/wasapi/resample.cpp

@@ -0,0 +1,437 @@
+#include "resample.h"
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
+#include <mutex>
+
+#include <audioclient.h>
+
+#include "util/logging.h"
+
+#include "util.h"
+
+namespace hooks::audio {
+
+    namespace {
+
+        constexpr double PI = 3.14159265358979323846;
+
+        // normalized sinc: sin(pi*x) / (pi*x), with the removable singularity at 0 filled in
+        inline double sinc(double x) {
+            if (x == 0.0) {
+                return 1.0;
+            }
+            const double px = PI * x;
+            return std::sin(px) / px;
+        }
+
+        // Blackman window across the kernel radius; zero at +/- radius
+        inline double blackman(double x, double radius) {
+            const double n = (x + radius) / (2.0 * radius);
+            if (n <= 0.0 || n >= 1.0) {
+                return 0.0;
+            }
+            return 0.42 - 0.5 * std::cos(2.0 * PI * n) + 0.08 * std::cos(4.0 * PI * n);
+        }
+    }
+
+    std::optional<uint32_t> Resampler::resolve(const WAVEFORMATEX *game_format) {
+        if (game_format == nullptr || !RESAMPLE_RATE.has_value()) {
+            return std::nullopt;
+        }
+        if (game_format->nSamplesPerSec == 0
+                || game_format->nSamplesPerSec == RESAMPLE_RATE.value()) {
+            return std::nullopt;
+        }
+        return RESAMPLE_RATE;
+    }
+
+    void Resampler::setup(const WAVEFORMATEX *game_format, WAVEFORMATEXTENSIBLE *device_out,
+            uint32_t target_rate) {
+        this->enabled = true;
+        this->channels = game_format->nChannels;
+        this->bytes_per_sample = game_format->wBitsPerSample / 8;
+        this->game_frame_size = this->channels * this->bytes_per_sample;
+
+        this->is_float = is_ieee_float(game_format);
+
+        const bool supported = this->is_float
+            ? this->bytes_per_sample == 4
+            : (this->bytes_per_sample >= 2 && this->bytes_per_sample <= 4);
+        if (!supported) {
+            log_fatal(
+                "audio::resample",
+                "unsupported sample format ({}-bit {}) for -resample",
+                game_format->wBitsPerSample, this->is_float ? "float" : "int");
+        }
+
+        this->src_rate = game_format->nSamplesPerSec;
+        this->dst_rate = target_rate;
+
+        // anti-alias cutoff: full bandwidth when upsampling, scaled down when decimating
+        this->cutoff = std::min(1.0, (double) this->dst_rate / (double) this->src_rate);
+        this->half_taps = 16;
+
+        // precompute the windowed-sinc kernel now that cutoff is known
+        this->build_kernel();
+
+        // prime the queue with half a window of silence so the first outputs have left history
+        this->in_queue.assign((size_t) this->half_taps * this->channels, 0.0f);
+        this->in_pos = this->half_taps;
+
+        this->make_device_format(game_format, device_out, target_rate);
+    }
+
+    void Resampler::make_device_format(const WAVEFORMATEX *game_format,
+            WAVEFORMATEXTENSIBLE *device_out, uint32_t target_rate) {
+        const size_t src_size = sizeof(WAVEFORMATEX) + game_format->cbSize;
+
+        memset(device_out, 0, sizeof(WAVEFORMATEXTENSIBLE));
+        memcpy(device_out, game_format, std::min(src_size, sizeof(WAVEFORMATEXTENSIBLE)));
+
+        device_out->Format.nSamplesPerSec = target_rate;
+        device_out->Format.nAvgBytesPerSec = target_rate * device_out->Format.nBlockAlign;
+    }
+
+    HRESULT Resampler::initialize(IAudioClient *real, AUDCLNT_SHAREMODE share_mode,
+            DWORD stream_flags, REFERENCE_TIME buffer_duration, REFERENCE_TIME periodicity,
+            const WAVEFORMATEX *device_format, LPCGUID session_guid) {
+
+        // the resampler bypasses the OS mixer and talks to the device directly, so it only makes
+        // sense (and only works) for exclusive streams. shared streams are already resampled by
+        // the Windows audio engine, so refuse loudly rather than silently doing nothing.
+        if (share_mode != AUDCLNT_SHAREMODE_EXCLUSIVE) {
+            log_fatal("audio::resample",
+                    "-resample requires WASAPI exclusive mode, but this stream is shared "
+                    "(Windows already resamples shared streams)");
+        }
+
+        // record the pacing model. event-driven streams fill the whole device buffer each period
+        // (produce_exact); timer-driven streams poll padding and write variable partial chunks, so
+        // they drain the pending output to the device's free space each call (flush_timer).
+        this->event_driven = (stream_flags & AUDCLNT_STREAMFLAGS_EVENTCALLBACK) != 0;
+
+        return initialize_with_alignment_retry(real, "audio::resample", share_mode, stream_flags,
+                buffer_duration, periodicity, device_format, session_guid);
+    }
+
+    UINT32 Resampler::frames_device_to_game(UINT32 device_frames) const {
+        if (this->dst_rate == 0) {
+            return device_frames;
+        }
+        // round down so the game never believes it has more room than the device can hold
+        return (UINT32) (((double) device_frames * this->src_rate) / this->dst_rate);
+    }
+
+    UINT32 Resampler::padding_device_to_game(UINT32 device_padding) const {
+        if (this->dst_rate == 0) {
+            return device_padding;
+        }
+        // round up so the reported free space stays conservative
+        return (UINT32) std::ceil(((double) device_padding * this->src_rate) / this->dst_rate);
+    }
+
+    HRESULT Resampler::get_buffer(UINT32 frames, BYTE **ppData) {
+        const size_t needed = (size_t) frames * this->game_frame_size;
+        if (this->scratch.size() < needed) {
+            this->scratch.resize(needed);
+        }
+
+        *ppData = this->scratch.data();
+
+        return S_OK;
+    }
+
+    void Resampler::enqueue_input(UINT32 frames, bool silent) {
+        const int bps = this->bytes_per_sample;
+        const int ch = this->channels;
+        const size_t base = this->in_queue.size();
+
+        this->in_queue.resize(base + (size_t) frames * ch);
+
+        if (silent || bps <= 0 || ch <= 0) {
+            std::fill(this->in_queue.begin() + base, this->in_queue.end(), 0.0f);
+            return;
+        }
+
+        const BYTE *src = this->scratch.data();
+        for (UINT32 f = 0; f < frames; f++) {
+            for (int c = 0; c < ch; c++) {
+                const size_t s = (size_t) f * ch + c;
+                this->in_queue[base + s] = read_sample(src + s * bps, bps, this->is_float);
+            }
+        }
+    }
+
+    void Resampler::build_kernel() {
+        const int taps = 2 * this->half_taps;
+        const int phases = this->kernel_phases;
+        const double cut = this->cutoff;
+        const double radius = (double) this->half_taps;
+
+        // one extra row at frac == 1.0 so emit_frame can interpolate against row p + 1 safely
+        this->kernel_table.resize((size_t) (phases + 1) * taps);
+
+        for (int p = 0; p <= phases; p++) {
+            const double frac = (double) p / (double) phases;
+            for (int k = 0; k < taps; k++) {
+                // tap k maps to input offset t = k - (half_taps - 1), matching emit_frame
+                const double x = frac - (double) (k - (this->half_taps - 1));
+                this->kernel_table[(size_t) p * taps + k] =
+                        (float) (cut * sinc(cut * x) * blackman(x, radius));
+            }
+        }
+    }
+
+    void Resampler::emit_frame() {
+        const int ch = this->channels;
+        const int radius = this->half_taps;
+        const int taps = 2 * radius;
+        const long avail = (long) (this->in_queue.size() / ch);
+        const long center = (long) std::floor(this->in_pos);
+
+        // pick the two kernel rows bracketing this fractional position and the blend between them
+        const double frac = this->in_pos - (double) center;
+        const double fp = frac * (double) this->kernel_phases;
+        const int p0 = (int) fp;
+        const float blend = (float) (fp - (double) p0);
+        const float *row0 = &this->kernel_table[(size_t) p0 * taps];
+        const float *row1 = &this->kernel_table[(size_t) (p0 + 1) * taps];
+
+        // base input index for tap 0 (t = -(radius - 1))
+        const long base = center - (radius - 1);
+
+        for (int c = 0; c < ch; c++) {
+            double acc = 0.0;
+            for (int k = 0; k < taps; k++) {
+                const long idx = base + k;
+                if (idx < 0 || idx >= avail) {
+                    continue;
+                }
+                const float w = row0[k] + blend * (row1[k] - row0[k]);
+                acc += (double) this->in_queue[(size_t) idx * ch + c] * w;
+            }
+            this->out_float.push_back((float) acc);
+        }
+    }
+
+    void Resampler::drop_consumed() {
+        const int ch = this->channels;
+        const long drop = (long) std::floor(this->in_pos) - this->half_taps;
+        if (drop > 0) {
+            const size_t drop_samples = (size_t) drop * ch;
+            if (drop_samples <= this->in_queue.size()) {
+                this->in_queue.erase(this->in_queue.begin(),
+                        this->in_queue.begin() + drop_samples);
+                this->in_pos -= drop;
+            }
+        }
+    }
+
+    UINT32 Resampler::produce_exact(UINT32 out_frames) {
+        const int ch = this->channels;
+        this->out_float.clear();
+        if (ch <= 0 || out_frames == 0) {
+            return 0;
+        }
+        this->out_float.reserve((size_t) out_frames * ch);
+
+        // resample ratio. drive it from the buffer size actually advertised to the game rather
+        // than the nominal src/dst ratio: GetBufferSize reports floor(dev_buf * src/dst) game
+        // frames, so the game only ever delivers that many input frames per device period.
+        // consuming at the nominal ratio would eat slightly more input than arrives on any device
+        // where dev_buf * src/dst is non-integer (e.g. 144 -> 132.3, floored to 132), slowly
+        // draining the queue until it underruns to permanent silence. using the advertised integer
+        // ratio keeps input and output exactly balanced; the resulting pitch error is below 0.3%
+        // and inaudible, and it collapses to the exact ratio when the division is integer (160 ->
+        // 147 stays 147/160 = 44100/48000).
+        const double step = (double) this->frames_device_to_game(this->device_buffer_frames)
+                / (double) this->device_buffer_frames;
+
+        // input frames the block will touch: from in_pos through the right edge of the sinc kernel
+        // at the final output sample. if the queue is short of this, the kernel tail reads past the
+        // end and distorts every buffer, so buffer one extra block of input before the first output
+        // (emitting silence without consuming) to build a cushion the kernel can always reach into.
+        const long avail = (long) (this->in_queue.size() / ch);
+        const long need = (long) std::ceil(this->in_pos + step * (double) out_frames)
+                + this->half_taps;
+
+        if (this->priming) {
+            if (avail < need + (long) out_frames) {
+                this->out_float.assign((size_t) out_frames * ch, 0.0f);
+                return out_frames;
+            }
+            this->priming = false;
+        }
+
+        for (UINT32 o = 0; o < out_frames; o++) {
+            this->emit_frame();
+            this->in_pos += step;
+        }
+
+        this->drop_consumed();
+        return out_frames;
+    }
+
+    UINT32 Resampler::produce_variable() {
+        const int ch = this->channels;
+        if (ch <= 0) {
+            return 0;
+        }
+
+        // input frames consumed per output frame. timer-driven streams write variable partial
+        // chunks, so produce however many output frames the currently queued input can fully
+        // support and leave the rest for the next call; this keeps input and output balanced at
+        // the exact src/dst ratio over time without depending on the device buffer size.
+        const double step = (double) this->src_rate / (double) this->dst_rate;
+        const long avail = (long) (this->in_queue.size() / ch);
+
+        // emit only while the sinc kernel's right edge stays within the queued input. the kernel
+        // reaches from in_pos out to half_taps frames ahead, so stop once that would read past the
+        // end; the remaining input becomes the next block's lookahead.
+        UINT32 produced = 0;
+        while ((long) std::ceil(this->in_pos) + this->half_taps < avail) {
+            this->emit_frame();
+            this->in_pos += step;
+            produced++;
+        }
+
+        this->drop_consumed();
+        return produced;
+    }
+
+    void Resampler::write_output(BYTE *dst, UINT32 frames, float gain) const {
+        const int bps = this->bytes_per_sample;
+        const int ch = this->channels;
+        const size_t count = (size_t) frames * ch;
+
+        for (size_t i = 0; i < count; i++) {
+            write_sample(dst + i * bps, bps, this->is_float, this->out_float[i] * gain);
+        }
+    }
+
+    HRESULT Resampler::flush(IAudioRenderClient *real, IAudioClient *client, UINT32 frames,
+            DWORD flags, float boost) {
+        if (!this->enabled) {
+            return S_OK;
+        }
+
+        // cache the device buffer size once
+        if (this->device_buffer_frames == 0) {
+            client->GetBufferSize(&this->device_buffer_frames);
+        }
+        if (this->device_buffer_frames == 0) {
+            return S_OK;
+        }
+
+        const bool silent = (flags & AUDCLNT_BUFFERFLAGS_SILENT) != 0;
+        this->enqueue_input(frames, silent);
+
+        // confirm once that conversion actually started producing output
+        static std::once_flag active_printed;
+        std::call_once(active_printed, [this]() {
+            log_info("audio::resample", "resample active: {} Hz -> {} Hz ({} ch, {})",
+                    this->src_rate, this->dst_rate, this->channels,
+                    this->event_driven ? "event-driven" : "timer-driven");
+        });
+
+        // the boost is applied here (inside write_output) rather than in the standard ReleaseBuffer
+        // path, so log it once for parity with that path's "volume boost active" line.
+        if (boost != 1.0f) {
+            static std::once_flag boost_printed;
+            std::call_once(boost_printed, [boost]() {
+                log_info("audio::resample", "volume boost active (resample): gain={}", boost);
+            });
+        }
+
+        return this->event_driven
+                ? this->flush_event(real, boost)
+                : this->flush_timer(real, client, boost);
+    }
+
+    HRESULT Resampler::flush_event(IAudioRenderClient *real, float boost) {
+
+        // event-driven exclusive streams must hand the device a full buffer every period and may
+        // not push partial counts. resample the whole input block into exactly the device buffer
+        // size.
+        const UINT32 produced = this->produce_exact(this->device_buffer_frames);
+        if (produced == 0) {
+            return S_OK;
+        }
+
+        BYTE *dev = nullptr;
+        HRESULT ret = real->GetBuffer(produced, &dev);
+        if (FAILED(ret) || dev == nullptr) {
+            return ret;
+        }
+
+        // mute the first few buffers to avoid a pop on stream start
+        float gain = boost;
+        if (this->buffers_to_mute > 0) {
+            gain = 0.0f;
+            this->buffers_to_mute--;
+        }
+
+        this->write_output(dev, produced, gain);
+
+        return real->ReleaseBuffer(produced, 0);
+    }
+
+    HRESULT Resampler::flush_timer(IAudioRenderClient *real, IAudioClient *client, float boost) {
+
+        // convert everything currently queued into the pending output FIFO (out_float). timer-
+        // driven games write variable partial chunks, so produce only what the queued input can
+        // fully support and keep the remainder for the next call.
+        this->produce_variable();
+
+        const int ch = this->channels;
+        if (ch <= 0) {
+            return S_OK;
+        }
+
+        const UINT32 pending = (UINT32) (this->out_float.size() / ch);
+        if (pending == 0) {
+            return S_OK;
+        }
+
+        // push as many frames as the device currently has free, keeping the rest queued for the
+        // next call. timer-driven games poll padding and write whenever there is room, so matching
+        // the device's free space here avoids overflowing the ring while staying device-paced.
+        UINT32 padding = 0;
+        if (FAILED(client->GetCurrentPadding(&padding))) {
+            return S_OK;
+        }
+        const UINT32 device_free = this->device_buffer_frames > padding
+                ? this->device_buffer_frames - padding
+                : 0;
+        if (device_free == 0) {
+            return S_OK;
+        }
+
+        const UINT32 to_write = std::min(pending, device_free);
+
+        BYTE *dev = nullptr;
+        HRESULT ret = real->GetBuffer(to_write, &dev);
+        if (FAILED(ret) || dev == nullptr) {
+            return ret;
+        }
+
+        // mute the first few buffers to avoid a pop on stream start
+        float gain = boost;
+        if (this->buffers_to_mute > 0) {
+            gain = 0.0f;
+            this->buffers_to_mute--;
+        }
+
+        this->write_output(dev, to_write, gain);
+        ret = real->ReleaseBuffer(to_write, 0);
+
+        // drop the frames just written from the front of the pending FIFO
+        this->out_float.erase(this->out_float.begin(),
+                this->out_float.begin() + (size_t) to_write * ch);
+
+        return ret;
+    }
+}

src/spice2x/hooks/audio/backends/wasapi/resample.h

@@ -0,0 +1,149 @@
+#pragma once
+
+#include <cstdint>
+#include <optional>
+#include <vector>
+
+#include <windows.h>
+#include <mmreg.h>
+#include <audioclient.h>
+
+#include "hooks/audio/audio.h"
+
+struct IAudioClient;
+struct IAudioRenderClient;
+
+namespace hooks::audio {
+
+    // Streaming sample-rate converter for the WASAPI render path. The real device is opened at the
+    // target rate while the game keeps writing its native-rate audio into a scratch buffer; on
+    // release that buffer is converted with a windowed-sinc kernel and pushed to the device.
+    // Channel count and sample format are preserved; only the sample rate changes.
+    //
+    // Frame counts differ between the two rates, so unlike the per-frame downmix this is stateful:
+    // a fractional read position and a window of input history carry across ReleaseBuffer calls,
+    // and the device buffer is only filled up to the space the device currently has free.
+    struct Resampler {
+
+        // whether the resampler is active for the current stream
+        bool enabled = false;
+
+        // whether the stream is event-driven (AUDCLNT_STREAMFLAGS_EVENTCALLBACK). timer-driven
+        // streams instead poll padding and write variable partial chunks, so they drain the
+        // pending output to the device's free space rather than pushing a full buffer per period.
+        bool event_driven = true;
+
+        // decide whether the stream should be resampled and to which rate. returns the target rate
+        // when RESAMPLE_RATE is set and differs from the game's rate, otherwise nullopt.
+        static std::optional<uint32_t> resolve(const WAVEFORMATEX *game_format);
+
+        // enable resampling for game_format and fill device_out with the equivalent format at the
+        // target rate to open the real device with.
+        void setup(const WAVEFORMATEX *game_format, WAVEFORMATEXTENSIBLE *device_out,
+                   uint32_t target_rate);
+
+        // build the device format equivalent to game_format at target_rate (same channels/depth).
+        static void make_device_format(const WAVEFORMATEX *game_format,
+                                       WAVEFORMATEXTENSIBLE *device_out, uint32_t target_rate);
+
+        // initialize the real device at the target rate, performing the standard WASAPI buffer
+        // realignment retry on AUDCLNT_E_BUFFER_SIZE_NOT_ALIGNED.
+        HRESULT initialize(IAudioClient *real, AUDCLNT_SHAREMODE share_mode, DWORD stream_flags,
+                           REFERENCE_TIME buffer_duration, REFERENCE_TIME periodicity,
+                           const WAVEFORMATEX *device_format, LPCGUID session_guid);
+
+        // translate a device-rate frame count to the equivalent game-rate count, so the buffer-size
+        // and padding values reported to the game stay paced at the game's native rate.
+        UINT32 frames_device_to_game(UINT32 device_frames) const;
+        UINT32 padding_device_to_game(UINT32 device_padding) const;
+
+        // hand the game a scratch buffer sized for `frames` of its native format to write into.
+        HRESULT get_buffer(UINT32 frames, BYTE **ppData);
+
+        // pointer to the input scratch (sized by get_buffer). when chained after the downmix, the
+        // downmix writes its stereo output here for the resampler to consume on the next flush.
+        BYTE *input_data() { return this->scratch.data(); }
+
+        // convert the `frames` the game wrote and push output to the real render client. `boost`
+        // is applied to the converted output. event-driven streams fill exactly one device buffer
+        // per period; timer-driven streams push as many converted frames as the device has free.
+        HRESULT flush(IAudioRenderClient *real, IAudioClient *client, UINT32 frames, DWORD flags,
+                      float boost);
+
+    private:
+
+        // append `frames` of the scratch buffer (native format), or silence, to the input queue
+        void enqueue_input(UINT32 frames, bool silent);
+
+        // event-driven path: produce exactly one full device buffer and push it.
+        HRESULT flush_event(IAudioRenderClient *real, float boost);
+
+        // timer-driven path: convert all queued input into the pending output FIFO, then push as
+        // many frames as the device currently has free, keeping the remainder for the next call.
+        HRESULT flush_timer(IAudioRenderClient *real, IAudioClient *client, float boost);
+
+        // produce exactly out_frames output frames using the fixed src/dst ratio. event-driven
+        // exclusive streams must fill the whole device buffer every period; a small input cushion
+        // is buffered first (see priming) so the sinc kernel always has lookahead.
+        UINT32 produce_exact(UINT32 out_frames);
+
+        // convert all input the kernel can fully support into the pending output FIFO (out_float),
+        // appending without clearing. returns the number of frames produced. used by the
+        // timer-driven path where output is drained to the device in device-paced chunks.
+        UINT32 produce_variable();
+
+        // convolve the windowed-sinc kernel at the current in_pos and append the resulting frame
+        // (one sample per channel) to out_float
+        void emit_frame();
+
+        // precompute the windowed-sinc kernel sampled at kernel_phases sub-sample positions, so
+        // emit_frame is a table lookup instead of recomputing sin/cos per tap (which is far too
+        // expensive to run per sample on the audio callback thread and causes underrun crackle).
+        void build_kernel();
+
+        // drop input frames that in_pos has advanced past, keeping a window of history for the
+        // next block's left context
+        void drop_consumed();
+
+        // convert the first `frames` of out_float to the device format, scaled by `gain`
+        void write_output(BYTE *dst, UINT32 frames, float gain) const;
+
+        // sample format of the stream
+        int channels = 0;
+        int bytes_per_sample = 0;
+        bool is_float = false;
+        int game_frame_size = 0;
+
+        uint32_t src_rate = 0;
+        uint32_t dst_rate = 0;
+
+        // sinc low-pass cutoff (1.0 when upsampling, dst/src when downsampling) and window radius
+        double cutoff = 1.0;
+        int half_taps = 16;
+
+        // precomputed kernel: (kernel_phases + 1) rows of 2*half_taps weights, indexed by the
+        // fractional sample position (linearly interpolated between adjacent rows in emit_frame)
+        std::vector<float> kernel_table;
+        int kernel_phases = 1024;
+
+        // interleaved float input queue and the fractional read position within it (in frames)
+        std::vector<float> in_queue;
+        double in_pos = 0.0;
+
+        // emit silence until a full block of input lookahead has accumulated, so the sinc kernel
+        // never reads past the end of the queue (which would distort the tail of every buffer)
+        bool priming = true;
+
+        // interleaved float scratch for produced output
+        std::vector<float> out_float;
+
+        // buffer the game writes its native-rate audio into between get_buffer / flush
+        std::vector<BYTE> scratch;
+
+        // cached device buffer size (frames); a full buffer is produced every period
+        UINT32 device_buffer_frames = 0;
+
+        // leading buffers to silence to avoid a pop on stream start
+        int buffers_to_mute = 16;
+    };
+}

src/spice2x/hooks/audio/backends/wasapi/util.cpp

@@ -2,7 +2,9 @@
 
 #include <audioclient.h>
 
+#include "hooks/audio/util.h"
 #include "util/flags_helper.h"
+#include "util/logging.h"
 
 #include "defs.h"
 
@@ -18,3 +20,46 @@ std::string stream_flags_str(DWORD flags) {
     FLAG(flags, AUDCLNT_STREAMFLAGS_SRC_DEFAULT_QUALITY);
     FLAGS_END(flags);
 }
+
+void print_format(AUDCLNT_SHAREMODE share_mode, DWORD stream_flags, REFERENCE_TIME buffer_duration,
+        REFERENCE_TIME periodicity, const WAVEFORMATEX *device_format) {
+    log_info("audio::wasapi", "... ShareMode         : {}", share_mode_str(share_mode));
+    log_info("audio::wasapi", "... StreamFlags       : {}", stream_flags_str(stream_flags));
+    log_info("audio::wasapi", "... hnsBufferDuration : {} ({:.3f} ms)",
+            buffer_duration, buffer_duration / 10000.0);
+    log_info("audio::wasapi", "... hnsPeriodicity    : {} ({:.3f} ms)",
+            periodicity, periodicity / 10000.0);
+    print_format(device_format);
+}
+
+void print_format(AUDCLNT_SHAREMODE share_mode, const WAVEFORMATEX *device_format) {
+    log_info("audio::wasapi", "... ShareMode         : {}", share_mode_str(share_mode));
+    print_format(device_format);
+}
+
+HRESULT initialize_with_alignment_retry(IAudioClient *client, const char *log_group,
+        AUDCLNT_SHAREMODE share_mode, DWORD stream_flags, REFERENCE_TIME buffer_duration,
+        REFERENCE_TIME periodicity, const WAVEFORMATEX *device_format, LPCGUID session_guid) {
+
+    HRESULT ret = client->Initialize(share_mode, stream_flags, buffer_duration, periodicity,
+            device_format, session_guid);
+
+    // the requested buffer size can end up unaligned for the device; recover by asking for the next
+    // aligned buffer size and re-initializing with a matching duration.
+    if (ret == AUDCLNT_E_BUFFER_SIZE_NOT_ALIGNED) {
+        UINT32 aligned_frames = 0;
+        if (SUCCEEDED(client->GetBufferSize(&aligned_frames)) && aligned_frames > 0) {
+            REFERENCE_TIME aligned_duration = (REFERENCE_TIME)
+                    (10000.0 * 1000 / device_format->nSamplesPerSec * aligned_frames + 0.5);
+
+            log_info(log_group, "buffer not aligned, retrying with {} frames ({} hns)",
+                    aligned_frames, aligned_duration);
+
+            ret = client->Initialize(share_mode, stream_flags, aligned_duration,
+                    periodicity != 0 ? aligned_duration : 0, device_format, session_guid);
+        }
+    }
+
+    return ret;
+}
+

src/spice2x/hooks/audio/backends/wasapi/util.h

@@ -1,8 +1,100 @@
 #pragma once
 
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+#include <cstring>
 #include <string>
 
 #include <windows.h>
 #include <mmreg.h>
+#include <audioclient.h>
 
 std::string stream_flags_str(DWORD flags);
+
+// log the stream parameters (share mode, flags, buffer duration, periodicity) followed by the wave
+// format, matching the block printed at the top of IAudioClient::Initialize.
+void print_format(AUDCLNT_SHAREMODE share_mode, DWORD stream_flags, REFERENCE_TIME buffer_duration,
+        REFERENCE_TIME periodicity, const WAVEFORMATEX *device_format);
+
+// log the share mode followed by the wave format, for paths that only have a share mode (e.g.
+// IAudioClient::IsFormatSupported).
+void print_format(AUDCLNT_SHAREMODE share_mode, const WAVEFORMATEX *device_format);
+
+// detect IEEE float samples: WAVE_FORMAT_IEEE_FLOAT, or WAVE_FORMAT_EXTENSIBLE whose SubFormat is
+// KSDATAFORMAT_SUBTYPE_IEEE_FLOAT (Data1 == 3; _PCM has Data1 == 1)
+inline bool is_ieee_float(const WAVEFORMATEX *fmt) {
+    return fmt->wFormatTag == WAVE_FORMAT_IEEE_FLOAT
+        || (fmt->wFormatTag == WAVE_FORMAT_EXTENSIBLE
+            && reinterpret_cast<const WAVEFORMATEXTENSIBLE *>(fmt)->SubFormat.Data1 == 0x00000003);
+}
+
+// read one sample at `p` as a normalized float in [-1, 1]
+inline float read_sample(const BYTE *p, int bytes, bool is_float) {
+    if (is_float) {
+        float v;
+        memcpy(&v, p, sizeof(float));
+        return v;
+    }
+    switch (bytes) {
+        case 2: {
+            int16_t v;
+            memcpy(&v, p, sizeof(v));
+            return v * (1.0f / 32768.0f);
+        }
+        case 3: {
+            int32_t v = p[0] | (p[1] << 8) | (p[2] << 16);
+            if (v & 0x800000) {
+                v |= ~0xFFFFFF; // sign extend
+            }
+            return v * (1.0f / 8388608.0f);
+        }
+        case 4: {
+            int32_t v;
+            memcpy(&v, p, sizeof(v));
+            return (float) (v * (1.0 / 2147483648.0));
+        }
+        default:
+            return 0.0f;
+    }
+}
+
+// write the normalized float `value` to the sample at `p`, clamping to the format's range
+inline void write_sample(BYTE *p, int bytes, bool is_float, float value) {
+    if (is_float) {
+        float v = std::clamp(value, -1.0f, 1.0f);
+        memcpy(p, &v, sizeof(v));
+        return;
+    }
+    switch (bytes) {
+        case 2: {
+            int16_t v = (int16_t) std::clamp(
+                (int) std::lround(value * 32768.0f), -32768, 32767);
+            memcpy(p, &v, sizeof(v));
+            break;
+        }
+        case 3: {
+            int32_t v = (int32_t) std::clamp<int64_t>(
+                std::llround((double) value * 8388608.0), -8388608, 8388607);
+            p[0] = v & 0xFF;
+            p[1] = (v >> 8) & 0xFF;
+            p[2] = (v >> 16) & 0xFF;
+            break;
+        }
+        case 4: {
+            int32_t v = (int32_t) std::clamp<int64_t>(
+                std::llround((double) value * 2147483648.0), INT32_MIN, INT32_MAX);
+            memcpy(p, &v, sizeof(v));
+            break;
+        }
+        default:
+            break;
+    }
+}
+
+// initialize the real audio client, recovering from AUDCLNT_E_BUFFER_SIZE_NOT_ALIGNED by asking the
+// device for the next aligned buffer size and re-initializing with a matching duration. log_group
+// names the subsystem in the retry log line.
+HRESULT initialize_with_alignment_retry(IAudioClient *client, const char *log_group,
+        AUDCLNT_SHAREMODE share_mode, DWORD stream_flags, REFERENCE_TIME buffer_duration,
+        REFERENCE_TIME periodicity, const WAVEFORMATEX *device_format, LPCGUID session_guid);

src/spice2x/hooks/audio/util.cpp

@@ -48,33 +48,33 @@ void copy_wave_format(WAVEFORMATEXTENSIBLE *destination, const WAVEFORMATEX *sou
 }
 
 void print_format(const WAVEFORMATEX *pFormat) {
-    log_info("audio", "Wave Format:");
+    log_info("audio::wasapi", "Wave Format:");
 
     // format specific
     if (pFormat->wFormatTag == WAVE_FORMAT_EXTENSIBLE) {
         auto format = reinterpret_cast<const WAVEFORMATEXTENSIBLE *>(pFormat);
-        log_info("audio", "... SubFormat           : {}", guid2s(format->SubFormat));
+        log_info("audio::wasapi", "... SubFormat           : {}", guid2s(format->SubFormat));
     } else {
-        log_info("audio", "... wFormatTag          : {}", pFormat->wFormatTag);
+        log_info("audio::wasapi", "... wFormatTag          : {}", pFormat->wFormatTag);
     }
 
     // generic
-    log_info("audio", "... nChannels           : {}", pFormat->nChannels);
-    log_info("audio", "... nSamplesPerSec      : {}", pFormat->nSamplesPerSec);
-    log_info("audio", "... nAvgBytesPerSec     : {}", pFormat->nAvgBytesPerSec);
-    log_info("audio", "... nBlockAlign         : {}", pFormat->nBlockAlign);
-    log_info("audio", "... wBitsPerSample      : {}", pFormat->wBitsPerSample);
+    log_info("audio::wasapi", "... nChannels           : {}", pFormat->nChannels);
+    log_info("audio::wasapi", "... nSamplesPerSec      : {}", pFormat->nSamplesPerSec);
+    log_info("audio::wasapi", "... nAvgBytesPerSec     : {}", pFormat->nAvgBytesPerSec);
+    log_info("audio::wasapi", "... nBlockAlign         : {}", pFormat->nBlockAlign);
+    log_info("audio::wasapi", "... wBitsPerSample      : {}", pFormat->wBitsPerSample);
 
     // format specific
     if (pFormat->wFormatTag == WAVE_FORMAT_EXTENSIBLE) {
         auto format = reinterpret_cast<const WAVEFORMATEXTENSIBLE *>(pFormat);
 
         if (pFormat->wBitsPerSample == 0) {
-            log_info("audio", "... wSamplesPerBlock    : {}", format->Samples.wSamplesPerBlock);
+            log_info("audio::wasapi", "... wSamplesPerBlock    : {}", format->Samples.wSamplesPerBlock);
         } else {
-            log_info("audio", "... wValidBitsPerSample : {}", format->Samples.wValidBitsPerSample);
+            log_info("audio::wasapi", "... wValidBitsPerSample : {}", format->Samples.wValidBitsPerSample);
         }
 
-        log_info("audio", "... dwChannelMask       : {}", channel_mask_str(format->dwChannelMask));
+        log_info("audio::wasapi", "... dwChannelMask       : {}", channel_mask_str(format->dwChannelMask));
     }
 }

src/spice2x/launcher/launcher.cpp

@@ -72,6 +72,7 @@
 #include "games/museca/museca.h"
 #include "hooks/avshook.h"
 #include "hooks/audio/audio.h"
+#include "hooks/audio/asio_proxy.h"
 #include "hooks/audio/backends/wasapi/downmix.h"
 #include "hooks/debughook.h"
 #include "hooks/devicehook.h"
@@ -496,6 +497,9 @@ int main_implementation(int argc, char *argv[]) {
     if (options[launcher::Options::spice2x_SDVXAsioDriver].is_active()) {
         games::sdvx::ASIO_DRIVER = options[launcher::Options::spice2x_SDVXAsioDriver].value_text();
     }
+    if (options[launcher::Options::SDVXAsioTwoChannel].value_bool()) {
+        WrappedAsio::FORCE_TWO_CHANNELS = true;
+    }
     if (options[launcher::Options::spice2x_SDVXSubPos].is_active()) {
         auto txt = options[launcher::Options::spice2x_SDVXSubPos].value_text();
         if (txt == "top") {
@@ -683,6 +687,9 @@ int main_implementation(int argc, char *argv[]) {
     if (options[launcher::Options::GitaDoraArenaAsioDriver].is_active()) {
         games::gitadora::ASIO_DRIVER = options[launcher::Options::GitaDoraArenaAsioDriver].value_text();
     }
+    if (options[launcher::Options::GitaDoraArenaRealtekAccess].value_bool()) {
+        games::gitadora::ALLOW_REALTEK_AUDIO = true;
+    }
     if (options[launcher::Options::LoadNostalgiaModule].value_bool()) {
         attach_nostalgia = true;
     }
@@ -1106,6 +1113,18 @@ int main_implementation(int argc, char *argv[]) {
             hooks::audio::VOLUME_BOOST = (float) std::pow(10.0, decibels / 20.0);
         }
     }
+    if (options[launcher::Options::AudioResample].is_active()) {
+        const uint32_t rate = options[launcher::Options::AudioResample].value_uint32();
+        if (rate > 0) {
+            hooks::audio::RESAMPLE_RATE = rate;
+        }
+    }
+    if (options[launcher::Options::AudioExclusiveBuffer].is_active()) {
+        const uint32_t ms = options[launcher::Options::AudioExclusiveBuffer].value_uint32();
+        if (ms > 0) {
+            hooks::audio::EXCLUSIVE_BUFFER_MS = ms;
+        }
+    }
 
     if (options[launcher::Options::AudioBackend].is_active()) {
         auto &name = options[launcher::Options::AudioBackend].value_text();

src/spice2x/launcher/options.cpp

@@ -948,6 +948,15 @@ static const std::vector<OptionDefinition> OPTION_DEFINITIONS = {
         .category = "Game Options",
         .picker = OptionPickerType::AsioDriver,
     },
+    {
+        // SDVXAsioTwoChannel
+        .title = "SDVX ASIO Two Channel Audio",
+        .name = "sdvxasio2ch",
+        .desc = "Force the game to use two channels for ASIO output.",
+        .type = OptionType::Bool,
+        .game_name = "Sound Voltex",
+        .category = "Game Options",
+    },
     {
         // spice2x_SDVXSubPos
         .title = "SDVX Subscreen Overlay Position",
@@ -1198,12 +1207,25 @@ static const std::vector<OptionDefinition> OPTION_DEFINITIONS = {
         .desc = "For Arena Model: ASIO driver name to use in place of XONAR. "
             "String should match a subkey under HKLM\\SOFTWARE\\ASIO\\\n\n"
             "Requires 7.1 @ 48kHz; if the game rejects your ASIO device, it will automatically "
-            "fall back to using WASAPI.",
+            "fall back to using WASAPI.\n\n"
+            "If the game crashes after successful ASIO init, try enabling -gdafakerealtek.",
         .type = OptionType::Text,
         .game_name = "GitaDora",
         .category = "Game Options",
         .picker = OptionPickerType::AsioDriver,
     },
+    {
+        // GitaDoraArenaRealtekAccess
+        .title = "GitaDora Arena ASIO Allow Headphones",
+        .name = "gdarealtek",
+        .desc = "For Arena Model: allow the game to access the Realtek audio for headphone output.\n\n"
+            "After opening ASIO device, the game then tries to look for audio devices named Realtek to open "
+            "a second audio stream. For compatibility, spice prevents the game from doing this, but enabling "
+            "this option will allow the game to access the Realtek audio device again.",
+        .type = OptionType::Bool,
+        .game_name = "GitaDora",
+        .category = "Game Options",
+    },
     {
         .title = "Force Load Jubeat Module",
         .name = "jb",
@@ -1965,7 +1987,7 @@ static const std::vector<OptionDefinition> OPTION_DEFINITIONS = {
     },
     {
         // VolumeBoost
-        .title = "WASAPI Boost Audio Volume",
+        .title = "WASAPI/ASIO Boost Audio Volume",
         .name = "volumeboost",
         .desc = "Artificially amplifies the hooked audio output by the selected amount, applied "
             "right before the audio reaches the device. Works regardless of channel layout or "
@@ -1984,6 +2006,37 @@ static const std::vector<OptionDefinition> OPTION_DEFINITIONS = {
             {"30", "+30 dB"},
         },
     },
+    {
+        // AudioResample
+        .title = "WASAPI Exclusive Mode Resampling",
+        .name = "resample",
+        .desc = "Resamples the hooked audio output to a fixed sample rate before it reaches the "
+            "device. Useful when a game requests a sample rate the device cannot output in "
+            "exclusive mode (e.g. the game wants 44100 Hz but the device is locked to 48000 Hz).\n\n"
+            "Select the TARGET sample rate (one that your audio card supports).\n\n"
+            "Will result in couple milliseconds of latency and increased CPU usage when active.",
+        .type = OptionType::Enum,
+        .category = "Audio",
+        .elements = {
+            {"44100", "44.1 kHz"},
+            {"48000", "48 kHz"},
+            {"88200", "88.2 kHz"},
+            {"96000", "96 kHz"},
+            {"176400", "176.4 kHz"},
+            {"192000", "192 kHz"},
+        },
+    },
+    {
+        // AudioExclusiveBuffer
+        .title = "WASAPI Exclusive Buffer Size",
+        .name = "exclusivebuffer",
+        .desc = "Enlarges the WASAPI exclusive-mode device buffer to the specified duration in milliseconds.\n\n"
+            "Try setting this to 10 or even 20 if you hear crackling or glitches "
+            "(at the cost of slightly increased latency).",
+        .type = OptionType::Integer,
+        .setting_name = "16",
+        .category = "Audio",
+    },
     {
         // DelayBy5Seconds
         .title = "Delay by 5 Seconds (DEPRECATED - use -sleepduration instead)",

src/spice2x/launcher/options.h

@@ -94,6 +94,7 @@ namespace launcher {
             spice2x_SDVXDigitalKnobSensitivity,
             SDVXDigitalKnobSocd,
             spice2x_SDVXAsioDriver,
+            SDVXAsioTwoChannel,
             spice2x_SDVXSubPos,
             SDVXSubMonitorOverride,
             LoadDDRModule,
@@ -116,6 +117,7 @@ namespace launcher {
             GitaDoraArenaSingleWindow,
             GitaDoraArenaWindowLayout,
             GitaDoraArenaAsioDriver,
+            GitaDoraArenaRealtekAccess,
             LoadJubeatModule,
             LoadReflecBeatModule,
             LoadShogikaiModule,
@@ -201,6 +203,8 @@ namespace launcher {
             AudioDummy,
             DownmixAudioToStereo,
             VolumeBoost,
+            AudioResample,
+            AudioExclusiveBuffer,
             DelayBy5Seconds,
             spice2x_DelayByNSeconds,
             LoadStubs,

src/spice2x/overlay/windows/config.cpp

@@ -12,9 +12,9 @@
 #include "build/resource.h"
 #include "cfg/config.h"
 #include "cfg/configurator.h"
-#include "external/asio/asiolist.h"
 #include "external/imgui/imgui_internal.h"
 #include "external/imgui/misc/cpp/imgui_stdlib.h"
+#include "hooks/audio/asio_driver_scan.h"
 #include "games/io.h"
 #include "games/sdvx/sdvx.h"
 #include "games/popn/popn.h"
@@ -63,7 +63,7 @@ namespace overlay::windows {
     constexpr ImVec4 TEXT_COLOR_RED(1.f, 0.f, 0.f, 1.f);
     constexpr uint32_t OPTION_INPUT_TEXT_WIDTH = 512;
 
-    std::unique_ptr<AsioDriverList> asio_driver_list;
+    std::optional<std::vector<hooks::audio::AsioDriverScanEntry>> asio_driver_list;
 
     Config::Config(overlay::SpiceOverlay *overlay) : Window(overlay) {
         this->title = "Configuration";
@@ -4426,30 +4426,24 @@ namespace overlay::windows {
     void Config::build_option_value_picker(Option& option) {
         auto &definition = option.get_definition();
         if (definition.picker == OptionPickerType::AsioDriver) {
-            if (asio_driver_list == nullptr) {
-                asio_driver_list = std::make_unique<AsioDriverList>();
+            if (!asio_driver_list.has_value()) {
+                asio_driver_list = hooks::audio::scan_asio_drivers();
             }
 
-            ImGui::TextUnformatted("If your ASIO driver is not shown here, close this");
-            ImGui::TextUnformatted("popup and enter the driver name manually.");
-            ImGui::SameLine();
-            ImGui::HelpMarker(
-                "This list is populated by scanning the registry for ASIO drivers.\n\n"
-                "If your driver is not showing up, it may be because it is not properly registered in the system.\n\n"
-                "For 64-bit games, check in HKLM\\SOFTWARE\\ASIO\\.\n\n"
-                "For 32-bit games on 64-bit Windows, check in HKLM\\SOFTWARE\\WOW6432Node\\ASIO\\.\n\n"
-                "spicecfg runs in 32-bit, so it may not see 64-bit-only drivers.");
-
-            ImGui::TextUnformatted("");
-            if (asio_driver_list->driver_list.empty()) {
+            if (asio_driver_list->empty()) {
                 ImGui::TextUnformatted("No ASIO drivers found.");
             } else {
                 ImGui::TextUnformatted("Pick from ASIO drivers:");
                 ImGui::SetNextItemWidth(300.f);
                 if (ImGui::BeginListBox("##asiodrivers")) {
-                    for (const auto &driver : asio_driver_list->driver_list) {
-                        const bool is_selected = option.value == std::string(driver.name);
-                        if (ImGui::Selectable(fmt::format("[{}] {}", driver.id, driver.name).c_str(), is_selected)) {
+                    for (const auto &driver : *asio_driver_list) {
+                        const bool is_selected = option.value == driver.name;
+                        const char *arch = (driver.found_32bit && driver.found_64bit)
+                            ? "32/64-bit"
+                            : (driver.found_64bit ? "64-bit" : "32-bit");
+                        if (ImGui::Selectable(
+                                fmt::format("{} ({})", driver.name, arch).c_str(),
+                                is_selected)) {
                             option.value = driver.name;
                         }
                     }