C++实战：如何高效获取Windows系统GPU占用率

1次阅读

共计 3292 个字符，预计需要花费 9 分钟才能阅读完成。

在 Windows 平台开发高性能应用时，实时监控 GPU 占用率是性能调优的关键指标。典型场景包括：

游戏引擎需要动态调整画质参数以避免帧率波动
AI 推理服务需平衡多 GPU 卡间的负载分配
视频处理软件要防止编码任务过载导致卡顿

现有方案普遍存在精度不足（如任务管理器仅显示整数百分比）、延迟高（WMI 查询响应慢）或兼容性差（厂商专属 API 绑定硬件）等问题。

方案	延迟	精度	兼容性	实现复杂度
WMI 查询	高(~50ms)	5% 粒度	全平台通用	低
NVAPI	低(~1ms)	0.1% 粒度	仅 NVIDIA 显卡	中
DirectX 计数器	中(~10ms)	0.5% 粒度	需 DX12 兼容硬件	高

#include <comdef.h>
#include <Wbemidl.h>
#pragma comment(lib, "wbemuuid.lib")

float GetGPULoadByWMI() {
    HRESULT hr;
    IWbemLocator* pLoc = nullptr;
    IWbemServices* pSvc = nullptr;
    IEnumWbemClassObject* pEnumerator = nullptr;

    // 初始化 COM（单线程套间）hr = CoInitializeEx(0, COINIT_APARTMENTTHREADED);

    // 连接 WMI 命名空间
    hr = CoCreateInstance(CLSID_WbemLocator, 0, CLSCTX_INPROC_SERVER,
                          IID_IWbemLocator, (LPVOID*)&pLoc);
    hr = pLoc->ConnectServer(_bstr_t(L"ROOT\\CIMV2"), NULL, NULL, 0, NULL, 0, 0, &pSvc);

    // 执行查询
    hr = pSvc->ExecQuery(bstr_t("WQL"), 
        bstr_t("SELECT LoadPercentage FROM Win32_VideoController"),
        WBEM_FLAG_FORWARD_ONLY, NULL, &pEnumerator);

    // 解析结果
    IWbemClassObject* pObj = nullptr;
    ULONG uReturn = 0;
    float load = 0.0f;

    while (pEnumerator->Next(WBEM_INFINITE, 1, &pObj, &uReturn) == S_OK) {
        VARIANT vtProp;
        hr = pObj->Get(L"LoadPercentage", 0, &vtProp, 0, 0);
        load = static_cast<float>(vtProp.intVal);
        VariantClear(&vtProp);
        pObj->Release();
        break; // 通常只取第一个 GPU
    }

    // 释放资源
    pSvc->Release();
    pLoc->Release();
    pEnumerator->Release();
    CoUninitialize();

    return load;
}

#include "nvapi.h"

float GetNvidiaGPULoad() {
    NvAPI_Status status;
    NV_GPU_DYNAMIC_PSTATES_INFO_EX info = {0};
    info.version = NV_GPU_DYNAMIC_PSTATES_INFO_EX_VER;

    status = NvAPI_Initialize();
    if (status != NVAPI_OK) return -1.0f;

    NvPhysicalGpuHandle hGpu;
    NvU32 gpuCount = 0;
    status = NvAPI_EnumPhysicalGPUs(&hGpu, &gpuCount);
    if (status != NVAPI_OK || gpuCount == 0) return -1.0f;

    status = NvAPI_GPU_GetDynamicPstatesInfoEx(hGpu, &info);
    if (status != NVAPI_OK) return -1.0f;

    // 查找 GPU 负载指标（通常索引 0）for (int i = 0; i < NVAPI_MAX_GPU_UTILIZATIONS; i++) {if (info.utilization[i].bIsPresent && 
            info.utilization[i].type == NVAPI_GPU_UTILIZATION_GPU) {return info.utilization[i].percentage;
        }
    }

    return -1.0f;
}

#include <dxgi.h>
#include <d3d12.h>

struct DXGIAdapterMonitor {
    IDXGIAdapter3* pAdapter;
    ID3D12Device* pDevice;
    ID3D12QueryHeap* pQueryHeap;
    ID3D12CommandQueue* pCmdQueue;
};

bool InitDXGIMonitor(DXGIAdapterMonitor* pMon) {
    // 创建 DXGI 工厂
    IDXGIFactory1* pFactory;
    CreateDXGIFactory1(IID_PPV_ARGS(&pFactory));

    // 枚举适配器
    IDXGIAdapter1* pAdapter;
    for (UINT i = 0; pFactory->EnumAdapters1(i, &pAdapter) != DXGI_ERROR_NOT_FOUND; ++i) {
        DXGI_ADAPTER_DESC1 desc;
        pAdapter->GetDesc1(&desc);
        if (desc.Flags & DXGI_ADAPTER_FLAG_SOFTWARE) continue;

        // 转换为 Adapter3 接口
        pAdapter->QueryInterface(IID_PPV_ARGS(&pMon->pAdapter));
        pAdapter->Release();
        break;
    }

    // 创建 D3D12 设备
    D3D12CreateDevice(pMon->pAdapter, D3D_FEATURE_LEVEL_12_0, 
                      IID_PPV_ARGS(&pMon->pDevice));

    // 创建查询堆和命令队列（代码略）return true;
}

float GetDXGIGPULoad(DXGIAdapterMonitor* pMon) {
    DXGI_QUERY_VIDEO_MEMORY_INFO info;
    pMon->pAdapter->QueryVideoMemoryInfo(0, DXGI_MEMORY_SEGMENT_GROUP_LOCAL, &info);

    // 计算使用率 = (预算内存 - 可用内存) / 预算内存
    return (info.Budget - info.AvailableForReservation) * 100.0f / info.Budget;
}