cuda下resize的性能是否有问题？或者是我的用法不正确

[xiangyuanzhi]

MNN组件版本：
MNN Version 3.0.1 with cuda

显卡型号：NVIDIA GeForce RTX 3060 Ti

MNN测试代码：
`void testMNN() {

MNN::BackendConfig backend_config;
MNNDeviceContext gpuDeviceConfig;
gpuDeviceConfig.deviceId = 0;
backend_config.sharedContext = &gpuDeviceConfig;
MNN::Express::Executor::getGlobalExecutor()->setGlobalExecutorConfig(MNN_FORWARD_CUDA, backend_config, MNN_GPU_TUNING_WIDE);
MNN::Express::ExecutorScope::Current()->lazyEval = 0;

LARGE_INTEGER frequency;
QueryPerformanceFrequency(&frequency);

for(int i=0; i<10; i++) {
    LARGE_INTEGER t0;
    QueryPerformanceCounter(&t0);

    float h_crop = 2944.f;
    float w_crop = 2520.f;
    MNN::Express::VARP inputVar = MNN::Express::_Input({1,8,2944,2520}, MNN::Express::NCHW, halide_type_of<float>());
    float *fp = inputVar->writeMap<float>();
    for(int j=0; j<8*2944*2520;j++)
        fp[j] = 0.618f;

    LARGE_INTEGER t1;
    QueryPerformanceCounter(&t1);

    float m_size = 640.f;
    MNN::Express::VARP scaleVar = MNN::Express::_Resize(inputVar, m_size/w_crop, m_size/h_crop);
    
    LARGE_INTEGER t2;
    QueryPerformanceCounter(&t2);

    scaleVar = MNN::Express::_Resize(scaleVar, w_crop/m_size, h_crop/m_size);

    LARGE_INTEGER t3;
    QueryPerformanceCounter(&t3);

    scaleVar->readMap<float>();

    LARGE_INTEGER t4;
    QueryPerformanceCounter(&t4);

    //double elapsed1 = (double)(t1.QuadPart - t0.QuadPart) / frequency.QuadPart;
    double elapsed2 = (double)(t2.QuadPart - t1.QuadPart) / frequency.QuadPart;
    double elapsed3 = (double)(t3.QuadPart - t2.QuadPart) / frequency.QuadPart;
    double elapsed4 = (double)(t4.QuadPart - t3.QuadPart) / frequency.QuadPart;
    printf("===scale_down:%fs, scale_up:%fs, copy:%fs===\n", elapsed2, elapsed3, elapsed4);
}`

PyTorch对比测试代码：
`from torch.nn import functional as F
import torch

for i in range(10):

    #t0 = time.time()
    h_crop = 2944
    w_crop = 2520
    input_var = torch.randn(1, 8, h_crop, w_crop).cuda()
    t1 = time.time()

    m_size = 640
    scale_var = F.interpolate(input_var, size=(m_size, m_size), mode='bilinear')
    torch.cuda.synchronize()

    t2 = time.time()
    scale_var = F.interpolate(scale_var, size=(h_crop, w_crop), mode='bilinear')
    torch.cuda.synchronize()

    t3 = time.time()
    res = scale_var.cpu().numpy()
    t4 = time.time()

    print("===scale_down:{}s, scale_up:{}s, copy_cpu:{}s===".format(round(t2-t1,4), round(t3-t2,4), round(t4-t3,4)))`

性能对比数据：