def gpu_remap(numpy_img,map_tensor):
    # 准备图像数据
    img_tensor = torch.from_numpy(numpy_img).contiguous().cuda(non_blocking=True)
    img_tensor = img_tensor.permute(2,0,1).unsqueeze(0).float()
    res = torch.nn.functional.grid_sample(img_tensor,map_tensor,
                                          mode='bilinear',
                                          padding_mode='zeros',
                                          align_corners=None)
    res = res.char()
    res = res[0].permute(1,2,0)
    res = res.cpu()
    res = res.numpy()
    res = np.uint8(res)
    return res

def cpu_remap(numpy_img,mapx,mapy):
    return cv2.remap(numpy_img,mapx,mapy,cv2.INTER_LINEAR)

if __name__=='__main__':
    mtx = np.asarray(
                      [[1.45400077e+03,0.00000000e+00,8.91370579e+02],
                       [0.00000000e+00,1.20949778e+03,3.22130436e+02],
                       [0.00000000e+00,0.00000000e+00,1.00000000e+00]])
    dist =np.asarray([[-0.45415106,0.42747461,0.04382156,0.003691,-0.29239333]])
    w,h = 1920,1080
    # 准备map
    mapx, mapy = cv2.initUndistortRectifyMap(mtx, dist, None, mtx, (w, h), 5)
        
    # 准备grid数据
    mapx_tensor = torch.from_numpy(mapx).unsqueeze(2)/1920*2-1
    mapy_tensor = torch.from_numpy(mapy).unsqueeze(2)/1080*2-1
    map_tensor = torch.cat([mapx_tensor,mapy_tensor],dim=2).unsqueeze(0).cuda() 
    img_org = cv2.imread('test.jpg')

    for i in range(100):
        cpu_res = cpu_remap(img_org,mapx,mapy)
   
    for i in range(100):
        cpu_res = gpu_remap(img_org,map_tensor)

计算了下耗时，结果发现还是CPU->GPU数据拷贝是瓶颈。。。但是计算耗时确实少很多，后续看看如何优化吧。。。