# 检查 CUDA 支持
nvidia-smi

# 检查 OpenCV CUDA 编译
python3 -c "
import cv2
count = cv2.cuda.getCudaEnabledDeviceCount()
print(f'CUDA 设备: {count}')
print(cv2.getBuildInformation())
"

import cv2
import numpy as np
import time

img = cv2.imread("photo.jpg")

# === CPU 路径 ===
start = time.perf_counter()
for _ in range(100):
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    blurred = cv2.GaussianBlur(gray, (15, 15), 0)
    edges = cv2.Canny(blurred, 50, 150)
cpu_time = (time.perf_counter() - start) / 100 * 1000

# === GPU 路径（UMat）===
img_gpu = cv2.UMat(img)  # 转换为 UMat

start = time.perf_counter()
for _ in range(100):
    gray_gpu = cv2.cvtColor(img_gpu, cv2.COLOR_BGR2GRAY)
    blurred_gpu = cv2.GaussianBlur(gray_gpu, (15, 15), 0)
    edges_gpu = cv2.Canny(blurred_gpu, 50, 150)
gpu_time = (time.perf_counter() - start) / 100 * 1000

print(f"CPU: {cpu_time:.2f} ms")
print(f"GPU (UMat): {gpu_time:.2f} ms")
print(f"加速比: {cpu_time / gpu_time:.1f}×")

# UMat → NumPy 转换
result = edges_gpu.get()  # 从 GPU 取回结果

import cv2

count = cv2.cuda.getCudaEnabledDeviceCount()
if count == 0:
    print("没有 CUDA 设备")
    exit()

for i in range(count):
    info = cv2.cuda.DeviceInfo(i)
    print(f"\nGPU #{i}: {info.name()}")
    print(f"  计算能力: {info.majorVersion()}.{info.minorVersion()}")
    print(f"  总内存: {info.totalMemory() / 1024**3:.1f} GB")
    print(f"  多处理器: {info.multiProcessorCount()}")
    print(f"  是否集成: {info.isIntegrated()}")
    print(f"  是否兼容: {info.isCompatible()}")

import cv2
import numpy as np
import time

img = cv2.imread("photo.jpg")
h, w = img.shape[:2]

# 上传到 GPU
gpu_img = cv2.cuda_GpuMat()
gpu_img.upload(img)

# === CUDA 高斯模糊 ===
# 创建 CUDA 高斯滤波器
gauss_filter = cv2.cuda.createGaussianFilter(
    cv2.CV_8UC3, cv2.CV_8UC3, (15, 15), 5.0
)

# 应用滤波器
start = time.perf_counter()
gpu_blurred = gauss_filter.apply(gpu_img)
gpu_time = (time.perf_counter() - start) * 1000

# CPU 对比
start = time.perf_counter()
cpu_blurred = cv2.GaussianBlur(img, (15, 15), 5.0)
cpu_time = (time.perf_counter() - start) * 1000

print(f"CUDA 高斯模糊: {gpu_time:.2f} ms")
print(f"CPU 高斯模糊: {cpu_time:.2f} ms")
print(f"加速比: {cpu_time / gpu_time:.1f}×")

# 下载结果
result = gpu_blurred.download()

import cv2

# 上传图像到 GPU
gpu_img = cv2.cuda_GpuMat()
gpu_img.upload(img)

# 颜色转换
gpu_gray = cv2.cuda.cvtColor(gpu_img, cv2.COLOR_BGR2GRAY)

# 调整大小
gpu_resized = cv2.cuda.resize(gpu_img, (640, 480))

# 阈值处理
_, gpu_thresh = cv2.cuda.threshold(gpu_gray, 127, 255,
                                    cv2.THRESH_BINARY)

# 边缘检测（Canny）
gpu_edges = cv2.cuda.Canny(gpu_gray, 50, 150)

# 滤波
# 创建滤波器
box_filter = cv2.cuda.createBoxFilter(cv2.CV_8UC1, cv2.CV_8UC1, (5, 5))
gpu_boxed = box_filter.apply(gpu_gray)

# 形态学操作
morph_filter = cv2.cuda.createMorphologyFilter(
    cv2.MORPH_DILATE, cv2.CV_8UC1,
    cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
)
gpu_dilated = morph_filter.apply(gpu_thresh)

# 算术运算
gpu_result = cv2.cuda.add(gpu_img, gpu_img)
gpu_result = cv2.cuda.multiply(gpu_img, 2.0)

# 下载结果
result = gpu_result.download()

import cv2

# 创建 CUDA 流（异步操作）
stream = cv2.cuda_Stream()

# 异步上传
gpu_img1 = cv2.cuda_GpuMat()
gpu_img1.upload(img1, stream)

gpu_img2 = cv2.cuda_GpuMat()
gpu_img2.upload(img2, stream)

# 异步处理
result = cv2.cuda.add(gpu_img1, gpu_img2, stream=stream)

# 等待完成
stream.waitForCompletion()

# 下载
output = result.download()

import cv2
import numpy as np

net = cv2.dnn.readNetFromONNX("model.onnx")

# 设置 CUDA 后端
net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA_FP16)

# 正常使用，OpenCV 自动 GPU 推理
blob = cv2.dnn.blobFromImage(image, 1/255.0, (640, 640),
                              swapRB=True, crop=False)
net.setInput(blob)
output = net.forward()