给一张测试图，对测试图分别cpu和GPU进行处理，进行时间统计，最后做展示。
环境：win10 + cuda11.3 +python3.7 + numba 等
硬件：cpu:i59400 ,gpu:RTX1650 4G

首先进行cuda安装，cuDNN等的安装，参考该博客https://blog.csdn.net/xiao_yun_zi/article/details/124046964?ops_request_misc=%257B%2522request%255Fid%2522%253A%2522167774321216800180677939%2522%252C%2522scm%2522%253A%252220140713.130102334.pc%255Fall.%2522%257D&request_id=167774321216800180677939&biz_id=0&utm_medium=distribute.pc_search_result.none-task-blog-2_allfirst_rank_ecpm_v1~pc_rank_34-1-124046964-null-null.142^v73pc_new_rank,201^v4add_ask,239^v2insert_chatgpt&utm_term=%E6%B8%85%E5%8D%8E%E6%BA%90%E9%95%9C%E5%83%8F%E5%AE%89%E8%A3%85cuda11.3&spm=1018.2226.3001.4187

安装成功后，可以看到自己GPU的信息，大小，传输带宽等

python GPU编程示例如下：

import cv2
import time
import numpy as np
import math
from numba import cuda


@cuda.jit
def process_gpu(img_gpu, channels):
    tx = cuda.blockIdx.x*cuda.blockDim.x + cuda.threadIdx.x
    ty = cuda.blockIdx.y*cuda.blockDim.y + cuda.threadIdx.y
    for c in range(channels):
        color = img_gpu[tx, ty][c]*2 +30
        if color > 255:
            img_gpu[tx, ty][c] = 255
        elif color < 255:
            img_gpu[tx, ty][c] = 0
        else:
            img_gpu[tx, ty][c] = color



def process_cpu(img_cpu, dst_cpu):
    rows, cols, channels = img_cpu.shape
    for j in range(cols):
        for i in range(rows):
            for c in range(channels):
                color = img[i, j][c]*2 + 30
                if color > 255:
                    dst_cpu[i, j][c] = 255
                elif color < 0:
                    dst_cpu[i, j][c] = 0
                else:
                    dst_cpu[i, j][c] = color




if __name__ == '__main__':
    img = cv2.imread("test.jpg")
    cv2.namedWindow("test", 1)
    try:
        cv2.imshow("test", img)
    except:
        print("load failed")

    rows, cols, channels = img.shape
    img_cpu = img.copy()
    dst_cpu = img.copy()
    start_cpu = time.time()

    process_cpu(img_cpu, dst_cpu)
    end_cpu = time.time()
    print("cpu_time:",str(end_cpu-start_cpu))
    cv2.imshow("cpu_photo", dst_cpu)

    ############### gpu process  #########
    img_gpu = cuda.to_device(img)
    threadperblock = (32, 32)
    blockpergrid_x = int(math.ceil(rows/threadperblock[0]))
    blockpergrid_y = int(math.ceil(cols/threadperblock[1]))
    blockpergrid = (blockpergrid_x, blockpergrid_x)
    cuda.syncthreads

    start_gpu = time.time()
    process_gpu[blockpergrid, threadperblock](img_gpu, channels)
    cuda.syncthreads
    end_gpu = time.time()
    print("gpu_time: ", str(end_gpu-start_gpu))

    dst_gpu = img_gpu.copy_to_host()
    cv2.imshow("img_gpu", dst_gpu)
    cv2.waitKey(0)

    cv2.destroyWindow("test")

运行结果展示：


通过结果可以看出，cuda编程加速很多，同时发现GPU得到的图片比cpu得到的图片失真很多。暂时，不知道原因，等待继续研究。

编程时遇到编译问题：numba.cuda.cudadrv.error.NvvmError: Failed to compile

**有校解决办法是用conda 安装numba ，**若用conda 安装遇到如下错误：

Preparing transaction: done
Verifying transaction: failed

RemoveError: 'requests' is a dependency of conda and cannot be removed from
conda's operating environment.

请先将conda 升级。强制升级命令为conda update --force conda