当前位置： 首页 > news >正文

上海优化外包公司浙江seo技术培训

news 2025/7/5 1:29:35
上海优化外包公司,浙江seo技术培训,大连网站制作网站,wordpress汉化音乐主题GPU-Puzzles项目可以让你学习到GPU编程和cuda核心并行编程的概念，通过一个个小问题让你理解cuda的编程和调用，创建共享显存空间，实现卷积和矩阵乘法等，通过每个小问题之后还会奖励一个狗狗小视频😁 下面是项目的仓库&…

GPU-Puzzles项目可以让你学习到GPU编程和cuda核心并行编程的概念，通过一个个小问题让你理解cuda的编程和调用，创建共享显存空间，实现卷积和矩阵乘法等，通过每个小问题之后还会奖励一个狗狗小视频😁

下面是项目的仓库：https://github.com/srush/GPU-Puzzlesicon-default.png?t=O83Ahttps://github.com/srush/GPU-Puzzles

我本人也是做完了所有的puzzles，特地做一份讲解供大家参考

Puzzle 1: Map 

Implement a "kernel" (GPU function) that adds 10 to each position of vector a and stores it in vector out. You have 1 thread per position.

题目的目的是让out输出为a中所有元素＋10

def map_spec(a):return a + 10def map_test(cuda):def call(out, a) -> None:local_i = cuda.threadIdx.x# FILL ME IN (roughly 1 lines)out[local_i] = a[local_i] + 10return callSIZE = 4
out = np.zeros((SIZE,))
a = np.arange(SIZE)
problem = CudaProblem("Map", map_test, [a], out, threadsperblock=Coord(SIZE, 1), spec=map_spec
)
problem.show()

这里不能完全用Python代码的思想去阅读，函数每次调用cuda.threadId.x的时候都会取一个新的核心，实现并行的效果，下面是可视化运行效果

# MapScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             1 |             1 |             0 |             0 |

Puzzle 2 - Zip

Implement a kernel that adds together each position of a and b and stores it in out. You have 1 thread per position.

在out中每个元素是a,b向量中对应位置元素和，这里不需要什么额外操作，直接local_i作为对应位置的索引即可

def zip_spec(a, b):return a + bdef zip_test(cuda):def call(out, a, b) -> None:local_i = cuda.threadIdx.x# FILL ME IN (roughly 1 lines)out[local_i] = a[local_i] + b[local_i]return callSIZE = 4
out = np.zeros((SIZE,))
a = np.arange(SIZE)
b = np.arange(SIZE)
problem = CudaProblem("Zip", zip_test, [a, b], out, threadsperblock=Coord(SIZE, 1), spec=zip_spec
)
problem.show()

可视化效果


# ZipScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             2 |             1 |             0 |             0 |

Puzzle 3 - Guards 

Implement a kernel that adds 10 to each position of a and stores it in out. You have more threads than positions.

这里是Map的升级版，我们拥有更多的cuda线程，但是这不影响，利用if判断使local_i索引有效即可

def map_guard_test(cuda):def call(out, a, size) -> None:local_i = cuda.threadIdx.x# FILL ME IN (roughly 2 lines)if (local_i < size):out[local_i] = a[local_i] + 10return callSIZE = 4
out = np.zeros((SIZE,))
a = np.arange(SIZE)
problem = CudaProblem("Guard",map_guard_test,[a],out,[SIZE],threadsperblock=Coord(8, 1),spec=map_spec,
)
problem.show()

可视化效果

# GuardScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             1 |             1 |             0 |             0 |

Puzzle 4 - Map 2D

Implement a kernel that adds 10 to each position of a and stores it in out. Input a is 2D and square. You have more threads than positions.

这里更进了一步，教我们cuda可以创建二维的线程，我们创建了多余的线程，需要更多边界判断

def map_2D_test(cuda):def call(out, a, size) -> None:local_i = cuda.threadIdx.xlocal_j = cuda.threadIdx.y# FILL ME IN (roughly 2 lines)if local_i < size and local_j < size:out[local_i, local_j] = a[local_i, local_j]+1return callSIZE = 2
out = np.zeros((SIZE, SIZE))
a = np.arange(SIZE * SIZE).reshape((SIZE, SIZE))
problem = CudaProblem("Map 2D", map_2D_test, [a], out, [SIZE], threadsperblock=Coord(3, 3), spec=map_spec
)
problem.show()

可视化效果

# Map 2DScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             1 |             1 |             0 |             0 |

Puzzle 5 - Broadcast

Implement a kernel that adds a and b and stores it in out. Inputs a and b are vectors. You have more threads than positions.

其实就是前面zip的二维版本，对行向量和列向量相同索引位置相加，线程比实际矩阵大，要注意边界

def broadcast_test(cuda):def call(out, a, b, size) -> None:local_i = cuda.threadIdx.xlocal_j = cuda.threadIdx.y# FILL ME IN (roughly 2 lines)if local_i < size and local_j < size:out[local_i, local_j] = a[local_i, 0] + b[0, local_j]return callSIZE = 2
out = np.zeros((SIZE, SIZE))
a = np.arange(SIZE).reshape(SIZE, 1)
b = np.arange(SIZE).reshape(1, SIZE)
problem = CudaProblem("Broadcast",broadcast_test,[a, b],out,[SIZE],threadsperblock=Coord(3, 3),spec=zip_spec,
)
problem.show()

可视化效果

# BroadcastScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             2 |             1 |             0 |             0 |

Puzzle 6 - Blocks

Implement a kernel that adds 10 to each position of a and stores it in out. You have fewer threads per block than the size of a.

这里和前面不同，每一块中线程数比矩阵要小，但是这也不影响，因为把块是把线程分组了，每个块有一定数量的线程，程序会循环遍历取出一个个线程块，判断好边界条件即可

def map_block_test(cuda):def call(out, a, size) -> None:i = cuda.blockIdx.x * cuda.blockDim.x + cuda.threadIdx.x# FILL ME IN (roughly 2 lines)if i < size:out[i] = a[i] + 10return callSIZE = 9
out = np.zeros((SIZE,))
a = np.arange(SIZE)
problem = CudaProblem("Blocks",map_block_test,[a],out,[SIZE],threadsperblock=Coord(4, 1),blockspergrid=Coord(3, 1),spec=map_spec,
)
problem.show()

可视化效果

# BlocksScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             1 |             1 |             0 |             0 |

Puzzle 7 - Blocks 2D 

Implement the same kernel in 2D. You have fewer threads per block than the size of a in both directions.

这里线程矩阵比实际的矩阵大小要小，但是没关系，设置好边界条件直接遍历即可

def map_block2D_test(cuda):def call(out, a, size) -> None:i = cuda.blockIdx.x * cuda.blockDim.x + cuda.threadIdx.x# FILL ME IN (roughly 4 lines)j = cuda.blockIdx.y * cuda.blockDim.y + cuda.threadIdx.yif i < size and j < size:out[i, j] = a[i, j] + 10return callSIZE = 5
out = np.zeros((SIZE, SIZE))
a = np.ones((SIZE, SIZE))problem = CudaProblem("Blocks 2D",map_block2D_test,[a],out,[SIZE],threadsperblock=Coord(3, 3),blockspergrid=Coord(2, 2),spec=map_spec,
)
problem.show()

可视化效果

# Blocks 2DScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             1 |             1 |             0 |             0 |

Puzzle 8 - Shared 

Implement a kernel that adds 10 to each position of a and stores it in out. You have fewer threads per block than the size of a.

这里是利用共享内存的教学，cuda申请共享内存只能用静态变量

TPB = 4
def shared_test(cuda):def call(out, a, size) -> None:shared = cuda.shared.array(TPB, numba.float32)i = cuda.blockIdx.x * cuda.blockDim.x + cuda.threadIdx.xlocal_i = cuda.threadIdx.xif i < size:shared[local_i] = a[i]cuda.syncthreads()# FILL ME IN (roughly 2 lines)if I < size:out[i] = shared[local_i] + 10return callSIZE = 8
out = np.zeros(SIZE)
a = np.ones(SIZE)
problem = CudaProblem("Shared",shared_test,[a],out,[SIZE],threadsperblock=Coord(TPB, 1),blockspergrid=Coord(2, 1),spec=map_spec,
)
problem.show()

可视化效果

# SharedScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             1 |             1 |             1 |             1 |

Puzzle 9 - Pooling

Implement a kernel that sums together the last 3 position of a and stores it in out. You have 1 thread per position. You only need 1 global read and 1 global write per thread.

这里教会我们要控制全局读写次数，如果把需要重复读的内容移动到共享内存，可以提升效率

def pool_spec(a):out = np.zeros(*a.shape)for i in range(a.shape[0]):out[i] = a[max(i - 2, 0) : i + 1].sum()return outTPB = 8
def pool_test(cuda):def call(out, a, size) -> None:shared = cuda.shared.array(TPB, numba.float32)i = cuda.blockIdx.x * cuda.blockDim.x + cuda.threadIdx.xlocal_i = cuda.threadIdx.x# FILL ME IN (roughly 8 lines)if i < size:shared[local_i] = a[i]cuda.syncthreads()if i == 0:out[i] = shared[local_i]elif i == 1:out[i] = shared[local_i] + shared[local_i - 1]else:out[i] = shared[local_i] + shared[local_i - 1] + shared[local_i - 2]return callSIZE = 8
out = np.zeros(SIZE)
a = np.arange(SIZE)
problem = CudaProblem("Pooling",pool_test,[a],out,[SIZE],threadsperblock=Coord(TPB, 1),blockspergrid=Coord(1, 1),spec=pool_spec,
)
problem.show()

可视化效果

# PoolingScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             1 |             1 |             3 |             1 |

Puzzle 10 - Dot Product

Implement a kernel that computes the dot-product of a and b and stores it in out. You have 1 thread per position. You only need 2 global reads and 1 global write per thread.

这里是手动实现向量点乘，我们可以先把结果暂存到一个共享内存中，然后最后用一个线程统一计算共享内存的数据并输出到结果，这样可以控制全局读写次数

def dot_spec(a, b):return a @ bTPB = 8
def dot_test(cuda):def call(out, a, b, size) -> None:shared = cuda.shared.array(TPB, numba.float32)i = cuda.blockIdx.x * cuda.blockDim.x + cuda.threadIdx.xlocal_i = cuda.threadIdx.x# FILL ME IN (roughly 9 lines)if i < size:shared[local_i] = a[i] * b[i]cuda.syncthreads()if local_i == 0:total = 0.0for j in range(TPB):total += shared[j]out[0] = totalreturn callSIZE = 8
out = np.zeros(1)
a = np.arange(SIZE)
b = np.arange(SIZE)
problem = CudaProblem("Dot",dot_test,[a, b],out,[SIZE],threadsperblock=Coord(SIZE, 1),blockspergrid=Coord(1, 1),spec=dot_spec,
)
problem.show()

可视化效果

# DotScore (Max Per Thread):|  Global Reads | Global Writes |  Shared Reads | Shared Writes ||             2 |             1 |             8 |             1 |

后面还有puzzle11-14，因为难度比较大，思路较为复杂，留在下一篇文章做讲解

​​​​​​​👉🏻GPU Puzzles讲解（二）-CSDN博客

http://www.tj-hxxt.cn/news/561.html

相关文章：

  • 牛商网招聘常州谷歌优化
  • 中山创海软件网站建设2024年新闻摘抄
  • 手机网站 table样式百度指数的网址
  • 网站建设标志设计天津seo数据监控
  • wordpress超级排版器插件免费seo排名网站
  • 网站建设培训公司如何弄一个自己的网站
  • 在哪个网站做一照一码seo页面优化公司
  • 手机微网站尺寸编写网站
  • 租整套房做民宿的网站合肥网站制作公司
  • 做app还是做网站外贸做网站公司哪家好
  • 做发包业务网站太原seo霸屏
  • 网站建设网络营销平台: 云搜系统广州seo关键词
  • 优秀企业网站欣赏seo优化网站词
  • 又顺又旺的公司名字大全成都百度搜索排名优化
  • 外贸网站seo招聘百度指数的搜索指数
  • 建设网站的课题百度推广方式
  • 宁波市住房和城乡建设培训中心网站百度搜索引擎下载免费
  • 网站建设系统chi系统自媒体平台app下载
  • 怎么做网站广告古董推广网站排名优化seo教程
  • 英特尔nuc做网站服务器百度搜索推广多少钱
  • 网网站站建建站站手机创建网站免费注册
  • 哪里有配音的网站seo诊断站长
  • 济宁500元网站建设德州网站建设优化
  • 电子商务网站的目的网上代写文章一般多少钱
  • 用html5设计个人网站好搜自然seo
  • 做网站用的什么语言seo推广案例
  • 做攻略的网站帮忙推广的平台
  • 深圳餐饮网站设计qq推广
  • 做漂亮的二维码网站seo关键词推广怎么做
  • 官网推广运营的步骤有哪些全国最好网络优化公司