共计 8603 个字符，预计需要花费 22 分钟才能阅读完成。
对于使用 C 语言和 C++ 来开发 GPU 加速应用程序的开发者来说，NVIDIA CUDA Toolkit 可提供一个综合的开发环境。CUDA Toolkit 包含一个针对英伟达 GPU 的编译程序、诸多数学库以及可用于调试和优化应用程序性能的各种工具。你还将找到编程指南、用户手册、API 参考、以及能够帮助你快速着手开发 GPU 加速应用程序的其它文档。
说明：由于 Nvidia 并未给出 Ubuntu 16.04 上面的 CUDA Toolkit，本文方法不一定可行，我这边安装成功，感觉完全是瞎猫碰死耗子了。不过没有安装 sample，只是其他程序可以使用显卡了。
1. 第一个网址，使用
sudo apt-get install nvidia-cuda-toolkit安装 cuda toolkit，要看网速，下载很慢。还有，网址中说重启 ubuntu 有问题（I can’t log in to my computer and end up in infinite login screen）。我这边安装了之后，正常登陆了，没有出现问题。
2. 安装完之后的信息：
装的是 7.5.17，不是最新的 7.5.18，但是能用就行。
3. 第二个网址中 qed 给出了在终端中持续显示 GPU 当前的使用率（仅限 nvidia 的显卡）：
nvidia-smi -l 1
结果：
说明：上面的命令貌似要显卡支持才行。也可以使用 Jonathan 提供的命令（目前没测试）：
watch -n0.1 "nvidia-settings -q GPUUtilization -q useddedicatedgpumemory"160713 说明 ：a. 这条命令显示信息如下：
b. 其实这条命令就是在终端中显示‘NVIDIA X serve settings’中的一些信息，如下（NVIDIA X serve settings 位置为 /usr/share/applications，也可以直接打开该软件查看）：
c. 由于这张图使用的 GPU 和之前使用的 GPU 不一样，因而参数不一致（比如显存）。
4. 安装完 cuda 之后，安装 cutorch，之后安装 cunn，都安装成功。使用 GPU 的程序也能正常运行。
5. 第三个参考网址中给出了测试程序，本处稍微进行了修改，打印出来每次循环执行的时间（CPU 版本和 GPU 版本代码实际上差不多）：
① CPU 版本：
require 'torch'
require 'nn'
require 'optim'
--require 'cunn'
--require 'cutorch'
mnist = require 'mnist'
fullset = mnist.traindataset()
testset = mnist.testdataset()
trainset = {size = 50000,
data = fullset.data[{{1,50000}}]:double(),
label = fullset.label[{{1,50000}}]
}
validationset = {size = 10000,
data = fullset.data[{{50001,60000}}]:double(),
label = fullset.label[{{50001,60000}}]
}
trainset.data = trainset.data - trainset.data:mean()
validationset.data = validationset.data - validationset.data:mean()
model = nn.Sequential()
model:add(nn.Reshape(1, 28, 28))
model:add(nn.MulConstant(1/256.0*3.2))
model:add(nn.SpatialConvolutionMM(1, 20, 5, 5, 1, 1, 0, 0))
model:add(nn.SpatialMaxPooling(2, 2 , 2, 2, 0, 0))
model:add(nn.SpatialConvolutionMM(20, 50, 5, 5, 1, 1, 0, 0))
model:add(nn.SpatialMaxPooling(2, 2 , 2, 2, 0, 0))
model:add(nn.Reshape(4*4*50))
model:add(nn.Linear(4*4*50, 500))
model:add(nn.ReLU())
model:add(nn.Linear(500, 10))
model:add(nn.LogSoftMax())
model = require('weight-init')(model, 'xavier')
criterion = nn.ClassNLLCriterion()
--model = model:cuda()
--criterion = criterion:cuda()
--trainset.data = trainset.data:cuda()
--trainset.label = trainset.label:cuda()
--validationset.data = validationset.data:cuda()
--validationset.label = validationset.label:cuda()--[[]]
sgd_params = {learningRate = 1e-2,
learningRateDecay = 1e-4,
weightDecay = 1e-3,
momentum = 1e-4
}
x, dl_dx = model:getParameters()
step = function(batch_size)
local current_loss = 0
local count = 0
local shuffle = torch.randperm(trainset.size)
batch_size = batch_size or 200
for t = 1,trainset.size,batch_size do
-- setup inputs and targets for this mini-batch
local size = math.min(t + batch_size - 1, trainset.size) - t
local inputs = torch.Tensor(size, 28, 28)--:cuda()
local targets = torch.Tensor(size)--:cuda()
for i = 1,size do
local input = trainset.data[shuffle[i+t]]
local target = trainset.label[shuffle[i+t]]
-- if target == 0 then target = 10 end
inputs[i] = input
targets[i] = target
end
targets:add(1)
local feval = function(x_new)
-- reset data
if x ~= x_new then x:copy(x_new) end
dl_dx:zero()
-- perform mini-batch gradient descent
local loss = criterion:forward(model:forward(inputs), targets)
model:backward(inputs, criterion:backward(model.output, targets))
return loss, dl_dx
end
_, fs = optim.sgd(feval, x, sgd_params)
-- fs is a table containing value of the loss function
-- (just 1 value for the SGD optimization)
count = count + 1
current_loss = current_loss + fs[1]
end
-- normalize loss
return current_loss / count
end
eval = function(dataset, batch_size)
local count = 0
batch_size = batch_size or 200
for i = 1,dataset.size,batch_size do
local size = math.min(i + batch_size - 1, dataset.size) - i
local inputs = dataset.data[{{i,i+size-1}}]--:cuda()
local targets = dataset.label[{{i,i+size-1}}]:long()--:cuda()
local outputs = model:forward(inputs)
local _, indices = torch.max(outputs, 2)
indices:add(-1)
local guessed_right = indices:eq(targets):sum()
count = count + guessed_right
end
return count / dataset.size
end
max_iters = 5
do
local last_accuracy = 0
local decreasing = 0
local threshold = 1 -- how many deacreasing epochs we allow
for i = 1,max_iters do
timer = torch.Timer()
local loss = step()
print(string.format('Epoch: %d Current loss: %4f', i, loss))
local accuracy = eval(validationset)
print(string.format('Accuracy on the validation set: %4f', accuracy))
if accuracy < last_accuracy then
if decreasing > threshold then break end
decreasing = decreasing + 1
else
decreasing = 0
end
last_accuracy = accuracy
print('Time elapsed: ' .. i .. 'iter: ' .. timer:time().real .. ' seconds')
end
end
testset.data = testset.data:double()
eval(testset)② GPU 版本：
require 'torch'
require 'nn'
require 'optim'
require 'cunn'
require 'cutorch'
mnist = require 'mnist'
fullset = mnist.traindataset()
testset = mnist.testdataset()
trainset = {size = 50000,
data = fullset.data[{{1,50000}}]:double(),
label = fullset.label[{{1,50000}}]
}
validationset = {size = 10000,
data = fullset.data[{{50001,60000}}]:double(),
label = fullset.label[{{50001,60000}}]
}
trainset.data = trainset.data - trainset.data:mean()
validationset.data = validationset.data - validationset.data:mean()
model = nn.Sequential()
model:add(nn.Reshape(1, 28, 28))
model:add(nn.MulConstant(1/256.0*3.2))
model:add(nn.SpatialConvolutionMM(1, 20, 5, 5, 1, 1, 0, 0))
model:add(nn.SpatialMaxPooling(2, 2 , 2, 2, 0, 0))
model:add(nn.SpatialConvolutionMM(20, 50, 5, 5, 1, 1, 0, 0))
model:add(nn.SpatialMaxPooling(2, 2 , 2, 2, 0, 0))
model:add(nn.Reshape(4*4*50))
model:add(nn.Linear(4*4*50, 500))
model:add(nn.ReLU())
model:add(nn.Linear(500, 10))
model:add(nn.LogSoftMax())
model = require('weight-init')(model, 'xavier')
criterion = nn.ClassNLLCriterion()
model = model:cuda()
criterion = criterion:cuda()
trainset.data = trainset.data:cuda()
trainset.label = trainset.label:cuda()
validationset.data = validationset.data:cuda()
validationset.label = validationset.label:cuda()--[[]]
sgd_params = {learningRate = 1e-2,
learningRateDecay = 1e-4,
weightDecay = 1e-3,
momentum = 1e-4
}
x, dl_dx = model:getParameters()
step = function(batch_size)
local current_loss = 0
local count = 0
local shuffle = torch.randperm(trainset.size)
batch_size = batch_size or 200
for t = 1,trainset.size,batch_size do
-- setup inputs and targets for this mini-batch
local size = math.min(t + batch_size - 1, trainset.size) - t
local inputs = torch.Tensor(size, 28, 28):cuda()
local targets = torch.Tensor(size):cuda()
for i = 1,size do
local input = trainset.data[shuffle[i+t]]
local target = trainset.label[shuffle[i+t]]
-- if target == 0 then target = 10 end
inputs[i] = input
targets[i] = target
end
targets:add(1)
local feval = function(x_new)
-- reset data
if x ~= x_new then x:copy(x_new) end
dl_dx:zero()
-- perform mini-batch gradient descent
local loss = criterion:forward(model:forward(inputs), targets)
model:backward(inputs, criterion:backward(model.output, targets))
return loss, dl_dx
end
_, fs = optim.sgd(feval, x, sgd_params)
-- fs is a table containing value of the loss function
-- (just 1 value for the SGD optimization)
count = count + 1
current_loss = current_loss + fs[1]
end
-- normalize loss
return current_loss / count
end
eval = function(dataset, batch_size)
local count = 0
batch_size = batch_size or 200
for i = 1,dataset.size,batch_size do
local size = math.min(i + batch_size - 1, dataset.size) - i
local inputs = dataset.data[{{i,i+size-1}}]:cuda()
local targets = dataset.label[{{i,i+size-1}}]:long():cuda()
local outputs = model:forward(inputs)
local _, indices = torch.max(outputs, 2)
indices:add(-1)
local guessed_right = indices:eq(targets):sum()
count = count + guessed_right
end
return count / dataset.size
end
max_iters = 5
do
local last_accuracy = 0
local decreasing = 0
local threshold = 1 -- how many deacreasing epochs we allow
for i = 1,max_iters do
timer = torch.Timer()
local loss = step()
print(string.format('Epoch: %d Current loss: %4f', i, loss))
local accuracy = eval(validationset)
print(string.format('Accuracy on the validation set: %4f', accuracy))
if accuracy < last_accuracy then
if decreasing > threshold then break end
decreasing = decreasing + 1
else
decreasing = 0
end
last_accuracy = accuracy
print('Time elapsed: ' .. i .. 'iter: ' .. timer:time().real .. ' seconds')
end
end
testset.data = testset.data:double()
eval(testset)6. CPU 和 GPU 使用率
① CPU 版本
CPU 情况：
GPU 情况：
② GPU 版本
CPU 情况：
GPU 情况：
7. 可以看出，CPU 版本的程序，CPU 全部使用上了，GPU 则基本没用。GPU 版本，只有一个核心（线程）的 CPU 完全是用上了，其他的则在围观。。。而 GPU 使用率已经很高了。
8. 时间比较
CPU 版本：
Epoch: 1 Current loss: 0.619644
Accuracy on the validation set: 0.924800
Time elapsed: 1iter: 895.69850516319 seconds
Epoch: 2 Current loss: 0.225129
Accuracy on the validation set: 0.949000
Time elapsed: 2iter: 914.15352702141 secondsGPU 版本：
Epoch: 1 Current loss: 0.687380
Accuracy on the validation set: 0.925300
Time elapsed: 1iter: 14.031280994415 seconds
Epoch: 2 Current loss: 0.231011
Accuracy on the validation set: 0.944000
Time elapsed: 2iter: 13.848378896713 seconds
Epoch: 3 Current loss: 0.167991
Accuracy on the validation set: 0.959800
Time elapsed: 3iter: 14.071791887283 seconds
Epoch: 4 Current loss: 0.135209
Accuracy on the validation set: 0.963700
Time elapsed: 4iter: 14.238609790802 seconds
Epoch: 5 Current loss: 0.113471
Accuracy on the validation set: 0.966800
Time elapsed: 5iter: 14.328102111816 seconds说明：① CPU 为 4790K@4.4GHZ（8 线程全开时，应该没有这么高的主频，具体多少没注意）；GPU 为 nvidia GTX 970。
② 由于 CPU 版本的执行时间实在太长，我都怀疑程序是否有问题了。。。但是看着 CPU 一直 100% 的全力工作，又不忍心暂停。直到第一次循环结束，用了将近 900s，才意识到，原来程序应该木有错误。。。等第二次循环结束，就直接停止测试了。。。GPU 版本的程序，每次循环则只用 14s，时间上差距。。。额，使用 CPU 执行时间是 GPU 执行时间的 64 倍。。。
更多 Ubuntu 相关信息见 Ubuntu 专题页面 http://www.linuxidc.com/topicnews.aspx?tid=2
本文永久更新链接地址 ：http://www.linuxidc.com/Linux/2016-07/133200.htm
正文完
星哥玩云-微信公众号
