MATLAB性能基准测试

方法1

从广义上讲，第一种方法包括两个部分。第一个为我们提供了一个很大的长字符串，该字符串具有单元格数组中的所有字符，并用下划线将两个单元格分隔开，这实际上是通过实现的cumsum。第二个基于的代码为bsxfun我们提供了所需格式的所需数字输出。

代码看起来像这样-

function out = solution_cumsum_bsxfun(list_of_words)

N = 1 + sum(list_of_words{1}=='_'); %// Number of "words" per cell
lens = cellfun('length',list_of_words); %// No. of chars [lengths] in each cell
tlens = sum(lens); %// Total number of characters [total of lengths]
cumlens = cumsum(lens); %// Cumulative lengths

%// Create an array of ones and zeros. 
%// The length of this array would be equal to sum of characters in all cells. 
%// The ones would be at the positions where new cells start, zeros otherwise
startpos_newcell(1,tlens)=0;
startpos_newcell(cumlens(1:end-1)+1)=1;

%// Calculate new positions for all characters in the cell array with 
%// places/positions left between two cells for putting underscores
newpos = (1:tlens) + cumsum(startpos_newcell);

%// Create one long string that has all the characters from the cell arary
%// with underscores separating cells
one_str(newpos) = [list_of_words{:}];
one_str(cumlens + [1:numel(cumlens)]') = '_'; %//'#

pos_us = find(one_str=='_'); %// positions of underscores
wordend_idx = pos_us - [1:numel(pos_us)]; %// word end indices
wordlens = [wordend_idx(1) diff(wordend_idx)]; %// word lengths
max_wordlens = max(wordlens); %// maximum word length

%// Create mask where digit characters are to be placed
mask = bsxfun(@ge,[1:max_wordlens]',[max_wordlens+1-wordlens]); %//'#

%// Create a numeric array with columns for each word and each row holding a digit.
%// The most significant digits to least significant going from top to bottom
num1(max_wordlens,size(mask,2)) = 0;
num1(mask) = one_str(one_str~='_') - 48;

%// Finally get the desired output converting each column to a single
%// number, reshaping to desired size and scaling down by a factor of 1000
out = reshape(10.^(max_wordlens-4:-1:-3)*num1,N,[]).'; %//'#

return;

方法＃2

方法＃1略有变化，因为它使用进行了第一部分的工作sprintf。现在，在@chappjc的解决方案中看到了这个想法后，我想到了这个想法。我感谢他让我在此修改版本中使用它。

这是代码-

function out = solution_sprintf_bsxfun(list_of_words)

N = 1 + sum(list_of_words{1}=='_'); %// Number of "words" per cell

%// Create one long string that has all the characters from the cell arary
%// with underscores separating cells
one_str = sprintf('%s_',list_of_words{:});

pos_us = find(one_str=='_'); %// positions of underscores
wordend_idx = pos_us - [1:numel(pos_us)]; %// word end indices
wordlens = [wordend_idx(1) diff(wordend_idx)]; %// word lengths
max_wordlens = max(wordlens); %// maximum word length

%// Create mask where digit characters are to be placed
mask = bsxfun(@ge,[1:max_wordlens]',[max_wordlens+1-wordlens]); %//'#

%// Create a numeric array with columns for each word and each row holding a digit.
%// The most significant digits to least significant going from top to bottom
num1(max_wordlens,size(mask,2)) = 0;
num1(mask) = one_str(one_str~='_') - 48;

%// Finally get the desired output converting each column to a single
%// number, reshaping to desired size and scaling down by a factor of 1000
out = reshape(10.^(max_wordlens-4:-1:-3)*num1,N,[]).'; %//'#

return;

方法＃3

这是一种基于以下方面的全新方法getByteStreamFromArray：

function out = solution_bytstrm_bsxfun(list_of_words)

allchars = [list_of_words{:}];
digits_array = getByteStreamFromArray(allchars);

%// At my 32-bit system getByteStreamFromArray gets the significant digits 
%// from index 65 onwards. Thus, we need to crop/index it accordingly
digits_array = digits_array(65:65+numel(allchars)-1) - 48;
% --------------------------------------------------------------------
N = 1 + sum(list_of_words{1}=='_'); %// Number of "words" per cell
lens = cellfun('length',list_of_words); %// No. of chars [lengths] in each cell
cumlens = cumsum(lens)'; %//'# Cumulative lengths
% ----------------------------------------------------------
pos_us = find(digits_array==47);
starts = [[1 cumlens(1:end-1)+1];reshape(pos_us+1,N-1,[])];
ends = [reshape(pos_us-1,N-1,[]) ; cumlens];
gaps = ends(:) - starts(:) + 1;

maxg = max(gaps);
mask1 = bsxfun(@lt,maxg-gaps',[1:maxg]');

num_array(maxg,numel(lens)*N)=0;
num_array(mask1) = digits_array(digits_array~=47);
out = reshape(10.^(maxg-4:-1:-3)*num_array,N,[]).';

return;

接下来列出上述方法的GPU版本。

方法＃4

function out = soln_sprintf_bsxfun_gpu(list_of_words)

N = 1 + sum(list_of_words{1}=='_'); %// Number of "words" per cell

%// Create one long string that has all the characters from the cell arary
%// with underscores separating cells. Now this one uses sprintf as
%// firstly proposed in @chappjc's solution and he has agreed to let me use
%// it, so appreciating his help on this.
digits_array = gpuArray(single(sprintf('%s_',list_of_words{:})));
digits_array = digits_array - 48;

mask_us = digits_array==47; %// mask of underscores
pos_us = find(mask_us);

wordend_idx = pos_us - gpuArray.colon(1,numel(pos_us)); %// word end indices
wordlens = [wordend_idx(1) diff(wordend_idx)]; %// word lengths
max_wordlens = max(wordlens); %// maximum word length

%// Create a numeric array with columns for each word and each row holding a digit.
%// The most significant digits to least significant going from top to bottom
num1 =  single(zeros(max_wordlens,numel(pos_us),'gpuArray')); %//'
num1(bsxfun(@ge,gpuArray.colon(1,max_wordlens)',...
    max_wordlens+1-single(wordlens))) = digits_array(~mask_us); %//'

%// Finally get the desired output converting each column to a single
%// number, reshaping to desired size and scaling down by a factor of 1000
outg = reshape(10.^(max_wordlens-4:-1:-3)*num1,N,[]).'; %//'#
out = gather(outg);

返回;

方法5

function out = soln_bytstrm_bsxfun_gpu(list_of_words)

us_ascii_num = 95;
allchars = [list_of_words{:}];

%// At my 32-bit system getByteStreamFromArray gets the significant digits 
%// from index 65 onwards. Thus, we need to crop/index it accordingly
alldigits = getByteStreamFromArray(allchars);
digits_array1 = gpuArray(alldigits(65:65+numel(allchars)-1));
% --------------------------------------------------------------------
lens = cellfun('length',list_of_words); %// No. of chars [lengths] in each cell
N = sum(digits_array1(1:lens(1))==us_ascii_num)+1; %// Number of "words" per cell
lens = gpuArray(lens);
cumlens = cumsum(lens)'; %//'# Cumulative lengths
% ----------------------------------------------------------
mask_us = digits_array1==us_ascii_num; %// mask of underscores
pos_us = find(mask_us);
starts = [[1 cumlens(1:end-1)+1];reshape(pos_us+1,N-1,[])];
ends = [reshape(pos_us-1,N-1,[]) ; cumlens];
gaps = ends(:) - starts(:) + 1;

maxg = max(gaps);
mask1 = bsxfun(@lt,maxg-gaps',[1:maxg]');

num_array = zeros(maxg,numel(lens)*N,'gpuArray'); %//'
num_array(mask1) = digits_array1(~mask_us)-48;
out = reshape(10.^(maxg-4:-1:-3)*num_array,N,[]).'; %//'
out = gather(out);

return;