What would be the most efficient way to expand a dense matrix with new columns in FORTRAN?
Say T is a dense matrix m by n
and I would like to make it m by n+1.
One strategy I could think of : Reallocate at each step and assign the last column or would there be some better ways, such as allocating some space before and checking if that is sufficient and if not do the reallocation kind of stuff? Any ideas?
Assuming m and n are in some sense not exceedingly large, so that your matrices fit into memory and what you're after is performance in time, what I'd do I'd allocate a large matrix and store the actual size separately. This is what, for example, BLAS libraries use as a 'leading dimension'. Then, when you need to add a column, you check if your actual size is still smaller than the maximum size, and reallocate memory if necessary.
If you have a Fortran 2003 Compiler you might make use of move_alloc: http://software.intel.com/sites/products/documentation/hpc/compilerpro/en-us/fortran/lin/compiler_f/lref_for/source_files/rfmvallo.htm
Related
Hi I am a graduate student studying scientific computing using c++. Some of our research focus on speed of an algorithm, therefore it is important to construct array structure that is fast enough.
I've seen two ways of constructing 3D Arrays.
First one is to use vector liblary.
vector<vector<vector<double>>> a (isize,vector<double>(jsize,vector<double>(ksize,0)))
This gives 3D array structure of size isize x jsize x ksize.
The other one is to construct a structure containing 1d array of size isize* jsize * ksize using
new double[isize*jsize*ksize]. To access the specific location of (i,j,k) easily, operator overloading is necessary(am I right?).
And from what I have experienced, first one is much faster since it can access to location (i,j,k) easily while latter one has to compute location and return the value. But I have seen some people preferring latter one over the first one. Why do they prefer the latter setting? and is there any disadvantage of using the first one?
Thanks in adavance.
Main difference between those will be the layout:
vector<vector<vector<T>>>
This will get you a 1D array of vector<vector<T>>.
Each item will be a 1D array of vector<T>.
And each item of those 1D array will be a 1D array of T.
The point is, vector itself does not store its content. It manages a chunk of memory, and stores the content there. This has a number of bad consequences:
For a matrix of dimension X·Y·Z, you will end up allocating 1 + X + X·Y memory chunks. That's horribly slow, and will trash the heap. Imagine: a cube matrix of size 20 would trigger 421 calls to new!
To access a cell, you have 3 levels of indirection:
You must access the vector<vector<vector<T>>> object to get pointer to top-level memory chunk.
You must then access the vector<vector<T>> object to get pointer to second-level memory chunk.
You must then access the vector<T> object to get pointer to the leaf memory chunk.
Only then you can access the T data.
Those memory chunks will be spread around the heap, causing a lot of cache misses and slowing the overall computation.
Should you get it wrong at some point, it is possible to end up with some lines in your matrix having different lengths. After all, they're independent 1-d arrays.
Having a contiguous memory block (like new T[X * Y * Z]) on the other hand gives:
You allocate 1 memory chunk. No heap trashing, O(1).
You only need to access the pointer to the memory chunk, then can go straight for desired element.
All matrix is contiguous in memory, which is cache-friendly.
Those days, a single cache miss means dozens or hundreds lost computing cycles, do not underestimate the cache-friendliness aspect.
By the way, there is a probably better way you didn't mention: using one of the numerous matrix libraries that will handle this for you automatically and provide nice support tools (like SSE-accelerated matrix operations). One such library is Eigen, but there are plenty others.
→ You want to do scientific computing? Let a lib handle the boilerplate and the basics so you can focus on the scientific computing part.
In my point of view, there are too much advantages std::vector's have over normal plain arrays.
In short here are some:
It is much harder to create memory leaks with std::vector. This point alone is one of the biggest advantages. This has nothing to do with performance, but should be considered all the time.
std::vector is part of the STL. This part of C++ is one of the most used one. Thousands of people use the STL and so they get "tested" every day. Over the last years they got optimized so radically, they don't lack any performance anymore. (pls correct me if i see this wrong)
Handling with std::vector is easy as 1, 2, 3. No pointer handling no nothing... Just accessing it via methods or with []-operator and more other methods.
First of all, the idea that you access (i,j,k) in your vec^3 directly is somewhat flawed. What you have is a structure of pointers where you need to dereference three pointers along the way. Note that I have no idea whether that is faster or slower than computing the position within a one-dimensional array, though. You'd need to test that and it might depend on the size of your data (especially whether it fits in a chunk).
Second, the vector^3 requires pointers and vector sizes, which require more memory. In many cases, this will be irrelevant (as the image grows cubically but the memory difference only quadratically) but if your algoritm is really going to fill out any memory available, that can matter.
Third, the raw array stores everything in consecutive memory, which is good for streaming and can be good for certain algorithms because of quick cache accesses. For example when you add one 3D image to another.
Note that all of this is about hyper-optimization that you might not need. The advantages of vectors that skratchi.at pointed out in his answer are quite strong, and I add the advantage that vectors usually increase readability. If you do not have very good reasons not to use vectors, then use them.
If you should decide for the raw array, in any case, make sure that you wrap it well and keep the class small and simple, in order to counter problems regarding leaks and such.
Welcome to SO.
If everything what you have are the two alternatives, then the first one could be better.
Prefer using STL array or vector instead of a C array
You should avoid to use C++ plain arrays since you need to manage yourself the memory allocating/deallocating with new/delete and other boilerplate code like keep track of the size/check bounds. In clearly words "C arrays are less safe, and have no advantages over array and vector."
However, there are some important drawbacks in the first alternative. Something I would like to highlight is that:
std::vector<std::vector<std::vector<T>>>
is not a 3-d matrix. In a matrix, all the rows must have the same size. On the other hand, in a "vector of vectors" there is no guarantee that all the nested vectors have the same length. The reason is that a vector is a linear 1-D structure as pointed out in the #spectras answer. Hence, to avoid all sort of bad or unexpected behaviours, you must to include guards in your code to obtain the rectangular invariant guaranteed.
Luckily, the first alternative is not the only one you may have in hands.
For example, you can replace the c-style array by a std::array:
const int n = i_size * j_size * k_size;
std::array<int, n> myFlattenMatrix;
or use std::vector in case your matrix dimensions can change.
Accessing element by its 3 coordinates
Regarding your question
To access the specific location of (i,j,k) easily, operator
overloading is necessary(am I right?).
Not exactly. Since there isn't a 3-parameter operator for neither std::vector nor array, you can't overload it. But you can create a template class or function to wrap it for you. In any case you will must to deference the 3 vectors or calculate the flatten index of the element in the linear storage.
Considering do not use a third part matrix library like Eigen for your experiments
You aren't coding it for production but for research purposes instead. Particularly, your research is exactly regarding the performance of algorithms. In that case, I prefer do not recommend to use a third part library, like Eigen, absolutely. Of course it depends a lot of what kind of "speed of an algorithm" metrics are you willing to gather, but Eigen, for instance, will do a lot of things under the hood (like vectorization) which will have a tremendous influence on your experiments. Since it will be hard for you to control those unseen optimizations, these library's features may lead you to wrong conclusions about your algorithms.
Algorithm's performance and big-o notation
Usually, the performance of algorithms are analysed by using the big-O approach where factors like the actual time spent, hardware speed or programming language traits aren't taken in account. The book "Data Structures and Algorithms in C++" by Adam Drozdek can provide more details about it.
Is it at all possible in the Eigen library to reserve space for a row-major sparse matrix row by row (different for each row)?
I am trying to optimize the memory consumtion of filling a pretty big sparse matrix (~70mio x 70mio with ~2 billion nnz is the biggest I could reach but I'd like go even further). To clearify the way I went:
First I was using the recommended setFromTriplets which is probably the fastest way filling the matrix, but checking on the memory consumption I found a peak about double the average memory at the point where I used this function, which makes sense since I - at some point - store the elments in both, the matrix and the vector of Triplets until the vector goes out of scope.
Using insert() improved the maximum memory consumption a lot obviously. Though, I still got a peak due to reallocation. I then also used reserve() so that there is no (or less) reallocation. It also lowers the peak quite a bit, but it is not completely gone, again due to some reallocation (if valgrind is correct). Since most of my rows have a lower NNZ than the max I get quite a few empty allocated entries in the storage which increases the average memory consumption. Using makeCompressed lowers the average again but obviously makes the peak higher again as well because more reallocation has to be done when calling it.
Why I asked the question above now is: I can calculate the NNZ for each row up front and also sort them so that I actually should be able to completely optimize this having a compressed matrix without any empty allocation and with no reallocation peak if I could reserve a different number of NNZ for each row.
I would be greatful if anybody would let me know if that is possible or not in Eigen and if not: do you know any library that supports it?
Thanks a bunch!
Yes, it is possible, and you perhaps could have found that function in the docu:
template<class SizesType>
void SparseMatrix::reserve(const SizesType & reserveSizes);
Note that SizesType can be, e.g., an std::vector (or std::deque) or an Eigen::VectorXi.
Also, if you are able to insert the elements to your matrix in order, you may also have a look at the (internal) function insertBack.
I have to read a file in which is stored a matrix with cars (1=BlueCar, 2=RedCar, 0=Empty).
I need to write an algorithm to move the cars of the matrix in that way:
blue ones move downward;
red ones move rightward;
there is a turn in which all the blue ones move and a turn to move all the red ones.
Before the file is read I don't know the matrix size and if it's dense or sparse, so I have to implement two data structures (one for dense and one for sparse) and two algorithms.
I need to reach the best time and space complexity possible.
Due to the unknown matrix size, I think to store the data on the heap.
If the matrix is dense, I think to use something like:
short int** M = new short int*[m];
short int* M_data = new short int[m*n];
for(int i=0; i< m; ++i)
{
M[i] = M_data + i * n;
}
With this structure I can allocate a contiguous space of memory and it is also simple to be accessed with M[i][j].
Now the problem is the structure to choose for the sparse case, and I have to consider also how I can move the cars through the algorithm in the simplest way: for example when I evaluate a car, I need to find easily if in the next position (downward or rightward) there is another car or if it's empty.
Initially I thought to define BlueCar and RedCar objects that inherits from the general Car object. In this objects I can save the matrix coordinates and then put them in:
std::vector<BluCar> sparseBlu;
std::vector<RedCar> sparseRed;
Otherwise I can do something like:
vector< tuple< row, column, value >> sparseMatrix
But the problem of finding what's in the next position still remains.
Probably this is not the best way to do it, so how can I implement the sparse case in a efficient way? (also using a unique structure for sparse)
Why not simply create a memory mapping directly over the file? (assuming your data 0,1,2 is stored in contiguous bytes (or bits) in the file, and the position of those bytes also represents the coordinates of the cars)
This way you don't need to allocate extra memory and read in all the data, and the data can simply and efficiently be accessed with M[i][j].
Going over the rows would be L1-cache friendly.
In case of very sparse data, you could scan through the data once and keep a list of the empty regions/blocks in memory (only need to store startpos and size), which you could then skip (and adjust where needed) in further runs.
With memory mapping, only frequently accessed pages are kept in memory. This means that once you have scanned for the empty regions, memory will only be allocated for the frequently accessed non-empty regions (all this will be done automagically by the kernel - no need to keep track of it yourself).
Another benefit is that you are accessing the OS disk cache directly. Thus no need to keep copying and moving data between kernel space and user space.
To further optimize space- and memory usage, the cars could be stored in 2 bits in the file.
Update:
I'll have to move cars with openMP and MPI... Will the memory mapping
work also with concurrent threads?
You could certainly use multithreading, but not sure if openMP would be the best solution here, because if you work on different parts of the data at the same time, you may need to check some overlapping regions (i.e. a car could move from one block to another).
Or you could let the threads work on the middle parts of the blocks, and then start other threads to do the boundaries (with red cars that would be one byte, with blue cars a full row).
You would also need a locking mechanism for adjusting the list of the sparse regions. I think the best way would be to launch separate threads (depending on the size of the data of course).
In a somewhat similar task, I simply made use of Compressed Row Storage.
The Compressed Row and Column (in the next section) Storage formats
are the most general: they make absolutely no assumptions about the
sparsity structure of the matrix, and they don't store any unnecessary
elements. On the other hand, they are not very efficient, needing an
indirect addressing step for every single scalar operation in a
matrix-vector product or preconditioner solve.
You will need to be a bit more specific about time and space complexity requirements. CSR requires an extra indexing step for simple operations, but that is a minor amount of overhead if you're just doing simple matrix operations.
There's already an existing C++ implementation available online as well.
I have the following situation solved with a vector, but one of my older colleagues told me in a discussion that it would be much faster with an array.
I calculate lots (and I mean lots!) of 12-dimensional vectors from lots of audio files and have to store them for processing. I really need all those vectors before I can start my calculation. Anyhow, I can not predict how many audios, and I can not predict how many vectors are extracted from each audio. Therefor I need a structure to hold the vectors dynamically.
Therefor I create a new double array for each vector and push it to a vector.
I now want to face and test, if my colleague is really right that the calculation can be boosted with using also an array instead of a vector for storing.
vector<double*>* Features = new vector<double*>();
double* feature = new double[12];
// adding elements
Features->push_back(features);
As far as i know to create dynamically 2d array I need to know the count of rows.
double* container = new double*[rows];
container[0] = new double[12];
// and so on..
I know rows after processing all audios, and I don't want to process the audio double times.
Anyone got any idea on how to solve this and append it, or is it just not possible in that way and I should use either vector or create own structure (which assumed may be slower than vector).
Unless have any strong reasons not to, I would suggest something like this:
std::vector<std::array<double, 12>> Features;
You get all the memory locality you could want, and all of the the automagic memory management you need.
You can certainly do this, but it would be much better if you perform this with std::vector. For dynamic growth of a 2D array, you would have to perform all these things.
Create a temporary 2D Array
Allocate memory to it.
Allocate memory to its each component array.
Copy data into its component arrays.
Delete each component array of the original 2D Array.
Delete the 2D Array.
Take new Input.
Add new item to the temporary 2D array.
Create the original 2D Array and allocate memory to it.
Allocate memory to its component arrays.
Copy temporary data into it again.
After doing this in each step, it is hardly acceptable that arrays would be any faster. Use std:vector. The above written answers explain that.
Using vector will make the problem easier because it makes growing the data automatic. Unfortunately due to how vectors grow, using vectors may not be the best solution because of the number of times required to grow for a large data set. On the other hand if you set the initial size of the vector quite large but only need a small number of 12 index arrays. You just wasted a large amount of memory. If there is someway to produce a guess of the size required you might use that guess value to dynamically allocate arrays or set the vector to that size initially.
If you are only going to calculate with the data once or twice, than maybe you should consider using map or list. These two structures for large arrays will create a memory structure that matches your exact needs, and bypass the extra time requirements for growing the arrays. On the other hand the calculations with these data structures will be slower.
I hope these thoughts add some alternative solutions to this discussion.
I need a way to represent a 2-D array (a dense matrix) of doubles in C++, with absolute minimum accessing overhead.
I've done some timing on various linux/unix machines and gcc versions. An STL vector of vectors, declared as:
vector<vector<double> > matrix(n,vector<double>(n));
and accessed through matrix[i][j] is between 5% and 100% slower to access than an array declared as:
double *matrix = new double[n*n];
accessed through an inlined index function matrix[index(i,j)], where index(i,j) evaluates to i+n*j. Other ways of arranging a 2-D array without STL - an array of n pointers to the start of each row, or defining the whole thing on the stack as a constant size matrix[n][n] - run at almost exactly the same speed as the index function method.
Recent GCC versions (> 4.0) seem to be able to compile the STL vector-of-vectors to nearly the same efficiency as the non-STL code when optimisations are turned on, but this is somewhat machine-dependent.
I'd like to use STL if possible, but will have to choose the fastest solution. Does anyone have any experience in optimising STL with GCC?
If you're using GCC the compiler can analyze your matrix accesses and change the order in memory in certain cases. The magic compiler flag is defined as:
-fipa-matrix-reorg
Perform matrix flattening and
transposing. Matrix flattening tries
to replace a m-dimensional matrix with
its equivalent n-dimensional matrix,
where n < m. This reduces the level of
indirection needed for accessing the
elements of the matrix. The second
optimization is matrix transposing
that attemps to change the order of
the matrix's dimensions in order to
improve cache locality. Both
optimizations need fwhole-program
flag. Transposing is enabled only if
profiling information is avaliable.
Note that this option is not enabled by -O2 or -O3. You have to pass it yourself.
My guess would be the fastest is, for a matrix, to use 1D STL array and override the () operator to use it as 2D matrix.
However, the STL also defines a type specifically for non-resizeable numerical arrays: valarray. You also have various optimisations for in-place operations.
valarray accept as argument a numerical type:
valarray<double> a;
Then, you can use slices, indirect arrays, ... and of course, you can inherit the valarray and define your own operator()(int i, int j) for 2D arrays ...
Very likely this is a locality-of-reference issue. vector uses new to allocate its internal array, so each row will be at least a little apart in memory due to each block's header; it could be a long distance apart if memory is already fragmented when you allocate them. Different rows of the array are likely to at least incur a cache-line fault and could incur a page fault; if you're really unlucky two adjacent rows could be on memory lines that share a TLB slot and accessing one will evict the other.
In contrast your other solutions guarantee that all the data is adjacent. It could help your performance if you align the structure so it crosses as few cache lines as possible.
vector is designed for resizable arrays. If you don't need to resize the arrays, use a regular C++ array. STL operations can generally operate on C++ arrays.
Do be sure to walk the array in the correct direction, i.e. across (consecutive memory addresses) rather than down. This will reduce cache faults.
My recommendation would be to use Boost.UBLAS, which provides fast matrix/vector classes.
To be fair depends on the algorithms you are using upon the matrix.
The double name[n*m] format is very fast when you are accessing data by rows both because has almost no overhead besides a multiplication and addition and because your rows are packed data that will be coherent in cache.
If your algorithms access column ordered data then other layouts might have much better cache coherence. If your algorithm access data in quadrants of the matrix even other layouts might be better.
Try to make some research directed at the type of usage and algorithms you are using. That is specially important if the matrix are very large, since cache misses may hurt your performance way more than needing 1 or 2 extra math operations to access each address.
You could just as easily do vector< double >( n*m );
You may want to look at the Eigen C++ template library at http://eigen.tuxfamily.org/ . It generates AltiVec or sse2 code to optimize the vector/matrix calculations.
There is the uBLAS implementation in Boost. It is worth a look.
http://www.boost.org/doc/libs/1_36_0/libs/numeric/ublas/doc/matrix.htm
Another related library is Blitz++: http://www.oonumerics.org/blitz/docs/blitz.html
Blitz++ is designed to optimize array manipulation.
I have done this some time back for raw images by declaring my own 2 dimensional array classes.
In a normal 2D array, you access the elements like:
array[2][3]. Now to get that effect, you'd have a class array with an overloaded
[] array accessor. But, this would essentially return another array, thereby giving
you the second dimension.
The problem with this approach is that it has a double function call overhead.
The way I did it was to use the () style overload.
So instead of
array[2][3], change I had it do this style array(2,3).
That () function was very tiny and I made sure it was inlined.
See this link for the general concept of that:
http://www.learncpp.com/cpp-tutorial/99-overloading-the-parenthesis-operator/
You can template the type if you need to.
The difference I had was that my array was dynamic. I had a block of char memory I'd declare. And I employed a column cache, so I knew where in my sequence of bytes the next row began. Access was optimized for accessing neighbouring values, because I was using it for image processing.
It's hard to explain without the code but essentially the result was as fast as C, and much easier to understand and use.