say I want to multiply two matrices together, 50 by 50. I have 2 ways to arrange threads and blocks.
a) one thread to calculate each element of the result matrix. So I have a loop in thread multiplies one row and one column.
b) one thread to do each multiplication. Each element of the result matrix requires 50 threads. After multiplications are done, I can use binary reduction to sum the results.
I wasn't sure which way to take, so I took b. It wasn't ideal. In fact it was slow. Any idea why? My guess would be there are just too many threads and they are waiting for resource most of time, is that true?
As with so many things in high performance computing, the key to understanding performance here is understanding the use of memory.
If you are using one thread do to do one multiplication, then for that thread you have to pull two pieces of data from memory, multiply them, then do some logarthmic number of adds. That's three memory accesses for a mult and an add and a bit - the arithmatic intensity is very low. The good news is that there are many many threads worth of tasks this way, each of which only needs a tiny bit of memory/registers, which is good for occupancy; but the memory access to work ratio is poor.
The simple one thread doing one dot product approach has the same sort of problem - each multiplication requires two memory accesses to load. The good news is that there's only one store to global memory for the whole dot product, and you avoid the binary reduction which doesn't scale as well and requires a lot of synchronization; the down side is there's way less threads now, which at least your (b) approach had working for you.
Now you know that there should be some way of doing more operations per memory access than this; for square NxN matricies, there's N^3 work to do the multiplication, but only 3xN^2 elements - so you should be able to find a way to do way more than 1 computation per 2ish memory accesses.
The approach taken in the CUDA SDK is the best way - the matricies are broken into tiles, and your (b) approach - one thread per output element - is used. But the key is in how the threads are arranged. By pulling in entire little sub-matricies from slow global memory into shared memory, and doing calculations from there, it's possible to do many multiplications and adds on each number you've read in from memory. This approach is the most successful approach in lots of applications, because getting data - whether it's over a network, or from main memory for a CPU, or off-chip access for a GPU - often takes much longer than processing the data.
There's documents in NVidia's CUDA pages (esp http://developer.nvidia.com/object/cuda_training.html ) which describe their SDK example very nicely.
Have you looked at the CUDA documentation: Cuda Programming Model
Also, sample source code: Matrix Multiplication
Did you look at
$SDK/nvidia-gpu-sdk-3.1/C/src/matrixMul
i.e. the matrix multiplication example in the SDK?
If you don't need to implement this yourself, just use a library -- CUBLAS, MAGMA, etc., provide tuned matrix multiplication implementations.
Related
To launch a CUDA kernel, we use dim3 to specify the dimensions, and I think the meaning of each dimension is opt to the user, for example, it could mean (width, height) or (rows, cols), which has the meaning reversed.
So I did an experiment with the CUDA sample in the SDK: 3_Imaging/convolutionSeparable, simply exchage .x and .y in the kernel function, and reverse the dimensions of blocks and threads used to launch the kernel, so the meaning changes from dim(width, height)/idx(x, y) to dim(rows, cols)/idx(row, col).
The result is the same, however, the performance decreases, the original one takes about 26ms, while the modified one takes about 40ms on my machine(SM 3.0).
My question is, what makes the difference? is (rows, cols) not feasible for CUDA?
P.S. I only modified convolutionRows, no convolutionColumns
EDIT: The change can be found here.
There are at least two potential consequences of your changes:
First, you are changing the memory access pattern to the main memory so the
access is as not coalesced as in the original case.
You should think about the GPU main memory in the same way as it was
a "CPU" memory, i.e., prefetching, blocking, sequential accesses...
techniques to applies in order to get performance.
If you want to know more about this topic, it is mandatory to read
this paper. What every programmer should know about memory.
You'll find an example a comparison between row and column major
access to the elements of a matrix there.
To get and idea on how important this is, think that most -if not
all- GPU high performance codes perform a matrix transposition
before any computation in order to achieve a more coalesced memory
access, and still this additional step worths in terms on
performance. (sparse matrix operations, for instance)
Second. This is more subtle, but in some scenarios it has a deep impact on the performance of a kernel; the launching configuration. It is not the same launching 20 blocks of 10 threads as launching 10 blocks of 20 threads. There is a big difference in the amount of resources a thread needs (shared memory, number of registers,...). The more resources a thread needs the less warps can be mapped on a single SM so the less occupancy... and the -most of the times- less performance.
This not applies to your question, since the number of blocks is equal to the number of threads.
When programming for GPUs you must be aware of the architecture in order to understand how that changes will modify the performance. Of course, I am not familiar with the code so there will be others factors among these two.
I am working on big datasets that are image cubes (450x450x1500). I have a kernel that works on individual data elements. Each data element produces 6 intermediate results (floats). My block consists of 1024 threads. The 6 intermediate results are stored in shared memory by each thread (6 float arrays). However, now I need to add each of the intermediate result to produce a sum (6 sum values). I do not have enough global memory to save these 6 float arrays to global memory and then run a reduction from thrust or any other library from the host code.
Are there any reduction routines that can be called from inside a kernel function on arrays in shared memory?
What will be the best way to solve this problem? I am a newbie to CUDA programming and would welcome any suggestions.
This seems unlikely:
I do not have enough global memory to save these 6 float arrays to global memory and then run a reduction from thrust or any other library from the host code.
I can't imagine how you have enough space to store your data in shared memory but not in global memory.
Anyway, CUB provides reduction routines that can be called from within a threadblock, and that can operate on data stored in shared memory.
Or you can write your own sum-reduction code. It's not terribly hard to do, there are many questions on SO about it, such as this one.
Or you could adapt the cuda sample code.
Update
After seeing all the comments, I understand that instead of doing 1 or a few times of reduction, you need to do the reductions for 450x450x6 times.
In this case there's simpler solution.
You don't need to implement relatively complex parallel reduction for each 1500-D vector。 Since you already have 450x450x6 vectors to reduce, you could reduce all these vectors in parallel using traditional serial reduction method.
You could use a block with 16x16 threads to process a particular region of the image, and a grid with 29x29 blocks to cover the whole 450x450 image.
In each thread, you could iterate over the 1500 frames. In each iterration, you coulde first compute the 6 intermediate results, then add them to the sums. When yo finish all the iterations, you could write the 6 sums to global mem.
That finishes the kernel design. And no shared mem is needed.
You wil find that the performance is very good. Since it is a memory bound operation,it won't be much longer than simply access all the image cube data once.
In case you don't have enough global mem for the whole cube, you could split it into 4 sub-cubes of [1500][225][225], and call the kernel routine on each sub-cube. The only thing you need to change is the grid size.
Have a look at this that explains parallel reduction in CUDA thoroughly.
If I understand it correctly, each thread should sum up "only" 6 floats.
I'm not sure if it is worth doing that by a parallel reduction in general, in the sense that you will experience performance gains.
If you are targeting a Kepler, you may try to use shuffle operations if you properly set the block size so that your intermediate results fit the Streaming Multiprocessor's registers in some way.
As also pointed out by Robert Crovella, your statement about the possibility of storing the intermediate results seems strange as the amount of global memory is certainly larger than the amount of shared memory.
I am searching for some special functions (CUDA) that dedicate to typical dense matrix multiplications, e.g. A*B, where the size of A is 6*n, the size of B is n*6 and n is very large (n=2^24). I have utilized CUBLAS and some other libraries to test this example, In CUBLAS, for this example, we use 6*6=36 threads, which is far from the total parallelism of GPU, so I split A and B into submatrices(vectors) and then implement dot product function for each of them and the performance has been quite well improved. The problem is, in this case, we need to launch 36 CUDA kernels and in between them there are a lot of same data footprints (same data has been accessed for several times from the global memory of GPU). So I am asking whether there exists any solution to this kind of problem.
I have recently written such a matrix multiplication routine for a client of mine. The trick is to extract more parallelism by splitting the long inner summation into several smaller ones. Then use a separate kernel launch to calculate the full sum from the partial ones.
I'm working on an algorithm that has to do a small number
of operations on a large numbers of small arrays, somewhat independently.
To give an idea:
1k sorting of arrays of length typically of 0.5k-1k elements.
1k of LU-solve of matrices that have rank 10-20.
everything is in floats.
Then, there is some horizontality to this problem: the above
operations have to be carried independently on 10k arrays.
Also, the intermediate results need not be stored: for example, i don't
need to keep the sorted arrays, only the sum of the smallest $m$ elements.
The whole thing has been programmed in c++ and runs. My question is:
would you expect a problem like this to enjoy significant speed ups
(factor 2 or more) with CUDA?
You can run this in 5 lines of ArrayFire code. I'm getting speedups of ~6X with this over the CPU. I'm getting speedups of ~4X with this over Thrust (which was designed for vectors, not matrices). Since you're only using a single GPU, you can run ArrayFire Free version.
array x = randu(512,1000,f32);
array y = sort(x); // sort each 512-element column independently
array x = randu(15,15,1000,f32), y;
gfor (array i, x.dim(2))
y(span,span,i) = lu(x(span,span,i)); // LU-decomposition of each 15x15 matrix
Keep in mind that GPUs perform best when memory accesses are aligned to multiples of 32, so a bunch of 32x32 matrices will perform better than a bunch of 31x31.
If you "only" need a factor of 2 speed up I would suggest looking at more straightforward optimisation possibilities first, before considering GPGPU/CUDA. E.g. assuming x86 take a look at using SSE for a potential 4x speed up by re-writing performance critical parts of your code to use 4 way floating point SIMD. Although this would tie you to x86 it would be more portable in that it would not require the presence of an nVidia GPU.
Having said that, there may even be simpler optimisation opportunities in your code base, such as eliminating redundant operations (useless copies and initialisations are a favourite) or making your memory access pattern more cache-friendly. Try profiling your code with a decent profiler to see where the bottlenecks are.
Note however that in general sorting is not a particularly good fit for either SIMD or CUDA, but other operations such as LU decomposition may well benefit.
Just a few pointers, you maybe already incorporated:
1) If you just need the m smallest elements, you are probably better of to just search the smallest element, remove it and repeat m - times.
2) Did you already parallelize the code on the cpu? OpenMP or so ...
3) Did you think about buying better hardware? (I know it´s not the nice think to do, but if you want to reach performance goals for a specific application it´s sometimes the cheapest possibility ...)
If you want to do it on CUDA, it should work conceptually, so no big problems should occur. However, there are always the little things, which depend on experience and so on.
Consider the thrust-library for the sorting thing, hopefully someone else can suggest some good LU-decomposition algorithm.
I got to solve a pretty standard problem on the GPU, but I'm quite new to practical GPGPU, so I'm looking for ideas to approach this problem.
I have many points in 3-space which are assigned to a very small number of groups (each point belongs to one group), specifically 15 in this case (doesn't ever change). Now I want to compute the mean and covariance matrix of all the groups. So on the CPU it's roughly the same as:
for each point p
{
mean[p.group] += p.pos;
covariance[p.group] += p.pos * p.pos;
++count[p.group];
}
for each group g
{
mean[g] /= count[g];
covariance[g] = covariance[g]/count[g] - mean[g]*mean[g];
}
Since the number of groups is extremely small, the last step can be done on the CPU (I need those values on the CPU, anyway). The first step is actually just a segmented reduction, but with the segments scattered around.
So the first idea I came up with, was to first sort the points by their groups. I thought about a simple bucket sort using atomic_inc to compute bucket sizes and per-point relocation indices (got a better idea for sorting?, atomics may not be the best idea). After that they're sorted by groups and I could possibly come up with an adaption of the segmented scan algorithms presented here.
But in this special case, I got a very large amount of data per point (9-10 floats, maybe even doubles if the need arises), so the standard algorithms using a shared memory element per thread and a thread per point might make problems regarding per-multiprocessor resources as shared memory or registers (Ok, much more on compute capability 1.x than 2.x, but still).
Due to the very small and constant number of groups I thought there might be better approaches. Maybe there are already existing ideas suited for these specific properties of such a standard problem. Or maybe my general approach isn't that bad and you got ideas for improving the individual steps, like a good sorting algorithm suited for a very small number of keys or some segmented reduction algorithm minimizing shared memory/register usage.
I'm looking for general approaches and don't want to use external libraries. FWIW I'm using OpenCL, but it shouldn't really matter as the general concepts of GPU computing don't really differ over the major frameworks.
Even though there are few groups, I don't think you will be able to avoid the initial sorting into groups while still keeping the reduction step efficient. You will probably also want to perform the full sort, not just sorting indexes, because that will help keep memory access efficient in the reduction step.
For sorting, read about general strategies here:
http://http.developer.nvidia.com/GPUGems2/gpugems2_chapter46.html
For reduction (old but still good):
http://developer.download.nvidia.com/compute/cuda/1.1-Beta/x86_website/projects/reduction/doc/reduction.pdf
For an example implementation of parallel reduction:
http://developer.nvidia.com/cuda-cc-sdk-code-samples#reduction