Related
I'm new to armadillo. I have the below code, which I assume is inefficient. Any suggestions to make it more memory efficient and/or speedy? Following the armadillo docs and Rcpp gallery, I was unable to get .colptr's, uvec's, or batch insertion to work. But I assume any of them would be improvements.
With an input of X (~100 x 30000), even my stupidly large work VM crashes.
Linux release 7.3.1611 (Core)
117GB RAM / 0GB SWAP
(24 x 2.494 GHz) processor(s)
R version 3.3.2 (2016-10-31)
Platform: x86_64-pc-linux-gnu (64-bit)
Running under: CentOS Linux 7 (Core)
code
#include <RcppArmadillo.h>
// [[Rcpp::depends(RcppArmadillo)]]
using namespace Rcpp;
using namespace arma;
// [[Rcpp::export]]
sp_mat arma_distmat_LT(const arma::mat& x) { // input expected X_{n x p} n << p
int nr, nc;
Col<double> col0, col1;
nr = x.n_rows;
nc = x.n_cols;
sp_mat out(nc, nc);
for (int i = 0; i < nc; i++) {
col0 = x.col(i);
for (int j = i + 1; j < nc; j++) {
col1 = x.col(j);
out(j, i) = as_scalar(col0.t() * col1);
}
}
return out;
}
Call: sourceCpp("<file>"); dist_x <- arma_distmat_LT(X)
Note: these are distances because I am calculating cosine similarities where I have set L2 norm == 1.
It looks to me as if you're just computing the (upper triangular) matrix product t(X)%*%X. You can actually do that directly in R with the underused crossprod function.
X <- matrix(rnorm(100*30000), ncol=30000)
res <- crossprod(X, X)
This takes a few minutes on my laptop. If you change your code to use the Armadillo library then you can use
sp_mat arma_distmat_LT2(const arma::mat& x) { // input expected X_{n x p} n << p
int nr, nc;
Col<double> col0, col1;
nr = x.n_rows;
nc = x.n_cols;
sp_mat out(nc, nc);
out = trimatl(x.t() * x, k=-1);
return out;
}
Still takes a few minutes. It uses an awful amount of memory though so I doubt you can have a lot of objects in memory at the same time.
The code could probably be optimized to only compute the lower/upper triangular matrix.
Just to show the speedup for a 100*800 matrix:
> microbenchmark(crossprod(X, X), arma_distmat_LT(X), arma_distmat_LT2(X))
Unit: milliseconds
expr min lq mean median uq
crossprod(X, X) 50.25574 53.72049 57.98812 56.29532 58.71277
arma_distmat_LT(X) 1331.83243 1471.42465 1523.74060 1492.84611 1512.45416
arma_distmat_LT2(X) 29.69420 33.23954 36.24613 35.54700 38.05208
max neval cld
160.81227 100 a
3080.37891 100 b
66.07351 100 a
As you can see there is a substantial speedup to be gained by brute-forcing it. That being said I'm sure that the cross product can be optimised further.
I just wanted to ask if someone could tell me the reason behind why my eig function computation time is different for the same size of matrices?
I am performing a lot of eig(A,B) operations on 4x4, 4x4 matrices to get eigenvalues and eigen vectors.
It happens 557 * 4 * 267 times.
std::pair<MatrixXcd, VectorXd> eig(const MatrixXcd& A, const MatrixXcd& B)
{
Eigen::GeneralizedSelfAdjointEigenSolver<MatrixXcd> solver(A, B);
MatrixXcd V = solver.eigenvectors();
VectorXd D = solver.eigenvalues();
return std::make_pair(V, D);
}
(...)
for (int t = 0; t < 557; t++)
{
(...)
for(int sig = 0; sig < 4; sig++)
{
for(int f = 0; f < 267; f++)
{
// calculation of Xi and Xfs
mRt = mXfs * mXfs.adjoint();
mRj = mXfi * mXfi.adjoint();
eValVec = eig(mRt, mRj);
(...)
}
(...)
}
}
For t<=7 eig computes 4 * 267 iterations in around 3 sec.
For t>7 it suddenly becomes really fast and computes 4 * 267 iterations in ~0.1 sec.
I checked for sure that its eig that slows my program by commenting other parts of code and leaving only that function. I always use 4x4 matrices.
I guess important notice might be that all values or Rj matrix in t<=7 equals 0. So its like:
0 0 0 0
0 0 0 0
0 0 0 0
0 0 0 0
While Rx has always values. After t>7 Rj also gain non zero values and computation becomes faster. So eig function might have problem to deal with zeros? I don't change anything related to eig function - only numbers in 4x4 matrices I insert to it changes. Still computation time differs. And this lag never happens after.
Do anyone have idea how could I fix it/debug it or make it more stable? I am doing processing in real time and 3 sec is quite too long for me.
Windows 7, NVidia GeForce 425M.
I wrote a simple CUDA code which calculates the row sums of a matrix.
The matrix has uni-dimensional representation (pointer to a float).
The serial version of code is below (it has 2 loops, as expected):
void serial_rowSum (float* m, float* output, int nrow, int ncol) {
float sum;
for (int i = 0 ; i < nrow ; i++) {
sum = 0;
for (int j = 0 ; j < ncol ; j++)
sum += m[i*ncol+j];
output[i] = sum;
}
}
Inside the CUDA code, I call the kernel function sweeping the matrix by rows. Below, the kernel call snippet:
dim3 threadsPerBlock((unsigned int) nThreadsPerBlock); // has to be multiple of 32
dim3 blocksPerGrid((unsigned int) ceil(nrow/(float) nThreadsPerBlock));
kernel_rowSum<<<blocksPerGrid, threadsPerBlock>>>(d_m, d_output, nrow, ncol);
and the kernel function which performs the parallel sum of the rows (still has 1 loop):
__global__ void kernel_rowSum(float *m, float *s, int nrow, int ncol) {
int rowIdx = threadIdx.x + blockIdx.x * blockDim.x;
if (rowIdx < nrow) {
float sum=0;
for (int k = 0 ; k < ncol ; k++)
sum+=m[rowIdx*ncol+k];
s[rowIdx] = sum;
}
}
So far so good. The serial and parallel (CUDA) results are equal.
The whole point is that the CUDA version takes almost twice the time of the serial one to compute, even if I change the nThreadsPerBlock parameter: I tested nThreadsPerBlock from 32 to 1024 (maximum number of threads per block allowed for my card).
IMO, the matrix dimension is big enough to justify parallelization: 90,000 x 1,000.
Below, I report the time elapsed for the serial and parallel versions using different nThreadsPerBlock. Time reported in msec over an average of 100 samples:
Matrix: nrow = 90000 x ncol = 1000
Serial: Average Time Elapsed per Sample in msec (100 samples): 289.18.
CUDA (32 ThreadsPerBlock): Average Time Elapsed per Sample in msec (100 samples): 497.11.
CUDA (1024 ThreadsPerBlock): Average Time Elapsed per Sample in msec (100 samples): 699.66.
Just in case, the version with 32/1024 nThreadsPerBlock is the fastest/slowest one.
I understand that there is a kind of overhead when copying from Host to Device and the other way around, but maybe the slowness is because I am not implementing the fastest code.
Since I am far from being a CUDA expert:
Am I coding the fastest version for this task? How could I improve my code?
Can I get rid of the loop in the kernel function?
Any thoughts appreciated.
EDIT 1
Although I describe a standard rowSum, I am interested in the AND/OR operation of rows which have (0;1} values, like rowAND/rowOR. That said, it doesn't allow me to exploit the cuBLAS multiply by 1's COL column vector trick, as suggested by some commentators.
EDIT 2
As suggest by users other users and here endorsed:
FORGET ABOUT TRYING TO WRITE YOUR OWN FUNCTIONS, use Thrust library instead and the magic comes.
Since you mentioned you need general reduction algorithm other than sum only. I will try to give 3 approaches here. kernel approach may have the highest performance. thrust approach is easiest to implement. cuBLAS approach works only with sum and have good performance.
Kernel Approach
Here's a very good doc introducing how to optimize standard parallel reduction. Standard reduction can be divide into 2 stages.
Multiple thread blocks each reduces one part of the data;
One thread block reduces from result of stage 1 to the final 1 element.
For your multi-reduction (reduce rows of mat) problem, only stage 1 is enough. The idea is to reduce 1 row per thread block. For further considerations like multi-row per thread block or 1 row per multiple thread blocks, you can refer to the paper provided by #Novak. This may improve the performance more, especially for matrices with bad shape.
Thrust Approach
General multi-reduction can be done by thrust::reduction_by_key in a few minutes. You can find some discussions here Determining the least element and its position in each matrix column with CUDA Thrust.
However thrust::reduction_by_key does not assume each row has the same length, so you will get performance penalty. Another post How to normalize matrix columns in CUDA with max performance? gives profiling comparison between thrust::reduction_by_key and cuBLAS approach on sum of rows. It may give you a basic understanding about the performance.
cuBLAS Approach
Sum of rows/cols of a matrix A can be seen as a matrix-vector multiplication where the elements of the vector are all ones. it can be represented by the following matlab code.
y = A * ones(size(A,2),1);
where y is the sum of rows of A.
cuBLAS libary provides a high performance matrix-vector multiplication function cublas<t>gemv() for this operation.
Timing result shows that this routine is only 10~50% slower than simply read all the elements of A once, which can be seen as the theoretical upper limit of the performance for this operation.
Reducing the rows of a matrix can be solved by using CUDA Thrust in three ways (they may not be the only ones, but addressing this point is out of scope). As also recognized by the same OP, using CUDA Thrust is preferable for such a kind of problem. Also, an approach using cuBLAS is possible.
APPROACH #1 - reduce_by_key
This is the approach suggested at this Thrust example page. It includes a variant using make_discard_iterator.
APPROACH #2 - transform
This is the approach suggested by Robert Crovella at CUDA Thrust: reduce_by_key on only some values in an array, based off values in a “key” array.
APPROACH #3 - inclusive_scan_by_key
This is the approach suggested by Eric at How to normalize matrix columns in CUDA with max performance?.
APPROACH #4 - cublas<t>gemv
It uses cuBLAS gemv to multiply the relevant matrix by a column of 1's.
THE FULL CODE
Here is the code condensing the two approaches. The Utilities.cu and Utilities.cuh files are mantained here and omitted here. The TimingGPU.cu and TimingGPU.cuh are maintained here and are omitted as well.
#include <cublas_v2.h>
#include <thrust/host_vector.h>
#include <thrust/device_vector.h>
#include <thrust/generate.h>
#include <thrust/reduce.h>
#include <thrust/functional.h>
#include <thrust/random.h>
#include <thrust/sequence.h>
#include <stdio.h>
#include <iostream>
#include "Utilities.cuh"
#include "TimingGPU.cuh"
// --- Required for approach #2
__device__ float *vals;
/**************************************************************/
/* CONVERT LINEAR INDEX TO ROW INDEX - NEEDED FOR APPROACH #1 */
/**************************************************************/
template <typename T>
struct linear_index_to_row_index : public thrust::unary_function<T,T> {
T Ncols; // --- Number of columns
__host__ __device__ linear_index_to_row_index(T Ncols) : Ncols(Ncols) {}
__host__ __device__ T operator()(T i) { return i / Ncols; }
};
/******************************************/
/* ROW_REDUCTION - NEEDED FOR APPROACH #2 */
/******************************************/
struct row_reduction {
const int Ncols; // --- Number of columns
row_reduction(int _Ncols) : Ncols(_Ncols) {}
__device__ float operator()(float& x, int& y ) {
float temp = 0.f;
for (int i = 0; i<Ncols; i++)
temp += vals[i + (y*Ncols)];
return temp;
}
};
/**************************/
/* NEEDED FOR APPROACH #3 */
/**************************/
template<typename T>
struct MulC: public thrust::unary_function<T, T>
{
T C;
__host__ __device__ MulC(T c) : C(c) { }
__host__ __device__ T operator()(T x) { return x * C; }
};
/********/
/* MAIN */
/********/
int main()
{
const int Nrows = 5; // --- Number of rows
const int Ncols = 8; // --- Number of columns
// --- Random uniform integer distribution between 10 and 99
thrust::default_random_engine rng;
thrust::uniform_int_distribution<int> dist(10, 99);
// --- Matrix allocation and initialization
thrust::device_vector<float> d_matrix(Nrows * Ncols);
for (size_t i = 0; i < d_matrix.size(); i++) d_matrix[i] = (float)dist(rng);
TimingGPU timerGPU;
/***************/
/* APPROACH #1 */
/***************/
timerGPU.StartCounter();
// --- Allocate space for row sums and indices
thrust::device_vector<float> d_row_sums(Nrows);
thrust::device_vector<int> d_row_indices(Nrows);
// --- Compute row sums by summing values with equal row indices
//thrust::reduce_by_key(thrust::make_transform_iterator(thrust::counting_iterator<int>(0), linear_index_to_row_index<int>(Ncols)),
// thrust::make_transform_iterator(thrust::counting_iterator<int>(0), linear_index_to_row_index<int>(Ncols)) + (Nrows*Ncols),
// d_matrix.begin(),
// d_row_indices.begin(),
// d_row_sums.begin(),
// thrust::equal_to<int>(),
// thrust::plus<float>());
thrust::reduce_by_key(
thrust::make_transform_iterator(thrust::make_counting_iterator(0), linear_index_to_row_index<int>(Ncols)),
thrust::make_transform_iterator(thrust::make_counting_iterator(0), linear_index_to_row_index<int>(Ncols)) + (Nrows*Ncols),
d_matrix.begin(),
thrust::make_discard_iterator(),
d_row_sums.begin());
printf("Timing for approach #1 = %f\n", timerGPU.GetCounter());
// --- Print result
for(int i = 0; i < Nrows; i++) {
std::cout << "[ ";
for(int j = 0; j < Ncols; j++)
std::cout << d_matrix[i * Ncols + j] << " ";
std::cout << "] = " << d_row_sums[i] << "\n";
}
/***************/
/* APPROACH #2 */
/***************/
timerGPU.StartCounter();
thrust::device_vector<float> d_row_sums_2(Nrows, 0);
float *s_vals = thrust::raw_pointer_cast(&d_matrix[0]);
gpuErrchk(cudaMemcpyToSymbol(vals, &s_vals, sizeof(float *)));
thrust::transform(d_row_sums_2.begin(), d_row_sums_2.end(), thrust::counting_iterator<int>(0), d_row_sums_2.begin(), row_reduction(Ncols));
printf("Timing for approach #2 = %f\n", timerGPU.GetCounter());
for(int i = 0; i < Nrows; i++) {
std::cout << "[ ";
for(int j = 0; j < Ncols; j++)
std::cout << d_matrix[i * Ncols + j] << " ";
std::cout << "] = " << d_row_sums_2[i] << "\n";
}
/***************/
/* APPROACH #3 */
/***************/
timerGPU.StartCounter();
thrust::device_vector<float> d_row_sums_3(Nrows, 0);
thrust::device_vector<float> d_temp(Nrows * Ncols);
thrust::inclusive_scan_by_key(
thrust::make_transform_iterator(thrust::make_counting_iterator(0), linear_index_to_row_index<int>(Ncols)),
thrust::make_transform_iterator(thrust::make_counting_iterator(0), linear_index_to_row_index<int>(Ncols)) + (Nrows*Ncols),
d_matrix.begin(),
d_temp.begin());
thrust::copy(
thrust::make_permutation_iterator(
d_temp.begin() + Ncols - 1,
thrust::make_transform_iterator(thrust::make_counting_iterator(0), MulC<int>(Ncols))),
thrust::make_permutation_iterator(
d_temp.begin() + Ncols - 1,
thrust::make_transform_iterator(thrust::make_counting_iterator(0), MulC<int>(Ncols))) + Nrows,
d_row_sums_3.begin());
printf("Timing for approach #3 = %f\n", timerGPU.GetCounter());
for(int i = 0; i < Nrows; i++) {
std::cout << "[ ";
for(int j = 0; j < Ncols; j++)
std::cout << d_matrix[i * Ncols + j] << " ";
std::cout << "] = " << d_row_sums_3[i] << "\n";
}
/***************/
/* APPROACH #4 */
/***************/
cublasHandle_t handle;
timerGPU.StartCounter();
cublasSafeCall(cublasCreate(&handle));
thrust::device_vector<float> d_row_sums_4(Nrows);
thrust::device_vector<float> d_ones(Ncols, 1.f);
float alpha = 1.f;
float beta = 0.f;
cublasSafeCall(cublasSgemv(handle, CUBLAS_OP_T, Ncols, Nrows, &alpha, thrust::raw_pointer_cast(d_matrix.data()), Ncols,
thrust::raw_pointer_cast(d_ones.data()), 1, &beta, thrust::raw_pointer_cast(d_row_sums_4.data()), 1));
printf("Timing for approach #4 = %f\n", timerGPU.GetCounter());
for(int i = 0; i < Nrows; i++) {
std::cout << "[ ";
for(int j = 0; j < Ncols; j++)
std::cout << d_matrix[i * Ncols + j] << " ";
std::cout << "] = " << d_row_sums_4[i] << "\n";
}
return 0;
}
TIMING RESULTS (tested on a Kepler K20c)
Matrix size #1 #1-v2 #2 #3 #4 #4 (no plan)
100 x 100 0.63 1.00 0.10 0.18 139.4 0.098
1000 x 1000 1.25 1.12 3.25 1.04 101.3 0.12
5000 x 5000 8.38 15.3 16.05 13.8 111.3 1.14
100 x 5000 1.25 1.52 2.92 1.75 101.2 0.40
5000 x 100 1.35 1.99 0.37 1.74 139.2 0.14
It seems that approaches #1 and #3 outperform approach #2, except in the cases of small numbers of columns. The best approach, however, is approach #4, which is significantly more convenient than the others, provided that the time needed to create the plan can be amortized during the computation.
If this is the extent (summing the rows) of the operations you need to do with this data, I wouldn't expect a sizable benefit from the GPU. You have exactly one arithmetic operation per data element, and for that you are paying the cost of transferring that data element to the GPU. And beyond a certain problem size (whatever it takes to keep the machine busy) you get no added benefit from larger problem sizes, because the arithmetic intensity is O(n).
So this isn't a particularly exciting problem to solve on the GPU.
But as talonmies has indicated, you have a coalescing problem in the way you have crafted it, which will further slow things down. Let's take a look at a small example:
C1 C2 C3 C4
R1 11 12 13 14
R2 21 22 23 24
R3 31 32 33 34
R4 41 42 43 44
Above is a simple pictorial example of a small portion of your matrix. The machine data storage is such that elements (11), (12), (13), and (14) are stored in adjacent memory locations.
For coalesced access, we want an access pattern such that adjacent memory locations are requested from the same instruction, executed across the warp.
We need to think about execution of your code from the standpoint of a warp, that is 32 threads executing in lock-step. What is your code doing? Which elements is it retrieving (asking for) at each step/instruction? Let's take a look at this line of code:
sum+=m[rowIdx*ncol+k];
Adjacent threads in the warp have adjacent (i.e. consecutive) values for rowIdx as you have created that variable. So when k = 0, which data element is being asked for by each thread when we try to retrieve the value m[rowIdx*ncol+k] ?
In block 0, thread 0 has a rowIdx of 0. Thread 1 has a rowIdx of 1, etc. So the values being asked for by each thread at this instruction are:
Thread: Memory Location: Matrix Element:
0 m[0] (11)
1 m[ncol] (21)
2 m[2*ncol] (31)
3 m[3*ncol] (41)
But this is not coalesced access! Elements (11), (21), etc. are not adjacent in memory. For coalesced access, we would like that Matrix Element row to read like this:
Thread: Memory Location: Matrix Element:
0 m[?] (11)
1 m[?] (12)
2 m[?] (13)
3 m[?] (14)
If you then work backwards to determine what the value of ? should be, you will come up with an instruction something like this:
sum+=m[k*ncol+rowIdx];
This will give coalesced access, but it will not give you the correct answer, because we are now summing matrix columns instead of matrix rows. We can fix this by re-organizing your data storage to be in column-major order rather than row-major order. (You should be able to google that for ideas, right?) Conceptually, this is equivalent to transposing your matrix m. Whether this is convenient for you to do or not is outside the scope of your question, as I see it, and not really a CUDA issue. It may be a simple thing for you to do as you are creating the matrix on the host or transferring the matrix from host to device. But in summary, I don't know of a way to sum the matrix rows with 100% coalesced access, if the matrix is stored in row-major order. (You could resort to a sequence of row-reductions but that looks painful to me.)
It's not uncommon, when we are thinking about ways to accelerate code on the GPU, to consider re-organizing our data storage to facilitate the GPU. This is one example.
And, yes, what I'm outlining here still retains a loop in the kernel.
As an additional comment, I would suggest timing the data copy portions, and kernel (compute) portions separately. I can't tell from your question whether you are timing just the kernel or the entire (GPU) operation, including the data copies. If you time the data copies separately, you may discover that just the data copy time exceeds your CPU time. Any effort put into optimizing your CUDA code will not affect the data copy time. This might be a useful data point before you spend much time on this.
I can see with the CPU profiler, that the compute_variances() is the bottleneck of my project.
% cumulative self self total
time seconds seconds calls ms/call ms/call name
75.63 5.43 5.43 40 135.75 135.75 compute_variances(unsigned int, std::vector<Point, std::allocator<Point> > const&, float*, float*, unsigned int*)
19.08 6.80 1.37 readDivisionSpace(Division_Euclidean_space&, char*)
...
Here is the body of the function:
void compute_variances(size_t t, const std::vector<Point>& points, float* avg,
float* var, size_t* split_dims) {
for (size_t d = 0; d < points[0].dim(); d++) {
avg[d] = 0.0;
var[d] = 0.0;
}
float delta, n;
for (size_t i = 0; i < points.size(); ++i) {
n = 1.0 + i;
for (size_t d = 0; d < points[0].dim(); ++d) {
delta = (points[i][d]) - avg[d];
avg[d] += delta / n;
var[d] += delta * ((points[i][d]) - avg[d]);
}
}
/* Find t dimensions with largest scaled variance. */
kthLargest(var, points[0].dim(), t, split_dims);
}
where kthLargest() doesn't seem to be a problem, since I see that:
0.00 7.18 0.00 40 0.00 0.00 kthLargest(float*, int, int, unsigned int*)
The compute_variances() takes a vector of vectors of floats (i.e. a vector of Points, where Points is a class I have implemented) and computes the variance of them, in each dimension (with regard to the algorithm of Knuth).
Here is how I call the function:
float avg[(*points)[0].dim()];
float var[(*points)[0].dim()];
size_t split_dims[t];
compute_variances(t, *points, avg, var, split_dims);
The question is, can I do better? I would really happy to pay the trade-off between speed and approximate computation of variances. Or maybe I could make the code more cache friendly or something?
I compiled like this:
g++ main_noTime.cpp -std=c++0x -p -pg -O3 -o eg
Notice, that before edit, I had used -o3, not with a capital 'o'. Thanks to ypnos, I compiled now with the optimization flag -O3. I am sure that there was a difference between them, since I performed time measurements with one of these methods in my pseudo-site.
Note that now, compute_variances is dominating the overall project's time!
[EDIT]
copute_variances() is called 40 times.
Per 10 calls, the following hold true:
points.size() = 1000 and points[0].dim = 10000
points.size() = 10000 and points[0].dim = 100
points.size() = 10000 and points[0].dim = 10000
points.size() = 100000 and points[0].dim = 100
Each call handles different data.
Q: How fast is access to points[i][d]?
A: point[i] is just the i-th element of std::vector, where the second [], is implemented as this, in the Point class.
const FT& operator [](const int i) const {
if (i < (int) coords.size() && i >= 0)
return coords.at(i);
else {
std::cout << "Error at Point::[]" << std::endl;
exit(1);
}
return coords[0]; // Clear -Wall warning
}
where coords is a std::vector of float values. This seems a bit heavy, but shouldn't the compiler be smart enough to predict correctly that the branch is always true? (I mean after the cold start). Moreover, the std::vector.at() is supposed to be constant time (as said in the ref). I changed this to have only .at() in the body of the function and the time measurements remained, pretty much, the same.
The division in the compute_variances() is for sure something heavy! However, Knuth's algorithm was a numerical stable one and I was not able to find another algorithm, that would de both numerical stable and without division.
Note that I am not interesting in parallelism right now.
[EDIT.2]
Minimal example of Point class (I think I didn't forget to show something):
class Point {
public:
typedef float FT;
...
/**
* Get dimension of point.
*/
size_t dim() const {
return coords.size();
}
/**
* Operator that returns the coordinate at the given index.
* #param i - index of the coordinate
* #return the coordinate at index i
*/
FT& operator [](const int i) {
return coords.at(i);
//it's the same if I have the commented code below
/*if (i < (int) coords.size() && i >= 0)
return coords.at(i);
else {
std::cout << "Error at Point::[]" << std::endl;
exit(1);
}
return coords[0]; // Clear -Wall warning*/
}
/**
* Operator that returns the coordinate at the given index. (constant)
* #param i - index of the coordinate
* #return the coordinate at index i
*/
const FT& operator [](const int i) const {
return coords.at(i);
/*if (i < (int) coords.size() && i >= 0)
return coords.at(i);
else {
std::cout << "Error at Point::[]" << std::endl;
exit(1);
}
return coords[0]; // Clear -Wall warning*/
}
private:
std::vector<FT> coords;
};
1. SIMD
One easy speedup for this is to use vector instructions (SIMD) for the computation. On x86 that means SSE, AVX instructions. Based on your word length and processor you can get speedups of about x4 or even more. This code here:
for (size_t d = 0; d < points[0].dim(); ++d) {
delta = (points[i][d]) - avg[d];
avg[d] += delta / n;
var[d] += delta * ((points[i][d]) - avg[d]);
}
can be sped-up by doing the computation for four elements at once with SSE. As your code really only processes one single element in each loop iteration, there is no bottleneck. If you go down to 16bit short instead of 32bit float (an approximation then), you can fit eight elements in one instruction. With AVX it would be even more, but you need a recent processor for that.
It is not the solution to your performance problem, but just one of them that can also be combined with others.
2. Micro-parallelizm
The second easy speedup when you have that many loops is to use parallel processing. I typically use Intel TBB, others might suggest OpenMP instead. For this you would probably have to change the loop order. So parallelize over d in the outer loop, not over i.
You can combine both techniques, and if you do it right, on a quadcore with HT you might get a speed-up of 25-30 for the combination without any loss in accuracy.
3. Compiler optimization
First of all maybe it is just a typo here on SO, but it needs to be -O3, not -o3!
As a general note, it might be easier for the compiler to optimize your code if you declare the variables delta, n within the scope where you actually use them. You should also try the -funroll-loops compiler option as well as -march. The option to the latter depends on your CPU, but nowadays typically -march core2 is fine (also for recent AMDs), and includes SSE optimizations (but I would not trust the compiler just yet to do that for your loop).
The big problem with your data structure is that it's essentially a vector<vector<float> >. That's a pointer to an array of pointers to arrays of float with some bells and whistles attached. In particular, accessing consecutive Points in the vector doesn't correspond to accessing consecutive memory locations. I bet you see tons and tons of cache misses when you profile this code.
Fix this before horsing around with anything else.
Lower-order concerns include the floating-point division in the inner loop (compute 1/n in the outer loop instead) and the big load-store chain that is your inner loop. You can compute the means and variances of slices of your array using SIMD and combine them at the end, for instance.
The bounds-checking once per access probably doesn't help, either. Get rid of that too, or at least hoist it out of the inner loop; don't assume the compiler knows how to fix that on its own.
Here's what I would do, in guesstimated order of importance:
Return the floating-point from the Point::operator[] by value, not by reference.
Use coords[i] instead of coords.at(i), since you already assert that it's within bounds. The at member checks the bounds. You only need to check it once.
Replace the home-baked error indication/checking in the Point::operator[] with an assert. That's what asserts are for. They are nominally no-ops in release mode - I doubt that you need to check it in release code.
Replace the repeated division with a single division and repeated multiplication.
Remove the need for wasted initialization by unrolling the first two iterations of the outer loop.
To lessen impact of cache misses, run the inner loop alternatively forwards then backwards. This at least gives you a chance at using some cached avg and var. It may in fact remove all cache misses on avg and var if prefetch works on reverse order of iteration, as it well should.
On modern C++ compilers, the std::fill and std::copy can leverage type alignment and have a chance at being faster than the C library memset and memcpy.
The Point::operator[] will have a chance of getting inlined in the release build and can reduce to two machine instructions (effective address computation and floating point load). That's what you want. Of course it must be defined in the header file, otherwise the inlining will only be performed if you enable link-time code generation (a.k.a. LTO).
Note that the Point::operator[]'s body is only equivalent to the single-line
return coords.at(i) in a debug build. In a release build the entire body is equivalent to return coords[i], not return coords.at(i).
FT Point::operator[](int i) const {
assert(i >= 0 && i < (int)coords.size());
return coords[i];
}
const FT * Point::constData() const {
return &coords[0];
}
void compute_variances(size_t t, const std::vector<Point>& points, float* avg,
float* var, size_t* split_dims)
{
assert(points.size() > 0);
const int D = points[0].dim();
// i = 0, i_n = 1
assert(D > 0);
#if __cplusplus >= 201103L
std::copy_n(points[0].constData(), D, avg);
#else
std::copy(points[0].constData(), points[0].constData() + D, avg);
#endif
// i = 1, i_n = 0.5
if (points.size() >= 2) {
assert(points[1].dim() == D);
for (int d = D - 1; d >= 0; --d) {
float const delta = points[1][d] - avg[d];
avg[d] += delta * 0.5f;
var[d] = delta * (points[1][d] - avg[d]);
}
} else {
std::fill_n(var, D, 0.0f);
}
// i = 2, ...
for (size_t i = 2; i < points.size(); ) {
{
const float i_n = 1.0f / (1.0f + i);
assert(points[i].dim() == D);
for (int d = 0; d < D; ++d) {
float const delta = points[i][d] - avg[d];
avg[d] += delta * i_n;
var[d] += delta * (points[i][d] - avg[d]);
}
}
++ i;
if (i >= points.size()) break;
{
const float i_n = 1.0f / (1.0f + i);
assert(points[i].dim() == D);
for (int d = D - 1; d >= 0; --d) {
float const delta = points[i][d] - avg[d];
avg[d] += delta * i_n;
var[d] += delta * (points[i][d] - avg[d]);
}
}
++ i;
}
/* Find t dimensions with largest scaled variance. */
kthLargest(var, D, t, split_dims);
}
for (size_t d = 0; d < points[0].dim(); d++) {
avg[d] = 0.0;
var[d] = 0.0;
}
This code could be optimized by simply using memset. The IEEE754 representation of 0.0 in 32bits is 0x00000000. If the dimension is big, it worth it.
Something like:
memset((void*)avg, 0, points[0].dim() * sizeof(float));
In your code, you have a lot of calls to points[0].dim(). It would be better to call once at the beginning of the function and store in a variable. Likely, the compiler already does this (since you are using -O3).
The division operations are a lot more expensive (from clock-cycle POV) than other operations (addition, subtraction).
avg[d] += delta / n;
It could make sense, to try to reduce the number of divisions: use partial non-cumulative average calculation, that would result in Dim division operation for N elements (instead of N x Dim); N < points.size()
Huge speedup could be achieved, using Cuda or OpenCL, since the calculation of avg and var could be done simultaneously for each dimension (consider using a GPU).
Another optimization is cache optimization including both data cache and instruction cache.
High level optimization techniques
Data Cache optimizations
Example of data cache optimization & unrolling
for (size_t d = 0; d < points[0].dim(); d += 4)
{
// Perform loading all at once.
register const float p1 = points[i][d + 0];
register const float p2 = points[i][d + 1];
register const float p3 = points[i][d + 2];
register const float p4 = points[i][d + 3];
register const float delta1 = p1 - avg[d+0];
register const float delta2 = p2 - avg[d+1];
register const float delta3 = p3 - avg[d+2];
register const float delta4 = p4 - avg[d+3];
// Perform calculations
avg[d + 0] += delta1 / n;
var[d + 0] += delta1 * ((p1) - avg[d + 0]);
avg[d + 1] += delta2 / n;
var[d + 1] += delta2 * ((p2) - avg[d + 1]);
avg[d + 2] += delta3 / n;
var[d + 2] += delta3 * ((p3) - avg[d + 2]);
avg[d + 3] += delta4 / n;
var[d + 3] += delta4 * ((p4) - avg[d + 3]);
}
This differs from classic loop unrolling in that loading from the matrix is performed as a group at the top of the loop.
Edit 1:
A subtle data optimization is to place the avg and var into a structure. This will ensure that the two arrays are next to each other in memory, sans padding. The data fetching mechanism in processors like datums that are very close to each other. Less chance for data cache miss and better chance to load all of the data into the cache.
You could use Fixed Point math instead of floating point math as an optimization.
Optimization via Fixed Point
Processors love to manipulate integers (signed or unsigned). Floating point may take extra computing power due to the extraction of the parts, performing the math, then reassemblying the parts. One mitigation is to use Fixed Point math.
Simple Example: meters
Given the unit of meters, one could express lengths smaller than a meter by using floating point, such as 3.14159 m. However, the same length can be expressed in a unit of finer detail like millimeters, e.g. 3141.59 mm. For finer resolution, a smaller unit is chosen and the value multiplied, e.g. 3,141,590 um (micrometers). The point is choosing a small enough unit to represent the floating point accuracy as an integer.
The floating point value is converted at input into Fixed Point. All data processing occurs in Fixed Point. The Fixed Point value is convert to Floating Point before outputting.
Power of 2 Fixed Point Base
As with converting from floating point meters to fixed point millimeters, using 1000, one could use a power of 2 instead of 1000. Selecting a power of 2 allows the processor to use bit shifting instead of multiplication or division. Bit shifting by a power of 2 is usually faster than multiplication or division.
Keeping with the theme and accuracy of millimeters, we could use 1024 as the base instead of 1000. Similarly, for higher accuracy, use 65536 or 131072.
Summary
Changing the design or implementation to used Fixed Point math allows the processor to use more integral data processing instructions than floating point. Floating point operations consume more processing power than integral operations in all but specialized processors. Using powers of 2 as the base (or denominator) allows code to use bit shifting instead of multiplication or division. Division and multiplication take more operations than shifting and thus shifting is faster. So rather than optimizing code for execution (such as loop unrolling), one could try using Fixed Point notation rather than floating point.
Point 1.
You're computing the average and the variance at the same time.
Is that right?
Don't you have to calculate the average first, then once you know it, calculate the sum of squared differences from the average?
In addition to being right, it's more likely to help performance than hurt it.
Trying to do two things in one loop is not necessarily faster than two consecutive simple loops.
Point 2.
Are you aware that there is a way to calculate average and variance at the same time, like this:
double sumsq = 0, sum = 0;
for (i = 0; i < n; i++){
double xi = x[i];
sum += xi;
sumsq += xi * xi;
}
double avg = sum / n;
double avgsq = sumsq / n
double variance = avgsq - avg*avg;
Point 3.
The inner loops are doing repetitive indexing.
The compiler might be able to optimize that to something minimal, but I wouldn't bet my socks on it.
Point 4.
You're using gprof or something like it.
The only reasonably reliable number to come out of it is self-time by function.
It won't tell you very well how time is spent inside the function.
I and many others rely on this method, which takes you straight to the heart of what takes time.
I've profiled my model and it seems that this kernel accounts for about 2/3 of my total runtime. I was looking for suggestions to optimize it. The code is as follows.
__global__ void calcFlux(double* concs, double* fluxes, double* dt)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
fluxes[idx]=knowles_flux(idx, concs);
//fluxes[idx]=flux(idx, concs);
}
__device__ double knowles_flux(int r, double *conc)
{
double frag_term = 0;
double flux = 0;
if (r == ((maxlength)-1))
{
//Calculation type : "Max"
flux = -km*(r)*conc[r]+2*(ka)*conc[r-1]*conc[0];
}
else if (r > ((nc)-1))
{
//Calculation type : "F"
//arrSum3(conc, &frag_term, r+1, maxlength-1);
for (int s = r+1; s < (maxlength); s++)
{
frag_term += conc[s];
}
flux = -(km)*(r)*conc[r] + 2*(km)*frag_term - 2*(ka)*conc[r]*conc[0] + 2*(ka)*conc[r-1]*conc[0];
}
else if (r == ((nc)-1))
{
//Calculation type : "N"
//arrSum3(conc, &frag_term, r+1, maxlength-1);
for (int s = r+1; s < (maxlength); s++)
{
frag_term += conc[s];
}
flux = (kn)*pow(conc[0],(nc)) + 2*(km)*frag_term - 2*(ka)*conc[r]*conc[0];
}
else if (r < ((nc)-1))
{
//Calculation type : "O"
flux = 0;
}
return flux;
}
Just to give you an idea of why the for loop is an issue, this kernel is launched on an array of about maxlength = 9000 elements. For our purposes now, nc is in the range of 2-6. Here's an illustration of how this kernel processes the incoming array (conc). For this array, five different types of calculations need to be applied to different groups of elements.
Array element : 0 1 2 3 4 5 6 7 8 9 ... 8955 8956 8957 8958 8959 8960
Type of calc : M O O O O O N F F F ... F F F F F Max
The potential problems I've been trying to deal with right now are branch divergence from the quadruple if-else and the for loop.
My idea for dealing with the branch divergence is to break this kernel down into four separate device functions or kernels that treat each region separately and all launch at the same time. I'm not sure this is significantly better than just letting the branch divergence take place, which if I'm not mistaken, would cause the four calculation types to be run in serial.
To deal with the for loop, you'll notice that there's a commented out arrSum3 function, which I wrote based off my previously (and probably poorly) written parallel reduction kernel. Using it in place of the for loop drastically increased my runtime. I feel like there's a clever way to accomplish what I'm trying to do with the for loop, but I'm just not that smart and my advisor is tired of me "wasting time" thinking about it.
Appreciate any help.
EDIT
Full code is located here : https://stackoverflow.com/q/21170233/1218689
Assuming sgn() and abs() are not derived from "if"s and "else"s
__device__ double knowles_flux(int r, double *conc)
{
double frag_term = 0;
double flux = 0;
//Calculation type : "Max"
//no divergence
//should prefer 20-30 extra cycles instead of a branching.
//may not be good for CPU
fluxA = (1-abs(sgn(r-(maxlength-1)))) * (-km*(r)*conc[r]+2*(ka)*conc[r-1]*conc[0]);
//is zero if r and maxlength-1 are not equal
//always compute this in shared memory so work will be equal for all cores, no divergence
// you should divide kernel into several pieces to do a reduction
// but if you dont want that, then you can try :
for (int s = 0;s<someLimit ; s++) // all count for same number of cycles so no divergence
{
frag_term += conc[s] * ( abs(sgn( s-maxlength ))*sgn(1- sgn( s-maxlength )) )* ( sgn(1+sgn(s-(r+1))) );
}
//but you can make easier of this using "add and assign" operation
// in local memory (was it __shared in CUDA?)
// global conc[] to local concL[] memory(using all cores)(100 cycles)
// for(others from zero to upper_limit)
// if(localID==0)
// {
// frag_termL[0]+=concL[s] // local to local (10 cycles/assign.)
// frag_termL[0+others]=frag_termL[0]; // local to local (10 cycles/assign.)
// } -----> uses nearly same number of cycles but uses much less energy
//using single core (2000 instr. with single core vs 1000 instr. with 2k cores)
// in local memory, then copy it to private registers accordingly using all cores
//Calculation type : "F"
fluxB = ( abs(sgn(r-(nc-1)))*sgn(1+sgn(r-(nc-1))) )*(-(km)*(r)*conc[r] + 2*(km)*frag_term - 2*(ka)*conc[r]*conc[0] + 2*(ka)*conc[r-1]*conc[0]);
// is zero if r is not greater than (nc-1)
//Calculation type : "N"
fluxC = ( 1-abs(sgn(r-(nc-1))) )*((kn)*pow(conc[0],(nc)) + 2*(km)*frag_term - 2*(ka)*conc[r]*conc[0]);
//zero if r and nc-1 are not equal
flux=fluxA+fluxB+fluxC; //only one of these can be different than zero
flux=flux*( -sgn(r-(nc-1))*sgn(1-sgn(r-(nc-1))) )
//zero if r > (nc-1)
return flux;
}
Okay, let me open a bit:
if(a>b) x+=y;
can be taken as
if a-b is negative sgn(a-b) is -1
then adding 1 to that -1 gives zero ==> satisfies lower part of comparison(a<b)
x+= (sgn(a-b) +1) = 0 if a<b (not a>b), x unchanged
if(a-b) is zero, sgn(a-b) is zero
then we should multiply the upper solution with sgn(a-b) too!
x+= y*(sgn(a-b) +1)*sgn(a-b)
means
x+= y*( 0 + 1) * 0 = 0 a==b is satisfied too!
lets check what happens if a>b
x+= y*(sgn(a-b) +1)*sgn(a-b)
x+= y*(1 +1)*1 ==> y*2 is not acceptable, needs another sgn on outherside
x+= y* sgn((sgn(a-b)+1)*sgn(a-b))
x+= y* sgn((1+1)*1)
x+= y* sgn(2)
x+= y only when a is greater than b
when there are too many
abs(sgn(r-(nc-1))
then you can re-use it as
tmp=abs(sgn(r-(nc-1))
..... *tmp*(tmp-1) ....
...... +tmp*zxc[s] .....
...... ......
to decrease total cycles even more! Register accessing can be in the level of terabytes/s so shouldnt be a problem. Just as doing that for global access:
tmpGlobal= conc[r];
...... tmpGlobal * tmp .....
.... tmpGlobal +x -y ....
all private registers doing stuff in terabytes per second.
Warning: reading from conc[-1] shouldnt cause any faults as long as it is multiplied by zero if the real address of conc[0] is not real zero already . But writing is hazardous.
if you need to escape from conc[-1] anyway, you can multiply the index with some absolut-ified value too! See:
tmp=conc[i-1] becomes tmp=conc[abs((i-1))] will always read from positive index, the value will be multiplied by zero later anyway. This was lower bound protection.
You can apply a higher bound protection too. Just this adds even more cycles.
Think about using vector-shuffle operations if working on a pure scalar values is not fast enough when accessing conc[r-1] and conc[r+1]. Shuffle operation between a vector's elements is faster than copying it through local mem to another core/thread.