I have sucessfully implemented a single threaded program in CUDA for Gaussian elimination and would like to achieve parallelism. Up to this point the parallel code looks like:
__global__ void ParallelGaussian(double* A)
{
int index = threadIdx.x;
int stride = blockDim.x;
if (index < ROWS) //Skip additional threads
{
for (unsigned int r = index; r < ROWS; r += stride)
{
//Forward elimination to reduce to row echelon form
for (unsigned int k = r + 1; k < ROWS; ++k)
{
double c = -A[(ROWS + 1) * k + r] / A[(ROWS + 1) * r + r];
for (unsigned int j = r; j < ROWS + 1; ++j)
{
if (r == j)
A[(ROWS + 1) * k + j] = 0.0;
else
A[(ROWS + 1) * k + j] += c * A[(ROWS + 1) * r + j];
}
}
}
}
}
As we can see the code on the GPU will transform the 1D-array (matrix) to a lower triangular matrix and then on the CPU I will continue with back substitution to get the final result. There is no pivoting done in this approach as it is not entirely needed but indeed improves the numerical stability of the algorithm.
Launching the kernel with a single thread and a block works and transforms the matrix into row echelon form:
ParallelGaussian << < 1, 1 >> >(dev_a);
However, if I would like to increase the number of threads, like
ParallelGaussian << < 1, 32 >> >(dev_a);
it will fail to produce the lower triangular matrix. Now adding __syncthreads() calls into the code in order to synchronize the threads in a block doesn't improve the situation what so ever and I can't figure out why.
Consider your inner loop. Every thread accesses A, and since k and j run from r to the end of the matrix, there is the potential for multiple threads to modify the same A[(ROWS + 1) * k + j] value.
You also potentially have some threads accessing A[(ROWS + 1) * r + j] while other threads are updating that value.
One possible solution is to have each thread accumulate into individual result arrays, then combine those at the end. This is memory intensive.
Another would be to restructure this so that only one thread will write to a particular value, and storing those values in a new matrix (so that you don't change any value that might be needed by a different thread).
Related
I have spent a lot of time to learn about implementing/visualizing dynamic programming problems using iteration but I find it very hard to understand, I can implement the same using recursion with memoization but it is slow when compared to iteration.
Can someone explain the same by a example of a hard problem or by using some basic concepts. Like the matrix chain multiplication, longest palindromic sub sequence and others. I can understand the recursion process and then memoize the overlapping sub problems for efficiency but I can't understand how to do the same using iteration.
Thanks!
Dynamic programming is all about solving the sub-problems in order to solve the bigger one. The difference between the recursive approach and the iterative approach is that the former is top-down, and the latter is bottom-up. In other words, using recursion, you start from the big problem you are trying to solve and chop it down to a bit smaller sub-problems, on which you repeat the process until you reach the sub-problem so small you can solve. This has an advantage that you only have to solve the sub-problems that are absolutely needed and using memoization to remember the results as you go. The bottom-up approach first solves all the sub-problems, using tabulation to remember the results. If we are not doing extra work of solving the sub-problems that are not needed, this is a better approach.
For a simpler example, let's look at the Fibonacci sequence. Say we'd like to compute F(101). When doing it recursively, we will start with our big problem - F(101). For that, we notice that we need to compute F(99) and F(100). Then, for F(99) we need F(97) and F(98). We continue until we reach the smallest solvable sub-problem, which is F(1), and memoize the results. When doing it iteratively, we start from the smallest sub-problem, F(1) and continue all the way up, keeping the results in a table (so essentially it's just a simple for loop from 1 to 101 in this case).
Let's take a look at the matrix chain multiplication problem, which you requested. We'll start with a naive recursive implementation, then recursive DP, and finally iterative DP. It's going to be implemented in a C/C++ soup, but you should be able to follow along even if you are not very familiar with them.
/* Solve the problem recursively (naive)
p - matrix dimensions
n - size of p
i..j - state (sub-problem): range of parenthesis */
int solve_rn(int p[], int n, int i, int j) {
// A matrix multiplied by itself needs no operations
if (i == j) return 0;
// A minimal solution for this sub-problem, we
// initialize it with the maximal possible value
int min = std::numeric_limits<int>::max();
// Recursively solve all the sub-problems
for (int k = i; k < j; ++k) {
int tmp = solve_rn(p, n, i, k) + solve_rn(p, n, k + 1, j) + p[i - 1] * p[k] * p[j];
if (tmp < min) min = tmp;
}
// Return solution for this sub-problem
return min;
}
To compute the result, we starts with the big problem:
solve_rn(p, n, 1, n - 1)
The key of DP is to remember all the solutions to the sub-problems instead of forgetting them, so we don't need to recompute them. It's trivial to make a few adjustments to the above code in order to achieve that:
/* Solve the problem recursively (DP)
p - matrix dimensions
n - size of p
i..j - state (sub-problem): range of parenthesis */
int solve_r(int p[], int n, int i, int j) {
/* We need to remember the results for state i..j.
This can be done in a matrix, which we call dp,
such that dp[i][j] is the best solution for the
state i..j. We initialize everything to 0 first.
static keyword here is just a C/C++ thing for keeping
the matrix between function calls, you can also either
make it global or pass it as a parameter each time.
MAXN is here too because the array size when doing it like
this has to be a constant in C/C++. I set it to 100 here.
But you can do it some other way if you don't like it. */
static int dp[MAXN][MAXN] = {{0}};
/* A matrix multiplied by itself has 0 operations, so we
can just return 0. Also, if we already computed the result
for this state, just return that. */
if (i == j) return 0;
else if (dp[i][j] != 0) return dp[i][j];
// A minimal solution for this sub-problem, we
// initialize it with the maximal possible value
dp[i][j] = std::numeric_limits<int>::max();
// Recursively solve all the sub-problems
for (int k = i; k < j; ++k) {
int tmp = solve_r(p, n, i, k) + solve_r(p, n, k + 1, j) + p[i - 1] * p[k] * p[j];
if (tmp < dp[i][j]) dp[i][j] = tmp;
}
// Return solution for this sub-problem
return dp[i][j];;
}
We start with the big problem as well:
solve_r(p, n, 1, n - 1)
Iterative solution is only to, well, iterate all the states, instead of starting from the top:
/* Solve the problem iteratively
p - matrix dimensions
n - size of p
We don't need to pass state, because we iterate the states. */
int solve_i(int p[], int n) {
// But we do need our table, just like before
static int dp[MAXN][MAXN];
// Multiplying a matrix by itself needs no operations
for (int i = 1; i < n; ++i)
dp[i][i] = 0;
// L represents the length of the chain. We go from smallest, to
// biggest. Made L capital to distinguish letter l from number 1
for (int L = 2; L < n; ++L) {
// This double loop goes through all the states in the current
// chain length.
for (int i = 1; i <= n - L + 1; ++i) {
int j = i + L - 1;
dp[i][j] = std::numeric_limits<int>::max();
for (int k = i; k <= j - 1; ++k) {
int tmp = dp[i][k] + dp[k+1][j] + p[i-1] * p[k] * p[j];
if (tmp < dp[i][j])
dp[i][j] = tmp;
}
}
}
// Return the result of the biggest problem
return dp[1][n-1];
}
To compute the result, just call it:
solve_i(p, n)
Explanation of the loop counters in the last example:
Let's say we need to optimize the multiplication of 4 matrices: A B C D. We are doing an iterative approach, so we will first compute the chains with the length of two: (A B) C D, A (B C) D, and A B (C D). And then chains of three: (A B C) D, and A (B C D). That is what L, i and j are for.
L represents the chain length, it goes from 2 to n - 1 (n is 4 in this case, so that is 3).
i and j represent the starting and ending position of the chain. In case L = 2, i goes from 1 to 3, and j goes from 2 to 4:
(A B) C D A (B C) D A B (C D)
^ ^ ^ ^ ^ ^
i j i j i j
In case L = 3, i goes from 1 to 2, and j goes from 3 to 4:
(A B C) D A (B C D)
^ ^ ^ ^
i j i j
So generally, i goes from 1 to n - L + 1, and j is i + L - 1.
Now, let's continue with the algorithm assuming that we are at the step where we have (A B C) D. We now need to take into account the sub-problems (which are already calculated): ((A B) C) D and (A (B C)) D. That is what k is for. It goes through all the positions between i and j and computes the sub problems.
I hope I helped.
The problem with recursion is the high number of stack frames that need to be pushed/popped. This can quickly become the bottle-neck.
The Fibonacci Series can be calculated with iterative DP or recursion with memoization. If we calculate F(100) in DP all we need is an array of length 100 e.g. int[100] and that's the guts of our used memory. We calculate all entries of the array pre-filling f[0] and f[1] as they are defined to be 1. and each value just depends on the previous two.
If we use a recursive solution we start at fib(100) and work down. Every method call from 100 down to 0 is pushed onto the stack, AND checked if it's memoized. These operations add up and iteration doesn't suffer from either of these. In iteration (bottom-up) we already know all of the previous answers are valid. The bigger impact is probably the stack frames; and given a larger input you may get a StackOverflowException for what was otherwise trivial with an iterative DP approach.
If multiple threads are simultaneously writing a single memory location.,there will be a race condition,right??
In my case same is happening..
Consider a module from 'reduce.cl'
int i = get_global_id(0);
int n,j;
n = keyMobj[i]; // this n is the key..It can be either 0 or 1.
for(j=0; j<2; j++)
sumMobj[n*2+j] += dataMobj[i].dattr[j]; //summing operation.
Here, The memory locations
sumMobj===> [...0..., ....1...] is accessed 4 threads simultaneously &
sumMobj===> [....3..., ....4...] is accessed 6 threads simultaneously..
Is there any way to still make it parallely,like using locking or semaphore? As this summing is a very big part in my algorithm...
I can give you some hint as I was also facing similar problem.
I can think of three different methods for achieving similar goal:
Consider a simple kernel, assuming you launched 4 (0-3) threads
_kernel void addition (int *p)
{
int i = get_local_id(0);
p[4]+= p[i];
}
You want to add values p[0], p[1], p[2], p[3], p[4], and store the final sum in p[4]. right? i.e:
p[4]= p[0] + p[1] + p[2] + p[3] + p[4]
Method -1 (no parallelism)
Assign this job to only 1 thread (no parallelism):
int i = get_local_id(0);
if (i==0)
{
p[4]+= p[i];
}
Method-2 (with parallelism)
Express your problem as follows:
p[4]= p[0] + p[1] + p[2] + p[3] + p[4] + 0
This is a reduction problem
So launch 3 threads: i=0 to i=2. In first iteration
i=0 finds p[0] + p[1]
i=1 finds p[2] + p[3]
i=2 finds p[4] + 0
Now you have three numbers, you apply the same logic as above and add these numbers (with suitable padding of 0 to make it in power of two)
Method -3 Atomic operations
If you still need to implement this atomically, you can use atomic_add():
int fsfunc atomic_add ( volatile __global int *p ,int val)
Description
Read the 32-bit value (referred to as old) stored at location pointed
by p. Compute (old + val) and store result at location pointed by p.
The function returns old.
This is assuming the data is int type. Otherwise you can see the link as suggested above.
I am trying to run the following piece of LUDecomposition code in an OpenCL kernel.
'A' below is a single precision floating point array.
for( k = 0; k < N; k++ )
{
for(j=k+1; j < N; j++) {
A[k * N + j] = A[k * N + j] / A[k * N + k];
}
for(i=k+1; i < N; i++) {
for (j=k+1; j < N; j++) {
A[i * N + j] = A[i * N + j] - (A[i * N + k] * A[k * N + j]);
}
}
}
I am running this code on the GPU on just a single GPU thread (completely sequential). So I have the global thread and local thread mapping for the kernel as follows.
globalthread[0] = 1;
globalthread[1] = 1;
localthread[0] = 1;
localthread[1] = 1;
But when I compare the GPU output to the output of the same function run on the CPU
(directly and not as an opencl device) I am seeing that the outputs dont match. (the mismatch is beyond the limits of floating points accuracy).
I found this unexplainable inspite of best efforts. While trying to narrow down the problem,
I found that the problem arises from the second statement. Specifically due to the subtraction operation and when the value of A[i][j] goes negative.
I have made sure that both CPU and GPU are working on the same inputs. But such a strange behavior for such a simple computation seems weird. Can anyone help explain why the outputs might be differing?
I also ran it on both NVIDIA and AMD Devices and I see the same behavior. (to rule out any platform specific issue)
Here is the sample output of the mismatch:
platform name is NVIDIA CUDA
platform version is OpenCL 1.1 CUDA 4.2.1
number of devices is 2
device name is Tesla C2050 / C2070
GPU Runtime: 0.023669s
CPU Runtime: 0.000123s
Values differ at index (45, 40): arr1=0.946256, arr2=0.963078
Values differ at index (60, 52): arr1=-9.348129, arr2=-9.483719
Values differ at index (61, 52): arr1=11.343384, arr2=11.093756
Non-Matching CPU-GPU Outputs Beyond Error Threshold of 1.05 Percent: 3
I tried to implement bessel function using that formula, this is the code:
function result=Bessel(num);
if num==0
result=bessel(0,1);
elseif num==1
result=bessel(1,1);
else
result=2*(num-1)*Bessel(num-1)-Bessel(num-2);
end;
But if I use MATLAB's bessel function to compare it with this one, I get too high different values.
For example if I type Bessel(20) it gives me 3.1689e+005 as result, if instead I type bessel(20,1) it gives me 3.8735e-025 , a totally different result.
such recurrence relations are nice in mathematics but numerically unstable when implementing algorithms using limited precision representations of floating-point numbers.
Consider the following comparison:
x = 0:20;
y1 = arrayfun(#(n)besselj(n,1), x); %# builtin function
y2 = arrayfun(#Bessel, x); %# your function
semilogy(x,y1, x,y2), grid on
legend('besselj','Bessel')
title('J_\nu(z)'), xlabel('\nu'), ylabel('log scale')
So you can see how the computed values start to differ significantly after 9.
According to MATLAB:
BESSELJ uses a MEX interface to a Fortran library by D. E. Amos.
and gives the following as references for their implementation:
D. E. Amos, "A subroutine package for Bessel functions of a complex
argument and nonnegative order", Sandia National Laboratory Report,
SAND85-1018, May, 1985.
D. E. Amos, "A portable package for Bessel functions of a complex
argument and nonnegative order", Trans. Math. Software, 1986.
The forward recurrence relation you are using is not stable. To see why, consider that the values of BesselJ(n,x) become smaller and smaller by about a factor 1/2n. You can see this by looking at the first term of the Taylor series for J.
So, what you're doing is subtracting a large number from a multiple of a somewhat smaller number to get an even smaller number. Numerically, that's not going to work well.
Look at it this way. We know the result is of the order of 10^-25. You start out with numbers that are of the order of 1. So in order to get even one accurate digit out of this, we have to know the first two numbers with at least 25 digits precision. We clearly don't, and the recurrence actually diverges.
Using the same recurrence relation to go backwards, from high orders to low orders, is stable. When you start with correct values for J(20,1) and J(19,1), you can calculate all orders down to 0 with full accuracy as well. Why does this work? Because now the numbers are getting larger in each step. You're subtracting a very small number from an exact multiple of a larger number to get an even larger number.
You can just modify the code below which is for the Spherical bessel function. It is well tested and works for all arguments and order range. I am sorry it is in C#
public static Complex bessel(int n, Complex z)
{
if (n == 0) return sin(z) / z;
if (n == 1) return sin(z) / (z * z) - cos(z) / z;
if (n <= System.Math.Abs(z.real))
{
Complex h0 = bessel(0, z);
Complex h1 = bessel(1, z);
Complex ret = 0;
for (int i = 2; i <= n; i++)
{
ret = (2 * i - 1) / z * h1 - h0;
h0 = h1;
h1 = ret;
if (double.IsInfinity(ret.real) || double.IsInfinity(ret.imag)) return double.PositiveInfinity;
}
return ret;
}
else
{
double u = 2.0 * abs(z.real) / (2 * n + 1);
double a = 0.1;
double b = 0.175;
int v = n - (int)System.Math.Ceiling((System.Math.Log(0.5e-16 * (a + b * u * (2 - System.Math.Pow(u, 2)) / (1 - System.Math.Pow(u, 2))), 2)));
Complex ret = 0;
while (v > n - 1)
{
ret = z / (2 * v + 1.0 - z * ret);
v = v - 1;
}
Complex jnM1 = ret;
while (v > 0)
{
ret = z / (2 * v + 1.0 - z * ret);
jnM1 = jnM1 * ret;
v = v - 1;
}
return jnM1 * sin(z) / z;
}
}
I have 2 tables of values and want to scale the first one so that it matches the 2nd one as good as possible. Both have the same length. If both are drawn as graphs in a diagram they should be as close to each other as possible. But I do not want quadratic, but simple linear weights.
My problem is, that I have no idea how to actually compute the best scaling factor because of the Abs function.
Some pseudocode:
//given:
float[] table1= ...;
float[] table2= ...;
//wanted:
float factor= ???; // I have no idea how to compute this
float remainingDifference=0;
for(int i=0; i<length; i++)
{
float scaledValue=table1[i] * factor;
//Sum up the differences. I use the Abs function because negative differences are differences too.
remainingDifference += Abs(scaledValue - table2[i]);
}
I want to compute the scaling factor so that the remainingDifference is minimal.
Simple linear weights is hard like you said.
a_n = first sequence
b_n = second sequence
c = scaling factor
Your residual function is (sums are from i=1 to N, the number of points):
SUM( |a_i - c*b_i| )
Taking the derivative with respect to c yields:
d/dc SUM( |a_i - c*b_i| )
= SUM( b_i * (a_i - c*b_i)/|a_i - c*b_i| )
Setting to 0 and solving for c is hard. I don't think there's an analytic way of doing that. You may want to try https://math.stackexchange.com/ to see if they have any bright ideas.
However if you work with quadratic weights, it becomes significantly simpler:
d/dc SUM( (a_i - c*b_i)^2 )
= SUM( 2*(a_i - c*b_i)* -c )
= -2c * SUM( a_i - c*b_i ) = 0
=> SUM(a_i) - c*SUM(b_i) = 0
=> c = SUM(a_i) / SUM(b_i)
I strongly suggest the latter approach if you can.
I would suggest trying some sort of variant on Newton Raphson.
Construct a function Diff(k) that looks at the difference in area between your two graphs between fixed markers A and B.
mathematically I guess it would be integral ( x = A to B ){ f(x) - k * g(x) }dx
anyway realistically you could just subtract the values,
like if you range from X = -10 to 10, and you have a data point for f(i) and g(i) on each integer i in [-10, 10], (ie 21 datapoints )
then you just sum( i = -10 to 10 ){ f(i) - k * g(i) }
basically you would expect this function to look like a parabola -- there will be an optimum k, and deviating slightly from it in either direction will increase the overall area difference
and the bigger the difference, you would expect the bigger the gap
so, this should be a pretty smooth function ( if you have a lot of data points )
so you want to minimise Diff(k)
so you want to find whether derivative ie d/dk Diff(k) = 0
so just do Newton Raphson on this new function D'(k)
kick it off at k=1 and it should zone in on a solution pretty fast
that's probably going to give you an optimal computation time
if you want something simpler, just start with some k1 and k2 that are either side of 0
so say Diff(1.5) = -3 and Diff(2.9) = 7
so then you would pick a k say 3/10 of the way (10 = 7 - -3) between 1.5 and 2.9
and depending on whether that yields a positive or negative value, use it as the new k1 or k2, rinse and repeat
In case anyone stumbles upon this in the future, here is some code (c++)
The trick is to first sort the samples by the scaling factor that would result in the best fit for the 2 samples each. Then start at both ends iterate to the factor that results in the minimum absolute deviation (L1-norm).
Everything except for the sort has a linear run time => Runtime is O(n*log n)
/*
* Find x so that the sum over std::abs(pA[i]-pB[i]*x) from i=0 to (n-1) is minimal
* Then return x
*/
float linearFit(const float* pA, const float* pB, int n)
{
/*
* Algebraic solution is not possible for the general case
* => iterative algorithm
*/
if (n < 0)
throw "linearFit has invalid argument: expected n >= 0";
if (n == 0)
return 0;//If there is nothing to fit, any factor is a perfect fit (sum is always 0)
if (n == 1)
return pA[0] / pB[0];//return x so that pA[0] = pB[0]*x
//If you don't like this , use a std::vector :P
std::unique_ptr<float[]> targetValues_(new float[n]);
std::unique_ptr<int[]> indices_(new int[n]);
//Get proper pointers:
float* targetValues = targetValues_.get();//The value for x that would cause pA[i] = pB[i]*x
int* indices = indices_.get(); //Indices of useful (not nan and not infinity) target values
//The code above guarantees n > 1, so it is safe to get these pointers:
int m = 0;//Number of useful target values
for (int i = 0; i < n; i++)
{
float a = pA[i];
float b = pB[i];
float targetValue = a / b;
targetValues[i] = targetValue;
if (std::isfinite(targetValue))
{
indices[m++] = i;
}
}
if (m <= 0)
return 0;
if (m == 1)
return targetValues[indices[0]];//If there is only one target value, then it has to be the best one.
//sort the indices by target value
std::sort(indices, indices + m, [&](int ia, int ib){
return targetValues[ia] < targetValues[ib];
});
//Start from the extremes and meet at the optimal solution somewhere in the middle:
int l = 0;
int r = m - 1;
// m >= 2 is guaranteed => l > r
float penaltyFactorL = std::abs(pB[indices[l]]);
float penaltyFactorR = std::abs(pB[indices[r]]);
while (l < r)
{
if (l == r - 1 && penaltyFactorL == penaltyFactorR)
{
break;
}
if (penaltyFactorL < penaltyFactorR)
{
l++;
if (l < r)
{
penaltyFactorL += std::abs(pB[indices[l]]);
}
}
else
{
r--;
if (l < r)
{
penaltyFactorR += std::abs(pB[indices[r]]);
}
}
}
//return the best target value
if (l == r)
return targetValues[indices[l]];
else
return (targetValues[indices[l]] + targetValues[indices[r]])*0.5;
}