Piecemeal processing of a matrix - CUDA - matrix

OK, so lets say I have an ( N x N ) matrix that I would like to process. This matrix is quite large for my computer, and if I try to send it to the device all at once I get a 'out of memory error.'
So is there a way to send sections of the matrix to the device? One way I can see to do it is copy portions of the matrix on the host, and then send these manageable copied portions from the host to the device, and then put them back together at the end.
Here is something I have tried, but the cudaMemcpy in the for loop returns error code 11, 'invalid argument.'
int h_N = 10000;
size_t h_size_m = h_N*sizeof(float);
h_A = (float*)malloc(h_size_m*h_size_m);
int d_N = 2500;
size_t d_size_m = d_N*sizeof(float);
InitializeMatrices(h_N);
int i;
int iterations = (h_N*h_N)/(d_N*d_N);
for( i = 0; i < iterations; i++ )
{
float* h_array_ref = h_A+(i*d_N*d_N);
cudasafe( cudaMemcpy(d_A, h_array_ref, d_size_m*d_size_m, cudaMemcpyHostToDevice), "cudaMemcpy");
cudasafe( cudaFree(d_A), "cudaFree(d_A)" );
}
What I'm trying to accomplish with the above code is this: instead of send the entire matrix to the device, I simply send a pointer to a place within that matrix and reserve enough space on the device to do the work, and then with the next iteration of the loop move the pointer forward within the matrix, etc. etc.

Not only can you do this (assuming your problem is easily decomposed this way into sub-arrays), it can be a very useful thing to do for performance; once you get the basic approach you've described working, you can start using asynchronous memory copies and double-buffering to overlap some of the memory transfer time with the time spent computing what is already on-card.
But first one gets the simple thing working. Below is a 1d example (multiplying a vector by a scalar and adding another scalar) but using a linearized 2d array would be the same; the key part is
CHK_CUDA( cudaMalloc(&xd, batchsize*sizeof(float)) );
CHK_CUDA( cudaMalloc(&yd, batchsize*sizeof(float)) );
tick(&gputimer);
int nbatches = 0;
for (int nstart=0; nstart < n; nstart+=batchsize) {
int size=batchsize;
if ((nstart + batchsize) > n) size = n - nstart;
CHK_CUDA( cudaMemcpy(xd, &(x[nstart]), size*sizeof(float), cudaMemcpyHostToDevice) );
blocksize = (size+nblocks-1)/nblocks;
cuda_saxpb<<<nblocks, blocksize>>>(xd, a, b, yd, size);
CHK_CUDA( cudaMemcpy(&(ycuda[nstart]), yd, size*sizeof(float), cudaMemcpyDeviceToHost) );
nbatches++;
}
gputime = tock(&gputimer);
CHK_CUDA( cudaFree(xd) );
CHK_CUDA( cudaFree(yd) );
You allocate the buffers at the start, and then loop through until you're done, each time doing the copy, starting the kernel, and then copying back. You free at the end.
The full code is
#include <stdio.h>
#include <stdlib.h>
#include <getopt.h>
#include <cuda.h>
#include <sys/time.h>
#include <math.h>
#define CHK_CUDA(e) {if (e != cudaSuccess) {fprintf(stderr,"Error: %s\n", cudaGetErrorString(e)); exit(-1);}}
__global__ void cuda_saxpb(const float *xd, const float a, const float b,
float *yd, const int n) {
int i = threadIdx.x + blockIdx.x*blockDim.x;
if (i<n) {
yd[i] = a*xd[i]+b;
}
return;
}
void cpu_saxpb(const float *x, float a, float b, float *y, int n) {
int i;
for (i=0;i<n;i++) {
y[i] = a*x[i]+b;
}
return;
}
int get_options(int argc, char **argv, int *n, int *s, int *nb, float *a, float *b);
void tick(struct timeval *timer);
double tock(struct timeval *timer);
int main(int argc, char **argv) {
int n=1000;
int nblocks=10;
int batchsize=100;
float a = 5.;
float b = -1.;
int err;
float *x, *y, *ycuda;
float *xd, *yd;
double abserr;
int blocksize;
int i;
struct timeval cputimer;
struct timeval gputimer;
double cputime, gputime;
err = get_options(argc, argv, &n, &batchsize, &nblocks, &a, &b);
if (batchsize > n) {
fprintf(stderr, "Resetting batchsize to size of vector, %d\n", n);
batchsize = n;
}
if (err) return 0;
x = (float *)malloc(n*sizeof(float));
if (!x) return 1;
y = (float *)malloc(n*sizeof(float));
if (!y) {free(x); return 1;}
ycuda = (float *)malloc(n*sizeof(float));
if (!ycuda) {free(y); free(x); return 1;}
/* run CPU code */
tick(&cputimer);
cpu_saxpb(x, a, b, y, n);
cputime = tock(&cputimer);
/* run GPU code */
/* only have to allocate once */
CHK_CUDA( cudaMalloc(&xd, batchsize*sizeof(float)) );
CHK_CUDA( cudaMalloc(&yd, batchsize*sizeof(float)) );
tick(&gputimer);
int nbatches = 0;
for (int nstart=0; nstart < n; nstart+=batchsize) {
int size=batchsize;
if ((nstart + batchsize) > n) size = n - nstart;
CHK_CUDA( cudaMemcpy(xd, &(x[nstart]), size*sizeof(float), cudaMemcpyHostToDevice) );
blocksize = (size+nblocks-1)/nblocks;
cuda_saxpb<<<nblocks, blocksize>>>(xd, a, b, yd, size);
CHK_CUDA( cudaMemcpy(&(ycuda[nstart]), yd, size*sizeof(float), cudaMemcpyDeviceToHost) );
nbatches++;
}
gputime = tock(&gputimer);
CHK_CUDA( cudaFree(xd) );
CHK_CUDA( cudaFree(yd) );
abserr = 0.;
for (i=0;i<n;i++) {
abserr += fabs(ycuda[i] - y[i]);
}
printf("Y = a*X + b, problemsize = %d\n", n);
printf("CPU time = %lg millisec.\n", cputime*1000.);
printf("GPU time = %lg millisec (done with %d batches of %d).\n",
gputime*1000., nbatches, batchsize);
printf("CUDA and CPU results differ by %lf\n", abserr);
free(x);
free(y);
free(ycuda);
return 0;
}
int get_options(int argc, char **argv, int *n, int *s, int *nb, float *a, float *b) {
const struct option long_options[] = {
{"nvals" , required_argument, 0, 'n'},
{"nblocks" , required_argument, 0, 'B'},
{"batchsize" , required_argument, 0, 's'},
{"a", required_argument, 0, 'a'},
{"b", required_argument, 0, 'b'},
{"help", no_argument, 0, 'h'},
{0, 0, 0, 0}};
char c;
int option_index;
int tempint;
while (1) {
c = getopt_long(argc, argv, "n:B:a:b:s:h", long_options, &option_index);
if (c == -1) break;
switch(c) {
case 'n': tempint = atoi(optarg);
if (tempint < 1 || tempint > 500000) {
fprintf(stderr,"%s: Cannot use number of points %s;\n Using %d\n", argv[0], optarg, *n);
} else {
*n = tempint;
}
break;
case 's': tempint = atoi(optarg);
if (tempint < 1 || tempint > 50000) {
fprintf(stderr,"%s: Cannot use number of points %s;\n Using %d\n", argv[0], optarg, *s);
} else {
*s = tempint;
}
break;
case 'B': tempint = atoi(optarg);
if (tempint < 1 || tempint > 1000 || tempint > *n) {
fprintf(stderr,"%s: Cannot use number of blocks %s;\n Using %d\n", argv[0], optarg, *nb);
} else {
*nb = tempint;
}
break;
case 'a': *a = atof(optarg);
break;
case 'b': *b = atof(optarg);
break;
case 'h':
puts("Calculates y[i] = a*x[i] + b on the GPU.");
puts("Options: ");
puts(" --nvals=N (-n N): Set the number of values in y,x.");
puts(" --batchsize=N (-s N): Set the number of values to transfer at a time.");
puts(" --nblocks=N (-B N): Set the number of blocks used.");
puts(" --a=X (-a X): Set the parameter a.");
puts(" --b=X (-b X): Set the parameter b.");
puts(" --niters=N (-I X): Set number of iterations to calculate.");
puts("");
return +1;
}
}
return 0;
}
void tick(struct timeval *timer) {
gettimeofday(timer, NULL);
}
double tock(struct timeval *timer) {
struct timeval now;
gettimeofday(&now, NULL);
return (now.tv_usec-timer->tv_usec)/1.0e6 + (now.tv_sec - timer->tv_sec);
}
Running this one gets:
$ ./batched-saxpb --nvals=10240 --batchsize=10240 --nblocks=20
Y = a*X + b, problemsize = 10240
CPU time = 0.072 millisec.
GPU time = 0.117 millisec (done with 1 batches of 10240).
CUDA and CPU results differ by 0.000000
$ ./batched-saxpb --nvals=10240 --batchsize=5120 --nblocks=20
Y = a*X + b, problemsize = 10240
CPU time = 0.066 millisec.
GPU time = 0.133 millisec (done with 2 batches of 5120).
CUDA and CPU results differ by 0.000000
$ ./batched-saxpb --nvals=10240 --batchsize=2560 --nblocks=20
Y = a*X + b, problemsize = 10240
CPU time = 0.067 millisec.
GPU time = 0.167 millisec (done with 4 batches of 2560).
CUDA and CPU results differ by 0.000000
The GPU time goes up in this case (we're doing more memory copies) but the answers stay the same.
Edited: The original version of this code had an option for running multiple iterations of the kernel for timing purposes, but that's unnecessarily confusing in this context so it's removed.

Related

Binary Matrix Reduction in CUDA

I have to traverse all cells of an imaginary matrix m * n and add + 1 for all cells that meet a certain condition.
My naive solution was as follows:
#include <stdio.h>
__global__ void calculate_pi(int center, int *count) {
int x = threadIdx.x;
int y = blockIdx.x;
if (x*x + y*y <= center*center) {
*count++;
}
}
int main() {
int interactions;
printf("Enter the number of interactions: ");
scanf("%d", &interactions);
int l = sqrt(interactions);
int h_count = 0;
int *d_count;
cudaMalloc(&d_count, sizeof(int));
cudaMemcpy(&d_count, &h_count, sizeof(int), cudaMemcpyHostToDevice);
calculate_pi<<<l,l>>>(l/2, d_count);
cudaMemcpy(&h_count, d_count, sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(d_count);
printf("Sum: %d\n", h_count);
return 0;
}
In my use case, the value of interactions can be very large, making it impossible to allocate l * l of space.
Can someone help me? Any suggestions are welcome.
There are at least 2 problems with your code:
Your kernel code will not work correctly with an ordinary add here:
*count++;
this is because multiple threads are trying to do this at the same time, and CUDA does not automatically sort that out for you. For the purpose of this explanation, we will fix this with an atomicAdd(), although other methods are possible.
The ampersand doesn't belong here:
cudaMemcpy(&d_count, &h_count, sizeof(int), cudaMemcpyHostToDevice);
^
I assume that is just a typo, since you did it correctly on the subsequent cudaMemcpy operation:
cudaMemcpy(&h_count, d_count, sizeof(int), cudaMemcpyDeviceToHost);
This methodology (effectively creating a square array of threads using threadIdx.x for one dimension and blockIdx.x for the other) will only work up to an interactions value that leads to an l value of 1024, or less, because CUDA threadblocks are limited to 1024 threads, and you are using l as the size of the threadblock in your kernel launch. To fix this you would want to learn how to create a CUDA 2D grid of arbitrary dimensions, and adjust your kernel launch and in-kernel indexing calculations appropriately. For now we will just make sure that the calculated l value is in range for your code design.
Here's an example addressing the above issues:
$ cat t1590.cu
#include <stdio.h>
__global__ void calculate_pi(int center, int *count) {
int x = threadIdx.x;
int y = blockIdx.x;
if (x*x + y*y <= center*center) {
atomicAdd(count, 1);
}
}
int main() {
int interactions;
printf("Enter the number of interactions: ");
scanf("%d", &interactions);
int l = sqrt(interactions);
if ((l > 1024) || (l < 1)) {printf("Error: interactions out of range\n"); return 0;}
int h_count = 0;
int *d_count;
cudaMalloc(&d_count, sizeof(int));
cudaMemcpy(d_count, &h_count, sizeof(int), cudaMemcpyHostToDevice);
calculate_pi<<<l,l>>>(l/2, d_count);
cudaMemcpy(&h_count, d_count, sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(d_count);
cudaError_t err = cudaGetLastError();
if (err == cudaSuccess){
printf("Sum: %d\n", h_count);
printf("fraction satisfying test: %f\n", h_count/(float)interactions);
}
else
printf("CUDA error: %s\n", cudaGetErrorString(err));
return 0;
}
$ nvcc -o t1590 t1590.cu
$ ./t1590
Enter the number of interactions: 1048576
Sum: 206381
fraction satisfying test: 0.196820
$
We see that the code indicates a calculated fraction of about 0.2. Does this appear to be correct? I claim that it does appear to be correct based on your test. You are effectively creating a grid that represents dimensions of lxl. Your test is asking, effectively, "which points in that grid are within a circle, with the center at the origin (corner) of the grid, and radius l/2 ?"
Pictorially, that looks something like this:
and it is reasonable to assume the red shaded area is somewhat less than 0.25 of the total area, so 0.2 is a reasonable estimate of that area.
As a bonus, here is a version of the code that reduces the restriction listed in item 3 above:
#include <stdio.h>
__global__ void calculate_pi(int center, int *count) {
int x = threadIdx.x+blockDim.x*blockIdx.x;
int y = threadIdx.y+blockDim.y*blockIdx.y;
if (x*x + y*y <= center*center) {
atomicAdd(count, 1);
}
}
int main() {
int interactions;
printf("Enter the number of interactions: ");
scanf("%d", &interactions);
int l = sqrt(interactions);
int h_count = 0;
int *d_count;
const int bs = 32;
dim3 threads(bs, bs);
dim3 blocks((l+threads.x-1)/threads.x, (l+threads.y-1)/threads.y);
cudaMalloc(&d_count, sizeof(int));
cudaMemcpy(d_count, &h_count, sizeof(int), cudaMemcpyHostToDevice);
calculate_pi<<<blocks,threads>>>(l/2, d_count);
cudaMemcpy(&h_count, d_count, sizeof(int), cudaMemcpyDeviceToHost);
cudaFree(d_count);
cudaError_t err = cudaGetLastError();
if (err == cudaSuccess){
printf("Sum: %d\n", h_count);
printf("fraction satisfying test: %f\n", h_count/(float)interactions);
}
else
printf("CUDA error: %s\n", cudaGetErrorString(err));
return 0;
}
This is launching a 2D grid based on l, and should work up to at least 1 billion interactions .

Dealing with matrices in CUDA: understanding basic concepts

I'm building a CUDA kernel to compute the numerical N*N jacobian of a function, using finite differences; in the example I provided, it is the square function (each entry of the vector is squared). The host coded allocates in linear memory, while I'm using a 2-dimensional indexing in the kernel.
My issue is that I haven't found a way to sum on the diagonal of the matrices cudaMalloc'ed. My attempt has been to use the statement threadIdx.x == blockIdx.x as a condition for the diagonal, but instead it evaluates to true only for them both at 0.
Here is the kernel and EDIT: I posted the whole code as an answer, based on the suggestions in the comments (the main() is basically the same, while the kernel is not)
template <typename T>
__global__ void jacobian_kernel (
T * J,
const T t0,
const T tn,
const T h,
const T * u0,
const T * un,
const T * un_old)
{
T cgamma = 2 - sqrtf(2);
const unsigned int t = threadIdx.x;
const unsigned int b = blockIdx.x;
const unsigned int tid = t + b * blockDim.x;
/*__shared__*/ T temp_sx[BLOCK_SIZE][BLOCK_SIZE];
/*__shared__*/ T temp_dx[BLOCK_SIZE][BLOCK_SIZE];
__shared__ T sm_temp_du[BLOCK_SIZE];
T* temp_du = &sm_temp_du[0];
if (tid < N )
{
temp_sx[b][t] = un[t];
temp_dx[b][t] = un[t];
if ( t == b )
{
if ( tn == t0 )
{
temp_du[t] = u0[t]*0.001;
temp_sx[b][t] += temp_du[t]; //(*)
temp_dx[b][t] -= temp_du[t];
temp_sx[b][t] += ( abs( temp_sx[b][t] ) < 10e-6 ? 0.1 : 0 );
temp_dx[b][t] += ( abs( temp_dx[b][t] ) < 10e-6 ? 0.1 : 0 );
temp_sx[b][t] = ( temp_sx[b][t] == 0 ? 0.1 : temp_sx[b][t] );
temp_dx[b][t] = ( temp_dx[b][t] == 0 ? 0.1 : temp_dx[b][t] );
}
else
{
temp_du[t] = MAX( un[t] - un_old[t], 10e-6 );
temp_sx[b][t] += temp_du[t];
temp_dx[b][t] -= temp_du[t];
}
}
__syncthreads();
//J = f(tn, un + du)
d_func(tn, (temp_sx[b]), (temp_sx[b]), 1.f);
d_func(tn, (temp_dx[b]), (temp_dx[b]), 1.f);
__syncthreads();
J[tid] = (temp_sx[b][t] - temp_dx[b][t]) * powf((2 * temp_du[t]), -1);
//J[tid]*= - h*cgamma/2;
//J[tid]+= ( t == b ? 1 : 0);
//J[tid] = temp_J[tid];
}
}
The general procedure for computing the jacobian is
Copy un into every row of temp_sx and temp_dx
Compute du as a 0.01 magnitude from u0
Sum du to the diagonal of temp_sx, subtract du from the diagonal of temp_dx
Compute the square function on each entry of temp_sx and temp_dx
Subtract them and divide every entry by 2*du
This procedure can be summarized with (f(un + du*e_i) - f(un - du*e_i))/2*du.
My problem is to sum du to the diagonal of the matrices of temp_sx and temp_dx like I tried in (*). How can I achieve that?
EDIT: Now calling 1D blocks and threads; in fact, .y axis wasn't used at all in the kernel. I'm calling the kernel with a fixed amount of shared memory
Note that in int main() I'm calling the kernel with
#define REAL sizeof(float)
#define N 32
#define BLOCK_SIZE 16
#define NUM_BLOCKS ((N*N + BLOCK_SIZE - 1)/ BLOCK_SIZE)
...
dim3 dimGrid(NUM_BLOCKS,);
dim3 dimBlock(BLOCK_SIZE);
size_t shm_size = N*N*REAL;
jacobian_kernel <<< dimGrid, dimBlock, size_t shm_size >>> (...);
So that I attempt to deal with block-splitting the function calls. In the kernel to sum on the diagonal I used if(threadIdx.x == blockIdx.x){...}. Why isn't this correct? I'm asking it because while debugging and making the code print the statement, It only evaluates true if they both are 0. Thus du[0] is the only numerical value and the matrix becomes nan. Note that this approach worked with the first code I built, where instead I called the kernel with
jacobian_kernel <<< N, N >>> (...)
So that when threadIdx.x == blockIdx.x the element is on the diagonal. This approach doesn't fit anymore though, since now I need to deal with larger N (possibly larger than 1024, which is the maximum number of threads per block).
What statement should I put there that works even if the matrices are split into blocks and threads?
Let me know if I should share some other info.
Here is how I managed to solve my problem, based on the suggestion in the comments on the answer. The example is compilable, provided you put helper_cuda.h and helper_string.h in the same directory or you add -I directive to the CUDA examples include path, installed along with the CUDA toolkit. The relevant changes are only in the kernel; there's a minor change in the main() though, since I was calling double the resources to execute the kernel, but the .y axis of the grid of thread blocks wasn't even used at all, so it didn't generate any error.
#include <stdio.h>
#include <stdlib.h>
#include <iostream>
#include <math.h>
#include <assert.h>
#include <cuda.h>
#include <cuda_runtime.h>
#include "helper_cuda.h"
#include "helper_string.h"
#include <fstream>
#ifndef MAX
#define MAX(a,b) ((a > b) ? a : b)
#endif
#define REAL sizeof(float)
#define N 128
#define BLOCK_SIZE 128
#define NUM_BLOCKS ((N*N + BLOCK_SIZE - 1)/ BLOCK_SIZE)
template <typename T>
inline void printmatrix( T mat, int rows, int cols);
template <typename T>
__global__ void jacobian_kernel ( const T * A, T * J, const T t0, const T tn, const T h, const T * u0, const T * un, const T * un_old);
template<typename T>
__device__ void d_func(const T t, const T u[], T res[], const T h = 1);
template<typename T>
int main ()
{
float t0 = 0.; //float tn = 0.;
float h = 0.1;
float* u0 = (float*)malloc(REAL*N); for(int i = 0; i < N; ++i){u0[i] = i+1;}
float* un = (float*)malloc(REAL*N); memcpy(un, u0, REAL*N);
float* un_old = (float*)malloc(REAL*N); memcpy(un_old, u0, REAL*N);
float* J = (float*)malloc(REAL*N*N);
float* A = (float*)malloc(REAL*N*N); host_heat_matrix(A);
float *d_u0;
float *d_un;
float *d_un_old;
float *d_J;
float *d_A;
checkCudaErrors(cudaMalloc((void**)&d_u0, REAL*N)); //printf("1: %p\n", d_u0);
checkCudaErrors(cudaMalloc((void**)&d_un, REAL*N)); //printf("2: %p\n", d_un);
checkCudaErrors(cudaMalloc((void**)&d_un_old, REAL*N)); //printf("3: %p\n", d_un_old);
checkCudaErrors(cudaMalloc((void**)&d_J, REAL*N*N)); //printf("4: %p\n", d_J);
checkCudaErrors(cudaMalloc((void**)&d_A, REAL*N*N)); //printf("4: %p\n", d_J);
checkCudaErrors(cudaMemcpy(d_u0, u0, REAL*N, cudaMemcpyHostToDevice)); assert(d_u0 != NULL);
checkCudaErrors(cudaMemcpy(d_un, un, REAL*N, cudaMemcpyHostToDevice)); assert(d_un != NULL);
checkCudaErrors(cudaMemcpy(d_un_old, un_old, REAL*N, cudaMemcpyHostToDevice)); assert(d_un_old != NULL);
checkCudaErrors(cudaMemcpy(d_J, J, REAL*N*N, cudaMemcpyHostToDevice)); assert(d_J != NULL);
checkCudaErrors(cudaMemcpy(d_A, A, REAL*N*N, cudaMemcpyHostToDevice)); assert(d_A != NULL);
dim3 dimGrid(NUM_BLOCKS); std::cout << "NUM_BLOCKS \t" << dimGrid.x << "\n";
dim3 dimBlock(BLOCK_SIZE); std::cout << "BLOCK_SIZE \t" << dimBlock.x << "\n";
size_t shm_size = N*REAL; //std::cout << shm_size << "\n";
//HERE IS A RELEVANT CHANGE OF THE MAIN, SINCE I WAS CALLING
//THE KERNEL WITH A 2D GRID BUT WITHOUT USING THE .y AXIS,
//WHILE NOW THE GRID IS 1D
jacobian_kernel <<< dimGrid, dimBlock, shm_size >>> (d_A, d_J, t0, t0, h, d_u0, d_un, d_un_old);
checkCudaErrors(cudaMemcpy(J, d_J, REAL*N*N, cudaMemcpyDeviceToHost)); //printf("4: %p\n", d_J);
printmatrix( J, N, N);
checkCudaErrors(cudaDeviceReset());
free(u0);
free(un);
free(un_old);
free(J);
}
template <typename T>
__global__ void jacobian_kernel (
const T * A,
T * J,
const T t0,
const T tn,
const T h,
const T * u0,
const T * un,
const T * un_old)
{
T cgamma = 2 - sqrtf(2);
const unsigned int t = threadIdx.x;
const unsigned int b = blockIdx.x;
const unsigned int tid = t + b * blockDim.x;
/*__shared__*/ T temp_sx[BLOCK_SIZE][BLOCK_SIZE];
/*__shared__*/ T temp_dx[BLOCK_SIZE][BLOCK_SIZE];
__shared__ T sm_temp_du;
T* temp_du = &sm_temp_du;
//HERE IS A RELEVANT CHANGE (*)
if ( t < BLOCK_SIZE && b < NUM_BLOCKS )
{
temp_sx[b][t] = un[t]; //printf("temp_sx[%d] = %f\n", t,(temp_sx[b][t]));
temp_dx[b][t] = un[t];
//printf("t = %d, b = %d, t + b * blockDim.x = %d \n",t, b, tid);
//HERE IS A NOTE (**)
if ( t == b )
{
//printf("t = %d, b = %d \n",t, b);
if ( tn == t0 )
{
*temp_du = u0[t]*0.001;
temp_sx[b][t] += *temp_du;
temp_dx[b][t] -= *temp_du;
temp_sx[b][t] += ( abs( temp_sx[b][t] ) < 10e-6 ? 0.1 : 0 );
temp_dx[b][t] += ( abs( temp_dx[b][t] ) < 10e-6 ? 0.1 : 0 );
temp_sx[b][t] = ( temp_sx[b][t] == 0 ? 0.1 : temp_sx[b][t] );
temp_dx[b][t] = ( temp_dx[b][t] == 0 ? 0.1 : temp_dx[b][t] );
}
else
{
*temp_du = MAX( un[t] - un_old[t], 10e-6 );
temp_sx[b][t] += *temp_du;
temp_dx[b][t] -= *temp_du;
}
;
}
//printf("du[%d] %f\n", tid, (*temp_du));
__syncthreads();
//printf("temp_sx[%d][%d] = %f\n", b, t, temp_sx[b][t]);
//printf("temp_dx[%d][%d] = %f\n", b, t, temp_dx[b][t]);
//d_func(tn, (temp_sx[b]), (temp_sx[b]), 1.f);
//d_func(tn, (temp_dx[b]), (temp_dx[b]), 1.f);
matvec_dev( tn, A, (temp_sx[b]), (temp_sx[b]), N, N, 1.f );
matvec_dev( tn, A, (temp_dx[b]), (temp_dx[b]), N, N, 1.f );
__syncthreads();
//printf("temp_sx_later[%d][%d] = %f\n", b, t, (temp_sx[b][t]));
//printf("temp_sx_later[%d][%d] - temp_dx_later[%d][%d] = %f\n", b,t,b,t, (temp_sx[b][t] - temp_dx[b][t]) / 2 * *temp_du);
//if (t == b ) printf( "2du[%d]^-1 = %f\n",t, powf((2 * *temp_du), -1));
J[tid] = (temp_sx[b][t] - temp_dx[b][t]) / (2 * *temp_du);
}
}
template<typename T>
__device__ void d_func(const T t, const T u[], T res[], const T h )
{
__shared__ float temp_u;
temp_u = u[threadIdx.x];
res[threadIdx.x] = h*powf( (temp_u), 2);
}
template <typename T>
inline void printmatrix( T mat, int rows, int cols)
{
std::ofstream matrix_out;
matrix_out.open( "heat_matrix.txt", std::ofstream::out);
for( int i = 0; i < rows; i++)
{
for( int j = 0; j <cols; j++)
{
double next = mat[i + N*j];
matrix_out << ( (next >= 0) ? " " : "") << next << " ";
}
matrix_out << "\n";
}
}
The relevant change is on (*). Before I used if (tid < N) which has two downsides:
First, it is wrong, since it should be tid < N*N, as my data is 2D, while tid is a global index which tracks all the data.
Even if I wrote tid < N*N, since I'm splitting the function calls into blocks, the t < BLOCK_SIZE && b < NUM_BLOCKS seems clearer to me in how the indexing is arranged in the code.
Moreover, the statement t == b in (**) is actually the right one to operate on the diagonal elements of the matrix. The fact that it was evaluated true only on 0 was because of my error right above.
Thanks for the suggestions!

Simple OpenCL program not working

This program is a simple parallel program which adds the elements of 2 vectors.
The program was error free and it was compiled successfully but the results are not right
#include <stdio.h>
#include <stdlib.h>
#include <iostream>
#include <iomanip>
#include <array>
#include <fstream>
#include <sstream>
#include <string>
#include <algorithm>
#include <iterator>
#ifdef __APPLE__
#include <OpenCL/opencl.h>
#else
#include <CL/cl.h>
#include <time.h>
#endif
#define MAX_SOURCE_SIZE (0x100000)
// number of points in Both A and B files (number of rows)
const int number_of_points = 11;
// number of points axis in Both A and B files (number of Columns)
const int number_of_axis = 3;
using namespace std;
int main(int argc, char *argv[]) {
clock_t tStart = clock();
// Create the two input vectors
// working variables
int i;
ifstream input_fileA, input_fileB; // input files
string line; // transfer row from file to array
float x; // transfer word from file to array
int row = 0; // number of rows of file A,B (= array)
int col = 0; // number of rows of file A,B (= array)
// working arrays
// array contains file A data
float arrayA[number_of_points][number_of_axis]={{0}};
// array contains file B data
float arrayB[number_of_points][number_of_axis]={{0}};
// float X1[number_of_points]; // X values of file A points
float Y1[number_of_points]; // Y values of file A points
// float X2[number_of_points]; // X values of file B points
float Y2[number_of_points]; // Y values of file B points
float *X1 = (float*)malloc(sizeof(float)*number_of_points);
float *X2 = (float*)malloc(sizeof(float)*number_of_points);
// import input files
input_fileA.open(argv[1]);
input_fileB.open(argv[2]);
// transfer input files data to array
// input file A to arrayA
row = 0;
while (getline(input_fileA, line))
{
istringstream streamA(line);
col = 0;
while(streamA >> x){
arrayA[row][col] = x;
col++;
}
row++;
}
// input file B to arrayB
row = 0;
while (getline(input_fileB, line))
{
istringstream streamB(line);
col = 0;
while(streamB >> x){
arrayB[row][col] = x;
col++;
}
row++;
}
// put Xs of points in X vectors and Ys of points in Y vectors
// input file A
for (int i = 0; i<number_of_points; i++){
X1[i] = arrayA[i][1];
Y1[i] = arrayA[i][2];
}
// input file B
for (int i = 0; i<number_of_points; i++){
X2[i] = arrayB[i][1];
Y2[i] = arrayB[i][2];
}
// int i;
// const int LIST_SIZE = 50;
// int *A = (int*)malloc(sizeof(int)*number_of_points);
// int *B = (int*)malloc(sizeof(int)*number_of_points);
// for(i = 0; i < number_of_points; i++) {
// A[i] = X1[i];
// B[i] = X2[i];
// }
// Load the kernel source code into the array source_str
FILE *fp;
char *source_str;
size_t source_size;
fp = fopen("vector_add_kernel.cl", "r");
if (!fp) {
fprintf(stderr, "Failed to load kernel.\n");
exit(1);
}
source_str = (char*)malloc(MAX_SOURCE_SIZE);
source_size = fread( source_str, 1, MAX_SOURCE_SIZE, fp);
fclose( fp );
// Get platform and device information
cl_platform_id platform_id = NULL;
cl_device_id device_id = NULL;
cl_uint ret_num_devices;
cl_uint ret_num_platforms;
cl_int ret = clGetPlatformIDs(1, &platform_id, &ret_num_platforms);
ret = clGetDeviceIDs( platform_id, CL_DEVICE_TYPE_ALL, 1,
&device_id, &ret_num_devices);
// Create an OpenCL context
cl_context context =
clCreateContext( NULL, 1, &device_id, NULL, NULL, &ret);
// Create a command queue
cl_command_queue command_queue =
clCreateCommandQueue(context, device_id, 0, &ret);
// Create memory buffers on the device for each vector
cl_mem x1_mem_obj = clCreateBuffer(context, CL_MEM_READ_ONLY,
number_of_points * sizeof(float), NULL, &ret);
cl_mem x2_mem_obj = clCreateBuffer(context, CL_MEM_READ_ONLY,
number_of_points * sizeof(float), NULL, &ret);
cl_mem c_mem_obj = clCreateBuffer(context, CL_MEM_WRITE_ONLY,
number_of_points * sizeof(float), NULL, &ret);
// Copy the lists A and B to their respective memory buffers
ret = clEnqueueWriteBuffer(command_queue, x1_mem_obj, CL_TRUE, 0,
number_of_points * sizeof(float), X1, 0, NULL, NULL);
ret = clEnqueueWriteBuffer(command_queue, x2_mem_obj, CL_TRUE, 0,
number_of_points * sizeof(float), X2, 0, NULL, NULL);
// Create a program from the kernel source
cl_program program = clCreateProgramWithSource(context, 1,
(const char **)&source_str, (const size_t *)&source_size, &ret);
// Build the program
ret = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL);
// Create the OpenCL kernel
cl_kernel kernel = clCreateKernel(program, "vector_add", &ret);
// Set the arguments of the kernel
ret = clSetKernelArg(kernel, 0, sizeof(cl_mem), (void *)&x1_mem_obj);
ret = clSetKernelArg(kernel, 1, sizeof(cl_mem), (void *)&x2_mem_obj);
ret = clSetKernelArg(kernel, 2, sizeof(cl_mem), (void *)&c_mem_obj);
// Execute the OpenCL kernel on the list
size_t global_item_size = number_of_points; // Process the entire lists
size_t local_item_size = 64; // Process in groups of 64
ret = clEnqueueNDRangeKernel(command_queue, kernel, 1, NULL,
&global_item_size, &local_item_size, 0, NULL, NULL);
// Read the memory buffer C on the device to the local variable C
// int *C = (int*)malloc(sizeof(int)*number_of_points);
float *C = (float*)malloc(sizeof(float)*number_of_points);
ret = clEnqueueReadBuffer(command_queue, c_mem_obj, CL_TRUE, 0,
number_of_points * sizeof(float), C, 0, NULL, NULL);
// Display the result to the screen
for(i = 0; i < number_of_points; i++)
printf("%f + %f = %f\n", X1[i], X2[i], C[i]);
// Clean up
ret = clFlush(command_queue);
ret = clFinish(command_queue);
ret = clReleaseKernel(kernel);
ret = clReleaseProgram(program);
ret = clReleaseMemObject(x1_mem_obj);
ret = clReleaseMemObject(x2_mem_obj);
ret = clReleaseMemObject(c_mem_obj);
ret = clReleaseCommandQueue(command_queue);
ret = clReleaseContext(context);
free(X1);
free(X2);
free(C);
printf("ALL Time taken: %.2fs\n", (double)(clock() - tStart)/CLOCKS_PER_SEC);
return 0;
}
and the kernel file
__kernel void vector_add(__global float *X1,
__global float *X2,
__global float *C) {
// Get the index of the current element
int i = get_global_id(0);
// Do the operation
C[i] = X1[i] + X2[i];
}
The result was
0.000000 + 0.000000 = 0.000000
1.000000 + 1.000000 = 0.000000
2.000000 + 2.000000 = 0.000000
3.000000 + 3.000000 = 0.000000
4.000000 + 4.000000 = 0.000000
5.000000 + 5.000000 = 0.000000
6.000000 + 6.000000 = 0.000000
7.000000 + 7.000000 = 0.000000
8.000000 + 8.000000 = 0.000000
9.000000 + 9.000000 = 0.000000
1.000000 + 1.000000 = 0.000000
ALL Time taken: 0.07s
You've committed one of the cardinal sins of OpenCL programming, in that you are not checking the error codes from any of your OpenCL API calls! You should always check the return code from every single OpenCL API call. If you did this, it would point you towards the problem very quickly.
The problem is in your kernel enqueue call. If you check the error code, you'll see that you are getting -54 back, which corresponds to CL_INVALID_WORK_GROUP_SIZE. Specifically, kernel invocations have the requirement that the work-group size (local size) exactly divides the global size. You are asking for a work-group size of 64 and a global size of 11, which does not fulfil this requirement.
You can also pass NULL as the work-group size parameter, and the OpenCL implementation will pick a work-group size that will definitely work on your behalf.

Julia Set - Cuda , improve the performance failed

Recently I am learning the examples in the book CUDA by JASON SANDERS.
the example of Juila Set makes a bad performance of 7032ms.
Here is the program:
#include <cuda.h>
#include <cuda_runtime.h>
#include <cpu_bitmap.h>
#include <book.h>
#define DIM 1024
struct cuComplex{
float r;
float i;
__device__ cuComplex(float a, float b) : r(a),i(b){
}
__device__ float magnitude2(void){
return r*r+i*i;
}
__device__ cuComplex operator *(const cuComplex& a){
return cuComplex(r*a.r-i*a.i, i*a.r+r*a.i);
}
__device__ cuComplex operator +(const cuComplex& a){
return cuComplex(r+a.r,i+a.i);
}
};
__device__ int julia(int x,int y){
const float scale = 1.5;
float jx = scale * (float)(DIM/2 - x)/(DIM/2);
float jy = scale * (float)(DIM/2 - y)/(DIM/2);
cuComplex c(-0.8,0.156);
cuComplex a(jx,jy);
int i = 0;
for(i = 0; i<200; i++){
a = a*a + c;
if(a.magnitude2() > 1000){
return 0;
}
}
return 1;
}
__global__ void kernel(unsigned char *ptr){
int x = blockIdx.x;
int y = blockIdx.y;
int offset = x + y*gridDim.x;
int juliaValue = julia(x,y);
ptr[offset*4 + 0] = 255*juliaValue;
ptr[offset*4 + 1] = 0;
ptr[offset*4 + 2] = 1;
ptr[offset*4 + 3] = 255;
}
int main(void){
CPUBitmap bitmap(DIM,DIM);
unsigned char * dev_bitmap;
dim3 grid(DIM,DIM);
dim3 blocks(DIM/16,DIM/16);
dim3 threads(16,16);
dim3 thread(DIM,DIM);
cudaEvent_t start,stop;
cudaEvent_t bitmapCpy_start,bitmapCpy_stop;
HANDLE_ERROR(cudaEventCreate(&start));
HANDLE_ERROR(cudaEventCreate(&stop));
HANDLE_ERROR(cudaEventCreate(&bitmapCpy_start));
HANDLE_ERROR(cudaEventCreate(&bitmapCpy_stop));
HANDLE_ERROR(cudaMalloc((void **)&dev_bitmap,bitmap.image_size()));
HANDLE_ERROR(cudaEventRecord(start,0));
kernel<<<grid,1>>>(dev_bitmap);
HANDLE_ERROR(cudaMemcpy(bitmap.get_ptr(),dev_bitmap,bitmap.image_size(),cudaMemcpyDeviceToHost));
//HANDLE_ERROR(cudaEventRecord(bitmapCpy_stop,0));
//HANDLE_ERROR(cudaEventSynchronize(bitmapCpy_stop));
// float copyTime;
// HANDLE_ERROR(cudaEventElapsedTime(&copyTime,bitmapCpy_start,bitmapCpy_stop));
HANDLE_ERROR(cudaEventRecord(stop,0));
HANDLE_ERROR(cudaEventSynchronize(stop));
float elapsedTime;
HANDLE_ERROR(cudaEventElapsedTime(&elapsedTime,start,stop));
//printf("Total time is %3.1f ms, time for copying is %3.1f ms \n",elapsedTime,copyTime);
printf("Total time is %3.1f ms\n",elapsedTime);
bitmap.display_and_exit();
HANDLE_ERROR(cudaEventDestroy(start));
HANDLE_ERROR(cudaEventDestroy(stop));
HANDLE_ERROR(cudaEventDestroy(bitmapCpy_start));
HANDLE_ERROR(cudaEventDestroy(bitmapCpy_stop));
HANDLE_ERROR(cudaFree(dev_bitmap));
}
I think the main factor that influences the performance is that the program above just run 1 thread in every block:
kernel<<<grid,1>>>(dev_bitmap);
so I change the kernel like the following:
__global__ void kernel(unsigned char *ptr){
int x = threadIdx.x + blockIdx.x*blockDim.x;
int y = threadIdx.y + blockIdx.y*blockDim.y;
int offset = x + y*gridDim.x*blockIdx.x;
int juliaValue = julia(x,y);
ptr[offset*4 + 0] = 255*juliaValue;
ptr[offset*4 + 1] = 0;
ptr[offset*4 + 2] = 1;
ptr[offset*4 + 3] = 255;
}
and call kernel:
dim3 blocks(DIM/16,DIM/16);
dim3 threads(16,16);
kernel<<<blocks,threads>>>(dev_bitmap);
I think this change is not a big deal, but when I ran it, it acted like that it ran into some endless loops, no image appeared and I couldn't do anything with my screen, just blocked there.
toolkit: cuda 5.5
system: ubuntu 12.04
When I run the original code you have posted here, I get a correct display and a time of ~340ms.
When I make your kernel change, I get an "unspecified launch error" on the kernel launch.
In your modified kernel, you have the following which is an incorrect computation:
int offset = x + y*gridDim.x*blockIdx.x;
When I change it to:
int offset = x + y*gridDim.x*blockDim.x;
I get normal execution and results, and an indicated time of ~10ms.

Cuda thrust global memory writing very slow

I am currently writing a code, that calculates a integral Histogram on the GPU using the Nvidia thrust library.
Therefore I allocate a continuous Block of device memory which I update with a custom functor all the time.
The problem is, that the write to the device memory is veeery slow, but the reads are actually ok.
The basic setup is the following:
struct HistogramCreation
{
HistogramCreation(
...
// pointer to memory
...
){}
/// The actual summation operator
__device__ void operator()(int index){
.. do the calculations ..
for(int j=0;j<30;j++){
(1) *_memoryPointer = values (also using reads to such locations) ;
}
}
}
void foo(){
cudaMalloc(_pointer,size);
HistogramCreation initialCreation( ... _pointer ...);
thrust::for_each(
thrust::make_counting_iterator(0),
thrust::make_counting_iterator(_imageSize),
initialCreation);
}
if I change the writing in (1) to the following>
unsigned int val = values;
The performance is much better. THis is the only global memory write I have.
Using the memory write I get about 2s for HD Footage.
using the local variable it takes about 50 ms so about a factor of 40 less.
Why is this so slow? how could I improve it?
Just as #OlegTitov said, frequent load/store with global
memory should be avoided as much as possible. When there's a
situation where it's inevitable, then coalesced memory
access can help the execution process not to get too slow;
however in most cases, histogram calculation is pretty tough
to realize the coalesced access.
While most of the above is basically just restating
#OlegTitov's answer, i'd just like to share about an
investigation i did about finding summation with NVIDIA
CUDA. Actually the result is pretty interesting and i hope
it'll be a helpful information for other xcuda developers.
The experiment was basically to run a speed test of finding
summation with various memory access patterns: using global
memory (1 thread), L2 cache (atomic ops - 128 threads), and
L1 cache (shared mem - 128 threads)
This experiment used:
Kepler GTX 680,
1546 cores # 1.06GHz
GDDR5 256-bit # 3GHz
Here are the kernels:
__global__
void glob(float *h) {
float* hist = h;
uint sd = SEEDRND;
uint random;
for (int i = 0; i < NUMLOOP; i++) {
if (i%NTHREADS==0) random = rnd(sd);
int rind = random % NBIN;
float randval = (float)(random % 10)*1.0f ;
hist[rind] += randval;
}
}
__global__
void atom(float *h) {
float* hist = h;
uint sd = SEEDRND;
for (int i = threadIdx.x; i < NUMLOOP; i+=NTHREADS) {
uint random = rnd(sd);
int rind = random % NBIN;
float randval = (float)(random % 10)*1.0f ;
atomicAdd(&hist[rind], randval);
}
}
__global__
void shm(float *h) {
int lid = threadIdx.x;
uint sd = SEEDRND;
__shared__ float shm[NTHREADS][NBIN];
for (int i = 0; i < NBIN; i++) shm[lid][i] = h[i];
for (int i = lid; i < NUMLOOP; i+=NTHREADS) {
uint random = rnd(sd);
int rind = random % NBIN;
float randval = (float)(random % 10)*1.0f ;
shm[lid][rind] += randval;
}
/* reduction here */
for (int i = 0; i < NBIN; i++) {
__syncthreads();
if (threadIdx.x < 64) {
shm[threadIdx.x][i] += shm[threadIdx.x+64][i];
}
__syncthreads();
if (threadIdx.x < 32) {
shm[threadIdx.x][i] += shm[threadIdx.x+32][i];
}
__syncthreads();
if (threadIdx.x < 16) {
shm[threadIdx.x][i] += shm[threadIdx.x+16][i];
}
__syncthreads();
if (threadIdx.x < 8) {
shm[threadIdx.x][i] += shm[threadIdx.x+8][i];
}
__syncthreads();
if (threadIdx.x < 4) {
shm[threadIdx.x][i] += shm[threadIdx.x+4][i];
}
__syncthreads();
if (threadIdx.x < 2) {
shm[threadIdx.x][i] += shm[threadIdx.x+2][i];
}
__syncthreads();
if (threadIdx.x == 0) {
shm[0][i] += shm[1][i];
}
}
for (int i = 0; i < NBIN; i++) h[i] = shm[0][i];
}
OUTPUT
atom: 102656.00 shm: 102656.00 glob: 102656.00
atom: 122240.00 shm: 122240.00 glob: 122240.00
... blah blah blah ...
One Thread: 126.3919 msec
Atomic: 7.5459 msec
Sh_mem: 2.2207 msec
The ratio between these kernels is 57:17:1. Many things can
be analyzed here, and it truly does not mean that using
L1 or L2 memory spaces will always give you more than 10
times speedup of the whole program.
And here's the main and other funcs:
#include <iostream>
#include <cstdlib>
#include <cstdio>
using namespace std;
#define NUMLOOP 1000000
#define NBIN 36
#define SEEDRND 1
#define NTHREADS 128
#define NBLOCKS 1
__device__ uint rnd(uint & seed) {
#if LONG_MAX > (16807*2147483647)
int const a = 16807;
int const m = 2147483647;
seed = (long(seed * a))%m;
return seed;
#else
double const a = 16807;
double const m = 2147483647;
double temp = seed * a;
seed = (int) (temp - m * floor(temp/m));
return seed;
#endif
}
... the above kernels ...
int main()
{
float *h_hist, *h_hist2, *h_hist3, *d_hist, *d_hist2,
*d_hist3;
h_hist = (float*)malloc(NBIN * sizeof(float));
h_hist2 = (float*)malloc(NBIN * sizeof(float));
h_hist3 = (float*)malloc(NBIN * sizeof(float));
cudaMalloc((void**)&d_hist, NBIN * sizeof(float));
cudaMalloc((void**)&d_hist2, NBIN * sizeof(float));
cudaMalloc((void**)&d_hist3, NBIN * sizeof(float));
for (int i = 0; i < NBIN; i++) h_hist[i] = 0.0f;
cudaMemcpy(d_hist, h_hist, NBIN * sizeof(float),
cudaMemcpyHostToDevice);
cudaMemcpy(d_hist2, h_hist, NBIN * sizeof(float),
cudaMemcpyHostToDevice);
cudaMemcpy(d_hist3, h_hist, NBIN * sizeof(float),
cudaMemcpyHostToDevice);
cudaEvent_t start, end;
float elapsed = 0, elapsed2 = 0, elapsed3;
cudaEventCreate(&start);
cudaEventCreate(&end);
cudaEventRecord(start, 0);
atom<<<NBLOCKS, NTHREADS>>>(d_hist);
cudaThreadSynchronize();
cudaEventRecord(end, 0);
cudaEventSynchronize(start);
cudaEventSynchronize(end);
cudaEventElapsedTime(&elapsed, start, end);
cudaEventRecord(start, 0);
shm<<<NBLOCKS, NTHREADS>>>(d_hist2);
cudaThreadSynchronize();
cudaEventRecord(end, 0);
cudaEventSynchronize(start);
cudaEventSynchronize(end);
cudaEventElapsedTime(&elapsed2, start, end);
cudaEventRecord(start, 0);
glob<<<1, 1>>>(d_hist3);
cudaThreadSynchronize();
cudaEventRecord(end, 0);
cudaEventSynchronize(start);
cudaEventSynchronize(end);
cudaEventElapsedTime(&elapsed3, start, end);
cudaMemcpy(h_hist, d_hist, NBIN * sizeof(float),
cudaMemcpyDeviceToHost);
cudaMemcpy(h_hist2, d_hist2, NBIN * sizeof(float),
cudaMemcpyDeviceToHost);
cudaMemcpy(h_hist3, d_hist3, NBIN * sizeof(float),
cudaMemcpyDeviceToHost);
/* print output */
for (int i = 0; i < NBIN; i++) {
printf("atom: %10.2f shm: %10.2f glob:
%10.2f¥n",h_hist[i],h_hist2[i],h_hist3[i]);
}
printf("%12s: %8.4f msec¥n", "One Thread", elapsed3);
printf("%12s: %8.4f msec¥n", "Atomic", elapsed);
printf("%12s: %8.4f msec¥n", "Sh_mem", elapsed2);
return 0;
}
When writing GPU code you should avoid reading and writing to/from global memory. Global memory is very slow on GPU. That's the hardware feature. The only thing you can do is to make neighboring treads read/write in neighboring adresses in global memory. This will cause coalescing and speed up the process. But in general read your data once, process it and write it out once.
Note that NVCC might optimize out a lot of your code after you make the modification - it detects that no write to global memory is made and just removes the "unneeded" code. So this speedup may not be coming out of the global writer per ce.
I would recommend using profiler on your actual code (the one with global write) to see if there's anything like unaligned access or other perf problem.

Resources