Getting stack overflows with a CUDA kernel
I have a huge problem with the code I am programming. I am not an expert, and i asked many people before coming here. corrected a lot of things, too. So, I guess I am ready to show you the code and ask you my questions. I will put the entire code here, as a way to make you understand well what my problem is. The thing i wanna do there is, if ARRAY_SIZE is too big for the THREAD_SIZE, so I put the data of the big array into a smaller array, specially-created with size THREAD_SIZE . Then, I send it to the kernel and do whatever I have to do. But I am having problem on the part
isub_matrix[x*THREAD_SIZE+y]=big_matrix[x*ARRAY_SIZE+y];
where the code stops, due to stack overflow. First, I made a double pointer of big_matrix. But people in the #cuda channel at freenode irc network told me it was too big for the CPU memory to handle it, that I should create a linear pointer. I did it, but I still have the same problem of stack overflow. So, here it goes... updated after some changes, that didnt work yet (the stack overflow stopped, but theres a linking and manifest update fail)
#define ARRAY_SIZE 2048
#define THREAD_SIZE 32
#define PI 3.14
int main(int argc, char** argv)
{
int array_plus=0,x,y;
float time;
//unsigned int memsize=sizeof(float)*THREAD_SIZE*THREAD_SIZE;
//bool array_rest;
cudaEvent_t start,stop;
float *d_isub_matrix;
float *big_matrix = new float[ARRAY_SIZE*ARRAY_SIZE];
float *big_matrix2 = new float[ARRAY_SIZE*ARRAY_SIZE];
float *isub_matrix = new float[THREAD_SIZE*THREAD_SIZE];
float *osub_matrix = new float[THREAD_SIZE*THREAD_SIZE];
//if the array's size is not compatible with the thread's size, it won't work.
//array_rest=(ARRAY_SIZE*ARRAY_SIZE)/(THREAD_SIZE*THREAD_SIZE);
//isub_matrix=(float*) malloc(memsize);
//osub_matrix=(float*) malloc(memsize);
if(((ARRAY_SIZE*ARRAY_SIZE)%(THREAD_SIZE*THREAD_SIZE)==0))
{
//allocating space in CPU memory and GPU memory for the big matrix and its sub matrixes
//it has to be like this (lots of loops)
//populating the big array
for(x=0;x<ARRAY_SIZE;x++)
{
for(y=0;y<ARRAY_SIZE;y++)
big_matrix[x*ARRAY_SIZE+y]=rand()%10000;
}
//kind of loop for the big array
//Start counting the time of processing (everything)
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start,0);
while(array_plus<ARRAY_SIZE)
{
//putting the big array's values into the sub-matrix
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
isub_matrix[x*THREAD_SIZE+y]=big_matrix[(x+array_plus)*ARRAY_SIZE+y];
}
cudaMalloc((void**)&d_isub_matrix,THREAD_SIZE*THREAD_SIZE*sizeof(float));
cudaMalloc((void**)&osub_matrix,THREAD_SIZE*THREAD_SIZE*sizeof(float));
cudaMemcpy(d_isub_matrix,isub_matrix,((THREAD_SIZE*THREAD_SIZE)*sizeof(float)),cudaMemcpyHostToDevice);
//call the cuda kernel
twiddle_factor<<<1,256>>>(isub_matrix,osub_matrix);//<----
cudaMemcpy(osub_matrix,isub_matrix,((THREAD_SIZE*THREAD_SIZE)*sizeof(float)),cudaMemcpyDeviceToHost);
array_plus=array_plus+THREAD_SIZE;
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
big_matrix2[x*THREAD_SIZE+array_plus+y]=osub_matrix[x*THREAD_SIZE+y];
}
array_rest=array_plus+(ARRAY_SIZE);
cudaFree(isub_matrix);
cudaFree(osub_matrix);
system("PAUSE");
}
//Stop the time
cudaEventRecord(stop,0);
cudaEventSynchronize(stop);
cudaEventElapsedTime(&time,start,stop);
//Free memory in GPU
printf("The processing time took... %fms to finish",time);
system("PAUSE");
}
printf("The processing time took...NAO ENTROU!");
system("PAUSE");
return 0;
}
//things to do: TRANSPOSITION!!!!
Another question is about the parallel part. The compiler (Visual Studio) says that I engaged too many pow() and exp() at once. How should I solve this problem?
if((xIndex<THREAD_SIZE)&&(yIndex<THREAD_SIZE))
{
block[xIndex][yIndex]=exp(sum_sin[xIndex][yIndex])+exp(sum_cos[xIndex][yIndex]);
}
The original code is down here. I commented it because i wanted to know if at least my code was taking some value in the GPU. But it wasnt even launching the Kernel... so sad)
__global__ void twiddle_factor(float *isub_matrix, float *osub_matrix)
{
__shared__ float block[THREAD_SIZE][THREAD_SIZE];
// int x,y,z;
unsigned int xIndex = threadIdx.x;
unsigned int yIndex = threadIdx.y;
/*
int sum_sines=0.0;
int sum_cosines=0.0;
float sum_sin[THREAD_SIZE],sum_cos[THREAD_SIZE];
float angle=(2*PI)/THREAD_SIZE;
//put into shared memory the FFT calculation (F(u))
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
{
for(z=0;z<THREAD_SIZE;z++)
{
sum_sines=sum_sin+sin(isub_matrix[y*THREAD_SIZE+z]*(angle*z));
sum_cosines=sum_cos+cos(isub_matrix[y*THREAD_SIZE+z]*(angle*z));
}
sum_sin[x][y]=sum_sines/THREAD_SIZE;
sum_cos[x][y]=sum_cosines/THREAD_SIZE;
}
}
*/
if((xIndex<THREAD_SIZE)&&(yIndex<THREAD_SIZE))
block[xIndex][yIndex]=pow(THREAD_SIZE,0.5);
//block[xIndex][yIndex]=pow(exp(sum_sin[xIndex*THREAD_SIZE+yIndex])+exp(sum_cos[xIndex*THREAD_SIZE+yIndex]),0.5);
__syncthreads();
//transposition X x Y
//transfer back the results into another sub-matrix that is allocated in CPU
if((xIndex<THREAD_SIZE)&&(yIndex<THREAD_SIZE))
osub_matrix[yIndex*THREAD_SIZE+xIndex]=block[xIndex][yIndex];
__syncthreads();
}
Thanks for reading it all!
Below is the entire code:
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>
#define ARRAY_SIZE 2048
#define THREAD_SIZE 32
#define PI 3.14
__global__ void twiddle_factor(float *isub_matrix, float *osub_matrix)
{
__shared__ float block[THREAD_SIZE][THREAD_SIZE];
int x,y,z;
unsigned int xIndex = threadIdx.x;
unsigned int yIndex = threadIdx.y;
float sum_sines=0.0;
//float expo_sums;
float sum_cosines=0.0;
float sum_sin[THREAD_SIZE][THREAD_SIZE],sum_cos[THREAD_SIZE][THREAD_SIZE];
float angle=(2*PI)/THREAD_SIZE;
//put into shared memory the FFT calculation (F(u))
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
{
for(z=0;z<THREAD_SIZE;z++)
{
sum_sines=sum_sines+sin(isub_matrix[y*THREAD_SIZE+z]*(angle*z));
sum_cosines=sum_cosines+cos(isub_matrix[y*THREAD_SIZE+z]*(angle*z));
}
sum_sin[x][y]=sum_sines/THREAD_SIZE;
sum_cos[x][y]=sum_cosines/THREAD_SIZE;
}
}
if((xIndex<THREAD_SIZE)&&(yIndex<THREAD_SIZE))
{
block[xIndex][yIndex]=exp(sum_sin[xIndex][yIndex])+exp(sum_cos[xIndex][yIndex]);
}
__syncthreads();
//transposition X x Y
//transfer back the results into another sub-matrix that is allocated in CPU
if((xIndex<THREAD_SIZE)&&(yIndex<THREAD_SIZE))
osub_matrix[yIndex*THREAD_SIZE+xIndex]=block[xIndex][yIndex];
__syncthreads();
}
int main(int argc, char** argv)
{
int array_plus=0,x,y;
float time;
//unsigned int memsize=sizeof(float)*THREAD_SIZE*THREAD_SIZE;
//bool array_rest;
cudaEvent_t start,stop;
float *d_isub_matrix,*d_osub_matrix;
float *big_matrix = new float[ARRAY_SIZE*ARRAY_SIZE];
float *big_matrix2 = new float[ARRAY_SIZE*ARRAY_SIZE];
float *isub_matrix = new float[THREAD_SIZE*THREAD_SIZE];
float *osub_matrix = new float[THREAD_SIZE*THREAD_SIZE];
//if the array's size is not compatible with the thread's size, it won't work.
//array_rest=(ARRAY_SIZE*ARRAY_SIZE)/(THREAD_SIZE*THREAD_SIZE);
//isub_matrix=(float*) malloc(memsize);
//osub_matrix=(float*) malloc(memsize);
if(((ARRAY_SIZE*ARRAY_SIZE)%(THREAD_SIZE*THREAD_SIZE)==0)&&(ARRAY_SIZE>=THREAD_SIZE))
{
//allocating space in CPU memory and GPU memory for the big matrix and its sub matrixes
//it has to be like this (lots of loops)
//populating the big array
for(x=0;x<ARRAY_SIZE;x++)
{
for(y=0;y<ARRAY_SIZE;y++)
big_matrix[x*ARRAY_SIZE+y]=rand()%10000;
}
//kind of loop for the big array
//Start counting the time of processing (everything)
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start,0);
while(array_plus<ARRAY_SIZE)
{
//putting the big array's values into the sub-matrix
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
isub_matrix[x*THREAD_SIZE+y]=big_matrix[x*ARRAY_SIZE+y];
}
cudaMalloc((void**)&d_isub_matrix,THREAD_SIZE*THREAD_SIZE*sizeof(float));
cudaMalloc((void**)&d_osub_matrix,THREAD_SIZE*THREAD_SIZE*sizeof(float));
cudaMemcpy(d_isub_matrix,isub_matrix,((THREAD_SIZE*THREAD_SIZE)*sizeof(float)),cudaMemcpyHostToDevice);
//call the cuda kernel
twiddle_factor<<<1,256>>>(d_isub_matrix,d_osub_matrix);//<----
cudaMemcpy(osub_matrix,d_osub_matrix,((THREAD_SIZE*THREAD_SIZE)*sizeof(float)),cudaMemcpyDeviceToHost);
array_plus=array_plus+THREAD_SIZE;
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
big_matrix2[x*THREAD_SIZE+array_plus+y]=osub_matrix[x*THREAD_SIZE+y];
}
cudaFree(isub_matrix);
cudaFree(osub_matrix);
cudaFree(d_osub_matrix);
cudaFree(d_isub_matrix);
}
//Stop the time
cudaEventRecord(stop,0);
cudaEventSynchronize(stop);
cudaEventElapsedTime(&time,start,stop);
//Free memory in GPU
I see loads of problem in this code.
You are not allocating memory for isub_matrix before copying the data from big_matrix to isub_matrix
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
isub_matrix[x*THREAD_SIZE+y]=big_matrix[x*ARRAY_SIZE+y];
}
You are not doing any cudaMemcpy from host to device for isub_matrix. After allocating memory on the device for isub_matrix, you need to copy the data.
I see that inside the while loop you are computing the same data.
//putting the big array's values into the sub-matrix
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
isub_matrix[x*THREAD_SIZE+y]=big_matrix[x*ARRAY_SIZE+y];
}
The for loop should be dependent on the array_plus.
I would suggest u to do this
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
isub_matrix[x*THREAD_SIZE+y]=big_matrix[(x+array_plus)*ARRAY_SIZE+y];
}
Based on the updated version:
The error I see is
float *d_osub_matrix;
cudaMalloc((void**)&d_osub_matrix,THREAD_SIZE*THREAD_SIZE*sizeof(float));
and call.
twiddle_factor<<<1,256>>>(d_isub_matrix,d_osub_matrix);
Then do
cudaMemcpy(osub_matrix,d_osub_matrix, ((THREAD_SIZE*THREAD_SIZE)*sizeof(float)),cudaMemcpyDeviceToHost);
By the way, it is not
twiddle_factor<<<1,256>>>(isub_matrix,osub_matrix);
It should be
twiddle_factor<<<1,256>>>(d_isub_matrix,osub_matrix);
Final and completed code:
int main(int argc, char** argv)
{
int array_plus=0,x,y;
int array_plus_x, array_plus_y;
float time;
//unsigned int memsize=sizeof(float)*THREAD_SIZE*THREAD_SIZE;
//bool array_rest;
cudaEvent_t start,stop;
float *d_isub_matrix,*d_osub_matrix;
float *big_matrix = new float[ARRAY_SIZE*ARRAY_SIZE];
float *big_matrix2 = new float[ARRAY_SIZE*ARRAY_SIZE];
float *isub_matrix = new float[THREAD_SIZE*THREAD_SIZE];
float *osub_matrix = new float[THREAD_SIZE*THREAD_SIZE];
//if the array's size is not compatible with the thread's size, it won't work.
//array_rest=(ARRAY_SIZE*ARRAY_SIZE)/(THREAD_SIZE*THREAD_SIZE);
//isub_matrix=(float*) malloc(memsize);
//osub_matrix=(float*) malloc(memsize);
if(((ARRAY_SIZE*ARRAY_SIZE)%(THREAD_SIZE*THREAD_SIZE)==0)&&(ARRAY_SIZE>=THREAD_SIZE))
{
//allocating space in CPU memory and GPU memory for the big matrix and its sub matrixes
//it has to be like this (lots of loops)
//populating the big array
for(x=0;x<ARRAY_SIZE;x++)
{
for(y=0;y<ARRAY_SIZE;y++)
big_matrix[x*ARRAY_SIZE+y]=rand()%10000;
}
//kind of loop for the big array
//Start counting the time of processing (everything)
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start,0);
for(array_plus_x = 0; array_plus_x < ARRAY_SIZE; array_plus_x += THREAD_SIZE)
for(array_plus_y = 0; array_plus_y < ARRAY_SIZE; array_plus_y += THREAD_SIZE)
{
//putting the big array's values into the sub-matrix
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
isub_matrix[x*THREAD_SIZE+y]=big_matrix[(x+array_plus_x)*ARRAY_SIZE+(y+array_plus_y)];
}
cudaMalloc((void**)&d_isub_matrix,THREAD_SIZE*THREAD_SIZE*sizeof(float));
cudaMalloc((void**)&d_osub_matrix,THREAD_SIZE*THREAD_SIZE*sizeof(float));
cudaMemcpy(d_isub_matrix,isub_matrix,((THREAD_SIZE*THREAD_SIZE)*sizeof(float)),cudaMemcpyHostToDevice);
//call the cuda kernel
dim3 block(32,32);
twiddle_factor<<<1,block>>>(d_isub_matrix,d_osub_matrix);//<----
cudaMemcpy(osub_matrix,d_osub_matrix,((THREAD_SIZE*THREAD_SIZE)*sizeof(float)),cudaMemcpyDeviceToHost);
for(x=0;x<THREAD_SIZE;x++)
{
for(y=0;y<THREAD_SIZE;y++)
big_matrix2[(x+array_plus_x)*ARRAY_SIZE+(y+array_plus_y)]=osub_matrix[x*THREAD_SIZE+y];
}
cudaFree(d_osub_matrix);
cudaFree(d_isub_matrix);
}
//Stop the time
cudaEventRecord(stop,0);
cudaEventSynchronize(stop);
cudaEventElapsedTime(&time,start,stop);
//Free memory in GPU
I think the problem is in the line.
cudaMemcpy(osub_matrix,isub_matrix,((THREAD_SIZE*THREAD_SIZE)*sizeof(float)),cudaMemcpyDeviceToHost);
This is because you allocate both osub_matrix and isub_matrix in the device.
下一篇: 使用CUDA内核获取堆栈溢出