I'm considering the following simple code in which I'm converting thrust::host_vector<int>::iterator h_temp_iterator = h_temp.begin(); and thrust::device_vector<int>::iterator d_temp_iterator = d_temp.begin(); to raw pointers.
To this end, I'm passing &(h_temp_iterator[0]) and &(d_temp_iterator[0]) to a function and a kernel, respectively. The former (CPU case) compiles, the latter (GPU case) not. The two cases should be in principle symmetric, so I do not understand the reason for the error message which is:
Error 1 error : no suitable conversion function from "thrust::device_ptr<int>" to "int *" exists
The configurations are:
Windows 7, Visual Studio 2010, CUDA 7.5, compiling for the 3.5 architecture.
Windows 10, Visual Studio 2013, CUDA 8.0, compiling for the 5.2 architecture.
CODE
#include <thrust\host_vector.h>
#include <thrust\device_vector.h>
__global__ void testKernel(int *a, const int N)
{
int i = threadIdx.x;
if (i >= N) return;
a[i] = 2;
}
void testFunction(int *a, const int N)
{
for (int i = 0; i < N; i++) a[i] = 2;
}
int main()
{
const int N = 10;
thrust::host_vector<int> h_temp(N);
thrust::device_vector<int> d_temp(N);
thrust::host_vector<int>::iterator h_temp_iterator = h_temp.begin();
thrust::device_vector<int>::iterator d_temp_iterator = d_temp.begin();
testFunction(&(h_temp_iterator[0]), N);
testKernel<<<1, N>>>(&(d_temp_iterator[0]), N);
for (int i = 0; i < N; i++) printf("%i %i\n", i, h_temp[i]);
return 0;
}
Following talonmies' comments, the solution is to pass
thrust::raw_pointer_cast(&d_temp_iterator[0])
and not
&d_temp_iterator[0]
In the following, the fully working code
#include <thrust\host_vector.h>
#include <thrust\device_vector.h>
__global__ void testKernel(int *a, const int N)
{
int i = threadIdx.x;
if (i >= N) return;
a[i] = 2;
printf("GPU %i %i\n", i, a[i]);
}
void testFunction(int *a, const int N)
{
for (int i = 0; i < N; i++) {
a[i] = 2;
printf("CPU %i %i\n", i, a[i]);
}
}
int main()
{
const int N = 10;
thrust::host_vector<int> h_temp(N);
thrust::device_vector<int> d_temp(N);
thrust::host_vector<int>::iterator h_temp_iterator = h_temp.begin();
thrust::device_vector<int>::iterator d_temp_iterator = d_temp.begin();
int *temp = thrust::raw_pointer_cast(&d_temp_iterator[0]);
testFunction(&(h_temp_iterator[0]), N);
testKernel<<<1, N>>>(temp, N);
return 0;
}
As you are using thrust, another prettier solution is to get the pointer with data() and cast it to a raw pointer:
thrust::raw_pointer_cast(d_temp_iterator.data())
Related
I met a problem about using cudaMemcpy with cudaMemcpyDeviceToHost.
There is a struct which have a pointer int* a, It will malloc in the kernel function.
And then I need copy this int* a to host memory.
My question is: I didn't know how it can not work by using cudaMemcpy.
There my codes:
#include <cuda_runtime.h>
#include <stdio.h>
typedef struct { int n, m; int *a; } myst;
__global__ void xthread(myst *st)
{
unsigned int idx = blockIdx.x*blockDim.x + threadIdx.x;
myst *mst = &st[idx];
mst->n = idx;
mst->m = idx+1;
mst->a = (int *)malloc((mst->m)*sizeof(int));
mst->a[0] = idx;
}
int main(int argc,char **argv)
{
dim3 dimGrid(1);
dim3 dimBlock(2);
myst *mst = NULL;
myst *hst = (myst *)malloc(2 * sizeof(myst));
cudaMalloc(&mst, 2 * sizeof(myst));
xthread<<<dimGrid, dimBlock>>>(mst);
cudaDeviceSynchronize();
cudaMemcpy(&hst[0],&mst[0],sizeof(myst),cudaMemcpyDeviceToHost);
cudaMemcpy(&hst[1],&mst[1],sizeof(myst),cudaMemcpyDeviceToHost);
int *pInt1 = (int *)malloc((hst[0].m)*sizeof(int)) ;
int *pInt2 = (int *)malloc((hst[1].m)*sizeof(int)) ;
cudaMemcpy(pInt1, hst[0].a, (hst[0].m)*sizeof(int), cudaMemcpyDeviceToHost);
cudaMemcpy(pInt2, hst[1].a, (hst[1].m)*sizeof(int), cudaMemcpyDeviceToHost);
printf("%d\t%d\t%d\n",hst[0].n,hst[0].m, pInt1[0]);
printf("%d\t%d\t%d\n",hst[1].n,hst[1].m, pInt2[0]);
free(pInt1);
free(pInt2);
return 0;
}
The codes will go warning about "Cuda API error detected: cudaMemcpy returned (0xb)"
I saw a similar question : copy data which is allocated in device from device to host
But it seem that can not solve my problem.
Thx.
Alright, I work it out with a stupid way (-.-!!).
While return form the kernel function, I count how many space I have to malloc in Host and Device, and cudaMalloc again a big space . Next, in other kernel function named ythread, copy the data which in the Heap to the big space.
typedef struct { int n, m; int *a; } myst;
__global__ void xthread(myst *st) {
unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
myst *mst = &st[idx];
mst->n = idx;
mst->m = idx + 1;
mst->a = (int *) malloc((mst->m) * sizeof(int));
for (int i = 0; i < mst->m; i++) {
mst->a[i] = idx + 900 + i * 10;
}
}
__global__ void ythread(myst *st, int *total_a) {
unsigned int idx = blockIdx.x*blockDim.x + threadIdx.x;
myst *mst = &st[idx];
int offset=0;
for(int i=0; i<idx; i++) {
offset += st[i].m;
}
for(int i=0; i<mst->m; i++) {
total_a[offset+i] = mst->a[i];
}
}
int main(int argc,char **argv) {
dim3 dimGrid(1);
dim3 dimBlock(2);
myst *mst = NULL;
cudaMalloc((void**)&mst, dimBlock.x * sizeof(myst));
xthread<<<dimGrid, dimBlock>>>(mst);
cudaDeviceSynchronize();
myst *hst = (myst *)malloc(dimBlock.x * sizeof(myst));
cudaMemcpy(hst, mst, dimBlock.x*sizeof(myst),cudaMemcpyDeviceToHost);
int t_size = 0;
for(int i=0; i<dimBlock.x; i++) {
t_size += hst[i].m;
}
printf("t_size:%d\n", t_size);
int * t_a_h = (int *)malloc(t_size*sizeof(int));
int * t_a_d = NULL;
cudaMalloc((void**)&t_a_d, t_size*sizeof(int));
ythread<<<dimGrid, dimBlock>>>(mst, t_a_d);
cudaDeviceSynchronize();
cudaMemcpy(t_a_h, t_a_d, t_size*sizeof(int),cudaMemcpyDeviceToHost);
for(int i=0; i<t_size; i++) {
printf("t_a_h[%d]:%d\n", i, t_a_h[i]);
}
free(t_a_h);
cudaFree(mst);
cudaFree(t_a_d);
return 0;
}
Emmmmmm, it work, but I think there is a better way to solve this problem.
I'm considering the following simple code in which I'm converting thrust::host_vector<int>::iterator h_temp_iterator = h_temp.begin(); and thrust::device_vector<int>::iterator d_temp_iterator = d_temp.begin(); to raw pointers.
To this end, I'm passing &(h_temp_iterator[0]) and &(d_temp_iterator[0]) to a function and a kernel, respectively. The former (CPU case) compiles, the latter (GPU case) not. The two cases should be in principle symmetric, so I do not understand the reason for the error message which is:
Error 1 error : no suitable conversion function from "thrust::device_ptr<int>" to "int *" exists
The configurations are:
Windows 7, Visual Studio 2010, CUDA 7.5, compiling for the 3.5 architecture.
Windows 10, Visual Studio 2013, CUDA 8.0, compiling for the 5.2 architecture.
CODE
#include <thrust\host_vector.h>
#include <thrust\device_vector.h>
__global__ void testKernel(int *a, const int N)
{
int i = threadIdx.x;
if (i >= N) return;
a[i] = 2;
}
void testFunction(int *a, const int N)
{
for (int i = 0; i < N; i++) a[i] = 2;
}
int main()
{
const int N = 10;
thrust::host_vector<int> h_temp(N);
thrust::device_vector<int> d_temp(N);
thrust::host_vector<int>::iterator h_temp_iterator = h_temp.begin();
thrust::device_vector<int>::iterator d_temp_iterator = d_temp.begin();
testFunction(&(h_temp_iterator[0]), N);
testKernel<<<1, N>>>(&(d_temp_iterator[0]), N);
for (int i = 0; i < N; i++) printf("%i %i\n", i, h_temp[i]);
return 0;
}
Following talonmies' comments, the solution is to pass
thrust::raw_pointer_cast(&d_temp_iterator[0])
and not
&d_temp_iterator[0]
In the following, the fully working code
#include <thrust\host_vector.h>
#include <thrust\device_vector.h>
__global__ void testKernel(int *a, const int N)
{
int i = threadIdx.x;
if (i >= N) return;
a[i] = 2;
printf("GPU %i %i\n", i, a[i]);
}
void testFunction(int *a, const int N)
{
for (int i = 0; i < N; i++) {
a[i] = 2;
printf("CPU %i %i\n", i, a[i]);
}
}
int main()
{
const int N = 10;
thrust::host_vector<int> h_temp(N);
thrust::device_vector<int> d_temp(N);
thrust::host_vector<int>::iterator h_temp_iterator = h_temp.begin();
thrust::device_vector<int>::iterator d_temp_iterator = d_temp.begin();
int *temp = thrust::raw_pointer_cast(&d_temp_iterator[0]);
testFunction(&(h_temp_iterator[0]), N);
testKernel<<<1, N>>>(temp, N);
return 0;
}
As you are using thrust, another prettier solution is to get the pointer with data() and cast it to a raw pointer:
thrust::raw_pointer_cast(d_temp_iterator.data())
I have 2000 2D-arrays (each array is 1000x1000). I need to compute the mean of each one and put the result in one 2000 vector.
I tried to do that by calling the kernel for each 2D-array, but it is naive, and I want to do the computation once.
What I have been doing is a kernel for one 2D-array. I want to make my kernel do this for 2000 2D-arrays, but in one kernel.
#include <stdio.h>
#include <cuda.h>
#include <time.h>
void init_mat(float *a, const int N, const int M);
void print_mat(float *a, const int N, const int M, char *d);
void print_array(float *a, const int N, char *d);
const int threadsPerBlock=256;
__global__
void kernel(float *mat, float *out, const int N, const int M){
__shared__ float cache[threadsPerBlock];
int tid=threadIdx.x+blockIdx.x*blockDim.x;
int cacheIndex = threadIdx.x;
float sum=0;
if (tid<M) {
for(int i=0; i<N; i++)
sum += mat[(i*M)+tid];
cache[cacheIndex] = sum;
out[tid] =cache[cacheIndex];
}
__syncthreads();
int i = blockDim.x/2;
while(i != 0) {
if(cacheIndex<i)
cache[cacheIndex]+= cache[cacheIndex +i];
__syncthreads();
I /= 2;
}
if (cacheIndex == 0)
out[blockIdx.x]=cache[0];
}
int main (void) {
srand( time(NULL) );
float *a, *b, *c;
float *dev_a, *dev_b, *dev_c;
int N=1000;
int M=1000;
b=(float*)malloc(sizeof(float)*N*M);
c=(float*)malloc(sizeof(float)*M);
init_mat(b, N, M);
printf("<<<<<<<<<< initial data:\n");
print_mat(b, N, M, "matrix");
cudaMalloc((void**)&dev_b, sizeof(float)*N*M);
cudaMalloc((void**)&dev_c, sizeof(float)*M);
cudaMemcpy(dev_b, b, sizeof(float)*N*M, cudaMemcpyHostToDevice);
printf("\n\nRunning Kernel...\n\n");
kernel<<<M/256+1, 256>>>(dev_b, dev_c, N, M);
cudaMemcpy(c, dev_c, sizeof(float)*M, cudaMemcpyDeviceToHost);
cudaFree(dev_a);
cudaFree(dev_b);
cudaFree(dev_c);
printf(">>>>>>>>>> final data:\n");
print_array(c, M, "out-vector");
};
void init_mat(float *a, const int N, const int M) {
int i, j;
for(i=0; i<N; i++)
for(j=0; j<M; j++)
a[i*M+j] = rand() % 100 + 1;
}
void print_mat(float *a, const int N, const int M, char *d) {
int i, j;
for(i=0; i<N; i++){
printf("\n%s[%d]:", d, i);
for (j=0; j<M; j++)
printf("\t%6.4f", a[i*M+j]);
}
printf("\n");
}
void print_array(float *a, const int N, char *d) {
int i;
for(i=0; i<N; i++)
printf("\n%s[%d]: %f",d, i, a[i]);
printf("\n");
}
For a reasonably large number of arrays (e.g. 2000) and reasonably large sized arrays (e.g. 2000), the GPU can be fairly efficient if we assign a block to perform the sum reduction (and mean calculation) for each array. This means if you have 2000 arrays we will launch 2000 blocks.
In order to handle arbitrary sized arrays with a fixed number of threads per block, we will use an idea like the grid-striding loop but instead we will cause each block to use a block-striding loop to load all the data associated with a particular array. This means the threads of each block will "stride" through the assigned array, to load all the elements of that array.
Apart from this, the main reduction operation is similar to what you have written, and calculation of the mean is trivial this way - we can calculate the mean before writing the result to global memory, once we have the sum calculated via reduction.
Here is a worked example. If you compile with -DMEAN the code will output the mean of each array. If you omit that compile switch, the code will output the sum of each array. Let N be the number of arrays, and let K be the size of each array.
$ cat t1285.cu
#include <stdio.h>
const size_t N = 1000; // number of arrays
const size_t K = 1000; // size of each array
const int nTPB = 256; // number of threads per block, must be a power-of-2
typedef float mytype; // type of data to be summed
// produce the sum or mean of each array
template <typename T>
__global__ void breduce(const T * __restrict__ idata, T * __restrict__ odata, const int bsize){
__shared__ T sdata[nTPB];
T sum = 0;
//block-striding loop
size_t offset = blockIdx.x*bsize + threadIdx.x;
while (offset < (blockIdx.x+1)*bsize){
sum += idata[offset];
offset += blockDim.x;}
sdata[threadIdx.x] = sum;
__syncthreads();
//shared memory reduction sweep
for (int i = nTPB>>1; i > 0; i>>=1){
if (threadIdx.x < i) sdata[threadIdx.x] += sdata[threadIdx.x+i];
__syncthreads();}
// write output sum for this block/array
#ifndef MEAN
if (!threadIdx.x) odata[blockIdx.x] = sdata[0];
#else
if (!threadIdx.x) odata[blockIdx.x] = sdata[0]/bsize;
#endif
}
int main(){
mytype *h_idata, *h_odata, *d_idata, *d_odata;
h_idata=(mytype *)malloc(N*K*sizeof(mytype));
h_odata=(mytype *)malloc(N*sizeof(mytype));
cudaMalloc(&d_idata, N*K*sizeof(mytype));
cudaMalloc(&d_odata, N*sizeof(mytype));
for (size_t i = 0; i < N; i++)
for (size_t j = 0; j < K; j++)
h_idata[i*K+j] = 1 + (i&1); // fill alternating arrays with 1 and 2
memset(h_odata, 0, N*sizeof(mytype)); // zero out
cudaMemset(d_odata, 0, N*sizeof(mytype)); // zero out
cudaMemcpy(d_idata, h_idata, N*K*sizeof(mytype), cudaMemcpyHostToDevice);
breduce<<<N, nTPB>>>(d_idata, d_odata, K);
cudaMemcpy(h_odata, d_odata, N*sizeof(mytype), cudaMemcpyDeviceToHost);
// validate
for (size_t i = 0; i < N; i++)
#ifndef MEAN
if (h_odata[i] != (K*(1 + (i&1)))) {printf("mismatch at %d, was: %f, should be: %f\n", i, (float)h_odata[i], (float)(K*(1 + (i&1)))); return 1;}
#else
if (h_odata[i] != ((1 + (i&1)))) {printf("mismatch at %d, was: %f, should be: %f\n", i, (float)h_odata[i], (float)((1 + (i&1)))); return 1;}
#endif
return 0;
}
$ nvcc -arch=sm_35 -o t1285 t1285.cu -DMEAN
$ cuda-memcheck ./t1285
========= CUDA-MEMCHECK
========= ERROR SUMMARY: 0 errors
$ nvcc -arch=sm_35 -o t1285 t1285.cu
$ cuda-memcheck ./t1285
========= CUDA-MEMCHECK
========= ERROR SUMMARY: 0 errors
$
Does anyone know what is the maximum supported size for cub::scan ? I got core dump for input sizes over 500 million. I wanted to make sure I'm not doing anything wrong...
Here is my code:
#define CUB_STDERR
#include <stdio.h>
#include "cub/util_allocator.cuh"
#include "cub/device/device_scan.cuh"
#include <sys/time.h>
using namespace cub;
bool g_verbose = false; // Whether to display input/output to console
CachingDeviceAllocator g_allocator(true); // Caching allocator for device memory
typedef int mytype;
/**
* Solve inclusive-scan problem
*/
static void solve(mytype *h_in, mytype *h_cpu, int n)
{
mytype inclusive = 0;
for (int i = 0; i < n; ++i) {
inclusive += h_in[i];
h_cpu[i] = inclusive;
}
}
static int compare(mytype *h_cpu, mytype *h_o, int n)
{
for (int i = 0; i < n; i++) {
if (h_cpu[i] != h_o[i]) {
return i + 1;
}
}
return 0;
}
/**
* Main
*/
int main(int argc, char** argv)
{
cudaSetDevice(0);
struct timeval start, end;
int num_items = 1073741824;
const int repetitions = 5;
mytype *h_in, *h_out, *h_cpu;
const int size = num_items * sizeof(mytype);
// Allocate host arrays
h_in = (mytype *)malloc(size);
h_out = (mytype *)malloc(size);
h_cpu = (mytype *)malloc(size);
// Initialize problem and solution
for (int i = 0; i < num_items; i++) {
h_in[i] = i;
h_out[i] = 0;
h_cpu[i] = 0;
}
solve(h_in, h_cpu, num_items);
// Allocate problem device arrays
mytype *d_in = NULL;
CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(mytype) * num_items));
// Initialize device input
CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(mytype) * num_items, cudaMemcpyHostToDevice));
// Allocate device output array
mytype *d_out = NULL;
CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(mytype) * num_items));
// Allocate temporary storage
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
CubDebugExit(DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items));
CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));
// Run
gettimeofday(&start, NULL);
for (long i = 0; i < repetitions; i++)
DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);
cudaThreadSynchronize();
gettimeofday(&end, NULL);
double ctime = end.tv_sec + end.tv_usec / 1000000.0 - start.tv_sec - start.tv_usec / 1000000.0;
cudaMemcpy(h_out, d_out, sizeof(mytype) * num_items, cudaMemcpyDeviceToHost);
int cmp = compare(h_cpu, h_out, num_items);
printf("%d\t", num_items);
if (!cmp)
printf("\t%7.4fs \n", ctime);
printf("\n");
if (h_in) delete[] h_in;
if (h_out) delete[] h_out;
if (h_cpu) delete[] h_cpu;
if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));
if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));
if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));
printf("\n\n");
return 0;
}
The problem is here:
const int size = num_items * sizeof(mytype);
And it can be fixed by changing it to:
const size_t size = num_items * sizeof(mytype);
The value of num_items in the code is over 1 Billion. When we multiply that by sizeof(mytype) we are multiplying it by 4, so the result is over 4 Billion. This value cannot be stored in an int variable. If you try to use it anyway like that, then your subsequent host code will do bad things. This problem (the core dump) actually has nothing to do with CUDA. The code would core dump if you removed all the CUB elements.
When I modify the line of code above, and compile for the correct GPU (e.g. -arch=sm_35 in my case, or -arch=sm_52 for a Titan X GPU), then I get the correct answer (and no seg fault/core dump).
In general, the correct starting point when chasing a seg fault/core dump type error, is to recognize that this error arises from host code and you should attempt to localize the exact line of source code that is generating this error. This can be done trivially/tediously by putting many printf statements in your code, until you identify the line of your code after which you don't see any printf output, or by using a host code debugger, such as gdb on linux.
Also note that this code as written will require slightly more than 12GB of memory on the host, and slightly more than 8GB of memory on the GPU, so it will only run properly in such settings.
For reference, here is the fixed code (based on what OP posted here):
#define CUB_STDERR
#include <stdio.h>
#include "cub/util_allocator.cuh"
#include "cub/device/device_scan.cuh"
#include <sys/time.h>
using namespace cub;
bool g_verbose = false; // Whether to display input/output to console
CachingDeviceAllocator g_allocator(true); // Caching allocator for device memory
typedef int mytype;
/**
* Solve inclusive-scan problem
*/
static void solve(mytype *h_in, mytype *h_cpu, int n)
{
mytype inclusive = 0;
for (int i = 0; i < n; ++i) {
inclusive += h_in[i];
h_cpu[i] = inclusive;
}
}
static int compare(mytype *h_cpu, mytype *h_o, int n)
{
for (int i = 0; i < n; i++) {
if (h_cpu[i] != h_o[i]) {
return i + 1;
}
}
return 0;
}
/**
* Main
*/
int main(int argc, char** argv)
{
cudaSetDevice(0);
struct timeval start, end;
int num_items = 1073741824;
const int repetitions = 5;
mytype *h_in, *h_out, *h_cpu;
const size_t size = num_items * sizeof(mytype);
// Allocate host arrays
h_in = (mytype *)malloc(size);
h_out = (mytype *)malloc(size);
h_cpu = (mytype *)malloc(size);
// Initialize problem and solution
for (int i = 0; i < num_items; i++) {
h_in[i] = i;
h_out[i] = 0;
h_cpu[i] = 0;
}
solve(h_in, h_cpu, num_items);
// Allocate problem device arrays
mytype *d_in = NULL;
CubDebugExit(g_allocator.DeviceAllocate((void**)&d_in, sizeof(mytype) * num_items));
// Initialize device input
CubDebugExit(cudaMemcpy(d_in, h_in, sizeof(mytype) * num_items, cudaMemcpyHostToDevice));
// Allocate device output array
mytype *d_out = NULL;
CubDebugExit(g_allocator.DeviceAllocate((void**)&d_out, sizeof(mytype) * num_items));
// Allocate temporary storage
void *d_temp_storage = NULL;
size_t temp_storage_bytes = 0;
CubDebugExit(DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items));
CubDebugExit(g_allocator.DeviceAllocate(&d_temp_storage, temp_storage_bytes));
// Run
gettimeofday(&start, NULL);
for (long i = 0; i < repetitions; i++)
DeviceScan::InclusiveSum(d_temp_storage, temp_storage_bytes, d_in, d_out, num_items);
cudaThreadSynchronize();
gettimeofday(&end, NULL);
double ctime = end.tv_sec + end.tv_usec / 1000000.0 - start.tv_sec - start.tv_usec / 1000000.0;
cudaMemcpy(h_out, d_out, sizeof(mytype) * num_items, cudaMemcpyDeviceToHost);
int cmp = compare(h_cpu, h_out, num_items);
printf("%d\t", num_items);
if (!cmp)
printf("\t%7.4fs \n", ctime);
printf("\n");
if (h_in) delete[] h_in;
if (h_out) delete[] h_out;
if (h_cpu) delete[] h_cpu;
if (d_in) CubDebugExit(g_allocator.DeviceFree(d_in));
if (d_out) CubDebugExit(g_allocator.DeviceFree(d_out));
if (d_temp_storage) CubDebugExit(g_allocator.DeviceFree(d_temp_storage));
printf("\n\n");
return 0;
}
I have an array of numbers as {1,2,3,4,5,6,7,8,9,10} and I want to separate even and odd numbers as:
even = {2,4,6,8}
and:
odd = {1,3,5,7}
I am aware of atomic operations in CUDA, and also aware that the output is not expected to suffer from race conditions. I don't want to use atomic operations. How can I achieve this without using atomic keywords?
CODE:
#include <stdio.h>
#include <cuda.h>
// Kernel that executes on the CUDA device
__global__ void square_array(float *total,float *even,float *odd, int N)
{
int idx = blockIdx.x * blockDim.x + threadIdx.x;
int a=total[idx];
if ((a%2)==0)
{
for (int i=0;i<=idx;i++)
{
int b = even[i];
if(b==0)
{
even[i] = total[idx];
break;
}
}
}
else
{
for (int i=0;i<idx;i++)
{
int c = odd[i];
odd[i] = total[idx];
break;
}
}
}
// main routine that executes on the host
int main(void)
{
float *total_h,*even_h, *odd_h,*total_d, *even_d,*odd_d; // Pointer to host & device arrays
const int N = 10; // Number of elements in arrays
size_t size = N * sizeof(float);
total_h = (float *)malloc(size); // Allocate array on host
even_h = (float *)malloc(size); // Allocate array on host
odd_h = (float *)malloc(size); // Allocate array on host
cudaMalloc((void **) &total_d, size);
cudaMalloc((void **) &even_d, size);
cudaMemset(even_d,0,size);
cudaMalloc((void **) &odd_d, size); // Allocate array on device
cudaMemset(odd_d,0,size);
// Initialize host array and copy it to CUDA device
for (int i=0; i<N; i++) total_h[i] = (float)i+1;
cudaMemcpy(total_d, total_h, size, cudaMemcpyHostToDevice);
// Do calculation on device:
square_array <<< 1,10 >>> (total_d,even_d,odd_d, N);
// Retrieve result from device and store it in host array
cudaMemcpy(even_h, even_d, sizeof(float)*N, cudaMemcpyDeviceToHost);
cudaMemcpy(odd_h, odd_d, sizeof(float)*N, cudaMemcpyDeviceToHost);
// Print results
printf("total Numbers\n");
for (int i=0; i<N; i++) printf("%f\n",total_h[i]);
printf("EVEN Numbers\n");
for (int i=0; i<N; i++) printf("%f\n",even_h[i]);
printf("ODD Numbers\n");
for (int i=0; i<N; i++) printf("%f\n",odd_h[i]);
// Cleanup
free(total_h);
free(even_h);
free(odd_h);
cudaFree(total_d);
cudaFree(even_d);
cudaFree(odd_d);
}
OUTPUT:
As suggested by Jared Hoberock, it would be much more easy to use the efficient partitioning algorithm available in CUDA Thrust instead of starting the development of a partitioning routine of your own. Below, please find a complete worked example.
#include <thrust\device_vector.h>
#include <thrust\partition.h>
#include <thrust\execution_policy.h>
struct is_even { __host__ __device__ bool operator()(const int &x) { return (x % 2) == 0; } };
void main() {
const int N = 10;
thrust::host_vector<int> h_data(N);
for (int i=0; i<N; i++) h_data[i] = i;
thrust::device_vector<int> d_data(h_data);
thrust::device_vector<int> d_evens(N/2);
thrust::device_vector<int> d_odds(N/2);
thrust::partition_copy(d_data.begin(), d_data.end(), d_evens.begin(), d_odds.begin(), is_even());
printf("Even numbers\n");
for (int i=0; i<N/2; i++) {
int val = d_evens[i];
printf("evens[%i] = %i\n",i,val);
}
printf("Odd numbers\n");
for (int i=0; i<N/2; i++) {
int val = d_odds[i];
printf("odds[%i] = %i\n",i,val);
}
}