attempts to free array inside kernel

Dear all;
please read my program

#include "cuda.h"
#include "cuda_runtime.h"
#include "device_functions.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <stdlib.h>
#include <math.h>
#define N  4096
#define P 4

__global__ void myKernel(int *a, int *temp, int maxId)
	int threadId = threadIdx.x + blockIdx.x*blockDim.x;
	int v1 = maxId;

	for (int i = (v1/2); i >0; i--){
		if (0 <= threadId && threadId < i){
			int rec_num = (2 ^ (v1 - i)) * 2048;
                int index1 =rec_num * (threadId);
		int endIndex1 = index1 + 1;
		int index2 = endIndex1 + 1;
		int endIndex2 = index2 + 1;
		int targetIndex = index1;
		int sortedSize = 0;
		while (index1 <= endIndex1 && index2 <= endIndex2){

			if (a[index1] <= a[index2]){
				temp[targetIndex] = a[index1];
				temp[targetIndex] = a[index2];




int main()
	int *a;
	int *dev_a;
	int *dev_temp;
        a = (int *)malloc(N*sizeof(int));
        cudaMalloc((void **)&dev_a, N*sizeof(int));
	cudaMalloc((void **)&dev_temp, N*sizeof(int));
	for (int i = 0; i < N; i++)
		int num = rand() % 100;
		a[i] = num;

cudaMemcpy(dev_a, a, N*sizeof(int), cudaMemcpyHostToDevice);
myKernel << <P + 1, 1024 >> >(dev_a, dev_temp,4);
cudaMemcpy(a, dev_temp, N*sizeof(int), cudaMemcpyDeviceToHost);

i had allocated memory(array) on device inside main, and i passed it through kernel program runs but not with the expected result.
when the kernel not using the ‘for loop’, it works well
but with wrong result with ‘for loop’.

from my trials on this program, I found that at each iteration of the ‘for loop’ must use an empty ‘temp’ array ‘as if it passed for the first time though the kernel call’.
how can i free temp at end of for loop before the next iteration??

Your goal doesn’t really make sense. The temp array is not initialized to anything. The concept of “empty” is undefined. If you’re depending on an uninitialized area to be in a certain state to make things work, your code is broken.

Furthermore, your code as posted will not compile. Even beyond the basic missing curly-braces at the end, your kernel definition does not contain an appropriate number of close braces. It’s not clear where you intend the if statement to end and where you intend the for-loop to end.