为什么这个cuda程序在运行后貌似没有更新h指向的数组值？

[superwavelet]

代码如下：

#include <X11/Xlib.h> // X11 library headers
#include <X11/Xutil.h>
#include <X11/Xos.h>
#include <stdio.h>

__global__ void ComputMatrix (float *g, float *h,int M) 
{
	int col = blockIdx.x*blockDim.x+threadIdx.x;
	int row =blockIdx.y*blockDim.y+threadIdx.y;

  if(col < M && row < M&&row!=0&&row!=M-1&&col!=0&&col!=M-1) 
  {
    g[row*M+col]=0.25*(h[row*M+col-1]+h[row*M+col+1]+h[(row-1)*M+col]+h[(row+1)*M+col]);
	h[row*M+col]=g[row*M+col];
  }

}


int main (int argc, char **argv ) {

	int limit=10000;
	float *h;
	float *g;
	float *dev_g, *dev_h;
	int iBlock=100;

    int N;
	printf("Please input the length:\n"); 
	scanf("%d",&N); 

	h = (float *)malloc(N*N * sizeof(float));
	g = (float *)malloc(N*N * sizeof(float));
    int X_RESN=N;/* x resolution */
    int Y_RESN=N; /* y resolution */

	/* --------------------------- X11 graphics setup ------------------------------ */
	Window win; /* initialization for a window */
	unsigned int width, height, /* window size */
		win_x,win_y, /* window position */
		border_width, /* border width in pixels */
		display_width, display_height, /* size of screen */
		screen; /* which screen */
	char *window_name = "My graphics program", *display_name = NULL;
	GC gc;
	unsigned long valuemask = 0;
	XGCValues values;
	Display *display;
	XSizeHints size_hints;
	Pixmap bitmap;
	//XPoint points[800];
	FILE *fp, *fopen ();
	char str[100];

	XSetWindowAttributes attr[1];
	if ( (display = XOpenDisplay (display_name)) == NULL ) { /* connect to Xserver */
		fprintf (stderr, "drawon: cannot connect to X server %s\n",XDisplayName (display_name) );
		exit (-1);
	}
	screen = DefaultScreen (display); /* get screen size */
	display_width = DisplayWidth (display, screen);
	display_height = DisplayHeight (display, screen);
	width = X_RESN; /* set window size */
	height = Y_RESN;
	win_x = 0; win_y = 0; /* set window position */
	border_width = 4; /* create opaque window */
	win = XCreateSimpleWindow (display, RootWindow (display, screen),
		win_x, win_y, width, height, border_width,
		BlackPixel (display, screen), WhitePixel (display, screen));
	size_hints.flags = USPosition|USSize;
	size_hints.x = win_x;
	size_hints.y = win_y;
	size_hints.width = width;
	size_hints.height = height;
	size_hints.min_width = 300;
	size_hints.min_height = 300;
	XSetNormalHints (display, win, &size_hints);
	XStoreName(display, win, window_name);
	gc = XCreateGC (display, win, valuemask, &values); /* create graphics context */
	XSetBackground (display, gc, WhitePixel (display, screen));
	XSetForeground (display, gc, BlackPixel (display, screen));
	XSetLineAttributes (display, gc, 1, LineSolid, CapRound, JoinRound);
	attr[0].backing_store = Always;
	attr[0].backing_planes = 1;
	attr[0].backing_pixel = BlackPixel(display, screen);
	XChangeWindowAttributes(display, win, CWBackingStore | CWBackingPlanes | CWBackingPixel, attr);
	XMapWindow (display, win);
	XSync(display, 0);

	cudaMalloc((void**)&dev_g,N *N* sizeof(float));
	cudaMalloc((void**)&dev_h,N *N* sizeof(float));

[superwavelet]

第二部分如下：

    int i=0;
	int j=0;

    for (i=0;i<N*N;i++)
    {
		h[i]=0;
		g[i]=0;
    }

	for (i=0;i<N;i++)
	{
		h[i]=20;
		g[i]=20;
		h[N*(N-1)+i]=20;
		g[N*(N-1)+i]=20;
		h[i*N]=20;
		g[i*N]=20;
		g[i*N+N-1]=20;
		h[i*N+N-1]=20;
	}

	for (i=N*3/10;i<(N-N*3/10+1);i++)
	{
		h[i]=100;
		g[i]=100;
	}

	cudaMemcpy(dev_g, g , N*N*sizeof(float),cudaMemcpyHostToDevice);
	cudaMemcpy(dev_h, h , N*N*sizeof(float),cudaMemcpyHostToDevice);

	dim3 dimBlock;
	dimBlock.x = iBlock;
	dimBlock.y = iBlock;

	dim3 dimGrid;
	dimGrid.x = N / dimBlock.x;
	dimGrid.y = N / dimBlock.y;

	for (i=0;i<limit;i++)
	{
		ComputMatrix<<<dimGrid,dimBlock>>>(dev_g,dev_h,N);
		if (i%100==0)
		{

        cudaMemcpy(h,dev_h,N*N*sizeof(float),cudaMemcpyDeviceToHost);

	XClearWindow(display,win);

	int iteration=0;
	for (iteration=0;iteration<limit;iteration++)
	{

		for (i=0;i<N*N;i++)
		{
				if (90<h[i]&&h[i]<=100)
				{
					XSetForeground(display,gc,(long) 0x7F0000);
				}
				if (80<h[i]&&h[i]<=90)
				{
					XSetForeground(display,gc,(long) 0xB22222);
				}
				if (70<h[i]&&h[i]<=80)
				{
					XSetForeground(display,gc,(long) 0xDC143C);
				}
				if (60<h[i]&&h[i]<=70)
				{
					XSetForeground(display,gc,(long) 0xFF0000);
				}
				if (50<h[i]&&h[i]<=60)
				{
					XSetForeground(display,gc,(long) 0xFF6347);
				}
				if (40<h[i]&&h[i]<=50)
				{
					XSetForeground(display,gc,(long) 0xFF4500);
				}
				if (30<h[i]&&h[i]<=40)
				{
					XSetForeground(display,gc,(long) 0xFFA500);

				}
				if (20<h[i]&&h[i]<=30)
				{
					XSetForeground(display,gc,(long) 0xFFD700);
				}
				if (10<h[i]&&h[i]<=20)
				{
					XSetForeground(display,gc,(long) 0xFFFF00);
				}

				if (0<=h[i]&&h[i]<=10)
				{
					XSetForeground(display,gc,(long) 0x00FF00);
				}

				XDrawPoint(display,win,gc,i%N,i/N);
				XFlush(display);
 		}
	//sleep(1);
     }
	}
	}
}

[superwavelet]

为什么？把block和grid中block的个数弄很大还是一样，最终是h指向的数组没得到更新。

[superwavelet]

还是没有解决，有哪位哥哥姐姐帮帮忙？谢啦