Managing global count variable can you tell me the GPU version code for this simple CPU code?

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

Belowis My CPU code. I dont know how to manage the numEdges variable in GPU code.

so please suggest me.

[codebox]int MF_GetEdgeCount_X(unsigned char* edge, int srcW, int srcH, int* xIndexes, int* yIndexes)

{

int numEdges=0;

for (int k = 0; k < srcH - 2; k++) 

{

	int j = 1;

	while (j < srcW - 2) 

	{

		int x = srcW * k + j;

		if (edge[x] != edge[x-1] && edge[x] != 0) 

		{

			xIndexes[numEdges] = j;

			yIndexes[numEdges] = k;

			numEdges++;

		}

		j++;

	}

}

return numEdges;

}[/codebox]

2

I have written the GPU version code, see below.

I think, it should work. can you please tell me how can I copy the value of “numEdgesX” ( this is device variable) into a host variable? so that I can test the code.

[codebox]device int numEdgesX=0;

global

void GetEdgeCount_X(unsigned char* edge, int srcW, int srcH, int* xIndexes, int* yIndexes)

{

int idx = __umul24(blockIdx.x,blockDim.x)+threadIdx.x;

int idy = __umul24(blockIdx.y,blockDim.y)+threadIdx.y;

if ( (idx < srcW-2) && (idy < srcH-2) )

{

	int index = __umul24(idy,srcW)+idx;

	if (edge[index] != edge[index-1] && edge[index] != 0) 

	{

		xIndexes[numEdgesX] = idx;

		yIndexes[numEdgesX] = idy;

		numEdgesX++;

	}

}

}[/codebox]

3

your native code program hard to get the high performance with CUDA, because all threads depend on the value of variable “numEdgesX”.
you should take a look at the atomic function() and the “histogram” tutorial in nvidia sdk to get more idea .
good luck.