Nsight visual studio edition, running extremely slow in debug mode

Developer Tools Nsight Visual Studio Edition
1

Hi, I’m using newest nsight 5.6 and vs2017.

I am doing a simple matrix multiply example, it is all fine when running the program normally.

However, when I tried to run nsight debug, the time cost is extremely high, actually it is too long that I’d never finish it.

I am using legacy mode with my geforce 1060, a normal run of 600ms cost, but never managed to finish it under debug mode.

2

Hello, does the break point work for you? Looks like it just hangs, can you tell me your os and driver? I guess performing a clean installation for you driver may solve this issue.

3

Hey, I’m in windows 10, my driver is 398.11, which is installed by geforece experience.

The break point works fine, and also the step and others. But if I clear the break point or start without break point, it just keep running and running.

however, I can press pause, and it will actually stop in according code line.

Thank you.

4

Which code line you stop at? Actually it’s impossible to always pause at a same CUDA code line. I think this may be the root cause. Just a guess, dose your app stop at “__syncthreads();” ?

5

The global function looks like this

__global__ void CuMatrixMult(float *A, float *B, float *C)
{
	__shared__ float N[BLOCKDIM][BLOCKDIM];
	__shared__ float M[BLOCKDIM][BLOCKDIM];
	int Tx = threadIdx.x;
	int Ty = threadIdx.y;
	int Bx = blockIdx.x;
	int By = blockIdx.y;
	int i, j;
	float rN = A[Ty * WIDTH + Bx * BLOCKDIM + Tx];
	float rM = B[(By*BLOCKDIM + Ty)*WIDTH + Tx];
	float sum = 0;

	for (i = 0; i < WIDTH / BLOCKDIM; i++) {
		N[Ty][Tx] = rN;
		M[Ty][Tx] = rM;
		__syncthreads();
		for (j = 0; j < BLOCKDIM; j++) {
			rN = N[j][Tx];
			rM = M[Ty][j];
			sum += N[j][Tx] * M[Ty][j];
		}
		rN = A[((i + 1)*BLOCKDIM + Ty)*WIDTH + Bx * BLOCKDIM + Tx];
		rM = B[(By*BLOCKDIM + Ty)*WIDTH + (i + 1) * BLOCKDIM + Tx];

		__syncthreads();
	}

	C[(By*BLOCKDIM + Ty)*WIDTH + Bx * BLOCKDIM + Tx] = sum;
}

so I tried 4 times, 2 stop at “rN = N[j][Tx];”, 1 stop at “sum += N[j][Tx] * M[Ty][j];”, 1 stop at “__syncthreads();”(second one)

So I do not really know is there a pattern or not…

6

well I actually wait it finished this time, it cost me 4 minutes, but the result is not right…

7

Can you try the default matrixMul sample in cuda samples?

8

Hi, sorry it cost me some time, the results remains similar…

I don’t really know where is the problem…

let me run some more test when I am free…