No double in program, but nvcc warns about it

Accelerated Computing CUDA CUDA Programming and Performance
1

Consider the following program:

#define DIMENSION 96

#define ITERATIONS 1e4

__constant__ float cmatrix[DIMENSION*DIMENSION];

__global__ void kernel_constant() {

	__shared__ float vin[DIMENSION];

	int tid = threadIdx.x;

	int iter, i, row;

	float temp;

	

	row = tid * DIMENSION;

	for(iter=0; iter < ITERATIONS; iter++) {

 	temp = 0;

 	for (i=0; i < DIMENSION; i++) {

 	temp += vin[i] * cmatrix[row + i];

 	}

 	__syncthreads();

	}

}

int main( void ) {

	kernel_constant<<<1, DIMENSION, 0, 0>>>();

}

when compiling it, nvcc complains:

$ nvcc constant.cu -o constant

ptxas /tmp/tmpxft_00000425_00000000-2_constant.ptx, line 61; warning : Double is not supported. Demoting to float

But there is no double anywhere to be seen. What am I doing wrong?

2

Must be something to do with the inbuilt functions you are using… Try sinf() instead of just sin()… Could be something on those lines…

3

Or any floating variable that you initialize must be explicitly initialized as float (ie float my_float=1.0f; instead as float my_float=1.0;)

4

Hmm. I’m not using any such functions.

Changing “temp = 0” to “temp = 0.0f” did not get rid of the warning either.

5

How about ITERATIONS? Why not write that out properly as an integer.

6

All the zeros looked just too ugly ;-). Thanks, that was the problem.