Date structure problem

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

I have a following program, and when I compile it, I get an error like “error: identifier “__T20” is undefined”. Similar code can be compiled for CPU. Any advices?

struct Vec
{
float x, y, z;
Vec( float x_ = 0, float y_ = 0, float z_ = 0 )
{
x = x_;
y = y_;
z = z_;
}
};

global void kernel()
{
Vec t[10];
return;
}

int main()
{
dim3 block( 16, 16 );
dim3 grid( 1, 1 );

kernel <<< grid, block>>> ();

return 0;

}

Thanks in advance!

2

Is this a nvcc bug? Or should I pass some parameters to nvcc?
Still in trouble. Any ideas please?

3

Thanks for reporting this. I’ve filed a bug report. As a workaround, try making an array of char large enough to hold the number of Vecs you want, and reinterpret_cast the array to Vec *.

This code compiles on my system:

[codebox]

struct Vec

{

float x, y, z;

Vec( float x_ = 0, float y_ = 0, float z_ = 0 )

{

x = x_;

y = y_;

z = z_;

}

};

global void kernel(void)

{

//Vec t[10];

char temp[10 * sizeof(Vec)];

Vec t = reinterpret_cast<Vec>(temp);

return;

}

int main()

{

dim3 block( 16, 16 );

dim3 grid( 1, 1 );

kernel <<< grid, block>>> ();

return 0;

}

[/codebox]

Of course with this approach, you’ll have to explicitly initialize each element of the array.