CODEWITHSUNDEEP
ccudagpunvidia
user25108
user25108

Reputation: 393

error in cuda shared memory static allocation

I wrote a CUDA code using shared memory like this:

__global__ void matrix_mul_shared(float *ad,float *bd,float *cd,int N)
{
    float pvalue=0;
    int TILE=blockDim.x;
    int ty=threadIdx.y;
    int tx=threadIdx.x;

    //allocate shared memory per block
    __shared__ float ads[1][1];
    __shared__ float bds[1][1];

 .

. . }

This code works , but the following code fails;

__global__ void matrix_mul_shared(float *ad,float *bd,float *cd,int N)
{
    float pvalue=0;
    int TILE=blockDim.x;
    int ty=threadIdx.y;
    int tx=threadIdx.x;

    //allocate shared memory per block
    __shared__ float ads[TILE][TILE];
    __shared__ float bds[TILE][TILE];

 .
. 
.
}

The compiler is expecting something constant at the lines where I am allocating shared memory. It says(I forgot the exact error but it is something like this):

The parameters should be a constant

I was able to use printf and print the value of TILE, and it is coming out 1. so why this error?

Upvotes: 0

Views: 227

Answers (1)

Vitality
Vitality

Reputation: 21475

I think the error you are receiving is

error: expression must have a constant value

The variable TILE is not a constant in the sense meant by the compiler. The compiler is asking something known at compile time as dimension of your shared memory array.

A possible solution:

#define TILE 16

__global__ void matrix_mul_shared(float *ad,float *bd,float *cd,int N)
{
    ...

    __shared__ float ads[TILE][TILE];

    ...
}

Upvotes: 1

Related Questions