What is the effective way to initialize shared memory

Accelerated Computing CUDA CUDA Programming and Performance
1

My program need to use shared memory with some initialization data.

In the begining I use:
shared unsigned int data2[4] = {1, 2, 3, 4};

But nvcc report error “initializer not allowed for shared variable”.

Is there any convenient way to initialize shared memory?

2

Let a single thread do it

if (threadIdx.x == 0) {

  data2[0] = 1;

  data2[1] = 2;

  // etc

}

__syncthreads();

or let multiple threads based on their index do it if it is many elements.

data2[threadIdx.x] = threadIdx.x;

__syncthreads();
3

nice answer. Thanks.

4

That is nice.

5

All depends on if the value you are putting into each array element can be derived from the threadIdx
if it can’t and there are only a few values then Bloodhunt’s code using thread 0 is way to go

but if it can then using multiple threads working in parrallel, as nachovall did, will do it faster.