Example of using static shared memory with array arithmetic

shared_memory_bounds.cu

// This is a small example based on a question on GPU mode #beginner
#include <cuda_runtime.h>
#include <stdio.h>

#define WARP_SIZE 32

#define ACTUAL_WARPS_PER_BLOCK 8

#define WARPS_EACH_BLK 4

__global__ void sharedMemoryErrorKernel() {
    __shared__ unsigned int sat[WARPS_EACH_BLK];
    __shared__ unsigned int commons[WARPS_EACH_BLK];

    unsigned int tid = threadIdx.x;
    unsigned int warp_id = tid / WARP_SIZE;
    unsigned int lane_id = tid % WARP_SIZE;

    // 'local_sat[0]' is equivalent to 'sat[warp_id]'
    // 'warp_id' shoud be  <= 'WARPS_EACH_BLK'.

    if (lane_id == 0) {
        sat[warp_id] = warp_id;
        commons[warp_id] = warp_id;
    }

    __syncthreads();

    // Read
    if (lane_id == 0) {
        printf("Thread %u (Warp %u, Lane %u):\n", tid, warp_id, lane_id);
        printf("  sat[%u]: %u\n", warp_id, sat[warp_id]);
        printf("  commons[%u]: %u\n", warp_id, commons[warp_id]);
    }
}

int main() {
    const int threadsPerBlock = ACTUAL_WARPS_PER_BLOCK * WARP_SIZE; // e.g., 8 * 32 = 256 threads

    sharedMemoryErrorKernel<<<1, threadsPerBlock>>>();

    cudaError_t err = cudaDeviceSynchronize();
    if (err != cudaSuccess) {
        fprintf(stderr, "CUDA Runtime Error: %s\n", cudaGetErrorString(err));
    }

    return 0;
}

You are probably either miscalculating warp_id or WARPS_EACH_BLK is ill defined. I'd start by printing those.

Also, once you master static shared memory, you probably want to eventually switch to use dynamic shared memory instead of static. It's a bit harder to use and reason but it's the normal CUDA way.