layout(local_size_x = 64) in;

layout(metal, binding = 0) buffer ssbo {
    atomicUint globalCounter;
};

workgroup atomicUint localCounter;

void main() {
    // Initialize the local counter.
    if (sk_LocalInvocationID.x == 0) {
        atomicStore(localCounter, 0);
    }

    // Synchronize the threads in the workgroup so they all see the initial value.
    workgroupBarrier();

    // All threads increment the counter.
    atomicAdd(localCounter, 1);

    // Synchronize the threads again to ensure they have all executed the increment
    // and the following load reads the same value across all threads in the
    // workgroup.
    workgroupBarrier();

    // Add the workgroup-only tally to the global counter.
    if (sk_LocalInvocationID.x == 0) {
        atomicAdd(globalCounter, atomicLoad(localCounter));
    }
}