#version 450

/* built in:
in uvec3 gl_NumWorkGroups;
in uvec3 gl_WorkGroupID;
in uvec3 gl_LocalInvocationID;
in uvec3 gl_GlobalInvocationID;
in uint  gl_LocalInvocationIndex;
*/

// Why did we not have conflicts in the Reduction? 
// Because of the sequential addressing (here we use interleaved => we have conflicts).
// TODO: tailor to your architecture (these parameter work for virtually all NVIDIA GPUs)
#define NUM_BANKS			32
#define NUM_BANKS_LOG		5
#define SIMD_GROUP_SIZE		32

layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;

layout(push_constant) uniform PushStruct {
    uint size;
} p;

layout(binding = 0) buffer inoutBufer {uint array[];};
layout(binding = 1) buffer offsetBufer {uint higherLevelArray[];};

// TODO: Shared variables

// Bank conflicts
#define AVOID_BANK_CONFLICTS
#ifdef AVOID_BANK_CONFLICTS
// TODO: define your conflict-free macro here
#else
#define OFFSET(A) (A)
#endif

//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void main()
{
    // TODO: Kernel implementation

    // Cache first half of elements in the local memory
    // Cache second half of elements

    // Perform up-sweep

    // Unroll the last steps when arrived at warp size
    // Set the last element to 0


    // Perform down-sweep
}