groupshared float scratch[64];

struct Params {
	uint count;
	float scale;
};
RWStructuredBuffer<Params> params : register(u0, space0);
struct DataBuffer {
	float values[64];
};
RWStructuredBuffer<DataBuffer> data : register(u1, space0);

[numthreads(64, 1, 1)]
void main(uint __luma_work_local_index : SV_GroupIndex, uint3 __luma_work_global_id : SV_DispatchThreadID, uint3 __luma_work_local_id : SV_GroupThreadID, uint3 __luma_work_group_id : SV_GroupID) {
	uint idx = __luma_work_local_index;
	uint gid = __luma_work_global_id.x;
	uint lid = __luma_work_local_id.x;
	uint group_x = __luma_work_group_id.x;
	if ((gid < params.count)) {
		scratch[idx] = (data.values[idx] * params.scale);
		GroupMemoryBarrierWithGroupSync();
		data.values[idx] = (scratch[idx] + (float)(lid + group_x));
	}
}