-- Compute shader example
-- Demonstrates: shared memory, barriers, compute builtins, buffer bindings

struct Params
	count: uint
end

uniform params: Params
buffer data: [1024]float

-- Workgroup-shared memory (visible to all threads in a workgroup)
shared tile: [64]float

@entry(compute) @workgroup_size(64, 1, 1)
function main(
	@builtin(global_invocation_id) gid: uvec3,
	@builtin(local_invocation_index) lid: uint
)
	-- Load data into shared memory
	if gid.x < params.count then
		tile[lid] = data[gid.x]
	else
		tile[lid] = 0.0
	end

	-- Synchronize all threads in the workgroup
	barrier()

	-- Write back (example: each thread reads neighbor's value)
	if gid.x < params.count then
		if lid > uint(0) then
			data[gid.x] = tile[lid] + tile[lid - uint(1)]
		else
			data[gid.x] = tile[lid] * 2.0
		end
	end
end