#version 450
#extension GL_EXT_buffer_reference : require
#extension GL_EXT_buffer_reference_uvec2 : require

layout(local_size_x_id = 0) in;

struct Command
{
	uint type;
	uint src_offset;
	uint dst_offset;
};

layout(buffer_reference, std430, buffer_reference_align = 4) readonly buffer Commands
{
	Command commands[];
};

layout(buffer_reference, std430, buffer_reference_align = 4) readonly buffer SrcBuffer {
	uint values[];
};

layout(buffer_reference, std430, buffer_reference_align = 4) writeonly buffer DstBuffer {
	uint values[];
};

layout(buffer_reference, std430, buffer_reference_align = 4) readonly buffer IndirectCount {
	uint count;
};

layout(buffer_reference, std430, buffer_reference_align = 4) writeonly buffer IndirectCountWrite {
	uint count;
};

layout(push_constant) uniform Registers
{
	Commands commands_va;
	SrcBuffer src_buffer_va;
	DstBuffer dst_buffer_va;
	uvec2 indirect_count_va;
	IndirectCountWrite dst_indirect_count_va;
	uint src_stride;
	uint dst_stride;
};

void main()
{
	Command cmd = commands_va.commands[gl_LocalInvocationIndex];

	uint draw_id = gl_WorkGroupID.x;
	uint max_draws = gl_NumWorkGroups.x;

	if (any(notEqual(indirect_count_va, uvec2(0))))
	{
		max_draws = min(max_draws, IndirectCount(indirect_count_va).count);
		if (gl_WorkGroupID.x == 0u)
			dst_indirect_count_va.count = max_draws;
	}

	if (draw_id < max_draws)
	{
		uint src_offset = src_stride * draw_id + cmd.src_offset;
		uint dst_offset = dst_stride * draw_id + cmd.dst_offset;
		uint src_value = src_buffer_va.values[src_offset];
		dst_buffer_va.values[dst_offset] = src_value;
	}
}