Why matrix multiplication when using queue?

I'm trying to get my head around the multiplication of embeddings by model.module.prototypes.weight.t() before passing them into distributed_sinkhorn. It is not clear why do we do this multiplication only when using a queue (Line 298). Moreover, why do we mix output prototypes with multiplied embeddings (Line 301)? What's the motivation behind this?

swav/main_swav.py

Lines 294 to 307 in 06b1b7c

    
           # time to use the queue 
        
           if queue is not None: 
        
               if use_the_queue or not torch.all(queue[i, -1, :] == 0): 
        
                   use_the_queue = True 
        
                   out = torch.cat((torch.mm( 
        
                       queue[i], 
        
                       model.module.prototypes.weight.t() 
        
                   ), out)) 
        
               # fill the queue 
        
               queue[i, bs:] = queue[i, :-bs].clone() 
        
               queue[i, :bs] = embedding[crop_id * bs: (crop_id + 1) * bs] 
        
           # get assignments 
        
           q = distributed_sinkhorn(out)[-bs:]

	# time to use the queue
	if queue is not None:
	if use_the_queue or not torch.all(queue[i, -1, :] == 0):
	use_the_queue = True
	out = torch.cat((torch.mm(
	queue[i],
	model.module.prototypes.weight.t()
	), out))
	# fill the queue
	queue[i, bs:] = queue[i, :-bs].clone()
	queue[i, :bs] = embedding[crop_id * bs: (crop_id + 1) * bs]

	# get assignments
	q = distributed_sinkhorn(out)[-bs:]