The copy-out operation is essentially the reverse of the copy-in operation. The operation is performed by the LHT_copy_out::copy_table_out() method, which is called after a loop containing a histogram operation.

Like copy-in, the table is read and written one 1024-bit line at a time. Each line is read from L1D as a pair of 512-bit vectors using two LUTRD (lookup read) instructions, rearranged using two VPERM instructions, and written using normal vector stores. The loop pipelines at an ii of 2, resulting in a throughput of 512 bits per cycle.

For purposes of the LUTRD used for the copy-out, the table is considered to be 16 parallel tables containing 32-bit elements. This enables each LUTRD instruction to read the maximal payload of 512 bits. The copy-out LTCR configuration is computed during the init() function by a call to the copy_out_config() method and stored in the tvals structure.