Looping LOON

The synthesis flow we have considered so far involves basically one BOOG and one LOON synthesis, and taking the fastest resulting netlist. Now we replace the script buf_loon with script loop_buf_loon which loops LOON until the fastest netlist has been found.

The looping is three deep inside a loop that continues until there is no further improvement in the delay. BOOG provides the starting netlist, and subsequent netlists are taken from the first synthesis loop which gives the fastest result at the end of three loops. Each synthesis loop then requires 84 LOON commands instead of 1, as is shown by the pseudo-code below.

min_delay=100000; old_min_delay=100001
while [ "$old_min_delay" -gt "$min_delay" ]; do
 old_min_delay=$min_delay
 for opt1 in 0 1 2 4; do
  ./buf_loon cct0 cct1                   #4 times 
  for opt2 in 0 1 2 4; do
   ./buf_loon cct1 cct2                  #16 times 
   for opt3 in 0 1 2 4; do
    ./buf_loon cct2 cct3 > buf_loon.log  #64 times 
    delay=$(grep '^Critical ' buf_loon.log | cut)
    if [ "$min_delay" -gt "$delay" ]; then
      min_delay=$delay
      cct0=cct1
    fi

Running loop_buf_loon with each BOOG netlist from the four min directories (x1, x2, x4, x8) and four opt levels (0, 1, 2, 4), so 16 netlists in all, on a Pentium M at 1.5GHz this takes nearly 4 hours to run. On a Core Duo T2600 at 2.16GHz it takes just over 2½ hours. (This ratio in job times implies that the T2600 at 2.16GHz matches the performance of a Pentium M at 2.33GHz…not much of an increase over the four year gap between the two machines.)

The script loop_buf_loon, like buf_loon is structured like loon, so that the synthesis job is similar to the previous one. In addition it starts with script ./find_fanin to check whether fanin reduction is requested and if so it runs script ./find_rin. In the job on the right, the script is looped for the four starting libraries and four opt levels for boog.

Events 1 and 2 set up the BOOG looping. Event 3 defines the BOOG library path and event 4 runs boog.

Event 5 defines the library path for LOON synthesis: the full library with macros and 45fF wireload timing. Event 6 runs loop_buf_loon starting with the netlist from boog, writing the fastest netlist to multi8.vst by looping loon until it has been reached. The opt level defined in the LAX file is not used since all opt levels are looped inside loop_buf_loon in order to choose the fastest netlist. What is needed from the LAX file are the input resistances and output loads.

The fastest netlist comes when lib=x4 and opt1=4. The job output for this condition is shown on the right. This corresponds to the boog and loon sequence below (ignoring the ./find_rin script at the start of loop_buf_loon which is used to buffer the inputs).

1 $ boog -l loon_0000_300_4 multi8 multi8_4
2 $ loon -l loon_1500_300_4 multi8_4 multi8_4_4
3 $ loon -l loon_1500_300_2 multi8_4_4 multi8_4_42
4 $ loon -l loon_1500_300_2 multi8_4_42 multi8_4_422
5 $ loon -l loon_1500_300_0 multi8_4_422 multi8_4_4220

Event 5 with opt level 0 flattens the netlist in case the previous synthesis used any macros. The netlist from the initial BOOG synthesis has a fanin of 69, which is reduced to 3 with the first LOON synthesis.

Note the sequence which gives a crash. Some synthesis flows have a lot of crashes!