Round compressed grid locations #2509
Conversation
github-actions
bot
added
VPR
|
Titan benchmarks
Link to QoR spreadsheet |
|
QoR looks promising ... will need to go through and check the various test failures to make sure they are small circuit QoR changes that are neutral or irrelevant. |
|
(Your idea): could also run a 3D arch regtest (your call). |
|
vtr_reg_nightly_test1 and vtr_reg_nightly_test1_odin failures are fixed. vtr_reg_nightly_test3_odin hits the 7 hours executin time limit, which is a little strange because vtr_reg_nightly_test3 finishes in ~1 hour. |
|
If this looks like a random failure we could move some test out of vtr_reg_nightly_test3_odin (and put it in another suite, or tweak the designs used). |
|
I ran nightly3_odin tasks on wintermute. There is no significant execution time difference between this branch and the master branch on wintermute. |
|
OK, if you resolve the conflicts we can merge. They look like a bunch of golden result conflicts, so hopefully not too hard to resolve. Maybe the odin II run is at the edge of 6 hours and we need to split it? |
When translating a grid location to a compressed grid location, we call std::lower_bound() to find the corresponding compressed location. Assume that we want to translate the grid lcoation (66, 67) for a DSP block in a grid with DSP columns at x=50 and x=70. The current grid_loc_to_compressed_loc_approx() would select the column at x=50, while the one at x=70 was closer. This PR changes grid_loc_to_compressed_loc_approx() to choose the closest compressed location instead.
In calculate_centroid_loc(), we cast float variables to int, which is equivalent to taking their floor. This means that the centroid layer is always zero unless all connected pins are on layer 1.
How Has This Been Tested?
I ran experiments on Titan benchmarks. I'll post QoR summary in a comment in this PR. If the results are promising, we can experiment with other benchmarks and architectures.