Background
partition_c is the densest NVDLA partition — 84 FakeRAM macro instances packed into an asap7 die. It landed in #68 at CORE_UTILIZATION=40 / PLACE_DENSITY_LB_ADDON=0.30 / MACRO_PLACE_HALO=3 3, with a 42% final utilization, which is the lowest of the 5 partitions.
The tuning journey that got us there was about trading utilization for flow completion, not finding a PPA optimum:
| # |
util / addon / halo |
Furthest stage |
Failure |
| 1 |
50 / 0.25 / 2 |
post-CTS DPL |
DPL-0036: clkbuf couldn't legalize near cbuf_ram_bank31 |
| 2 |
45 / 0.30 / 5 |
MPL floorplan |
MPL-0003: no valid tilings (halo×84 ate tiling budget) |
| 3 |
48 / 0.25 / 3 |
GRT |
GRT-0116: placed but routed-congested |
| 4 |
40 / 0.30 / 3 |
6_final |
Succeeded (42% util) |
Task
Run a proper PPA sweep on partition_c to reclaim utilization without regressing any flow stage. Suggested grid:
MACRO_PLACE_HALO ∈ {3, 4}
CORE_UTILIZATION ∈ {40, 42, 44, 46}
PLACE_DENSITY_LB_ADDON ∈ {0.25, 0.28, 0.30}
Acceptance
- Clean
bazel build //designs/asap7/NVDLA/partition_c:partition_c_final end-to-end (no DPL-0036, MPL-0003, or GRT-0116).
- Final utilization > 42% (stretch: 48–50%).
- No timing / power regression vs. the current configuration.
Related
Background
partition_cis the densest NVDLA partition — 84 FakeRAM macro instances packed into an asap7 die. It landed in #68 atCORE_UTILIZATION=40 / PLACE_DENSITY_LB_ADDON=0.30 / MACRO_PLACE_HALO=3 3, with a 42% final utilization, which is the lowest of the 5 partitions.The tuning journey that got us there was about trading utilization for flow completion, not finding a PPA optimum:
DPL-0036: clkbuf couldn't legalize nearcbuf_ram_bank31MPL-0003: no valid tilings (halo×84 ate tiling budget)GRT-0116: placed but routed-congestedTask
Run a proper PPA sweep on
partition_cto reclaim utilization without regressing any flow stage. Suggested grid:MACRO_PLACE_HALO ∈ {3, 4}CORE_UTILIZATION ∈ {40, 42, 44, 46}PLACE_DENSITY_LB_ADDON ∈ {0.25, 0.28, 0.30}Acceptance
bazel build //designs/asap7/NVDLA/partition_c:partition_c_finalend-to-end (noDPL-0036,MPL-0003, orGRT-0116).Related
partition_cconfig)optimize-ppaskill is a good starting point for the sweep harness.