Access Matrix

[liaopeiyuan]

Access matrix on ComputeDAG for all tensors, following a Tiramisu-style construction and flattened, then applied to each TIR-buffer during prediction.

Results on a 1000-subset in e5:

Model: xgb
RMSE: 0.0793
R^2: 0.7567
pairwise comparision accuracy: 0.8504
mape: 10638767510.5327
average peak score@1: 0.8312
average peak score@5: 0.9035

Model: lgbm
RMSE: 0.0793
R^2: 0.7565
pairwise comparision accuracy: 0.8510
mape: 11206471241.5266
average peak score@1: 0.8373
average peak score@5: 0.9073

e2e Runs:

LGBM:

Estimated total latency: 17.081 ms Trials: 102 Used time : 229 s Next ID: -1
Mean inference time (std dev): 14.80 ms (0.00 ms)

XGB:

Estimated total latency: 16.960 ms Trials: 102 Used time : 221 s Next ID: -1
Mean inference time (std dev): 14.60 ms (0.02 ms

Ablation studies wrt naive features:

Model: xgb
RMSE: 0.0790
R^2: 0.7572
pairwise comparision accuracy: 0.8511
mape: 10192227945.6599
average peak score@1: 0.8331
average peak score@5: 0.9058

Model: lgbm
RMSE: 0.0682
R^2: 0.8189
pairwise comparision accuracy: 0.8729
mape: 10005960995.9234
average peak score@1: 0.8736
average peak score@5: 0.9355

Full-scale test results:

Model: xgb
RMSE: 0.0801
R^2: 0.7496
pairwise comparision accuracy: 0.8494
mape: 10025578232.0939
average peak score@1: 0.8284
average peak score@5: 0.9036

Model: lgbm
RMSE: 0.0770
R^2: 0.7686
pairwise comparision accuracy: 0.8564
mape: 10181567914.0740
average peak score@1: 0.8415
average peak score@5: 0.9137

[liaopeiyuan]

This PR is ready for review.