fix(optimizer): fix leveled noise propagation #933

aPere3 · 2024-07-08T14:56:26Z

No description provided.

github-actions

Benchmark

Benchmark suite	Current: `af7d986`	Previous: `dd8cb06`	Ratio
`v0 PBS table generation`	`60049852` ns/iter (`± 376730`)	`60058969` ns/iter (`± 940607`)	`1.00`
`v0 PBS simulate dag table generation`	`37623699` ns/iter (`± 443648`)	`37806462` ns/iter (`± 292750`)	`1.00`
`v0 WoP-PBS table generation`	`68694502` ns/iter (`± 2861756`)	`68533645` ns/iter (`± 388289`)	`1.00`

This comment was automatically generated by workflow using github-action-benchmark.

compilers/concrete-compiler/compiler/lib/Dialect/FHE/Analysis/ConcreteOptimizer.cpp

compilers/concrete-compiler/compiler/lib/Dialect/FHE/Analysis/MANP.cpp

compilers/concrete-optimizer/concrete-optimizer/src/dag/unparametrized.rs

...ilers/concrete-optimizer/concrete-optimizer/src/optimization/dag/multi_parameters/analyze.rs

...rete-optimizer/concrete-optimizer/src/optimization/dag/multi_parameters/symbolic_variance.rs

compilers/concrete-optimizer/concrete-optimizer/src/optimization/dag/solo_key/analyze.rs

...ilers/concrete-optimizer/concrete-optimizer/src/optimization/dag/multi_parameters/analyze.rs

compilers/concrete-optimizer/concrete-optimizer/src/dag/unparametrized.rs

BourgerieQuentin

I'm not sure about the conversion between the manp and the average weight for levelled op, it feels to me it's not always right

aPere3 · 2024-08-19T13:15:22Z

...rete-optimizer/concrete-optimizer/src/optimization/dag/multi_parameters/symbolic_variance.rs

-    #[allow(clippy::float_cmp)]
-    pub fn after_levelled_op(&self, manp: f64) -> Self {
-        let new_coeff = manp * manp;
-        // detect the previous base manp level
-        // this is the maximum value of fresh base noise and pbs base noise
-        let mut current_max: f64 = 0.0;
-        for partition in PartitionIndex::range(0, self.nb_partitions()) {
-            let fresh_coeff = self.coeff_input(partition);
-            let pbs_noise_coeff = self.coeff_pbs(partition);
-            current_max = current_max.max(fresh_coeff).max(pbs_noise_coeff);
-        }
-        // assert!(1.0 <= current_max);
-        // assert!(
-        //     current_max <= new_coeff,
-        //     "Non monotonious levelled op: {current_max} <= {new_coeff}"
-        // );
-        // replace all current_max by new_coeff
-        // multiply everything else by new_coeff / current_max
-        let mut new = self.clone();
-        if current_max == 0.0 {
-            return new;
-        }
-        for cell in &mut new.coeffs.values {
-            if *cell == current_max {
-                *cell = new_coeff;
-            } else {
-                *cell *= new_coeff / current_max;
-            }
-        }
-        new
-    }
-


@BourgerieQuentin

cla-bot bot added the cla-signed label Jul 8, 2024

github-actions bot reviewed Jul 8, 2024

View reviewed changes

aPere3 force-pushed the alex/fix_leveled_noise branch 2 times, most recently from 376a65d to e6912b4 Compare July 9, 2024 15:06

aPere3 requested a review from rudy-6-4 July 15, 2024 07:41

aPere3 force-pushed the alex/fix_leveled_noise branch 3 times, most recently from a0eae73 to 4ab336f Compare July 15, 2024 13:49