Use recursion factor of 4 for all proofs (#1440)

ipa-core/src/helpers/gateway/send.rs

@@ -25,6 +25,7 @@ use crate::{
         labels::{ROLE, STEP},
         metrics::{BYTES_SENT, RECORDS_SENT},
     },
+    utils::non_zero_prev_power_of_two,
 };
 
 /// Sending end of the gateway channel.
@@ -256,14 +257,6 @@ impl SendChannelConfig {
         total_records: TotalRecords,
         record_size: usize,
     ) -> Self {
-        // this computes the greatest positive power of 2 that is
-        // less than or equal to target.
-        fn non_zero_prev_power_of_two(target: usize) -> usize {
-            let bits = usize::BITS - target.leading_zeros();
-
-            1 << (std::cmp::max(1, bits) - 1)
-        }
-
         assert!(record_size > 0, "Message size cannot be 0");
 
         let total_capacity = gateway_config.active.get() * record_size;

ipa-core/src/protocol/context/dzkp_validator.rs

@@ -20,7 +20,7 @@ use crate::{
         },
         ipa_prf::{
             validation_protocol::{proof_generation::ProofBatch, validation::BatchToVerify},
-            LargeProofGenerator, SmallProofGenerator,
+            CompressedProofGenerator, FirstProofGenerator,
         },
         Gate, RecordId, RecordIdRange,
     },
@@ -50,6 +50,9 @@ const BIT_ARRAY_SHIFT: usize = BIT_ARRAY_LEN.ilog2() as usize;
 // A smaller value is used for tests, to enable covering some corner cases with a
 // reasonable runtime. Some of these tests use TARGET_PROOF_SIZE directly, so for tests
 // it does need to be a power of two.
+//
+// TARGET_PROOF_SIZE is closely related to MAX_PROOF_RECURSION; see the assertion that
+// `uv_values.len() <= max_uv_values` in `ProofBatch` for more detail.
 #[cfg(test)]
 pub const TARGET_PROOF_SIZE: usize = 8192;
 #[cfg(not(test))]
@@ -73,7 +76,7 @@ pub const TARGET_PROOF_SIZE: usize = 50_000_000;
 // to blocks of 256), leaving some margin is advised.
 //
 // The implementation requires that MAX_PROOF_RECURSION is at least 2.
-pub const MAX_PROOF_RECURSION: usize = 9;
+pub const MAX_PROOF_RECURSION: usize = 14;
 
 /// `MultiplicationInputsBlock` is a block of fixed size of intermediate values
 /// that occur duringa multiplication.
@@ -601,8 +604,8 @@ impl Batch {
         ctx: Base<'_, B>,
         batch_index: usize,
     ) -> Result<(), Error> {
-        const PRSS_RECORDS_PER_BATCH: usize = LargeProofGenerator::PROOF_LENGTH
-            + (MAX_PROOF_RECURSION - 1) * SmallProofGenerator::PROOF_LENGTH
+        const PRSS_RECORDS_PER_BATCH: usize = FirstProofGenerator::PROOF_LENGTH
+            + (MAX_PROOF_RECURSION - 1) * CompressedProofGenerator::PROOF_LENGTH
             + 2; // P and Q masks
 
         let proof_ctx = ctx.narrow(&Step::GenerateProof);

ipa-core/src/protocol/hybrid/oprf.rs

@@ -1,3 +1,5 @@
+use std::cmp::max;
+
 use futures::{stream, StreamExt, TryStreamExt};
 use typenum::Const;
 
@@ -17,8 +19,9 @@ use crate::{
     protocol::{
         basics::{BooleanProtocols, Reveal},
         context::{
-            dzkp_validator::DZKPValidator, reshard_try_stream, DZKPUpgraded, MacUpgraded,
-            MaliciousProtocolSteps, ShardedContext, UpgradableContext, Validator,
+            dzkp_validator::{DZKPValidator, TARGET_PROOF_SIZE},
+            reshard_try_stream, DZKPUpgraded, MacUpgraded, MaliciousProtocolSteps, ShardedContext,
+            UpgradableContext, Validator,
         },
         hybrid::step::HybridStep,
         ipa_prf::{
@@ -34,7 +37,9 @@ use crate::{
         Vectorizable,
     },
     seq_join::seq_join,
+    utils::non_zero_prev_power_of_two,
 };
+
 // In theory, we could support (runtime-configured breakdown count) ≤ (compile-time breakdown count)
 // ≤ 2^|bk|, with all three values distinct, but at present, there is no runtime configuration and
 // the latter two must be equal. The implementation of `move_single_value_to_bucket` does support a
@@ -64,13 +69,17 @@ pub const CONV_CHUNK: usize = 256;
 /// Vectorization dimension for PRF
 pub const PRF_CHUNK: usize = 16;
 
-// We expect 2*256 = 512 gates in total for two additions per conversion. The vectorization factor
-// is CONV_CHUNK. Let `len` equal the number of converted shares. The total amount of
-// multiplications is CONV_CHUNK*512*len. We want CONV_CHUNK*512*len ≈ 50M, or len ≈ 381, for a
-// reasonably-sized proof. There is also a constraint on proof chunks to be powers of two, so
-// we pick the closest power of two close to 381 but less than that value. 256 gives us around 33M
-// multiplications per batch
-const CONV_PROOF_CHUNK: usize = 256;
+/// Returns a suitable proof chunk size (in records) for use with `convert_to_fp25519`.
+///
+/// We expect 2*256 = 512 gates in total for two additions per conversion. The
+/// vectorization factor is `CONV_CHUNK`. Let `len` equal the number of converted
+/// shares. The total amount of multiplications is `CONV_CHUNK`*512*len. We want
+/// `CONV_CHUNK`*512*len ≈ 50M for a reasonably-sized proof. There is also a constraint
+/// on proof chunks to be powers of two, and we don't want to compute a proof chunk
+/// of zero when `TARGET_PROOF_SIZE` is smaller for tests.
+fn conv_proof_chunk() -> usize {
+    non_zero_prev_power_of_two(max(2, TARGET_PROOF_SIZE / CONV_CHUNK / 512))
+}
 
 /// This computes the Dodis-Yampolsky PRF value on every match key from input,
 /// and reshards the reports according to the computed PRF. At the end, reports with the
@@ -101,7 +110,7 @@ where
             protocol: &HybridStep::ConvertFp25519,
             validate: &HybridStep::ConvertFp25519Validate,
         },
-        CONV_PROOF_CHUNK,
+        conv_proof_chunk(),
     );
     let m_ctx = validator.context();
 

ipa-core/src/protocol/ipa_prf/aggregation/mod.rs

@@ -25,6 +25,7 @@ use crate::{
         replicated::semi_honest::AdditiveShare as Replicated, BitDecomposed, FieldSimd,
         SharedValue, TransposeFrom, Vectorizable,
     },
+    utils::non_zero_prev_power_of_two,
 };
 
 pub(crate) mod breakdown_reveal;
@@ -96,8 +97,14 @@ pub type AggResult<const B: usize> = Result<BitDecomposed<Replicated<Boolean, B>
 /// saturating the output) is:
 ///
 /// $\sum_{i = 1}^k 2^{k - i} (b + i - 1) \approx 2^k (b + 1) = N (b + 1)$
+///
+/// We set a floor of 2 to avoid computing a chunk of zero when `TARGET_PROOF_SIZE` is
+/// smaller for tests.
 pub fn aggregate_values_proof_chunk(input_width: usize, input_item_bits: usize) -> usize {
-    max(2, TARGET_PROOF_SIZE / input_width / (input_item_bits + 1)).next_power_of_two()
+    non_zero_prev_power_of_two(max(
+        2,
+        TARGET_PROOF_SIZE / input_width / (input_item_bits + 1),
+    ))
 }
 
 // This is the step count for AggregateChunkStep. We need it to size RecordId arrays.

ipa-core/src/protocol/ipa_prf/boolean_ops/share_conversion_aby.rs

@@ -381,7 +381,7 @@ mod tests {
         helpers::stream::process_slice_by_chunks,
         protocol::{
             context::{dzkp_validator::DZKPValidator, UpgradableContext, TEST_DZKP_STEPS},
-            ipa_prf::{CONV_CHUNK, CONV_PROOF_CHUNK, PRF_CHUNK},
+            ipa_prf::{conv_proof_chunk, CONV_CHUNK, PRF_CHUNK},
         },
         rand::thread_rng,
         secret_sharing::SharedValue,
@@ -415,7 +415,7 @@ mod tests {
             let [res0, res1, res2] = world
                 .semi_honest(records.into_iter(), |ctx, records| async move {
                     let c_ctx = ctx.set_total_records((COUNT + CONV_CHUNK - 1) / CONV_CHUNK);
-                    let validator = &c_ctx.dzkp_validator(TEST_DZKP_STEPS, CONV_PROOF_CHUNK);
+                    let validator = &c_ctx.dzkp_validator(TEST_DZKP_STEPS, conv_proof_chunk());
                     let m_ctx = validator.context();
                     seq_join(
                         m_ctx.active_work(),

ipa-core/src/protocol/ipa_prf/malicious_security/prover.rs

@@ -1,14 +1,15 @@
 use std::{borrow::Borrow, iter::zip, marker::PhantomData};
 
-#[cfg(all(test, unit_test))]
-use crate::ff::Fp31;
 use crate::{
     error::Error::{self, DZKPMasks},
     ff::{Fp61BitPrime, PrimeField},
     helpers::hashing::{compute_hash, hash_to_field},
     protocol::{
         context::Context,
-        ipa_prf::malicious_security::lagrange::{CanonicalLagrangeDenominator, LagrangeTable},
+        ipa_prf::{
+            malicious_security::lagrange::{CanonicalLagrangeDenominator, LagrangeTable},
+            CompressedProofGenerator,
+        },
         prss::SharedRandomness,
         RecordId, RecordIdRange,
     },
@@ -84,8 +85,8 @@ where
         // compute final uv values
         let (u_values, v_values) = &mut self.uv_chunks[0];
         // shift first element to last position
-        u_values[SmallProofGenerator::RECURSION_FACTOR - 1] = u_values[0];
-        v_values[SmallProofGenerator::RECURSION_FACTOR - 1] = v_values[0];
+        u_values[CompressedProofGenerator::RECURSION_FACTOR - 1] = u_values[0];
+        v_values[CompressedProofGenerator::RECURSION_FACTOR - 1] = v_values[0];
         // set masks in first position
         u_values[0] = my_p_mask;
         v_values[0] = my_q_mask;
@@ -105,15 +106,11 @@ pub struct ProofGenerator<F: PrimeField, const L: usize, const P: usize, const M
     phantom_data: PhantomData<F>,
 }
 
-#[cfg(all(test, unit_test))]
-pub type TestProofGenerator = ProofGenerator<Fp31, 4, 7, 3>;
-
 // Compression Factor is L
 // P, Proof size is 2*L - 1
 // M, the number of interpolated points is L - 1
 // The reason we need these is that Rust doesn't support basic math operations on const generics
-pub type SmallProofGenerator = ProofGenerator<Fp61BitPrime, 8, 15, 7>;
-pub type LargeProofGenerator = ProofGenerator<Fp61BitPrime, 32, 63, 31>;
+pub type SmallProofGenerator = ProofGenerator<Fp61BitPrime, 4, 7, 3>;
 
 impl<F: PrimeField, const L: usize, const P: usize, const M: usize> ProofGenerator<F, L, P, M> {
     // define constants such that they can be used externally
@@ -265,7 +262,7 @@ mod test {
             context::Context,
             ipa_prf::malicious_security::{
                 lagrange::{CanonicalLagrangeDenominator, LagrangeTable},
-                prover::{LargeProofGenerator, SmallProofGenerator, TestProofGenerator, UVValues},
+                prover::{ProofGenerator, SmallProofGenerator, UVValues},
             },
             RecordId, RecordIdRange,
         },
@@ -274,6 +271,9 @@ mod test {
         test_fixture::{Runner, TestWorld},
     };
 
+    type TestProofGenerator = ProofGenerator<Fp31, 4, 7, 3>;
+    type LargeProofGenerator = ProofGenerator<Fp61BitPrime, 32, 63, 31>;
+
     fn zip_chunks<F: PrimeField, const U: usize, I, J>(a: I, b: J) -> UVValues<F, U>
     where
         I: IntoIterator<Item = u128>,

ipa-core/src/protocol/ipa_prf/mod.rs

@@ -1,4 +1,4 @@
-use std::{convert::Infallible, iter::zip, num::NonZeroU32, ops::Add};
+use std::{cmp::max, convert::Infallible, iter::zip, num::NonZeroU32, ops::Add};
 
 use futures::{stream, StreamExt, TryStreamExt};
 use generic_array::{ArrayLength, GenericArray};
@@ -24,8 +24,8 @@ use crate::{
     protocol::{
         basics::{BooleanArrayMul, BooleanProtocols, Reveal},
         context::{
-            dzkp_validator::DZKPValidator, DZKPUpgraded, MacUpgraded, MaliciousProtocolSteps,
-            UpgradableContext,
+            dzkp_validator::{DZKPValidator, TARGET_PROOF_SIZE},
+            DZKPUpgraded, MacUpgraded, MaliciousProtocolSteps, UpgradableContext,
         },
         ipa_prf::{
             boolean_ops::convert_to_fp25519,
@@ -44,6 +44,7 @@ use crate::{
         BitDecomposed, FieldSimd, SharedValue, TransposeFrom, Vectorizable,
     },
     seq_join::seq_join,
+    utils::non_zero_prev_power_of_two,
 };
 
 pub(crate) mod aggregation;
@@ -58,7 +59,9 @@ pub(crate) mod shuffle;
 pub(crate) mod step;
 pub mod validation_protocol;
 
-pub use malicious_security::prover::{LargeProofGenerator, SmallProofGenerator};
+pub type FirstProofGenerator = malicious_security::prover::SmallProofGenerator;
+pub type CompressedProofGenerator = malicious_security::prover::SmallProofGenerator;
+
 pub use shuffle::Shuffle;
 
 /// Match key type
@@ -409,13 +412,17 @@ where
     Ok(noisy_output_histogram)
 }
 
-// We expect 2*256 = 512 gates in total for two additions per conversion. The vectorization factor
-// is CONV_CHUNK. Let `len` equal the number of converted shares. The total amount of
-// multiplications is CONV_CHUNK*512*len. We want CONV_CHUNK*512*len ≈ 50M, or len ≈ 381, for a
-// reasonably-sized proof. There is also a constraint on proof chunks to be powers of two, so
-// we pick the closest power of two close to 381 but less than that value. 256 gives us around 33M
-// multiplications per batch
-const CONV_PROOF_CHUNK: usize = 256;
+/// Returns a suitable proof chunk size (in records) for use with `convert_to_fp25519`.
+///
+/// We expect 2*256 = 512 gates in total for two additions per conversion. The
+/// vectorization factor is `CONV_CHUNK`. Let `len` equal the number of converted
+/// shares. The total amount of multiplications is `CONV_CHUNK`*512*len. We want
+/// `CONV_CHUNK`*512*len ≈ 50M for a reasonably-sized proof. There is also a constraint
+/// on proof chunks to be powers of two, and we don't want to compute a proof chunk
+/// of zero when `TARGET_PROOF_SIZE` is smaller for tests.
+fn conv_proof_chunk() -> usize {
+    non_zero_prev_power_of_two(max(2, TARGET_PROOF_SIZE / CONV_CHUNK / 512))
+}
 
 #[tracing::instrument(name = "compute_prf_for_inputs", skip_all)]
 async fn compute_prf_for_inputs<C, BK, TV, TS>(
@@ -443,7 +450,7 @@ where
             protocol: &Step::ConvertFp25519,
             validate: &Step::ConvertFp25519Validate,
         },
-        CONV_PROOF_CHUNK,
+        conv_proof_chunk(),
     );
     let m_ctx = validator.context();
 

ipa-core/src/protocol/ipa_prf/prf_sharding/mod.rs

@@ -52,6 +52,7 @@ use crate::{
         replicated::{semi_honest::AdditiveShare as Replicated, ReplicatedSecretSharing},
         BitDecomposed, FieldSimd, SharedValue, TransposeFrom, Vectorizable,
     },
+    utils::non_zero_prev_power_of_two,
 };
 
 pub mod feature_label_dot_product;
@@ -515,7 +516,10 @@ where
         // TODO: this override was originally added to work around problems with
         // read_size vs. batch size alignment. Those are now fixed (in #1332), but this
         // is still observed to help performance (see #1376), so has been retained.
-        std::cmp::min(sh_ctx.active_work().get(), chunk_size.next_power_of_two()),
+        std::cmp::min(
+            sh_ctx.active_work().get(),
+            non_zero_prev_power_of_two(chunk_size),
+        ),
     );
     dzkp_validator.set_total_records(TotalRecords::specified(histogram[1]).unwrap());
     let ctx_for_row_number = set_up_contexts(&dzkp_validator.context(), histogram)?;

ipa-core/src/protocol/ipa_prf/quicksort.rs

@@ -30,6 +30,7 @@ use crate::{
         Vectorizable,
     },
     seq_join::seq_join,
+    utils::non_zero_prev_power_of_two,
 };
 
 impl<K> ChunkBuffer<SORT_CHUNK> for (Vec<AdditiveShare<K>>, Vec<AdditiveShare<K>>)
@@ -98,7 +99,7 @@ where
 }
 
 fn quicksort_proof_chunk(key_bits: usize) -> usize {
-    (TARGET_PROOF_SIZE / key_bits / SORT_CHUNK).next_power_of_two()
+    non_zero_prev_power_of_two(TARGET_PROOF_SIZE / key_bits / SORT_CHUNK)
 }
 
 /// Insecure quicksort using MPC comparisons and a key extraction function `get_key`.