Fix witness generation logic, update circuits with missing gates

circuit_definitions/Cargo.toml

@@ -24,6 +24,7 @@ zk_evm = {git = "https://github.com/matter-labs/era-zk_evm.git", branch = "v1.4.
 derivative = "*"
 serde = {version = "1", features = ["derive"]}
 crossbeam = "0.8"
+seq-macro = "0.3.5"
 
 [features]
 default = []

circuit_definitions/src/circuit_definitions/base_layer/ecrecover.rs

@@ -15,8 +15,11 @@ pub struct ECRecoverFunctionInstanceSynthesisFunction<
     _marker: std::marker::PhantomData<(F, R)>,
 }
 
-use zkevm_circuits::ecrecover::ecrecover_function_entry_point;
 use zkevm_circuits::ecrecover::input::*;
+use zkevm_circuits::ecrecover::{
+    decomp_table::*, ecrecover_function_entry_point, naf_abs_div2_table::*,
+    secp256k1::fixed_base_mul_table::*,
+};
 
 impl<
         F: SmallField,
@@ -69,7 +72,10 @@ where
                 share_constants: false,
             },
         );
-
+        let builder = U8x4FMAGate::configure_builder(
+            builder,
+            GatePlacementStrategy::UseGeneralPurposeColumns,
+        );
         let builder = ZeroCheckGate::configure_builder(
             builder,
             GatePlacementStrategy::UseGeneralPurposeColumns,
@@ -87,6 +93,10 @@ where
             builder,
             GatePlacementStrategy::UseGeneralPurposeColumns,
         );
+        let builder = UIntXAddGate::<8>::configure_builder(
+            builder,
+            GatePlacementStrategy::UseGeneralPurposeColumns,
+        );
         let builder = DotProductGate::<4>::configure_builder(
             builder,
             GatePlacementStrategy::UseGeneralPurposeColumns,
@@ -147,6 +157,31 @@ where
         let table = create_and8_table();
         cs.add_lookup_table::<And8Table, 3>(table);
 
+        let table = create_naf_abs_div2_table();
+        cs.add_lookup_table::<NafAbsDiv2Table, 3>(table);
+
+        let table = create_wnaf_decomp_table();
+        cs.add_lookup_table::<WnafDecompTable, 3>(table);
+
+        seq_macro::seq!(C in 0..32 {
+            let table = create_fixed_base_mul_table::<F, 0, C>();
+            cs.add_lookup_table::<FixedBaseMulTable<0, C>, 3>(table);
+            let table = create_fixed_base_mul_table::<F, 1, C>();
+            cs.add_lookup_table::<FixedBaseMulTable<1, C>, 3>(table);
+            let table = create_fixed_base_mul_table::<F, 2, C>();
+            cs.add_lookup_table::<FixedBaseMulTable<2, C>, 3>(table);
+            let table = create_fixed_base_mul_table::<F, 3, C>();
+            cs.add_lookup_table::<FixedBaseMulTable<3, C>, 3>(table);
+            let table = create_fixed_base_mul_table::<F, 4, C>();
+            cs.add_lookup_table::<FixedBaseMulTable<4, C>, 3>(table);
+            let table = create_fixed_base_mul_table::<F, 5, C>();
+            cs.add_lookup_table::<FixedBaseMulTable<5, C>, 3>(table);
+            let table = create_fixed_base_mul_table::<F, 6, C>();
+            cs.add_lookup_table::<FixedBaseMulTable<6, C>, 3>(table);
+            let table = create_fixed_base_mul_table::<F, 7, C>();
+            cs.add_lookup_table::<FixedBaseMulTable<7, C>, 3>(table);
+        });
+
         let table = create_byte_split_table::<F, 1>();
         cs.add_lookup_table::<ByteSplitTable<1>, 3>(table);
         let table = create_byte_split_table::<F, 2>();

circuit_definitions/src/circuit_definitions/base_layer/keccak256_round_function.rs

@@ -87,6 +87,10 @@ where
             builder,
             GatePlacementStrategy::UseGeneralPurposeColumns,
         );
+        let builder = UIntXAddGate::<8>::configure_builder(
+            builder,
+            GatePlacementStrategy::UseGeneralPurposeColumns,
+        );
         let builder = SelectionGate::configure_builder(
             builder,
             GatePlacementStrategy::UseGeneralPurposeColumns,

src/external_calls.rs

@@ -229,7 +229,7 @@ round_function: R, // used for all queues implementation
     // dbg!(tools.witness_tracer.vm_snapshots.len());
 
     let (instance_oracles, artifacts) = create_artifacts_from_tracer(
-        out_of_circuit_vm.witness_tracer,
+        &mut out_of_circuit_vm.witness_tracer,
         &round_function,
         &geometry,
         (

src/witness/individual_circuits/events_sort_dedup.rs

@@ -32,7 +32,7 @@ pub fn compute_events_dedup_and_sort<
         // return singe dummy witness
         use crate::boojum::gadgets::queue::QueueState;
 
-        let initial_fsm_state = EventsDeduplicatorFSMInputOutput::<F>::placeholder_witness();
+        let mut initial_fsm_state = EventsDeduplicatorFSMInputOutput::<F>::placeholder_witness();
 
         assert_eq!(
             take_queue_state_from_simulator(&unsorted_simulator),
@@ -44,7 +44,13 @@ pub fn compute_events_dedup_and_sort<
             take_queue_state_from_simulator(&unsorted_simulator);
         passthrough_input.intermediate_sorted_queue_state = QueueState::placeholder_witness();
 
-        let final_fsm_state = EventsDeduplicatorFSMInputOutput::<F>::placeholder_witness();
+        let mut final_fsm_state = EventsDeduplicatorFSMInputOutput::<F>::placeholder_witness();
+        // NOTE: little hack to make the trivial case work. we should really however remove the
+        // requirement to have trivial events sorter circuits in the scheduler.
+        initial_fsm_state.lhs_accumulator = [F::ONE; 2];
+        initial_fsm_state.rhs_accumulator = [F::ONE; 2];
+        final_fsm_state.lhs_accumulator = [F::ONE; 2];
+        final_fsm_state.rhs_accumulator = [F::ONE; 2];
 
         let mut passthrough_output = EventsDeduplicatorOutputData::placeholder_witness();
         passthrough_output.final_queue_state = QueueState::placeholder_witness();

src/witness/individual_circuits/keccak256_round_function.rs

@@ -47,9 +47,11 @@ pub fn keccak256_decompose_into_per_circuit_witness<
             } = el;
 
             // we read, then write
-            if let Some(reads) = reads.as_ref() {
-                artifacts.keccak_256_memory_queries.extend_from_slice(reads);
-            }
+            reads.iter().for_each(|read| {
+                if let Some(read) = read {
+                    artifacts.keccak_256_memory_queries.push(*read);
+                }
+            });
 
             if let Some(writes) = writes.as_ref() {
                 artifacts
@@ -204,9 +206,14 @@ pub fn keccak256_decompose_into_per_circuit_witness<
         assert_eq!(request, _req);
 
         // those are refreshed every cycle
-        let mut input_buffer = zk_evm::zk_evm_abstractions::precompiles::keccak256::Buffer::new();
-        use crate::zk_evm::zk_evm_abstractions::precompiles::keccak256::NEW_WORDS_PER_CYCLE;
-        let mut words_buffer = [0u64; NEW_WORDS_PER_CYCLE];
+        use crate::zk_evm::zk_evm_abstractions::precompiles::keccak256::KECCAK_PRECOMPILE_BUFFER_SIZE;
+        let mut input_buffer = zk_evm::zk_evm_abstractions::precompiles::keccak256::ByteBuffer {
+            bytes: [0u8; KECCAK_PRECOMPILE_BUFFER_SIZE],
+            filled: 0,
+        };
+        use crate::zk_evm::zk_evm_abstractions::precompiles::keccak256::MEMORY_READS_PER_CYCLE;
+        use crate::zk_evm::zk_evm_abstractions::precompiles::keccak256::NUM_WORDS_PER_QUERY;
+        let mut words_buffer = [0u8; NUM_WORDS_PER_QUERY * MEMORY_READS_PER_CYCLE * 8];
 
         use crate::zk_evm::zk_evm_abstractions::precompiles::precompile_abi_in_log;
         let mut precompile_request = precompile_abi_in_log(request);
@@ -219,25 +226,25 @@ pub fn keccak256_decompose_into_per_circuit_witness<
 
         precompile_state = Keccak256PrecompileState::RunRoundFunction;
 
+        let needs_full_padding_round =
+            precompile_request.input_memory_length as usize % KECCAK_PRECOMPILE_BUFFER_SIZE == 0;
+
         for (round_idx, round) in round_witness.into_iter().enumerate() {
             // we proceed the request as long as we can
             if round_idx == 0 {
                 assert!(round.new_request.is_some());
             }
 
             // simulate absorb
-            if input_buffer.can_read_into() {
+            if input_buffer.can_fill_bytes(NUM_WORDS_PER_QUERY * MEMORY_READS_PER_CYCLE * 8) {
                 use crate::zk_evm::zk_evm_abstractions::precompiles::keccak256::NUM_WORDS_PER_QUERY;
-                assert!(round.reads.is_some());
-                let reads = round.reads.unwrap();
-                for (query_index, read) in reads.into_iter().enumerate() {
-                    let data = read.value;
-                    let mut bytes32_buffer = [0u8; 32];
-                    data.to_big_endian(&mut bytes32_buffer[..]);
-                    for (i, chunk) in bytes32_buffer.chunks(8).enumerate() {
-                        let as_u64 = u64::from_le_bytes(chunk.try_into().unwrap());
-                        words_buffer[query_index * NUM_WORDS_PER_QUERY + i] = as_u64;
+                for (query_index, read) in round.reads.into_iter().enumerate() {
+                    if read.is_none() {
+                        continue;
                     }
+                    let read = read.unwrap();
+                    let data = read.value;
+                    data.to_big_endian(&mut words_buffer[..]);
 
                     let read_query = memory_queries_it.next().unwrap();
                     assert_eq!(read, read_query);
@@ -255,10 +262,14 @@ pub fn keccak256_decompose_into_per_circuit_witness<
                     precompile_request.input_memory_offset += 1;
                 }
 
-                input_buffer.append(&words_buffer);
+                input_buffer.fill_with_bytes(&words_buffer, 0, 32);
             }
 
-            let words = input_buffer.consume_rate();
+            let words: Vec<u64> = input_buffer
+                .consume::<KECCAK_PRECOMPILE_BUFFER_SIZE>()
+                .chunks(8)
+                .map(|chunk| u64::from_le_bytes(chunk.try_into().unwrap()))
+                .collect();
             use crate::zk_evm::zk_evm_abstractions::precompiles::keccak256::KECCAK_RATE_IN_U64_WORDS;
             let mut block = [0u8; KECCAK_RATE_IN_U64_WORDS * 8];
 
@@ -327,8 +338,8 @@ pub fn keccak256_decompose_into_per_circuit_witness<
                     for el in u64_words_buffer_markers.iter_mut() {
                         *el = false;
                     }
-                    for el in input_buffer.words.iter_mut() {
-                        *el = 0u64;
+                    for el in input_buffer.bytes.iter_mut() {
+                        *el = 0u8;
                     }
                     // internal state is a bit more tricky, it'll be a round over empty input
                     let mut internal_state_over_empty_buffer = Keccak256::default();
@@ -354,21 +365,29 @@ pub fn keccak256_decompose_into_per_circuit_witness<
                 let read_precompile_call =
                     precompile_state == Keccak256PrecompileState::GetRequestFromQueue;
 
+                let padding_round = read_unaligned_words_for_round
+                    && precompile_request.input_memory_length == 0
+                    && needs_full_padding_round;
+
                 let hidden_fsm_output_state = Keccak256RoundFunctionFSMWitness::<F> {
                     completed,
                     read_unaligned_words_for_round,
+                    padding_round,
                     keccak_internal_state,
                     read_precompile_call,
                     timestamp_to_use_for_read: request.timestamp.0,
                     timestamp_to_use_for_write: request.timestamp.0 + 1,
-                    u8_words_buffer: buffer_to_bytes(&input_buffer),
-                    u64_words_buffer_markers,
                     precompile_call_params: Keccak256PrecompileCallParamsWitness::<F> {
                         input_page: precompile_request.memory_page_to_read,
-                        input_offset: precompile_request.input_memory_offset,
+                        input_memory_byte_offset: precompile_request.input_memory_offset,
+                        input_memory_byte_length: precompile_request.input_memory_length,
                         output_page: precompile_request.memory_page_to_write,
-                        output_offset: precompile_request.output_memory_offset,
-                        num_rounds: num_rounds_left as u32,
+                        output_word_offset: precompile_request.output_memory_offset,
+                        needs_full_padding_round,
+                    },
+                    buffer: zkevm_circuits::keccak256_round_function::buffer::ByteBufferWitness {
+                        bytes: input_buffer.bytes,
+                        filled: input_buffer.filled as u8,
                     },
                 };
 
@@ -474,22 +493,6 @@ pub(crate) fn encode_kecca256_inner_state(state: [u64; 25]) -> [[[u8; 8]; 5]; 5]
     result
 }
 
-fn buffer_to_bytes(
-    buffer: &zk_evm::zk_evm_abstractions::precompiles::keccak256::Buffer,
-) -> [u8; zkevm_circuits::keccak256_round_function::BYTES_BUFFER_SIZE] {
-    assert_eq!(
-        zkevm_circuits::keccak256_round_function::BUFFER_SIZE_IN_U64_WORDS,
-        buffer.words.len()
-    );
-
-    let mut result = [0u8; zkevm_circuits::keccak256_round_function::BYTES_BUFFER_SIZE];
-    for (dst, src) in result.array_chunks_mut::<8>().zip(buffer.words.iter()) {
-        *dst = src.to_le_bytes();
-    }
-
-    result
-}
-
 // #[cfg(test)]
 // mod test {
 //     use super::*;

src/witness/individual_circuits/storage_application.rs

@@ -4,7 +4,9 @@ use crate::boojum::sha3::digest::{FixedOutput, Update};
 use crate::boojum::sha3::Keccak256;
 use crate::witness::individual_circuits::keccak256_round_function::encode_kecca256_inner_state;
 use crate::witness::tree::*;
-use crate::zk_evm::zk_evm_abstractions::precompiles::keccak256::{transmute_state, BUFFER_SIZE};
+use crate::zk_evm::zk_evm_abstractions::precompiles::keccak256::{
+    transmute_state, KECCAK_PRECOMPILE_BUFFER_SIZE,
+};
 use crate::zkevm_circuits::base_structures::state_diff_record::NUM_KECCAK256_ROUNDS_PER_RECORD_ACCUMULATION;
 use crate::zkevm_circuits::storage_application::input::*;
 use blake2::Blake2s256;

src/witness/oracle.rs

@@ -173,7 +173,7 @@ pub fn create_artifacts_from_tracer<
     F: SmallField,
     R: BuildableCircuitRoundFunction<F, 8, 12, 4> + AlgebraicRoundFunction<F, 8, 12, 4>,
 >(
-    tracer: WitnessTracer,
+    tracer: &mut WitnessTracer,
     round_function: &R,
     geometry: &GeometryConfig,
     entry_point_decommittment_query: (DecommittmentQuery, Vec<U256>),
@@ -933,11 +933,11 @@ pub fn create_artifacts_from_tracer<
     // when it potentially came into and out of scope
 
     let mut artifacts = FullBlockArtifacts::<F>::default();
-    artifacts.vm_memory_queries_accumulated = vm_memory_queries_accumulated;
-    artifacts.all_decommittment_queries = decommittment_queries;
-    artifacts.keccak_round_function_witnesses = keccak_round_function_witnesses;
-    artifacts.sha256_round_function_witnesses = sha256_round_function_witnesses;
-    artifacts.ecrecover_witnesses = ecrecover_witnesses;
+    artifacts.vm_memory_queries_accumulated = vm_memory_queries_accumulated.to_vec();
+    artifacts.all_decommittment_queries = decommittment_queries.to_vec();
+    artifacts.keccak_round_function_witnesses = keccak_round_function_witnesses.to_vec();
+    artifacts.sha256_round_function_witnesses = sha256_round_function_witnesses.to_vec();
+    artifacts.ecrecover_witnesses = ecrecover_witnesses.to_vec();
     artifacts.original_log_queue = original_log_queue;
     artifacts.original_log_queue_simulator =
         original_log_queue_simulator.unwrap_or(LogQueueSimulator::empty());