chore: Address feedback

Cargo.toml

@@ -78,7 +78,7 @@ getrandom = "=0.2.14" # 0.2.15 depends on yanked libc 0.2.154
 hashbrown = { version = "0.14.5", features = ["serde"] }
 hex = "0.4.3"
 home = "0.5.9"
-hybrid-array = "0.2.0-rc.9"
+hybrid-array = "0.2.0-rc"
 indicatif = "0.17.8"
 itertools = "0.12.1"
 k256 = "0.13.3"

core/src/stark/air.rs

@@ -161,6 +161,9 @@ impl<F: PrimeField32> RiscvAir<F> {
         chips.push(RiscvAir::Bls12381Fp2Op(bls12381_fp2_op));
         let bls12381_g1_decompress = Bls12381G1DecompressChip::new();
         chips.push(RiscvAir::Bls12381G1Decompress(bls12381_g1_decompress));
+        let blake_2s_round = Blake2sRoundChip::new();
+        chips.push(RiscvAir::Blake2sRound(blake_2s_round));
+
         let div_rem = DivRemChip;
         chips.push(RiscvAir::DivRem(div_rem));
 
@@ -184,8 +187,6 @@ impl<F: PrimeField32> RiscvAir<F> {
         chips.push(RiscvAir::ProgramMemory(program_memory_init));
         let byte = ByteChip::default();
         chips.push(RiscvAir::ByteLookup(byte));
-        let blake_2s_round = Blake2sRoundChip::new();
-        chips.push(RiscvAir::Blake2sRound(blake_2s_round));
 
         chips
     }

core/src/syscall/precompiles/blake2s/air.rs

@@ -191,8 +191,6 @@ where
             );
         }
 
-        self.constrain_shuffled_indices(builder, &local.shuffled_indices, local.is_real);
-
         for i in 0..4usize {
             // 3x
 

core/src/syscall/precompiles/blake2s/columns.rs

@@ -15,8 +15,6 @@ pub struct Blake2sRoundCols<T> {
     pub a_ptr: T,
     pub b_ptr: T,
 
-    pub shuffled_indices: [T; 16],
-
     /// Memory layout:
     /// a: v0[0] || v0[1] || v0[2] || v0[3] ||
     ///    v1[0] || v1[1] || v1[2] || v1[3] ||

core/src/syscall/precompiles/blake2s/mod.rs

@@ -6,9 +6,6 @@ mod trace;
 pub use columns::*;
 
 use crate::runtime::{MemoryReadRecord, MemoryWriteRecord};
-use crate::stark::SphinxAirBuilder;
-use p3_air::AirBuilder;
-use p3_field::AbstractField;
 use serde::Deserialize;
 use serde::Serialize;
 
@@ -31,23 +28,6 @@ impl Blake2sRoundChip {
     pub fn new() -> Self {
         Blake2sRoundChip
     }
-    pub fn constrain_shuffled_indices<AB: SphinxAirBuilder>(
-        &self,
-        builder: &mut AB,
-        indices: &[AB::Var],
-        is_real: AB::Var,
-    ) {
-        for index in 0..4 {
-            builder
-                .when(is_real)
-                .assert_eq(indices[index], AB::F::from_canonical_usize(0));
-        }
-        for index in 4..indices.len() {
-            builder
-                .when(is_real)
-                .assert_eq(indices[index], AB::F::from_canonical_usize(1));
-        }
-    }
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]

core/src/syscall/precompiles/blake2s/trace.rs

@@ -153,10 +153,6 @@ impl<F: PrimeField32> MachineAir<F> for Blake2sRoundChip {
 
             // 3x (m2, R1, R2)
             for i in 0..4usize {
-                cols.shuffled_indices[i + 4] = F::from_canonical_u32(1);
-                cols.shuffled_indices[i + 8] = F::from_canonical_u32(1);
-                cols.shuffled_indices[i + 12] = F::from_canonical_u32(1);
-
                 let v0 = v0_outer[i];
                 let v1 = v1_outer[i];
                 let v2 = v2_outer[i];

core/src/utils/prove.rs

@@ -356,11 +356,11 @@ where
     OpeningProof<SC>: Send + Sync,
     ShardMainData<SC>: Serialize + DeserializeOwned,
 {
-    //#[cfg(debug_assertions)]
-    //{
-    let record_clone = record.clone();
-    machine.debug_constraints(pk, record_clone);
-    //}
+    #[cfg(debug_assertions)]
+    {
+        let record_clone = record.clone();
+        machine.debug_constraints(pk, record_clone);
+    }
     let stats = record.stats().clone();
     let cycles = stats.get("cpu_events").unwrap();
 

zkvm/precompiles/src/lib.rs

@@ -47,13 +47,5 @@ extern "C" {
     pub fn sys_alloc_aligned(bytes: usize, align: usize) -> *mut u8;
     pub fn syscall_bls12381_g2_add(p: *mut u32, q: *const u32);
     pub fn syscall_bls12381_g2_double(p: *mut u32);
-    pub fn syscall_blake2s_xor_rotate_16(w: *mut u32);
-
-    pub fn syscall_blake2s_add_2(left: *mut u32, right: *const u32);
-    pub fn syscall_blake2s_add_3(left: *mut u32, right: *const u32);
-    pub fn syscall_blake2s_xor_rotate_right_16(left: *mut u32, right: *const u32);
-    pub fn syscall_blake2s_xor_rotate_right_12(left: *mut u32, right: *const u32);
-    pub fn syscall_blake2s_xor_rotate_right_8(left: *mut u32, right: *const u32);
-    pub fn syscall_blake2s_xor_rotate_right_7(left: *mut u32, right: *const u32);
-    pub fn syscall_blake2s_quarter_round(left: *mut u32, right: *const u32);
+    pub fn syscall_blake2s_round(left: *mut u32, right: *const u32);
 }