Sign shreds on the GPU (#6595)

* sign gpu shreds

* wip

* checks

* tests build

* test

* tests

* test

* nits

* sign cpu test

* write out the sigs in parallel

* clippy

* cpu test

* prepare secret for gpu

* woot!

* update

* bump perf libs

Cargo.lock

@@ -3287,6 +3287,7 @@ dependencies = [
  "serde_json 1.0.41 (registry+https://github.com/rust-lang/crates.io-index)",
  "serial_test 0.2.0 (registry+https://github.com/rust-lang/crates.io-index)",
  "serial_test_derive 0.2.0 (registry+https://github.com/rust-lang/crates.io-index)",
+ "sha2 0.8.0 (registry+https://github.com/rust-lang/crates.io-index)",
  "solana-budget-api 0.21.0",
  "solana-budget-program 0.21.0",
  "solana-chacha-sys 0.21.0",

core/Cargo.toml

@@ -44,6 +44,7 @@ rayon = "1.2.0"
 serde = "1.0.102"
 serde_derive = "1.0.102"
 serde_json = "1.0.41"
+sha2 = "0.8.0"
 solana-budget-api = { path = "../programs/budget_api", version = "0.21.0" }
 solana-budget-program = { path = "../programs/budget_program", version = "0.21.0" }
 solana-chacha-sys = { path = "../chacha-sys", version = "0.21.0" }

core/src/cuda_runtime.rs

@@ -67,15 +67,9 @@ pub struct PinnedVec<T> {
     pinnable: bool,
 }
 
-impl Reset for PinnedVec<u8> {
+impl<T: Default + Clone> Reset for PinnedVec<T> {
     fn reset(&mut self) {
-        self.resize(0, 0u8);
-    }
-}
-
-impl Reset for PinnedVec<u32> {
-    fn reset(&mut self) {
-        self.resize(0, 0u32);
+        self.resize(0, T::default());
     }
 }
 

core/src/packet.rs

@@ -62,12 +62,6 @@ impl Reset for Packets {
     }
 }
 
-impl Reset for PinnedVec<Packet> {
-    fn reset(&mut self) {
-        self.resize(0, Packet::default());
-    }
-}
-
 //auto derive doesn't support large arrays
 impl Default for Packets {
     fn default() -> Packets {

core/src/sigverify_shreds.rs

@@ -2,20 +2,22 @@
 use crate::cuda_runtime::PinnedVec;
 use crate::packet::{Packet, Packets};
 use crate::recycler::Recycler;
-use crate::recycler::Reset;
 use crate::sigverify::{self, TxOffset};
 use crate::sigverify_stage::SigVerifier;
 use bincode::deserialize;
+use rayon::iter::IndexedParallelIterator;
 use rayon::iter::IntoParallelIterator;
+use rayon::iter::IntoParallelRefMutIterator;
 use rayon::iter::ParallelIterator;
 use rayon::ThreadPool;
+use sha2::{Digest, Sha512};
+use solana_ed25519_dalek::{Keypair, PublicKey, SecretKey};
 use solana_ledger::bank_forks::BankForks;
 use solana_ledger::leader_schedule_cache::LeaderScheduleCache;
 use solana_ledger::perf_libs;
 use solana_ledger::shred::ShredType;
 use solana_metrics::inc_new_counter_debug;
 use solana_rayon_threadlimit::get_thread_count;
-use solana_sdk::pubkey::Pubkey;
 use solana_sdk::signature::Signature;
 use std::collections::{HashMap, HashSet};
 use std::mem::size_of;
@@ -28,7 +30,7 @@ pub struct ShredSigVerifier {
     bank_forks: Arc<RwLock<BankForks>>,
     leader_schedule_cache: Arc<LeaderScheduleCache>,
     recycler_offsets: Recycler<TxOffset>,
-    recycler_pubkeys: Recycler<PinnedVec<Pubkey>>,
+    recycler_keys: Recycler<PinnedVec<[u8; 32]>>,
     recycler_out: Recycler<PinnedVec<u8>>,
 }
 
@@ -42,7 +44,7 @@ impl ShredSigVerifier {
             bank_forks,
             leader_schedule_cache,
             recycler_offsets: Recycler::default(),
-            recycler_pubkeys: Recycler::default(),
+            recycler_keys: Recycler::default(),
             recycler_out: Recycler::default(),
         }
     }
@@ -70,23 +72,22 @@ impl SigVerifier for ShredSigVerifier {
     fn verify_batch(&self, mut batches: Vec<Packets>) -> Vec<Packets> {
         let r_bank = self.bank_forks.read().unwrap().working_bank();
         let slots: HashSet<u64> = Self::read_slots(&batches);
-        let mut leader_slots: HashMap<u64, Pubkey> = slots
+        let mut leader_slots: HashMap<u64, [u8; 32]> = slots
             .into_iter()
             .filter_map(|slot| {
-                Some((
-                    slot,
-                    self.leader_schedule_cache
-                        .slot_leader_at(slot, Some(&r_bank))?,
-                ))
+                let key = self
+                    .leader_schedule_cache
+                    .slot_leader_at(slot, Some(&r_bank))?;
+                Some((slot, key.to_bytes()))
             })
             .collect();
-        leader_slots.insert(std::u64::MAX, Pubkey::default());
+        leader_slots.insert(std::u64::MAX, [0u8; 32]);
 
         let r = verify_shreds_gpu(
             &batches,
             &leader_slots,
             &self.recycler_offsets,
-            &self.recycler_pubkeys,
+            &self.recycler_keys,
             &self.recycler_out,
         );
         sigverify::mark_disabled(&mut batches, &r);
@@ -100,12 +101,6 @@ thread_local!(static PAR_THREAD_POOL: RefCell<ThreadPool> = RefCell::new(rayon::
                     .build()
                     .unwrap()));
 
-impl Reset for PinnedVec<Pubkey> {
-    fn reset(&mut self) {
-        self.resize(0, Pubkey::default());
-    }
-}
-
 /// Assuming layout is
 /// signature: Signature
 /// signed_msg: {
@@ -114,7 +109,7 @@ impl Reset for PinnedVec<Pubkey> {
 ///   ...
 /// }
 /// Signature is the first thing in the packet, and slot is the first thing in the signed message.
-fn verify_shred_cpu(packet: &Packet, slot_leaders: &HashMap<u64, Pubkey>) -> Option<u8> {
+fn verify_shred_cpu(packet: &Packet, slot_leaders: &HashMap<u64, [u8; 32]>) -> Option<u8> {
     let sig_start = 0;
     let sig_end = size_of::<Signature>();
     let slot_start = sig_end + size_of::<ShredType>();
@@ -127,19 +122,19 @@ fn verify_shred_cpu(packet: &Packet, slot_leaders: &HashMap<u64, Pubkey>) -> Opt
     }
     let slot: u64 = deserialize(&packet.data[slot_start..slot_end]).ok()?;
     trace!("slot {}", slot);
-    let pubkey: &Pubkey = slot_leaders.get(&slot)?;
+    let pubkey = slot_leaders.get(&slot)?;
     if packet.meta.size < sig_end {
         return Some(0);
     }
     let signature = Signature::new(&packet.data[sig_start..sig_end]);
     trace!("signature {}", signature);
-    if !signature.verify(pubkey.as_ref(), &packet.data[msg_start..msg_end]) {
+    if !signature.verify(pubkey, &packet.data[msg_start..msg_end]) {
         return Some(0);
     }
     Some(1)
 }
 
-fn verify_shreds_cpu(batches: &[Packets], slot_leaders: &HashMap<u64, Pubkey>) -> Vec<Vec<u8>> {
+fn verify_shreds_cpu(batches: &[Packets], slot_leaders: &HashMap<u64, [u8; 32]>) -> Vec<Vec<u8>> {
     use rayon::prelude::*;
     let count = sigverify::batch_size(batches);
     debug!("CPU SHRED ECDSA for {}", count);
@@ -149,7 +144,7 @@ fn verify_shreds_cpu(batches: &[Packets], slot_leaders: &HashMap<u64, Pubkey>) -
                 .into_par_iter()
                 .map(|p| {
                     p.packets
-                        .par_iter()
+                        .iter()
                         .map(|p| verify_shred_cpu(p, slot_leaders).unwrap_or(0))
                         .collect()
                 })
@@ -160,14 +155,17 @@ fn verify_shreds_cpu(batches: &[Packets], slot_leaders: &HashMap<u64, Pubkey>) -
     rv
 }
 
-fn shred_gpu_pubkeys(
+fn slot_key_data_for_gpu<
+    T: Sync + Sized + Default + std::fmt::Debug + Eq + std::hash::Hash + Clone + Copy,
+>(
+    offset_start: usize,
     batches: &[Packets],
-    slot_leaders: &HashMap<u64, Pubkey>,
+    slot_keys: &HashMap<u64, T>,
     recycler_offsets: &Recycler<TxOffset>,
-    recycler_pubkeys: &Recycler<PinnedVec<Pubkey>>,
-) -> (PinnedVec<Pubkey>, TxOffset, usize) {
+    recycler_keys: &Recycler<PinnedVec<T>>,
+) -> (PinnedVec<T>, TxOffset, usize) {
     //TODO: mark Pubkey::default shreds as failed after the GPU returns
-    assert_eq!(slot_leaders.get(&std::u64::MAX), Some(&Pubkey::default()));
+    assert_eq!(slot_keys.get(&std::u64::MAX), Some(&T::default()));
     let slots: Vec<Vec<u64>> = PAR_THREAD_POOL.with(|thread_pool| {
         thread_pool.borrow().install(|| {
             batches
@@ -184,7 +182,7 @@ fn shred_gpu_pubkeys(
                             let slot: Option<u64> =
                                 deserialize(&packet.data[slot_start..slot_end]).ok();
                             match slot {
-                                Some(slot) if slot_leaders.get(&slot).is_some() => slot,
+                                Some(slot) if slot_keys.get(&slot).is_some() => slot,
                                 _ => std::u64::MAX,
                             }
                         })
@@ -193,20 +191,20 @@ fn shred_gpu_pubkeys(
                 .collect()
         })
     });
-    let mut keys_to_slots: HashMap<Pubkey, Vec<u64>> = HashMap::new();
+    let mut keys_to_slots: HashMap<T, Vec<u64>> = HashMap::new();
     for batch in slots.iter() {
         for slot in batch.iter() {
-            let key = slot_leaders.get(slot).unwrap();
+            let key = slot_keys.get(slot).unwrap();
             keys_to_slots
                 .entry(*key)
                 .or_insert_with(|| vec![])
                 .push(*slot);
         }
     }
-    let mut pubkeys: PinnedVec<Pubkey> = recycler_pubkeys.allocate("shred_gpu_pubkeys");
+    let mut keyvec = recycler_keys.allocate("shred_gpu_pubkeys");
     let mut slot_to_key_ix = HashMap::new();
     for (i, (k, slots)) in keys_to_slots.iter().enumerate() {
-        pubkeys.push(*k);
+        keyvec.push(*k);
         for s in slots {
             slot_to_key_ix.insert(s, i);
         }
@@ -214,33 +212,34 @@ fn shred_gpu_pubkeys(
     let mut offsets = recycler_offsets.allocate("shred_offsets");
     slots.iter().for_each(|packet_slots| {
         packet_slots.iter().for_each(|slot| {
-            offsets.push((slot_to_key_ix.get(slot).unwrap() * size_of::<Pubkey>()) as u32);
+            offsets
+                .push((offset_start + (slot_to_key_ix.get(slot).unwrap() * size_of::<T>())) as u32);
         });
     });
     //HACK: Pubkeys vector is passed along as a `Packets` buffer to the GPU
     //TODO: GPU needs a more opaque interface, which can handle variable sized structures for data
     //Pad the Pubkeys buffer such that it is bigger than a buffer of Packet sized elems
     let num_in_packets =
-        (pubkeys.len() * size_of::<Pubkey>() + (size_of::<Packet>() - 1)) / size_of::<Packet>();
+        (keyvec.len() * size_of::<T>() + (size_of::<Packet>() - 1)) / size_of::<Packet>();
     trace!("num_in_packets {}", num_in_packets);
     //number of bytes missing
-    let missing = num_in_packets * size_of::<Packet>() - pubkeys.len() * size_of::<Pubkey>();
+    let missing = num_in_packets * size_of::<Packet>() - keyvec.len() * size_of::<T>();
     trace!("missing {}", missing);
     //extra Pubkeys needed to fill the buffer
-    let extra = (missing + size_of::<Pubkey>() - 1) / size_of::<Pubkey>();
+    let extra = (missing + size_of::<T>() - 1) / size_of::<T>();
     trace!("extra {}", extra);
-    trace!("pubkeys {}", pubkeys.len());
+    trace!("keyvec {}", keyvec.len());
     for _ in 0..extra {
-        pubkeys.push(Pubkey::default());
-        trace!("pubkeys {}", pubkeys.len());
+        keyvec.push(T::default());
+        trace!("keyvec {}", keyvec.len());
     }
-    trace!("pubkeys {:?}", pubkeys);
+    trace!("keyvec {:?}", keyvec);
     trace!("offsets {:?}", offsets);
-    (pubkeys, offsets, num_in_packets)
+    (keyvec, offsets, num_in_packets)
 }
 
 fn shred_gpu_offsets(
-    mut pubkeys_end: u32,
+    mut pubkeys_end: usize,
     batches: &[Packets],
     recycler_offsets: &Recycler<TxOffset>,
 ) -> (TxOffset, TxOffset, TxOffset, Vec<Vec<u32>>) {
@@ -252,19 +251,19 @@ fn shred_gpu_offsets(
         let mut sig_lens = Vec::new();
         for packet in &batch.packets {
             let sig_start = pubkeys_end;
-            let sig_end = sig_start + size_of::<Signature>() as u32;
+            let sig_end = sig_start + size_of::<Signature>();
             let msg_start = sig_end;
-            let msg_end = sig_start + packet.meta.size as u32;
-            signature_offsets.push(sig_start);
-            msg_start_offsets.push(msg_start);
+            let msg_end = sig_start + packet.meta.size;
+            signature_offsets.push(sig_start as u32);
+            msg_start_offsets.push(msg_start as u32);
             let msg_size = if msg_end < msg_start {
                 0
             } else {
                 msg_end - msg_start
             };
-            msg_sizes.push(msg_size);
+            msg_sizes.push(msg_size as u32);
             sig_lens.push(1);
-            pubkeys_end += size_of::<Packet>() as u32;
+            pubkeys_end += size_of::<Packet>();
         }
         v_sig_lens.push(sig_lens);
     }
@@ -273,9 +272,9 @@ fn shred_gpu_offsets(
 
 fn verify_shreds_gpu(
     batches: &[Packets],
-    slot_leaders: &HashMap<u64, Pubkey>,
+    slot_leaders: &HashMap<u64, [u8; 32]>,
     recycler_offsets: &Recycler<TxOffset>,
-    recycler_pubkeys: &Recycler<PinnedVec<Pubkey>>,
+    recycler_pubkeys: &Recycler<PinnedVec<[u8; 32]>>,
     recycler_out: &Recycler<PinnedVec<u8>>,
 ) -> Vec<Vec<u8>> {
     let api = perf_libs::api();
@@ -288,10 +287,10 @@ fn verify_shreds_gpu(
     let mut rvs = Vec::new();
     let count = sigverify::batch_size(batches);
     let (pubkeys, pubkey_offsets, mut num_packets) =
-        shred_gpu_pubkeys(batches, slot_leaders, recycler_offsets, recycler_pubkeys);
+        slot_key_data_for_gpu(0, batches, slot_leaders, recycler_offsets, recycler_pubkeys);
     //HACK: Pubkeys vector is passed along as a `Packets` buffer to the GPU
     //TODO: GPU needs a more opaque interface, which can handle variable sized structures for data
-    let pubkeys_len = (num_packets * size_of::<Packet>()) as u32;
+    let pubkeys_len = num_packets * size_of::<Packet>();
     trace!("num_packets: {}", num_packets);
     trace!("pubkeys_len: {}", pubkeys_len);
     let (signature_offsets, msg_start_offsets, msg_sizes, v_sig_lens) =
@@ -355,6 +354,217 @@ fn verify_shreds_gpu(
     rvs
 }
 
+/// Assuming layout is
+/// signature: Signature
+/// signed_msg: {
+///   type: ShredType
+///   slot: u64,
+///   ...
+/// }
+/// Signature is the first thing in the packet, and slot is the first thing in the signed message.
+fn sign_shred_cpu(
+    packet: &mut Packet,
+    slot_leaders_pubkeys: &HashMap<u64, [u8; 32]>,
+    slot_leaders_privkeys: &HashMap<u64, [u8; 32]>,
+) {
+    let sig_start = 0;
+    let sig_end = sig_start + size_of::<Signature>();
+    let slot_start = sig_end + size_of::<ShredType>();
+    let slot_end = slot_start + size_of::<u64>();
+    let msg_start = sig_end;
+    let msg_end = packet.meta.size;
+    trace!("slot start and end {} {}", slot_start, slot_end);
+    assert!(
+        packet.meta.size >= slot_end,
+        "packet is not large enough for a slot"
+    );
+    let slot: u64 =
+        deserialize(&packet.data[slot_start..slot_end]).expect("can't deserialize slot");
+    trace!("slot {}", slot);
+    let pubkey = slot_leaders_pubkeys
+        .get(&slot)
+        .expect("slot pubkey missing");
+    let privkey = slot_leaders_privkeys
+        .get(&slot)
+        .expect("slot privkey missing");
+    let keypair = Keypair {
+        secret: SecretKey::from_bytes(&privkey[0..32]).expect("dalek privkey parser"),
+        public: PublicKey::from_bytes(&pubkey[0..32]).expect("dalek pubkey parser"),
+    };
+    assert!(
+        packet.meta.size >= sig_end,
+        "packet is not large enough for a signature"
+    );
+    let signature = keypair.sign(&packet.data[msg_start..msg_end]);
+    trace!("signature {:?}", signature);
+    packet.data[0..sig_end].copy_from_slice(&signature.to_bytes());
+}
+
+fn sign_shreds_cpu(
+    batches: &mut [Packets],
+    slot_leaders_pubkeys: &HashMap<u64, [u8; 32]>,
+    slot_leaders_privkeys: &HashMap<u64, [u8; 32]>,
+) {
+    use rayon::prelude::*;
+    let count = sigverify::batch_size(batches);
+    debug!("CPU SHRED ECDSA for {}", count);
+    PAR_THREAD_POOL.with(|thread_pool| {
+        thread_pool.borrow().install(|| {
+            batches.par_iter_mut().for_each(|p| {
+                p.packets.iter_mut().for_each(|mut p| {
+                    sign_shred_cpu(&mut p, slot_leaders_pubkeys, slot_leaders_privkeys)
+                });
+            });
+        })
+    });
+    inc_new_counter_debug!("ed25519_shred_verify_cpu", count);
+}
+
+pub fn sign_shreds_gpu(
+    batches: &mut [Packets],
+    slot_leaders_pubkeys: &HashMap<u64, [u8; 32]>,
+    slot_leaders_privkeys: &HashMap<u64, [u8; 32]>,
+    recycler_offsets: &Recycler<TxOffset>,
+    recycler_pubkeys: &Recycler<PinnedVec<[u8; 32]>>,
+    recycler_secrets: &Recycler<PinnedVec<Signature>>,
+    recycler_out: &Recycler<PinnedVec<u8>>,
+) {
+    let sig_size = size_of::<Signature>();
+    let api = perf_libs::api();
+    if api.is_none() {
+        return sign_shreds_cpu(batches, slot_leaders_pubkeys, slot_leaders_privkeys);
+    }
+    let slot_leaders_secrets: HashMap<u64, Signature> = slot_leaders_privkeys
+        .iter()
+        .map(|(k, v)| {
+            if *k == std::u64::MAX {
+                (*k, Signature::default())
+            } else {
+                let mut hasher = Sha512::default();
+                hasher.input(&v);
+                let mut result = hasher.result();
+                result[0] &= 248;
+                result[31] &= 63;
+                result[31] |= 64;
+                let sig = Signature::new(result.as_slice());
+                (*k, sig)
+            }
+        })
+        .collect();
+    let api = api.unwrap();
+
+    let mut elems = Vec::new();
+    let count = sigverify::batch_size(batches);
+    let mut offset: usize = 0;
+    let mut num_packets = 0;
+    let (pubkeys, pubkey_offsets, num_pubkey_packets) = slot_key_data_for_gpu(
+        offset,
+        batches,
+        slot_leaders_pubkeys,
+        recycler_offsets,
+        recycler_pubkeys,
+    );
+    offset += num_pubkey_packets * size_of::<Packet>();
+    num_packets += num_pubkey_packets;
+    trace!("offset: {}", offset);
+    let (secrets, secret_offsets, num_secret_packets) = slot_key_data_for_gpu(
+        offset,
+        batches,
+        &slot_leaders_secrets,
+        recycler_offsets,
+        recycler_secrets,
+    );
+    offset += num_secret_packets * size_of::<Packet>();
+    num_packets += num_secret_packets;
+    //HACK: Pubkeys vector is passed along as a `Packets` buffer to the GPU
+    //TODO: GPU needs a more opaque interface, which can handle variable sized structures for data
+    trace!("offset: {}", offset);
+    let (signature_offsets, msg_start_offsets, msg_sizes, _v_sig_lens) =
+        shred_gpu_offsets(offset, batches, recycler_offsets);
+    let total_sigs = signature_offsets.len();
+    let mut signatures_out = recycler_out.allocate("ed25519 signatures");
+    signatures_out.resize(total_sigs * sig_size, 0);
+    elems.push(
+        perf_libs::Elems {
+            #![allow(clippy::cast_ptr_alignment)]
+            elems: pubkeys.as_ptr() as *const solana_sdk::packet::Packet,
+            num: num_pubkey_packets as u32,
+        },
+    );
+
+    elems.push(
+        perf_libs::Elems {
+            #![allow(clippy::cast_ptr_alignment)]
+            elems: secrets.as_ptr() as *const solana_sdk::packet::Packet,
+            num: num_secret_packets as u32,
+        },
+    );
+
+    for p in batches.iter() {
+        elems.push(perf_libs::Elems {
+            elems: p.packets.as_ptr(),
+            num: p.packets.len() as u32,
+        });
+        let mut v = Vec::new();
+        v.resize(p.packets.len(), 0);
+        num_packets += p.packets.len();
+    }
+
+    trace!("Starting verify num packets: {}", num_packets);
+    trace!("elem len: {}", elems.len() as u32);
+    trace!("packet sizeof: {}", size_of::<Packet>() as u32);
+    const USE_NON_DEFAULT_STREAM: u8 = 1;
+    unsafe {
+        let res = (api.ed25519_sign_many)(
+            elems.as_mut_ptr(),
+            elems.len() as u32,
+            size_of::<Packet>() as u32,
+            num_packets as u32,
+            total_sigs as u32,
+            msg_sizes.as_ptr(),
+            pubkey_offsets.as_ptr(),
+            secret_offsets.as_ptr(),
+            msg_start_offsets.as_ptr(),
+            signatures_out.as_mut_ptr(),
+            USE_NON_DEFAULT_STREAM,
+        );
+        if res != 0 {
+            trace!("RETURN!!!: {}", res);
+        }
+    }
+    trace!("done sign");
+    let mut sizes: Vec<usize> = vec![0];
+    sizes.extend(batches.iter().map(|b| b.packets.len()));
+    PAR_THREAD_POOL.with(|thread_pool| {
+        thread_pool.borrow().install(|| {
+            batches
+                .par_iter_mut()
+                .enumerate()
+                .for_each(|(batch_ix, batch)| {
+                    let num_packets = sizes[batch_ix];
+                    batch
+                        .packets
+                        .iter_mut()
+                        .enumerate()
+                        .for_each(|(packet_ix, packet)| {
+                            let sig_ix = packet_ix + num_packets;
+                            let sig_start = sig_ix * sig_size;
+                            let sig_end = sig_start + sig_size;
+                            packet.data[0..sig_size]
+                                .copy_from_slice(&signatures_out[sig_start..sig_end]);
+                        });
+                });
+        });
+    });
+    inc_new_counter_debug!("ed25519_shred_sign_gpu", count);
+    recycler_out.recycle(signatures_out);
+    recycler_offsets.recycle(signature_offsets);
+    recycler_offsets.recycle(pubkey_offsets);
+    recycler_offsets.recycle(msg_sizes);
+    recycler_offsets.recycle(msg_start_offsets);
+    recycler_pubkeys.recycle(pubkeys);
+}
+
 #[cfg(test)]
 pub mod tests {
     use super::*;
@@ -375,12 +585,18 @@ pub mod tests {
         packet.data[0..shred.payload.len()].copy_from_slice(&shred.payload);
         packet.meta.size = shred.payload.len();
 
-        let leader_slots = [(slot, keypair.pubkey())].iter().cloned().collect();
+        let leader_slots = [(slot, keypair.pubkey().to_bytes())]
+            .iter()
+            .cloned()
+            .collect();
         let rv = verify_shred_cpu(&packet, &leader_slots);
         assert_eq!(rv, Some(1));
 
         let wrong_keypair = Keypair::new();
-        let leader_slots = [(slot, wrong_keypair.pubkey())].iter().cloned().collect();
+        let leader_slots = [(slot, wrong_keypair.pubkey().to_bytes())]
+            .iter()
+            .cloned()
+            .collect();
         let rv = verify_shred_cpu(&packet, &leader_slots);
         assert_eq!(rv, Some(0));
 
@@ -401,12 +617,18 @@ pub mod tests {
         batch[0].packets[0].data[0..shred.payload.len()].copy_from_slice(&shred.payload);
         batch[0].packets[0].meta.size = shred.payload.len();
 
-        let leader_slots = [(slot, keypair.pubkey())].iter().cloned().collect();
+        let leader_slots = [(slot, keypair.pubkey().to_bytes())]
+            .iter()
+            .cloned()
+            .collect();
         let rv = verify_shreds_cpu(&batch, &leader_slots);
         assert_eq!(rv, vec![vec![1]]);
 
         let wrong_keypair = Keypair::new();
-        let leader_slots = [(slot, wrong_keypair.pubkey())].iter().cloned().collect();
+        let leader_slots = [(slot, wrong_keypair.pubkey().to_bytes())]
+            .iter()
+            .cloned()
+            .collect();
         let rv = verify_shreds_cpu(&batch, &leader_slots);
         assert_eq!(rv, vec![vec![0]]);
 
@@ -414,7 +636,10 @@ pub mod tests {
         let rv = verify_shreds_cpu(&batch, &leader_slots);
         assert_eq!(rv, vec![vec![0]]);
 
-        let leader_slots = [(slot, keypair.pubkey())].iter().cloned().collect();
+        let leader_slots = [(slot, keypair.pubkey().to_bytes())]
+            .iter()
+            .cloned()
+            .collect();
         batch[0].packets[0].meta.size = 0;
         let rv = verify_shreds_cpu(&batch, &leader_slots);
         assert_eq!(rv, vec![vec![0]]);
@@ -436,10 +661,13 @@ pub mod tests {
         batch[0].packets[0].data[0..shred.payload.len()].copy_from_slice(&shred.payload);
         batch[0].packets[0].meta.size = shred.payload.len();
 
-        let leader_slots = [(slot, keypair.pubkey()), (std::u64::MAX, Pubkey::default())]
-            .iter()
-            .cloned()
-            .collect();
+        let leader_slots = [
+            (slot, keypair.pubkey().to_bytes()),
+            (std::u64::MAX, [0u8; 32]),
+        ]
+        .iter()
+        .cloned()
+        .collect();
         let rv = verify_shreds_gpu(
             &batch,
             &leader_slots,
@@ -451,8 +679,8 @@ pub mod tests {
 
         let wrong_keypair = Keypair::new();
         let leader_slots = [
-            (slot, wrong_keypair.pubkey()),
-            (std::u64::MAX, Pubkey::default()),
+            (slot, wrong_keypair.pubkey().to_bytes()),
+            (std::u64::MAX, [0u8; 32]),
         ]
         .iter()
         .cloned()
@@ -466,10 +694,7 @@ pub mod tests {
         );
         assert_eq!(rv, vec![vec![0]]);
 
-        let leader_slots = [(std::u64::MAX, Pubkey::default())]
-            .iter()
-            .cloned()
-            .collect();
+        let leader_slots = [(std::u64::MAX, [0u8; 32])].iter().cloned().collect();
         let rv = verify_shreds_gpu(
             &batch,
             &leader_slots,
@@ -480,10 +705,13 @@ pub mod tests {
         assert_eq!(rv, vec![vec![0]]);
 
         batch[0].packets[0].meta.size = 0;
-        let leader_slots = [(slot, keypair.pubkey()), (std::u64::MAX, Pubkey::default())]
-            .iter()
-            .cloned()
-            .collect();
+        let leader_slots = [
+            (slot, keypair.pubkey().to_bytes()),
+            (std::u64::MAX, [0u8; 32]),
+        ]
+        .iter()
+        .cloned()
+        .collect();
         let rv = verify_shreds_gpu(
             &batch,
             &leader_slots,
@@ -494,6 +722,101 @@ pub mod tests {
         assert_eq!(rv, vec![vec![0]]);
     }
 
+    #[test]
+    fn test_sigverify_shreds_sign_gpu() {
+        solana_logger::setup();
+        let recycler_offsets = Recycler::default();
+        let recycler_pubkeys = Recycler::default();
+        let recycler_secrets = Recycler::default();
+        let recycler_out = Recycler::default();
+
+        let mut batch = [Packets::default()];
+        let slot = 0xdeadc0de;
+        let keypair = Keypair::new();
+        let shred = Shred::new_from_data(slot, 0xc0de, 0xdead, Some(&[1, 2, 3, 4]), true, true);
+        batch[0].packets.resize(1, Packet::default());
+        batch[0].packets[0].data[0..shred.payload.len()].copy_from_slice(&shred.payload);
+        batch[0].packets[0].meta.size = shred.payload.len();
+        let pubkeys = [
+            (slot, keypair.pubkey().to_bytes()),
+            (std::u64::MAX, [0u8; 32]),
+        ]
+        .iter()
+        .cloned()
+        .collect();
+        let privkeys = [
+            (slot, keypair.secret.to_bytes()),
+            (std::u64::MAX, [0u8; 32]),
+        ]
+        .iter()
+        .cloned()
+        .collect();
+        //unsigned
+        let rv = verify_shreds_gpu(
+            &batch,
+            &pubkeys,
+            &recycler_offsets,
+            &recycler_pubkeys,
+            &recycler_out,
+        );
+        assert_eq!(rv, vec![vec![0]]);
+        //signed
+        sign_shreds_gpu(
+            &mut batch,
+            &pubkeys,
+            &privkeys,
+            &recycler_offsets,
+            &recycler_pubkeys,
+            &recycler_secrets,
+            &recycler_out,
+        );
+        let rv = verify_shreds_cpu(&batch, &pubkeys);
+        assert_eq!(rv, vec![vec![1]]);
+
+        let rv = verify_shreds_gpu(
+            &batch,
+            &pubkeys,
+            &recycler_offsets,
+            &recycler_pubkeys,
+            &recycler_out,
+        );
+        assert_eq!(rv, vec![vec![1]]);
+    }
+
+    #[test]
+    fn test_sigverify_shreds_sign_cpu() {
+        solana_logger::setup();
+
+        let mut batch = [Packets::default()];
+        let slot = 0xdeadc0de;
+        let keypair = Keypair::new();
+        let shred = Shred::new_from_data(slot, 0xc0de, 0xdead, Some(&[1, 2, 3, 4]), true, true);
+        batch[0].packets.resize(1, Packet::default());
+        batch[0].packets[0].data[0..shred.payload.len()].copy_from_slice(&shred.payload);
+        batch[0].packets[0].meta.size = shred.payload.len();
+        let pubkeys = [
+            (slot, keypair.pubkey().to_bytes()),
+            (std::u64::MAX, [0u8; 32]),
+        ]
+        .iter()
+        .cloned()
+        .collect();
+        let privkeys = [
+            (slot, keypair.secret.to_bytes()),
+            (std::u64::MAX, [0u8; 32]),
+        ]
+        .iter()
+        .cloned()
+        .collect();
+        //unsigned
+        let rv = verify_shreds_cpu(&batch, &pubkeys);
+        assert_eq!(rv, vec![vec![0]]);
+        //signed
+        sign_shreds_cpu(&mut batch, &pubkeys, &privkeys);
+        let rv = verify_shreds_cpu(&batch, &pubkeys);
+        assert_eq!(rv, vec![vec![1]]);
+    }
+
     #[test]
     fn test_sigverify_shreds_read_slots() {
         solana_logger::setup();

fetch-perf-libs.sh

@@ -1,6 +1,6 @@
 #!/usr/bin/env bash
 
-PERF_LIBS_VERSION=v0.15.0
+PERF_LIBS_VERSION=v0.16.1
 VERSION=$PERF_LIBS_VERSION-1
 
 set -e

ledger/src/perf_libs.rs

@@ -39,6 +39,24 @@ pub struct Api<'a> {
         ) -> u32,
     >,
 
+    #[allow(clippy::type_complexity)]
+    pub ed25519_sign_many: Symbol<
+        'a,
+        unsafe extern "C" fn(
+            vecs: *mut Elems,
+            num: u32,          //number of vecs
+            message_size: u32, //size of each element inside the elems field of the vec
+            total_packets: u32,
+            total_signatures: u32,
+            message_lens: *const u32,
+            pubkey_offsets: *const u32,
+            privkey_offsets: *const u32,
+            signed_message_offsets: *const u32,
+            sgnatures_out: *mut u8, //combined length of all the items in vecs
+            use_non_default_stream: u8,
+        ) -> u32,
+    >,
+
     pub chacha_cbc_encrypt_many_sample: Symbol<
         'a,
         unsafe extern "C" fn(

sdk/src/pubkey.rs

@@ -58,6 +58,9 @@ impl Pubkey {
             sol_log_64(0, 0, 0, i as u64, u64::from(*k));
         }
     }
+    pub fn to_bytes(self) -> [u8; 32] {
+        self.0
+    }
 }
 
 impl AsRef<[u8]> for Pubkey {