Add GPU based PoH verification (#4524)

* Add GPU poh verify

* Switch to single PoH verify function

* Add EntrySlice verify tests with hashes and txs

* Add poh-verify benchmarks

core/benches/poh_verify.rs

@@ -0,0 +1,48 @@
+#![feature(test)]
+extern crate test;
+
+use solana::entry::EntrySlice;
+use solana::entry::{next_entry_mut, Entry};
+use solana_sdk::hash::{hash, Hash};
+use solana_sdk::signature::{Keypair, KeypairUtil};
+use solana_sdk::system_transaction;
+use test::Bencher;
+
+const NUM_HASHES: u64 = 400;
+const NUM_ENTRIES: usize = 800;
+
+#[bench]
+fn bench_poh_verify_ticks(bencher: &mut Bencher) {
+    let zero = Hash::default();
+    let mut cur_hash = hash(&zero.as_ref());
+    let start = *&cur_hash;
+
+    let mut ticks: Vec<Entry> = Vec::with_capacity(NUM_ENTRIES);
+    for _ in 0..NUM_ENTRIES {
+        ticks.push(next_entry_mut(&mut cur_hash, NUM_HASHES, vec![]));
+    }
+
+    bencher.iter(|| {
+        ticks.verify(&start);
+    })
+}
+
+#[bench]
+fn bench_poh_verify_transaction_entries(bencher: &mut Bencher) {
+    let zero = Hash::default();
+    let mut cur_hash = hash(&zero.as_ref());
+    let start = *&cur_hash;
+
+    let keypair1 = Keypair::new();
+    let pubkey1 = keypair1.pubkey();
+
+    let mut ticks: Vec<Entry> = Vec::with_capacity(NUM_ENTRIES);
+    for _ in 0..NUM_ENTRIES {
+        let tx = system_transaction::create_user_account(&keypair1, &pubkey1, 42, cur_hash);
+        ticks.push(next_entry_mut(&mut cur_hash, NUM_HASHES, vec![tx]));
+    }
+
+    bencher.iter(|| {
+        ticks.verify(&start);
+    })
+}

core/src/entry.rs

@@ -16,6 +16,13 @@ use std::borrow::Borrow;
 use std::sync::mpsc::{Receiver, Sender};
 use std::sync::{Arc, RwLock};
 
+#[cfg(feature = "cuda")]
+use crate::sigverify::poh_verify_many;
+#[cfg(feature = "cuda")]
+use std::sync::Mutex;
+#[cfg(feature = "cuda")]
+use std::thread;
+
 pub type EntrySender = Sender<Vec<Entry>>;
 pub type EntryReceiver = Receiver<Vec<Entry>>;
 
@@ -212,6 +219,7 @@ pub trait EntrySlice {
 }
 
 impl EntrySlice for [Entry] {
+    #[cfg(not(feature = "cuda"))]
     fn verify(&self, start_hash: &Hash) -> bool {
         let genesis = [Entry {
             num_hashes: 0,
@@ -233,6 +241,78 @@ impl EntrySlice for [Entry] {
         })
     }
 
+    #[cfg(feature = "cuda")]
+    fn verify(&self, start_hash: &Hash) -> bool {
+        let genesis = [Entry {
+            num_hashes: 0,
+            hash: *start_hash,
+            transactions: vec![],
+        }];
+
+        let hashes: Vec<Hash> = genesis
+            .par_iter()
+            .chain(self)
+            .map(|entry| entry.hash)
+            .take(self.len())
+            .collect();
+
+        let num_hashes_vec: Vec<u64> = self
+            .into_iter()
+            .map(|entry| entry.num_hashes.saturating_sub(1))
+            .collect();
+
+        let length = self.len();
+        let hashes = Arc::new(Mutex::new(hashes));
+        let hashes_clone = hashes.clone();
+
+        let gpu_verify_thread = thread::spawn(move || {
+            let mut hashes = hashes_clone.lock().unwrap();
+            let res;
+            unsafe {
+                res = poh_verify_many(
+                    hashes.as_mut_ptr() as *mut u8,
+                    num_hashes_vec.as_ptr(),
+                    length,
+                    1,
+                );
+            }
+            if res != 0 {
+                panic!("GPU PoH verify many failed");
+            }
+        });
+
+        let tx_hashes: Vec<Option<Hash>> = self
+            .into_par_iter()
+            .map(|entry| {
+                if entry.transactions.is_empty() {
+                    None
+                } else {
+                    Some(hash_transactions(&entry.transactions))
+                }
+            })
+            .collect();
+
+        gpu_verify_thread.join().unwrap();
+
+        let hashes = Arc::try_unwrap(hashes).unwrap().into_inner().unwrap();
+        hashes
+            .into_par_iter()
+            .zip(tx_hashes)
+            .zip(self)
+            .all(|((hash, tx_hash), answer)| {
+                if answer.num_hashes == 0 {
+                    hash == answer.hash
+                } else {
+                    let mut poh = Poh::new(hash, None);
+                    if let Some(mixin) = tx_hash {
+                        poh.record(mixin).unwrap().hash == answer.hash
+                    } else {
+                        poh.tick().unwrap().hash == answer.hash
+                    }
+                }
+            })
+    }
+
     fn to_blobs(&self) -> Vec<Blob> {
         split_serializable_chunks(
             &self,
@@ -556,6 +636,53 @@ mod tests {
         assert!(!bad_ticks.verify(&zero)); // inductive step, bad
     }
 
+    #[test]
+    fn test_verify_slice_with_hashes() {
+        solana_logger::setup();
+        let zero = Hash::default();
+        let one = hash(&zero.as_ref());
+        let two = hash(&one.as_ref());
+        assert!(vec![][..].verify(&one)); // base case
+        assert!(vec![Entry::new_tick(1, &two)][..].verify(&one)); // singleton case 1
+        assert!(!vec![Entry::new_tick(1, &two)][..].verify(&two)); // singleton case 2, bad
+
+        let mut ticks = vec![next_entry(&one, 1, vec![])];
+        ticks.push(next_entry(&ticks.last().unwrap().hash, 1, vec![]));
+        assert!(ticks.verify(&one)); // inductive step
+
+        let mut bad_ticks = vec![next_entry(&one, 1, vec![])];
+        bad_ticks.push(next_entry(&bad_ticks.last().unwrap().hash, 1, vec![]));
+        bad_ticks[1].hash = one;
+        assert!(!bad_ticks.verify(&one)); // inductive step, bad
+    }
+
+    #[test]
+    fn test_verify_slice_with_hashes_and_transactions() {
+        solana_logger::setup();
+        let zero = Hash::default();
+        let one = hash(&zero.as_ref());
+        let two = hash(&one.as_ref());
+        let alice_pubkey = Keypair::default();
+        let tx0 = create_sample_payment(&alice_pubkey, one);
+        let tx1 = create_sample_timestamp(&alice_pubkey, one);
+        assert!(vec![][..].verify(&one)); // base case
+        assert!(vec![next_entry(&one, 1, vec![tx0.clone()])][..].verify(&one)); // singleton case 1
+        assert!(!vec![next_entry(&one, 1, vec![tx0.clone()])][..].verify(&two)); // singleton case 2, bad
+
+        let mut ticks = vec![next_entry(&one, 1, vec![tx0.clone()])];
+        ticks.push(next_entry(
+            &ticks.last().unwrap().hash,
+            1,
+            vec![tx1.clone()],
+        ));
+        assert!(ticks.verify(&one)); // inductive step
+
+        let mut bad_ticks = vec![next_entry(&one, 1, vec![tx0])];
+        bad_ticks.push(next_entry(&bad_ticks.last().unwrap().hash, 1, vec![tx1]));
+        bad_ticks[1].hash = one;
+        assert!(!bad_ticks.verify(&one)); // inductive step, bad
+    }
+
     fn blob_sized_entries(num_entries: usize) -> Vec<Entry> {
         // rough guess
         let mut magic_len = BLOB_DATA_SIZE

core/src/sigverify.rs

@@ -17,6 +17,9 @@ use solana_sdk::signature::Signature;
 use solana_sdk::transaction::Transaction;
 use std::mem::size_of;
 
+#[cfg(feature = "cuda")]
+use std::os::raw::c_int;
+
 pub const NUM_THREADS: u32 = 10;
 use std::cell::RefCell;
 
@@ -68,6 +71,13 @@ extern "C" {
 
     pub fn chacha_init_sha_state(sha_state: *mut u8, num_keys: u32);
     pub fn chacha_end_sha_state(sha_state_in: *const u8, out: *mut u8, num_keys: u32);
+
+    pub fn poh_verify_many(
+        hashes: *mut u8,
+        num_hashes_arr: *const u64,
+        num_elems: usize,
+        use_non_default_stream: u8,
+    ) -> c_int;
 }
 
 #[cfg(not(feature = "cuda"))]

fetch-perf-libs.sh

@@ -17,7 +17,7 @@ if [[ ! -d target/perf-libs ]]; then
   (
     set -x
     cd target/perf-libs
-    curl https://solana-perf.s3.amazonaws.com/v0.12.1/x86_64-unknown-linux-gnu/solana-perf.tgz | tar zxvf -
+    curl https://solana-perf.s3.amazonaws.com/v0.13.2/x86_64-unknown-linux-gnu/solana-perf.tgz | tar zxvf -
   )
 fi
 

sdk/src/hash.rs

@@ -1,15 +1,15 @@
 //! The `hash` module provides functions for creating SHA-256 hashes.
 
 use bs58;
-use generic_array::typenum::U32;
-use generic_array::GenericArray;
 use sha2::{Digest, Sha256};
+use std::convert::TryFrom;
 use std::fmt;
 use std::mem;
 use std::str::FromStr;
 
 #[derive(Serialize, Deserialize, Clone, Copy, Default, Eq, PartialEq, Ord, PartialOrd, Hash)]
-pub struct Hash(GenericArray<u8, U32>);
+#[repr(transparent)]
+pub struct Hash([u8; 32]);
 
 #[derive(Clone, Default)]
 pub struct Hasher {
@@ -28,9 +28,7 @@ impl Hasher {
     pub fn result(self) -> Hash {
         // At the time of this writing, the sha2 library is stuck on an old version
         // of generic_array (0.9.0). Decouple ourselves with a clone to our version.
-        Hash(GenericArray::clone_from_slice(
+        Hash(<[u8; 32]>::try_from(self.hasher.result().as_slice()).unwrap())
-            self.hasher.result().as_slice(),
-        ))
     }
 }
 
@@ -75,7 +73,7 @@ impl FromStr for Hash {
 
 impl Hash {
     pub fn new(hash_slice: &[u8]) -> Self {
-        Hash(GenericArray::clone_from_slice(&hash_slice))
+        Hash(<[u8; 32]>::try_from(hash_slice).unwrap())
     }
 }