Merge pull request #6080 from Vurich/ethash

Unroll loops in light_compute

Cargo.lock

@@ -316,6 +316,11 @@ name = "crossbeam"
 version = "0.2.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 
+[[package]]
+name = "crunchy"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+
 [[package]]
 name = "crypt32-sys"
 version = "0.2.0"
@@ -445,6 +450,7 @@ dependencies = [
 name = "ethash"
 version = "1.8.0"
 dependencies = [
+ "crunchy 0.1.3 (registry+https://github.com/rust-lang/crates.io-index)",
  "log 0.3.7 (registry+https://github.com/rust-lang/crates.io-index)",
  "parking_lot 0.4.0 (registry+https://github.com/rust-lang/crates.io-index)",
  "primal 0.2.3 (registry+https://github.com/rust-lang/crates.io-index)",
@@ -3180,6 +3186,7 @@ dependencies = [
 "checksum core-foundation 0.2.2 (registry+https://github.com/rust-lang/crates.io-index)" = "20a6d0448d3a99d977ae4a2aa5a98d886a923e863e81ad9ff814645b6feb3bbd"
 "checksum core-foundation-sys 0.2.2 (registry+https://github.com/rust-lang/crates.io-index)" = "05eed248dc504a5391c63794fe4fb64f46f071280afaa1b73308f3c0ce4574c5"
 "checksum crossbeam 0.2.10 (registry+https://github.com/rust-lang/crates.io-index)" = "0c5ea215664ca264da8a9d9c3be80d2eaf30923c259d03e870388eb927508f97"
+"checksum crunchy 0.1.3 (registry+https://github.com/rust-lang/crates.io-index)" = "e6aa9cb5f2d7bffc4eecfaf924fe450549dc4f0c3a6502298dc24f968b1eabbe"
 "checksum crypt32-sys 0.2.0 (registry+https://github.com/rust-lang/crates.io-index)" = "e34988f7e069e0b2f3bfc064295161e489b2d4e04a2e4248fb94360cdf00b4ec"
 "checksum ctrlc 1.1.1 (git+https://github.com/paritytech/rust-ctrlc.git)" = "<none>"
 "checksum custom_derive 0.1.7 (registry+https://github.com/rust-lang/crates.io-index)" = "ef8ae57c4978a2acd8b869ce6b9ca1dfe817bff704c220209fdef2c0b75a01b9"

ethash/Cargo.toml

@@ -10,3 +10,7 @@ log = "0.3"
 sha3 = { path = "../util/sha3" }
 primal = "0.2.3"
 parking_lot = "0.4"
+crunchy = "0.1.0"
+
+[features]
+benches = []

ethash/src/compute.rs

@@ -289,24 +289,55 @@ fn hash_compute(light: &Light, full_size: usize, header_hash: &H256, nonce: u64)
 		let num_full_pages = (full_size / page_size) as u32;
 		let cache: &[Node] = &light.cache;  // deref once for better performance
 
-		for i in 0..(ETHASH_ACCESSES as u32) {
+		debug_assert_eq!(ETHASH_ACCESSES, 64);
-			let index = fnv_hash(f_mix.get_unchecked(0).as_words().get_unchecked(0) ^ i, *mix.get_unchecked(0).as_words().get_unchecked((i as usize) % MIX_WORDS)) % num_full_pages;
+		debug_assert_eq!(MIX_NODES, 2);
-			for n in 0..MIX_NODES {
+		debug_assert_eq!(NODE_WORDS, 16);
-				let tmp_node = calculate_dag_item(index * MIX_NODES as u32 + n as u32, cache);
+
-				for w in 0..NODE_WORDS {
+		unroll! {
-					*mix.get_unchecked_mut(n).as_words_mut().get_unchecked_mut(w) = fnv_hash(*mix.get_unchecked(n).as_words().get_unchecked(w), *tmp_node.as_words().get_unchecked(w));
+			// ETHASH_ACCESSES
+			for i_usize in 0..64 {
+				let i = i_usize as u32;
+
+				let index = fnv_hash(
+					f_mix.get_unchecked(0).as_words().get_unchecked(0) ^ i,
+					*mix.get_unchecked(0).as_words().get_unchecked(i as usize % MIX_WORDS)
+				) % num_full_pages;
+
+				unroll! {
+					// MIX_NODES
+					for n in 0..2 {
+						let tmp_node = calculate_dag_item(
+							index * MIX_NODES as u32 + n as u32,
+							cache,
+						);
+
+						unroll! {
+							// NODE_WORDS
+							for w in 0..16 {
+								*mix.get_unchecked_mut(n).as_words_mut().get_unchecked_mut(w) =
+									fnv_hash(
+										*mix.get_unchecked(n).as_words().get_unchecked(w),
+										*tmp_node.as_words().get_unchecked(w),
+									);
+							}
+						}
+					}
 				}
 			}
 		}
 
+		debug_assert_eq!(MIX_WORDS / 4, 8);
+
 		// compress mix
-		for i in 0..(MIX_WORDS / 4) {
+		unroll! {
-			let w = i * 4;
+			for i in 0..8 {
-			let mut reduction = *mix.get_unchecked(0).as_words().get_unchecked(w + 0);
+				let w = i * 4;
-			reduction = reduction.wrapping_mul(FNV_PRIME) ^ *mix.get_unchecked(0).as_words().get_unchecked(w + 1);
+				let mut reduction = *mix.get_unchecked(0).as_words().get_unchecked(w + 0);
-			reduction = reduction.wrapping_mul(FNV_PRIME) ^ *mix.get_unchecked(0).as_words().get_unchecked(w + 2);
+				reduction = reduction.wrapping_mul(FNV_PRIME) ^ *mix.get_unchecked(0).as_words().get_unchecked(w + 1);
-			reduction = reduction.wrapping_mul(FNV_PRIME) ^ *mix.get_unchecked(0).as_words().get_unchecked(w + 3);
+				reduction = reduction.wrapping_mul(FNV_PRIME) ^ *mix.get_unchecked(0).as_words().get_unchecked(w + 2);
-			*mix.get_unchecked_mut(0).as_words_mut().get_unchecked_mut(i) = reduction;
+				reduction = reduction.wrapping_mul(FNV_PRIME) ^ *mix.get_unchecked(0).as_words().get_unchecked(w + 3);
+				*mix.get_unchecked_mut(0).as_words_mut().get_unchecked_mut(i) = reduction;
+			}
 		}
 
 		let mut mix_hash = [0u8; 32];
@@ -331,13 +362,22 @@ fn calculate_dag_item(node_index: u32, cache: &[Node]) -> Node {
 		*ret.as_words_mut().get_unchecked_mut(0) ^= node_index;
 		sha3::sha3_512(ret.bytes.as_mut_ptr(), ret.bytes.len(), ret.bytes.as_ptr(), ret.bytes.len());
 
-		for i in 0..ETHASH_DATASET_PARENTS {
+		debug_assert_eq!(NODE_WORDS, 16);
+		for i in 0..ETHASH_DATASET_PARENTS as u32 {
 			let parent_index = fnv_hash(node_index ^ i, *ret.as_words().get_unchecked(i as usize % NODE_WORDS)) % num_parent_nodes as u32;
 			let parent = cache.get_unchecked(parent_index as usize);
-			for w in 0..NODE_WORDS {
+
-				*ret.as_words_mut().get_unchecked_mut(w) = fnv_hash(*ret.as_words().get_unchecked(w), *parent.as_words().get_unchecked(w));
+			unroll! {
+				for w in 0..16 {
+					*ret.as_words_mut().get_unchecked_mut(w) =
+						fnv_hash(
+							*ret.as_words().get_unchecked(w),
+							*parent.as_words().get_unchecked(w)
+						);
+				}
 			}
 		}
+
 		sha3::sha3_512(ret.bytes.as_mut_ptr(), ret.bytes.len(), ret.bytes.as_ptr(), ret.bytes.len());
 		ret
 	}
@@ -359,13 +399,20 @@ fn light_new<T: AsRef<Path>>(cache_dir: T, block_number: u64) -> Light {
 			sha3::sha3_512(nodes.get_unchecked_mut(i).bytes.as_mut_ptr(), NODE_BYTES, nodes.get_unchecked(i - 1).bytes.as_ptr(), NODE_BYTES);
 		}
 
+		debug_assert_eq!(NODE_WORDS, 16);
+
+		// This _should_ get unrolled by the compiler, since it's not using the loop variable.
 		for _ in 0..ETHASH_CACHE_ROUNDS {
 			for i in 0..num_nodes {
 				let idx = *nodes.get_unchecked_mut(i).as_words().get_unchecked(0) as usize % num_nodes;
 				let mut data = nodes.get_unchecked((num_nodes - 1 + i) % num_nodes).clone();
-				for w in 0..NODE_WORDS {
+
-					*data.as_words_mut().get_unchecked_mut(w) ^= *nodes.get_unchecked(idx).as_words().get_unchecked(w);
+				unroll! {
+					for w in 0..16 {
+						*data.as_words_mut().get_unchecked_mut(w) ^= *nodes.get_unchecked(idx).as_words().get_unchecked(w);
+					}
 				}
+
 				sha3_512(&data.bytes, &mut nodes.get_unchecked_mut(i).bytes);
 			}
 		}

ethash/src/lib.rs

@@ -16,10 +16,15 @@
 
 //! Ethash implementation
 //! See https://github.com/ethereum/wiki/wiki/Ethash
+
+#![cfg_attr(feature = "benches", feature(test))]
+
 extern crate primal;
 extern crate sha3;
 extern crate parking_lot;
 
+#[macro_use]
+extern crate crunchy;
 #[macro_use]
 extern crate log;
 mod compute;
@@ -128,3 +133,27 @@ fn test_lru() {
 	assert_eq!(ethash.cache.lock().recent_epoch.unwrap(), 2);
 	assert_eq!(ethash.cache.lock().prev_epoch.unwrap(), 0);
 }
+
+#[cfg(feature = "benches")]
+mod benchmarks {
+	extern crate test;
+
+	use compute::{Light, light_compute, SeedHashCompute};
+	use self::test::Bencher;
+
+	#[bench]
+	fn bench_light_compute(b: &mut Bencher) {
+		let hash = [0xf5, 0x7e, 0x6f, 0x3a, 0xcf, 0xc0, 0xdd, 0x4b, 0x5b, 0xf2, 0xbe, 0xe4, 0x0a, 0xb3, 0x35, 0x8a, 0xa6, 0x87, 0x73, 0xa8, 0xd0, 0x9f, 0x5e, 0x59, 0x5e, 0xab, 0x55, 0x94, 0x05, 0x52, 0x7d, 0x72];
+		let nonce = 0xd7b3ac70a301a249;
+		let light = Light::new(486382);
+
+		b.iter(|| light_compute(&light, &hash, nonce));
+	}
+
+	#[bench]
+	fn bench_seedhash(b: &mut Bencher) {
+		let seed_compute = SeedHashCompute::new();
+
+		b.iter(|| seed_compute.get_seedhash(486382));
+	}
+}