Optimize Keccak implementation and remove Crypto++ dependency

CMakeLists.txt

@@ -4,10 +4,6 @@ project(ethash)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_SOURCE_DIR}/cmake/modules/")
 set(ETHHASH_LIBS ethash)
 
-if (WIN32 AND WANT_CRYPTOPP)
-	add_subdirectory(cryptopp)
-endif()
-
 add_subdirectory(src/libethash)
 # bin2h.cmake doesn't work
 if (NOT OpenCL_FOUND)

MANIFEST.in

@@ -1,8 +1,8 @@
 include setup.py
 
-# C sources
+# C/C++ sources
 include src/libethash/internal.c
-include src/libethash/sha3.c
+include src/libethash/keccak.cpp
 include src/libethash/util.c
 include src/python/core.c
 
@@ -13,5 +13,5 @@ include src/libethash/endian.h
 include src/libethash/ethash.h
 include src/libethash/fnv.h
 include src/libethash/internal.h
-include src/libethash/sha3.h
+include src/libethash/keccak.h
 include src/libethash/util.h

setup.py

@@ -5,7 +5,7 @@
     'src/python/core.c',
     'src/libethash/io.c',
     'src/libethash/internal.c',
-    'src/libethash/sha3.c']
+    'src/libethash/keccak.cpp']
 if os.name == 'nt':
     sources += [
         'src/libethash/util.c',
@@ -25,13 +25,13 @@
     'src/libethash/io.h',
     'src/libethash/fnv.h',
     'src/libethash/internal.h',
-    'src/libethash/sha3.h',
+    'src/libethash/keccak.h',
     'src/libethash/util.h',
 ]
 pyethash = Extension('pyethash',
                      sources=sources,
                      depends=depends,
-                     extra_compile_args=["-Isrc/", "-std=gnu99", "-Wall"])
+                     extra_compile_args=["-Isrc/", "-std=gnu99", "-std=c++11", "-Wall"])
 
 setup(
     name='pyethash',

src/benchmark/benchmark.cpp

@@ -29,14 +29,7 @@
 #endif
 #include <vector>
 #include <algorithm>
-
-#ifdef WITH_CRYPTOPP
-#include <libethash/sha3_cryptopp.h>
-#include <string>
-
-#else
-#include "libethash/sha3.h"
-#endif // WITH_CRYPTOPP
+#include <libethash/keccak.h>
 
 #undef min
 #undef max
@@ -135,7 +128,7 @@ extern "C" int main(void)
 		auto time = std::chrono::duration_cast<std::chrono::milliseconds>(high_resolution_clock::now() - startTime).count();
 
 		ethash_h256_t cache_hash;
-		SHA3_256(&cache_hash, (uint8_t const*)cache_mem, params.cache_size);
+		keccak256(&cache_hash, (uint8_t const*)cache_mem, params.cache_size);
 		debugf("ethash_mkcache: %ums, sha3: %s\n", (unsigned)((time*1000)/CLOCKS_PER_SEC), bytesToHexString(&cache_hash, sizeof(cache_hash)).data());
 
 		// print a couple of test hashes

src/libethash/CMakeLists.txt

@@ -8,9 +8,11 @@ set(CMAKE_BUILD_TYPE Release)
 
 if (NOT MSVC)
 	set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} -std=gnu99")
+	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11")
 endif()
 
 set(FILES 	util.h
+			keccak.h keccak.cpp
           	io.c
           	internal.c
           	ethash.h
@@ -25,18 +27,6 @@ else()
 	list(APPEND FILES io_posix.c)
 endif()
 
-if (NOT CRYPTOPP_FOUND)
-	find_package(CryptoPP 5.6.2)
-endif()
-
-if (CRYPTOPP_FOUND)
-	add_definitions(-DWITH_CRYPTOPP)
-	include_directories( ${CRYPTOPP_INCLUDE_DIRS} )
-	list(APPEND FILES sha3_cryptopp.cpp sha3_cryptopp.h)
-else()
-	list(APPEND FILES sha3.c sha3.h)
-endif()
-
 add_library(${LIBRARY} ${FILES})
 
 if (CRYPTOPP_FOUND)

src/libethash/internal.c

@@ -32,14 +32,7 @@
 #include "internal.h"
 #include "data_sizes.h"
 #include "io.h"
-
-#ifdef WITH_CRYPTOPP
-
-#include "sha3_cryptopp.h"
-
-#else
-#include "sha3.h"
-#endif // WITH_CRYPTOPP
+#include "keccak.h"
 
 uint64_t ethash_get_datasize(uint32_t const block_number)
 {
@@ -67,10 +60,10 @@ bool static ethash_compute_cache_nodes(
 	}
 	uint32_t const num_nodes = (uint32_t) (params->cache_size / sizeof(node));
 
-	SHA3_512(nodes[0].bytes, (uint8_t*)seed, 32);
+	keccak512(nodes[0].bytes, (uint8_t*)seed, 32);
 
 	for (unsigned i = 1; i != num_nodes; ++i) {
-		SHA3_512(nodes[i].bytes, nodes[i - 1].bytes, 64);
+		keccak512(nodes[i].bytes, nodes[i - 1].bytes, 64);
 	}
 
 	for (unsigned j = 0; j != ETHASH_CACHE_ROUNDS; j++) {
@@ -81,7 +74,7 @@ bool static ethash_compute_cache_nodes(
 			for (unsigned w = 0; w != NODE_WORDS; ++w) {
 				data.words[w] ^= nodes[idx].words[w];
 			}
-			SHA3_512(nodes[i].bytes, data.bytes, sizeof(data));
+			keccak512(nodes[i].bytes, data.bytes, sizeof(data));
 		}
 	}
 
@@ -133,7 +126,7 @@ void ethash_calculate_dag_item(
 	node const* init = &cache_nodes[node_index % num_parent_nodes];
 	memcpy(ret, init, sizeof(node));
 	ret->words[0] ^= node_index;
-	SHA3_512(ret->bytes, ret->bytes, sizeof(node));
+	keccak512(ret->bytes, ret->bytes, sizeof(node));
 #if defined(_M_X64) && ENABLE_SSE
 	__m128i const fnv_prime = _mm_set1_epi32(FNV_PRIME);
 	__m128i xmm0 = ret->xmm[0];
@@ -171,7 +164,7 @@ void ethash_calculate_dag_item(
 		}
 #endif
 	}
-	SHA3_512(ret->bytes, ret->bytes, sizeof(node));
+	keccak512(ret->bytes, ret->bytes, sizeof(node));
 }
 
 bool ethash_compute_full_data(
@@ -213,7 +206,7 @@ static bool ethash_hash(
 	fix_endian64(s_mix[0].double_words[4], nonce);
 
 	// compute sha3-512 hash and replicate across mix
-	SHA3_512(s_mix->bytes, s_mix->bytes, 40);
+	keccak512(s_mix->bytes, s_mix->bytes, 40);
 	fix_endian_arr32(s_mix[0].words, 16);
 
 	node* const mix = s_mix + 1;
@@ -279,7 +272,7 @@ static bool ethash_hash(
 	fix_endian_arr32(mix->words, MIX_WORDS / 4);
 	memcpy(&ret->mix_hash, mix->bytes, 32);
 	// final Keccak hash
-	SHA3_256(&ret->result, s_mix->bytes, 64 + 32); // Keccak-256(s + compressed_mix)
+	keccak256(&ret->result, s_mix->bytes, 64 + 32); // Keccak-256(s + compressed_mix)
 	return true;
 }
 
@@ -294,17 +287,17 @@ void ethash_quick_hash(
 	memcpy(buf, header_hash, 32);
 	fix_endian64_same(nonce);
 	memcpy(&(buf[32]), &nonce, 8);
-	SHA3_512(buf, buf, 40);
+	keccak512(buf, buf, 40);
 	memcpy(&(buf[64]), mix_hash, 32);
-	SHA3_256(return_hash, buf, 64 + 32);
+	keccak256(return_hash, buf, 64 + 32);
 }
 
 void ethash_get_seedhash(ethash_h256_t* seedhash, const uint32_t block_number)
 {
 	ethash_h256_reset(seedhash);
 	const uint32_t epochs = block_number / ETHASH_EPOCH_LENGTH;
 	for (uint32_t i = 0; i < epochs; ++i)
-		SHA3_256(seedhash, (uint8_t*)seedhash, 32);
+		keccak256(seedhash, (uint8_t*)seedhash, 32);
 }
 
 int ethash_quick_check_difficulty(

src/libethash/keccak.cpp

@@ -0,0 +1,121 @@
+
+/*
+  Core Keccack implementation devided from https://github.com/coruus/keccak-tiny by David Leon Gil
+*/
+
+#include "keccak.h"
+
+#include <cstring>
+
+namespace
+{
+    uint8_t  const rho[] = {  1,  3,  6, 10, 15, 21, 28, 36, 45, 55,  2, 14,
+                             27, 41, 56,  8, 25, 43, 62, 18, 39, 61, 20, 44 };
+    uint8_t  const pi[]  = { 10,  7, 11, 17, 18,  3,  5, 16,  8, 21, 24,  4,
+                             15, 23, 19, 13, 12,  2, 20, 14, 22,  9,  6,  1 };
+    uint64_t const RC[]  = { 1ULL, 0x8082ULL, 0x800000000000808aULL, 0x8000000080008000ULL,
+                             0x808bULL, 0x80000001ULL, 0x8000000080008081ULL, 0x8000000000008009ULL,
+                             0x8aULL, 0x88ULL, 0x80008009ULL, 0x8000000aULL,
+                             0x8000808bULL, 0x800000000000008bULL, 0x8000000000008089ULL, 0x8000000000008003ULL,
+                             0x8000000000008002ULL, 0x8000000000000080ULL, 0x800aULL, 0x800000008000000aULL,
+                             0x8000000080008081ULL, 0x8000000000008080ULL, 0x80000001ULL, 0x8000000080008008ULL};
+
+    inline uint64_t rot(uint64_t x, uint64_t s) { return x << s | x >> (64 - s); }
+
+	#define REPEAT6(e)  e e e e e e
+	#define REPEAT24(e) REPEAT6(e e e e)
+	#define REPEAT5(e)  e e e e e
+	#define FOR5(v, s, e) v = 0; REPEAT5(e; v += s;)
+
+    /// Keccak-f[1600]
+    inline void keccakf1600(uint64_t* a)
+    {
+        uint64_t b[5] = {0};
+        uint64_t t = 0;
+        uint8_t x, y;
+
+        for (size_t i = 0; i < 24; ++i)
+        {
+            // Round[1600](a,RC):
+            //  theta step:
+            FOR5(x, 1,
+                 b[x] = 0;
+                         FOR5(y, 5,
+                              b[x] ^= a[x + y]; ))
+            FOR5(x, 1,
+                 FOR5(y, 5,
+                      a[y + x] ^= b[(x + 4) % 5] ^ rot(b[(x + 1) % 5], 1); ))
+
+            //  rho and pi steps
+            t = a[1];
+            x = 0;
+            REPEAT24(b[0] = a[pi[x]];
+                             a[pi[x]] = rot(t, rho[x]);
+                             t = b[0];
+                             x++; )
+            // chi step
+            FOR5(y,
+                 5,
+                 FOR5(x, 1,
+                      b[x] = a[y + x];)
+                         FOR5(x, 1,
+                         a[y + x] = b[x] ^ ((~b[(x + 1) % 5]) & b[(x + 2) % 5]); ))
+            // iota step
+            a[0] ^= RC[i];
+        }
+    }
+
+    inline void xorin(uint64_t* dst, const uint64_t* src, size_t size)
+    {
+        for (size_t i = 0; i < size; ++i)
+            dst[i] ^= src[i];
+    }
+
+    inline void keccak(uint64_t* out, size_t outSize, const uint64_t* data, size_t size)
+    {
+        static const size_t stateSize = 200;
+        const auto r = stateSize - (2 * outSize); // 256: 136, 512: 72
+        uint64_t a[stateSize / 8] = {0};
+        // Absorb input.
+        while (size >= r)
+        {
+            xorin(a, data, r / 8);
+            keccakf1600(a);
+            data += r / 8;
+            size -= r;
+        }
+        // Xor in the DS and pad frame.
+        a[size / 8] ^= 0x01; // 0x01: Keccak, 0x06: SHA3
+        a[(r - 1) / 8] ^= 0x8000000000000000;
+        // Xor in the last block.
+        xorin(a, data, size / 8);
+        keccakf1600(a);
+        std::memcpy(out, a, outSize); // TODO: How about using out as a state
+    }
+}
+
+extern "C"
+{
+
+void keccak256(ethash_h256_t* out, uint8_t const* data, size_t size)
+{
+	// FIXME: What with unaligned memory?
+	keccak((uint64_t*)out, 32, (uint64_t*)data, size);
+}
+
+void keccak512(uint8_t* out, uint8_t const* data, size_t size)
+{
+	keccak((uint64_t*)out, 64, (uint64_t*)data, size);
+}
+
+void keccak256_96(ethash_h256_t* out, uint8_t const* data)
+{
+	keccak((uint64_t*)out, 32, (uint64_t*)data, 96 / 8);
+}
+
+void keccak512_40(uint8_t* out, uint8_t const* data)
+{
+	keccak((uint64_t*)out, 64, (uint64_t*)data, 40 / 8);
+}
+
+} // extern "C"

src/libethash/keccak.h

@@ -0,0 +1,23 @@
+#pragma once
+
+#include <stdint.h>
+#include <stddef.h>
+
+struct ethash_h256;
+typedef struct ethash_h256 ethash_h256_t;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void keccak256(ethash_h256_t* out, uint8_t const* data, size_t size);
+
+void keccak512(uint8_t* out, uint8_t const* data, size_t size);
+
+void keccak256_96(ethash_h256_t* out, uint8_t const* data);
+
+void keccak512_40(uint8_t* out, uint8_t const* data);
+
+#ifdef __cplusplus
+}
+#endif