github-wownero/RandomX
1
0
Fork 0
mirror of https://github.com/wownero/RandomX.git synced 2026-10-20 00:38:48 +02:00
Code Issues Packages Projects Releases Wiki Activity

Combined hash and fill AES loop (#166)

Browse Source 
Adds more parallelizm into AES loop so modern CPUs can take advantage of it. Also, scratchpad data moves between L1 and L3 caches only one time which saves time and energy per hash.
This commit is contained in:
SChernykh
2019-12-01 16:58:38 +01:00
committed by tevador
parent e3561d661e
commit 219c02e1e5
8 changed files with 140 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/aes_hash.cpp
+81
View File
@@ -239,3 +239,84 @@ void fillAes4Rx4(void *state, size_t outputSize, void *buffer) {
template void fillAes4Rx4<true>(void *state, size_t outputSize, void *buffer);
template void fillAes4Rx4<false>(void *state, size_t outputSize, void *buffer);
template<bool softAes>
void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state) {
uint8_t* scratchpadPtr = (uint8_t*)scratchpad;
const uint8_t* scratchpadEnd = scratchpadPtr + scratchpadSize;
// initial state
rx_vec_i128 hash_state0 = rx_set_int_vec_i128(AES_HASH_1R_STATE0);
rx_vec_i128 hash_state1 = rx_set_int_vec_i128(AES_HASH_1R_STATE1);
rx_vec_i128 hash_state2 = rx_set_int_vec_i128(AES_HASH_1R_STATE2);
rx_vec_i128 hash_state3 = rx_set_int_vec_i128(AES_HASH_1R_STATE3);
const rx_vec_i128 key0 = rx_set_int_vec_i128(AES_GEN_1R_KEY0);
const rx_vec_i128 key1 = rx_set_int_vec_i128(AES_GEN_1R_KEY1);
const rx_vec_i128 key2 = rx_set_int_vec_i128(AES_GEN_1R_KEY2);
const rx_vec_i128 key3 = rx_set_int_vec_i128(AES_GEN_1R_KEY3);
rx_vec_i128 fill_state0 = rx_load_vec_i128((rx_vec_i128*)fill_state + 0);
rx_vec_i128 fill_state1 = rx_load_vec_i128((rx_vec_i128*)fill_state + 1);
rx_vec_i128 fill_state2 = rx_load_vec_i128((rx_vec_i128*)fill_state + 2);
rx_vec_i128 fill_state3 = rx_load_vec_i128((rx_vec_i128*)fill_state + 3);
constexpr int PREFETCH_DISTANCE = 4096;
const char* prefetchPtr = ((const char*)scratchpad) + PREFETCH_DISTANCE;
scratchpadEnd -= PREFETCH_DISTANCE;
for (int i = 0; i < 2; ++i) {
//process 64 bytes at a time in 4 lanes
while (scratchpadPtr < scratchpadEnd) {
hash_state0 = aesenc<softAes>(hash_state0, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 0));
hash_state1 = aesdec<softAes>(hash_state1, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 1));
hash_state2 = aesenc<softAes>(hash_state2, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 2));
hash_state3 = aesdec<softAes>(hash_state3, rx_load_vec_i128((rx_vec_i128*)scratchpadPtr + 3));
fill_state0 = aesdec<softAes>(fill_state0, key0);
fill_state1 = aesenc<softAes>(fill_state1, key1);
fill_state2 = aesdec<softAes>(fill_state2, key2);
fill_state3 = aesenc<softAes>(fill_state3, key3);
rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 0, fill_state0);
rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 1, fill_state1);
rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 2, fill_state2);
rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + 3, fill_state3);
rx_prefetch_t0(prefetchPtr);
scratchpadPtr += 64;
prefetchPtr += 64;
}
prefetchPtr = (const char*) scratchpad;
scratchpadEnd += PREFETCH_DISTANCE;
}
rx_store_vec_i128((rx_vec_i128*)fill_state + 0, fill_state0);
rx_store_vec_i128((rx_vec_i128*)fill_state + 1, fill_state1);
rx_store_vec_i128((rx_vec_i128*)fill_state + 2, fill_state2);
rx_store_vec_i128((rx_vec_i128*)fill_state + 3, fill_state3);
//two extra rounds to achieve full diffusion
rx_vec_i128 xkey0 = rx_set_int_vec_i128(AES_HASH_1R_XKEY0);
rx_vec_i128 xkey1 = rx_set_int_vec_i128(AES_HASH_1R_XKEY1);
hash_state0 = aesenc<softAes>(hash_state0, xkey0);
hash_state1 = aesdec<softAes>(hash_state1, xkey0);
hash_state2 = aesenc<softAes>(hash_state2, xkey0);
hash_state3 = aesdec<softAes>(hash_state3, xkey0);
hash_state0 = aesenc<softAes>(hash_state0, xkey1);
hash_state1 = aesdec<softAes>(hash_state1, xkey1);
hash_state2 = aesenc<softAes>(hash_state2, xkey1);
hash_state3 = aesdec<softAes>(hash_state3, xkey1);
//output hash
rx_store_vec_i128((rx_vec_i128*)hash + 0, hash_state0);
rx_store_vec_i128((rx_vec_i128*)hash + 1, hash_state1);
rx_store_vec_i128((rx_vec_i128*)hash + 2, hash_state2);
rx_store_vec_i128((rx_vec_i128*)hash + 3, hash_state3);
}
template void hashAndFillAes1Rx4<false>(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);
template void hashAndFillAes1Rx4<true>(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);