fzorb/REDACTED-rig
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

soft_aes: fix previous optimization

Browse source 
Previously removed unrolled variant is faster on some CPUs
Some CPUs are faster with added unrolled variant
The best variant depends on number of threads on some CPUs
This commit is contained in:
cohcho 2020-10-04 14:47:58 +00:00
parent 31e896feef
commit 5f0f2506e8
6 changed files with 116 additions and 72 deletions
Download patch file Download diff file Expand all files Collapse all files

110
src/crypto/randomx/aes_hash.cpp
View file

@ -26,8 +26,13 @@ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#include <thread>
#include <vector>
#include "crypto/randomx/aes_hash.hpp"
#include "crypto/randomx/soft_aes.h"
#include "crypto/randomx/randomx.h"
#include "base/tools/Chrono.h"
#include "base/tools/Profiler.h"
#define AES_HASH_1R_STATE0 0xd7983aad, 0xcc82db47, 0x9fa856de, 0x92b52c0d
@ -214,7 +219,7 @@ 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<int softAes>
template<int softAes, int unroll>
void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state) {
PROFILE_SCOPE(RandomX_AES);
@ -260,7 +265,7 @@ void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, voi
rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + k * 4 + 2, fill_state2); \
rx_store_vec_i128((rx_vec_i128*)scratchpadPtr + k * 4 + 3, fill_state3);
switch(softAes) {
switch (softAes) {
case 0:
HASH_STATE(0);
HASH_STATE(1);
@ -277,13 +282,51 @@ void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, voi
break;
default:
HASH_STATE(0);
FILL_STATE(0);
rx_prefetch_t0(prefetchPtr);
switch (unroll) {
case 4:
HASH_STATE(0);
FILL_STATE(0);
rx_prefetch_t0(prefetchPtr);
scratchpadPtr += 64;
prefetchPtr += 64;
HASH_STATE(1);
FILL_STATE(1);
rx_prefetch_t0(prefetchPtr + 64);
HASH_STATE(2);
FILL_STATE(2);
rx_prefetch_t0(prefetchPtr + 64 * 2);
HASH_STATE(3);
FILL_STATE(3);
rx_prefetch_t0(prefetchPtr + 64 * 3);
scratchpadPtr += 64 * 4;
prefetchPtr += 64 * 4;
break;
case 2:
HASH_STATE(0);
FILL_STATE(0);
rx_prefetch_t0(prefetchPtr);
HASH_STATE(1);
FILL_STATE(1);
rx_prefetch_t0(prefetchPtr + 64);
scratchpadPtr += 64 * 2;
prefetchPtr += 64 * 2;
break;
default:
HASH_STATE(0);
FILL_STATE(0);
rx_prefetch_t0(prefetchPtr);
scratchpadPtr += 64;
prefetchPtr += 64;
break;
}
break;
}
}
@ -317,6 +360,53 @@ void hashAndFillAes1Rx4(void *scratchpad, size_t scratchpadSize, void *hash, voi
rx_store_vec_i128((rx_vec_i128*)hash + 3, hash_state3);
}
template void hashAndFillAes1Rx4<0>(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);
template void hashAndFillAes1Rx4<1>(void *scratchpad, size_t scratchpadSize, void *hash, void* fill_state);
template void hashAndFillAes1Rx4<2>(void* scratchpad, size_t scratchpadSize, void* hash, void* fill_state);
template void hashAndFillAes1Rx4<0,2>(void* scratchpad, size_t scratchpadSize, void* hash, void* fill_state);
template void hashAndFillAes1Rx4<1,1>(void* scratchpad, size_t scratchpadSize, void* hash, void* fill_state);
template void hashAndFillAes1Rx4<2,1>(void* scratchpad, size_t scratchpadSize, void* hash, void* fill_state);
template void hashAndFillAes1Rx4<2,2>(void* scratchpad, size_t scratchpadSize, void* hash, void* fill_state);
template void hashAndFillAes1Rx4<2,4>(void* scratchpad, size_t scratchpadSize, void* hash, void* fill_state);
hashAndFillAes1Rx4_impl* softAESImpl = &hashAndFillAes1Rx4<1,1>;
void SelectSoftAESImpl(size_t threadsCount)
{
constexpr int test_length_ms = 100;
const std::vector<hashAndFillAes1Rx4_impl *> impl = {
&hashAndFillAes1Rx4<1,1>,
&hashAndFillAes1Rx4<2,1>,
&hashAndFillAes1Rx4<2,2>,
&hashAndFillAes1Rx4<2,4>,
};
size_t fast_idx = 0;
double fast_speed = 0.0;
for (size_t run = 0; run < 3; ++run) {
for (size_t i = 0; i < impl.size(); ++i) {
const uint64_t t1 = xmrig::Chrono::highResolutionMSecs();
std::vector<uint32_t> count(threadsCount, 0);
std::vector<std::thread> threads;
for (size_t t = 0; t < threadsCount; ++t) {
threads.emplace_back([&, t]() {
std::vector<uint8_t> scratchpad(10 * 1024);
uint8_t hash[64] = {};
uint8_t state[64] = {};
do {
(*impl[i])(scratchpad.data(), scratchpad.size(), hash, state);
++count[t];
} while (xmrig::Chrono::highResolutionMSecs() - t1 < test_length_ms);
});
}
uint32_t total = 0;
for (size_t t = 0; t < threadsCount; ++t) {
threads[t].join();
total += count[t];
}
const uint64_t t2 = xmrig::Chrono::highResolutionMSecs();
const double speed = total * 1e3 / (t2 - t1);
if (speed > fast_speed) {
fast_idx = i;
fast_speed = speed;
}
}
}
softAESImpl = impl[fast_idx];
}