mirror of
https://github.com/monero-project/monero.git
synced 2024-12-23 12:09:54 +00:00
multiexp: cut down on memory allocations
This commit is contained in:
parent
1b867e7f40
commit
1ed0ed4de4
1 changed files with 13 additions and 15 deletions
|
@ -394,37 +394,35 @@ rct::key straus(const std::vector<MultiexpData> &data, const std::shared_ptr<str
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
MULTIEXP_PERF(PERF_TIMER_START_UNIT(digits, 1000000));
|
MULTIEXP_PERF(PERF_TIMER_START_UNIT(digits, 1000000));
|
||||||
std::vector<std::vector<uint8_t>> digits;
|
std::unique_ptr<uint8_t[]> digits{new uint8_t[256 * data.size()]};
|
||||||
digits.resize(data.size());
|
|
||||||
for (size_t j = 0; j < data.size(); ++j)
|
for (size_t j = 0; j < data.size(); ++j)
|
||||||
{
|
{
|
||||||
digits[j].resize(256);
|
|
||||||
unsigned char bytes33[33];
|
unsigned char bytes33[33];
|
||||||
memcpy(bytes33, data[j].scalar.bytes, 32);
|
memcpy(bytes33, data[j].scalar.bytes, 32);
|
||||||
bytes33[32] = 0;
|
bytes33[32] = 0;
|
||||||
|
const unsigned char *bytes = bytes33;
|
||||||
#if 1
|
#if 1
|
||||||
static_assert(STRAUS_C == 4, "optimized version needs STRAUS_C == 4");
|
static_assert(STRAUS_C == 4, "optimized version needs STRAUS_C == 4");
|
||||||
const unsigned char *bytes = bytes33;
|
|
||||||
unsigned int i;
|
unsigned int i;
|
||||||
for (i = 0; i < 256; i += 8, bytes++)
|
for (i = 0; i < 256; i += 8, bytes++)
|
||||||
{
|
{
|
||||||
digits[j][i] = bytes[0] & 0xf;
|
digits[j*256+i] = bytes[0] & 0xf;
|
||||||
digits[j][i+1] = (bytes[0] >> 1) & 0xf;
|
digits[j*256+i+1] = (bytes[0] >> 1) & 0xf;
|
||||||
digits[j][i+2] = (bytes[0] >> 2) & 0xf;
|
digits[j*256+i+2] = (bytes[0] >> 2) & 0xf;
|
||||||
digits[j][i+3] = (bytes[0] >> 3) & 0xf;
|
digits[j*256+i+3] = (bytes[0] >> 3) & 0xf;
|
||||||
digits[j][i+4] = ((bytes[0] >> 4) | (bytes[1]<<4)) & 0xf;
|
digits[j*256+i+4] = ((bytes[0] >> 4) | (bytes[1]<<4)) & 0xf;
|
||||||
digits[j][i+5] = ((bytes[0] >> 5) | (bytes[1]<<3)) & 0xf;
|
digits[j*256+i+5] = ((bytes[0] >> 5) | (bytes[1]<<3)) & 0xf;
|
||||||
digits[j][i+6] = ((bytes[0] >> 6) | (bytes[1]<<2)) & 0xf;
|
digits[j*256+i+6] = ((bytes[0] >> 6) | (bytes[1]<<2)) & 0xf;
|
||||||
digits[j][i+7] = ((bytes[0] >> 7) | (bytes[1]<<1)) & 0xf;
|
digits[j*256+i+7] = ((bytes[0] >> 7) | (bytes[1]<<1)) & 0xf;
|
||||||
}
|
}
|
||||||
#elif 1
|
#elif 1
|
||||||
for (size_t i = 0; i < 256; ++i)
|
for (size_t i = 0; i < 256; ++i)
|
||||||
digits[j][i] = ((bytes[i>>3] | (bytes[(i>>3)+1]<<8)) >> (i&7)) & mask;
|
digits[j*256+i] = ((bytes[i>>3] | (bytes[(i>>3)+1]<<8)) >> (i&7)) & mask;
|
||||||
#else
|
#else
|
||||||
rct::key shifted = data[j].scalar;
|
rct::key shifted = data[j].scalar;
|
||||||
for (size_t i = 0; i < 256; ++i)
|
for (size_t i = 0; i < 256; ++i)
|
||||||
{
|
{
|
||||||
digits[j][i] = shifted.bytes[0] & 0xf;
|
digits[j*256+i] = shifted.bytes[0] & 0xf;
|
||||||
shifted = div2(shifted, (256-i)>>3);
|
shifted = div2(shifted, (256-i)>>3);
|
||||||
}
|
}
|
||||||
#endif
|
#endif
|
||||||
|
@ -470,7 +468,7 @@ skipfirst:
|
||||||
if (skip[j])
|
if (skip[j])
|
||||||
continue;
|
continue;
|
||||||
#endif
|
#endif
|
||||||
const uint8_t digit = digits[j][i];
|
const uint8_t digit = digits[j*256+i];
|
||||||
if (digit)
|
if (digit)
|
||||||
{
|
{
|
||||||
ge_add(&p1, &band_p3, &CACHE_OFFSET(local_cache, j, digit));
|
ge_add(&p1, &band_p3, &CACHE_OFFSET(local_cache, j, digit));
|
||||||
|
|
Loading…
Reference in a new issue