Simplify Pow2Montgomery GPU executor branching
This commit is contained in:
1 parent
865a88fac0
commit
b08175dadf
1 file changed
+26
-23
@@ -135,43 +135,46 @@ public static partial class ULongExtensions
|
|||||||
{
|
{
|
||||||
ulong currentBit = (exponent >> index) & 1UL;
|
ulong currentBit = (exponent >> index) & 1UL;
|
||||||
ulong squared = MontgomeryMultiply(result, result, modulus, nPrime);
|
ulong squared = MontgomeryMultiply(result, result, modulus, nPrime);
|
||||||
result = currentBit == 0UL ? squared : result;
|
bool processWindow = currentBit != 0UL;
|
||||||
index -= (int)(currentBit ^ 1UL);
|
|
||||||
if (currentBit == 0UL)
|
|
||||||
{
|
|
||||||
continue;
|
|
||||||
}
|
|
||||||
|
|
||||||
int windowStart = index - windowSize + 1;
|
result = processWindow ? result : squared;
|
||||||
// if (windowStart < 0)
|
index -= (int)(currentBit ^ 1UL);
|
||||||
|
|
||||||
|
int windowStartCandidate = index - windowSize + 1;
|
||||||
|
// if (windowStartCandidate < 0)
|
||||||
// {
|
// {
|
||||||
// windowStart = 0;
|
// windowStartCandidate = 0;
|
||||||
// }
|
// }
|
||||||
// Clamp the negative offset without branching so the GPU loop stays divergence-free.
|
// Clamp the negative offset without branching so the GPU loop stays divergence-free.
|
||||||
// This still handles windows that would otherwise extend past the most significant bit.
|
// This still handles windows that would otherwise extend past the most significant bit.
|
||||||
int negativeMask = windowStart >> 31;
|
int negativeMask = windowStartCandidate >> 31;
|
||||||
windowStart &= ~negativeMask;
|
windowStartCandidate &= ~negativeMask;
|
||||||
|
|
||||||
windowStart = GetNextSetBitIndexGpu(exponent, windowStart);
|
int windowStart = processWindow ? GetNextSetBitIndexGpu(exponent, windowStartCandidate) : windowStartCandidate;
|
||||||
|
int windowLength = processWindow ? index - windowStart + 1 : 0;
|
||||||
int windowLength = index - windowStart + 1;
|
|
||||||
for (int square = 0; square < windowLength; square++)
|
for (int square = 0; square < windowLength; square++)
|
||||||
{
|
{
|
||||||
result = MontgomeryMultiply(result, result, modulus, nPrime);
|
result = MontgomeryMultiply(result, result, modulus, nPrime);
|
||||||
}
|
}
|
||||||
|
|
||||||
ulong mask = (1UL << windowLength) - 1UL;
|
result = processWindow
|
||||||
ulong windowValue = (exponent >> windowStart) & mask;
|
? MontgomeryMultiply(
|
||||||
ulong multiplier = ComputeMontgomeryOddPowerGpu(windowValue, divisor, modulus, nPrime);
|
result,
|
||||||
result = MontgomeryMultiply(result, multiplier, modulus, nPrime);
|
ComputeMontgomeryOddPowerGpu(
|
||||||
|
(exponent >> windowStart) & ((1UL << windowLength) - 1UL),
|
||||||
|
divisor,
|
||||||
|
modulus,
|
||||||
|
nPrime),
|
||||||
|
modulus,
|
||||||
|
nPrime)
|
||||||
|
: result;
|
||||||
|
|
||||||
index = windowStart - 1;
|
index = processWindow ? windowStart - 1 : index;
|
||||||
}
|
}
|
||||||
|
|
||||||
if (!keepMontgomery)
|
result = keepMontgomery
|
||||||
{
|
? result
|
||||||
result = MontgomeryMultiply(result, 1UL, modulus, nPrime);
|
: MontgomeryMultiply(result, 1UL, modulus, nPrime);
|
||||||
}
|
|
||||||
|
|
||||||
return result;
|
return result;
|
||||||
}
|
}
|
||||||
|
|||||||
Reference in new issue
Block a user