Azahar-Mirror/dynarmic
1
0
Fork 0
mirror of https://github.com/azahar-emu/dynarmic synced 2025-11-11 17:39:59 +01:00
Code Issues Packages Projects Releases Wiki Activity

common/fp/op: Add half-precision specialization for FPRecipStepFused

Browse Source
This commit is contained in:
Lioncash 2019-04-13 18:47:37 -04:00 committed by MerryMage
parent c6e6ec0e69
commit 68d8cd2b13
2 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
src/common/fp/info.h
View File

@ -88,7 +88,7 @@ struct FPInfo<u64> {
template<typename FPT, bool sign, int exponent, FPT value> template<typename FPT, bool sign, int exponent, FPT value>
constexpr FPT FPValue() { constexpr FPT FPValue() {
if constexpr (value == 0) { if constexpr (value == 0) {
return FPInfo<FPT>::Zero(sign); return FPT(FPInfo<FPT>::Zero(sign));
} }
constexpr int point_position = static_cast<int>(FPInfo<FPT>::explicit_mantissa_width); constexpr int point_position = static_cast<int>(FPInfo<FPT>::explicit_mantissa_width);
@ -100,7 +100,7 @@ constexpr FPT FPValue() {
constexpr FPT mantissa = (value << offset) & FPInfo<FPT>::mantissa_mask; constexpr FPT mantissa = (value << offset) & FPInfo<FPT>::mantissa_mask;
constexpr FPT biased_exponent = static_cast<FPT>(normalized_exponent + FPInfo<FPT>::exponent_bias); constexpr FPT biased_exponent = static_cast<FPT>(normalized_exponent + FPInfo<FPT>::exponent_bias);
return FPInfo<FPT>::Zero(sign) | mantissa | (biased_exponent << FPInfo<FPT>::explicit_mantissa_width); return FPT(FPInfo<FPT>::Zero(sign) | mantissa | (biased_exponent << FPInfo<FPT>::explicit_mantissa_width));
} }
} // namespace Dynarmic::FP } // namespace Dynarmic::FP

7
src/common/fp/op/FPRecipStepFused.cpp
View File

@ -37,18 +37,19 @@ FPT FPRecipStepFused(FPT op1, FPT op2, FPCR fpcr, FPSR& fpsr) {
} }
if (inf1 || inf2) { if (inf1 || inf2) {
return FPInfo<FPT>::Infinity(sign1 != sign2); return FPT(FPInfo<FPT>::Infinity(sign1 != sign2));
} }
// result_value = 2.0 + (value1 * value2) // result_value = 2.0 + (value1 * value2)
FPUnpacked result_value = FusedMulAdd(ToNormalized(false, 0, 2), value1, value2); const FPUnpacked result_value = FusedMulAdd(ToNormalized(false, 0, 2), value1, value2);
if (result_value.mantissa == 0) { if (result_value.mantissa == 0) {
return FPInfo<FPT>::Zero(fpcr.RMode() == RoundingMode::TowardsMinusInfinity); return FPT(FPInfo<FPT>::Zero(fpcr.RMode() == RoundingMode::TowardsMinusInfinity));
} }
return FPRound<FPT>(result_value, fpcr, fpsr); return FPRound<FPT>(result_value, fpcr, fpsr);
} }
template u16 FPRecipStepFused<u16>(u16 op1, u16 op2, FPCR fpcr, FPSR& fpsr);
template u32 FPRecipStepFused<u32>(u32 op1, u32 op2, FPCR fpcr, FPSR& fpsr); template u32 FPRecipStepFused<u32>(u32 op1, u32 op2, FPCR fpcr, FPSR& fpsr);
template u64 FPRecipStepFused<u64>(u64 op1, u64 op2, FPCR fpcr, FPSR& fpsr); template u64 FPRecipStepFused<u64>(u64 op1, u64 op2, FPCR fpcr, FPSR& fpsr);