dynarmic/src/common/fp/op/FPRSqrtEstimate.cpp
MerryMage 7a673a8a43 fp: Change FPUnpacked to a normalized representation
Having a known position for the highest set bit makes writing algorithms easier
2020-04-22 20:46:22 +01:00

96 lines
3.1 KiB
C++

/* This file is part of the dynarmic project.
* Copyright (c) 2018 MerryMage
* This software may be used and distributed according to the terms of the GNU
* General Public License version 2 or any later version.
*/
#include <array>
#include "common/assert.h"
#include "common/bit_util.h"
#include "common/common_types.h"
#include "common/fp/fpcr.h"
#include "common/fp/fpsr.h"
#include "common/fp/info.h"
#include "common/fp/op/FPRSqrtEstimate.h"
#include "common/fp/process_exception.h"
#include "common/fp/process_nan.h"
#include "common/fp/unpacked.h"
#include "common/safe_ops.h"
namespace Dynarmic::FP {
/// Input is a u0.9 fixed point number. Only values in [0.25, 1.0) are valid.
/// Output is a u0.8 fixed point number, with an implied 1 prefixed.
/// i.e.: The output is a value in [1.0, 2.0).
static u8 RecipSqrtEstimate(u64 a) {
using LUT = std::array<u8, 512>;
static const LUT lut = [] {
LUT result{};
for (u64 i = 128; i < result.size(); i++) {
u64 a = i;
// Convert to u.10 (with 8 significant bits), force to odd
if (a < 256) {
// [0.25, 0.5)
a = a * 2 + 1;
} else {
// [0.5, 1.0)
a = (a | 1) * 2;
}
// Calculate largest b which for which b < 1.0 / sqrt(a).
// Start from b = 1.0 (in u.9) since b cannot be smaller.
u64 b = 512;
// u.10 * u.9 * u.9 -> u.28
while (a * (b + 1) * (b + 1) < (1u << 28)) {
b++;
}
// Round to nearest u0.8 (with implied set integer bit).
result[i] = static_cast<u8>((b + 1) / 2);
}
return result;
}();
return lut[a & 0x1FF];
}
template<typename FPT>
FPT FPRSqrtEstimate(FPT op, FPCR fpcr, FPSR& fpsr) {
auto [type, sign, value] = FPUnpack<FPT>(op, fpcr, fpsr);
if (type == FPType::SNaN || type == FPType::QNaN) {
return FPProcessNaN(type, op, fpcr, fpsr);
}
if (type == FPType::Zero) {
FPProcessException(FPExc::DivideByZero, fpcr, fpsr);
return FPInfo<FPT>::Infinity(sign);
}
if (sign) {
FPProcessException(FPExc::InvalidOp, fpcr, fpsr);
return FPInfo<FPT>::DefaultNaN();
}
if (type == FPType::Infinity) {
return FPInfo<FPT>::Zero(false);
}
const int result_exponent = (-(value.exponent + 1)) >> 1;
const bool was_exponent_odd = (value.exponent) % 2 == 0;
const u64 scaled = Safe::LogicalShiftRight(value.mantissa, normalized_point_position - (was_exponent_odd ? 7 : 8));
const u64 estimate = RecipSqrtEstimate(scaled);
const FPT bits_exponent = static_cast<FPT>(result_exponent + FPInfo<FPT>::exponent_bias);
const FPT bits_mantissa = static_cast<FPT>(estimate << (FPInfo<FPT>::explicit_mantissa_width - 8));
return (bits_exponent << FPInfo<FPT>::explicit_mantissa_width) | (bits_mantissa & FPInfo<FPT>::mantissa_mask);
}
template u32 FPRSqrtEstimate<u32>(u32 op, FPCR fpcr, FPSR& fpsr);
template u64 FPRSqrtEstimate<u64>(u64 op, FPCR fpcr, FPSR& fpsr);
} // namespace Dynarmic::FP