Interpreter: make fres match hardware.

New table-based implementation written based on actual hardware behavior.
(hwtest coming soon).
This commit is contained in:
magumagu 2014-05-22 01:06:30 -07:00
parent ad4ad7c1ed
commit 2f8a147eda
3 changed files with 61 additions and 29 deletions

View File

@ -271,3 +271,61 @@ inline u64 ConvertToDouble(u32 _x)
return ((x & 0xc0000000) << 32) | z | ((x & 0x3fffffff) << 29);
}
}
// Used by fres and ps_res.
inline double ApproximateReciprocal(double val)
{
static const int expected_base[] = {
0x7ff800, 0x783800, 0x70ea00, 0x6a0800,
0x638800, 0x5d6200, 0x579000, 0x520800,
0x4cc800, 0x47ca00, 0x430800, 0x3e8000,
0x3a2c00, 0x360800, 0x321400, 0x2e4a00,
0x2aa800, 0x272c00, 0x23d600, 0x209e00,
0x1d8800, 0x1a9000, 0x17ae00, 0x14f800,
0x124400, 0x0fbe00, 0x0d3800, 0x0ade00,
0x088400, 0x065000, 0x041c00, 0x020c00,
};
static const int expected_dec[] = {
0x3e1, 0x3a7, 0x371, 0x340,
0x313, 0x2ea, 0x2c4, 0x2a0,
0x27f, 0x261, 0x245, 0x22a,
0x212, 0x1fb, 0x1e5, 0x1d1,
0x1be, 0x1ac, 0x19b, 0x18b,
0x17c, 0x16e, 0x15b, 0x15b,
0x143, 0x143, 0x12d, 0x12d,
0x11a, 0x11a, 0x108, 0x106,
};
union {
double valf;
long long vali;
};
valf = val;
long long mantissa = vali & ((1LL << 52) - 1);
long long sign = vali & (1ULL << 63);
long long exponent = vali & (0x7FFLL << 52);
// Special case 0
if (mantissa == 0 && exponent == 0)
return sign ? -INFINITY : INFINITY;
// Special case NaN-ish numbers
if (exponent == (0x7FFLL << 52))
{
if (mantissa == 0)
return sign ? -0.0 : 0.0;
return 0.0 + valf;
}
// Special case small inputs
if (exponent < (895LL << 52))
return sign ? -FLT_MAX : FLT_MAX;
// Special case large inputs
if (exponent >= (1149LL << 52))
return sign ? -0.0f : 0.0f;
exponent = (0x7FDLL << 52) - exponent;
int i = (int)(mantissa >> 37);
vali = sign | exponent;
vali |= (long long)(expected_base[i / 1024] - (expected_dec[i / 1024] * (i % 1024) + 1) / 2) << 29;
return valf;
}

View File

@ -388,19 +388,7 @@ void Interpreter::fdivsx(UGeckoInstruction _inst)
void Interpreter::fresx(UGeckoInstruction _inst)
{
double b = rPS0(_inst.FB);
double one_over = ForceSingle(1.0 / b);
// this is based on the real hardware tests
if (b != 0.0 && IsINF(one_over))
{
if (one_over > 0)
riPS0(_inst.FD) = riPS1(_inst.FD) = MAX_SINGLE;
else
riPS0(_inst.FD) = riPS1(_inst.FD) = MIN_SINGLE;
}
else
{
rPS0(_inst.FD) = rPS1(_inst.FD) = one_over;
}
rPS0(_inst.FD) = rPS1(_inst.FD) = ApproximateReciprocal(b);
if (b == 0.0)
{
SetFPException(FPSCR_ZX);

View File

@ -178,22 +178,8 @@ void Interpreter::ps_res(UGeckoInstruction _inst)
{
SetFPException(FPSCR_ZX);
}
rPS0(_inst.FD) = ForceSingle(1.0 / a);
if (a != 0.0 && IsINF(rPS0(_inst.FD)))
{
if (rPS0(_inst.FD) > 0)
riPS0(_inst.FD) = MAX_SINGLE; // largest finite single
else
riPS0(_inst.FD) = MIN_SINGLE; // most negative finite single
}
rPS1(_inst.FD) = ForceSingle(1.0 / b);
if (b != 0.0 && IsINF(rPS1(_inst.FD)))
{
if (rPS1(_inst.FD) > 0)
riPS1(_inst.FD) = MAX_SINGLE;
else
riPS1(_inst.FD) = MIN_SINGLE;
}
rPS0(_inst.FD) = ApproximateReciprocal(a);
rPS1(_inst.FD) = ApproximateReciprocal(b);
UpdateFPRF(rPS0(_inst.FD));
if (_inst.Rc) Helper_UpdateCR1();
}