More accurate impl of FMINNM/FMAXNM, add vector variants (#296)
* More accurate impl of FMINNM/FMAXNM, add vector variants * Optimize for the 0 case when op1 != op2 * Address PR feedback
This commit is contained in:
gdkchan
GitHub
parent
eeb626947e
commit
221270db90
5 changed files with 336 additions and 174 deletions
|
|
@ -79,7 +79,7 @@ namespace ChocolArm64.Instruction
|
|||
if (scaled == 0)
|
||||
{
|
||||
// Zero -> Infinity
|
||||
return BitConverter.Int64BitsToDouble((long)(x_sign | 0x7ff0000000000000));
|
||||
return BitConverter.Int64BitsToDouble((long)(x_sign | 0x7FF0000000000000));
|
||||
}
|
||||
|
||||
// Denormal
|
||||
|
|
@ -94,7 +94,7 @@ namespace ChocolArm64.Instruction
|
|||
if (x_sign != 0)
|
||||
{
|
||||
// Negative -> NaN
|
||||
return BitConverter.Int64BitsToDouble((long)0x7ff8000000000000);
|
||||
return BitConverter.Int64BitsToDouble((long)0x7FF8000000000000);
|
||||
}
|
||||
|
||||
if (x_exp == 0x7ff && scaled == 0)
|
||||
|
|
@ -153,7 +153,7 @@ namespace ChocolArm64.Instruction
|
|||
if (scaled == 0)
|
||||
{
|
||||
// Zero -> Infinity
|
||||
return BitConverter.Int64BitsToDouble((long)(x_sign | 0x7ff0000000000000));
|
||||
return BitConverter.Int64BitsToDouble((long)(x_sign | 0x7FF0000000000000));
|
||||
}
|
||||
|
||||
// Denormal
|
||||
|
|
@ -208,8 +208,8 @@ namespace ChocolArm64.Instruction
|
|||
ulong op1_other = op1_bits & 0x7FFFFFFFFFFFFFFF;
|
||||
ulong op2_other = op2_bits & 0x7FFFFFFFFFFFFFFF;
|
||||
|
||||
bool inf1 = op1_other == 0x7ff0000000000000;
|
||||
bool inf2 = op2_other == 0x7ff0000000000000;
|
||||
bool inf1 = op1_other == 0x7FF0000000000000;
|
||||
bool inf2 = op2_other == 0x7FF0000000000000;
|
||||
bool zero1 = op1_other == 0;
|
||||
bool zero2 = op2_other == 0;
|
||||
|
||||
|
|
@ -220,7 +220,7 @@ namespace ChocolArm64.Instruction
|
|||
else if (inf1 || inf2)
|
||||
{
|
||||
// Infinity
|
||||
return BitConverter.Int64BitsToDouble((long)(0x7ff0000000000000 | (op1_sign ^ op2_sign)));
|
||||
return BitConverter.Int64BitsToDouble((long)(0x7FF0000000000000 | (op1_sign ^ op2_sign)));
|
||||
}
|
||||
|
||||
return 2.0 + op1 * op2;
|
||||
|
|
@ -261,5 +261,277 @@ namespace ChocolArm64.Instruction
|
|||
uint new_exp = (uint)((exponent + 127) & 0xFF) << 23;
|
||||
return BitConverter.Int32BitsToSingle((int)((x_sign << 31) | new_exp | (x_mantissa << 13)));
|
||||
}
|
||||
|
||||
public static float MaxNum(float op1, float op2)
|
||||
{
|
||||
uint op1_bits = (uint)BitConverter.SingleToInt32Bits(op1);
|
||||
uint op2_bits = (uint)BitConverter.SingleToInt32Bits(op2);
|
||||
|
||||
if (IsQNaN(op1_bits) && !IsQNaN(op2_bits))
|
||||
{
|
||||
op1 = float.NegativeInfinity;
|
||||
}
|
||||
else if (!IsQNaN(op1_bits) && IsQNaN(op2_bits))
|
||||
{
|
||||
op2 = float.NegativeInfinity;
|
||||
}
|
||||
|
||||
return Max(op1, op2);
|
||||
}
|
||||
|
||||
public static double MaxNum(double op1, double op2)
|
||||
{
|
||||
ulong op1_bits = (ulong)BitConverter.DoubleToInt64Bits(op1);
|
||||
ulong op2_bits = (ulong)BitConverter.DoubleToInt64Bits(op2);
|
||||
|
||||
if (IsQNaN(op1_bits) && !IsQNaN(op2_bits))
|
||||
{
|
||||
op1 = double.NegativeInfinity;
|
||||
}
|
||||
else if (!IsQNaN(op1_bits) && IsQNaN(op2_bits))
|
||||
{
|
||||
op2 = double.NegativeInfinity;
|
||||
}
|
||||
|
||||
return Max(op1, op2);
|
||||
}
|
||||
|
||||
public static float Max(float op1, float op2)
|
||||
{
|
||||
// Fast path
|
||||
if (op1 > op2)
|
||||
{
|
||||
return op1;
|
||||
}
|
||||
|
||||
if (op1 < op2 || (op1 == op2 && op2 != 0))
|
||||
{
|
||||
return op2;
|
||||
}
|
||||
|
||||
uint op1_bits = (uint)BitConverter.SingleToInt32Bits(op1);
|
||||
uint op2_bits = (uint)BitConverter.SingleToInt32Bits(op2);
|
||||
|
||||
// Handle NaN cases
|
||||
if (ProcessNaNs(op1_bits, op2_bits, out uint op_bits))
|
||||
{
|
||||
return BitConverter.Int32BitsToSingle((int)op_bits);
|
||||
}
|
||||
|
||||
// Return the most positive zero
|
||||
if ((op1_bits & op2_bits) == 0x80000000u)
|
||||
{
|
||||
return BitConverter.Int32BitsToSingle(int.MinValue);
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
public static double Max(double op1, double op2)
|
||||
{
|
||||
// Fast path
|
||||
if (op1 > op2)
|
||||
{
|
||||
return op1;
|
||||
}
|
||||
|
||||
if (op1 < op2 || (op1 == op2 && op2 != 0))
|
||||
{
|
||||
return op2;
|
||||
}
|
||||
|
||||
ulong op1_bits = (ulong)BitConverter.DoubleToInt64Bits(op1);
|
||||
ulong op2_bits = (ulong)BitConverter.DoubleToInt64Bits(op2);
|
||||
|
||||
// Handle NaN cases
|
||||
if (ProcessNaNs(op1_bits, op2_bits, out ulong op_bits))
|
||||
{
|
||||
return BitConverter.Int64BitsToDouble((long)op_bits);
|
||||
}
|
||||
|
||||
// Return the most positive zero
|
||||
if ((op1_bits & op2_bits) == 0x8000000000000000ul)
|
||||
{
|
||||
return BitConverter.Int64BitsToDouble(long.MinValue);
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
public static float MinNum(float op1, float op2)
|
||||
{
|
||||
uint op1_bits = (uint)BitConverter.SingleToInt32Bits(op1);
|
||||
uint op2_bits = (uint)BitConverter.SingleToInt32Bits(op2);
|
||||
|
||||
if (IsQNaN(op1_bits) && !IsQNaN(op2_bits))
|
||||
{
|
||||
op1 = float.PositiveInfinity;
|
||||
}
|
||||
else if (!IsQNaN(op1_bits) && IsQNaN(op2_bits))
|
||||
{
|
||||
op2 = float.PositiveInfinity;
|
||||
}
|
||||
|
||||
return Max(op1, op2);
|
||||
}
|
||||
|
||||
public static double MinNum(double op1, double op2)
|
||||
{
|
||||
ulong op1_bits = (ulong)BitConverter.DoubleToInt64Bits(op1);
|
||||
ulong op2_bits = (ulong)BitConverter.DoubleToInt64Bits(op2);
|
||||
|
||||
if (IsQNaN(op1_bits) && !IsQNaN(op2_bits))
|
||||
{
|
||||
op1 = double.PositiveInfinity;
|
||||
}
|
||||
else if (!IsQNaN(op1_bits) && IsQNaN(op2_bits))
|
||||
{
|
||||
op2 = double.PositiveInfinity;
|
||||
}
|
||||
|
||||
return Min(op1, op2);
|
||||
}
|
||||
|
||||
public static float Min(float op1, float op2)
|
||||
{
|
||||
// Fast path
|
||||
if (op1 < op2)
|
||||
{
|
||||
return op1;
|
||||
}
|
||||
|
||||
if (op1 > op2 || (op1 == op2 && op2 != 0))
|
||||
{
|
||||
return op2;
|
||||
}
|
||||
|
||||
uint op1_bits = (uint)BitConverter.SingleToInt32Bits(op1);
|
||||
uint op2_bits = (uint)BitConverter.SingleToInt32Bits(op2);
|
||||
|
||||
// Handle NaN cases
|
||||
if (ProcessNaNs(op1_bits, op2_bits, out uint op_bits))
|
||||
{
|
||||
return BitConverter.Int32BitsToSingle((int)op_bits);
|
||||
}
|
||||
|
||||
// Return the most negative zero
|
||||
if ((op1_bits | op2_bits) == 0x80000000u)
|
||||
{
|
||||
return BitConverter.Int32BitsToSingle(int.MinValue);
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
public static double Min(double op1, double op2)
|
||||
{
|
||||
// Fast path
|
||||
if (op1 < op2)
|
||||
{
|
||||
return op1;
|
||||
}
|
||||
|
||||
if (op1 > op2 || (op1 == op2 && op2 != 0))
|
||||
{
|
||||
return op2;
|
||||
}
|
||||
|
||||
ulong op1_bits = (ulong)BitConverter.DoubleToInt64Bits(op1);
|
||||
ulong op2_bits = (ulong)BitConverter.DoubleToInt64Bits(op2);
|
||||
|
||||
// Handle NaN cases
|
||||
if (ProcessNaNs(op1_bits, op2_bits, out ulong op_bits))
|
||||
{
|
||||
return BitConverter.Int64BitsToDouble((long)op_bits);
|
||||
}
|
||||
|
||||
// Return the most negative zero
|
||||
if ((op1_bits | op2_bits) == 0x8000000000000000ul)
|
||||
{
|
||||
return BitConverter.Int64BitsToDouble(long.MinValue);
|
||||
}
|
||||
|
||||
return 0;
|
||||
}
|
||||
|
||||
private static bool ProcessNaNs(uint op1_bits, uint op2_bits, out uint op_bits)
|
||||
{
|
||||
if (IsSNaN(op1_bits))
|
||||
{
|
||||
op_bits = op1_bits | (1u << 22); // op1 is SNaN, return QNaN op1
|
||||
}
|
||||
else if (IsSNaN(op2_bits))
|
||||
{
|
||||
op_bits = op2_bits | (1u << 22); // op2 is SNaN, return QNaN op2
|
||||
}
|
||||
else if (IsQNaN(op1_bits))
|
||||
{
|
||||
op_bits = op1_bits; // op1 is QNaN, return QNaN op1
|
||||
}
|
||||
else if (IsQNaN(op2_bits))
|
||||
{
|
||||
op_bits = op2_bits; // op2 is QNaN, return QNaN op2
|
||||
}
|
||||
else
|
||||
{
|
||||
op_bits = 0;
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
private static bool ProcessNaNs(ulong op1_bits, ulong op2_bits, out ulong op_bits)
|
||||
{
|
||||
if (IsSNaN(op1_bits))
|
||||
{
|
||||
op_bits = op1_bits | (1ul << 51); // op1 is SNaN, return QNaN op1
|
||||
}
|
||||
else if (IsSNaN(op2_bits))
|
||||
{
|
||||
op_bits = op2_bits | (1ul << 51); // op2 is SNaN, return QNaN op2
|
||||
}
|
||||
else if (IsQNaN(op1_bits))
|
||||
{
|
||||
op_bits = op1_bits; // op1 is QNaN, return QNaN op1
|
||||
}
|
||||
else if (IsQNaN(op2_bits))
|
||||
{
|
||||
op_bits = op2_bits; // op2 is QNaN, return QNaN op2
|
||||
}
|
||||
else
|
||||
{
|
||||
op_bits = 0;
|
||||
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
private static bool IsQNaN(uint op_bits)
|
||||
{
|
||||
return (op_bits & 0x007FFFFF) != 0 &&
|
||||
(op_bits & 0x7FC00000) == 0x7FC00000;
|
||||
}
|
||||
|
||||
private static bool IsQNaN(ulong op_bits)
|
||||
{
|
||||
return (op_bits & 0x000FFFFFFFFFFFFF) != 0 &&
|
||||
(op_bits & 0x7FF8000000000000) == 0x7FF8000000000000;
|
||||
}
|
||||
|
||||
private static bool IsSNaN(uint op_bits)
|
||||
{
|
||||
return (op_bits & 0x007FFFFF) != 0 &&
|
||||
(op_bits & 0x7FC00000) == 0x7F800000;
|
||||
}
|
||||
|
||||
private static bool IsSNaN(ulong op_bits)
|
||||
{
|
||||
return (op_bits & 0x000FFFFFFFFFFFFF) != 0 &&
|
||||
(op_bits & 0x7FF8000000000000) == 0x7FF0000000000000;
|
||||
}
|
||||
}
|
||||
}
|
||||
Loading…
Add table
Add a link
Reference in a new issue