deadcade/Ryujinx
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Improve CountLeadingZeros() algorithm, nits. (#219)

Browse source 
* Update AInstEmitSimdArithmetic.cs

* Update ASoftFallback.cs

* Update ASoftFallback.cs

* Update ASoftFallback.cs

* Update AInstEmitSimdArithmetic.cs
This commit is contained in:
LDj3SNuD 2018-07-14 20:07:44 +02:00 committed by gdkchan
parent fc12fca962
commit be31f5b46d
2 changed files with 37 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
ChocolArm64/Instruction/AInstEmitSimdArithmetic.cs
View file

@ -101,11 +101,13 @@ namespace ChocolArm64.Instruction
int Bytes = Op.GetBitsCount() >> 3; int Bytes = Op.GetBitsCount() >> 3;
int Elems = Bytes >> Op.Size; int Elems = Bytes >> Op.Size;
int ESize = 8 << Op.Size;
for (int Index = 0; Index < Elems; Index++) for (int Index = 0; Index < Elems; Index++)
{ {
EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size); EmitVectorExtractZx(Context, Op.Rn, Index, Op.Size);
Context.EmitLdc_I4(8 << Op.Size); Context.EmitLdc_I4(ESize);
Emit(); Emit();

46
ChocolArm64/Instruction/ASoftFallback.cs
View file

@ -12,22 +12,42 @@ namespace ChocolArm64.Instruction
public static ulong CountLeadingSigns(ulong Value, int Size) public static ulong CountLeadingSigns(ulong Value, int Size)
{ {
return CountLeadingZeros((Value >> 1) ^ Value, Size - 1); Value ^= Value >> 1;
}
public static ulong CountLeadingZeros(ulong Value, int Size) int HighBit = Size - 2;
{
int HighBit = Size - 1;
for (int Bit = HighBit; Bit >= 0; Bit--) for (int Bit = HighBit; Bit >= 0; Bit--)
{ {
if (((Value >> Bit) & 1) != 0) if (((Value >> Bit) & 0b1) != 0)
{ {
return (ulong)(HighBit - Bit); return (ulong)(HighBit - Bit);
} }
} }
return (ulong)Size; return (ulong)(Size - 1);
}
private static readonly byte[] ClzNibbleTbl = { 4, 3, 2, 2, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0 };
public static ulong CountLeadingZeros(ulong Value, int Size)
{
if (Value == 0)
{
return (ulong)Size;
}
int NibbleIdx = Size;
int PreCount, Count = 0;
do
{
NibbleIdx -= 4;
PreCount = ClzNibbleTbl[(Value >> NibbleIdx) & 0b1111];
Count += PreCount;
}
while (PreCount == 4);
return (ulong)Count;
} }
public static uint CountSetBits8(uint Value) public static uint CountSetBits8(uint Value)
@ -61,8 +81,8 @@ namespace ChocolArm64.Instruction
private static uint Crc32w(uint Crc, uint Poly, uint Val) private static uint Crc32w(uint Crc, uint Poly, uint Val)
{ {
Crc = Crc32(Crc, Poly, (byte)(Val >> 0)); Crc = Crc32(Crc, Poly, (byte)(Val >> 0 ));
Crc = Crc32(Crc, Poly, (byte)(Val >> 8)); Crc = Crc32(Crc, Poly, (byte)(Val >> 8 ));
Crc = Crc32(Crc, Poly, (byte)(Val >> 16)); Crc = Crc32(Crc, Poly, (byte)(Val >> 16));
Crc = Crc32(Crc, Poly, (byte)(Val >> 24)); Crc = Crc32(Crc, Poly, (byte)(Val >> 24));
@ -71,8 +91,8 @@ namespace ChocolArm64.Instruction
private static uint Crc32x(uint Crc, uint Poly, ulong Val) private static uint Crc32x(uint Crc, uint Poly, ulong Val)
{ {
Crc = Crc32(Crc, Poly, (byte)(Val >> 0)); Crc = Crc32(Crc, Poly, (byte)(Val >> 0 ));
Crc = Crc32(Crc, Poly, (byte)(Val >> 8)); Crc = Crc32(Crc, Poly, (byte)(Val >> 8 ));
Crc = Crc32(Crc, Poly, (byte)(Val >> 16)); Crc = Crc32(Crc, Poly, (byte)(Val >> 16));
Crc = Crc32(Crc, Poly, (byte)(Val >> 24)); Crc = Crc32(Crc, Poly, (byte)(Val >> 24));
Crc = Crc32(Crc, Poly, (byte)(Val >> 32)); Crc = Crc32(Crc, Poly, (byte)(Val >> 32));
@ -168,9 +188,10 @@ namespace ChocolArm64.Instruction
public static long SMulHi128(long LHS, long RHS) public static long SMulHi128(long LHS, long RHS)
{ {
long Result = (long)UMulHi128((ulong)(LHS), (ulong)(RHS)); long Result = (long)UMulHi128((ulong)LHS, (ulong)RHS);
if (LHS < 0) Result -= RHS; if (LHS < 0) Result -= RHS;
if (RHS < 0) Result -= LHS; if (RHS < 0) Result -= LHS;
return Result; return Result;
} }
@ -187,6 +208,7 @@ namespace ChocolArm64.Instruction
ulong Z1 = T & 0xFFFFFFFF; ulong Z1 = T & 0xFFFFFFFF;
ulong Z0 = T >> 32; ulong Z0 = T >> 32;
Z1 += LLow * RHigh; Z1 += LLow * RHigh;
return LHigh * RHigh + Z0 + (Z1 >> 32); return LHigh * RHigh + Z0 + (Z1 >> 32);
} }
} }