Add Flush-to-zero mode (input, output) to FP instructions (slow paths); update FP Tests. Update Naming Conventions for Tests project. (#489)
* Update SoftFloat.cs * Update SoftFallback.cs * Update InstEmitSimdShift.cs * Update InstEmitSimdCvt.cs * Update InstEmitSimdArithmetic.cs * Update CryptoHelper.cs * Update CpuTestSimd.cs * Update CpuTestSimdReg.cs * Update CpuThreadState.cs * Update OpCodeTable.cs * Add files via upload * Nit. * Remove unused using. Nit. * Remove unused using. FZ update. * Nit. * Remove unused using.
This commit is contained in:
parent
d527632d1c
commit
ecf67bdcef
8 changed files with 499 additions and 295 deletions
|
@ -185,10 +185,10 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
int idx = columns << 2;
|
||||
|
||||
byte row0 = inState[idx + 0]; // A, E, I, M: [Row0, Col0-Col3]
|
||||
byte row1 = inState[idx + 1]; // B, F, J, N: [Row1, Col0-Col3]
|
||||
byte row2 = inState[idx + 2]; // C, G, K, O: [Row2, Col0-Col3]
|
||||
byte row3 = inState[idx + 3]; // D, H, L, P: [Row3, Col0-Col3]
|
||||
byte row0 = inState[idx + 0]; // A, E, I, M: [row0, col0-col3]
|
||||
byte row1 = inState[idx + 1]; // B, F, J, N: [row1, col0-col3]
|
||||
byte row2 = inState[idx + 2]; // C, G, K, O: [row2, col0-col3]
|
||||
byte row3 = inState[idx + 3]; // D, H, L, P: [row3, col0-col3]
|
||||
|
||||
outState[idx + 0] = (byte)((uint)_gfMul0E[row0] ^ _gfMul0B[row1] ^ _gfMul0D[row2] ^ _gfMul09[row3]);
|
||||
outState[idx + 1] = (byte)((uint)_gfMul09[row0] ^ _gfMul0E[row1] ^ _gfMul0B[row2] ^ _gfMul0D[row3]);
|
||||
|
@ -246,10 +246,10 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
int idx = columns << 2;
|
||||
|
||||
byte row0 = inState[idx + 0]; // A, E, I, M: [Row0, Col0-Col3]
|
||||
byte row1 = inState[idx + 1]; // B, F, J, N: [Row1, Col0-Col3]
|
||||
byte row2 = inState[idx + 2]; // C, G, K, O: [Row2, Col0-Col3]
|
||||
byte row3 = inState[idx + 3]; // D, H, L, P: [Row3, Col0-Col3]
|
||||
byte row0 = inState[idx + 0]; // A, E, I, M: [row0, col0-col3]
|
||||
byte row1 = inState[idx + 1]; // B, F, J, N: [row1, col0-col3]
|
||||
byte row2 = inState[idx + 2]; // C, G, K, O: [row2, col0-col3]
|
||||
byte row3 = inState[idx + 3]; // D, H, L, P: [row3, col0-col3]
|
||||
|
||||
outState[idx + 0] = (byte)((uint)_gfMul02[row0] ^ _gfMul03[row1] ^ row2 ^ row3);
|
||||
outState[idx + 1] = (byte)((uint)row0 ^ _gfMul02[row1] ^ _gfMul03[row2] ^ row3);
|
||||
|
|
|
@ -304,7 +304,7 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
EmitVectorZero32_128(context, op.Rd);
|
||||
}
|
||||
else /* if (Op.Size == 1) */
|
||||
else /* if (op.Size == 1) */
|
||||
{
|
||||
Type[] typesMulAdd = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
|
||||
|
||||
|
@ -516,7 +516,7 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
EmitVectorZero32_128(context, op.Rd);
|
||||
}
|
||||
else /* if (Op.Size == 1) */
|
||||
else /* if (op.Size == 1) */
|
||||
{
|
||||
Type[] typesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
|
||||
|
||||
|
@ -715,7 +715,7 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
EmitVectorZero32_128(context, op.Rd);
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
Type[] typesSsv = new Type[] { typeof(double) };
|
||||
Type[] typesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
|
||||
|
@ -772,7 +772,7 @@ namespace ChocolArm64.Instructions
|
|||
EmitVectorZeroUpper(context, op.Rd);
|
||||
}
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
Type[] typesSav = new Type[] { typeof(double) };
|
||||
Type[] typesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
|
||||
|
@ -1016,7 +1016,7 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
EmitVectorZero32_128(context, op.Rd);
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
Type[] typesSsv = new Type[] { typeof(double) };
|
||||
Type[] typesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
|
||||
|
@ -1043,7 +1043,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
EmitScalarBinaryOpF(context, () =>
|
||||
{
|
||||
EmitSoftFloatCall(context, nameof(SoftFloat32.FprSqrtStepFused));
|
||||
EmitSoftFloatCall(context, nameof(SoftFloat32.FPRSqrtStepFused));
|
||||
});
|
||||
}
|
||||
}
|
||||
|
@ -1081,7 +1081,7 @@ namespace ChocolArm64.Instructions
|
|||
EmitVectorZeroUpper(context, op.Rd);
|
||||
}
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
Type[] typesSav = new Type[] { typeof(double) };
|
||||
Type[] typesMulSub = new Type[] { typeof(Vector128<double>), typeof(Vector128<double>) };
|
||||
|
@ -1106,7 +1106,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
EmitVectorBinaryOpF(context, () =>
|
||||
{
|
||||
EmitSoftFloatCall(context, nameof(SoftFloat32.FprSqrtStepFused));
|
||||
EmitSoftFloatCall(context, nameof(SoftFloat32.FPRSqrtStepFused));
|
||||
});
|
||||
}
|
||||
}
|
||||
|
|
|
@ -89,9 +89,9 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
context.EmitLdarg(TranslatedSub.StateArgIdx);
|
||||
|
||||
context.EmitCall(typeof(SoftFloat1632), nameof(SoftFloat1632.FPConvert));
|
||||
context.EmitCall(typeof(SoftFloat16_32), nameof(SoftFloat16_32.FPConvert));
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
EmitVectorExtractF(context, op.Rn, part + index, 0);
|
||||
|
||||
|
@ -139,12 +139,12 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
context.EmitLdarg(TranslatedSub.StateArgIdx);
|
||||
|
||||
context.EmitCall(typeof(SoftFloat3216), nameof(SoftFloat3216.FPConvert));
|
||||
context.EmitCall(typeof(SoftFloat32_16), nameof(SoftFloat32_16.FPConvert));
|
||||
|
||||
context.Emit(OpCodes.Conv_U8);
|
||||
EmitVectorInsertTmp(context, part + index, 1);
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
context.Emit(OpCodes.Conv_R4);
|
||||
|
||||
|
@ -354,7 +354,7 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
context.Emit(OpCodes.Conv_U8);
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
VectorHelper.EmitCall(context, signed
|
||||
? nameof(VectorHelper.SatF64ToS64)
|
||||
|
@ -516,7 +516,7 @@ namespace ChocolArm64.Instructions
|
|||
? nameof(VectorHelper.SatF32ToS32)
|
||||
: nameof(VectorHelper.SatF32ToU32));
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
VectorHelper.EmitCall(context, signed
|
||||
? nameof(VectorHelper.SatF64ToS64)
|
||||
|
@ -565,7 +565,7 @@ namespace ChocolArm64.Instructions
|
|||
? nameof(VectorHelper.SatF32ToS32)
|
||||
: nameof(VectorHelper.SatF32ToU32));
|
||||
}
|
||||
else /* if (SizeF == 1) */
|
||||
else /* if (sizeF == 1) */
|
||||
{
|
||||
VectorHelper.EmitCall(context, signed
|
||||
? nameof(VectorHelper.SatF64ToS64)
|
||||
|
@ -601,7 +601,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
VectorHelper.EmitCall(context, nameof(VectorHelper.SatF32ToS32));
|
||||
}
|
||||
else /* if (Size == 1) */
|
||||
else /* if (size == 1) */
|
||||
{
|
||||
VectorHelper.EmitCall(context, nameof(VectorHelper.SatF64ToS32));
|
||||
}
|
||||
|
@ -612,7 +612,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
VectorHelper.EmitCall(context, nameof(VectorHelper.SatF32ToS64));
|
||||
}
|
||||
else /* if (Size == 1) */
|
||||
else /* if (size == 1) */
|
||||
{
|
||||
VectorHelper.EmitCall(context, nameof(VectorHelper.SatF64ToS64));
|
||||
}
|
||||
|
@ -634,7 +634,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
VectorHelper.EmitCall(context, nameof(VectorHelper.SatF32ToU32));
|
||||
}
|
||||
else /* if (Size == 1) */
|
||||
else /* if (size == 1) */
|
||||
{
|
||||
VectorHelper.EmitCall(context, nameof(VectorHelper.SatF64ToU32));
|
||||
}
|
||||
|
@ -645,7 +645,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
VectorHelper.EmitCall(context, nameof(VectorHelper.SatF32ToU64));
|
||||
}
|
||||
else /* if (Size == 1) */
|
||||
else /* if (size == 1) */
|
||||
{
|
||||
VectorHelper.EmitCall(context, nameof(VectorHelper.SatF64ToU64));
|
||||
}
|
||||
|
|
|
@ -658,9 +658,9 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
context.Emit(signed ? OpCodes.Shr : OpCodes.Shr_Un);
|
||||
}
|
||||
else /* if (Op.Size == 3) */
|
||||
else /* if (op.Size == 3) */
|
||||
{
|
||||
EmitShrImm_64(context, signed, round ? roundConst : 0L, shift);
|
||||
EmitShrImm64(context, signed, round ? roundConst : 0L, shift);
|
||||
}
|
||||
|
||||
if (accumulate)
|
||||
|
@ -795,9 +795,9 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
context.Emit(signedSrc ? OpCodes.Shr : OpCodes.Shr_Un);
|
||||
}
|
||||
else /* if (Op.Size == 2 && Round) */
|
||||
else /* if (op.Size == 2 && round) */
|
||||
{
|
||||
EmitShrImm_64(context, signedSrc, roundConst, shift); // Shift <= 32
|
||||
EmitShrImm64(context, signedSrc, roundConst, shift); // shift <= 32
|
||||
}
|
||||
|
||||
EmitSatQ(context, op.Size, signedSrc, signedDst);
|
||||
|
@ -814,8 +814,8 @@ namespace ChocolArm64.Instructions
|
|||
}
|
||||
}
|
||||
|
||||
// Dst_64 = (Int(Src_64, Signed) + RoundConst) >> Shift;
|
||||
private static void EmitShrImm_64(
|
||||
// dst64 = (Int(src64, signed) + roundConst) >> shift;
|
||||
private static void EmitShrImm64(
|
||||
ILEmitterCtx context,
|
||||
bool signed,
|
||||
long roundConst,
|
||||
|
@ -825,8 +825,8 @@ namespace ChocolArm64.Instructions
|
|||
context.EmitLdc_I4(shift);
|
||||
|
||||
SoftFallback.EmitCall(context, signed
|
||||
? nameof(SoftFallback.SignedShrImm_64)
|
||||
: nameof(SoftFallback.UnsignedShrImm_64));
|
||||
? nameof(SoftFallback.SignedShrImm64)
|
||||
: nameof(SoftFallback.UnsignedShrImm64));
|
||||
}
|
||||
|
||||
private static void EmitVectorShImmWidenBinarySx(ILEmitterCtx context, Action emit, int imm)
|
||||
|
|
|
@ -16,8 +16,8 @@ namespace ChocolArm64.Instructions
|
|||
context.EmitCall(typeof(SoftFallback), mthdName);
|
||||
}
|
||||
|
||||
#region "ShrImm_64"
|
||||
public static long SignedShrImm_64(long value, long roundConst, int shift)
|
||||
#region "ShrImm64"
|
||||
public static long SignedShrImm64(long value, long roundConst, int shift)
|
||||
{
|
||||
if (roundConst == 0L)
|
||||
{
|
||||
|
@ -25,7 +25,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
return value >> shift;
|
||||
}
|
||||
else /* if (Shift == 64) */
|
||||
else /* if (shift == 64) */
|
||||
{
|
||||
if (value < 0L)
|
||||
{
|
||||
|
@ -37,7 +37,7 @@ namespace ChocolArm64.Instructions
|
|||
}
|
||||
}
|
||||
}
|
||||
else /* if (RoundConst == 1L << (Shift - 1)) */
|
||||
else /* if (roundConst == 1L << (shift - 1)) */
|
||||
{
|
||||
if (shift <= 63)
|
||||
{
|
||||
|
@ -52,14 +52,14 @@ namespace ChocolArm64.Instructions
|
|||
return add >> shift;
|
||||
}
|
||||
}
|
||||
else /* if (Shift == 64) */
|
||||
else /* if (shift == 64) */
|
||||
{
|
||||
return 0L;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
public static ulong UnsignedShrImm_64(ulong value, long roundConst, int shift)
|
||||
public static ulong UnsignedShrImm64(ulong value, long roundConst, int shift)
|
||||
{
|
||||
if (roundConst == 0L)
|
||||
{
|
||||
|
@ -67,12 +67,12 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
return value >> shift;
|
||||
}
|
||||
else /* if (Shift == 64) */
|
||||
else /* if (shift == 64) */
|
||||
{
|
||||
return 0UL;
|
||||
}
|
||||
}
|
||||
else /* if (RoundConst == 1L << (Shift - 1)) */
|
||||
else /* if (roundConst == 1L << (shift - 1)) */
|
||||
{
|
||||
ulong add = value + (ulong)roundConst;
|
||||
|
||||
|
@ -82,7 +82,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
return (add >> shift) | (0x8000000000000000UL >> (shift - 1));
|
||||
}
|
||||
else /* if (Shift == 64) */
|
||||
else /* if (shift == 64) */
|
||||
{
|
||||
return 1UL;
|
||||
}
|
||||
|
@ -93,7 +93,7 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
return add >> shift;
|
||||
}
|
||||
else /* if (Shift == 64) */
|
||||
else /* if (shift == 64) */
|
||||
{
|
||||
return 0UL;
|
||||
}
|
||||
|
@ -285,8 +285,8 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
if (op1 <= (ulong)long.MaxValue)
|
||||
{
|
||||
// Op1 from ulong.MinValue to (ulong)long.MaxValue
|
||||
// Op2 from long.MinValue to long.MaxValue
|
||||
// op1 from ulong.MinValue to (ulong)long.MaxValue
|
||||
// op2 from long.MinValue to long.MaxValue
|
||||
|
||||
long add = (long)op1 + op2;
|
||||
|
||||
|
@ -303,8 +303,8 @@ namespace ChocolArm64.Instructions
|
|||
}
|
||||
else if (op2 >= 0L)
|
||||
{
|
||||
// Op1 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
|
||||
// Op2 from (long)ulong.MinValue to long.MaxValue
|
||||
// op1 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
|
||||
// op2 from (long)ulong.MinValue to long.MaxValue
|
||||
|
||||
state.SetFpsrFlag(Fpsr.Qc);
|
||||
|
||||
|
@ -312,8 +312,8 @@ namespace ChocolArm64.Instructions
|
|||
}
|
||||
else
|
||||
{
|
||||
// Op1 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
|
||||
// Op2 from long.MinValue to (long)ulong.MinValue - 1L
|
||||
// op1 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
|
||||
// op2 from long.MinValue to (long)ulong.MinValue - 1L
|
||||
|
||||
ulong add = op1 + (ulong)op2;
|
||||
|
||||
|
@ -334,8 +334,8 @@ namespace ChocolArm64.Instructions
|
|||
{
|
||||
if (op1 >= 0L)
|
||||
{
|
||||
// Op1 from (long)ulong.MinValue to long.MaxValue
|
||||
// Op2 from ulong.MinValue to ulong.MaxValue
|
||||
// op1 from (long)ulong.MinValue to long.MaxValue
|
||||
// op2 from ulong.MinValue to ulong.MaxValue
|
||||
|
||||
ulong add = (ulong)op1 + op2;
|
||||
|
||||
|
@ -352,15 +352,15 @@ namespace ChocolArm64.Instructions
|
|||
}
|
||||
else if (op2 > (ulong)long.MaxValue)
|
||||
{
|
||||
// Op1 from long.MinValue to (long)ulong.MinValue - 1L
|
||||
// Op2 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
|
||||
// op1 from long.MinValue to (long)ulong.MinValue - 1L
|
||||
// op2 from (ulong)long.MaxValue + 1UL to ulong.MaxValue
|
||||
|
||||
return (ulong)op1 + op2;
|
||||
}
|
||||
else
|
||||
{
|
||||
// Op1 from long.MinValue to (long)ulong.MinValue - 1L
|
||||
// Op2 from ulong.MinValue to (ulong)long.MaxValue
|
||||
// op1 from long.MinValue to (long)ulong.MinValue - 1L
|
||||
// op2 from ulong.MinValue to (ulong)long.MaxValue
|
||||
|
||||
long add = op1 + (long)op2;
|
||||
|
||||
|
@ -379,7 +379,7 @@ namespace ChocolArm64.Instructions
|
|||
#endregion
|
||||
|
||||
#region "Count"
|
||||
public static ulong CountLeadingSigns(ulong value, int size) // Size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
|
||||
public static ulong CountLeadingSigns(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
|
||||
{
|
||||
value ^= value >> 1;
|
||||
|
||||
|
@ -398,7 +398,7 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
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) // Size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
|
||||
public static ulong CountLeadingZeros(ulong value, int size) // size is 8, 16, 32 or 64 (SIMD&FP or Base Inst.).
|
||||
{
|
||||
if (value == 0ul)
|
||||
{
|
||||
|
@ -419,7 +419,7 @@ namespace ChocolArm64.Instructions
|
|||
return (ulong)count;
|
||||
}
|
||||
|
||||
public static ulong CountSetBits8(ulong value) // "Size" is 8 (SIMD&FP Inst.).
|
||||
public static ulong CountSetBits8(ulong value) // "size" is 8 (SIMD&FP Inst.).
|
||||
{
|
||||
if (value == 0xfful)
|
||||
{
|
||||
|
@ -531,72 +531,72 @@ namespace ChocolArm64.Instructions
|
|||
#endregion
|
||||
|
||||
#region "Sha1"
|
||||
public static Vector128<float> HashChoose(Vector128<float> hashAbcd, uint hashE, Vector128<float> wk)
|
||||
public static Vector128<float> HashChoose(Vector128<float> hash_abcd, uint hash_e, Vector128<float> wk)
|
||||
{
|
||||
for (int e = 0; e <= 3; e++)
|
||||
{
|
||||
uint t = ShaChoose((uint)VectorExtractIntZx(hashAbcd, (byte)1, 2),
|
||||
(uint)VectorExtractIntZx(hashAbcd, (byte)2, 2),
|
||||
(uint)VectorExtractIntZx(hashAbcd, (byte)3, 2));
|
||||
uint t = ShaChoose((uint)VectorExtractIntZx(hash_abcd, (byte)1, 2),
|
||||
(uint)VectorExtractIntZx(hash_abcd, (byte)2, 2),
|
||||
(uint)VectorExtractIntZx(hash_abcd, (byte)3, 2));
|
||||
|
||||
hashE += Rol((uint)VectorExtractIntZx(hashAbcd, (byte)0, 2), 5) + t;
|
||||
hashE += (uint)VectorExtractIntZx(wk, (byte)e, 2);
|
||||
hash_e += Rol((uint)VectorExtractIntZx(hash_abcd, (byte)0, 2), 5) + t;
|
||||
hash_e += (uint)VectorExtractIntZx(wk, (byte)e, 2);
|
||||
|
||||
t = Rol((uint)VectorExtractIntZx(hashAbcd, (byte)1, 2), 30);
|
||||
hashAbcd = VectorInsertInt((ulong)t, hashAbcd, (byte)1, 2);
|
||||
t = Rol((uint)VectorExtractIntZx(hash_abcd, (byte)1, 2), 30);
|
||||
hash_abcd = VectorInsertInt((ulong)t, hash_abcd, (byte)1, 2);
|
||||
|
||||
Rol32_160(ref hashE, ref hashAbcd);
|
||||
Rol32_160(ref hash_e, ref hash_abcd);
|
||||
}
|
||||
|
||||
return hashAbcd;
|
||||
return hash_abcd;
|
||||
}
|
||||
|
||||
public static uint FixedRotate(uint hashE)
|
||||
public static uint FixedRotate(uint hash_e)
|
||||
{
|
||||
return hashE.Rol(30);
|
||||
return hash_e.Rol(30);
|
||||
}
|
||||
|
||||
public static Vector128<float> HashMajority(Vector128<float> hashAbcd, uint hashE, Vector128<float> wk)
|
||||
public static Vector128<float> HashMajority(Vector128<float> hash_abcd, uint hash_e, Vector128<float> wk)
|
||||
{
|
||||
for (int e = 0; e <= 3; e++)
|
||||
{
|
||||
uint t = ShaMajority((uint)VectorExtractIntZx(hashAbcd, (byte)1, 2),
|
||||
(uint)VectorExtractIntZx(hashAbcd, (byte)2, 2),
|
||||
(uint)VectorExtractIntZx(hashAbcd, (byte)3, 2));
|
||||
uint t = ShaMajority((uint)VectorExtractIntZx(hash_abcd, (byte)1, 2),
|
||||
(uint)VectorExtractIntZx(hash_abcd, (byte)2, 2),
|
||||
(uint)VectorExtractIntZx(hash_abcd, (byte)3, 2));
|
||||
|
||||
hashE += Rol((uint)VectorExtractIntZx(hashAbcd, (byte)0, 2), 5) + t;
|
||||
hashE += (uint)VectorExtractIntZx(wk, (byte)e, 2);
|
||||
hash_e += Rol((uint)VectorExtractIntZx(hash_abcd, (byte)0, 2), 5) + t;
|
||||
hash_e += (uint)VectorExtractIntZx(wk, (byte)e, 2);
|
||||
|
||||
t = Rol((uint)VectorExtractIntZx(hashAbcd, (byte)1, 2), 30);
|
||||
hashAbcd = VectorInsertInt((ulong)t, hashAbcd, (byte)1, 2);
|
||||
t = Rol((uint)VectorExtractIntZx(hash_abcd, (byte)1, 2), 30);
|
||||
hash_abcd = VectorInsertInt((ulong)t, hash_abcd, (byte)1, 2);
|
||||
|
||||
Rol32_160(ref hashE, ref hashAbcd);
|
||||
Rol32_160(ref hash_e, ref hash_abcd);
|
||||
}
|
||||
|
||||
return hashAbcd;
|
||||
return hash_abcd;
|
||||
}
|
||||
|
||||
public static Vector128<float> HashParity(Vector128<float> hashAbcd, uint hashE, Vector128<float> wk)
|
||||
public static Vector128<float> HashParity(Vector128<float> hash_abcd, uint hash_e, Vector128<float> wk)
|
||||
{
|
||||
for (int e = 0; e <= 3; e++)
|
||||
{
|
||||
uint t = ShaParity((uint)VectorExtractIntZx(hashAbcd, (byte)1, 2),
|
||||
(uint)VectorExtractIntZx(hashAbcd, (byte)2, 2),
|
||||
(uint)VectorExtractIntZx(hashAbcd, (byte)3, 2));
|
||||
uint t = ShaParity((uint)VectorExtractIntZx(hash_abcd, (byte)1, 2),
|
||||
(uint)VectorExtractIntZx(hash_abcd, (byte)2, 2),
|
||||
(uint)VectorExtractIntZx(hash_abcd, (byte)3, 2));
|
||||
|
||||
hashE += Rol((uint)VectorExtractIntZx(hashAbcd, (byte)0, 2), 5) + t;
|
||||
hashE += (uint)VectorExtractIntZx(wk, (byte)e, 2);
|
||||
hash_e += Rol((uint)VectorExtractIntZx(hash_abcd, (byte)0, 2), 5) + t;
|
||||
hash_e += (uint)VectorExtractIntZx(wk, (byte)e, 2);
|
||||
|
||||
t = Rol((uint)VectorExtractIntZx(hashAbcd, (byte)1, 2), 30);
|
||||
hashAbcd = VectorInsertInt((ulong)t, hashAbcd, (byte)1, 2);
|
||||
t = Rol((uint)VectorExtractIntZx(hash_abcd, (byte)1, 2), 30);
|
||||
hash_abcd = VectorInsertInt((ulong)t, hash_abcd, (byte)1, 2);
|
||||
|
||||
Rol32_160(ref hashE, ref hashAbcd);
|
||||
Rol32_160(ref hash_e, ref hash_abcd);
|
||||
}
|
||||
|
||||
return hashAbcd;
|
||||
return hash_abcd;
|
||||
}
|
||||
|
||||
public static Vector128<float> Sha1SchedulePart1(Vector128<float> w03, Vector128<float> w47, Vector128<float> w811)
|
||||
public static Vector128<float> Sha1SchedulePart1(Vector128<float> w0_3, Vector128<float> w4_7, Vector128<float> w8_11)
|
||||
{
|
||||
if (!Sse.IsSupported)
|
||||
{
|
||||
|
@ -605,16 +605,16 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
Vector128<float> result = new Vector128<float>();
|
||||
|
||||
ulong t2 = VectorExtractIntZx(w47, (byte)0, 3);
|
||||
ulong t1 = VectorExtractIntZx(w03, (byte)1, 3);
|
||||
ulong t2 = VectorExtractIntZx(w4_7, (byte)0, 3);
|
||||
ulong t1 = VectorExtractIntZx(w0_3, (byte)1, 3);
|
||||
|
||||
result = VectorInsertInt((ulong)t1, result, (byte)0, 3);
|
||||
result = VectorInsertInt((ulong)t2, result, (byte)1, 3);
|
||||
|
||||
return Sse.Xor(result, Sse.Xor(w03, w811));
|
||||
return Sse.Xor(result, Sse.Xor(w0_3, w8_11));
|
||||
}
|
||||
|
||||
public static Vector128<float> Sha1SchedulePart2(Vector128<float> tw03, Vector128<float> w1215)
|
||||
public static Vector128<float> Sha1SchedulePart2(Vector128<float> tw0_3, Vector128<float> w12_15)
|
||||
{
|
||||
if (!Sse2.IsSupported)
|
||||
{
|
||||
|
@ -623,8 +623,8 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
Vector128<float> result = new Vector128<float>();
|
||||
|
||||
Vector128<float> t = Sse.Xor(tw03, Sse.StaticCast<uint, float>(
|
||||
Sse2.ShiftRightLogical128BitLane(Sse.StaticCast<float, uint>(w1215), (byte)4)));
|
||||
Vector128<float> t = Sse.Xor(tw0_3, Sse.StaticCast<uint, float>(
|
||||
Sse2.ShiftRightLogical128BitLane(Sse.StaticCast<float, uint>(w12_15), (byte)4)));
|
||||
|
||||
uint tE0 = (uint)VectorExtractIntZx(t, (byte)0, 2);
|
||||
uint tE1 = (uint)VectorExtractIntZx(t, (byte)1, 2);
|
||||
|
@ -676,28 +676,28 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
#region "Sha256"
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining)]
|
||||
public static Vector128<float> HashLower(Vector128<float> hashAbcd, Vector128<float> hashEfgh, Vector128<float> wk)
|
||||
public static Vector128<float> HashLower(Vector128<float> hash_abcd, Vector128<float> hash_efgh, Vector128<float> wk)
|
||||
{
|
||||
return Sha256Hash(hashAbcd, hashEfgh, wk, true);
|
||||
return Sha256Hash(hash_abcd, hash_efgh, wk, true);
|
||||
}
|
||||
|
||||
[MethodImpl(MethodImplOptions.AggressiveInlining)]
|
||||
public static Vector128<float> HashUpper(Vector128<float> hashEfgh, Vector128<float> hashAbcd, Vector128<float> wk)
|
||||
public static Vector128<float> HashUpper(Vector128<float> hash_efgh, Vector128<float> hash_abcd, Vector128<float> wk)
|
||||
{
|
||||
return Sha256Hash(hashAbcd, hashEfgh, wk, false);
|
||||
return Sha256Hash(hash_abcd, hash_efgh, wk, false);
|
||||
}
|
||||
|
||||
public static Vector128<float> Sha256SchedulePart1(Vector128<float> w03, Vector128<float> w47)
|
||||
public static Vector128<float> Sha256SchedulePart1(Vector128<float> w0_3, Vector128<float> w4_7)
|
||||
{
|
||||
Vector128<float> result = new Vector128<float>();
|
||||
|
||||
for (int e = 0; e <= 3; e++)
|
||||
{
|
||||
uint elt = (uint)VectorExtractIntZx(e <= 2 ? w03 : w47, (byte)(e <= 2 ? e + 1 : 0), 2);
|
||||
uint elt = (uint)VectorExtractIntZx(e <= 2 ? w0_3 : w4_7, (byte)(e <= 2 ? e + 1 : 0), 2);
|
||||
|
||||
elt = elt.Ror(7) ^ elt.Ror(18) ^ elt.Lsr(3);
|
||||
|
||||
elt += (uint)VectorExtractIntZx(w03, (byte)e, 2);
|
||||
elt += (uint)VectorExtractIntZx(w0_3, (byte)e, 2);
|
||||
|
||||
result = VectorInsertInt((ulong)elt, result, (byte)e, 2);
|
||||
}
|
||||
|
@ -705,11 +705,11 @@ namespace ChocolArm64.Instructions
|
|||
return result;
|
||||
}
|
||||
|
||||
public static Vector128<float> Sha256SchedulePart2(Vector128<float> w03, Vector128<float> w811, Vector128<float> w1215)
|
||||
public static Vector128<float> Sha256SchedulePart2(Vector128<float> w0_3, Vector128<float> w8_11, Vector128<float> w12_15)
|
||||
{
|
||||
Vector128<float> result = new Vector128<float>();
|
||||
|
||||
ulong t1 = VectorExtractIntZx(w1215, (byte)1, 3);
|
||||
ulong t1 = VectorExtractIntZx(w12_15, (byte)1, 3);
|
||||
|
||||
for (int e = 0; e <= 1; e++)
|
||||
{
|
||||
|
@ -717,8 +717,8 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
elt = elt.Ror(17) ^ elt.Ror(19) ^ elt.Lsr(10);
|
||||
|
||||
elt += (uint)VectorExtractIntZx(w03, (byte)e, 2);
|
||||
elt += (uint)VectorExtractIntZx(w811, (byte)(e + 1), 2);
|
||||
elt += (uint)VectorExtractIntZx(w0_3, (byte)e, 2);
|
||||
elt += (uint)VectorExtractIntZx(w8_11, (byte)(e + 1), 2);
|
||||
|
||||
result = VectorInsertInt((ulong)elt, result, (byte)e, 2);
|
||||
}
|
||||
|
@ -731,8 +731,8 @@ namespace ChocolArm64.Instructions
|
|||
|
||||
elt = elt.Ror(17) ^ elt.Ror(19) ^ elt.Lsr(10);
|
||||
|
||||
elt += (uint)VectorExtractIntZx(w03, (byte)e, 2);
|
||||
elt += (uint)VectorExtractIntZx(e == 2 ? w811 : w1215, (byte)(e == 2 ? 3 : 0), 2);
|
||||
elt += (uint)VectorExtractIntZx(w0_3, (byte)e, 2);
|
||||
elt += (uint)VectorExtractIntZx(e == 2 ? w8_11 : w12_15, (byte)(e == 2 ? 3 : 0), 2);
|
||||
|
||||
result = VectorInsertInt((ulong)elt, result, (byte)e, 2);
|
||||
}
|
||||
|
|
File diff suppressed because it is too large
Load diff
Loading…
Reference in a new issue