Sse optimized the 32-bit Vector & Scalar integer-to-fp conversion instructions (signed & unsigned); added the related Gp & V_Fixed Tests (signed & unsigned). (#662)

* Update CpuTestSimdCvt.cs

* Update CpuTestSimd.cs

* Update CpuTestSimdShImm.cs

* Update InstEmitSimdCvt.cs

* Update OpCodeTable.cs

* Update InstEmitSimdCvt.cs
This commit is contained in:
LDj3SNuD 2019-04-21 04:07:35 +02:00 committed by gdkchan
parent c33c574a24
commit f62521ef11
2 changed files with 173 additions and 22 deletions

View file

@ -363,7 +363,7 @@ namespace ChocolArm64.Instructions
if (context.CurrOp.RegisterSize == RegisterSize.Int32) if (context.CurrOp.RegisterSize == RegisterSize.Int32)
{ {
context.Emit(OpCodes.Conv_U4); context.Emit(OpCodes.Conv_I4);
} }
EmitFloatCast(context, op.Size); EmitFloatCast(context, op.Size);
@ -393,11 +393,20 @@ namespace ChocolArm64.Instructions
{ {
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp; OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
EmitVectorExtractSx(context, op.Rn, 0, op.Size + 2); int sizeF = op.Size & 1;
EmitFloatCast(context, op.Size); if (Optimizations.UseSse2 && sizeF == 0)
{
EmitSse2cvtF_Signed(context, scalar: true);
}
else
{
EmitVectorExtractSx(context, op.Rn, 0, sizeF + 2);
EmitScalarSetF(context, op.Rd, op.Size); EmitFloatCast(context, sizeF);
EmitScalarSetF(context, op.Rd, sizeF);
}
} }
public static void Scvtf_V(ILEmitterCtx context) public static void Scvtf_V(ILEmitterCtx context)
@ -408,18 +417,24 @@ namespace ChocolArm64.Instructions
if (Optimizations.UseSse2 && sizeF == 0) if (Optimizations.UseSse2 && sizeF == 0)
{ {
Type[] typesCvt = new Type[] { typeof(Vector128<int>) }; EmitSse2cvtF_Signed(context, scalar: false);
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
context.EmitStvec(op.Rd);
if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
} }
else
{
EmitVectorCvtf(context, signed: true);
}
}
public static void Scvtf_V_Fixed(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
// sizeF == ((OpCodeSimdShImm64)op).Size - 2
int sizeF = op.Size & 1;
if (Optimizations.UseSse2 && sizeF == 0)
{
EmitSse2cvtF_Signed(context, scalar: false);
} }
else else
{ {
@ -469,19 +484,56 @@ namespace ChocolArm64.Instructions
{ {
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp; OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
EmitVectorExtractZx(context, op.Rn, 0, op.Size + 2); int sizeF = op.Size & 1;
if (Optimizations.UseSse2 && sizeF == 0)
{
EmitSse2cvtF_Unsigned(context, scalar: true);
}
else
{
EmitVectorExtractZx(context, op.Rn, 0, sizeF + 2);
context.Emit(OpCodes.Conv_R_Un); context.Emit(OpCodes.Conv_R_Un);
EmitFloatCast(context, op.Size); EmitFloatCast(context, sizeF);
EmitScalarSetF(context, op.Rd, op.Size); EmitScalarSetF(context, op.Rd, sizeF);
}
} }
public static void Ucvtf_V(ILEmitterCtx context) public static void Ucvtf_V(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
int sizeF = op.Size & 1;
if (Optimizations.UseSse2 && sizeF == 0)
{
EmitSse2cvtF_Unsigned(context, scalar: false);
}
else
{ {
EmitVectorCvtf(context, signed: false); EmitVectorCvtf(context, signed: false);
} }
}
public static void Ucvtf_V_Fixed(ILEmitterCtx context)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
// sizeF == ((OpCodeSimdShImm64)op).Size - 2
int sizeF = op.Size & 1;
if (Optimizations.UseSse2 && sizeF == 0)
{
EmitSse2cvtF_Unsigned(context, scalar: false);
}
else
{
EmitVectorCvtf(context, signed: false);
}
}
private static void EmitFcvtn(ILEmitterCtx context, bool signed, bool scalar) private static void EmitFcvtn(ILEmitterCtx context, bool signed, bool scalar)
{ {
@ -838,7 +890,7 @@ namespace ChocolArm64.Instructions
int fBits = GetImmShr(fixedOp); int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits) // BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits)
int fpScaled = 0x40000000 + (fBits - 1) * 0x800000; int fpScaled = 0x3F800000 + fBits * 0x800000;
context.EmitLdc_I4(fpScaled); context.EmitLdc_I4(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav)); context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
@ -894,7 +946,7 @@ namespace ChocolArm64.Instructions
int fBits = GetImmShr(fixedOp); int fBits = GetImmShr(fixedOp);
// BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits) // BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits)
long fpScaled = 0x4000000000000000L + (fBits - 1) * 0x10000000000000L; long fpScaled = 0x3FF0000000000000L + fBits * 0x10000000000000L;
context.EmitLdc_I8(fpScaled); context.EmitLdc_I8(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav)); context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
@ -972,7 +1024,7 @@ namespace ChocolArm64.Instructions
int fBits = GetImmShr(fixedOp); int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits) // BitConverter.Int32BitsToSingle(fpScaled) == MathF.Pow(2f, fBits)
int fpScaled = 0x40000000 + (fBits - 1) * 0x800000; int fpScaled = 0x3F800000 + fBits * 0x800000;
context.EmitLdc_I4(fpScaled); context.EmitLdc_I4(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav)); context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
@ -1060,7 +1112,7 @@ namespace ChocolArm64.Instructions
int fBits = GetImmShr(fixedOp); int fBits = GetImmShr(fixedOp);
// BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits) // BitConverter.Int64BitsToDouble(fpScaled) == Math.Pow(2d, fBits)
long fpScaled = 0x4000000000000000L + (fBits - 1) * 0x10000000000000L; long fpScaled = 0x3FF0000000000000L + fBits * 0x10000000000000L;
context.EmitLdc_I8(fpScaled); context.EmitLdc_I8(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav)); context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
@ -1158,6 +1210,101 @@ namespace ChocolArm64.Instructions
} }
} }
private static void EmitSse2cvtF_Signed(ILEmitterCtx context, bool scalar)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
Type[] typesMul = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
Type[] typesCvt = new Type[] { typeof(Vector128<int>) };
Type[] typesSav = new Type[] { typeof(int) };
context.EmitLdvec(op.Rn);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
if (op is OpCodeSimdShImm64 fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == 1f / MathF.Pow(2f, fBits)
int fpScaled = 0x3F800000 - fBits * 0x800000;
context.EmitLdc_I4(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesMul));
}
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZero32_128(context, op.Rd);
}
else if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
private static void EmitSse2cvtF_Unsigned(ILEmitterCtx context, bool scalar)
{
OpCodeSimd64 op = (OpCodeSimd64)context.CurrOp;
Type[] typesMulAdd = new Type[] { typeof(Vector128<float>), typeof(Vector128<float>) };
Type[] typesSrlSll = new Type[] { typeof(Vector128<int>), typeof(byte) };
Type[] typesCvt = new Type[] { typeof(Vector128<int>) };
Type[] typesSav = new Type[] { typeof(int) };
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(16);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrlSll));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
context.EmitLdc_I4(0x47800000); // 65536.0f (1 << 16)
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesMulAdd));
context.EmitLdvec(op.Rn);
context.EmitLdc_I4(16);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftLeftLogical), typesSrlSll));
context.EmitLdc_I4(16);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ShiftRightLogical), typesSrlSll));
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.ConvertToVector128Single), typesCvt));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Add), typesMulAdd));
if (op is OpCodeSimdShImm64 fixedOp)
{
int fBits = GetImmShr(fixedOp);
// BitConverter.Int32BitsToSingle(fpScaled) == 1f / MathF.Pow(2f, fBits)
int fpScaled = 0x3F800000 - fBits * 0x800000;
context.EmitLdc_I4(fpScaled);
context.EmitCall(typeof(Sse2).GetMethod(nameof(Sse2.SetAllVector128), typesSav));
context.EmitCall(typeof(Sse).GetMethod(nameof(Sse.Multiply), typesMulAdd));
}
context.EmitStvec(op.Rd);
if (scalar)
{
EmitVectorZero32_128(context, op.Rd);
}
else if (op.RegisterSize == RegisterSize.Simd64)
{
EmitVectorZeroUpper(context, op.Rd);
}
}
private static string GetSse41NameRnd(RoundMode roundMode) private static string GetSse41NameRnd(RoundMode roundMode)
{ {
switch (roundMode) switch (roundMode)

View file

@ -439,6 +439,8 @@ namespace ChocolArm64
SetA64(">00111100x000010>xxxxxxxxxxxxxxx", InstEmit.Scvtf_Gp_Fixed, typeof(OpCodeSimdCvt64)); SetA64(">00111100x000010>xxxxxxxxxxxxxxx", InstEmit.Scvtf_Gp_Fixed, typeof(OpCodeSimdCvt64));
SetA64("010111100x100001110110xxxxxxxxxx", InstEmit.Scvtf_S, typeof(OpCodeSimd64)); SetA64("010111100x100001110110xxxxxxxxxx", InstEmit.Scvtf_S, typeof(OpCodeSimd64));
SetA64("0>0011100<100001110110xxxxxxxxxx", InstEmit.Scvtf_V, typeof(OpCodeSimd64)); SetA64("0>0011100<100001110110xxxxxxxxxx", InstEmit.Scvtf_V, typeof(OpCodeSimd64));
SetA64("0x001111001xxxxx111001xxxxxxxxxx", InstEmit.Scvtf_V_Fixed, typeof(OpCodeSimdShImm64));
SetA64("0100111101xxxxxx111001xxxxxxxxxx", InstEmit.Scvtf_V_Fixed, typeof(OpCodeSimdShImm64));
SetA64("01011110000xxxxx000000xxxxxxxxxx", InstEmit.Sha1c_V, typeof(OpCodeSimdReg64)); SetA64("01011110000xxxxx000000xxxxxxxxxx", InstEmit.Sha1c_V, typeof(OpCodeSimdReg64));
SetA64("0101111000101000000010xxxxxxxxxx", InstEmit.Sha1h_V, typeof(OpCodeSimd64)); SetA64("0101111000101000000010xxxxxxxxxx", InstEmit.Sha1h_V, typeof(OpCodeSimd64));
SetA64("01011110000xxxxx001000xxxxxxxxxx", InstEmit.Sha1m_V, typeof(OpCodeSimdReg64)); SetA64("01011110000xxxxx001000xxxxxxxxxx", InstEmit.Sha1m_V, typeof(OpCodeSimdReg64));
@ -548,6 +550,8 @@ namespace ChocolArm64
SetA64(">00111100x000011>xxxxxxxxxxxxxxx", InstEmit.Ucvtf_Gp_Fixed, typeof(OpCodeSimdCvt64)); SetA64(">00111100x000011>xxxxxxxxxxxxxxx", InstEmit.Ucvtf_Gp_Fixed, typeof(OpCodeSimdCvt64));
SetA64("011111100x100001110110xxxxxxxxxx", InstEmit.Ucvtf_S, typeof(OpCodeSimd64)); SetA64("011111100x100001110110xxxxxxxxxx", InstEmit.Ucvtf_S, typeof(OpCodeSimd64));
SetA64("0>1011100<100001110110xxxxxxxxxx", InstEmit.Ucvtf_V, typeof(OpCodeSimd64)); SetA64("0>1011100<100001110110xxxxxxxxxx", InstEmit.Ucvtf_V, typeof(OpCodeSimd64));
SetA64("0x101111001xxxxx111001xxxxxxxxxx", InstEmit.Ucvtf_V_Fixed, typeof(OpCodeSimdShImm64));
SetA64("0110111101xxxxxx111001xxxxxxxxxx", InstEmit.Ucvtf_V_Fixed, typeof(OpCodeSimdShImm64));
SetA64("0x101110<<1xxxxx000001xxxxxxxxxx", InstEmit.Uhadd_V, typeof(OpCodeSimdReg64)); SetA64("0x101110<<1xxxxx000001xxxxxxxxxx", InstEmit.Uhadd_V, typeof(OpCodeSimdReg64));
SetA64("0x101110<<1xxxxx001001xxxxxxxxxx", InstEmit.Uhsub_V, typeof(OpCodeSimdReg64)); SetA64("0x101110<<1xxxxx001001xxxxxxxxxx", InstEmit.Uhsub_V, typeof(OpCodeSimdReg64));
SetA64("0x101110<<1xxxxx011001xxxxxxxxxx", InstEmit.Umax_V, typeof(OpCodeSimdReg64)); SetA64("0x101110<<1xxxxx011001xxxxxxxxxx", InstEmit.Umax_V, typeof(OpCodeSimdReg64));