drashna/ryujinx
1
0
Fork 0
Code Activity Actions

Add SSE4.2 Path for CRC32, add A32 variant, add tests for non-castagnoli variants. (#1328)

Browse source 
* Add CRC32 A32 instructions.

* Fix CRC32 instructions.

* Add CRC intrinsic and fast path.

Loop is currently unrolled, will look into adding temp vars after tests are added.

* Begin work on Crc tests

* Fix SSE4.2 path for CRC32C, finialize tests.

* Remove unused IR path.

* Fix spacing between prefix checks.

* This should be Src.

* PTC Version

* OpCodeTable Order

* Integer check improvement. Value and Crc can be either 32 or 64 size.

* This wasn't necessary...

* If size is 3, value type must be I64.

* Fix same src+dest handling for non crc intrinsics.

* Pre-fix (ha) issue with vex encodings
This commit is contained in:
riperiperi 2020-07-13 11:48:14 +01:00 • committed by GitHub
parent 30d4f752f4
commit d7044b10a2
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
15 changed files with 448 additions and 161 deletions
Download patch file Download diff file Expand all files Collapse all files

147
ARMeilleure/Instructions/InstEmitHash.cs
View file

@ -1,182 +1,67 @@
// https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/fast-crc-computation-generic-polynomials-pclmulqdq-paper.pdf
using ARMeilleure.Decoders;
using ARMeilleure.IntermediateRepresentation;
using ARMeilleure.Translation;
using static ARMeilleure.Instructions.InstEmitHashHelper;
using static ARMeilleure.Instructions.InstEmitHelper;
using static ARMeilleure.Instructions.InstEmitSimdHelper;
using static ARMeilleure.IntermediateRepresentation.OperandHelper;
namespace ARMeilleure.Instructions
{
static partial class InstEmit
{
private const int ByteSizeLog2 = 0;
private const int HWordSizeLog2 = 1;
private const int WordSizeLog2 = 2;
private const int DWordSizeLog2 = 3;
public static void Crc32b(ArmEmitterContext context)
{
if (Optimizations.UsePclmulqdq)
{
EmitCrc32Optimized(context, false, 8);
}
else
{
EmitCrc32Call(context, nameof(SoftFallback.Crc32b));
}
EmitCrc32Call(context, ByteSizeLog2, false);
}
public static void Crc32h(ArmEmitterContext context)
{
if (Optimizations.UsePclmulqdq)
{
EmitCrc32Optimized(context, false, 16);
}
else
{
EmitCrc32Call(context, nameof(SoftFallback.Crc32h));
}
EmitCrc32Call(context, HWordSizeLog2, false);
}
public static void Crc32w(ArmEmitterContext context)
{
if (Optimizations.UsePclmulqdq)
{
EmitCrc32Optimized(context, false, 32);
}
else
{
EmitCrc32Call(context, nameof(SoftFallback.Crc32w));
}
EmitCrc32Call(context, WordSizeLog2, false);
}
public static void Crc32x(ArmEmitterContext context)
{
if (Optimizations.UsePclmulqdq)
{
EmitCrc32Optimized64(context, false);
}
else
{
EmitCrc32Call(context, nameof(SoftFallback.Crc32x));
}
EmitCrc32Call(context, DWordSizeLog2, false);
}
public static void Crc32cb(ArmEmitterContext context)
{
if (Optimizations.UsePclmulqdq)
{
EmitCrc32Optimized(context, true, 8);
}
else
{
EmitCrc32Call(context, nameof(SoftFallback.Crc32cb));
}
EmitCrc32Call(context, ByteSizeLog2, true);
}
public static void Crc32ch(ArmEmitterContext context)
{
if (Optimizations.UsePclmulqdq)
{
EmitCrc32Optimized(context, true, 16);
}
else
{
EmitCrc32Call(context, nameof(SoftFallback.Crc32ch));
}
EmitCrc32Call(context, HWordSizeLog2, true);
}
public static void Crc32cw(ArmEmitterContext context)
{
if (Optimizations.UsePclmulqdq)
{
EmitCrc32Optimized(context, true, 32);
}
else
{
EmitCrc32Call(context, nameof(SoftFallback.Crc32cw));
}
EmitCrc32Call(context, WordSizeLog2, true);
}
public static void Crc32cx(ArmEmitterContext context)
{
if (Optimizations.UsePclmulqdq)
{
EmitCrc32Optimized64(context, true);
}
else
{
EmitCrc32Call(context, nameof(SoftFallback.Crc32cx));
}
EmitCrc32Call(context, DWordSizeLog2, true);
}
private static void EmitCrc32Optimized(ArmEmitterContext context, bool castagnoli, int bitsize)
{
OpCodeAluBinary op = (OpCodeAluBinary)context.CurrOp;
long mu = castagnoli ? 0x0DEA713F1 : 0x1F7011641; // mu' = floor(x^64/P(x))'
long polynomial = castagnoli ? 0x105EC76F0 : 0x1DB710641; // P'(x) << 1
Operand crc = GetIntOrZR(context, op.Rn);
Operand data = GetIntOrZR(context, op.Rm);
crc = context.VectorInsert(context.VectorZero(), crc, 0);
switch (bitsize)
{
case 8: data = context.VectorInsert8(context.VectorZero(), data, 0); break;
case 16: data = context.VectorInsert16(context.VectorZero(), data, 0); break;
case 32: data = context.VectorInsert(context.VectorZero(), data, 0); break;
}
Operand tmp = context.AddIntrinsic(Intrinsic.X86Pxor, crc, data);
tmp = context.AddIntrinsic(Intrinsic.X86Psllq, tmp, Const(64 - bitsize));
tmp = context.AddIntrinsic(Intrinsic.X86Pclmulqdq, tmp, X86GetScalar(context, mu), Const(0));
tmp = context.AddIntrinsic(Intrinsic.X86Pclmulqdq, tmp, X86GetScalar(context, polynomial), Const(0));
if (bitsize < 32)
{
crc = context.AddIntrinsic(Intrinsic.X86Pslldq, crc, Const((64 - bitsize) / 8));
tmp = context.AddIntrinsic(Intrinsic.X86Pxor, tmp, crc);
}
SetIntOrZR(context, op.Rd, context.VectorExtract(OperandType.I32, tmp, 2));
}
private static void EmitCrc32Optimized64(ArmEmitterContext context, bool castagnoli)
{
OpCodeAluBinary op = (OpCodeAluBinary)context.CurrOp;
long mu = castagnoli ? 0x0DEA713F1 : 0x1F7011641; // mu' = floor(x^64/P(x))'
long polynomial = castagnoli ? 0x105EC76F0 : 0x1DB710641; // P'(x) << 1
Operand crc = GetIntOrZR(context, op.Rn);
Operand data = GetIntOrZR(context, op.Rm);
crc = context.VectorInsert(context.VectorZero(), crc, 0);
data = context.VectorInsert(context.VectorZero(), data, 0);
Operand tmp = context.AddIntrinsic(Intrinsic.X86Pxor, crc, data);
Operand res = context.AddIntrinsic(Intrinsic.X86Pslldq, tmp, Const(4));
tmp = context.AddIntrinsic(Intrinsic.X86Pclmulqdq, res, X86GetScalar(context, mu), Const(0));
tmp = context.AddIntrinsic(Intrinsic.X86Pclmulqdq, tmp, X86GetScalar(context, polynomial), Const(0));
tmp = context.AddIntrinsic(Intrinsic.X86Pxor, tmp, res);
tmp = context.AddIntrinsic(Intrinsic.X86Psllq, tmp, Const(32));
tmp = context.AddIntrinsic(Intrinsic.X86Pclmulqdq, tmp, X86GetScalar(context, mu), Const(1));
tmp = context.AddIntrinsic(Intrinsic.X86Pclmulqdq, tmp, X86GetScalar(context, polynomial), Const(0));
SetIntOrZR(context, op.Rd, context.VectorExtract(OperandType.I32, tmp, 2));
}
private static void EmitCrc32Call(ArmEmitterContext context, string name)
private static void EmitCrc32Call(ArmEmitterContext context, int size, bool c)
{
OpCodeAluBinary op = (OpCodeAluBinary)context.CurrOp;
Operand n = GetIntOrZR(context, op.Rn);
Operand m = GetIntOrZR(context, op.Rm);
Operand d = context.Call(typeof(SoftFallback).GetMethod(name), n, m);
Operand d = EmitCrc32(context, n, m, size, c);
SetIntOrZR(context, op.Rd, d);
}