aaru-dps/AaruBenchmark
1
0
Fork 0
mirror of https://github.com/aaru-dps/AaruBenchmark.git synced 2025-12-16 19:24:36 +00:00
Code Issues Packages Projects Releases Wiki Activity

Reformat and cleanup.

Browse Source
This commit is contained in:
Natalia Portillo
2023-10-03 21:49:17 +01:00
parent a957e741ae
commit dbc5be31a2
32 changed files with 2729 additions and 2498 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
Aaru6.Checksums/CRC32/arm_simd.cs
View File

@@ -50,13 +50,13 @@ using System.Runtime.Intrinsics.Arm;
namespace Aaru6.Checksums.CRC32;
internal static class ArmSimd
static class ArmSimd
{
internal static uint Step64(byte[] buf, long len, uint crc)
{
uint c = crc;
int bufPos = 0;
var bufPos = 0;
while(len >= 64)
{
@@ -87,9 +87,7 @@ internal static class ArmSimd
}
while(len-- > 0)
{
c = Crc32.ComputeCrc32(c, buf[bufPos++]);
}
return c;
}
@@ -98,7 +96,7 @@ internal static class ArmSimd
{
uint c = crc;
int bufPos = 0;
var bufPos = 0;
while(len >= 32)
{
@@ -129,9 +127,7 @@ internal static class ArmSimd
}
while(len-- > 0)
{
c = Crc32.ComputeCrc32(c, buf[bufPos++]);
}
return c;
}

54
Aaru6.Checksums/CRC32/clmul.cs
View File

@@ -54,7 +54,7 @@ using System.Runtime.Intrinsics.X86;
namespace Aaru6.Checksums.CRC32;
internal static class Clmul
static class Clmul
{
static readonly uint[] _crcK =
{
@@ -64,18 +64,18 @@ internal static class Clmul
static readonly Vector128<uint>[] _pshufbShfTable =
{
Vector128.Create(0x84838281, 0x88878685, 0x8c8b8a89, 0x008f8e8d), /* shl 15 (16 - 1)/shr1 */
Vector128.Create(0x85848382, 0x89888786, 0x8d8c8b8a, 0x01008f8e), /* shl 14 (16 - 3)/shr2 */
Vector128.Create(0x86858483, 0x8a898887, 0x8e8d8c8b, 0x0201008f), /* shl 13 (16 - 4)/shr3 */
Vector128.Create(0x87868584, 0x8b8a8988, 0x8f8e8d8c, 0x03020100), /* shl 12 (16 - 4)/shr4 */
Vector128.Create(0x88878685, 0x8c8b8a89, 0x008f8e8d, 0x04030201), /* shl 11 (16 - 5)/shr5 */
Vector128.Create(0x89888786, 0x8d8c8b8a, 0x01008f8e, 0x05040302), /* shl 10 (16 - 6)/shr6 */
Vector128.Create(0x8a898887, 0x8e8d8c8b, 0x0201008f, 0x06050403), /* shl 9 (16 - 7)/shr7 */
Vector128.Create(0x8b8a8988, 0x8f8e8d8c, 0x03020100, 0x07060504), /* shl 8 (16 - 8)/shr8 */
Vector128.Create(0x8c8b8a89, 0x008f8e8d, 0x04030201, 0x08070605), /* shl 7 (16 - 9)/shr9 */
Vector128.Create(0x8d8c8b8a, 0x01008f8e, 0x05040302, 0x09080706), /* shl 6 (16 -10)/shr10*/
Vector128.Create(0x8e8d8c8b, 0x0201008f, 0x06050403, 0x0a090807), /* shl 5 (16 -11)/shr11*/
Vector128.Create(0x8f8e8d8c, 0x03020100, 0x07060504, 0x0b0a0908), /* shl 4 (16 -12)/shr12*/
Vector128.Create(0x84838281, 0x88878685, 0x8c8b8a89, 0x008f8e8d), /* shl 15 (16 - 1)/shr1 */
Vector128.Create(0x85848382, 0x89888786, 0x8d8c8b8a, 0x01008f8e), /* shl 14 (16 - 3)/shr2 */
Vector128.Create(0x86858483, 0x8a898887, 0x8e8d8c8b, 0x0201008f), /* shl 13 (16 - 4)/shr3 */
Vector128.Create(0x87868584, 0x8b8a8988, 0x8f8e8d8c, 0x03020100), /* shl 12 (16 - 4)/shr4 */
Vector128.Create(0x88878685, 0x8c8b8a89, 0x008f8e8d, 0x04030201), /* shl 11 (16 - 5)/shr5 */
Vector128.Create(0x89888786, 0x8d8c8b8a, 0x01008f8e, 0x05040302), /* shl 10 (16 - 6)/shr6 */
Vector128.Create(0x8a898887, 0x8e8d8c8b, 0x0201008f, 0x06050403), /* shl 9 (16 - 7)/shr7 */
Vector128.Create(0x8b8a8988, 0x8f8e8d8c, 0x03020100, 0x07060504), /* shl 8 (16 - 8)/shr8 */
Vector128.Create(0x8c8b8a89, 0x008f8e8d, 0x04030201, 0x08070605), /* shl 7 (16 - 9)/shr9 */
Vector128.Create(0x8d8c8b8a, 0x01008f8e, 0x05040302, 0x09080706), /* shl 6 (16 -10)/shr10*/
Vector128.Create(0x8e8d8c8b, 0x0201008f, 0x06050403, 0x0a090807), /* shl 5 (16 -11)/shr11*/
Vector128.Create(0x8f8e8d8c, 0x03020100, 0x07060504, 0x0b0a0908), /* shl 4 (16 -12)/shr12*/
Vector128.Create(0x008f8e8du, 0x04030201, 0x08070605, 0x0c0b0a09), /* shl 3 (16 -13)/shr13*/
Vector128.Create(0x01008f8eu, 0x05040302, 0x09080706, 0x0d0c0b0a), /* shl 2 (16 -14)/shr14*/
Vector128.Create(0x0201008fu, 0x06050403, 0x0a090807, 0x0e0d0c0b) /* shl 1 (16 -15)/shr15*/
@@ -85,7 +85,7 @@ internal static class Clmul
static void Fold4(ref Vector128<uint> xmmCRC0, ref Vector128<uint> xmmCRC1, ref Vector128<uint> xmmCRC2,
ref Vector128<uint> xmmCRC3)
{
Vector128<uint> xmmFold4 = Vector128.Create(0xc6e41596, 0x00000001, 0x54442bd4, 0x00000001);
var xmmFold4 = Vector128.Create(0xc6e41596, 0x00000001, 0x54442bd4, 0x00000001);
Vector128<uint> xTmp0 = xmmCRC0;
Vector128<uint> xTmp1 = xmmCRC1;
@@ -93,25 +93,25 @@ internal static class Clmul
Vector128<uint> xTmp3 = xmmCRC3;
xmmCRC0 = Pclmulqdq.CarrylessMultiply(xmmCRC0.AsUInt64(), xmmFold4.AsUInt64(), 0x01).AsUInt32();
xTmp0 = Pclmulqdq.CarrylessMultiply(xTmp0.AsUInt64(), xmmFold4.AsUInt64(), 0x10).AsUInt32();
xTmp0 = Pclmulqdq.CarrylessMultiply(xTmp0.AsUInt64(), xmmFold4.AsUInt64(), 0x10).AsUInt32();
Vector128<float> psCRC0 = xmmCRC0.AsSingle();
Vector128<float> psT0 = xTmp0.AsSingle();
Vector128<float> psRes0 = Sse.Xor(psCRC0, psT0);
xmmCRC1 = Pclmulqdq.CarrylessMultiply(xmmCRC1.AsUInt64(), xmmFold4.AsUInt64(), 0x01).AsUInt32();
xTmp1 = Pclmulqdq.CarrylessMultiply(xTmp1.AsUInt64(), xmmFold4.AsUInt64(), 0x10).AsUInt32();
xTmp1 = Pclmulqdq.CarrylessMultiply(xTmp1.AsUInt64(), xmmFold4.AsUInt64(), 0x10).AsUInt32();
Vector128<float> psCRC1 = xmmCRC1.AsSingle();
Vector128<float> psT1 = xTmp1.AsSingle();
Vector128<float> psRes1 = Sse.Xor(psCRC1, psT1);
xmmCRC2 = Pclmulqdq.CarrylessMultiply(xmmCRC2.AsUInt64(), xmmFold4.AsUInt64(), 0x01).AsUInt32();
xTmp2 = Pclmulqdq.CarrylessMultiply(xTmp2.AsUInt64(), xmmFold4.AsUInt64(), 0x10).AsUInt32();
xTmp2 = Pclmulqdq.CarrylessMultiply(xTmp2.AsUInt64(), xmmFold4.AsUInt64(), 0x10).AsUInt32();
Vector128<float> psCRC2 = xmmCRC2.AsSingle();
Vector128<float> psT2 = xTmp2.AsSingle();
Vector128<float> psRes2 = Sse.Xor(psCRC2, psT2);
xmmCRC3 = Pclmulqdq.CarrylessMultiply(xmmCRC3.AsUInt64(), xmmFold4.AsUInt64(), 0x01).AsUInt32();
xTmp3 = Pclmulqdq.CarrylessMultiply(xTmp3.AsUInt64(), xmmFold4.AsUInt64(), 0x10).AsUInt32();
xTmp3 = Pclmulqdq.CarrylessMultiply(xTmp3.AsUInt64(), xmmFold4.AsUInt64(), 0x10).AsUInt32();
Vector128<float> psCRC3 = xmmCRC3.AsSingle();
Vector128<float> psT3 = xTmp3.AsSingle();
Vector128<float> psRes3 = Sse.Xor(psCRC3, psT3);
@@ -130,13 +130,13 @@ internal static class Clmul
Vector128<uint> xmmCRC1 = Vector128<uint>.Zero;
Vector128<uint> xmmCRC2 = Vector128<uint>.Zero;
Vector128<uint> xmmCRC3 = Vector128<uint>.Zero;
int bufPos = 0;
var bufPos = 0;
bool first = true;
var first = true;
/* fold 512 to 32 step variable declarations for ISO-C90 compat. */
Vector128<uint> xmmMask = Vector128.Create(0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000);
Vector128<uint> xmmMask2 = Vector128.Create(0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
var xmmMask = Vector128.Create(0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000);
var xmmMask2 = Vector128.Create(0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
while((len -= 64) >= 0)
{
@@ -155,10 +155,10 @@ internal static class Clmul
bufPos += 16;
Vector128<uint> xmmT3 = Vector128.Create(BitConverter.ToUInt32(src, bufPos),
BitConverter.ToUInt32(src, bufPos + 4),
BitConverter.ToUInt32(src, bufPos + 8),
BitConverter.ToUInt32(src, bufPos + 12));
var xmmT3 = Vector128.Create(BitConverter.ToUInt32(src, bufPos),
BitConverter.ToUInt32(src, bufPos + 4),
BitConverter.ToUInt32(src, bufPos + 8),
BitConverter.ToUInt32(src, bufPos + 12));
bufPos += 16;
@@ -181,7 +181,7 @@ internal static class Clmul
/*
* k1
*/
Vector128<uint> crcFold = Vector128.Create(_crcK[0], _crcK[1], _crcK[2], _crcK[3]);
var crcFold = Vector128.Create(_crcK[0], _crcK[1], _crcK[2], _crcK[3]);
Vector128<uint> xTmp0 = Pclmulqdq.CarrylessMultiply(xmmCRC0.AsUInt64(), crcFold.AsUInt64(), 0x10).AsUInt32();

42
Aaru6.Checksums/CRC32/vmull.cs
View File

@@ -54,7 +54,7 @@ using System.Runtime.Intrinsics.Arm;
namespace Aaru6.Checksums.CRC32;
internal static class Vmull
static class Vmull
{
static readonly uint[] _crcK =
{
@@ -64,18 +64,18 @@ internal static class Vmull
static readonly Vector128<uint>[] _pshufbShfTable =
{
Vector128.Create(0x84838281, 0x88878685, 0x8c8b8a89, 0x008f8e8d), /* shl 15 (16 - 1)/shr1 */
Vector128.Create(0x85848382, 0x89888786, 0x8d8c8b8a, 0x01008f8e), /* shl 14 (16 - 3)/shr2 */
Vector128.Create(0x86858483, 0x8a898887, 0x8e8d8c8b, 0x0201008f), /* shl 13 (16 - 4)/shr3 */
Vector128.Create(0x87868584, 0x8b8a8988, 0x8f8e8d8c, 0x03020100), /* shl 12 (16 - 4)/shr4 */
Vector128.Create(0x88878685, 0x8c8b8a89, 0x008f8e8d, 0x04030201), /* shl 11 (16 - 5)/shr5 */
Vector128.Create(0x89888786, 0x8d8c8b8a, 0x01008f8e, 0x05040302), /* shl 10 (16 - 6)/shr6 */
Vector128.Create(0x8a898887, 0x8e8d8c8b, 0x0201008f, 0x06050403), /* shl 9 (16 - 7)/shr7 */
Vector128.Create(0x8b8a8988, 0x8f8e8d8c, 0x03020100, 0x07060504), /* shl 8 (16 - 8)/shr8 */
Vector128.Create(0x8c8b8a89, 0x008f8e8d, 0x04030201, 0x08070605), /* shl 7 (16 - 9)/shr9 */
Vector128.Create(0x8d8c8b8a, 0x01008f8e, 0x05040302, 0x09080706), /* shl 6 (16 -10)/shr10*/
Vector128.Create(0x8e8d8c8b, 0x0201008f, 0x06050403, 0x0a090807), /* shl 5 (16 -11)/shr11*/
Vector128.Create(0x8f8e8d8c, 0x03020100, 0x07060504, 0x0b0a0908), /* shl 4 (16 -12)/shr12*/
Vector128.Create(0x84838281, 0x88878685, 0x8c8b8a89, 0x008f8e8d), /* shl 15 (16 - 1)/shr1 */
Vector128.Create(0x85848382, 0x89888786, 0x8d8c8b8a, 0x01008f8e), /* shl 14 (16 - 3)/shr2 */
Vector128.Create(0x86858483, 0x8a898887, 0x8e8d8c8b, 0x0201008f), /* shl 13 (16 - 4)/shr3 */
Vector128.Create(0x87868584, 0x8b8a8988, 0x8f8e8d8c, 0x03020100), /* shl 12 (16 - 4)/shr4 */
Vector128.Create(0x88878685, 0x8c8b8a89, 0x008f8e8d, 0x04030201), /* shl 11 (16 - 5)/shr5 */
Vector128.Create(0x89888786, 0x8d8c8b8a, 0x01008f8e, 0x05040302), /* shl 10 (16 - 6)/shr6 */
Vector128.Create(0x8a898887, 0x8e8d8c8b, 0x0201008f, 0x06050403), /* shl 9 (16 - 7)/shr7 */
Vector128.Create(0x8b8a8988, 0x8f8e8d8c, 0x03020100, 0x07060504), /* shl 8 (16 - 8)/shr8 */
Vector128.Create(0x8c8b8a89, 0x008f8e8d, 0x04030201, 0x08070605), /* shl 7 (16 - 9)/shr9 */
Vector128.Create(0x8d8c8b8a, 0x01008f8e, 0x05040302, 0x09080706), /* shl 6 (16 -10)/shr10*/
Vector128.Create(0x8e8d8c8b, 0x0201008f, 0x06050403, 0x0a090807), /* shl 5 (16 -11)/shr11*/
Vector128.Create(0x8f8e8d8c, 0x03020100, 0x07060504, 0x0b0a0908), /* shl 4 (16 -12)/shr12*/
Vector128.Create(0x008f8e8du, 0x04030201, 0x08070605, 0x0c0b0a09), /* shl 3 (16 -13)/shr13*/
Vector128.Create(0x01008f8eu, 0x05040302, 0x09080706, 0x0d0c0b0a), /* shl 2 (16 -14)/shr14*/
Vector128.Create(0x0201008fu, 0x06050403, 0x0a090807, 0x0e0d0c0b) /* shl 1 (16 -15)/shr15*/
@@ -84,9 +84,7 @@ internal static class Vmull
static Vector128<ulong> vmull_p64(Vector64<ulong> a, Vector64<ulong> b)
{
if(Aes.IsSupported)
{
return Aes.PolynomialMultiplyWideningLower(a, b);
}
// Masks
Vector128<byte> k4832 = Vector128.Create(Vector64.Create(0x0000fffffffffffful),
@@ -199,8 +197,8 @@ internal static class Vmull
// Accumulate the products
Vector128<byte> cross1 = AdvSimd.Xor(t0Shift, t1Shift);
Vector128<byte> cross2 = AdvSimd.Xor(t2Shift, t3Shift);
Vector128<byte> mix = AdvSimd.Xor(d, cross1);
Vector128<byte> r = AdvSimd.Xor(mix, cross2);
Vector128<byte> mix = AdvSimd.Xor(d, cross1);
Vector128<byte> r = AdvSimd.Xor(mix, cross2);
return r.AsUInt64();
}
@@ -269,13 +267,13 @@ internal static class Vmull
Vector128<ulong> qCRC1 = Vector128<ulong>.Zero;
Vector128<ulong> qCRC2 = Vector128<ulong>.Zero;
Vector128<ulong> qCRC3 = Vector128<ulong>.Zero;
int bufPos = 0;
var bufPos = 0;
bool first = true;
var first = true;
/* fold 512 to 32 step variable declarations for ISO-C90 compat. */
Vector128<uint> qMask = Vector128.Create(0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000);
Vector128<uint> qMask2 = Vector128.Create(0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
var qMask = Vector128.Create(0xFFFFFFFF, 0xFFFFFFFF, 0x00000000, 0x00000000);
var qMask2 = Vector128.Create(0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
uint crc;
Vector128<ulong> xTmp0;
@@ -356,7 +354,7 @@ internal static class Vmull
Vector128<byte> qCRC0B = qCRC0.AsByte();
qCRC0 = Vector128.Create(AdvSimd.Extract(qCRC0B, 8), AdvSimd.Extract(qCRC0B, 9), AdvSimd.Extract(qCRC0B, 10),
qCRC0 = Vector128.Create(AdvSimd.Extract(qCRC0B, 8), AdvSimd.Extract(qCRC0B, 9), AdvSimd.Extract(qCRC0B, 10),
AdvSimd.Extract(qCRC0B, 11), AdvSimd.Extract(qCRC0B, 12), AdvSimd.Extract(qCRC0B, 13),
AdvSimd.Extract(qCRC0B, 14), AdvSimd.Extract(qCRC0B, 15), 0, 0, 0, 0, 0, 0, 0, 0).
AsUInt64();