发布 VectorTraits v2.0(支持 x86的Sse系列指令集等)

  • 发布 VectorTraits v2.0(支持 x86的Sse系列指令集等)已关闭评论
  • 65 次浏览
  • A+
所属分类:.NET技术
摘要

VectorTraits已更新至 v2.0版。支持 x86的Sse系列指令集; 为 Vector128/Vector256 补充全部的向量方法; 还提供了 浮点数判断(YIsNaN, YIsinfinity)、符号判断(YCopySign, YSign) 等原创的向量方法。

目录

VectorTraits已更新至 v2.0版。支持 x86的Sse系列指令集; 为 Vector128/Vector256 补充全部的向量方法; 还提供了 浮点数判断(YIsNaN, YIsinfinity)、符号判断(YCopySign, YSign) 等原创的向量方法。

变更日志如下。

  • Major: Support for the X86 Sse family instruction set; supplement all vector methods for Vector128/Vector256; also provides innovative vector methods such as check floating number (YIsNaN, YIsInfinity), sign (YCopySign, YSign) (支持 x86的Sse系列指令集; 为 Vector128/Vector256 补充全部的向量方法; 还提供了 浮点数判断(YIsNaN, YIsinfinity)、符号判断(YCopySign, YSign) 等原创的向量方法).
  • Provides the CPU model info (提供CPU型号信息). VectorEnvironment adds members: CpuModelName, CpuFlags, CpuDetectionCommand, CpuDetectionException, CpuDetectionResult .
  • Provides information about the supported instruction set (提供所支持的指令集信息). e.g. VectorEnvironment.SupportedInstructionSets, IBaseTraits.UsedInstructionSets
  • Supplement all vector methods for Vector128/Vector256 (为 Vector128/Vector256 补充全部的向量方法): Dot, Equals, EqualsAll, EqualsAny, GreaterThanAll, GreaterThanAny, GreaterThanOrEqual, GreaterThanOrEqualAll, GreaterThanOrEqualAny, LessThanAll, LessThanAny, LessThanOrEqual, LessThanOrEqualAll, LessThanOrEqualAny, Sqrt .
  • Provides the vector methods of bitwise operations (提供位运算的向量方法): YBitToByte, YBitToInt16, YBitToInt32, YBitToInt64, YOrNot .
  • Provides the vector methods of check floating number (提供浮点数判断的向量方法): YIsEvenInteger, YIsFinite, YIsInfinity, YIsInfinityOrNaN, YIsInteger, YIsNaN, YIsNegative, YIsNegativeZero, YIsNegativeInfinity, YIsNormal, YIsNotNaN, YIsOddInteger, YIsPositive, YIsPositiveInfinity, YIsSubnormal, YIsZero, YIsZeroOrSubnormal.
  • Provides the vector methods of check sign (提供符号判断的向量方法): YCopySign, YSign, YSignFloat .
  • Provides the vector methods of clamp (提供限制的向量方法): YMaxNumber, YMinNumber .
  • Provides the vector methods of compare (提供比较的向量方法): YIsAllTrue, YIsAnyTrue, YIsNotEquals.
  • VectorTraits.dll: Add TargetFrameworks net8.0 and netstandard2.1 (增加目标框架 net8.0netstandard2.1 ).
  • Provides arrays of fixed length (提供固定长度的数组). e.g. FixedArray2, FixedArray4 ...
  • BitMath changed from static to abstract class. Add namespace Zyl.VectorTraits.Numerics, add some math functions (BitMath从静态类改为抽象类. 新增名称空间 Zyl.VectorTraits.Numerics, 增加一些的数学函数).
  • Benchmark programs add command line parameter (基准测试程序增加命令行参数): -accelerated0 -allowFakeBenchmark -cpuDetection0 -fixedVector0 -test0
  • Add tool program (增加工具程序): UpdateBenchmarkResults.
  • Optimized hardware acceleration of ExtractMostSignificantBits methods in the Arm architecture (优化ExtractMostSignificantBits方法在Arm架构的硬件加速). For 8~64 bit types.

完整列表: ChangeLog

支持 x86的Sse系列指令集

本库已经支持了x86架构的Sse系列指令集。既:Sse, Sse2, Sse3, Ssse3, Sse41, Sse42。
在X86架构上使用128位向量(Vector128、128位时的Vector)时,现在能充分获得硬件加速。

为 Vector128/Vector256 补充全部的向量方法

相关日志:

  • Supplement all vector methods for Vector128/Vector256 (为 Vector128/Vector256 补充全部的向量方法): Dot, Equals, EqualsAll, EqualsAny, GreaterThanAll, GreaterThanAny, GreaterThanOrEqual, GreaterThanOrEqualAll, GreaterThanOrEqualAny, LessThanAll, LessThanAny, LessThanOrEqual, LessThanOrEqualAll, LessThanOrEqualAny, Sqrt .

对于固定长度的向量类型(Vector128/Vector256),在v1.0版时只是提供了常用的向量方法。
而现在v2.0版,已经补充全部的向量方法。已经覆盖了 .NET 7.0中所有运算类的向量方法。例如 Dot, Equals, EqualsAll, EqualsAny 等.

这些向量方法的清单如下。

  • Dot①: Computes the dot product of two vectors (计算两个向量的点积).
    Mnemonic: rt := left[0]*right[0] + left[1]*right[1] + left[2]*right[2] + ... + left[Count-1]*right[Count-1] .
  • Equals: Compares two vectors to determine if they are equal on a per-element basis (比较两个向量,确定它们每个元素是否相等).
    Mnemonic: rt[i] := to_mask(left[i] == right[i]).
  • EqualsAll: Compares two vectors to determine if all elements are equal (比较两个向量以判定所有元素是否相等).
    Mnemonic: rt := (left[0] == right[0]) && (left[1] == right[1]) && ... && (left[Count-1] == right[Count-1]).
  • EqualsAny: Compares two vectors to determine if any elements are equal (比较两个向量以判定任一元素是否相等).
    Mnemonic: rt := (left[0] == right[0]) || (left[1] == right[1]) || ... || (left[Count-1] == right[Count-1]).
  • GreaterThanAll: Compares two vectors to determine if all elements are greater (比较两个向量以判定所有元素是否大于).
    Mnemonic: rt := (left[0] > right[0]) && (left[1] > right[1]) && ... && (left[Count-1] > right[Count-1]).
  • GreaterThanAny: Compares two vectors to determine if any elements are greater (比较两个向量以判定任一元素是否大于).
    Mnemonic: rt := (left[0] > right[0]) || (left[1] > right[1]) || ... || (left[Count-1] > right[Count-1]).
  • GreaterThanOrEqual: Compares two vectors to determine which is greater or equal on a per-element basis (比较两个向量,在每个元素的基础上确定哪个更大或等于).
    Mnemonic: rt[i] := to_mask(left[i] >= right[i]).
  • GreaterThanOrEqualAll: Compares two vectors to determine if all elements are greater or equal (比较两个向量以判定所有元素是否大于或等于).
    Mnemonic: rt := (left[0] >= right[0]) && (left[1] >= right[1]) && ... && (left[Count-1] >= right[Count-1]).
  • GreaterThanOrEqualAny: Compares two vectors to determine if any elements are greater or equal (比较两个向量以判定任一元素是否大于或等于).
    Mnemonic: rt := (left[0] >= right[0]) || (left[1] >= right[1]) || ... || (left[Count-1] >= right[Count-1]).
  • LessThanAll: Compares two vectors to determine if all elements are less (比较两个向量以判定所有元素是否小于).
    Mnemonic: rt := (left[0] < right[0]) && (left[1] < right[1]) && ... && (left[Count-1] < right[Count-1]).
  • LessThanAny: Compares two vectors to determine if any elements are less (比较两个向量以判定任一元素是否小于).
    Mnemonic: rt := (left[0] < right[0]) || (left[1] < right[1]) || ... || (left[Count-1] < right[Count-1]).
  • LessThanOrEqual: Compares two vectors to determine which is less or equal on a per-element basis (比较两个向量,在每个元素的基础上确定哪个更小或等于).
    Mnemonic: rt[i] := to_mask(left[i] <= right[i]).
  • LessThanOrEqualAll: Compares two vectors to determine if all elements are less or equal (比较两个向量以判定所有元素是否小于或等于).
    Mnemonic: rt := (left[0] <= right[0]) && (left[1] <= right[1]) && ... && (left[Count-1] <= right[Count-1]).
  • LessThanOrEqualAny: Compares two vectors to determine if any elements are less or equal (比较两个向量以判定任一元素是否小于或等于).
    Mnemonic: rt := (left[0] <= right[0]) || (left[1] <= right[1]) || ... || (left[Count-1] <= right[Count-1]).
  • Sqrt①: Computes the square root of a vector on a per-element basis (计算向量中每个元素的平方根).
    Mnemonic: rt[i] := sqrt(vector[i]) = pow(vector[i], 1.0/2). When x is less than 0, floating-point types return NaN, integer types return 0.

注:①表示这些方法也适用于不定长度的向量类型Vector. 因发现Vector的这些方法中, 部分元素类型没能提供硬件加速。

提供CPU型号信息

相关日志:

  • Provides the CPU model info (提供CPU型号信息). VectorEnvironment adds members: CpuModelName, CpuFlags, CpuDetectionCommand, CpuDetectionException, CpuDetectionResult .

在使用向量方法时,很多时候我们想知道CPU型号信息,但BCL未提供办法。于是v2.0版提供了查看CPU型号信息的属性,且支持 Windows/Linux/Mac 这3大主流操作系统。

VectorEnvironment类提供了这些成员。

  • CpuModelName: CPU型号名。例如 Intel(R) Core(TM) i5-8250U CPU @ 1.60GHz, Neoverse-N1
  • CpuFlags: CPU特征标志。目前仅Linux操作系统上有值,它就是lscpu命令所返回的flags字段。例如Neoverse-N1的该属性为 fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp ssbs.
  • CpuDetectionCommand: CPU检测所用命令。不同操作系统(Windows/Linux/Mac)上,所用的命令不同。例如wmic
  • CpuDetectionException: CPU检测时的最新异常。
  • CpuDetectionResult: CPU检测的原始返回值. 它与 CpuDetectionCommand 有关。

一般情况下直接使用VectorEnvironment的属性就行。范例代码如下。

writer.WriteLine("CpuModelName: {0}", VectorEnvironment.CpuModelName); writer.WriteLine("CpuFlags: {0}", VectorEnvironment.CpuFlags); writer.WriteLine("CpuDetectionException: {0}", VectorEnvironment.CpuDetectionException); writer.WriteLine("CpuDetectionCommand: {0}", VectorEnvironment.CpuDetectionCommand); writer.Write("CpuDetectionResult:t"); VectorTextUtil.WriteLines(writer, VectorEnvironment.CpuDetectionResult); writer.WriteLine(); 

只有有一种情况需要特殊处理。就是在 .NET Framework 4.5等低版本(低于.NET Framework 4.6.1)程序里使用本库时,此时使用的是 netstandard1.1的库,而 netstandard1.1未提供 System.Diagnostics.Process 类,会导致无法获取CPU信息的情况。
解决办法是是提前将 System.Diagnostics.Process 的类型赋值给一个属性。随后本库便可通过反射来使用该类,从而能正常获取CPU信息了。范例代码如下。

#if NETSTANDARD1_3_OR_GREATER || NETCOREAPP2_0_OR_GREATER || NET461_OR_GREATER     // No need to set up `ProcessUtil.TypeOfProcess` properties.  #else     Zyl.VectorTraits.Impl.Util.ProcessUtil.TypeOfProcess = typeof(System.Diagnostics.Process); #endif 

结果范例1: X86 CPU on Windows

CpuModelName: Intel(R) Core(TM) i5-8250U CPU @ 1.60GHz CpuFlags:  CpuDetectionException: CpuDetectionCommand: wmic   AddressWidth=64 Architecture=9 AssetTag=None Availability=3 Caption=Intel64 Family 6 Model 142 Stepping 10 Characteristics=252 ConfigManagerErrorCode= ConfigManagerUserConfig= CpuStatus=1 CreationClassName=Win32_Processor CurrentClockSpeed=1600 CurrentVoltage=11 DataWidth=64 Description=Intel64 Family 6 Model 142 Stepping 10 DeviceID=CPU0 ErrorCleared= ErrorDescription= ExtClock=100 Family=205 InstallDate= L2CacheSize=1024 L2CacheSpeed= L3CacheSize=6144 L3CacheSpeed=0 LastErrorCode= Level=6 LoadPercentage=7 Manufacturer=GenuineIntel MaxClockSpeed=1800 Name=Intel(R) Core(TM) i5-8250U CPU @ 1.60GHz NumberOfCores=4 NumberOfEnabledCore=4 NumberOfLogicalProcessors=8 OtherFamilyDescription= PartNumber=None PNPDeviceID= PowerManagementCapabilities= PowerManagementSupported=FALSE ProcessorId=BFEBFBFF000806EA ProcessorType=3 Revision= Role=CPU SecondLevelAddressTranslationExtensions=TRUE SerialNumber=None SocketDesignation=U3E1 Status=OK StatusInfo=3 Stepping= SystemCreationClassName=Win32_ComputerSystem SystemName=THINK1621 ThreadCount=8 UniqueId= UpgradeMethod=51 Version= VirtualizationFirmwareEnabled=TRUE VMMonitorModeExtensions=TRUE VoltageCaps= 

结果范例2: Arm CPU on Linux

CpuModelName: Neoverse-N1 CpuFlags: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp ssbs CpuDetectionException:  CpuDetectionCommand: lscpu Architecture:                    aarch64 CPU op-mode(s):                  32-bit, 64-bit Byte Order:                      Little Endian CPU(s):                          2 On-line CPU(s) list:             0,1 Vendor ID:                       ARM Model name:                      Neoverse-N1 Model:                           1 Thread(s) per core:              1 Core(s) per socket:              2 Socket(s):                       1 Stepping:                        r3p1 BogoMIPS:                        243.75 Flags:                           fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp ssbs L1d cache:                       128 KiB (2 instances) L1i cache:                       128 KiB (2 instances) L2 cache:                        2 MiB (2 instances) L3 cache:                        32 MiB (1 instance) NUMA node(s):                    1 NUMA node0 CPU(s):               0,1 Vulnerability Itlb multihit:     Not affected Vulnerability L1tf:              Not affected Vulnerability Mds:               Not affected Vulnerability Meltdown:          Not affected Vulnerability Mmio stale data:   Not affected Vulnerability Retbleed:          Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1:        Mitigation; __user pointer sanitization Vulnerability Spectre v2:        Mitigation; CSV2, BHB Vulnerability Srbds:             Not affected Vulnerability Tsx async abort:   Not affected 

结果范例3: Arm CPU on Mac OS

CpuModelName: Apple M2 CpuFlags:  CpuDetectionException:  CpuDetectionCommand: sysctl kern.sched_rt_avoid_cpu0: 0 kern.cpu_checkin_interval: 4000 hw.ncpu: 8 hw.activecpu: 8 hw.perflevel0.physicalcpu: 4 hw.perflevel0.physicalcpu_max: 4 hw.perflevel0.logicalcpu: 4 hw.perflevel0.logicalcpu_max: 4 hw.perflevel0.cpusperl2: 4 hw.perflevel1.physicalcpu: 4 hw.perflevel1.physicalcpu_max: 4 hw.perflevel1.logicalcpu: 4 hw.perflevel1.logicalcpu_max: 4 hw.perflevel1.cpusperl2: 4 hw.physicalcpu: 8 hw.physicalcpu_max: 8 hw.logicalcpu: 8 hw.logicalcpu_max: 8 hw.cputype: 16777228 hw.cpusubtype: 2 hw.cpu64bit_capable: 1 hw.cpufamily: -634136515 hw.cpusubfamily: 2 machdep.cpu.cores_per_package: 8 machdep.cpu.core_count: 8 machdep.cpu.logical_per_package: 8 machdep.cpu.thread_count: 8 machdep.cpu.brand_string: Apple M2 

提供所支持的指令集信息

相关日志:

  • Provides information about the supported instruction set (提供所支持的指令集信息). e.g. VectorEnvironment.SupportedInstructionSets, IBaseTraits.UsedInstructionSets

有了这些属性后,便能查看所支持的指令集了。

  • VectorEnvironment类提供了SupportedInstructionSets属性。用于获取本机所支持的指令。
  • 许多类型(Vectors、Vector128s、Vector256s与IBaseTraits的派生类)提供了UsedInstructionSets属性。用于获取该向量类型所使用的指令。

注:多个指令集时,会使用分隔符。分隔符是逗号 ','。

结果范例1: X86 CPU on Windows

VectorEnvironment.CpuModelName:	Intel(R) Core(TM) i5-8250U CPU @ 1.60GHz VectorEnvironment.SupportedInstructionSets:	Aes, Avx, Avx2, Bmi1, Bmi2, Fma, Lzcnt, Pclmulqdq, Popcnt, Sse, Sse2, Sse3, Ssse3, Sse41, Sse42, X86Base Vector128s.Instance:	WVectorTraits128Avx2	// Sse, Sse2, Sse3, Ssse3, Sse41, Sse42, Avx, Avx2 Vector256s.Instance:	WVectorTraits256Avx2	// Avx, Avx2, Sse, Sse2 Vectors.Instance:	VectorTraits256Avx2	// Avx, Avx2, Sse, Sse2 

注:“//” 后面的就是UsedInstructionSets属性的值。

结果范例2: Arm CPU on Linux

VectorEnvironment.CpuModelName:	Neoverse-N1 VectorEnvironment.CpuFlags:	fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp ssbs VectorEnvironment.SupportedInstructionSets:	AdvSimd, Aes, ArmBase, Crc32, Sha1, Sha256 Vector128s.Instance:	WVectorTraits128AdvSimdB64	// AdvSimd Vectors.Instance:	VectorTraits128AdvSimdB64	// AdvSimd 

结果范例3: Arm CPU on Mac OS

VectorEnvironment.CpuModelName:	Apple M2 VectorEnvironment.SupportedInstructionSets:	AdvSimd, Aes, ArmBase, Crc32, Dp, Rdm, Sha1, Sha256 Vector128s.Instance:	WVectorTraits128AdvSimdB64	// AdvSimd Vectors.Instance:	VectorTraits128AdvSimdB64	// AdvSimd 

新增了向量方法

相关日志:

  • Provides the vector methods of bitwise operations (提供位运算的向量方法): YBitToByte, YBitToInt16, YBitToInt32, YBitToInt64, YOrNot .
  • Provides the vector methods of check floating number (提供浮点数判断的向量方法): YIsEvenInteger, YIsFinite, YIsInfinity, YIsInfinityOrNaN, YIsInteger, YIsNaN, YIsNegative, YIsNegativeZero, YIsNegativeInfinity, YIsNormal, YIsNotNaN, YIsOddInteger, YIsPositive, YIsPositiveInfinity, YIsSubnormal, YIsZero, YIsZeroOrSubnormal.
  • Provides the vector methods of check sign (提供符号判断的向量方法): YCopySign, YSign, YSignFloat .
  • Provides the vector methods of clamp (提供限制的向量方法): YMaxNumber, YMinNumber .
  • Provides the vector methods of compare (提供比较的向量方法): YIsAllTrue, YIsAnyTrue, YIsNotEquals.

位运算的向量方法

  • YBitToByte: Converts binary bits to each element of the Byte vector (将各个二进制位转换为 Byte 向量的每个元素). Bit 0 meaning is 0, bit 1 meaning is 1 for all bits (byte.MaxValue).
    Mnemonic: rt[i] := to_mask(0 != ((value>>i)&1)) .
  • YBitToInt16: Converts binary bits to each element of the Int16 vector (将各个二进制位转换为 Int16 向量的每个元素). Bit 0 meaning is 0, bit 1 meaning is 1 for all bits (-1).
    Mnemonic: rt[i] := to_mask(0 != ((value>>i)&1)) .
  • YBitToInt32: Converts binary bits to each element of the Int32 vector (将各个二进制位转换为 Int32 向量的每个元素). Bit 0 meaning is 0, bit 1 meaning is 1 for all bits (-1).
    Mnemonic: rt[i] := to_mask(0 != ((value>>i)&1)) .
  • YBitToInt64: Converts binary bits to each element of the Int64 vector (将各个二进制位转换为 Int64 向量的每个元素). Bit 0 meaning is 0, bit 1 meaning is 1 for all bits (-1).
    Mnemonic: rt[i] := to_mask(0 != ((value>>i)&1)) .

注:YBitToByte 等方法,是 ExtractMostSignificantBits 方法的逆运算。

浮点数判断的向量方法

  • YIsEvenInteger: Determines if a element represents an even integral number (确定元素是否为偶数整数).
    Mnemonic: rt[i] := to_mask(isEvenInteger(value[i])) .
  • YIsFinite: Determines if a element is finite. It contains zero, subnormal, and normal. It does not contain infinity, NaN (确定元素是否为有限值. 它包含 零、次正规数、正规数. 它不含无穷大、非数).
    Mnemonic: rt[i] := to_mask(isFinite(value[i])) .
  • YIsInfinity: Determines if a element is infinite (确定元素是否为无穷大).
    Mnemonic: rt[i] := to_mask(isInfinity(value[i])) .
  • YIsInfinityOrNaN: Determines if a element is infinite or NaN (确定元素是否为无穷大或非数).
    Mnemonic: rt[i] := to_mask(isInfinityOrNaN(value[i])) .
  • YIsInteger: Determines if a element represents an integral number (确定元素是否为整数).
    Mnemonic: rt[i] := to_mask(isInteger(value[i])) .
  • YIsNaN: Determines if a element is NaN (确定元素是否为非数).
    Mnemonic: rt[i] := to_mask(isNaN(value[i])) .
  • YIsNegative: Determines if a element represents a negative number or negative zero (确定元素是否为负数或负零).
    Mnemonic: rt[i] := to_mask(isNegative(value[i])) = to_mask((value[i]<0) || isNegativeZero(value[i])) .
  • YIsNegativeInfinity: Determines if a element is negative infinity (确定元素是否为负无穷大).
    Mnemonic: rt[i] := to_mask(isNegativeInfinity(value[i])) .
  • YIsNegativeZero: Determines if a element represents a negative zero (确定元素是否为负零).
    Mnemonic: rt[i] := to_mask(isNegativeZero(value[i])) .
  • YIsNormal: Determines if a element is normal (确定元素是否为正规数).
    Mnemonic: rt[i] := to_mask(isNormal(value[i])) .
  • YIsNotNaN: Determines if a element is not NaN (确定元素是否不为非数).
    Mnemonic: rt[i] := to_mask(isNotNaN(value[i])) = to_mask(!isNaN(value[i])) .
  • YIsOddInteger: Determines if a element represents an odd integral number (确定元素是否为奇数整数).
    Mnemonic: rt[i] := to_mask(isOddInteger(value[i])) .
  • YIsPositive: Determines if a element represents zero or a positive number (确定元素是否为零或正数).
    Mnemonic: rt[i] := to_mask(isPositive(value[i])) .
  • YIsPositiveInfinity: Determines if a element is positive infinity (确定元素是否为正无穷大).
    Mnemonic: rt[i] := to_mask(isPositiveInfinity(value[i])) .
  • YIsSubnormal: Determines if a element is subnormal (确定元素是否为次正规数).
    Mnemonic: rt[i] := to_mask(isSubnormal(value[i])) .
  • YIsZero: Determines if a element is zero (确定元素是否为零).
    Mnemonic: rt[i] := to_mask(0==value[i]) .
  • YIsZeroOrSubnormal: Determines if a element is zero or subnormal (确定元素是否为零或次正规数).
    Mnemonic: rt[i] := to_mask(isZeroOrSubnormal(value[i])) .

符号判断的向量方法

  • YCopySign: Copies the sign of a value to the sign of another value (将一个值的符号复制到另一个值).
    Mnemonic: rt[i] := copySign(value[i], sign[i]).
  • YSign: Determine the sign of each element (判断各个元素的符号). 其元素取决于value的对应元素的符号情况: 值为正数时返回1, 值为0或NaN时返回0, 值为负数时返回-1.
    Mnemonic: rt[i] := sign(value[i]).
  • YSignFloat: Determine the sign of each element and returns a floating point number (判断各个元素的符号并返回浮点数). 其元素取决于value的对应元素的符号情况: 值为正数时返回1, 值为0时返回0, 值为负数时返回-1, 值为NaN时返回NaN.
    Mnemonic: rt[i] := signFloat(value[i]).

限制的向量方法

  • YMaxNumber: Computes the maximum number of two vectors on a per-element basis (在每个元素的基础上计算两个向量的最大数值). The maxNumber method matches the IEEE 754:2019 maximumNumber function. This requires NaN inputs to not be propagated back to the caller and for -0.0 to be treated as less than +0.0 (maxNumber方法与 IEEE 754:2019 maximumNumber 函数匹配。 这要求 NaN 输入不传播回调用方,且 -0.0 被视为小于 +0.0).
    Mnemonic: `rt[i] := maxNumber(left[i], right[i]).
  • YMinNumber: Computes the minimum number of two vectors on a per-element basis (在每个元素的基础上计算两个向量的最小数值). The minNumber method matches the IEEE 754:2019 minimumNumber function. This requires NaN inputs to not be propagated back to the caller and for -0.0 to be treated as less than +0.0 (minNumber方法与 IEEE 754:2019 minimumNumber 函数匹配。 这要求 NaN 输入不传播回调用方,且 -0.0 被视为小于 +0.0).
    Mnemonic: `rt[i] := minNumber(left[i], right[i]).

比较的向量方法

  • YIsAllTrue: Checks if all elements of the vector is true (检查向量中所有元素是不是都为true).
    Mnemonic: rt := value[0] && value[1] && value[2] && ... && value[Count-1] . The element of value must be 0 or AllBitsSet (Signed integer value -1).
  • YIsAnyTrue: Checks if any elements of the vector is true (检查向量中任一元素是不是为true).
    Mnemonic: rt := value[0] || value[1] || value[2] || ... || value[Count-1] . The element of value must be 0 or AllBitsSet (Signed integer value -1).
  • YIsNotEquals: Compares two vectors to determine if they are not equal on a per-element basis (比较两个向量,确定它们每个元素是否不相等).
    Mnemonic: rt[i] := to_mask(left[i] != right[i]).

增加目标框架 net8.0netstandard2.1

相关日志:

  • VectorTraits.dll: Add TargetFrameworks net8.0 and netstandard2.1 (增加目标框架 net8.0netstandard2.1 ).

本库增加目标框架—— net8.0netstandard2.1。能更好的利用这些目标框架的一些新的方法来优化性能。
对于net8.0 新增的512位向量(Vector512)与Avx512指令集,本库的3.0版将支持它们。

提供固定长度的数组

相关日志:

  • Provides arrays of fixed length (提供固定长度的数组). e.g. FixedArray2, FixedArray4 ...

这一项功能主要是给本库使用的。Unsafe.Add的地址计算,存在写法繁琐,很难利用“寄存器相对寻址”、“相对基址变址寻址”指令问题。
使用“固定长度的数组”后,能解决这些问题,且具有编译时范围检查。
“固定长度的数组”的原理比较简单,就是利用值类型的结构体的一些特点。将数组下标寻址,换成访问结构体的成员。当数组下标是常数时,使用“固定长度的数组”会更方便。

“固定长度的数组”的范例

例如下面是一个计算float的256位向量相等(Vector256.Equals<float>)的函数。

public static Vector256<float> Equals_Basic(Vector256<float> left, Vector256<float> right) {     UnsafeUtil.SkipInit(out Vector256<float> rt);     ref int prt = ref Unsafe.As<Vector256<float>, int>(ref rt);     ref float pleft = ref Unsafe.As<Vector256<float>, float>(ref left);     ref float pright = ref Unsafe.As<Vector256<float>, float>(ref right);     prt = BitMath.ToInt32Mask(pleft == pright);     Unsafe.Add(ref prt, 1) = BitMath.ToInt32Mask(Unsafe.Add(ref pleft, 1) == Unsafe.Add(ref pright, 1));     Unsafe.Add(ref prt, 2) = BitMath.ToInt32Mask(Unsafe.Add(ref pleft, 2) == Unsafe.Add(ref pright, 2));     Unsafe.Add(ref prt, 3) = BitMath.ToInt32Mask(Unsafe.Add(ref pleft, 3) == Unsafe.Add(ref pright, 3));     Unsafe.Add(ref prt, 4) = BitMath.ToInt32Mask(Unsafe.Add(ref pleft, 4) == Unsafe.Add(ref pright, 4));     Unsafe.Add(ref prt, 5) = BitMath.ToInt32Mask(Unsafe.Add(ref pleft, 5) == Unsafe.Add(ref pright, 5));     Unsafe.Add(ref prt, 6) = BitMath.ToInt32Mask(Unsafe.Add(ref pleft, 6) == Unsafe.Add(ref pright, 6));     Unsafe.Add(ref prt, 7) = BitMath.ToInt32Mask(Unsafe.Add(ref pleft, 7) == Unsafe.Add(ref pright, 7));     return rt; } 

说明一下,它是一个用标量算法实现函数,用于在不支持向量指令时进行回退。所以需要分别对每一个元素来做相等比较,并将比较结果转为掩码。
float是32位的,256位向量里能放下8个float。于是上面的代码使用Unsafe.Add,分别对8个元素进行了计算。

下面是使用“固定长度的数组”改写后的代码。

public static Vector256<float> Equals_Basic(Vector256<float> left, Vector256<float> right) {     UnsafeUtil.SkipInit(out Vector256<float> rt);     ref FixedArray8<int> p = ref Unsafe.As<Vector256<float>, FixedArray8<int>>(ref rt);     ref FixedArray8<float> pleft = ref Unsafe.As<Vector256<float>, FixedArray8<float>>(ref left);     ref FixedArray8<float> pright = ref Unsafe.As<Vector256<float>, FixedArray8<float>>(ref right);     p.I0 = BitMath.ToInt32Mask(pleft.I0 == pright.I0);     p.I1 = BitMath.ToInt32Mask(pleft.I1 == pright.I1);     p.I2 = BitMath.ToInt32Mask(pleft.I2 == pright.I2);     p.I3 = BitMath.ToInt32Mask(pleft.I3 == pright.I3);     p.I4 = BitMath.ToInt32Mask(pleft.I4 == pright.I4);     p.I5 = BitMath.ToInt32Mask(pleft.I5 == pright.I5);     p.I6 = BitMath.ToInt32Mask(pleft.I6 == pright.I6);     p.I7 = BitMath.ToInt32Mask(pleft.I7 == pright.I7);     return rt; } 

可见,摆脱了冗长的Unsafe.Add后,代码简洁了很多。
FixedArray8等“固定长度的数组”是结构体,且元素类型也是值类型,故它可以安全使用Unsafe.As来转换引用的类型,从而直接操作变量的内存。

备注:寻址方式说明

X86架构支持2种相对寻址。资料摘录如下。

寄存器相对寻址方式 格式:操作码 寄存器,DISP+基址或变址寄存器 有效地址=寄存器+8/16位相对值DISP MOV AX, [SI+06H] ;AX←DS:[SI+06H] MOV AX, 06H[SI] ;AX←DS:[SI+06H]  相对基址变址寻址方式 格式:操作码 寄存器,DISP+(基址寄存器)+(变址寄存器) 有效地址=BX/BP+SI/DI+8/16位相对值DISP MOV AX, [BX+DI+6] ;AX←DS:[BX+DI+6] MOV AX, 6[BX+DI] ; MOV AX, 6[BX][DI] 

相对值DISP,是立即数(Immediate)。如上面范例里的6(或06H),它会直接编译到机器码里。
当使用这2种相对寻址时,能在同一指令内就能完成“地址计算”与实际的“数据搬运”。
而不使用它们时,得花3条指令才能达到同样的效果。分别是 “将立即数加载到寄存器”(相对值)、“加法”(基址+相对值)、“寄存器间接寻址”。
目前发现直至.NET8.0,Unsafe.Add的地址计算代码在JIT编译为机器码运行时,不会使用相对寻址。即原本1条指令就能寻址,但JIT只会编译为3条指令的,影响了性能。
于是本库使用“固定长度的数组”来解决这一性能问题。

BitMath从静态类改为抽象类. 新增名称空间 Zyl.VectorTraits.Numerics

相关日志:

  • BitMath changed from static to abstract class. Add namespace Zyl.VectorTraits.Numerics, add some math functions (BitMath从静态类改为抽象类. 新增名称空间 Zyl.VectorTraits.Numerics, 增加一些的数学函数).

这次改善了标量版数学函数的组织。向量方法是基于这些标量方法的。
原来的 BitMath 类太太了。且2.0版增加了浮点类型判断等多个函数,若仍然放在BitMath里,会使该类便的更大,不易维护。
看了一下 .NET 7.0 的泛型数学,感觉它的分类方案不错。于是新增名称空间 Zyl.VectorTraits.Numerics,随后按照泛型数学的分类,建立了各个静态类。目前有这些类。

  • BitMathCore: 本库新增的函数。例如 BitSelect 等。
  • MathBitConverter: BitConverter(位转换) 的数学函数。例如 SingleToInt32Bits 等。
  • MathIFloatingPoint: IFloatingPoint(浮点数接口) 的数学函数。例如 Truncate 等。
  • MathINumber: INumber(数字性接口) 的数学函数。例如 Clamp 等。
  • MathINumberBase: INumberBase(数字基本性接口) 的数学函数。例如 Abs 等。
  • MathIRootFunctions: IRootFunctions(根函数接口) 的数学函数。例如 Sqrt 等。
  • MathOperators: 运算符的数学函数。例如 BigMul 等。

目前BitMath是一个转发汇总的作用,便于简单的使用标量版数学函数。

基准测试程序增加命令行参数

相关日志:

  • Benchmark programs add command line parameter (基准测试程序增加命令行参数): -accelerated0 -allowFakeBenchmark -cpuDetection0 -fixedVector0 -test0

手动运行基准测试程序时,往往希望能显示更多的信息,便于分析数据。
而用脚本自动运行时,一般不希望显示次要信息。
于是用命令行参数来控制比较好。

  • accelerated[=n]: 是否显示各个向量方式的加速情况(例如 Ceiling_AcceleratedTypes 等)。
  • accelerated0: 相当于“accelerated=0”,禁止显示各个向量方式的加速情况。
  • allowFakeBenchmark[=n]: 是否对 FakeBenchmarkAttribute 特性的函数也进行基准测试。一般情况下无需使用它,仅在脚本自动测试时,有时会用到。
  • allowFakeBenchmark0: 相当于“allowFakeBenchmark=0”,禁止FakeBenchmarkAttribute的基准测试。
  • cpuDetection[=n]: 是否显示CPU检测信息(例如 CpuDetectionCommand 等)。
  • cpuDetection0: 相当于“cpuDetection=0”,禁止显示CPU检测信息。
  • fixedVector[=n]: 是否显示定长向量的信息(例如 Vector128/Vector256 等)。
  • fixedVector0: 相当于“fixedVector=0”,禁止显示定长向量的信息。
  • test[=n]: 是否进行特殊测试。详见AloneTestUtil类。
  • test0: 相当于“test=0”,禁止进行特殊测试。

增加工具程序UpdateBenchmarkResults

相关日志:

  • Add tool program (增加工具程序): UpdateBenchmarkResults.

为了能将基准测试结果进行自动归类与更新,开发了这一工具程序。
编译好的程序已放在 /tools/build/ 文件夹,可以这样使用。

  1. 在各台计算机上运行基准测试的脚本,并将结果分别放到 /articles/BenchmarkResultsRaw文件夹。
  2. 运行脚本 /tools/build/UpdateBenchmarkResults.bat。它会对评测数据进行汇总, 随后更新到 /articles/BenchmarkResults文件夹。

该工具程序的源代码在 /tools/UpdateBenchmarkResults 文件夹。

算法优化

相关日志:

  • Optimized hardware acceleration of ExtractMostSignificantBits methods in the Arm architecture (优化ExtractMostSignificantBits方法在Arm架构的硬件加速). For 8~64 bit types.

附录

以前的发布日志: