Command simd
Package simd implements portable and vector-size-agnostic SIMD types, and functions and methods for working with these types. SIMD types are either implemented in hardware (for example, arm64 "Neon" or amd64 "AVX/AVX2/AVX512") using the corresponding types in the simd/archsimd package, or emulated in pure Go. In all cases, the vector length is at least 128 bits, and within a given program execution, all vectors have the same length.
SIMD Types
There is a simd type corresponding to each primitive numeric type, except for complex64 and complex128. Each of these is the type name, capitalized, with an "s" suffix, for example Int8s, Uint16s, or Float64s.
There are also simd "mask" types that abstract the mask registers present in some architectures, otherwise these will be implemented as bit masks.
Obtaining SIMD values
The zero value of a SIMD vector type is valid and represents a zero vector.
For each SIMD type, "Load<Types>(s []<type>) <Types>" loads a full verctor of the type from a long-enough slice.
For slices that are not long enough, "Load<Types>Part(s []<type>) (<Types>, int)" will load as many elements as are available from the slice, fill the remainder with zero, and also return the number that were loaded.
For each SIMD type, "Broadcast<Types>(x type) <Types>" returns a vector whose elements are all initialized to x.
Examples:
Operations
SIMD types provide methods for unary (Float32s.Abs, Int16s.Not), binary (Float64s.Add, Uint32s.GreaterEqual), and ternary operations (Int64s.IfElse, Float32s.MulAdd). Relational operations produce masks.
SIMD types also support conversions between types, both those that are mostly value-preserving (Int32s.ConvertToFloat32, Float32s.ConvertToInt32) and those that change types without altering the underlying vector bit pattern.
Signed integer types convert to mask types (Int16s.ToMask), but this is a comparison against zero, not a simple bitwise conversion. Mask types convert to signed integers in an operation (Mask32s.ToInt32s) that may be a zero-cost bitwise conversion, or not, depending on the underlying hardware.
Storing
SIMD vector types have two methods, one for storing the entire vector into a slice "Store([]<type>)"" and a second for storing part of a vector into a slice "StorePart([]<type>) int". StorePart returns the number of elements actually stored.
String conversion
Vectors and masks provide a String method for conversion to strings.
Conversion to and from simd/archsimd types.
Each SIMD vector type has a "ToArch() any" method that returns the type supported by the current hardware as an "any". Code using these methods must be build-tagged to the relevant architecture(s) and type-assert the returned value to the appropriate type.
The simd package also includes generic functions for converting an architecture-dependent simd/archsimd value (e.g. archsimd.Float32x4) into the corresponding simd type. This function will panic if the correspondence is incorrect.
For an example of converting between simd and arch/simd types, see the test file sum_amd64_test.go.
Bugs
- ☞
Calls won't work, and there may be other bugs.
- ☞
SIMD-dependent var initializers don't work.
- ☞
Modified names may appear in stack traces and debugging.