Simd

Library compatible with GCC\Clang vector extensions.

This is a small single-header library, just include Simd.h to use it.

It defines the following SIMD floating point types: float4, int4, float8, int8 and some useful mathematical functions: Floor, Round, Truncate, CastToFloat, TruncateToInt, RoundToInt, Exp, Log, Min, Max, HorSum and MultiplyAccumulate. All types, functions and operators are defined in a Simd namespace.

Example of library usage:

Simd::float8 c1 = Simd::Set<Simd::float8>(coeffs[0]);
Simd::float8 c2 = Simd::Set<Simd::float8>(coeffs[1]);
Simd::float8 c3 = Simd::Set<Simd::float8>(coeffs[2]);
auto dstEnd = dst + n;
while(dst != dstEnd)
{
	Simd::float8 a = Simd::Load8Aligned(dst);
	Simd::float8 b1 = Simd::Load8Aligned(srcs[0]);
	Simd::float8 b2 = Simd::Load8Aligned(srcs[1]);
	Simd::float8 b3 = Simd::Load8Aligned(srcs[2]);
	Simd::StoreAligned(a + b1 * c1 + b2 * c2 + b3 * c3, dst);

	dst += 8;
	srcs[0] += 8;
	srcs[1] += 8;
	srcs[2] += 8;
}
Simd::End();

The code above calculates weighted sum of three float arrays together and adds the result to dst. It assumes that all arrays are 32-byte aligned, and array size n is a multiple of 8.

For GCC and Clang these types are just typedef of corresponding built-in vector types without any wrappers around them.

For MSVC compiler that don't have support of these extensions, library implements these types as wrappers with overloaded operators +, -, *, /,=, +=, -=, *=, /=, <<, >>, &, |, &&, || and so on. In release builds these operators are inlined by compiler and don't affect performance. float4 type is implemented via SSE intrinsics, int4 - SSE2, float8 - AVX, int8 - AVX2. Compile code with corresponding /arch:(SSE|SSE2|AVX|AVX2) key to enable them, otherwise these types will not be defined (for x86) or only float4 and int4 types will be defined in x64. As a alternative you can #define SIMD_SSE_LEVEL SIMD_SSE_LEVEL_* before including Simd.h. Also some functions use SSE3/SSSE3/SSE4.1/SSE4.2 for better performance when their generation is enabled with SIMD_SSE_LEVEL or /arch:AVX or /arch:AVX2 compiler key.

The library was not tested well and probably contain bugs, so be ready to look into Simd.h source code to fix bugs.

Performance tests in this project are messy and some of them have incorrect estimated value of GFlops. My future plan is to make this library a part of Intra project, so I will be able to write tests using Intra testing features.

devoln/Simd

Simd