nimd

NIMD: A Write Once SIMD Library for Nim wrapping libsimdpp.

NIMD supports a large part of libsimdpp. See its docs: libsimdpp docs

NIMD allows you to use SIMD vectorization without the need for specific intrinsics.

Courtesy of libsimdpp:

On architectures that support different SIMD instruction sets the library allows the same source code files to be compiled for each SIMD instruction set and then hooked into an internal or third-party dynamic dispatch mechanism. This allows the capabilities of the processor to be queried on runtime and the most efficient implementation to be selected.

Simple example

import nimd

var  x = Vector(1.float32, 2, 3, 4)
let 
  y = Vector(4.float32, 3, 2, 1)
  z = x * y / x # standard ops
x[3]=9 #insertion
echo z # float32x4[4.0, 3.0, 2.0, 1.0]
echo x # float32x4[1.0, 2.0, 3.0, 9.0]
echo y.min(x) # float32x4[1.0, 2.0, 2.0, 1.0]

Installation

nimble install https://github.com/CircArgs/nimd.git

You will also need libsimdpp (As of this writing NIMD works with f7ab03f). Clone it somewhere git clone https://github.com/p12tic/libsimdpp.git. Lets call the absolute path to libsimdpp: MY_LIBSIMDPP_PATH.

We can build projects derivative of NIMD by using the following nim compiler flags:

• cpp: use the c++ backend
• -t:-I$MY_LIBSIMDPP_PATH: a NIMD compile-time flag to state where libsimdpp is located

Note: you could also add these to a config.nims

Example:

nim cpp -t:"-I/home/nick/Projects/testnimd/libsimdpp" -r src/testnimd.nim

Types

NIMD supports the following vector types

base_type	size_type	simd_width	simd_type
uint	uint8	16	uint8x16
int	int8	16	int8x16
uint	uint16	8	uint16x8
int	int16	8	int16x8
uint	uint32	4	uint32x4
int	int32	4	int32x4
uint	uint64	2	uint64x2
int	int64	2	int64x2
float	float32	4	float32x4
float	float64	2	float64x2
uint	uint8	32	uint8x32
int	int8	32	int8x32
uint	uint16	16	uint16x16
int	int16	16	int16x16
uint	uint32	8	uint32x8
int	int32	8	int32x8
uint	uint64	4	uint64x4
int	int64	4	int64x4
float	float32	8	float32x8
float	float64	4	float64x4
uint	uint8	64	uint8x64
int	int8	64	int8x64
uint	uint16	32	uint16x32
int	int16	32	int16x32
uint	uint32	16	uint32x16
int	int32	16	int32x16
uint	uint64	8	uint64x8
int	int64	8	int64x8
float	float32	16	float32x16
float	float64	8	float64x8
mask_uint	mask_uint8	16	mask_uint8x16
mask_int	mask_int8	16	mask_int8x16
mask_uint	mask_uint16	8	mask_uint16x8
mask_int	mask_int16	8	mask_int16x8
mask_uint	mask_uint32	4	mask_uint32x4
mask_int	mask_int32	4	mask_int32x4
mask_uint	mask_uint64	2	mask_uint64x2
mask_int	mask_int64	2	mask_int64x2
mask_float	mask_float32	4	mask_float32x4
mask_float	mask_float64	2	mask_float64x2
mask_uint	mask_uint8	32	mask_uint8x32
mask_int	mask_int8	32	mask_int8x32
mask_uint	mask_uint16	16	mask_uint16x16
mask_int	mask_int16	16	mask_int16x16
mask_uint	mask_uint32	8	mask_uint32x8
mask_int	mask_int32	8	mask_int32x8
mask_uint	mask_uint64	4	mask_uint64x4
mask_int	mask_int64	4	mask_int64x4
mask_float	mask_float32	8	mask_float32x8
mask_float	mask_float64	4	mask_float64x4
mask_uint	mask_uint8	64	mask_uint8x64
mask_int	mask_int8	64	mask_int8x64
mask_uint	mask_uint16	32	mask_uint16x32
mask_int	mask_int16	32	mask_int16x32
mask_uint	mask_uint32	16	mask_uint32x16
mask_int	mask_int32	16	mask_int32x16
mask_uint	mask_uint64	8	mask_uint64x8
mask_int	mask_int64	8	mask_int64x8
mask_float	mask_float32	16	mask_float32x16
mask_float	mask_float64	8	mask_float64x8