Tools to work with ISPC in Julia



Build Status

Note: recent changes in the way Julia represents functions and closures require a redesign of this package, which will be published when the 0.5 IR stabilises a bit more.

ISPC.jl uses the Intel ISPC compiler to compile fragments of ISPC-C or Julia code to vector code (eg. SSE/AVX on Intel CPUs) at runtime. It is similar to the standard @simd macro but supports control statements and a number of math functions.

Speedups of 2x-8x compared to plain Julia code can be achieved (single core).

  • high-level interface: translates Julia to ISPC
  • low-level interface: compiles and loads ISPC programs


  • Julia 0.4 or 0.5-dev up to cf93d6f (2016-01-29 02:19 UTC)
  • ispc (must be found in $PATH)
  • libtool or g++ (must be found in $PATH)
  • (optional) llvm_dis for looking at ISPC llvm assembly.

High-level interface

Work in progress! The high-level interface is fairly functional already, but bugs and the odd missing math function should be expected. More documentation to come.

See the example notebook for an overview.

The high-level interface translates Julia code that has been annotated with a set of macros into code compiled by ISPC on-the-fly. Example:

using ISPC

@inline function mandel(c_re, c_im, count)
    z_re = c_re
    z_im = c_im
    i = 0
    while i < count
        if (z_re * z_re + z_im * z_im > 4.0f0)
        new_re = z_re*z_re - z_im*z_im
        new_im = 2.0f0 * z_re * z_im
        z_re = c_re + new_re
        z_im = c_im + new_im
        i += 1
    return i

@ispc function mandelbrot_ispc(x0, y0, x1, y1, output, max_iters)
    height, width = size(output)
    dx = (x1 - x0) / width
    dy = (y1 - y0) / height
    @kernel(`--target=avx1-i32x8`) do
        for i = 1:width
            @foreach(1:height) do j
                x = x0 + i * dx
                y = y0 + j * dy
                output[j,i] = mandel(x, y, max_iters)

Supported ISPC constructs:

  • @foreach
  • @foreach(:active) (untested)
  • @unmasked (untested)
  • @coherent (untested)


The high-level interface is meant to be used with numerical kernels only. Not all of Julia's syntax and types are supported:

  • Arrays can only be indexed with integers, providing either a single index (linear indexing) or one index per dimension (multi-dimensional arrays). Indexing must yield an array element, not a sub-array.

  • All outer variables ("kernel arguments") must be primitive types or arrays of primitive types.

  • Simple composite types like UnitRange are supported inside kernels (eg. in for loops) but not as kernel arguments at the moment.

  • Only functions that have a direct translation to ISPC are supported inside kernels. User-defined functions may be used if they are declared @inline. This restriction may be lifted in the future.

  • Inner functions, exceptions, task switching, I/O etc. are not supported.

ISPC task-level constructs are not yet supported.

Low-level interface

The low-level interface allows you to load and call fragments of ISPC code:

using ISPC

# A basic ISPC kernel:
code = """
export void simple(uniform float vin[], uniform float vout[],
                   uniform int count) {
    foreach (index = 0 ... count) {
        float v = vin[index];
        if (v < 0.5)
            v = v * v;
            v = sqrt(v);
        vout[index] = v;

# Compile the code and get a function pointer to our kernel:
lib = load_ispc(code, `--target=avx1-i32x8`)
fptr = Libdl.dlsym(lib, "simple")

# Call the kernel:
vin = rand(Float32, 1000);
vout = zeros(Float32, 1000);
ccall(fptr, Void, (Ref{Float32}, Ref{Float32}, UInt64), vin, vout, length(vout))

How it works

  1. The @kernel macro creates a lambda function containing the kernel code.
  2. code_lowered() or code_typed() is run on the main @ispc function to get information about closure variables captured by the kernel lambda. These become kernel arguments.
  3. Kernel fragments in the main function are replaced by calls to a @generated kernel_call function.
  4. When kernel_call is called for the first time, type inference is run on the kernel fragment. This gives us types for local variables and inlines functions that can be inlined.
  5. The lowered and typed AST is transformed to un-lower gotos back into if and while statements (ISPC does not support varying gotos)
  6. The transformed AST is translated to ISPC C
  7. The resulting code is compiled with ispc, loaded with Libdl and called with ccall.

First Commit


Last Touched

8 months ago


86 commits

Used By: