clj-vapi

Vectorised array operations leveraging Java Vector API for SIMD parallelism.

Prerequisites

• Clojure >= 1.12.0
• jdk.incubator.vector

Add the following module to the list of JVM options inside your deps.edn:

:jvm-opts ["--add-modules" "jdk.incubator.vector"]

Installation

Clojure CLI/deps.edn

com.github.igmonk/clj-vapi {:mvn/version "0.1.0-SNAPSHOT"}

Leiningen/Boot

[com.github.igmonk/clj-vapi "0.1.0-SNAPSHOT"]

Usage

The top level interface is in clj-vapi.core.

(use 'clj-vapi.core)

Use vadd to add two arrays element-wise:

(def sum
  (vadd (int-array (range 1000))
        (int-array (range 1000))))

Refer to examples/clj_vapi for more examples.

Benchmarks

Execution time estimation using criterium:

(use 'criterium.core)

(def a (int-array (range 1000)))
(def b (int-array (range 1000)))

(quick-bench (vadd a b))

Expected output for a CPU with SIMD registers limited to 128 bit-size:

Evaluation count : 1632240 in 6 samples of 272040 calls.
             Execution time mean : 397.445113 ns
    Execution time std-deviation : 48.570087 ns
   Execution time lower quantile : 355.713814 ns ( 2.5%)
   Execution time upper quantile : 460.866929 ns (97.5%)
                   Overhead used : 7.036000 ns

Refer to bench/clj_vapi for more benches.

License

MIT