overdrive

Overdrive is a high-performance SIMD abstraction library for Nim that aims to abstract away SIMD operations behind a neat, (mostly) safe Nim-like API.

The gist is that it lets you write SIMD code that runs on a lot of stuff, ranging from your desktop/laptop to your Android phone, to your tablet, to your smart TV.

features

• Store
• Indexing
• Masking
• Averages
• Insertion
• Zero/one testing

roadmap

• AVX2 support
• SSE4.1 support
• SSE3 support
• SSE2 support
• Scalar fallback support (see section "Scalar Fallback")
• Multi-arch CI setup (chore)
• NEON support
• RVV support
• WASM SIMD support (for Emscripten target)

basic example

Here, we can see a basic example of using Overdrive to write a simple string-find function.

func find(s: string, c: char): int =
  var target: Vector[char]
  target.store(c) # Load the character we're searching for into a register

  var i = 0
  let cap = target.capacity

  while i + cap <= s.len:
    # If there are more than or 32 bytes/characters ahead,
    var blk: Vector[char]
    blk.store(s[i].addr) # Load the string from i ..< i + 32

    let masked = blk.mask(target) # Mask the block against the target
    if masked != 0: # If masked were to be zero, it'd mean `c` is not in this block.
      let offset = countTrailingZeroBits(masked)
        # Get the first offset at which `c` occurs in the block
      return i + offset

    i += cap

  # Process the bytes that didn't fit in the register
  while i < s.len:
    if s[i] == c:
      return i

    inc i
  
  # We couldn't find c
  return -1

Overdrive neatly abstracts away the subtleties and pain-points of using vector intrinsics away from the programmer, acting as a write-once, run-anywhere SIMD library like Google's Highway but in pure, idiomatic Nim with next to zero performance cost thanks to the use of templates and compile-time ISA selection.

installation

To use Overdrive in your program, run the following:

neo

Neo can be downloaded here. However, you aren't forced to use it.

$ neo add gh:xTrayambak/overdrive

nimble

$ nimble add https://github.com/xTrayambak/overdrive

scalar fallback

Overdrive can utilize a scalar backend when no supported ISA is usable, but this is more of a last-resort and should be avoided as much as possible. It is horribly slow and is not guaranteed to be correct (for now). You are recommended to utilize the getBackend() function to implement your own scalar algorithm implementations when SIMD acceleration is not present.

when getBackend() != VInstSet.Scalar:
  # SIMD accelerated path
  # ...
else:
  # Scalar path
  # ...

The scalar fallback uses an array[N, U]. N can be specified via the compile-time flag -d:OverdriveScalarRegSize, and it is 32 by default.