Convert [UInt32] -> [UInt8] -> [[UInt8]] in Swift

Question

I am trying to speed up my current implementation of a function that converts [UInt32] to a [UInt8] which in turn is split up into [[UInt8]] with 6 arrays at each index.

My implementation:

extension Array {
func splitBy(subSize: Int) -> [[Element]] {
    return 0.stride(to: self.count, by: subSize).map { startIndex in
        let endIndex = startIndex.advancedBy(subSize, limit: self.count)
        return Array(self[startIndex ..< endIndex])
    }
  }
}



func convertWordToBytes(fullW : [UInt32]) -> [[UInt8]] {
    var combined8 = [UInt8]()

    //Convert 17 [UInt32] to 68 [UInt8]
    for i in 0...16{
        _ = 24.stride(through: 0, by: -8).map {
            combined8.append(UInt8(truncatingBitPattern: fullW[i] >> UInt32($0)))
        }
    }

    //Split [UInt8] to [[UInt8]] with 6 values at each index.
    let combined48 = combined8.splitBy(6) 

    return combined48
}

This function will be iterated millions of times in my program and its speed is a huge burden.

Anyone got any ideas? Thanks

Answer 1

If you profile (Cmd + I) your code, you will see that most of the time is on various "copy to buffer" functions. This happens when you append a new element to the array but it has run out of its initial allocated space so it must be moved to a location on the heap with more memory. Morals of the lesson: heap allocation is slow but unavoidable with arrays. Do it as few times as possible.

Try this:

func convertWordToBytes2(fullW: [UInt32]) -> [[UInt8]] {
    let subSize = 6

    // We allocate the array only once per run since allocation is so slow
    // There will only be assignment to it after
    var combined48 = [UInt8](count: fullW.count * 4, repeatedValue: 0).splitBy(subSize)

    var row = 0
    var col = 0

    for i in 0...16 {
        for j in 24.stride(through: 0, by: -8) {
            let value = UInt8(truncatingBitPattern: fullW[i] >> UInt32(j))
            combined48[row][col] = value

            col += 1
            if col >= subSize {
                row += 1
                col = 0
            }
        }
    }

    return combined48
}

Benchmark code:

let testCases = (0..<1_000_000).map { _ in
    (0..<17).map { _ in arc4random() }
}

testCases.forEach {
    convertWordToBytes($0)
    convertWordToBytes2($0)
}

Result (on my 2012 iMac)

Weight          Self Weight         Symbol Name
9.35 s   53.2%  412.00 ms           specialized convertWordToBytes([UInt32]) -> [[UInt8]]
3.28 s   18.6%  344.00 ms           specialized convertWordToBytes2([UInt32]) -> [[UInt8]]

By eliminating multiple allocations, we already reduced the run time by 60%. But each test case is independent, which lends itself perfectly to parallel processing with today's multi-core CPU. A modified loop...:

dispatch_apply(testCases.count, dispatch_get_global_queue(DISPATCH_QUEUE_PRIORITY_HIGH, 0)) { i in
    convertWordToBytes2(testCases[i])
}

... will shave about 1 second off the wall time when executed on my quad-core i7 with 8 threads:

Weight    Self Weight       Symbol Name
2.28 s    6.4%  0 s         _dispatch_worker_thread3  0x58467
2.24 s    6.3%  0 s         _dispatch_worker_thread3  0x58463
2.22 s    6.2%  0 s         _dispatch_worker_thread3  0x58464
2.21 s    6.2%  0 s         _dispatch_worker_thread3  0x58466
2.21 s    6.2%  0 s         _dispatch_worker_thread3  0x58465
2.21 s    6.2%  0 s         _dispatch_worker_thread3  0x58461
2.18 s    6.1%  0 s         _dispatch_worker_thread3  0x58462

The time saving is not as much as I hoped for. Apparently there's some contention when accessing the heap memory. For anything even faster, you should explore a C-based solution.