Class variables are deallocated after calling function - swift

Question

I have a function in swift, as below. There is a loop with a reference to variables existing within an instance of this class. (fftfilterbankReal is an array in the class). However after one pass, I get an error with 'index out of range' at the code line 'for i in 0..
In the debugger it seems that on the 2nd iteration of this loop, there are no variables under the 'self' drop down.
If I comment out the line 'vDSP_zvmul(&kernel!, 1, &fft1Input, 1, &result, 1, vDSP_Length(r.count), 1)' then the loop runs and I can debug at any time and visually see the self variables in the debugger.
What am I missing that seems to make these variables disappear? I have read into memory allocation and such, and my class variables are declared using 'var' and nothing more, as that should default to strong in swift.
func convolveInput(realsamples:[Float], imagsamples:[Float]) -> [Float]{
realResult = Array(repeating: [], count: filterbankReal.count)
imagResult = Array(repeating: [], count: filterbankReal.count)
let x = realsamples
let y = imagsamples
var N = x.count
var logN = 16
var fft1Setup = vDSP_create_fftsetup(UInt(logN), FFTRadix(FFT_RADIX2))!
var paddedLength = x.count + filterbankReal.count - 1
var halfPaddedLength = paddedLength/2
var halfKernelLength = kernelLength/2
//setup Complex Buffer 1
var reals = [Float]()
var imags = [Float]()
for i in 0..<x.count{
reals.append(x[i])
imags.append(y[i])
}
var complexBuffer1 = DSPSplitComplex(realp: UnsafeMutablePointer(mutating: reals), imagp: UnsafeMutablePointer(mutating: imags))
//Perform FFT on incoming samples
var re = [Float](repeating:0.0, count: N)
var im = [Float](repeating:0.0, count: N)
var fft1Input = DSPSplitComplex(realp: UnsafeMutablePointer(mutating: re), imagp: UnsafeMutablePointer(mutating: im))
var fftlength = 10
vDSP_fft_zop(fft1Setup, &(complexBuffer1), 1, &fft1Input, 1, UInt(fftlength), Int32(FFT_FORWARD))
//Remove DC from FFT Signal
re.remove(at: 0)
im.remove(at: 0)
for i in 0..<self.fftfilterbankReal.count {
var r:[Float] = self.fftfilterbankReal[i]
var im:[Float] = self.fftfilterbankImag[i]
var kernel:DSPSplitComplex? = DSPSplitComplex(realp: &r, imagp: &im)
var res:Float = 0
var ims:Float = 0
var result:DSPSplitComplex = DSPSplitComplex(realp: &res, imagp: &ims)
vDSP_zvmul(&kernel!, 1, &fft1Input, 1, &result, 1, vDSP_Length(r.count), 1)
self.realResult[i].append(res)
self.imagResult[i].append(ims)
}

Answer 1

Your code is sort of a showcase of bad usages when working with Arrays and pointers.
For example:
var complexBuffer1 = DSPSplitComplex(realp: UnsafeMutablePointer(mutating: reals), imagp: UnsafeMutablePointer(mutating: imags))
or:
var kernel:DSPSplitComplex? = DSPSplitComplex(realp: &r, imagp: &im)
DSPSplitComplex holds two pointers for real part and imaginary part separately and does not copy the contents. You should not pass Swift Arrays for such parameters.
And the most critical part in your code is...
var res:Float = 0
var ims:Float = 0
var result:DSPSplitComplex = DSPSplitComplex(realp: &res, imagp: &ims)
vDSP_zvmul(&kernel!, 1, &fft1Input, 1, &result, 1, vDSP_Length(r.count), 1)
vDSP_zvmul generates N (in your code N = vDSP_Length(r.count)) complex numbers, so you need to prepare a region which can hold N elements.
Once you call vDSP_zvmul with your current code, you break whole stack contents which causes what you have experienced:
In the debugger it seems that on the 2nd iteration of this loop, there
are no variables under the 'self' drop down.
You are hiding many parts of your code, so it is very hard to guess what you really want to do, but if I re-write your code in safer manner, it would be something like this:
func convolveInput(realsamples:[Float], imagsamples:[Float]) -> [Float]{
realResult = Array(repeating: [], count: filterbankReal.count)
imagResult = Array(repeating: [], count: filterbankReal.count)
let x = realsamples
let y = imagsamples
var N = x.count
var logN = 16
var fft1Setup = vDSP_create_fftsetup(UInt(logN), FFTRadix(FFT_RADIX2))!
var paddedLength = x.count + filterbankReal.count - 1
var halfPaddedLength = paddedLength/2
var halfKernelLength = kernelLength/2
//setup Complex Buffer 1
var reals = UnsafeMutableBufferPointer<Float>.allocate(capacity: x.count)
defer {reals.deallocate()}
var imags = UnsafeMutableBufferPointer<Float>.allocate(capacity: y.count)
defer {imags.deallocate()}
_ = reals.initialize(from: x)
_ = imags.initialize(from: y)
var complexBuffer1 = DSPSplitComplex(realp: reals.baseAddress!, imagp: imags.baseAddress!)
//Perform FFT on incoming samples
var re = UnsafeMutableBufferPointer<Float>.allocate(capacity: N)
defer {re.deallocate()}
var im = UnsafeMutableBufferPointer<Float>.allocate(capacity: N)
defer {im.deallocate()}
var fft1Input = DSPSplitComplex(realp: re.baseAddress!, imagp: im.baseAddress!)
let fftlength = 10
vDSP_fft_zop(fft1Setup, &complexBuffer1, 1, &fft1Input, 1, UInt(fftlength), Int32(FFT_FORWARD))
//Remove DC from FFT Signal
fft1Input = DSPSplitComplex(realp: re.baseAddress!+1, imagp: im.baseAddress!+1)
for i in 0..<self.fftfilterbankReal.count {
self.fftfilterbankReal[i].withUnsafeMutableBufferPointer {rBuf in
self.fftfilterbankImag[i].withUnsafeMutableBufferPointer {imBuf in
var kernel = DSPSplitComplex(realp: rBuf.baseAddress!, imagp: imBuf.baseAddress!)
var res = UnsafeMutableBufferPointer<Float>.allocate(capacity: rBuf.count)
defer {res.deallocate()}
var ims = UnsafeMutableBufferPointer<Float>.allocate(capacity: rBuf.count)
defer {ims.deallocate()}
var result:DSPSplitComplex = DSPSplitComplex(realp: res.baseAddress!, imagp: ims.baseAddress!)
vDSP_zvmul(&kernel, 1, &fft1Input, 1, &result, 1, vDSP_Length(rBuf.count), 1)
//vDSP_zvmul generates `N` complex numbers,
// I do not understand what you really want to do...
self.realResult[i].append(res[0])
self.imagResult[i].append(ims[0])
}
}
}
//...
}
There may be other parts to fix, but anyway, please try and see what you get.

Answer 2

Foreword: Boy, sifting through all this took me almost 2 hours, and it's still not even perfect, but boy is this much nicer. I hope it helps!
Your code is suffer massively because the Accelerate APIs are C APIs that are not adapted to take advantage of Swift features. You will have much more readable code if you make yourself some nice wrappers for the Accelerate API, which lets you tuck all the "Ugly stuff" away into a corner you seldom have to see or edit.
I did this by creating a new type, ComplexFloatArray, which is similar to DSPSplitComplex, but actually encapsulates and owns its buffers. This prevents the dangling buffers that DSPSplitComplex is susceptible to.
After dining the ComplexFloatArray types, it's time to define some wrappers for the the Accelerate functions you used. In this case, vDSP_zvmul and vDSP_fft_zop. Since C doesn't have tuples, returning multiple values from a C function usually requires that you use out-parameters, which are used pervasively in the Accelerate framework. We can re-design these as Swift functions with regular return types. These APIs are very naturally expressed as instance methods that operate on ComplexFloatArray, so we'll put them there.
Additionally, your code is made much more complex by its dependance on external state. Convolution is a function, there's no reason why it does anything besides taking in input (via parameters, and not via instance variables) and returns the result (via return value, and not via instance variables).
import Accelerate
class ComplexFloatArray {
var reals: [Float]
var imaginaries: [Float]
init(reals: [Float], imaginaries: [Float]) {
self.reals = reals
self.imaginaries = imaginaries
}
}
extension ComplexFloatArray { // Core features
var count: Int {
assert(reals.count == imaginaries.count)
return reals.count
}
static let stride = 1
func append(real: Float, imaginary: Float) {
self.reals.append(real)
self.imaginaries.append(imaginary)
}
func useAsDSPSplitComplex<R>(_ closure: (inout DSPSplitComplex) -> R) -> R {
return reals.withUnsafeMutableBufferPointer { realBufferPointer in
return imaginaries.withUnsafeMutableBufferPointer { imaginaryBufferPointer in
var dspSplitComplex = DSPSplitComplex(realp: realBufferPointer.baseAddress!, imagp: imaginaryBufferPointer.baseAddress!)
return closure(&dspSplitComplex)
}
}
}
}
extension ComplexFloatArray { // Convenience utilities
convenience init() {
self.init(reals: [], imaginaries: [])
}
static func zeros(count: Int) -> ComplexFloatArray {
return ComplexFloatArray(reals: Array(repeating: 0, count: count), imaginaries:Array(repeating: 0, count: count))
}
}
extension ComplexFloatArray { // Vector multiplciation extensions
enum ComplexMultiplicationType: Int32 { case normal = 1, conjugate = -1 }
func complexMultiply(
with other: ComplexFloatArray,
multiplicationType: ComplexMultiplicationType = .normal
) -> ComplexFloatArray {
assert(self.count == other.count, "Multiplied vectors must have the same size!")
let result = ComplexFloatArray.zeros(count: self.count)
self.useAsDSPSplitComplex { selfPointer in
other.useAsDSPSplitComplex { otherPointer in
result.useAsDSPSplitComplex { resultPointer in
vDSP_zvmul(
&selfPointer, ComplexFloatArray.stride,
&otherPointer, ComplexFloatArray.stride,
&resultPointer, ComplexFloatArray.stride, vDSP_Length(result.count),
multiplicationType.rawValue)
}
}
}
return result
}
}
extension ComplexFloatArray { // FFT extensions
enum FourierTransformDirection: Int32 { case forward = 1, inverse = -1 }
//TODO: name log2n label better
func outOfPlaceComplexFourierTransform(
setup: FFTSetup,
resultSize: Int,
log2n: UInt,
direction: FourierTransformDirection
) -> ComplexFloatArray {
let result = ComplexFloatArray.zeros(count: resultSize)
self.useAsDSPSplitComplex { selfPointer in
result.useAsDSPSplitComplex{ resultPointer in
vDSP_fft_zop(
setup,
&selfPointer, ComplexFloatArray.stride,
&resultPointer, ComplexFloatArray.stride,
log2n,
direction.rawValue
)
}
}
return result
}
}
extension FFTSetup {
enum FourierTransformRadix: Int32 {
case radix2 = 0, radix3 = 1, radix5 = 2
// Static let constants are only initialized once
// This function's intent to to make sure this enum stays in sync with the raw constants the Accelerate framework uses
static let assertRawValuesAreCorrect: Void = {
func assertRawValue(for actual: FourierTransformRadix, isEqualTo expected: Int) {
assert(actual.rawValue == expected, "\(actual) has a rawValue of \(actual.rawValue), but expected \(expected).")
}
assertRawValue(for: .radix2, isEqualTo: kFFTRadix2)
assertRawValue(for: .radix3, isEqualTo: kFFTRadix3)
assertRawValue(for: .radix5, isEqualTo: kFFTRadix5)
}()
}
init(log2n: Int, _ radix: FourierTransformRadix) {
_ = FourierTransformRadix.assertRawValuesAreCorrect
guard let setup = vDSP_create_fftsetup(vDSP_Length(log2n), FFTRadix(radix.rawValue)) else {
fatalError("vDSP_create_fftsetup(\(log2n), \(radix)) returned nil")
}
self = setup
}
}
struct NameMe {
// I don't know what this is, but if it can somehow be removed,
// the whole convolveInput method could be moved into an extension on ComplexFloatArray.
var fftFilterBank: [ComplexFloatArray]
func convolve(samples: ComplexFloatArray) -> [ComplexFloatArray] {
// TODO: rework reimplement this code to remove the DC from samples, and add it back in
// //Remove DC from FFT Signal
// re.remove(at: 0)
// im.remove(at: 0)
let fftlength: UInt = 10 // Todo: what is this, exactly?
let fft1Input = samples.outOfPlaceComplexFourierTransform( // Rename me to something better
setup: FFTSetup(log2n: 16, .radix2),
resultSize: samples.count,
log2n: fftlength,
direction: .forward
)
return self.fftFilterBank.map { kernel in kernel.complexMultiply(with: fft1Input) }
}
// Stub for compatibility with the old API. Deprecate it and move to the
// `convolve(samples: ComplexFloatArray) -> [ComplexFloatArray]` as soon as possible.
func convolveInput(realsamples: [Float], imagsamples: [Float]) -> [ComplexFloatArray] {
return self.convolve(samples: ComplexFloatArray(reals: realsamples, imaginaries: imagsamples))
}
}
I have some notes along the way
This function is WAAAAAAAAAAAY too long. If you have a function that's over 10 lines long, there's a fairly strong indicator that it's growing too large, does to many things, and could benefit from being broken down into simpler steps.
You have lots of redundant variables. You don't need more than 1 copy of any given immutable value. You have all these different names, referring to the same thing, which just complicates things. There might be an argument to be made that this can be useful if the new names have significance, but names like x, y, re, im are near-useless in their communicative ability, and should almost-always be avoided entirely.
Arrays are value types with Copy-on-Write. You can make copies of them by simply assigning to them to a new variable, so code like:
var reals = [Float]()
var imags = [Float]()
for i in 0..<x.count{
reals.append(x[i])
imags.append(y[i])
}
Is both slow, and visually cumbersome. This could be simply: let (reals, imags) = (x, y). But again, these copies are unnecessary (as are x and y). Remove them, and just use realsamples and imagsamples directly.
When you find yourself frequently passing multiple pieces of data together, that's a very strong indication that you should define a new aggregate type to wrap them. For example, if you're passing two Array<Float> to represent complex vectors, you should define a ComplexVector type. This can let you enforce invariants (e.g. there are always as many real values as imaginary values), and add convenient operations (e.g. a func append(real: Float, imaginary: Float), which operates on both simultaneously, ensuring you can never forget to append to one of the arrays).
In closing,
There's a lot going on here, so I can't possible pre-empt every question and explain it ahead of time. I encourage you to take some time, read through this, and feel free to ask me any follow up questions.
I suspect I've made mistakes during my refactor (because I had no test cases to work with), but the code is modular enough that it should be very simple to isolate and fix and bugs.