Skip to content

[5.1] Collected Accelerate overlay work #24269

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 11 commits into from
Apr 26, 2019

Conversation

stephentyrone
Copy link
Contributor

This batches up all of @FlexMonkey's Accelerate overlay changes for 5.1 branch.

rdar://problem/50203684

FlexMonkey and others added 7 commits April 25, 2019 09:13
* [Accelerate] [Quadrature] New Quadrature Overlay

A class that simplifies approximating the definite integral of a function.

* Fixes in response to PR Review.

Change `@_exported import Accelerate` to `import Accelerate`.
Correct date in copyright comment.

* Code Review Fixes.

* Remove mutable integrator, simply set options in init().
* Remove mutable `maxIntervals` and `qagPointsPerInterval`.
* Update tests.

* Code Review Fixes.

* Use standard library `Result`.
* Implement `QAGPointsPerInterval` as a struct.

* Tidy up Passing Integrand to `quadrature_integrate_function`.

Pass the integrand closure directly to the `quadrature_integrate_function` initialiser rather than attaching as a property of the `Quadrature` class.

* Make `Quadrature` a struct, and remove requirement for `integrand` to be escaping.

* Code Review Changes

* Add long-form integrator algorithm aliases (update tests to use these aliases).
* Make quadrature error description public.
* New tests for error description and `QAGPointsPerInterval`.

* Vectorized Integrand for Quadrature

Create an alternative implementation of `integrate` that accepts an integrand of type `(_ input: UnsafeBufferPointer<Double>, _ result: UnsafeMutableBufferPointer<Double>)`.

* Vectorized Integrand for Quadrature

Create an alternative implementation of `integrate` that accepts an integrand of type `(_ input: UnsafeBufferPointer<Double>, _ result: UnsafeMutableBufferPointer<Double>)`.

* Vectorized Integrand for Quadrature

Remove _ContiguousCollection - it's no longer used.

* Delete ContiguousCollection.swift

Not required.

* Refactor tests.
* [Accelerate] [vForce] New vForce Overlay

This PR contains a suite of overlays to the vForce transcendental and trigonometric functions on vectors of any length. The overlays simplify the API to the existing functions and accept collections that implement a new protocol called AccelerateBuffer. Conformances are provided for the most useful stdlib Collections.
* Add missing cast to new vDSP_FIR overlay.

* Use macOS instead of OSX in availability annotations.

* Reorder Accelerate/CMakeLists.txt alphabetically.
* Accelerate vImage Swift Overlays

A suite of functions, enumerations, and option sets to make working with vImage in Swift simpler.

* vImage_Buffer - new initializers to instantiate and initialize buffers with a single function call.
* vImage_Buffer - new functions to copy buffers and create CGImage instances from contents.
* vImage_CGImageFormat - new initializers.
* vImage_CGImageFormat - new equivalence operator.
* vImageConverter - methods to wrap free functions.
* vImageConverter - new make and convert functions.
* vImageCVImageFormat - new make functions.
* vImageCVImageFormat - methods to wrap free functions.
* vImage_Error - errorDescription function.
* vImage flags as an option set.

* Add new methods for generating CV -> CG and CG -> CV converters.

* update comments: `height` and `width` to `size`.

* `vImage_CGImageFormat.componentCount` should be `Int`.

* `vImage_CGImageFormat.componentCount` should be `Int`

* Move `self.init()` to after the size check for kvImageNoAllocate init.

* Buffer initializers to width and height rather than size.

* change vImage_CGImageFormat lightweight initializer to accept Int for `bitsPerComponent` and `bitsPerPixel`.

* Flesh out docs for vImage_Buffer.size

* Remove faux initializer in favor of new static function: `preferredAlignmentAndRowBytes`.

* Change functions to use proper error handling rather than inout error codes.

* Removed `flags` from basic init.

The only real flag to pass here is print diagnostics to console, and I now throw proper errors.

* Tests to check error throwing for buffer copy.

* remove unnecessary import, add missing docs.

* Add comments to error enums.

* Fix bug creating string from format code.

* Make `vImageCVImageFormat.formatCode` a `UInt32`.

* Remove equivalence operator from `CGImageFormat`.
* Swift overlay to vDSP cascaded biquad IIR filter.

* update comment in test

* Swift overlay to vDSP Discrete Cosine Transform

* Swift Overlays to vDSP Fast Fourier Transform and Discrete Fourier Transform

* Tests for DFT and FFT

* Some refactoring and begin documentation.

* Swift overlays to FFT and DFT operations.

* Expand docs.
* Implement `vDSP_destroy_fftsetup` and `vDSP_destroy_fftsetupD`

* Refactor Tests

* Refactor Tests

* Provide transform function that returns the result. Refactoring: parameter names more consistent with other vDSP operations, push `fileprivate` symbols to bottom of file.

* Add apply() function that returns the result.

* Added a version of DFT transform that returns the result.

* update some DFT comments

* Remove FFT overlays from this branch to separate DFT and FFT pull requests.

* [Accelerate] [vDSP] Swift Overlays to FFT Operations

This PR contains Swift overlays to simplify performing 1D and 2D fast Fourier transforms on single- and double-precision data.

This PR includes a subset of the operations (e.g. vDSP_fft_zrop for FFT). My plan is to add the additional operations (e.g. multiple signals, in-place operations, etc.) in a subsequent PR once this is approved.

cc: @moiseev @airspeedswift @stephentyrone

* Composite Branch

Contains FFT, DFT, DCT, and Biquad branches.
@stephentyrone
Copy link
Contributor Author

@swift-ci please test

@stephentyrone
Copy link
Contributor Author

@swift-ci please test source compatibility

@swift-ci
Copy link
Contributor

Build failed
Swift Test OS X Platform
Git Sha - 1508934

eeckstein and others added 2 commits April 25, 2019 15:45
…vImage-test

Temporarily disable a new vImage test that's failing for some platforms.
@stephentyrone
Copy link
Contributor Author

@swift-ci please test

@stephentyrone
Copy link
Contributor Author

@swift-ci please test source compatibility

@swift-ci
Copy link
Contributor

Build failed
Swift Test Linux Platform
Git Sha - 1508934

@swift-ci
Copy link
Contributor

Build failed
Swift Test OS X Platform
Git Sha - 1508934

@stephentyrone
Copy link
Contributor Author

Turns out the GYB_SOURCES change never made it onto 5.1 branch. Alas.

@stephentyrone
Copy link
Contributor Author

@swift-ci please test

@swift-ci
Copy link
Contributor

Build failed
Swift Test Linux Platform
Git Sha - 10f301c

@stephentyrone
Copy link
Contributor Author

@swift-ci please test

@swift-ci
Copy link
Contributor

Build failed
Swift Test Linux Platform
Git Sha - c6335c9

/// - Returns: The `c` in `c[i] = a[i] + b`.
@inline(__always)
@available(iOS 9999, macOS 9999, tvOS 9999, watchOS 9999, *)
public static func add<U>(_ scalar: Float,
Copy link
Contributor

@karwa karwa Apr 26, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a design point, and I know that this is an Apple framework and not necessarily subject to SE or any community involvement, but I'll posit that lots of developers who could use these functions won't use them because they're hard to discover. vDSP.add(6, myArray) is certainly nicer than the C name vDSP_vsadd but still, if you didn't already know it exists, you'd never find it.

Did you consider putting these functions on AccelerateBuffer directly? Alternatively, namespacing them would also be more discoverable than hanging them off of vDSP as static functions, IMO:

extension AccelerateBuffer {
  func vectorised() -> vDSP<Self> { return vDSP(data: self) }
}

struct vDSP<T: AccelerateBuffer> {
  var data: T
}
extension vDSP where T.Element == Double {
  func add(scalar: T.Element) {}
}

[1.0, 2, 3].vectorised().add(scalar: 6)

Just a thought. I'm happy to see these APIs improve 👍

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@karwa We definitely wouldn't want them on AB directly because we don't want every [Float] to pop up ~500 autocomplete suggestions unrelated to what most people are doing (worse, in some editors, the autocomplete suggestions don't take into account conditional conformance, so they'd appear on every Array; that's an editor bug, but we don't need to compound the pain).

Your vDSP wrapping function is plausible, but it's also worth considering that most of these operations do not have a primary source that makes a good LHS. Ultimately, this design was up to the Accelerate team, but I think they struck a reasonable compromise.

@swift-ci
Copy link
Contributor

Build failed
Swift Test OS X Platform
Git Sha - 10f301c

@stephentyrone stephentyrone merged commit 2802df1 into swiftlang:swift-5.1-branch Apr 26, 2019
@stephentyrone stephentyrone deleted the accelerate-5.1 branch April 26, 2019 11:49
stephentyrone added a commit to stephentyrone/swift that referenced this pull request Apr 26, 2019
* [Accelerate] [Quadrature] New Quadrature Overlay (swiftlang#23127)

* [Accelerate] [Quadrature] New Quadrature Overlay

A class that simplifies approximating the definite integral of a function.

* Fixes in response to PR Review.

Change `@_exported import Accelerate` to `import Accelerate`.
Correct date in copyright comment.

* Code Review Fixes.

* Remove mutable integrator, simply set options in init().
* Remove mutable `maxIntervals` and `qagPointsPerInterval`.
* Update tests.

* Code Review Fixes.

* Use standard library `Result`.
* Implement `QAGPointsPerInterval` as a struct.

* Tidy up Passing Integrand to `quadrature_integrate_function`.

Pass the integrand closure directly to the `quadrature_integrate_function` initialiser rather than attaching as a property of the `Quadrature` class.

* Make `Quadrature` a struct, and remove requirement for `integrand` to be escaping.

* Code Review Changes

* Add long-form integrator algorithm aliases (update tests to use these aliases).
* Make quadrature error description public.
* New tests for error description and `QAGPointsPerInterval`.

* Vectorized Integrand for Quadrature

Create an alternative implementation of `integrate` that accepts an integrand of type `(_ input: UnsafeBufferPointer<Double>, _ result: UnsafeMutableBufferPointer<Double>)`.

* Vectorized Integrand for Quadrature

Create an alternative implementation of `integrate` that accepts an integrand of type `(_ input: UnsafeBufferPointer<Double>, _ result: UnsafeMutableBufferPointer<Double>)`.

* Vectorized Integrand for Quadrature

Remove _ContiguousCollection - it's no longer used.

* Delete ContiguousCollection.swift

Not required.

* Refactor tests.

* Accelerate vForce (swiftlang#24152)

* [Accelerate] [vForce] New vForce Overlay

This PR contains a suite of overlays to the vForce transcendental and trigonometric functions on vectors of any length. The overlays simplify the API to the existing functions and accept collections that implement a new protocol called AccelerateBuffer. Conformances are provided for the most useful stdlib Collections.

* Merge pull request swiftlang#24184 from FlexMonkey/Accelerate_Swift-vDSP-Overlays

* Fix a case where vDSP is missing a cast (swiftlang#24197)

* Add missing cast to new vDSP_FIR overlay.

* Use macOS instead of OSX in availability annotations.

* Reorder Accelerate/CMakeLists.txt alphabetically.

* [Accelerate] [vImage] Swift Overlays (swiftlang#23592)

* Accelerate vImage Swift Overlays

A suite of functions, enumerations, and option sets to make working with vImage in Swift simpler.

* vImage_Buffer - new initializers to instantiate and initialize buffers with a single function call.
* vImage_Buffer - new functions to copy buffers and create CGImage instances from contents.
* vImage_CGImageFormat - new initializers.
* vImage_CGImageFormat - new equivalence operator.
* vImageConverter - methods to wrap free functions.
* vImageConverter - new make and convert functions.
* vImageCVImageFormat - new make functions.
* vImageCVImageFormat - methods to wrap free functions.
* vImage_Error - errorDescription function.
* vImage flags as an option set.

* Add new methods for generating CV -> CG and CG -> CV converters.

* update comments: `height` and `width` to `size`.

* `vImage_CGImageFormat.componentCount` should be `Int`.

* `vImage_CGImageFormat.componentCount` should be `Int`

* Move `self.init()` to after the size check for kvImageNoAllocate init.

* Buffer initializers to width and height rather than size.

* change vImage_CGImageFormat lightweight initializer to accept Int for `bitsPerComponent` and `bitsPerPixel`.

* Flesh out docs for vImage_Buffer.size

* Remove faux initializer in favor of new static function: `preferredAlignmentAndRowBytes`.

* Change functions to use proper error handling rather than inout error codes.

* Removed `flags` from basic init.

The only real flag to pass here is print diagnostics to console, and I now throw proper errors.

* Tests to check error throwing for buffer copy.

* remove unnecessary import, add missing docs.

* Add comments to error enums.

* Fix bug creating string from format code.

* Make `vImageCVImageFormat.formatCode` a `UInt32`.

* Remove equivalence operator from `CGImageFormat`.

* [Accelerate] [vDSP] DFT, FFT, DCT, and Biquad (swiftlang#24207)

* Swift overlay to vDSP cascaded biquad IIR filter.

* update comment in test

* Swift overlay to vDSP Discrete Cosine Transform

* Swift Overlays to vDSP Fast Fourier Transform and Discrete Fourier Transform

* Tests for DFT and FFT

* Some refactoring and begin documentation.

* Swift overlays to FFT and DFT operations.

* Expand docs.
* Implement `vDSP_destroy_fftsetup` and `vDSP_destroy_fftsetupD`

* Refactor Tests

* Refactor Tests

* Provide transform function that returns the result. Refactoring: parameter names more consistent with other vDSP operations, push `fileprivate` symbols to bottom of file.

* Add apply() function that returns the result.

* Added a version of DFT transform that returns the result.

* update some DFT comments

* Remove FFT overlays from this branch to separate DFT and FFT pull requests.

* [Accelerate] [vDSP] Swift Overlays to FFT Operations

This PR contains Swift overlays to simplify performing 1D and 2D fast Fourier transforms on single- and double-precision data.

This PR includes a subset of the operations (e.g. vDSP_fft_zrop for FFT). My plan is to add the additional operations (e.g. multiple signals, in-place operations, etc.) in a subsequent PR once this is approved.

cc: @moiseev @airspeedswift @stephentyrone

* Composite Branch

Contains FFT, DFT, DCT, and Biquad branches.

* Merge pull request swiftlang#24232 from stephentyrone/veclib-enum-availability

* Merge pull request swiftlang#24275 from stephentyrone/disable-broken-vImage-test

Temporarily disable a new vImage test that's failing for some platforms.

* Remove magic-symbols-for-install-name.c from Accelerate CMakeLists.txt

This file exists on master but not on 5.1.

* Apparently GYB_SOURCES hasn't made it onto 5.1.

* Slice should only conform to AB when the Base conforms to AB.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

6 participants