Swift-to-Rust CoreML Bridge#

This directory contains the Swift code that wraps Apple's CoreML framework for depth estimation, making it callable from Rust via FFI.

Architecture#

Rust (src/depth_coreml.rs)
    ↕ FFI (extern "C")
Swift (CoreMLDepth.swift)
    ↕ Apple Frameworks
CoreML + Neural Engine

Files#

CoreMLDepth.swift — Swift wrapper providing C-compatible API for CoreML
README.md — This file

API#

coreml_load_model#

@_cdecl("coreml_load_model")
public func loadModel(_ pathPtr: UnsafePointer<CChar>) -> UnsafeMutableRawPointer?

Loads and compiles a CoreML model from an .mlpackage file.

Returns: Opaque pointer to MLModel, or NULL on error

coreml_infer_depth#

@_cdecl("coreml_infer_depth")
public func inferDepth(
    _ modelPtr: UnsafeMutableRawPointer,
    _ rgbData: UnsafePointer<Float>,
    _ width: Int32,
    _ height: Int32,
    _ outputPtr: UnsafeMutablePointer<Float>
) -> Int32

Runs depth estimation inference.

Input:

modelPtr: Model from coreml_load_model
rgbData: NCHW Float32 array [1, 3, H, W]
width, height: Input dimensions
outputPtr: Buffer for output [H * W floats]

Returns: 0 on success, negative error code on failure

coreml_unload_model#

@_cdecl("coreml_unload_model")
public func unloadModel(_ modelPtr: UnsafeMutableRawPointer)

Releases the model.

Data Types#

Input#

Format: NCHW (batch, channels, height, width)
Type: Float32
Shape: [1, 3, 518, 518]
Normalized: ImageNet mean/std

Output#

Format: HW (height, width)
Type: Float16 (model) → Float32 (converted for Rust)
Shape: [518, 518]
Range: Raw depth values (not normalized)

Compilation#

This Swift code is compiled by build.rs when building the Rust crate:

swiftc -emit-library -static -module-name CoreMLDepth -O \
    swift-bridge/CoreMLDepth.swift -o libCoreMLDepth.a

Linked frameworks:

CoreML.framework
CoreVideo.framework
Accelerate.framework
Foundation.framework

Error Handling#

Error codes returned by coreml_infer_depth:

0: Success
-1: Failed to create input MLMultiArray
-2: Inference failed
-3: Failed to extract depth output
-4: General error (see stdout for details)

Performance Notes#

First Load#

Compiles .mlpackage to .mlmodelc (~7.5s)
Cached in /var/folders/.../T/

Subsequent Loads#

Uses cached .mlmodelc
Fast (~10ms)

Inference#

Base model: ~128ms on M4 Pro
Uses Apple Neural Engine + GPU + CPU
Configured via config.computeUnits = .all

Testing#

Test Swift code directly:

import CoreML

let path = "checkpoints/DepthAnythingV2BaseF16.mlpackage"
let url = URL(fileURLWithPath: path)

let compiledURL = try MLModel.compileModel(at: url)
let config = MLModelConfiguration()
config.computeUnits = .all
let model = try MLModel(contentsOf: compiledURL, configuration: config)

print("Model loaded: \(model.modelDescription)")

Debugging#

Enable CoreML logging:

export COREML_VERBOSE=1
cargo run --example photo_coreml

Requirements#

macOS 15+ (for Float16 support)
Xcode Command Line Tools
Swift 5.0+
Apple Silicon (for Neural Engine)

Configure Feed