Enable CoreML, with fallback

[venkr]

This PR:

• Copies the coreml directory from whisper.cpp over.
• Adds flags WHISPER_USE_COREML and WHISPER_COREML_ALLOW_FALLBACK

As a result, this changes behavior of the .spm package such that if you have the CoreML model present, it'll use CoreML, otherwise the functionality will remain unchanged.

[vade]

Hey this is interesting. Few quick questions:

• Why do we need to have the implementation file which, I presume, is compiled by the CoreML runtime? My understanding is depending on the model you load, the compiled implementation may differ slightly, as well as if your runtime is Swift? Or is this required by Whisper.cpp for C++ interop?

[vizaiapp]

Hi folks, @venkr thank you for this change! I just tried to update and test this out, though I'm running into an issue. When I try to load whisper_init_from_file, the function breaks out of whatever context I call it in whenever a CoreML model is available.

Using the example in test-swift in the SPM package:

import Foundation
import whisper

let ctx = whisper_init_from_file("models/for-tests-ggml-base.en.bin")

var params = whisper_full_default_params(WHISPER_SAMPLING_GREEDY)

params.print_realtime   = true
params.print_progress   = false
params.print_timestamps = true
params.print_special    = false
params.translate        = false
//params.language         = "en";
params.n_threads        = 4
params.offset_ms        = 0

let n_samples = Int32(WHISPER_SAMPLE_RATE)
let pcmf32 = [Float](repeating: 0, count: Int(n_samples))

let ret = whisper_full(ctx, params, pcmf32, n_samples)
assert(ret == 0, "Failed to run the model")

let n_segments = whisper_full_n_segments(ctx)

for i in 0..<n_segments {
    let text_cur = whisper_full_get_segment_text(ctx, i)
    print(text_cur as Any)
}

whisper_print_timings(ctx)
whisper_free(ctx)

It doesn't execute past the initial whisper_init_from_file call. Is this the expected behavior? I'm able to run whisper.cpp fine with CoreML enabled.