Neural Engine, CoreML not utilized for Apple Silicon

[shiyuwang-jamk]

GPU was used in spite of the -ng parameter. Zero usage in ANE.

I have followed the steps in README for CoreML, and the log looks like this:

make clean
WHISPER_COREML=1 make -j

I whisper.cpp build info: 
I UNAME_S:  Darwin
I UNAME_P:  arm
I UNAME_M:  arm64
I CFLAGS:   -I.              -O3 -DNDEBUG -std=c11   -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_METAL
I CXXFLAGS: -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DWHISPER_USE_COREML -DGGML_USE_METAL
I LDFLAGS:   -framework Accelerate -framework Foundation -framework CoreML -framework Foundation -framework Metal -framework MetalKit
I CC:       Apple clang version 15.0.0 (clang-1500.3.9.4)
I CXX:      Apple clang version 15.0.0 (clang-1500.3.9.4)

cc  -I.              -O3 -DNDEBUG -std=c11   -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_METAL   -c ggml.c -o ggml.o
cc  -I.              -O3 -DNDEBUG -std=c11   -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_METAL   -c ggml-alloc.c -o ggml-alloc.o
cc  -I.              -O3 -DNDEBUG -std=c11   -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_METAL   -c ggml-backend.c -o ggml-backend.o
cc  -I.              -O3 -DNDEBUG -std=c11   -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_METAL   -c ggml-quants.c -o ggml-quants.o
c++ -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DWHISPER_USE_COREML -DGGML_USE_METAL -c whisper.cpp -o whisper.o
c++ -O3 -I . -fobjc-arc -c coreml/whisper-encoder.mm -o whisper-encoder.o
c++ -O3 -I . -fobjc-arc -c coreml/whisper-encoder-impl.m -o whisper-encoder-impl.o
cc -I.              -O3 -DNDEBUG -std=c11   -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DGGML_USE_ACCELERATE -DACCELERATE_NEW_LAPACK -DACCELERATE_LAPACK_ILP64 -DGGML_USE_METAL -c ggml-metal.m -o ggml-metal.o
c++ -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DWHISPER_USE_COREML -DGGML_USE_METAL examples/main/main.cpp examples/common.cpp examples/common-ggml.cpp examples/grammar-parser.cpp ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o whisper.o whisper-encoder.o whisper-encoder-impl.o ggml-metal.o -o main  -framework Accelerate -framework Foundation -framework CoreML -framework Foundation -framework Metal -framework MetalKit
c++ -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DWHISPER_USE_COREML -DGGML_USE_METAL examples/bench/bench.cpp ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o whisper.o whisper-encoder.o whisper-encoder-impl.o ggml-metal.o -o bench  -framework Accelerate -framework Foundation -framework CoreML -framework Foundation -framework Metal -framework MetalKit
c++ -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DWHISPER_USE_COREML -DGGML_USE_METAL examples/quantize/quantize.cpp examples/common.cpp examples/common-ggml.cpp examples/grammar-parser.cpp ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o whisper.o whisper-encoder.o whisper-encoder-impl.o ggml-metal.o -o quantize  -framework Accelerate -framework Foundation -framework CoreML -framework Foundation -framework Metal -framework MetalKit
c++ -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC -D_XOPEN_SOURCE=600 -D_DARWIN_C_SOURCE -pthread -DWHISPER_USE_COREML -DGGML_USE_METAL examples/server/server.cpp examples/common.cpp examples/common-ggml.cpp examples/grammar-parser.cpp ggml.o ggml-alloc.o ggml-backend.o ggml-quants.o whisper.o whisper-encoder.o whisper-encoder-impl.o ggml-metal.o -o server  -framework Accelerate -framework Foundation -framework CoreML -framework Foundation -framework Metal -framework MetalKit 
./main -h

./main -ng -l fi -otxt -ovtt -osrt -olrc -m "./models/ggml-large-v3.bin" -pp -f "../Henkilöauton ajokoe.wav" -of "whcpp-ml"

whisper_init_from_file_with_params_no_state: loading model from './models/ggml-large-v3.bin'
whisper_model_load: loading model
whisper_model_load: n_vocab       = 51866
whisper_model_load: n_audio_ctx   = 1500
whisper_model_load: n_audio_state = 1280
whisper_model_load: n_audio_head  = 20
whisper_model_load: n_audio_layer = 32
whisper_model_load: n_text_ctx    = 448
whisper_model_load: n_text_state  = 1280
whisper_model_load: n_text_head   = 20
whisper_model_load: n_text_layer  = 32
whisper_model_load: n_mels        = 128
whisper_model_load: ftype         = 1
whisper_model_load: qntvr         = 0
whisper_model_load: type          = 5 (large v3)
whisper_model_load: adding 1609 extra tokens
whisper_model_load: n_langs       = 100
whisper_model_load:      CPU total size =  3094.36 MB
whisper_model_load: model size    = 3094.36 MB
whisper_init_state: kv self size  =  220.20 MB
whisper_init_state: kv cross size =  245.76 MB
whisper_init_state: loading Core ML model from './models/ggml-large-v3-encoder.mlmodelc'
whisper_init_state: first run on a device may take a while ...
whisper_init_state: Core ML model loaded
whisper_init_state: compute buffer (conv)   =   10.92 MB
whisper_init_state: compute buffer (cross)  =    9.38 MB
whisper_init_state: compute buffer (decode) =  209.26 MB

system_info: n_threads = 4 / 8 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | METAL = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 | CUDA = 0 | COREML = 1 | OPENVINO = 0

main: processing '../Henkilöauton ajokoe.wav' (20244723 samples, 1265.3 sec), 4 threads, 1 processors, 5 beams + best of 5, lang = fi, task = transcribe, timestamps = 1 ...


[00:00:00.000 --> 00:00:02.000]   Terve. Moikka.
[00:00:02.000 --> 00:00:04.000]   Mä voin mennä sajelemaan. Joo.
[00:00:04.000 --> 00:00:06.000]   Joo, Mika. Petra, moi.
[00:00:06.000 --> 00:00:08.000]   Okei, mä voin ottaa sulta ne paperit sieltä.
[00:00:08.000 --> 00:00:10.000]   Siitä se ja se.
[00:00:10.000 --> 00:00:12.000]   All right, mennään tonne autoon. Joo.
[00:00:12.000 --> 00:00:14.000]   Katsotaan siellä loppuun.
[00:00:14.000 --> 00:00:23.000]   All right, sit mä tarkastan, et mul on oikee puski mukana.

[ggerganov]

Don't add -ng and adjust the following parameter in the code to enable ANE and CPU-only:

whisper.cpp/coreml/whisper-encoder.mm

Lines 25 to 29 in 3928dbd

	// select which device to run the Core ML model on
	MLModelConfiguration *config = [[MLModelConfiguration alloc] init];
	// config.computeUnits = MLComputeUnitsCPUAndGPU;
	//config.computeUnits = MLComputeUnitsCPUAndNeuralEngine;
	config.computeUnits = MLComputeUnitsAll;

[shiyuwang-jamk]

Thanks for the compiling tip. It does not seem I can monitor neural engine usage with powermetrics as it does for CPUs and GPUs.