GGML_CPU_ALL_VARIANTS causes a silent exception

[peardox]

Building on Win11 using VC like this (CMAKE_TOOLCHAIN_FILE is for SDL2)

cmake -B build -DCMAKE_TOOLCHAIN_FILE=C:\git\vcpkg\scripts\buildsystems\vcpkg.cmake -DGGML_CCACHE=OFF -DGGML_CUDA=1 -DGGML_VULKAN=1 -DWHISPER_SDL2=ON -DGGML_BACKEND_DL=ON -DBUILD_SHARED_LIBS=ON -DGGML_CPU_ALL_VARIANTS=ON

This should (and does) build CPU/CUDA/Vulkan for all x64 CPUs back to Haswell, the aim, of course, being to allow for older CPU architectures so the same apps will run on a wide range of end-user configs

However, when it comes to loading GGML it fails silently. Looking into this but my C++ is rusty

[peardox]

PR #3042

Finally traced down the issue.

If e.g. whisper_init_from_file_with_params is called in a GGML_BACKEND_DL build the examples have no backend yet but something within the init is calling the backend

I presume there's a list of backends which is empty when called via GGML_BACKEND_DL as a null pointer is eventually returned where it shouldn't be.

Not sure why it works without GGML_BACKEND_DL but getting it working was the big goal for me.

I've modified bench, cli and stream to call whisper_load_backends() which I have added to whisper.h

If whisper_load_backends() is called without GGML_BACKEND_DL it does nothing (existing code) so the extra header function is benign

[peardox]

PR failed so made the whisper.h change conditional as well so as not to confuse things

[peardox]

As suggested tried ggml_backend_load_all then added ggml_backend_load_best.
The benefit of adding a load_best is that it may be preferable to use e.g. vulkan over CUDA in very specific circumstances - namely one of the new AMD APUs with a Radeon 8060S Graphics subsystem as it MAY be faster than CUDA if in a machine with both (8060S = 4060 so they say)
A side-effect of load_best is that a command-line switch could specify desired backend - cpu/vulkan/cuda etc