[STABLE ABI] Porting audio to Torch Stable ABI

[pearu]

This is a tracking issue for porting torchaudio extension modules to use torch stable ABI.

Must have:

• AT_DISPATCH_FLOATING_TYPES_AND_HALF, AT_DISPATCH_FLOATING_TYPES. Prototypes of these macros are implemented in [STABLE ABI] Port forced_align #4079 (see dispatch.h) and moved to upstream, see [STABLE ABI] Add STABLE_DISPATCH_... CPP macros pytorch#163973. The plan is to move these macros to headeronly, see ghstack starting from Move AT_FORALL_... macros and ScalarTypeToCPPTypeT to headeronly pytorch#164350 and Refactor AT_DISPATCH_CASE_... macros to headeronly pytorch#165695

  UPDATE: when Move AT_DISPATCH_V2 helper macros to headeronly and add THO_DISPATCH_V2_TMPL pytorch#165856 lands, we'll use AT_DISPATCH V2 macros.

• mutable/const_data_ptr methods and templates. Prototypes are implemented in [STABLE ABI] Port forced_align #4079 (see ops.h) and moved to upstream, see [STABLE ABI] Add mutable_data_ptr() and const_data_ptr() methods to torch::stable::Tensor. pytorch#161891

  • Why we need this: dependency of accessors

• tensor accessors for both CPU and CUDA. Prototypes of these templates are implemented in [STABLE ABI] Port forced_align #4079 (see TensorAccessor.h) and moved to upstream, see [STABLE ABI] Add tensor accessors. pytorch#164123, landing requires ArrayRef

  • Why we need this: The codebase currently relies on accessors [with up to 3 dimensions] (

    audio/src/libtorchaudio/forced_align/cpu/compute.cpp

    Lines 40 to 42 in 87ff22e

    	auto logProbs_a = logProbs.accessor<scalar_t, 3>();
    	auto targets_a = targets.accessor<target_t, 2>();
    	auto paths_a = paths.accessor<target_t, 2>();

    ). We technically could work around this by relying on raw pointers and enforcing contiguity, but that would obfuscate what would otherwise be simple indexing calls.

  UPDATE: Refactor TensorAccessor for headeronly. pytorch#166855 refactors TensorAccessor to headeronly

• parallel_for, in-progress by torch stable ABI team: Add stable parallel_for pytorch#161320

  • Alternative is to not use parallel_for, for decreased performance. Acceptable, but not ideal.

Nice to have:

We think we can reasonably work around these without incurring too much debt in TorchAudio, so we can treat them as nice-to-have for now.

• Tensor operations as stable ABI operations:
  • item<T>() template, prototype available in [STABLE ABI] Port forced_align #4079 (see ops.h)
  • to function (to cuda, to cpu, as an alternative to cpu and cuda functions below)
  • cpu function, prototype available in [STABLE ABI] Port forced_align #4079 (see ops.h), moved to upstream, see [STABLE ABI] Add cpu operation. pytorch#161911
  • cuda function, prototype available in [STABLE ABI] Port forced_align #4079 (see ops.h)
  • copy_ function, done in [STABLE ABI] Add clone method to torch::stable::Tensor pytorch#161896
  • clone function, done in [STABLE ABI] Add copy_ operation. pytorch#161895
  • index function, requires Slice, workaround using other aten methods
  • new_zeros function, prototype available in [STABLE ABI] Port forced_align #4079 (see ops.h)
  • new_empty with device support, prototypes in [STABLE ABI] Port forced_align #4079 and [STABLE ABI] Add device_type and device_index optional arguments to new_empty. pytorch#161894
  • tensor function
  • max function
  • select function
  • unsqueeze function
  • squeeze function