Updated C wrapper wrt. Torch v1.10

[stemann]

Updates the C wrapper based on ocaml-torch @ 0.14 - matching Torch v1.10 (current JLL-build)

Contributes to #54 - follow-up for #56

Notable included changes:

• Updated C wrapper generator based on ocaml-torch @ 0.14 - matching Torch v1.10.
• Updated C wrapper (generated part)
• Updated C wrapper (manual part), torch_api.{cpp, h}
  • Re-implemented change of return type of functions to status code etc.
• Renamed C wrapper from doeye_caml to torch_c_api
• Restored Julia-specific additions
• Conditional CUDA build - C wrapper can be built without CUDA.
• Added buildkite stepsGitHub Actions workflow for building C wrapper
• Dev. container: Added support for CUDA and CUDNN

The last two changes could be moved to a separate PR (to reduce number of changes in this PR).

To-do:

• [x] Clean-up the automatic part (remove comments - fix indentation)
• [x] Adapt the manual part, torch_api.{cpp, h}
• [ ] The commit Changed torch_api.cpp to reduce diff should be removed before merging: It is meant to reduce the diff when reviewing - it's only a bunch of indentation changes etc. to make the diff smaller.

[zsz00]

What's the status now?

[stemann]

It's been a idle for a while, but status is summarised by these comments:

• https://github.com/FluxML/Torch.jl/pull/59#issuecomment-1663745853
• https://github.com/FluxML/Torch.jl/pull/60#issuecomment-1670840261

[stemann]

See #54 for status.

[DhairyaLGandhi]

Perhaps it would make sense to start merging some of the excellent changes here?

[stemann]

Yes! :-)

Please give torch_api.{cpp,h} a thorough review - I have made some changes in an effort to make things a bit more consistent - e.g. wrt. always returning an int status code.

Edit: I'll try to go over it as well and try to make a re-cap of the changes.

[stemann]

This is the current main diff which covers the hand-written part (torch_api.{cpp,h}): https://github.com/FluxML/Torch.jl/compare/7828132d..ece96546

[stemann]

The overall aim was to to update for Torch v1.10.2 - but also to make it easier to apply a diff of changes for subsequent version updates...

[stemann]

Recap:

General

• Removed argument names from function definitions in torch_api.h.
• Definitions in torch_api.h were updated to match the implementation in torch_api.cpp - e.g. wrt. return value in the first argument, and return code as return value.

Modified function definitions

int at_float_vec(double *values, int value_len, int type);

int at_int_vec(int64_t *values, int value_len, int type);

int at_grad_set_enabled(int);

int at_int64_value_at_indexes(double *i, tensor, int *indexes, int indexes_len);

tensor at_load(char *filename);

int ato_adam(optimizer *, double learning_rate,
                   double beta1,
                   double beta2,
                   double weight_decay);

int atm_load(char *, module *);

Added function definitions

int at_is_sparse(int *, tensor)
int at_device(int *, tensor)

int at_stride(tensor, int *)

int at_autocast_clear_cache();
int at_autocast_decrement_nesting(int *);
int at_autocast_increment_nesting(int *);
int at_autocast_is_enabled(int *);
int at_autocast_set_enabled(int *, int b);

int at_to_string(char **, tensor, int line_size)

int at_get_num_threads(int *);
int at_set_num_threads(int n_threads);

int ati_none(ivalue *);
int ati_bool(ivalue *, int);
int ati_string(ivalue *, char *);
int ati_tuple(ivalue *, ivalue *, int);
int ati_generic_list(ivalue *, ivalue *, int);
int ati_generic_dict(ivalue *, ivalue *, int);
int ati_int_list(ivalue *, int64_t *, int);
int ati_double_list(ivalue *, double *, int);
int ati_bool_list(ivalue *, char *, int);
int ati_string_list(ivalue *, char **, int);
int ati_tensor_list(ivalue *, tensor *, int);

int ati_to_string(char **, ivalue);
int ati_to_bool(int *, ivalue);
int ati_length(int *, ivalue);
int ati_to_generic_list(ivalue, ivalue *, int);
int ati_to_generic_dict(ivalue, ivalue *, int);
int ati_to_int_list(ivalue, int64_t *, int);
int ati_to_double_list(ivalue, double *, int);
int ati_to_bool_list(ivalue, char *, int);
int ati_to_tensor_list(ivalue, tensor *, int);

[stemann]

@DhairyaLGandhi Do you know if anyone is available for reviewing these changes?

I'm at JuliaCon, FYI.