Updated C wrapper wrt. Torch v1.10 #61

stemann · 2023-08-11T14:59:32Z

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 steps~~GitHub 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:

Clean-up the automatic part (remove comments - fix indentation)
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 · 2023-10-21T10:14:01Z

What's the status now?

stemann · 2023-10-22T11:48:27Z

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

stemann · 2023-11-11T09:44:33Z

See #54 for status.

Copied from https://github.com/LaurentMazare/ocaml-torch/tree/a6499811f40282a071179d4306afbbb6023dcc4a/src/gen/gen.ml Also updated dune-project accordingly.

…a torch wrapper

A. Non-void, non-* methods. Search/replace: 1. torch_api.h: ^(?!void)(\w+) (at.+)$ -> int $2($1 *, 2. torch_api.h: , $;$ -> ); 3. torch_api.cpp: ^(?!void)(\w+) (at.+)$ -> int $2($1 *out__, 4. torch_api.cpp: , $ \{$ -> ) { B. void-methods. Search/replace 1. torch_api.{h,cpp}: ^void (at.+)$ -> int $1( C. *-methods. Search/replace 1. torch.api.h: ^(\w+ \*)(at.+)\( -> int $2($1*, 2. torch_api.cpp: ^(\w+ \*)(at.+)\( -> int $2($1*out__, D. Implemented return status code Replaced ``` ^(\s*) return new (.+) \s*$ \s*return nullptr; ``` with ``` $1 out__[0] = new $2 $1 return 0; $1) $1return 1; ``` E. Implemented return status code 1. Replaced ``` ^(\s*)PROTECT$return new (.+)$ \s*return nullptr; ``` with ``` $1PROTECT( $1 out__[0] = new $2 $1 return 0; $1) $1return 1; ``` F. Implemented return status code 1. Replaced ``` ^(\s*)PROTECT$return (.+)$ ``` with ``` $1PROTECT( $1 out__[0] = $2 $1 return 0; $1) ``` G. Implemented return status code 1. Replaced ``` ^(\s*) return (.+) \s*\) \s*return nullptr; ``` with ``` $1 out__[0] = $2 $1 return 0; $1) $1return 1; ``` H. Restored error handling Handled caml_failwith by search/replace: Replaced: ``` $ ^(\s*) caml_failwith$(.+) ``` with: ``` { $1 myerr = strdup($2 $1 return 1; $1} ``` I. Replaced ``` ^(\s+)PROTECT\( return (.+) $ ``` with ``` $1PROTECT( $1 out__[0] = $2 $1 return 0; $1) ``` J. Manual implement return status code K. Changed return code from -1 to 1 to reduce diff L. Fixed a couple of warnings

Also, made CUDA build optional. int at_empty_cache(); int at_no_grad(int flag); int at_sync(); int at_from_blob(tensor *, void *data, int64_t *dims, int ndims, int64_t *strides, int nstrides, int dev);

Also: * Dev. container: Updated for Torch 1.10.2 * Added /build to .gitignore

Factored out scripts - for re-use in CI and dev. container.

DhairyaLGandhi · 2024-04-29T12:28:28Z

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

stemann · 2024-04-29T12:39:22Z

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 · 2024-04-29T12:47:35Z

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 · 2024-04-29T13:05:43Z

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 · 2024-04-29T13:22:46Z

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 · 2024-07-09T11:24:57Z

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

I'm at JuliaCon, FYI.

stemann force-pushed the stemann/ocaml_torch_1.10 branch from acb50ff to 6d3200a Compare August 11, 2023 15:05

stemann mentioned this pull request Nov 10, 2023

Fixed compat to enable tests on CI #60

Merged

stemann force-pushed the stemann/ocaml_torch_1.10 branch from 6d3200a to 6afc28e Compare November 11, 2023 09:36

stemann mentioned this pull request Nov 11, 2023

Update wrapper #54

Open

stemann force-pushed the stemann/ocaml_torch_1.10 branch from 6afc28e to f407437 Compare November 12, 2023 11:44

stemann added 3 commits November 12, 2023 12:45

Added wrapper generator from ocaml-torch 0.14

28550c8

Copied from https://github.com/LaurentMazare/ocaml-torch/tree/a6499811f40282a071179d4306afbbb6023dcc4a/src/gen/gen.ml Also updated dune-project accordingly.

Adapted wrapper_generator to produce equivalent code to existing juli…

57ded88

…a torch wrapper

Removed writing of OCaml stubs and wrapper

f374bf6

stemann force-pushed the stemann/ocaml_torch_1.10 branch 19 times, most recently from 6f18889 to beb24f2 Compare November 17, 2023 10:38

stemann force-pushed the stemann/ocaml_torch_1.10 branch 2 times, most recently from 090f09e to aa302ad Compare April 29, 2024 10:28

stemann added 2 commits April 29, 2024 12:29

Restored Julia-specific additions

ece9654

Also, made CUDA build optional. int at_empty_cache(); int at_no_grad(int flag); int at_sync(); int at_from_blob(tensor *, void *data, int64_t *dims, int ndims, int64_t *strides, int nstrides, int dev);

Dev. container: Added support for CUDA and CUDNN

c8f1e9c

Also: * Dev. container: Updated for Torch 1.10.2 * Added /build to .gitignore

stemann force-pushed the stemann/ocaml_torch_1.10 branch 12 times, most recently from 7f5dbfa to aa10911 Compare April 29, 2024 12:13

stemann added 2 commits April 29, 2024 14:18

Added GitHub Actions workflow for building C Wrapper

f799133

Factored out scripts - for re-use in CI and dev. container.

Fixed C wrapper wrt. formatting

37c6056

stemann force-pushed the stemann/ocaml_torch_1.10 branch from aa10911 to 37c6056 Compare April 29, 2024 12:18

stemann marked this pull request as ready for review April 29, 2024 12:42

stemann mentioned this pull request May 7, 2024

CUDA: Introduce arm64 subarchitecture. JuliaPackaging/Yggdrasil#8622

Merged

ToucheSir approved these changes Aug 8, 2024

View reviewed changes

ToucheSir merged commit d1711d7 into FluxML:master Aug 8, 2024
4 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Updated C wrapper wrt. Torch v1.10 #61

Updated C wrapper wrt. Torch v1.10 #61

stemann commented Aug 11, 2023 •

edited

Loading

zsz00 commented Oct 21, 2023

stemann commented Oct 22, 2023

stemann commented Nov 11, 2023

DhairyaLGandhi commented Apr 29, 2024

stemann commented Apr 29, 2024 •

edited

Loading

stemann commented Apr 29, 2024

stemann commented Apr 29, 2024

stemann commented Apr 29, 2024 •

edited

Loading

stemann commented Jul 9, 2024 •

edited

Loading

Updated C wrapper wrt. Torch v1.10 #61

Updated C wrapper wrt. Torch v1.10 #61

Conversation

stemann commented Aug 11, 2023 • edited Loading

zsz00 commented Oct 21, 2023

stemann commented Oct 22, 2023

stemann commented Nov 11, 2023

DhairyaLGandhi commented Apr 29, 2024

stemann commented Apr 29, 2024 • edited Loading

stemann commented Apr 29, 2024

stemann commented Apr 29, 2024

stemann commented Apr 29, 2024 • edited Loading

General

Modified function definitions

Added function definitions

stemann commented Jul 9, 2024 • edited Loading

stemann commented Aug 11, 2023 •

edited

Loading

stemann commented Apr 29, 2024 •

edited

Loading

stemann commented Apr 29, 2024 •

edited

Loading

stemann commented Jul 9, 2024 •

edited

Loading