Skip to content

[ET-VK] Enable dynamic operator registration #2305

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 4 commits into from

Conversation

jorgep31415
Copy link
Contributor

@jorgep31415 jorgep31415 commented Mar 7, 2024

Stack from ghstack (oldest at bottom):

This change follows

  1. in the footsteps of Implement global shader registry #2222 for static initialization and
  2. the popular TorchLibraryImpl for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using REGISTER_OPERATORS + VK_REGISTER_OP(), as shown in Arithmetic.h/cpp.

Typically in Linux/Android C++ environments, the symbols corresponding to OperatorRegisterInit static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to link_whole = True for the vulkan_graph_runtime library.

We update our Compute API tests to verify we can go through OperatorRegistry with proper static initialization.

Differential Revision: D54641117

This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)

[ghstack-poisoned]
Copy link

pytorch-bot bot commented Mar 7, 2024

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/2305

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure, 1 Unrelated Failure

As of commit 5675406 with merge base 47b837b (image):

NEW FAILURE - The following job has failed:

BROKEN TRUNK - The following job failed but were present on the merge base:

👉 Rebase onto the `viable/strict` branch to avoid these failures

This comment was automatically generated by Dr. CI and updates every 15 minutes.

@facebook-github-bot facebook-github-bot added the CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. label Mar 7, 2024
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D54641117

jorgep31415 added a commit that referenced this pull request Mar 7, 2024
This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)

ghstack-source-id: 217857140
Pull Request resolved: #2305
This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)

[ghstack-poisoned]
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D54641117

jorgep31415 added a commit that referenced this pull request Mar 7, 2024
Pull Request resolved: #2305

This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.
ghstack-source-id: 217867140
@exported-using-ghexport

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)
This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)

[ghstack-poisoned]
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D54641117

jorgep31415 added a commit that referenced this pull request Mar 7, 2024
Pull Request resolved: #2305

This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.
ghstack-source-id: 217871298
@exported-using-ghexport

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)
jorgep31415 added a commit that referenced this pull request Mar 8, 2024
This must land before #2305.

Otherwise, we will have two definitions of the same function signature in `executorch/backends/vulkan/runtime/graph/ops/OpUtils.cpp` and `ATen/native/vulkan/impl/Common.cpp`, resulting in a linker error, when building `vulkan_compute_api_test_bin`:
```
ld.lld: error: duplicate symbol: at::native::vulkan::adaptive_work_group_size(at::native::vulkan::api::utils::detail::vec<unsigned int, 3u> const&)
>>> defined at Common.cpp:8 (./xplat/caffe2/aten/src/ATen/native/vulkan/impl/Common.cpp:8)
>>>            __objects__/aten/src/ATen/native/vulkan/impl/Common.cpp.o:(at::native::vulkan::adaptive_work_group_size(at::native::vulkan::api::utils::detail::vec<unsigned int, 3u> const&)) in archive buck-out/v2/gen/fbsource/a9b839d0bc77bff6/xplat/caffe2/__torch_vulkan_ops__/libtorch_vulkan_ops.a
>>> defined at OpUtils.cpp:16 (./xplat/executorch/backends/vulkan/runtime/graph/ops/OpUtils.cpp:16)
>>>            __objects__/runtime/graph/ops/OpUtils.cpp.o:(.text._ZN2at6native6vulkan24adaptive_work_group_sizeERKNS1_3api5utils6detail3vecIjLj3EEE+0x0) in archive buck-out/v2/gen/fbsource/a9b839d0bc77bff6/xplat/executorch/backends/vulkan/__vulkan_graph_runtime__/libvulkan_graph_runtime.a
clang-15: error: linker command failed with exit code 1 (use -v to see invocation)
```

These files are part of the `torch_vulkan_ops` and `vulkan_runtime_graph` libraries, respectively.

We resolve the issue by removing the dependency of `vulkan_compute_api_test_bin` to on `torch_vulkan_ops`. Unfortunately, this requires copy-pasting more code to `VulkanBackend.cpp`.

Differential Revision: [D54659269](https://our.internmc.facebook.com/intern/diff/D54659269/)

[ghstack-poisoned]
This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)

[ghstack-poisoned]
@facebook-github-bot
Copy link
Contributor

This pull request was exported from Phabricator. Differential Revision: D54641117

jorgep31415 added a commit that referenced this pull request Mar 8, 2024
Pull Request resolved: #2305

This change follows
1. in the footsteps of #2222 for static initialization and
2. the popular `TorchLibraryImpl` for wrapping with macros.

https://www.internalfb.com/code/fbsource/[b6860acf0fd7a95224f2ed3f6fe48f699a9a45c0]/fbcode/caffe2/torch/library.h?lines=1004%2C1012-1026

Contributors can now write their operator and register them within the same file using `REGISTER_OPERATORS` + `VK_REGISTER_OP()`, as shown in `Arithmetic.h/cpp`.

Typically in Linux/Android C++ environments, the symbols corresponding to `OperatorRegisterInit` static instances are discarded since they aren't used for anything other than static initialization. Hence, we need to `link_whole = True` for the `vulkan_graph_runtime` library.

We update our Compute API tests to verify we can go through `OperatorRegistry` with proper static initialization.
ghstack-source-id: 217884083
@exported-using-ghexport

Differential Revision: [D54641117](https://our.internmc.facebook.com/intern/diff/D54641117/)
facebook-github-bot pushed a commit that referenced this pull request Mar 8, 2024
Summary:
bypass-github-export-checks

Pull Request resolved: #2312

This must land before #2305.

Otherwise, we will have two definitions of the same function signature in `executorch/backends/vulkan/runtime/graph/ops/OpUtils.cpp` and `ATen/native/vulkan/impl/Common.cpp`, resulting in a linker error, when building `vulkan_compute_api_test_bin`:
```
ld.lld: error: duplicate symbol: at::native::vulkan::adaptive_work_group_size(at::native::vulkan::api::utils::detail::vec<unsigned int, 3u> const&)
>>> defined at Common.cpp:8 (./xplat/caffe2/aten/src/ATen/native/vulkan/impl/Common.cpp:8)
>>>            __objects__/aten/src/ATen/native/vulkan/impl/Common.cpp.o:(at::native::vulkan::adaptive_work_group_size(at::native::vulkan::api::utils::detail::vec<unsigned int, 3u> const&)) in archive buck-out/v2/gen/fbsource/a9b839d0bc77bff6/xplat/caffe2/__torch_vulkan_ops__/libtorch_vulkan_ops.a
>>> defined at OpUtils.cpp:16 (./xplat/executorch/backends/vulkan/runtime/graph/ops/OpUtils.cpp:16)
>>>            __objects__/runtime/graph/ops/OpUtils.cpp.o:(.text._ZN2at6native6vulkan24adaptive_work_group_sizeERKNS1_3api5utils6detail3vecIjLj3EEE+0x0) in archive buck-out/v2/gen/fbsource/a9b839d0bc77bff6/xplat/executorch/backends/vulkan/__vulkan_graph_runtime__/libvulkan_graph_runtime.a
clang-15: error: linker command failed with exit code 1 (use -v to see invocation)
```

These files are part of the `torch_vulkan_ops` and `vulkan_runtime_graph` libraries, respectively.

We resolve the issue by removing the dependency of `vulkan_compute_api_test_bin` to on `torch_vulkan_ops`. Unfortunately, this requires copy-pasting more code to `VulkanBackend.cpp`.
ghstack-source-id: 217882620
exported-using-ghexport

Reviewed By: SS-JIA

Differential Revision: D54659269

fbshipit-source-id: 4d780c08d2a75b9e36ca1d646f5794a285b8086b
@facebook-github-bot
Copy link
Contributor

This pull request has been merged in f3487ed.

jorgep31415 added a commit that referenced this pull request Mar 8, 2024
In #2305, I added the `soname = "libvulkan_graph_runtime.$(ext)"` to the wrong library. Noob mistake. Also, removed an unneeded dependency.

Differential Revision: [D54688153](https://our.internmc.facebook.com/intern/diff/D54688153/)

[ghstack-poisoned]
jorgep31415 added a commit that referenced this pull request Mar 8, 2024
In #2305, I added the `soname = "libvulkan_graph_runtime.$(ext)"` to the wrong library. Noob mistake. Also, removed an unneeded dependency.

Differential Revision: [D54688153](https://our.internmc.facebook.com/intern/diff/D54688153/)

ghstack-source-id: 217969105
Pull Request resolved: #2319
facebook-github-bot pushed a commit that referenced this pull request Mar 8, 2024
Summary:
bypass-github-export-checks

Pull Request resolved: #2319

In #2305, I added the `soname = "libvulkan_graph_runtime.$(ext)"` to the wrong library. Noob mistake. Also, I'd forgotten to remove an unneeded dependency.

Reviewed By: SS-JIA

Differential Revision: D54688153

fbshipit-source-id: 67a1316c76180030d4ae5225f16f01dcc750220a
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
CLA Signed This label is managed by the Facebook bot. Authors need to sign the CLA before a PR can be reviewed. fb-exported Merged
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants