[ET-VK][Ops] aten.index_select #3744
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
## The Operator `nn.Module` invocations of [`torch.index_select`](https://pytorch.org/docs/stable/generated/torch.index_select.html) get compiled to `aten.index_select.default` in the Edge Dialect, which carries the following signature. ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` ## Implementation This is a C-packing-only implementation. It is very similar to `aten.slice`: #3171 ``` - func: slice.Tensor(Tensor(a) self, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a) ``` It features a similar split between a shader for N,H,W and a shader for C, because copying from the C-dimension is more difficult due to C-packing. Both `index_select` and `slice` copy specific indices across 1 dimension. The difference is in the way these indices are specified. - `slice` uses `start=1`/`end=5`/`step=2` as three scalars for indices `1,3`. - `index_select` lists the exact indices inside a tensor e.g. `index=torch.tensor([1,3])`. Hence, `slice` uses a `offset=1` and `step=2` to compute input position. In `index_select`, we read the index tensor to compute input position. Differential Revision: [D57745489](https://our.internmc.facebook.com/intern/diff/D57745489/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/3744
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 7b36eff with merge base 1343224 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D57745489 |
jorgep31415
added a commit
that referenced
this pull request
May 24, 2024
## The Operator `nn.Module` invocations of [`torch.index_select`](https://pytorch.org/docs/stable/generated/torch.index_select.html) get compiled to `aten.index_select.default` in the Edge Dialect, which carries the following signature. ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` ## Implementation This is a C-packing-only implementation. It is very similar to `aten.slice`: #3171 ``` - func: slice.Tensor(Tensor(a) self, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a) ``` It features a similar split between a shader for N,H,W and a shader for C, because copying from the C-dimension is more difficult due to C-packing. Both `index_select` and `slice` copy specific indices across 1 dimension. The difference is in the way these indices are specified. - `slice` uses `start=1`/`end=5`/`step=2` as three scalars for indices `1,3`. - `index_select` lists the exact indices inside a tensor e.g. `index=torch.tensor([1,3])`. Hence, `slice` uses a `offset=1` and `step=2` to compute input position. In `index_select`, we read the index tensor to compute input position. Differential Revision: [D57745489](https://our.internmc.facebook.com/intern/diff/D57745489/) ghstack-source-id: 227736336 Pull Request resolved: #3744
copyrightly
approved these changes
May 28, 2024
|
This pull request has been merged in c665c17. |
jorgep31415
added a commit
that referenced
this pull request
May 29, 2024
## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D57745489 |
jorgep31415
added a commit
that referenced
this pull request
May 29, 2024
## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/) ghstack-source-id: 228038402 Pull Request resolved: #3762
jorgep31415
added a commit
that referenced
this pull request
May 29, 2024
## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/) [ghstack-poisoned]
jorgep31415
added a commit
that referenced
this pull request
May 29, 2024
## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/) [ghstack-poisoned]
jorgep31415
added a commit
that referenced
this pull request
May 29, 2024
## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/) [ghstack-poisoned]
jorgep31415
added a commit
that referenced
this pull request
May 29, 2024
## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/) [ghstack-poisoned]
jorgep31415
added a commit
that referenced
this pull request
May 29, 2024
## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/) [ghstack-poisoned]
jorgep31415
added a commit
that referenced
this pull request
May 29, 2024
## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/) [ghstack-poisoned]
facebook-github-bot
pushed a commit
that referenced
this pull request
May 30, 2024
Summary: Pull Request resolved: #3762 ## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: #3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` ghstack-source-id: 228201965 Reviewed By: copyrightly Differential Revision: D57880520 fbshipit-source-id: 67da04bcbb2b36ce2c1ec2c8f7ccf59ed512547c
kedarnath03
pushed a commit
to kedarnath03/executorch
that referenced
this pull request
Jun 25, 2025
Pull Request resolved: pytorch/executorch#3762 ## The Operator `nn.Module` invocations on the embedding returned by [`torch.nn.Embedding`](https://pytorch.org/docs/stable/generated/torch.nn.Embedding.html) get compiled to `aten.embedding.default` in the Edge Dialect, which carries the following signature. ``` - func: embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor ``` ## Implementation This is a C-packing-only implementation. Interestingly, the 1D-`indices` case is equivalent to the `dim=0` case of the preceding `aten.index_select`: pytorch/executorch#3744 ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` I naïvely thought the rest of the operator would be similarly easy but it wasn't. The 2D and 3D-`indices` cases are more involved to the extent that we require a standalone `cpp`/`glsl` file. ## Codegen We add support for making 2D and 3D index tensors. This requires new generation functions as well as renaming of the `case_name` string to recursively handle list `pylist`s. ``` // 1D Test(weight=[10, 9], indices=[0, 2]), // 2D Test(weight=[10, 9], indices=[[0, 2], [1, 4], [7, 7]]), // 3D Test(weight=[10, 9], indices=[[[3, 1, 4], [1, 5, 9]], [[2, 6, 5], [3, 5, 8]]]), ``` ghstack-source-id: 228201965 Differential Revision: [D57880520](https://our.internmc.facebook.com/intern/diff/D57880520/)
kedarnath03
pushed a commit
to kedarnath03/executorch
that referenced
this pull request
Jun 25, 2025
Pull Request resolved: pytorch/executorch#3744 ## The Operator `nn.Module` invocations of [`torch.index_select`](https://pytorch.org/docs/stable/generated/torch.index_select.html) get compiled to `aten.index_select.default` in the Edge Dialect, which carries the following signature. ``` - func: index_select(Tensor self, int dim, Tensor index) -> Tensor ``` ## Implementation This is a C-packing-only implementation. It is very similar to `aten.slice`: pytorch/executorch#3171 ``` - func: slice.Tensor(Tensor(a) self, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a) ``` It features a similar split between a shader for N,H,W and a shader for C, because copying from the C-dimension is more difficult due to C-packing. Both `index_select` and `slice` copy specific indices across 1 dimension. The difference is in the way these indices are specified. - `slice` uses `start=1`/`end=5`/`step=2` as three scalars for indices `1,3`. - `index_select` lists the exact indices inside a tensor e.g. `index=torch.tensor([1,3])`. Hence, `slice` uses a `offset=1` and `step=2` to compute input position. In `index_select`, we read the index tensor to compute input position. Differential Revision: [D57745489](https://our.internmc.facebook.com/intern/diff/D57745489/) ghstack-source-id: 227954599
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
The Operator
nn.Moduleinvocations oftorch.index_selectget compiled toaten.index_select.defaultin the Edge Dialect, which carries the following signature.Implementation
This is a C-packing-only implementation. It is very similar to
aten.slice: #3171It features a similar split between a shader for N,H,W and a shader for C, because copying from the C-dimension is more difficult due to C-packing.
Both
index_selectandslicecopy specific indices across 1 dimension. The difference is in the way these indices are specified.sliceusesstart=1/end=5/step=2as three scalars for indices1,3.index_selectlists the exact indices inside a tensor e.g.index=torch.tensor([1,3]).Hence,
sliceuses aoffset=1andstep=2to compute input position. Inindex_select, we read the index tensor to compute input position.Differential Revision: D57745489