[ET-VK] Introduce AOT operator registry #6488
Merged
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
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
@update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6488
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit aa0d67f with merge base 16b633b ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
SS-JIA
added a commit
that referenced
this pull request
Oct 24, 2024
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
ghstack-source-id: 249892740
Pull Request resolved: #6488
|
This pull request was exported from Phabricator. Differential Revision: D64915640 |
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
[ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64915640 |
SS-JIA
added a commit
that referenced
this pull request
Oct 25, 2024
Pull Request resolved: #6488 ## Changes Move the following files to the root directory of Vulkan backend: * `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py` * `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py` In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is ``` @update_features( [ exir_ops.edge.aten._log_softmax.default, exir_ops.edge.aten._softmax.default, exir_ops.edge.aten.mean.dim, exir_ops.edge.aten.sum.dim_IntList, exir_ops.edge.aten.amax.default, exir_ops.edge.aten.amin.default, ] ) def register_reduce_op(features: OpFeatures): features.texture_impl = TextureImplFeatures( uses_packed_dim=True, ) features.resize_fn = True def check_reduce_node(node: torch.fx.Node) -> bool: dim_list = node.args[1] assert isinstance(dim_list, list) if len(dim_list) != 1: return False keepdim = node.args[2] assert isinstance(keepdim, bool) if not keepdim: return False return True features.check_node_fn = check_reduce_node return features ``` ## Rationale The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`). ghstack-source-id: 250098667 @exported-using-ghexport Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
junpi3
approved these changes
Oct 25, 2024
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
[ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Oct 25, 2024
Pull Request resolved: #6488 ## Changes Move the following files to the root directory of Vulkan backend: * `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py` * `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py` In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is ``` @update_features( [ exir_ops.edge.aten._log_softmax.default, exir_ops.edge.aten._softmax.default, exir_ops.edge.aten.mean.dim, exir_ops.edge.aten.sum.dim_IntList, exir_ops.edge.aten.amax.default, exir_ops.edge.aten.amin.default, ] ) def register_reduce_op(features: OpFeatures): features.texture_impl = TextureImplFeatures( uses_packed_dim=True, ) features.resize_fn = True def check_reduce_node(node: torch.fx.Node) -> bool: dim_list = node.args[1] assert isinstance(dim_list, list) if len(dim_list) != 1: return False keepdim = node.args[2] assert isinstance(keepdim, bool) if not keepdim: return False return True features.check_node_fn = check_reduce_node return features ``` ## Rationale The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`). ghstack-source-id: 250109412 @exported-using-ghexport Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
|
This pull request was exported from Phabricator. Differential Revision: D64915640 |
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
[ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Oct 25, 2024
Pull Request resolved: #6488 ## Changes Move the following files to the root directory of Vulkan backend: * `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py` * `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py` In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is ``` @update_features( [ exir_ops.edge.aten._log_softmax.default, exir_ops.edge.aten._softmax.default, exir_ops.edge.aten.mean.dim, exir_ops.edge.aten.sum.dim_IntList, exir_ops.edge.aten.amax.default, exir_ops.edge.aten.amin.default, ] ) def register_reduce_op(features: OpFeatures): features.texture_impl = TextureImplFeatures( uses_packed_dim=True, ) features.resize_fn = True def check_reduce_node(node: torch.fx.Node) -> bool: dim_list = node.args[1] assert isinstance(dim_list, list) if len(dim_list) != 1: return False keepdim = node.args[2] assert isinstance(keepdim, bool) if not keepdim: return False return True features.check_node_fn = check_reduce_node return features ``` ## Rationale The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`). ghstack-source-id: 250122044 @exported-using-ghexport Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
|
This pull request was exported from Phabricator. Differential Revision: D64915640 |
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
[ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Oct 25, 2024
Pull Request resolved: #6488 ## Changes Move the following files to the root directory of Vulkan backend: * `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py` * `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py` In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is ``` @update_features( [ exir_ops.edge.aten._log_softmax.default, exir_ops.edge.aten._softmax.default, exir_ops.edge.aten.mean.dim, exir_ops.edge.aten.sum.dim_IntList, exir_ops.edge.aten.amax.default, exir_ops.edge.aten.amin.default, ] ) def register_reduce_op(features: OpFeatures): features.texture_impl = TextureImplFeatures( uses_packed_dim=True, ) features.resize_fn = True def check_reduce_node(node: torch.fx.Node) -> bool: dim_list = node.args[1] assert isinstance(dim_list, list) if len(dim_list) != 1: return False keepdim = node.args[2] assert isinstance(keepdim, bool) if not keepdim: return False return True features.check_node_fn = check_reduce_node return features ``` ## Rationale The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`). ghstack-source-id: 250128087 @exported-using-ghexport Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
|
This pull request was exported from Phabricator. Differential Revision: D64915640 |
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
[ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Oct 25, 2024
Pull Request resolved: #6488 ## Changes Move the following files to the root directory of Vulkan backend: * `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py` * `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py` In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is ``` @update_features( [ exir_ops.edge.aten._log_softmax.default, exir_ops.edge.aten._softmax.default, exir_ops.edge.aten.mean.dim, exir_ops.edge.aten.sum.dim_IntList, exir_ops.edge.aten.amax.default, exir_ops.edge.aten.amin.default, ] ) def register_reduce_op(features: OpFeatures): features.texture_impl = TextureImplFeatures( uses_packed_dim=True, ) features.resize_fn = True def check_reduce_node(node: torch.fx.Node) -> bool: dim_list = node.args[1] assert isinstance(dim_list, list) if len(dim_list) != 1: return False keepdim = node.args[2] assert isinstance(keepdim, bool) if not keepdim: return False return True features.check_node_fn = check_reduce_node return features ``` ## Rationale The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`). ghstack-source-id: 250136695 @exported-using-ghexport Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
|
This pull request was exported from Phabricator. Differential Revision: D64915640 |
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
[ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D64915640 |
SS-JIA
added a commit
that referenced
this pull request
Oct 25, 2024
Pull Request resolved: #6488 ## Changes Move the following files to the root directory of Vulkan backend: * `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py` * `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py` In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is ``` @update_features( [ exir_ops.edge.aten._log_softmax.default, exir_ops.edge.aten._softmax.default, exir_ops.edge.aten.mean.dim, exir_ops.edge.aten.sum.dim_IntList, exir_ops.edge.aten.amax.default, exir_ops.edge.aten.amin.default, ] ) def register_reduce_op(features: OpFeatures): features.texture_impl = TextureImplFeatures( uses_packed_dim=True, ) features.resize_fn = True def check_reduce_node(node: torch.fx.Node) -> bool: dim_list = node.args[1] assert isinstance(dim_list, list) if len(dim_list) != 1: return False keepdim = node.args[2] assert isinstance(keepdim, bool) if not keepdim: return False return True features.check_node_fn = check_reduce_node return features ``` ## Rationale The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`). ghstack-source-id: 250165806 @exported-using-ghexport Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
## Changes
Move the following files to the root directory of Vulkan backend:
* `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py`
* `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py`
In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is
```
update_features(
[
exir_ops.edge.aten._log_softmax.default,
exir_ops.edge.aten._softmax.default,
exir_ops.edge.aten.mean.dim,
exir_ops.edge.aten.sum.dim_IntList,
exir_ops.edge.aten.amax.default,
exir_ops.edge.aten.amin.default,
]
)
def register_reduce_op(features: OpFeatures):
features.texture_impl = TextureImplFeatures(
uses_packed_dim=True,
)
features.resize_fn = True
def check_reduce_node(node: torch.fx.Node) -> bool:
dim_list = node.args[1]
assert isinstance(dim_list, list)
if len(dim_list) != 1:
return False
keepdim = node.args[2]
assert isinstance(keepdim, bool)
if not keepdim:
return False
return True
features.check_node_fn = check_reduce_node
return features
```
## Rationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`).
Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
[ghstack-poisoned]
SS-JIA
added a commit
that referenced
this pull request
Oct 26, 2024
Pull Request resolved: #6488 ## Changes Move the following files to the root directory of Vulkan backend: * `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py` * `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py` In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is ``` @update_features( [ exir_ops.edge.aten._log_softmax.default, exir_ops.edge.aten._softmax.default, exir_ops.edge.aten.mean.dim, exir_ops.edge.aten.sum.dim_IntList, exir_ops.edge.aten.amax.default, exir_ops.edge.aten.amin.default, ] ) def register_reduce_op(features: OpFeatures): features.texture_impl = TextureImplFeatures( uses_packed_dim=True, ) features.resize_fn = True def check_reduce_node(node: torch.fx.Node) -> bool: dim_list = node.args[1] assert isinstance(dim_list, list) if len(dim_list) != 1: return False keepdim = node.args[2] assert isinstance(keepdim, bool) if not keepdim: return False return True features.check_node_fn = check_reduce_node return features ``` ## Rationale The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`). ghstack-source-id: 250279709 @exported-using-ghexport Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/)
|
This pull request was exported from Phabricator. Differential Revision: D64915640 |
facebook-github-bot
merged commit 610dac2
into
gh/SS-JIA/128/base
45 of 48 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
SS-JIA
added a commit
that referenced
this pull request
Oct 28, 2024
Pull Request resolved: #6488 ## Changes Move the following files to the root directory of Vulkan backend: * `backends/vulkan/partitioner/supported_ops.py` -> `backends/vulkan/op_registry.py` * `backends/vulkan/_passes/custom_ops_defs.py` -> `backends/vulkan/custom_ops_lib.py` In the new `op_registry.py` file, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the new `OpFeatures` class for more details. An example of registering a new operator to the export flow is ``` @update_features( [ exir_ops.edge.aten._log_softmax.default, exir_ops.edge.aten._softmax.default, exir_ops.edge.aten.mean.dim, exir_ops.edge.aten.sum.dim_IntList, exir_ops.edge.aten.amax.default, exir_ops.edge.aten.amin.default, ] ) def register_reduce_op(features: OpFeatures): features.texture_impl = TextureImplFeatures( uses_packed_dim=True, ) features.resize_fn = True def check_reduce_node(node: torch.fx.Node) -> bool: dim_list = node.args[1] assert isinstance(dim_list, list) if len(dim_list) != 1: return False keepdim = node.args[2] assert isinstance(keepdim, bool) if not keepdim: return False return True features.check_node_fn = check_reduce_node return features ``` ## Rationale The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e. `is_valid_to_copy`) and graph transforms do not have to maintain their own operator metadata (`USES_WEIGHTS` in `insert_prepack_nodes`). ghstack-source-id: 250279709 @exported-using-ghexport Differential Revision: [D64915640](https://our.internmc.facebook.com/intern/diff/D64915640/) Co-authored-by: Stephen Jia <ssjia@meta.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
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):
Changes
Move the following files to the root directory of Vulkan backend:
backends/vulkan/partitioner/supported_ops.py->backends/vulkan/op_registry.pybackends/vulkan/_passes/custom_ops_defs.py->backends/vulkan/custom_ops_lib.pyIn the new
op_registry.pyfile, the way operator features are specified is reworked to provide much more detail about the features of the operator implementation in Vulkan. See the newOpFeaturesclass for more details. An example of registering a new operator to the export flow isRationale
The purpose of these changes is to centralize operator definitions so that there is a common source of truth about the capabilities of operator implementation in Vulkan. This way, the partitioner does not have to implement ad-hoc functions for specific operators (i.e.
is_valid_to_copy) and graph transforms do not have to maintain their own operator metadata (USES_WEIGHTSininsert_prepack_nodes).Differential Revision: D64915640