Convert contiguous select_copy to zero-copy view in ReplaceViewCopyWithViewPass (#19198)#19198
Open
JacobSzwejbka wants to merge 1 commit intopytorch:mainfrom
Open
Convert contiguous select_copy to zero-copy view in ReplaceViewCopyWithViewPass (#19198)#19198JacobSzwejbka wants to merge 1 commit intopytorch:mainfrom
JacobSzwejbka wants to merge 1 commit intopytorch:mainfrom
Conversation
JacobSzwejbka
requested review from
larryliu0820 and
manuelcandales
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19198
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@JacobSzwejbka has exported this pull request. If you are a Meta employee, you can view the originating Diff in D102396195. |
This PR needs a
|
meta-codesync
Bot
changed the title
JacobSzwejbka
added a commit
to JacobSzwejbka/executorch-1
that referenced
this pull request
Apr 29, 2026
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. It uses the base spec dim_order to compute actual memory strides, so this works for any contiguous layout (C-contiguous, channels-last, etc.). For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Differential Revision: D102396195
JacobSzwejbka
force-pushed
the
export-D102396195
branch
from
6aeff0a to
340c71b
Compare
JacobSzwejbka
added a commit
to JacobSzwejbka/executorch-1
that referenced
this pull request
Apr 29, 2026
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. It uses the base spec dim_order to compute actual memory strides, so this works for any contiguous layout (C-contiguous, channels-last, etc.). For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Differential Revision: D102396195
JacobSzwejbka
force-pushed
the
export-D102396195
branch
from
340c71b to
2e6d5ea
Compare
JacobSzwejbka
added a commit
to JacobSzwejbka/executorch-1
that referenced
this pull request
Apr 29, 2026
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Differential Revision: D102396195
JacobSzwejbka
force-pushed
the
export-D102396195
branch
from
2e6d5ea to
ed21ca2
Compare
JacobSzwejbka
added a commit
to JacobSzwejbka/executorch-1
that referenced
this pull request
Apr 29, 2026
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Differential Revision: D102396195
JacobSzwejbka
force-pushed
the
export-D102396195
branch
from
ed21ca2 to
fa78ec0
Compare
JacobSzwejbka
added a commit
to JacobSzwejbka/executorch-1
that referenced
this pull request
Apr 29, 2026
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Differential Revision: D102396195
JacobSzwejbka
force-pushed
the
export-D102396195
branch
from
fa78ec0 to
199c9fe
Compare
JacobSzwejbka
added a commit
to JacobSzwejbka/executorch-1
that referenced
this pull request
Apr 29, 2026
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Differential Revision: D102396195
JacobSzwejbka
force-pushed
the
export-D102396195
branch
2 times, most recently
from
7177f19 to
d546d84
Compare
JacobSzwejbka
added a commit
to JacobSzwejbka/executorch-1
that referenced
this pull request
Apr 29, 2026
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Reviewed By: metascroy Differential Revision: D102396195
JacobSzwejbka
added a commit
to JacobSzwejbka/executorch-1
that referenced
this pull request
Apr 29, 2026
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Reviewed By: metascroy Differential Revision: D102396195
JacobSzwejbka
force-pushed
the
export-D102396195
branch
from
d546d84 to
2b185e0
Compare
…thViewPass (pytorch#19198) Summary: Extends the ReplaceViewCopyWithViewPass to convert `select_copy` ops to zero-copy `memory.select` views when the output is a contiguous sub-region of the base tensor. This is the same pattern used for `view_copy` -> `memory.view`, but for select operations. The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing. For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with `mem_offset = base_offset + byte_delta`. For dynamic shapes, a new `executorch_prim::et_select` runtime op sets the output data pointer to `self.data_ptr + offset`. Changes: - `exir/memory.py`: Added `memory.select` function - `exir/passes/replace_view_copy_with_view_pass.py`: Extended `_ViewSpec` with `byte_offset`, `stride`, `dim_order` params; added contiguity check using `stride_from_dim_order`; added select_copy handling in the pass - Pipeline integration: memory planner, to_out_var skiplist, emitter, serialization - `kernels/prim_ops/et_select.{h,cpp}`: C++ runtime op for dynamic select views - Tests: 5 new Python tests + 1 C++ test Authored with Claude. Reviewed By: metascroy Differential Revision: D102396195
JacobSzwejbka
force-pushed
the
export-D102396195
branch
from
2b185e0 to
6387dc4
Compare
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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.
Summary:
Extends the ReplaceViewCopyWithViewPass to convert
select_copyops to zero-copymemory.selectviews when the output is a contiguous sub-region of the base tensor. This is the same pattern used forview_copy->memory.view, but for select operations.The pass checks that the base is densely packed, non-constant, and static, and that the selected output forms a dense packing.
For static memory-planned subviews, the emitter elides the op entirely (no runtime instruction) by serializing tensor metadata with
mem_offset = base_offset + byte_delta. For dynamic shapes, a newexecutorch_prim::et_selectruntime op sets the output data pointer toself.data_ptr + offset.Changes:
exir/memory.py: Addedmemory.selectfunctionexir/passes/replace_view_copy_with_view_pass.py: Extended_ViewSpecwithbyte_offset,stride,dim_orderparams; added contiguity check usingstride_from_dim_order; added select_copy handling in the passkernels/prim_ops/et_select.{h,cpp}: C++ runtime op for dynamic select viewsAuthored with Claude.
Reviewed By: metascroy
Differential Revision: D102396195