[tensorexp] ExternalCallWithAlloc (take ownership of aten Tensor, no memcpy ) #72225
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
[ghstack-poisoned]
CI Flow Status⚛️ CI FlowRuleset - Version:
|
Contributor
🔗 Helpful links
💊 CI failures summary and remediationsAs of commit 854ac78 (more details on the Dr. CI page): 💚 💚 Looks good so far! There are no failures yet. 💚 💚 This comment was automatically generated by Dr. CI (expand for details).Please report bugs/suggestions to the (internal) Dr. CI Users group. |
facebook-github-bot
added
the
oncall: jit
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) [ghstack-poisoned]
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) Result: quantized segmentation model mean inference goes down from 49ms to 22ms, the same as original model :) The ultimate goal is to reduce memcpy in external_functions for the at operators without out variants (no control for caller on where output tensor will be placed). 1. Introducing ExternalCall2 which has bufs_out and bufs_in. The size of buf_ptrs is bufs_out + bufs_in + bufs_out, first bufs_out slots are for output values of the result tensor buffers ptrs, bufs_in - tensor buffer ptrs of input arguments last bufs_out slots to store TensorImpl* as void* to release buffer ptrs after usage of buf_out ptrs. In external_functions implementation we are doing c10::intrusive_ptr::inc_ref to keep result tensor alive, it will be freed with FreeExt IR statement calling external function nnc_aten_free 2. Changing logic of memReuse: - No allocate Bufs that will be allocated in external function calls - Adding FreeExt for Bufs allocated with external function calls, 3. As output buffer are preallocated and we can not return buffers that are external Allocated and we did incref - decref => Transformation ExternalCall -> ExternalCall2 happens in codegen; ExtCalls which result buffer is output buffer are not converted to ExternalCall2 Before:  After: (no memcpy) <img width="1791" alt="Screen Shot 2022-02-11 at 11 34 49 AM" src="https://user-images.githubusercontent.com/6638825/153664528-edaf1205-b98c-40e3-b4bb-c9815bff24fe.png"> [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
priyaramani
approved these changes
Mar 2, 2022
torch/csrc/jit/tensorexpr/stmt.h
Outdated
| std::vector<ExprPtr> args_; | ||
| }; | ||
|
|
||
| class TORCH_API ExternalCall2 : public StmtNode<ExternalCall2> { |
Contributor
There was a problem hiding this comment.
Can we call this ExternalCallMemResuse or something like that instead of ExternalCall2, and add a comment to say how its different from ExternalCall ?
Contributor
Author
There was a problem hiding this comment.
Thanks, let it be ExternalCallWithAlloc
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) Result: quantized segmentation model mean inference goes down from 49ms to 22ms, the same as original model :) The ultimate goal is to reduce memcpy in external_functions for the at operators without out variants (no control for caller on where output tensor will be placed). 1. Introducing ExternalCall2 which has bufs_out and bufs_in. The size of buf_ptrs is bufs_out + bufs_in + bufs_out, first bufs_out slots are for output values of the result tensor buffers ptrs, bufs_in - tensor buffer ptrs of input arguments last bufs_out slots to store TensorImpl* as void* to release buffer ptrs after usage of buf_out ptrs. In external_functions implementation we are doing c10::intrusive_ptr::inc_ref to keep result tensor alive, it will be freed with FreeExt IR statement calling external function nnc_aten_free 2. Changing logic of memReuse: - No allocate Bufs that will be allocated in external function calls - Adding FreeExt for Bufs allocated with external function calls, 3. As output buffer are preallocated and we can not return buffers that are external Allocated and we did incref - decref => Transformation ExternalCall -> ExternalCall2 happens in codegen; ExtCalls which result buffer is output buffer are not converted to ExternalCall2 Before:  After: (no memcpy) <img width="1791" alt="Screen Shot 2022-02-11 at 11 34 49 AM" src="https://user-images.githubusercontent.com/6638825/153664528-edaf1205-b98c-40e3-b4bb-c9815bff24fe.png"> [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) Result: quantized segmentation model mean inference goes down from 49ms to 22ms, the same as original model :) The ultimate goal is to reduce memcpy in external_functions for the at operators without out variants (no control for caller on where output tensor will be placed). 1. Introducing ExternalCall2 which has bufs_out and bufs_in. The size of buf_ptrs is bufs_out + bufs_in + bufs_out, first bufs_out slots are for output values of the result tensor buffers ptrs, bufs_in - tensor buffer ptrs of input arguments last bufs_out slots to store TensorImpl* as void* to release buffer ptrs after usage of buf_out ptrs. In external_functions implementation we are doing c10::intrusive_ptr::inc_ref to keep result tensor alive, it will be freed with FreeExt IR statement calling external function nnc_aten_free 2. Changing logic of memReuse: - No allocate Bufs that will be allocated in external function calls - Adding FreeExt for Bufs allocated with external function calls, 3. As output buffer are preallocated and we can not return buffers that are external Allocated and we did incref - decref => Transformation ExternalCall -> ExternalCall2 happens in codegen; ExtCalls which result buffer is output buffer are not converted to ExternalCall2 Before:  After: (no memcpy) <img width="1791" alt="Screen Shot 2022-02-11 at 11 34 49 AM" src="https://user-images.githubusercontent.com/6638825/153664528-edaf1205-b98c-40e3-b4bb-c9815bff24fe.png"> [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) Result: quantized segmentation model mean inference goes down from 49ms to 22ms, the same as original model :) The ultimate goal is to reduce memcpy in external_functions for the at operators without out variants (no control for caller on where output tensor will be placed). 1. Introducing ExternalCall2 which has bufs_out and bufs_in. The size of buf_ptrs is bufs_out + bufs_in + bufs_out, first bufs_out slots are for output values of the result tensor buffers ptrs, bufs_in - tensor buffer ptrs of input arguments last bufs_out slots to store TensorImpl* as void* to release buffer ptrs after usage of buf_out ptrs. In external_functions implementation we are doing c10::intrusive_ptr::inc_ref to keep result tensor alive, it will be freed with FreeExt IR statement calling external function nnc_aten_free 2. Changing logic of memReuse: - No allocate Bufs that will be allocated in external function calls - Adding FreeExt for Bufs allocated with external function calls, 3. As output buffer are preallocated and we can not return buffers that are external Allocated and we did incref - decref => Transformation ExternalCall -> ExternalCall2 happens in codegen; ExtCalls which result buffer is output buffer are not converted to ExternalCall2 Before:  After: (no memcpy) <img width="1791" alt="Screen Shot 2022-02-11 at 11 34 49 AM" src="https://user-images.githubusercontent.com/6638825/153664528-edaf1205-b98c-40e3-b4bb-c9815bff24fe.png"> [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) Result: quantized segmentation model mean inference goes down from 49ms to 22ms, the same as original model :) The ultimate goal is to reduce memcpy in external_functions for the at operators without out variants (no control for caller on where output tensor will be placed). 1. Introducing ExternalCall2 which has bufs_out and bufs_in. The size of buf_ptrs is bufs_out + bufs_in + bufs_out, first bufs_out slots are for output values of the result tensor buffers ptrs, bufs_in - tensor buffer ptrs of input arguments last bufs_out slots to store TensorImpl* as void* to release buffer ptrs after usage of buf_out ptrs. In external_functions implementation we are doing c10::intrusive_ptr::inc_ref to keep result tensor alive, it will be freed with FreeExt IR statement calling external function nnc_aten_free 2. Changing logic of memReuse: - No allocate Bufs that will be allocated in external function calls - Adding FreeExt for Bufs allocated with external function calls, 3. As output buffer are preallocated and we can not return buffers that are external Allocated and we did incref - decref => Transformation ExternalCall -> ExternalCall2 happens in codegen; ExtCalls which result buffer is output buffer are not converted to ExternalCall2 Before:  After: (no memcpy) <img width="1791" alt="Screen Shot 2022-02-11 at 11 34 49 AM" src="https://user-images.githubusercontent.com/6638825/153664528-edaf1205-b98c-40e3-b4bb-c9815bff24fe.png"> [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
IvanKobzarev
changed the title
…Tensor, no memcpy )" Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) Result: quantized segmentation model mean inference goes down from 49ms to 22ms, the same as original model :) The ultimate goal is to reduce memcpy in external_functions for the at operators without out variants (no control for caller on where output tensor will be placed). 1. Introducing ExternalCall2 which has bufs_out and bufs_in. The size of buf_ptrs is bufs_out + bufs_in + bufs_out, first bufs_out slots are for output values of the result tensor buffers ptrs, bufs_in - tensor buffer ptrs of input arguments last bufs_out slots to store TensorImpl* as void* to release buffer ptrs after usage of buf_out ptrs. In external_functions implementation we are doing c10::intrusive_ptr::inc_ref to keep result tensor alive, it will be freed with FreeExt IR statement calling external function nnc_aten_free 2. Changing logic of memReuse: - No allocate Bufs that will be allocated in external function calls - Adding FreeExt for Bufs allocated with external function calls, 3. As output buffer are preallocated and we can not return buffers that are external Allocated and we did incref - decref => Transformation ExternalCall -> ExternalCall2 happens in codegen; ExtCalls which result buffer is output buffer are not converted to ExternalCall2 Before:  After: (no memcpy) <img width="1791" alt="Screen Shot 2022-02-11 at 11 34 49 AM" src="https://user-images.githubusercontent.com/6638825/153664528-edaf1205-b98c-40e3-b4bb-c9815bff24fe.png"> [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
…Tensor, no memcpy )" Differential Revision: [D33960933](https://our.internmc.facebook.com/intern/diff/D33960933) Result: quantized segmentation model mean inference goes down from 49ms to 22ms, the same as original model :) The ultimate goal is to reduce memcpy in external_functions for the at operators without out variants (no control for caller on where output tensor will be placed). 1. Introducing ExternalCall2 which has bufs_out and bufs_in. The size of buf_ptrs is bufs_out + bufs_in + bufs_out, first bufs_out slots are for output values of the result tensor buffers ptrs, bufs_in - tensor buffer ptrs of input arguments last bufs_out slots to store TensorImpl* as void* to release buffer ptrs after usage of buf_out ptrs. In external_functions implementation we are doing c10::intrusive_ptr::inc_ref to keep result tensor alive, it will be freed with FreeExt IR statement calling external function nnc_aten_free 2. Changing logic of memReuse: - No allocate Bufs that will be allocated in external function calls - Adding FreeExt for Bufs allocated with external function calls, 3. As output buffer are preallocated and we can not return buffers that are external Allocated and we did incref - decref => Transformation ExternalCall -> ExternalCall2 happens in codegen; ExtCalls which result buffer is output buffer are not converted to ExternalCall2 Before:  After: (no memcpy) <img width="1791" alt="Screen Shot 2022-02-11 at 11 34 49 AM" src="https://user-images.githubusercontent.com/6638825/153664528-edaf1205-b98c-40e3-b4bb-c9815bff24fe.png"> [ghstack-poisoned]
Contributor
Author
|
@IvanKobzarev has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
facebook-github-bot
pushed a commit
that referenced
this pull request
Mar 9, 2022
Summary: Pull Request resolved: #72225 Test Plan: Imported from OSS Reviewed By: dagitses Differential Revision: D33960933 Pulled By: IvanKobzarev fbshipit-source-id: fc73a3de9e5150919e3806516065b4a6c8316000
Contributor
|
Hey @IvanKobzarev. |
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):
Differential Revision: D33960933
Result:
quantized segmentation model mean inference goes down from 49ms to 22ms, the same as original model :)
The ultimate goal is to reduce memcpy in external_functions for the at operators without out variants (no control for caller on where output tensor will be placed).
The size of buf_ptrs is bufs_out + bufs_in + bufs_out,
first bufs_out slots are for output values of the result tensor buffers ptrs,
bufs_in - tensor buffer ptrs of input arguments
last bufs_out slots to store TensorImpl* as void* to release buffer ptrs after usage of buf_out ptrs.
In external_functions implementation we are doing c10::intrusive_ptr::inc_ref to keep result tensor alive, it will be freed with FreeExt IR statement calling external function nnc_aten_free
Before:
After: (no memcpy)
