[ET][Portable][Build Size] Move to elementwise_utils as apply_tritensor_elementwise_fn #6005
+271
−177
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
…or_elementwise_fn @swolchok's technique is superior to the TensorReader/TensorWriter approach I introduced in D63703174. So, I am rewriting my build size reduction stack on top of his approach. Superior how? - It should lead to smaller overall build size. Current measurements indicate this. Complete data will be published after the stack is complete. - It is better suited for dtype selective build, since it passes the op name to all of the ET_SWITCHes involved. - It is more performant. Current measurements to clamp.Tensor_out indicate this. Note that in the data below, my stack is marginally more performant for the vanilla case (no broadcast & all dtypes equal), but this is only because I added a "fast path" in my code for such vanilla case, which can be trivially added to Scott's approach as well. It is more relevant to compare numbers for mixed dtype or broadcasting. ``` Baseline clamp.Tensor_out no broadcast float: 25451 [23423 - 28839] microseconds clamp.Tensor_out no broadcast double: 25461 [23377 - 50940] microseconds clamp.Tensor_out no broadcast mixed dtype: 23367 [21353 - 27022] microseconds clamp.Tensor_out broadcast: 702529 [679667 - 742005] microseconds Manuel C clamp.Tensor_out no broadcast float: 22919 [21333 - 27140] microseconds clamp.Tensor_out no broadcast double: 23095 [21472 - 27462] microseconds clamp.Tensor_out no broadcast mixed dtype: 35042 [32875 - 42491] microseconds clamp.Tensor_out broadcast: 936541 [916437 - 971499] microseconds Scott W clamp.Tensor_out no broadcast float: 28263 [26458 - 32832] microseconds clamp.Tensor_out no broadcast double: 27442 [25548 - 39417] microseconds clamp.Tensor_out no broadcast mixed dtype: 25592 [23620 - 30148] microseconds clamp.Tensor_out broadcast: 695399 [674244 - 738919] microseconds ``` Build size reduction after Scott's diffs touching clamp.Tensor_out and where.self_out: - clamp: 7.42 MB -> 119 KB - where: 106 KB -> 16 KB Differential Revision: [D63838072](https://our.internmc.facebook.com/intern/diff/D63838072/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/6005
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 2b59ac2 with merge base cb12061 ( 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: D63838072 |
This was referenced Oct 8, 2024
[ET][Portable][Build Size] Introduce comparison op pattern. Binary ops: eq, ge, gt, le, lt, ne
#6013
Closed
swolchok
approved these changes
Oct 8, 2024
…ply_tritensor_elementwise_fn" swolchok's technique is superior to the TensorReader/TensorWriter approach I introduced in D63703174. So, I am rewriting my build size reduction stack on top of his approach. Superior how? - It should lead to smaller overall build size. Current measurements indicate this. Complete data will be published after the stack is complete. - It is better suited for dtype selective build, since it passes the op name to all of the ET_SWITCHes involved. - It is more performant. Current measurements to clamp.Tensor_out indicate this. Note that in the data below, my stack is marginally more performant for the vanilla case (no broadcast & all dtypes equal), but this is only because I added a "fast path" in my code for such vanilla case, which can be trivially added to Scott's approach as well. It is more relevant to compare numbers for mixed dtype or broadcasting. ``` Baseline clamp.Tensor_out no broadcast float: 25451 [23423 - 28839] microseconds clamp.Tensor_out no broadcast double: 25461 [23377 - 50940] microseconds clamp.Tensor_out no broadcast mixed dtype: 23367 [21353 - 27022] microseconds clamp.Tensor_out broadcast: 702529 [679667 - 742005] microseconds Manuel C clamp.Tensor_out no broadcast float: 22919 [21333 - 27140] microseconds clamp.Tensor_out no broadcast double: 23095 [21472 - 27462] microseconds clamp.Tensor_out no broadcast mixed dtype: 35042 [32875 - 42491] microseconds clamp.Tensor_out broadcast: 936541 [916437 - 971499] microseconds Scott W clamp.Tensor_out no broadcast float: 28263 [26458 - 32832] microseconds clamp.Tensor_out no broadcast double: 27442 [25548 - 39417] microseconds clamp.Tensor_out no broadcast mixed dtype: 25592 [23620 - 30148] microseconds clamp.Tensor_out broadcast: 695399 [674244 - 738919] microseconds ``` Build size reduction after Scott's diffs touching clamp.Tensor_out and where.self_out: - clamp: 7.42 MB -> 119 KB - where: 106 KB -> 16 KB Differential Revision: [D63838072](https://our.internmc.facebook.com/intern/diff/D63838072/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D63838072 |
malfet
approved these changes
Oct 8, 2024
…ply_tritensor_elementwise_fn" swolchok's technique is superior to the TensorReader/TensorWriter approach I introduced in D63703174. So, I am rewriting my build size reduction stack on top of his approach. Superior how? - It should lead to smaller overall build size. Current measurements indicate this. Complete data will be published after the stack is complete. - It is better suited for dtype selective build, since it passes the op name to all of the ET_SWITCHes involved. - It is more performant. Current measurements to clamp.Tensor_out indicate this. Note that in the data below, my stack is marginally more performant for the vanilla case (no broadcast & all dtypes equal), but this is only because I added a "fast path" in my code for such vanilla case, which can be trivially added to Scott's approach as well. It is more relevant to compare numbers for mixed dtype or broadcasting. ``` Baseline clamp.Tensor_out no broadcast float: 25451 [23423 - 28839] microseconds clamp.Tensor_out no broadcast double: 25461 [23377 - 50940] microseconds clamp.Tensor_out no broadcast mixed dtype: 23367 [21353 - 27022] microseconds clamp.Tensor_out broadcast: 702529 [679667 - 742005] microseconds Manuel C clamp.Tensor_out no broadcast float: 22919 [21333 - 27140] microseconds clamp.Tensor_out no broadcast double: 23095 [21472 - 27462] microseconds clamp.Tensor_out no broadcast mixed dtype: 35042 [32875 - 42491] microseconds clamp.Tensor_out broadcast: 936541 [916437 - 971499] microseconds Scott W clamp.Tensor_out no broadcast float: 28263 [26458 - 32832] microseconds clamp.Tensor_out no broadcast double: 27442 [25548 - 39417] microseconds clamp.Tensor_out no broadcast mixed dtype: 25592 [23620 - 30148] microseconds clamp.Tensor_out broadcast: 695399 [674244 - 738919] microseconds ``` Build size reduction after Scott's diffs touching clamp.Tensor_out and where.self_out: - clamp: 7.42 MB -> 119 KB - where: 106 KB -> 16 KB Differential Revision: [D63838072](https://our.internmc.facebook.com/intern/diff/D63838072/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D63838072 |
…ply_tritensor_elementwise_fn" swolchok's technique is superior to the TensorReader/TensorWriter approach I introduced in D63703174. So, I am rewriting my build size reduction stack on top of his approach. Superior how? - It should lead to smaller overall build size. Current measurements indicate this. Complete data will be published after the stack is complete. - It is better suited for dtype selective build, since it passes the op name to all of the ET_SWITCHes involved. - It is more performant. Current measurements to clamp.Tensor_out indicate this. Note that in the data below, my stack is marginally more performant for the vanilla case (no broadcast & all dtypes equal), but this is only because I added a "fast path" in my code for such vanilla case, which can be trivially added to Scott's approach as well. It is more relevant to compare numbers for mixed dtype or broadcasting. ``` Baseline clamp.Tensor_out no broadcast float: 25451 [23423 - 28839] microseconds clamp.Tensor_out no broadcast double: 25461 [23377 - 50940] microseconds clamp.Tensor_out no broadcast mixed dtype: 23367 [21353 - 27022] microseconds clamp.Tensor_out broadcast: 702529 [679667 - 742005] microseconds Manuel C clamp.Tensor_out no broadcast float: 22919 [21333 - 27140] microseconds clamp.Tensor_out no broadcast double: 23095 [21472 - 27462] microseconds clamp.Tensor_out no broadcast mixed dtype: 35042 [32875 - 42491] microseconds clamp.Tensor_out broadcast: 936541 [916437 - 971499] microseconds Scott W clamp.Tensor_out no broadcast float: 28263 [26458 - 32832] microseconds clamp.Tensor_out no broadcast double: 27442 [25548 - 39417] microseconds clamp.Tensor_out no broadcast mixed dtype: 25592 [23620 - 30148] microseconds clamp.Tensor_out broadcast: 695399 [674244 - 738919] microseconds ``` Build size reduction after Scott's diffs touching clamp.Tensor_out and where.self_out: - clamp: 7.42 MB -> 119 KB - where: 106 KB -> 16 KB Differential Revision: [D63838072](https://our.internmc.facebook.com/intern/diff/D63838072/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D63838072 |
…ply_tritensor_elementwise_fn" swolchok's technique is superior to the TensorReader/TensorWriter approach I introduced in D63703174. So, I am rewriting my build size reduction stack on top of his approach. Superior how? - It should lead to smaller overall build size. Current measurements indicate this. Complete data will be published after the stack is complete. - It is better suited for dtype selective build, since it passes the op name to all of the ET_SWITCHes involved. - It is more performant. Current measurements to clamp.Tensor_out indicate this. Note that in the data below, my stack is marginally more performant for the vanilla case (no broadcast & all dtypes equal), but this is only because I added a "fast path" in my code for such vanilla case, which can be trivially added to Scott's approach as well. It is more relevant to compare numbers for mixed dtype or broadcasting. ``` Baseline clamp.Tensor_out no broadcast float: 25451 [23423 - 28839] microseconds clamp.Tensor_out no broadcast double: 25461 [23377 - 50940] microseconds clamp.Tensor_out no broadcast mixed dtype: 23367 [21353 - 27022] microseconds clamp.Tensor_out broadcast: 702529 [679667 - 742005] microseconds Manuel C clamp.Tensor_out no broadcast float: 22919 [21333 - 27140] microseconds clamp.Tensor_out no broadcast double: 23095 [21472 - 27462] microseconds clamp.Tensor_out no broadcast mixed dtype: 35042 [32875 - 42491] microseconds clamp.Tensor_out broadcast: 936541 [916437 - 971499] microseconds Scott W clamp.Tensor_out no broadcast float: 28263 [26458 - 32832] microseconds clamp.Tensor_out no broadcast double: 27442 [25548 - 39417] microseconds clamp.Tensor_out no broadcast mixed dtype: 25592 [23620 - 30148] microseconds clamp.Tensor_out broadcast: 695399 [674244 - 738919] microseconds ``` Build size reduction after Scott's diffs touching clamp.Tensor_out and where.self_out: - clamp: 7.42 MB -> 119 KB - where: 106 KB -> 16 KB Differential Revision: [D63838072](https://our.internmc.facebook.com/intern/diff/D63838072/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D63838072 |
|
This pull request has been merged in a79caab. |
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):
@swolchok's technique is superior to the TensorReader/TensorWriter approach I introduced in D63703174. So, I am rewriting my build size reduction stack on top of his approach.
Superior how?
Build size reduction after Scott's diffs touching clamp.Tensor_out and where.self_out:
Differential Revision: D63838072