Consistent compute numel/contiguous strategy with SymInts #85858

ezyang · 2022-09-28T22:00:18Z

Stack from ghstack (oldest at bottom):

Previously, our handling for contiguity was inconsistent in the following ways:

is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed
based on sizes_and_strides_, even if you had symbolic ints
Furthermore, even if you set custom policy for strides, these quantities were
not overridable by subclasses
Furthermore, we didn't even store these fields on ExtraMeta
We duplicate implementations of compute_contiguous (plain, channels last,
channels last 3d)
We inconsistently called refresh_numel()/refresh_contiguous(), versus
recomputing it ourselves

This factor makes a consistent strategy for all of the boolean fields, and
for numel computation. After this refactor:

All layout boolean fields are interposable via strides policy
and can be overridden from Python; you will never access a garbage field
All layout boolean fields are on ExtraMeta
You can always call refresh_numel/contiguous, no matter if your Tensor is
contiguous or not
The numel/layout boolean fields are always populated consistently with
the sizes strides fields (either on Tensor or ExtraMeta), even if you
have custom policy
There is only one implementation of the actual computation logic

Signed-off-by: Edward Z. Yang ezyang@fb.com

Differential Revision: D39907696

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyang@fb.com> [ghstack-poisoned]

pytorch-bot · 2022-09-28T22:00:20Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/85858

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 Failures, 1 Pending

As of commit 7d04800:

The following jobs have failed:

macos-12-py3-x86-64 / test (default, 2, 2, macos-12)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 00ffd47c46fc5f0909689cb190b61036328d0ee5 Pull Request resolved: #85858

ezyang · 2022-09-28T22:16:44Z

@ezyang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 1a51a32b8b1694d240ac0762e361d06e4a60eac8 Pull Request resolved: #85858

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 1eea645bc4a86916b59e61277da1d1dbc9ad4484 Pull Request resolved: #85858

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 48c083df3960094757bbab811c9b37b2deaffeb1 Pull Request resolved: #85858

ezyang · 2022-09-28T23:19:53Z

@ezyang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 48d338ff4d42acce606905b045ac61c1520f2874 Pull Request resolved: #85858

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 0280430740448d9dd8243360b4424a222af63e19 Pull Request resolved: #85858

ezyang · 2022-09-28T23:38:57Z

@ezyang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 5933aa01242c6fedb5047324c5dce3e191662232 Pull Request resolved: #85858

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: f87e66f70fc8bebd2507ff2d9ea5c1768e3bcaa7 Pull Request resolved: #85858

ezyang · 2022-09-30T02:38:46Z

@ezyang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

wconstab · 2022-09-30T03:29:47Z

c10/core/TensorImpl.cpp

+}
+
+template <typename T>
+bool_is_channels_last_contiguous _compute_channels_last_contiguous_2d(


just out of curiosity, what problem is the use of strong types solving here? accidentally putting the wrong bools in the wrong slots? or implicit conversions?

Putting bools in wrong slots.

wconstab · 2022-09-30T03:46:45Z

c10/core/TensorImpl.h

+        set_fields(
+            is_contiguous,
+            is_channels_last_contiguous,
+            bool_is_channels_last_3d_contiguous(false),


ok yea the strong bools are pretty nice here

wconstab · 2022-09-30T03:49:31Z

c10/core/TensorImpl.h

+            is_contiguous || is_channels_last_contiguous ||
+            is_channels_last_3d_contiguous ||
+            compute_non_overlapping_and_dense());
+        set_fields(


would these be more readable if you did

set_fields( bool_is_channels_last_contiguous_2d( compute_channels_last_contiguous_2d()), bool_is_channels_last_3d_contiguous( !is_channels_last_contiguous && compute_channels_last_contiguous_3d()), ...

No, because if I do it directly in set_fields I cannot reference the intermediate values

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 723fc848d9dc509ba99872c0b95b2a0563221365 Pull Request resolved: #85858

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: fba0ecf9947811a27f7aab81e480db76cce9d639 Pull Request resolved: #85858

ezyang · 2022-09-30T04:40:44Z

@ezyang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> ghstack-source-id: 2d575f9826676431fe3415920bda65ede82c0f00 Pull Request resolved: #85858

ezyang · 2022-09-30T04:41:49Z

@ezyang has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

albanD

SGTM

ezyang · 2022-09-30T16:58:42Z

@pytorchbot merge -f "spurious oom from infra flakiness"

pytorchmergebot · 2022-09-30T17:00:04Z

@pytorchbot successfully started a merge job. Check the current status here.
The merge job was triggered with the force (-f) flag. This means your change will be merged immediately, bypassing any CI checks (ETA: 1-5 minutes). If this is not the intended behavior, feel free to use some of the other merge options in the wiki.
Please reach out to the PyTorch DevX Team with feedback or questions!

pytorchmergebot · 2022-09-30T17:00:08Z

Merge failed

Reason: Command git -C /home/runner/work/pytorch/pytorch cherry-pick -x 6c11c6f28669db9869ac52c8d12dabb49b606bd4 returned non-zero exit code 1

Auto-merging c10/core/TensorImpl.cpp
CONFLICT (content): Merge conflict in c10/core/TensorImpl.cpp
error: could not apply 6c11c6f286... Consistent compute numel/contiguous strategy with SymInts
hint: After resolving the conflicts, mark them with
hint: "git add/rm <pathspec>", then run
hint: "git cherry-pick --continue".
hint: You can instead skip this commit with "git cherry-pick --skip".
hint: To abort and get back to the state before "git cherry-pick",
hint: run "git cherry-pick --abort".

Details for Dev Infra team

Raised by workflow job

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyangfb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) [ghstack-poisoned]

Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyang@fb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) Pull Request resolved: #85858 Approved by: https://github.com/albanD

…ytorch#85858)" Summary: Original commit changeset: 02df5806208b Original Phabricator Diff: D39907696 Differential Revision: D40105192 fbshipit-source-id: d5663085c7274623b12d3fe1574f0f3c36107342

) Previously, our handling for contiguity was inconsistent in the following ways: - is_strides_like 2d/3d and is_non_overlapping_and_dense always were computed based on sizes_and_strides_, even if you had symbolic ints - Furthermore, even if you set custom policy for strides, these quantities were not overridable by subclasses - Furthermore, we didn't even store these fields on ExtraMeta - We duplicate implementations of compute_contiguous (plain, channels last, channels last 3d) - We inconsistently called refresh_numel()/refresh_contiguous(), versus recomputing it ourselves This factor makes a consistent strategy for all of the boolean fields, and for numel computation. After this refactor: - All layout boolean fields are interposable via strides policy and can be overridden from Python; you will never access a garbage field - All layout boolean fields are on ExtraMeta - You can always call refresh_numel/contiguous, no matter if your Tensor is contiguous or not - The numel/layout boolean fields are always populated consistently with the sizes strides fields (either on Tensor or ExtraMeta), even if you have custom policy - There is only one implementation of the actual computation logic Signed-off-by: Edward Z. Yang <ezyang@fb.com> Differential Revision: [D39907696](https://our.internmc.facebook.com/intern/diff/D39907696) Pull Request resolved: pytorch#85858 Approved by: https://github.com/albanD

ezyang requested review from albanD and soulitzer as code owners September 28, 2022 22:00

pytorch-bot bot added the release notes: jit release notes category label Sep 28, 2022

facebook-github-bot added the cla signed label Sep 28, 2022

facebook-github-bot added the oncall: jit Add this issue/PR to JIT oncall triage queue label Sep 28, 2022

ezyang added ciflow/trunk Trigger trunk jobs on your pull request and removed oncall: jit Add this issue/PR to JIT oncall triage queue labels Sep 28, 2022

github-actions bot requested review from anjali411, antoniojkim, Chillee, Krovatkin, miladm and wconstab September 28, 2022 22:03

ezyang added the with-ssh label Sep 28, 2022

ezyang removed with-ssh ciflow/trunk Trigger trunk jobs on your pull request labels Sep 30, 2022

wconstab reviewed Sep 30, 2022

View reviewed changes

ezyang mentioned this pull request Sep 30, 2022

Make FunctionalTensorWrapper correctly handle symbolic shapes #85975

Closed

albanD approved these changes Sep 30, 2022

View reviewed changes

pytorch-bot bot added the ciflow/trunk Trigger trunk jobs on your pull request label Sep 30, 2022

pytorchmergebot closed this in 3b6588a Sep 30, 2022

facebook-github-bot deleted the gh/ezyang/1419/head branch October 4, 2022 14:19

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Consistent compute numel/contiguous strategy with SymInts #85858

Consistent compute numel/contiguous strategy with SymInts #85858

ezyang commented Sep 28, 2022 •

edited

pytorch-bot bot commented Sep 28, 2022 •

edited

ezyang commented Sep 28, 2022

ezyang commented Sep 28, 2022

ezyang commented Sep 28, 2022

ezyang commented Sep 30, 2022

wconstab Sep 30, 2022

ezyang Sep 30, 2022

wconstab Sep 30, 2022

wconstab Sep 30, 2022 •

edited

ezyang Sep 30, 2022

wconstab Sep 30, 2022

ezyang commented Sep 30, 2022

ezyang commented Sep 30, 2022

albanD left a comment

ezyang commented Sep 30, 2022

pytorchmergebot commented Sep 30, 2022

pytorchmergebot commented Sep 30, 2022

Consistent compute numel/contiguous strategy with SymInts #85858

Consistent compute numel/contiguous strategy with SymInts #85858

Conversation

ezyang commented Sep 28, 2022 • edited

pytorch-bot bot commented Sep 28, 2022 • edited

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/85858

❌ 1 Failures, 1 Pending

ezyang commented Sep 28, 2022

ezyang commented Sep 28, 2022

ezyang commented Sep 28, 2022

ezyang commented Sep 30, 2022

wconstab Sep 30, 2022

Choose a reason for hiding this comment

ezyang Sep 30, 2022

Choose a reason for hiding this comment

wconstab Sep 30, 2022

Choose a reason for hiding this comment

wconstab Sep 30, 2022 • edited

Choose a reason for hiding this comment

ezyang Sep 30, 2022

Choose a reason for hiding this comment

wconstab Sep 30, 2022

Choose a reason for hiding this comment

ezyang commented Sep 30, 2022

ezyang commented Sep 30, 2022

albanD left a comment

Choose a reason for hiding this comment

ezyang commented Sep 30, 2022

pytorchmergebot commented Sep 30, 2022

pytorchmergebot commented Sep 30, 2022

Merge failed

ezyang commented Sep 28, 2022 •

edited

pytorch-bot bot commented Sep 28, 2022 •

edited

wconstab Sep 30, 2022 •

edited