Add vectorization in elementwise_util #9432
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
swolchok
requested review from
JacobSzwejbka,
lucylq and
manuelcandales
as code owners
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/9432
Note: Links to docs will display an error until the docs builds have been completed. ⏳ No Failures, 27 PendingAs of commit 4a0d1db with merge base bc42d8d ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Mar 20, 2025
Merged
[ghstack-poisoned]
[ghstack-poisoned]
manuelcandales
approved these changes
Jun 4, 2025
|
starting stacked land, noting that CI is green |
swolchok
added a commit
that referenced
this pull request
Jun 9, 2025
Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432.
swolchok
added a commit
that referenced
this pull request
Jun 9, 2025
Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432.
JacobSzwejbka
added a commit
that referenced
this pull request
Jun 10, 2025
This reverts commit 4c35fe0.
swolchok
added a commit
that referenced
this pull request
Jun 12, 2025
…nels test (#11205)", and "Add vectorization in elementwise_util (#9432)" Stack was reverted due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: [D76467389](https://our.internmc.facebook.com/intern/diff/D76467389/) [ghstack-poisoned]
swolchok
added a commit
that referenced
this pull request
Jun 12, 2025
…nels test (#11205)", and "Add vectorization in elementwise_util (#9432)" Stack was reverted due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: [D76467389](https://our.internmc.facebook.com/intern/diff/D76467389/) ghstack-source-id: 289985405 Pull Request resolved: #11604
swolchok
added a commit
that referenced
this pull request
Jun 12, 2025
…table_kernels test (#11205)", and "Add vectorization in elementwise_util (#9432)"" Stack was reverted due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: [D76467389](https://our.internmc.facebook.com/intern/diff/D76467389/) [ghstack-poisoned]
swolchok
added a commit
that referenced
this pull request
Jun 12, 2025
…nels test (#11205)", and "Add vectorization in elementwise_util (#9432)" Pull Request resolved: #11604 Stack was reverted due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: [D76467389](https://our.internmc.facebook.com/intern/diff/D76467389/) ghstack-source-id: 289996914
swolchok
added a commit
that referenced
this pull request
Jun 13, 2025
…nels test (#11205)", and "Add vectorization in elementwise_util (#9432)" Pull Request resolved: #11604 Stack was reverted due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. ghstack-source-id: 290334876 Differential Revision: [D76467389](https://our.internmc.facebook.com/intern/diff/D76467389/)
swolchok
added a commit
that referenced
this pull request
Jun 13, 2025
…"Add optimized_portable_kernels test (#11205)", and "Add vectorization in elementwise_util (#9432)"" Stack was reverted due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: [D76467389](https://our.internmc.facebook.com/intern/diff/D76467389/) [ghstack-poisoned]
swolchok
added a commit
that referenced
this pull request
Jun 13, 2025
…ortable_kernels test (#11205)", and "Add vectorization in elementwise_util (#9432)"" Stack was reverted due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: [D76467389](https://our.internmc.facebook.com/intern/diff/D76467389/) [ghstack-poisoned]
DariusHolmgren
pushed a commit
to DariusHolmgren/executorch
that referenced
this pull request
Jun 14, 2025
…ble_kernels test (pytorch#11205)", and "Add vectorization in elementwise_util (pytorch#9432)" Summary: Stack was reverted due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if defined(...) && ...` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) Original summary for pytorch#11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in pytorch#9432. Original summary for pytorch#11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support pytorch#9432. Original summary for pytorch#9432: This is a first cut at pytorch#9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: D76467389 *** fix ET_USE_PYTORCH_HEADERS
swolchok
added a commit
that referenced
this pull request
Jun 18, 2025
…kernels test (#11205)", and "Add vectorization in elementwise_util (#9432)" Stack was reverted (again! I bypassed some broken jobs and it turns out this re-broke them) due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes in first reapply: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) New fixes in second reapply: - So far, none; D76843086 and D76857541 fix things up in preparation for this diff. (some rebase conflict fixes though) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: [D76754826](https://our.internmc.facebook.com/intern/diff/D76754826/) [ghstack-poisoned]
swolchok
added a commit
that referenced
this pull request
Jun 18, 2025
…kernels test (#11205)", and "Add vectorization in elementwise_util (#9432)" Stack was reverted (again! I bypassed some broken jobs and it turns out this re-broke them) due to internal CI failures. Reapplying as an exported internal diff so that we make sure to catch any more of those. New fixes in first reapply: - straightforward op_sub build fixes - s/EXPECT_EQ/EXPECT_FLOAT_EQ/ in vectorized_math_test - define ET_USE_PYTORCH_HEADERS to detect whether exceptions are enabled, and use `#if` instead of `#ifdef` to check the macro so that we don't use PyTorch headers if exceptions are disabled. (otherwise, we might have problems with e.g. TORCH_CHECK) New fixes in second reapply: - So far, none; D76843086 and D76857541 fix things up in preparation for this diff. (some rebase conflict fixes though) Original summary for #11204: Set of math functions that work on both scalars and at::vec::Vectorized, to be used in #9432. Original summary for #11205: Make sure we test the optimized versions of portable kernels even if they are shadowed by optimized implementations. Intended to support #9432. Original summary for #9432: This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the `unary_ufunc_*` utilities in [pattern.h](https://github.com/pytorch/executorch/blob/main/kernels/portable/cpu/pattern/pattern.h) for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util. This PR adds an interesting testing problem: in theory, *all* operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above. Differential Revision: [D76754826](https://our.internmc.facebook.com/intern/diff/D76754826/) ghstack-source-id: 291370586 Pull Request resolved: #11802
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.
This is a first cut at #9241 . In this PR I've vectorized a small initial set of ops: atan2, clamp, fmod_Scalar, maximum, minimum, mul, pow, and sigmoid. In addition, the following ops should have gotten vectorized automatically because they already used generic lambdas: add, div, rsub, sub. I've left covering ops that use the
unary_ufunc_*utilities in pattern.h for a follow-up push, because pattern.h and elementwise_util need some work before we can migrate pattern.h's utilities to be backed by elementwise_util.This PR adds an interesting testing problem: in theory, all operators might need test cases long enough to tickle vectorization, because we might accidentally vectorize ops unexpectedly and break their lambdas due to anticipated differences in semantics. I address this issue by using Vectorized for the scalar prologue/epilogue in debug mode (we run tests in both debug and release) so that we can detect broken lambdas. I additionally intentionally introduced a bug in the vectorized path in elementwise_util and manually verified that we saw test failures for each vectorized op called out above.