Use std::fma in addcmul and foreach pointwise ops for FMA parity with CUDA#3275
Open
AKloniecki wants to merge 10 commits intomainfrom
Open
Use std::fma in addcmul and foreach pointwise ops for FMA parity with CUDA#3275AKloniecki wants to merge 10 commits intomainfrom
AKloniecki wants to merge 10 commits intomainfrom
Conversation
… CUDA The CUDA pointwise_op_impl (DeviceAddCmulCdiv.cuh) uses std::fma to guarantee consistent fused multiply-add behavior, particularly when alpha=1 where it emits std::fma(tensor1, tensor2, input). The XPU kernels were instead computing input + alpha * op(tensor1, tensor2) without any FMA guarantee, causing bitwise differences between addcmul and add-with-alpha operations. Port the same FMA logic to XPU: add a pointwise_op_impl helper in ForeachFunctors.h that mirrors CUDA's implementation, and update AddcmulFunctor, PointwiseOpScalarFunctor, and PointwiseOpScalarListFunctor to use it. This fixes test_addcmul_alpha_one_fma_parity on XPU. Signed-off-by: Artur Kłoniecki <arturx.kloniecki@intel.com>
AKloniecki
requested review from
EikanWang,
astachowiczhabana,
Copilot and
guangyey
Contributor
There was a problem hiding this comment.
Pull request overview
This PR aligns XPU addcmul and foreach pointwise operations with CUDA’s fused multiply-add (FMA) behavior by using std::fma for real floating-point math, addressing bitwise parity failures when alpha == 1 (issue #2759).
Changes:
- Updated XPU
addcmulkernel to usestd::fma(with analpha == 1fast-path) for floating-point accumulator types. - Introduced a
pointwise_op_implhelper inForeachFunctors.hto centralize FMA behavior forinput + alpha * op(tensor1, tensor2). - Switched foreach pointwise scalar and scalar-list functors to call
pointwise_op_impl.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.
| File | Description |
|---|---|
| src/ATen/native/xpu/sycl/PointwiseOpsKernels.cpp | Adds std::fma usage in AddcmulFunctor to match CUDA fused behavior. |
| src/ATen/native/xpu/sycl/ForeachFunctors.h | Adds pointwise_op_impl and routes foreach pointwise ops through it for FMA parity. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
…ication in AddcmulFunctor Agent-Logs-Url: https://github.com/intel/torch-xpu-ops/sessions/93af4d41-5f18-47eb-9154-be673b0da4a7 Co-authored-by: AKloniecki <188310598+AKloniecki@users.noreply.github.com>
Contributor
|
What's the issue this PR is fixing? |
Contributor
|
+please fix the linter |
guangyey
reviewed
Apr 7, 2026
Contributor
|
@AKloniecki could you please fix the lint issue. |
Signed-off-by: Artur Kłoniecki <arturx.kloniecki@intel.com>
Contributor
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Contributor
|
@AKloniecki please fix the linter issues + address all comments. Then we can auto-merge this PR |
Signed-off-by: Artur Kłoniecki <arturx.kloniecki@intel.com>
Contributor
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 3 out of 3 changed files in this pull request and generated no new comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Performance outliers, please check!
|
Performance outliers, please check!
|
Contributor
There was a problem hiding this comment.
Pull request overview
Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
+29
to
+36
| if (alpha == opmath_t(1)) { | ||
| if constexpr ( | ||
| std::is_same_v<Op, std::multiplies<opmath_t>> && | ||
| std::is_floating_point_v<opmath_t>) { | ||
| return std::fma(tensor1, tensor2, input); | ||
| } else { | ||
| return input + op(tensor1, tensor2); | ||
| } |
There was a problem hiding this comment.
The FMA fast-path is guarded by an exact type match (std::is_same_v<Op, std::multiplies<opmath_t>>). This is fragile: if callers pass an equivalent multiply op with a different type (e.g., std::multiplies<void>, custom functor/wrapper, or a different instantiation), the FMA path won’t trigger and you’ll lose the intended CUDA parity. Consider making the multiply/FMA intent explicit (e.g., separate overload/tag for multiply, or a small trait like is_multiply_op<Op, opmath_t> that can recognize expected multiply functors).
tszulist-hbn
approved these changes
Apr 16, 2026
astachowiczhabana
approved these changes
Apr 16, 2026
Performance outliers, please check!
|
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Comment on lines
+16
to
+17
| #include <functional> | ||
| #include <ATen/native/xpu/sycl/DeviceAddCmulCdiv.h> |
There was a problem hiding this comment.
The PR description says PointwiseOpsKernels.cpp removes unused #include <functional>, but the diff adds it back. Since DeviceAddCmulCdiv.h already includes <functional>, this include is likely redundant here—either drop #include <functional> from this file or update the PR description to reflect the new requirement.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
pointwise_op_implhelper into a sharedDeviceAddCmulCdiv.hheaderForeachFunctors.hto include the shared header instead of definingpointwise_op_impldirectlyPointwiseOpsKernels.cppto use the shared helper inAddcmulFunctor, removing duplicated FMA logic and unused#include <functional>/#include <type_traits>Fixes: #2759