Optimize math related build option #2462
Conversation
There was a problem hiding this comment.
Pull request overview
This PR optimizes math-related compiler flags for GNU compiler builds by simplifying floating-point behavior control. The change replaces four specific floating-point flags with two more comprehensive options.
Key Changes:
- Consolidated multiple floating-point flags into
-fno-fast-mathand-fmafor simpler and more predictable floating-point behavior - Maintained strict floating-point semantics while enabling fused multiply-add optimizations
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
chunhuanMeng
changed the title
|
@chunhuanMeng , I suppose CUDA also enables |
|
|
Please help collect the performance data. |
| set(SYCL_KERNEL_OPTIONS ${SYCL_KERNEL_OPTIONS} -fno-associative-math) | ||
| set(SYCL_KERNEL_OPTIONS ${SYCL_KERNEL_OPTIONS} -fno-approx-func) | ||
| set(SYCL_KERNEL_OPTIONS ${SYCL_KERNEL_OPTIONS} -Wno-absolute-value) | ||
| set(SYCL_KERNEL_OPTIONS ${SYCL_KERNEL_OPTIONS} -fno-fast-math) |
There was a problem hiding this comment.
@chunhuanMeng , should we better enable -ffp-contract=fast?
There was a problem hiding this comment.
-ffp-contract=fast enables floating-point contraction, allowing FMA formation, and -fma does the same.
|
Before landing the PR, let's collect performance data to obtain performance insights. |
|
No regression for performance PR run: https://github.com/intel/torch-xpu-ops/actions/runs/20703251204 |
Performance outliers, please check!
|
Updates the SYCL kernel compilation flags in the
set_build_flagsmacro to better control floating-point behavior and enable fused multiply-add (FMA) optimizations for both MSVC and GNU compilers.Compiler flag changes for floating-point behavior and FMA:
/Qfmato enable FMA instructions, and/Qftz-to disable flush-to-zero mode.-fno-fast-mathfor strict floating-point compliance and-fmato enable FMA instructions.