Skip to content

Enable -fno-reciprocal-math to fix div accuracy. #2413

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kdrozd-dev wants to merge 5 commits into
base: main
Choose a base branch
from
Open

Enable -fno-reciprocal-math to fix div accuracy. #2413

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
0101740
add -fno-reciprocal-math to SYCL_KERNEL_OPTIONS
kdrozd-dev Nov 26, 2025
4445f8e
Merge branch 'main' into kdrozd/div_accuracy
kdrozd-dev Dec 10, 2025
78e1175
Merge branch 'main' into kdrozd/div_accuracy
kdrozd-dev Dec 11, 2025
5992505
Merge branch 'main' into kdrozd/div_accuracy
kdrozd-dev Dec 12, 2025
ed404b1
Merge branch 'main' into kdrozd/div_accuracy
kdrozd-dev Dec 16, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions cmake/BuildFlags.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -104,6 +104,7 @@ macro(set_build_flags)
# gcc -shared host.o kernel.o device-code.o -o libxxx.so
set(SYCL_KERNEL_OPTIONS ${SYCL_KERNEL_OPTIONS} -fno-sycl-unnamed-lambda)
set(SYCL_KERNEL_OPTIONS ${SYCL_KERNEL_OPTIONS} -sycl-std=2020)
set(SYCL_KERNEL_OPTIONS ${SYCL_KERNEL_OPTIONS} -fno-reciprocal-math)
Copy link
Contributor

@EikanWang EikanWang Nov 30, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Let's study the NVCC behaviors first. I think reciprocal-math is a general optimization, and most the compilers enable it by default.

Copy link
Contributor Author

@kdrozd-dev kdrozd-dev Dec 1, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@EikanWang I believe --prec-div https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html?highlight=reciprocal#prec-div-true-false-prec-div is the corresponding NVCC flag by default set to True unless --use_fast_math (https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html?highlight=reciprocal#use-fast-math-use-fast-math) is set to True. I do not believe pytorch is built with --use_fast_math correct me if I'm wrong.

Copy link
Contributor

@EikanWang EikanWang Dec 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the information. Some HW platforms like CPU turn on this option by default. For CUDA, it keeps its default, saying use_fast_math is False. So, it makes sense to add no-reciprocal-math. I'm curious why SYCL turns on this option by default.

@fengyuan14 do you know the background?

torch-xpu-ops/cmake/BuildFlags.cmake

Line 77 in 20baf08

# The fast-math will be enabled by default in SYCL compiler.

Copy link
Contributor

@EikanWang EikanWang Dec 3, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Per my understanding, we should disable fast-math to align with CUDA

if(CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
set(SYCL_KERNEL_OPTIONS ${SYCL_KERNEL_OPTIONS} /fp:strict)
# Suppress warnings about dllexport.
Expand Down
Loading