Skip to content

Revert "[DEBUG] Revert "Enable SPV_INTEL_fp_fast_math_mode (#4058)"(#4473)" #4576

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

Merged
merged 9 commits into from
Jul 22, 2025
Merged

Revert "[DEBUG] Revert "Enable SPV_INTEL_fp_fast_math_mode (#4058)"(#4473)" #4576

Changes from 7 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
3ed2624
Revert "[DEBUG] Revert "Enable `SPV_INTEL_fp_fast_math_mode` (#4058)"…
anmyachev Jun 26, 2025
02d3986
Merge branch 'main' into amyachev/issue4514
anmyachev Jul 11, 2025
97550fc
Merge branch 'main' into amyachev/issue4514
anmyachev Jul 14, 2025
889b024
Merge branch 'main' into amyachev/issue4514
anmyachev Jul 17, 2025
5b7762a
Merge branch 'main' into amyachev/issue4514
anmyachev Jul 18, 2025
f3f090f
inst.hasNoSignedZeros && inst.hasNoNaNs && inst.hasNoInfs
anmyachev Jul 18, 2025
e294d73
leave only 'inst.hasNoNaNs'
anmyachev Jul 18, 2025
22fbc12
Update third_party/intel/triton_xpu.cc
anmyachev Jul 21, 2025
7c9cd30
Merge branch 'main' into amyachev/issue4514
anmyachev Jul 22, 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
6 changes: 3 additions & 3 deletions third_party/intel/triton_xpu.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -300,9 +300,9 @@ void init_triton_intel(py::module &&m) {
for (Instruction &inst : instructions(func)) {
if (auto *op = dyn_cast<FPMathOperator>(&inst)) {
FastMathFlags FMF;
// Allow contract when default fp fusion is enabled.
if ((enableFpFusion.has_value() && enableFpFusion.value()) &&
!fastMath.has_value()) {
// Default to allow contract when default fp fusion is not disabled.
if ((!enableFpFusion.has_value() || enableFpFusion.value()) &&
!fastMath.has_value() && inst.hasNoNaNs()) {
Copy link
Contributor Author

@anmyachev anmyachev Jul 18, 2025

Choose a reason for hiding this comment

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

@whitneywhtsang with hasNoNaNs change detectron2_fasterrcnn_r_50_fpn passes accuracy check locally. This change is inspired by

intel-xpu-backend-for-triton/python/src/llvm.cc

Line 58 in b361546

opt.NoNaNsFPMath = true;

Could you take a look?

Here is

Copy link
Contributor

@whitneywhtsang whitneywhtsang Jul 18, 2025

Choose a reason for hiding this comment

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

Is the performance regression recovered?

Copy link
Contributor

@whitneywhtsang whitneywhtsang Jul 18, 2025

Choose a reason for hiding this comment

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

Started a BMG performance run: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/16380688900

Copy link
Contributor

@whitneywhtsang whitneywhtsang Jul 19, 2025

Choose a reason for hiding this comment

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

From the above BMG run result, the geomean for ATTN D_HEAD=128 CAUSAL=1 is 60TFlops, which is the bad performance.

Copy link
Contributor Author

@anmyachev anmyachev Jul 19, 2025

Choose a reason for hiding this comment

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

From the above BMG run result, the geomean for ATTN D_HEAD=128 CAUSAL=1 is 60TFlops, which is the bad performance.

Hm, how many teraflops do you expect to see? I don't remember

Copy link
Contributor

@whitneywhtsang whitneywhtsang Jul 19, 2025

Choose a reason for hiding this comment

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

Expect to see above 80TFLops, #4514 (comment)

Copy link
Contributor Author

@anmyachev anmyachev Jul 19, 2025

Choose a reason for hiding this comment

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

I'll look into why that is. However, I have doubts that adding contract to all operations regardless of their safety in terms of accuracy is what is expected from the default behavior when the fast-math option is not enabled. I suspect that when AllowFPOpFusion option is enabled, some analysis is performed that adds it only where it is safe. Maybe you've already tested this and know exactly how it works?

if (op->getOpcode() == Instruction::FAdd ||
op->getOpcode() == Instruction::FMul)
FMF.setAllowContract(true);
Expand Down
Loading