[sshlyapn]

This PR fixes an exponent underflow bug in the FP8 E8M0 conversion path for doubles

Motivation

Currently, __hip_cvt_double_to_e8m0() computes the float-domain exponent using unsigned short arithmetic. For small normal double exponents (non-zero but less than the bias delta), this subtraction underflowed, producing a large value and incorrectly taking NaN path

Technical Details

• Compute the exponent delta in signed arithmetic and clamp float_exp <= 0 to 0x0U (underflow to float subnormal/zero domain)
• Preserve overflow/NaN/Inf handling via float_exp > 0xFF -> e8m0_NaN, and leave the existing rounding/saturation post-processing intact
• Updated Catch2 matchers as Catch::Matchers::WithinAbs/WithinRel to match current Catch2 API

Test Plan

• Extended projects/hip-tests/catch/unit/deviceLib/fp8_e8m0.cc with underflow-focused inputs
• Note: this test remains disabled (make hip-tests compileable with TheRock #1624 ), so it cannot currently be exercised in CI

Test Result

• Local: verified by running the FP8 E8M0 unit test locally; all checks passed
• CI: not run (test is currently disabled)

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[sshlyapn]

@AidanBeltonS I noticed you've worked on the FP8_E8M0 data type. Could you please review these changes?

[Copilot]

Pull request overview

This PR fixes an exponent underflow bug in the double-to-E8M0 conversion function where unsigned arithmetic caused wraparound for small normal double exponents, incorrectly classifying them as NaN.

Key Changes:

• Replace unsigned exponent arithmetic with signed arithmetic to prevent underflow wraparound
• Add explicit clamping logic for underflow (float_exp <= 0), valid range, and overflow cases
• Extend test coverage with edge cases for subnormal and small normal doubles
• Update Catch2 matcher syntax to current API (Catch::Matchers::WithinAbs/WithinRel)

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
projects/clr/hipamd/include/hip/amd_detail/amd_hip_fp8.h	Refactored __hip_cvt_double_to_e8m0() to use signed arithmetic for exponent conversion, preventing unsigned underflow and correctly handling small normal doubles
projects/hip-tests/catch/unit/deviceLib/fp8_e8m0.cc	Added comprehensive underflow test cases (1e-50, denorm_min, double::min, float::min cast to double) with correct expected values, and updated Catch2 matcher syntax

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[AidanBeltonS]

Thanks for the patch and regression test. This looks good to me, thanks for catching this! One minor comment.

[sshlyapn]

@AidanBeltonS @b-sumner thanks for the review!
Btw, regarding the disabled tests - is there a plan to re‑enable it soon?

[geomin12]

Force merging as hip-test errors occur in various other places:

2026-02-06T20:58:14.3141786Z C:/home/runner/_work/rocm-systems/rocm-systems/projects/hip-tests/catch/unit/streamperthread/hipStreamPerThread_Basic.cc(60): FAILED:
2026-02-06T20:58:14.3141891Z   CHECK( hostMem[i] == (i + 1 + 5) )
2026-02-06T20:58:14.3141973Z with expansion:
2026-02-06T20:58:14.3142060Z   0 == 33432087 (0x1fe2217)

occurs here as well: https://github.com/ROCm/rocm-systems/actions/runs/21760049052/job/62785319334
Will open issue with hip-tests team