Repository navigation
fix(rocprim): Add proper gfx950 config support, and fix launch_manager Wavefront support - #13251
Merged
stanleytsang-amd merged 1 commit intoOct 8, 2026
Conversation
…handle Wavefront size
✅ All Checks Passed — Ready for Review
📖 Need help? See the Policy FAQ for details on every check and how to fix failures. |
|
🎉 All checks passed! This PR is ready for review. |
amd-jmahovsky
approved these changes
Oct 7, 2026
amd-jmahovsky
left a comment
amd-jmahovsky
left a comment
Contributor
There was a problem hiding this comment.
LGTM, but onesweep radix sort performance may be changed.
1 task done
stanleytsang-amd
deleted the
users/stanleytsang-amd/fix-gfx950-configs-rocprim
branch
October 8, 2026 14:50
Contributor
TheRock Submodule Bump ActivityNewest first
|
stanleytsang-amd
added a commit
that referenced
this pull request
Oct 17, 2026
…r wavefront support (legacy) (#13881) ## Motivation Legacy \`projects/rocprim\` has the same bug fixed for hipccl3 in #13251. \`rocprim.linking\` fails in TheRock CI on gfx950 (https://github.com/ROCm/TheRock/actions/runs/38904910579) with: \`hip_global.cpp: Cannot find Symbol with name: _ZN7rocprim17ROCPRIM_400800_NS6detail17trampoline_kernelI...launch_manager...\` gfx950 config support also needed fixing: it was listed as both CDNA4 and CDNA3, and MI355X was not supported. ## Technical Details \`visit_config\` resolves size64/size32 for the fallback (dynamic wavefront) target, but \`launch_manager::make_launch_plan\` ignored it and called \`trampoline_kernel<..., Target, LaunchSelector>\`, so \`TargetWaveSize\` fell back to \`get_wavefront_size(Target::g)\`. For the fallback target that is \`dynamic\` in the host pass but a concrete size in the gfx950 device pass, so the host looked up a kernel symbol that is not in the code object. - \`launch_manager\` takes a \`Wavefront\` template parameter and forwards it to \`trampoline_kernel\`. - \`visit_config\` passes the resolved wavefront explicitly in the fallback branch. - Removed gfx950 from \`ROCPRIM_IS_CDNA3()\`. - Added \`gpu::mi355x\` / \`"MI355X"\`. ## Test Plan Rerun \`rocprim.linking\` and the rocPRIM tests on gfx950 in CI. Not built or tested locally. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
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.
Motivation
#12972 introduced a bug - the new launch_manager loses the wavefront size for fallback configs. We are seeing the following failure in CI:
88/91 Test #88: rocprim.linking .............................Subprocess aborted***Exception: 0.26 sec Running main() from /googletest-1.14.0/googletest/src/gtest_main.cc [==========] Running 3 tests from 1 test suite. [----------] Global test environment set-up. [----------] 3 tests from TestLinking [ RUN ] TestLinking.Test0 :0:/TheRock/rocm-systems/projects/clr/hipamd/src/hip_global.cpp:66 : 14973002789391 us: Cannot find Symbol with name: _ZN7rocprim17ROCPRIM_500000_NS6detail17trampoline_kernelINS0_14default_configENS1_20scan_config_selectorIiEEZZZNS1_9scan_implILNS1_25lookback_scan_determinismE0ELb0ELb0ES3_PiS8_iNS0_4plusIiEEiEEDaPvRmT3_T4_T5_mT6_P12ihipStream_tbENKUlT_T0_E_clISt17integral_constantIbLb0EESN_IbLb1EEEEDaSJ_SK_ENKUlSJ_SK_T1_T2_E_clINS1_14launch_managerIS3_NS1_11comp_targetILNS1_3genE0ELNS1_11target_archE4294967295ELNS1_3gpuE0ELNS1_3repE0EEEEES3_S10_SN_INS0_4arch9wavefront6targetELS14_1EEEEDaSJ_SK_SQ_SR_EUlSJ_E0_S10_NS1_30default_config_static_selectorELS14_2EEEvSQ_Also, gfx950 config support needed some fixing - it was incorrectly listed as both CDNA4 and CDNA3, and also did not support MI355X.
Technical Details
The old
make_launch_planin config_types handled the dynamic wavefront case explicitly. It instantiatedtrampoline_kernel<..., Target, LaunchSelector, size64>or size32 depending on target_current.warp_size.visit_configstill works out size64/size32 for the fallback branch and passes it to the visitor. But launch_manager::make_launch_plan ignores it. It callstrampoline_kernel<Config, ConfigSelector, Kernel, Target, LaunchSelector>, so TargetWaveSize falls back to its default, get_wavefront_size(Target::g).For the fallback target, that default differs between host and device. On the host, max_size() is 64 and min_size() is not, so it resolves to dynamic. In the gfx950 device pass, max_size() equals min_size(), so it resolves to size64.
The host therefore asks for a kernel mangled with dynamic. The code object only contains the size64 instantiation, which gives "Cannot find Symbol".
Test Plan
Re ran all tests on gfx950
Test Result
Pass
Submission Checklist