[PGO][HIP] Linux HSA-introspection device drain on the profile_rocm reland - #2808
Closed
lfmeadow wants to merge 4 commits into
Closed
[PGO][HIP] Linux HSA-introspection device drain on the profile_rocm reland#2808lfmeadow wants to merge 4 commits into
lfmeadow wants to merge 4 commits into
Conversation
…eland amd-staging now carries the relanded HIP offload PGO runtime as the opt-in clang_rt.profile_rocm library (llvm#201606), but lacks the host-link wiring (llvm#201607, still open upstream), so a HIP+PGO host link leaves the __llvm_profile_offload_* symbols unresolved. This restarts the device-drain work cleanly on top of that reland instead of rebasing the prior branch. Drain runtime (split by platform, mutually exclusive guards so exactly one provides the __llvm_profile_offload_* / __llvm_profile_hip_collect_device_data symbols): - InstrProfilingPlatformROCm.cpp: Linux HSA introspection. Walks every GPU agent's loaded code objects at exit, so it drains device counters even for device modules with no host-side shadow (RCCL-style synthesized kernels) and for kernels on a non-current device. - InstrProfilingPlatformROCmWindows.cpp: the reland host-shadow drain, unchanged, guarded to _WIN32 (Windows has no HSA introspection). - CMakeLists.txt wires both into clang_rt.profile_rocm. Driver host-link wiring (net-new part of llvm#201607, adapted to amd-staging): - Linux.cpp / MSVC.cpp link clang_rt.profile_rocm ahead of the base profile for HIP host links when profiling is requested. - hip-profile-rocm-runtime.hip covers the link ordering. Standalone (non-TheRock) build + test recipe: - compiler-rt/test/profile/device-pgo/{toolchain-cache.cmake,build.sh,README.md} build the toolchain + clang_rt.profile_rocm + the amdgcn device profile runtime in one cmake configure (LLVM_RUNTIME_TARGETS=default;amdgcn-amd-amdhsa). - run_gpu_tests.py (lit-lite runner) derives the multi-device feature from the runtime-visible GPU count via amdgpu-arch; GPU/ and AMDGPU/ .hip tests use Linux HSA CHECK strings and mark agent-walk-only cases UNSUPPORTED: windows. Co-authored-by: Cursor <cursoragent@cursor.com>
Instrumented HIP device code references the amdgcn device profile runtime (__llvm_profile_instrument_gpu and the __llvm_profile_sections bounds table emitted by InstrProfilingPlatformGPU). Nothing on the offload device link pulls it in, so a plain `clang -x hip -fprofile-instr-generate -fcoverage-mapping` link fails with `undefined symbol: __llvm_profile_instrument_gpu` on both Linux and Windows. Forward the arch-suffixed static device profile runtime to the GPU device linker, in both device-link paths: - Clang.cpp LinkerWrapper::ConstructJob, next to -lompdevice, for the --offload-new-driver path HIP uses on Linux. - HIPAMD.cpp constructLldCommand, after --no-whole-archive, for the traditional offload path (e.g. Windows, which does not route device linking through clang-linker-wrapper). The VFS/fs exists() guard keeps this a no-op when the device profile runtime is not installed. Co-authored-by: Cursor <cursoragent@cursor.com>
Validated the standalone recipe end-to-end on a 4x gfx90a host (built the toolchain + clang_rt.profile_rocm + amdgcn libclang_rt.profile.a, then ran the GPU/ and AMDGPU/ suites: 11 passed, 0 failed, with multi-device autodetected). Fixes found while validating: - Build the offload toolchain tools the HIP compile/link actually exec's: clang-offload-bundler, clang-linker-wrapper, llvm-link, llvm-offload-binary, and offload-arch (also installed as amdgpu-arch, used for --offload-arch=native and the multi-device feature). Drop the non-existent clang++ ninja target (the clang target produces the clang++ symlink). - Runner invocation docs: the flag is --hip-lib-path (not --hip-path), and --toolchain-bin must be absolute because RUN lines execute from a temp dir. Co-authored-by: Cursor <cursoragent@cursor.com>
lfmeadow
force-pushed
the
device-pgo-hsa-drain-v3
branch
from
June 9, 2026 20:15
0a4bdac to
5de6939
Compare
ronlieb
approved these changes
Jun 9, 2026
There was a problem hiding this comment.
Pull request overview
This PR relands HIP device PGO / coverage support on top of current amd-staging by ensuring both the device profile runtime and the host ROCm drain runtime (clang_rt.profile_rocm) are linked appropriately, and by adding a Linux HSA-introspection-based exit drain (with a Windows-only legacy host-shadow implementation). It also adds a standalone build-and-run recipe plus targeted tests for the end-to-end workflow.
Changes:
- Add/adjust driver logic to forward the AMDGPU device profile runtime into the offload device link, and link
clang_rt.profile_rocmfor HIP host links when profiling is enabled. - Implement Linux HSA-introspection device draining in
InstrProfilingPlatformROCm.cppand add a Windows-specific legacy drain TU. - Add a standalone “lit-lite” GPU test runner plus new/updated HIP tests and a CMake/Ninja recipe to build the required toolchain + runtimes.
Reviewed changes
Copilot reviewed 18 out of 18 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| compiler-rt/test/profile/run_gpu_tests.py | New minimal lit-style runner to execute HIP .hip tests without llvm-lit/FileCheck installed. |
| compiler-rt/test/profile/GPU/instrprof-hip-nondefault-device.hip | Update expectations for Linux HSA agent-walk drain; gate off Windows. |
| compiler-rt/test/profile/GPU/instrprof-hip-multiple-kernels.hip | Fix expected mangled names for kernels. |
| compiler-rt/test/profile/GPU/instrprof-hip-multi-gpu.hip | Update expectations for multi-GPU drain via HSA agent walk; gate off Windows. |
| compiler-rt/test/profile/GPU/instrprof-hip-coverage.hip | Tighten FileCheck pattern to avoid false matches on “80.00%” etc. |
| compiler-rt/test/profile/device-pgo/toolchain-cache.cmake | New cache file to build host toolchain + profile_rocm + amdgcn device profile runtime in one configure. |
| compiler-rt/test/profile/device-pgo/README.md | New documentation for standalone build and GPU test execution. |
| compiler-rt/test/profile/device-pgo/build.sh | New convenience script to build required tools and runtimes for device-PGO tests. |
| compiler-rt/test/profile/AMDGPU/lit.local.cfg.py | New local lit config gating AMDGPU drain tests on amdgpu feature. |
| compiler-rt/test/profile/AMDGPU/device-symbols.hip | New test asserting __llvm_profile_sections is present in device ELF dynsym. |
| compiler-rt/test/profile/AMDGPU/device-no-kernel.hip | New robustness test for no-kernel-launch scenario (no crash / appropriate no-op). |
| compiler-rt/test/profile/AMDGPU/device-early-collect.hip | New regression test for early explicit collection not poisoning later atexit drain. |
| compiler-rt/test/profile/AMDGPU/device-basic.hip | New end-to-end test covering host+device profraw merge and llvm-cov reporting. |
| compiler-rt/lib/profile/InstrProfilingPlatformROCmWindows.cpp | New Windows-only legacy host-shadow + hipMemcpy drain implementation. |
| compiler-rt/lib/profile/InstrProfilingPlatformROCm.cpp | Linux/Unix HSA-introspection drain implementation + refactor and safety checks. |
| compiler-rt/lib/profile/CMakeLists.txt | Wire both ROCm platform TUs into clang_rt.profile_rocm and adjust removal logic. |
| clang/lib/Driver/ToolChains/HIPAMD.cpp | Forward device profile runtime archive into traditional device link path. |
| clang/lib/Driver/ToolChains/Clang.cpp | Forward device profile runtime archive into linker-wrapper offload device link. |
- ROCm drain: split alreadySeenBounds() into a pure seenBounds() check and a separate recordBounds(); record a bounds tuple only after a successful drain (>0) or a confirmed-empty section (0). A transient drain failure (<0) is left unrecorded so a later agent or a subsequent collect call can retry instead of silently dropping device profiling data. - ROCm drain: fix the DrainCompleted latch comment to match the code (latch only when data was actually drained, never on the "no instrumented code object" walk, which would otherwise suppress a real atexit drain after a later launch). - HIPAMD.cpp: probe the device profile runtime via TC.getVFS().exists() (matching the new-offload-driver path in Clang.cpp) so overlay/virtual filesystems are honored instead of bypassing them with llvm::sys::fs::exists(). - run_gpu_tests.py: add a per-RUN-line timeout (default 600s, <=0 disables) so a hung GPU/compiler invocation can no longer wedge the whole run. Co-authored-by: Cursor <cursoragent@cursor.com>
Author
|
Superseded by the upstream llvm#203056, which landed the additive Strategy-B device-profile drain (supplemental HSA-introspection pass with content-dedup on top of upstream's host-shadow drain from llvm#202095) and is now merged down into This PR's Strategy-A (HSA-sole on Linux) design plus its driver hunks were superseded during upstream review (the new-offload-driver path links the device RT without the |
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.
Summary
Restarts the HIP device profile-drain work cleanly on top of the current
amd-staging, which now carries the relanded opt-inclang_rt.profile_rocmruntime (llvm#201606). Supersedes #2743 (closed).End-to-end chain for
clang -x hip -fprofile-instr-generate -fcoverage-mapping:__llvm_profile_instrument_gpu+__llvm_profile_sections) -- already in amd-staging.libclang_rt.profile.a-- added here (Clang.cppnew-offload-driver path +HIPAMD.cpptraditional path).__llvm_profile_offload_register_shadow_variable-- already in amd-staging (CGCUDANV).clang_rt.profile_rocm-- added here (Linux.cpp/MSVC.cpp, the net-new part of Reland HIP offload PGO compiler support and link the device-profile runtime llvm/llvm-project#201607).Details
InstrProfilingPlatformROCm.cpp-> Linux HSA introspection: walks every GPU agent's loaded code objects at exit, draining device counters even for device modules with no host-side shadow (RCCL-style synthesized kernels) and for kernels left on a non-current device.InstrProfilingPlatformROCmWindows.cpp(new) -> the reland host-shadow drain, unchanged, guarded_WIN32.compiler-rt/lib/profile/CMakeLists.txtwires both intoclang_rt.profile_rocm.__llvm_profile_instrument_gpu):Clang.cppLinkerWrapper::ConstructJob(Linux/new-offload-driver) andHIPAMD.cppconstructLldCommand(Windows/traditional). Without it the device link failsundefined symbol: __llvm_profile_instrument_gpu.Linux.cpp/MSVC.cpplinkclang_rt.profile_rocmahead of the base profile for HIP host links when profiling is requested;clang/test/Driver/hip-profile-rocm-runtime.hipcovers ordering.compiler-rt/test/profile/device-pgo/(toolchain-cache.cmake,build.sh,README.md): toolchain +clang_rt.profile_rocm+ amdgcn device runtime in one configure (LLVM_RUNTIME_TARGETS=default;amdgcn-amd-amdhsa).run_gpu_tests.py(lit-lite) derivesmulti-devicefrom the runtime-visible GPU count viaamdgpu-arch;GPU/+AMDGPU/.hiptests use Linux HSA CHECK strings and mark agent-walk-only casesUNSUPPORTED: windows.Validation
Built and run end-to-end on a 4x gfx90a host via the standalone recipe:
build.shproduced the toolchain + hostlibclang_rt.profile_rocm.a(contains the HSA drain:drainDevices,__llvm_profile_hip_collect_device_data; the Windows TU compiled empty, no duplicate symbols) + amdgcnlibclang_rt.profile.a.run_gpu_tests.pyoverGPU/+AMDGPU/: 11 passed, 0 failed, withmulti-deviceauto-detected (so the multi-GPU and non-default-device agent-walk tests actually ran).libclang_rt.profile_rocm.abeforelibclang_rt.profile.a; non-HIP PGO links neitherprofile_rocm.clang-linker-wrapperdevice link receives--device-linker=amdgcn-amd-amdhsa=.../libclang_rt.profile.a(theClang.cppchange), confirming the device-runtime link.Remaining: Windows host-shadow path not built/run here (no Windows box); CI wiring is out of scope for this PR.