[rocprofiler-compute] Roofline: Use gfx950 builtins for MFMA by vedithal-amd · Pull Request #3886 · ROCm/rocm-systems

vedithal-amd · 2026-03-09T19:48:23Z

Motivation

Resolves #3506

MFMA tests for F16, BF16, and I8 were far below spec on gfx950.

Copy of #3837 to remove unintended rccl changes being merged.
Original author @benrichard-amd

Technical Details

gfx950 introduced new MFMA instructions with higher throughput for F16, BF16 and I8. Use the builtins for these on gfx950.

type	peak	actual
FP16	2.3 PFLOPS	1.8 PFLOPS
BF16	2.3 PFLOPS	2.0 PFLOPS
I8	4.6 POPS	4.2POPS

JIRA ID

Test Plan

Run roofline on MI350. Verify expected performance improvement.
Verify reported MI350 TFLOPs using rocprof-compute.
Run roofline on MI100/MI250/MI300. Verify works as expected.

Test Result

Submission Checklist

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

Copilot

Pull request overview

Updates profiling/benchmarking infrastructure to improve MFMA roofline accuracy on gfx950 and expands ROCprofiler-SDK/rocprofv3 functionality (pause/resume, topology overrides, firmware restrictions), along with several build/test/config maintenance tweaks across the repo.

Changes:

Update rocprofiler-compute MFMA microbench kernels to use gfx950-specific MFMA builtins and adjust peak-rate tables.
Add rocprofiler-sdk rocprofv3 roctx pause/resume integration tests and implement pause/resume behavior using context start/stop.
Introduce firmware restriction parsing/checking (YAML) and new dynamic-library path helpers; update metrics file naming/lookup and related docs/build scripts.

Reviewed changes

Copilot reviewed 158 out of 162 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
projects/rocshmem/CMakeLists.txt	Disable examples when building tests-only.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/tracing/validate.py	New pytest validator for tracing output.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/tracing/pytest.ini	New pytest configuration for tracing validator.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/tracing/conftest.py	Fixture for loading JSON output for tracing tests.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/tracing/CMakeLists.txt	New CTest integration tests for tracing (execute + validate).
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/pc-sampling/validate.py	New pytest validator for PC sampling output.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/pc-sampling/pytest.ini	New pytest configuration for PC sampling validator.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/pc-sampling/conftest.py	Fixture for loading JSON output; skips if unavailable.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/pc-sampling/CMakeLists.txt	New CTest integration tests for PC sampling (execute + validate).
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/nested-pause-resume/validate.py	New pytest validator for nested pause/resume behavior.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/nested-pause-resume/pytest.ini	New pytest configuration for nested pause/resume validator.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/nested-pause-resume/conftest.py	Fixture for loading JSON output for nested pause/resume tests.
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/nested-pause-resume/CMakeLists.txt	New CTest integration tests for nested pause/resume (execute + validate).
projects/rocprofiler-sdk/tests/rocprofv3/roctx-pause-resume/CMakeLists.txt	Add roctx pause/resume test subdirectories.
projects/rocprofiler-sdk/tests/rocprofv3/CMakeLists.txt	Register roctx-pause-resume tests in rocprofv3 suite.
projects/rocprofiler-sdk/tests/bin/roctx-pause-resume/roctx-pause-resume.cpp	Add HIP+ROCTX test binary exercising pause/resume + nested behavior.
projects/rocprofiler-sdk/tests/bin/roctx-pause-resume/CMakeLists.txt	Build/test binary configuration for roctx-pause-resume.
projects/rocprofiler-sdk/tests/bin/CMakeLists.txt	Add roctx-pause-resume test binary to build.
projects/rocprofiler-sdk/source/share/rocprofiler-sdk/counter_defs.yaml	Add firmware restriction metadata to counters YAML.
projects/rocprofiler-sdk/source/share/rocprofiler-sdk/CMakeLists.txt	Change installed share YAML from `counter_defs.yaml` to `config.yaml`.
projects/rocprofiler-sdk/source/lib/tests/common/dl.cpp	Add unit tests for new dl path helpers.
projects/rocprofiler-sdk/source/lib/tests/common/CMakeLists.txt	Build/link new `dl.cpp` test.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/0/name	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/0/mem_banks/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/0/io_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/0/io_links/1/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/0/io_links/2/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/0/gpu_id	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/1/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/1/name	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/1/mem_banks/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/1/mem_banks/0/used_memory	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/1/io_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/1/p2p_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/1/p2p_links/1/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/1/gpu_id	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/2/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/2/name	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/2/mem_banks/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/2/mem_banks/0/used_memory	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/2/io_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/2/p2p_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/2/p2p_links/1/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/2/gpu_id	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/3/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/3/name	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/3/mem_banks/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/3/mem_banks/0/used_memory	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/3/io_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/3/p2p_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/3/p2p_links/1/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/3/gpu_id	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/name	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/mem_banks/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/mem_banks/0/used_memory	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/io_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/io_links/1/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/io_links/2/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/io_links/3/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/io_links/4/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/io_links/5/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/io_links/6/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/io_links/7/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/p2p_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/4/gpu_id	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/5/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/5/name	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/5/mem_banks/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/5/mem_banks/0/used_memory	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/5/io_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/5/gpu_id	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/6/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/6/name	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/6/mem_banks/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/6/mem_banks/0/used_memory	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/6/io_links/0/properties	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/data/topology/nodes/6/gpu_id	Add local topology fixture data for agent tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/agent.cpp	Add local-topology verification and improve agent type/name handling.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/tests/CMakeLists.txt	Copy topology fixtures into unit test runtime dir + trigger reconfigure on updates.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/registration.cpp	Set `ROCPROFILER_REGISTER_LIBRARY` to current SDK path using new dl helpers.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hip/hip.def.cpp	Remove hipModuleGetLoadingMode instrumentation entry.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/hip/abi.cpp	Update ABI enforcement to remove hipModuleGetLoadingMode.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/firmware_restrictions.cpp	Add unit tests for firmware restriction YAML parsing/checking.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/tests/CMakeLists.txt	Build firmware restriction tests.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/metrics.cpp	Switch metrics YAML lookup from counter_defs.yaml to config.yaml.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/firmware_restrictions.hpp	New firmware restriction API.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/firmware_restrictions.cpp	Implement YAML parsing + installed-file check.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/controller.cpp	Run firmware restriction check during controller init.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/CMakeLists.txt	Build/link firmware restriction implementation.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/agent.cpp	Add topology path overrides and default agent naming fallback.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk-tool/tool.cpp	Implement pause/resume by starting/stopping relevant contexts.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk-tool/config.hpp	Add selected-regions reference counting config flag.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk-rocattach/symbol_lookup.cpp	Use new dl helpers to locate rocattach library path.
projects/rocprofiler-sdk/source/lib/common/logging.hpp	Include fmt helpers used by logging.
projects/rocprofiler-sdk/source/lib/common/environment.hpp	Extend get_env template to floating-point.
projects/rocprofiler-sdk/source/lib/common/environment.cpp	Implement floating-point get_env specialization + expand set_env instantiations.
projects/rocprofiler-sdk/source/lib/common/dl.hpp	New dl helper declarations.
projects/rocprofiler-sdk/source/lib/common/dl.cpp	New dl helper implementations (iterate loaded libs, dladdr helpers).
projects/rocprofiler-sdk/source/lib/common/CMakeLists.txt	Build/install new dl helpers.
projects/rocprofiler-sdk/source/include/rocprofiler-sdk/hip/runtime_api_id.h	Remove hipModuleGetLoadingMode enum ID.
projects/rocprofiler-sdk/source/include/rocprofiler-sdk/hip/api_args.h	Remove hipModuleGetLoadingMode args.
projects/rocprofiler-sdk/source/include/rocprofiler-sdk/cxx/enum_string.hpp	Remove hipModuleGetLoadingMode label + adjust asserts.
projects/rocprofiler-sdk/source/docs/how-to/using-rocprofv3.rst	Update docs to refer to config.yaml.
projects/rocprofiler-sdk/source/docs/api-reference/counter_collection_services.rst	Update docs to refer to config.yaml and document firmware restrictions.
projects/rocprofiler-sdk/source/bin/rocprofv3.py	Add CLI flag/env plumb for selected-regions ref counting.
projects/rocprofiler-sdk/cmake/rocprofiler_utilities.cmake	Extend unit-test helper to install/copy data and configure file triggers.
projects/rocprofiler-sdk/cmake/rocprofiler_config_packaging.cmake	Include CPackComponent for packaging config.
projects/rocprofiler-sdk/CHANGELOG.md	Add/update changelog entries for multiple features/fixes.
projects/rocprofiler-sdk/.cmake-format.yaml	Teach cmake-format about new macro args (DATA/CONFIGURE_FILES).
projects/rocprofiler-register/cmake/rocprofiler_register_config_packaging.cmake	Include CPackComponent for packaging config.
projects/rocprofiler-compute/src/utils/benchmark.py	Use gfx950 MFMA builtins and adjust peak tables for FP16/BF16/I8.
projects/rocprofiler-compute/CHANGELOG.md	Note MFMA roofline fix for MI350.
projects/rocprof-trace-decoder/test/CMakeLists.txt	Add dependency ordering between execute and check tests.
projects/rocminfo/.github/dependabot.yml	Remove per-project dependabot config (now centralized).
projects/rocm-smi-lib/.github/dependabot.yml	Remove per-project dependabot config (now centralized).
projects/rdc/.github/dependabot.yml	Remove per-project dependabot config (now centralized).
projects/rccl/tools/topo_expl/stubs.cc	Add AINIC-related stub symbols for topo_expl.
projects/rccl/src/transport/net.cc	Add NIC detection + set primary NIC details + expose API.
projects/rccl/src/plugin/net.cc	Select ROCm internal IB plugin depending on detected NIC.
projects/rccl/src/misc/ionicdvwrap.cc	Gate ionicdv symbol wrapping on detected AINIC usage.
projects/rccl/src/include/plugin/tuner/tuner_v5.h	Add bwRatio field to tuner constants struct.
projects/rccl/src/include/net.h	Add rcclPrimaryNic/rcclUseAinic API + NIC type/info structs.
projects/rccl/src/graph/tuning.cc	Add tuning model, wire bwRatio/hwLatencies into tuner constants, apply NIC-based overrides.
projects/rccl/src/graph/topo.h	Declare `rcclApplyTuningOverrides`.
projects/rccl/src/graph/rome_models.cc	Centralize gfx950 tuning override logic.
projects/rccl/ext-tuner/example/plugin.c	Update example tuner to use bwRatio and hwLatencies.
projects/rccl/ext-tuner/example/nccl/tuner.h	Update example tuner header struct layout.
projects/rccl/ext-src/rocm_netib.patch	Update patch to use rcclUseAinic instead of param directly.
projects/hipother/hipnv/include/hip/nvidia_detail/nvidia_hip_runtime_api.h	Remove hipModuleGetLoadingMode wrapper (NV path).
projects/hip/include/hip/hip_runtime_api.h	Remove hipModuleGetLoadingMode API + associated enum type.
projects/hip/.github/dependabot.yml	Remove per-project dependabot config (now centralized).
projects/hip-tests/catch/unit/module/hipModuleGetLoadingMode.cc	Remove hipModuleGetLoadingMode test coverage (API removed).
projects/hip-tests/catch/unit/module/CMakeLists.txt	Stop building removed hipModuleGetLoadingMode test.
projects/clr/rocclr/utils/flags.hpp	Remove HIP_MODULE_LOADING flag definition.
projects/clr/rocclr/device/devprogram.cpp	Refactor symbol enumeration; always load vars/funcs for DynCO.
projects/clr/hipamd/src/hip_table_interface.cpp	Remove hipModuleGetLoadingMode dispatch wrapper.
projects/clr/hipamd/src/hip_platform.hpp	Remove PlatformState::GetLoadingMode declaration.
projects/clr/hipamd/src/hip_platform.cpp	Remove module loading mode handling in platform state.
projects/clr/hipamd/src/hip_module.cpp	Remove hipModuleGetLoadingMode API implementation.
projects/clr/hipamd/src/hip_hcc.map.in	Stop exporting hipModuleGetLoadingMode symbol.
projects/clr/hipamd/src/hip_graph_internal.hpp	Update leaf node sync detection logic.
projects/clr/hipamd/src/hip_code_object.hpp	Remove lazy-loading state flags; adjust managed-var accessor.
projects/clr/hipamd/src/hip_code_object.cpp	Always populate globals at load; remove lazy flags.
projects/clr/hipamd/src/hip_api_trace.cpp	Remove hipModuleGetLoadingMode tracing entry + drop step version to 25.
projects/clr/hipamd/src/amdhip.def	Stop exporting hipModuleGetLoadingMode from def file.
projects/clr/hipamd/include/hip/amd_detail/hip_prof_str.h	Remove hipModuleGetLoadingMode from profiler strings/args.
projects/clr/hipamd/include/hip/amd_detail/hip_api_trace.hpp	Drop runtime API step version to 25 and remove loading-mode pointer.
projects/clr/CHANGELOG.md	Remove changelog mention of hipModuleGetLoadingMode being added.
projects/amdsmi/tests/amd_smi_test/CMakeLists.txt	Update test link target to use selected AMD SMI library target.
projects/amdsmi/src/nic/ai-nic/amdsmi_unified/CMakeLists.txt	Improve include interface for NIC static lib for build/install.
projects/amdsmi/src/CMakeLists.txt	Build both shared/static, restrict exported symbols via version script, and update install/export logic.
projects/amdsmi/example/CMakeLists.txt	Link examples against static library when available.
projects/amdsmi/amdsmi_cli/amdsmi_init.py	Adjust init error handling; prefer driver-reload guidance.
projects/amdsmi/amdsmi_cli/amdsmi_helpers.py	Update partition-mode warning text and remove reload-driver prompt.
projects/amdsmi/amdsmi_cli/amdsmi_commands.py	Delay default info collection until amdgpu initialized; update reload message.
projects/amdsmi/CMakeLists.txt	Add options to auto-build static libs and build both libs.
projects/amdsmi/CHANGELOG.md	Document removal of reload-driver CLI option(s).
projects/amdsmi/.github/dependabot.yml	Remove per-project dependabot config (now centralized).
.github/dependabot.yml	Centralize pip dependabot configs and reduce PR rate via grouping.

Comments suppressed due to low confidence (6)

projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/metrics.cpp:1

loadMetrics now hardcodes config.yaml, but this diff does not add a config.yaml file (and still modifies counter_defs.yaml). This will reliably fail at runtime with the ROCP_FATAL_IF when config.yaml is not present. Either add/rename the YAML file to config.yaml (and ensure it’s installed to the searched location), or keep the filename as counter_defs.yaml consistently across lookup, install rules, and documentation.
projects/rocprofiler-sdk/source/lib/rocprofiler-sdk/counters/firmware_restrictions.cpp:1
Both helpers return std::string but currently return path expressions (path / ...). Unless common::filesystem::path provides an implicit conversion to std::string (unusual), this is a compile error. Convert the resulting path explicitly (e.g., .string()) before returning, and consider normalizing to absolute/canonical if that’s what downstream expects.
projects/rocprofiler-sdk/source/docs/api-reference/counter_collection_services.rst:1
The documented key fw_restriction_schema_version uses underscores, but the implementation in parse_firmware_restrictions requires fw-restriction-schema-version (hyphens). This mismatch will cause users to author YAML that the parser rejects. Update the docs to match the implemented key (or update the parser to accept both for backward/forward compatibility).
projects/rocprofiler-sdk/source/lib/common/dl.cpp:1
In the ROCPROFILER_SYMBOL_PATH_USE_DLOPEN branch, the function never returns a resolved path even if dlsym succeeds (it always falls through to return std::nullopt). If this code path is meant to be usable for debugging, it should translate the found symbol address (_fn) into a library path (e.g., via dladdr) and return it, or else the branch is misleading and ineffective.
projects/rocprofiler-sdk/source/lib/common/dl.cpp:1
In the ROCPROFILER_SYMBOL_PATH_USE_DLOPEN branch, the function never returns a resolved path even if dlsym succeeds (it always falls through to return std::nullopt). If this code path is meant to be usable for debugging, it should translate the found symbol address (_fn) into a library path (e.g., via dladdr) and return it, or else the branch is misleading and ineffective.
projects/rocprofiler-sdk/source/lib/common/dl.cpp:1
In the ROCPROFILER_SYMBOL_PATH_USE_DLOPEN branch, the function never returns a resolved path even if dlsym succeeds (it always falls through to return std::nullopt). If this code path is meant to be usable for debugging, it should translate the found symbol address (_fn) into a library path (e.g., via dladdr) and return it, or else the branch is misleading and ineffective.

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

Copilot · 2026-03-09T19:54:39Z

projects/clr/rocclr/device/devprogram.cpp

 amd_comgr_status_t getSymbolFromModule(amd_comgr_symbol_t symbol, void* userData) {
+  size_t nlen = 0;
+  size_t* userDataInfo = nullptr;
  amd_comgr_status_t status;
  amd_comgr_symbol_type_t type;
+  std::vector<std::string>* var_names = nullptr;


This introduces (1) unused locals (userDataInfo, var_names) which can break builds under -Werror, and (2) a leak: if symbol_get_info(...NAME...) fails, the function returns without delete[] name. Prefer removing the unused variables and using RAII (e.g., std::string/std::vector<char>/std::unique_ptr<char[]>) so early returns don’t leak.

Copilot · 2026-03-09T19:54:39Z

projects/clr/rocclr/device/devprogram.cpp

+  /* Retrieve the symbol name */
+  char* name = new char[nlen + 1];
+  status = amd::Comgr::symbol_get_info(symbol, AMD_COMGR_SYMBOL_INFO_NAME, name);
+  if (status != AMD_COMGR_STATUS_SUCCESS) {
+    return status;
+  }


This introduces (1) unused locals (userDataInfo, var_names) which can break builds under -Werror, and (2) a leak: if symbol_get_info(...NAME...) fails, the function returns without delete[] name. Prefer removing the unused variables and using RAII (e.g., std::string/std::vector<char>/std::unique_ptr<char[]>) so early returns don’t leak.

Copilot · 2026-03-09T19:54:40Z

projects/clr/rocclr/device/devprogram.cpp

+    sym_info->var_names->push_back(std::string(name));
  }

+  delete[] name;


This introduces (1) unused locals (userDataInfo, var_names) which can break builds under -Werror, and (2) a leak: if symbol_get_info(...NAME...) fails, the function returns without delete[] name. Prefer removing the unused variables and using RAII (e.g., std::string/std::vector<char>/std::unique_ptr<char[]>) so early returns don’t leak.

* Clear CHANGELOG indicating we resolved roofline peaks for MI350 * Fix typo in pre-processor guard preventing roofline from running on MI300 * Ruff formatting

* Use gfx950 builtin for MFMA FP16 * Use gfx950 builtin for MFMA BF16 * Use gfx950 builtin for MFMA I8 * Fix comments * Update copyright * Update CHANGELOG * Fix formatting * Review comments * Clear CHANGELOG indicating we resolved roofline peaks for MI350 * Fix typo in pre-processor guard preventing roofline from running on MI300 * Ruff formatting * Fix uninitialized variables --------- Co-authored-by: benrichard-amd <ben.richard@amd.com>

Copilot AI review requested due to automatic review settings March 9, 2026 19:48

vedithal-amd requested review from a team and prbasyal-amd as code owners March 9, 2026 19:48

vedithal-amd changed the base branch from develop to users/vedithal/rocprofiler-compute-temp-develop March 9, 2026 19:48

vedithal-amd requested a review from a team as a code owner March 9, 2026 19:48

github-actions bot added project: rocprofiler-compute project: rccl labels Mar 9, 2026

vedithal-amd force-pushed the benrichard-amd/roofline-mfma-gfx950 branch from fc5d17f to 8d9f4c8 Compare March 9, 2026 19:49

vedithal-amd requested review from a team, dayatsin-amd and shwetagkhatri as code owners March 9, 2026 19:49

github-actions bot added documentation Improvements or additions to documentation project: amdsmi project: rocm-smi-lib project: rocprofiler-register project: rocprofiler-sdk project: rdc github actions Pull requests that update GitHub Actions code labels Mar 9, 2026

github-actions bot added project: clr project: hip project: hipother project: hip-tests project: rocminfo project: rocshmem project: rocprof-trace-decoder labels Mar 9, 2026

Copilot AI reviewed Mar 9, 2026

View reviewed changes

vedithal-amd force-pushed the benrichard-amd/roofline-mfma-gfx950 branch from 8d9f4c8 to ed700d7 Compare March 9, 2026 19:56

Use gfx950 builtin for MFMA FP16

1716c94

vedithal-amd force-pushed the benrichard-amd/roofline-mfma-gfx950 branch from ed700d7 to 1716c94 Compare March 9, 2026 19:59

benrichard-amd and others added 8 commits March 9, 2026 19:59

Use gfx950 builtin for MFMA BF16

b3216fb

Use gfx950 builtin for MFMA I8

9db42af

Fix comments

18cd682

Update copyright

754cf4c

Update CHANGELOG

c9873a6

Fix formatting

1515ee7

Review comments

90220a0

* Clear CHANGELOG indicating we resolved roofline peaks for MI350 * Fix typo in pre-processor guard preventing roofline from running on MI300 * Ruff formatting

Fix uninitialized variables

a102094

feizheng10 approved these changes Mar 9, 2026

View reviewed changes

vedithal-amd merged commit 9e338f9 into users/vedithal/rocprofiler-compute-temp-develop Mar 9, 2026
14 of 16 checks passed

vedithal-amd deleted the benrichard-amd/roofline-mfma-gfx950 branch March 9, 2026 20:08

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[rocprofiler-compute] Roofline: Use gfx950 builtins for MFMA#3886

[rocprofiler-compute] Roofline: Use gfx950 builtins for MFMA#3886
vedithal-amd merged 9 commits intousers/vedithal/rocprofiler-compute-temp-developfrom
benrichard-amd/roofline-mfma-gfx950

vedithal-amd commented Mar 9, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Mar 9, 2026

Uh oh!

Copilot AI Mar 9, 2026

Uh oh!

Copilot AI Mar 9, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

vedithal-amd commented Mar 9, 2026

Motivation

Technical Details

JIRA ID

Test Plan

Test Result

Submission Checklist

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Mar 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 9, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Mar 9, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants