mirror of https://github.com/wolfpld/tracy.git synced 2026-10-23 01:00:54 +00:00

Files

Basil Milanich 030e699eb5 Move rocprof on-demand repro to tests/ and convert to CMake

Conform to the new repo layout (master moved repros under tests/, e.g.
tests/cuda/repro/graph). Relocate examples/RocprofOnDemandRepro to
tests/rocprof/repro/on_demand and replace the hand-written Makefile with
a CMake build mirroring the CUDA repro: builds the HIP reproducer, wires
it as a ctest target, and optionally builds the check_gpu_ctx_name
verification helper against the Tracy server library.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

2026-06-08 09:47:22 -05:00

check_gpu_ctx_name.cpp

Move rocprof on-demand repro to tests/ and convert to CMake

2026-06-08 09:47:22 -05:00

CMakeLists.txt

Move rocprof on-demand repro to tests/ and convert to CMake

2026-06-08 09:47:22 -05:00

README.md

Move rocprof on-demand repro to tests/ and convert to CMake

2026-06-08 09:47:22 -05:00

repro.cpp

Move rocprof on-demand repro to tests/ and convert to CMake

2026-06-08 09:47:22 -05:00

README.md

Rocprofiler On-Demand Profiling Repro

Demonstrates that unpatched Tracy crashes when a profiler connects to a HIP application built with TRACY_ON_DEMAND and TRACY_ROCPROF.

Root cause

Three bugs in TracyRocprof.cpp break on-demand profiling:

GpuNewContext not deferred. gpu_context_allocate() writes a GpuNewContext queue item but does not call DeferItem(). When a Tracy client connects late, the context creation message is never replayed. The server then receives GpuZoneBegin events for a context it has never seen, triggering:
```
Assertion `ctx' failed in ProcessGpuZoneBeginImplCommon
```
GpuContextName not deferred. Same function writes the context name ("rocprofv3") without calling DeferItem(). Even after fixing bug 1, a late-connecting client sees the GPU context but it appears unnamed in the profiler. Use check_gpu_ctx_name to verify.
Kernel symbols dropped before init. The data->init guard at the top of tool_callback_tracing_callback() blocks all callbacks before the GPU context is allocated. Kernel symbol registrations (CODE_OBJECT_DEVICE_KERNEL_SYMBOL_REGISTER) happen at HIP init time — before data->init is true — so they are silently dropped. Even if the crash is worked around, kernel names would be missing.

Prerequisites

AMD GPU with working ROCm driver
librocprofiler-sdk.so available (typically at /opt/rocm/lib/)
/opt/rocm/bin/hipcc

Build and run

cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build
./build/repro &
tracy-capture -o repro.tracy -s 5

If ROCm is not under /opt/rocm, pass -DROCM_PATH=/path/to/rocm.

The reproducer is also registered as a ctest target:

ctest --test-dir build -R repro

Verifying the context name

check_gpu_ctx_name loads a .tracy file and prints the GPU context names. It links the Tracy server library, so it is built only on request:

cmake -B build -DBUILD_CHECK_TOOL=ON
cmake --build build --target check_gpu_ctx_name
./build/check_gpu_ctx_name repro.tracy
# Expected (patched):   "GPU context 0: rocprofv3"
# Expected (unpatched): "GPU context 0: (unnamed)"

Exit codes: 0 = all contexts named, 2 = unnamed context found.

What to expect

Tracy version	Result
Unpatched	`tracy-capture` crashes: `Assertion 'ctx' failed`
Patched (GpuNewContext only)	Capture succeeds but GPU context is unnamed
Fully patched	Capture succeeds with ~50 GPU zones and context named "rocprofv3"