Changelog

v0.3.5

Fix 7 documentation inconsistencies with codebase (#89): remove fictional RTL_NO_INJECT, document all 8 DB tables, add missing roctx functions, fix stale rpd-lite/rpd_lite references

Dispatch info: Record hardware queue ID, workgroup size, and grid dimensions per kernel in the completionSignal column of rocpd_op
Gzip Perfetto output: rtl trace now produces compressed .json.gz files for Perfetto timeline visualization
Build fix: prefer repo-root librtl.so over system-installed version during development
Add ROCR interceptible queue SEGFAULT reproducer (repro/)
Sync __version__ with pyproject.toml

Fix roctx support: rtl trace now sets LD_PRELOAD automatically so roctx markers work without manual setup
Add tutorial: profiling prefill vs decode with roctx markers
Add MI355X benchmark results and hot trace viewer as sample pages

Fix SQLite concurrency crash: flush() and close() now hold g_db_mutex, preventing races with record_kernel() batch commits. GLM-5 TP=8 lite mode was crashing with 1788 SQLite errors leading to GPU memory fault.
Default mode changed to lite: Lite mode is now the default (RTL_MODE unset = lite). Safe for all ROCm versions including 7.2 which has the ROCR InterceptQueue::staging_buffer_ heap overflow bug. Use RTL_MODE=default for full count==1 profiling (safe on ROCm 7.13+).
Removed redundant atexit handler for DB flush (shutdown already handles it).

Three modes: default (signal injection, skip graph replay), lite (also skip has-signal packets, ~0% overhead), full (profile everything including graph replay, requires ROCm 7.13+)
Set via RTL_MODE env var or rtl trace --mode CLI flag
Default mode: profiles all count==1 kernel dispatches with real GPU timing, skips graph replay batches
Lite mode: additionally skips NCCL and other kernels with existing completion signals, matching v0.1.1 behavior for near-zero overhead
Full mode: profiles graph replay batches (count > 1). Requires ROCm 7.13+ with ROCR fix to avoid InterceptQueue::staging_buffer_ heap overflow

Root cause identified: hsa_amd_queue_intercept_create has a heap overflow bug in InterceptQueue::staging_buffer_ (hardcoded to 256 entries). Fixed upstream in rocm-systems commit 559d48b1.
Default and lite modes skip batch submissions (count > 1) as a workaround
RTL_DEBUG=1/2: Packet-level diagnostic logging. Level 1 logs per-call summary. Level 2 adds per-packet details.

Original completion_signal is saved before injection and forwarded via hsa_signal_subtract_screlease after timestamp collection. Packets with app-provided signals are now profiled correctly instead of being skipped.

Added TestBatchSkip, TestSignalForwarding, TestNoInjectMode, TestDebugLogging source-code audit tests
Added TestHipGraph GPU E2E tests (basic, multi-stream, large, stress, batch skip logging, no-0x1009)
HIP graph workloads added to gpu_workload binary (hipgraph, hipgraph_ms, hipgraph_large, hipgraph_stress)

Fixed profiling_set_profiler_enabled return value not checked in RTL_NO_INJECT path
Fixed version mismatch between pyproject.toml (0.2.1) and __init__.py (0.3.0)

Breaking: Replaced observe-only profiling with signal injection (#31)
- HIP runtime (ROCm 7.2) does not set completion_signal on kernel dispatch packets
- Signal pool (64 pre-allocated, 4096 max) avoids per-dispatch allocation overhead
- No extra HSA queues (avoids TP=8 OOM from barrier-packet approach)
Fix: batch dispatch (count > 1) no longer silently dropped
Added diagnostic counters printed at shutdown for each process