v0.7.1

• trace_count and trace_query read an empty table whose export xctrace 27.1 cuts right after the schema
  (time-profile under Wine with PERF_INSTRUMENTS="Time Profiler") as 0 rows instead of failing with
  exit -5: no message.

v0.7.0

• PERF_INSTRUMENTS="Time Profiler;Metal GPU Counters" adds Instruments to the Metal System Trace recording; their
  tables live in the .trace, so it pairs with KEEP_TRACE=1.
• trace_count: rows of several tables of a kept .trace, in all and per process; by default the tables that
  matter for a translation layer. Measured on drawstress d3d12 under Mullion (M1 Pro, 30 s): 2559
  graphics-compiler-spill-events and 115 compiled shaders for the Wine process, 0 rows in the GPU counter and
  shader profiler tables.
• perf_baseline_record: records the same program count times and saves the runs as a reference in one call.
• xctrace 27.1 writes metal-resource-allocations and time-profile whole and then dies with SIGTRAP (exit 133):
  an export cut at a row boundary is read to its last row instead of failing.

v0.6.0

• perf_baseline and perf_check: a regression reference. perf_baseline saves copies of several equal recordings
  (ideally 3 or more) as a named reference in baselines.json beside the runs, so it outlives KEEP_RUNS, and
  returns the range each number spans across them. perf_check judges a recording against that range: regressed
  when FPS falls under it by more than margin_pct (default 3), GPU ms per frame rises over it, or encoders or
  command buffers per frame exceed it; improved when a number leaves the range the other way; same otherwise.
  The range, not one run, is the reference because 5 equal runs of drawstress d3d12 on an M1 Pro differed by 4.4 %
  in FPS and 6.1 % in GPU ms per frame. Frame pacing, short encoders and GPU idle time are reported but do not
  decide: they varied 13 % and 7 points between equal runs.
• perf_compare says what changed in the GPU work and where: encoders and command buffers per frame, the share of
  encoders under short_us, the GPU's idle share and idle ms per frame, and labels, each encoder label's count and
  GPU ms per frame in both runs, largest change in ms per frame first. Labels compare per frame because two
  recordings of different lengths do not compare by total time; a label missing from a run that holds the full 20
  labels is outside_top_20, not gone. A run summarized before 0.4.0 is named in encoders.missing_in, and the
  headline numbers still compare. Shares (short encoders, GPU idle) give their change in points, without a %.
• The compare-layer prompt reports perf_compare's table and labels instead of rebuilding them from two
  perf_diagnose calls.
• Each recording names its translation layer in layer: the Wine version, the overrides in force (environment and
  the prefix's registry), the size and SHA-256 of d3d12, d3d12core, d3d11, d3d10core and dxgi in Wine and
  in the prefix, a 12-character fingerprint of all of it, and the program's variables (MULLION_DIAG=profile) with
  paths cut to file names. Native runs say they have no layer. perf_runs shows the fingerprint and Wine;
  perf_compare says whether both runs share a layer, which DLLs differ, and that different layers compare layers,
  not one change. Which copy of a DLL Wine loads is not recorded: Instruments does not keep Wine's own log.
• perf_record.sh keeps the caller's WINEDLLOVERRIDES (a Mullion built for system32 needs d3d12,...=n) and
  adds mscoree=;mshtml= after it; before, it replaced them.
• A prefix copy of a Direct3D DLL identical to Wine's own (Wine's first boot makes one) leaves the layer
  fingerprint unchanged: the fingerprint and differs look at the copy Wine loads, the prefix's when it exists.
  Measured on an M1 Pro: the first drawstress run on a new prefix got a different fingerprint from the next
  three, with identical DLLs.
• Downloads are hashed in 1 MiB blocks.

v0.5.0

• workload_list and workload_fetch: download a test program once from this project's releases, check the
  SHA-256 this package pins and keep it in ~/Library/Caches/metal-gpu-trace. Claude Code asks before every
  download (anthropic/requiresUserInteraction).
• perf_record records a fetched test program with workload, api and dim.
• First test program: bgfx 17-drawstress for macOS arm64 and Windows x64, built from source by
  .github/workflows/workloads.yml (release workloads-1). Recorded on an M1 Pro, it gives the same encoders per
  frame as a local build (2 in native Metal, 7 in D3D12 under Mullion) and FPS within 4 %.
• compare-layer prompt: records a test program in native Metal and under a translation layer, then diagnoses and
  compares both runs.

v0.4.0

• perf_diagnose: says whether a recording splits its GPU work into too many short encoders, as translation layers
  do, with encoders and command buffers per frame, the labels with the most GPU time and the GPU's idle time.
  Defaults (4 encoders per frame, 25 µs, 25 % short) come from bgfx 17-drawstress on an M1 Pro: 7 encoders per frame
  and 50 % under 25 µs under Mullion, 2 and 2 % in native Metal.
• Summaries gain encoders and gpu_idle, from metal-gpu-intervals and metal-application-encoders-list.
  Summaries written by older versions have neither: record again to diagnose them.
• Every tool parameter has a description; metal-gpu-trace --version.
• Registry files: server.json (MCP Registry), glama.json; CHANGELOG.md.
• trace_tables and trace_query take a run name; perf_runs says whether each run kept its .trace
  (trace_kept).
• A trace without metal-gpu-intervals fails with a hint; summaries list other missing tables in missing_tables.
• capture_summary returns calls_total.
• Tools carry title, readOnlyHint, destructiveHint, idempotentHint and openWorldHint; initialize
  returns usage instructions.
• A recording's deadline grows with its length (10 s per recorded second plus 120 s), and one that overruns leaves
  no files behind.
• perf-runs/ is created only when a recording needs it. Tests run under pytest; CI on macOS with Python 3.9 and
  3.13.

Workloads workloads-1

Test programs for metal-gpu-trace's workload_fetch. Built from source by the workloads workflow; BUILD.txt lists the commits.

v0.2.0

• First release on GitHub: perf_record, perf_attach, perf_runs, perf_compare, trace_tables,
  trace_query, capture_summary, capture_records, gpu_family, and the GPU family resources.