memory-stats reads /proc/self/smaps by default on Linux: every sample formats and parses one
block per mapping while holding the mmap lock. The benches' RssSampler polls it every 2 ms, so
on a profile of bevel_bench --size 4000x3000 the rss-sampler thread was ~22% of all samples
(smaps_account 14.7%, smaps_pte_entry 6.3%), contending with the measured threads.
Enable the crate's always_use_statm feature: one short read of the kernel's RSS counters, the
same number VmRSS reports. The measured code is unchanged; only the harness overhead goes.
On Windows, GPU UI test binaries crashed with STATUS_ACCESS_VIOLATION
whenever two of their tests ran at once: arbitrary_rotation_preview 3
runs in 4 on main, and color_managed_canvas, crop_default_mode,
drag_preview_canvas and full_res_preview at random. egui_kittest
prefers a CPU adapter, which on Windows is WARP over DX12, and a dump
puts the fault inside d3d10warp.dll with two WARP devices rendering on
parallel test threads. Bare egui harnesses with no PhotoCraft code
crash the same way, and Vulkan or GL never does. CI's Windows runners
have no GPU, so they run on WARP too.
Three binaries already held a per-binary lock for the same class of
driver bug (Mesa llvmpipe and RADV, #194; libvulkan with NVIDIA and
lavapipe). That lock now lives once in photocraft-testkit
(`gpu_lock`, with the reasons) and every test in the twelve GPU test
binaries without it takes it on its first line. Under nextest each
test is its own process and the lock is free.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
xsvm