GPU-PCIe-Test v3.4.0

Fixed

• Fence timeout was treated as a successful batch (all variants) - a single
  8 s fence timeout returned from WaitForBenchFenceEx while the GPU work was
  still in flight. The CPU-timed bandwidth loops recorded the timeout as a real
  sample, and the next batch reset the command buffer / allocator underneath a
  live submission and re-submitted a fence still in use. Timeouts are now
  retried on the same submission (up to MAX_FENCE_RETRIES, ~24 s of grace
  for a slow GPU) and any non-Success result aborts the benchmark. The
  dual-queue bidirectional loop no longer resets fences that never signaled,
  and the latency tests no longer block indefinitely in
  vkGetQueryPoolResults after a timeout.
• Rated RAM latency was 1000x too high whenever the memory speed was not an
  exact table entry (DDR5-6200 showed ~12900 ns instead of 12.9 ns).
• Bidirectional result was mislabeled after the v3.0.7 halved-buffer
  fallback - the results table showed the requested size, not the size that
  was actually measured.
• Benchmark summary window could walk results while the worker inserted
  into it - results are now published before the window is told to open,
  and the window holds the results mutex for its whole body.
• Vulkan memory-latency test crashed instead of skipping on allocation
  failure - buffer, memory-type, allocation, bind and map results are now
  checked (matters on eGPUs with tight OS memory budgets).
• Render-finished semaphores are now per swapchain image (Vulkan
  variants) - a per-frame semaphore could be re-signaled while an earlier
  present still referenced it (VUID-vkQueueSubmit-pSignalSemaphores-00067).
• Linux validation builds did nothing - -DENABLE_VULKAN_VALIDATION=ON
  defined the macro but the Linux source never enabled
  VK_LAYER_KHRONOS_validation; it now mirrors the Windows Vulkan variant.
• Soldered LPDDR reported as single-channel RAM (Strix Halo and similar
  APUs) - channels were guessed from the SMBIOS device count, so firmware
  that describes a 256-bit LPDDR5X bus as one device produced a
  "single-channel" warning and a theoretical bandwidth 4x too low, which
  inflated the iGPU percent-of-RAM figures. Soldered memory now uses the
  SMBIOS per-device data widths when present, is never warned about, and
  the iGPU comparison raises the channel estimate when the measured
  bandwidth proves it too low (logged as inferred).
• Linux build failed with CMake < 3.24 - DOWNLOAD_EXTRACT_TIMESTAMP
  is now only passed to CMake versions that know it (the stated minimum
  is 3.16).
• Linux build broke with GCC 16 - a dead successfulRuns counter tripped
  -Werror=unused-but-set-variable.
• build_gui.bat was stored LF despite the v3.0.7 CRLF fix (the commit
  was normalized by core.autocrlf=input). .gitattributes now pins *.bat
  to CRLF on checkout so it cannot regress.

Added

• GitHub Actions release workflow: the Windows D3D12 and Vulkan executables
  and a Linux AppImage are built and attached to every v* tag.
• packaging/arch/PKGBUILD for Arch Linux / CachyOS (makepkg -si), plus a
  desktop entry and icon installed by CMake (used by the AppImage too).

Changed

• Version scheme unified: tag, changelog and in-app version now all read
  3.4.0 (earlier commits used 3.0.x while tags were 3.x.0).

Downloads

• Windows (Direct3D 12): GPU-PCIe-Test_GUI.exe
• Windows (Vulkan): GPU-PCIe-Test-Vulkan.exe
• Linux (any distro): GPU-PCIe-Test-3.4.0-x86_64.AppImage - chmod +x and run (needs your distro Vulkan loader + GPU driver)
• Arch Linux / CachyOS: gpu-pcie-test-3.4.0-1-x86_64.pkg.tar.zst - sudo pacman -U

Windows executables and the AppImage are produced by the release workflow from this tag; the Arch package is built natively on CachyOS from packaging/arch/PKGBUILD.