[Trycky64]

Update

Rebased the buffer/image copy span fix onto current main (ac3bd4400e2e4eb82aaa18fd8207df642a37a6ed). The production change remains limited to gpu/vk/vk_xfer.cpp: the bound span includes row, layer, and depth-slice padding between accessed data, but stops after the final texel block Vulkan reads or writes. The #1082 and #1084 changes and tests remain in the base.

Following review feedback, the opt-in GPU fixture now selects a compatible Vulkan physical device by queue, format, and memory capabilities instead of requiring an AMD RX 9070. --gpu-device NAME remains available for diagnostics. A new runner-discovered compile-only check builds the GPU fixture when a compiler and Vulkan headers are available; it does not load Vulkan or run GPU work.

Validation

• Windows targeted runner: 5 passed, 0 failed, 0 skipped (copy-span, GPU-fixture compile, partial-clear mip, and program-cache index/reader checks).
• Linux targeted runner: 5 passed, 0 failed, 0 skipped for the same checks.
• RX 9070 Vulkan probe after rebase: passed on Vulkan 1.4.359, driver 0x8001a0; four texels matched and padding remained untouched. Khronos validation layers were unavailable. An explicit RX 9070 selector also passed; a non-matching selector returned SKIP cleanly.
• Full Windows Vulkan backend build: hookshot_vk.dll linked successfully from an isolated temporary build directory; 185 compile jobs, 0 failed. No game files or output directory were used.
• Full Windows runner on current main: 156 checks, 128 passed, 5 failed, 23 skipped. On this branch: 156 checks, 126 passed, the same 5 failed, and 25 skipped. The failures were nx2/test_nup_read_name.py, nx2/test_nup_paths.py, tests/test_fps_overlay_order.py, tests/test_fpx_sticky.py, and tests/test_precompile_compute_guard.py; no additional failing check appeared on this branch. nx2/test_nup_paths.py also causes the runner's final Windows summary to raise UnicodeEncodeError while printing a replacement character, on both base and branch.

No in-game test was performed, and the real-device run covers only the RX 9070. This change does not claim to fix issues #696, #723, or #708.

This resubmission updates the existing PR #1083 after the reviewer requested portable GPU-device selection and automatic fixture compilation. It remains independent of the partial-clear mip fix.

[hckrman101]

Went through the span maths and it checks out. With bufferRowLength = (row/bpb)bw and bufferImageHeight = hbbh, Vulkan's slice stride is hbrow and layers are addressed like depth slices, so the last block touched ends at row(hblayersd - 1) + wb*bpb. hb/layers/d are all rejected when zero in resolve(), so the -1 can't underflow, and the uint64 cast comes before the multiply. The base address doesn't change, so bufferOffset alignment is the same as before. Each call is a single region.

The smaller size also flows into memory_stamp_gpu / texture_gpu_wrote on the image->buffer path, which is what you want: write-back no longer covers the trailing stride padding past the last row. The CPU fallback (texture_read_texels) isn't affected.

The fixture is registered in run_all.py, uses the existing {vulkan_headers} placeholder, and builds from the production functions, so it can't drift from vk_xfer.cpp. The 8 cases cover the padded/unpadded, array, 3D, BC edge, truncated pool and bad stride paths, both directions.

Two small things, neither blocking:

• vulkan_copy_span_gpu.cpp only runs with --gpu, so CI never compiles it and it can bitrot without anyone noticing. It also only accepts an RX 9070 (see inline).
• #1084 also touches vk_xfer.cpp and run_all.py. Different hunks, but whichever lands second may need a small rebase on run_all.py.

[hckrman101]

nit: as written, this fails with CHECK(physical) on any machine without an RX 9070. Nothing here looks AMD-specific, so maybe take the first device that has a graphics queue (with an optional name filter), or print a skip message when there's no match, so other people can run it with --gpu.