Repository navigation
fix(agc): preserve infinite depth bounds - #7627
Conversation
|
Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.
|
|
On Windows the new Its registration has no AI-assisted |
|
Windows issue at head The test logic is fine: running Suggested fix: add Windows 11, MinGW (WinLibs GCC 15.2.0), Release. |
|
Fixed the registration in Thanks LeandroLP23 and akkadbakkad89 for isolating this. The graphics target rebuild and both focused CTests pass on Linux (2/2, no skips, 2.99 s); conventions and whitespace pass. I checked the generated Linux CTest properties, but cannot claim a corrected Windows run locally. Please rerun the focused CTest on Windows when convenient. The description records the original Windows failure and the remaining rerun requirement. |
|
Cross-linking the systematic fix that grew out of this report: #7645 (tracks issue #7637). The narrow registration fix you pushed covers this test; #7645 additionally stages the |
…on Windows A test registration that omits the WIN32 PATH prepend for libs/unpatched makes the executable block on the Windows entry-point dialog (0xC0000139) until ctest reports a timeout; the dialog is invisible to headless and CI runs. The prepend is per-test boilerplate repeated across ~95 CMakeLists files, so new registrations keep missing it (boykopovar#7627). add_test_executable now copies $<TARGET_RUNTIME_DLLS> next to the exe at link time (copy_if_different rather than copy -t, whose -t option older cmake -E lacks; the exe itself is listed as a source so a target with no shared-library dependencies still receives a valid command and copies nothing). The loader resolves the .prx imports from the executable directory ahead of PATH, so no registration needs the prepend and a test exe can be run by hand with no environment setup. Fixes boykopovar#7637. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
github-actions
Bot
added
the
unblocks title
…on Windows A test registration that omits the WIN32 PATH prepend for libs/unpatched makes the executable block on the Windows entry-point dialog (0xC0000139) until ctest reports a timeout; the dialog is invisible to headless and CI runs. The prepend is per-test boilerplate repeated across ~95 CMakeLists files, so new registrations keep missing it (boykopovar#7627). add_test_executable now copies $<TARGET_RUNTIME_DLLS> next to the exe at link time (copy_if_different rather than copy -t, whose -t option older cmake -E lacks; the exe itself is listed as a source so a target with no shared-library dependencies still receives a valid command and copies nothing). The loader resolves the .prx imports from the executable directory ahead of PATH, so no registration needs the prepend and a test exe can be run by hand with no environment setup. Fixes boykopovar#7637. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
|
Intel and AMD Windows data: with this PR the 19 tests built from Windows 11 Pro 26300, WinLibs GCC 15.2.0 r7, Release, every execution test built,
Under CTest I also ran full CTest with this PR and the 14 other open AGC PRs I tested in the same batch merged together (#7647, #7641, #7630, #7603, #7600, #7591, #7579, #7535, #7533, #7474, #7462, #7050, #7048, #7650). Neither GPU showed a failure beyond AI-assisted: yes (Claude Code) |
33f65a7 to
8bd880c
Compare
|
Checked the decoder change against current main (NaN still rejected, infinities preserved only for |
What
Preserve IEEE infinities when decoding
DB_DEPTH_BOUNDS_MIN/MAX. Crimson Spires PPSA03568 writes positive infinity (0x7f800000) to DWORD 9; main aborts withAGC graphics: non-finite register at DWORD 9before its main menu.Use a dedicated depth-bound decoder that rejects NaNs while preserving finite values and infinities. Other floating-point register checks stay unchanged.
ValidateDepthBoundsstill refuses bounds outside [0,1] unlessVK_EXT_depth_range_unrestrictedis enabled; no clamping, title-specific path, extension bypass or fake success is added. The Vulkan extension removes the depth-bound range restriction.Decoder commit
964dcbef47cfd38c89523eb6dd202c0d3e55c199plus test-registration follow-up33f65a7eba, on mainc34e9d855cb496ef1525b3ccbddebdba10fe6541.Tested
2026-10-19 current-main composition check: on pinned main
9d2d40f5eedfa8c5f4312517175147eb0181001a, #7817 alone removes the earlier full-optimizer BDA ABI failure but the native title then aborts atnon-finite register at DWORD 9. Adding only this PR produces0f6f2897cc2d2ae81cd30cefc5ad7b3134a53f71: libraries build and depth-bounds state, unused-viewport depth and graphics checks pass 3/3, with required Vulkan, shader cache disabled and core validation (zero VUIDs or missing-layer messages). A warm native PPSA03568 run reaches driving dialogue and closes normally after 249.10 s, exit 0, PID gone and settings restored. The driving scene is still too dark. Adding the published #6959 stack (including its #6934 prerequisite) separately restores visible scene lighting. These are explicitly composed results, not a new isolated bring-up or speed claim. Sanitized run summary · Focused checks · Baseline driving scene · Layered-rendering driving sceneLatest rebase: two commits, no merge commits, on main
3fdd1f52e968c57d2b6890b11863f96aee0c11ce; head8bd880ceb1b358e6e58bb5e9337eff01e34388a3. Resolved the additive CMake conflict while retaining upstream’s unused-viewport-depth test and this PR’s depth-bounds runtime properties, including Windows PATH setup.agc_driver_graphics_testsand its runtime dependencies build. Depth-bounds state, unused viewport depth and graphics checks pass 3/3, no skips, 3.29 seconds on the Linux configuration below. Conventions and whitespace checks pass. This refresh adds no new native-title or performance claim.Isolated current-main follow-up on main
676d423968a1a8fcc79c6247020ce22090a8e0d4: only this PR's decoder and CTest-registration changes are applied, producing5aede09b0711f27a449e77cd9bfed444417f5a58. Full libraries and graphics tests build;agc_driver_depth_bounds_stateandagc_driver_graphicspass 2/2, 2.65 s. With the same dump option (APS5_PLAYGO_UNLISTED_CHUNKS=installed), unchanged main reaches 574 frames and aborts atAGC graphics: non-finite register at DWORD 9; main plus this PR reaches driving dialogue and 1,648 completed frames, then closes normally after 132.32 s (exit 0). Both processes are gone and the udmabuf limit is restored to 64 MiB. The driving screenshot is visibly too dark, also confirmed by the user; this establishes startup progress, not visual correctness or a speed gain. Isolated-run summary, tests and dark-scene screenshot, assets prefixedcrimson-depthbounds-sync76-.Current-main integration check on
a6dd29b3daf3ba25bf4c2e138ad380b887ad4d49: exact #7681 plus this PR and #7630, integration headce17127384d4436a092b623236945d9fe3385e81. Full libraries and focused test executables build; initialization, commands, guest shader linking, depth-bounds state, graphics and three control-flow modes pass 8/8 without skips (6.17 seconds, RADV). A 228.91-second native run reaches the Crimson Spires menu, records 1,451 completed frames and closes normally (exit 0); PID gone and udmabuf limit restored to 64 MiB. Menu screenshots still show an excessively dark 3D scene with bright UI. Vulkan validation reports image-view usage, blend attachment, index/vertex-buffer and barrier-mask errors; this is startup/integration evidence, not visual correctness, full gameplay or performance evidence. The current-main-only run hits the independent versioned-register-defaults sentinel regression before this depth-bounds path; current main plus only #7681 reaches 514 frames then reproducesnon-finite register at DWORD 9.Arch Linux 7.2.9-arch1-1, Ryzen 9 7950X, GCC Release, bundled source-built dependencies, RX 6650 XT with RADV Mesa 26.2.4.
libsandagc_driver_graphics_testsbuild successfully.agc_driver_depth_bounds_stateandagc_driver_graphics: 2/2 pass, no skips. These are CPU/mock graphics checks. Regression coverage includes finite bounds, each infinite endpoint, both infinities, NaN rejection at both endpoints, preservation of infinity rejection for other registers, and rejection without / acceptance with the unrestricted-depth capability.DB_DEPTH_BOUNDS_MAX. The diagnostic-only native build reports0x7f800000; that diagnostic is excluded from the PR.c777a0f1941con main218aa8f00babrenders the menu at 30 and 130 seconds, closes normally after 148.49 seconds (exit 0), and restores its temporary kernel setting. Main218aa8f00bababorts at the bound read; the intervening main merge feat(agc): link calls through the AGC FS table #6160 does not change State.cpp. The final rebased head964dcbef47cfalso passes the original bound-read failure. Its 30-second snapshot still shows the publisher logo; Cross was sent at 46.72 seconds, then it aborts at 48.72 seconds with a separateunsupported dynamic s_setpc_b64 at program counter 0x000003e0(shader stage 5). This final run does not independently establish a menu screenshot or gameplay. Its process exited and the temporary kernel setting was restored.APS5_PLAYGO_UNLISTED_CHUNKS=installedoption. This is an installation assumption, not proof of complete chunk metadata.33f65a7ebaadds this test to its siblings' shared shader-cache setting, 30-second timeout, source working directory and Windows runtime-library PATH. Local rebuild and both focused CTests pass (2/2, no skips, 2.99 s); generated CTest properties checked. Independent Windows validation of the corrected registration is recorded below.Evidence: test logs, negative control, native summary and pre-rebase menu screenshot, assets prefixed
crimson-depthbounds-.savvaniss independently reports all 19 Graphics.cpp tests passing three runs each, including
agc_driver_depth_bounds_state, on Windows 11 Pro 26300, Intel UHD 770 (32.0.101.7088) and AMD RX 9070 XT (32.0.31041.1004), WinLibs GCC 15.2.0 r7 Release. The tested composition merges exact head33f65a7ebinto mainf3059421f, with shader caching disabled and serial CTest. The control without this implementation failsagc_driver_graphicsat non-finite DWORD 9 andagc_driver_covered_centroid_stateon both GPUs. The corrected depth-bounds CTest registration passes on Windows. These decoder/mock tests do not establish GPU rendering or title performance.Checklist
main; no other open PR implements the same functionsNotImplemented_nid_no_patch); silent stubs are listed under Silent stubs in their library's technical debt file.mdfiles or images added to the repository (attach them to this PR)