[bug] agc_driver_dcc_keys_first_write fails on main on a real NVIDIA GPU #2188

Activity

1. Dextroid17 commented on Oct 9, 2026

   

   Dextroid17

   on Oct 9, 2026

   ContributorAuthor

   More actions

   Confirmed device-specific, as the body suspected. On lavapipe — the same software device the linux CI job installs, Mesa 26.2.4 / LLVM 23.1.1 / 256 bits — the test passes; on the RTX 3080 Ti it fails with the same message. So the lavapipe job cannot see this.

   To re-check on a machine with both: install or extract vulkan-swrast, then

   VK_ICD_FILENAMES=<…>/lvp_icd.json \
   LD_LIBRARY_PATH=<…>/usr/lib \
   ANYPS5_REQUIRE_VULKAN=1 GALLIUM_OVERRIDE_CPU_CAPS=avx \
   ctest --test-dir build -R ^agc_driver_dcc_keys_first_write$ --output-on-failure
   

   That reproduces the CI device on a workstation without touching the NVIDIA install.

   Written with AI assistance.
2. Dextroid17 commented on Oct 9, 2026

   

   Dextroid17

   on Oct 9, 2026

   ContributorAuthor

   More actions

   Closing this myself, because by this project's own ground rules it should not have been filed: the failure reproduces only on the NVIDIA proprietary driver, that path is not expected to work on Linux yet, and the same test passes on lavapipe, which is the device CI actually runs. My framing above — a CI blind spot — was wrong; this is an unsupported device, not a defect.

   Leaving it visible rather than deleting it, since the lavapipe half is the useful part: on the CI device the test passes, so nothing is hidden from CI.

   Written with AI assistance.
3. Dextroid17

   closed this as completedon Oct 9, 2026
4. SleglJan commented on Oct 9, 2026

   

   SleglJan

   on Oct 9, 2026

   Contributor

   More actions

   The bisect the report left open. git bisect run between f145139 (the base of #1979, assumed good; the assertion does not exist there) and d70b899, 95 commits, 9 steps, 7 of them built and tested once each on an RTX 3080 (615.71.09, Linux, CachyOS kernel 7.2.9, GCC 16.2.1, Release) and 2 skipped because they do not build (6cc0503, 0ab2a04): the first bad commit is 0e1d25a ("fix(agc): avoid redundant transfers for untracked imported textures", merged through #2096), the commit that added the assertion an unchanged pending image was uploaded again to DccKeysFirstWrite.cpp. Tested directly with the same build and GPU, its parent a8b032b passes 3 of 3 and 0e1d25a fails. At d70b899 the test fails 3 of 3 runs on the RTX 3080 and passes 3 of 3 on RADV (Mesa 26.2.4, Ryzen 7950X3D iGPU) on the same machine. On Windows, popcoder565's run on #2096 (RTX 3060 Laptop, 616.92, head 3ed0d03) has agc_driver_dcc_keys_first_write passing 30 of 30, and #2155 attributes the storageRefreshTests failure of agc_driver_recorder_tests on that machine to the same commit.

   AI-assisted: yes (Claude Code).
5. fma-rc commented on Oct 10, 2026

   

   fma-rc

   on Oct 10, 2026

   Last edited by fma-rc

   Contributor

   More actions

   This reproduces on main 0c0dc1e with an RTX 5080 (615.71.09, Linux 7.2.9, GCC 16.2.1, Release), and the trigger is not the NVIDIA driver as such: it is host imports staying write-watched.

   Device	Setting	main	main + patch below
   RTX 5080	default	fails, 3 of 3	passes, 3 of 3
   RTX 5080	APS5_WRITE_WATCH_IMPORTS=watch	fails, 3 of 3	passes, 3 of 3
   RTX 5080	APS5_WRITE_WATCH_IMPORTS=unwatch	passes, 3 of 3	passes, 3 of 3
   RADV (Radeon 610M, Mesa 26.1.4)	default	passes, 3 of 3	passes, 3 of 3
   RADV	APS5_WRITE_WATCH_IMPORTS=watch	fails, 3 of 3, same message	fails, 3 of 3, at "the first write: the write-back stored 0 written and 0 cleared bytes"

   The probe reports 0 of 4 scratch pages written after a GPU read ... imported ranges stay watched on NVIDIA and 4 of 4 ... imported ranges are compared on RADV.

   In the code: GuestMemory::ImportWatched stamps the blocks with StampKind::ImportWindow, which ChangedBlocks reports as BlockMaybeWritten. StorageTexture::Refresh treats any non-zero stamp as changed (Texture.cpp L1289), a changed pending unit blocks the equalsOriginal() rescue (L1326), and it is not droppable because it is not BlockWritten (L1388). So the unit is written back and the image uploaded again. The unwatched path instead compares against original (L1271).

   Candidate (7 lines, in scanBlocks): for a tracked block unit stamped BlockMaybeWritten, compare it against original with compareUntracked and mark it unchanged when equal.

   With it, on the RTX 5080 the full ctest goes from 564 of 565 to 565 of 565; on RADV ctest -R ^agc_ is 266 of 266 before and after. tools/check_conventions.py passes.

   Retested on main ee50132: same results in the table; full ctest on the RTX 5080 583 of 584 → 584 of 584, RADV ctest -R ^agc_ 269 of 269 before and after.

   Not established: any effect on a title; the non-block-unit branch (L1298–1305); Windows. The RADV forced-watch row still fails with the patch, at a later assertion. compareUntracked calls FlushGpuWrites, which this adds to the tracked path.

   Patch against main 0c0dc1e

   diff --git a/core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp b/core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
   index 26911542..4b626008 100644
   --- a/core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
   +++ b/core/libs/prx/libSceAgcDriver/Graphics/src/Texture.cpp
   @@ -1270,6 +1270,13 @@ bool StorageTexture::Refresh() {
                tracked = GuestMemory::ChangedBlocks(descriptor.baseAddress, static_cast<std::size_t>(guestBytes), generations, stampedBlocks, cpuBlocks);
                compared = !tracked && compareUntracked(descriptor.baseAddress, static_cast<std::size_t>(guestBytes), stampedBlocks, true);
                if (compared) cpuBlocks = stampedBlocks;
   +            for (std::uint32_t unit = 0; tracked && blockUnits && unit < trackedLayers; ++unit) {
   +                if (stampedBlocks[unit] != GuestMemory::BlockMaybeWritten) continue;
   +                std::array<std::uint8_t, 1> block{};
   +                if (!compareUntracked(layerBegin(unit), static_cast<std::size_t>(layerBytes(unit)), block) || block[0] != GuestMemory::BlockUnchanged) continue;
   +                stampedBlocks[unit] = GuestMemory::BlockUnchanged;
   +                cpuBlocks[unit] = 0;
   +            }
            };
            const auto keysStart = profile ? std::chrono::steady_clock::now() : std::chrono::steady_clock::time_point{};
            keys = ProvedKeys();

   AI-assisted: yes.