#1594
Tests: synchronize Vulkan occlusion results before CPU readback
Claimed in #1548. Found during the investigation of #1540.
What and why
Both occlusion tests read their query results from a mapped buffer, but neither marked that buffer host_readable. That skipped the replay queue’s transfer-write → host-read barrier. A queue-idle wait completes execution; Vulkan’s fence synchronization note also calls for a memory dependency to make writes visible to the host.
Set host_readable in both test bindings so they use the runtime’s existing readback path.
Tested
vk_overlapping_occlusionandvk_binary_occlusionboth passed on Linux llvmpipe with Mesa 25.0.7 and 26.1.6, with synchronization validation enabled.- A local
vkCmdPipelineBarrier2trace on Mesa 26.1.6 measured 0 before / 1 after transfer-to-host-read barriers in the first overlapping-occlusion submission. All pixel counts and sentinels agreed in both runs. - The reported Windows Radeon timeout did not reproduce on llvmpipe. Its cause remains unconfirmed.
git diff --checkpassed.
No game data. AI-assisted implementation and testing.