Commit Graph
11 Commits
Author SHA1 Message Date
Ahmad Sayed Alsayed EbrahimandClaude Opus 5.5 a4bd2eb271 perf(vkbridge): async requests with ordering, batched descriptor sets, buffered server reads
- Ordered async protocol: vkb_msg_hdr seq/barrier, per-process ordering gate on
  the server; void commands and VkResult commands without outputs (submits,
  resets, binds, vkEndCommandBuffer replay) no longer wait for a reply.
  VKBRIDGE_SYNC=1 restores synchronous calls for debugging.
- Client async queue batched across threads with a ~100us background flusher.
- Descriptor sets allocated in batches per (pool, layout); the client keeps the
  pool's books (layout descriptor counts vs pool sizes) so batches ask only for
  what fits and a full pool answers OUT_OF_POOL_MEMORY without a round trip.
  Caches for a destroyed layout are dropped (its handle may be reused).
- Server parses requests out of one receive buffer (descriptor FIFO for
  SCM_RIGHTS) instead of two reads per message.
- Server logs async VkResult failures (generated vkb_cmd_result table) and
  per-command timing with VKBRIDGE_STATS=<s>.
- tests/d3d11_bench.c: game-like D3D11 frame (1000 draws, Map DISCARD, rebinds).

d3d11_bench 1000 draws on RADV: native 0.30 ms/frame; bridge 26.3 ms (sync)
-> 1.43 ms. Host suite 14/14 in plain and Mali-imitation modes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 c4ac049007 feat(vkbridge): VK_EXT_depth_clip_enable emulation via depth clamp; D3D11 DepthClipEnable check
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 02a06bec3f feat(vkbridge): vertex attribute divisor emulation; host regression suite; dispatch/version fixes
- divisor != 1 bindings get stride 0 and are rebound per draw; d > 1 splits draws, with a
  BaseInstance-as-0 vertex shader variant so DXVK's SV_InstanceID stays exact
- device-level dispatch never falls back to a loader trampoline (crashed in vkCmdSetLineStipple)
- core commands newer than the device's version are not handed out; the app's instance
  version is passed through as asked (Zink 26.2 aborted on a 1.4 entry point)
- run_tests.sh: smoke, bcn_check, DXVK d3d11_check, vkcube x3, gamescope+Zink, in plain and
  Mali-imitation modes - 14/14 pass

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 92c94c5c15 feat(vkbridge): BC1-BC7 texture emulation (GPU decode on upload)
BC images are created decoded (RGBA8/R8/RG8/RGBA16F, keeping sRGB/SNORM meaning) and every
buffer->image upload is a compute dispatch (shaders/bc_decode.comp, a GLSL port of bcdec with
BC6H mode layouts generated from its source). The app's compute bindings are restored after.
bcn_check: all 10 formats incl. every BC6H/BC7 mode bit-exact with bcdec; DXVK accepts the
Mali-like device (BC, clip/cull, maintenance5 emulated) and its BC/clip checks pass.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 2e569b850a feat(vkbridge): SPIR-V editor; clip/cull distance and GS/TES point size emulation
The last pre-raster stage's ClipDistance becomes a varying and the fragment shader discards on a
negative distance; cull distance and GS/TES PointSize become unread varyings. Rewritten modules
pass spirv-val; DXVK's SV_ClipDistance check passes with the GPU's clip distance hidden.
The client reports features by what the server emulates and what the GPU lacks.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 4a96c62f93 test(vkbridge): D3D11 checker (BCn, clip distance, step rate) under Proton's DXVK; gamescope/zink host scripts
DXVK (Proton Experimental 11) passes 9/9 through the bridge on RADV.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 bf6dbb642a feat(vkbridge): emulation layer + VK_KHR_maintenance5 for drivers without it
Zink (Mesa 26.2) and DXVK 2.7 refuse a Vulkan 1.3 device without maintenance5. The server
installs dispatch-table wrappers: flags2 folded into 32-bit flags, inline shader modules,
BindIndexBuffer2/area granularity/subresource layout 2, VK_REMAINING_ARRAY_LAYERS and
VK_WHOLE_SIZE resolved from tracked images/buffers. Host: Zink glxgears in gamescope at
144 fps with the driver's maintenance5 hidden and the API capped at 1.3.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 c3f02222fc feat(vkbridge): Vulkan 1.4.341 registry, stand-in DRM node for gamescope, test switches
- registry moved from 1.3.247 to 1.4.341: current Zink and DXVK need post-1.3.247
  extensions (maintenance5); the bridge only exposes what the driver reports
- gamescope on the bridge: Mali's device stands in as the render node, the ICD reports its
  numbers (VKBRIDGE_DRM_RENDER), the session shim answers PRIME on it - linux-dmabuf works
- host: gamescope (Arch, no GPU device) composites vkcube through the bridge, zero-copy;
  Zink 26.2 runs GL 4.6 on the bridge
- VKBRIDGE_TRACE / HIDE_EXTS / MAX_API for diagnosis and imitating Mali on the host

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 4050270511 feat(vkbridge): Phase 1 integration - Setup toggle, server process, session wiring, builds
- Setup: 'Mali bridge' switch on the GPU row (non-Adreno only); the non-Adreno gate relaxes when on
- VkBridgeComponent starts libvkbridge_server.so before the guest and waits for its socket
- session script points the loader at the bridge ICD and logs the self-test + vulkaninfo
- Flatpak sandboxes get the bridge ICD and socket; the compositor uses the system driver without KGSL
- Gradle builds the server with the NDK; CI/build_local cross-build the ICD + smoke test

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:28:13 +03:00
ahmadandClaude Opus 5.5 d82c42d818 feat(vkbridge): self-checking smoke test, frame dump, WSI extensions and crash log on the server
Smoke test passes 9/9 through the bridge on RADV (dmabuf) and lavapipe (hostptr, UMA model);
vkcube renders correctly over Wayland dma-buf, wl_shm and X11.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:27:54 +03:00
ahmadandClaude Opus 5.5 a8fd343fc1 feat(vkbridge): Vulkan bridge core - codegen, wire runtime, server, client ICD, WSI
Generated from vk.xml 1.3.247; server shares host-visible memory via a
self-tested strategy (hostptr/dmabuf/ahb). vulkaninfo runs through it on host.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KZqR8XUE8czvGzEpgSCVQR
2026-10-10 14:27:54 +03:00