[Kim2091]

Bridge: cut per-frame IPC cost of the D3D9 bridge

Most of a bridged frame went to IPC overhead:

• The once-per-frame event query GetData was a blocking round trip.
  It drained the whole command queue and forced lockstep between client
  and server.
• Each of the ~28,000 commands per frame in HL2 RTX paid the full
  per-command cost.

Measured in HL2 RTX:

• Frame time: ~14-15 ms -> ~8-8.6 ms
• Commands per frame: ~28,000 -> ~4,000

Changes:

• Non-blocking queries (client.asyncQueries, default True)

  • Event queries are answered on the client from the UID of the last
    Issue the server executed, published in shared memory.
  • Other small query types send GetData without waiting. The server
    writes the result to a shared per-query slot.
  • D3DQUERYTYPE_RESOURCEMANAGER and slot exhaustion keep the blocking
    path.
  • Fixes: the blocking path reads the data blob even when pData is
    null, and the server no longer crashes on queries whose create
    failed.

• Redundant state filtering (client.dropRedundantState, default True)

  • The client tracks the state it sent and skips setters whose value
    the server already holds (21 setters).
  • Shader constants are narrowed to the registers that changed.
  • Tracking is reset on state block Apply, Reset and
    SetRenderTarget(0).
  • Replaces eliminateRedundantSetterCalls, which compared against
    game-visible state.

• State packets (client.batchState, default True)

  • Consecutive setters travel as records of one Bridge_StatePacket.
  • The packet closes before any other command, so ordering is
    unchanged.

• Spin-then-block queues (ipcSpinMicroseconds, default 100)

  • Queue waits spin, then block on a named event that the other side
    signals only while the waiter sleeps.
  • pull() copies the element out before releasing its slot.
  • The unused semaphore-based queue is removed.

• Present semaphore fix (presentSemaphoreMaxFrames, default 1)

  • Both sides create it with presentSemaphoreMaxFrames as the
    initial count, so the setting takes effect. Before, it was always
    lockstep.

• Source engine defaults (hl2.exe)

  • Creates no longer wait for responses.
  • Fixes the trailing space in the presentSemaphoreMaxFrames key.
  • The client sends its response policy at the handshake so both
    sides agree.
  • Failed creates are logged, and uploads to them are skipped.

• Perf counters (perfLogIntervalFrames, default 0 = off)

  • Logs per-frame command, wait and query statistics on both sides.

• Unit tests: sent_state, query_results, ipc_queue_wait,
  upload_parts.

Data queue fixes

These fix hangs and crashes on large texture uploads. The first two come
from xoxor4d's fork; the squashed commit credits them as co-author.

• No stall at the batch start (xoxor4d)
  • The client no longer stalls when its data ring wraps while the
    server waits at the current batch start. A server waiting there has
    consumed everything, so nothing can be overwritten.
• Absolute ring positions (xoxor4d)
  • The overwrite protection compares absolute word positions instead
    of ring offsets plus a reset flag.
  • The flag could strand both processes in a wait that never ended.
• Surface uploads in 1 MB blobs
  • A blob that does not fit before the end of the ring restarts at
    word 0, and the skipped tail counts against its command.
  • A single 64 MB blob (4096x4096 A8R8G8B8) could not fit in the
    default 96 MB ring. Before the fix, the queue deadlocked and the
    client then overwrote its own command header.
• Uploads over 16 MB split into several commands
  • A lock larger than the ring (for example a 256 MB A32B32G32R32F
    level) used to need a bigger clientChannelMemSize.
  • The server locks the surface on the first part and unlocks it on
    the last, so the runtime still sees one upload.
  • Pending locks are tracked per surface.
• Not covered: volume texture uploads and large server-to-client
  readbacks still go in one piece.

The protocol changed: deploy the client d3d9.dll and the server
NvRemixBridge.exe together.

Co-Authored-By: Claude Opus 5.5 (1M context) noreply@anthropic.com

[MarkEHenderson]

REMIX-6297 for tracking