Count what the shadow copy-ins overwrite and cost, and keep stores through the device address table behind a switch.

NX2NAVI_SHADOW_CHECK=1 logs with the shadow line, every 60 frames: the whole copy-ins that overwrote a store a
shader made through the device address table, and the pools they fell on; the write-backs that moved bytes into
a page the CPU had written since its copy-in; the bytes both sides changed since their base copy; what the
copy-ins cost by kind, and which pools and flag sets took them. The stores are found by content: a whole
copy-in leaves a hash of the page, and the next one compares the shadow page with it first.

NX2NAVI_SHADOW_TABLE_STORES=1 keeps such stores: a page found stored into gets a base copy, so copy-ins take
only the CPU's bytes, and the fence of every submission that uses the table writes the stores back to guest
memory. --probe-vkmem checks it with a GPU write that nothing notes. It is off by default. The overwrite it
removes was the suspected cause of the flicker on Linux and is not: standing outdoors (RTX 5090 with
NX2NAVI_SHADOW=1) it happens exactly four times a frame, on ten pages in pools of 20 and 44 KiB, with the copy
that precedes the frame's stores, and no frame came out wrong with it or without it in 1,800 frames each; the
hashes cost 1.8 ms a frame of the replay (67 against 75 fps); and a count the CPU resets before the write-back
has landed would keep counting, where the whole copy-in resets it every time.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
This commit is contained in:
chrg6
2026-11-10 23:15:02 -05:00
parent 175c1947de
commit 5b1413d817
7 changed files with 449 additions and 14 deletions

View File

@@ -283,9 +283,20 @@ pointers into either. A shader that reads through the table resolves nothing, so
the copy-in: a part whose pipelines declare the table is marked at its draws and dispatches, and its prologue brings
the shadow of every pool in the table up to date with the pages the CPU wrote and queues the same pages for those
pools' VRAM mirrors (`vk_memory.c` `vk_mem_gpa_sync`). That copies the written pages of every CPU-visible pool in
each such part, which the cost above does not include. What a shader stores through the table into a shadow copy does
not come back to guest memory. `--probe-vkmem` checks the read: a shader follows the table to a word of a pool that is
never bound, and must read `0x13579BDF` and, after a plain CPU write, `0x2468ACE0`. On Linux, run it for each backing:
each such part, which the cost above does not include. What a shader stores through the table lands in the shadow copy
and in the pool's VRAM mirror, and no binding names it, so nothing notes it as a GPU write: it does not come back to
guest memory, and the next whole copy-in of its page (the CPU rewrites other words of the same records) puts guest
memory's stale bytes back over it in both copies. In the runs measured so far that copy-in always came before the
frame's stores, which does no harm. `NX2NAVI_SHADOW_TABLE_STORES=1` keeps the stores instead: the copy-ins then keep a
hash of each page they write, a page whose shadow no longer matches its hash was stored into, keeps a base copy from
then on like a page with a noted GPU write, and has its stores written back to guest memory at the fence of every
submission that uses the table. It is off by default (`runtime/vk_memory.c`, "Pages a shader stores into", says why).
`NX2NAVI_SHADOW_CHECK=1` logs every 60 frames how often a copy-in overwrote such a store and for which pools, how many
write-backs landed on pages the CPU had written since their copy-in, and what the copy-ins cost and where they went.
`--probe-vkmem` checks the read: a shader follows the table to a word of a pool that is never bound, and must read
`0x13579BDF` and, after a plain CPU write, `0x2468ACE0`; and with `NX2NAVI_SHADOW_TABLE_STORES=1` the store: a GPU
write into a shadow page that nothing notes must survive the CPU's writes of that page and reach guest memory. On
Linux, run it for each backing:
```sh
for shadow in 0 1; do for memory in import mirror; do