Skip to content

Macro DMA reads wait only for compute/copy writers (env-gated experiment) - #16

Merged
NxtDoorActivity merged 1 commit into
nx2from
chata/dma-writer-wait
Nov 7, 2026
Merged

NxtDoorActivity merged 1 commit into
nx2from
chata/dma-writer-wait

Conversation

@NxtDoorActivity

Copy link
Copy Markdown
Owner

The GPU profile shows the front GPU thread blocked on the back thread most of each second: split drain ~450 ms/s, with ~5,400–8,700 macro DMA reads per second (macro 0xe2c) each draining the whole queue.

With EDEN_NX2_DMA_WRITER_WAIT=1, FrontDmaRead calls WaitForWriters (compute and copy work sent so far) instead of Drain. Draws are not tracked as writers, so a draw that writes the macro's source memory would be missed. Off by default.

Test: Kokiri idle with and without the variable; compare FPS and the split drain / macro dma read numbers from EDEN_NX2_GPU_PROFILE=1.


Generated by Claude Code

…WRITER_WAIT=1)

The GPU profile shows the front waiting for the back thread's whole queue on every
macro memory read (about 600 per frame). With the env var set it waits only for the
compute and copy work sent so far. Off by default.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RmvPX2tqfoEsDjDfWTAmSf
@NxtDoorActivity
NxtDoorActivity merged commit 44d6bc2 into nx2 Nov 7, 2026
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants