diff options
| author | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-11-15 18:53:00 -0500 |
|---|---|---|
| committer | Arthur Hanson Rasmusson <arthur@vgpu.io> | 2026-11-15 18:53:00 -0500 |
| commit | cb786b9bdb188b49614ebba77d8359f74af5e416 (patch) | |
| tree | dd52ac81f61ba63e28360545fa60b5d1925a0192 /tools/cuda-ab/perf/PLAN-AND-HARNESS.md | |
| parent | 01c22a8d96d9de45a02d7556178445b3eed89f2b (diff) | |
| download | open-iov-cb786b9bdb188b49614ebba77d8359f74af5e416.tar.gz open-iov-cb786b9bdb188b49614ebba77d8359f74af5e416.zip | |
vgsp: eng71 - scan the tracked data pages for the vertex payload and bind VS0
c812/c813: the 151-word DRAW-window release has VS0=VS1=VS3=0 and
SZ0=SZ1=SZ3=0; the DRAW_IAB emits its 8x3-vertex draws (VA=0xd74=0x15)
against no vertex stream, so nothing rasterizes. The client's vertex
payload is a 24-vertex, 20-byte-stride float2 buffer (FORMAT 0x1014,
ATTR(0) float2 offset 0) written into the launch's tracked data pages.
sema_gl_vs_autobind (0=off, 1=scan+log, 2=bind; params=2): scan the
0x120400000..0x120800000 data pages for a 512-byte block of 24
plausible float2 vertices (score>=8 nice values), then prepend to the
saved release window: VS0 location [fmt=0x1014, upper, lower=reloc_va
+off], and the Turing stream SIZE pair (0x600<-0, 0x604<-0x100000).
VS-SCAN/MME-VS-BIND/MME-VS-SIZE receipts; the starts 0..21 x stride
20 resolve inside the 1 MB bound.
Diffstat (limited to 'tools/cuda-ab/perf/PLAN-AND-HARNESS.md')
0 files changed, 0 insertions, 0 deletions
