[DjDeveloperr]

G17P ray queries run in candidate mode (control 0x9142), so the hardware queues every triangle hit, opaque ones included, for the software candidate loop. Each hit then costs a software pass of dependent loads plus a hardware round trip.

With field 12:11 set (0x9942), the hardware commits opaque triangle hits itself and queues only the hits software has to decide. This change uses 0x9942 when the ray flags leave opaque triangle hits to the hardware. That means none of these flags is set:

• Opaque. Such rays commit every hit anyway, and incoherent ones traced slower with hardware commits.
• NoOpaque.
• A facing or opacity cull.
• SkipTriangles.

After each trace or refill, the lowering takes the hardware's committed hit when it is closer than the current committed one. It builds the same packets a software commit would:

• the distance comes from the tmax slot;
• the hit words come from query words 256/512;
• the instance comes from the committed-instance pointer array at query offset 8960.

The current-instance array at 9152, which candidates use, holds the instance where traversal stopped, which can be past the committed hit. The merge only looks up the committed fields the shader reads (node, instance IDs, barycentrics).

How the layout was established:

• With 0x9942 alone, the CTS failures (procedural triangle_in_between, callable_shader, tri_and_box) came from the lowering ignoring hardware-committed hits. Merging the hardware record fixes them.
• An image-CRC comparison then exposed a stale instance for some reflection hits.
• A shader-printf scan of the query block, at a pixel with a wrong image, located the committed-instance array: it matched the pointer the software path commits.

Mirror of aurora-silicon#100 (merged as 3295ac7). Qualification on J700 (Neo) against aurora-wip 86ffb65:

• gpubench-rt RT plan, alternated ×2. Every rt.scene image CRC is identical to the baseline.

  Row	Change
  Solar Bay, ray query	+8.3%
  Solar Bay, ray-tracing pipeline	+7.7%
  extreme	+11.0%
  heavy	+5.9%
  light	+3.7–4.2%
  medium	+3.4–3.8%
  rt.asbuild	unchanged
  rt.rays, coherent and incoherent, all depths	−0.5…+1.2% (noise)

• Full ray-query + ray-tracing-pipeline CTS: 47102 pass / 10088 not supported / 0 fail, identical to the baseline.

• Host asahi meson suite: 75/75 on aurora-wip.

• No kernel errors during the runs.

This omacom tree has the identical agx_nir_gen1_ray_query.{c,h}. The cherry-pick builds with -Dwerror=true and passes the asahi meson suite 76/76.

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; the previous first-hit stopping issue is fixed.

Summary

Lets G17P hardware commit eligible opaque triangle hits and copies those hits into the query’s saved result.

• Eligible G17P ray queries now commit opaque triangle hits in hardware.

_{Reviews (2) · Last reviewed commit: "agx: end G17P ray queries with Terminate..." · Reviewed by Greptile}