Add YAML-driven acceleration structure definitions for .test files#1214
Conversation
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
from
25f1db9 to
9bc46ea
Compare
This comment was marked as resolved.
This comment was marked as resolved.
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
3 times, most recently
from
5327bdc to
fe1a6e6
Compare
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
from
fe1a6e6 to
a60120c
Compare
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
from
a60120c to
8d00c38
Compare
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
2 times, most recently
from
e2af1c3 to
a7d8a28
Compare
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
3 times, most recently
from
de55a15 to
c075d64
Compare
bogner
left a comment
bogner
left a comment
There was a problem hiding this comment.
My comments on the one test apply to all three, but otherwise this LGTM.
| # CHECK: Name: Output | ||
| # CHECK: Data: [ 1 ] |
There was a problem hiding this comment.
Why are we doing both a Results block and FileCheck? The FileCheck part is redundant.
There was a problem hiding this comment.
Interesting. Many other tests still seem to use a stdout-readback + FileCheck/CHECK, and even though FileCheck is more featureful (regex, CHECK-NEXT/NOT, ordering assertions on log output) most of those are also "just compare a buffer" and could equally well use a direct buffer comparison through Results:. Any clue why they don't — historical inertia from before Results: existed, or is there a deliberate reason I'm missing? Happy to do a consistency pass in a separate PR if you'd like them migrated.
There was a problem hiding this comment.
It's just historical inertia from before Results existed. If you feel like doing a consistency pass I wouldn't say no to it :)
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
from
c075d64 to
acf1e6e
Compare
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
from
acf1e6e to
6c4fecc
Compare
Add AccelerationStructure to ResourceKind and introduce pipeline structs (TriangleGeometry, AABBGeometry, BLASDesc, InstanceDesc, TLASDesc) with YAML parsing and name resolution for buffer/BLAS/TLAS references. Backend switch statements are updated to handle the new enum value. Includes three placeholder InlineRT test files exercising the new YAML format (gated behind REQUIRES: raytracing). Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
MarijnS95
force-pushed
the
cherry-pick-yaml-as-defs
branch
from
6c4fecc to
8ef7553
Compare
|
Test failures are unrelated to this PR, merging. |
4 participants
For #1158
Summary
AccelerationStructuretoResourceKindand introduce pipeline structs (TriangleGeometry,AABBGeometry,BLASDesc,InstanceDesc,TLASDesc) ininclude/Support/Pipeline.hwith YAML parsing and name-based resolution for buffer/BLAS/TLAS references inlib/Support/Pipeline.cpp.RaytracingTiercapability enum (NotSupported/1.0/1.1) wired throughDXFeatures.def/.h/.cppso the DX device can report a raytracing tier.acceleration-structurelit feature intest/lit.cfg.py, derived per-backend from DXRaytracingTier != NotSupported, MetalsupportsRaytracing, and VulkanVK_KHR_acceleration_structure.AccelerationStructurecases to the resource-kind switch statements in the DX, Metal, and Vulkan device backends (no execution yet — just enum coverage).triangle-setup,indexed-triangle-setup,multi-instance) exercising the new YAML format, gated behindREQUIRES: acceleration-structureand markedXFAIL: *because backend execution of acceleration structures is not implemented yet.This PR only introduces the YAML surface area, parsing, and capability detection — the runtime side that actually builds/binds acceleration structures is intentionally out of scope and will follow.
Test plan
check-hlsl-offloadpasses on DX, Metal, and Vulkan with the three new tests reported asXFAIL(or skipped whereacceleration-structureisn't available).