[SPIR-V] Add SPV_EXT_descriptor_heap + SPV_KHR_untyped_pointers codegen by jzakharovnv · Pull Request #8517 · microsoft/DirectXShaderCompiler

jzakharovnv · 2026-06-04T21:43:25Z

Building off of #8281, this PR adds a native lowering via SPV_EXT_descriptor_heap and SPV_KHR_untyped_pointers and is part 1/4 in a series.

ResourceDescriptorHeap and SamplerDescriptorHeap are lowered to untyped variables decorated with ResourceHeapEXT and SamplerHeapEXT. Each heap access emits OpUntypedAccessChainKHR into a runtime array of the appropriate descriptor type. Buffer-like resources (StructuredBuffer, ByteAddressBuffer, ConstantBuffer, TextureBuffer) use OpTypeBufferEXT and OpBufferPointerEXT; image and sampler resources use OpLoad. Interlocked operations on RWTexture use OpUntypedImageTexelPointerEXT.

Requires -fspv-target-env=vulkan1.3.

Assisted by an AI agent.

@dnovillo

github-actions · 2026-06-04T21:45:20Z

✅ With the latest revision this PR passed the C/C++ code formatter.

Building off of microsoft#8281, this commit adds a native lowering via SPV_EXT_descriptor_heap and SPV_KHR_untyped_pointers. ResourceDescriptorHeap and SamplerDescriptorHeap are lowered to untyped variables decorated with ResourceHeapEXT and SamplerHeapEXT. Each heap access emits OpUntypedAccessChainKHR into a runtime array of the appropriate descriptor type. Buffer-like resources (StructuredBuffer, ByteAddressBuffer, ConstantBuffer, TextureBuffer) use OpTypeBufferEXT and OpBufferPointerEXT; image and sampler resources use OpLoad. Interlocked operations on RWTexture use OpUntypedImageTexelPointerEXT. Requires -fspv-target-env=vulkan1.3. Assisted-by: Claude.

jzakharovnv · 2026-06-04T23:02:24Z

@microsoft-github-policy-service agree company="NVIDIA"

dnovillo

Thanks for this! I just started looking at it and have a couple of questions. I'll add more as I read the PRs.

dnovillo · 2026-06-08T21:05:41Z

+SpirvEmitter::getDescriptorHeapRuntimeArrayType(const SpirvType *elemType,
+                                                bool onSamplerHeap) {
+  constexpr uint32_t kDefaultResourceHeapStride = 64;
+  constexpr uint32_t kDefaultSamplerHeapStride = 32;


I think I would float these defaults to SpirvEmitter.h and document where the seemingly magic values 32 and 64 come from.

Definitely agree that these need some better documentation. I placed them at this scope because they seem a little niche, and I want to try to avoid polluting very wide scopes, but I can move them up.

dnovillo · 2026-06-08T21:54:36Z

+  constexpr uint32_t kDefaultSamplerHeapStride = 32;
+  const uint32_t stride =
+      onSamplerHeap ? kDefaultSamplerHeapStride : kDefaultResourceHeapStride;
+  return spvContext.getRuntimeArrayType(elemType, stride);


It doesn't seem that there are tests for OpDecorate ArrayStride N in this PR. I haven't checked the others. Could you add one (unless I missed it?)

Yes, there is a healthy amount of testing in the last 2 PRs of this series, but I neglected to have for this one. I can fix that.

llvm-beanz · 2026-06-23T15:28:07Z

-    addExtension(Extension::EXT_descriptor_heap, "DescriptorHeap", {});
-    addExtension(Extension::KHR_untyped_pointers, "DescriptorHeap", {});
+    const llvm::StringRef feature = "DescriptorHeap";
+    featureManager.requestTargetEnv(SPV_ENV_VULKAN_1_3, feature, {});


This function can return failure (which is getting dropped here), and there is no test verifying its error handling.

Thanks for catching that, I'll add a test.

llvm-beanz · 2026-06-23T15:32:16Z

+            tryToAssignToDescriptorHeapBuffer(expr))
+      return aliasResult.getValue();
+
+    auto *rhs = loadIfGLValue(expr->getRHS());


LLVM's coding standards which DXC adopts (although not historically well enforced), have an "almost never auto" policy:

https://github.com/microsoft/DirectXShaderCompiler/blob/main/docs/CodingStandards.rst#use-auto-type-deduction-to-make-code-more-readable

The heavy use of auto in this code makes it significantly harder to review this PR in a browser without an IDE telling me the types of things.

Noted, thanks for tip!

llvm-beanz · 2026-06-23T15:33:30Z

+  if (result && !result->isRValue()) {
+    result =
+        spvBuilder.createLoad(structType, result, buffer->getExprLoc(), range);
+  }


Suggested change

if (result && !result->isRValue()) {

result =

spvBuilder.createLoad(structType, result, buffer->getExprLoc(), range);

}

if (result && !result->isRValue())

result =

spvBuilder.createLoad(structType, result, buffer->getExprLoc(), range);

nit: https://llvm.org/docs/CodingStandards.html#don-t-use-braces-on-simple-single-statement-bodies-of-if-else-loop-statements

Will fix in next commit

llvm-beanz · 2026-06-23T15:35:31Z

+  }
+
+  // ConstantBuffer -> Uniform (UBO); all others -> StorageBuffer (SSBO)
+  // TODO: Remove this manual override once LowerTypeVisitor returns the


Should this be fixed before we merge this change? Seems like you're working around a clear bug here.

To get the storage class here rather than in LowerTypeVisitor is bit of a bandage but provides functionally correct output. I had left this TODO in with the aim of making this PR as small/unintrusive as possible and with room to improve with future commits. If you're okay with filing an issue, let me know, otherwise I can go about cleaning up this fix.

llvm-beanz · 2026-06-23T15:41:19Z

+
+float4 main(uint idx : A) : SV_Target {
+  Texture2D<float4> tex = ResourceDescriptorHeap[NonUniformResourceIndex(idx)];
+  SamplerState samp = SamplerDescriptorHeap[NonUniformResourceIndex(idx + 1)];


This seems like something we should have the compiler issue a diagnostic on. Silently dropping something the user explicitly wrote seems unfortunate.

I agree, this is a bit esoteric but correct according to @Tobski. How should we handle a diagnostic here, just a simple warning?

Oh... Actually I think the correct solution for HLSL is that in the absence of NonUniformResourceIndex an access should be marked as Uniform, and NonUniformResourceIndex suppresses that.

It should have an effect, so no diagnostic should be required.

github-project-automation Bot added this to HLSL Roadmap Jun 4, 2026

github-project-automation Bot moved this to New in HLSL Roadmap Jun 4, 2026

jzakharovnv mentioned this pull request Jun 4, 2026

[SPIR-V] Add descriptor heap RaytracingAccelerationStructure support #8518

Open

jzakharovnv force-pushed the pr1-descriptor-heap-core branch from 4b4b143 to f640327 Compare June 4, 2026 21:58

jzakharovnv force-pushed the pr1-descriptor-heap-core branch from f640327 to cbcae38 Compare June 4, 2026 22:42

dnovillo suggested changes Jun 8, 2026

View reviewed changes

github-project-automation Bot moved this from New to In progress in HLSL Roadmap Jun 8, 2026

llvm-beanz reviewed Jun 23, 2026

View reviewed changes

Uh oh!

Conversation

jzakharovnv commented Jun 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions Bot commented Jun 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

jzakharovnv commented Jun 4, 2026

Uh oh!

dnovillo left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

jzakharovnv commented Jun 4, 2026 •

edited

Loading

github-actions Bot commented Jun 4, 2026 •

edited

Loading