[DX][VK] Implement SV_Depth pixel-shader output and Depth32 SRV mapping (#1046) by alsepkow · Pull Request #1229 · llvm/offload-test-suite

alsepkow · 2026-05-27T01:21:04Z

Add SV_Depth pixel-shader output and Depth32-as-SRV support to the D3D12 and Vulkan backends.

Introduces a DepthBuffer YAML binding with an explicit Format field so tests can declare a CPU-readable depth attachment. The D3D12 backend creates a DSV, transitions to COPY_SOURCE for readback, and maps Depth32 to R32_FLOAT for SRV reads. Vulkan has a symmetric implementation; Metal returns an unsupported error for now.

Tests: SVDepth.test (pixel-shader depth output) and Texture2D.Load.Depth32.test.yaml (compute SRV read).

Resolves #1046

Assisted by Claude Opus 4.7

Adds the missing case for DataFormat::Depth32 in DX getDXFormat() so that SRV/UAV reads of textures declared with a depth-compatible format succeed. The depth-stencil view path (createTexture) uses getDXGIFormat() which already maps to D32_FLOAT. Adds Texture2D.Load.Depth32.test.yaml exercising the new path with a compute Load() of a 2x2 single-channel depth-format texture. Issue: llvm#1046 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Adds end-to-end coverage for SV_Depth by introducing a new `Bindings.DepthBuffer` schema key that lets a test bind a CPU-readable depth attachment, draws four single-pixel triangles each writing a distinct exact-binary-fraction depth via SV_Depth, copies the depth target back to a host buffer, and compares element-by-element against the expected values on D3D12, WARP, and Vulkan. Schema/API plumbing ------------------- * `IOBindings` gains `DepthBuffer` (name) + `DepthBufferPtr` (resolved CPUBuffer*), mirroring the existing RenderTarget pattern, with YAML parsing and name resolution in `lib/Support/Pipeline.cpp`. * New cross-backend helper `createDepthBufferFromCPUBuffer` in `lib/API/Device.cpp` builds a D32_FLOAT DepthStencil texture from a Depth32/1-channel CPUBuffer and validates the texture description against the buffer. * `isFloatingPointFormat` now accepts `DataFormat::Depth32` so the existing `BufferFloatULP` comparator can be used for depth read-back, matching the existing `testBufferFloatULP` case mapping Depth32 to a float compare. DirectX backend (`lib/API/DX/Device.cpp`) ------------------------------------------- * `createDepthStencil` now branches on `Bindings.DepthBufferPtr`: when set it builds the depth target from the helper and allocates a paired `DSReadback` buffer; otherwise it falls back to the existing default depth target (no read back). * After the draw, the depth target is transitioned from DEPTH_WRITE to COPY_SOURCE and copied to `DSReadback` using a placed footprint whose format matches the resource's actual DXGI format (D32_FLOAT), not the SRV-cast (R32_FLOAT) used for shader reads. * `readBack` maps `DSReadback`, gets copyable footprints, and copies into the user's `DepthBufferPtr` via `copyFromTexture`. * PSO DSVFormat is now derived from the actual depth target (`State.DepthStencil->getDesc().Fmt`) instead of being hard-coded to D32_FLOAT_S8X24_UINT, so depth-only formats are valid here. Vulkan backend (`lib/API/VK/Device.cpp`) ------------------------------------------ * InvocationState gains the same `DSReadback` buffer. * `createDepthStencil` mirrors the DX branch. * After `endEncoding`, `copyTextureToReadback` is reused for the depth attachment with the depth aspect/layout/stage masks (`LAYOUT_DEPTH_STENCIL_ATTACHMENT_OPTIMAL`, `ACCESS_DEPTH_STENCIL_ATTACHMENT_WRITE_BIT`, `LATE_FRAGMENT_TESTS_BIT`). * `readBackData` maps and copies depth bytes into the `DepthBufferPtr` using its image row stride. Metal backend (`lib/API/MTL/MTLDevice.cpp`) --------------------------------------------- * `createDepthStencil` now fail-fasts with `not_supported` if a test binds `Bindings.DepthBuffer` so the next contributor knows to wire it up instead of getting silently undefined behavior. Test ---- * New `test/Feature/Semantics/SVDepth.test` runs on D3D12, WARP, and VK and verifies four per-pixel SV_Depth writes (0.125, 0.25, 0.375, 0.5) against a Depth32 read-back buffer with `BufferFloatULP` ULP 0. `XFAIL: Clang` because DXC's HLSL -> DXIL lowering of SV_Depth is not yet implemented in clang. * Fix a YAML-schema regression in the previously-committed `Texture2D.Load.Depth32.test.yaml`: `DispatchSize` -> top-level `DispatchParameters: { DispatchGroupCount: ... }` to track the upstream schema rename. SV_StencilRef is intentionally out of scope until a stencil-bearing `DataFormat` (e.g. D24S8) is added; `createDepthBufferFromCPUBuffer` deliberately uses `Format::D32Float`. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Apply clang-format 19.1.6 to changed regions per pr-code-format CI.

The Metal offload backend cannot bind a Depth32 texture as an SRV through this path (rejects with Metal does not support buffer robustness). Mark the test UNSUPPORTED on Metal until llvm#1046 grows a Metal story; D3D12 and Vulkan continue to PASS. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Mac MTL CI reports `Bindings.DepthBuffer is not yet supported on the Metal backend.` for `Feature/Semantics/SVDepth.test`. The Metal device backend has no handling for the `DepthBuffer` binding introduced alongside this test, so mark it `UNSUPPORTED: Metal` until the MTL implementation lands. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Picks up the DriverVer fix (upstream llvm#1228) so clang-tidy stops failing on Device.cpp. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Fixes misc-const-correctness warnings-as-errors from clang-tidy in CI. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

manon-traverse · 2026-05-28T14:49:46Z

+  auto TexFmtOrErr = toFormat(Buf.Format, Buf.Channels);
+  if (!TexFmtOrErr)
+    return TexFmtOrErr.takeError();
+  if (*TexFmtOrErr != Format::D32Float)
+    return llvm::createStringError(
+        std::errc::invalid_argument,
+        "Depth buffer binding requires DataFormat::Depth32 with 1 channel.");


Why do we parse a format from the .yaml file if we only accept one format?
Also, we should consider using the Format enum in the .yaml so we don't have to do this conversion in the first place. We can then just check isDepth instead, which will also allow us to support other depth formats like those that include a stencil buffer.

Followed your suggestion.

# Conflicts: # lib/API/DX/Device.cpp # lib/Support/Pipeline.cpp

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Address review feedback on the SV_Depth path: the YAML depth-buffer binding parsed its format from a CPUBuffer's DataFormat+Channels and then errored out unless that resolved to Format::D32Float, which was a needless round-trip through the transitional toFormat() bridge and hard-coded a single accepted format. Take the format directly from the YAML instead. Bindings.DepthBuffer is now a struct rather than a bare buffer name: DepthBuffer: Name: DepthTarget Format: D32Float # optional; defaults to D32Float createDepthBufferFromCPUBuffer takes the Format as an argument, drops the toFormat() call, and validates with isDepthFormat() so depth-stencil formats like D32FloatS8Uint can be plumbed through once a backend wires them up. The CPUBuffer remains the readback destination and is no longer used as the format source; a new validateTextureDimsMatchCPUBuffer helper checks width/height/size without re-deriving Format via toFormat(). DX, VK, and MTL backends, the IOBindings YAML mapping, and the SVDepth lit test are updated for the new binding shape. SVDepth.test still passes on D3D12 and Vulkan. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

# Conflicts: # include/API/FormatConversion.h

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Fixes crash on depth-only formats like D32Float.

alsepkow · 2026-06-18T02:15:38Z

+# RUN: %dxc_target -T cs_6_0 -Fo %t.o %t/source.hlsl
+# RUN: %offloader %t/pipeline.yaml %t.o
+
+# Metal's offload backend cannot bind a depth-format texture as an SRV


nit: comment not needed

alsepkow · 2026-06-22T22:53:54Z

Moving this back to draft. Some NFC changes were factored out into a separate PR. #1327
When thats merged I'll merge it into this branch and remove the PR from draft again.

manon-traverse · 2026-06-23T06:26:20Z

+    if (IS.DSReadback) {
+      void *DSMapped = nullptr;
+      auto &DSReadback = llvm::cast<DXBuffer>(*IS.DSReadback);
+      if (auto Err = HR::toError(DSReadback.Buffer->Map(0, nullptr, &DSMapped),
+                                 "Failed to map depth buffer readback"))
+        return Err;
+
+      auto &DS = llvm::cast<DXTexture>(*IS.DepthStencil);
+      const D3D12_RESOURCE_DESC DSDesc = DS.Resource->GetDesc();
+      D3D12_PLACED_SUBRESOURCE_FOOTPRINT DSPlaced = {};
+      uint32_t DSNumRows = 0;
+      uint64_t DSRowSizeInBytes = 0;
+      uint64_t DSTotalBytes = 0;
+      Device->GetCopyableFootprints(&DSDesc, 0u, 1u, 0u, &DSPlaced, &DSNumRows,
+                                    &DSRowSizeInBytes, &DSTotalBytes);
+
+      P.Bindings.DepthBuffer.Ptr->copyFromTexture(DSMapped,
+                                                  DSPlaced.Footprint.RowPitch);
+      DSReadback.Buffer->Unmap(0, nullptr);
+    }


Let's use the copyTextureToBuffer function on the ComputeEncoder instead ^.^
This got recently added to main for the DX12 backend and the VK and MTL backends will use that path soon too.

alsepkow and others added 7 commits May 26, 2026 17:24

[NFC] clang-format

2dff94e

Apply clang-format 19.1.6 to changed regions per pr-code-format CI.

Merge upstream/main into alex/dx-depth-1046

e35a7e2

Picks up the DriverVer fix (upstream llvm#1228) so clang-tidy stops failing on Device.cpp. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

[NFC] Add const to ReadbackDX references for clang-tidy

b2306cc

Fixes misc-const-correctness warnings-as-errors from clang-tidy in CI. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

manon-traverse reviewed May 28, 2026

View reviewed changes

Comment thread lib/API/DX/Device.cpp Outdated

alsepkow and others added 4 commits May 29, 2026 17:34

Merge remote-tracking branch 'upstream/main' into alex/dx-depth-1046

bd29d97

# Conflicts: # lib/API/DX/Device.cpp # lib/Support/Pipeline.cpp

Trim verbose comments on depth-buffer paths

a7b1b16

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Merge remote-tracking branch 'upstream/main' into alex/dx-depth-1046

ec15b5a

alsepkow requested a review from manon-traverse June 15, 2026 22:51

alsepkow commented Jun 16, 2026

View reviewed changes

Comment thread lib/API/DX/Device.cpp Outdated

alsepkow and others added 4 commits June 15, 2026 18:47

Cleanup

5e9a623

Merge branch 'main' into alex/dx-depth-1046

814cc33

# Conflicts: # include/API/FormatConversion.h

[NFC] clang-format

b2a28e4

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

[NFC] Add const to Fmt for clang-tidy

db5af03

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

alsepkow marked this pull request as ready for review June 17, 2026 04:08

alsepkow added 3 commits June 16, 2026 21:20

Merge remote-tracking branch 'upstream/main' into alex/dx-depth-1046

ccf9f96

[VK] Only set stencil aspect bit when format has stencil

c515247

Fixes crash on depth-only formats like D32Float.

Merge upstream/main to pick up transpose XFAIL removal

d42b739

alsepkow commented Jun 18, 2026

View reviewed changes

alsepkow mentioned this pull request Jun 22, 2026

Express GPU texture formats directly via a CPUBuffer GpuFormat field #1326

Open

alsepkow marked this pull request as draft June 22, 2026 22:52

manon-traverse reviewed Jun 23, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[DX][VK] Implement SV_Depth pixel-shader output and Depth32 SRV mapping (#1046)#1229

[DX][VK] Implement SV_Depth pixel-shader output and Depth32 SRV mapping (#1046)#1229
alsepkow wants to merge 18 commits into
llvm:mainfrom
alsepkow:alex/dx-depth-1046

alsepkow commented May 27, 2026 •

edited

Loading

Uh oh!

manon-traverse May 28, 2026

Uh oh!

alsepkow Jun 16, 2026

Uh oh!

Uh oh!

Uh oh!

alsepkow Jun 18, 2026

Uh oh!

alsepkow commented Jun 22, 2026

Uh oh!

manon-traverse Jun 23, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

alsepkow commented May 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

manon-traverse May 28, 2026

Choose a reason for hiding this comment

Uh oh!

alsepkow Jun 16, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

alsepkow Jun 18, 2026

Choose a reason for hiding this comment

Uh oh!

alsepkow commented Jun 22, 2026

Uh oh!

manon-traverse Jun 23, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

alsepkow commented May 27, 2026 •

edited

Loading