[VK] Add multi-slice + multi-mip Texture2DArray support (stacked on #1256)#1257
Draft
alsepkow wants to merge 17 commits into
Draft
[VK] Add multi-slice + multi-mip Texture2DArray support (stacked on #1256)#1257alsepkow wants to merge 17 commits into
alsepkow wants to merge 17 commits into
Conversation
Lifts the early-return guard on MipLevels > 1 for SRV textures and threads the mip count through to the D3D12_RESOURCE_DESC and the SRV's MipLevels field. Adds a per-mip upload path using GetCopyableFootprints so each subresource is copied with the correct D3D12_TEXTURE_DATA_PITCH_ALIGNMENT row pitch. The single-mip path is unchanged. RWTexture2D mips are still rejected with not_supported (no per-mip UAV support yet). Adds a new Texture2D.Load.MipMaps.test.yaml covering the 4x4 + 3-mip layout documented in docs/MipMappedTextures.md. Re-enables Texture2D.mips.OperatorIndex / Texture2D.OperatorIndex / Texture2D.GetDimensions on dxc-d3d12 and dxc-warp-d3d12 by qualifying the XFAIL to 'Clang && DirectX' for the unrelated Clang-DX path tracked in llvm/llvm-project#101558. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Addresses code review feedback on the multi-mip support change: the relaxed guard (changed from `!= 1` to `> 1 && !SRV`) silently accepted MipLevels values of 0 or negative integers. Since OutputProps.MipLevels is signed int, 0 would cast to a UINT(0) in the SRV desc (which D3D12 interprets as the magic "use all mip levels" value) and negative values would wrap to a huge uint16_t in the RESOURCE_DESC. Add an explicit lower-bound check so invalid input is rejected with a clear error instead of silently producing surprising behavior. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Apply clang-format 19.1.6 to changed regions per pr-code-format CI.
Adds a new `Texture2DArray` `ResourceKind` and wires it through the DirectX backend so shaders can `Load` from layered SRVs: * `OutputProperties` gains an `ArraySize` field (default 1) and the YAML size validation now accounts for it. * `getResourceDescription` sets `DepthOrArraySize` from `ArraySize` for `Texture2DArray` resources and rejects mismatched uses of `ArraySize > 1` on non-array kinds. * `getSRVDescription` emits a `TEXTURE2DARRAY` view covering all slices. * The upload path enumerates slices via `CopyTextureRegion` with a per-slice `D3D12_PLACED_SUBRESOURCE_FOOTPRINT`; the host-side data layout is tightly packed slice-major (slice0 rows | slice1 rows | ...). Vulkan and Metal switches gain `Texture2DArray` cases for build completeness; their device implementations remain TODO and the new test is gated `UNSUPPORTED: Vulkan || Metal`. Adds `Feature/Textures/Texture2DArray.Load.test.yaml` exercising a 2x2x3 array with `Load(int4)` and offset variants. Verified passing on d3d12 (NVIDIA RTX 5060 Ti) and warp-d3d12. 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>
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>
Extends Texture2DArray support to RWTexture2DArray: - Pipeline.h: add ResourceKind::RWTexture2DArray; update isTexture, isUAV, getDXKind, getDXDimension switches and YAML enum trait. - Enums.h: API enum value. - DX/Device.cpp: createUAV uses D3D12_UAV_DIMENSION_TEXTURE2DARRAY with FirstArraySlice=0 and ArraySize from OutputProperties. Multi-slice readback via per-slice CD3DX12_TEXTURE_COPY_LOCATION with placed footprint offset (note: requires Width*ElementSize to be 256-aligned; documented in issue plan). - VK/Device.cpp, MTL/MTLDevice.cpp: llvm_unreachable placeholders. Test: RWTexture2DArray.Store.test.yaml verifies a 16x4x2 compute write/readback round-trip. 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>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
# Conflicts: # include/Support/Pipeline.h
Clang HLSL does not yet implement the Texture2DArray / RWTexture2DArray templates. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Keep mip-aware getResourceDescription from llvm#1230 alongside Tex2DArray plumbing from llvm#1239. The upload helper preserves both paths conditionally: mips path when MipFootprints non-empty (Texture2D mip-chain), slice path otherwise (Texture2DArray flat). Multi-mip + multi-slice support layers on top in a follow-up commit. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…i-mip support Combines the previously-separate per-mip and per-slice texture-upload paths into a single shared path driven by ID3D12Device::GetCopyableFootprints so that Texture2DArray resources can now carry per-slice mip chains. Changes ------- * Texture2DArray SRV/UAV view descriptions now use `MipLevels` from the CPU buffer (rather than hardcoding 1) so multi-mip array textures expose the full mip chain to the shader. * `addResourceUploadCommands` now accepts an optional `ArrayRef<D3D12_PLACED_SUBRESOURCE_FOOTPRINT>` and, when supplied, iterates all subresources in D3D12 ordering (`Sub = Slice * MipLevels + Mip`) instead of only the base subresource. The single-subresource fast path is preserved for unaffected callers. * `createSRV` / `createUAV` route through `GetCopyableFootprints` whenever `MipLevels > 1 || ArraySize > 1` to obtain D3D12-aligned per-subresource row pitches for the upload buffer. This subsumes the prior bespoke per-mip and per-slice handling. * Memcpy loop copies each subresource's tightly packed CPU rows into the appropriately padded D3D12 footprint, advancing the source pointer by the tight row size for each subresource. CPU data layout is the natural D3D12 subresource order: slice-major, mip-major within slice. This consolidates the per-slice 256-byte-alignment helper requested as a follow-up on llvm#1077 -- there is now exactly one path for all multi-subresource textures. Test ---- New `Texture2DArray.Load.MipMaps.test.yaml` exercises a 2-slice, 2-mip Texture2DArray and reads back one texel from each (slice, mip) pair via `Texture2DArray::Load(int4(x, y, slice, mip))`. Existing tests verified locally on d3d12: PASS Texture2DArray.Load PASS Texture2DArray.Load.MipMaps <-- new PASS RWTexture2DArray.Store PASS Texture2D.Load.MipMaps PASS Texture2D.mips.OperatorIndex VK/MTL scope: the new test is gated `UNSUPPORTED: Vulkan || Metal` to match the existing Texture2DArray tests. Multi-slice + multi-mip support on the Vulkan and Metal backends remains a follow-up (tracked under llvm#1077 backend-parity ACs). Refs llvm#1077 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Lifts the hardcoded �rrayLayers = 1 / �aseArrayLayer = 0 / layerCount = 1 sites in the Vulkan backend so that Texture2DArray (and RWTexture2DArray) resources now route through actual multi-layer VkImage / VkImageView / VkBufferImageCopy paths. Combined with the multi-subresource upload changes on DX in llvm#1256, this closes the Tex2DArray cross-backend convergence gap. Changes ------- * createImage: arrayLayers = max(1, OutputProps.ArraySize) * Descriptor view: layerCount = max(1, ArraySize), levelCount = MipLevels * copyResourceDataToDevice: VkBufferImageCopy regions now iterate (Slice, Mip) in D3D12 subresource order (slice-major, mip-major within slice). baseArrayLayer = Slice. SubRange.layerCount = ArraySize. * copyResourceDataToHost: same expansion on the readback path. Subresource buffer-offset convention matches the existing DX layout in llvm#1256 (the YAML Data array is laid out slice-major, mip-major within slice; per-subresource bufferOffset is computed accumulatively). Tests ----- Existing Tex2DArray tests are un-gated on Vulkan: Texture2DArray.Load.test.yaml (UNSUPPORTED: Vulkan -> Metal only) RWTexture2DArray.Store.test.yaml (UNSUPPORTED: Vulkan -> Metal only) Texture2DArray.Load.MipMaps.test.yaml (UNSUPPORTED: Vulkan -> Metal only) Local lit (serial, d3d12 + vk): Feature/Textures full sweep -- 30 PASS / 10 XFAIL / 0 FAIL This includes 3 new VK PASSes (Tex2DArray.Load, RWTex2DArray.Store, Tex2DArray.Load.MipMaps) without regressing any of the 27 prior PASSes or 10 XFAILs. Refs llvm#1077 Stacked on llvm#1256 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
6 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked draft PR -- depends on #1256 (DX multi-subresource upload). Target should be re-pointed once #1256 lands.
What this adds
Lifts the hardcoded �rrayLayers = 1 / �aseArrayLayer = 0 / layerCount = 1 sites in the Vulkan backend so that
Texture2DArray(andRWTexture2DArray) resources route through actual multi-layerVkImage/VkImageView/VkBufferImageCopypaths. Combined with the multi-subresource upload changes on DX in #1256, this closes theTexture2DArraycross-backend convergence gap.createImage:arrayLayers = max(1, OutputProps.ArraySize)layerCount = max(1, ArraySize),levelCount = MipLevelscopyResourceDataToDevice):VkBufferImageCopyregions now iterate(Slice, Mip)in D3D12 subresource order (slice-major, mip-major within slice).baseArrayLayer = Slice. Subresource barrier rangelayerCount = ArraySize.copyResourceDataToHost): same expansion.Subresource buffer-offset convention matches the DX layout in #1256 (the YAML
Dataarray is laid out slice-major, mip-major within slice; per-subresourcebufferOffsetis computed accumulatively).Test
Three existing tests are un-gated on Vulkan and now PASS on both backends:
Texture2DArray.Load.test.yamlRWTexture2DArray.Store.test.yamlTexture2DArray.Load.MipMaps.test.yaml(new in #1256)Full
Feature/Textureslit sweep (d3d12 + vk, serial): 30 PASS / 10 XFAIL / 0 FAIL -- 3 new VK PASSes without regressing any of the 27 prior PASSes or 10 XFAILs.Convergence note
This is the natural backend-parity follow-up to #1256 and was kept as its own PR for reviewability -- the DX changes in #1256 are conceptually distinct (unified
GetCopyableFootprintsupload path) from the VK changes here (lift hardcoded singletons + iterate slices/mips). Metal still needs the same treatment (currentlyllvm_unreachable); tracked as backend-parity work under #1077.Refs #1077
Stacked on #1256