[HLSL] Add LinAlg FP8 input-vector and bias role coverage - #8869
Merged
Jack Elliott (JoeCitizen) merged 1 commit intoSep 2, 2026
Merged
Conversation
The existing FP8 MatVec coverage only exercises FP8 as the *matrix* component type. Proposal 0035 requires all three roles: "The DirectX API specification requires that all implementations support both FP8 formats for matrices, bias, and input vectors" (proposals/0035-linalg-matrix.md:2049-2053) This adds the two remaining roles. They are not the same kind of thing, and the difference is why they are implemented differently. Input vector: a real driver capability. The interpretation is an immarg on the DXIL MatVec op and VectorInputType is its own field in the D3D12 capability query, so this is expressible through the existing private-builtin shader. A D3D12 probe confirms WARP answers supported=1 for both FP8 formats as the vector type with an F16 matrix. Bias: not a distinct MatVec capability. Verified from two independent directions before writing any test. The lowered DXIL shows the bias arriving at the builtin as *native F16* after a separate dx.op.linAlgConvert, so by the time MatVecMulAdd sees it there is no FP8 left to support. The API agrees: querying an FP8 bias type returns supported=0, while an F32 bias control also returns 0, so the field discriminates rather than being ignored. The value is therefore in exercising the header's VectorRef load -> Convert -> truncate chain end to end, not in a new capability. The private builtin shares one component type between bias and result, so an interpreted bias is only expressible through the public dx::linalg header. Hence the new MatVecMulAddMemoryBiasShader and the BiasFromMemory flag. It reuses the existing bindings and root signature, so only the bias path differs from the register-bias cases. querySupport asks the driver about ResultType rather than BiasInputType for these cases, because that is what the lowered DXIL actually carries. Both roles are classified Mandatory. FP8 is required functionally even where it is not native: "the driver is required to emulate them. This emulation is required" (D3D12LinearAlgebraRuntimeFeatureSupport.md:427). Closing a harness gap first. encodeByte handled only I8/U8 and returned nullopt for FP8, so packing an FP8 input vector failed outright, while encodeComponents carried a separate duplicate FP8 codec. The codec now lives in encodeByte and encodeComponents delegates, which both fixes the gap and removes the duplication. The round-trip check is load-bearing: the oracle is an exact integer dot product and cannot absorb a rounded input, so a value that is not exactly representable now reports loudly instead of silently producing a wrong expectation. All chosen values stay within the set that is exact in E5M2, which has only two mantissa bits. No convert capability gate is applied. There is no convert entry in D3D12_LINEAR_ALGEBRA_OPERATION_TYPE to gate on, and FP8 conversion support is mandatory regardless. Validation on WARP: 82/80/0/2, against a 78/76/0/2 baseline on upstream/main. Compared per test rather than by totals: four tests added, zero outcome changes among the pre-existing 78. The two non-passing are the pre-existing I32 contention skips. Assisted-by: GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 83725f5d-8e98-4c1d-91ee-ad47629e007b
Copilot started reviewing on behalf of
Jack Elliott (JoeCitizen)
September 1, 2026 18:45
View session
Contributor
There was a problem hiding this comment.
Pull request overview
Adds runtime coverage for FP8 input-vector and memory-bias roles in linear algebra matrix-vector operations.
Changes:
- Extends byte encoding to FP8 components.
- Adds public-header shader coverage for FP8 memory bias.
- Adds mandatory tests for both FP8 formats.
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
| 1, 0, -1, 2, -2, 3, -3, 1, 0, 1, 2, -1, 3, -2, 1, -3, | ||
| -1, 2, 0, 1, -2, 1, 3, -1, 2, -1, 1, 0, 1, -3, -2, 3, | ||
| }; | ||
| // Only 0 through 8 are exactly representable in both FP8 formats. |
Alex Sepkowski (alsepkow)
approved these changes
Sep 1, 2026
Alex Sepkowski (alsepkow)
left a comment
Alex Sepkowski (alsepkow)
left a comment
Contributor
There was a problem hiding this comment.
LGTUS: Joshua Batista (@bob80905)
One minor nit.
| const std::optional<BYTE> Encoded = encodeHalfToFP8(Half, Type); | ||
| // The oracle is an exact integer dot product, so reject rounded inputs. | ||
| const std::optional<float> RoundTrip = | ||
| Encoded ? decodeFP8ToFloat(*Encoded, Type) : std::optional<float>(); |
Contributor
There was a problem hiding this comment.
nit:
Suggested change
| Encoded ? decodeFP8ToFloat(*Encoded, Type) : std::optional<float>(); | |
| Encoded ? decodeFP8ToFloat(*Encoded, Type) : std::nullopt; |
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
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.
The existing FP8 MatVec coverage only exercises FP8 as the matrix component type, but proposal 0035 requires every implementation to support both FP8 formats for matrices, bias and input vectors. This adds the two remaining roles, which are not the same kind of thing: the input-vector cells go through the existing private builtin because the interpretation is an immarg on the DXIL op and
VectorInputTypeis its own capability field, while the bias cells need the publicdx::linalgheader because the private builtin shares one component type between bias and result.Assisted-by: GitHub Copilot