[HLSL] Require FP8 conversion support in the LinAlg convert gate by JoeCitizen · Pull Request #8868 · microsoft/DirectXShaderCompiler

Jack Elliott (JoeCitizen) · 2026-09-01T17:20:19Z

FP8 conversion destinations were classified CapabilityGated, but there is no capability query for the convert operation — D3D12_LINEAR_ALGEBRA_OPERATION_TYPE has six values and none is convert, so the gate proxies through THREAD_VECTOR_MATRIX_MULTIPLY, and proposal 0035 requires every implementation to support both FP8 formats for matrices, bias and input vectors. This makes those destinations Mandatory, with TierSupported and SourceSupported out-params so a device advertising no LinAlg tier, or no native 16-bit shader ops, still skips rather than fails.

Assisted-by: GitHub Copilot

The LinAlg convert tests gate on a capability query and classify every outcome as CapabilityGated, so a device that reports no FP8 support skips rather than fails. That is wrong for FP8: proposal 0035 "Emulating FP" states that "the DirectX API specification requires that all implementations support both FP8 formats for matrices, bias, and input vectors", and permits emulation where the hardware lacks native support. The D3D12 runtime feature document agrees - the "Optional" entries in the Tier 1 vector-matrix table are the Native column only, L427 makes the emulation itself required, and the version history records "FP8 is required for tier 1". An implementation that cannot convert to FP8 is non-conformant, so the suite must report it. There is no capability query for the convert operation; the operation type enum covers matrix construction, the three multiply scopes, outer product and atomic accumulate store. The gate therefore proxies through THREAD_VECTOR_MATRIX_MULTIPLY, asking whether any supported configuration advertises FP8 as its VectorInputType. That proxy is sound precisely because 0035 makes FP8-as-input-vector mandatory. queryConvertSupport previously collapsed three distinct outcomes into a single unsupported answer: no linear algebra tier at all, an unsupported source component type, and an unsupported destination. Only the last is a conformance failure. The tier and source results are now reported separately so the caller can downgrade to CapabilityGated for the first two, mirroring matVecMulApplicable. Without that downgrade this change would fail devices that have no linear algebra, or that lack native 16-bit shader operations and so cannot perform an F16 source conversion at all - both strictly worse than the gap being closed. The requirement is passed at the call site, matching how the rest of the file declares mandatory cases. Only the two FP8 destination gates become Mandatory; the I16, I32 and F16 destinations keep their existing CapabilityGated classification, so their behaviour is unchanged. Validation is a per-test differential on preview WARP: 78/76/0/2 with zero per-test movement against the parent commit. That is the expected result and it is not evidence the change works - WARP enumerates VectorInputType=20 and 21, so it never reaches the classification. The evidence is a discriminating control that forces the FP8 destination query to report unsupported: this commit fails both FP8 conversion tests, while the parent skips them. Two further controls confirm the downgrades, with an absent tier and absent native 16-bit support each still skipping. No IHV driver in the loop reports FP8 vector input as unsupported, so the mandatory path is exercised only by forced control on WARP and never by a real driver refusing. This builds on the FP8 datatype mapping fix; without it the gate asserts in toLinAlgDataType before the classification is reached. Assisted-by: GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: 83725f5d-8e98-4c1d-91ee-ad47629e007b

Copilot

Pull request overview

Updates LinAlg conversion applicability so Tier 1 devices must advertise FP8 destinations while unsupported tiers or source types still skip.

Changes:

Exposes tier and source support from conversion queries.
Adds configurable mandatory versus capability-gated checks.
Marks both FP8 conversion destinations mandatory.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Joshua Batista (bob80905)

LGTM, though for consistency it would be nice if queryTierSupport could set the TierSupported bool itself like how queryConvertSourceSupport sets SourceSupported.

Alex Sepkowski (alsepkow)

LGTM. Just some naming/cleanliness nits

Alex Sepkowski (alsepkow) · 2026-09-02T02:44:43Z

+  // A device without linear algebra is outside the Tier 1 requirements, and a
+  // mandatory destination is still unreachable when the source type itself is
+  // unsupported, so both skip rather than failing.
+  const bool QueryAnswered = SUCCEEDED(QueryResult);
+  const bool NoLinearAlgebra = QueryAnswered && !TierSupported;
+  const bool UnsupportedSource = QueryAnswered && !SourceSupported;
+
+  linalg_test::CapabilityRequirement Effective = Requirement;
+  if (NoLinearAlgebra || UnsupportedSource)
+    Effective = linalg_test::CapabilityRequirement::CapabilityGated;
+
+  return applyApplicability(
+      linalg_test::classifyApplicability(QueryResult, Supported, Effective),
+      CaseName);


Nit: simplify

Suggested change

// A device without linear algebra is outside the Tier 1 requirements, and a

// mandatory destination is still unreachable when the source type itself is

// unsupported, so both skip rather than failing.

const bool QueryAnswered = SUCCEEDED(QueryResult);

const bool NoLinearAlgebra = QueryAnswered && !TierSupported;

const bool UnsupportedSource = QueryAnswered && !SourceSupported;

linalg_test::CapabilityRequirement Effective = Requirement;

if (NoLinearAlgebra || UnsupportedSource)

Effective = linalg_test::CapabilityRequirement::CapabilityGated;

return applyApplicability(

linalg_test::classifyApplicability(QueryResult, Supported, Effective),

CaseName);

// A device without linear algebra is outside the Tier 1 requirements, and a

// mandatory destination is unreachable when the source type is unsupported.

// Only successfully queried missing prerequisites should downgrade to a skip.

const bool MissingPrerequisite =

SUCCEEDED(QueryResult) && (!TierSupported || !SourceSupported);

const linalg_test::CapabilityRequirement Effective =

MissingPrerequisite

? linalg_test::CapabilityRequirement::CapabilityGated

: Requirement

Alex Sepkowski (alsepkow) · 2026-09-02T02:46:16Z

                                   bool &Supported) {
+  TierSupported = false;
+  SourceSupported = false;
  Supported = false;


nit:

Suggested change

Supported = false;

DestinationSupported = false;

Alex Sepkowski (alsepkow) · 2026-09-02T02:46:38Z

+                       LPCWSTR CaseName) {
+  bool TierSupported = false;
+  bool SourceSupported = false;
  bool Supported = false;


nit:

Suggested change

bool Supported = false;

bool DestinationSupported = false;

Alex Sepkowski (alsepkow) · 2026-09-02T02:48:14Z

  if (FAILED(HR) || !SourceSupported)
    return HR;

  HR = queryConvertDestinationEnumeration(Device, *DestinationType, Supported);


nit:

Suggested change

HR = queryConvertDestinationEnumeration(Device, *DestinationType, DestinationSupported);

Alex Sepkowski (alsepkow) · 2026-09-02T02:48:32Z

    hlsl_test::LogCommentFmt(
        L"Convert operation enumeration is unavailable; falling back to the "
        L"granular vector-input query");
    HR = queryConvertDestinationGranular(Device, *DestinationType, Supported);


nit:

Suggested change

HR = queryConvertDestinationGranular(Device, *DestinationType, DestinationSupported);

Alex Sepkowski (alsepkow) · 2026-09-02T02:48:43Z

        L"granular vector-input query");
    HR = queryConvertDestinationGranular(Device, *DestinationType, Supported);
  }
  if (SUCCEEDED(HR) && !Supported)


nit:

Suggested change

if (SUCCEEDED(HR) && !DestinationSupported)

Alex Sepkowski (alsepkow) · 2026-09-02T02:49:06Z

-    return false;
+  const HRESULT QueryResult =
+      queryConvertSupport(Device, SourceCompType, DestinationCompType,
+                          TierSupported, SourceSupported, Supported);


nit:

Suggested change

TierSupported, SourceSupported, Supported);

TierSupported, SourceSupported, DestinationSupported);

Alex Sepkowski (alsepkow) · 2026-09-02T02:49:21Z

+    Effective = linalg_test::CapabilityRequirement::CapabilityGated;
+
+  return applyApplicability(
+      linalg_test::classifyApplicability(QueryResult, Supported, Effective),


nit:

Suggested change

linalg_test::classifyApplicability(QueryResult, Supported, Effective),

linalg_test::classifyApplicability(QueryResult, DestinationSupported, Effective),

Alex Sepkowski (alsepkow) · 2026-09-02T02:50:42Z

                                   ComponentType SourceCompType,
                                   ComponentType DestinationCompType,
+                                   bool &TierSupported, bool &SourceSupported,
                                   bool &Supported) {


nit:

Suggested change

bool &Supported) {

bool &DestinationSupported) {

Copilot AI balanced review requested due to automatic review settings September 1, 2026 17:20

github-project-automation Bot added this to HLSL Roadmap Sep 1, 2026

github-project-automation Bot moved this to New in HLSL Roadmap Sep 1, 2026

Copilot started reviewing on behalf of Jack Elliott (JoeCitizen) September 1, 2026 17:20 View session

Copilot AI reviewed Sep 1, 2026

View reviewed changes

Joshua Batista (bob80905) approved these changes Sep 1, 2026

View reviewed changes

Alex Sepkowski (alsepkow) approved these changes Sep 2, 2026

View reviewed changes

Jack Elliott (JoeCitizen) merged commit 9cbbf12 into microsoft:main Sep 2, 2026
15 checks passed

github-project-automation Bot moved this from New to Done in HLSL Roadmap Sep 2, 2026

Navigation Menu

Uh oh!

[HLSL] Require FP8 conversion support in the LinAlg convert gate - #8868

[HLSL] Require FP8 conversion support in the LinAlg convert gate#8868
Jack Elliott (JoeCitizen) merged 1 commit into
microsoft:mainfrom
JoeCitizen:linalg-hlk-fp8-convert-mandatory

Jack Elliott (JoeCitizen) commented Sep 1, 2026 •

edited

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Joshua Batista (bob80905) left a comment

Uh oh!

Alex Sepkowski (alsepkow) left a comment

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Alex Sepkowski (alsepkow) Sep 2, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants


	HR = queryConvertDestinationEnumeration(Device, *DestinationType, DestinationSupported);


	HR = queryConvertDestinationGranular(Device, *DestinationType, DestinationSupported);

	TierSupported, SourceSupported, Supported);
	TierSupported, SourceSupported, DestinationSupported);

	linalg_test::classifyApplicability(QueryResult, Supported, Effective),
	linalg_test::classifyApplicability(QueryResult, DestinationSupported, Effective),

Uh oh!

Conversation

Jack Elliott (JoeCitizen) commented Sep 1, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Joshua Batista (bob80905) left a comment

Choose a reason for hiding this comment

Uh oh!

Alex Sepkowski (alsepkow) 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!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Jack Elliott (JoeCitizen) commented Sep 1, 2026 •

edited

Loading