Module bfp_experiment 

AMD bfp16ebs8 vendor experiment (block-8): the XBFP artifact container.

This is a backend-local experiment, not a TOSA capability: it is never advertised through TosaCapabilityProvider, never accepted under the stable TOSA artifact format, and creates no protocol value. Design and boundaries: docs/plans/issue-148-bfp16ebs8-vendor-tier.md; the silicon-characterized numerical contract it exposes: docs/research/amdxdna-bfp16ebs8-characterization.md (issue #146).

Flavor 1 is a block-scaled MATMUL with MXINT8 semantics executed on the proven block-8 decomposition: C[8,N=8] (FP32) = A[8,K] · B[8,K]ᵀ, K ∈ {32, 64, …, 512}. Operand planes are streams of 72-byte v64bfp16ebs8 units (64 two’s-complement int8 mantissas, then 8 exponent bytes); MXINT8 semantics require the four exponent bytes of each 32-group to be equal, and e = 255 (the hardware’s structural Inf/NaN space) is outside the contract. Accumulation is FP32 in ascending-k chain order — the guest-visible oracle is an FP32 fold in exactly that order, proven bit-exact on the reference NPU.

Structs

BfpExperimentArtifact

A parsed, envelope-validated XBFP container.

Constants

UNIT_BYTES

One 72-byte v64bfp16ebs8 unit: 64 mantissa bytes then 8 exponent bytes.

XDNA_BFP_EXPERIMENT_FORMAT

The experiment’s artifact format word ("XBFP" read as a big-endian u32). Deliberately distinct from crate::XDNA_PRECOMPILED_FORMAT and from released TOSA.

XDNA_BFP_EXPERIMENT_TARGET_IDENTITY

The experiment’s own target identity. A load must present exactly this identity: the numerical label is immutable, and no TOSA target may alias it.

Functions

encode

Build a flavor-1 XBFP container. Offline tooling and tests only; the serving path never encodes.