Backend coverage matrix (T1.5)

May 19, 2026 · View on GitHub

Combined snapshot of T1.2 (device.Backend inventory), T1.3 (CPU per-domain ISA dispatch), and T1.4 (Metal / CUDA / XLA kernel registration). Counts are registration status from filesystem and kernels.Default; they do not assert full R1 implementation, parity, or legality on every required operation.

Machine-checkable source: pkg/backend/device/coverageaudit/, validated by coverageaudit_test.go.

Detail documents

TaskDocumentPackage
T1.2backend-inventory.mdpkg/backend/device/inventory*.go
T1.3cpu-dispatch-matrix.mdpkg/backend/device/cpu/dispatchaudit/
T1.4device-backend-matrix.mdpkg/backend/device/backendaudit/

R1 execution targets (registration snapshot)

Equal standing per AGENTS.md §1. Applicable is the audit denominator for that row; registered is what the combined audit currently finds.

TargetScopeRegisteredApplicableNotes
Go scalarCPU domains3030Go reference in every operation domain
AVX-512CPU domains (amd64)230Per-domain assembly/dispatch registration
AVX2CPU domains (amd64)230Per-domain assembly/dispatch registration
SSE2CPU domains (amd64)230Per-domain assembly/dispatch registration
NEONCPU domains (arm64)2030Per-domain assembly/dispatch registration
MetalRequired IR operations68119462 kernels.Default registrations total
CUDARequired IR operations0119Tensor upload/download; 0 kernel registrations
XLARequired IR operations0119Tensor upload/download; 0 kernel registrations

Backend inventory (T1.2)

ItemCount
device.Backend methods151
ir.RequiredOperationIDs()119
Required ops → direct Backend method74
Required ops → composite Backend methods3
Required ops → kernel registry32
Required ops → graph-only10
Backend methods with no required-op ID88

CPU dispatch (T1.3)

ISA pathDomains registered
Scalar (Go)30 / 30
AVX-512 (amd64)8 / 30
AVX2 (amd64)2 / 30
SSE2 (amd64)2 / 30
NEON (arm64)20 / 30

AMD64 AVX-512 registered on: activation, dot, dropout, elementwise, matmul, pool, pospop, reduction.

Per-domain table: cpu-dispatch-matrix.md.

Device backends (T1.4)

BackendKernel registrationsRequired ops with Metal kernel
Metal462 (158 unique names)68 / 119
CUDA0
XLA0

Per-kernel and per-operation tables: device-backend-matrix.md.