Evaluating Metal acceleration and CPU fallback

Start with the intended workflow

Assess whether a target Apple workstation can use the optional Metal DOT path.

ABI reports platform availability, whether native kernels are linked, and whether a Metal pipeline initialized. These fields have different meanings. If the native path is unavailable, the vector-operation implementation retains a CPU fallback with accelerated=false.

Local entry points

Run from the ABI checkout with the CLI built and available on PATH. Use disposable data for evaluation; do not connect a test to a live memory store.

abi backends
abi wdbx gpu info

Acceptance and limits

Inspect backend status on the intended machine and compare native and fallback results. Publish latency or throughput only with a reproducible harness, workload, hardware and methodology.

CUDA and Vulkan dispatch are not linked in the inspected ABI GPU implementation.

A compiled kernel is not proof of acceleration on a particular machine or workload.

Interpret the report correctly

available describes platform availability, native_kernels describes linkage and accelerated describes an initialized native path. On macOS, Metal can be the preferred backend while accelerated remains false. CUDA and Vulkan names are catalog entries in this implementation, not working native dispatch. Preserve the reported fallback message in an evaluation log so results from different machines are not silently combined.

Check native and fallback parity

Run the crate’s vector-operation tests from the ABI checkout using its pinned toolchain wrapper. The following command runs local tests and redirects stdin; it does not publish benchmark results. Review skips and backend initialization failures explicitly. For performance work, add a separately reviewed harness that declares vector dimensions, repetition count and timing boundary.

./tools/cargo.sh test -p abi-gpu < /dev/null

Acceptance for an integration

Use equal-length, empty and zero-norm vectors as appropriate to the operation, plus lengths that exercise a remainder beyond a SIMD block. Compare results under a declared numeric tolerance; matching output is a correctness observation, not a speedup. Check the exact caller path because backend status alone does not prove that a particular operation dispatched to the GPU. Record the source revision, machine and backend report with any local result. No hardware-specific speed or power outcome is supplied by this guide.