BruceDai
aeb7574c54
Automatic update from web-platform-tests webnn: generate required data types and ranks JSON from the Spec Previously, the per-operator minimum data type / rank limits used by the WebNN WPT conformance harness were hand-maintained in `minimum_datatype_set.json`, this CL replaces the hand-written file with a generated one and adds the tooling to regenerate it: - Add `tools/generate-required-datatypes-ranks.py`, which fetches `index.bs` from the WebNN spec (the `minimum_datatype_rankrange` branch) and parses every `tensor-limits-*` table to extract each operand's required data types and rank range, then writes `resources/required_datatypes_ranks.json`. - Rename `minimum_datatype_set.json` -> `required_datatypes_ranks.json` and regenerate its contents from the spec. The new file covers 95 operators / 184 operands. - In `resources/utils.js`, rename `minimumDataTypeSet` -> `requiredDataTypesAndRanks` and `getMinimumDataTypeSetJson` -> `getRequiredDataTypesAndRanks`, and point the fetch at the new resource path. - In `conformance_tests/is_infinite.https.any.js` and `is_nan.https.any.js`, update the operator argument key from `'input'` to `'a'` to match the spec's actual parameter name for these element-wise unary ops. Bug: 434735286 Change-Id: Ib5b83c568ae34d4d8bb7c0df52f90f6f7565ef4d Cq-Include-Trybots: luci.chromium.try:win11-blink-rel, mac14.arm64-blink-rel, mac14-blink-rel, mac15.arm64-blink-rel, mac15-blink-rel, linux-blink-rel Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/7936012 Commit-Queue: Hu, Ningxin <ningxin.hu@intel.com> Reviewed-by: Reilly Grant <reillyg@chromium.org> Reviewed-by: Hu, Ningxin <ningxin.hu@intel.com> Cr-Commit-Position: refs/heads/main@{#1654974} -- wpt-commits: 6237d26f48fa1059a5fe6855a92b9343cbc05021 wpt-pr: 61003
Web Neural Network API Test Suite
Spec: https://www.w3.org/TR/webnn/
Repo: https://github.com/webmachinelearning/webnn
Web IDL Tests
The idlharness.https.any.js test validates that the API implements
the interfaces as defined in the Web IDL in the specification. A
snapshot of the spec's Web IDL resides in ../interfaces/webnn.idl
and is automatically daily by shared WPT infrastructure.
Validation Tests
The tests in validation_tests go beyond what can be validated from
the Web IDL definitions in the spec. They ensure that parameters to
the various WebNN methods are validated according to the algorithms in
the spec, and that invalid inputs are rejected as specified. For
example, validation_tests/matmul.https.any.js verifies that an
exception is thrown if the rank of either of the input tensors is less
than 2, or if the input tensor shapes are not broadcastable, etc.
Conformance Tests
The tests in conformance_tests exercise the various WebNN operators
and ensure that they are behaving as expected. For example,
conformance_tests/matmul.https.any.js verifies that multiplying
N-dimensional matrices actually produces the expected output.
Variations
WebNN supports execution on various compute devices, e.g. CPU, GPU and
NPU. The various tests make use of META: variant=?... to enable the
test harness to execute the same test multiple times, with parameters
specifying each device type in turn.
NOTE: The specific API for device selection is currently under debate. See: https://github.com/webmachinelearning/webnn/blob/main/device-selection-explainer.md
Globals
WebNN interfaces are exposed to both Window and Worker global contexts
via the navigator.ml member. These tests make use of the .any.js
convention and META: global=window,dedicatedworker to allow the test
harness to execute the same test multiple times, once for each global
context type.
Since executing a large number of tests consumes significant compute
cycles, and since unexpected behavior differences across contexts
(e.g. window vs. dedicatedworker) for individual WebNN operators
(e.g. matmul()) is unlikely, most operator-specific tests only
specify global=window.