[webnn] Add float32 tests for Element-wise logical operations #43304
Conversation
| @@ -277,6 +277,14 @@ const PrecisionMetrics = { | |||
| min: {ULP: {float32: 0, float16: 0}}, | |||
| pow: {ULP: {float32: 32, float16: 2}}, | |||
| // End Element-wise binary operations | |||
| // Begin Element-wise logical operations | |||
| equal: {ULP: {uint8: 0}}, | |||
There was a problem hiding this comment.
For logical operations, no mater input data type is float32 or float16 and etc. , the output type should be uint8.
And here uint8 is align with expected data type in json file, likes
"expected": {
"name": "output",
"shape": [2, 1, 4, 1, 3],
"data": [
0,
0,
0,
1,
0,
...
0
],
"type": "uint8"
}
We also use expected data type (precisionType) as a key to get tolerance:
// L#347
let tolerance = PrecisionMetrics[operationName][metricType][precisionType];@fdwr Is it OK for this addition code? Thanks.
| @@ -277,6 +277,14 @@ const PrecisionMetrics = { | |||
| min: {ULP: {float32: 0, float16: 0}}, | |||
| pow: {ULP: {float32: 32, float16: 2}}, | |||
| // End Element-wise binary operations | |||
| // Begin Element-wise logical operations | |||
| equal: {ULP: {uint8: 0}}, | |||
|
Hi @fdwr, I synced with @huningxin, we decided to rename not op tests to previous logicalNot op tests, since Chromium WebNN API implementation has already supported logicalNot op, then these WPT logicalNot tests can be ran, please take another look at this update, thanks. |
|
@BruceDai : Ok, well whatever the final outcome of the spec is, it's a simple rename either way. 👍 |
equal / greater / greaterOrEqual
lesser / lesserOrEqual / not
BruceDai
force-pushed
the
add_webnn_logical_tests
branch
from
75c0395
to
80d3ab8
Compare
|
Rebase code to check whether last two fail CIs can pass, then go to ask for merging it. |
@fdwr @Honry PTAL, thanks.