Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[webnn] Add float32 tests for WebNN hardSwish op #38712

Merged
merged 2 commits into from
Mar 1, 2023
Merged

[webnn] Add float32 tests for WebNN hardSwish op #38712

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
31cf5ea
[webnn] Add float32 tests for WebNN hardSwish op
BruceDai Feb 27, 2023
d29a6df
[webnn] Optimize assert_array_approx_equals_ulp method by firstly check
BruceDai Feb 28, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
10 changes: 10 additions & 0 deletions webnn/hard_swish.https.any.js
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,10 @@
// META: title=test WebNN API tanh operation
// META: global=window,dedicatedworker
// META: script=./resources/utils.js
// META: timeout=long

'use strict';

// https://webmachinelearning.github.io/webnn/#api-mlgraphbuilder-hard-swish

testWebNNOperation('hardSwish', buildOperationWithSingleInput);
334 changes: 334 additions & 0 deletions webnn/resources/test_data/hard_swish.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,334 @@
{
"tests": [
{
"name": "hardSwish float32 1D tensor",
"inputs": {
"x": {
"shape": [24],
"data": [
0.7341583533045579,
9.118859151005996,
3.545238531520827,
2.621943879280181,
-6.445507690595167,
-1.6835596550754381,
5.523179785756591,
-5.958856051028132,
-9.169189933081544,
6.420943542920213,
-3.293031330275471,
1.0410166785810624,
-7.246322671816956,
-0.9472730969847909,
-5.778352255817807,
3.185229125228698,
-7.261818072290236,
4.174602615173795,
3.7802628241590686,
-6.07124038718255,
-9.909919471919547,
-7.744259390113584,
-8.286120816748381,
8.083491160956697
],
"type": "float32"
}
},
"expected": {
"name": "output",
"shape": [24],
"data": [
0.4569105803966522,
9.11885929107666,
3.545238494873047,
2.4567370414733887,
0,
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Fail to run these hardSwish tests by Chromium integrated WebNN API with XNNPack API back, since XNNPack API can distinguish 0.0 and -0.0.
@fdwr @huningxin Any suggestion? Thanks.

Copy link

@fdwr fdwr Feb 27, 2023 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Signed and unsigned zero can be treated as equivalent (IEEE does https://en.wikipedia.org/wiki/Signed_zero#Comparisons). I recommend that before checking for ULP differences in assert_array_approx_equals_ulp, first check if expected == actual, and only if !=, would you need to check ULP too.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@fdwr Thanks for your suggestion! Yes, if actual was as same as expect one, no need to measure the ULP distance which would save testing time. Please take another look, thanks.

-0.3693843185901642,
5.52318000793457,
0,
0,
6.420943737030029,
0,
0.7011276483535767,
0,
-0.3240821659564972,
0,
3.1852290630340576,
0,
4.174602508544922,
3.7802627086639404,
0,
0,
0,
0,
8.083491325378418
],
"type": "float32"
}
},
{
"name": "hardSwish float32 2D tensor",
"inputs": {
"x": {
"shape": [4, 6],
"data": [
0.7341583533045579,
9.118859151005996,
3.545238531520827,
2.621943879280181,
-6.445507690595167,
-1.6835596550754381,
5.523179785756591,
-5.958856051028132,
-9.169189933081544,
6.420943542920213,
-3.293031330275471,
1.0410166785810624,
-7.246322671816956,
-0.9472730969847909,
-5.778352255817807,
3.185229125228698,
-7.261818072290236,
4.174602615173795,
3.7802628241590686,
-6.07124038718255,
-9.909919471919547,
-7.744259390113584,
-8.286120816748381,
8.083491160956697
],
"type": "float32"
}
},
"expected": {
"name": "output",
"shape": [4, 6],
"data": [
0.4569105803966522,
9.11885929107666,
3.545238494873047,
2.4567370414733887,
0,
-0.3693843185901642,
5.52318000793457,
0,
0,
6.420943737030029,
0,
0.7011276483535767,
0,
-0.3240821659564972,
0,
3.1852290630340576,
0,
4.174602508544922,
3.7802627086639404,
0,
0,
0,
0,
8.083491325378418
],
"type": "float32"
}
},
{
"name": "hardSwish float32 3D tensor",
"inputs": {
"x": {
"shape": [2, 3, 4],
"data": [
0.7341583533045579,
9.118859151005996,
3.545238531520827,
2.621943879280181,
-6.445507690595167,
-1.6835596550754381,
5.523179785756591,
-5.958856051028132,
-9.169189933081544,
6.420943542920213,
-3.293031330275471,
1.0410166785810624,
-7.246322671816956,
-0.9472730969847909,
-5.778352255817807,
3.185229125228698,
-7.261818072290236,
4.174602615173795,
3.7802628241590686,
-6.07124038718255,
-9.909919471919547,
-7.744259390113584,
-8.286120816748381,
8.083491160956697
],
"type": "float32"
}
},
"expected": {
"name": "output",
"shape": [2, 3, 4],
"data": [
0.4569105803966522,
9.11885929107666,
3.545238494873047,
2.4567370414733887,
0,
-0.3693843185901642,
5.52318000793457,
0,
0,
6.420943737030029,
0,
0.7011276483535767,
0,
-0.3240821659564972,
0,
3.1852290630340576,
0,
4.174602508544922,
3.7802627086639404,
0,
0,
0,
0,
8.083491325378418
],
"type": "float32"
}
},
{
"name": "hardSwish float32 4D tensor",
"inputs": {
"x": {
"shape": [2, 2, 2, 3],
"data": [
0.7341583533045579,
9.118859151005996,
3.545238531520827,
2.621943879280181,
-6.445507690595167,
-1.6835596550754381,
5.523179785756591,
-5.958856051028132,
-9.169189933081544,
6.420943542920213,
-3.293031330275471,
1.0410166785810624,
-7.246322671816956,
-0.9472730969847909,
-5.778352255817807,
3.185229125228698,
-7.261818072290236,
4.174602615173795,
3.7802628241590686,
-6.07124038718255,
-9.909919471919547,
-7.744259390113584,
-8.286120816748381,
8.083491160956697
],
"type": "float32"
}
},
"expected": {
"name": "output",
"shape": [2, 2, 2, 3],
"data": [
0.4569105803966522,
9.11885929107666,
3.545238494873047,
2.4567370414733887,
0,
-0.3693843185901642,
5.52318000793457,
0,
0,
6.420943737030029,
0,
0.7011276483535767,
0,
-0.3240821659564972,
0,
3.1852290630340576,
0,
4.174602508544922,
3.7802627086639404,
0,
0,
0,
0,
8.083491325378418
],
"type": "float32"
}
},
{
"name": "hardSwish float32 5D tensor",
"inputs": {
"x": {
"shape": [2, 1, 4, 1, 3],
"data": [
0.7341583533045579,
9.118859151005996,
3.545238531520827,
2.621943879280181,
-6.445507690595167,
-1.6835596550754381,
5.523179785756591,
-5.958856051028132,
-9.169189933081544,
6.420943542920213,
-3.293031330275471,
1.0410166785810624,
-7.246322671816956,
-0.9472730969847909,
-5.778352255817807,
3.185229125228698,
-7.261818072290236,
4.174602615173795,
3.7802628241590686,
-6.07124038718255,
-9.909919471919547,
-7.744259390113584,
-8.286120816748381,
8.083491160956697
],
"type": "float32"
}
},
"expected": {
"name": "output",
"shape": [2, 1, 4, 1, 3],
"data": [
0.4569105803966522,
9.11885929107666,
3.545238494873047,
2.4567370414733887,
0,
-0.3693843185901642,
5.52318000793457,
0,
0,
6.420943737030029,
0,
0.7011276483535767,
0,
-0.3240821659564972,
0,
3.1852290630340576,
0,
4.174602508544922,
3.7802627086639404,
0,
0,
0,
0,
8.083491325378418
],
"type": "float32"
}
}
]
}
18 changes: 12 additions & 6 deletions webnn/resources/utils.js
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -268,6 +268,7 @@ const PrecisionMetrics = {
sin: {ATOL: {float32: 1/1024, float16: 1/512}},
tan: {ATOL: {float32: 1/1024, float16: 1/512}},
// End Element-wise unary operations
hardSwish: {ULP: {float32: 4, float16: 4}},
gemm: {ULP: {float32: getGemmPrecisionTolerance, float16: getGemmPrecisionTolerance}},
leakyRelu: {ULP: {float32: 1, float16: 1}},
matmul: {ULP: {float32: getMatmulPrecisionTolerance, float16: getMatmulPrecisionTolerance}},
Expand Down Expand Up @@ -353,12 +354,17 @@ const assert_array_approx_equals_ulp = (actual, expected, nulp, dataType, descri
`assert_array_approx_equals_ulp: ${description} lengths differ, expected ${expected.length} but got ${actual.length}`);
let actualBitwise, expectedBitwise, distance;
for (let i = 0; i < actual.length; i++) {
actualBitwise = getBitwise(actual[i], dataType);
expectedBitwise = getBitwise(expected[i], dataType);
distance = actualBitwise - expectedBitwise;
distance = distance >= 0 ? distance : -distance;
assert_true(distance <= nulp,
`assert_array_approx_equals_ulp: ${description} actual ${actual[i]} should be close enough to expected ${expected[i]} by the acceptable ${nulp} ULP distance, but they have ${distance} ULP distance`);
if (actual[i] === expected[i]) {
continue;
} else {
// measure the ULP distance
actualBitwise = getBitwise(actual[i], dataType);
expectedBitwise = getBitwise(expected[i], dataType);
distance = actualBitwise - expectedBitwise;
distance = distance >= 0 ? distance : -distance;
assert_true(distance <= nulp,
`assert_array_approx_equals_ulp: ${description} actual ${actual[i]} should be close enough to expected ${expected[i]} by the acceptable ${nulp} ULP distance, but they have ${distance} ULP distance`);
}
}
};

Expand Down