WebNN: Add IDL and mojo definitions for scatterND operator

The `scatterND` operator is proposed by WebML WG [1] for improving
performance of passing partial MLBuffer/MLTensor between transformers
decoder iterations for key-value reuse.

This CL also implements data type limits, inputs validation and adds
validation tests for `scatterND` operator.

[1]: webmachinelearning/webnn#375 (comment)

Bug: 363677531
Change-Id: Ib68db63c7b51b99f9976b1eb3c06f8b7f0de9f97
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/+/5829103
Reviewed-by: Alex Gough <[email protected]>
Reviewed-by: Weizhong Xia <[email protected]>
Reviewed-by: Reilly Grant <[email protected]>
Commit-Queue: ningxin hu <[email protected]>
Cr-Commit-Position: refs/heads/main@{#1353781}

webnn/validation_tests/scatterND.https.any.js

@@ -0,0 +1,102 @@
+// META: title=validation tests for WebNN API scatterND operation
+// META: global=window,dedicatedworker
+// META: variant=?cpu
+// META: variant=?gpu
+// META: variant=?npu
+// META: script=../resources/utils_validation.js
+
+'use strict';
+
+const tests = [
+  {
+    name: '[scatterND] Test scatterND with valid tensors',
+    input: {dataType: 'float32', dimensions: [4, 4, 4]},
+    indices: {dataType: 'int32', dimensions: [2, 1]},
+    updates: {dataType: 'float32', dimensions: [2, 4, 4]},
+    output: {dataType: 'float32', dimensions: [4, 4, 4]}
+  },
+  {
+    name:
+        '[scatterND] Throw if updates tensor data type is not the same as input data type',
+    input: {dataType: 'float32', dimensions: [4, 4, 4]},
+    indices: {dataType: 'int32', dimensions: [2, 1]},
+    updates: {dataType: 'float16', dimensions: [2, 4, 4]},
+  },
+  {
+    name: '[scatterND] Throw if input is a scalar',
+    input: {dataType: 'float32', dimensions: []},
+    indices: {dataType: 'int32', dimensions: [2, 1]},
+    updates: {dataType: 'float32', dimensions: [2, 4, 4]},
+  },
+  {
+    name: '[scatterND] Throw if indices is a scalar',
+    input: {dataType: 'float32', dimensions: [4, 4, 4]},
+    indices: {dataType: 'int32', dimensions: []},
+    updates: {dataType: 'float32', dimensions: [2, 4, 4]},
+  },
+  {
+    name:
+        '[scatterND] Throw if the size of last dimension of indices tensor is greater than input rank',
+    input: {dataType: 'float32', dimensions: [4, 4, 4]},
+    indices: {dataType: 'int32', dimensions: [2, 4]},
+    updates: {dataType: 'float32', dimensions: [2, 4, 4]},
+  },
+  {
+    name: '[scatterND] Throw if updates tensor shape is invalid.',
+    input: {dataType: 'float32', dimensions: [4, 4, 4]},
+    indices: {dataType: 'int32', dimensions: [2, 1]},
+    // Updates tensor shape should be [2, 4, 4].
+    updates: {dataType: 'float32', dimensions: [2, 3, 4]},
+  }
+];
+
+tests.forEach(test => promise_test(async t => {
+                const builder = new MLGraphBuilder(context);
+                const input = builder.input('input', test.input);
+                const indices = builder.input('indices', test.indices);
+                const updates = builder.input('updates', test.updates);
+
+                if (test.output) {
+                  const output = builder.scatterND(input, indices, updates);
+                  assert_equals(output.dataType(), test.output.dataType);
+                  assert_array_equals(output.shape(), test.output.dimensions);
+                } else {
+                  const label = 'a_scatter_nd'
+                  const options = {label};
+                  const regexp = new RegExp('\\[' + label + '\\]');
+                  assert_throws_with_label(
+                      () => builder.scatterND(input, indices, updates, options),
+                      regexp);
+                }
+              }, test.name));
+
+multi_builder_test(async (t, builder, otherBuilder) => {
+  const input =
+      otherBuilder.input('input', {dataType: 'float32', dimensions: [8]});
+  const indices =
+      builder.input('indices', {dataType: 'int32', dimensions: [4, 1]});
+  const updates =
+      builder.input('indices', {dataType: 'int32', dimensions: [4]});
+
+  assert_throws_js(TypeError, () => builder.scatterND(input, indices, updates));
+}, '[scatterND] Throw if input is from another builder');
+
+multi_builder_test(async (t, builder, otherBuilder) => {
+  const input = builder.input('input', {dataType: 'float32', dimensions: [8]});
+  const indices =
+      otherBuilder.input('indices', {dataType: 'int32', dimensions: [4, 1]});
+  const updates =
+      builder.input('indices', {dataType: 'int32', dimensions: [4]});
+
+  assert_throws_js(TypeError, () => builder.scatterND(input, indices, updates));
+}, '[scatterND] Throw if indcies is from another builder');
+
+multi_builder_test(async (t, builder, otherBuilder) => {
+  const input = builder.input('input', {dataType: 'float32', dimensions: [8]});
+  const indices =
+      builder.input('indices', {dataType: 'int32', dimensions: [4, 1]});
+  const updates =
+      otherBuilder.input('indices', {dataType: 'int32', dimensions: [4]});
+
+  assert_throws_js(TypeError, () => builder.scatterND(input, indices, updates));
+}, '[scatterND] Throw if updates is from another builder');