Merge pull request #63 from mei1127/add_cast

Implement cast

.eslintrc.js

@@ -7,6 +7,7 @@ module.exports = {
     'chai': 'readonly',
     'BigInt': 'readonly',
     'BigInt64Array': 'readonly',
+    'BigUint64Array': 'readonly',
   },
   rules: {
     'semi': 'error',

src/cast.js

@@ -0,0 +1,46 @@
+'use strict';
+
+import {Tensor} from '../src/lib/tensor.js';
+
+/**
+ * Cast each element in the input tensor to the target data type.
+ * @param {Tensor} input
+ * @return {Tensor}
+ */
+
+export function cast(input, type) {
+  let outputArray;
+  switch (type) {
+    case 'int8':
+      outputArray = new Int8Array(input.data);
+      break;
+    case 'uint8':
+      outputArray = new Uint8Array(input.data);
+      break;
+    case 'int32':
+      outputArray = new Int32Array(input.data);
+      break;
+    case 'uint32':
+      outputArray = new Uint32Array(input.data);
+      break;
+    case 'int64':
+      outputArray = new BigInt64Array(Array.from(input.data, (num) => BigInt(Math.trunc(num))));
+      break;
+    case 'float32':
+      outputArray = new Float32Array(input.data);
+      break;
+    case 'float64':
+      outputArray = new Float64Array(input.data);
+      break;
+    case 'float16':
+      // TODO: https://github.com/webmachinelearning/webnn-baseline/issues/66
+      throw new Error('Unsupported output type: float16');
+    case 'uint64':
+      // TODO: https://github.com/webmachinelearning/webnn-baseline/issues/67
+      throw new Error('Unsupported output type: uint64');
+    default:
+      throw new Error('Unsupported output type: ' + type);
+  }
+  const output = new Tensor(input.shape, outputArray);
+  return output;
+}

test/cast_test.js

@@ -0,0 +1,175 @@
+'use strict';
+
+import {Tensor} from '../src/lib/tensor.js';
+import {cast} from '../src/cast.js';
+import * as utils from './utils.js';
+
+describe('test cast', function() {
+  const InputDataType = {
+    int8: Int8Array,
+    uint8: Uint8Array,
+    int32: Int32Array,
+    uint32: Uint32Array,
+    int64: BigInt64Array,
+    float32: Float32Array,
+    float64: Float64Array,
+  };
+  function testCast(input, type, expected) {
+    let tensorInput;
+    if (input.type) {
+      tensorInput = new Tensor(input.shape, new InputDataType[input.type](input.data));
+    } else {
+      tensorInput = new Tensor(input.shape, input.data);
+    }
+    const outputTensor = cast(tensorInput, type);
+    utils.checkShape(outputTensor, expected.shape);
+    utils.checkValue(outputTensor, expected.data);
+  }
+
+  it('cast float64 to int8', function() {
+    const input = {
+      shape: [5],
+      data: [
+        -0.25, 0.25, 3.75, 14, -14,
+      ],
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        0, 0, 3, 14, -14,
+      ],
+    };
+    testCast(input, 'int8', expected);
+  });
+
+  it('cast float64 to uint8', function() {
+    const input = {
+      shape: [5],
+      data: [
+        0.25, 0.75, 3.75, 14, 15,
+      ],
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        0, 0, 3, 14, 15,
+      ],
+    };
+    testCast(input, 'uint8', expected);
+  });
+
+  it('cast float64 to int32', function() {
+    const input = {
+      shape: [5],
+      data: [
+        -0.25, 0.75, 3.21, 1234, -1234,
+      ],
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        0, 0, 3, 1234, -1234,
+      ],
+    };
+    testCast(input, 'int32', expected);
+  });
+
+  it('cast float64 to uint32', function() {
+    const input = {
+      shape: [5],
+      data: [
+        0.75, 0.25, 3.21, 14, 15,
+      ],
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        0, 0, 3, 14, 15,
+      ],
+    };
+    testCast(input, 'uint32', expected);
+  });
+
+  it('cast float64 to int64', function() {
+    const input = {
+      shape: [5],
+      data: [
+        -0.25, 0.75, 3.21, 1234, -1234,
+      ],
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        0n, 0n, 3n, 1234n, -1234n,
+      ],
+    };
+    testCast(input, 'int64', expected);
+  });
+
+  it('cast float64 to float32', function() {
+    const input = {
+      shape: [5],
+      data: [
+        -0.25, 0.25, 3.21, 1234, -1234,
+      ],
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        -0.25, 0.25, 3.2100000381469727, 1234, -1234,
+      ],
+    };
+    testCast(input, 'float32', expected);
+  });
+
+  it('cast int32 to float32', function() {
+    const input = {
+      shape: [5],
+      data: [
+        0, 1, -2, -3, 3,
+      ],
+      type: 'int32',
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        0, 1, -2, -3, 3,
+      ],
+    };
+    testCast(input, 'float32', expected);
+  });
+
+  it('cast uint32 to float64', function() {
+    const input = {
+      shape: [5],
+      data: [
+        0, 1, 22, 33, 33,
+      ],
+      type: 'uint32',
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        0, 1, 22, 33, 33,
+      ],
+    };
+    testCast(input, 'float64', expected);
+  });
+
+  it('cast float32 to float64', function() {
+    const input = {
+      shape: [5],
+      data: [
+        0, 0.1, 0.2, -300, 993,
+      ],
+      type: 'float32',
+    };
+    const expected = {
+      shape: [5],
+      data: [
+        0, 0.10000000149011612, 0.20000000298023224, -300, 993,
+      ],
+    };
+    testCast(input, 'float64', expected);
+  });
+});

test/utils.js

@@ -27,11 +27,17 @@ function getBitwise(value) {
  *     false: The distance between a and b is far away from the given ULP distance.
  */
 assert.isAlmostEqualUlp = function(a, b, nulp, message) {
-  const aBitwise = getBitwise(a);
-  const bBitwise = getBitwise(b);
-  let distance = aBitwise - bBitwise;
-  distance = distance >= 0 ? distance : -distance;
-  return assert.isTrue(distance <= nulp, message);
+  if (typeof(a) == 'number') {
+    const aBitwise = getBitwise(a);
+    const bBitwise = getBitwise(b);
+    let distance = aBitwise - bBitwise;
+    distance = distance >= 0 ? distance : -distance;
+    return assert.isTrue(distance <= nulp, message);
+  } else {
+    let distance = a - b;
+    distance = distance >= 0n ? distance : -distance;
+    return assert.isTrue(distance <= nulp, message);
+  }
 };
 
 export function checkValue(tensor, expected, nulp = 0) {