forked from tensorflow/tfjs
-
Notifications
You must be signed in to change notification settings - Fork 0
/
BatchMatMul.ts
139 lines (113 loc) · 5.27 KB
/
BatchMatMul.ts
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
/**
* @license
* Copyright 2020 Google LLC. All Rights Reserved.
* Licensed under the Apache License, Version 2.0 (the License);
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an AS IS BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
* =============================================================================
*/
import {BatchMatMul, BatchMatMulAttrs, BatchMatMulInputs, broadcast_util, buffer, KernelConfig, KernelFunc, TypedArray, util} from '@tensorflow/tfjs-core';
import {MathBackendCPU} from '../backend_cpu';
import {assertNotComplex} from '../cpu_util';
import {reshape} from './Reshape';
export function batchMatMul(args: {
inputs: BatchMatMulInputs,
attrs: BatchMatMulAttrs,
backend: MathBackendCPU
}) {
const {inputs, backend, attrs} = args;
const {a, b} = inputs;
const {transposeA, transposeB} = attrs;
assertNotComplex([a, b], 'matMul');
const aRank = a.shape.length;
const bRank = b.shape.length;
const innerShapeA = transposeA ? a.shape[aRank - 2] : a.shape[aRank - 1];
const innerShapeB = transposeB ? b.shape[bRank - 1] : b.shape[bRank - 2];
const outerShapeA = transposeA ? a.shape[aRank - 1] : a.shape[aRank - 2];
const outerShapeB = transposeB ? b.shape[bRank - 2] : b.shape[bRank - 1];
const outerDimsA = a.shape.slice(0, -2);
const outerDimsB = b.shape.slice(0, -2);
const batchDimA = util.sizeFromShape(outerDimsA);
const batchDimB = util.sizeFromShape(outerDimsB);
const outShapeOuterDims = broadcast_util.assertAndGetBroadcastShape(
a.shape.slice(0, -2), b.shape.slice(0, -2));
const outShape = outShapeOuterDims.concat([outerShapeA, outerShapeB]);
util.assert(
innerShapeA === innerShapeB,
() => `Error in matMul: inner shapes (${innerShapeA}) and (` +
`${innerShapeB}) of Tensors with shapes ${a.shape} and ` +
`${b.shape} and transposeA=${transposeA}` +
` and transposeB=${transposeB} must match.`);
const a3dShape = transposeA ? [batchDimA, innerShapeA, outerShapeA] :
[batchDimA, outerShapeA, innerShapeA];
const b3dShape = transposeB ? [batchDimB, outerShapeB, innerShapeB] :
[batchDimB, innerShapeB, outerShapeB];
// The rest of the implementation is designed to operate on rank-3 tensors
const a3d = reshape({inputs: {x: a}, backend, attrs: {shape: a3dShape}});
const b3d = reshape({inputs: {x: b}, backend, attrs: {shape: b3dShape}});
const sharedDim = transposeA ? a3d.shape[1] : a3d.shape[2];
const leftDim = transposeA ? a3d.shape[2] : a3d.shape[1];
const rightDim = transposeB ? b3d.shape[1] : b3d.shape[2];
const batchDim = Math.max(batchDimA, batchDimB);
const a3dValues = backend.data.get(a3d.dataId).values as TypedArray;
const b3dValues = backend.data.get(b3d.dataId).values as TypedArray;
const a3dStrides = util.computeStrides(a3d.shape);
const b3dStrides = util.computeStrides(b3d.shape);
const [aBatch, aOuterStep, aInnerStep] = transposeA ?
[a3dStrides[0], 1, a3dStrides[1]] :
[a3dStrides[0], a3dStrides[1], 1];
const [bInnerStep, bOuterStep, bBatch] = transposeB ?
[1, b3dStrides[1], b3dStrides[0]] :
[b3dStrides[1], 1, b3dStrides[0]];
const size = leftDim * rightDim;
const result = buffer([batchDim, leftDim, rightDim], a3d.dtype);
const resVals = result.values as TypedArray;
const blockSize = backend.blockSize;
for (let bi = 0; bi < batchDim; bi++) {
const batchIndexA = bi % batchDimA;
const batchIndexB = bi % batchDimB;
for (let i0 = 0; i0 < leftDim; i0 += blockSize) {
// for when blockSize doesn't evenly divide the input
const iBlock = Math.min(i0 + blockSize, leftDim);
for (let j0 = 0; j0 < rightDim; j0 += blockSize) {
const jBlock = Math.min(j0 + blockSize, rightDim);
for (let k0 = 0; k0 < sharedDim; k0 += blockSize) {
const kBlock = Math.min(k0 + blockSize, sharedDim);
for (let i = i0; i < iBlock; i++) {
for (let j = j0; j < jBlock; j++) {
let sum = 0.0;
for (let k = k0; k < kBlock; k++) {
const aVal =
// tslint:disable-next-line: max-line-length
a3dValues[batchIndexA * aBatch + i * aOuterStep + k * aInnerStep];
const bVal =
// tslint:disable-next-line: max-line-length
b3dValues[k * bInnerStep + j * bOuterStep + batchIndexB * bBatch];
sum += aVal * bVal;
}
resVals[bi * size + (i * rightDim + j)] += sum;
}
}
}
}
}
}
backend.disposeIntermediateTensorInfo(a3d);
backend.disposeIntermediateTensorInfo(b3d);
// set correct shape on output.
return backend.makeTensorInfo(
outShape, result.dtype, result.values as TypedArray);
}
export const batchMatMulConfig: KernelConfig = {
kernelName: BatchMatMul,
backendName: 'cpu',
kernelFunc: batchMatMul as unknown as KernelFunc,
};