[js/web] adding conv fuse logic (microsoft#7604)

* adding conv fuse logic

* fixing merge

* fix file name in kebab case

* fix lint error

Author: xzhu1900

web/lib/onnxjs/backends/webgl/ops/conv-pack.ts

@@ -1,6 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 
+import {Attribute} from '../../../attribute';
 import {Logger} from '../../../instrument';
 import {Conv} from '../../../ops/conv';
 import {Tensor} from '../../../tensor';
@@ -36,6 +37,11 @@ export class WebGLConvPacked extends Conv {
     const outputShape = WebGLConv.calcOutputShape(xshape, kshape, this.dilations, this.pads, this.strides);
     const im2col = new WebGLIm2ColPacked(outputShape, kshape, this.dilations, this.pads, this.strides);
     const matmul = new WebGLMatMulPacked();
+    if (this.activation) {
+      const attributes = new Attribute(undefined);
+      attributes.set('__internal_activation', 'string', (this.activation));
+      matmul.initialize(attributes);
+    }
     const reshape = new WebGLReshapePacked();
     // shape for kernel reshape
     const shape =

web/lib/onnxjs/backends/webgl/ops/conv.ts

@@ -10,7 +10,9 @@ import {getGlsl} from '../glsl-source';
 import {WebGLInferenceHandler} from '../inference-handler';
 import {Artifact, ProgramInfo, RunData, TextureLayout, WebGLOperator} from '../types';
 import {WebGLContext} from '../webgl-context';
+
 import {WebGLConvPacked} from './conv-pack';
+import {getActicationSnippet} from './fuse-utils';
 
 export class WebGLConv extends Conv {
   unpackedGroupedConvImpl: WebGLUnpackedGroupedConv;
@@ -66,7 +68,7 @@ export class WebGLUnpackedGroupedConv extends Conv implements WebGLOperator {
 
   createProgramInfo(handler: WebGLInferenceHandler, inputs: Tensor[]): ProgramInfo {
     const hasBias = inputs.length > 2;
-    const processBias = hasBias ? 'dotProd += getBias(output_channel);' : '';
+    const processBias = hasBias ? 'value += getBias(output_channel);' : '';
     const xShape = inputs[0].dims.slice();
     const wShape = inputs[1].dims.slice();
     const outputChannelsPerGroup = wShape[0] / this.group;
@@ -85,18 +87,20 @@ export class WebGLUnpackedGroupedConv extends Conv implements WebGLOperator {
     const outputShape = WebGLConv.calcOutputShape(xShape, wShape, this.dilations, this.pads, this.strides);
     const glsl = getGlsl(handler.session.backend.glContext.version);
 
+    const {activationFunction, applyActivation} = getActicationSnippet(this.activation);
+
     const shaderSource = `
     const ivec2 strides = ivec2(${this.strides[0]}, ${this.strides[1]});
     const ivec2 pads = ivec2(${this.pads[0]}, ${this.pads[1]});
-
+    ${activationFunction}
     void main() {
       ivec4 coords = getOutputCoords();
       int batch = coords.x;
       int output_channel = coords.y;
       ivec2 xRCCorner = coords.zw * strides - pads;
       int group_id = output_channel / ${outputChannelsPerGroup};
 
-      float dotProd = 0.0;
+      float value = 0.0;
       for (int wInChannel = 0; wInChannel < ${wShape[1]}; wInChannel++) {
         int input_channel = group_id * ${wShape[1]} + wInChannel;
         for (int wHeight = 0; wHeight < ${wShape[2]}; wHeight++) {
@@ -114,12 +118,13 @@ export class WebGLUnpackedGroupedConv extends Conv implements WebGLOperator {
 
             float xVal = getX(batch, input_channel, xWidth, xHeight);
             float wVal = getW(output_channel, wInChannel, wWidth, wHeight);
-            dotProd += xVal*wVal;
+            value += xVal*wVal;
           }
         }
       }
       ${processBias}
-      ${glsl.output} = vec4(dotProd, .0, .0, .0);
+      ${applyActivation}
+      ${glsl.output} = vec4(value, .0, .0, .0);
     }
 `;
     return {
@@ -215,7 +220,6 @@ export class WebGLUnpackedConv extends Conv {
         let blend = false;
         for (let k = 0; k < sharedDim; k += sharedDimReadSize) {
           Logger.verbose('MatMul2D', `k = ${k}, sharedDim: ${sharedDim}, readSize = ${sharedDimReadSize}`);
-
           if (k === sharedDimReadSize) {
             blend = true;
             gl.enable(gl.BLEND);
@@ -248,6 +252,7 @@ export class WebGLUnpackedConv extends Conv {
     const im2colDims = WebGLUnpackedConv.calcIm2ColDims(xshape, kshape, outputShape, 4);
     const outputLayout = inferenceHandler.createTextureLayoutFromShape(
         im2colDims, 4, [im2colDims[0], im2colDims[1], im2colDims[2], im2colDims[3] * 4], {breakAxis: 3});
+
     const shaderSource = `
       const int XC = ${xshape[1]};
       const int XH = ${xshape[2]};
@@ -263,13 +268,12 @@ export class WebGLUnpackedConv extends Conv {
       const int KHKW = KH*KW;
       const int XCKHKW = XC * KHKW;
       const int outputChannels = 4;
-
       vec4 process(int indices[${rank}]) {
         int b  = indices[0]; // batch size
         int oh = indices[1] * strideH - padH; //output height
         int ow = indices[2] * strideW - padW; //output width
         int p = indices[3] * outputChannels; //patch
-        vec4 v = vec4(0.0);
+        vec4 value = vec4(0.0);
         for(int i=0; i < outputChannels; ++i) {
           if(p < XCKHKW) {
             int patchC = p / KHKW;
@@ -286,12 +290,12 @@ export class WebGLUnpackedConv extends Conv {
                 xh2 < XH &&
                 xw2 >= 0 &&
                 xw2 < XW) {
-              v[i] = _X(x);
+              value[i] = _X(x);
             }
           }
           ++p;
         }
-        return v;
+        return value;
       }
       `;
     return {
@@ -321,16 +325,20 @@ export class WebGLUnpackedConv extends Conv {
     const outputLayout = inferenceHandler.createTextureLayoutFromShape(outputShape);
     const initValue = (inputs.length < 3) ? '0.0' : '_B(b)';
     const sharedDim = im2colLayout.shape[3];
-    const blendEnabled = inferenceHandler.session.backend.glContext.isBlendSupported;
+    const blendEnabled = inferenceHandler.session.backend.glContext.isBlendSupported && !this.activation;
     const sharedDimReadSize = blendEnabled && inferenceHandler.session.backend.matmulMaxBatchSize ?
         this.calcSharedDimReadSize(inferenceHandler.session.backend.matmulMaxBatchSize, sharedDim) :
         sharedDim;
     const samplers = ['Im2Col', 'K'];
     if (inputs.length === 3) {
       samplers.push('B');
     }
+
+    const {activationFunction, applyActivation} = getActicationSnippet(this.activation);
+
     const glsl = getGlsl(inferenceHandler.session.backend.glContext.version);
     const shaderSource = `
+    ${activationFunction}
     float process(int indices[${rank}]) {
       int b[1];
       b[0] = indices[1];
@@ -341,15 +349,16 @@ export class WebGLUnpackedConv extends Conv {
       int im2colOffset = im2col[0] * ${im2colLayout.strides[0]} + im2col[1] * ${
         im2colLayout.strides[1]} + im2col[2] * ${im2colLayout.strides[2]} + sharedDimOffset;
       int kernelOffset = indices[1] * ${kLayout.strides[0]} + sharedDimOffset;
-      float sum = sharedDimOffset == 0 ? ${initValue} : 0.0;
+      float value = sharedDimOffset == 0 ? ${initValue} : 0.0;
       for (int i = 0; i < ${sharedDimReadSize}; ++i) {
         vec2 im2colCoords = offsetToCoords(im2colOffset, ${im2colLayout.width}, ${im2colLayout.height});
         vec2 kernelCoords = offsetToCoords(kernelOffset, ${kLayout.width}, ${kLayout.height});
-        sum += dot(${glsl.texture2D}(Im2Col, im2colCoords), ${glsl.texture2D}(K, kernelCoords));
+        value += dot(${glsl.texture2D}(Im2Col, im2colCoords), ${glsl.texture2D}(K, kernelCoords));
         ++im2colOffset;
         ++kernelOffset;
       }
-      return sum;
+      ${applyActivation}
+      return value;
     }`;
     return {
       inputLayouts: inputs.length === 3 ? [im2colLayout, kLayout, bLayout!] : [im2colLayout, kLayout],

web/lib/onnxjs/backends/webgl/ops/fuse-utils.ts

@@ -0,0 +1,27 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+import {glslRelu, glslSigmoid} from './unary-op';
+
+export function getActicationSnippet(activation: string) {
+  let activationFunction = '';
+  let activationName = '';
+  switch (activation) {
+    case 'Relu':
+      activationName = glslRelu().name;
+      activationFunction = glslRelu().body;
+      break;
+    case 'Sigmoid':
+      activationName = glslSigmoid().name;
+      activationFunction = glslSigmoid().body;
+      break;
+    default:
+      // TODO: adding other activations that can be fused.
+      activationName = '';
+      activationFunction = '';
+  }
+  const applyActivation = activation ? `
+  value = ${activationName}(value);` :
+                                       '';
+  return {activationFunction, applyActivation};
+}

web/lib/onnxjs/backends/webgl/ops/matmul-pack.ts

@@ -6,6 +6,7 @@ import {Tensor} from '../../../tensor';
 import {BroadcastUtil} from '../../../util';
 import {WebGLInferenceHandler} from '../inference-handler';
 import {ProgramInfo, RunData, WebGLOperator} from '../types';
+import {getActicationSnippet} from './fuse-utils';
 
 export class WebGLMatMulPacked extends MatMul implements WebGLOperator {
   run(inferenceHandler: WebGLInferenceHandler, inputs: Tensor[]): Tensor[] {
@@ -25,8 +26,11 @@ export class WebGLMatMulPacked extends MatMul implements WebGLOperator {
     const aRank = aShape.length;
     const bRank = bShape.length;
     const sharedDim = aShape[aShape.length - 1];
+
+    const {activationFunction, applyActivation} = getActicationSnippet(this.activation);
     // TODO:fix broadcasting
     const shaderSource = `
+      ${activationFunction}
       vec4 process(int indices[${rank}]) {
           int a[${aRank}];
           int b[${bRank}];
@@ -41,6 +45,7 @@ export class WebGLMatMulPacked extends MatMul implements WebGLOperator {
               value += ${getA(aRank)}.ggaa * ${getB(bRank)}.baba;
           }
           ${processBias}
+          ${applyActivation}
           return value;
       }`;
     return {

web/lib/onnxjs/graph.ts

@@ -737,4 +737,27 @@ class GraphImpl implements Graph, Graph.Transformer {
       nodeIndex++;
     }
   }
-}
+
+  isActivation(n: Node): boolean {
+    switch (n.opType) {
+      // TODO: add other activation methods
+      case 'Relu':
+      case 'Sigmoid':
+        return true;
+      default:
+        return false;
+    }
+  }
+
+  fuseConvActivationNodes() {
+    for (const node of this._nodes) {
+      if (node.opType === 'Conv') {
+        const next = this._allData[node.outputs[0]]._to;
+        if (next.length === 1 && this.isActivation(this._nodes[next[0]])) {
+          node.attributes.set('__internal_activation', 'string', (this._nodes[next[0]].opType));
+          this.deleteNode(next[0]);
+        }
+      }
+    }
+  }
+}

web/lib/onnxjs/ops/conv.ts

@@ -17,6 +17,7 @@ export abstract class Conv implements Operator {
     this.kernelShape = attributes.getInts('kernel_shape', []);
     this.pads = attributes.getInts('pads', [0, 0, 0, 0]);
     this.strides = attributes.getInts('strides', [1, 1]);
+    this.activation = attributes.getString('__internal_activation', '');
   }
 
   checkInputs(inputs: Tensor[]): boolean {
@@ -88,4 +89,5 @@ export abstract class Conv implements Operator {
   protected kernelShape: number[];
   protected pads: number[];
   protected strides: number[];
+  protected activation: string;
 }

web/lib/onnxjs/ops/matmul.ts

@@ -9,7 +9,9 @@ import {Tensor} from '../tensor';
 export abstract class MatMul implements Operator {
   abstract run(inferenceHandler: InferenceHandler, inputs: Tensor[]): Tensor[]|Promise<Tensor[]>;
 
-  initialize(_attributes: Attribute): void {}
+  initialize(attributes: Attribute): void {
+    this.activation = attributes.getString('__internal_activation', '');
+  }
 
   checkInputs(inputs: Tensor[]): boolean {
     if (!inputs || inputs.length !== 2) {
@@ -38,4 +40,5 @@ export abstract class MatMul implements Operator {
 
     return true;
   }
+  protected activation: string;
 }