Merge pull request #4 from wangqunbaidu/master

[D][change]update NHWC and NCHW fomat

README-2.md

-[中文版](https://github.com/PaddlePaddle/Paddle-Lite/blob/develop/web/README_cn.md)
-
-# Web
-
-Web project is an open source deep learning framework designed to work on web browser. It could run on nearly every browser with WebGL support.
-
-## Key Features
-
-### Modular
-
-Web project is built on Atom system which is a versatile framework to support GPGPU operation on WebGL. It is quite modular and could be used to make computation tasks faster by utilizing WebGL.
-
-### High Performance
-
-Web project could run TinyYolo model in less than 30ms on chrome. This is fast enough to run deep learning models in many realtime scenarios.
-
-### Browser Coverage
-
-* PC: Chrome
-* Mac: Chrome
-* Android: Baidu App and QQ Browser
-
-## How To Build & Deploy Demo
-
-```bash
-cd web                        # enter root directory for web project
-npm i                         # install dependencies for npm
-mkdir dist                    # create deployment directory
-cd dist                       # enter deployment directory
-git clone https://github.com/DerekYangMing/Paddle-Web-Models.git # get models
-mv Paddle-Web-Models/separablemodel .                            # move models to specific directory
-cd ..                         # return to root directory for web project
-npm run testVideoDemo         # start demo
-```
-
-## How To Preview Demo
-
-1. Open chrome with url: https://localhost:8123/
-2. Start demo by pressing the 【start detection】 button.
-3. Ensure at least one face is recorded by the camera. The face detection rectangle should be displayed if everything goes fine.
-
-## Feedback and Community Support
-
-- Questions, reports, and suggestions are welcome through Github Issues!
-- Forum: Opinions and questions are welcome at our [PaddlePaddle Forum](https://ai.baidu.com/forum/topic/list/168)！
-- QQ group chat: 696965088

README_cn 2.md

-# PaddleJS Examples
-
-百度PaddleJS使用现成的或者通过转换工具转换的 JavaScript 友好的paddle模型以在浏览器中运行，在浏览器中实现在线推理能力。
-
-## 演示
-
-目前Web项目运行TinyYolo模型可以达到30ms以内，对于一般的实时场景已经足够应对。
-
-### 模块化
-
-## 浏览器覆盖面
-
-* PC: Chrome
-* Mac: Chrome
-* Android: Baidu App and QQ Browser
-
-## 构建部署
-
-```bash
-cd web                        # 进入根目录
-npm i                         # 安装依赖
-mkdir dist                    # 创建资源目录
-cd dist                       # 进入资源目录
-git clone https://github.com/DerekYangMing/Paddle-Web-Models.git # 获取模型
-mv Paddle-Web-Models/separablemodel .                            # 移动模型到制定地点
-cd ..                         # 返回根目录
-npm run tinyYolo              # 启动 tinyYolo 在线推理服务
-```
-
-## 如何预览 demo
-
-1. 在浏览器中打开url: https://localhost:8123/
-2. 点击【开始检测】按钮。
-3. 将人脸对准摄像头，没有问题的话，可以正常检测到人脸。
-
-## 效果
-
-![image](./tinyYolo/demoshow.png)

package.json

 {
-  "name": "paddle-web-demo",
+  "name": "paddel-web",
   "version": "1.0.0",
   "description": "paddle",
   "main": "index.js",

src/executor/executor.es6

@@ -32,8 +32,11 @@ export default class GraphExecutor {
         else if (this.type === 'depthwise_conv2d') {
             return this.inputs.Input;
         }
+        else if (this.type === 'conv2d_transpose') {
+			return this.inputs.Input;
+		}
         else if (this.type === 'elementwise_add') {
-            return this.inputs.X;
+            return this.inputs.X.concat(this.inputs.Y);
         }
         else if (this.type === 'relu' || this.type === 'leaky_relu') {
             return this.inputs.X;
@@ -65,8 +68,13 @@ export default class GraphExecutor {
         }
         else if (this.type === 'batchnorm' || this.type === 'batch_norm') {
             this.outputs.out = this.outputs.Y;
-            return this.outputs.Y;
+            delete this.outputs.Y;
+            return this.outputs.out;
         }
+		else if (this.outputs.Y) {
+			this.outpus.out = this.outputs.Y;
+			return this.outputs.out;
+		}
         else {
             return this.outputs.Out || this.outputs.Output;
         }
@@ -83,7 +91,6 @@ export default class GraphExecutor {
         if (this.type !== 'feed') {
             // let time = +Date.now();
             // log.start(this.opData.iLayer + '-' + this.type);
-            console.log(this.type, this.opData);
             runtime.run(this.type, this.opData, isRendered);
             // log.end(this.opData.iLayer + '-' + this.type);
             // if (runtime.gpu.frameBufferIsComplete().isComplete) {

src/executor/postProcess.es6

 /* eslint-disable */
 
-/* 后处理图片 by zhangmiao06 */
-// let preTestRun = index => {
-//     let img = document.getElementById('image');
-//     img.src = tempPic[index];
-//     img.onload = function () {
-//         testRun(testOutput.data[index], img);
-//     };
-// };
-
+/** 后处理图片 by zhangmiao06, wangqun
+ * let preTestRun = index => {
+ *     let img = document.getElementById('image');
+ *     img.src = tempPic[index];
+ *     img.onload = function () {
+ *         testRun(testOutput.data[index], img);
+ *    };
+ * };
+ **/
 import models from '../utils/models';
 
 const isSimilar = (r1, r2, threshold = 5) => {
@@ -132,11 +132,11 @@ const reshapeMany = data => {
 
 export default class PostProcess {
     constructor(options) {
-        this.modelConfig = models[options.modelName];
+        this.modelConfig = options;
         this.count = 0;
         this.lastRect = [0, 0, 0, 0]
     }
-    
+
     run(data, img, callback, canavs) {
         let {from, to} = this.modelConfig.outputShapes;
         let shape = [].concat(from).reverse();

src/executor/runner.es6

 /**
  * @file Runner 整个流程封装一下
- * @author hantian(hantianjiao@baidu.com)
+ * @author hantian(hantianjiao@baidu.com), wangqungit push origin master:refs/for/master
  * 使用方法：
  * const runner = new Runner({
  *      modelName: 'separate' // '608' | '320' | '320fused' | 'separate'
@@ -11,16 +11,15 @@
  */
 import IO from '../feed/ImageFeed';
 import DataFeed from '../feed/dataFeed';
-import Graph from './loader';
 import PostProcess from './postProcess';
-import models from '../utils/models';
 import Logger from '../../tools/logger';
+import Paddle from '../paddle/paddle';
 window.log = new Logger();
 
 export default class Runner {
     // 加载模型&预热
     constructor(options) {
-        this.modelConfig = models[options.modelName];
+        this.modelConfig = options; // models[options.modelName];
         this.flags = {
             isRunning: false,
             isPreheating: false,
@@ -41,28 +40,42 @@ export default class Runner {
             name: 'image',
             shape: [1, 3, fh, fw]
         }];
-        const MODEL_URL = `/${path}/model.json`;
-        let dir = `https://mms-graph.cdn.bcebos.com/activity/facegame/paddle/${path}/`;
-        if (location.href.indexOf('test=1') > -1) {
-            dir = `/src/view/common/lib/paddle/${path}/`;
-        }
+
         const MODEL_CONFIG = {
-            dir: dir,
-            main: 'model.json' // 主文件
+            dir: `/${path}/`, // 存放模型的文件夹
+            main: 'model.json', // 主文件
         };
-        const graphModel = new Graph();
-        this.model = await graphModel.loadGraphModel(MODEL_CONFIG, {
-            multipart: true,
-            dataType: 'binary',
-            binaryOption: {
-                fileCount: 1, // 切成了多少文件
-                getFileName(i) { // 获取第i个文件的名称
-                    return 'chunk_0.dat';
+        // const graphModel = new Graph();
+        // this.model = await graphModel.loadGraphModel(MODEL_CONFIG, {
+        //     multipart: true,
+        //     dataType: 'binary',
+        //     binaryOption: {
+        //         fileCount: 1, // 切成了多少文件
+        //         getFileName(i) { // 获取第i个文件的名称
+        //             return 'chunk_0.dat';
+        //         }
+        //     },
+        //     feed
+        // });
+
+
+        const paddle = new Paddle({
+            urlConf: MODEL_CONFIG,
+            options: {
+                multipart: true,
+                dataType: 'binary',
+                options: {
+                    fileCount: 1, // 切成了多少文件
+                    getFileName(i) { // 获取第i个文件的名称
+                        return 'chunk_0.dat';
+                    }
                 }
-            },
-            feed
+            }
         });
-        this.model.execute({
+        this.model = await paddle.load();
+
+
+        let inst = this.model.execute({
             input: feed
         });
         this.flags.isPreheating = false;
@@ -149,6 +162,11 @@ export default class Runner {
 
     startStream(getMedia, callback) {
         this.flags.runVideoPaused = false;
-        this.runStream(getMedia, callback);
+        if (typeof getMedia === 'function') {
+            this.runStream(getMedia(), callback);
+        } else {
+            this.runStream(getMedia, callback);
+        }
+
     }
 }

src/factory/fshader/factory.es6

 import ops from './ops';
 /**
  * @file 工厂类，生成fragment shader
- * @author wangqun
+ * @author yangmingming
  */
 export default class Factory {
     constructor(opts) {
@@ -17,10 +17,11 @@ export default class Factory {
         }
     }
 
-    buildShader(opName, data) {
+    buildShader(opName, data, runtime = undefined) {
         let result = '';
         result = this.buildPrefix(opName);
         result += this.buildCommon(opName);
+        result += runtime !== undefined ? this.buildRuntime(runtime) : '';
         result += this.buildOp(opName);
         data.texture2d = this.texture2d;
         result = this.populateData(result, data);
@@ -38,6 +39,12 @@ export default class Factory {
         return ops.common.params + ops.common.func;
     }
 
+    buildRuntime(runtime) {
+        return `
+            int layer_run_time = ${runtime};
+        `;
+    }
+
     buildOp(opName) {
         let code = ops.ops[opName].params;
         // 依赖的方法
@@ -50,7 +57,7 @@ export default class Factory {
             let snippet = atoms[func];
             code += this.populateData(snippet, data);
         });
-        // suffix
+        // 引入 suffix 方法
         code += this.buildSuffix(opName);
         // main方法
         code += ops.ops[opName].func;

src/factory/fshader/ops.es6

@@ -12,6 +12,9 @@ import conv2d_conf from '../../shader/conv2d/conf';
 import conv2d_depthwise_params from '../../shader/conv2d_depthwise/params';
 import conv2d_depthwise_func from '../../shader/conv2d_depthwise/main';
 import conv2d_depthwise_conf from '../../shader/conv2d_depthwise/conf';
+import conv2d_transpose_params from '../../shader/conv2d_transpose/params';
+import conv2d_transpose_func from '../../shader/conv2d_transpose/main';
+import conv2d_transpose_conf from '../../shader/conv2d_transpose/conf';
 import dynamic_params from '../../shader/dynamic/params';
 import dynamic_func from '../../shader/dynamic/main';
 import dynamic_conf from '../../shader/dynamic/conf';
@@ -36,6 +39,12 @@ import softmax_conf from '../../shader/softmax/conf';
 import batchnorm_params from '../../shader/batchnorm/params';
 import batchnorm_func from '../../shader/batchnorm/main';
 import batchnorm_conf from '../../shader/batchnorm/conf';
+import reshape_params from '../../shader/reshape/params';
+import reshape_func from '../../shader/reshape/main';
+import reshape_conf from '../../shader/reshape/conf';
+import transpose_params from '../../shader/transpose/params';
+import transpose_func from '../../shader/transpose/main';
+import transpose_conf from '../../shader/transpose/conf';
 
 import conv2d_elementwise_add_params from '../../shader/conv2d_elementwise_add/params';
 import conv2d_elementwise_add_func from '../../shader/conv2d_elementwise_add/main';
@@ -45,6 +54,14 @@ import conv2d_elementwise_add_winograd_params from '../../shader/conv2d_elementw
 import conv2d_elementwise_add_winograd_func from '../../shader/conv2d_elementwise_add_winograd/main';
 import conv2d_elementwise_add_winograd_conf from '../../shader/conv2d_elementwise_add_winograd/conf';
 
+import concat_params from '../../shader/concat/params';
+import concat_func from '../../shader/concat/main';
+import concat_conf from '../../shader/concat/conf';
+
+import split_params from '../../shader/split/params';
+import split_func from '../../shader/split/main';
+import split_conf from '../../shader/split/conf';
+
 import getArrayIndexFromTensorPos from '../../shader/atom/getArrayIndexFromTensorPos';
 import getArrayIndexFromTexturePos from '../../shader/atom/getArrayIndexFromTexturePos';
 import getTensorPosFromArrayIndex from '../../shader/atom/getTensorPosFromArrayIndex';
@@ -56,6 +73,7 @@ import moveTexture2PosToReal from '../../shader/atom/moveTexture2PosToReal';
 import getPixelsFromTexturePos from '../../shader/atom/getPixelsFromTexturePos';
 import getRangePowSumFromArrayIndex from '../../shader/atom/getRangePowSumFromArrayIndex';
 import getRangeSumFromArrayIndex from '../../shader/atom/getRangeSumFromArrayIndex';
+import transferFromNHWCtoNCHW from '../../shader/atom/transferFromNHWCtoNCHW';
 import sigmoid from '../../shader/atom/sigmoid';
 import prelu from '../../shader/atom/prelu';
 import scale from '../../shader/atom/scale';
@@ -75,6 +93,11 @@ export default {
         ivec56
     },
     ops: {
+    	conv2d_transpose:{
+            params: conv2d_transpose_params,
+            func: conv2d_transpose_func,
+            confs: conv2d_transpose_conf
+		},
         conv2d: {
             params: conv2d_params,
             func: conv2d_func,
@@ -125,6 +148,16 @@ export default {
             func: mul_func,
             confs: mul_conf
         },
+        concat: {
+            params: concat_params,
+            func: concat_func,
+            confs: concat_conf
+        },
+        split: {
+            params: split_params,
+            func: split_func,
+            confs: split_conf
+        },
         relu: {
             params: dynamic_params,
             func: dynamic_func,
@@ -149,7 +182,17 @@ export default {
             params: batchnorm_params,
             func: batchnorm_func,
             confs: batchnorm_conf
-        }
+        },
+		reshape: {
+			params: reshape_params,
+			func: reshape_func,
+			confs: reshape_conf
+		},
+		transpose: {
+			params: transpose_params,
+			func: transpose_func,
+			confs: transpose_conf
+		}
     },
     atoms: {
         getArrayIndexFromTensorPos,
@@ -166,6 +209,7 @@ export default {
         sigmoid,
         prelu,
         scale,
-        softmax
+        softmax,
+        transferFromNHWCtoNCHW
     }
 };

src/feed/ImageFeed.es6

@@ -13,6 +13,7 @@ export default class imageFeed {
         this.pixels = '';
         this.defaultParams = {
             gapFillWith: '#000',
+            mean: [0, 0, 0],
             std: [1, 1, 1]
         };
     };
@@ -61,7 +62,8 @@ export default class imageFeed {
         let std = opt.std;
         // 考虑channel因素获取数据
         for (let i = 0; i < data.length; i += 4) {
-
+            // img_mean 0.485, 0.456, 0.406
+            //img_std 0.229, 0.224, 0.225
             let index = i / 4;
             let vIndex = Math.floor(index / sw);
             let hIndex = index - (vIndex * sw) - 1;
@@ -82,10 +84,12 @@ export default class imageFeed {
      * @param shape
      */
     allReshapeToRGB(imageData, opt, scaleSize) {
-        const {sw, sh} = scaleSize;
+        //const {sw, sh} = scaleSize;
         const [b, c, h, w] = opt.targetShape;
         let data = imageData.data || imageData;
+        // mean和std是介于0-1之间的
         let mean = opt.mean;
+        let std = opt.std;
         let dataLength = data.length;
         // let result = new Float32Array(dataLength * 3);
         let result = this.result;
@@ -101,10 +105,14 @@ export default class imageFeed {
                 let iwj = iw + j;
                 for (let k = 0; k < c; ++k) {
                     let a = iwj * 4 + k;
-                    result[offset++] = (data[a] - mean[k]) / 256;
+                    result[offset] = data[a] / 255;
+                    result[offset] -= mean[k];
+                    result[offset] /= std[k];
+                    offset++;
                 }
             }
         }
+        
         return result;
     };
 
@@ -115,6 +123,7 @@ export default class imageFeed {
      * @return {Object} 缩放后的尺寸
      */
     reSize(image, params) {
+        console.log('execute resize!!');
         // 原始图片宽高
         const width = this.pixelWidth;
         const height = this.pixelHeight;
@@ -136,7 +145,40 @@ export default class imageFeed {
         this.setInputCanvas(image);
         return {sw, sh};
     };
+    /**
+     * 根据scale缩放图像并且缩放成目标尺寸并居中
+     */
+    resizeAndFitTargetSize(image, params){
+        console.log('execute resizeAndFitTargetSize!!');
+        // 原始图片宽高
+        const width = this.pixelWidth;
+        const height = this.pixelHeight;
+        // 缩放后的宽高
+        let sw = width;
+        let sh = height;
+        // 最小边缩放到scale
+        if (width < height) {
+            sw = params.scale;
+            sh = Math.round(sw * height / width);
+        } else {
+            sh = params.scale;
+            sw = Math.round(sh * width / height);
+        }
 
+        this.fromPixels2DContext.canvas.width = sw;
+        this.fromPixels2DContext.canvas.height = sh;
+        const targetWidth = params.targetSize.width;
+        const targetHeight = params.targetSize.height;
+        this.fromPixels2DContext.drawImage(
+            image, 0, 0, sw, sh);
+        let x = (sw - targetWidth)/2;
+        let y = (sh - targetHeight)/2;
+        sw = targetWidth;
+        sh = targetHeight;
+        let data = this.getImageData(params, x, y, {sw, sh});
+        this.setInputCanvas(image);
+        return data;
+}
 
     /**
      * 缩放成目标尺寸并居中
@@ -199,14 +241,16 @@ export default class imageFeed {
      * @param pixels
      * @returns {Uint8ClampedArray}
      */
-    getImageData(pixels, scaleSize) {
+    getImageData(pixels, x, y, scaleSize) {
+
         const {sw, sh} = scaleSize;
         // 复制画布上指定矩形的像素数据
         let vals = this.fromPixels2DContext
-            .getImageData(0, 0, sw, sh);
+            .getImageData(x, y, sw, sh);
         // crop图像
         // const width = pixels.width;
         // const height = pixels.height;
+
         return vals;
     };
 
@@ -236,18 +280,27 @@ export default class imageFeed {
         if (pixels instanceof HTMLImageElement || pixels instanceof HTMLVideoElement) {
             this.pixelWidth = pixels.naturalWidth || pixels.width;
             this.pixelHeight = pixels.naturalHeight || pixels.height;
-            if (opt.scale) { // 兼容以前的，如果有scale就是短边缩放到scale模式
+            if (opt.scale && opt.targetSize){ // Moblienet的情况
+                data = this.resizeAndFitTargetSize(pixels, opt);
+                data2 = this.fromPixels2DContext2.getImageData(0, 0, this.pixelWidth, this.pixelHeight);
+            }
+            else if (opt.scale) { // 兼容以前的，如果有scale就是短边缩放到scale模式
+
                 scaleSize = this.reSize(pixels, opt);
-                data = this.getImageData(opt, scaleSize);
+                console.dir(scaleSize);
+                console.dir(pixels);
+                data = this.getImageData(opt, 0, 0, scaleSize);
                 data2 = this.fromPixels2DContext2.getImageData(0, 0, this.pixelWidth, this.pixelHeight);
             }
-            else if (opt.targetSize) { // 如果有targetSize，就是装在目标宽高里的模式
+            else if (opt.targetSize) { // 如果有targetSize，就是装在目标宽高里的模式 TinyYolo的情况
+
                 scaleSize = this.fitToTargetSize(pixels, opt);
-                data = this.getImageData(opt, scaleSize);
+                data = this.getImageData(opt, 0, 0, scaleSize);
                 data2 = this.fromPixels2DContext2.getImageData(0, 0, this.pixelWidth, this.pixelHeight);
             }
         }
 
+
         if (opt.gray) {
             data = grayscale(data);
         }
@@ -259,7 +312,6 @@ export default class imageFeed {
         if (opt.targetShape) {
             data = this.allReshapeToRGB(data, opt, scaleSize);
         }
-        
         return [{data: data, shape: opt.shape || opt.targetShape, name: 'image', canvas: data2}];
     }
 }

src/gpu/gpu.es6

@@ -23,6 +23,7 @@ export default class gpu {
         opts.width_raw_canvas = Number(opts.width_raw_canvas) || 512;
         opts.height_raw_canvas = Number(opts.height_raw_canvas) || 512;
         const canvas = opts.el ? opts.el : document.createElement('canvas');
+
         canvas.addEventListener('webglcontextlost', evt => {
             evt.preventDefault();
             console.log('webgl context is lost~');
@@ -103,7 +104,7 @@ export default class gpu {
         this.cacheTextures = {};
         this.uniformLocations = {};
         // texture buffer
-        this.outTextures = [];
+        this.texturesMap = {};
         // pbo
         this.pbo = gl.createBuffer();
     }
@@ -155,7 +156,7 @@ export default class gpu {
             gl.FLOAT,          // Data type for each chanel.
             null);
         gl.bindTexture(gl.TEXTURE_2D, null);
-        this.outTextures.push(texture);
+        this.texturesMap[out.tensorId] = texture;
         return program;
     }
 
@@ -263,19 +264,19 @@ export default class gpu {
      * @param {WebGLTexture} texture 材质
      * @returns {WebGLFramebuffer} The framebuffer
      */
-    attachFrameBuffer(iLayer) {
+    attachFrameBuffer(iLayer, tensorId) {
         this.prevTexture = this.currentTexture;
         // this.currentTexture = this.textureBuffer[this.textureBufferIndex % 2];
         // this.textureBufferIndex = (this.textureBufferIndex + 1) >= 2 ? 0 : 1;
-        this.currentTexture = this.outTextures[iLayer];
-        console.log('this.currentTexture', this.currentTexture);
+        this.currentTexture = this.texturesMap[tensorId];
         const gl = this.gl;
         gl.framebufferTexture2D(gl.FRAMEBUFFER, // The target is always a FRAMEBUFFER.
-            gl.COLOR_ATTACHMENT0, // We are providing the color buffer.
+            gl.COLOR_ATTACHMENT0, // We are providing the color buffer.表示texture是颜色关联对象
             gl.TEXTURE_2D, // This is a 2D image texture.
             this.currentTexture, // The texture.
             0 // 0, we aren't using MIPMAPs
         );
+
         gl.viewport(
             0,
             0,
@@ -340,7 +341,8 @@ export default class gpu {
         const gl = this.gl;
         let texture;
         if (!item.data) {
-            texture = this.prevTexture;
+            // texture = this.prevTexture;
+            texture = this.texturesMap[item.tensorId];
         } else {
             // texture = gl.createTexture();
             if (isRendered && (iLayer > 0 || (iLayer === 0 && item.tensor !== 'origin'))) {
@@ -348,9 +350,7 @@ export default class gpu {
                 texture = tData[item.variable + '_' + item.tensor];
             } else {
                 texture = gl.createTexture();
-                if (index === 0) {
-                    this.cacheTextures['' + iLayer] = this.cacheTextures['' + iLayer] || {};
-                }
+                this.cacheTextures['' + iLayer] = this.cacheTextures['' + iLayer] || {};
                 this.cacheTextures['' + iLayer][item.variable + '_' + item.tensor] = texture;
             }
         }

src/graph/graph.es6

@@ -10,7 +10,6 @@ import Utils from '../utils/utils';
  * @file Graph，绘制生成model网络
  * @author wangqun@baidu.com
  */
-let start = 0;
 // 生成factory实例
 const factory = new Factory({});
 // 获取op的输入配置
@@ -48,16 +47,19 @@ export default class Graph {
         const executor = this.constructExecutor(op);
         const opData = new OpData(op.type, executor.inputs, executor.outputs, executor.attrs);
         const name = opData.name;
-        const fsCode = factory.buildShader(name, opData.data);
 
-        opData.fsCode = fsCode;
-        opData.program = this.inst.createProgram(fsCode, opData.tensor['out']);
+        opData.program = [];
+        opData.program = opData.outputTensors.map((outTensor, index) => {
+            const fsCode = factory.buildShader(name, opData.fShaderParams[index], index);
+            return this.inst.createProgram(fsCode, outTensor);
+        });
+
         opData.renderData = opConfs[name].map(elem => {
             let item = Object.assign({}, elem);
-            const tensorData = opData.tensor[item.tensor];
+            const tensorData = opData.inputTensors.find(tensor => tensor.name === item.tensor);
             if (item.type === 'texture') {
+                item.tensorId = tensorData.opts.type;
                 item.data = tensorData.data;
-
                 if (this.feedOp.id === op.id && item.tensor === 'origin') {
                     item.shape = tensorData.shape;
                     this.feedItem = item;
@@ -70,7 +72,6 @@ export default class Graph {
             }
             return item;
         });
-
         // console.timeEnd('opData.renderData');
         opData.iLayer = this.iLayer++;
         op.opData = opData;
@@ -83,7 +84,6 @@ export default class Graph {
             return;
         }
         executor.execute(this.inst, this.isExecuted);
-        // if (executor.next && start++ < 2) {
         if (executor.next) {
             const id = executor.next;
             const next = this.getTensor(id);
@@ -136,7 +136,9 @@ export default class Graph {
         const input = executor.inputs;
         const output = executor.outputs;
         Object.keys(output).forEach(function(key){
-            output[key] = that.getTensorAttr(output[key][0]);
+            output[key].forEach((item, index) => {
+                output[key][index] = that.getTensorAttr(item)[0];
+            });
         });
         Object.keys(input).forEach(function(key){
             if (that.test && ((key === 'Input') || (key === 'X'))) {
@@ -240,8 +242,10 @@ export default class Graph {
      */
     getNextExecutor(ops, id) {
         return ops.filter((item, key) => {
-            if (id === item.inputsName[0]) {
-                return true;
+            for (let i = 0; i < item.inputsName.length; i++) {
+                if (id === item.inputsName[i]) {
+                    return true;
+                }
             }
         });
     }

src/loader/loader.es6

@@ -29,6 +29,24 @@ export default class Loader  {
         if (!this.loadOptions) {
             this.loadOptions = {};
         }
+        else {
+            // this.fetchJson(this.modelGonfig.dir + 'x.json').then(data => {
+            //     const [b, c, h, w] = [1, 3, 320, 320];
+            //     const size = data.length;
+            //     const total = 3 * 320 * 320;
+            //     this.testData = new Float32Array(total);
+            //     for (let i = 0; i < size; i++) {
+            //         let j = i / (c * w) | 0;
+            //         let k = i % (c * w);
+            //         let b1 = j / h | 0;
+            //         let h1 = j % h;
+            //         let c1 = k % c;
+            //         let w1 = k / c | 0;
+            //         let l = b1 * (c * h * w) + c1 * (h * w) + h1 * (w) + w1;
+            //         this.testData[i] = data[l];
+            //     }
+            // });
+        }
     }
 
     fetchOneChunk(path) {
@@ -110,10 +128,22 @@ export default class Loader  {
                 && item.name.match(TMP_SCHEME_REGEX) === null
                 && item.name.match(TMP_REGEX) === null;
             })
+            // .sort((a, b) => {
+            //     if (a.name > b.name) {
+            //         return 1;
+            //     }
+            //     if (a.name < b.name) {
+            //         return -1;
+            //     }
+            //     return 0;
+            // }) // 按字母顺序排列 在model.json里
             .forEach(item => {
                 len = item.shape.reduce((a, b) => a * b); // 长度为shape的乘积
-                item.data = this.allData.slice(marker, marker + len);
-                marker += len;
+                // 为了减少模型体积，模型转换工具不会导出非persistable的数据，这里只需要读取persistable的数据
+                if (item.persistable) {
+                    item.data = this.allData.slice(marker, marker + len);
+                    marker += len;
+                }
             });
     }
 

src/paddle/paddle.es6

@@ -3,7 +3,7 @@ import 'babel-polyfill';
 import Loader from '../loader/loader';
 import Graph from '../graph/graph';
 /**
- * @file paddle对象，负责加载模型和执行在线推理
+ * @file GraphModel，绘制生成model网络
  * @author wangqun@baidu.com
  */
 
@@ -28,6 +28,7 @@ export default class Paddle {
     }
 
     async load() {
+
         if (this.options === null) {
             // todo saniac 报错提示修改
             throw new Error(
@@ -56,7 +57,6 @@ export default class Paddle {
      * @returns {*}
      */
     execute(inputs) {
-        debugger;
         let that = this;
         this.feed = this.graph.feed = inputs;
         // 生成op数据

src/runtime/runtime.es6

 /* eslint-disable */
 import Gpu from '../gpu/gpu';
 import getMaxUniforms from '../test/getMaxUniforms';
+import Factory from '../factory/fshader/factory';
+import {getTextureShapeInfo} from '../utils/opData';
+// 生成factory实例
+const factory = new Factory({});
 /**
  * @file gpu运行时
  * @author wangqun@baidu.com, yangmingming@baidu.com
@@ -36,26 +40,30 @@ export default {
         }
         // 设置gpu参数
         const gpu = this.gpu;
-        gpu.setOutProps(opData.tensor['out']);
-        // 生成帧缓存材质
-        gpu.attachFrameBuffer(opData.iLayer);
-        // let end = +Date.now();
-        let bufferStatus = gpu.frameBufferIsComplete();
-        if (bufferStatus.isComplete) {
-            // start = +Date.now();
-            // timeObj['buferstatus-time'] = start - end;
-            // gpu.attachShader(opData.fshader);
-            gpu.setProgram(opData.program, isRendered);
-            // end = +Date.now();
-            // timeObj['createshader-time'] = end - start;
-            // timeObj['jsTime'] = end - time;
-            // statistic.push(timeObj);
-            // 开始计算
-            this.gpu.render(opData.renderData, opData.iLayer, isRendered);
-            return this;
-        } else {
-            return bufferStatus.message;
-        }
+        opData.program.forEach((program, index) => {
+            const outTensor = opData.outputTensors[index];
+            const outTensorId = outTensor.tensorId;
+            gpu.setOutProps(outTensor);
+            // 生成帧缓存材质
+            gpu.attachFrameBuffer(opData.iLayer, outTensorId);
+            // let end = +Date.now();
+            let bufferStatus = gpu.frameBufferIsComplete();
+            if (bufferStatus.isComplete) {
+                // start = +Date.now();
+                // timeObj['buferstatus-time'] = start - end;
+                // gpu.attachShader(opData.fshader);
+
+                gpu.setProgram(program, isRendered);
+                // end = +Date.now();
+                // timeObj['createshader-time'] = end - start;
+                // timeObj['jsTime'] = end - time;
+                // statistic.push(timeObj);
+
+
+                // 开始计算，执行 gl.drawArrays
+                this.gpu.render(opData.renderData, opData.iLayer, isRendered);
+            }
+        });
     },
 
     /**

src/shader/atom/getArrayIndexFromTensorPos.es6

@@ -2,7 +2,6 @@
 /**
  * @file 公共方法
  * @author yangmingming
- *
  */
 export default `
 

src/shader/atom/prefix.es6

@@ -11,7 +11,7 @@ export default `
     precision mediump float;
     precision mediump int;
 #endif
-    varying vec2 vCoord;
+
     void setOutput(float result) {
         gl_FragColor.r = result;
     }

src/shader/atom/transferFromNHWCtoNCHW.es6

+/* eslint-disable */
+/**
+ * @file 公共方法
+ * @author chenhaoze
+ */
+// TEXTURE_NAME, tensor name
+// 获取材质中的数据
+// uniform sampler2D TEXTURE_NAME;
+export default `
+ivec4 transferFromNHWCtoNCHW( int sumVal,  const int channel, const int width_shape, const int height_shape, const int total_shape) {
+
+	int n_origin = int(total_shape/(channel * width_shape * height_shape));
+	int new_a = sumVal % width_shape;
+	sumVal = int((sumVal - new_a) / width_shape);
+	int new_b = sumVal % height_shape;
+	sumVal = int((sumVal - new_b) / height_shape);
+	int new_g = sumVal % channel;
+	sumVal = int((sumVal - new_g) / channel);
+	int new_r = sumVal % n_origin;
+	return ivec4(new_r,new_g,new_b,new_a);
+}
+`;

src/shader/batchnorm/conf.es6

 /* eslint-disable */
 /**
  * @file batchnorm的配置文件
- * @author yangmingming
+ * @author wangqun
  */
 export default {
     dep: [
@@ -16,6 +16,30 @@ export default {
             conf: {
                 TEXTURE_NAME: 'texture_scale'
             }
+        },
+        {
+            func: 'getPixelsFromTexturePos',
+            conf: {
+                TEXTURE_NAME: 'texture_bias'
+            }
+        },
+        {
+            func: 'getPixelsFromTexturePos',
+            conf: {
+                TEXTURE_NAME: 'texture_variance'
+            }
+        },
+        {
+            func: 'getPixelsFromTexturePos',
+            conf: {
+                TEXTURE_NAME: 'texture_mean'
+            }
+        },
+        {
+            func: 'getPixelsFromTexturePos',
+            conf: {
+                TEXTURE_NAME: 'texture_origin'
+            }
         }
     ],
     conf: [
@@ -26,15 +50,23 @@ export default {
         'HEIGHT_TEXTURE_ORIGIN',
         'CHANNEL_ORIGIN',
         'TOTAL_SHAPE_ORIGIN',
+
         'WIDTH_SHAPE_OUT',
         'HEIGHT_SHAPE_OUT',
         'WIDTH_TEXTURE_OUT',
         'HEIGHT_TEXTURE_OUT',
         'CHANNEL_OUT',
         'OFFSET_Y_OUT',
+
         'EPSILON',
         'WIDTH_TEXTURE_SCALE',
         'HEIGHT_TEXTURE_SCALE',
+        'WIDTH_TEXTURE_BIAS',
+        'HEIGHT_TEXTURE_BIAS',
+        'WIDTH_TEXTURE_MEAN',
+        'HEIGHT_TEXTURE_MEAN',
+        'WIDTH_TEXTURE_VARIANCE',
+        'HEIGHT_TEXTURE_VARIANCE',
         'MULTI_VALUE',
         'BIAS_VALUE',
         'ACTIVE_FUNCTION'
@@ -46,6 +78,24 @@ export default {
             setter: 'initTexture',
             type: 'texture'
         },
+        {
+            tensor: 'bias',
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
+        },
+        {
+            tensor: 'mean',
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
+        },
+        {
+            tensor: 'variance',
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
+        },
         {
             tensor: 'origin',
             variable: 'texture',
@@ -53,4 +103,4 @@ export default {
             type: 'texture'
         }
     ]
-};
\ No newline at end of file
+};

src/shader/batchnorm/main.es6

@@ -9,11 +9,15 @@ void main(void) {
     // 输出数据
     ivec4 oPos = getOutputTensorPos();
     float o = getValueFromTensorPos_origin(oPos.r, oPos.g, oPos.b, oPos.a);
+
     // 归一化数据
-    vec4 scale = getPixelsFromTexturePos_texture_scale(vec2((float(int(oPos.g)) + 0.5) / float(width_texture_scale), 0.0));
-    float x = (o - scale[3]) / sqrt(scale[2] + epsilon);
-    float res = scale[0] * x + scale[1];
-    
+    vec4 scale = getPixelsFromTexturePos_texture_scale(vec2( float(oPos.g) / float(width_texture_scale), 0.0));
+    vec4 bias = getPixelsFromTexturePos_texture_bias(vec2((float(oPos.g)) / float(width_texture_bias), 0.0));
+    vec4 mean = getPixelsFromTexturePos_texture_mean(vec2((float(oPos.g)) / float(width_texture_mean), 0.0));
+    vec4 variance = getPixelsFromTexturePos_texture_variance(vec2((float(oPos.g)) / float(width_texture_variance), 0.0));
+
+    float x = (o - mean[0]) / sqrt(variance[0] + epsilon);
+    float res = scale[0] * x + bias[0];
     setOutput(res);
 }
-`;
\ No newline at end of file
+`;

src/shader/batchnorm/params.es6

 /* eslint-disable */
 /**
  * @file batchnorm参数文件
- * @author yangmingming
+ * @author wangqun
  */
 export default `
 // 输入数据
@@ -14,9 +14,18 @@ const int channel_origin = CHANNEL_ORIGIN;
 const int total_shape_origin = TOTAL_SHAPE_ORIGIN;
 // 计算数据
 const float epsilon = float(EPSILON);
+const int width_texture_bias = WIDTH_TEXTURE_BIAS;
+const int height_texture_bias = HEIGHT_TEXTURE_BIAS;
+const int width_texture_variance = WIDTH_TEXTURE_VARIANCE;
+const int height_texture_variance = HEIGHT_TEXTURE_VARIANCE;
+const int width_texture_mean = WIDTH_TEXTURE_MEAN;
+const int height_texture_mean = HEIGHT_TEXTURE_MEAN;
 const int width_texture_scale = WIDTH_TEXTURE_SCALE;
 const int height_texture_scale = HEIGHT_TEXTURE_SCALE;
 // 输入数据
 uniform sampler2D texture_origin;
 uniform sampler2D texture_scale;
-`;
\ No newline at end of file
+uniform sampler2D texture_bias;
+uniform sampler2D texture_variance;
+uniform sampler2D texture_mean;
+`;

src/shader/concat/conf.es6

+/* eslint-disable */
+/**
+ * @file concat的配置文件
+ * @author zhangjingyuan02
+ */
+export default {
+    dep: [
+        {
+            func: 'getValueFromTensorPos',
+            conf: {
+                TENSOR_NAME: 'origin'
+            }
+        },
+        {
+            func: 'getValueFromTensorPos',
+            conf: {
+                TENSOR_NAME: 'counter'
+            }
+        },
+        {
+            func: 'transferFromNHWCtoNCHW'
+        }
+    ],
+    conf: [
+        'LENGTH_SHAPE_COUNTER',
+        'WIDTH_SHAPE_COUNTER',
+        'HEIGHT_SHAPE_COUNTER',
+        'WIDTH_TEXTURE_COUNTER',
+        'HEIGHT_TEXTURE_COUNTER',
+        'CHANNEL_COUNTER',
+
+        'WIDTH_SHAPE_ORIGIN',
+        'HEIGHT_SHAPE_ORIGIN',
+        'LENGTH_SHAPE_ORIGIN',
+        'WIDTH_TEXTURE_ORIGIN',
+        'HEIGHT_TEXTURE_ORIGIN',
+        'CHANNEL_ORIGIN',
+
+        'WIDTH_SHAPE_OUT',
+        'HEIGHT_SHAPE_OUT',
+        'WIDTH_TEXTURE_OUT',
+        'HEIGHT_TEXTURE_OUT',
+        'CHANNEL_OUT',
+        'OFFSET_Y_OUT'
+    ],
+    input: [
+        {
+            tensor: 'origin',
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
+        },
+        {
+            tensor: 'counter',
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
+        }
+    ]
+};

src/shader/concat/main.es6

+/* eslint-disable */
+/**
+ * @file concat主函数
+ * @author zhangjingyuan02
+ */
+export default `
+// start函数
+void main(void) {
+    ivec4 oPos = getOutputTensorPos();
+    // 输出坐标转换为输入坐标
+	int sumVal = oPos.g + oPos.a * channel_out + oPos.b * channel_out * width_shape_out + oPos.r * channel_out * width_shape_out * height_shape_out;
+    ivec4 new_oPos = transferFromNHWCtoNCHW(sumVal, channel_out, width_shape_out, height_shape_out, total_shape_out);
+    float o = 0.0;
+    if (new_oPos[dim] > inputs_dim[0] - 1) {
+        new_oPos[dim] = new_oPos[dim] - inputs_dim[0];
+        o = getValueFromTensorPos_counter(new_oPos.r, new_oPos.g, new_oPos.b, new_oPos.a);
+    }
+    else {
+        o = getValueFromTensorPos_origin(new_oPos.r, new_oPos.g, new_oPos.b, new_oPos.a);
+    }
+	setOutput(float(o));
+}
+`;

src/shader/concat/params.es6

+/* eslint-disable */
+/**
+ * @file concat 参数文件
+ * @author zhangjingyuan02
+ */
+export default `
+// mul的input数据
+const int axis = AXIS;
+
+// 常量
+// 输入数据
+const int length_shape_counter = LENGTH_SHAPE_COUNTER;
+const int width_shape_counter = WIDTH_SHAPE_COUNTER;
+const int height_shape_counter = HEIGHT_SHAPE_COUNTER;
+const int width_texture_counter = WIDTH_TEXTURE_COUNTER;
+const int height_texture_counter = HEIGHT_TEXTURE_COUNTER;
+const int channel_counter = CHANNEL_COUNTER;
+const int total_shape_counter = TOTAL_SHAPE_COUNTER;
+
+const int width_shape_origin = WIDTH_SHAPE_ORIGIN;
+const int height_shape_origin = HEIGHT_SHAPE_ORIGIN;
+const int length_shape_origin = LENGTH_SHAPE_ORIGIN;
+const int width_texture_origin = WIDTH_TEXTURE_ORIGIN;
+const int height_texture_origin = HEIGHT_TEXTURE_ORIGIN;
+const int channel_origin = CHANNEL_ORIGIN;
+const int total_shape_origin = TOTAL_SHAPE_ORIGIN;
+
+const int total_shape_out = TOTAL_SHAPE_OUT;
+
+const int dim = DIM;
+const int inputs_dim[1] = int[](INPUTS_DIM);
+
+
+// uniform变量
+// 输入数据
+uniform sampler2D texture_counter;
+uniform sampler2D texture_origin;
+`;

src/shader/conv2d/conf.es6

@@ -16,7 +16,13 @@ export default {
             conf: {
                 TENSOR_NAME: 'filter'
             }
-        }
+        },
+		{
+			func: 'transferFromNHWCtoNCHW',
+			conf:{
+
+			}
+		}
     ],
     conf: [
         'LENGTH_SHAPE_FILTER',

src/shader/conv2d/params.es6

@@ -22,7 +22,7 @@ export default `
     const int width_texture_origin = WIDTH_TEXTURE_ORIGIN;
     const int height_texture_origin = HEIGHT_TEXTURE_ORIGIN;
     const int channel_origin = CHANNEL_ORIGIN;
-    
+
     // 计算相关
     // 拆分步长
     const int stride_h = STRIDES_X;

src/shader/conv2d_depthwise/conf.es6

@@ -16,7 +16,13 @@ export default {
             conf: {
                 TENSOR_NAME: 'filter'
             }
-        }
+        },
+        		{
+			func: 'transferFromNHWCtoNCHW',
+			conf:{
+
+			}
+		}
     ],
     conf: [
         'LENGTH_SHAPE_FILTER',

src/shader/conv2d_depthwise/main.es6

@@ -12,7 +12,6 @@ export default `
         int y = oPos.b;
         int b = oPos.r; 
         float res = 0.0;
-
         int top = y * stride_v - padTop;
         int left = x * stride_h - padLeft;
         for (int fy = 0; fy < height_shape_filter; fy++) {

src/shader/conv2d_elementwise_add/conf.es6

@@ -16,7 +16,13 @@ export default {
             conf: {
                 TENSOR_NAME: 'filter'
             }
-        }
+        },
+		{
+			func: 'transferFromNHWCtoNCHW',
+			conf:{
+
+			}
+		}
     ],
     conf: [
         'LENGTH_SHAPE_FILTER',

src/shader/conv2d_elementwise_add/main.es6

@@ -7,6 +7,8 @@ export default `
     // start函数
     void main(void) {
         ivec4 oPos = getOutputTensorPosLIMIT_OUT();
+		int sumVal = oPos.g + oPos.a * channel_out + oPos.b * channel_out * width_shape_out;
+        ivec4 new_oPos = transferFromNHWCtoNCHW(sumVal, channel_out, width_shape_out, height_shape_out, total_shape_out);
         int x = oPos.a;
         int c = oPos.g;
         int y = oPos.b;

src/shader/conv2d_elementwise_add/params.es6

@@ -32,7 +32,8 @@ export default `
     const int dilation_v = DILATIONS_Y;
     // groups
     const int groups = GROUPS;
-   
+
+   const int total_shape_out = TOTAL_SHAPE_OUT;
     // 加法
     const int axis = AXIS;
      

src/shader/conv2d_transpose/conf.es6

+/* eslint-disable */
+/**
+ * @file conv2d的配置文件
+ * @author chenhaoze
+ */
+export default {
+    dep: [
+        {
+            func: 'getValueFromTensorPos',
+            conf: {
+                TENSOR_NAME: 'origin'
+            }
+        },
+        {
+            func: 'getValueFromTensorPos',
+            conf: {
+                TENSOR_NAME: 'filter'
+            }
+        }
+    ],
+    conf: [
+        'LENGTH_SHAPE_FILTER',
+        'WIDTH_SHAPE_FILTER',
+        'HEIGHT_SHAPE_FILTER',
+        'WIDTH_TEXTURE_FILTER',
+        'HEIGHT_TEXTURE_FILTER',
+        'CHANNEL_FILTER',
+
+        'WIDTH_SHAPE_ORIGIN',
+        'HEIGHT_SHAPE_ORIGIN',
+        'LENGTH_SHAPE_ORIGIN',
+        'WIDTH_TEXTURE_ORIGIN',
+        'HEIGHT_TEXTURE_ORIGIN',
+        'CHANNEL_ORIGIN',
+
+        'WIDTH_SHAPE_OUT',
+        'HEIGHT_SHAPE_OUT',
+        'WIDTH_TEXTURE_OUT',
+        'HEIGHT_TEXTURE_OUT',
+        'CHANNEL_OUT',
+        'OFFSET_Y_OUT',
+
+        'STRIDE_HORIZONTAL',
+        'STRIDE_VERTICAL',
+        'PAD_LEFT',
+        'PAD_TOP',
+        'DILATION_HORIZONTAL',
+        'DILATION_VERTICAL',
+        'GROUPS',
+        'MULTI_VALUE',
+        'BIAS_VALUE',
+        'ACTIVE_FUNCTION'
+    ],
+    input: [
+        // {
+        //     tensor: 'filter',
+        //     variable: 'numbers_shape',
+        //     setter: 'uniform1iv',
+        //     type: 'uniform'
+        // },
+        {
+            tensor: 'filter',
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
+        },
+        {
+            tensor: 'origin',
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
+        }
+        // {
+        //     tensor: 'origin',
+        //     variable: 'numbers_shape',
+        //     setter: 'uniform1iv',
+        //     type: 'uniform'
+        // },
+        // {
+        //     tensor: 'out',
+        //     variable: 'numbers_shape',
+        //     setter: 'uniform1iv',
+        //     type: 'uniform'
+        // }
+    ]
+};

src/shader/conv2d_transpose/main.es6

+/* eslint-disable */
+/**
+ * @file 主函数
+ * @author chenhaoze
+ */
+export default `
+    // start函数
+    void main(void) {
+        ivec4 oPos = getOutputTensorPosLIMIT_OUT();
+        int x = oPos.a;
+        int c = oPos.g;
+        int y = oPos.b;
+        int b = oPos.r; 
+        float res = 0.0;
+
+// 重排遍历顺序
+//int sumVal = oPos.g + oPos.a * channel_out + oPos.b * channel_out * width_shape_out;
+//int new_a = sumVal % width_shape_out;
+//int new_b = int((sumVal - new_a) / width_shape_out) % height_shape_out;
+//int new_g = int((((sumVal - new_a) / width_shape_out) - new_b) / height_shape_out);
+//int x = new_a;
+//int c = new_g;
+//int y = new_b;
+        // 获取output的坐标
+        int oTensorChannel = (c / (channel_out / groups)) * channel_filter;
+
+        int oy = y * 1 - padTop;
+        for (int fy = 0; fy < height_shape_filter; fy++) {
+            if (oy >= height_shape_origin) {
+                break;
+            }
+            if (oy < 0) {
+                oy += dilation_v;
+                continue;
+            }
+            int ox = x * 1 - padLeft;
+            for (int fx = 0; fx < width_shape_filter; fx++) {
+                if (ox >= width_shape_origin) {
+                    break;
+                }
+                if (ox < 0) {
+                    ox += dilation_h;
+                    continue;
+                }
+                // channel计算
+                for (int j = 0; j < channel_filter; j++) {
+					float o = 0.0;
+                	if (ox % stride_h == 0 && oy % stride_v == 0) {
+						int temp_x = int(ox / stride_h);
+						int temp_y = int(oy / stride_v);
+						o = getValueFromTensorPosLIMIT_ORIGIN_origin(b, oTensorChannel + j, temp_y, temp_x);
+					}
+                    float f = getValueFromTensorPosLIMIT_FILTER_filter(c, j, fy, fx);
+                    res += f * o;
+                }
+                ox += dilation_h;
+            }
+            oy += dilation_v;
+        }
+        setOutput(res);
+    }
+`;

src/shader/conv2d_transpose/params.es6

+/* eslint-disable */
+/**
+ * @file 参数文件
+ * @author chenhaoze
+ */
+export default `
+    // conv2d的input数据
+    
+    // 常量
+    // 卷积核
+    const int length_shape_filter = LENGTH_SHAPE_FILTER;
+    const int width_shape_filter = WIDTH_SHAPE_FILTER;
+    const int height_shape_filter = HEIGHT_SHAPE_FILTER;
+    const int width_texture_filter = WIDTH_TEXTURE_FILTER;
+    const int height_texture_filter = HEIGHT_TEXTURE_FILTER;
+    const int channel_filter = CHANNEL_FILTER;
+    
+    // 输入数据
+    const int width_shape_origin = WIDTH_SHAPE_ORIGIN;
+    const int height_shape_origin = HEIGHT_SHAPE_ORIGIN;
+    const int length_shape_origin = LENGTH_SHAPE_ORIGIN;
+    const int width_texture_origin = WIDTH_TEXTURE_ORIGIN;
+    const int height_texture_origin = HEIGHT_TEXTURE_ORIGIN;
+    const int channel_origin = CHANNEL_ORIGIN;
+    
+    // 计算相关
+    // 拆分步长
+    const int stride_h = int(STRIDES_X);
+    const int stride_v = int(STRIDES_Y);
+    // padding的数目
+    const int padLeft = width_shape_filter - PADDINGS_X - 1;
+    const int padTop = height_shape_filter - PADDINGS_Y - 1;
+    // dilation膨胀系数
+    const int dilation_h = DILATION_H;
+    const int dilation_v = DILATION_V;
+    // groups
+    const int groups = GROUPS;
+    
+    // uniform变量
+    // 卷积核
+    uniform sampler2D texture_filter;
+    
+    // 输入数据
+    uniform sampler2D texture_origin;
+`;

src/shader/elementwise_add/conf.es6

@@ -5,16 +5,22 @@
  */
 export default {
     dep: [
-        {
-            func: 'getPixelsFromTexturePos',
+		{
+            func: 'getValueFromTensorPos',
             conf: {
-                TEXTURE_NAME: 'texture_origin'
+                TENSOR_NAME: 'origin'
             }
         },
         {
-            func: 'getPixelsFromTexturePos',
+			func: 'transferFromNHWCtoNCHW',
+			conf:{
+
+			}
+		},
+		{
+            func: 'getValueFromTensorPos',
             conf: {
-                TEXTURE_NAME: 'texture_counter'
+                TENSOR_NAME: 'counter'
             }
         }
     ],
@@ -26,7 +32,12 @@ export default {
         'HEIGHT_TEXTURE_ORIGIN',
         'CHANNEL_ORIGIN',
 
-        'TOTAL_SHAPE_COUNTER',
+		'WIDTH_SHAPE_COUNTER',
+        'HEIGHT_SHAPE_COUNTER',
+        'LENGTH_SHAPE_COUNTER',
+        'WIDTH_TEXTURE_COUNTER',
+        'HEIGHT_TEXTURE_COUNTER',
+        'CHANNEL_COUNTER',
 
         'WIDTH_SHAPE_OUT',
         'HEIGHT_SHAPE_OUT',
@@ -35,10 +46,7 @@ export default {
         'CHANNEL_OUT',
         'OFFSET_Y_OUT',
 
-        'AXIS',
-        'MULTI_VALUE',
-        'BIAS_VALUE',
-        'ACTIVE_FUNCTION'
+        'AXIS'
     ],
     input: [
         {
@@ -47,11 +55,11 @@ export default {
             setter: 'initTexture',
             type: 'texture'
         },
-        {
+		{
             tensor: 'counter',
-            variable: 'data',
-            setter: 'uniform1fv',
-            type: 'uniform'
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
         }
     ]
 };

src/shader/elementwise_add/main.es6

@@ -7,11 +7,23 @@ export default `
 // start函数
 void main(void) {
     // 输出数据
-    ivec4 oPos = getOutputTensorPosLIMIT_OUT();
-    int index = oPos[axis];
-    float o = getPixelsFromTexturePos_texture_origin(vCoord).r;
-    float c = getValueFromCounter(index);
-    float res = ACTIVE_FUNCTION(o + c, multi_value, bias_value);
-    setOutput(res);
+    ivec4 oPos = getOutputTensorPos();
+    int sumVal = oPos.g + oPos.a * channel_origin + oPos.b * channel_origin * width_shape_origin + oPos.r * channel_origin * width_shape_origin * height_shape_origin;
+    float o = getValueFromTensorPos_origin(oPos.r, oPos.g, oPos.b, oPos.a);
+	ivec4 pos_counter;
+	pos_counter.r = channel_out;
+	pos_counter.g = height_shape_origin;
+	pos_counter.b = width_shape_origin;
+	pos_counter.a = 1;
+	int index = 0;
+	for (int i = 4 - length_shape_origin + axis; i < 4 - length_shape_origin + axis + length_shape_counter; i++ ){
+		if (index > 0) {
+			index = index * pos_counter[i];
+		}
+		index += oPos[i];
+	}
+	float c = getValueFromTensorPos_counter(oPos.r, oPos.g, oPos.b, oPos.a);
+	float res = c + o;
+	setOutput(res);
 }
 `;

src/shader/elementwise_add/params.es6

@@ -6,15 +6,21 @@
 export default `
     // 输入数据
     const int axis = AXIS;
-    // const int total_shape_counter = TOTAL_SHAPE_COUNTER;
-    uniform float data_counter[TOTAL_SHAPE_COUNTER];
+    const int width_shape_origin = WIDTH_SHAPE_ORIGIN;
+    const int height_shape_origin = HEIGHT_SHAPE_ORIGIN;
+    const int length_shape_origin = LENGTH_SHAPE_ORIGIN;
+    const int width_texture_origin = WIDTH_TEXTURE_ORIGIN;
+    const int height_texture_origin = HEIGHT_TEXTURE_ORIGIN;
+    const int channel_origin = CHANNEL_ORIGIN;
+
+    const int height_shape_counter = HEIGHT_SHAPE_COUNTER;
+    const int width_shape_counter = WIDTH_SHAPE_COUNTER;
+    const int length_shape_counter = LENGTH_SHAPE_COUNTER;
+    const int width_texture_counter = WIDTH_TEXTURE_COUNTER;
+    const int height_texture_counter = HEIGHT_TEXTURE_COUNTER;
+    const int channel_counter = CHANNEL_COUNTER;
+    
     uniform sampler2D texture_origin;
-    float getValueFromCounter(int index) {
-        for (int i = 0; i < TOTAL_SHAPE_COUNTER; i++) {
-            if (i == index) {
-                return data_counter[i];
-            }
-        }
-        return 0.0;
-    }
+	uniform sampler2D texture_counter;
+
 `;

src/shader/mul/conf.es6

-graph.es6/* eslint-disable */
+/* eslint-disable */
 /**
  * @file mul的配置文件
  * @author yangmingming zhangmiao06

src/shader/pool2d/main.es6

@@ -5,7 +5,7 @@
 export default `
 // start函数
 void main(void) {
-    float res = (-1.0 / exp(-20.0));
+	float res = 0.0;
     // 获取output的坐标
     ivec4 out_pos = getOutputTensorPosLIMIT_OUT();
     // X、Y方向的移动步长

src/shader/reshape/conf.es6

+/* eslint-disable */
+/**
+ * @file batchnorm的配置文件
+ * @author chenhaoze
+ */
+export default {
+    dep: [
+		{
+			func: 'getValueFromTensorPos',
+			conf: {
+				TENSOR_NAME: 'origin'
+			}
+		},
+		{
+			func: 'transferFromNHWCtoNCHW',
+			conf:{
+
+			}
+		}
+    ],
+    conf: [
+		'WIDTH_SHAPE_ORIGIN',
+		'HEIGHT_SHAPE_ORIGIN',
+		'LENGTH_SHAPE_ORIGIN',
+		'WIDTH_TEXTURE_ORIGIN',
+		'HEIGHT_TEXTURE_ORIGIN',
+		'CHANNEL_ORIGIN',
+
+		'WIDTH_SHAPE_OUT',
+		'HEIGHT_SHAPE_OUT',
+		'WIDTH_TEXTURE_OUT',
+		'HEIGHT_TEXTURE_OUT',
+		'CHANNEL_OUT',
+		'OFFSET_Y_OUT',
+
+		'MULTI_VALUE',
+		'BIAS_VALUE',
+		'ACTIVE_FUNCTION'
+    ],
+    input: [
+			{
+				tensor: 'origin',
+				variable: 'texture',
+				setter: 'initTexture',
+				type: 'texture'
+			}
+    ]
+};

src/shader/reshape/main.es6

+/* eslint-disable */
+/**
+ * @file reshape主函数
+ * @author chenhaoze
+ */
+export default `
+// start函数
+void main(void) {
+    // 输出数据
+	ivec4 oPos = getOutputTensorPos();
+    // 输出坐标转换为输入坐标
+	int sumVal = oPos.g + oPos.a * channel_out + oPos.b * channel_out * width_shape_out + oPos.r * channel_out * width_shape_out * height_shape_out;
+ 	ivec4 new_oPos = transferFromNHWCtoNCHW(sumVal, channel_origin, width_shape_origin, height_shape_origin, total_shape_origin);
+	float o = getValueFromTensorPos_origin(new_oPos.r, new_oPos.g, new_oPos.b, new_oPos.a);
+	setOutput(float(o));
+}
+`;

src/shader/reshape/params.es6

+/* eslint-disable */
+/**
+ * @file batchnorm参数文件
+ * @author chenhaoze
+ */
+export default `
+// 输入数据
+const int width_shape_origin = WIDTH_SHAPE_ORIGIN;
+const int height_shape_origin = HEIGHT_SHAPE_ORIGIN;
+const int length_shape_origin = LENGTH_SHAPE_ORIGIN;
+const int width_texture_origin = WIDTH_TEXTURE_ORIGIN;
+const int height_texture_origin = HEIGHT_TEXTURE_ORIGIN;
+const int channel_origin = CHANNEL_ORIGIN;
+const int total_shape_origin = TOTAL_SHAPE_ORIGIN;
+
+
+
+// 输入数据
+ uniform sampler2D texture_origin;
+`;

src/shader/softmax/conf.es6

@@ -6,17 +6,19 @@
 export default {
     dep: [
         {
-            func: 'getPixelsFromTexturePos',
+            func: 'getValueFromTensorPos',
             conf: {
-                TEXTURE_NAME: 'texture_origin'
+                TENSOR_NAME: 'origin'
             }
         }
     ],
     conf: [
+        'WIDTH_SHAPE_ORIGIN',
+        'HEIGHT_SHAPE_ORIGIN',
+        'LENGTH_SHAPE_ORIGIN',
         'WIDTH_TEXTURE_ORIGIN',
         'HEIGHT_TEXTURE_ORIGIN',
-        'TOTAL_SHAPE_ORIGIN',
-        'OFFSET_Y_OUT'
+        'CHANNEL_ORIGIN',
     ],
     input: [
         {

src/shader/softmax/main.es6

@@ -6,55 +6,40 @@
 export default `
 // start函数
 void main(void) {
-    float res = 0.0;
-    vec4 v4 = getPixelsFromTexturePos_texture_origin(vCoord);
-    vec2 onePixel = vec2(1.0 / float(width_texture_origin), 1.0 / float(height_texture_origin));
+    ivec4 oPos = getOutputTensorPos();
+    const int n = int(total_shape_origin/channel_origin/height_shape_origin/width_shape_origin);
+    float o = getValueFromTensorPos_origin(oPos[0], oPos[1], oPos[2], oPos[3]);
+    // 输出坐标转换为输入坐标
     float total = 0.0;
-    float maxValue = getPixelsFromTexturePos_texture_origin(onePixel).r;
-    int number = 0;
-    vec4 pixels;
-    vec4 result;
-    // 求最大
-    for (int i = 0; i < height_texture_origin; i++) {
-        for (int j = 0; j < width_texture_origin; j++) {
-            pixels = getPixelsFromTexturePos_texture_origin(onePixel * vec2(float(j), float(i)));
-            number = i * width_texture_origin + j;
-            if ((number * 4 + 1) < total_shape_origin) {
-                maxValue = max(pixels.r, maxValue);
-            }
-            if ((number * 4 + 2) < total_shape_origin) {
-                maxValue = max(pixels.g, maxValue);
-            }
-            if ((number * 4 + 3) < total_shape_origin) {
-                maxValue = max(pixels.b, maxValue);
-            }
-            if ((number * 4 + 4) < total_shape_origin) {
-                maxValue = max(pixels.a, maxValue);
-            }
+    float res = 0.0;
+    if (axis == 0) {
+        for (int i = 0; i < n; i++){
+        float temp = getValueFromTensorPos_origin(i, oPos[1], oPos[2], oPos[3]);
+        total += exp(temp);
+        }
+        res = exp(o) / total;
+    }
+    else if (axis == 1) {
+        for (int i = 0; i < channel_origin; i++){
+        float temp = getValueFromTensorPos_origin(oPos[0], i, oPos[2], oPos[3]);
+        total += exp(temp);
+        }
+        res = exp(o) / total;
+    }
+    else if (axis == 2) {
+        for (int i = 0; i < height_shape_origin; i++){
+        float temp = getValueFromTensorPos_origin(oPos[0], oPos[1], i, oPos[3]);
+        total += exp(temp);
         }
+        res = exp(o) / total;
     }
-    // 求和
-    for (int i = 0; i < height_texture_origin; i++) {
-        for (int j = 0; j < width_texture_origin; j++) {
-            pixels = getPixelsFromTexturePos_texture_origin(onePixel * vec2(float(j), float(i)));
-            number = i * width_texture_origin + j;
-            if ((number * 4 + 1) < total_shape_origin) {
-                total += exp(pixels.r - maxValue);
-            }
-            if ((number * 4 + 2) < total_shape_origin) {
-                total += exp(pixels.g - maxValue);
-            }
-            if ((number * 4 + 3) < total_shape_origin) {
-                total += exp(pixels.b - maxValue);
-            }
-            if ((number * 4 + 4) < total_shape_origin) {
-                total += exp(pixels.a - maxValue);
-            }
+    else {
+        for (int i = 0; i < width_shape_origin; i++){
+        float temp = getValueFromTensorPos_origin(oPos[0], oPos[1], oPos[2], i);
+        total += exp(temp);
         }
+        res = exp(o) / total;
     }
-    outColor = exp(v4 - vec4(maxValue, maxValue, maxValue, maxValue)) / vec4(total, total, total, total);
-    
-    // res = result.a;
-    // setOutput(res);
+    setOutput(res);
 }
 `;

src/shader/softmax/params.es6

@@ -5,10 +5,13 @@
  */
 export default `
 // 输入数据
+const int width_shape_origin = WIDTH_SHAPE_ORIGIN;
+const int height_shape_origin = HEIGHT_SHAPE_ORIGIN;
 const int width_texture_origin = WIDTH_TEXTURE_ORIGIN;
 const int height_texture_origin = HEIGHT_TEXTURE_ORIGIN;
 const int total_shape_origin = TOTAL_SHAPE_ORIGIN;
-
+const int channel_origin = CHANNEL_ORIGIN;
+const int axis = AXIS;
 // uniform变量
 // 输入数据
 uniform sampler2D texture_origin;

src/shader/split/conf.es6

+/* eslint-disable */
+/**
+ * @file split的配置文件
+ * @author zhangjingyuan02
+ */
+export default {
+    dep: [
+		{
+			func: 'getValueFromTensorPos',
+			conf: {
+				TENSOR_NAME: 'origin'
+			}
+		},
+		{
+			func: 'transferFromNHWCtoNCHW'
+		}
+    ],
+    conf: [
+		'WIDTH_SHAPE_ORIGIN',
+		'HEIGHT_SHAPE_ORIGIN',
+		'LENGTH_SHAPE_ORIGIN',
+		'WIDTH_TEXTURE_ORIGIN',
+		'HEIGHT_TEXTURE_ORIGIN',
+		'CHANNEL_ORIGIN',
+
+		'WIDTH_SHAPE_OUT',
+		'HEIGHT_SHAPE_OUT',
+		'WIDTH_TEXTURE_OUT',
+		'HEIGHT_TEXTURE_OUT',
+		'CHANNEL_OUT',
+		'OFFSET_Y_OUT',
+    ],
+    input: [
+        {
+            tensor: 'origin',
+            variable: 'texture',
+            setter: 'initTexture',
+            type: 'texture'
+        }
+    ]
+};

src/shader/split/main.es6

+/* eslint-disable */
+/**
+ * @file split主函数
+ * @author zhangjingyuan02
+ */
+export default `
+// start函数
+void main(void) {
+    int length = int(target_value.length() / num);
+    ivec4 oPos = getOutputTensorPos();
+    // 输出坐标转换为输入坐标
+	int sumVal = oPos.g + oPos.a * channel_out + oPos.b * channel_out * width_shape_out + oPos.r * channel_out * width_shape_out * height_shape_out;
+    ivec4 new_oPos = transferFromNHWCtoNCHW(sumVal, channel_out, width_shape_out, height_shape_out, total_shape_out);
+    new_oPos[dim] = new_oPos[dim] + layer_run_time * length;
+	float o = getValueFromTensorPos_origin(new_oPos.r, new_oPos.g, new_oPos.b, new_oPos.a);
+	setOutput(float(o));
+}
+`;

src/shader/split/params.es6

+/* eslint-disable */
+/**
+ * @file split参数文件
+ * @author zhangjingyuan02
+ */
+export default `
+// 常量
+
+const int width_shape_origin = WIDTH_SHAPE_ORIGIN;
+const int height_shape_origin = HEIGHT_SHAPE_ORIGIN;
+const int length_shape_origin = LENGTH_SHAPE_ORIGIN;
+const int width_texture_origin = WIDTH_TEXTURE_ORIGIN;
+const int height_texture_origin = HEIGHT_TEXTURE_ORIGIN;
+const int channel_origin = CHANNEL_ORIGIN;
+const int total_shape_origin = TOTAL_SHAPE_ORIGIN;
+
+const int total_shape_out = TOTAL_SHAPE_OUT;
+
+const int dim = DIM;
+const int num = NUM;
+const int target_value[TARGET_LENGTH] = int[](TARGET_VALUE);
+
+
+// 输入数据
+uniform sampler2D texture_origin;
+`;

src/shader/transpose/conf.es6

+/* eslint-disable */
+/**
+ * @file batchnorm的配置文件
+ * @author chenhaoze
+ */
+export default {
+    dep: [
+		{
+			func: 'getValueFromTensorPos',
+			conf: {
+				TENSOR_NAME: 'origin'
+			}
+		},
+		{
+			func: 'transferFromNHWCtoNCHW',
+			conf:{
+
+			}
+		}
+    ],
+    conf: [
+		'WIDTH_SHAPE_ORIGIN',
+		'HEIGHT_SHAPE_ORIGIN',
+		'LENGTH_SHAPE_ORIGIN',
+		'WIDTH_TEXTURE_ORIGIN',
+		'HEIGHT_TEXTURE_ORIGIN',
+		'CHANNEL_ORIGIN',
+
+		'WIDTH_SHAPE_OUT',
+		'HEIGHT_SHAPE_OUT',
+		'WIDTH_TEXTURE_OUT',
+		'HEIGHT_TEXTURE_OUT',
+		'CHANNEL_OUT',
+		'OFFSET_Y_OUT',
+
+		'MULTI_VALUE',
+		'BIAS_VALUE',
+		'ACTIVE_FUNCTION'
+    ],
+    input: [
+			{
+				tensor: 'origin',
+				variable: 'texture',
+				setter: 'initTexture',
+				type: 'texture'
+			}
+    ]
+};

src/shader/transpose/main.es6

+/* eslint-disable */
+/**
+ * @file reshape主函数
+ * @author chenhaoze
+ */
+export default `
+// start函数
+void main(void) {
+    // 输出数据
+	ivec4 oPos = getOutputTensorPos();
+	    // 重排遍历顺序
+	int sumVal = oPos.g + oPos.a * channel_out + oPos.b * channel_out * width_shape_out + oPos.r * channel_out * width_shape_out * height_shape_out;
+	ivec4 new_oPos = transferFromNHWCtoNCHW(sumVal, channel_out, width_shape_out, height_shape_out, total_shape_origin);
+
+	// 转置 坐标变换
+	oPos = new_oPos;
+	float o = 0.0;
+	if (perm_size == 1) {
+		o = getValueFromTensorPos_origin(oPos[0], oPos[1], oPos[2], oPos[3]);
+	}
+	else if (perm_size == 2) {
+		o = getValueFromTensorPos_origin(oPos[0], oPos[1], oPos[min(2 + perm_0, 3)], oPos[min(2 + perm_1, 3)]);
+	}
+	else if (perm_size == 3) {
+		o = getValueFromTensorPos_origin(oPos[0], oPos[min(1 + perm_0, 3)], oPos[min(1 + perm_1, 3)], oPos[min(1 + perm_2, 3)]);
+	}
+	else if (perm_size == 4) {
+		o = getValueFromTensorPos_origin(oPos[perm_0], oPos[perm_1], oPos[perm_2], oPos[perm_3]);
+	}
+
+
+	setOutput(float(o));
+}
+`;

src/shader/transpose/params.es6

+/* eslint-disable */
+/**
+ * @file batchnorm参数文件
+ * @author chenhaoze
+ */
+export default `
+// 输入数据
+const int width_shape_origin = WIDTH_SHAPE_ORIGIN;
+const int height_shape_origin = HEIGHT_SHAPE_ORIGIN;
+const int length_shape_origin = LENGTH_SHAPE_ORIGIN;
+const int width_texture_origin = WIDTH_TEXTURE_ORIGIN;
+const int height_texture_origin = HEIGHT_TEXTURE_ORIGIN;
+const int channel_origin = CHANNEL_ORIGIN;
+const int total_shape_origin = TOTAL_SHAPE_ORIGIN;
+
+
+const int perm_size = PERM_SIZE;
+const int perm_0 = PERM_0;
+const int perm_1 = PERM_1;
+const int perm_2 = PERM_2;
+const int perm_3 = PERM_3;
+
+// 输入数据
+ uniform sampler2D texture_origin;
+`;

src/shader/v_shader.es6

 /* eslint-disable */
 /**
  * @file 顶点文件
- * @author wangqun
- * @desc  顶点坐标系转换，适配webgl1
+ * @author yangmingming
  */
 export default `
 attribute vec4 position;
 varying vec2 vCoord;
+
 void main() {
     vCoord.x = (position.x + 1.0) / 2.0;
     vCoord.y = (position.y + 1.0) / 2.0;

src/shader/v_shader2.es6

 /* eslint-disable */
 /**
  * @file 顶点文件,webgl 2.0
- * @author wangqun
- * @desc  顶点坐标系转换，适配webgl2
+ * @author yangmingming
  */
 export default `#version 300 es
 in vec4 position;
 out vec2 vCoord;
+
 void main() {
     vCoord.x = (position.x + 1.0) / 2.0;
     vCoord.y = (position.y + 1.0) / 2.0;

src/utils/opData.es6

@@ -3,7 +3,7 @@ import Utils from './utils';
 import Tensor from './tensor';
 /**
  * @file op的数据对象
- * @author wangqun, yangmingming
+ * @author yangmingming
  *
  */
 const keys = [
@@ -57,13 +57,16 @@ const opBehavior = {
         'needBatch',
         'isApplySeparableConv'
     ],
+	conv2d_transpose: [
+		'needBatch'
+	],
     batchnorm: [
         'needBatch',
         'mergeTensor'
     ],
     elementwise_add: [
-        'broadcast',
-        'needBatch'
+        'needBatch',
+		'processAxis'
     ],
     conv2d_elementwise_add: [
         'mergeAttrs',
@@ -91,18 +94,34 @@ const opBehavior = {
     mul: [
         'reshape',
         'needBatch'
+    ],
+	reshape: [
+		'needBatch',
+		'inferShape'
+	],
+	transpose: [
+		'needBatch',
+		'setPerm'
+	],
+    concat: [
+        'normalizeDim',
+        'needBatch'
+    ],
+    split: [
+        'normalizeDim',
+        'needBatch'
     ],
     softmax: [
-    ]
+        'needBatch'
+    ],
+    scale: [
+        'needBatch'
+    ],
 };
 const mergeType = 'conv2d-elementwise_add';
+
 export default class OpData {
     constructor(name, input = {}, output = {}, attrs = {}) {
-    console.log('now in constructor');
-    console.dir(name);
-    console.dir(input);
-    console.dir(output);
-
         this.realName = name;
         this.name = name;
         this.attrs = attrs;
@@ -120,14 +139,46 @@ export default class OpData {
                 'multi_value': '1.0',
                 'bias_value': '0.0'
             };
+
             // tensor数据
-            this.tensor = {};
+            this.inputTensors = [];
+            this.outputTensors = [];
+            this.fShaderParams = [];
             this.buildTensor();
-            this.buildAttrs();
+            this.buildShaderParams();
         }
     }
 
+    inferShape(){
+		if (this.name == 'reshape'){
+			let inputShape = this.input.X[0].shape;
+			let targetShape = this.attrs.new_shape;
+			for (let i = 0; i < targetShape.length; i++){
+				if (targetShape[i] == 0) {
+					targetShape[i] = inputShape[i];
+            	}
+        	}
+			let total_length = 1;
+			for (let j = 0;j < inputShape.length; j++){
+				total_length *= inputShape[j];
+			}
+			let minusPos = -1;
+			for (let i = 0; i < targetShape.length; i++){
+				if (targetShape[i] == -1) {
+					minusPos = i;
+					continue;
+            	}
+            	total_length /= targetShape[i];
+        	}
+        	if (minusPos != -1) {
+				targetShape[minusPos] = total_length;
+			}
+			this.output.Out[0].shape = targetShape;
+		}
+	}
+
     buildTensor() {
+
         // todo: 是否需要形状对齐
         // todo: 是否需要广播tensor
         const tensorData = [];
@@ -150,8 +201,10 @@ export default class OpData {
                 // 默认取第一个数据
                 const data = this.output[key] || [{}];
                 if (tensorName[key.toLowerCase()]) {
-                    data[0].tensorName = tensorName[key.toLowerCase()];
-                    tensorData.push(data[0]);
+                    data.forEach(item => {
+                        item.tensorName = tensorName[key.toLowerCase()];
+                        tensorData.push(item);
+                    });
                 }
             }
         }
@@ -160,20 +213,22 @@ export default class OpData {
         behavior.forEach(behavior => {
             this[behavior](tensorData);
         });
+
         // 生成tensor对象
         tensorData.forEach(data => {
-            // console.log(data);
             if (data) {
+                let tensor = null;
+                const tensorName = data.tensorName;
                 if (data.notTensor) {
-                    this.tensor[data.tensorName] = {
-                        name: data.tensorName,
+                    tensor = {
+                        name: tensorName,
                         data: new Float32Array(data.data),
                         total_shape: data.data.length
                     };
                 } else {
-                    this.tensor[data.tensorName] = new Tensor({
+                    tensor = new Tensor({
                         type: data.name,
-                        name: data.tensorName,
+                        name: tensorName,
                         shape: data.shape,
                         data: data.data,
                         needBatch: data.needBatch || false,
@@ -181,24 +236,25 @@ export default class OpData {
                         isPacked: data.isPacked || false
                     });
                 }
+
+                if (tensorName === 'out') {
+                    this.outputTensors.push(tensor);
+                }
+                else {
+                    this.inputTensors.push(tensor);
+                }
             }
         });
-        // console.dir(['tensors', this.tensor]);
-        // console.log('now in buildTensor show this and tensorData');
-        // console.log(this);
-        // console.log(tensorData);
     }
 
-    buildAttrs() {
+    buildShaderParams() {
         // 计算属性
         for (let key in this.attrs) {
             if (this.attrs.hasOwnProperty(key)) {
                 const item = this.attrs[key];
-                if (Object.prototype.toString.call(item) === '[object Array]') {
-                    if (keys.indexOf(key) > -1) {
-                        this.data[key + '_x'] = item[0];
-                        this.data[key + '_y'] = item[1];
-                    }
+                if (Object.prototype.toString.call(item) === '[object Array]' && keys.indexOf(key) > -1) {
+                    this.data[key + '_x'] = item[0];
+                    this.data[key + '_y'] = item[1];
                 } else {
                     this.data[key] = item;
                     // 获取shader所需的数据
@@ -209,19 +265,66 @@ export default class OpData {
                 }
             }
         }
-        // 获取tensor的数据
-        for (let key in this.tensor) {
-            const tensor = this.tensor[key];
+        // 遍历 获取input tensor的数据
+        this.inputTensors.forEach(inputTensor => {
             tensorAttrs.forEach(attr => {
-                this.data[attr+ '_' + tensor.name] = tensor[attr];
+                this.data[attr+ '_' + inputTensor.name] = inputTensor[attr];
             });
-        }
+        });
+
+        // 根据out tensor 个数 生成对应的 fShader 个数
+        this.outputTensors.forEach(outTensor => {
+            const params = JSON.parse(JSON.stringify(this.data));
+            // 获取output tensor的数据
+            tensorAttrs.forEach(attr => {
+                params[attr+ '_' + outTensor.name] = outTensor[attr];
+            });
+            this.fShaderParams.push(params);
+        });
+
     }
 
     needBatch(tensorData = []) {
         tensorData.forEach(data => (data.needBatch = true));
     }
 
+	setPerm(tensorData = []){
+		let arrayPerm = this.attrs['perm'];
+		let l = arrayPerm.length;
+		if (l == 3) {
+			if (arrayPerm == [2,0,1]) {
+				arrayPerm = [1,2,0];
+		}
+			else if (arrayPerm == [1,2,0]){
+				arrayPerm = [2,0,1];
+		}
+		}
+		else if (l == 4){
+			let temp = [0,0,0,0];
+			for (let i = 0; i < 4; i++){
+				temp[[arrayPerm[i]]] = i;
+			}
+			arrayPerm = temp;
+		}
+		this.data['perm_0'] = 0;
+		this.data['perm_1'] = 0;
+		this.data['perm_2'] = 0;
+		this.data['perm_3'] = 0;
+		if (l >= 1) {
+			this.data['perm_0'] = arrayPerm[0];
+		}
+		if (l >= 2) {
+			this.data['perm_1'] = arrayPerm[1];
+		}
+		if (l >= 3) {
+			this.data['perm_2'] = arrayPerm[2];
+		}
+		if (l >= 4) {
+			this.data['perm_3'] = arrayPerm[3];
+		}
+		this.data['perm_size'] = l;
+	}
+
     isGlobalPooling(tensorData = []) {
         let counter = tensorData.filter(tensor => (tensor.tensorName === 'origin'))[0] || {};
         let length = counter.shape && counter.shape.length || 0;
@@ -236,6 +339,7 @@ export default class OpData {
         }, {});
     }
 
+
     isApplyWinoGrad(tensorData = []) {
         const filter = tensorData.filter(item => {
             const [b, c, h, w] = item.shape;
@@ -292,36 +396,6 @@ export default class OpData {
         });
     }
 
-    broadcast(tensorData = []) {
-        tensorData.forEach(item => {
-            if (item.tensorName === 'counter') {
-                item.notTensor = true;
-            }
-        });
-
-        return;
-
-        // mobilenet model
-        // todo: 默认y的shape length是1, 以后需要实现通用版本
-console.log('2. x and y is ');
-console.log(x);
-console.log(y);
-        let shape = Utils.getBroadcastShapeInPaddle(x.shape, y.shape, this.attrs['axis']);
-        // 填充shape数据
-        if (small.shape.length === 1) {
-            const result = [];
-            small.shape = shape;
-            let total = shape.reduce((all, num) => all * num);
-            for (let i = 0; i < small.shape[0]; i++) {
-                let item = small.data[i];
-                for (let j = 0; j < total / shape[0]; j++) {
-                    result.push(item);
-                }
-            }
-            small.data = result;
-        }
-    }
-
     isMax(tensorData = []) {
         const type = this.attrs['pooling_type'] === 'max' ? 1 : 0;
         this.attrs['pooling_type'] = type;
@@ -355,6 +429,37 @@ console.log(y);
         }
     }
 
+    normalizeDim() {
+        const origin_shape = this.input.X[0].shape;
+        const axis = this.attrs.axis > -1 ? this.attrs.axis : origin_shape.length + this.attrs.axis;
+        const dim_value = [];
+        for (let index = 0; index < origin_shape[axis]; index++) {
+            dim_value[index] = index;
+        }
+        this.attrs.target_length = dim_value.length;
+        this.attrs.target_value = dim_value;
+        // 保存 输入 tensor 对应dim 的长度
+        this.attrs.inputs_dim = [origin_shape[axis]];
+        this.attrs.dim = 4 - origin_shape.length + axis;
+    }
+
+    processAxis() {
+		let shape_x = this.input.X[0].shape;
+		let shape_y = this.input.Y[0].shape;
+		let y_length = shape_y.length;
+		for (let i = shape_y.length - 1; i >=0 ;i--){
+			if (shape_y[i] == 1) {
+				y_length -= 1;
+			}
+		}
+		let axis_temp = this.attrs['axis'];
+		if (axis_temp == -1) {
+			this.attrs['axis'] = shape_x.length - y_length;
+		}
+		this.attrs['shape_length_origin'] = shape_x.length;
+		this.attrs['shape_length_counter'] = y_length;
+	}
+
     reshape(tensorData = []) {
         let input = tensorData[0];
         let counter = tensorData[1];
@@ -380,23 +485,23 @@ console.log(y);
             result[tensor.tensorName + 'Index'] = index;
         });
 
-       for (let i = 0; i < result[constants[0]].shape[0]; i++) {
-           data.push(result[constants[0]].data[i]);
-           data.push(result[constants[1]].data[i]);
-           data.push(result[constants[2]].data[i]);
-           data.push(result[constants[3]].data[i]);
-       }
+     //   for (let i = 0; i < result[constants[0]].shape[0]; i++) {
+      //      data.push(result[constants[0]].data[i]);
+      //      data.push(result[constants[1]].data[i]);
+     //       data.push(result[constants[2]].data[i]);
+      //      data.push(result[constants[3]].data[i]);
+      //  }
 
-        tensorData[result[constants[0] + 'Index']].data = data;
+        // tensorData[result[constants[0] + 'Index']].data = data;
         for (let i = 0; i < constants.length; i++){
             tensorData[result[constants[i] + 'Index']].data = result[constants[i]].data;
         }
         // 充分利用shader空间
-        tensorData[result[constants[0] + 'Index']].notCompressed = true;
-        tensorData[result[constants[0] + 'Index']].shape[0] *= 4;
-        tensorData.splice(result[constants[1] + 'Index'], 1, 0);
-        tensorData.splice(result[constants[2] + 'Index'], 1, 0);
-        tensorData.splice(result[constants[3] + 'Index'], 1, 0);
+        //tensorData[result[constants[0] + 'Index']].notCompressed = true;
+       // tensorData[result[constants[0] + 'Index']].shape[0] *= 4;
+        //tensorData.splice(result[constants[1] + 'Index'], 1, 0);
+        //tensorData.splice(result[constants[2] + 'Index'], 1, 0);
+        //tensorData.splice(result[constants[3] + 'Index'], 1, 0);
     }
 
     checkIsMerge() {

src/utils/tensor.es6

+/* eslint-disable */
 import Utils from './utils';
 /**
  * @file Tensor类
- * @author wangqun, yangmingming
+ * @author yangmingming
  */
 export default class Tensor {
     constructor(opts = {}) {
@@ -10,6 +11,8 @@ export default class Tensor {
         this.isPacked = this.isPacked || false;
         // 设置tensor名字
         this.name = opts.name;
+        // 设置 tensorId
+        this.tensorId = opts.type;
         // tensor的形状
         let shape = this.shape = opts.shape;
         // 原始数据个数
@@ -32,7 +35,6 @@ export default class Tensor {
         // tensor数据
         let data;
         if (opts.type === 'image' || opts.type === 'x') {
-            console.log('image', this.data);
             this.data = opts.data;
         }
         else if (opts.data && opts.data.length) {
@@ -167,3 +169,4 @@ export default class Tensor {
         }
     }
 }
+/* eslint-enable */

src/utils/utils.es6

 /**
  * @file 工具类
- * @author wangqun, yangmingming
+ * @author yangmingming
  */
+/* eslint-disable */
 export default {
     // todo: 适用2维矩阵乘法，以后实现通用版本
     getReshapeInPaddle(inputShape = [], counterShape = [], outShape = []) {
@@ -108,21 +109,22 @@ export default {
      * @return {{shape: *[], zeroNumber: number}} {Object} texture信息
      */
     getTextureInfoFromTensorShape(shape = [], isPacked = false) {
-        let b = shape[0] || 1;
-        let c = shape[1] || 1;
-        let h = shape[2] || 1;
-        let w = shape[3] || 1;
+        let b = shape[0];
+        let c = shape[1];
+        let h = shape[2];
+        let w = shape[3];
         let height = b * h;
         let width = c * w;
         let offsetX = 0;
         let offsetY = 0;
         // 安卓和ios的max texture size是4096, 改造存储空间(2bh, cw / 2)
         let exceedMax = false;
-        if (height > 4096 || width > 4096) {
-            height *= 2;
-            width = c * (Math.ceil(w / 2));
-            exceedMax = true;
-        }
+        // FIXME:为了让mobilenet能正常执行，这里先注释掉，待群哥修复
+        // if (height > MAX_TEXTURE_SIZE || width > MAX_TEXTURE_SIZE) {
+        //     height *= 2;
+        //     width = c * (Math.ceil(w / 2));
+        //     exceedMax = true;
+        // }
         if (isPacked) {
             // 紧凑布局
             height = b * c * Math.ceil(h / 2);
@@ -178,7 +180,109 @@ export default {
             let l = b1 * (c * h * w) + c1 * (h * w) + h1 * (w) + w1;
             data[offset] = renderData.data[l];
             offset += 4;
+            // data.push(renderData.data[l]);
+            // data.push(0);
+            // data.push(0);
+            // data.push(0);
         }
         renderData.data = data;
+    },
+    /*
+     * 将shape扩充到4维，在shape前补1
+     */
+    padToFourDimShape(shape) {
+        let fourDimShape = [];
+        if (shape.length == 4) {
+            fourDimShape = shape;
+        } else if (shape.length < 4) {
+            for (let i = 0; i < 4 - shape.length; i++) {
+                fourDimShape.push(1);
+            }
+            fourDimShape = fourDimShape.concat(shape);
+        }
+        return fourDimShape;
+    },
+
+    /* 
+     * 将nhwc排布数据转为nchw排布数据
+     */
+    nhwc2nchw(data, shape) {
+        let N = shape[0];
+        let H = shape[1];
+        let W = shape[2];
+        let C = shape[3];
+        let WXC = W * C;
+        let HXWXC = H * W * C;
+        let nchwData = [];
+        for (let n = 0; n < N; n++) {
+            for (let c = 0; c < C; c++) {
+                for (let h = 0; h < H; h++) {
+                    for (let w = 0; w < W; w++) {
+                        nchwData.push(data[n * HXWXC + h * WXC + w * C + c]);
+                    }
+                }
+            }
+        }
+        return nchwData;
+    },
+
+    /* 
+     * 将nchw排布数据转为nhwc排布数据
+     */
+    nchw2nhwc(data, shape) {
+        let N = shape[0];
+        let C = shape[1];
+        let H = shape[2];
+        let W = shape[3];
+        let HXW = H * W;
+        let CXHXW = C * H * W;
+        let nhwcData = [];
+        for (let n = 0; n < N; n++) {
+            for (let h = 0; h < H; h++) {
+                for (let w = 0; w < W; w++) {
+                    for (let c = 0; c < C; c++) {
+                        nhwcData.push(data[n * CXHXW + c * HXW + h * W + w]);
+                    }
+                }
+            }
+        }
+        return nhwcData;
+    },
+
+    /* 
+     * 等距间隔打印数据
+     */ 
+    stridePrint(data, count = 20) {
+        let realPrintCount = count;
+        if (data.length <= realPrintCount) {
+            this.continuousPrint(data, realPrintCount);
+            return;
+        }
+        let numbers = [];
+        let stride = Math.floor(data.length / realPrintCount);
+        if (stride == 0) {
+            stride = 1;
+        }
+        realPrintCount = Math.floor(data.length / stride)
+        for (let i = 0; i < realPrintCount; i++) {
+            numbers.push(i * stride + ": " + data[i * stride]);
+        }
+        console.log(numbers)
+    },
+
+    /* 
+     * 连续打印数据
+     */
+    continuousPrint(data, count = 100) {
+        let numbers = [];
+        let realPrintCount = count;
+        if (data.length <= realPrintCount) {
+            realPrintCount = data.length;
+        }
+        for (let i = 0; i < realPrintCount; i++) {
+            numbers.push(i + ": " + data[i]);
+        }
+        console.log(numbers)
     }
 };
+/* eslint-enable */