feat(core & gl): 适配face detect & face landmark model

packages/paddlejs-backend-webgl/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@paddlejs/paddlejs-backend-webgl",
-  "version": "1.2.9",
+  "version": "1.2.10-beta.0",
   "description": "",
   "main": "lib/index",
   "scripts": {

packages/paddlejs-backend-webgl/src/backend.ts

@@ -409,7 +409,13 @@ export default class WebGLBackend extends PaddlejsBackend {
         return loc;
     }
 
-    dispose() {
-
+    dispose({ inputTensors }) {
+        for (const tensor of inputTensors) {
+            const tensorName = tensor.tensorId;
+            this.texturesMap[tensorName] && this.gl.deleteTexture(this.texturesMap[tensorName]);
+            this.cacheTextures[tensorName] && this.gl.deleteTexture(this.cacheTextures[tensorName]);
+            this.cacheTextures[tensorName] = null;
+            this.texturesMap[tensorName] = null;
+        }
     }
 }

packages/paddlejs-backend-webgl/src/ops/atom/common_func.ts

@@ -75,6 +75,11 @@ float exp_func(float x, float y, float z) {
     return result;
 }`;
 
+const abs_func = `
+float abs_func(float x, float y, float z) {
+    return abs(x);
+}`;
+
 export {
     prelu,
     relu6,
@@ -87,6 +92,7 @@ export {
     sqrt,
     pow_func,
     tanh_func,
+    abs_func,
     transferFromNHWCtoNCHW
 };
 

packages/paddlejs-backend-webgl/src/ops/atom/fuse_ops.ts

@@ -93,6 +93,10 @@ export default function genFuseOpCode(params: OpParams) {
                     act_name = 'exp_func';
                     break;
 
+                case 'abs':
+                    act_name = 'abs_func';
+                    break;
+
                 default:
                     break;
             }
@@ -127,7 +131,7 @@ export default function genFuseOpCode(params: OpParams) {
 
     return `
         ${activation_func}
-        
+
         float fuse_op(float x) {
             float res = x;
             ${calculation_str}

packages/paddlejs-backend-webgl/src/ops/index.ts

@@ -26,6 +26,8 @@ import batchnorm from './shader/batchnorm';
 import reshape2 from './shader/reshape2';
 import bilinear_interp from './shader/bilinear_interp';
 import bilinear_interp_v2 from './shader/bilinear_interp_v2';
+import linear_interp from './shader/linear_interp';
+import linear_interp_v2 from './shader/linear_interp_v2';
 import transpose2 from './shader/transpose2';
 import softmax from './shader/softmax';
 import dynamic from './shader/dynamic';
@@ -36,11 +38,12 @@ import arg_min from './shader/arg_min';
 import unsqueeze2 from './shader/unsqueeze2';
 import flatten_contiguous_range from './shader/flatten_contiguous_range';
 import greater_than from './shader/greater_than';
-import reduce_sum from './shader/reduce_sum';
 import where from './shader/where';
 import connect from './shader/connect';
 import squeeze2 from './shader/squeeze2';
 import pad3d from './shader/pad3d';
+import reduce_sum from './shader/reduce_sum';
+import reduce_max from './shader/reduce_max';
 import reduce_mean from './shader/reduce_mean';
 import shuffle_channel from './shader/shuffle_channel';
 import hard_swish from './shader/hard_swish';
@@ -97,6 +100,7 @@ const ops = {
     reshape: reshape2,
     reshape2,
     bilinear_interp,
+    linear_interp,
     transpose2,
     unpacked_2_packed,
     packed_2_unpacked,
@@ -105,9 +109,10 @@ const ops = {
     flatten2: reshape2,
     greater_than,
     reduce_sum,
+    reduce_mean,
+    reduce_max,
     where,
     connect,
-    reduce_mean,
     hard_swish,
     nearest_interp,
     nearest_interp_v2,
@@ -130,9 +135,11 @@ const ops = {
     sqrt: dynamic('sqrt'),
     tanh: dynamic('tanh'),
     exp: dynamic('exp'),
+    abs: dynamic('abs'),
     squeeze2,
     pad3d,
     bilinear_interp_v2,
+    linear_interp_v2,
     shuffle_channel,
     pack_out,
     nhwc_2_nchw,

packages/paddlejs-backend-webgl/src/ops/shader/batchnorm.ts

@@ -3,9 +3,13 @@
  */
 
 function mainFunc(
-    { bias, scale, mean, variance },
+    { bias, scale, mean, variance, origin },
     { epsilon }
 ) {
+    // calc the real pos of channel
+    const len = origin.length_unformatted_shape;
+    const realCPos = 4 - len + 1;
+    const posStr = `oPos[${realCPos}]`;
     return `
     // start函数
     void main(void) {
@@ -14,11 +18,14 @@ function mainFunc(
         float o = getValueFromTensorPos_origin(oPos.r, oPos.g, oPos.b, oPos.a);
 
         // 归一化数据
-        vec4 scale = getPixelsFromTexturePos_scale(vec2( float(oPos.g) / float(${scale.width_texture}) + 0.00001, 0.0));
-        vec4 bias = getPixelsFromTexturePos_bias(vec2( float(oPos.g) / float(${bias.width_texture}) + 0.00001, 0.0));
-        vec4 mean = getPixelsFromTexturePos_mean(vec2((float(oPos.g)) / float(${mean.width_texture})  + 0.00001, 0.0));
+        vec4 scale = getPixelsFromTexturePos_scale(
+            vec2(float(${posStr}) / float(${scale.width_texture}) + 0.00001, 0.0));
+        vec4 bias = getPixelsFromTexturePos_bias(
+            vec2(float(${posStr}) / float(${bias.width_texture}) + 0.00001, 0.0));
+        vec4 mean = getPixelsFromTexturePos_mean(
+            vec2((float(${posStr})) / float(${mean.width_texture}) + 0.00001, 0.0));
         vec4 variance = getPixelsFromTexturePos_variance(
-            vec2((float(oPos.g)) / float(${variance.width_texture}) + 0.00001,
+            vec2((float(${posStr})) / float(${variance.width_texture}) + 0.00001,
             0.0)
         );
 

packages/paddlejs-backend-webgl/src/ops/shader/dynamic.ts

@@ -13,7 +13,8 @@ const commonFuncBehaviors = {
     pow: ['transToPow'],
     exp: ['transToExp'],
     sqrt: ['transToSqrt'],
-    tanh: ['transToTanh']
+    tanh: ['transToTanh'],
+    abs: ['transToAbs']
 };
 
 function mainFunc(

packages/paddlejs-backend-webgl/src/ops/shader/linear_interp.ts

@@ -0,0 +1,58 @@
+/**
+ * @file bilinear_interp
+ */
+
+function mainFunc(
+    { out, origin },
+    { align_mode = 1, align_corners = true }
+) {
+    const outW = out.width_shape;
+    const inW = origin.width_shape;
+    const scale = align_corners ? (outW - 1) / (inW - 1) : (outW / inW);
+
+    return `
+    // start函数
+
+    vec4 getData(float n, float scale, bool align_flag, int in_len) {
+        float m = align_flag ? ((n + 0.5) / scale - 0.5) : (n / scale);
+        // int a1 = int(floor(m + 0.5));
+        int a1 = int(m);
+        a1 = a1 > 0 ? a1 : 0;
+        int a2 = a1 < (in_len - 1) ? (a1 + 1) : in_len;
+
+        float idx_src = (n + 0.5) / scale - 0.5;
+        idx_src = idx_src > 0.0 ? idx_src : 0.0;
+        float b1 = align_flag ? (idx_src - float(a1)) : (n / scale - float(a1));
+        float b2 = 1.0 - b1;
+        return vec4(float(a1), float(a2), b1, b2);
+    }
+
+    void main(void) {
+        // 输出数据
+        ivec4 oPos = getOutputTensorPos();
+
+        bool align_flag = ${align_mode} == 0 && !${align_corners};
+
+        float scale = float(${scale});
+
+        vec4 v = getData(float(oPos.a), scale, align_flag, ${inW});
+
+        float v1 = v.r;
+        float v2 = v.g;
+        float v3 = v.b;
+        float v4 = v.a;
+
+        float value1 = getValueFromTensorPos_origin(oPos.r, oPos.g, oPos.b, int(v1));
+        float value2 = getValueFromTensorPos_origin(oPos.r, oPos.g, oPos.b, int(v2));
+        float value = v4 * value1 + v3 * value2;
+        // setOutput(float(int(v1)));
+        setOutput(float(value));
+    }
+    `;
+}
+export default {
+    mainFunc,
+    textureFuncConf: {
+        origin: ['getValueFromTensorPos']
+    }
+};

packages/paddlejs-backend-webgl/src/ops/shader/linear_interp_v2.ts

@@ -0,0 +1,8 @@
+/**
+ * @file bilinear_interp_v2
+ * @implements 差值实现方式（BilinearInterpolation）与 bilinear_interp 一致，唯一的区别是 outshape 计算
+ */
+
+import linear_interp from './linear_interp';
+
+export default linear_interp;

packages/paddlejs-backend-webgl/src/ops/shader/pool2d_max.ts

@@ -22,7 +22,7 @@ function mainFunc(
                 index = ${originShape[2] * originShape[3]} * out_pos[1] + ${originShape[3]} * oy + ox;
             }
         `;
-        outputCode = 'setOutput(float(index));';
+        outputCode = 'setOutput(float(res));';
     }
     return `
     // start函数

packages/paddlejs-backend-webgl/src/ops/shader/reduce_max.ts

@@ -0,0 +1,32 @@
+
+/**
+ * @file concat
+ */
+
+function mainFunc(
+    {},
+    { inputs_dim, dim }
+) {
+    return `
+    // start函数
+    void main(void) {
+        ivec4 oPos = getOutputTensorPos();
+        // 输出坐标转换为输入坐标
+        float o = 0.0;
+        for (int i = 0; i < ${inputs_dim}; i++) {
+            oPos[${dim}] = i;
+            o = max(o, getValueFromTensorPos_origin(oPos.r, oPos.g, oPos.b, oPos.a));
+        }
+        setOutput(float(o));
+    }
+    `;
+}
+export default {
+    mainFunc,
+    textureFuncConf: {
+        origin: ['getValueFromTensorPos']
+    },
+    behaviors: [
+        'normalizeDim'
+    ]
+};

packages/paddlejs-backend-webgl/src/ops/shader/reshape2.ts

@@ -3,23 +3,35 @@
  */
 
 function mainFunc(
-    { out },
+    { origin, out },
     {}
 ) {
     return `
     // start函数
     void main(void) {
-        vec2 outCoord = vCoord.xy * (_2d_shape_texture_out);
-        int index = int(outCoord.x) + int(outCoord.y) * int(${out.width_texture});
-        ivec4 originPos = getTensorPosFromArrayIndex_origin(index);
-        float res = getValueFromTensorPos_origin(originPos[0], originPos[1], originPos[2], originPos[3]);
-        setOutput(res);
+        // 输出数据
+        ivec4 oPos = getOutputTensorPos();
+        // 输出坐标转换为输入坐标
+        int sumVal = oPos.a
+            + oPos.b * ${out.width_shape}
+            + oPos.g * ${out.height_shape} * ${out.width_shape}
+            + oPos.r * ${out.channel} * ${out.width_shape} * ${out.height_shape};
+        ivec4 new_oPos = transferFromNHWCtoNCHW(
+            sumVal,
+            ${origin.channel},
+            ${origin.width_shape},
+            ${origin.height_shape},
+            ${origin.total_shape}
+        );
+        float o = getValueFromTensorPos_origin(new_oPos.r, new_oPos.g, new_oPos.b, new_oPos.a);
+        setOutput(float(o));
     }
     `;
 }
 export default {
     mainFunc,
     textureFuncConf: {
-        origin: ['getTensorPosFromArrayIndex', 'getValueFromTensorPos']
-    }
+        origin: ['getValueFromTensorPos']
+    },
+    commonFuncConf: ['transferFromNHWCtoNCHW']
 };