Add code below into mobilenet/index.js to test.

let x = tf.tensor2d([1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24],[4,6]);
let k = tf.tensor2d([1,0,0,1,1,0,1,0,0,1,0,1],[6,2]);
x.matMul(k).print(1);

An incorrect result will occur.

Tensor
  dtype: float32
  rank: 2
  shape: [4,2]
  values:
    [[32, 13],
     [50, 31],
     [20, 49],
     [38, 67]]

TEST ENV:
Windows10, Chrome Canary(Version 77.0.3819.0), D3D11 backend.

CASE:

const inputShape = [1,3,3,1]; /*Ni,Hi,Wi,Ci*/
const filterShape = [2,2,1,1]; /*Hk,Wk,Ci,Co*/
const stride = 1;
const pad = 'valid';

const x = tf.tensor4d([1,2,3,4,5,6,7,8,9], inputShape);
const w = tf.tensor4d([1,1,1,1], filterShape);

const result = tf.conv2d(x, w, stride, pad);
result.print(1);

this case will use 3 ops:

• im2col: https://github.com/qjia7/tfjs-core/blob/compileAndRunCS/src/kernels/backend_webgl.ts#L1857
• matmul: https://github.com/qjia7/tfjs-core/blob/compileAndRunCS/src/kernels/backend_webgl.ts#L1895
• packTensor: https://github.com/qjia7/tfjs-core/blob/compileAndRunCS/src/kernels/backend_webgl.ts#L2302

if changing im2col or packTensor to render pipeline, result will be wrong.

wrong result:

Tensor
  dtype: float32
  rank: 4
  shape: [1,2,2,1]
  values:
    [[[[0],
       [0]],

      [[0],
       [0]]]]

In this branch https://github.com/GreyZzzzzzXh/tfjs-core/commits/webgl_pack, I switched all packed ops used in MobileNet to run on CS backend, and reverted MatMulPackedProgram, then we could get the correct result.
(Same with https://storage.googleapis.com/tfjs-examples/mobilenet/dist/index.html?tfjsflags=WEBGL_PACK:true).

But if we use the optimized packed mulmat, a wrong result will occur.

tfjs-examples/mnist could not run correctly on compute-context. It is related with that some operators could not work normally on compute-context.
These operators are as bleow,
batchMatMul, reduce, realDivide, add, subtract, exp, relu, conv2d, unpackTensor

Errors:
Uncaught (in promise) TypeError: Cannot read property '0' of null
at getSampler2DCS (shader_compiler.ts:1577)
at getSamplerFromInInfoCS (shader_compiler.ts:172)
at getSampler4DCS (shader_compiler.ts:2006)
at getSamplerFromInInfoCS (shader_compiler.ts:176)
at getInputSamplingSnippetCS (shader_compiler.ts:227)
at shader_compiler.ts:107
at Array.map ()
at makeCSShader (shader_compiler.ts:106)
at compileCSProgram (gpgpu_math.ts:135)
at backend_webgl.ts:2435

Sample code:
async function depthwiseConv2dTestSingleDepth_() {
const fSize = 2;
const pad = 'valid';
const stride = 1;
const chMul = 1;
const inDepth = 1;
ENV.set('WEBGL_PACK', false);
console.log(ENV.get('WEBGL_PACK'));
const x = tf.tensor4d(
[
0, 1, 2, 5, 0, 5,
0, 1, 5
],
[1, 3, 3, inDepth]);
const w = tf.tensor4d(
[1.0, 2.0, 2.0, 2.0],
[fSize, fSize, inDepth, chMul],
);

const result = tf.depthwiseConv2d(x, w, stride, pad);
result.print();
}

depthwiseConv2dTestSingleDepth_();