render pipeline + cs pipeline makes result incorrect

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]]]]

qjia7 / tfjs-core

render pipeline + cs pipeline makes result incorrect #17