Broken tensor on specific GPU

System information

• Have I written custom code: https://codesandbox.io/s/tfjs-unstack-issue-loz5r?file=/src/index.js

• OS Platform and Distribution: Windows 10 GPU: WebGL:

• TensorFlow.js version: 3.0.0

• Browser version: Chrome 88.0.4324.146

Describe the current behavior With WebGL backend this code returns broken tensor:

  const frameSourceRaw = tf.tensor3d(
    new Int32Array(Array(128 * 224 * 4).fill(225)),
    [128, 224, 4],
    "int32"
  );
  const unstacked = tf.unstack(frameSourceRaw, -1);
  unstacked.forEach((a) => a.print());

is printed:

    [[225, 225, 225, ..., 225, 225, 225],
     [225, 225, 225, ..., 225, 225, 225],
     [225, 225, 225, ..., 225, 225, 225],
     ...,
     [0  , 0  , 0  , ..., 0  , 0  , 0  ],
     [0  , 0  , 0  , ..., 0  , 0  , 0  ],
     [0  , 0  , 0  , ..., 0  , 0  , 0  ]]

while frameSourceRaw is correct. This behavior does not appear only after unstacked operation, but also after other (slice, div). More examples: https://codesandbox.io/s/tfjs-unstack-issue-loz5r?file=/src/index.js

Zeros appear after 65536 (2^16) element. With CPU backend or with other GPU or with other array dimensions (eg. 144 * 256) output tensor is correct.

[pyu10055]

Looks like the tensor texture might be exceeding the max texture limit 16384, which translate to 16384 * 4 = 65536, since we are packing the tensor into 4 channels of the texture. Can you try to disable the packing and see if the 0 values appears at 1/4 of the size.

tf.env().set('WEBGL_PACK', false)

[google-ml-butler[bot]]

Are you satisfied with the resolution of your issue? Yes No