[WebGL] Failed to compile fragment shader

[Honry]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow.js): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): All platforms
• TensorFlow.js version (use command below): tf.js 3.7.0 webgl backend.
• Browser version: Chrome 93.0.4545.0

Describe the current behavior
WebGL backend throws Failed to compile fragment shader error while using depthwiseConv2d op with specific options. While WebGPU backend doesn't reproduce.

Describe the expected behavior
No error.

Standalone code to reproduce the issue

CodePen: https://codepen.io/honry/pen/wvdMmva

    const input = tf.fill([1, 65, 65, 960], 0);
    const weights = tf.fill([3, 3, 960, 1], 1);
    const result = tf.depthwiseConv2d(input, weights, 1, 'same', 'NHWC', 4);

Other info / logs Include any logs or source code that would be helpful to
diagnose the problem. If including tracebacks, please include the full
traceback. Large logs and files should be attached.

Please check error log from https://codepen.io/honry/pen/wvdMmva

[huningxin]

Thanks @Honry , how about the cpu backend and wasm backend?

[Honry]

@huningxin, this issue only reproduce on webgl backend.

[ColorfulCakeChen]

A similiar situation could be occurred by the following codes (i.e. depthwiseCon2d with ( strides == 2 ) ):

//await tf.setBackend( "cpu" ); // Error not happend in CPU.
await tf.setBackend( "webgl" ); // Error happend in WebGL.

let a = [
  111, 112, 113, 114, 121, 122, 123, 124, 131, 132, 133, 134,
  141, 142, 143, 144, 151, 152, 153, 154, 211, 212, 213, 214,
  221, 222, 223, 224, 231, 232, 233, 234, 241, 242, 243, 244,
  251, 252, 253, 254, 311, 312, 313, 314, 321, 322, 323, 324,
  331, 332, 333, 334, 341, 342, 343, 344, 351, 352, 353, 354];
let shape = [ 3, 5, 4 ];
let x = tf.tensor( a, shape );

let filters = tf.tensor( [ 11, 22, 33, 44 ], [ 1, 1, 4, 1 ] );

// When ( strides == 1 ), No problem.
let c = tf.depthwiseConv2d( x, filters, 1, "valid" );
c.print();
let d = tf.depthwiseConv2d( x, filters, 1, "same" );
d.print();

// When ( strides == 2 ), error: "Failed to compile fragment shader."
let e = tf.depthwiseConv2d( x, filters, 2, "valid" );
e.print();

// When ( strides == 2 ), error: "Failed to compile fragment shader."
let f = tf.depthwiseConv2d( x, filters, 2, "same" );
f.print();

[google-ml-butler]

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

[Honry]

@pyu10055, many thanks for fixing this issue! The compile error had been fixed, but the computed result looks wrong.

Here's a simple test, I just compared the first value of the output buffer, which is obviously different from result from wasm and cpu backends.

https://codepen.io/honry/pen/wvdMmva

[Honry]

@pyu10055, friendly ping. :) Should I create a separate issue for incorrect computational result?

[rthadur]

@Honry please open a new issue. Thank you

[Honry]

Thanks @rthadur, I've created a new issue at #5378.