enable unpacked conv to work in mixed texture layout inferencing

microsoft / onnxjs

ONNX.js: run ONNX models using JavaScript

Other

1.75k stars 130 forks source link

enable unpacked conv to work in mixed texture layout inferencing #284

Closed xzhu1900 closed 3 years ago

xzhu1900 commented 3 years ago

This PR adds logic to enable unpacked conv op to work in an inference session where both packed kernel and unpacked kernel are used. Now we can inference the teams model with all supported packed kernel with correct result.

This PR also fixed a kernel cache bug.

xzhu1900 commented 3 years ago

reopen to trigger CI.