Hyeongseok Oh

### What Let's move libraries for test in `runtime/libs` into `tests/libs`. - [x] `runtime/libs/nnapi` - [x] Move - [x] `runtime/libs/rua` - [x] Move - [ ] `runtime/libs/tflite` - [x] Move...

Create and register observers on execution, not compilation. It is more flexible and allows for more dynamic observation (ex. profile, dump on demand) ONE-DCO-1.0-Signed-off-by: Hyeongseok Oh

Let's support full quantization on runtime - [x] Introduce full quantization type: https://github.com/Samsung/ONE/pull/11497 - [ ] Full quantization from circle model including minmax data & weight quantization - [ ]...

[onert/test] Generate random input on each run

1

### What Let's generate input random data per each run as default. And we can introduce additional argument to test driver to run with fixed input data on random input...

Call full quantization path on on-device quantization (use #11497)

[infra] Update HDF5 version

1

This commit updates HDF5 version to 1.14.4.2. This version does not require additional build tool execution for build, so we don't need patch for cross build. ONE-DCO-1.0-Signed-off-by: Hyeongseok Oh

[onert] Quantization type kernel for transformer

1

Below is required I/O quantization type (uint8/uint16) kernel for quantized transformer model - [ ] MUL - [x] UINT8 - [ ] INT16 - [ ] ADD - [x] UINT8...

[onert] Support multi-batch input inference for single batch model

2

Currently we are supporting multi-batch input inference for single batch model on shape inference routine. This feature is used for trix backend multi-batch inference and training input. But at first,...

[onert] In-place Reshape and ExpandDims operation

2

We are always using `memcpy` for `Reshape` and `ExpandDims` operation. If these operation's input is used only once and updated shape does not require padding, we can skip copy operation....

This commit allows to set input and output buffer data type to float when model is quantized. ONE-DCO-1.0-Signed-off-by: Hyeongseok Oh --- Draft: #13284