Thibaut Goetghebuer-Planchon

[TFLite] Move AVERAGE_POOL_3D and MAX_POOL_3D to builtin ops

25

Hi, This PR moves the custom `AvgPool3D` and `MaxPool3D` ops to bulitin `AVERAGE_POOL_3D` and `MAX_POOL_3D` ops (`AVG` is changed to `AVERAGE` for coherence with `AVERAGE_POOL_2D`). As discussed previously @jianlijianli we...

Add a CMake `TSL_ROBIN_MAP_ENABLE_INSTALL` option to explicitly enable/disable the install target. The option is OFF by default when robin-map is used as subproject. Fix issue #66

[TOSA] Limit ReLU clamp_max attribute to the max value of the quantized type

9

Hi, The `max_val` of [tosa.clamp](https://www.mlplatform.org/tosa/tosa_spec.html#_clamp) is an int8/int16 value, it should thus be restricted to the int8/int16 max value during legalization which the legalization does.

[TOSA] On BatchToSpace legalization, use slice input shape to fix slicing dynamic size

9

Hi, This PR improves the `tfl.batch_to_space_nd` TFL -> TOSA legalization when the input has dynamic shapes by propagating the inferred input size of the slicing.

[TOSA] Separate left shifting in SQUARED_DIFFERENCE to avoid precision loss

12

Hi, This PR moves the left shifting in the `tfl.squared_difference` TFL -> TOSA legalization in a separate `tosa.rescale` to avoid any precision loss which could create non bit-exact results.

[TFLite] Quantization and unfolding of the ADD_N operator

19

Hello, When adding multiple tensors (> 2) in a NN we can either use `tf.keras.layers.Add` or `tf.math.add_n`. The first one is quantizable but not the second one as unfortunatly even...

Add support for upsample_nearest2d op in the Arm backend

6

Add support for the `torch.ops.aten.upsample_nearest2d.vec` operator in the Arm backend. The `torch.ops.aten.upsample_nearest2d.vec` operator is added to the operators to not decompose so that it can directly be legalized to a...