Age | Commit message (Collapse) | Author |
|
*Add the call to ReorderWeightChannelsForAcl for QuantizedSymm8PerAxis weights
Signed-off-by: Teresa Charlin <teresa.charlinreyes@arm.com>
Change-Id: I0d2e4c8328753ad15858130945b4fbb193943dfa
|
|
* Add conversion method to reverse bits in Shrink_Axis_Mask
* Add Unit tests for Neon, CL and Reference backends
* Fix supportedness of constant layer which is causing error
in DeepSpeech Uint8
* Also convert the Begin_Mask and End_Mask
Change-Id: I448b083c3463558e8fb5204923ab554cd43264ba
Signed-off-by: Francis Murtagh <francis.murtagh@arm.com>
|
|
* Variable-length array replaced with std::vector
* "and" replaced with "&&" (OK, this one is standard but MSVC doesn't
like it and it's inconsistent with the rest of the codebase)
* Replace u_int8_t with uint8_t
Change-Id: I7b968a2cfa85a5492a2a547364464c594efb067b
Signed-off-by: Rob Hughes <robert.hughes@arm.com>
|
|
Failing VTS tests were "NeuralnetworksHidlTest.depthwise_conv2d_*"
In depthwise convolution there was a difference in weight tensor channel
order between the reference and ACL implementations. This specifically related
to NCHW. This commit:
* Adds ReorderWeightChannelsForAcl to WorkloadUtils which will correct the weight tensor channel order.
* Add unit tests to detect this problem.
Signed-off-by: Colm Donelan <Colm.Donelan@arm.com>
Change-Id: Icaeac08e14b3d5da9e222ad2f118db55ebb15d09
|
|
* Unified ArmNN's weight format to [ M, I, H, W ] for the depthwise convolution
* Added conversion utilities to permute/reshape the weights as appropriate
when using CL and Neon backends
* Updated the reference implementation of the convolution
* Updated the relevant unit tests accordingly
!android-nn-driver:459
Change-Id: I07d0818efa9d1ca1e5dad82983aac1fe78eadb18
|