Create libneuron_performance.a
dos2unix and clang-format
Squashed commit of the following:
update tensorflow to 2.8.0
Ignore non-existing target devices
Support multi accelerators and setting accelerator by option
Ahwb performance improvement
1. set input/output from memory once.
2. lock/unlock once and out of invoke time.
add Eval Trace
add D1300 related change
add sustained paramters
Introduce Tensorflow builtin farmhash
reset compilation power hal and set sustained power param empty
make neuron delegate default low latency and use turbo boost
use tf-2.7.0 and make options more consistent
dos2unix and clang-format
remove unnecessary workspace update
remove redundant --opt-accuracy
remove intval-coloring-coverage=0.01
make default compile options empty for set parameter from command line
Fix compile_option null pointer warning
Fix typo by Using android_sdk_version instead of neuron_sdk_version
Support compile option
Reject Reduce OPs with zero-dimension axis tensor (Any/Min/Max/Prod/Sum)
remove float 32 support of gather
support int32 reshape and remove gather allocation type check
Fix build error on NNAPI delegate and AHardwareBuffer
Resolving undelegated OPs for ETHZ_v5 models
Apply clang-format
Support redirecting to NNAPI delegate if Neuron U-SDK in not available
add ahwb option
Support HardSwish directly
Do not lower HardSwith to atomic OPs.
Support ETHZv5 models
use TF 2.7.0-rc1 for better TFLite NMS
Add AHardWareBuffer to neuron delegate
Apply Dequantize on uint8 input -> int8 input scenario
Fix type
Update Neuron delegate provider to support turbo boost preference
Print Neuron implementation logs to both Android and stderr
Introduce new execution preference: turbo boost
Support setting from optimization string
Introduce performance helper classes to acquire performance lock
Introduce new APUWareUtils API to acquire performance lock by parameters
Support Neuron adapter new APIs
Fix wrong API declaration to get Neuron verison
Support reusable Neuron execution
Do not print failure message if the failure vector is empty
Co-authored-by: Code Lin <code.lin@mediatek.com>
Co-authored-by: Lauren Chen <Lauren.Chen@mediatek.com>
Co-authored-by: Leo Hsieh <leo.hsieh@mediatek.com>