surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 1 | ------ ArmNN for Android NNAPI supported operations ------ |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 2 | |
| 3 | This release of ArmNN for Android supports use as a driver for the Android Neural Networks API. It implements the android.hardware.neuralnetworks@1.0 interface. |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 4 | android.hardware.neuralnetworks@1.1 models available in Android P can be executed if they return true from the utility method compliantWithV1_0(model). |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 5 | |
| 6 | For more information on the Android Neural Networks API, see https://developer.android.com/ndk/guides/neuralnetworks/index.html |
| 7 | |
| 8 | For integration and usage documentation, please see README.md. |
| 9 | |
| 10 | --- Support for Android Neural Networks HAL operations --- |
| 11 | |
| 12 | The following AndroidNN operations are currently supported. |
| 13 | |
| 14 | AndroidNN operator Tensor type supported |
Matthew Bentham | 378333d | 2018-10-01 16:07:27 +0100 | [diff] [blame] | 15 | ADD (FLOAT32,QUANT8_ASYMM) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 16 | AVERAGE_POOL_2D (FLOAT32,QUANT8_ASYMM) |
| 17 | CONCATENATION (FLOAT32) |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 18 | CONV_2D (FLOAT32,QUANT8_ASYMM) |
| 19 | DEPTHWISE_CONV_2D* (FLOAT32,QUANT8_ASYMM) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 20 | FLOOR (FLOAT32) |
Matthew Bentham | 378333d | 2018-10-01 16:07:27 +0100 | [diff] [blame] | 21 | FULLY_CONNECTED (FLOAT32,QUANT8_ASYMM) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 22 | L2_NORMALIZATION (FLOAT32) |
| 23 | L2_POOL_2D (FLOAT32) |
| 24 | LOCAL_RESPONSE_NORMALIZATION (FLOAT32) |
| 25 | LOGISTIC (FLOAT32,QUANT8_ASYMM) |
| 26 | MAX_POOL_2D (FLOAT32,QUANT8_ASYMM) |
Matthew Bentham | 378333d | 2018-10-01 16:07:27 +0100 | [diff] [blame] | 27 | MUL (FLOAT32,QUANT8_ASYMM) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 28 | RELU (FLOAT32,QUANT8_ASYMM) |
| 29 | RELU1 (FLOAT32,QUANT8_ASYMM) |
| 30 | RELU6 (FLOAT32,QUANT8_ASYMM) |
| 31 | RESHAPE (FLOAT32,QUANT8_ASYMM) |
| 32 | RESIZE_BILINEAR (FLOAT32) |
| 33 | SOFTMAX (FLOAT32,QUANT8_ASYMM) |
saoste01 | b847148 | 2018-10-10 09:44:51 +0100 | [diff] [blame^] | 34 | SQUEEZE (FLOAT32,QUANT8_ASYMM) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 35 | TANH (FLOAT32) |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 36 | LSTM (FLOAT32) |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 37 | |
surmeh01 | 49b9e10 | 2018-05-17 14:11:25 +0100 | [diff] [blame] | 38 | * Depthwise convolution only supports a value of 1 for the depth multiplier. In addition, the QUANT8_ASYMM version only supports 3x3 kernels. |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 39 | |
| 40 | --- Unsupported operators --- |
| 41 | |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 42 | The following AndroidNN 1.0 operations are currently not supported. |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 43 | |
| 44 | DEPTH_TO_SPACE |
| 45 | DEQUANTIZE |
| 46 | EMBEDDING_LOOKUP |
| 47 | HASHTABLE_LOOKUP |
| 48 | LSH_PROJECTION |
telsoa01 | 5307bc1 | 2018-03-09 13:51:08 +0000 | [diff] [blame] | 49 | RNN |
| 50 | SPACE_TO_DEPTH |
| 51 | SVDF |
| 52 | |
| 53 | Where operations are not supported by the ArmNN Android NN Driver, the driver indicates this to the framework appropriately and the framework implements those operations using a CPU implementation. |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 54 | |
| 55 | The following AndroidNN 1.1 operations are currently not supported. |
| 56 | |
| 57 | BATCH_TO_SPACE_ND |
| 58 | DIV |
| 59 | MEAN |
| 60 | PAD |
| 61 | SPACE_TO_BATCH_ND |
telsoa01 | ce3e84a | 2018-08-31 09:31:35 +0100 | [diff] [blame] | 62 | STRIDED_SLICE |
| 63 | SUB |
| 64 | TRANSPOSE |
| 65 | |
| 66 | Where any of these operations are present the model will return false from the compliantWithV1_0(model) and the driver will not currently be utilised. |