tequilaOS/platform_hardware_interfaces
2
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Merge "Add PACK operation to NNAPI feature level 6." am: 6e74ca65a7 am: 9fc0aae600 am: 6232d05bf5

Browse source 
Original change: https://android-review.googlesource.com/c/platform/hardware/interfaces/+/1835817

Change-Id: I94d84ef899e8c23efaafba6c6e841281fb3f45fa
This commit is contained in:
David Gross 2021-10-19 21:54:20 +00:00 committed by Automerger Merge Worker
commit 9e836f5709
2 changed files with 50 additions and 0 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

1
neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperationType.aidl
View file

@ -137,4 +137,5 @@ enum OperationType {
FILL = 100,
RANK = 101,
BATCH_MATMUL = 102,
PACK = 103,
}

49
neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl
View file

@ -5273,4 +5273,53 @@ enum OperationType {
* c_o = r_y if adj_y else c_y
*/
BATCH_MATMUL = 102,
/**
* Packs N input tensors (N >= 1) of rank R into one output tensor of rank R+1.
* The tensors are packed along a given axis.
*
* The input tensors must have identical {@link OperandType} and dimensions.
*
* For example, suppose there are N input tensors of shape (A, B, C).
* If axis is 0, the output tensor will have shape (N, A, B, C).
* If axis is 1, the output tensor will have shape (A, N, B, C).
*
* All dimensions through the axis dimension determine the output tile count;
* the remaining dimensions determine the tile shape.
*
* Return to the example of N input tensors of shape (A, B, C).
* If axis is 0, there are N tiles in the output, each of shape (A, B, C).
* If axis is 1, there are A*N tiles in the output, each of shape (B, C).
*
* The coordinates of a tile within the output tensor are (t[0],...,t[axis]).
* The coordinates of a tile within an input tensor are (t[0],...,t[axis-1]).
* (If axis is 0, an input tensor consists of a single tile.)
* If we index input tensors starting with 0 (rather than by operand number),
* then output_tile[t[0],...,t[axis]] = input_tile[t[axis]][t[0],...,t[axis-1]].
* That is, all output tile coordinates except for the axis coordinate select
* the corresponding location within some input tensor; and the axis coordinate
* selects the input tensor.
*
* Supported tensor {@link OperandType}:
* * {@link OperandType::TENSOR_FLOAT16}
* * {@link OperandType::TENSOR_FLOAT32}
* * {@link OperandType::TENSOR_QUANT8_ASYMM}
* * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}
* * {@link OperandType::TENSOR_INT32}
*
* Supported input tensor rank: from 1
*
* Inputs:
* * 0: A scalar of type {@link OperandType::INT32}, specifying
* the axis along which to pack. The valid range is [0, R+1).
* * 1 ~ N: Input tensors to be packed together.
* For {@link OperandType::TENSOR_QUANT8_ASYMM} and
* {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensors,
* the scales and zeroPoint must be the same for all input tensors,
* and will be the same for the output tensor.
*
* Outputs:
* * 0: The packed tensor.
*/
PACK = 103,
}