From 8090245174e252697a406852d302fc30ad97d5db Mon Sep 17 00:00:00 2001 From: Lev Proleev Date: Mon, 20 Jul 2020 20:01:54 +0100 Subject: [PATCH] Create first version of NNAPI AIDL interface Bug: 161428342 Test: m android.hardware.neuralnetworks-update-api && m Change-Id: Icf8123746def6f4c654dc3e413e5169ab020c8b4 --- .../compatibility_matrix.current.xml | 7 + neuralnetworks/aidl/Android.bp | 19 + .../hardware/neuralnetworks/BufferDesc.aidl | 22 + .../hardware/neuralnetworks/BufferRole.aidl | 24 + .../hardware/neuralnetworks/Capabilities.aidl | 26 + .../hardware/neuralnetworks/DataLocation.aidl | 24 + .../hardware/neuralnetworks/DeviceBuffer.aidl | 23 + .../hardware/neuralnetworks/DeviceType.aidl | 25 + .../hardware/neuralnetworks/ErrorStatus.aidl | 30 + .../neuralnetworks/ExecutionPreference.aidl | 24 + .../neuralnetworks/ExecutionResult.aidl | 24 + .../hardware/neuralnetworks/Extension.aidl | 23 + .../ExtensionNameAndPrefix.aidl | 23 + .../ExtensionOperandTypeInformation.aidl | 24 + .../neuralnetworks/FusedActivationFunc.aidl | 25 + .../hardware/neuralnetworks/IBuffer.aidl | 23 + .../hardware/neuralnetworks/IDevice.aidl | 36 + .../IFencedExecutionCallback.aidl | 22 + .../neuralnetworks/IPreparedModel.aidl | 25 + .../IPreparedModelCallback.aidl | 22 + .../neuralnetworks/IPreparedModelParcel.aidl | 22 + .../hardware/neuralnetworks/Memory.aidl | 24 + .../hardware/neuralnetworks/Model.aidl | 27 + .../neuralnetworks/NumberOfCacheFiles.aidl | 23 + .../hardware/neuralnetworks/Operand.aidl | 28 + .../neuralnetworks/OperandExtraParams.aidl | 23 + .../neuralnetworks/OperandLifeTime.aidl | 28 + .../neuralnetworks/OperandPerformance.aidl | 23 + .../hardware/neuralnetworks/OperandType.aidl | 37 + .../hardware/neuralnetworks/Operation.aidl | 24 + .../neuralnetworks/OperationType.aidl | 123 + .../hardware/neuralnetworks/OutputShape.aidl | 23 + .../neuralnetworks/PerformanceInfo.aidl | 23 + .../hardware/neuralnetworks/Priority.aidl | 24 + .../hardware/neuralnetworks/Request.aidl | 24 + .../neuralnetworks/RequestArgument.aidl | 24 + .../neuralnetworks/RequestMemoryPool.aidl | 23 + .../hardware/neuralnetworks/Subgraph.aidl | 25 + .../SymmPerChannelQuantParams.aidl | 23 + .../hardware/neuralnetworks/Timing.aidl | 23 + .../hardware/neuralnetworks/BufferDesc.aidl | 31 + .../hardware/neuralnetworks/BufferRole.aidl | 40 + .../hardware/neuralnetworks/Capabilities.aidl | 63 + .../hardware/neuralnetworks/DataLocation.aidl | 37 + .../hardware/neuralnetworks/DeviceBuffer.aidl | 36 + .../hardware/neuralnetworks/DeviceType.aidl | 45 + .../hardware/neuralnetworks/ErrorStatus.aidl | 52 + .../neuralnetworks/ExecutionPreference.aidl | 41 + .../neuralnetworks/ExecutionResult.aidl | 47 + .../hardware/neuralnetworks/Extension.aidl | 42 + .../ExtensionNameAndPrefix.aidl | 49 + .../ExtensionOperandTypeInformation.aidl | 38 + .../neuralnetworks/FusedActivationFunc.aidl | 30 + .../hardware/neuralnetworks/IBuffer.aidl | 58 + .../hardware/neuralnetworks/IDevice.aidl | 431 ++ .../IFencedExecutionCallback.aidl | 56 + .../neuralnetworks/IPreparedModel.aidl | 173 + .../IPreparedModelCallback.aidl | 51 + .../neuralnetworks/IPreparedModelParcel.aidl | 28 + .../hardware/neuralnetworks/Memory.aidl | 31 + .../hardware/neuralnetworks/Model.aidl | 70 + .../neuralnetworks/NumberOfCacheFiles.aidl | 27 + .../hardware/neuralnetworks/Operand.aidl | 113 + .../neuralnetworks/OperandExtraParams.aidl | 40 + .../neuralnetworks/OperandLifeTime.aidl | 63 + .../neuralnetworks/OperandPerformance.aidl | 31 + .../hardware/neuralnetworks/OperandType.aidl | 154 + .../hardware/neuralnetworks/Operation.aidl | 46 + .../neuralnetworks/OperationType.aidl | 5132 +++++++++++++++++ .../hardware/neuralnetworks/OutputShape.aidl | 33 + .../neuralnetworks/PerformanceInfo.aidl | 37 + .../hardware/neuralnetworks/Priority.aidl | 29 + .../hardware/neuralnetworks/Request.aidl | 55 + .../neuralnetworks/RequestArgument.aidl | 53 + .../neuralnetworks/RequestMemoryPool.aidl | 36 + .../hardware/neuralnetworks/Subgraph.aidl | 51 + .../SymmPerChannelQuantParams.aidl | 33 + .../hardware/neuralnetworks/Timing.aidl | 37 + 78 files changed, 8484 insertions(+) create mode 100644 neuralnetworks/aidl/Android.bp create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/BufferDesc.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/BufferRole.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Capabilities.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DataLocation.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DeviceBuffer.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DeviceType.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ErrorStatus.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionPreference.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionResult.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Extension.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/FusedActivationFunc.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IBuffer.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IDevice.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IFencedExecutionCallback.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModel.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModelCallback.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModelParcel.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Memory.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Model.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/NumberOfCacheFiles.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Operand.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandExtraParams.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandLifeTime.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandPerformance.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandType.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Operation.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperationType.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OutputShape.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PerformanceInfo.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Priority.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Request.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/RequestArgument.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/RequestMemoryPool.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Subgraph.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/SymmPerChannelQuantParams.aidl create mode 100644 neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Timing.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/BufferDesc.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/BufferRole.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Capabilities.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/DataLocation.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/DeviceBuffer.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/DeviceType.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/ErrorStatus.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionPreference.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionResult.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Extension.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/FusedActivationFunc.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/IBuffer.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/IFencedExecutionCallback.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModelCallback.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModelParcel.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Memory.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Model.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/NumberOfCacheFiles.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Operand.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/OperandExtraParams.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/OperandLifeTime.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/OperandPerformance.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/OperandType.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Operation.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/OutputShape.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/PerformanceInfo.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Priority.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Request.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/RequestArgument.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/RequestMemoryPool.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Subgraph.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/SymmPerChannelQuantParams.aidl create mode 100644 neuralnetworks/aidl/android/hardware/neuralnetworks/Timing.aidl diff --git a/compatibility_matrices/compatibility_matrix.current.xml b/compatibility_matrices/compatibility_matrix.current.xml index 91ca312d45..fd255551bf 100644 --- a/compatibility_matrices/compatibility_matrix.current.xml +++ b/compatibility_matrices/compatibility_matrix.current.xml @@ -393,6 +393,13 @@ .* + + android.hardware.neuralnetworks + + IDevice + .* + + android.hardware.nfc 1.2 diff --git a/neuralnetworks/aidl/Android.bp b/neuralnetworks/aidl/Android.bp new file mode 100644 index 0000000000..308f89f663 --- /dev/null +++ b/neuralnetworks/aidl/Android.bp @@ -0,0 +1,19 @@ +aidl_interface { + name: "android.hardware.neuralnetworks", + vendor_available: true, + srcs: [ + "android/hardware/neuralnetworks/*.aidl", + ], + stability: "vintf", + imports: [ + "android.hardware.common", + ], + backend: { + java: { + enabled: false, + }, + cpp: { + enabled: false, + }, + }, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/BufferDesc.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/BufferDesc.aidl new file mode 100644 index 0000000000..2074a2ad4d --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/BufferDesc.aidl @@ -0,0 +1,22 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable BufferDesc { + int[] dimensions; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/BufferRole.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/BufferRole.aidl new file mode 100644 index 0000000000..97f748bcf8 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/BufferRole.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable BufferRole { + int modelIndex; + int ioIndex; + float frequency; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Capabilities.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Capabilities.aidl new file mode 100644 index 0000000000..31afafc7df --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Capabilities.aidl @@ -0,0 +1,26 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Capabilities { + android.hardware.neuralnetworks.PerformanceInfo relaxedFloat32toFloat16PerformanceScalar; + android.hardware.neuralnetworks.PerformanceInfo relaxedFloat32toFloat16PerformanceTensor; + android.hardware.neuralnetworks.OperandPerformance[] operandPerformance; + android.hardware.neuralnetworks.PerformanceInfo ifPerformance; + android.hardware.neuralnetworks.PerformanceInfo whilePerformance; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DataLocation.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DataLocation.aidl new file mode 100644 index 0000000000..5b03ba038e --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DataLocation.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable DataLocation { + int poolIndex; + long offset; + long length; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DeviceBuffer.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DeviceBuffer.aidl new file mode 100644 index 0000000000..9cff6db999 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DeviceBuffer.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable DeviceBuffer { + android.hardware.neuralnetworks.IBuffer buffer; + int token; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DeviceType.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DeviceType.aidl new file mode 100644 index 0000000000..dd4dae7d0e --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/DeviceType.aidl @@ -0,0 +1,25 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@Backing(type="int") @VintfStability +enum DeviceType { + OTHER = 1, + CPU = 2, + GPU = 3, + ACCELERATOR = 4, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ErrorStatus.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ErrorStatus.aidl new file mode 100644 index 0000000000..ba18c3801e --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ErrorStatus.aidl @@ -0,0 +1,30 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@Backing(type="int") @VintfStability +enum ErrorStatus { + NONE = 0, + DEVICE_UNAVAILABLE = 1, + GENERAL_FAILURE = 2, + OUTPUT_INSUFFICIENT_SIZE = 3, + INVALID_ARGUMENT = 4, + MISSED_DEADLINE_TRANSIENT = 5, + MISSED_DEADLINE_PERSISTENT = 6, + RESOURCE_EXHAUSTED_TRANSIENT = 7, + RESOURCE_EXHAUSTED_PERSISTENT = 8, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionPreference.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionPreference.aidl new file mode 100644 index 0000000000..cccae5403d --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionPreference.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@Backing(type="int") @VintfStability +enum ExecutionPreference { + LOW_POWER = 0, + FAST_SINGLE_ANSWER = 1, + SUSTAINED_SPEED = 2, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionResult.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionResult.aidl new file mode 100644 index 0000000000..c17ddb9116 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExecutionResult.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable ExecutionResult { + boolean outputSufficientSize; + android.hardware.neuralnetworks.OutputShape[] outputShapes; + android.hardware.neuralnetworks.Timing timing; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Extension.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Extension.aidl new file mode 100644 index 0000000000..9eb8896af7 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Extension.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Extension { + String name; + android.hardware.neuralnetworks.ExtensionOperandTypeInformation[] operandTypes; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl new file mode 100644 index 0000000000..a271a63128 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable ExtensionNameAndPrefix { + String name; + char prefix; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.aidl new file mode 100644 index 0000000000..d1c3f099b0 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable ExtensionOperandTypeInformation { + char type; + boolean isTensor; + int byteSize; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/FusedActivationFunc.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/FusedActivationFunc.aidl new file mode 100644 index 0000000000..ddd3c2abd7 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/FusedActivationFunc.aidl @@ -0,0 +1,25 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@Backing(type="int") @VintfStability +enum FusedActivationFunc { + NONE = 0, + RELU = 1, + RELU1 = 2, + RELU6 = 3, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IBuffer.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IBuffer.aidl new file mode 100644 index 0000000000..a297a6bb31 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IBuffer.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +interface IBuffer { + void copyFrom(in android.hardware.neuralnetworks.Memory src, in int[] dimensions); + void copyTo(in android.hardware.neuralnetworks.Memory dst); +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IDevice.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IDevice.aidl new file mode 100644 index 0000000000..38fda16b56 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IDevice.aidl @@ -0,0 +1,36 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +interface IDevice { + android.hardware.neuralnetworks.DeviceBuffer allocate(in android.hardware.neuralnetworks.BufferDesc desc, in android.hardware.neuralnetworks.IPreparedModelParcel[] preparedModels, in android.hardware.neuralnetworks.BufferRole[] inputRoles, in android.hardware.neuralnetworks.BufferRole[] outputRoles); + android.hardware.neuralnetworks.Capabilities getCapabilities(); + android.hardware.neuralnetworks.NumberOfCacheFiles getNumberOfCacheFilesNeeded(); + android.hardware.neuralnetworks.Extension[] getSupportedExtensions(); + boolean[] getSupportedOperations(in android.hardware.neuralnetworks.Model model); + android.hardware.neuralnetworks.DeviceType getType(); + String getVersionString(); + void prepareModel(in android.hardware.neuralnetworks.Model model, in android.hardware.neuralnetworks.ExecutionPreference preference, in android.hardware.neuralnetworks.Priority priority, in long deadline, in ParcelFileDescriptor[] modelCache, in ParcelFileDescriptor[] dataCache, in byte[] token, in android.hardware.neuralnetworks.IPreparedModelCallback callback); + void prepareModelFromCache(in long deadline, in ParcelFileDescriptor[] modelCache, in ParcelFileDescriptor[] dataCache, in byte[] token, in android.hardware.neuralnetworks.IPreparedModelCallback callback); + const int BYTE_SIZE_OF_CACHE_TOKEN = 32; + const int MAX_NUMBER_OF_CACHE_FILES = 32; + const int EXTENSION_TYPE_HIGH_BITS_PREFIX = 15; + const int EXTENSION_TYPE_LOW_BITS_TYPE = 16; + const int OPERAND_TYPE_BASE_MAX = 65535; + const int OPERATION_TYPE_BASE_MAX = 65535; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IFencedExecutionCallback.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IFencedExecutionCallback.aidl new file mode 100644 index 0000000000..a7cf90690e --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IFencedExecutionCallback.aidl @@ -0,0 +1,22 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +interface IFencedExecutionCallback { + android.hardware.neuralnetworks.ErrorStatus getExecutionInfo(out android.hardware.neuralnetworks.Timing timingLaunched, out android.hardware.neuralnetworks.Timing timingFenced); +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModel.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModel.aidl new file mode 100644 index 0000000000..87677122e9 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModel.aidl @@ -0,0 +1,25 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +interface IPreparedModel { + android.hardware.neuralnetworks.ExecutionResult executeSynchronously(in android.hardware.neuralnetworks.Request request, in boolean measureTiming, in long deadline, in long loopTimeoutDuration); + android.hardware.neuralnetworks.IFencedExecutionCallback executeFenced(in android.hardware.neuralnetworks.Request request, in ParcelFileDescriptor[] waitFor, in boolean measureTiming, in long deadline, in long loopTimeoutDuration, in long duration, out @nullable ParcelFileDescriptor syncFence); + const long DEFAULT_LOOP_TIMEOUT_DURATION_NS = 2000000000; + const long MAXIMUM_LOOP_TIMEOUT_DURATION_NS = 15000000000; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModelCallback.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModelCallback.aidl new file mode 100644 index 0000000000..d1ae2eb72b --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModelCallback.aidl @@ -0,0 +1,22 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +interface IPreparedModelCallback { + void notify(in android.hardware.neuralnetworks.ErrorStatus status, in android.hardware.neuralnetworks.IPreparedModel preparedModel); +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModelParcel.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModelParcel.aidl new file mode 100644 index 0000000000..048251a361 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/IPreparedModelParcel.aidl @@ -0,0 +1,22 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable IPreparedModelParcel { + android.hardware.neuralnetworks.IPreparedModel preparedModel; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Memory.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Memory.aidl new file mode 100644 index 0000000000..aa735c02d0 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Memory.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Memory { + android.hardware.common.NativeHandle handle; + long size; + String name; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Model.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Model.aidl new file mode 100644 index 0000000000..944bd7f5ed --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Model.aidl @@ -0,0 +1,27 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Model { + android.hardware.neuralnetworks.Subgraph main; + android.hardware.neuralnetworks.Subgraph[] referenced; + byte[] operandValues; + android.hardware.neuralnetworks.Memory[] pools; + boolean relaxComputationFloat32toFloat16; + android.hardware.neuralnetworks.ExtensionNameAndPrefix[] extensionNameToPrefix; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/NumberOfCacheFiles.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/NumberOfCacheFiles.aidl new file mode 100644 index 0000000000..ca5f917578 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/NumberOfCacheFiles.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable NumberOfCacheFiles { + int numModelCache; + int numDataCache; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Operand.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Operand.aidl new file mode 100644 index 0000000000..6615b9b42c --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Operand.aidl @@ -0,0 +1,28 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Operand { + android.hardware.neuralnetworks.OperandType type; + int[] dimensions; + float scale; + int zeroPoint; + android.hardware.neuralnetworks.OperandLifeTime lifetime; + android.hardware.neuralnetworks.DataLocation location; + @nullable android.hardware.neuralnetworks.OperandExtraParams extraParams; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandExtraParams.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandExtraParams.aidl new file mode 100644 index 0000000000..20317c7016 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandExtraParams.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +union OperandExtraParams { + android.hardware.neuralnetworks.SymmPerChannelQuantParams channelQuant; + byte[] extension; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandLifeTime.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandLifeTime.aidl new file mode 100644 index 0000000000..1082f9ee1f --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandLifeTime.aidl @@ -0,0 +1,28 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@Backing(type="int") @VintfStability +enum OperandLifeTime { + TEMPORARY_VARIABLE = 0, + SUBGRAPH_INPUT = 1, + SUBGRAPH_OUTPUT = 2, + CONSTANT_COPY = 3, + CONSTANT_POOL = 4, + NO_VALUE = 5, + SUBGRAPH = 6, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandPerformance.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandPerformance.aidl new file mode 100644 index 0000000000..9232b4c70e --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandPerformance.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable OperandPerformance { + android.hardware.neuralnetworks.OperandType type; + android.hardware.neuralnetworks.PerformanceInfo info; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandType.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandType.aidl new file mode 100644 index 0000000000..bd95fab52a --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperandType.aidl @@ -0,0 +1,37 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@Backing(type="int") @VintfStability +enum OperandType { + FLOAT32 = 0, + INT32 = 1, + UINT32 = 2, + TENSOR_FLOAT32 = 3, + TENSOR_INT32 = 4, + TENSOR_QUANT8_ASYMM = 5, + BOOL = 6, + TENSOR_QUANT16_SYMM = 7, + TENSOR_FLOAT16 = 8, + TENSOR_BOOL8 = 9, + FLOAT16 = 10, + TENSOR_QUANT8_SYMM_PER_CHANNEL = 11, + TENSOR_QUANT16_ASYMM = 12, + TENSOR_QUANT8_SYMM = 13, + TENSOR_QUANT8_ASYMM_SIGNED = 14, + SUBGRAPH = 15, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Operation.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Operation.aidl new file mode 100644 index 0000000000..383eba4a15 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Operation.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Operation { + android.hardware.neuralnetworks.OperationType type; + int[] inputs; + int[] outputs; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperationType.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperationType.aidl new file mode 100644 index 0000000000..f786829eb9 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OperationType.aidl @@ -0,0 +1,123 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@Backing(type="int") @VintfStability +enum OperationType { + ADD = 0, + AVERAGE_POOL_2D = 1, + CONCATENATION = 2, + CONV_2D = 3, + DEPTHWISE_CONV_2D = 4, + DEPTH_TO_SPACE = 5, + DEQUANTIZE = 6, + EMBEDDING_LOOKUP = 7, + FLOOR = 8, + FULLY_CONNECTED = 9, + HASHTABLE_LOOKUP = 10, + L2_NORMALIZATION = 11, + L2_POOL_2D = 12, + LOCAL_RESPONSE_NORMALIZATION = 13, + LOGISTIC = 14, + LSH_PROJECTION = 15, + LSTM = 16, + MAX_POOL_2D = 17, + MUL = 18, + RELU = 19, + RELU1 = 20, + RELU6 = 21, + RESHAPE = 22, + RESIZE_BILINEAR = 23, + RNN = 24, + SOFTMAX = 25, + SPACE_TO_DEPTH = 26, + SVDF = 27, + TANH = 28, + BATCH_TO_SPACE_ND = 29, + DIV = 30, + MEAN = 31, + PAD = 32, + SPACE_TO_BATCH_ND = 33, + SQUEEZE = 34, + STRIDED_SLICE = 35, + SUB = 36, + TRANSPOSE = 37, + ABS = 38, + ARGMAX = 39, + ARGMIN = 40, + AXIS_ALIGNED_BBOX_TRANSFORM = 41, + BIDIRECTIONAL_SEQUENCE_LSTM = 42, + BIDIRECTIONAL_SEQUENCE_RNN = 43, + BOX_WITH_NMS_LIMIT = 44, + CAST = 45, + CHANNEL_SHUFFLE = 46, + DETECTION_POSTPROCESSING = 47, + EQUAL = 48, + EXP = 49, + EXPAND_DIMS = 50, + GATHER = 51, + GENERATE_PROPOSALS = 52, + GREATER = 53, + GREATER_EQUAL = 54, + GROUPED_CONV_2D = 55, + HEATMAP_MAX_KEYPOINT = 56, + INSTANCE_NORMALIZATION = 57, + LESS = 58, + LESS_EQUAL = 59, + LOG = 60, + LOGICAL_AND = 61, + LOGICAL_NOT = 62, + LOGICAL_OR = 63, + LOG_SOFTMAX = 64, + MAXIMUM = 65, + MINIMUM = 66, + NEG = 67, + NOT_EQUAL = 68, + PAD_V2 = 69, + POW = 70, + PRELU = 71, + QUANTIZE = 72, + QUANTIZED_16BIT_LSTM = 73, + RANDOM_MULTINOMIAL = 74, + REDUCE_ALL = 75, + REDUCE_ANY = 76, + REDUCE_MAX = 77, + REDUCE_MIN = 78, + REDUCE_PROD = 79, + REDUCE_SUM = 80, + ROI_ALIGN = 81, + ROI_POOLING = 82, + RSQRT = 83, + SELECT = 84, + SIN = 85, + SLICE = 86, + SPLIT = 87, + SQRT = 88, + TILE = 89, + TOPK_V2 = 90, + TRANSPOSE_CONV_2D = 91, + UNIDIRECTIONAL_SEQUENCE_LSTM = 92, + UNIDIRECTIONAL_SEQUENCE_RNN = 93, + RESIZE_NEAREST_NEIGHBOR = 94, + QUANTIZED_LSTM = 95, + IF = 96, + WHILE = 97, + ELU = 98, + HARD_SWISH = 99, + FILL = 100, + RANK = 101, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OutputShape.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OutputShape.aidl new file mode 100644 index 0000000000..1300c49b7a --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/OutputShape.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable OutputShape { + int[] dimensions; + boolean isSufficient; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PerformanceInfo.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PerformanceInfo.aidl new file mode 100644 index 0000000000..b5dc179943 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/PerformanceInfo.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable PerformanceInfo { + float execTime; + float powerUsage; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Priority.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Priority.aidl new file mode 100644 index 0000000000..980bee328f --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Priority.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@Backing(type="int") @VintfStability +enum Priority { + LOW = 0, + MEDIUM = 1, + HIGH = 2, +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Request.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Request.aidl new file mode 100644 index 0000000000..6f77066fa7 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Request.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Request { + android.hardware.neuralnetworks.RequestArgument[] inputs; + android.hardware.neuralnetworks.RequestArgument[] outputs; + android.hardware.neuralnetworks.RequestMemoryPool[] pools; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/RequestArgument.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/RequestArgument.aidl new file mode 100644 index 0000000000..c9560efe4b --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/RequestArgument.aidl @@ -0,0 +1,24 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable RequestArgument { + boolean hasNoValue; + android.hardware.neuralnetworks.DataLocation location; + int[] dimensions; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/RequestMemoryPool.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/RequestMemoryPool.aidl new file mode 100644 index 0000000000..123e4b0af4 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/RequestMemoryPool.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +union RequestMemoryPool { + android.hardware.neuralnetworks.Memory pool; + int token; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Subgraph.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Subgraph.aidl new file mode 100644 index 0000000000..771d15a21d --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Subgraph.aidl @@ -0,0 +1,25 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Subgraph { + android.hardware.neuralnetworks.Operand[] operands; + android.hardware.neuralnetworks.Operation[] operations; + int[] inputIndexes; + int[] outputIndexes; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/SymmPerChannelQuantParams.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/SymmPerChannelQuantParams.aidl new file mode 100644 index 0000000000..2282febed2 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/SymmPerChannelQuantParams.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable SymmPerChannelQuantParams { + float[] scales; + int channelDim; +} diff --git a/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Timing.aidl b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Timing.aidl new file mode 100644 index 0000000000..b08d34acc1 --- /dev/null +++ b/neuralnetworks/aidl/aidl_api/android.hardware.neuralnetworks/current/android/hardware/neuralnetworks/Timing.aidl @@ -0,0 +1,23 @@ +/////////////////////////////////////////////////////////////////////////////// +// THIS FILE IS IMMUTABLE. DO NOT EDIT IN ANY CASE. // +/////////////////////////////////////////////////////////////////////////////// + +// This file is a snapshot of an AIDL interface (or parcelable). Do not try to +// edit this file. It looks like you are doing that because you have modified +// an AIDL interface in a backward-incompatible way, e.g., deleting a function +// from an interface or a field from a parcelable and it broke the build. That +// breakage is intended. +// +// You must not make a backward incompatible changes to the AIDL files built +// with the aidl_interface module type with versions property set. The module +// type is used to build AIDL files in a way that they can be used across +// independently updatable components of the system. If a device is shipped +// with such a backward incompatible change, it has a high risk of breaking +// later when a module using the interface is updated, e.g., Mainline modules. + +package android.hardware.neuralnetworks; +@VintfStability +parcelable Timing { + long timeOnDevice; + long timeInDriver; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/BufferDesc.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/BufferDesc.aidl new file mode 100644 index 0000000000..1b92ebc988 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/BufferDesc.aidl @@ -0,0 +1,31 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * A buffer descriptor. Describes the properties of a buffer. + */ +@VintfStability +parcelable BufferDesc { + /** + * Dimensions of the buffer. May have unknown dimensions or rank. A buffer with some number of + * unspecified dimensions is represented by setting each unspecified dimension to 0. A buffer + * with unspecified rank is represented by providing an empty dimensions vector. + */ + int[] dimensions; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/BufferRole.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/BufferRole.aidl new file mode 100644 index 0000000000..7877bc0180 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/BufferRole.aidl @@ -0,0 +1,40 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Describes a role of an input or output to a prepared model. + */ +@VintfStability +parcelable BufferRole { + /** + * The index of the IPreparedModel within the "preparedModel" argument passed in + * IDevice::allocate. + */ + int modelIndex; + /** + * The index of the input or output operand. + */ + int ioIndex; + /** + * A floating-point value within the range (0.0, 1.0]. Describes how likely the buffer is to be + * used in the specified role. This is provided as a hint to optimize the case when multiple + * roles prefer different buffer locations or data layouts. + */ + float frequency; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Capabilities.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Capabilities.aidl new file mode 100644 index 0000000000..5ce78ee96f --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Capabilities.aidl @@ -0,0 +1,63 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.OperandPerformance; +import android.hardware.neuralnetworks.PerformanceInfo; + +/** + * The capabilities of a driver. + * + * This represents performance of non-extension operations. + * + * Performance of an operation other than {@link OperationType::IF} and {@link OperationType::WHILE} + * comes from the type of its first operand. + */ +@VintfStability +parcelable Capabilities { + /** + * Driver performance when operating on float32 data but performing calculations with range + * and/or precision as low as that of the IEEE 754 16-bit floating-point format. + */ + PerformanceInfo relaxedFloat32toFloat16PerformanceScalar; + PerformanceInfo relaxedFloat32toFloat16PerformanceTensor; + /** + * Performance by operand type. Must be sorted by OperandType. + * + * If a particular {@link OperandType} is not present in operandPerformance, its performance is + * treated as { .execTime = FLT_MAX, .powerUsage = FLT_MAX }. + * + * Performance does not apply to {@link OperandType::SUBGRAPH}, and a driver must not report + * operand performance for {@link OperandType::SUBGRAPH}. + */ + OperandPerformance[] operandPerformance; + /** + * Performance of an {@link OperationType::IF} operation is the sum of + * {@link Capabilities::ifPerformance} and the mean of performance for the two branch subgraphs, + * where performance for a subgraph is the sum of the performance of all operations within the + * subgraph. + */ + PerformanceInfo ifPerformance; + /** + * Performance of a {@link OperationType::WHILE} operation is the sum of + * {@link Capabilities::whilePerformance}, performance for the condition subgraph and + * performance for the body subgraph, where performance for a subgraph is the sum of the + * performance of all operations within the subgraph. + */ + PerformanceInfo whilePerformance; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/DataLocation.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/DataLocation.aidl new file mode 100644 index 0000000000..57e3f4ade6 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/DataLocation.aidl @@ -0,0 +1,37 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Describes the location of a data object. + */ +@VintfStability +parcelable DataLocation { + /** + * The index of the memory pool where this location is found. + */ + int poolIndex; + /** + * Offset in bytes from the start of the pool. + */ + long offset; + /** + * The length of the data in bytes. + */ + long length; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/DeviceBuffer.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/DeviceBuffer.aidl new file mode 100644 index 0000000000..d51e1b2881 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/DeviceBuffer.aidl @@ -0,0 +1,36 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.IBuffer; + +/** + * A type that is used to represent a driver allocated buffer and token that corresponds to it. + */ + @VintfStability + parcelable DeviceBuffer { + /** + * An IBuffer object used to interact with the device allocated buffer. + */ + IBuffer buffer; + /** + * A positive token identifying the allocated buffer. The token is provided when referencing the + * buffer as one of the memory pools in the request of an execution. The token must not collide + * with the tokens of other IBuffer objects that are currently alive in the same driver service. + */ + int token; + } \ No newline at end of file diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/DeviceType.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/DeviceType.aidl new file mode 100644 index 0000000000..8399d504ce --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/DeviceType.aidl @@ -0,0 +1,45 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Device types. + * + * The type of NNAPI device. + */ +@VintfStability +@Backing(type="int") +enum DeviceType { + /** + * The device does not fall into any category below. + */ + OTHER = 1, + /** + * The device runs NNAPI models on single or multi-core CPU. + */ + CPU = 2, + /** + * The device can run NNAPI models and also accelerate graphics APIs such as OpenGL ES and + * Vulkan. + */ + GPU = 3, + /** + * Dedicated accelerator for Machine Learning workloads. + */ + ACCELERATOR = 4, +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/ErrorStatus.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/ErrorStatus.aidl new file mode 100644 index 0000000000..860f86a156 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/ErrorStatus.aidl @@ -0,0 +1,52 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Calls to neural networks AIDL interfaces may return a ServiceSpecificException with the following + * error codes. + */ +@VintfStability +@Backing(type="int") +enum ErrorStatus { + NONE, + DEVICE_UNAVAILABLE, + GENERAL_FAILURE, + OUTPUT_INSUFFICIENT_SIZE, + INVALID_ARGUMENT, + /** + * Failure because a deadline could not be met for a task, but future deadlines may still be met + * for the same task after a short delay. + */ + MISSED_DEADLINE_TRANSIENT, + /** + * Failure because a deadline could not be met for a task, and future deadlines will likely also + * not be met for the same task even after a short delay. + */ + MISSED_DEADLINE_PERSISTENT, + /** + * Failure because of a resource limitation within the driver, but future calls for the same + * task may still succeed after a short delay. + */ + RESOURCE_EXHAUSTED_TRANSIENT, + /** + * Failure because of a resource limitation within the driver, and future calls for the same + * task will likely also fail even after a short delay. + */ + RESOURCE_EXHAUSTED_PERSISTENT, +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionPreference.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionPreference.aidl new file mode 100644 index 0000000000..901cb384c9 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionPreference.aidl @@ -0,0 +1,41 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Execution preferences. + */ +@VintfStability +@Backing(type="int") +enum ExecutionPreference { + /** + * Prefer executing in a way that minimizes battery drain. This is desirable for compilations + * that will be executed often. + */ + LOW_POWER, + /** + * Prefer returning a single answer as fast as possible, even if this causes more power + * consumption. + */ + FAST_SINGLE_ANSWER, + /** + * Prefer maximizing the throughput of successive frames, for example when processing successive + * frames coming from the camera. + */ + SUSTAINED_SPEED, +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionResult.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionResult.aidl new file mode 100644 index 0000000000..403fe097ee --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExecutionResult.aidl @@ -0,0 +1,47 @@ +/* + * Copyright (C) 2021 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.ErrorStatus; +import android.hardware.neuralnetworks.OutputShape; +import android.hardware.neuralnetworks.Timing; + +/** + * A result from running a synchronous execution of a prepared model. + */ +@VintfStability +parcelable ExecutionResult { + /** + * A value of "true" indicates that the execution was successful. A value of "false" indicates + * the execution failed because at least one output operand buffer was not large enough to store + * the corresponding output. + */ + boolean outputSufficientSize; + /** + * A list of shape information of model output operands. The index in "outputShapes" corresponds + * to the index of the output operand in the Request outputs vector. + */ + OutputShape[] outputShapes; + /** + * Duration of execution. Unless measure is true and the execution is successful, all times must + * be reported as -1. A driver may choose to report any time as -1, indicating that measurement + * is not available. + */ + Timing timing; +} + diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Extension.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Extension.aidl new file mode 100644 index 0000000000..159e3c15aa --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Extension.aidl @@ -0,0 +1,42 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.ExtensionOperandTypeInformation; + +/** + * Information about an extension. + */ +@VintfStability +parcelable Extension { + /** + * The extension name. + * + * The name must consist of lowercase latin letters, numbers, periods, and underscore signs. The + * name must contain at least one period. + * + * The name must start with the reverse domain name of the vendor. + * + * Example: com.google.test_extension + */ + String name; + /** + * Information about operand types defined by the extension. + */ + ExtensionOperandTypeInformation[] operandTypes; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl new file mode 100644 index 0000000000..76074bf416 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionNameAndPrefix.aidl @@ -0,0 +1,49 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * The mapping between extension names and prefixes of operand and operation type values. + * + * An operand or operation whose numeric type value is above {@link IDevice::OPERAND_TYPE_BASE_MAX} + * or {@link IDevice::OPERATION_TYPE_BASE_MAX} respectively should be interpreted as an extension + * operand/operation. The low {@link IDevice::EXTENSION_TYPE_LOW_BITS_TYPE} bits of the value + * correspond to the type ID within the extension and the high + * {@link IDevice::EXTENSION_TYPE_HIGH_BITS_PREFIX} bits encode the "prefix", which maps uniquely to + * the extension name. The sign bit is always 0. + * + * For example, if a model contains an operation whose value is 0x7AAABBBB and extensionNameToPrefix + * contains an entry with prefix=0x7AAA and name="vendor.test.test_extension", then the operation + * should be interpreted as the operation 0xBBBB of the extension named vendor.test.test_extension. + * + * This is a one-to-one correspondence. That is, there must be at most one prefix corresponding to + * each extension name and at most one extension name corresponding to each prefix. + */ +@VintfStability +parcelable ExtensionNameAndPrefix { + /** + * The extension name. + * + * See {@link Extension::name} for the format specification. + */ + String name; + /** + * The extension prefix. Only the lowest 15 bits are used, so the value must be less than 32768. + */ + char prefix; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.aidl new file mode 100644 index 0000000000..d7f93c10b0 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/ExtensionOperandTypeInformation.aidl @@ -0,0 +1,38 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Information about an extension operand type. + */ +@VintfStability +parcelable ExtensionOperandTypeInformation { + /** + * The extension operand type. + */ + char type; + /** + * Indicates whether the extension operand type represents a tensor or a scalar. + */ + boolean isTensor; + /** + * The byte size of the operand (if scalar) or of a single element (if tensor). + */ + int byteSize; +} + diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/FusedActivationFunc.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/FusedActivationFunc.aidl new file mode 100644 index 0000000000..40f1053f41 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/FusedActivationFunc.aidl @@ -0,0 +1,30 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Fused activation function types. + */ +@VintfStability +@Backing(type="int") +enum FusedActivationFunc { + NONE, + RELU, + RELU1, + RELU6, +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IBuffer.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IBuffer.aidl new file mode 100644 index 0000000000..eb3dec6e4f --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IBuffer.aidl @@ -0,0 +1,58 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.Memory; + +/** + * This interface represents a device memory buffer. + */ +@VintfStability +interface IBuffer { + /** + * Sets the content of this buffer from a shared memory region. + * + * @param src The source shared memory region. + * @param dimensions Updated dimensional information. If the dimensions of the IBuffer object + * are not fully specified, then the dimensions must be fully specified here. + * If the dimensions of the IBuffer object are fully specified, then the + * dimensions may be empty here. If dimensions.size() > 0, then all dimensions + * must be specified here, and any dimension that was specified in the IBuffer + * object must have the same value here. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if there is an unspecified error + * - INVALID_ARGUMENT if provided memory is invalid, or if the dimensions is invalid + */ + void copyFrom(in Memory src, in int[] dimensions); + + /** + * Retrieves the content of this buffer to a shared memory region. + * + * The IBuffer object must have been initialized before the call to IBuffer::copyTo. For more + * information on the state of the IBuffer object, refer to IDevice::allocate. + * + * @param dst The destination shared memory region. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if the IBuffer object is uninitialized, or there is an unspecified + * error + * - INVALID_ARGUMENT if provided memory is invalid + */ + void copyTo(in Memory dst); +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl new file mode 100644 index 0000000000..0c4954c1b8 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IDevice.aidl @@ -0,0 +1,431 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.BufferDesc; +import android.hardware.neuralnetworks.BufferRole; +import android.hardware.neuralnetworks.Capabilities; +import android.hardware.neuralnetworks.DeviceBuffer; +import android.hardware.neuralnetworks.DeviceType; +import android.hardware.neuralnetworks.ExecutionPreference; +import android.hardware.neuralnetworks.Extension; +import android.hardware.neuralnetworks.IPreparedModel; +import android.hardware.neuralnetworks.IPreparedModelCallback; +import android.hardware.neuralnetworks.IPreparedModelParcel; +import android.hardware.neuralnetworks.Model; +import android.hardware.neuralnetworks.NumberOfCacheFiles; +import android.hardware.neuralnetworks.Priority; + +/** + * This interface represents a device driver. + */ +@VintfStability +interface IDevice { + /** + * The byte size of the cache token. + */ + const int BYTE_SIZE_OF_CACHE_TOKEN = 32; + /** + * The maximum number of files for each type of cache in compilation caching. + */ + const int MAX_NUMBER_OF_CACHE_FILES = 32; + + /** + * Numeric values of extension operand and operation types have the following structure: + * - The sign bit is always 0. + * - 15 high bits represent the "prefix", which corresponds uniquely to the extension name. + * - 16 low bits represent the type ID within the extension. + */ + const int EXTENSION_TYPE_HIGH_BITS_PREFIX = 15; + const int EXTENSION_TYPE_LOW_BITS_TYPE = 16; + /** + * OperandType with any value above {@link IDevice::OPERAND_TYPE_BASE_MAX} must be interpreted + * as an extension type according to {@link Model::extensionNameToPrefix}. + */ + const int OPERAND_TYPE_BASE_MAX = 0xFFFF; + /** + * OperationType with any value above {@link IDevice::OPERATION_TYPE_BASE_MAX} must be + * interpreted as an extension type according to {@link Model::extensionNameToPrefix}. + */ + const int OPERATION_TYPE_BASE_MAX = 0xFFFF; + + /** + * Allocates a driver-managed buffer with the properties specified by the buffer descriptor as + * well as the input and output roles. + * + * The allocate function must verify its inputs are correct. If there is an error, or if a + * certain role or property is not supported by the driver, the allocate function must return a + * service specific exception with an appropriate ErrorStatus. If the allocation is successful, + * this method must return a DeviceBuffer object with the produced IBuffer and a positive token + * identifying the allocated buffer. A successful allocation must accommodate all of the + * specified roles and buffer properties. + * + * The buffer is allocated to an uninitialized state. An uninitialized buffer may only be used + * in ways that are specified by outputRoles. A buffer is initialized after it is used as an + * output in a successful execution, or after a successful invocation of IBuffer::copyFrom on + * the buffer. An initialized buffer may be used according to all roles specified in inputRoles + * and outputRoles. A buffer will return to the uninitialized state if it is used as an output + * in a failed execution, or after a failed invocation of IBuffer::copyFrom on the buffer. + * + * The dimensions of the buffer can be deduced from the buffer descriptor as well as the + * dimensions of the corresponding model operands of the input and output roles. The dimensions + * or rank of the buffer may be unknown at this stage. As such, some driver services may only + * create a placeholder and defer the actual allocation until execution time. Note that the same + * buffer may be used for different shapes of outputs on different executions. When the buffer + * is used as an input, the input shape must be the same as the output shape from the last + * execution using this buffer as an output. + * + * The driver must apply proper validatation upon every usage of the buffer, and must fail the + * execution immediately if the usage is illegal. + * + * @param desc A buffer descriptor specifying the properties of the buffer to allocate. + * @param preparedModels A vector of IPreparedModel objects. Must only contain IPreparedModel + * objects from the same IDevice as this method is being invoked on. + * @param inputRoles A vector of roles with each specifying an input to a prepared model. + * @param outputRoles A vector of roles with each specifying an output to a prepared model. Each + * role specified in inputRoles and outputRoles must be unique. The + * corresponding model operands of the roles must have the same OperandType, + * scale, zero point, and ExtraParams. The dimensions of the operands and the + * dimensions specified in the buffer descriptor must be compatible with each + * other. Two dimensions are incompatible if there is at least one axis that + * is fully specified in both but has different values. + * @return DeviceBuffer object containing the allocated IBuffer object and a positive token that + * can be used to reference the buffer as one of the memory pools. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if a certain buffer property or a certain role is not supported, + * or if there is an unspecified error + * - INVALID_ARGUMENT if one of the input arguments is invalid + * - RESOURCE_EXHAUSTED_* if the task was aborted by the driver + */ + DeviceBuffer allocate(in BufferDesc desc, in IPreparedModelParcel[] preparedModels, + in BufferRole[] inputRoles, in BufferRole[] outputRoles); + + /** + * Gets the capabilities of a driver. + * + * @return Capabilities of the driver. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if there is an unspecified error + */ + Capabilities getCapabilities(); + + /** + * Gets the caching requirements of the driver implementation. + * + * There are two types of cache file descriptors provided to the driver: model cache and data + * cache. + * + * The data cache is for caching constant data, possibly including preprocessed and transformed + * tensor buffers. Any modification to the data cache should have no worse effect than + * generating bad output values at execution time. + * + * The model cache is for caching security-sensitive data such as compiled executable machine + * code in the device's native binary format. A modification to the model cache may affect the + * driver's execution behavior, and a malicious client could make use of this to execute beyond + * the granted permission. Thus, the driver must always check whether the model cache is + * corrupted before preparing the model from cache. + * + * getNumberOfCacheFilesNeeded returns how many of each type of cache files the driver + * implementation needs to cache a single prepared model. Returning 0 for both types indicates + * compilation caching is not supported by this driver. The driver may still choose not to cache + * certain compiled models even if it reports that caching is supported. + * + * If the device reports that caching is not supported, the user may avoid calling + * IDevice::prepareModelFromCache or providing cache file descriptors to + * IDevice::prepareModel. + * + * @return NumberOfCacheFiles structure indicating how many files for model and data cache the + * driver needs to cache a single prepared model. It must be less than or equal to + * MAX_NUMBER_OF_CACHE_FILES. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if there is an unspecified error + */ + NumberOfCacheFiles getNumberOfCacheFilesNeeded(); + + /** + * Gets information about extensions supported by the driver implementation. + * + * All extension operations and operands must be fully supported for the extension to appear in + * the list of supported extensions. + * + * @return A list of supported extensions. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if there is an unspecified error + */ + Extension[] getSupportedExtensions(); + + /** + * Gets the supported operations in a model. + * + * getSupportedOperations indicates which operations of the top-level subgraph are fully + * supported by the vendor driver. If an operation may not be supported for any reason, + * getSupportedOperations must return false for that operation. + * + * The {@link OperationType::IF} and {@link OperationType::WHILE} operations may only be fully + * supported if the vendor driver fully supports all operations in the referenced subgraphs. + * + * @param model A model whose operations -- and their corresponding operands -- are to be + * verified by the driver. + * @return A list of supported operations, where true indicates the operation is supported and + * false indicates the operation is not supported. The index of "supported" corresponds with + * the index of the operation it is describing in the main subgraph. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if there is an unspecified error + * - INVALID_ARGUMENT if provided model is invalid + */ + boolean[] getSupportedOperations(in Model model); + + /** + * Get the type of a given device. + * + * The device type can be used to help application developers to distribute Machine Learning + * workloads and other workloads such as graphical rendering. E.g., for an app which renders AR + * scenes based on real time object detection results, the developer could choose an ACCELERATOR + * type device for ML workloads, and reserve GPU for graphical rendering. + * + * @return The DeviceType of the device. Please note, this is not a bitfield of DeviceTypes. + * Each device must only be of a single DeviceType. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if the query resulted in an unspecified error + */ + DeviceType getType(); + + /** + * Get the version string of the driver implementation. + * + * The version string must be a unique token among the set of version strings of drivers of a + * specific device. The token identifies the device driver's implementation. The token must not + * be confused with the feature level which is solely defined by the interface version. This API + * is opaque to the Android framework, but the Android framework may use the information for + * debugging or to pass on to NNAPI applications. + * + * Application developers sometimes have specific requirements to ensure good user experiences, + * and they need more information to make intelligent decisions when the Android framework + * cannot. For example, combined with the device name and other information, the token can help + * NNAPI applications filter devices based on their needs: + * - An application demands a certain level of performance, but a specific version of the + * driver cannot meet that requirement because of a performance regression. + * The application can disallow the driver based on the version provided. + * - An application has a minimum precision requirement, but certain versions of + * the driver cannot meet that requirement because of bugs or certain optimizations. + * The application can filter out versions of these drivers. + * + * @return The version string of the device implementation. Must have nonzero length. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if the query resulted in an unspecified error + */ + String getVersionString(); + + /** + * Asynchronously creates a prepared model for execution and optionally saves it into cache + * files. + * + * prepareModel is used to make any necessary transformations to or alternative representations + * to a model for execution, possibly including transformations on the constant data, + * optimization on the model's graph, or compilation into the device's native binary format. The + * model itself is not changed. + * + * Optionally, caching information may be provided for the driver to save the prepared model to + * cache files for faster model compilation time when the same model preparation is requested in + * the future. There are two types of cache file descriptors provided to the driver: model cache + * and data cache. For more information on the two types of cache, refer to + * getNumberOfCacheFilesNeeded. + * + * The file descriptors must be opened with read and write permission. A file may have any size, + * and the corresponding file descriptor may have any offset. The driver must truncate a file to + * zero size before writing to that file. The file descriptors may be closed by the client once + * the asynchronous preparation has finished. The driver must dup a file descriptor if it wants + * to get access to the cache file later. + * + * The model is prepared asynchronously with respect to the caller. The prepareModel function + * must verify the inputs to the preparedModel function related to preparing the model (as + * opposed to saving the prepared model to cache) are correct. If there is an error, + * prepareModel must immediately invoke the callback with the appropriate ErrorStatus value and + * nullptr for the IPreparedModel, then return a status with a service specific exception with + * the same ErrorStatus. If the inputs to the prepareModel function that are related to + * preparing the model are valid and there is no error, prepareModel must launch an asynchronous + * task to prepare the model in the background, and immediately return from prepareModel. If the + * asynchronous task fails to launch, prepareModel must immediately invoke the callback with + * ErrorStatus::GENERAL_FAILURE and nullptr for the IPreparedModel, then return a service + * specific exception with ErrorStatus::GENERAL_FAILURE. + * + * When the asynchronous task has finished preparing the model, it must immediately invoke the + * callback function provided as an input to prepareModel. If the model was prepared + * successfully, the callback object must be invoked with an error status of ErrorStatus::NONE + * and the produced IPreparedModel object. If an error occurred preparing the model, the + * callback object must be invoked with the appropriate ErrorStatus value and nullptr for the + * IPreparedModel. + * + * The model is prepared with a priority. This priority is relative to other prepared models + * owned by the same client. Higher priority executions may use more compute resources than + * lower priority executions, and may preempt or starve lower priority executions. + * + * prepareModel can be called with an optional deadline. If the model is not able to be prepared + * before the provided deadline, the model preparation may be aborted, and either + * {@link ErrorStatus::MISSED_DEADLINE_TRANSIENT} or {@link + * ErrorStatus::MISSED_DEADLINE_PERSISTENT} may be returned. The error due to an abort must be + * sent the same way as other errors, described above. + * + * Optionally, the driver may save the prepared model to cache during the asynchronous + * preparation. Any error that occurs when saving to cache must not affect the status of + * preparing the model. Even if the input arguments related to the cache may be invalid, or the + * driver may fail to save to cache, the prepareModel function must finish preparing the model. + * The driver may choose not to save to cache even if the caching information is provided and + * valid. + * + * The only information that may be unknown to the model at this stage is the shape of the + * tensors, which may only be known at execution time. As such, some driver services may return + * partially prepared models, where the prepared model may only be finished when it is paired + * with a set of inputs to the model. Note that the same prepared model object may be used with + * different shapes of inputs on different (possibly concurrent) executions. + * + * Multiple threads may call prepareModel on the same model concurrently. + * + * @param model The model to be prepared for execution. + * @param preference Indicates the intended execution behavior of a prepared model. + * @param priority The priority of the prepared model relative to other prepared models owned by + * the client. + * @param deadline The time by which the model is expected to be prepared. The time is measured + * in nanoseconds since epoch of the steady clock (as from + * std::chrono::steady_clock). If the model cannot be prepared by the deadline, + * the preparation may be aborted. Passing -1 means the deadline is omitted. + * Other negative values are invalid. + * @param modelCache A vector of file descriptors for the security-sensitive cache. The length + * of the vector must either be 0 indicating that caching information is not + * provided, or match the numModelCache returned from + * getNumberOfCacheFilesNeeded. The cache file descriptors will be provided in + * the same order when retrieving the preparedModel from cache files with + * prepareModelFromCache. + * @param dataCache A vector of file descriptors for the constants' cache. The length of the + * vector must either be 0 indicating that caching information is not provided, + * or match the numDataCache returned from getNumberOfCacheFilesNeeded. The + * cache file descriptors will be provided in the same order when retrieving + * the preparedModel from cache files with prepareModelFromCache. + * @param token A caching token of length BYTE_SIZE_OF_CACHE_TOKEN identifying the prepared + * model. The same token will be provided when retrieving the prepared model from + * the cache files with prepareModelFromCache. Tokens should be chosen to have a + * low rate of collision for a particular application. The driver cannot detect a + * collision; a collision will result in a failed execution or in a successful + * execution that produces incorrect output values. If both modelCache and + * dataCache are empty indicating that caching information is not provided, this + * token must be ignored. + * @param callback A callback object used to return the error status of preparing the model for + * execution and the prepared model if successful, nullptr otherwise. The + * callback object's notify function must be called exactly once, even if the + * model could not be prepared. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if there is an unspecified error + * - INVALID_ARGUMENT if one of the input arguments related to preparing the model is + * invalid + * - MISSED_DEADLINE_* if the preparation is aborted because the model cannot be prepared by + * the deadline + * - RESOURCE_EXHAUSTED_* if the task was aborted by the driver + */ + void prepareModel(in Model model, in ExecutionPreference preference, in Priority priority, + in long deadline, in ParcelFileDescriptor[] modelCache, in ParcelFileDescriptor[] dataCache, + in byte[] token, in IPreparedModelCallback callback); + + /** + * Creates a prepared model from cache files for execution. + * + * prepareModelFromCache is used to retrieve a prepared model directly from cache files to avoid + * slow model compilation time. There are two types of cache file descriptors provided to the + * driver: model cache and data cache. For more information on the two types of cache files, + * refer to getNumberOfCacheFilesNeeded. + * + * The file descriptors must be opened with read and write permission. A file may have any size, + * and the corresponding file descriptor may have any offset. The driver must truncate a file to + * zero size before writing to that file. The file descriptors may be closed by the client once + * the asynchronous preparation has finished. The driver must dup a file descriptor if it wants + * to get access to the cache file later. + * + * The model is prepared asynchronously with respect to the caller. The prepareModelFromCache + * function must verify the inputs to the prepareModelFromCache function are correct, and that + * the security-sensitive cache has not been modified since it was last written by the driver. + * If there is an error, or if compilation caching is not supported, or if the + * security-sensitive cache has been modified, prepareModelFromCache must immediately invoke the + * callback with the appropriate ErrorStatus value and nullptr for the IPreparedModel, then + * return a status with a service specific exception with the same ErrorStatus. If the inputs to + * the prepareModelFromCache function are valid, the security-sensitive cache is not modified, + * and there is no error, prepareModelFromCache must launch an asynchronous task to prepare the + * model in the background, and immediately return from prepareModelFromCache. If the + * asynchronous task fails to launch, prepareModelFromCache must immediately invoke the callback + * with ErrorStatus::GENERAL_FAILURE and nullptr for the IPreparedModel, then return a service + * specific exception with ErrorStatus::GENERAL_FAILURE. + * + * When the asynchronous task has finished preparing the model, it must immediately invoke the + * callback function provided as an input to prepareModelFromCache. If the model was prepared + * successfully, the callback object must be invoked with an error status of ErrorStatus::NONE + * and the produced IPreparedModel object. If an error occurred preparing the model, the + * callback object must be invoked with the appropriate ErrorStatus value and nullptr for the + * IPreparedModel. + * + * prepareModelFromCache can be called with an optional deadline. If the model is not able to + * prepared before the provided deadline, the model preparation may be aborted, and either + * {@link ErrorStatus::MISSED_DEADLINE_TRANSIENT} or + * {@link ErrorStatus::MISSED_DEADLINE_PERSISTENT} may be returned. The error due to an abort + * must be sent the same way as other errors, described above. + * + * The only information that may be unknown to the model at this stage is the shape of the + * tensors, which may only be known at execution time. As such, some driver services may return + * partially prepared models, where the prepared model may only be finished when it is paired + * with a set of inputs to the model. Note that the same prepared model object may be used with + * different shapes of inputs on different (possibly concurrent) executions. + * + * @param deadline The time by which the model is expected to be prepared. The time is measured + * in nanoseconds since epoch of the steady clock (as from + * std::chrono::steady_clock). If the model cannot be prepared by the deadline, + * the preparation may be aborted. Passing -1 means the deadline is omitted. + * Other negative values are invalid. + * @param modelCache A vector of file descriptors for the security-sensitive cache. The length + * of the vector must match the numModelCache returned from + * getNumberOfCacheFilesNeeded. The cache file descriptors will be provided in + * the same order as with prepareModel. + * @param dataCache A vector of file descriptors for the constants' cache. The length of the + * vector must match the numDataCache returned from + * getNumberOfCacheFilesNeeded. The cache file descriptors will be provided in + * the same order as with prepareModel. + * @param token A caching token of length BYTE_SIZE_OF_CACHE_TOKEN identifying the prepared + * model. It is the same token provided when saving the cache files with + * prepareModel. Tokens should be chosen to have a low rate of collision for a + * particular application. The driver cannot detect a collision; a collision will + * result in a failed execution or in a successful execution that produces + * incorrect output values. + * @param callback A callback object used to return the error status of preparing the model for + * execution and the prepared model if successful, nullptr otherwise. The + * callback object's notify function must be called exactly once, even if the + * model could not be prepared. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if caching is not supported or if there is an unspecified error + * - INVALID_ARGUMENT if one of the input arguments is invalid + * - MISSED_DEADLINE_* if the preparation is aborted because the model cannot be prepared by + * the deadline + * - RESOURCE_EXHAUSTED_* if the task was aborted by the driver + */ + void prepareModelFromCache(in long deadline, in ParcelFileDescriptor[] modelCache, + in ParcelFileDescriptor[] dataCache, in byte[] token, in IPreparedModelCallback callback); +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IFencedExecutionCallback.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IFencedExecutionCallback.aidl new file mode 100644 index 0000000000..47e5916665 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IFencedExecutionCallback.aidl @@ -0,0 +1,56 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.ErrorStatus; +import android.hardware.neuralnetworks.Timing; + +/** + * IFencedExecutionCallback can be used to query the error status result and duration information + * from an IPreparedModel::executeFenced call. + */ +@VintfStability +interface IFencedExecutionCallback { + /** + * The getExecutionInfo method is used by the clients to query error status result and duration + * information. The method must only be called after the actual evaluation has finished or + * resulted in an runtime error, as indicated by the status of the sync fence returned by the + * IPreparedModel::executeFenced call, otherwise GENERAL_FAILURE must be returned. + * + * @param out timingLaunched The duration starts when executeFenced is called and ends when + * executeFenced signals the returned syncFence. Unless measureTiming + * was set to true when launching the execution and status is NONE, + * all times must be reported as -1. A driver may choose to report any + * time as -1, indicating that particular measurement is not + * available. + * @param out timingFenced The duration starts when all waitFor sync fences have been signaled + * and ends when executeFenced signals the returned syncFence. Unless + * measureTiming was set to true when launching the execution and status + * is NONE, all times must be reported as -1. A driver may choose to + * report any time as -1, indicating that particular measurement is not + * available. + * @return Error status returned from the asynchronously dispatched execution must be: + * - NONE if the asynchronous execution was successful + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if the asynchronous task resulted in an unspecified error + * - MISSED_DEADLINE_* if the execution is aborted because it cannot be completed by the + * deadline + * - RESOURCE_EXHAUSTED_* if the task was aborted by the driver + */ + ErrorStatus getExecutionInfo(out Timing timingLaunched, out Timing timingFenced); +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl new file mode 100644 index 0000000000..c1b2992010 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModel.aidl @@ -0,0 +1,173 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.common.NativeHandle; +import android.hardware.neuralnetworks.ErrorStatus; +import android.hardware.neuralnetworks.ExecutionResult; +import android.hardware.neuralnetworks.IFencedExecutionCallback; +import android.hardware.neuralnetworks.Request; + +/** + * IPreparedModel describes a model that has been prepared for execution and is used to launch + * executions. + */ +@VintfStability +interface IPreparedModel { + /** + * Each {@link OperationType::WHILE} operation in the model has an implicit execution timeout + * duration associated with it ("loop timeout duration"). This duration is configurable on a + * per-execution basis and must not exceed 15 seconds. The default value is 2 seconds. The units + * are nanoseconds. + */ + const long DEFAULT_LOOP_TIMEOUT_DURATION_NS = 2000000000; + const long MAXIMUM_LOOP_TIMEOUT_DURATION_NS = 15000000000; + + /** + * Performs a synchronous execution on a prepared model. + * + * The execution is performed synchronously with respect to the caller. executeSynchronously + * must verify the inputs to the function are correct, and the usages of memory pools allocated + * by IDevice::allocate are valid. If there is an error, executeSynchronously must immediately + * return a service specific exception with the appropriate ErrorStatus value. If the inputs to + * the function are valid and there is no error, executeSynchronously must perform the + * execution, and must not return until the execution is complete. + * + * The caller must not change the content of any data object referenced by 'request' (described + * by the {@link DataLocation} of a {@link RequestArgument}) until executeSynchronously returns. + * executeSynchronously must not change the content of any of the data objects corresponding to + * 'request' inputs. + * + * If the prepared model was prepared from a model wherein all tensor operands have fully + * specified dimensions, and the inputs to the function are valid, and at execution time every + * operation's input operands have legal values, then the execution should complete + * successfully: there must be no failure unless the device itself is in a bad state. + * + * executeSynchronously may be called with an optional deadline. If the execution is not able to + * be completed before the provided deadline, the execution may be aborted, and either + * {@link ErrorStatus::MISSED_DEADLINE_TRANSIENT} or {@link + * ErrorStatus::MISSED_DEADLINE_PERSISTENT} may be returned. The error due to an abort must be + * sent the same way as other errors, described above. + * + * Any number of calls to the execute* functions, in any combination, may be made concurrently, + * even on the same IPreparedModel object. + * + * @param request The input and output information on which the prepared model is to be + * executed. + * @param measure Specifies whether or not to measure duration of the execution. The duration + * runs from the time the driver sees the call to the executeSynchronously + * function to the time the driver returns from the function. + * @param deadline The time by which the execution is expected to complete. The time is measured + * in nanoseconds since epoch of the steady clock (as from + * std::chrono::steady_clock). If the execution cannot be finished by the + * deadline, the execution may be aborted. Passing -1 means the deadline is + * omitted. Other negative values are invalid. + * @param loopTimeoutDuration The maximum amount of time in nanoseconds that should be spent + * executing a {@link OperationType::WHILE} operation. If a loop + * condition model does not output false within this duration, the + * execution must be aborted. If -1 is provided, the maximum amount + * of time is {@link DEFAULT_LOOP_TIMEOUT_DURATION_NS}. Other + * negative values are invalid. When provided, the duration must not + * exceed {@link MAXIMUM_LOOP_TIMEOUT_DURATION_NS}. + * @return ExecutionResult parcelable, containing the status of the execution, output shapes and + * timing information. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if there is an unspecified error + * - INVALID_ARGUMENT if one of the input arguments is invalid + * - MISSED_DEADLINE_* if the execution is aborted because it cannot be completed by the + * deadline + * - RESOURCE_EXHAUSTED_* if the task was aborted by the driver + */ + ExecutionResult executeSynchronously(in Request request, in boolean measureTiming, + in long deadline, in long loopTimeoutDuration); + + /** + * Launch a fenced asynchronous execution on a prepared model. + * + * The execution is performed asynchronously with respect to the caller. executeFenced must + * verify the inputs to the function are correct, and the usages of memory pools allocated by + * IDevice::allocate are valid. If there is an error, executeFenced must immediately return a + * service specific exception with the corresponding ErrorStatus. If the inputs to the function + * are valid and there is no error, executeFenced must dispatch an asynchronous task to perform + * the execution in the background, assign a sync fence that will be signaled once the execution + * is completed and immediately return a callback that can be used by the client to query the + * duration and runtime error status. If the task has finished before the call returns, + * syncFence file descriptor may be set to -1. The execution must wait for all the sync fences + * (if any) in waitFor to be signaled before starting the actual execution. + * + * When the asynchronous task has finished its execution, it must immediately signal the + * syncFence returned from the executeFenced call. After the syncFence is signaled, the task + * must not modify the content of any data object referenced by 'request' (described by the + * {@link DataLocation} of a {@link RequestArgument}). + * + * executeFenced may be called with an optional deadline and an optional duration. If the + * execution is not able to be completed before the provided deadline or within the timeout + * duration (measured from when all sync fences in waitFor are signaled), whichever comes + * earlier, the execution may be aborted, and either + * {@link ErrorStatus::MISSED_DEADLINE_TRANSIENT} or {@link + * ErrorStatus::MISSED_DEADLINE_PERSISTENT} may be returned. The error due to an abort must be + * sent the same way as other errors, described above. + * + * If any of the sync fences in waitFor changes to error status after the executeFenced call + * succeeds, or the execution is aborted because it cannot finish before the deadline has been + * reached or the duration has elapsed, the driver must immediately set the returned syncFence + * to error status. + * + * Any number of calls to the execute* functions, in any combination, may be made concurrently, + * even on the same IPreparedModel object. + * + * @param request The input and output information on which the prepared model is to be + * executed. The outputs in the request must have fully specified dimensions. + * @param waitFor A vector of sync fence file descriptors. Execution must not start until all + * sync fences have been signaled. + * @param measure Specifies whether or not to measure duration of the execution. + * @param deadline The time by which the execution is expected to complete. The time is measured + * in nanoseconds since epoch of the steady clock (as from + * std::chrono::steady_clock).If the execution cannot be finished by the + * deadline, the execution may be aborted. Passing -1 means the deadline is + * omitted. Other negative values are invalid. + * @param loopTimeoutDuration The maximum amount of time in nanoseconds that should be spent + * executing a {@link OperationType::WHILE} operation. If a loop + * condition model does not output false within this duration, the + * execution must be aborted. If -1 is provided, the maximum amount + * of time is {@link DEFAULT_LOOP_TIMEOUT_DURATION_NS}. Other + * negative values are invalid. When provided, the duration must not + * exceed {@link MAXIMUM_LOOP_TIMEOUT_DURATION_NS}. + * @param duration The length of time in nanoseconds within which the execution is expected to + * complete after all sync fences in waitFor are signaled. If the execution + * cannot be finished within the duration, the execution may be aborted. Passing + * -1 means the duration is omitted. Other negative values are invalid. + * @param out syncFence The sync fence that will be signaled when the task is completed. The + * sync fence will be set to error if a critical error, e.g. hardware + * failure or kernel panic, occurs when doing execution. + * @return The IFencedExecutionCallback can be used to query information like duration and error + * status when the execution is completed. + * @throws ServiceSpecificException with one of the following ErrorStatus values: + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if there is an unspecified error + * - INVALID_ARGUMENT if one of the input arguments is invalid, including fences in error + * states. + * - MISSED_DEADLINE_* if the execution is aborted because it cannot be completed by the + * deadline + * - RESOURCE_EXHAUSTED_* if the task was aborted by the driver + */ + IFencedExecutionCallback executeFenced(in Request request, in ParcelFileDescriptor[] waitFor, + in boolean measureTiming, in long deadline, in long loopTimeoutDuration, in long duration, + out @nullable ParcelFileDescriptor syncFence); +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModelCallback.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModelCallback.aidl new file mode 100644 index 0000000000..adb421830c --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModelCallback.aidl @@ -0,0 +1,51 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.ErrorStatus; +import android.hardware.neuralnetworks.IPreparedModel; + +/** + * IPreparedModelCallback must be used to return a prepared model produced by an asynchronous task + * launched from IDevice::prepareModel*. + */ +@VintfStability +interface IPreparedModelCallback { + /** + * Notify must be invoked immediately after the asynchronous task holding this callback has + * finished preparing the model. If the model was successfully prepared, the method must be + * invoked with ErrorStatus::NONE and the prepared model. If the model was not able to be + * successfully prepared, the method must be invoked with the appropriate ErrorStatus and + * nullptr as the IPreparedModel. If the asynchronous task holding this callback fails to launch + * or if the model provided to IDevice::prepareModel is invalid, notify method must be invoked + * with the appropriate error as well as nullptr for the IPreparedModel. + * + * @param status Error status returned from the asynchronous model preparation task; must be: + * - NONE if the asynchronous task successfully prepared the model + * - DEVICE_UNAVAILABLE if driver is offline or busy + * - GENERAL_FAILURE if the asynchronous task resulted in an unspecified error + * - INVALID_ARGUMENT if one of the input arguments to prepareModel is invalid + * - MISSED_DEADLINE_* if the preparation is aborted because the model cannot be + * prepared by the deadline + * - RESOURCE_EXHAUSTED_* if the task was aborted by the driver + * @param preparedModel A model that has been asynchronously prepared for execution. If the + * model was unable to be prepared due to an error, nullptr must be passed + * in place of the IPreparedModel object. + */ + void notify(in ErrorStatus status, in IPreparedModel preparedModel); +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModelParcel.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModelParcel.aidl new file mode 100644 index 0000000000..f198c3f056 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/IPreparedModelParcel.aidl @@ -0,0 +1,28 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.IPreparedModel; + +/** + * A parcelable for passing a vector of IPreparedModel objects. + */ +@VintfStability +parcelable IPreparedModelParcel { + IPreparedModel preparedModel; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Memory.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Memory.aidl new file mode 100644 index 0000000000..8ecb067f30 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Memory.aidl @@ -0,0 +1,31 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package android.hardware.neuralnetworks; +import android.hardware.common.NativeHandle; + +import android.os.ParcelFileDescriptor; + +/** + * A type that is used to pass pieces of shared memory between processes. + * The type structure mimics hidl_memory type from HIDL. + */ +@VintfStability +parcelable Memory { + NativeHandle handle; + long size; + String name; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Model.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Model.aidl new file mode 100644 index 0000000000..3bb73185f4 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Model.aidl @@ -0,0 +1,70 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.ExtensionNameAndPrefix; +import android.hardware.neuralnetworks.Subgraph; +import android.hardware.neuralnetworks.Memory; + +/** + * A Neural Network Model. + * + * This includes not only the execution graph, but also constant data such as weights or scalars + * added at construction time. The only information that may not be known is the shape of the input + * tensors. + */ +@VintfStability +parcelable Model { + /** + * The top-level subgraph. + */ + Subgraph main; + /** + * Referenced subgraphs. + * + * Each subgraph is referenced by the main subgraph or at least one other referenced subgraph. + * + * There must be no reference cycles. + */ + Subgraph[] referenced; + /** + * A byte buffer containing operand data that were copied into the model. + * + * An operand's value must be located here if and only if Operand::lifetime equals + * OperandLifeTime::CONSTANT_COPY. + */ + byte[] operandValues; + /** + * A collection of shared memory pools containing operand values. + * + * An operand's value must be located here if and only if Operand::lifetime equals + * OperandLifeTime::CONSTANT_POOL. + */ + Memory[] pools; + /** + * 'true' indicates TENSOR_FLOAT32 may be calculated with range and/or precision as low as that + * of the IEEE 754 16-bit floating-point format. + * 'false' indicates TENSOR_FLOAT32 must be calculated using at least the range and precision of + * the IEEE 754 32-bit floating-point format. + */ + boolean relaxComputationFloat32toFloat16; + /** + * The mapping between extension names and prefixes of operand and operation type values. + */ + ExtensionNameAndPrefix[] extensionNameToPrefix; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/NumberOfCacheFiles.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/NumberOfCacheFiles.aidl new file mode 100644 index 0000000000..1ca2676646 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/NumberOfCacheFiles.aidl @@ -0,0 +1,27 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package android.hardware.neuralnetworks; + +/** + * Structure indicating how many files for model and numDataCache cache the driver needs to cache a + * single prepared model. + */ +@VintfStability +parcelable NumberOfCacheFiles { + int numModelCache; + int numDataCache; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Operand.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Operand.aidl new file mode 100644 index 0000000000..243a89d719 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Operand.aidl @@ -0,0 +1,113 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.DataLocation; +import android.hardware.neuralnetworks.OperandExtraParams; +import android.hardware.neuralnetworks.OperandLifeTime; +import android.hardware.neuralnetworks.OperandType; + +/** + * Describes one operand of the model's graph. + */ +@VintfStability +parcelable Operand { + /** + * The data type. + * + * Besides the values listed in {@link OperandType}, any value above + * {@link IDevice::OPERAND_TYPE_BASE_MAX} is possible and should be interpreted as an extension + * type according to {@link Model::extensionNameToPrefix}. + */ + OperandType type; + /** + * Dimensions of the operand. + * + * For a scalar operand, dimensions.size() must be 0. + * + * A tensor operand with all dimensions specified has "fully specified" dimensions. Whenever + * possible (i.e., whenever the dimensions are known at model construction time), a tensor + * operand should have (but is not required to have) fully specified dimensions, in order to + * enable the best possible performance. + * + * If a tensor operand's dimensions are not fully specified, the dimensions of the operand are + * deduced from the operand dimensions and values of the operation for which that operand is an + * output or from the corresponding {@link OperationType::IF} or {@link OperationType::WHILE} + * operation input operand dimensions in the case of referenced subgraph input operands. + * + * In the following situations, a tensor operand's dimensions must be fully specified: + * + * . The operand has lifetime CONSTANT_COPY or CONSTANT_POOL. + * + * . The operand has lifetime SUBGRAPH_INPUT and belongs to the main subgraph. Fully + * specified dimensions must either be present in the Operand or they must be provided in + * the corresponding RequestArgument. + * EXCEPTION: If the input is optional and omitted (by setting the hasNoValue field of the + * corresponding RequestArgument to true) then it need not have fully specified + * dimensions. + * + * A tensor operand with some number of unspecified dimensions is represented by setting each + * unspecified dimension to 0. + * + * A tensor operand with unspecified rank is represented by providing an empty dimensions + * vector. + */ + int[] dimensions; + /** + * Quantized scale of the operand. + * + * Must be 0 when not applicable to an operand type. + * + * See {@link OperandType}. + */ + float scale; + /** + * Quantized zero-point offset of the operand. + * + * Must be 0 when not applicable to an operand type. + * + * See {@link OperandType}. + */ + int zeroPoint; + /** + * How the operand is used. + */ + OperandLifeTime lifetime; + /** + * Where to find the data for this operand. + * If the lifetime is TEMPORARY_VARIABLE, SUBGRAPH_INPUT, SUBGRAPH_OUTPUT, or NO_VALUE: + * - All the fields must be 0. + * If the lifetime is CONSTANT_COPY: + * - location.poolIndex is 0. + * - location.offset is the offset in bytes into Model.operandValues. + * - location.length is set. + * If the lifetime is CONSTANT_POOL: + * - location.poolIndex is set. + * - location.offset is the offset in bytes into the specified pool. + * - location.length is set. + * If the lifetime is SUBGRAPH: + * - location.poolIndex is 0. + * - location.offset is the index of the referenced subgraph in {@link Model::referenced}. + * - location.length is 0. + */ + DataLocation location; + /** + * Additional parameters specific to a particular operand type. + */ + @nullable OperandExtraParams extraParams; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandExtraParams.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandExtraParams.aidl new file mode 100644 index 0000000000..b0112aea0c --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandExtraParams.aidl @@ -0,0 +1,40 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.SymmPerChannelQuantParams; + +/** + * Parameters specific to a particular operand type. + */ +@VintfStability +union OperandExtraParams { + /** + * Symmetric per-channel quantization parameters. + * + * Only applicable to operands of type TENSOR_QUANT8_SYMM_PER_CHANNEL. + */ + SymmPerChannelQuantParams channelQuant; + /** + * Extension operand parameters. + * + * The framework treats this as an opaque data blob. + * The format is up to individual extensions. + */ + byte[] extension; +} \ No newline at end of file diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandLifeTime.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandLifeTime.aidl new file mode 100644 index 0000000000..63d1971717 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandLifeTime.aidl @@ -0,0 +1,63 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * How an operand is used. + */ +@VintfStability +@Backing(type="int") +enum OperandLifeTime { + /** + * The operand is internal to the model. It's created by an operation and consumed by other + * operations. It must be an output operand of exactly one operation. + */ + TEMPORARY_VARIABLE, + /** + * The operand is an input of a subgraph. It must not be an output operand of any operation. + * + * An operand can't be both input and output of a subgraph. + */ + SUBGRAPH_INPUT, + /** + * The operand is an output of a subgraph. It must be an output operand of exactly one + * operation. + * + * An operand can't be both input and output of a subgraph. + */ + SUBGRAPH_OUTPUT, + /** + * The operand is a constant found in Model.operandValues. It must not be an output operand of + * any operation. + */ + CONSTANT_COPY, + /** + * The operand is a constant that was specified via a Memory object. It must not be an output + * operand of any operation. + */ + CONSTANT_POOL, + /** + * The operand does not have a value. This is valid only for optional arguments of operations. + */ + NO_VALUE, + /** + * The operand is a reference to a subgraph. It must be an input to one or more + * {@link OperationType::IF} or {@link OperationType::WHILE} operations. + */ + SUBGRAPH, +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandPerformance.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandPerformance.aidl new file mode 100644 index 0000000000..9a8c2cca23 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandPerformance.aidl @@ -0,0 +1,31 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.OperandType; +import android.hardware.neuralnetworks.PerformanceInfo; + +/** + * Driver performance when operating on a particular data type. In the case of float32 data, this is + * used when the calculations are not relaxed. + */ +@VintfStability +parcelable OperandPerformance { + OperandType type; + PerformanceInfo info; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandType.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandType.aidl new file mode 100644 index 0000000000..9274b6f97e --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperandType.aidl @@ -0,0 +1,154 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Operand types. + * + * The type of an operand in a model. + * + * Types prefaced with TENSOR_* must be used for tensor data (i.e., tensors + * with at least one dimension). Types not prefaced by TENSOR_* represent + * scalar values and must have no dimensions. + */ +@VintfStability +@Backing(type="int") +enum OperandType { + /** + * A 32 bit floating point scalar value. + */ + FLOAT32 = 0, + /** + * A signed 32 bit integer scalar value. + */ + INT32 = 1, + /** + * An unsigned 32 bit integer scalar value. + */ + UINT32 = 2, + /** + * A tensor of 32 bit floating point values. + */ + TENSOR_FLOAT32 = 3, + /** + * A tensor of 32 bit integer values. + */ + TENSOR_INT32 = 4, + /** + * A tensor of 8 bit unsigned integers that represent real numbers. + * + * Attached to this tensor are two numbers that can be used to convert the 8 bit integer to the + * real value and vice versa. These two numbers are: + * - scale: a 32 bit floating point value greater than zero. + * - zeroPoint: a 32 bit integer, in range [0, 255]. + * + * The formula is: + * real_value = (integer_value - zeroPoint) * scale. + */ + TENSOR_QUANT8_ASYMM = 5, + /** + * An 8 bit boolean scalar value. + * + * Values of this operand type are either true or false. A zero value represents false; any + * other value represents true. + */ + BOOL = 6, + /** + * A tensor of 16 bit signed integers that represent real numbers. + * + * Attached to this tensor is a number representing real value scale that is used to convert the + * 16 bit number to a real value in the following way: + * realValue = integerValue * scale. + * + * scale is a 32 bit floating point with value greater than zero. + */ + TENSOR_QUANT16_SYMM = 7, + /** + * A tensor of IEEE 754 16 bit floating point values. + */ + TENSOR_FLOAT16 = 8, + /** + * A tensor of 8 bit boolean values. + * + * Values of this operand type are either true or false. A zero value represents false; any + * other value represents true. + */ + TENSOR_BOOL8 = 9, + /** + * An IEEE 754 16 bit floating point scalar value. + */ + FLOAT16 = 10, + /** + * A tensor of 8 bit signed integers that represent real numbers. + * + * This tensor is associated with additional fields that can be used to convert the 8 bit signed + * integer to the real value and vice versa. These fields are: + * - channelDim: a 32 bit unsigned integer indicating channel dimension. + * - scales: an array of positive 32 bit floating point values. + * The size of the scales array must be equal to dimensions[channelDim]. + * + * {@link SymmPerChannelQuantParams} must hold the parameters for an Operand of this type. + * The channel dimension of this tensor must not be unknown (dimensions[channelDim] != 0). + * + * The formula is: + * realValue[..., C, ...] = + * integerValue[..., C, ...] * scales[C] + * where C is an index in the Channel dimension. + */ + TENSOR_QUANT8_SYMM_PER_CHANNEL = 11, + /** + * A tensor of 16 bit unsigned integers that represent real numbers. + * + * Attached to this tensor are two numbers that can be used to convert the 16 bit integer to the + * real value and vice versa. These two numbers are: + * - scale: a 32 bit floating point value greater than zero. + * - zeroPoint: a 32 bit integer, in range [0, 65535]. + * + * The formula is: + * real_value = (integer_value - zeroPoint) * scale. + */ + TENSOR_QUANT16_ASYMM = 12, + /** + * A tensor of 8 bit signed integers that represent real numbers. + * + * Attached to this tensor is a number representing real value scale that is used to convert the + * 8 bit number to a real value in the following way: + * realValue = integerValue * scale. + * + * scale is a 32 bit floating point with value greater than zero. + */ + TENSOR_QUANT8_SYMM = 13, + /** + * A tensor of 8 bit signed integers that represent real numbers. + * + * Attached to this tensor are two numbers that can be used to convert the 8 bit integer to the + * real value and vice versa. These two numbers are: + * - scale: a 32 bit floating point value greater than zero. + * - zeroPoint: a 32 bit integer, in range [-128, 127]. + * + * The formula is: + * real_value = (integer_value - zeroPoint) * scale. + */ + TENSOR_QUANT8_ASYMM_SIGNED = 14, + /** + * A reference to a subgraph. + * + * Must have the lifetime {@link OperandLifeTime::SUBGRAPH}. + */ + SUBGRAPH = 15, +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Operation.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Operation.aidl new file mode 100644 index 0000000000..acfb4b779f --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Operation.aidl @@ -0,0 +1,46 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.OperationType; + +/** + * Describes one operation of the model's graph. + */ +@VintfStability +parcelable Operation { + /** + * The operation type. + * + * Besides the values listed in {@link OperationType}, any value above + * {@link IDevice::OPERATION_TYPE_BASE_MAX} is possible and should be interpreted as an + * extension type according to {@link Model::extensionNameToPrefix}. + */ + OperationType type; + /** + * Describes the table that contains the indexes of the inputs of the operation. The offset is + * the index in the operandIndexes table. + */ + int[] inputs; + /** + * Describes the table that contains the indexes of the outputs of the operation. The offset is + * the index in the operandIndexes table. + */ + int[] outputs; +} + diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl new file mode 100644 index 0000000000..fd9da67bce --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/OperationType.aidl @@ -0,0 +1,5132 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Operation types. + * + * The type of an operation in a model. + */ +@VintfStability +@Backing(type="int") +enum OperationType { + /** + * Adds two tensors, element-wise. + * + * Takes two input tensors of identical {@link OperandType} and compatible + * dimensions. The output is the sum of both input tensors, optionally + * modified by an activation function. + * + * Two dimensions are compatible when: + * 1. they are equal, or + * 2. one of them is 1 + * + * The size of the output is the maximum size along each dimension of the + * input operands. It starts with the trailing dimensions, and works its + * way forward. + * + * Example: + * + * input1.dimension = {4, 1, 2} + * input2.dimension = {5, 4, 3, 1} + * output.dimension = {5, 4, 3, 2} + * + * Since HAL version 1.2, generic zero-sized input tensor is supported. Zero + * dimension is only compatible with 0 or 1. The size of the output + * dimension is zero if either of corresponding input dimension is zero. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * * {@link OperandType::TENSOR_INT32} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType}, and compatible dimensions + * as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scales and zeroPoint can be different from input0 scale and zeroPoint. + * * 2: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * For a {@link OperandType::TENSOR_INT32} tensor, + * the {@link FusedActivationFunc} must be "NONE". + * + * Outputs: + * * 0: The sum, a tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint can be different from inputs' scale and zeroPoint. + */ + ADD = 0, + /** + * Performs a 2-D average pooling operation. + * + * The output dimensions are functions of the filter dimensions, stride, and + * padding. + * + * The values in the output tensor are computed as: + * + * output[b, i, j, channel] = + * sum_{di, dj}( + * input[b, strides[1] * i + di, strides[2] * j + dj, channel] + * ) / sum(1) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Both explicit padding and implicit padding are supported. + * + * Inputs (explicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: An {@link OperandType::INT32} scalar, specifying the padding on + * the left, in the ‘width’ dimension. + * * 2: An {@link OperandType::INT32} scalar, specifying the padding on + * the right, in the ‘width’ dimension. + * * 3: An {@link OperandType::INT32} scalar, specifying the padding on + * the top, in the ‘height’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the padding on + * the bottom, in the ‘height’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 7: An {@link OperandType::INT32} scalar, specifying the filter + * width. + * * 8: An {@link OperandType::INT32} scalar, specifying the filter + * height. + * * 9: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 10: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Inputs (implicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: An {@link OperandType::INT32} scalar, specifying the implicit + * padding scheme, has to be one of the + * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. + * * 2: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 3: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the filter + * width. + * * 5: An {@link OperandType::INT32} scalar, specifying the filter + * height. + * * 6: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 7: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, out_height, out_width, depth]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + AVERAGE_POOL_2D = 1, + /** + * Concatenates the input tensors along the given dimension. + * + * The input tensors must have identical {@link OperandType} and the same + * dimensions except the dimension along the concatenation axis. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * (full support since HAL version 1.2, see the input section) + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0 ~ n-1: The list of n input tensors, of shape + * [D0, D1, ..., Daxis(i), ..., Dm]. + * Before HAL version 1.2, all input tensors of + * {@link OperandType::TENSOR_QUANT8_ASYMM} + * must have the same scale and zeroPoint as the output tensor. + * Input tensors of + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * are allowed to have different scale and zeroPoint. + * Since HAL version 1.2, zero-sized tensors are supported. + * * n: An {@link OperandType::INT32} scalar, specifying the + * concatenation axis. + * + * Outputs: + * * 0: The output, a tensor of the same {@link OperandType} as the input + * tensors. The output shape is [D0, D1, ..., sum(Daxis(i)), ..., Dm]. + * Since HAL version 1.2, for a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, + * the scale and zeroPoint values can be different from + * input tensors. Before HAL version 1.2 they have to be the same as for the input tensors. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint values can be different from input tensors. + */ + CONCATENATION = 2, + /** + * Performs a 2-D convolution operation. + * + * The CONV_2D op sweeps a 2-D filter that can mix channels together over a + * batch of images, applying the filter to each window of each image of the + * appropriate size. + * + * The output dimensions are functions of the filter dimensions, stride, and + * padding. + * + * The values in the output tensor are computed as: + * + * output[b, i, j, channel] = + * sum_{di, dj, k} ( + * input[b, strides[1] * i + di, strides[2] * j + dj, k] * + * filter[channel, di, dj, k] + * ) + bias[channel] + * + * Supported tensor {@link OperandType} configurations: + * * 32 bit floating point: + * * * {@link OperandType::TENSOR_FLOAT32} for input, filter, output, and bias. + * + * * Quantized: + * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, filter, and output. + * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to + * * * input.scale * filter.scale). + * + * Available since HAL version 1.2: + * * 16 bit floating point: + * * * {@link OperandType::TENSOR_FLOAT16} for input, filter, output, and bias. + * + * * Quantized with symmetric per channel quantization for the filter: + * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, and output. + * * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} for filter. + * * * {@link OperandType::TENSOR_INT32} for bias (scale set to 0.0, + * * * each value scaling is separate and equal to input.scale * filter.scales[channel]). + * + * Available since HAL version 1.3: + * * Quantized signed (since HAL version 1.3): + * * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} for input, filter, and output. + * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to + * * * input.scale * filter.scale). + * + * * Quantized signed with filter symmetric per channel quantization (since HAL version 1.3): + * * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} for input, and output. + * * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} for filter. + * * * {@link OperandType::TENSOR_INT32} for bias (scale set to 0.0, + * * * each value scaling is separate and equal to input.scale * filter.scales[channel]). + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Both explicit padding and implicit padding are supported. + * + * Inputs (explicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: A 4-D tensor, of shape + * [depth_out, filter_height, filter_width, depth_in], specifying the + * filter. + * For tensor of type {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} + * the channel dimension (SymmPerChannelQuantParams::channelDim) + * must be set to 0. + * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input + * tensor of type {@link OperandType::TENSOR_FLOAT32} + * or {@link OperandType::TENSOR_FLOAT16} the bias must be of the same type. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint + * of 0 and bias_scale == input_scale * filter_scale. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL}, + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint of 0 + * and bias_scale of 0. The actual scale of each value 'i' is equal to + * bias_scale[i] = input_scale * filter_scale[i]. + * * 3: An {@link OperandType::INT32} scalar, specifying the padding on + * the left, in the ‘width’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the padding on + * the right, in the ‘width’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the padding on + * the top, in the ‘height’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the padding on + * the bottom, in the ‘height’ dimension. + * * 7: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 8: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 9: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 10: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * * 11: An optional {@link OperandType::INT32} scalar, specifying the dilation + * factor for width. Defaults to 1. If set to k > 1, there will be k-1 skipped + * cells between each filter element on width dimension. If this input is set, + * input 12 (dilation factor for height) must be specified as well. + * Available since HAL version 1.2. + * * 12: An optional {@link OperandType::INT32} scalar, specifying the dilation + * factor for height. Defaults to 1. If set to k > 1, there will be k-1 skipped + * cells between each filter element on height dimension. If this input is set, + * input 11 (dilation factor for width) must be specified as well. + * Available since HAL version 1.2. + * + * Inputs (implicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: A 4-D tensor, of shape + * [depth_out, filter_height, filter_width, depth_in], specifying the + * filter. + * For tensor of type {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} + * the channel dimension (SymmPerChannelQuantParams::channelDim) + * must be set to 0. + * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input + * tensor of type {@link OperandType::TENSOR_FLOAT32} + * or {@link OperandType::TENSOR_FLOAT16} the bias must be of the same + * type. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint + * of 0 and bias_scale == input_scale * filter_scale. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL}, + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint of 0 + * and bias_scale of 0. The actual scale of each value 'i' is equal to + * bias_scale[i] = input_scale * filter_scale[i]. + * * 3: An {@link OperandType::INT32} scalar, specifying the implicit + * padding scheme, has to be one of the + * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. + * * 4: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 6: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 7: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * * 8: An optional {@link OperandType::INT32} scalar, specifying the dilation + * factor for width. Defaults to 1. If set to k > 1, there will be k-1 skipped + * cells between each filter element on width dimension. If this input is set, + * input 9 (dilation factor for height) must be specified as well. + * Available since HAL version 1.2. + * * 9: An optional {@link OperandType::INT32} scalar, specifying the dilation + * factor for height. Defaults to 1. If set to k > 1, there will be k-1 skipped + * cells between each filter element on height dimension. If this input is set, + * input 8 (dilation factor for width) must be specified as well. + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, out_height, out_width, depth_out]. + * Before HAL version 1.2, for output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, + * the following condition must be satisfied: output_scale > input_scale * filter_scale + */ + CONV_2D = 3, + /** + * Performs a depthwise 2-D convolution operation. + * + * Given an input tensor of shape [batches, height, width, depth_in] and a + * filter tensor of shape [1, filter_height, filter_width, depth_out] + * containing depth_out convolutional filters of depth 1, DEPTHWISE_CONV + * applies a different filter to each input channel (expanding from 1 + * channel to channel_multiplier channels for each), then concatenates the + * results together. + * + * The output has depth_out = depth_in * depth_multiplier channels. + * The output dimensions are functions of the filter dimensions, stride, and + * padding. + * + * The values in the output tensor are computed as: + * + * output[b, i, j, k * channel_multiplier + q] = + * sum_{di, dj} ( + * input[b, strides[1] * i + di, strides[2] * j + dj, k] * + * filter[1, di, dj, k * channel_multiplier + q] + * ) + bias[k * channel_multiplier + q] + * + * Supported tensor {@link OperandType} configurations: + * * 32 bit floating point: + * * * {@link OperandType::TENSOR_FLOAT32} for input, filter, output, and bias. + * + * * Quantized: + * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, filter, and output. + * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to + * * * input.scale * filter.scale). + * + * Available since HAL version 1.2: + * * 16 bit floating point: + * * * {@link OperandType::TENSOR_FLOAT16} for input, filter, output, and bias. + * + * * Quantized with symmetric per channel quantization for the filter: + * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, and output. + * * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} for filter. + * * * {@link OperandType::TENSOR_INT32} for bias (scale set to 0.0, + * * * each value scaling is separate and equal to input.scale * filter.scales[channel]). + * + * Available since HAL version 1.3: + * * Quantized signed (since HAL version 1.3): + * * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} for input, filter, and output. + * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to + * * * input.scale * filter.scale). + * + * * Quantized signed with filter symmetric per channel quantization (since HAL version 1.3): + * * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} for input, and output. + * * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} for filter. + * * * {@link OperandType::TENSOR_INT32} for bias (scale set to 0.0, + * * * each value scaling is separate and equal to input.scale * filter.scales[channel]). + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Both explicit padding and implicit padding are supported. + * + * Inputs (explicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input. + * * 1: A 4-D tensor, of shape [1, filter_height, filter_width, depth_out], + * specifying the filter. + * For tensor of type {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} + * the channel dimension (SymmPerChannelQuantParams::channelDim) + * must be set to 3. + * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input + * tensor of type {@link OperandType::TENSOR_FLOAT32} + * or {@link OperandType::TENSOR_FLOAT16} the bias must be of the same type. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint + * of 0 and bias_scale == input_scale * filter_scale. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL}, + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint of 0 + * and bias_scale of 0. The actual scale of each value 'i' is equal to + * bias_scale[i] = input_scale * filter_scale[i]. + * * 3: An {@link OperandType::INT32} scalar, specifying the padding on + * the left, in the ‘width’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the padding on + * the right, in the ‘width’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the padding on + * the top, in the ‘height’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the padding on + * the bottom, in the ‘height’ dimension. + * * 7: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 8: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 9: An {@link OperandType::INT32} scalar, specifying the depthwise + * multiplier. + * * 10: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 11: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * * 12: An optional {@link OperandType::INT32} scalar, specifying the dilation + * factor for width. Defaults to 1. If set to k > 1, there will be k-1 skipped + * cells between each filter element on width dimension. If this input is set, + * input 13 (dilation factor for height) must be specified as well. + * Available since HAL version 1.2. + * * 13: An optional {@link OperandType::INT32} scalar, specifying the dilation + * factor for height. Defaults to 1. If set to k > 1, there will be k-1 skipped + * cells between each filter element on height dimension. If this input is set, + * input 12 (dilation factor for width) must be specified as well. + * Available since HAL version 1.2. + * + * Inputs (implicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input. + * * 1: A 4-D tensor, of shape [1, filter_height, filter_width, depth_out], + * specifying the filter. + * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input + * tensor of type {@link OperandType::TENSOR_FLOAT32} + * or {@link OperandType::TENSOR_FLOAT16} the bias must be of the same type. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint + * of 0 and bias_scale == input_scale * filter_scale. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL}, + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint of 0 + * and bias_scale of 0. The actual scale of each value 'i' is equal to + * bias_scale[i] = input_scale * filter_scale[i]. + * * 3: An {@link OperandType::INT32} scalar, specifying the implicit + * padding scheme, has to be one of the + * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. + * * 4: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the depthwise + * multiplier. + * * 7: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 8: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * * 9: An optional {@link OperandType::INT32} scalar, specifying the dilation + * factor for width. Defaults to 1. If set to k > 1, there will be k-1 skipped + * cells between each filter element on width dimension. If this input is set, + * input 10 (dilation factor for height) must be specified as well. + * Available since HAL version 1.2. + * * 10: An optional {@link OperandType::INT32} scalar, specifying the dilation + * factor for height. Defaults to 1. If set to k > 1, there will be k-1 skipped + * cells between each filter element on height dimension. If this input is set, + * input 9 (dilation factor for width) must be specified as well. + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, out_height, out_width, depth_out]. Before HAL version 1.2, for + * output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, + * the following condition must be satisfied: + * output_scale > input_scale * filter_scale + */ + DEPTHWISE_CONV_2D = 4, + /** + * Rearranges data from depth into blocks of spatial data. + * + * More specifically, this op outputs a copy of the input tensor where + * values from the depth dimension are moved in spatial blocks to the height + * and width dimensions. The value block_size indicates the input block size + * and how the data is moved. + * + * Chunks of data of size block_size * block_size from depth are rearranged + * into non-overlapping blocks of size block_size x block_size. + * + * The width of the output tensor is input_depth * block_size, whereas the + * height is input_height * block_size. The depth of the input tensor must + * be divisible by block_size * block_size + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Inputs: + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input. + * * 1: An {@link OperandType::INT32} scalar, specifying the block_size. + * block_size must be >=1 and block_size * block_size must be a divisor + * of the input depth. + * * 2: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output 4-D tensor, of shape [batch, height*block_size, + * width*block_size, depth/(block_size*block_size)]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + DEPTH_TO_SPACE = 5, + /** + * Dequantizes the input tensor. + * + * The formula is: + * + * output = (input - zeroPoint) * scale. + * + * Supported input tensor {@link OperandType}: + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_SYMM} (since HAL version 1.2) + * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} (since HAL version 1.2) + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported output tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32}. + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: A tensor. + * Since HAL version 1.2, this tensor may be zero-sized. + * + * Outputs: + * * 0: A tensor with the same shape as input0. + */ + DEQUANTIZE = 6, + /** + * Looks up sub-tensors in the input tensor. + * + * This operator takes for input a tensor of values (Values) and + * a one-dimensional tensor of selection indices (Lookups). + * The output tensor is the concatenation of sub-tensors of Values as + * selected by Lookups. + * + * Think of Values as being sliced along its first dimension: + * The entries in Lookups select which slices are concatenated together + * to create the output tensor. + * + * For example, if Values has shape of [40, 200, 300] and + * Lookups has shape of [3], all three values found in Lookups are + * expected to be between 0 and 39. The resulting tensor must + * have shape of [3, 200, 300]. + * + * If a value in Lookups is out of bounds, the operation must fail + * and an error must be reported. + * + * Supported value tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.3) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} (since HAL version 1.2) + * * {@link OperandType::TENSOR_QUANT8_ASYMM} (since HAL version 1.2) + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported value tensor rank: from 2 + * + * Inputs: + * * 0: Lookups. A 1-D tensor of {@link OperandType::TENSOR_INT32}. + * The values are indices into the first dimension of Values. + * * 1: Values. An n-D tensor, where n >= 2, from which sub-tensors are + * extracted. + * + * Output: + * * 0: A n-D tensor with the same rank and shape as the Values + * tensor, except for the first dimension which has the same size + * as Lookups' only dimension. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input1. + */ + EMBEDDING_LOOKUP = 7, + /** + * Computes element-wise floor() on the input tensor. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor, of the same {@link OperandType} and dimensions as + * the input tensor. + */ + FLOOR = 8, + /** + * Denotes a fully (densely) connected layer, which connects all elements + * in the input tensor with each element in the output tensor. + * + * This layer implements the operation: + * + * outputs = activation(inputs * weights’ + bias) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4. + * + * Inputs: + * * 0: A tensor of at least rank 2, specifying the input. If rank is + * greater than 2, then it gets flattened to a 2-D Tensor. The + * (flattened) 2-D Tensor is reshaped (if necessary) to + * [batch_size, input_size], where "input_size" corresponds to the + * number of inputs to the layer, matching the second dimension of + * weights, and "batch_size" is calculated by dividing the number of + * elements by "input_size". + * Since HAL version 1.2, zero batch_size is supported for this tensor. + * * 1: A 2-D tensor, specifying the weights, of shape + * [num_units, input_size], where "num_units" corresponds to the number + * of output nodes. + * * 2: A 1-D tensor, of shape [num_units], specifying the bias. For input + * tensor of {@link OperandType::TENSOR_FLOAT32}, the bias should + * also be of {@link OperandType::TENSOR_FLOAT32}. + * For input tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the bias should be of {@link OperandType::TENSOR_INT32}, + * with zeroPoint of 0 and bias_scale == input_scale * filter_scale. + * * 3: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * + * Outputs: + * * 0: The output tensor, of shape [batch_size, num_units]. Before HAL version 1.2, for + * output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM}, the following + * condition must be satisfied: output_scale > input_scale * filter_scale. + */ + FULLY_CONNECTED = 9, + /** + * Looks up sub-tensors in the input tensor using a key-value map. + * + * This operator takes for input a tensor of values (Values), + * a one-dimensional tensor of selection values (Lookups) and + * a one-dimensional tensor that maps these values to Values + * indexes. The output tensor is the concatenation of sub-tensors of + * Values as selected by Lookups via Keys. + * + * Think of Values as being sliced along its outer-most dimension. + * The output is a concatenation of selected slices, with one slice + * for each entry of Lookups. The slice selected is the one at the + * same index as the Maps entry that matches the value in Lookups. + * + * For a hit, the corresponding sub-tensor of Values is included + * in the Output tensor. For a miss, the corresponding sub-tensor in + * Output must have zero values. + * + * For example, if Values has shape of [40, 200, 300], + * Keys should have a shape of [40]. If Lookups tensor has shape + * of [3], three slices are being concatenated, so the resulting tensor + * must have the shape of [3, 200, 300]. If the first entry in Lookups + * has the value 123456, that value must be located in Keys tensor. + * If the sixth entry of Keys contains 123456, the sixth slice of Values + * must be selected. If no entry in Keys has 123456, a slice of zeroes + * must be concatenated. + * + * Supported value tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * + * Supported value tensor rank: from 2 + * + * Inputs: + * * 0: Lookups. A 1-D {@link OperandType::TENSOR_INT32} tensor with + * shape [ k ]. + * * 1: Keys. A 1-D {@link OperandType::TENSOR_INT32} tensor with shape + * [ n ]; Keys and Values pair represent a map, i.e., the ith element + * in Keys (Keys[i]) is the key to select the ith sub-tensor in Values + * (Values[i]), where 0 <= i <= n-1. Keys tensor *MUST* be sorted in + * ascending order. + * * 2: Values. A tensor with shape of [ n, … ]; i.e., the first dimension + * must be n. + * + * Outputs: + * * 0: Output. A tensor with shape [ k …]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, + * the scale and zeroPoint must be the same as input2. + * * 1: Hits. A boolean tensor with shape [ k ] indicates whether the lookup + * hits (True) or not (False). + * Stored as {@link OperandType::TENSOR_QUANT8_ASYMM} with offset 0 + * and scale 1.0f. + * A non-zero byte represents True, a hit. A zero indicates otherwise. + */ + HASHTABLE_LOOKUP = 10, + /** + * Applies L2 normalization along the axis dimension. + * + * The values in the output tensor are computed as: + * + * output[batch, row, col, channel] = + * input[batch, row, col, channel] / + * sqrt(sum_{c} pow(input[batch, row, col, c], 2)) + * + * By default the axis dimension is the last dimension of the input tensor. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} (since HAL version 1.2) + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * Tensors with rank less than 4 are only supported since HAL version 1.2. + * + * Inputs: + * * 0: An n-D tensor, specifying the tensor to be normalized. + * * 1: An optional {@link OperandType::INT32} scalar, default to -1, + * specifying the dimension normalization would be performed on. + * Negative index is used to specify axis from the end (e.g. -1 for + * the last axis). Must be in the range [-n, n). + * Available since HAL version 1.2. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} and same shape as input0. + * For {@link OperandType::TENSOR_QUANT8_ASYMM}, + * the scale must be 1.f / 128 and the zeroPoint must be 128. + * For {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the scale must be 1.f / 128 and the zeroPoint must be 0. + * + * NOTE: Before HAL version 1.3, if the elements along an axis are all zeros, + * the result is undefined. Since HAL version 1.3, if the elements along an axis + * are all zeros, the result is logical zero. + */ + L2_NORMALIZATION = 11, + /** + * Performs an 2-D L2 pooling operation. + * + * The output dimensions are functions of the filter dimensions, stride, and + * padding. + * + * The values in the output tensor are computed as: + * + * output[b, i, j, c] = + * sqrt(sum_{di, dj} pow(input[b, strides[1] * i + di, strides[2] * j + dj, c], 2) / + * sum(1)) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Both explicit padding and implicit padding are supported. + * + * Inputs (explicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: An {@link OperandType::INT32} scalar, specifying the padding on + * the left, in the ‘width’ dimension. + * * 2: An {@link OperandType::INT32} scalar, specifying the padding on + * the right, in the ‘width’ dimension. + * * 3: An {@link OperandType::INT32} scalar, specifying the padding on + * the top, in the ‘height’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the padding on + * the bottom, in the ‘height’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 7: An {@link OperandType::INT32} scalar, specifying the filter + * width. + * * 8: An {@link OperandType::INT32} scalar, specifying the filter + * height. + * * 9: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 10: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Inputs (implicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: An {@link OperandType::INT32} scalar, specifying the implicit + * padding scheme, has to be one of the + * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. + * * 2: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 3: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the filter + * width. + * * 5: An {@link OperandType::INT32} scalar, specifying the filter + * height. + * * 6: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 7: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, out_height, out_width, depth]. + */ + L2_POOL_2D = 12, + /** + * Applies Local Response Normalization along the depth dimension. + * + * The 4-D input tensor is treated as a 3-D array of 1-D vectors (along the + * last dimension), and each vector is normalized independently. Within a + * given vector, each component is divided by the weighted, squared sum of + * inputs within depth_radius. + * + * The output is calculated using this formula: + * + * sqr_sum[a, b, c, d] = sum( + * pow(input[a, b, c, d - depth_radius : d + depth_radius + 1], 2)) + * output = input / pow((bias + alpha * sqr_sum), beta) + * + * For input tensor with rank less than 4, independently normalizes each + * 1-D slice along specified dimension. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: up to 4 + * Tensors with rank less than 4 are only supported since HAL version 1.2. + * + * Inputs: + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. + * * 1: An {@link OperandType::INT32} scalar, specifying the radius of + * the normalization window. + * * 2: A scalar, specifying the bias, must not be zero. + * For input tensor of {@link OperandType::TENSOR_FLOAT16}, the bias + * value must be of {@link OperandType::FLOAT16}. + * For input tensor of {@link OperandType::TENSOR_FLOAT32}, the bias + * value must be of {@link OperandType::FLOAT32}. + * * 3: A scalar, specifying the scale factor, alpha. + * For input tensor of {@link OperandType::TENSOR_FLOAT16}, the + * alpha value must be of {@link OperandType::FLOAT16}. + * For input tensor of {@link OperandType::TENSOR_FLOAT32}, the + * alpha value must be of {@link OperandType::FLOAT32}. + * * 4: A scalar, specifying the exponent, beta. + * For input tensor of {@link OperandType::TENSOR_FLOAT16}, the beta + * value must be of {@link OperandType::FLOAT16}. + * For input tensor of {@link OperandType::TENSOR_FLOAT32}, the beta + * value must be of {@link OperandType::FLOAT32}. + * * 5: An optional {@link OperandType::INT32} scalar, default to -1, + * specifying the dimension normalization would be performed on. + * Negative index is used to specify axis from the end (e.g. -1 for + * the last axis). Must be in the range [-n, n). + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + LOCAL_RESPONSE_NORMALIZATION = 13, + /** + * Computes sigmoid activation on the input tensor element-wise. + * + * The output is calculated using this formula: + * + * output = 1 / (1 + exp(-input)) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4. + * + * Inputs: + * * 0: A tensor, specifying the input. + * Since HAL version 1.2, this tensor may be zero-sized. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + * For {@link OperandType::TENSOR_QUANT8_ASYMM}, + * the scale must be 1.f / 256 and the zeroPoint must be 0. + * For {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the scale must be 1.f / 256 and the zeroPoint must be -128. + */ + LOGISTIC = 14, + /** + * Projects an input to a bit vector via locality senstive hashing. + * + * Supported input tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * + * Supported input tensor rank: from 1 + * + * Inputs: + * * 0: Hash functions. Dim.size == 2, DataType: Float. + * Tensor[0].Dim[0]: 15 of hash functions. + * Tensor[0].Dim[1]: 16 of projected output bits generated by each + * hash function. + * If the projection type is Sparse: + * Tensor[0].Dim[1] + ceil(log2(Tensor[0].Dim[0])) <= 32 + * + * * 1: Input. Dim.size >= 1, no restriction on DataType. + * * 2: Weight. Optional. Dim.size == 1, DataType: Float. + * If not set, each input element is considered to have the same weight + * of 1.0. + * Tensor[1].Dim[0] == Tensor[2].Dim[0] + * * 3: Type: + * Sparse: + * Value LSHProjectionType_SPARSE(=3) (since HAL version 1.2). + * Computed bit vector is considered to be sparse. + * Each output element is an int32 made up of multiple bits + * computed from hash functions. + * + * NOTE: To avoid collisions across hash functions, an offset value + * of k * (1 << Tensor[0].Dim[1]) will be added to each signature, + * where k is the index of the hash function. + * + * Value LSHProjectionType_SPARSE_DEPRECATED(=1). + * Legacy behavior that does not include the offset value. + * + * Dense: + * Value LSHProjectionType_DENSE(=2). + * Computed bit vector is considered to be dense. Each output + * element represents a bit and can take the value of either + * 0 or 1. + * + * Outputs: + * * 0: If the projection type is Sparse: + * Output.Dim == { Tensor[0].Dim[0] } + * A tensor of int32 that represents hash signatures. + * + * If the projection type is Dense: + * Output.Dim == { Tensor[0].Dim[0] * Tensor[0].Dim[1] } + * A flattened tensor that represents projected bit vectors. + * The offset value for sparse projections was added in HAL version 1.2. + */ + LSH_PROJECTION = 15, + /** + * Performs a single time step in a Long Short-Term Memory (LSTM) layer + * + * The LSTM operation is described by the following equations. + * + * \f{eqnarray*}{ + * i_t =& \sigma(W_{xi}x_t+W_{hi}h_{t-1}+W_{ci}C_{t-1}+b_i) & \\ + * f_t =& \sigma(W_{xf}x_t+W_{hf}h_{t-1}+W_{cf}C_{t-1}+b_f) & \\ + * C_t =& clip(f_t \odot C_{t-1} + i_t \odot + * g(W_{xc}x_t+W_{hc}h_{t-1}+b_c),\ t_{cell}) & \\ + * o_t =& \sigma(W_{xo}x_t+W_{ho}h_{t-1}+W_{co}C_t+b_o) & \\ + * & & \\ + * & clip(W_{proj}(o_t \odot g(C_t))+b_{proj},\ t_{proj}) + * & if\ there\ is\ a\ projection; \\ + * h_t =& & \\ + * & o_t \odot g(C_t) & otherwise. \\ + * \f} + * Where: + * * \f$x_t\f$ is the input, + * * \f$i_t\f$ is the input gate, + * * \f$f_t\f$ is the forget gate, + * * \f$C_t\f$ is the cell state, + * * \f$o_t\f$ is the output, + * * \f$h_t\f$ is the output state, + * * \f$\sigma\f$ is the logistic sigmoid function, + * * \f$g\f$ is the cell input and cell output activation function, usually + * \f$tahn\f$, + * * \f$W_{xi}\f$ is the input-to-input weight matrix, + * * \f$W_{hi}\f$ is the recurrent to input weight matrix, + * * \f$W_{ci}\f$ is the cell-to-input weight matrix, + * * \f$b_i\f$ is the input gate bias, + * * \f$W_{xf}\f$ is the input-to-forget weight matrix, + * * \f$W_{hf}\f$ is the recurrent-to-forget weight matrix, + * * \f$W_{cf}\f$ is the cell-to-forget weight matrix, + * * \f$b_f\f$ is the forget gate bias, + * * \f$W_{xc}\f$ is the input-to-cell weight matrix, + * * \f$W_{hc}\f$ is the recurrent-to-cell weight matrix, + * * \f$b_c\f$ is the cell bias, + * * \f$W_{xo}\f$ is the input-to-output weight matrix, + * * \f$W_{ho}\f$ is the recurrent-to-output weight matrix, + * * \f$W_{co}\f$ is the cell-to-output weight matrix, + * * \f$b_o\f$ is the output gate bias, + * * \f$W_{proj}\f$ is the projection weight matrix, + * * \f$b_{proj}\f$ is the projection bias, + * * \f$t_{cell}\f$ is the threshold for clipping the cell state, and + * * \f$t_{proj}\f$ is the threshold for clipping the projected output. + * * \f$\odot\f$ is the + * + * Hadamard product that takes two matrices and produces another + * matrix, each element of which is the product of the corresponding + * elements of the input matrices. + * + * Since HAL version 1.2 LSTM supports layer normalization. + * In case layer normalization is used, the inputs to internal activation + * functions (sigmoid and \f$g\f$) are normalized, rescaled and recentered + * following an approach from section 3.1 from + * https://arxiv.org/pdf/1607.06450.pdf + * + * The operation has the following independently optional inputs: + * * The cell-to-input weights (\f$W_{ci}\f$), cell-to-forget weights + * (\f$W_{cf}\f$) and cell-to-output weights (\f$W_{co}\f$) either all + * have values or neither of them have values (i.e., all set to null). If + * they have values, the peephole optimization is used. + * * The input-to-input weights (\f$W_{xi}\f$), recurrent-to-input weights + * (\f$W_{hi}\f$) and input gate bias (\f$b_i\f$) either all have values, + * or none of them have values. If they have no values, coupling of input + * and forget gates (CIFG) is used, in which case the input gate + * (\f$i_t\f$) is calculated using the following equation instead. + * \f{eqnarray*}{ + * i_t = 1 - f_t + * \f} + * In case peephole optimization is used and CIFG is not used + * cell-to-input (\f$W_{ci}\f$) weights must be present. Otherwise, the + * cell-to-input weights must have no value. + * * The projection weights (\f$W_{proj}\f$) is required only for the + * recurrent projection layer, and should otherwise have no value. + * * The projection bias (\f$b_{proj}\f$) may (but not required to) have a + * value if the recurrent projection layer exists, and should otherwise + * have no value. + * * (HAL version 1.2 or later) The four layer normalization weights either all have + * values or none of them have values. Additionally, if CIFG is used, + * input layer normalization weights tensor is omitted and the other layer + * normalization weights either all have values or none of them have + * values. Layer normalization is used when the values of all the layer + * normalization weights are present. + * + * References: + * + * The default non-peephole non-CIFG implementation is based on: + * http://www.bioinf.jku.at/publications/older/2604.pdf + * S. Hochreiter and J. Schmidhuber. "Long Short-Term Memory". Neural + * Computation, 9(8):1735-1780, 1997. + * + * The peephole implementation and projection layer is based on: + * https://research.google.com/pubs/archive/43905.pdf + * Hasim Sak, Andrew Senior, and Francoise Beaufays. "Long short-term memory + * recurrent neural network architectures for large scale acoustic + * modeling." INTERSPEECH, 2014. + * (However, the concept of peephole optimization was introduced in work + * prior to this paper.) + * + * The coupling of input and forget gate (CIFG) is based on: + * http://arxiv.org/pdf/1503.04069.pdf + * Greff et al. "LSTM: A Search Space Odyssey" + * + * The layer normalization is based on: + * https://arxiv.org/pdf/1607.06450.pdf + * Jimmy Ba et al. "Layer Normalization" + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * + * All input and output tensors must be of the same type. + * + * Inputs: + * * 0: The input (\f$x_t\f$). + * A 2-D tensor of shape [batch_size, input_size], where “batch_size” + * corresponds to the batching dimension, and “input_size” is the size + * of the input. + * * 1: The input-to-input weights (\f$W_{xi}\f$). Optional. + * A 2-D tensor of shape [num_units, input_size], where “num_units” + * corresponds to the number of cell units. + * * 2: The input-to-forget weights (\f$W_{xf}\f$). + * A 2-D tensor of shape [num_units, input_size]. + * * 3: The input-to-cell weights (\f$W_{xc}\f$). + * A 2-D tensor of shape [num_units, input_size]. + * * 4: The input-to-output weights (\f$W_{xo}\f$). + * A 2-D tensor of shape [num_units, input_size]. + * * 5: The recurrent-to-input weights (\f$W_{hi}\f$). Optional. + * A 2-D tensor of shape [num_units, output_size], where “output_size” + * corresponds to either the number of cell units (i.e., “num_units”), + * or the second dimension of the “projection_weights”, if defined. + * * 6: The recurrent-to-forget weights (\f$W_{hf}\f$). + * A 2-D tensor of shape [num_units, output_size]. + * * 7: The recurrent-to-cell weights (\f$W_{hc}\f$). + * A 2-D tensor of shape [num_units, output_size]. + * * 8: The recurrent-to-output weights (\f$W_{ho}\f$). + * A 2-D tensor of shape [num_units, output_size]. + * * 9: The cell-to-input weights (\f$W_{ci}\f$). Optional. + * A 1-D tensor of shape [num_units]. + * * 10:The cell-to-forget weights (\f$W_{cf}\f$). Optional. + * A 1-D tensor of shape [num_units]. + * * 11:The cell-to-output weights (\f$W_{co}\f$). Optional. + * A 1-D tensor of shape [num_units]. + * * 12:The input gate bias (\f$b_i\f$). Optional. + * A 1-D tensor of shape [num_units]. + * * 13:The forget gate bias (\f$b_f\f$). + * A 1-D tensor of shape [num_units]. + * * 14:The cell bias (\f$b_c\f$). + * A 1-D tensor of shape [num_units]. + * * 15:The output gate bias (\f$b_o\f$). + * A 1-D tensor of shape [num_units]. + * * 16:The projection weights (\f$W_{proj}\f$). Optional. + * A 2-D tensor of shape [output_size, num_units]. + * * 17:The projection bias (\f$b_{proj}\f$). Optional. + * A 1-D tensor of shape [output_size]. + * * 18:The output state (in) (\f$h_{t-1}\f$). + * A 2-D tensor of shape [batch_size, output_size]. + * * 19:The cell state (in) (\f$C_{t-1}\f$). + * A 2-D tensor of shape [batch_size, num_units]. + * * 20:The activation function (\f$g\f$). + * A value indicating the activation function: + * + * * 21:The clipping threshold (\f$t_{cell}\f$) for the cell state, such + * that values are bound within [-cell_clip, cell_clip]. If set to 0.0 + * then clipping is disabled. + * Until HAL version 1.2 this scalar must be of type {@link + * OperandType::FLOAT32}. Since HAL version 1.2, if all the input + * tensors have type {@link OperandType::TENSOR_FLOAT32}, this + * scalar must be of the type {@link OperandType::FLOAT32}, + * otherwise if all the input tensors have the type {@link + * OperandType::TENSOR_FLOAT16}, this scalar must be of type {@link + * OperandType::FLOAT16}. + * * 22:The clipping threshold (\f$t_{proj}\f$) for the output from the + * projection layer, such that values are bound within + * [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. + * Until HAL version 1.2 this scalar must be of type {@link + * OperandType::FLOAT32}. Since HAL version 1.2, if all the input + * tensors have type {@link OperandType::TENSOR_FLOAT32}, this + * scalar must be of the type {@link OperandType::FLOAT32}, + * otherwise if all the input tensors have the type {@link + * OperandType::TENSOR_FLOAT16}, this scalar must be of type {@link + * OperandType::FLOAT16}. + * Since HAL version 1.2 there are additional inputs to this op: + * * 23:The input layer normalization weights. + * A 1-D tensor of shape [num_units]. Used to rescale normalized inputs + * to activation at input gate. + * * 24:The forget layer normalization weights. + * A 1-D tensor of shape [num_units]. Used to rescale normalized inputs + * to activation at forget gate. + * * 25:The cell layer normalization weights. + * A 1-D tensor of shape [num_units]. Used to rescale normalized inputs + * to activation at cell gate. + * * 26:The output layer normalization weights. + * A 1-D tensor of shape [num_units]. Used to rescale normalized inputs + * to activation at output gate. + * + * Outputs: + * * 0: The scratch buffer. + * A 2-D tensor of shape [batch_size, num_units * 3] with CIFG, or + * [batch_size, num_units * 4] without CIFG. + * * 1: The output state (out) (\f$h_t\f$). + * A 2-D tensor of shape [batch_size, output_size]. + * * 2: The cell state (out) (\f$C_t\f$). + * A 2-D tensor of shape [batch_size, num_units]. + * * 3: The output (\f$o_t\f$). + * A 2-D tensor of shape [batch_size, output_size]. This is effectively + * the same as the current “output state (out)” value. + */ + LSTM = 16, + /** + * Performs an 2-D max pooling operation. + * + * The output dimensions are functions of the filter dimensions, stride, and + * padding. + * + * The values in the output tensor are computed as: + * + * output[b, i, j, channel] = + * max_{di, dj} ( + * input[b, strides[1] * i + di, strides[2] * j + dj, channel] + * ) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Both explicit padding and implicit padding are supported. + * + * Inputs (explicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: An {@link OperandType::INT32} scalar, specifying the padding on + * the left, in the ‘width’ dimension. + * * 2: An {@link OperandType::INT32} scalar, specifying the padding on + * the right, in the ‘width’ dimension. + * * 3: An {@link OperandType::INT32} scalar, specifying the padding on + * the top, in the ‘height’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the padding on + * the bottom, in the ‘height’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 7: An {@link OperandType::INT32} scalar, specifying the filter + * width. + * * 8: An {@link OperandType::INT32} scalar, specifying the filter + * height. + * * 9: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 10: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Inputs (implicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: An {@link OperandType::INT32} scalar, specifying the implicit + * padding scheme, has to be one of the + * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. + * * 2: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 3: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the filter + * width. + * * 5: An {@link OperandType::INT32} scalar, specifying the filter + * height. + * * 6: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 7: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, out_height, out_width, depth]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + MAX_POOL_2D = 17, + /** + * Multiplies two tensors, element-wise. + * + * Takes two input tensors of identical {@link OperandType} and compatible + * dimensions. The output is the product of both input tensors, optionally + * modified by an activation function. + * + * Two dimensions are compatible when: + * 1. they are equal, or + * 2. one of them is 1 + * + * The size of the resulting output is the maximum size along each dimension + * of the input operands. It starts with the trailing dimensions, and works + * its way forward. + * + * Since HAL version 1.2, generic zero-sized input tensor is supported. Zero + * dimension is only compatible with 0 or 1. The size of the output + * dimension is zero if either of corresponding input dimension is zero. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * * {@link OperandType::TENSOR_INT32} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType}, and compatible dimensions + * as input0. + * * 2: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * For a {@link OperandType::TENSOR_INT32} tensor, + * the {@link FusedActivationFunc} must be "NONE". + * + * Outputs: + * * 0: The product, a tensor of the same {@link OperandType} as input0. + * For output tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the following condition must be satisfied: + * output_scale > input1_scale * input2_scale. + */ + MUL = 18, + /** + * Computes rectified linear activation on the input tensor element-wise. + * + * The output is calculated using this formula: + * + * output = max(0, input) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4. + * + * Inputs: + * * 0: A tensor, specifying the input. + * Since HAL version 1.2, this tensor may be zero-sized. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + RELU = 19, + /** + * Computes rectified linear 1 activation on the input tensor element-wise. + * + * The output is calculated using this formula: + * + * output = min(1.f, max(-1.f, input)) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4. + * + * Inputs: + * * 0: A tensor, specifying the input. + * Since HAL version 1.2, this tensor may be zero-sized. + * + * Outputs: + * * 0: The output tensor of the same shape as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + RELU1 = 20, + /** + * Computes rectified linear 6 activation on the input tensor element-wise. + * + * The output is calculated using this formula: + * + * output = min(6, max(0, input)) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4. + * + * Inputs: + * * 0: A tensor, specifying the input. + * Since HAL version 1.2, this tensor may be zero-sized. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + RELU6 = 21, + /** + * Reshapes a tensor. + * + * Given tensor, this operation returns a tensor that has the same values as + * tensor, but with a newly specified shape. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4. + * + * Inputs: + * * 0: A tensor, specifying the tensor to be reshaped. + * * 1: A 1-D tensor of {@link OperandType::TENSOR_INT32}, defining the + * shape of the output tensor. The number of elements implied by shape + * must be the same as the number of elements in the input tensor. + * + * If one component of shape is the special value -1, the size of that + * dimension is computed so that the total size remains constant. In + * particular, a shape of [-1] flattens into 1-D. At most one component + * of shape can be -1. + * + * Outputs: + * * 0: The output tensor, of shape specified by the input shape. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + RESHAPE = 22, + /** + * Resizes images to given size using the bilinear interpretation. + * + * Resized images must be distorted if their output aspect ratio is not the + * same as input aspect ratio. The corner pixels of output may not be the + * same as corner pixels of input. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} (since HAL version 1.2) + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Both resizing by shape and resizing by scale are supported. + * + * Inputs (resizing by shape): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. + * Since HAL version 1.2, zero batches is supported for this tensor. + * * 1: An {@link OperandType::INT32} scalar, specifying the output + * width of the output tensor. + * * 2: An {@link OperandType::INT32} scalar, specifying the output + * height of the output tensor. + * * 3: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * * 4: Align corners. An optional {@link OperandType::BOOL} + * scalar, default to false. If True, the centers of the 4 corner + * pixels of the input and output tensors are aligned, preserving the + * values at the corner pixels. + * Available since HAL version 1.3. + * * 5: Half pixel centers. An optional {@link OperandType::BOOL} + * scalar, default to false. If True, the pixel centers are assumed to + * be at (0.5, 0.5). This is the default behavior of image.resize in + * TF 2.0. If this parameter is True, then align_corners parameter + * must be False. + * Available since HAL version 1.3. + * + * Inputs (resizing by scale, since HAL version 1.2): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. Zero batches is supported for this tensor. + * * 1: A scalar, specifying width_scale, the scaling factor of the width + * dimension from the input tensor to the output tensor. The output + * width is calculated as new_width = floor(width * width_scale). + * The scalar must be of {@link OperandType::FLOAT16} if input0 is + * of {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} otherwise. + * * 2: A scalar, specifying height_scale, the scaling factor of the height + * dimension from the input tensor to the output tensor. The output + * height is calculated as new_height = floor(height * height_scale). + * The scalar must be of {@link OperandType::FLOAT16} if input0 is + * of {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} otherwise. + * * 3: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * * 4: Align corners. An optional {@link OperandType::BOOL} + * scalar, default to false. If True, the centers of the 4 corner + * pixels of the input and output tensors are aligned, preserving the + * values at the corner pixels. + * Available since HAL version 1.3. + * * 5: Half pixel centers. An optional {@link OperandType::BOOL} + * scalar, default to false. If True, the pixel centers are assumed to + * be at (0.5, 0.5). This is the default behavior of image.resize in + * TF 2.0. If this parameter is True, then align_corners parameter + * must be False. + * Available since HAL version 1.3. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, new_height, new_width, depth]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + RESIZE_BILINEAR = 23, + /** + * A basic recurrent neural network layer. + * + * This layer implements the operation: + * outputs = state = activation(inputs * input_weights + + * state * recurrent_weights + bias) + * + * Where: + * * “input_weights” is a weight matrix that multiplies the inputs; + * * “recurrent_weights” is a weight matrix that multiplies the current + * “state” which itself is the output from the previous time step + * computation; + * * “bias” is a bias vector (added to each output vector in the batch); + * * “activation” is the function passed as the “fused_activation_function” + * argument (if not “NONE”). + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * + * The input tensors must all be the same type. + * + * Inputs: + * * 0: input. + * A 2-D tensor of shape [batch_size, input_size], where “batch_size” + * corresponds to the batching dimension, and “input_size” is the size + * of the input. + * * 1: weights. + * A 2-D tensor of shape [num_units, input_size], where “num_units” + * corresponds to the number of units. + * * 2: recurrent_weights. + * A 2-D tensor of shape [num_units, num_units], with columns + * corresponding to the weights from each unit. + * * 3: bias. + * A 1-D tensor of shape [num_units]. + * * 4: hidden state (in). + * A 2-D tensor of shape [batch_size, num_units]. + * * 5: fused_activation_function. + * An optional {@link FusedActivationFunc} value indicating the + * activation function. If “NONE” is specified then it results in a + * linear activation. + * + * Outputs: + * * 0: hidden state (out). + * A 2-D tensor of shape [batch_size, num_units]. + * + * * 1: output. + * A 2-D tensor of shape [batch_size, num_units]. This is effectively + * the same as the current state value. + */ + RNN = 24, + /** + * Computes the softmax activation on the input tensor element-wise, per + * batch, by normalizing the input vector so the maximum coefficient is + * zero. + * + * The output is calculated using this formula: + * + * output[batch, i] = + * exp((input[batch, i] - max(input[batch, :])) * beta) / + * sum_{k}{exp((input[batch, k] - max(input[batch, :])) * beta)} + * + * For input tensor with rank other than 2, the activation will be applied + * independently on each 1-D slice along specified dimension. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4. + * Tensors with rank other than 2 or 4 are only supported since HAL version 1.2. + * + * Inputs: + * * 0: A 2-D or 4-D tensor, specifying the tensor to be reshaped. + * Since HAL version 1.2, this tensor may be zero-sized. + * * 1: A scalar, specifying the positive scaling factor for the exponent, + * beta. If input0 is of {@link OperandType::TENSOR_FLOAT32}, + * {@link OperandType::TENSOR_QUANT8_ASYMM} or + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, the scalar + * must be of {@link OperandType::FLOAT32}. + * If input0 is of {@link OperandType::TENSOR_FLOAT16}, then the + * scalar must be of {@link OperandType::FLOAT16}. + * * 2: An optional {@link OperandType::INT32} scalar, default to -1, + * specifying the dimension the activation would be performed on. + * Negative index is used to specify axis from the end (e.g. -1 for + * the last axis). Must be in the range [-n, n). + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + * For {@link OperandType::TENSOR_QUANT8_ASYMM}, + * the scale must be 1.f / 256 and the zeroPoint must be 0. + * For {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the scale must be 1.f / 256 and the zeroPoint must be -128. + */ + SOFTMAX = 25, + /** + * Rearranges blocks of spatial data, into depth. + * + * More specifically, this op outputs a copy of the input tensor where + * values from the height and width dimensions are moved to the depth + * dimension. The value block_size indicates the input block size and how + * the data is moved. + * + * Chunks of data of size block_size * block_size from depth are rearranged + * into non-overlapping blocks of size block_size x block_size. + * + * The depth of the output tensor is input_depth * block_size * block_size. + * The input tensor's height and width must be divisible by block_size. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Inputs: + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input. + * * 1: An {@link OperandType::INT32} scalar, specifying the block_size. + * block_size must be >=1 and block_size must be a divisor of both the + * input height and width. + * * 2: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Outputs: + * * 0: The output 4-D tensor, of shape [batches, height/block_size, + * width/block_size, depth_in*block_size*block_size]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + SPACE_TO_DEPTH = 26, + /** + * SVDF op is a kind of stateful layer derived from the notion that a + * densely connected layer that's processing a sequence of input frames can + * be approximated by using a singular value decomposition of each of its + * nodes. The implementation is based on: + * + * https://research.google.com/pubs/archive/43813.pdf + * + * P. Nakkiran, R. Alvarez, R. Prabhavalkar, C. Parada. + * “Compressing Deep Neural Networks using a Rank-Constrained Topology”. + * INTERSPEECH, 2015. + * + * It processes the incoming input using a 2-stage filtering mechanism: + * * stage 1 performs filtering on the "features" dimension, whose outputs + * get pushed into a memory of fixed-size memory_size. + * * stage 2 performs filtering on the "time" dimension of the memory_size + * memoized outputs of stage 1. + * + * Specifically, for rank 1, this layer implements the operation: + * + * memory = push(conv1d(inputs, weights_feature, feature_dim, + * "PADDING_VALID")); + * outputs = activation(memory * weights_time + bias); + * + * Where: + * * “weights_feature” is a weights matrix that processes the inputs (by + * convolving the input with every “feature filter”), and whose outputs + * get pushed, stacked in order, into the fixed-size “memory” (the oldest + * entry gets dropped); + * * “weights_time” is a weights matrix that processes the “memory” (by a + * batched matrix multiplication on the num_units); + * * “bias” is an optional bias vector (added to each output vector in the + * batch); and + * * “activation” is the function passed as the “fused_activation_function” + * argument (if not “NONE”). + * + * Each rank adds a dimension to the weights matrices by means of stacking + * the filters. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * + * All input tensors must be the same type. + * + * Inputs: + * * 0: input. + * A 2-D tensor of shape [batch_size, input_size], where “batch_size” + * corresponds to the batching dimension, and “input_size” is the size + * of the input. + * * 1: weights_feature. + * A 2-D tensor of shape [num_units, input_size], where “num_units” + * corresponds to the number of units. + * * 2: weights_time. + * A 2-D tensor of shape [num_units, memory_size], where “memory_size” + * corresponds to the fixed-size of the memory. + * * 3: bias. + * An optional 1-D tensor of shape [num_units]. + * * 4: state (in). + * A 2-D tensor of shape [batch_size, (memory_size - 1) * num_units * rank]. + * * 5: rank. + * The rank of the SVD approximation. + * * 6: fused_activation_function. + * An optional {@link FusedActivationFunc} value indicating the + * activation function. If “NONE” is specified then it results in a + * linear activation. + * + * Outputs: + * * 0: state (out). + * A 2-D tensor of the same {@link OperandType} as the inputs, with shape + * [batch_size, (memory_size - 1) * num_units * rank]. + * * 1: output. + * A 2-D tensor of the same {@link OperandType} as the inputs, with shape + * [batch_size, num_units]. + */ + SVDF = 27, + /** + * Computes hyperbolic tangent of input tensor element-wise. + * + * The output is calculated using this formula: + * + * output = tanh(input) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} (since HAL version 1.2) + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4. + * + * Inputs: + * * 0: A tensor, specifying the input. + * Since HAL version 1.2, this tensor may be zero-sized. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + * For {@link OperandType::TENSOR_QUANT8_ASYMM}, + * the scale must be 1.f / 128 and the zeroPoint must be 128. + * For {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the scale must be 1.f / 128 and the zeroPoint must be 0. + */ + TANH = 28, + /** + * BatchToSpace for N-dimensional tensors. + * + * This operation reshapes the batch dimension (dimension 0) into M + 1 + * dimensions of shape block_shape + [batch], interleaves these blocks back + * into the grid defined by the spatial dimensions [1, ..., M], to obtain a + * result with the same rank as the input. + * + * This is the reverse of SpaceToBatch. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Inputs: + * * 0: An n-D tensor, specifying the tensor to be reshaped + * * 1: A 1-D Tensor of {@link OperandType::TENSOR_INT32}, the block + * sizes for each spatial dimension of the input tensor. All values + * must be >= 1. + * * 2: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since API level 29. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + BATCH_TO_SPACE_ND = 29, + /** + * Element-wise division of two tensors. + * + * Takes two input tensors of identical {@link OperandType} and compatible + * dimensions. The output is the result of dividing the first input tensor + * by the second, optionally modified by an activation function. + * + * For inputs of {@link OperandType::TENSOR_INT32}, performs + * "floor division" ("//" in Python). For example, + * 5 // 2 = 2 + * -5 // 2 = -3 + * + * Two dimensions are compatible when: + * 1. they are equal, or + * 2. one of them is 1 + * + * The size of the output is the maximum size along each dimension of the + * input operands. It starts with the trailing dimensions, and works its way + * forward. + * + * Example: + * input1.dimension = {4, 1, 2} + * input2.dimension = {5, 4, 3, 1} + * output.dimension = {5, 4, 3, 2} + * + * Since HAL version 1.2, generic zero-sized input tensor is supported. Zero + * dimension is only compatible with 0 or 1. The size of the output + * dimension is zero if either of corresponding input dimension is zero. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor, specifying the first input. + * * 1: A tensor of the same {@link OperandType}, and compatible dimensions + * as input0. + * * 2: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * For a {@link OperandType::TENSOR_INT32} tensor, + * the {@link FusedActivationFunc} must be "NONE". + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + */ + DIV = 30, + /** + * Computes the mean of elements across dimensions of a tensor. + * + * Reduces the input tensor along the given dimensions to reduce. Unless + * keep_dims is true, the rank of the tensor is reduced by 1 for each entry + * in axis. If keep_dims is true, the reduced dimensions are retained with + * length 1. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: A tensor, specifying the input. + * * 1: A 1-D Tensor of {@link OperandType::TENSOR_INT32}. The dimensions + * to reduce. Must be in the range + * [-rank(input_tensor), rank(input_tensor)). + * + * NOTE: When the operation was introduced, the documentation + * incorrectly stated that if dimensions were empty, the operation + * would reduce across all dimensions. This behavior was never + * implemented. + * + * * 2: An {@link OperandType::INT32} scalar, keep_dims. If positive, + * retains reduced dimensions with length 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + * If all dimensions are reduced and keep_dims is false, the output + * shape is [1]. + */ + MEAN = 31, + /** + * Pads a tensor. + * + * This operation pads a tensor according to the specified paddings. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * (full support since HAL version 1.2, see the output section) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor, specifying the tensor to be padded. + * * 1: A 2-D Tensor of {@link OperandType::TENSOR_INT32}, the paddings + * for each spatial dimension of the input tensor. The shape of the + * tensor must be {rank(input0), 2}. + * padding[i, 0] specifies the number of elements to be padded in the + * front of dimension i. + * padding[i, 1] specifies the number of elements to be padded after the + * end of dimension i. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. The + * output tensor has the same rank as input0, and each + * dimension of the output tensor has the same size as the + * corresponding dimension of the input tensor plus the size + * of the padding: + * output0.dimension[i] = + * padding[i, 0] + input0.dimension[i] + padding[i, 1] + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + * + * NOTE: Before HAL version 1.2, the pad value for + * {@link OperandType::TENSOR_QUANT8_ASYMM} is undefined. + * Since HAL version 1.2, the pad value is always the logical zero. + */ + PAD = 32, + /** + * SpaceToBatch for N-Dimensional tensors. + * + * This operation divides "spatial" dimensions [1, ..., M] of the input into + * a grid of blocks of shape block_shape, and interleaves these blocks with + * the "batch" dimension (0) such that in the output, the spatial dimensions + * [1, ..., M] correspond to the position within the grid, and the batch + * dimension combines both the position within a spatial block and the + * original batch position. Prior to division into blocks, the spatial + * dimensions of the input are optionally zero padded according to paddings. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * (full support since HAL version 1.2, see the output section) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * NCHW is supported since HAL version 1.2. + * + * Inputs: + * * 0: An n-D tensor, specifying the input. + * * 1: A 1-D Tensor of {@link OperandType::TENSOR_INT32}, the block + * sizes for each spatial dimension of the input tensor. All values + * must be >= 1. + * * 2: A 2-D Tensor of {@link OperandType::TENSOR_INT32}, the paddings + * for each spatial dimension of the input tensor. All values must be + * >= 0. The shape of the tensor must be {M, 2}, where M is the number + * of spatial dimensions. + * padding[i, 0] specifies the number of element to be padded in the + * front of dimension i. + * padding[i, 1] specifies the number of element to be padded after the + * end of dimension i. + * * 3: An optional {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * Available since HAL version 1.2. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + * + * NOTE: Before HAL version 1.2, the pad value for + * {@link OperandType::TENSOR_QUANT8_ASYMM} is undefined. + * Since HAL version 1.2, the pad value is always the logical zero. + */ + SPACE_TO_BATCH_ND = 33, + /** + * Removes dimensions of size 1 from the shape of a tensor. + * + * Given a tensor input, this operation returns a tensor of the same + * {@link OperandType} with all dimensions of size 1 removed. If you don't + * want to remove all size 1 dimensions, you can remove specific size 1 + * dimensions by specifying the axes (input1). + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor, the tensor to be squeezed. + * * 1: An optional 1-D tensor of {@link OperandType::TENSOR_INT32}. The + * dimensions to squeeze. If specified only squeezes the dimensions + * listed. Otherwise, squeezes all dimensions. The dimension index + * starts at 0. An error must be reported if squeezing a dimension that + * is not 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. Contains the + * same data as input, but has one or more dimensions of size 1 + * removed. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + * If all input dimensions are equal to 1 and are to be squeezed, the + * output shape is [1]. + */ + SQUEEZE = 34, + /** + * Extracts a strided slice of a tensor. + * + * Roughly speaking, this op extracts a slice of size (end - begin) / stride + * from the given input tensor. Starting at the location specified by begin + * the slice continues by adding stride to the index until all dimensions + * are not less than end. Note that a stride can be negative, which causes a + * reverse slice. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor, specifying the tensor to be sliced. + * * 1: begin, a 1-D tensor of {@link OperandType::TENSOR_INT32}. The + * starts of the dimensions of the input tensor to be sliced. The + * length must be of rank(input0). + * * 2: end, a 1-D tensor of {@link OperandType::TENSOR_INT32}. The + * ends of the dimensions of the input tensor to be sliced. The length + * must be of rank(input0). + * * 3: strides, a 1-D tensor of {@link OperandType::TENSOR_INT32}. The + * strides of the dimensions of the input tensor to be sliced. The + * length must be of rank(input0). The entries must be non-zero. + * * 4: begin_mask, an {@link OperandType::INT32} scalar. If the ith bit + * of begin_mask is set, begin[i] is ignored and the fullest possible + * range in that dimension is used instead. + * * 5: end_mask, an {@link OperandType::INT32} scalar. If the ith bit of + * end_mask is set, end[i] is ignored and the fullest possible range in + * that dimension is used instead. + * * 6: shrink_axis_mask, an {@link OperandType::INT32} scalar. If the + * ith bit of shrink_axis_mask is set, the ith dimension specification + * shrinks the dimensionality by 1, taking on the value at index + * begin[i]. In this case, the ith specification must define a + * slice of size 1, e.g. begin[i] = x, end[i] = x + 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0 and rank (n - k), + * where k is the number of bits set in shrink_axis_mask. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + * If shrink_axis_mask is true for all input dimensions, the output + * shape is [1]. + */ + STRIDED_SLICE = 35, + /** + * Element-wise subtraction of two tensors. + * + * Takes two input tensors of identical {@link OperandType} and compatible + * dimensions. The output is the result of subtracting the second input + * tensor from the first one, optionally modified by an activation function. + * + * Two dimensions are compatible when: + * 1. they are equal, or + * 2. one of them is 1 + * + * The size of the output is the maximum size along each dimension of the + * input operands. It starts with the trailing dimensions, and works its way + * forward. + * + * Example: + * input1.dimension = {4, 1, 2} + * input2.dimension = {5, 4, 3, 1} + * output.dimension = {5, 4, 3, 2} + * + * Since HAL version 1.2, generic zero-sized input tensor is supported. Zero + * dimension is only compatible with 0 or 1. The size of the output + * dimension is zero if either of corresponding input dimension is zero. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} (since HAL version 1.2) + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * * {@link OperandType::TENSOR_INT32} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor, specifying the first input. + * * 1: A tensor of the same {@link OperandType}, and compatible dimensions + * as input0. + * * 2: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * For a {@link OperandType::TENSOR_INT32} tensor, + * the {@link FusedActivationFunc} must be "NONE". + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint can be different from inputs' scale and zeroPoint. + */ + SUB = 36, + /** + * Transposes the input tensor, permuting the dimensions according to the + * perm tensor. + * + * The returned tensor's dimension i corresponds to the input dimension + * perm[i]. If perm is not given, it is set to (n-1...0), where n is the + * rank of the input tensor. Hence by default, this operation performs a + * regular matrix transpose on 2-D input Tensors. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} (since HAL version 1.2) + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor, specifying the tensor to be transposed. + * Since HAL version 1.2, this tensor may be zero-sized. + * * 1: An optional 1-D Tensor of {@link OperandType::TENSOR_INT32}, + * the permutation of the dimensions of the input tensor. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + TRANSPOSE = 37, + /** + * Computes the absolute value of a tensor, element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} (since HAL version 1.3) + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + ABS = 38, + /** + * Returns the index of the largest element along an axis. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: An n-D tensor specifying the input. Must be non-empty. + * * 1: An {@link OperandType::INT32} scalar specifying the axis to + * reduce across. Negative index is used to specify axis from the + * end (e.g. -1 for the last axis). Must be in the range [-n, n). + * + * Outputs: + * * 0: An (n - 1)-D {@link OperandType::TENSOR_INT32} tensor. + * If input is 1-dimensional, the output shape is [1]. + */ + ARGMAX = 39, + /** + * Returns the index of the smallest element along an axis. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: An n-D tensor specifying the input. Must be non-empty. + * * 1: An {@link OperandType::INT32} scalar specifying the axis to + * reduce across. Negative index is used to specify axis from the + * end (e.g. -1 for the last axis). Must be in the range [-n, n). + * + * Outputs: + * * 0: An (n - 1)-D {@link OperandType::TENSOR_INT32} tensor. + * If input is 1-dimensional, the output shape is [1]. + */ + ARGMIN = 40, + /** + * Transform axis-aligned bounding box proposals using bounding box deltas. + * + * Given the positions of bounding box proposals and the corresponding + * bounding box deltas for each class, return the refined bounding box + * regions. The resulting bounding boxes are cliped against the edges of + * the image. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT16_ASYMM} + * + * Inputs: + * * 0: A 2-D Tensor of shape [num_rois, 4], specifying the locations of the + * bounding box proposals, each line with format [x1, y1, x2, y2]. + * For tensor of type {@link OperandType::TENSOR_QUANT16_ASYMM}, + * the zeroPoint must be 0 and the scale must be 0.125. Zero num_rois + * is supported for this tensor. + * * 1: A 2-D Tensor of shape [num_rois, num_classes * 4], specifying the + * bounding box delta for each region of interest and each class. The + * bounding box deltas are organized in the following order + * [dx, dy, dw, dh], where dx and dy is the relative correction factor + * for the center position of the bounding box with respect to the width + * and height, dw and dh is the log-scale relative correction factor + * for the width and height. For input0 of type + * {@link OperandType::TENSOR_QUANT16_ASYMM}, this tensor should be + * of {@link OperandType::TENSOR_QUANT8_ASYMM} or + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}. Zero num_rois is + * supported for this tensor. + * * 2: An 1-D {@link OperandType::TENSOR_INT32} tensor, of shape + * [num_rois], specifying the batch index of each box. Boxes with + * the same batch index are grouped together. Zero num_rois is + * supported for this tensor. + * * 3: A 2-D Tensor of shape [batches, 2], specifying the information of + * each image in the batch, each line with format + * [image_height, image_width]. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0, with shape + * [num_rois, num_classes * 4], specifying the coordinates of each + * output bounding box for each class, with format [x1, y1, x2, y2]. + * For type of {@link OperandType::TENSOR_QUANT16_ASYMM}, the + * scale must be 0.125 and the zero point must be 0. + */ + AXIS_ALIGNED_BBOX_TRANSFORM = 41, + /** + * A recurrent neural network layer that applies an LSTM cell to a + * sequence of inputs in forward and backward directions. + * + * The op supports cross-linking via an auxiliary input. Regular cell feeds + * one input into the two RNN cells in the following way: + * + * INPUT (INPUT_REVERSED) + * | | + * --------------------- + * | FW_LSTM BW_LSTM | + * --------------------- + * | | + * FW_OUT BW_OUT + * + * An op with cross-linking takes two inputs and feeds them into the RNN + * cells in the following way: + * + * AUX_INPUT (AUX_INPUT_REVERSED) + * | | + * INPUT | (INPUT_R'D.)| + * | | | | + * ----------------------- + * | \ / \ / | + * | FW_LSTM BW_LSTM | + * ----------------------- + * | | + * FW_OUT BW_OUT + * + * The cross-linking mode is enabled iff auxiliary input and auxiliary + * weights are present. While stacking this op on top of itself, this + * allows to connect both forward and backward outputs from previous cell + * to the next cell's input. + * + * Since HAL version 1.3 parallel linking mode is supported. The mode is + * enabled if auxiliary input is present but auxiliary weights are omitted. + * In this case, the cell feeds inputs into the RNN in the following way: + * + * INPUT (AUX_INPUT_REVERSED) + * | | + * --------------------- + * | FW_LSTM BW_LSTM | + * --------------------- + * | | + * FW_OUT BW_OUT + * + * While stacking this op on top of itself, this allows to connect both + * forward and backward outputs from previous cell to the next cell's + * corresponding inputs. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: 3, either time-major or batch-major. + * + * All input and output tensors must be of the same type. + * + * Inputs: + * * 0: The input. + * A 3-D tensor of shape: + * If time-major: [max_time, batch_size, input_size] + * If batch-major: [batch_size, max_time, input_size] + * where "max_time" is the number of timesteps (sequence length), + * "batch_size" corresponds to the batching dimension, and + * "input_size" is the size of the input. + * * 1: The forward input-to-input weights. Optional. + * A 2-D tensor of shape [fw_num_units, input_size], where “fw_num_units” + * corresponds to the number of forward cell units. + * * 2: The forward input-to-forget weights. + * A 2-D tensor of shape [fw_num_units, input_size]. + * * 3: The forward input-to-cell weights. + * A 2-D tensor of shape [fw_num_units, input_size]. + * * 4: The forward input-to-output weights. + * A 2-D tensor of shape [fw_num_units, input_size]. + * * 5: The forward recurrent-to-input weights. Optional. + * A 2-D tensor of shape [fw_num_units, fw_output_size], where “fw_output_size” + * corresponds to either the number of cell units (i.e., fw_num_units), + * or the second dimension of the “fw_projection_weights”, if defined. + * * 6: The forward recurrent-to-forget weights. + * A 2-D tensor of shape [fw_num_units, fw_output_size]. + * * 7: The forward recurrent-to-cell weights. + * A 2-D tensor of shape [fw_num_units, fw_output_size]. + * * 8: The forward recurrent-to-output weights. + * A 2-D tensor of shape [fw_num_units, fw_output_size]. + * * 9: The forward cell-to-input weights. Optional. + * A 1-D tensor of shape [fw_num_units]. + * * 10: The forward cell-to-forget weights. Optional. + * A 1-D tensor of shape [fw_num_units]. + * * 11: The forward cell-to-output weights. Optional. + * A 1-D tensor of shape [fw_num_units]. + * * 12: The forward input gate bias. Optional. + * A 1-D tensor of shape [fw_num_units]. + * * 13: The forward forget gate bias. + * A 1-D tensor of shape [fw_num_units]. + * * 14: The forward cell gate bias. + * A 1-D tensor of shape [fw_num_units]. + * * 15: The forward output gate bias. + * A 1-D tensor of shape [fw_num_units]. + * * 16: The forward projection weights. Optional. + * A 2-D tensor of shape [fw_output_size, fw_num_units]. + * * 17: The forward projection bias. Optional. + * A 1-D tensor of shape [fw_output_size]. + * * 18: The backward input-to-input weights. Optional. + * A 2-D tensor of shape [bw_num_units, input_size], where “bw_num_units” + * corresponds to the number of backward cell units. + * * 19: The backward input-to-forget weights. + * A 2-D tensor of shape [bw_num_units, input_size]. + * * 20: The backward input-to-cell weights. + * A 2-D tensor of shape [bw_num_units, input_size]. + * * 21: The backward input-to-output weights. + * A 2-D tensor of shape [bw_num_units, input_size]. + * * 22: The backward recurrent-to-input weights. Optional. + * A 2-D tensor of shape [bw_num_units, bw_output_size], where “bw_output_size” + * corresponds to either the number of cell units (i.e., “bw_num_units”), + * or the second dimension of the “bw_projection_weights”, if defined. + * * 23: The backward recurrent-to-forget weights. + * A 2-D tensor of shape [bw_num_units, bw_output_size]. + * * 24: The backward recurrent-to-cell weights. + * A 2-D tensor of shape [bw_num_units, bw_output_size]. + * * 25: The backward recurrent-to-output weights. + * A 2-D tensor of shape [bw_num_units, bw_output_size]. + * * 26: The backward cell-to-input weights. Optional. + * A 1-D tensor of shape [bw_num_units]. + * * 27: The backward cell-to-forget weights. Optional. + * A 1-D tensor of shape [bw_num_units]. + * * 28: The backward cell-to-output weights. Optional. + * A 1-D tensor of shape [bw_num_units]. + * * 29: The backward input gate bias. Optional. + * A 1-D tensor of shape [bw_num_units]. + * * 30: The backward forget gate bias. + * A 1-D tensor of shape [bw_num_units]. + * * 31: The backward cell gate bias. + * A 1-D tensor of shape [bw_num_units]. + * * 32: The backward output gate bias. + * A 1-D tensor of shape [bw_num_units]. + * * 33: The backward projection weights. Optional. + * A 2-D tensor of shape [bw_output_size, bw_num_units]. + * * 34: The backward projection bias. Optional. + * A 1-D tensor of shape [bw_output_size]. + * * 35: The forward input activation state. + * A 2-D tensor of shape [batch_size, bw_output_size]. + * * 36: The forward input cell state. + * A 2-D tensor of shape [batch_size, bw_num_units]. + * * 37: The backward input activation state. + * A 2-D tensor of shape [batch_size, bw_output_size]. + * * 38: The backward input cell state. + * A 2-D tensor of shape [batch_size, bw_num_units]. + * * 39: The auxiliary input. Optional. + * A 3-D tensor of shape [max_time, batch_size, aux_input_size], + * where “batch_size” corresponds to the batching dimension, and + * “aux_input_size” is the size of the auxiliary input. Optional. See + * the docs above for the usage modes explanation. + * * 40: The forward auxiliary input-to-input weights. + * Optional. See the docs above for the usage modes explanation. + * A 2-D tensor of shape [fw_num_units, aux_input_size]. + * * 41: The forward auxiliary input-to-forget weights. + * Optional. See the docs above for the usage modes explanation. + * A 2-D tensor of shape [fw_num_units, aux_input_size]. + * * 42: The forward auxiliary input-to-cell weights. + * Optional. See the docs above for the usage modes explanation. + * A 2-D tensor of shape [fw_num_units, aux_input_size]. + * * 43: The forward auxiliary input-to-output weights. + * Optional. See the docs above for the usage modes explanation. + * A 2-D tensor of shape [fw_num_units, aux_input_size]. + * * 44: The backward auxiliary input-to-input weights. + * Optional. See the docs above for the usage modes explanation. + * A 2-D tensor of shape [bw_num_units, aux_input_size]. + * * 45: The backward auxiliary input-to-forget weights. + * Optional. See the docs above for the usage modes explanation. + * A 2-D tensor of shape [bw_num_units, aux_input_size]. + * * 46: The backward auxiliary input-to-cell weights. + * Optional. See the docs above for the usage modes explanation. + * A 2-D tensor of shape [bw_num_units, aux_input_size]. + * * 47: The backward auxiliary input-to-output weights. + * Optional. See the docs above for the usage modes explanation. + * A 2-D tensor of shape [bw_num_units, aux_input_size]. + * * 48: The activation function. + * A value indicating the activation function: + * + * * 49: The clipping threshold for the cell state, such + * that values are bound within [-cell_clip, cell_clip]. If set to 0.0 + * then clipping is disabled. + * If all the input tensors have type {@link OperandType::TENSOR_FLOAT32}, + * this scalar must be of the type {@link OperandType::FLOAT32}, + * otherwise if all the input tensors have the type + * {@link OperandType::TENSOR_FLOAT16}, this scalar must be + * of type {@link OperandType::FLOAT16}. + * * 50: The clipping threshold for the output from the + * projection layer, such that values are bound within + * [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. + * If all the input tensors have type {@link OperandType::TENSOR_FLOAT32}, + * this scalar must be of the type {@link OperandType::FLOAT32}, + * otherwise if all the input tensors have the type + * {@link OperandType::TENSOR_FLOAT16}, this scalar must be + * of type {@link OperandType::FLOAT16}. + * * 51: merge_outputs + * An {@link OperandType::BOOL} scalar specifying if the outputs + * from forward and backward cells should be merged. + * * 52: time_major + * An {@link OperandType::BOOL} scalar specifying the shape format + * of input and output tensors. + * * 53: The forward input layer normalization weights. Optional. + * A 1-D tensor of shape [fw_num_units]. Used to rescale normalized inputs + * to activation at input gate. + * * 54: The forward forget layer normalization weights. Optional. + * A 1-D tensor of shape [fw_num_units]. Used to rescale normalized inputs + * to activation at forget gate. + * * 55: The forward cell layer normalization weights. Optional. + * A 1-D tensor of shape [fw_num_units]. Used to rescale normalized inputs + * to activation at cell gate. + * * 56: The forward output layer normalization weights. Optional. + * A 1-D tensor of shape [fw_num_units]. Used to rescale normalized inputs + * to activation at output gate. + * * 57: The backward input layer normalization weights. Optional. + * A 1-D tensor of shape [bw_num_units]. Used to rescale normalized inputs + * to activation at input gate. + * * 58: The backward forget layer normalization weights. Optional. + * A 1-D tensor of shape [bw_num_units]. Used to rescale normalized inputs + * to activation at forget gate. + * * 59: The backward cell layer normalization weights. Optional. + * A 1-D tensor of shape [bw_num_units]. Used to rescale normalized inputs + * to activation at cell gate. + * * 60: The backward output layer normalization weights. Optional. + * A 1-D tensor of shape [bw_num_units]. Used to rescale normalized inputs + * to activation at output gate. + * + * Outputs: + * * 0: The forward output. + * A 3-D tensor of shape: + * If time-major and not merge_outputs: + * [max_time, batch_size, fw_output_size] + * If time-major and merge_outputs: + * [max_time, batch_size, fw_output_size + bw_output_size] + * If batch-major and not merge_outputs: + * [batch_size, max_time, fw_output_size] + * If batch-major and merge_outputs: + * [batch_size, max_time, fw_output_size + bw_output_size] + * * 1: The backward output. Unused if merge_outputs is true. + * A 3-D tensor of shape: + * If time-major: [max_time, batch_size, bw_output_size] + * If batch-major: [batch_size, max_time, bw_output_size] + * * 2: The forward activation state output. + * A 2-D tensor of shape [batch_size, fw_output_size] containing an + * activation state from the last time step in the sequence. This + * output is optional and can be omitted. If this output is present + * then outputs 3-5 must be present as well. + * Available since HAL version 1.3. + * * 3: The forward cell state output. + * A tensor of shape [batch_size, fw_cell_size] containing a cell state + * from the last time step in the sequence. This output is optional + * and can be omitted. If this output is present + * then outputs 2, 4, 5 must be present as well. + * Available since HAL version 1.3. + * * 4: The backward activation state output. + * A 2-D tensor of shape [batch_size, bw_output_size] containing an + * activation state from the last time step in the sequence. This + * output is optional and can be omitted. If this output is present + * then outputs 2, 3, 5 must be present as well. + * Available since HAL version 1.3. + * * 5: The backward cell state output. + * A tensor of shape [batch_size, bw_cell_size] containing a cell state + * from the last time step in the sequence. This output is optional + * and can be omitted. If this output is present + * then outputs 2-4 must be present as well. + * Available since HAL version 1.3. + */ + BIDIRECTIONAL_SEQUENCE_LSTM = 42, + /** + * A recurrent neural network layer that applies a basic RNN cell to a + * sequence of inputs in forward and backward directions. + * + * This Op unrolls the input along the sequence dimension, and implements + * the following operation for each element in the sequence s = + * 1...sequence_length: + * fw_outputs[s] = fw_state = activation(inputs[s] * fw_input_weights’ + + * fw_state * fw_recurrent_weights’ + fw_bias) + * + * And for each element in sequence t = sequence_length : 1 + * bw_outputs[t] = bw_state = activation(inputs[t] * bw_input_weights’ + + * bw_state * bw_recurrent_weights’ + bw_bias) + * + * Where: + * * “{fw,bw}_input_weights” is a weight matrix that multiplies the inputs; + * * “{fw,bw}_recurrent_weights” is a weight matrix that multiplies the + * current “state” which itself is the output from the previous time step + * computation; + * * “{fw,bw}_bias” is a bias vector (added to each output vector in the + * batch); + * * “activation” is the function passed as the “fused_activation_function” + * argument (if not “NONE”). + * + * The op supports cross-linking via an auxiliary input. Regular cell feeds + * one input into the two RNN cells in the following way: + * + * INPUT (INPUT_REVERSED) + * | | + * --------------------- + * | FW_RNN BW_RNN | + * --------------------- + * | | + * FW_OUT BW_OUT + * + * An op with cross-linking takes two inputs and feeds them into the RNN + * cells in the following way: + * + * AUX_INPUT (AUX_INPUT_REVERSED) + * | | + * INPUT | (INPUT_R'D.)| + * | | | | + * ----------------------- + * | \ / \ / | + * | FW_RNN BW_RNN | + * ----------------------- + * | | + * FW_OUT BW_OUT + * + * The cross-linking mode is enabled iff auxiliary input and auxiliary + * weights are present. While stacking this op on top of itself, this + * allows to connect both forward and backward outputs from previous cell + * to the next cell's input. + * + * Since HAL version 1.3 parallel linking mode is supported. The mode is + * enabled if auxiliary input is present but auxiliary weights are omitted. + * In this case, the cell feeds inputs into the RNN in the following way: + * + * INPUT (AUX_INPUT_REVERSED) + * | | + * --------------------- + * | FW_RNN BW_RNN | + * --------------------- + * | | + * FW_OUT BW_OUT + * + * While stacking this op on top of itself, this allows to connect both + * forward and backward outputs from previous cell to the next cell's + * corresponding inputs. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * The input tensors must all be the same type. + * + * Inputs: + * * 0: input. + * A 3-D tensor. The shape is defined by the input 6 (timeMajor). If + * it is set to true, then the input has a shape [maxTime, batchSize, + * inputSize], otherwise the input has a shape [batchSize, maxTime, + * inputSize]. + * * 1: fwWeights. + * A 2-D tensor of shape [fwNumUnits, inputSize]. + * * 2: fwRecurrentWeights. + * A 2-D tensor of shape [fwNumUnits, fwNumUnits]. + * * 3: fwBias. + * A 1-D tensor of shape [fwNumUnits]. + * * 4: fwHiddenState. + * A 2-D tensor of shape [batchSize, fwNumUnits]. Specifies a hidden + * state input for the first time step of the computation. + * * 5: bwWeights. + * A 2-D tensor of shape [bwNumUnits, inputSize]. + * * 6: bwRecurrentWeights. + * A 2-D tensor of shape [bwNumUnits, bwNumUnits]. + * * 7: bwBias. + * A 1-D tensor of shape [bwNumUnits]. + * * 8: bwHiddenState + * A 2-D tensor of shape [batchSize, bwNumUnits]. Specifies a hidden + * state input for the first time step of the computation. + * * 9: auxInput. + * A 3-D tensor. The shape is defined by the input 6 (timeMajor). If + * it is set to true, then the input has a shape [maxTime, batchSize, + * auxInputSize], otherwise the input has a shape [batchSize, maxTime, + * auxInputSize]. Can be omitted. See the docs above for the usage + * modes explanation. + * * 10:fwAuxWeights. + * A 2-D tensor of shape [fwNumUnits, auxInputSize]. Can be omitted. + * See the docs above for the usage modes explanation. + * * 11:bwAuxWeights. + * A 2-D tensor of shape [bwNumUnits, auxInputSize]. Can be omitted. + * See the docs above for the usage modes explanation. + * * 12:fusedActivationFunction. + * A {@link FusedActivationFunc} value indicating the activation function. If + * “NONE” is specified then it results in a linear activation. + * * 13:timeMajor + * An {@link OperandType::BOOL} scalar specifying the shape format + * of input and output tensors. + * * 14:mergeOutputs + * An {@link OperandType::BOOL} scalar specifying if the outputs + * from forward and backward cells are separate (if set to false) or + * concatenated (if set to true). + * Outputs: + * * 0: fwOutput. + * A 3-D tensor. The first two dimensions of the shape are defined by + * the input 6 (timeMajor) and the third dimension is defined by the + * input 14 (mergeOutputs). If timeMajor is set to true, then the first + * two dimensions are [maxTime, batchSize], otherwise they are set to + * [batchSize, maxTime]. If mergeOutputs is set to true, then the third + * dimension is equal to (fwNumUnits + bwNumUnits), otherwise it is set + * to fwNumUnits. + * * 1: bwOutput. + * A 3-D tensor. If the input 14 (mergeOutputs) is set to true, then + * this tensor is not produced. The shape is defined by the input 6 + * (timeMajor). If it is set to true, then the shape is set to + * [maxTime, batchSize, bwNumUnits], otherwise the shape is set to + * [batchSize, maxTime, bwNumUnits]. + * * 2: The forward hidden state output. + * A 2-D tensor of shape [batchSize, fwNumUnits] containing a hidden + * state from the last time step in the sequence. This output is + * optional and can be omitted. If this output is present then output + * 3 must be present as well. + * Available since HAL version 1.3. + * * 3: The backward hidden state output. + * A 2-D tensor of shape [batchSize, bwNumUnits] containing a hidden + * state from the last time step in the sequence. This output is + * optional and can be omitted. If this output is present then output + * 2 must be present as well. + * Available since HAL version 1.3. + */ + BIDIRECTIONAL_SEQUENCE_RNN = 43, + /** + * Greedily selects a subset of bounding boxes in descending order of score. + * + * This op applies NMS algorithm to each class. In each loop of execution, + * the box with maximum score gets selected and removed from the pending set. + * The scores of the rest of boxes are lowered according to the + * intersection-over-union (IOU) overlapping with the previously selected + * boxes and a specified NMS kernel method. Any boxes with score less + * than a threshold are removed from the pending set. + * + * Three NMS kernels are supported: + * * Hard: score_new = score_old * (1 if IoU < threshold else 0) + * * Linear: score_new = score_old * (1 if IoU < threshold else 1 - IoU) + * * Gaussian: score_new = score_old * exp(- IoU^2 / sigma) + * + * Axis-aligned bounding boxes are represented by its upper-left corner + * coordinate (x1,y1) and lower-right corner coordinate (x2,y2). A valid + * bounding box should satisfy x1 <= x2 and y1 <= y2. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Inputs: + * * 0: A 2-D Tensor of shape [num_rois, num_classes], specifying the score + * of each bounding box proposal. The boxes are grouped by batches in the + * first dimension. Zero num_rois is supported for this tensor. + * * 1: A 2-D Tensor specifying the bounding boxes of shape + * [num_rois, num_classes * 4], organized in the order [x1, y1, x2, y2]. + * The boxes are grouped by batches in the first dimension. The sequential + * order of the boxes corresponds with input0. For input0 of type + * {@link OperandType::TENSOR_QUANT8_ASYMM}, this tensor should be of + * {@link OperandType::TENSOR_QUANT16_ASYMM}, with zeroPoint of 0 and + * scale of 0.125. + * For input0 of type {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * this tensor should be of {@link OperandType::TENSOR_QUANT16_ASYMM}, + * with zeroPoint of -128 and scale of 0.125. + * Zero num_rois is supported for this tensor. + * * 2: A 1-D {@link OperandType::TENSOR_INT32} tensor, of shape + * [num_rois], specifying the batch index of each box. Boxes with + * the same batch index are grouped together. + * * 3: An {@link OperandType::FLOAT32} scalar, score_threshold. Boxes + * with scores lower than the threshold are filtered before sending + * to the NMS algorithm. + * * 4: An {@link OperandType::INT32} scalar, specifying the maximum + * number of selected bounding boxes for each image. Set to a negative + * value for unlimited number of output bounding boxes. + * * 5: An {@link OperandType::INT32} scalar, specifying the NMS + * kernel method, options are 0:hard, 1:linear, 2:gaussian. + * * 6: An {@link OperandType::FLOAT32} scalar, specifying the IoU + * threshold in hard and linear NMS kernel. This field is ignored if + * gaussian kernel is selected. + * * 7: An {@link OperandType::FLOAT32} scalar, specifying the sigma in + * gaussian NMS kernel. This field is ignored if gaussian kernel is + * not selected. + * * 8: An {@link OperandType::FLOAT32} scalar, nms_score_threshold. + * Boxes with scores lower than the threshold are dropped during the + * score updating phase in soft NMS. + * + * Outputs: + * * 0: A 1-D Tensor of the same {@link OperandType} as input0, with shape + * [num_output_rois], specifying the score of each output box. The boxes + * are grouped by batches, but the sequential order in each batch is not + * guaranteed. For type of {@link OperandType::TENSOR_QUANT8_ASYMM}, + * guaranteed. For type of {@link OperandType::TENSOR_QUANT8_ASYMM} + * or {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the scale and zero point must be the same as input0. + * * 1: A 2-D Tensor of the same {@link OperandType} as input1, with shape + * [num_output_rois, 4], specifying the coordinates of each + * output bounding box with the same format as input1. The sequential + * order of the boxes corresponds with output0. For type of + * {@link OperandType::TENSOR_QUANT16_ASYMM}, the scale must be + * 0.125 and the zero point must be 0. + * * 2: A 1-D {@link OperandType::TENSOR_INT32} tensor, of shape + * [num_output_rois], specifying the class of each output box. The + * sequential order of the boxes corresponds with output0. + * * 3: A 1-D {@link OperandType::TENSOR_INT32} tensor, of shape + * [num_output_rois], specifying the batch index of each box. Boxes + * with the same batch index are grouped together. + */ + BOX_WITH_NMS_LIMIT = 44, + /** + * Casts a tensor to a type. + * + * This operation ignores the scale and zeroPoint of quanized tensors, + * e.g. it treats a {@link OperandType::TENSOR_QUANT8_ASYMM} input + * as a tensor of uint8 values. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * Since HAL version 1.3, casting tensors of the following + * {@link OperandType} to the same {@link OperandType} is supported: + * * {@link OperandType::TENSOR_BOOL8} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT16_ASYMM} + * * {@link OperandType::TENSOR_QUANT16_SYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * * {@link OperandType::TENSOR_QUANT8_SYMM} + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: A tensor with the same shape as input0. + */ + CAST = 45, + /** + * Shuffle the channels of the input tensor. + * + * Given an input tensor and a integer value of num_groups, CHANNEL_SHUFFLE + * divide the channel dimension into num_groups groups, and reorganize the + * channels by grouping channels with the same index in each group. + * + * Along the channel dimension, the output is calculated using this formula: + * + * output_channel[k * num_groups + g] = input_channel[g * group_size + k] + * + * where group_size = num_channels / num_groups + * + * The number of channels must be divisible by num_groups. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor, specifying the tensor to be shuffled. + * * 1: An {@link OperandType::INT32} scalar, specifying the number of + * groups. + * * 2: An {@link OperandType::INT32} scalar, specifying the dimension + * channel shuffle would be performed on. Negative index is used to + * specify axis from the end (e.g. -1 for the last axis). Must be in + * the range [-n, n). + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} and same shape as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + CHANNEL_SHUFFLE = 46, + /** + * Apply postprocessing steps to bounding box detections. + * + * Bounding box detections are generated by applying transformation on a set + * of predefined anchors with the bounding box deltas from bounding box + * regression. A final step of hard NMS is applied to limit the number of + * returned boxes. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Inputs: + * * 0: A 3-D Tensor of shape [batches, num_anchors, num_classes], specifying + * the score of each anchor with each class. Class 0 for each + * [batches, num_anchors, 0] is background and will be ignored. + * * 1: A 3-D Tensor of shape [batches, num_anchors, length_box_encoding], with + * the first four values in length_box_encoding specifying the bounding + * box deltas. The box deltas are encoded in the order of [dy, dx, dh, dw], + * where dy and dx is the linear-scale relative correction factor for the + * center position of the bounding box with respect to the width and height, + * dh and dw is the log-scale relative correction factor for the width and + * height. All the entries in length_box_encoding beyond the first four + * values are ignored in this operation. + * * 2: A 2-D Tensor of shape [num_anchors, 4], specifying the shape of each + * predefined anchor, with format [ctr_y, ctr_x, h, w], where ctr_y and + * ctr_x are the center position of the box, and h and w are the height + * and the width. + * * 3: An {@link OperandType::FLOAT32} scalar, specifying the scaling + * factor for dy in bounding box deltas. + * * 4: An {@link OperandType::FLOAT32} scalar, specifying the scaling + * factor for dx in bounding box deltas. + * * 5: An {@link OperandType::FLOAT32} scalar, specifying the scaling + * factor for dh in bounding box deltas. + * * 6: An {@link OperandType::FLOAT32} scalar, specifying the scaling + * factor for dw in bounding box deltas. + * * 7: An {@link OperandType::BOOL} scalar, set to true to use regular + * multi-class NMS algorithm that do NMS separately for each class, + * set to false for a faster algorithm that only do one single NMS + * using the highest class score.. + * * 8: An {@link OperandType::INT32} scalar, max_num_detections, specifying + * the maximum number of boxes for the output. Boxes with the lowest + * scores are discarded to meet the limit. + * * 9: An {@link OperandType::INT32} scalar, only used when input7 is + * set to false, specifying the maximum number of classes per detection. + * * 10: An {@link OperandType::INT32} scalar, only used when input7 is + * set to true, specifying the maximum number of detections when + * applying NMS algorithm for each single class. + * * 11: A scalar, score_threshold. Boxes with scores lower than the + * threshold are filtered before sending to the NMS algorithm. The + * scalar must be of {@link OperandType::FLOAT16} if input0 is of + * {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} if input0 is of + * {@link OperandType::TENSOR_FLOAT32}. + * * 12: A scalar, specifying the IoU threshold for hard NMS. The scalar + * must be of {@link OperandType::FLOAT16} if input0 is of + * {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} if input0 is of + * {@link OperandType::TENSOR_FLOAT32}. + * * 13: An {@link OperandType::BOOL} scalar, set to true to include + * background class in the list of label map for the output, set + * to false to not include the background. When the background + * class is included, it has label 0 and the output classes start + * at 1 in the label map, otherwise, the output classes start at 0. + * + * Outputs: + * * 0: A 2-D tensor of the same {@link OperandType} as input0, with shape + * [batches, max_num_detections], specifying the score of each output + * detections. + * * 1: A 3-D tensor of shape [batches, max_num_detections, 4], specifying the + * coordinates of each output bounding box, with format + * [y1, x1, y2, x2]. + * * 2: A 2-D {@link OperandType::TENSOR_INT32} tensor, of shape + * [batches, max_num_detections], specifying the class label for each + * output detection. + * * 3: An 1-D {@link OperandType::TENSOR_INT32} tensor, of shape [batches], + * specifying the number of valid output detections for each batch. + */ + DETECTION_POSTPROCESSING = 47, + /** + * For input tensors x and y, computes x == y elementwise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * This operation supports broadcasting. + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType} and dimensions compatible + * with input0. + * + * Outputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + */ + EQUAL = 48, + /** + * Computes exponential of x element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + EXP = 49, + /** + * Inserts a dimension of 1 into a tensor's shape. + * + * Given a tensor input, this operation inserts a dimension of 1 at the + * given dimension index of input's shape. The dimension index starts at + * zero; if you specify a negative dimension index, it is counted backward + * from the end. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: An n-D tensor. + * * 1: An {@link OperandType::INT32} scalar specifying the dimension + * index to expand. Must be in the range [-(n + 1), (n + 1)). + * + * Outputs: + * * 0: An (n + 1)-D tensor with the same {@link OperandType} and data as + * input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + EXPAND_DIMS = 50, + /** + * Gathers values along an axis. + * + * Produces an output tensor with shape + * input0.dimension[:axis] + indices.dimension + input0.dimension[axis + 1:] + * where: + * # Vector indices (output is rank(input0)). + * output[a_0, ..., a_n, i, b_0, ..., b_n] = + * input0[a_0, ..., a_n, indices[i], b_0, ..., b_n] + * + * # Higher rank indices (output is rank(input0) + rank(indices) - 1). + * output[a_0, ..., a_n, i, ..., j, b_0, ... b_n] = + * input0[a_0, ..., a_n, indices[i, ..., j], b_0, ..., b_n] + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: An n-D tensor from which to gather values. + * * 1: An {@link OperandType::INT32} scalar specifying the axis. + * Negative index is used to specify axis from the end + * (e.g. -1 for the last axis). Must be in the range [-n, n). + * * 2: A k-D tensor {@link OperandType::TENSOR_INT32} of indices. + * The values must be in the bounds of the corresponding dimensions + * of input0. + * + * Outputs: + * * 0: An (n + k - 1)-D tensor with the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + GATHER = 51, + /** + * Generate aixs-aligned bounding box proposals. + * + * Bounding box proposals are generated by applying transformation on a set + * of predefined anchors with the bounding box deltas from bounding box + * regression. A final step of hard NMS is applied to limit the number of + * returned boxes. + * + * Axis-aligned bounding boxes are represented by its upper-left corner + * coordinate (x1,y1) and lower-right corner coordinate (x2,y2). A valid + * bounding box should satisfy x1 <= x2 and y1 <= y2. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Inputs: + * * 0: A 4-D Tensor specifying the score of each anchor at each + * location. With "NHWC" data layout, the tensor shape is + * [batches, height, width, num_anchors]. With "NCHW" data layout, + * the tensor shape is [batches, num_anchors, height, width]. + * * 1: A 4-D Tensor specifying the bounding box deltas. With "NHWC" data + * layout, the tensor shape is [batches, height, width, num_anchors * 4]. + * With "NCHW" data layout, the tensor shape is + * [batches, num_anchors * 4, height, width]. The box deltas are encoded + * in the order of [dx, dy, dw, dh], where dx and dy is the linear-scale + * relative correction factor for the center position of the bounding box + * with respect to the width and height, dw and dh is the log-scale + * relative correction factor for the width and height. The last + * dimensions is the channel dimension. + * * 2: A 2-D Tensor of shape [num_anchors, 4], specifying the shape of each + * predefined anchor, with format [x1, y1, x2, y2]. For input0 of type + * {@link OperandType::TENSOR_QUANT8_ASYMM} or + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, this tensor should be of + * {@link OperandType::TENSOR_QUANT16_SYMM}, with scale of 0.125. + * * 3: A 2-D Tensor of shape [batches, 2], specifying the size of + * each image in the batch, with format [image_height, image_width]. + * For input0 of type {@link OperandType::TENSOR_QUANT8_ASYMM} or + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, this + * tensor should be of {@link OperandType::TENSOR_QUANT16_SYMM}, with + * scale of 0.125. + * * 4: An {@link OperandType::FLOAT32} scalar, specifying the ratio + * from the height of original image to the height of feature map. + * * 5: An {@link OperandType::FLOAT32} scalar, specifying the ratio + * from the width of original image to the width of feature map. + * * 6: An {@link OperandType::INT32} scalar, specifying the maximum + * number of boxes before going into the hard NMS algorithm. Boxes + * with the lowest scores are discarded to meet the limit. Set to + * a non-positive value for unlimited number. + * * 7: An {@link OperandType::INT32} scalar, specifying the maximum + * number of boxes returning from the hard NMS algorithm. Boxes + * with the lowest scores are discarded to meet the limit. Set to + * a non-positive value for unlimited number. + * * 8: An {@link OperandType::FLOAT32} scalar, specifying the IoU + * threshold for hard NMS. + * * 9: An {@link OperandType::FLOAT32} scalar, min_size. Boxes with + * height or width lower than the absolute threshold are filtered out. + * * 10: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0 and input1. Set to false for NHWC. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0, of shape + * [num_output_rois], specifying the score of each output box. + * The boxes are grouped by batches, but the sequential order in + * each batch is not guaranteed. For type of + * {@link OperandType::TENSOR_QUANT8_ASYMM} or + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, the scale and zero + * point must be the same as input0. + * * 1: A tensor of the same {@link OperandType} as input3, of shape + * [num_output_rois, 4], specifying the coordinates of each output + * bounding box for each class, with format [x1, y1, x2, y2]. + * The sequential order of the boxes corresponds with output0. + * For type of {@link OperandType::TENSOR_QUANT16_ASYMM}, the + * scale must be 0.125 and the zero point must be 0. + * * 2: A 1-D {@link OperandType::TENSOR_INT32} tensor, of shape + * [num_output_rois], specifying the batch index of each box. Boxes + * with the same batch index are grouped together. + */ + GENERATE_PROPOSALS = 52, + /** + * For input tensors x and y, computes x > y elementwise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * This operation supports broadcasting. + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType} and dimensions compatible + * with input0. + * + * Outputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + */ + GREATER = 53, + /** + * For input tensors x and y, computes x >= y elementwise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * This operation supports broadcasting. + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType} and dimensions compatible + * with input0. + * + * Outputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + */ + GREATER_EQUAL = 54, + /** + * Performs a grouped 2-D convolution operation. + * + * Given an input tensor of shape [batches, height, width, depth_in] and a + * filter tensor of shape [depth_out, filter_height, filter_width, depth_group] + * containing depth_out convolutional filters of depth depth_group, GROUPED_CONV + * applies a group of different filters to each input channel group, then + * concatenates the results together. + * + * Specifically, the input channels are divided into num_groups groups, each with + * depth depth_group, i.e. depth_in = num_groups * depth_group. The convolutional + * filters are also divided into num_groups groups, i.e. depth_out is divisible + * by num_groups. GROUPED_CONV applies each group of filters to the corresponding + * input channel group, and the result are concatenated together. + * + * The output dimensions are functions of the filter dimensions, stride, and + * padding. + * + * The values in the output tensor are computed as: + * + * output[b, i, j, g * channel_multiplier + q] = + * sum_{di, dj, dk} ( + * input[b, strides[1] * i + di, strides[2] * j + dj, + * g * depth_group + dk] * + * filter[g * channel_multiplier + q, di, dj, dk] + * ) + bias[channel] + * + * where channel_multiplier = depth_out / num_groups + * + * Supported tensor {@link OperandType} configurations: + * * 16 bit floating point: + * * * {@link OperandType::TENSOR_FLOAT16} for input, filter, output, and bias. + * + * * 32 bit floating point: + * * * {@link OperandType::TENSOR_FLOAT32} for input, filter, output, and bias. + * + * * Quantized: + * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, filter, and output. + * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to + * * * input.scale * filter.scale). + * + * * Quantized signed (since HAL version 1.3): + * * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} for input, filter, and output. + * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to + * * * input.scale * filter.scale). + * + * * Quantized with symmetric per channel quantization for the filter: + * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, and output. + * * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} for filter. + * * * {@link OperandType::TENSOR_INT32} for bias (scale set to 0.0, + * * * each value scaling is separate and equal to input.scale * filter.scales[channel]). + * + * * Quantized signed with filter symmetric per channel quantization (since HAL version 1.3): + * * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} for input, and output. + * * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} for filter. + * * * {@link OperandType::TENSOR_INT32} for bias (scale set to 0.0, + * * * each value scaling is separate and equal to input.scale * filter.scales[channel]). + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * + * Both explicit padding and implicit padding are supported. + * + * Inputs (explicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input, where depth_in = num_groups * depth_group. + * * 1: A 4-D tensor, of shape + * [depth_out, filter_height, filter_width, depth_group], specifying + * the filter, where depth_out must be divisible by num_groups. For + * tensor of type {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} + * the channel dimension (channelDim at + * {@link SymmPerChannelQuantParams}) must be set to 0. + * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input + * tensor of type {@link OperandType::TENSOR_FLOAT32} or + * {@link OperandType::TENSOR_FLOAT16}, the bias must be of the same type. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint + * of 0 and bias_scale == input_scale * filter_scale. For filter tensor + * of {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL}, the bias + * should be of {@link OperandType::TENSOR_INT32}, with zeroPoint of + * 0 and bias_scale of 0. The actual scale of each value 'i' is equal to + * bias_scale[i] = input_scale * filter_scale[i]. + * * 3: An {@link OperandType::INT32} scalar, specifying the padding on + * the left, in the ‘width’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the padding on + * the right, in the ‘width’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the padding on + * the top, in the ‘height’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the padding on + * the bottom, in the ‘height’ dimension. + * * 7: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 8: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 9: An {@link OperandType::INT32} scalar, specifying the number of + * groups. + * * 10: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 11: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0 and output0. Set to false for NHWC. + * + * Inputs (implicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input, where depth_in = num_groups * depth_group. + * * 1: A 4-D tensor, of shape + * [depth_out, filter_height, filter_width, depth_group], specifying + * the filter, where depth_out must be divisible by num_groups. For + * tensor of type {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} + * the channel dimension (SymmPerChannelQuantParams::channelDim) + * must be set to 0. + * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input + * tensor of type {@link OperandType::TENSOR_FLOAT32} or + * {@link OperandType::TENSOR_FLOAT16}, the bias must be of the same + * {@link OperandType::TENSOR_FLOAT16}, the bias must be of the same type. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * the bias should be of {@link OperandType::TENSOR_INT32}, with zeroPoint + * of 0 and bias_scale == input_scale * filter_scale. For filter tensor + * of {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL}, the bias + * should be of {@link OperandType::TENSOR_INT32}, with zeroPoint of + * 0 and bias_scale of 0. The actual scale of each value 'i' is equal to + * bias_scale[i] = input_scale * filter_scale[i]. + * * 3: An {@link OperandType::INT32} scalar, specifying the implicit + * padding scheme, has to be one of the + * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. + * * 4: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the number of + * groups. + * * 7: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 8: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0 and output0. Set to false for NHWC. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, out_height, out_width, depth_out]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint can be different from inputs' scale and zeroPoint. + */ + GROUPED_CONV_2D = 55, + /** + * Localize the maximum keypoints from heatmaps. + * + * This operation approximates the accurate maximum keypoint scores and + * indices after bicubic upscaling by using Taylor expansion up to the + * quadratic term. + * + * The bounding box is represented by its upper-left corner coordinate + * (x1,y1) and lower-right corner coordinate (x2,y2) in the original image. + * A valid bounding box should satisfy x1 <= x2 and y1 <= y2. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * + * Inputs: + * * 0: A 4-D Tensor of shape + * [num_boxes, heatmap_size, heatmap_size, num_keypoints], + * specifying the heatmaps, the height and width of heatmaps should + * be the same, and must be greater than or equal to 2. + * * 1: A 2-D Tensor of shape [num_boxes, 4], specifying the bounding boxes, + * each with format [x1, y1, x2, y2]. For input0 of type + * {@link OperandType::TENSOR_QUANT8_ASYMM}, this tensor should + * be of {@link OperandType::TENSOR_QUANT16_ASYMM}, with zeroPoint + * of 0 and scale of 0.125. + * For input0 of type + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, this tensor + * should be of {@link OperandType::TENSOR_QUANT16_ASYMM}, with + * zeroPoint of -128 and scale of 0.125. + * * 2: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0. Set to false for NHWC. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0, with shape + * [num_boxes, num_keypoints], specifying score of the keypoints. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} or + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint can be different from input0 scale and zeroPoint. + * * 1: A tensor of the same {@link OperandType} as input1, with shape + * [num_boxes, num_keypoints, 2], specifying the location of + * the keypoints, the second dimension is organized as + * [keypoint_x, keypoint_y]. + * For type of {@link OperandType::TENSOR_QUANT16_ASYMM}, the + * scale must be 0.125 and the zero point must be 0. + */ + HEATMAP_MAX_KEYPOINT = 56, + /** + * Applies instance normalization to the input tensor. + * + * The values in the output tensor are computed as: + * + * output[b, h, w, c] = + * (input[b, h, w, c] - mean[b, c]) * gamma / + * sqrt(var[b, c] + epsilon) + beta + * + * Where the mean and variance are computed across the spatial dimensions: + * + * mean[b, c] = + * sum_{h, w}(input[b, h, w, c]) / sum(1) + * + * var[b, c] = + * sum_{h, w}(pow(input[b, h, w, c] - mean[b, c], 2)) / sum(1) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * + * Inputs: + * * 0: An n-D tensor, specifying the tensor to be normalized. + * * 1: A scalar, specifying gamma, the scale applied to the normalized + * tensor. The scalar must be of {@link OperandType::FLOAT16} if + * input0 is of {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} if input0 is of + * {@link OperandType::TENSOR_FLOAT32}. + * * 2: A scalar, specifying beta, the offset applied to the normalized + * tensor. The scalar must be of {@link OperandType::FLOAT16} if + * input0 is of {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} if input0 is of + * {@link OperandType::TENSOR_FLOAT32}. + * * 3: A scalar, specifying epsilon, the small value added to variance to + * avoid dividing by zero. The scalar must be of {@link OperandType::FLOAT16} if + * input0 is of {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} if input0 is of + * {@link OperandType::TENSOR_FLOAT32}. + * * 4: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0 and output0. Set to false for NHWC. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} and same shape as input0. + */ + INSTANCE_NORMALIZATION = 57, + /** + * For input tensors x and y, computes x < y elementwise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * This operation supports broadcasting. + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType} and dimensions compatible + * with input0. + * + * Outputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + */ + LESS = 58, + /** + * For input tensors x and y, computes x <= y elementwise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * This operation supports broadcasting. + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType} and dimensions compatible + * with input0. + * + * Outputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + */ + LESS_EQUAL = 59, + /** + * Computes natural logarithm of x element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + LOG = 60, + /** + * Returns the truth value of x AND y element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * + * Supported tensor rank: from 1 + * + * This operation supports broadcasting. + * + * Inputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + * * 1: A tensor of {@link OperandType::TENSOR_BOOL8} and dimensions + * compatible with input0. + * + * Outputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + */ + LOGICAL_AND = 61, + /** + * Computes the truth value of NOT x element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + LOGICAL_NOT = 62, + /** + * Returns the truth value of x OR y element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * + * Supported tensor rank: from 1 + * + * This operation supports broadcasting. + * + * Inputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + * * 1: A tensor of {@link OperandType::TENSOR_BOOL8} and dimensions + * compatible with input0. + * + * Outputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + */ + LOGICAL_OR = 63, + /** + * Computes the log softmax activations given logits. + * + * The output is calculated using this formula: + * + * output = logits * beta - log(reduce_sum(exp(logits * beta), axis)) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor specifying the input logits. + * * 1: A scalar, specifying the positive scaling factor for the exponent, + * beta. + * For input tensor of {@link OperandType::TENSOR_FLOAT16}, the beta + * value must be of {@link OperandType::FLOAT16}. + * For input tensor of {@link OperandType::TENSOR_FLOAT32}, the beta + * value must be of {@link OperandType::FLOAT32}. + * * 2: An {@link OperandType::INT32} scalar specifying the axis to + * reduce across. Negative index is used to specify axis from the + * end (e.g. -1 for the last axis). Must be in the range [-n, n). + * + * Outputs: + * * 0: The output tensor of the same {@link OperandType} and shape as + * input0. + */ + LOG_SOFTMAX = 64, + /** + * Returns the element-wise maximum of two tensors. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType} and compatible dimensions + * with input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, + * the scales and zeroPoint can be different from input0 scale and zeroPoint. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, + * the scale and zeroPoint can be different from inputs' scale and zeroPoint. + */ + MAXIMUM = 65, + /** + * Returns the element-wise minimum of two tensors. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType} and compatible dimensions + * with input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, + * the scales and zeroPoint can be different from input0 scale and zeroPoint. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, + * the scale and zeroPoint can be different from inputs' scale and zeroPoint. + */ + MINIMUM = 66, + /** + * Computes numerical negative value element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + NEG = 67, + /** + * For input tensors x and y, computes x != y elementwise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * This operation supports broadcasting. + * + * Inputs: + * * 0: A tensor. + * * 1: A tensor of the same {@link OperandType} and dimensions compatible + * with input0. + * + * Outputs: + * * 0: A tensor of {@link OperandType::TENSOR_BOOL8}. + */ + NOT_EQUAL = 68, + /** + * Pads a tensor with the given constant value according to the specified + * paddings. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor, specifying the tensor to be padded. + * * 1: A 2-D Tensor of {@link OperandType::TENSOR_INT32}, the paddings + * for each spatial dimension of the input tensor. The shape of the + * tensor must be {rank(input0), 2}. + * padding[i, 0] specifies the number of elements to be padded in the + * front of dimension i. + * padding[i, 1] specifies the number of elements to be padded after + * the end of dimension i. + * * 2: An scalar specifying the value to use for padding input0. + * For input tensor of {@link OperandType::TENSOR_FLOAT16}, the + * pad value must be of {@link OperandType::FLOAT16}. + * For input tensor of {@link OperandType::TENSOR_FLOAT32}, the + * pad value must be of {@link OperandType::FLOAT32}. + * For input tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the pad value must be of {@link OperandType::INT32}. The + * scale and zeroPoint are assumed to be the same as in input0. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. The + * output tensor has the same rank as input0, and each + * dimension of the output tensor has the same size as the + * corresponding dimension of the input tensor plus the size + * of the padding: + * output0.dimension[i] = + * padding[i, 0] + input0.dimension[i] + padding[i, 1] + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + PAD_V2 = 69, + /** + * Computes the power of one value to another. + * + * Given a tensor base and a tensor exponent, this operation computes + * base^exponent elementwise. + * + * This operations supports broadcasting. The size of the output is the + * maximum size along each dimension of the input operands. It starts with + * the trailing dimensions, and works its way forward. + * + * For example: + * base.dimension = {4, 1, 2} + * exponent.dimension = {5, 4, 3, 1} + * output.dimension = {5, 4, 3, 2} + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: A tensor specifying the base. + * * 1: A tensor specifying the exponent. + * + * Outputs: + * * 0: An output tensor. + */ + POW = 70, + /** + * Parametric Rectified Linear Unit. + * + * It follows: f(x) = alpha * x for x < 0, f(x) = x for x >= 0, where alpha + * is a learned array with the same {@link OperandType} and compatible + * dimensions as input x. + * + * Two dimensions are compatible when: + * 1. they are equal, or + * 2. one of them is 1 + * + * The size of the output is the maximum size along each dimension of the + * input operands. It starts with the trailing dimensions, and works its way + * forward. + * + * Example: + * input.dimension = {4, 1, 2} + * alpha.dimension = {5, 4, 3, 1} + * output.dimension = {5, 4, 3, 2} + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: A tensor, specifying the input. + * * 1: A tensor of the same {@link OperandType}, and compatible dimensions + * as input0, specifying the alpha. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scales and zeroPoint can be different from input0 scale and zeroPoint. + */ + PRELU = 71, + /** + * Quantizes the input tensor. + * + * The formula for {@link OperandType::TENSOR_QUANT8_ASYMM} output tensor is: + * + * output = max(0, min(255, round(input / scale) + zeroPoint) + * + * The formula for {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} output + * tensor is: + * + * output = max(-128, min(127, round(input / scale) + zeroPoint) + * + * Supported input tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported output tensor {@link OperandType}: + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: A tensor, may be zero-sized. + * + * Outputs: + * * 0: The output tensor of same shape as input0, but with + * {@link OperandType::TENSOR_QUANT8_ASYMM} or. + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}. + */ + QUANTIZE = 72, + /** + * A version of quantized LSTM, using 16 bit quantization for internal + * state. + * + * There is no projection layer, so cell state size is equal to the output + * size. + * + * Inputs: + * * 0: A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [numBatches, inputSize] specifying the input to the LSTM + * cell. Tensor is quantized with a fixed quantization range of + * [-1, 127/128] (scale = 1/128, zeroPoint = 128). + * * 1: The input-to-input weights. + * A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [outputSize, inputSize] specifying input-to-input part of + * weights for fully-connected layer inside the LSTM cell. + * Quantization zero point and scale must be the same across all the + * weights. + * * 2: The input-to-forget weights. + * A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [outputSize, inputSize] specifying input-to-forget part of + * weights for fully-connected layer inside the LSTM cell. + * Quantization zero point and scale must be the same across all the + * weights. + * * 3: The input-to-cell weights. + * A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [outputSize, inputSize] specifying input-to-cell part of + * weights for fully-connected layer inside the LSTM cell. + * Quantization zero point and scale must be the same across all the + * weights. + * * 4: The input-to-output weights. + * A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [outputSize, inputSize] specifying input-to-output part of + * weights for fully-connected layer inside the LSTM cell. + * Quantization zero point and scale must be the same across all the + * weights. + * * 5: The recurrent-to-input weights. + * A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [outputSize, outputSize] specifying recurrent-to-input part + * of weights for fully-connected layer inside the LSTM cell. + * Quantization zero point and scale must be the same across all the + * weights. + * * 6: The recurrent-to-forget weights. + * A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [outputSize, outputSize] specifying recurrent-to-forget + * part of weights for fully-connected layer inside the LSTM cell. + * Quantization zero point and scale must be the same across all the + * weights. + * * 7: The recurrent-to-cell weights. + * A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [outputSize, outputSize] specifying recurrent-to-cell part + * of weights for fully-connected layer inside the LSTM cell. + * Quantization zero point and scale must be the same across all the + * weights. + * * 8: The recurrent-to-output weights. + * A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [outputSize, outputSize] specifying recurrent-to-output + * part of weights for fully-connected layer inside the LSTM cell. + * Quantization zero point and scale must be the same across all the + * weights. + * * 9: The input gate bias. + * A 1-D tensor of type {@link OperandType::TENSOR_INT32} and shape + * [outputSize] specifying the bias for the fully-connected layer + * inside the LSTM cell. Bias is quantized with scale being a product + * of input and weights scales and zeroPoint equal to 0. + * * 10:The forget gate bias. + * A 1-D tensor of type {@link OperandType::TENSOR_INT32} and shape + * [outputSize] specifying the bias for the fully-connected layer + * inside the LSTM cell. Bias is quantized with scale being a product + * of input and weights scales and zeroPoint equal to 0. + * * 11:The cell bias. + * A 1-D tensor of type {@link OperandType::TENSOR_INT32} and shape + * [outputSize] specifying the bias for the fully-connected layer + * inside the LSTM cell. Bias is quantized with scale being a product + * of input and weights scales and zeroPoint equal to 0. + * * 12:The output gate bias. + * A 1-D tensor of type {@link OperandType::TENSOR_INT32} and shape + * [outputSize] specifying the bias for the fully-connected layer + * inside the LSTM cell. Bias is quantized with scale being a product + * of input and weights scales and zeroPoint equal to 0. + * * 13: A 2-D tensor of type {@link OperandType::TENSOR_QUANT16_SYMM} + * and shape [numBatches, outputSize] specifying the cell state from the + * previous time step of the LSTM cell. It is quantized using a + * quantization range of [-2^4, 2^4 * 32767/32768] (scale = 2^4 / + * 32768, zeroPoint = 0). + * * 14: A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [numBathes, outputSize] specifying the output of the LSTM + * cell from previous time-step. Tensor is quantized with a fixed + * quantization range of [-1, 127/128] (scale = 1/128, zeroPoint = + * 128). + * + * + * Outputs: + * * 0: A 2-D tensor of type {@link OperandType::TENSOR_QUANT16_SYMM} + * and shape [numBatches, outputSize] which contains a cell state from + * the current time step. Tensor is quantized using a quantization + * range of [-2^4, 2^4 * 32767/32768] (scale = 2^4 / 32768, zeroPoint = + * 0). + * * 1: A 2-D tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM} + * and shape [numBathes, outputSize] which contains the output value. + * Tensor is quantized with a fixed quantization range of [-1, 127/128] + * (scale = 1/128, zeroPoint = 128). + */ + QUANTIZED_16BIT_LSTM = 73, + /** + * Draws samples from a multinomial distribution. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Inputs: + * * 0: A 2-D tensor with shape [batches, classes], specifying the + * unnormalized log-probabilities for all classes. + * * 1: A scalar {@link OperandType::INT32}, specifying the number of + * independent samples to draw for each row slice. + * * 2: A 1-D {@link OperandType::TENSOR_INT32} tensor with shape [2], + * specifying seeds used to initialize the random distribution. If both + * provided seeds are 0, both will be randomly generated. + * Outputs: + * * 0: A 2-D {@link OperandType::TENSOR_INT32} tensor with shape + * [batches, samples], containing the drawn samples. + */ + RANDOM_MULTINOMIAL = 74, + /** + * Reduces a tensor by computing the "logical and" of elements along given + * dimensions. + * + * If keep_dims is true, the reduced dimensions are + * retained with length 1. Otherwise, the rank of the tensor is reduced by + * 1 for each entry in dimensions. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor. + * * 1: A 1-D tensor of {@link OperandType::TENSOR_INT32}. The dimensions + * to reduce. Dimension values must be in the range [-n, n). + * * 2: An {@link OperandType::BOOL} scalar, keep_dims. If true, + * retains reduced dimensions with length 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * If all dimensions are reduced and keep_dims is false, the output + * shape is [1]. + */ + REDUCE_ALL = 75, + /** + * Reduces a tensor by computing the "logical or" of elements along given + * dimensions. + * + * If keep_dims is true, the reduced dimensions are + * retained with length 1. Otherwise, the rank of the tensor is reduced by + * 1 for each entry in dimensions. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_BOOL8} + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor. + * * 1: A 1-D tensor of {@link OperandType::TENSOR_INT32}. The dimensions + * to reduce. Dimension values must be in the range [-n, n). + * * 2: An {@link OperandType::BOOL} scalar, keep_dims. If true, + * retains reduced dimensions with length 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * If all dimensions are reduced and keep_dims is false, the output + * shape is [1]. + */ + REDUCE_ANY = 76, + /** + * Reduces a tensor by computing the maximum of elements along given + * dimensions. + * + * If keep_dims is true, the reduced dimensions are + * retained with length 1. Otherwise, the rank of the tensor is reduced by + * 1 for each entry in dimensions. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor. + * * 1: A 1-D tensor of {@link OperandType::TENSOR_INT32}. The dimensions + * to reduce. Dimension values must be in the range [-n, n). + * * 2: An {@link OperandType::BOOL} scalar, keep_dims. If true, + * retains reduced dimensions with length 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * If all dimensions are reduced and keep_dims is false, the output + * shape is [1]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + REDUCE_MAX = 77, + /** + * Reduces a tensor by computing the minimum of elements along given + * dimensions. + * + * If keep_dims is true, the reduced dimensions are + * retained with length 1. Otherwise, the rank of the tensor is reduced by + * 1 for each entry in dimensions. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor. + * * 1: A 1-D tensor of {@link OperandType::TENSOR_INT32}. The dimensions + * to reduce. Dimension values must be in the range [-n, n). + * * 2: An {@link OperandType::BOOL} scalar, keep_dims. If true, + * retains reduced dimensions with length 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * If all dimensions are reduced and keep_dims is false, the output + * shape is [1]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + REDUCE_MIN = 78, + /** + * Reduces a tensor by multiplying elements along given dimensions. + * + * If keep_dims is true, the reduced dimensions are + * retained with length 1. Otherwise, the rank of the tensor is reduced by + * 1 for each entry in dimensions. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor. + * * 1: A 1-D tensor of {@link OperandType::TENSOR_INT32}. The dimensions + * to reduce. Dimension values must be in the range [-n, n). + * * 2: An {@link OperandType::BOOL} scalar, keep_dims. If true, + * retains reduced dimensions with length 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * If all dimensions are reduced and keep_dims is false, the output + * shape is [1]. + */ + REDUCE_PROD = 79, + /** + * Reduces a tensor by summing elements along given dimensions. + * + * If keep_dims is true, the reduced dimensions are + * retained with length 1. Otherwise, the rank of the tensor is reduced by + * 1 for each entry in dimensions. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: up to 4 + * + * Inputs: + * * 0: An n-D tensor. + * * 1: A 1-D tensor of {@link OperandType::TENSOR_INT32}. The dimensions + * to reduce. Dimension values must be in the range [-n, n). + * * 2: An {@link OperandType::BOOL} scalar, keep_dims. If true, + * retains reduced dimensions with length 1. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. + * If all dimensions are reduced and keep_dims is false, the output + * shape is [1]. + */ + REDUCE_SUM = 80, + /** + * Select and scale the feature map of each region of interest to a unified + * output size by average pooling sampling points from bilinear interpolation. + * + * The region of interest is represented by its upper-left corner coordinate + * (x1,y1) and lower-right corner coordinate (x2,y2) in the original image. + * A spatial scaling factor is applied to map into feature map coordinate. + * A valid region of interest should satisfy x1 <= x2 and y1 <= y2. + * + * No rounding is applied in this operation. The sampling points are unified + * distributed in the pooling bin and their values are calculated by bilinear + * interpolation. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * + * Inputs: + * * 0: A 4-D tensor, specifying the feature map. + * * 1: A 2-D Tensor of shape [num_rois, 4], specifying the locations of + * the regions of interest, each line with format [x1, y1, x2, y2]. + * For input0 of type {@link OperandType::TENSOR_QUANT8_ASYMM}, + * this tensor should be of {@link OperandType::TENSOR_QUANT16_ASYMM}, + * with zeroPoint of 0 and scale of 0.125. Zero num_rois is + * supported for this tensor. + * * 2: An 1-D {@link OperandType::TENSOR_INT32} tensor, of shape + * [num_rois], specifying the batch index of each box. Boxes with + * the same batch index are grouped together. Zero num_rois is + * supported for this tensor. + * * 3: An {@link OperandType::INT32} scalar, specifying the output + * height of the output tensor. + * * 4: An {@link OperandType::INT32} scalar, specifying the output + * width of the output tensor. + * * 5: An {@link OperandType::FLOAT32} scalar, specifying the ratio + * from the height of original image to the height of feature map. + * * 6: An {@link OperandType::FLOAT32} scalar, specifying the ratio + * from the width of original image to the width of feature map. + * * 7: An {@link OperandType::INT32} scalar, specifying the number of + * sampling points in height dimension used to compute the output. + * Set to 0 for adaptive value of ceil(roi_height/out_height). + * * 8: An {@link OperandType::INT32} scalar, specifying the number of + * sampling points in width dimension used to compute the output. + * Set to 0 for adaptive value of ceil(roi_width/out_width). + * * 9: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0 and output0. Set to false for NHWC. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. The output + * shape is [num_rois, out_height, out_width, depth]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint can be different from the input0 scale and zeroPoint. + */ + ROI_ALIGN = 81, + /** + * Select and scale the feature map of each region of interest to a unified + * output size by max-pooling. + * + * The region of interest is represented by its upper-left corner coordinate + * (x1,y1) and lower-right corner coordinate (x2,y2) in the original image. + * A spatial scaling factor is applied to map into feature map coordinate. + * A valid region of interest should satisfy x1 <= x2 and y1 <= y2. + * + * Rounding is applied in this operation to ensure integer boundary for + * regions of interest and pooling bins. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * + * Inputs: + * * 0: A 4-D tensor, specifying the feature map. + * * 1: A 2-D Tensor of shape [num_rois, 4], specifying the locations of + * the regions of interest, each line with format [x1, y1, x2, y2]. + * For input0 of type {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * this tensor should be of {@link OperandType::TENSOR_QUANT16_ASYMM}, + * with zeroPoint of 0 and scale of 0.125. + * * 2: An 1-D {@link OperandType::TENSOR_INT32} tensor, of shape + * [num_rois], specifying the batch index of each box. Boxes with + * the same batch index are grouped together. + * * 3: An {@link OperandType::INT32} scalar, specifying the output + * height of the output tensor. + * * 4: An {@link OperandType::INT32} scalar, specifying the output + * width of the output tensor. + * * 5: An {@link OperandType::FLOAT32} scalar, specifying the ratio + * from the height of original image to the height of feature map. + * * 6: An {@link OperandType::FLOAT32} scalar, specifying the ratio + * from the width of original image to the width of feature map. + * * 7: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0 and output0. Set to false for NHWC. + * + * Outputs: + * * 0: A tensor of the same {@link OperandType} as input0. The output + * shape is [num_rois, out_height, out_width, depth]. + * For input0 of type {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + ROI_POOLING = 82, + /** + * Computes reciprocal of square root of x element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + RSQRT = 83, + /** + * Using a tensor of booleans c and input tensors x and y select values + * elementwise from both input tensors: + * + * O[i] = C[i] ? x[i] : y[i]. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: A tensor of type {@link OperandType::TENSOR_BOOL8} acting as a + * mask that chooses, based on the value at each element, whether the + * corresponding element in the output should be taken from input1 (if + * true) or input2 (if false). + * * 1: An input tensor of the same shape as input0. + * * 2: An input tensor of the same shape and type as input1. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scales and zeroPoint can be different from input1 scale and zeroPoint. + * + * Outputs: + * * 0: A tensor of the same type and shape as input1 and input2. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} tensor, + * the scale and zeroPoint can be different from inputs' scale and zeroPoint. + */ + SELECT = 84, + /** + * Computes sin of x element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + SIN = 85, + /** + * Extracts a slice of specified size from the input tensor starting at a + * specified location. + * + * The starting location is specified as a 1-D tensor containing offsets + * for each dimension. The size is specified as a 1-D tensor containing + * either size of a slice along corresponding dimension or -1. In the latter + * case, all the remaining elements in dimension are included in the slice. + * + * A sum of begin offset and a size of a slice must not exceed size of a + * corresponding dimension. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: An n-D tensor to take slice from, may be zero-sized. + * * 1: A 1-D tensor of type {@link OperandType::TENSOR_INT32} specifying + * the beginning indices of the slice in each dimension. + * * 2: A 1-D tensor of type {@link OperandType::TENSOR_INT32} specifying + * the size of the slice in each dimension. + * + * Outputs: + * * 0: An n-D tensor of the same type as the input containing the slice. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * its scale and zeroPoint has to be same as the input0 scale and zeroPoint. + */ + SLICE = 86, + /** + * Splits a tensor along a given axis into num_splits subtensors. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: An n-D tensor to split. + * * 1: An {@link OperandType::INT32} scalar specifying the axis along + * which to split. + * * 2: An {@link OperandType::INT32} scalar indicating the number of + * splits along given axis. Must evenly divide axis size. + * + * Outputs: + * * 0 ~ (num_splits - 1): Resulting subtensors. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + SPLIT = 87, + /** + * Computes square root of x element-wise. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor. + * + * Outputs: + * * 0: The output tensor of same shape as input0. + */ + SQRT = 88, + /** + * Constructs a tensor by tiling a given tensor. + * + * This operation creates a new tensor by replicating `input` `multiples` + * times. The output tensor's i-th dimension has `input.dims(i) * multiples[i]` + * elements, and the values of `input` are replicated `multiples[i]` times + * along the i-th dimension. + * For example, tiling `[a b c d]` by `[2]` produces `[a b c d a b c d]`. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: input, an n-D tensor specifying the input. + * * 1: multiples, a 1-D tensor of {@link OperandType::TENSOR_INT32}. + * The length of multiples must be n. + * + * Outputs: + * * 0: A tiled tensor of the same {@link OperandType} and rank as `input`. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + TILE = 89, + /** + * Finds values and indices of the k largest entries for the last dimension. + * + * Resulting values in each dimensions are sorted in descending order. If + * two values are equal, the one with larger index appears first. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: from 1 + * + * Inputs: + * * 0: input, an n-D tensor specifying the input. + * * 1: k, an {@link OperandType::INT32} scalar, specifying the number of + * top elements to look for along the last dimension. + * + * Outputs: + * * 0: An n-D tensor of the same type as the input, containing the k + * largest elements along each last dimensional slice. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + * * 1: An n-D tensor of type {@link OperandType::TENSOR_INT32} + * containing the indices of values within the last dimension of input. + */ + TOPK_V2 = 90, + /** + * Performs the transpose of 2-D convolution operation. + * + * This operation is sometimes called "deconvolution" after Deconvolutional + * Networks, but is actually the transpose (gradient) of + * {@link OperandType::CONV_2D} rather than an actual deconvolution. + * + * The output dimensions are functions of the filter dimensions, stride, and + * padding. + * + * Supported tensor {@link OperandType} configurations: + * * 16 bit floating point: + * * * {@link OperandType::TENSOR_FLOAT16} for input, filter, output, and bias. + * + * * 32 bit floating point: + * * * {@link OperandType::TENSOR_FLOAT32} for input, filter, output, and bias. + * + * * Quantized: + * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, filter, and output. + * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to + * * * input.scale * filter.scale). + * + * * Quantized with symmetric per channel quantization for the filter: + * * * {@link OperandType::TENSOR_QUANT8_ASYMM} for input, and output. + * * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} for filter. + * * * {@link OperandType::TENSOR_INT32} for bias (scale set to 0.0, + * * * each value scaling is separate and equal to input.scale * filter.scales[channel]). + * + * Available since HAL version 1.3: + * * Quantized signed (since HAL version 1.3): + * * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} for input, filter, and output. + * * * {@link OperandType::TENSOR_INT32} for bias (with scale set to + * * * input.scale * filter.scale). + * + * * Quantized signed with filter symmetric per channel quantization (since HAL version 1.3): + * * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} for input, and output. + * * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} for filter. + * * * {@link OperandType::TENSOR_INT32} for bias (scale set to 0.0, + * * * each value scaling is separate and equal to input.scale * filter.scales[channel]). + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * + * Both explicit padding and implicit padding are supported. + * + * Inputs (explicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input. + * * 1: A 4-D tensor, of shape + * [depth_out, filter_height, filter_width, depth_in], specifying the + * filter. For tensor of type + * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} the channel + * dimension (SymmPerChannelQuantParams::channelDim) must be set to 0. + * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input + * tensor of type {@link OperandType::TENSOR_FLOAT32} or + * {@link OperandType::TENSOR_FLOAT16}, the bias must be of the + * same type. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the bias should be of {@link OperandType::TENSOR_INT32}, + * with zeroPoint of 0 and bias_scale == input_scale * filter_scale. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL}, + * the bias must be of {@link OperandType::TENSOR_INT32}, with zeroPoint of 0 + * and bias_scale of 0. The actual scale of each value 'i' is equal to + * bias_scale[i] = input_scale * filter_scale[i]. + * * 3: An {@link OperandType::INT32} scalar, specifying the padding on + * the left, in the ‘width’ dimension. + * * 4: An {@link OperandType::INT32} scalar, specifying the padding on + * the right, in the ‘width’ dimension. + * * 5: An {@link OperandType::INT32} scalar, specifying the padding on + * the top, in the ‘height’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the padding on + * the bottom, in the ‘height’ dimension. + * * 7: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 8: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 9: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 10: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0 and output0. Set to false for NHWC. + * + * Inputs (implicit padding): + * * 0: A 4-D tensor, of shape [batches, height, width, depth_in], + * specifying the input. + * * 1: A 4-D tensor, of shape + * [depth_out, filter_height, filter_width, depth_in], specifying the + * filter. For tensor of type + * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} the channel + * dimension (SymmPerChannelQuantParams::channelDim) must be set to 0. + * * 2: A 1-D tensor, of shape [depth_out], specifying the bias. For input + * tensor of type {@link OperandType::TENSOR_FLOAT32} or + * {@link OperandType::TENSOR_FLOAT16}, the bias should be of the + * same type. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} + * and {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED}, + * the bias should be of {@link OperandType::TENSOR_INT32}, + * with zeroPoint of 0 and bias_scale == input_scale * filter_scale. + * For filter tensor of {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL}, + * the bias must be of {@link OperandType::TENSOR_INT32}, with zeroPoint of 0 + * and bias_scale of 0. The actual scale of each value 'i' is equal to + * bias_scale[i] = input_scale * filter_scale[i]. + * * 3: An {@link OperandType::TENSOR_INT32} tensor, specifying the output + * tensor shape. + * * 4: An {@link OperandType::INT32} scalar, specifying the implicit + * padding scheme, has to be one of the + * following values: {0 (NONE), 1 (SAME), 2 (VALID)}. + * * 5: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘width’ dimension. + * * 6: An {@link OperandType::INT32} scalar, specifying the stride when + * walking through input in the ‘height’ dimension. + * * 7: An {@link OperandType::INT32} scalar, and has to be one of the + * {@link FusedActivationFunc} values. Specifies the activation to + * invoke on the result. + * * 8: An {@link OperandType::BOOL} scalar, set to true to specify + * NCHW data layout for input0 and output0. Set to false for NHWC. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, out_height, out_width, depth_out]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint can be different from inputs' scale and zeroPoint. + */ + TRANSPOSE_CONV_2D = 91, + /** + * A recurrent neural network specified by an LSTM cell. + * + * Performs (fully) dynamic unrolling of input. + * + * This Op unrolls the input along the time dimension, and implements the + * following operation for each element in the sequence + * s = 1...sequence_length: + * outputs[s] = projection(state = activation(LSTMOp(inputs[s]))) + * + * Where LSTMOp is the LSTM op as in {@link OperandType::LSTM}, + * the "projection" is an optional projection layer from state and output + * and the “activation” is the function passed as the + * “fused_activation_function” argument (if not “NONE”). + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: 3, either time-major or batch-major. + * + * All input and output tensors must be of the same type. + * + * Inputs: + * * 0: The input (\f$x_t\f$). + * A 3-D tensor of shape: + * If time-major: [max_time, batch_size, input_size] + * If batch-major: [batch_size, max_time, input_size] + * where “max_time” is the number of timesteps (sequence length), + * “batch_size” corresponds to the batching dimension, and + * “input_size” is the size of the input. + * * 1: The input-to-input weights (\f$W_{xi}\f$). Optional. + * A 2-D tensor of shape [num_units, input_size], where “num_units” + * corresponds to the number of cell units. + * * 2: The input-to-forget weights (\f$W_{xf}\f$). + * A 2-D tensor of shape [num_units, input_size]. + * * 3: The input-to-cell weights (\f$W_{xc}\f$). + * A 2-D tensor of shape [num_units, input_size]. + * * 4: The input-to-output weights (\f$W_{xo}\f$). + * A 2-D tensor of shape [num_units, input_size]. + * * 5: The recurrent-to-input weights (\f$W_{hi}\f$). Optional. + * A 2-D tensor of shape [num_units, output_size], where “output_size” + * corresponds to either the number of cell units (i.e., “num_units”), + * or the second dimension of the “projection_weights”, if defined. + * * 6: The recurrent-to-forget weights (\f$W_{hf}\f$). + * A 2-D tensor of shape [num_units, output_size]. + * * 7: The recurrent-to-cell weights (\f$W_{hc}\f$). + * A 2-D tensor of shape [num_units, output_size]. + * * 8: The recurrent-to-output weights (\f$W_{ho}\f$). + * A 2-D tensor of shape [num_units, output_size]. + * * 9: The cell-to-input weights (\f$W_{ci}\f$). Optional. + * A 1-D tensor of shape [num_units]. + * * 10:The cell-to-forget weights (\f$W_{cf}\f$). Optional. + * A 1-D tensor of shape [num_units]. + * * 11:The cell-to-output weights (\f$W_{co}\f$). Optional. + * A 1-D tensor of shape [num_units]. + * * 12:The input gate bias (\f$b_i\f$). Optional. + * A 1-D tensor of shape [num_units]. + * * 13:The forget gate bias (\f$b_f\f$). + * A 1-D tensor of shape [num_units]. + * * 14:The cell bias (\f$b_c\f$). + * A 1-D tensor of shape [num_units]. + * * 15:The output gate bias (\f$b_o\f$). + * A 1-D tensor of shape [num_units]. + * * 16:The projection weights (\f$W_{proj}\f$). Optional. + * A 2-D tensor of shape [output_size, num_units]. + * * 17:The projection bias (\f$b_{proj}\f$). Optional. + * A 1-D tensor of shape [output_size]. + * * 18:The output state (in) (\f$h_{t-1}\f$). + * A 2-D tensor of shape [batch_size, output_size]. + * * 19:The cell state (in) (\f$C_{t-1}\f$). + * A 2-D tensor of shape [batch_size, num_units]. + * * 20:The activation function (\f$g\f$). + * A value indicating the activation function: + * + * * 21:The clipping threshold (\f$t_{cell}\f$) for the cell state, such + * that values are bound within [-cell_clip, cell_clip]. If set to 0.0 + * then clipping is disabled. + * * 22:The clipping threshold (\f$t_{proj}\f$) for the output from the + * projection layer, such that values are bound within + * [-proj_clip, proj_clip]. If set to 0.0 then clipping is disabled. + * * 23:Time-major if true, batch-major if false. + * * 24:The input layer normalization weights. Optional. + * A 1-D tensor of shape [num_units]. Used to rescale normalized inputs + * to activation at input gate. + * * 25:The forget layer normalization weights. Optional. + * A 1-D tensor of shape [num_units]. Used to rescale normalized inputs + * to activation at forget gate. + * * 26:The cell layer normalization weights. Optional. + * A 1-D tensor of shape [num_units]. Used to rescale normalized inputs + * to activation at cell gate. + * * 27:The output layer normalization weights. Optional. + * A 1-D tensor of shape [num_units]. Used to rescale normalized inputs + * to activation at output gate. + * + * Outputs: + * * 0: The output (\f$o_t\f$). + * A 3-D tensor of shape: + * If time-major: [max_time, batch_size, output_size] + * If batch-major: [batch_size, max_time, output_size] + * * 1: A tensor of shape [batch_size, output_size] containing a hidden + * state from the last time step in the sequence. This output is + * optional and can be omitted. If this output is present then + * output #2 must be present as well. + * Available since HAL version 1.3. + * * 2: A tensor of shape [batch_size, cell_size] containing a cell state + * from the last time step in the sequence. This output is optional + * and can be omitted. + * Available since HAL version 1.3. + */ + UNIDIRECTIONAL_SEQUENCE_LSTM = 92, + /** + * A recurrent neural network layer that applies a basic RNN cell to a + * sequence of inputs. + * + * This layer unrolls the input along the sequence dimension, and implements + * the following operation + * for each element in the sequence s = 1...sequence_length: + * outputs[s] = state = activation(inputs[s] * input_weights’ + state * + * recurrent_weights’ + bias) + * + * Where: + * * “input_weights” is a weight matrix that multiplies the inputs; + * * “recurrent_weights” is a weight matrix that multiplies the current + * “state” which itself is the output from the previous time step + * computation; + * * “bias” is a bias vector (added to each output vector in the batch); + * * “activation” is the function passed as the “fused_activation_function” + * argument (if not “NONE”). + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * The input tensors must all be the same type. + * + * Inputs: + * * 0: input. + * A 3-D tensor. The shape is defined by the input 6 (timeMajor). If + * it is set to 1, then the input has a shape [maxTime, batchSize, + * inputSize], otherwise the input has a shape [batchSize, maxTime, + * inputSize]. + * * 1: weights. + * A 2-D tensor of shape [numUnits, inputSize]. + * * 2: recurrent_weights. + * A 2-D tensor of shape [numUnits, numUnits]. + * * 3: bias. + * A 1-D tensor of shape [numUnits]. + * * 4: hidden state + * A 2-D tensor of shape [batchSize, numUnits]. Specifies a hidden + * state input for the first time step of the computation. + * * 5: fusedActivationFunction. + * A {@link FusedActivationFunc} value indicating the activation function. If + * “NONE” is specified then it results in a linear activation. + * * 6: timeMajor + * An {@link OperandType::INT32} scalar specifying the shape format + * of input and output tensors. Must be set to either 0 or 1. + * Outputs: + * * 0: output. + * A 3-D tensor. The shape is defined by the input 6 (timeMajor). If + * it is set to 1, then the output has a shape [maxTime, batchSize, + * numUnits], otherwise the output has a shape [batchSize, maxTime, + * numUnits]. + * * 1: A tensor of shape [batchSize, numUnits] containing hidden state + * from the last time step in the sequence. This output is optional + * and can be omitted. + * Available since HAL version 1.3. + */ + UNIDIRECTIONAL_SEQUENCE_RNN = 93, + /** + * Resizes images to given size using the nearest neighbor interpretation. + * + * Resized images must be distorted if their output aspect ratio is not the + * same as input aspect ratio. The corner pixels of output may not be the + * same as corner pixels of input. + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} (since HAL version 1.3) + * + * Supported tensor rank: 4, with "NHWC" or "NCHW" data layout. + * With the default data layout NHWC, the data is stored in the order of: + * [batch, height, width, channels]. Alternatively, the data layout could + * be NCHW, the data storage order of: [batch, channels, height, width]. + * + * Both resizing by shape and resizing by scale are supported. + * + * Inputs (resizing by shape): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. Zero batches is supported for this tensor. + * * 1: An {@link OperandType::INT32} scalar, specifying the output + * width of the output tensor. + * * 2: An {@link OperandType::INT32} scalar, specifying the output + * height of the output tensor. + * * 3: An {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * * 4: Align corners. An optional {@link OperandType::BOOL} + * scalar, default to false. If True, the centers of the 4 corner + * pixels of the input and output tensors are aligned, preserving the + * values at the corner pixels. + * Available since HAL version 1.3. + * * 5: Half pixel centers. An optional {@link OperandType::BOOL} + * scalar, default to false. If True, the pixel centers are assumed to + * be at (0.5, 0.5). This is the default behavior of image.resize in + * TF 2.0. If this parameter is True, then align_corners parameter + * must be False. + * Available since HAL version 1.3. + * + * Inputs (resizing by scale): + * * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying + * the input. Zero batches is supported for this tensor. + * * 1: A scalar, specifying width_scale, the scaling factor of the width + * dimension from the input tensor to the output tensor. The output + * width is calculated as new_width = floor(width * width_scale). + * The scalar must be of {@link OperandType::FLOAT16} if input0 is + * of {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} otherwise. + * * 2: A scalar, specifying height_scale, the scaling factor of the height + * dimension from the input tensor to the output tensor. The output + * height is calculated as new_height = floor(height * height_scale). + * The scalar must be of {@link OperandType::FLOAT16} if input0 is + * of {@link OperandType::TENSOR_FLOAT16} and of + * {@link OperandType::FLOAT32} otherwise. + * * 3: An {@link OperandType::BOOL} scalar, default to false. + * Set to true to specify NCHW data layout for input0 and output0. + * * 4: Align corners. An optional {@link OperandType::BOOL} + * scalar, default to false. If True, the centers of the 4 corner + * pixels of the input and output tensors are aligned, preserving the + * values at the corner pixels. + * Available since HAL version 1.3. + * * 5: Half pixel centers. An optional {@link OperandType::BOOL} + * scalar, default to false. If True, the pixel centers are assumed to + * be at (0.5, 0.5). This is the default behavior of image.resize in + * TF 2.0. If this parameter is True, then align_corners parameter + * must be False. + * Available since HAL version 1.3. + * + * Outputs: + * * 0: The output 4-D tensor, of shape + * [batches, new_height, new_width, depth]. + * For a {@link OperandType::TENSOR_QUANT8_ASYMM} and + * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} tensor, + * the scale and zeroPoint must be the same as input0. + */ + RESIZE_NEAREST_NEIGHBOR = 94, + /** + * Quantized version of {@link OperationType::LSTM}. + * + * The input and the output use asymmetric quantized types, while the rest + * use symmetric ones. + * + * Inputs: + * * 0: The input to the LSTM cell. + * Type: {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * Shape: [batchSize, inputSize] + * * 1: The input-to-input weights. Optional. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [numUnits, inputSize] + * * 2: The input-to-forget weights. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [numUnits, inputSize] + * * 3: The input-to-cell weights. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [numUnits, inputSize] + * * 4: The input-to-output weights. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [numUnits, inputSize] + * * 5: The recurrent-to-input weights. Optional. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [numUnits, outputSize] + * * 6: The recurrent-to-forget weights. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [numUnits, outputSize] + * * 7: The recurrent-to-cell weights. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [numUnits, outputSize] + * * 8: The recurrent-to-output weights. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [numUnits, outputSize] + * * 9: The cell-to-input weights (for peephole). Optional. + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [numUnits] + * * 10: The cell-to-forget weights (for peephole). Optional. + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [numUnits] + * * 11: The cell-to-output weights (for peephole). Optional. + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [numUnits] + * * 12: The input gate bias. Quantized with scale being the + * product of input and weights scales and zeroPoint equal to 0. + * Optional. + * Type: {@link OperandType::TENSOR_INT32} + * Shape: [numUnits] + * * 13: The forget gate bias. Quantized with scale being the + * product of input and weights scales and zeroPoint equal to 0. + * Type: {@link OperandType::TENSOR_INT32} + * Shape: [numUnits] + * * 14: The cell bias. Quantized with scale being the + * product of input and weights scales and zeroPoint equal to 0. + * Type: {@link OperandType::TENSOR_INT32} + * Shape: [numUnits] + * * 15: The output gate bias. Quantized with scale being the + * product of input and weights scales and zeroPoint equal to 0. + * Type: {@link OperandType::TENSOR_INT32} + * Shape: [numUnits] + * * 16: The projection weights. Optional. + * Type: {@link OperandType::TENSOR_QUANT8_SYMM} + * Shape: [outputSize, numUnits] + * * 17: The projection bias. Quantized with scale being the + * product of input and weights scales and zeroPoint equal to 0. + * Optional. + * Type: {@link OperandType::TENSOR_INT32} + * Shape: [outputSize] + * * 18: The output from the previous time step. + * Type: {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * Shape: [batchSize, outputSize] + * * 19: The cell state from the previous time step. + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [batchSize, numUnits] + * * 20: The input layer normalization weights. Used to rescale + * normalized inputs to activation at input gate. Optional. + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [numUnits] + * * 21: The forget layer normalization weights. Used to + * rescale normalized inputs to activation at forget gate. Optional. + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [numUnits] + * * 22: The cell layer normalization weights. Used to rescale + * normalized inputs to activation at cell gate. Optional. + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [numUnits] + * * 23: The output layer normalization weights. Used to + * rescale normalized inputs to activation at output gate. Optional. + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [numUnits] + * * 24: The cell clip. If provided the cell state is clipped + * by this value prior to the cell output activation. Optional. + * Type: {@link OperandType::FLOAT32}. + * * 25: The projection clip. If provided and projection is enabled, + * this is used for clipping the projected values. Optional. + * Type: {@link OperandType::FLOAT32}. + * * 26: The scale of the intermediate result of matmul, + * i.e. input to layer normalization, at input gate. + * Type: {@link OperandType::FLOAT32}. + * * 27: The scale of the intermediate result of matmul, + * i.e. input to layer normalization, at forget gate. + * Type: {@link OperandType::FLOAT32}. + * * 28: The scale of the intermediate result of matmul, + * i.e. input to layer normalization, at cell gate. + * Type: {@link OperandType::FLOAT32}. + * * 29: The scale of the intermediate result of matmul, + * i.e. input to layer normalization, at output gate. + * Type: {@link OperandType::FLOAT32}. + * * 30: The zero point of the hidden state, i.e. input to + * projection. + * Type: {@link OperandType::INT32}. + * * 31: The scale of the hidden state, i.e. input to + * projection. + * Type: {@link OperandType::FLOAT32}. + * + * Outputs: + * * 0: The output state (out). + * Type: {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * Shape: [batchSize, outputSize] + * * 1: The cell state (out). + * Type: {@link OperandType::TENSOR_QUANT16_SYMM} + * Shape: [batchSize, numUnits] + * * 2: The output. This is effectively the same as the current + * "output state (out)" value. + * Type: {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * Shape: [batchSize, outputSize] + */ + QUANTIZED_LSTM = 95, + /** + * Executes one of the two referenced subgraphs as determined by a boolean + * value. + * + * The inputs and outputs of the two referenced subgraphs must agree with the + * signature of this operation. That is, if the operation has (3 + n) inputs + * and m outputs, both subgraphs must have n inputs and m outputs with the same + * types, ranks, dimensions, scales, + * zeroPoints, and extraParams as the corresponding operation + * inputs and outputs. + * All of the operands mentioned must have fully specified dimensions. + * + * Inputs: + * * 0: A value of type {@link OperandType::TENSOR_BOOL8} and shape [1] + * that determines which of the two referenced subgraphs to execute. + * The operand must have fully specified dimensions. + * * 1: A {@link OperandType::SUBGRAPH} reference to the subgraph to be + * executed if the condition is true. + * * 2: A {@link OperandType::SUBGRAPH} reference to the subgraph to be + * executed if the condition is false. + * * 3 ~ (n + 2): Inputs to be passed to the subgraph selected for execution. + * + * Outputs: + * * 0 ~ (m - 1): Outputs produced by the selected subgraph. + */ + IF = 96, + /** + * Executes the body subgraph until the condition subgraph outputs false. + * + * The inputs to this operation are the condition subgraph, the body subgraph, + * and operand values for the first iteration of the loop. The values are + * implicitly split into three groups of input-output, state-only, and + * input-only values, as described below. + * + * The outputs of this operation are the final values of input-output + * operands. + * + * Both the condition and body subgraph receive (m + k + n) inputs. + * * The first m (m >= 1) inputs are input-output operands. For the first + * iteration, these are initialized from the corresponding inputs of the + * WHILE operation. In subsequent iterations, their values come from the + * corresponding outputs of the body subgraph produced during the previous + * iteration. + * * The next k (k >= 0) inputs are state-only operands. They are similar to + * the input-output operands, except that their values are no longer + * available after the loop terminates. + * * The last n (n >= 0) inputs are input-only operands. Their values come + * from the corresponding inputs of the WHILE operation. + * + * The body subgraph produces (m + k) outputs. + * * The first m outputs are input-output operands. They become the outputs + * of the WHILE operation when a termination condition is reached. + * * The last k outputs are state-only operands. Their values are no longer + * available after the loop terminates. + * + * The numbers m, k, and n are inferred by the driver as follows: + * m = (WHILE operation output count) + * k = (body subgraph output count) - m + * n = (body subgraph input count) - m - k + * + * The pseudo-code below illustrates the flow of a WHILE operation with + * inputs condition, body, initial_input_output, initial_state, input_only + * (m = 1, k = 1, n = 1): + * + * input_output = initial_input_output + * state = initial_state + * while condition(input_output, state, input_only): + * input_output, state = body(input_output, state, input_only) + * return input_output + * + * Inputs: + * * 0: A {@link OperandType::SUBGRAPH} reference to the condition + * subgraph. The subgraph must have (m + k + n) inputs with + * the same types, ranks, dimensions, + * scales, zeroPoints, and extraParams as the + * corresponding inputs of the WHILE operation and exactly one output + * of {@link OperandType::TENSOR_BOOL8} and shape [1]. + * All of the operands mentioned must have fully specified dimensions. + * * 1: A {@link OperandType::SUBGRAPH} reference to the body subgraph. + * The subgraph must have (m + k + n) inputs and (m + k) outputs with + * the same types, ranks, dimensions, + * scales, zeroPoints, and extraParams as the + * corresponding inputs and outputs of the WHILE operation. + * All of the operands mentioned must have fully specified dimensions. + * * (m inputs): Initial values for input-output operands. + * * (k inputs): Initial values for state-only operands. + * * (n inputs): Values for input-only operands. + * + * Outputs: + * * 0 ~ (m - 1): Outputs produced by the loop. + */ + WHILE = 97, + /** + * Computes exponential linear activation on the input tensor element-wise. + * + * The output is calculated using the following formula: + * + * ELU(x) = max(0, x) + min(0, alpha * (exp(x) - 1)) + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor, specifying the input. May be zero-sized. + * * 1: A scalar, specifying the alpha parameter. + * For input tensor of {@link OperandType::TENSOR_FLOAT16}, + * the alpha value must be of {@link OperandType::FLOAT16}. + * For input tensor of {@link OperandType::TENSOR_FLOAT32}, + * the alpha value must be of {@link OperandType::FLOAT32}. + * + * Outputs: + * * 0: The output tensor of same shape and type as input0. + */ + ELU = 98, + /** + * Computes hard-swish activation on the input tensor element-wise. + * + * Hard swish activation is introduced in + * https://arxiv.org/pdf/1905.02244.pdf + * + * The output is calculated using the following formula: + * + * h-swish(x) = x * max(0, min(6, (x + 3))) / 6 + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A tensor, specifying the input. May be zero-sized. + * + * Outputs: + * * 0: The output tensor of same shape and type as input0. + * Scale and zero point of this tensor may be different from the input + * tensor's parameters. + */ + HARD_SWISH = 99, + /** + * Creates a tensor filled with a scalar value. + * + * Supported output tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: A 1-D tensor, specifying the desired output tensor shape. + * * 1: A scalar, specifying the value to fill the output tensors with. + * For output tensor of {@link OperandType::TENSOR_FLOAT16}, + * the scalar must be of {@link OperandType::FLOAT16}. + * For output tensor of {@link OperandType::TENSOR_FLOAT32}, + * the scalar must be of {@link OperandType::FLOAT32}. + * For output tensor of {@link OperandType::TENSOR_INT32}, + * the scalar must be of {@link OperandType::INT32}. + * + * Outputs: + * * 0: The output tensor. + */ + FILL = 100, + /** + * Returns the rank of a tensor. + * + * The rank of a tensor is the number of dimensions in it. Also known as + * "order", "degree", "ndims". + * + * Supported tensor {@link OperandType}: + * * {@link OperandType::TENSOR_FLOAT16} + * * {@link OperandType::TENSOR_FLOAT32} + * * {@link OperandType::TENSOR_INT32} + * * {@link OperandType::TENSOR_QUANT8_ASYMM} + * * {@link OperandType::TENSOR_QUANT16_SYMM} + * * {@link OperandType::TENSOR_BOOL8} + * * {@link OperandType::TENSOR_QUANT8_SYMM_PER_CHANNEL} + * * {@link OperandType::TENSOR_QUANT16_ASYMM} + * * {@link OperandType::TENSOR_QUANT8_SYMM} + * * {@link OperandType::TENSOR_QUANT8_ASYMM_SIGNED} + * + * Supported tensor rank: from 1. + * + * Inputs: + * * 0: The input tensor. + * + * Outputs: + * * 0: A scalar of {@link OperandType::INT32}, specifying the rank + * of the input tensor. + */ + RANK = 101, +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/OutputShape.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/OutputShape.aidl new file mode 100644 index 0000000000..d206a2559c --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/OutputShape.aidl @@ -0,0 +1,33 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Describes the shape information of an output operand after execution. + */ +@VintfStability +parcelable OutputShape { + /** + * Dimensions of the operand. + */ + int[] dimensions; + /** + * Whether the provided buffer size is sufficient for the output. + */ + boolean isSufficient; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/PerformanceInfo.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/PerformanceInfo.aidl new file mode 100644 index 0000000000..6ee29c2502 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/PerformanceInfo.aidl @@ -0,0 +1,37 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Performance information for the reference workload. + * + * Used by a driver to report its performance characteristics. + */ +@VintfStability +parcelable PerformanceInfo { + /** + * Ratio of the time taken by the driver to execute the workload compared to the time the CPU + * would take for the same workload. A lower number is better. + */ + float execTime; + /** + * Ratio of the energy used by the driver compared to what the CPU would use for doing the same + * workload. A lower number is better. + */ + float powerUsage; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Priority.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Priority.aidl new file mode 100644 index 0000000000..fe87598829 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Priority.aidl @@ -0,0 +1,29 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Priority given to a prepared model for execution. + */ +@VintfStability +@Backing(type="int") +enum Priority { + LOW, + MEDIUM, + HIGH, +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Request.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Request.aidl new file mode 100644 index 0000000000..396ff30758 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Request.aidl @@ -0,0 +1,55 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.RequestArgument; +import android.hardware.neuralnetworks.RequestMemoryPool; + +/** + * Inputs to be sent to and outputs to be retrieved from a prepared model. + * + * A Request serves two primary tasks: + * 1) Provides the input and output data to be used when executing the model. + * 2) Specifies any updates to the input operand metadata that were left unspecified at model + * preparation time. + * + * An output must not overlap with any other output, with an input, or with an operand of lifetime + * CONSTANT_POOL. + */ +@VintfStability +parcelable Request { + /** + * Input data and information to be used in the execution of a prepared model. + * + * The index of the input corresponds to the index in Model.main.inputIndexes. + * E.g., input[i] corresponds to Model.main.inputIndexes[i]. + */ + RequestArgument[] inputs; + /** + * Output data and information to be used in the execution of a prepared model. + * + * The index of the output corresponds to the index in Model.main.outputIndexes. + * E.g., output[i] corresponds to Model.main.outputIndexes[i]. + */ + RequestArgument[] outputs; + /** + * A collection of memory pools containing operand data for both the inputs and the outputs to a + * model. + */ + RequestMemoryPool[] pools; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/RequestArgument.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/RequestArgument.aidl new file mode 100644 index 0000000000..e615fa62b1 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/RequestArgument.aidl @@ -0,0 +1,53 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.DataLocation; + +/** + * Metadata information specifying the location of the input or output data and any updates to the + * input or output operand. + */ +@VintfStability +parcelable RequestArgument { + /** + * If true, the argument does not have a value. This can be used for operations that take + * optional arguments. If true, the fields of location are set to 0 and the dimensions vector is + * left empty. + */ + boolean hasNoValue; + /** + * The location within one of the memory pools passed in the Request. + */ + DataLocation location; + /** + * Updated dimension information. + * + * If dimensions.size() > 0, dimension information was provided along with the argument. This + * can be the case for models that accept inputs of varying size. This can't change the rank, + * just the value of the dimensions that were unspecified in the model. If dimensions.size() > + * 0, then all dimensions must be specified here; and any dimension that was specified in the + * model must have the same value here. + * + * If the dimensions in the model are not fully specified, then they must be fully specified + * here, unless hasNoValue is set to true. If the dimensions in the model are fully specified, + * then either dimensions.size() may be 0, or the dimensions in the model must be identical to + * the dimensions here. + */ + int[] dimensions; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/RequestMemoryPool.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/RequestMemoryPool.aidl new file mode 100644 index 0000000000..166746d388 --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/RequestMemoryPool.aidl @@ -0,0 +1,36 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.Memory; + +/** + * A memory pool. + */ +@VintfStability +union RequestMemoryPool { + /** + * Specifies a client-managed shared memory pool. + */ + Memory pool; + /** + * Specifies a driver-managed buffer. It is the token returned from IDevice::allocate, and is + * specific to the IDevice object. + */ + int token; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Subgraph.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Subgraph.aidl new file mode 100644 index 0000000000..0a76285fca --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Subgraph.aidl @@ -0,0 +1,51 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +import android.hardware.neuralnetworks.Operand; +import android.hardware.neuralnetworks.Operation; + +/** + * An excerpt of the execution graph. + */ +@VintfStability +parcelable Subgraph { + /** + * All operands included in the subgraph. + */ + Operand[] operands; + /** + * All operations included in the subgraph. + * + * The operations are sorted into execution order. Every operand with lifetime SUBGRAPH_OUTPUT + * or TEMPORARY_VARIABLE must be written before it is read. + */ + Operation[] operations; + /** + * Input indexes of the subgraph. There must be at least one. + * + * Each value corresponds to the index of the operand in "operands". + */ + int[] inputIndexes; + /** + * Output indexes of the subgraph. There must be at least one. + * + * Each value corresponds to the index of the operand in "operands". + */ + int[] outputIndexes; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/SymmPerChannelQuantParams.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/SymmPerChannelQuantParams.aidl new file mode 100644 index 0000000000..8ae41a4d3e --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/SymmPerChannelQuantParams.aidl @@ -0,0 +1,33 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Parameters for TENSOR_QUANT8_SYMM_PER_CHANNEL operand. + */ +@VintfStability +parcelable SymmPerChannelQuantParams { + /** + * Array of scaling values for each channel. Each value must be greater than zero. + */ + float[] scales; + /** + * Index of the channel dimension + */ + int channelDim; +} diff --git a/neuralnetworks/aidl/android/hardware/neuralnetworks/Timing.aidl b/neuralnetworks/aidl/android/hardware/neuralnetworks/Timing.aidl new file mode 100644 index 0000000000..b04f74e4ee --- /dev/null +++ b/neuralnetworks/aidl/android/hardware/neuralnetworks/Timing.aidl @@ -0,0 +1,37 @@ +/* + * Copyright (C) 2020 The Android Open Source Project + * + * Licensed under the Apache License, Version 2.0 (the "License"); + * you may not use this file except in compliance with the License. + * You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + + +package android.hardware.neuralnetworks; + +/** + * Timing information measured during execution. Each time is a duration from the beginning of some + * task to the end of that task, including time when that task is not active (for example, preempted + * by some other task, or waiting for some resource to become available). + * + * Times are measured in nanoseconds. When a time is not available, it must be reported as -1. + */ +@VintfStability +parcelable Timing { + /** + * Execution time on device (not driver, which runs on host processor). + */ + long timeOnDevice; + /** + * Execution time in driver (including time on device). + */ + long timeInDriver; +}