* limitations under the License.
*/
-/* This HAL is a work in progress */
-
package android.hardware.neuralnetworks@1.0;
-// The types an operand can have.
-// These values are the same as found in the NeuralNetworks.h and NeuralNetworksOEM.h files.
+/**
+ * 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.
+ */
enum OperandType : int32_t {
- FLOAT32 = 0,
- INT32 = 1,
- UINT32 = 2,
- TENSOR_FLOAT32 = 3,
- TENSOR_INT32 = 4,
- TENSOR_QUANT8_ASYMM = 5,
-
- OEM = 10000,
- TENSOR_OEM_BYTE = 10001,
+ /**
+ * The following entries are used to declare scalars.
+ */
+ FLOAT32 = 0,
+ INT32 = 1,
+ UINT32 = 2,
+
+ /**
+ * The following entries are used to declare tensors.
+ */
+ TENSOR_FLOAT32 = 3,
+ TENSOR_INT32 = 4,
+
+ /**
+ * A tensor of 8 bit 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
+ * - zero_value: a 32 bit integer
+ *
+ * The formula is:
+ * real_value = (integer_value - zero_value) * scale.
+ */
+ TENSOR_QUANT8_ASYMM = 5,
+
+ /**
+ * The following entries are OEM specific operand types.
+ */
+ OEM = 10000,
+ TENSOR_OEM_BYTE = 10001,
};
-// The type of operations. Unlike the operation types found in the
-// NeuralNetworks.h and NeuralNetworksOEM.h files, these specify the data type they operate on.
-// This is done to simplify the work of drivers.
-// TODO: Currently they are the same. Add a conversion when finalizing the model.
+/**
+ * Operation types.
+ *
+ * The type of an operation in a model.
+ */
enum OperationType : int32_t {
- 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,
+ /**
+ * Adds two tensors, elment-wise.
+ *
+ * Takes two input tensors of identical type 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}
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * Supported tensor rank: up to 4
+ *
+ * Inputs:
+ * 0: A tensor.
+ * 1: A tensor of the same type, and compatible dimensions as input0.
+ * 2: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 0: The sum, a tensor of the same type as input0.
+ */
+ ADD = 0,
+
+ /**
+ * Performs a 2-D average pooling operation.
+ *
+ * The output dimensions are functions of the filter dimensions, stride, and padding.
+ *
+ * The values in output Tensor is computed as:
+ * output[batch, row, col, channel] =
+ * sum_{i, j}(input[batch, row + i, col + j, channel]) / sum(1)
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
+ * 1: An INT32 value, specifying the padding on the left, in the ‘width’ dimension.
+ * 2: An INT32 value, specifying the padding on the right,in the ‘width’ dimension.
+ * 3: An INT32 value, specifying the padding on the top, in the ‘height’ dimension.
+ * 4: An INT32 value, specifying the padding on the bottom, in the ‘height’ dimension.
+ * 5: An INT32 value, specifying the output stride in the ‘width’ dimension.
+ * 6: An INT32 value, specifying the output stride in the ‘height’ dimension.
+ * 7: An INT32 value, specifying the filter width.
+ * 8: An INT32 value, specifying the filter height.
+ * 9: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth].
+ */
+ AVERAGE_POOL_2D = 1,
+
+ /**
+ * Concatenates the input tensors along the given dimension.
+ *
+ * The input tensors must have identical type and the same dimensions except the
+ * dimension along the concatenation axis.
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: up to 4
+ *
+ * Inputs:
+ * 0 ~ n: The list on n input tensors, of shape [D0, D1, ..., Daxis(i), ..., Dm]
+ * n+1: An INT32 value, specifying the concatenation axis.
+ * n+2: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 0: The output, a tensor of the same type as the input tensors.
+ The output shape is [D0, D1, ..., sum(Daxis(i)), ..., Dm].
+ */
+ CONCATENATION = 2,
+
+ /**
+ * Performs an 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 output Tensor is computed as:
+ * output[batch, row, col, channel] =
+ * sum_{i, j} (
+ * input[batch, row + i, col + j, k] *
+ * filter[channel, row + i, col + j, k] +
+ * bias[channel]
+ * )
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 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.
+ * 2: A 1-D tensor, of shape [depth_out], specifying the bias.
+ * For input tensor of {@link OperandType::TENSOR_FLOAT32} type, the bias should
+ * also be of {@link OperandType::TENSOR_FLOAT32}.
+ * For input tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} type, the bias
+ * should be of {@link OperandType::TENSOR_INT32}.
+ * 3: An INT32 value, specifying the padding on the left, in the ‘width’ dimension.
+ * 4: An INT32 value, specifying the padding on the right,in the ‘width’ dimension.
+ * 5: An INT32 value, specifying the padding on the top, in the ‘height’ dimension.
+ * 6: An INT32 value, specifying the padding on the bottom, in the ‘height’ dimension.
+ * 7: An INT32 value, specifying the output stride in the ‘width’ dimension.
+ * 8: An INT32 value, specifying the output stride in the ‘height’ dimension.
+ * 9: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth_out].
+ */
+ CONV_2D = 3,
+
+ /**
+ * Performs an depthwise 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_in] containing
+ * in_channels 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 output Tensor is 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[di, dj, k, q]
+ * )
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 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.
+ * 2: A 1-D tensor, of shape [depth_out], specifying the bias.
+ * For input tensor of {@link OperandType::TENSOR_FLOAT32} type, the bias should
+ * also be of {@link OperandType::TENSOR_FLOAT32}.
+ * For input tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} type, the bias
+ * should be of {@link OperandType::TENSOR_INT32}.
+ * 3: An INT32 value, specifying the padding on the left, in the ‘width’ dimension.
+ * 4: An INT32 value, specifying the padding on the right,in the ‘width’ dimension.
+ * 5: An INT32 value, specifying the padding on the top, in the ‘height’ dimension.
+ * 6: An INT32 value, specifying the padding on the bottom, in the ‘height’ dimension.
+ * 7: An INT32 value, specifying the output stride in the ‘width’ dimension.
+ * 8: An INT32 value, specifying the output stride in the ‘height’ dimension.
+ * 9: An INT32 value, specifying the depthwise multiplier.
+ * 10: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth_out].
+ */
+ 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 types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
+ * 1: An INT32 value, specifying the block_size. block_size must be >=1 and
+ * block_size * block_size must be a divisor of the input depth.
+ *
+ * Ouputs:
+ * 0: The output 4-D tensor, of shape [batch, height*block_size, width*block_size,
+ * depth/(block_size*block_size)].
+ */
+ DEPTH_TO_SPACE = 5,
+
+ /**
+ * Dequantizes the input tensor.
+ *
+ * The formula is:
+ * output = (input - zero_value) * scale.
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: up to 4
+ *
+ * Inputs:
+ * 0: A tensor of type {@link OperandType::TENSOR_QUANT8_ASYMM}.
+ *
+ * Ouputs:
+ * 0: The output tensor of same shape as input0, but with type
+ {@link OperandType::TENSOR_FLOAT32}.
+ */
+ DEQUANTIZE = 6,
+
+ /**
+ * Looks up items from a given tensor.
+ *
+ * Each item in the output is a raw copy of the corresponding item in
+ * the input “values”. If the the given “lookup” indices are out of bounds,
+ * the op will fail and an error will be reported.
+ *
+ * Inputs:
+ * * 0: Values. An n-D tensor of any type X (where n >= 2). E.g., if n is 2,
+ * then the shape would be [lookup_dimension, values_dimension], where
+ * “lookup_dimension” corresponds to the indexing dimension in the lookup
+ * table, and “values_dimension” to the contents.
+ * * 1: Lookups. An 1-D tensor of type T, of shape [lookup_size], where
+ * “lookup_size” is the number of elements to look for, and each entry
+ * corresponds to the first dimension of the “values” tensor.
+ *
+ * Output:
+ * * 0: A n-D tensor of type X and the same rank and shape as the “values”
+ * tensor, except for the first dimension which has size “lookup_size”.
+ */
+ EMBEDDING_LOOKUP = 7,
+
+ /**
+ * Computes element-wise floor() on the input tensor.
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * Supported tensor rank: up to 4
+ *
+ * Inputs:
+ * 0: A tensor.
+ *
+ * Ouputs:
+ * 0: The output, a tensor of the same type and dimensions as input0.
+ */
+ 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 types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: up to 4.
+ *
+ * Inputs:
+ * 0: A tensor, specifying the input. If rank is greater than 2, then it gets flattened to
+ * a 2-D Tensor. The 2-D Tensor is handled as if dimensions corresponded to shape
+ * [batch_size, input_size], where “batch_size” corresponds to the batching dimension,
+ * and “input_size” is the size of the input.
+ * 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} type, the bias should
+ * also be of {@link OperandType::TENSOR_FLOAT32}.
+ * For input tensor of {@link OperandType::TENSOR_QUANT8_ASYMM} type, the bias
+ * should be of {@link OperandType::TENSOR_INT32}.
+ * 3: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 0: The output tensor, of shape [batch_size, num_units].
+ */
+ FULLY_CONNECTED = 9,
+
+ /**
+ * Looks up values of a hash table with given keys.
+ *
+ * Inputs:
+ * * 0: Lookups. A 1-D int32 tensor with shape [ k ].
+ * * 1: Keys. A 1-D int32 tensor with shape [ n ], *MUST* be sorted in
+ * ascending order.
+ * * 2: Values. A tensor with shape [ n … ].
+ *
+ * Outputs:
+ * * 0: Output. A tensor with shape [ k …].
+ * * 1: Hits. A uint8 tensor with shape [ k ] indicates whether the lookup
+ * hits or not.
+ */
+ HASHTABLE_LOOKUP = 10,
+
+ /**
+ * Applies L2 normalization along a the depth dimension.
+ *
+ * The values in output Tensor is computed as:
+ * output[batch, row, col, channel] =
+ * input[batch, row, col, channel] /
+ * sqrt(sum_{c} pow(input[batch, row, col, c], 2))
+ *
+ * For x with more dimensions, independently normalizes each 1-D slice along dimension dim.
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
+ *
+ * Ouputs:
+ * 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth].
+ */
+ 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 output Tensor is computed as:
+ * output[batch, row, col, channel] =
+ * sqrt(sum_{i, j} pow(input[batch, row + i, col + j, channel], 2) / sum(1))
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
+ * 1: An INT32 value, specifying the padding on the left, in the ‘width’ dimension.
+ * 2: An INT32 value, specifying the padding on the right,in the ‘width’ dimension.
+ * 3: An INT32 value, specifying the padding on the top, in the ‘height’ dimension.
+ * 4: An INT32 value, specifying the padding on the bottom, in the ‘height’ dimension.
+ * 5: An INT32 value, specifying the output stride in the ‘width’ dimension.
+ * 6: An INT32 value, specifying the output stride in the ‘height’ dimension.
+ * 7: An INT32 value, specifying the filter width.
+ * 8: An INT32 value, specifying the filter height.
+ * 9: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 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.
+ *
+ * In details:
+ * 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)
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
+ * 1: An INT32 value, specifying the radius of the normalization window.
+ * 2: A FLOAT32 value, specifying the bias, must not be zero.
+ * 3: A FLOAT32 value, specifying the scale factor, alpha.
+ * 4: A FLOAT32 value, specifying the exponent, beta.
+ *
+ * Ouputs:
+ * 0: The output tensor of same shape as input0.
+ */
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,
-
- OEM_OPERATION = 10000,
+
+ /**
+ * Computes sigmoid activation on the input tensor element-wise.
+ *
+ * In details:
+ * output = 1 / (1 + exp(-input))
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: up to 4.
+ *
+ * Inputs:
+ * 0: A tensor, specifying the input.
+ *
+ * Ouputs:
+ * 0: The output tensor of same shape as input0.
+ */
+ LOGISTIC = 14,
+
+ /**
+ * Projects an input to a bit vector via locality senstive hashing.
+ *
+ * Inputs:
+ * * 0: Hash functions. Dim.size == 2, DataType: Float.
+ * Tensor[0].Dim[0]: Number of hash functions.
+ * Tensor[0].Dim[1]: Number of seeds per hash functions.
+ * Tensor[0].Dim[1] <= 32 in sparse case.
+ *
+ * * 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(=1).
+ * Computed bit vector is considered to be sparse.
+ * Each output element is an int32 made up of multiple bits computed from
+ * hash functions.
+ *
+ * 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.
+ */
+ LSH_PROJECTION = 15,
+
+ /**
+ * Long short-term memory unit (LSTM) recurrent network layer.
+ *
+ * The default non-peephole implementation is based on:
+ * http://deeplearning.cs.cmu.edu/pdfs/Hochreiter97_lstm.pdf
+ * S. Hochreiter and J. Schmidhuber. "Long Short-Term Memory". Neural
+ * Computation, 9(8):1735-1780, 1997.
+ *
+ * The peephole implementation 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.
+ *
+ * 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 class has the following independently optional inputs:
+ * * If input gate (if CIFG): “input_to_forget_weights”,
+ * “recurrent_to_input_weights”, “cell_to_input_weights”, “input_gate_bias”.
+ * * If no peephole connections: “cell_to_input_weights”,
+ * “cell_to_forget_weights”, “cell_to_output_weights”.
+ * * If no projection layer: “projection_weights” and “projection_bias”.
+ * * If no projection bias: “projection_bias”.
+ *
+ * Supported tensor types:
+ * * {@link OperandType::TENSOR_FLOAT32}
+ *
+ * Inputs:
+ * * 0: Input.
+ * A 2-D tensor of type T, of shape [batch_size, input_size], where
+ * “batch_size” corresponds to the batching dimension, and “input_size”
+ * is the size of the input.
+ * * 1: input_to_input_weights.
+ * A 2-D tensor of type T, of shape [num_units, input_size], where
+ * “num_units” corresponds to the number of cell units.
+ * * 2: input_to_forget_weights.
+ * A 2-D tensor of type T, of shape [num_units, input_size].
+ * * 3: input_to_cell_weights.
+ * A 2-D tensor of type T, of shape [num_units, input_size].
+ * * 4: input_to_output_weights.
+ * A 2-D tensor of type T, of shape [num_units, input_size].
+ * * 5: recurrent_to_input_weights.
+ * A 2-D tensor of type T, 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: recurrent_to_forget_weights.
+ * A 2-D tensor of type T, of shape [num_units, output_size].
+ * * 7: recurrent_to_cell_weights.
+ * A 2-D tensor of type T, of shape [num_units, output_size].
+ * * 8: recurrent_to_output_weights.
+ * A 2-D tensor of type T, of shape [num_units, output_size].
+ * * 9: cell_to_input_weights.
+ * A 1-D tensor of type T, of shape [num_units].
+ * * 10:cell_to_forget_weights.
+ * A 1-D tensor of type T, of shape [num_units].
+ * * 11:cell_to_output_weights.
+ * A 1-D tensor of type T, of shape [num_units].
+ * * 12:input_gate_bias.
+ * A 1-D tensor of type T, of shape [num_units].
+ * * 13:forget_gate_bias.
+ * A 1-D tensor of type T, of shape [num_units].
+ * * 14:cell_bias.
+ * A 1-D tensor of type T, of shape [num_units].
+ * * 15:output_gate_bias.
+ * A 1-D tensor of type T, of shape [num_units].
+ * * 16:projection_weights.
+ * A 2-D tensor of type T, of shape [output_size, num_units].
+ * * 17:projection_bias.
+ * A 1-D tensor of type T, of shape [output_size].
+ *
+ * Parameters:
+ * * 18:fused_activation_function.
+ * An (optional) ActivationFunctionType indicating the activation
+ * function.
+ * If “NONE” is specified then it results in a linear activation.
+ * * 19:cell_clip.
+ * A 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.
+ * * 20:proj_clip.
+ * A 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.
+ *
+ * Outputs:
+ * * 0: scratch_buffer.
+ * A 3-D tensor of type T, of shape [batch_size, num_cell, 4].
+ * * 1: output_state.
+ * A 2-D tensor of type T, of shape [batch_size, output_size].
+ * * 2: cell_state.
+ * A 2-D tensor of type T, of shape [batch_size, num_units].
+ * * 3: output.
+ * A 2-D tensor of type T, of shape [batch_size, output_size]. This is
+ * effectively the same as the current “output_state” 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 output Tensor is computed as:
+ * output[batch, row, col, channel] =
+ * max_{i, j} (input[batch, row + i, col + j, channel])
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
+ * 1: An INT32 value, specifying the padding on the left, in the ‘width’ dimension.
+ * 2: An INT32 value, specifying the padding on the right,in the ‘width’ dimension.
+ * 3: An INT32 value, specifying the padding on the top, in the ‘height’ dimension.
+ * 4: An INT32 value, specifying the padding on the bottom, in the ‘height’ dimension.
+ * 5: An INT32 value, specifying the output stride in the ‘width’ dimension.
+ * 6: An INT32 value, specifying the output stride in the ‘height’ dimension.
+ * 7: An INT32 value, specifying the filter width.
+ * 8: An INT32 value, specifying the filter height.
+ * 9: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 0: The output 4-D tensor, of shape [batches, out_height, out_width, depth].
+ */
+ MAX_POOL_2D = 17,
+
+ /**
+ * Multiplies two tensors, elment-wise.
+ *
+ * Takes two input tensors of identical type 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.
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * Supported tensor rank: up to 4
+ *
+ * Inputs:
+ * 0: A tensor.
+ * 1: A tensor of the same type, and compatible dimensions as input0.
+ * 2: An INT32 value, and has to be one of the {@link FusedActivationFunc} values.
+ * Specifies the activation to invoke on the result of each addition.
+ *
+ * Ouputs:
+ * 0: The product, a tensor of the same type as input0.
+ */
+ MUL = 18,
+
+ /**
+ * Computes rectified linear activation on the input tensor element-wise.
+ *
+ * In details:
+ * output = max(0, input)
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: up to 4.
+ *
+ * Inputs:
+ * 0: A tensor, specifying the input.
+ *
+ * Ouputs:
+ * 0: The output tensor of same shape as input0.
+ */
+ RELU = 19,
+
+ /**
+ * Computes rectified linear 1 activation on the input tensor element-wise.
+ *
+ * In details:
+ * output = min(1.f, max(-1.f, input))
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: up to 4.
+ *
+ * Inputs:
+ * 0: A tensor, specifying the input.
+ *
+ * Ouputs:
+ * 0: The output tensor of same shape as input0.
+ */
+ RELU1 = 20,
+
+ /**
+ * Computes rectified linear 6 activation on the input tensor element-wise.
+ *
+ * In details:
+ * output = min(6, max(0, input))
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: up to 4.
+ *
+ * Inputs:
+ * 0: A tensor, specifying the input.
+ *
+ * Ouputs:
+ * 0: The output tensor of same shape 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 types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: up to 4.
+ *
+ * Inputs:
+ * 0: A tensor, specifying the tensor to be reshaped.
+ * 1: A 1-D tensor of type {@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.
+ *
+ * Ouputs:
+ * 0: The output tensor, of shape specified by the input shape.
+ */
+ RESHAPE = 22,
+
+ /**
+ * Resizes images to given size using the bilinear interpretation.
+ *
+ * Resized images will be distorted if their original aspect ratio is not the same as input.
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 0: A 4-D tensor, of shape [batches, height, width, depth], specifying the input.
+ * 1: An INT32 value, specifying the output width of the output tensor.
+ * 2: An INT32 value, specifying the output height of the output tensor.
+ *
+ * Ouputs:
+ * 0: The output 4-D tensor, of shape [batches, new_height, new_width, depth].
+ */
+ 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 types:
+ * * {@link OperandType::TENSOR_FLOAT32}
+ *
+ * Inputs:
+ * * 0: input.
+ * A 2-D tensor of type T, 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 type T, of shape [num_units, input_size], where
+ * “num_units” corresponds to the number of units.
+ * * 2: recurrent_weights.
+ * A 2-D tensor of type T, of shape [num_units, num_units], with columns
+ * corresponding to the weights from each unit.
+ * * 3: bias.
+ * A 1-D tensor of type T, of shape [num_units].
+ *
+ * For FLOAT32 input tensor, bias must also be FLOAT32.
+ * For UINT8 input tensor, bias must be INT32.
+ *
+ * Parameters
+ * * 4: fused_activation_function.
+ * An (optional) ActivationFunctionType indicating the activation
+ * function. If “NONE” is specified then it results in a linear
+ * activation.
+ *
+ * * 5: Hidden state.
+ * A 2-D tensor of type T, of shape [batch_size, num_units].
+ *
+ * Outputs:
+ * * 0: output.
+ * A 2-D tensor of type T, 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.
+ *
+ * In details:
+ * output[batch, i] =
+ * exp((input[batch, i] - max(input[batch, :])) * beta) /
+ * sum_{k}{exp((input[batch, k] - max(input[batch, :])) * beta)}
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: 2 or 4.
+ *
+ * Inputs:
+ * 0: A 2-D or 4-D tensor, specifying the tensor to be reshaped.
+ * 1: A FLOAT32 value, specifying the scaling factor for the exponent, beta.
+ *
+ * Ouputs:
+ * 0: The output tensor of same shape as input0.
+ */
+ 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 types: {@link OperandType::TENSOR_FLOAT32}
+ * {@link OperandType::TENSOR_QUANT8_ASYMM}
+ * Supported tensor rank: 4, with "NHWC" data layout.
+ *
+ * Inputs:
+ * 0: A 4-D tensor, of shape [batches, height, width, depth_in], specifying the input.
+ * 1: An INT32 value, specifying the block_size. block_size must be >=1 and
+ * block_size must be a divisor of both the input height and width.
+ *
+ * Ouputs:
+ * 0: The output 4-D tensor, of shape [batch, height/block_size, width/block_size,
+ * depth*block_size*block_size].
+ */
+ 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, "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 types:
+ * * {@link OperandType::TENSOR_FLOAT32}
+ *
+ * Inputs:
+ * * 0: input.
+ * A 2-D tensor of type T, 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 type T, of shape [num_units, input_size], where
+ * “num_units” corresponds to the number of units.
+ * * 2: weights_time.
+ * A 2-D tensor of type T, of shape [num_units, memory_size], where
+ * “memory_size” corresponds to the fixed-size of the memory.
+ * * 3: bias.
+ * A optional 1-D tensor of type T, of shape [num_units].
+ *
+ * For FLOAT32 input tensor, bias must also be FLOAT32.
+ * For UINT8 input tensor, bias must be INT32.
+ *
+ * Parameters:
+ * * 4: rank.
+ * The rank of the SVD approximation.
+ * * 5: fused_activation_function.
+ * An (optional) ActivationFunctionType indicating the activation function.
+ * If “NONE” is specified then it results in a linear activation.
+ *
+ * Outputs:
+ * * 0: state.
+ * A 2-D tensor of type T, of shape [batch_size, (memory_size - 1) * num_units * rank].
+ * * 1: output.
+ * A 2-D tensor of type T, of shape [batch_size, num_units].
+ */
+ SVDF = 27,
+
+ /**
+ * Computes hyperbolic tangent of input tensor element-wise.
+ *
+ * In details:
+ * output = tanh(input)
+ *
+ * Supported tensor types: {@link OperandType::TENSOR_FLOAT32}
+ * Supported tensor rank: up to 4.
+ *
+ * Inputs:
+ * 0: A tensor, specifying the input.
+ *
+ * Ouputs:
+ * 0: The output tensor of same shape as input0.
+ */
+ TANH = 28,
+
+ /**
+ * OEM specific operation.
+ *
+ * This operation is OEM specific. It should only be used for OEM applications.
+ */
+ OEM_OPERATION = 10000,
};
-// Fused activation functions
+/**
+ * Fused activation function types.
+ */
enum FusedActivationFunc : int32_t {
NONE = 0,
RELU = 1,
RELU6 = 3,
};
-// How an operand is used.
+/**
+ * How an operand is used.
+ */
enum OperandLifeTime : int32_t {
- // The operand is internal to the model. It's created by an operation
- // and consumed by other operations.
+ /**
+ * The operand is internal to the model. It's created by an operation
+ * and consumed by other operations.
+ */
TEMPORARY_VARIABLE,
- // The operand is an input of the model. An operand can't be both
- // input and output of a model.
+
+ /**
+ * The operand is an input of the model. An operand can't be both
+ * input and output of a model.
+ */
MODEL_INPUT,
- // The operand is an output of the model.
+
+ /**
+ * The operand is an output of the model.
+ */
MODEL_OUTPUT,
- // The operand is a constant found in Model.operandValues.
+
+ /**
+ * The operand is a constant found in Model.operandValues.
+ */
CONSTANT_COPY,
- // The operand is a constant that was specified via a Memory object.
- CONSTANT_REFERENCE
+
+ /**
+ * The operand is a constant that was specified via a Memory object.
+ */
+ CONSTANT_REFERENCE,
};
-// Status of a device.
+/**
+ * Status of a device.
+ */
enum DeviceStatus : int32_t {
AVAILABLE,
BUSY,
OFFLINE,
- UNKNOWN // Do we need this?
-};
-
-// For the reference workload
-// Used by a driver to report its performance characteristics.
-// TODO revisit the data types and scales.
-struct PerformanceInfo {
- float execTime; // in nanoseconds
- float powerUsage; // in picoJoules
+ UNKNOWN,
};
+/**
+ * A typed operation.
+ */
struct OperationTuple {
- // The type of operation.
+ /**
+ * The type of operation.
+ */
OperationType operationType;
- // The input data type of operation.
+
+ /**
+ * The input data type of operation.
+ */
OperandType operandType;
};
-// The capabilities of a driver.
+/**
+ * Performance information for the reference workload.
+ *
+ * Used by a driver to report its performance characteristics.
+ */
+struct PerformanceInfo {
+ /**
+ * Execution time in nanoseconds.
+ */
+ float execTime;
+
+ /**
+ * Power usage in picoJoules.
+ */
+ float powerUsage;
+};
+
+/**
+ * The capabilities of a driver.
+ */
struct Capabilities {
+ /**
+ * A collection of typed operations supported by the driver.
+ */
vec<OperationTuple> supportedOperationTuples;
+
+ /**
+ * Indicates whether a driver caches its prepared model for reuse the next
+ * time the application begins. This is useful because the model may have
+ * been prepared in a previous run.
+ *
+ * True if caching is supported, false otherwise.
+ */
bool cachesCompilation;
- // TODO revisit the data types and scales.
+
+ /**
+ * Driver performance when operating on float32 data.
+ */
PerformanceInfo float32Performance;
+
+ /**
+ * Driver performance when operating on asymmetric 8-bit quantized data.
+ */
PerformanceInfo quantized8Performance;
};
-// Describes the location of a data object.
+/**
+ * Describes the location of a data object.
+ */
struct DataLocation {
- // The index of the memory pool where this location is found.
- // Two special values can also be used. See the LOCATION_* constants above.
+ /**
+ * The index of the memory pool where this location is found.
+ */
uint32_t poolIndex;
- // Offset in bytes from the start of the pool.
+
+ /**
+ * Offset in bytes from the start of the pool.
+ */
uint32_t offset;
- // The length of the data, in bytes.
+
+ /**
+ * The length of the data in bytes.
+ */
uint32_t length;
};
+/**
+ * Describes one operand of the model's graph.
+ */
struct Operand {
+ /**
+ * Data type of the operand.
+ */
OperandType type;
+
+ /**
+ * Dimensions of the operand.
+ */
vec<uint32_t> dimensions;
- // The number of operations that uses this operand as input.
- // TODO It would be nice to track the actual consumers, e.g. vec<uint32_t> consumers;
+ /**
+ * The number of operations that use this operand as input.
+ */
uint32_t numberOfConsumers;
+ /**
+ * Quantized scale of the operand.
+ *
+ * Only applicable if the operand is of type TENSOR_QUANT8_ASYMM.
+ */
float scale;
+
+ /**
+ * Quantized zero-point offset of the operand.
+ *
+ * Only applicable if the operand is of type TENSOR_QUANT8_ASYMM.
+ */
int32_t zeroPoint;
- // How the operand is used.
+ /**
+ * How the operand is used.
+ */
OperandLifeTime lifetime;
- // Where to find the data for this operand.
- // If the lifetime is TEMPORARY_VARIABLE, MODEL_INPUT, or MODEL_OUTPUT:
- // - All the fields will 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_REFERENCE:
- // - location.poolIndex is set.
- // - location.offset is the offset in bytes into the specified pool.
- // - location.length is set.
+ /**
+ * Where to find the data for this operand.
+ * If the lifetime is TEMPORARY_VARIABLE, MODEL_INPUT, or MODEL_OUTPUT:
+ * - All the fields will 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_REFERENCE:
+ * - location.poolIndex is set.
+ * - location.offset is the offset in bytes into the specified pool.
+ * - location.length is set.
+ */
DataLocation location;
};
-// Describes one operation of the graph.
+/**
+ * Describes one operation of the model's graph.
+ */
struct Operation {
- // The tuple describing the operation type and input type.
+ /**
+ * The tuple describing the operation type and input type.
+ */
OperationTuple opTuple;
- // Describes the table that contains the indexes of the inputs of the
- // operation. The offset is the index in the operandIndexes table.
+
+ /**
+ * Describes the table that contains the indexes of the inputs of the
+ * operation. The offset is the index in the operandIndexes table.
+ */
vec<uint32_t> inputs;
- // Describes the table that contains the indexes of the outputs of the
- // operation. The offset is the index in the operandIndexes table.
+
+ /**
+ * Describes the table that contains the indexes of the outputs of the
+ * operation. The offset is the index in the operandIndexes table.
+ */
vec<uint32_t> outputs;
};
+/**
+ * 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
+ * might not be known is the shape of the input tensors.
+ */
struct Model {
+ /**
+ * All operands included in the model.
+ */
vec<Operand> operands;
+
+ /**
+ * All operations included in the model.
+ *
+ * The operations are sorted into execution order.
+ */
vec<Operation> operations;
+
+ /**
+ * Input indexes of the model.
+ *
+ * Each value corresponds to the index of the operand in "operands".
+ */
vec<uint32_t> inputIndexes;
+
+ /**
+ * Output indexes of the model.
+ *
+ * Each value corresponds to the index of the operand in "operands".
+ */
vec<uint32_t> outputIndexes;
+
+ /**
+ * A byte buffer containing operand data that were copied into the model.
+ */
vec<uint8_t> operandValues;
+
+ /**
+ * A collection of shared memory pools containing operand data that were
+ * registered by the model.
+ */
vec<memory> pools;
};
+/**
+ * Metadata information specifying the location of the input or output data and
+ * any updates to the input or output operand.
+ */
struct RequestArgument {
- // The location within one of the memory pools
+ /**
+ * The location within one of the memory pools passed in the Request.
+ */
DataLocation location;
- // 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.
+
+ /**
+ * 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.
+ */
vec<uint32_t> dimensions;
};
+/**
+ * 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.
+ */
struct 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.inputIndexes.
+ * E.g., input[i] corresponds to Model.inputIndexes[i].
+ */
vec<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.outputIndexes.
+ * E.g., output[i] corresponds to Model.outputIndexes[i].
+ */
vec<RequestArgument> outputs;
+
+ /**
+ * A collection of shared memory pools containing operand data for both the
+ * inputs and the outputs to a model.
+ */
vec<memory> pools;
};
+/**
+ * Return status of a function.
+ */
enum ErrorStatus : int32_t {
NONE,
DEVICE_UNAVAILABLE,