common/cpu_operations/Pack.cpp

/*
 * 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.
 */

#define LOG_TAG "Operations"

#include "Pack.h"

#include "OperationResolver.h"
#include "OperationsExecutionUtils.h"

#ifdef NN_INCLUDE_CPU_IMPLEMENTATION
#include <limits>
#include <vector>

#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wunused-parameter"
#pragma clang diagnostic ignored "-Wsign-compare"
#include <tensorflow/lite/kernels/internal/reference/reference_ops.h>
#pragma clang diagnostic pop

#include "CpuOperationUtils.h"
#endif  // NN_INCLUDE_CPU_IMPLEMENTATION

namespace android {
namespace nn {
namespace pack_op {

#ifdef NN_INCLUDE_CPU_IMPLEMENTATION
bool prepare(IOperationExecutionContext* context) {
    // All input tensors must have the same dimensions and be of rank 1 or higher.
    const Shape firstInputTensorShape = context->getInputShape(kInputFirstTensor);
    const uint32_t firstInputTensorRank = getNumberOfDimensions(firstInputTensorShape);
    NN_RET_CHECK_GE(firstInputTensorRank, 1U);
    for (uint32_t inputTensorNum = 1, inputTensorCount = context->getNumInputs() - 1;
         inputTensorNum < inputTensorCount; ++inputTensorNum) {
        NN_RET_CHECK(SameShape(firstInputTensorShape,
                               context->getInputShape(kInputFirstTensor + inputTensorNum)))
                << "Input tensor #" << inputTensorNum
                << " dimensions do not match input tensor #0 dimensions";
    }

    // Fetch the axis dimension value.
    const int32_t axisDimension = context->getInputValue<int32_t>(kInputAxisScalar);
    NN_RET_CHECK_GE(axisDimension, 0);
    NN_RET_CHECK_LT(uint32_t(axisDimension), firstInputTensorRank + 1);

    // TODO: http://b/78268320 validate that output shape is consistent with input rather than
    // blindly overwriting it. Output tensor is of rank 1 higher than input tensors.
    const uint32_t outputTensorRank = firstInputTensorRank + 1;
    // For the (j)th output dimension:
    // - If (j) is less than the axis dimension, the (j)th output dimension must match the (j)th
    // input dimension.
    // - If (j) is the axis dimension, the (j)th output dimension must equal the number of input
    // tensors.
    // - If (j) is greater than the axis dimension, the (j)th output dimension must match the
    // (j-1)th input dimension.
    Shape outputShape = context->getOutputShape(kOutputTensor);
    outputShape.dimensions.resize(outputTensorRank);
    for (int32_t j = 0; j < axisDimension; ++j) {
        outputShape.dimensions[j] = firstInputTensorShape.dimensions[j];
    }
    outputShape.dimensions[axisDimension] = context->getNumInputs() - 1;
    for (int32_t j = axisDimension + 1; j < int32_t(outputTensorRank); ++j) {
        outputShape.dimensions[j] = firstInputTensorShape.dimensions[j - 1];
    }
    return context->setOutputShape(kOutputTensor, outputShape);
}

bool packParams(IOperationExecutionContext* context, tflite::PackParams* params) {
    const int32_t axis = context->getInputValue<int32_t>(kInputAxisScalar);
    NN_RET_CHECK_LE(axis, std::numeric_limits<typeof(params->axis)>().max())
            << "axis value out of range";
    params->axis = axis;

    const uint32_t inputTensorCount = context->getNumInputs() - 1;
    NN_RET_CHECK_LE(inputTensorCount, std::numeric_limits<typeof(params->inputs_count)>().max())
            << "input count out of range";
    params->inputs_count = inputTensorCount;

    // Note that the NNAPI PACK operation specification requires all input
    // tensors and the output tensor to have the same zeroPoint and scale.
    const Shape tensorShape = context->getInputShape(kInputFirstTensor);

    const std::vector<int32_t> paramsInputZeroPoint(inputTensorCount, tensorShape.offset);
    params->input_zeropoint = paramsInputZeroPoint.data();
    const std::vector<float> paramsInputScale(inputTensorCount, tensorShape.scale);
    params->input_scale = paramsInputScale.data();
    params->output_zeropoint = tensorShape.offset;
    params->output_scale = tensorShape.scale;

    return true;
}

template <typename T>
bool pack(IOperationExecutionContext* context) {
    tflite::PackParams params;
    NN_RET_CHECK(packParams(context, &params));

    const uint32_t inputTensorCount = context->getNumInputs() - 1;

    // Note that the NNAPI PACK operation specification requires all input
    // tensors to have the same dimensions.
    const tflite::RuntimeShape inputTensorShapes =
            convertShapeToTflshape(context->getInputShape(kInputFirstTensor));
    const std::vector<const tflite::RuntimeShape*> inputShapesPtrs(inputTensorCount,
                                                                   &inputTensorShapes);

    std::vector<const T*> inputData(inputTensorCount);
    for (uint32_t inputTensorNum = 0; inputTensorNum < inputTensorCount; ++inputTensorNum) {
        inputData[inputTensorNum] = context->getInputBuffer<T>(kInputFirstTensor + inputTensorNum);
    }

    tflite::reference_ops::Pack(params, inputShapesPtrs.data(), inputData.data(),
                                convertShapeToTflshape(context->getOutputShape(kOutputTensor)),
                                context->getOutputBuffer<T>(kOutputTensor));
    return true;
}

bool execute(IOperationExecutionContext* context) {
    switch (context->getInputType(kInputFirstTensor)) {
        case OperandType::TENSOR_FLOAT16:
            return pack<_Float16>(context);
        case OperandType::TENSOR_FLOAT32:
            return pack<float>(context);
        case OperandType::TENSOR_QUANT8_ASYMM:
            return pack<uint8_t>(context);
        case OperandType::TENSOR_QUANT8_ASYMM_SIGNED:
            return pack<int8_t>(context);
        case OperandType::TENSOR_INT32:
            return pack<int32_t>(context);
        default:
            NN_RET_CHECK_FAIL() << "Unsupported tensor type for operation " << kOperationName;
    }
}
#endif  // NN_INCLUDE_CPU_IMPLEMENTATION

}  // namespace pack_op

NN_REGISTER_OPERATION_DEFAULT_VALIDATION(PACK, pack_op::prepare, pack_op::execute);

}  // namespace nn
}  // namespace android