common/cpu_operations/LSHProjection.cpp

/*
 * Copyright (C) 2017 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 "LSHProjection.h"

#include <utils/hash/farmhash.h>

#include <memory>

#include "CpuExecutor.h"
#include "LegacyUtils.h"
#include "Tracing.h"
#include "nnapi/Types.h"

namespace android {
namespace nn {

LSHProjection::LSHProjection(const Operation& operation, RunTimeOperandInfo* operands) {
    input_ = GetInput(operation, operands, kInputTensor);
    weight_ = GetInput(operation, operands, kWeightTensor);
    hash_ = GetInput(operation, operands, kHashTensor);

    type_ = static_cast<LSHProjectionType>(
            getScalarData<int32_t>(*GetInput(operation, operands, kTypeParam)));

    output_ = GetOutput(operation, operands, kOutputTensor);
}

bool LSHProjection::Prepare(const Operation& operation, RunTimeOperandInfo* operands,
                            Shape* outputShape) {
    // Check that none of the required inputs are omitted.
    constexpr int requiredInputs[] = {kHashTensor, kInputTensor, kTypeParam};
    for (const int requiredInput : requiredInputs) {
        NN_RET_CHECK(!IsNullInput(GetInput(operation, operands, requiredInput)))
                << "required input " << requiredInput << " is omitted";
    }
    NN_CHECK_EQ(NumOutputs(operation), 1);

    const RunTimeOperandInfo* hash = GetInput(operation, operands, kHashTensor);
    NN_CHECK_EQ(NumDimensions(hash), 2u);
    // Support up to 32 bits.
    NN_CHECK(SizeOfDimension(hash, 1) <= 32);

    const RunTimeOperandInfo* input = GetInput(operation, operands, kInputTensor);
    NN_CHECK(NumDimensions(input) >= 1);

    const auto& typeOperand = operands[operation.inputs[kTypeParam]];
    NN_RET_CHECK(typeOperand.length >= sizeof(int32_t));
    auto type = static_cast<LSHProjectionType>(getScalarData<int32_t>(typeOperand));
    switch (type) {
        case LSHProjectionType_SPARSE:
        case LSHProjectionType_SPARSE_DEPRECATED:
            NN_CHECK(NumInputsWithValues(operation, operands) == 3);
            outputShape->dimensions = {SizeOfDimension(hash, 0)};
            break;
        case LSHProjectionType_DENSE: {
            RunTimeOperandInfo* weight = GetInput(operation, operands, kWeightTensor);
            NN_CHECK_EQ(NumInputsWithValues(operation, operands), 4);
            NN_CHECK_EQ(NumDimensions(weight), 1u);
            NN_CHECK_EQ(SizeOfDimension(weight, 0), SizeOfDimension(input, 0));
            outputShape->dimensions = {SizeOfDimension(hash, 0) * SizeOfDimension(hash, 1)};
            break;
        }
        default:
            return false;
    }

    outputShape->type = OperandType::TENSOR_INT32;
    outputShape->offset = 0;
    outputShape->scale = 0.f;

    return true;
}

// Compute sign bit of dot product of hash(seed, input) and weight.
// NOTE: use float as seed, and convert it to double as a temporary solution
//       to match the trained model. This is going to be changed once the new
//       model is trained in an optimized method.
//
template <typename T>
int runningSignBit(const RunTimeOperandInfo* input, const RunTimeOperandInfo* weight, float seed) {
    double score = 0.0;
    int input_item_bytes = nonExtensionOperandSizeOfData(input->type, input->dimensions) /
                           SizeOfDimension(input, 0);
    char* input_ptr = (char*)(input->buffer);

    const size_t seed_size = sizeof(seed);
    const size_t key_bytes = seed_size + input_item_bytes;
    std::unique_ptr<char[]> key(new char[key_bytes]);

    for (uint32_t i = 0; i < SizeOfDimension(input, 0); ++i) {
        // Create running hash id and value for current dimension.
        memcpy(key.get(), &seed, seed_size);
        memcpy(key.get() + seed_size, input_ptr, input_item_bytes);

        int64_t hash_signature = farmhash::Fingerprint64(key.get(), key_bytes);
        double running_value = static_cast<double>(hash_signature);
        input_ptr += input_item_bytes;
        if (weight->lifetime == Operand::LifeTime::NO_VALUE) {
            score += running_value;
        } else {
            score += static_cast<double>(reinterpret_cast<T*>(weight->buffer)[i]) * running_value;
        }
    }

    return (score > 0) ? 1 : 0;
}

template <typename T>
void SparseLshProjection(LSHProjectionType type, const RunTimeOperandInfo* hash,
                         const RunTimeOperandInfo* input, const RunTimeOperandInfo* weight,
                         int32_t* out_buf) {
    int num_hash = SizeOfDimension(hash, 0);
    int num_bits = SizeOfDimension(hash, 1);
    for (int i = 0; i < num_hash; i++) {
        int32_t hash_signature = 0;
        for (int j = 0; j < num_bits; j++) {
            T seed = reinterpret_cast<T*>(hash->buffer)[i * num_bits + j];
            int bit = runningSignBit<T>(input, weight, static_cast<float>(seed));
            hash_signature = (hash_signature << 1) | bit;
        }
        if (type == LSHProjectionType_SPARSE_DEPRECATED) {
            *out_buf++ = hash_signature;
        } else {
            *out_buf++ = hash_signature + i * (1 << num_bits);
        }
    }
}

template <typename T>
void DenseLshProjection(const RunTimeOperandInfo* hash, const RunTimeOperandInfo* input,
                        const RunTimeOperandInfo* weight, int32_t* out_buf) {
    int num_hash = SizeOfDimension(hash, 0);
    int num_bits = SizeOfDimension(hash, 1);
    for (int i = 0; i < num_hash; i++) {
        for (int j = 0; j < num_bits; j++) {
            T seed = reinterpret_cast<T*>(hash->buffer)[i * num_bits + j];
            int bit = runningSignBit<T>(input, weight, static_cast<float>(seed));
            *out_buf++ = bit;
        }
    }
}

template <typename T>
bool LSHProjection::Eval() {
    NNTRACE_COMP("LSHProjection::Eval");

    int32_t* out_buf = reinterpret_cast<int32_t*>(output_->buffer);

    switch (type_) {
        case LSHProjectionType_DENSE:
            DenseLshProjection<T>(hash_, input_, weight_, out_buf);
            break;
        case LSHProjectionType_SPARSE:
        case LSHProjectionType_SPARSE_DEPRECATED:
            SparseLshProjection<T>(type_, hash_, input_, weight_, out_buf);
            break;
        default:
            return false;
    }
    return true;
}

template bool LSHProjection::Eval<float>();
template bool LSHProjection::Eval<_Float16>();

template int runningSignBit<float>(const RunTimeOperandInfo* input,
                                   const RunTimeOperandInfo* weight, float seed);
template int runningSignBit<_Float16>(const RunTimeOperandInfo* input,
                                      const RunTimeOperandInfo* weight, float seed);

template void SparseLshProjection<float>(LSHProjectionType type, const RunTimeOperandInfo* hash,
                                         const RunTimeOperandInfo* input,
                                         const RunTimeOperandInfo* weight, int32_t* outBuffer);
template void SparseLshProjection<_Float16>(LSHProjectionType type, const RunTimeOperandInfo* hash,
                                            const RunTimeOperandInfo* input,
                                            const RunTimeOperandInfo* weight, int32_t* outBuffer);

template void DenseLshProjection<float>(const RunTimeOperandInfo* hash,
                                        const RunTimeOperandInfo* input,
                                        const RunTimeOperandInfo* weight, int32_t* outBuffer);
template void DenseLshProjection<_Float16>(const RunTimeOperandInfo* hash,
                                           const RunTimeOperandInfo* input,
                                           const RunTimeOperandInfo* weight, int32_t* outBuffer);

}  // namespace nn
}  // namespace android