TheRock/rocm-libraries/projects/miopen/test/rnn_util.hpp

/*******************************************************************************
 *
 * MIT License
 *
 * Copyright (c) 2023 Advanced Micro Devices, Inc.
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to deal
 * in the Software without restriction, including without limitation the rights
 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
 * copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 *
 *******************************************************************************/

#ifndef MIOPEN_RNN_UTIL_H_
#define MIOPEN_RNN_UTIL_H_

#include <cfloat>
#include <cmath>
#include <initializer_list>
#include <set>
#include <vector>
#include <cstdlib>
#include <numeric>

#include "gemm.hpp"
#include "random.hpp"

#include <miopen/tensor.hpp>

// complexity O(NlogN)
inline std::vector<int> GetReverseOrderIndex(const std::vector<int>& base_index)
{
    std::vector<int> reverse_index(base_index.size());
    unsigned next_rev_index = 0;
    for(auto id : base_index)
        reverse_index[id] = next_rev_index++;
    return reverse_index;
};

inline std::vector<int> GetSamplesIndexDescendingOrder(const std::vector<size_t>& unsorted_seq_lens)
{
    const auto sample_count = unsorted_seq_lens.size();

    std::vector<int> index_v(sample_count);
    std::iota(index_v.begin(), index_v.end(), 0);

    auto seq_len_cmp = [&unsorted_seq_lens](unsigned a_id, unsigned b_id) {
        return unsorted_seq_lens[a_id] > unsorted_seq_lens[b_id];
    };

    std::stable_sort(index_v.begin(), index_v.end(), seq_len_cmp);

    return index_v;
}

template <typename Tgpu>
inline void HiddenTensorReorder(const std::vector<Tgpu>& src_array,
                                std::vector<Tgpu>& dst_array,
                                const std::vector<int>& batch_order,
                                const std::vector<size_t> hid_len,
                                bool is_dst_direct_order)
{
    const size_t copy_size = hid_len[2];

    const size_t batch_stride = hid_len[2];
    const size_t layer_stride = batch_stride * hid_len[1];

    for(size_t batch_id = 0; batch_id < hid_len[1]; batch_id++)
    {
        const auto src_batch_off =
            batch_stride * (is_dst_direct_order ? batch_order[batch_id] : batch_id);
        const auto dst_batch_off =
            batch_stride * (is_dst_direct_order ? batch_id : batch_order[batch_id]);

        for(size_t layer_id = 0; layer_id < hid_len[0]; layer_id++)
        {
            const auto dst_offset = dst_batch_off + layer_id * layer_stride;
            const auto src_offset = src_batch_off + layer_id * layer_stride;

            std::copy(src_array.begin() + src_offset,
                      src_array.begin() + src_offset + copy_size,
                      dst_array.begin() + dst_offset);
        }
    }
}

inline void createTensorDescArray(std::vector<miopen::TensorDescriptor>& td,
                                  std::vector<miopenTensorDescriptor_t>& ptd,
                                  const std::vector<int> bs,
                                  const int secondDim,
                                  miopenDataType_t dataType)
{

    std::transform(bs.begin(), bs.end(), std::back_inserter(td), [&](int x) {
        return miopen::TensorDescriptor(
            dataType, {static_cast<std::size_t>(x), static_cast<std::size_t>(secondDim)});
    });
    std::transform(td.begin(), td.end(), std::back_inserter(ptd), [](miopen::TensorDescriptor& x) {
        return &x;
    });
}

inline std::tuple<size_t, size_t>
GetTempPackedBuffersSize(std::vector<int> batchs, int in_vec, int out_vec)
{
    size_t total_batch = std::accumulate(batchs.begin(), batchs.end(), 0ULL);

    size_t in_buff_size  = total_batch * in_vec;
    size_t out_buff_size = total_batch * out_vec;
    return {in_buff_size, out_buff_size};
}

inline size_t getSuperTensorSize(const std::vector<int>& bs,
                                 int seqLength,
                                 int inputSize,
                                 int hiddenSize,
                                 int maxPaddingVal,
                                 bool isBidirect,
                                 bool isInput,
                                 bool isPadded)
{
    return (isPadded //
                ? static_cast<size_t>(seqLength) * maxPaddingVal
                : std::accumulate(bs.begin(), bs.end(), 0ULL)) //
           * (isInput                                          //
                  ? static_cast<size_t>(inputSize)
                  : static_cast<size_t>(hiddenSize) * (isBidirect ? 2 : 1));
}

template <typename Tgpu>
void ChangeDataPadding(const std::vector<Tgpu>& src_array,
                       std::vector<Tgpu>& dst_array,
                       const std::vector<int>& batch_list,
                       int max_batch,
                       int sample_size,
                       bool is_src_packed)
{
    auto seq_len = batch_list.size();

    auto scr_ptr = &src_array[0];
    auto dst_ptr = &dst_array[0];

    for(int seq_id = 0; seq_id < seq_len; seq_id++)
    {
        auto packed_size = batch_list[seq_id] * sample_size;

        std::copy(scr_ptr, scr_ptr + packed_size, dst_ptr);

        if(is_src_packed)
        {
            dst_ptr += max_batch * sample_size;
            scr_ptr += packed_size;
        }
        else
        {
            scr_ptr += max_batch * sample_size;
            dst_ptr += packed_size;
        }
    }
}

// RNN VANILLA configs
inline std::vector<int> get_rnn_num_layers() { return {{1, 3}}; }

inline std::vector<int> get_rnn_batchSize() { return {{1, 17}}; }

inline std::vector<int> get_rnn_seq_len() { return {{1, 3, 51}}; }

inline std::vector<int> get_rnn_vector_len() { return {31}; }

inline std::vector<int> get_rnn_hidden_size() { return {127}; }

// LSTM configs
inline std::vector<int> get_lstm_num_layers() { return {{1, 3}}; }

inline std::vector<int> get_lstm_batchSize() { return {{1, 17}}; }

inline std::vector<int> get_lstm_seq_len() { return {{1, 25}}; }

inline std::vector<int> get_lstm_vector_len() { return {17}; }

inline std::vector<int> get_lstm_hidden_size() { return {67}; }

// GRU configs
inline std::vector<int> get_gru_num_layers() { return {{1, 3}}; }

inline std::vector<int> get_gru_batchSize() { return {{1, 17}}; }

inline std::vector<int> get_gru_seq_len() { return {{1, 23}}; }

inline std::vector<int> get_gru_vector_len() { return {13}; }

inline std::vector<int> get_gru_hidden_size() { return {67}; }

inline std::vector<std::vector<int>> generate_batchSeq(const int batchSize, const int seqLength)
{

    static constexpr int modval = 3;

    int currentval = batchSize;
    std::vector<int> batchSeq;
    batchSeq.reserve(seqLength);
    for(int i = 0; i < seqLength; i++)
    {
        if(i > 0)
        {
            int nvalue = currentval - prng::gen_0_to_B(modval);
            currentval = (nvalue < 1) ? 1 : nvalue;
            // printf("current value: %d\n", currentval);
        }
        // printf("adding a value to batch sequence: %d\n", currentval);
        batchSeq.push_back(currentval);
    }
    return {batchSeq};
}

inline int sumvc(const std::vector<int>& x) { return std::accumulate(x.begin(), x.end(), 0); }

template <typename T>
inline T activfunc(T x, int actvf)
{
    T alpha = static_cast<T>(1), beta0 = static_cast<T>(0), beta1 = static_cast<T>(1);
    if(actvf == 0)
    {
        return (x > 0) ? x : x * beta0;
    }
    else if(actvf == 2)
    {
        return static_cast<T>(1 / (1 + std::exp(-x)));
    }
    return static_cast<T>(alpha * std::tanh(beta1 * x));
}

template <typename T>
inline T dervactivfunc(T x, int actvf)
{
    if(actvf == 0)
    {
        return static_cast<T>(x > 0 ? 1 : 0);
    }
    else if(actvf == 2)
    {
        return static_cast<T>(std::exp(-x) / (1 + std::exp(-x)) / (1 + std::exp(-x)));
    }

    return static_cast<T>(1 / std::cosh(x) / std::cosh(x));
}

template <typename Dtype>
void RNN_mm_cpu_batched(const Dtype* a_ptr,
                        size_t a_cols,
                        size_t a_rows,
                        size_t lda,
                        size_t a_stride,
                        int a_flags,
                        const Dtype* b_ptr,
                        size_t b_cols,
                        size_t b_rows,
                        size_t ldb,
                        size_t b_stride,
                        int b_flags,
                        Dtype* c_ptr,
                        size_t c_cols,
                        size_t c_rows,
                        size_t ldc,
                        size_t c_stride,
                        int batchCount,
                        double alpha,
                        double beta)
{
    for(int i = 0; i < batchCount; ++i)
    {
        gemm_cpu(a_ptr + a_stride * i,
                 a_cols,
                 a_rows,
                 lda,
                 a_flags == 1 ? true : false,
                 b_ptr + b_stride * i,
                 b_cols,
                 b_rows,
                 ldb,
                 b_flags == 1 ? true : false,
                 c_ptr + c_stride * i,
                 c_cols,
                 c_rows,
                 ldc,
                 alpha,
                 beta);
    }
}

#endif