/* ************************************************************************
 * Copyright 2016-2021 Advanced Micro Devices, Inc.
 *
 * ************************************************************************ */

#include <fstream>
#include <iostream>
#include <stdlib.h>
#include <vector>

#include "testing_common.hpp"

using namespace std;

/* ============================================================================================ */

template <typename T>
hipblasStatus_t testing_trtri_batched(const Arguments& argus)
{
    bool FORTRAN = argus.fortran;
    auto hipblasTrtriBatchedFn
        = FORTRAN ? hipblasTrtriBatched<T, true> : hipblasTrtriBatched<T, false>;

    const double rel_error = get_epsilon<T>() * 1000;

    int N = argus.N;
    int lda;
    int ldinvA = lda = argus.lda;
    int batch_count  = argus.batch_count;

    size_t A_size = size_t(lda) * N;

    // check here to prevent undefined memory allocation error
    if(N < 0 || lda < 0 || lda < N || batch_count < 0)
    {
        return HIPBLAS_STATUS_INVALID_VALUE;
    }
    // Naming: dK is in GPU (device) memory. hK is in CPU (host) memory
    host_batch_vector<T> hA(A_size, 1, batch_count);
    host_batch_vector<T> hB(A_size, 1, batch_count);

    device_batch_vector<T> dA(A_size, 1, batch_count);
    device_batch_vector<T> dinvA(A_size, 1, batch_count);

    CHECK_HIP_ERROR(dA.memcheck());
    CHECK_HIP_ERROR(dinvA.memcheck());

    double             gpu_time_used, hipblas_error;
    hipblasLocalHandle handle(argus);

    char char_uplo = argus.uplo_option;
    char char_diag = argus.diag_option;

    hipblasFillMode_t uplo = char2hipblas_fill(char_uplo);
    hipblasDiagType_t diag = char2hipblas_diagonal(char_diag);

    hipblas_init(hA, true);

    for(int b = 0; b < batch_count; b++)
    {
        // proprocess the matrix to avoid ill-conditioned matrix
        for(int i = 0; i < N; i++)
        {
            for(int j = 0; j < N; j++)
            {
                hA[b][i + j * lda] *= 0.01;

                if(j % 2)
                    hA[b][i + j * lda] *= -1;
                if(uplo == HIPBLAS_FILL_MODE_LOWER && j > i)
                    hA[b][i + j * lda] = 0.0f;
                else if(uplo == HIPBLAS_FILL_MODE_UPPER && j < i)
                    hA[b][i + j * lda] = 0.0f;
                if(i == j)
                {
                    if(diag == HIPBLAS_DIAG_UNIT)
                        hA[b][i + j * lda] = 1.0;
                    else
                        hA[b][i + j * lda] *= 100.0;
                }
            }
        }
    }

    hB.copy_from(hA);
    CHECK_HIP_ERROR(dA.transfer_from(hA));
    CHECK_HIP_ERROR(dinvA.transfer_from(hA));

    if(argus.unit_check || argus.norm_check)
    {
        /* =====================================================================
            HIPBLAS
        =================================================================== */
        CHECK_HIPBLAS_ERROR(hipblasTrtriBatchedFn(handle,
                                                  uplo,
                                                  diag,
                                                  N,
                                                  dA.ptr_on_device(),
                                                  lda,
                                                  dinvA.ptr_on_device(),
                                                  ldinvA,
                                                  batch_count));

        // copy output from device to CPU
        CHECK_HIP_ERROR(hA.transfer_from(dinvA));

        /* =====================================================================
           CPU BLAS
        =================================================================== */
        for(int b = 0; b < batch_count; b++)
        {
            cblas_trtri<T>(char_uplo, char_diag, N, hB[b], lda);
        }

        if(argus.unit_check)
        {
            for(int b = 0; b < batch_count; b++)
                near_check_general<T>(N, N, lda, hB[b], hA[b], rel_error);
        }
        if(argus.norm_check)
        {
            hipblas_error = norm_check_general<T>('F', N, N, lda, hB, hA, batch_count);
        }
    }

    if(argus.timing)
    {
        hipStream_t stream;
        CHECK_HIPBLAS_ERROR(hipblasGetStream(handle, &stream));

        int runs = argus.cold_iters + argus.iters;
        for(int iter = 0; iter < runs; iter++)
        {
            if(iter == argus.cold_iters)
                gpu_time_used = get_time_us_sync(stream);

            CHECK_HIPBLAS_ERROR(hipblasTrtriBatchedFn(handle,
                                                      uplo,
                                                      diag,
                                                      N,
                                                      dA.ptr_on_device(),
                                                      lda,
                                                      dinvA.ptr_on_device(),
                                                      ldinvA,
                                                      batch_count));
        }
        gpu_time_used = get_time_us_sync(stream) - gpu_time_used;

        ArgumentModel<e_uplo_option, e_diag_option, e_N, e_lda, e_batch_count>{}.log_args<T>(
            std::cout,
            argus,
            gpu_time_used,
            trtri_gflop_count<T>(N),
            trtri_gbyte_count<T>(N),
            hipblas_error);
    }

    return HIPBLAS_STATUS_SUCCESS;
}