/* ************************************************************************
 * Copyright 2018-2021 Advanced Micro Devices, Inc.
 * ************************************************************************ */

#pragma once

#include "cblas_interface.hpp"
#include "flops.hpp"
#include "norm.hpp"
#include "rocblas.hpp"
#include "rocblas_datatype2string.hpp"
#include "rocblas_init.hpp"
#include "rocblas_math.hpp"
#include "rocblas_random.hpp"
#include "rocblas_test.hpp"
#include "rocblas_vector.hpp"
#include "unit.hpp"
#include "utility.hpp"

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

template <typename T>
void testing_geam_bad_arg(const Arguments& arg)
{
    auto rocblas_geam_fn = arg.fortran ? rocblas_geam<T, true> : rocblas_geam<T, false>;

    const rocblas_int M = 100;
    const rocblas_int N = 100;

    const rocblas_int lda = 100;
    const rocblas_int ldb = 100;
    const rocblas_int ldc = 100;

    const T h_alpha = 1.0;
    const T h_beta  = 1.0;

    const rocblas_operation transA = rocblas_operation_none;
    const rocblas_operation transB = rocblas_operation_none;

    rocblas_local_handle handle{arg};

    size_t size_A = N * size_t(lda);
    size_t size_B = N * size_t(ldb);
    size_t size_C = N * size_t(ldc);

    host_vector<T> hA(size_A);
    host_vector<T> hB(size_B);
    host_vector<T> hC(size_C);

    // allocate memory on device
    device_vector<T> dA(size_A);
    device_vector<T> dB(size_B);
    device_vector<T> dC(size_C);
    CHECK_DEVICE_ALLOCATION(dA.memcheck());
    CHECK_DEVICE_ALLOCATION(dB.memcheck());
    CHECK_DEVICE_ALLOCATION(dC.memcheck());

    EXPECT_ROCBLAS_STATUS(
        rocblas_geam_fn(
            handle, transA, transB, M, N, &h_alpha, nullptr, lda, &h_beta, dB, ldb, dC, ldc),
        rocblas_status_invalid_pointer);

    EXPECT_ROCBLAS_STATUS(
        rocblas_geam_fn(
            handle, transA, transB, M, N, &h_alpha, dA, lda, &h_beta, nullptr, ldb, dC, ldc),
        rocblas_status_invalid_pointer);

    EXPECT_ROCBLAS_STATUS(
        rocblas_geam_fn(
            handle, transA, transB, M, N, &h_alpha, dA, lda, &h_beta, dB, ldb, nullptr, ldc),
        rocblas_status_invalid_pointer);

    EXPECT_ROCBLAS_STATUS(
        rocblas_geam_fn(handle, transA, transB, M, N, nullptr, dA, lda, &h_beta, dB, ldb, dC, ldc),
        rocblas_status_invalid_pointer);

    EXPECT_ROCBLAS_STATUS(
        rocblas_geam_fn(handle, transA, transB, M, N, &h_alpha, dA, lda, nullptr, dB, ldb, dC, ldc),
        rocblas_status_invalid_pointer);

    EXPECT_ROCBLAS_STATUS(
        rocblas_geam_fn(
            nullptr, transA, transB, M, N, &h_alpha, dA, lda, &h_beta, dB, ldb, dC, ldc),
        rocblas_status_invalid_handle);
}

template <typename T>
void testing_geam(const Arguments& arg)
{
    auto rocblas_geam_fn = arg.fortran ? rocblas_geam<T, true> : rocblas_geam<T, false>;

    rocblas_operation transA = char2rocblas_operation(arg.transA);
    rocblas_operation transB = char2rocblas_operation(arg.transB);

    rocblas_int M = arg.M;
    rocblas_int N = arg.N;

    rocblas_int lda = arg.lda;
    rocblas_int ldb = arg.ldb;
    rocblas_int ldc = arg.ldc;

    T alpha = arg.get_alpha<T>();
    T beta  = arg.get_beta<T>();

    T* dC_in_place;

    rocblas_int A_row, A_col, B_row, B_col;
    rocblas_int inc1_A, inc2_A, inc1_B, inc2_B;

    double gpu_time_used, cpu_time_used;
    gpu_time_used = cpu_time_used = 0.0;

    double rocblas_error_1 = std::numeric_limits<double>::max();
    double rocblas_error_2 = std::numeric_limits<double>::max();
    double rocblas_error   = std::numeric_limits<double>::max();

    rocblas_local_handle handle{arg};

    if(transA == rocblas_operation_none)
    {
        A_row  = M;
        A_col  = N;
        inc1_A = 1;
        inc2_A = lda;
    }
    else
    {
        A_row  = N;
        A_col  = M;
        inc1_A = lda;
        inc2_A = 1;
    }
    if(transB == rocblas_operation_none)
    {
        B_row  = M;
        B_col  = N;
        inc1_B = 1;
        inc2_B = ldb;
    }
    else
    {
        B_row  = N;
        B_col  = M;
        inc1_B = ldb;
        inc2_B = 1;
    }

    size_t size_A = size_t(lda) * size_t(A_col);
    size_t size_B = size_t(ldb) * size_t(B_col);
    size_t size_C = size_t(ldc) * size_t(N);

    // argument sanity check before allocating invalid memory
    bool invalid_size = M < 0 || N < 0 || lda < A_row || ldb < B_row || ldc < M;
    if(invalid_size || !M || !N)
    {
        EXPECT_ROCBLAS_STATUS(rocblas_geam_fn(handle,
                                              transA,
                                              transB,
                                              M,
                                              N,
                                              nullptr,
                                              nullptr,
                                              lda,
                                              nullptr,
                                              nullptr,
                                              ldb,
                                              nullptr,
                                              ldc),
                              invalid_size ? rocblas_status_invalid_size : rocblas_status_success);
        return;
    }

    // Naming: dX is in GPU (device) memory. hK is in CPU (host) memory
    host_vector<T> h_alpha(1);
    host_vector<T> h_beta(1);
    host_vector<T> hA(size_A), hA_copy(size_A);
    host_vector<T> hB(size_B), hB_copy(size_B);
    host_vector<T> hC_1(size_C);
    host_vector<T> hC_2(size_C);
    host_vector<T> hC_gold(size_C);
    CHECK_HIP_ERROR(h_alpha.memcheck());
    CHECK_HIP_ERROR(h_beta.memcheck());
    CHECK_HIP_ERROR(hA.memcheck());
    CHECK_HIP_ERROR(hB.memcheck());
    CHECK_HIP_ERROR(hC_1.memcheck());
    CHECK_HIP_ERROR(hC_2.memcheck());
    CHECK_HIP_ERROR(hC_gold.memcheck());

    // Initial Data on CPU
    h_alpha[0] = alpha;
    h_beta[0]  = beta;

    // Initialize data on host memory
    rocblas_init_matrix(hA, arg, A_row, A_col, lda, 0, 1, rocblas_client_alpha_sets_nan, true);
    rocblas_init_matrix(hB, arg, B_row, B_col, ldb, 0, 1, rocblas_client_beta_sets_nan);

    // allocate memory on device
    device_vector<T> dA(size_A);
    device_vector<T> dB(size_B);
    device_vector<T> dC(size_C);
    device_vector<T> d_alpha(1);
    device_vector<T> d_beta(1);
    CHECK_DEVICE_ALLOCATION(dA.memcheck());
    CHECK_DEVICE_ALLOCATION(dB.memcheck());
    CHECK_DEVICE_ALLOCATION(dC.memcheck());
    CHECK_DEVICE_ALLOCATION(d_alpha.memcheck());
    CHECK_DEVICE_ALLOCATION(d_beta.memcheck());

    // copy data from CPU to device
    CHECK_HIP_ERROR(d_alpha.transfer_from(h_alpha));
    CHECK_HIP_ERROR(d_beta.transfer_from(h_beta));
    CHECK_HIP_ERROR(dA.transfer_from(hA));
    CHECK_HIP_ERROR(dB.transfer_from(hB));

    if(arg.unit_check || arg.norm_check)
    {
        // ROCBLAS
        CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_host));

        CHECK_ROCBLAS_ERROR(rocblas_geam_fn(
            handle, transA, transB, M, N, &alpha, dA, lda, &beta, dB, ldb, dC, ldc));

        CHECK_HIP_ERROR(hC_1.transfer_from(dC));

        CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_device));
        CHECK_ROCBLAS_ERROR(rocblas_geam_fn(
            handle, transA, transB, M, N, d_alpha, dA, lda, d_beta, dB, ldb, dC, ldc));

        // reference calculation for golden result
        cpu_time_used = get_time_us_no_sync();

        cblas_geam(transA,
                   transB,
                   M,
                   N,
                   (T*)h_alpha,
                   (T*)hA,
                   lda,
                   (T*)h_beta,
                   (T*)hB,
                   ldb,
                   (T*)hC_gold,
                   ldc);

        cpu_time_used = get_time_us_no_sync() - cpu_time_used;

        // fetch GPU
        CHECK_HIP_ERROR(hC_2.transfer_from(dC));

        if(arg.unit_check)
        {
            unit_check_general<T>(M, N, ldc, hC_gold, hC_1);
            unit_check_general<T>(M, N, ldc, hC_gold, hC_2);
        }

        if(arg.norm_check)
        {
            rocblas_error_1 = norm_check_general<T>('F', M, N, ldc, hC_gold, hC_1);
            rocblas_error_2 = norm_check_general<T>('F', M, N, ldc, hC_gold, hC_2);
        }

        // inplace check for dC == dA
        {

            if(arg.alpha_isnan<T>()) // need no NaN A
            {
                rocblas_init<T>(hA_copy, A_row, A_col, lda);
                CHECK_HIP_ERROR(dA.transfer_from(hA_copy));
            }
            else
                hA_copy = hA;

            dC_in_place = dA;

            CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_host));
            auto status_h = rocblas_geam_fn(
                handle, transA, transB, M, N, &alpha, dA, lda, &beta, dB, ldb, dC_in_place, ldc);

            if(lda != ldc || transA != rocblas_operation_none)
            {
                EXPECT_ROCBLAS_STATUS(status_h, rocblas_status_invalid_size);
            }
            else
            {
                CHECK_HIP_ERROR(
                    hipMemcpy(hC_1, dC_in_place, sizeof(T) * size_C, hipMemcpyDeviceToHost));
                // dA was clobbered by dC_in_place, so copy hA back to dA
                CHECK_HIP_ERROR(dA.transfer_from(hA));

                // reference calculation
                cblas_geam(transA,
                           transB,
                           M,
                           N,
                           (T*)h_alpha,
                           (T*)hA_copy,
                           lda,
                           (T*)h_beta,
                           (T*)hB,
                           ldb,
                           (T*)hC_gold,
                           ldc);

                if(arg.unit_check)
                {
                    unit_check_general<T>(M, N, ldc, hC_gold, hC_1);
                }

                if(arg.norm_check)
                {
                    rocblas_error = norm_check_general<T>('F', M, N, ldc, hC_gold, hC_1);
                }
            }
        }

        // inplace check for dC == dB
        {
            if(arg.beta_isnan<T>()) // need no NaN B
            {
                rocblas_init<T>(hB_copy, B_row, B_col, ldb);
                CHECK_HIP_ERROR(dB.transfer_from(hB_copy));
            }
            else
                hB_copy = hB;

            dC_in_place = dB;

            CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_host));
            auto status_h = rocblas_geam_fn(
                handle, transA, transB, M, N, &alpha, dA, lda, &beta, dB, ldb, dC_in_place, ldc);

            if(ldb != ldc || transB != rocblas_operation_none)
            {
                EXPECT_ROCBLAS_STATUS(status_h, rocblas_status_invalid_size);
            }
            else
            {
                CHECK_ROCBLAS_ERROR(status_h);

                CHECK_HIP_ERROR(
                    hipMemcpy(hC_1, dC_in_place, sizeof(T) * size_C, hipMemcpyDeviceToHost));

                // reference calculation
                cblas_geam(transA,
                           transB,
                           M,
                           N,
                           (T*)h_alpha,
                           (T*)hA_copy,
                           lda,
                           (T*)h_beta,
                           (T*)hB_copy,
                           ldb,
                           (T*)hC_gold,
                           ldc);

                if(arg.unit_check)
                {
                    unit_check_general<T>(M, N, ldc, hC_gold, hC_1);
                }

                if(arg.norm_check)
                {
                    rocblas_error = norm_check_general<T>('F', M, N, ldc, hC_gold, hC_1);
                }
            }
        }
    } // end of if unit/norm check

    if(arg.timing)
    {
        int number_cold_calls = arg.cold_iters;
        int number_hot_calls  = arg.iters;

        CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_host));

        for(int i = 0; i < number_cold_calls; i++)
        {
            rocblas_geam_fn(handle, transA, transB, M, N, &alpha, dA, lda, &beta, dB, ldb, dC, ldc);
        }

        hipStream_t stream;
        CHECK_ROCBLAS_ERROR(rocblas_get_stream(handle, &stream));
        gpu_time_used = get_time_us_sync(stream); // in microseconds
        for(int i = 0; i < number_hot_calls; i++)
        {
            rocblas_geam_fn(handle, transA, transB, M, N, &alpha, dA, lda, &beta, dB, ldb, dC, ldc);
        }
        gpu_time_used = get_time_us_sync(stream) - gpu_time_used;

        ArgumentModel<e_transA, e_transB, e_M, e_N, e_alpha, e_lda, e_beta, e_ldb, e_ldc>{}
            .log_args<T>(rocblas_cout,
                         arg,
                         gpu_time_used,
                         geam_gflop_count<T>(M, N),
                         ArgumentLogging::NA_value,
                         cpu_time_used,
                         rocblas_error_1,
                         rocblas_error_2);
    }
}