/**
 * This program is free software, you can redistribute it and/or modify.
 * Copyright (c) 2026 Huawei Technologies Co., Ltd.
 * This file is a part of the CANN Open Software.
 * Licensed under CANN Open Software License Agreement Version 2.0 (the "License").
 * Please refer to the License for details. You may not use this file except in compliance with the License.
 * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING
 * BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE. See LICENSE in the root of
 * the software repository for the full text of the License.
 */

// By setting the K_MAX_SHAPE_DIM macro, the dimension of the AscendC Tensor's ShapeInfo is configured to 0,
// optimizing stack space. If you need to use the ShapeInfo of the AscendC Tensor, please undefine this macro.
#ifndef K_MAX_SHAPE_DIM
#define K_MAX_SHAPE_DIM 0
#endif

#include "catlass/gemm/kernel/basic_matmul_tla.hpp"

#include "catlass/arch/arch.hpp"
#include "catlass/catlass.hpp"
#include "catlass/gemm/block/block_mmad.hpp"
#include "catlass/gemm/block/block_swizzle.hpp"
#include "catlass/gemm/device/device_gemm.hpp"
#include "catlass/gemm/dispatch_policy.hpp"
#include "catlass/gemm/gemm_type.hpp"
#include "catlass/layout/layout.hpp"
#include "catlass/status.hpp"
#include "tla/layout.hpp"

#include "golden.hpp"
#include "helper.hpp"

using namespace Catlass;
using namespace tla;

using Options = GemmOptions;

static void Run(const Options &options)
{
    aclrtStream stream{nullptr};

    ACL_CHECK(aclInit(nullptr));
    ACL_CHECK(aclrtSetDevice(options.deviceId));
    ACL_CHECK(aclrtCreateStream(&stream));

    uint32_t m = options.problemShape.m();
    uint32_t n = options.problemShape.n();
    uint32_t k = options.problemShape.k();

    using ElementA = float;
    using ElementB = float;
    using ElementC = float;
    // if no bias, set ElementBias to void
    using ElementBias = void;

    using ElementBiasType = std::conditional_t<std::is_void_v<ElementBias>, uint8_t, ElementBias>;

    using LayoutTagA = layout::VectorLayout;
    using LayoutTagB = layout::ColumnMajor;
    using LayoutTagC = layout::RowMajor;
    LayoutTagA tagA = layout::VectorLayout(k);
    LayoutTagB tagB = LayoutTagB::MakeLayout<ElementB>(k, n);
    LayoutTagC tagC = LayoutTagC::MakeLayout<ElementC>(m, n);

    size_t lenA = k;
    size_t lenB = tagB.Capacity();
    size_t lenC = tagC.Capacity();
    size_t lenBias = static_cast<size_t>(n);

    size_t sizeA = lenA * sizeof(ElementA);
    size_t sizeB = lenB * sizeof(ElementB);
    size_t sizeC = lenC * sizeof(ElementC);
    size_t sizeBias = lenBias * sizeof(ElementBiasType);
    size_t sizeWorkspace;

    std::vector<ElementA> hostA(lenA);
    std::vector<ElementB> hostB(lenB);
    std::vector<ElementBiasType> hostBias(lenBias);
    golden::FillRandomData<ElementA>(hostA, -5.0f, 5.0f);
    golden::FillRandomData<ElementB>(hostB, -5.0f, 5.0f);
    if constexpr (!std::is_void_v<ElementBias>) {
        golden::FillRandomData<ElementBiasType>(hostBias, -5.0f, 5.0f);
    }

    uint8_t *deviceA{nullptr};
    ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceA), sizeA, ACL_MEM_MALLOC_HUGE_FIRST));
    ACL_CHECK(aclrtMemcpy(deviceA, sizeA, hostA.data(), sizeA, ACL_MEMCPY_HOST_TO_DEVICE));

    uint8_t *deviceB{nullptr};
    ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceB), sizeB, ACL_MEM_MALLOC_HUGE_FIRST));
    ACL_CHECK(aclrtMemcpy(deviceB, sizeB, hostB.data(), sizeB, ACL_MEMCPY_HOST_TO_DEVICE));

    uint8_t *deviceC{nullptr};
    ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceC), sizeC, ACL_MEM_MALLOC_HUGE_FIRST));

    uint8_t *deviceBias{nullptr};
    if constexpr (!std::is_void_v<ElementBias>) {
        ACL_CHECK(aclrtMalloc(reinterpret_cast<void **>(&deviceBias), sizeBias, ACL_MEM_MALLOC_HUGE_FIRST));
        ACL_CHECK(aclrtMemcpy(deviceBias, sizeBias, hostBias.data(), sizeBias, ACL_MEMCPY_HOST_TO_DEVICE));
    }

    uint8_t *deviceWorkspace{nullptr};

    // Get the number of cube cores of the current hardware
    auto aicCoreNum = platform_ascendc::PlatformAscendCManager::GetInstance()->GetCoreNumAic();

    using ArchTag = Arch::Ascend950;
    constexpr bool enableUnitFlag = false;
    constexpr bool useHF32 = false;
    constexpr bool enableL1Resident = false;
    constexpr uint32_t l0CStages = 1;
    constexpr uint32_t l1AStages = 1;
    constexpr uint32_t l1BStages = 1;
    constexpr uint32_t l0AStages = 1;
    constexpr uint32_t l0BStages = 1;
    using DispatchPolicy = Gemm::MmadPingpong<
        ArchTag,
        enableUnitFlag, useHF32, l0CStages, enableL1Resident,
        l1AStages, l1BStages, l0AStages, l0BStages
    >;

    using L1TileShape = Shape<Int<128>, Int<128>, Int<256>>;
    using L0TileShape = Shape<Int<128>, Int<128>, Int<128>>;

    auto layoutA = tla::MakeLayout<ElementA, LayoutTagA>(m, k);
    auto layoutB = tla::MakeLayout<ElementB, LayoutTagB>(k, n);
    auto layoutC = tla::MakeLayout<ElementC, LayoutTagC>(m, n);

    using TileCopy = Gemm::Tile::PackedTileCopyTla<
        ArchTag, ElementA, LayoutTagA, ElementB, LayoutTagB, ElementC, LayoutTagC, ElementBias>;
    using BlockMmad = Gemm::Block::BlockMmadTla<
        DispatchPolicy, L1TileShape, L0TileShape, ElementA, ElementB, ElementC, ElementBias, TileCopy>;
    using BlockEpilogue = void;

    uint32_t taskNum = CeilDiv(options.problemShape.m(), tla::get<0>(L1TileShape{})) *
                       CeilDiv(options.problemShape.n(), tla::get<1>(L1TileShape{}));
    uint32_t aicCoreUsed = min(aicCoreNum, taskNum);

    // Swizzle offset is 3 and direction is 1.
    using BlockScheduler = typename Gemm::Block::GemmIdentityBlockSwizzle<3, 1>;

    // kernel level
    using MatmulKernel = Gemm::Kernel::BasicMatmulTla<BlockMmad, BlockEpilogue, BlockScheduler>;

    using MatmulAdapter = Gemm::Device::DeviceGemm<MatmulKernel>;

    MatmulKernel::Arguments arguments{
        options.problemShape, deviceA, layoutA, deviceB, layoutB, deviceC, layoutC, deviceBias
    };

    MatmulAdapter matmulOp;
    matmulOp.CanImplement(arguments);
    sizeWorkspace = matmulOp.GetWorkspaceSize(arguments);
    if (sizeWorkspace > 0) {
        ACL_CHECK(
            aclrtMalloc(reinterpret_cast<void **>(&deviceWorkspace), sizeWorkspace, ACL_MEM_MALLOC_HUGE_FIRST)
        );
    }
    matmulOp.Initialize(arguments, deviceWorkspace);
    matmulOp(stream, aicCoreUsed);

    ACL_CHECK(aclrtSynchronizeStream(stream));

    std::vector<ElementC> hostC(lenC);
    ACL_CHECK(aclrtMemcpy(hostC.data(), sizeC, deviceC, sizeC, ACL_MEMCPY_DEVICE_TO_HOST));

    std::vector<float> hostGolden(lenC);
    if constexpr (!std::is_void_v<ElementBias>) {
        golden::ComputeMatmulBias(options.problemShape, hostA, tagA, hostB, tagB, hostBias, hostGolden, tagC);
    } else {
        golden::ComputeMatmul(options.problemShape, hostA, tagA, hostB, tagB, hostGolden, tagC);
    }

    std::vector<uint64_t> errorIndices = golden::CompareData(hostC, hostGolden, k);
    if (errorIndices.empty()) {
        std::cout << "Compare success." << std::endl;
    } else {
        std::cerr << "Compare failed. Error count: " << errorIndices.size() << std::endl;
    }

    ACL_CHECK(aclrtFree(deviceA));
    ACL_CHECK(aclrtFree(deviceB));
    ACL_CHECK(aclrtFree(deviceC));
    if constexpr (!std::is_void_v<ElementBias>) {
        ACL_CHECK(aclrtFree(deviceBias));
    }
    if (sizeWorkspace > 0) {
        ACL_CHECK(aclrtFree(deviceWorkspace));
    }

    ACL_CHECK(aclrtDestroyStream(stream));
    ACL_CHECK(aclrtResetDevice(options.deviceId));
    ACL_CHECK(aclFinalize());
}

int main(int argc, const char **argv)
{
    Options options;
    if (options.Parse(argc, argv) != 0) {
        return -1;
    }
    Run(options);
    return 0;
}