/**
 * Copyright (c) 2025 Huawei Technologies Co., Ltd.
 * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
 * 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/small_matmul.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 "golden.hpp"
#include "helper.hpp"

using namespace Catlass;

using ArchTag = Arch::AtlasA2;
static constexpr uint32_t STAGES = 2;
constexpr bool ENABLE_UNIT_FLAG = true;
constexpr bool ENABLE_SHUFFLE_K = false;
using DispatchPolicy = Gemm::MmadAtlasA2Small<STAGES, ENABLE_UNIT_FLAG, ENABLE_SHUFFLE_K>;
using L1TileShape = GemmShape<128, 256, 256>;
using L0TileShape = GemmShape<128, 256, 64>;

template <class LayoutA, class LayoutB, class LayoutC>
CATLASS_GLOBAL void SmallMatmul(
    GemmCoord problemShape,
    GM_ADDR gmA,
    LayoutA layoutA,
    GM_ADDR gmB,
    LayoutB layoutB,
    GM_ADDR gmC,
    LayoutC layoutC
) {
    using AType = Gemm::GemmType<half, LayoutA>;
    using BType = Gemm::GemmType<half, LayoutB>;
    using CType = Gemm::GemmType<half, LayoutC>;

    using BlockMmad = Gemm::Block::BlockMmad<DispatchPolicy, L1TileShape, L0TileShape, AType, BType, CType>;
    using BlockEpilogue = void;

    // Swizzle offset is 3 and direction is 0.
    using BlockScheduler = typename Gemm::Block::GemmIdentityBlockSwizzle<3, 0>;

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

    typename MatmulKernel::Params params{problemShape, gmA, layoutA, gmB, layoutB, gmC, layoutC};

    // call a kernel
    MatmulKernel matmul;
    matmul(params);
}

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();

    // Get the number of cube cores of the current hardware
    auto aicCoreNum = platform_ascendc::PlatformAscendCManager::GetInstance()->GetCoreNumAic();
    if ((CeilDiv(m, L1TileShape::M) * CeilDiv(n, L1TileShape::N) > aicCoreNum) || (k > L1TileShape::K)) {
        std::cout << "Not satisfy the constraints of small_matmul." << std::endl;
        std::cout << "Please see ReadMe or code for more details." << std::endl;
        ACL_CHECK(aclrtDestroyStream(stream));
        ACL_CHECK(aclrtResetDevice(options.deviceId));
        ACL_CHECK(aclFinalize());
        return;
    }

    size_t lenA = static_cast<size_t>(m) * k;
    size_t lenB = static_cast<size_t>(k) * n;
    size_t lenC = static_cast<size_t>(m) * n;

    size_t sizeA = lenA * sizeof(fp16_t);
    size_t sizeB = lenB * sizeof(fp16_t);
    size_t sizeC = lenC * sizeof(fp16_t);

    using LayoutA = layout::RowMajor;
    using LayoutB = layout::RowMajor;
    using LayoutC = layout::RowMajor;
    LayoutA layoutA{m, k};
    LayoutB layoutB{k, n};
    LayoutC layoutC{m, n};

    std::vector<fp16_t> hostA(lenA);
    std::vector<fp16_t> hostB(lenB);
    golden::FillRandomData<fp16_t>(hostA, -5.0f, 5.0f);
    golden::FillRandomData<fp16_t>(hostB, -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));

    SmallMatmul<<<aicCoreNum, nullptr, stream>>>(
        options.problemShape, deviceA, layoutA, deviceB, layoutB, deviceC, layoutC
    );
    ACL_CHECK(aclrtSynchronizeStream(stream));

    std::vector<fp16_t> hostC(lenC);
    ACL_CHECK(aclrtMemcpy(hostC.data(), sizeC, deviceC, sizeC, ACL_MEMCPY_DEVICE_TO_HOST));

    std::vector<float> hostGolden(lenC);
    golden::ComputeMatmul(options.problemShape, hostA, layoutA, hostB, layoutB, hostGolden, layoutC);

    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));

    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;
}