#include "ggml-backend.h"
#include "ggml-backend-impl.h"
#include "ggml-cpu.h"
#include "ggml-impl.h"
#include <cctype>
#include <string>
#include <vector>

#if defined(__APPLE__)
#include <sys/types.h>
#include <sys/sysctl.h>
#endif

#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
    #define NOMINMAX
#endif
#include <windows.h>
#endif

// ggml-backend interface

#ifdef GGML_USE_CPU_HBM

// buffer type HBM

#include <hbwmalloc.h>

static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
    return "CPU_HBM";

    GGML_UNUSED(buft);
}

static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
    hbw_free(buffer->context);
}

static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
    void * ptr;
    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
    if (result != 0) {
        GGML_LOG_ERROR("failed to allocate HBM buffer of size %zu\n", size);
        return NULL;
    }

    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
    buffer->buft = buft;
    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;

    return buffer;
}

ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
        /* .iface    = */ {
            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
        },
        /* .context  = */ NULL,
    };

    return &ggml_backend_cpu_buffer_type_hbm;
}
#endif

static ggml_backend_buffer_type_t * ggml_backend_cpu_get_extra_bufts(ggml_backend_dev_t device) {
    static ggml_backend_buffer_type_t bufts[] = {
#ifdef GGML_USE_CPU_HBM
        ggml_backend_cpu_hbm_buffer_type(),
#endif
        NULL
    };

    return bufts;

    GGML_UNUSED(device);
}

// CPU backend - backend (stream)

struct ggml_backend_cpu_context {
    int                 n_threads;
    ggml_threadpool_t   threadpool;

    uint8_t *           work_data;
    size_t              work_size;

    ggml_abort_callback abort_callback;
    void *              abort_callback_data;
};

static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
    return "CPU";

    GGML_UNUSED(backend);
}

static void ggml_backend_cpu_free(ggml_backend_t backend) {
    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
    delete[] cpu_ctx->work_data;
    delete cpu_ctx;
    delete backend;
}

struct ggml_backend_plan_cpu {
    struct ggml_cplan cplan;
    struct ggml_cgraph cgraph;
};

static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;

    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;

    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
    cpu_plan->cgraph = *cgraph; // FIXME: deep copy

    if (cpu_plan->cplan.work_size > 0) {
        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
        if (cpu_plan->cplan.work_data == NULL) {
            delete cpu_plan;
            return NULL;
        }
    }

    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;

    return cpu_plan;
}

static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;

    delete[] cpu_plan->cplan.work_data;
    delete cpu_plan;

    GGML_UNUSED(backend);
}

static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;

    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);

    GGML_UNUSED(backend);
}

static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;

    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);

    if (cpu_ctx->work_size < cplan.work_size) {
        delete[] cpu_ctx->work_data;
        cpu_ctx->work_data = new uint8_t[cplan.work_size];
        if (cpu_ctx->work_data == NULL) {
            cpu_ctx->work_size = 0;
            return GGML_STATUS_ALLOC_FAILED;
        }
        cpu_ctx->work_size = cplan.work_size;
    }
    cplan.work_data = (uint8_t *)cpu_ctx->work_data;

    cplan.abort_callback      = cpu_ctx->abort_callback;
    cplan.abort_callback_data = cpu_ctx->abort_callback_data;

    return ggml_graph_compute(cgraph, &cplan);
}

static const struct ggml_backend_i ggml_backend_cpu_i = {
    /* .get_name                = */ ggml_backend_cpu_get_name,
    /* .free                    = */ ggml_backend_cpu_free,
    /* .set_tensor_async        = */ NULL,
    /* .get_tensor_async        = */ NULL,
    /* .cpy_tensor_async        = */ NULL,
    /* .synchronize             = */ NULL,
    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
    /* .graph_plan_update       = */ NULL,
    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
};

static ggml_guid_t ggml_backend_cpu_guid(void) {
    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
    return &guid;
}

ggml_backend_t ggml_backend_cpu_init(void) {
    // initialize CPU backend now to avoid slowing the first graph computation
    ggml_cpu_init();

    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
    if (ctx == NULL) {
        return NULL;
    }

    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
    ctx->threadpool          = NULL;
    ctx->work_data           = NULL;
    ctx->work_size           = 0;
    ctx->abort_callback      = NULL;
    ctx->abort_callback_data = NULL;

    ggml_backend_t cpu_backend = new ggml_backend {
        /* .guid      = */ ggml_backend_cpu_guid(),
        /* .interface = */ ggml_backend_cpu_i,
        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
        /* .context   = */ ctx,
    };

    if (cpu_backend == NULL) {
        delete ctx;
        return NULL;
    }

    return cpu_backend;
}

bool ggml_backend_is_cpu(ggml_backend_t backend) {
    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
}

void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));

    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
    ctx->n_threads = n_threads;
}

void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));

    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;

    if (ctx->threadpool && ctx->threadpool != threadpool) {
        // already had a different threadpool, pause/suspend it before switching
        ggml_threadpool_pause(ctx->threadpool);
    }
    ctx->threadpool = threadpool;
}

void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));

    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
    ctx->abort_callback = abort_callback;
    ctx->abort_callback_data = abort_callback_data;
}

// CPU backend - device

struct ggml_backend_cpu_device_context {
    std::string description = "CPU";

    ggml_backend_cpu_device_context() {
#ifdef __APPLE__
        size_t len = 0;
        if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
            description.resize(len);
            sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
        }
#elif defined(__linux__)
        FILE * f = fopen("/proc/cpuinfo", "r");
        if (f) {
            char buf[1024];
            while (fgets(buf, sizeof(buf), f)) {
                if (strncmp(buf, "model name", 10) == 0) {
                    char * p = strchr(buf, ':');
                    if (p) {
                        p++;
                        while (std::isspace(*p)) {
                            p++;
                        }
                        while (std::isspace(p[strlen(p) - 1])) {
                            p[strlen(p) - 1] = '\0';
                        }
                        description = p;
                        break;
                    }
                }
            }
            fclose(f);
        }
#elif defined(_WIN32)
        HKEY hKey;
        if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
                        TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
                        0,
                        KEY_READ,
                        &hKey) == ERROR_SUCCESS) {
            DWORD cpu_brand_size = 0;
            if (RegQueryValueExA(hKey,
                                TEXT("ProcessorNameString"),
                                NULL,
                                NULL,
                                NULL,
                                &cpu_brand_size) == ERROR_SUCCESS) {
                description.resize(cpu_brand_size);
                if (RegQueryValueExA(hKey,
                                    TEXT("ProcessorNameString"),
                                    NULL,
                                    NULL,
                                    (LPBYTE)&description[0], // NOLINT
                                    &cpu_brand_size) == ERROR_SUCCESS) {
                    if (description.find('\0') != std::string::npos) {
                        description.resize(description.find('\0'));
                    }
                }
            }
            RegCloseKey(hKey);
        }
#endif
    }
};

static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
    return "CPU";

    GGML_UNUSED(dev);
}

static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
    struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;

    return ctx->description.c_str();
}

static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
    // TODO
    *free = 0;
    *total = 0;

    GGML_UNUSED(dev);
}

static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
    return GGML_BACKEND_DEVICE_TYPE_CPU;

    GGML_UNUSED(dev);
}

static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
    props->name        = ggml_backend_cpu_device_get_name(dev);
    props->description = ggml_backend_cpu_device_get_description(dev);
    props->type        = ggml_backend_cpu_device_get_type(dev);
    ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
    props->caps = {
        /* .async                 = */ false,
        /* .host_buffer           = */ false,
        /* .buffer_from_host_ptr  = */ true,
        /* .events                = */ false,
    };
}

static ggml_backend_t ggml_backend_cpu_device_init_backend(ggml_backend_dev_t dev, const char * params) {
    return ggml_backend_cpu_init();

    GGML_UNUSED(dev);
    GGML_UNUSED(params);
}

static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
    return ggml_backend_cpu_buffer_type();

    GGML_UNUSED(dev);
}

static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
    return ggml_backend_cpu_buffer_from_ptr(ptr, size);

    GGML_UNUSED(dev);
    GGML_UNUSED(max_tensor_size);
}

static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
    switch (op->op) {
        case GGML_OP_CPY:
            return
                op->type != GGML_TYPE_IQ2_XXS &&
                op->type != GGML_TYPE_IQ2_XS  &&
                op->type != GGML_TYPE_IQ1_S   &&
                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
        case GGML_OP_MUL_MAT:
            return op->src[1]->type == GGML_TYPE_F32;// FIXME || op->src[1]->type == ggml_get_type_traits(op->src[0]->type)->vec_dot_type;
        case GGML_OP_ROPE_BACK:
            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
        case GGML_OP_IM2COL_BACK:
            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
        case GGML_OP_OUT_PROD:
            return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
        default:
            return true;
    }

    GGML_UNUSED(dev);
}

static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
    return ggml_backend_buft_is_host(buft);

    GGML_UNUSED(dev);
}

static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
    /* .get_name             = */ ggml_backend_cpu_device_get_name,
    /* .get_description      = */ ggml_backend_cpu_device_get_description,
    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
    /* .get_type             = */ ggml_backend_cpu_device_get_type,
    /* .get_props            = */ ggml_backend_cpu_device_get_props,
    /* .init_backend         = */ ggml_backend_cpu_device_init_backend,
    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
    /* .get_host_buffer_type = */ NULL,
    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_host_ptr,
    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
    /* .offload_op           = */ NULL,
    /* .event_new            = */ NULL,
    /* .event_free           = */ NULL,
    /* .event_synchronize    = */ NULL,
};

// CPU backend - backend (reg)

static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
    return "CPU";

    GGML_UNUSED(reg);
}

static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
    return 1;

    GGML_UNUSED(reg);
}

static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
    GGML_ASSERT(index == 0);

    static ggml_backend_cpu_device_context ctx;
    static ggml_backend_device ggml_backend_cpu_device = {
        /* .iface   = */ ggml_backend_cpu_device_i,
        /* .reg     = */ reg,
        /* .context = */ &ctx,
    };

    return &ggml_backend_cpu_device;
}

struct ggml_backend_feature {
    const char * name;
    const char * value;
};

// Not used yet
// This is intended to replace the the ggml_cpu_has_* functions when loading the CPU backend dynamically,
// and additionally to allow other backends to expose their own list of features that applications can query using the same API.
static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t reg) {
    static std::vector<ggml_backend_feature> features = []() {
        std::vector<ggml_backend_feature> features;
        if (ggml_cpu_has_sse3()) {
            features.push_back({ "SSE3", "1" });
        }
        if (ggml_cpu_has_ssse3()) {
            features.push_back({ "SSSE3", "1" });
        }
        if (ggml_cpu_has_avx()) {
            features.push_back({ "AVX", "1" });
        }
        if (ggml_cpu_has_avx2()) {
            features.push_back({ "AVX2", "1" });
        }
        if (ggml_cpu_has_f16c()) {
            features.push_back({ "F16C", "1" });
        }
        if (ggml_cpu_has_fma()) {
            features.push_back({ "FMA", "1" });
        }
        if (ggml_cpu_has_avx_vnni()) {
            features.push_back({ "AVX_VNNI", "1" });
        }
        if (ggml_cpu_has_avx512()) {
            features.push_back({ "AVX512", "1" });
        }
        if (ggml_cpu_has_avx512_vbmi()) {
            features.push_back({ "AVX512_VBMI", "1" });
        }
        if (ggml_cpu_has_avx512_vnni()) {
            features.push_back({ "AVX512_VNNI", "1" });
        }
        if (ggml_cpu_has_avx512_bf16()) {
            features.push_back({ "AVX512_BF16", "1" });
        }
        if (ggml_cpu_has_amx_int8()) {
            features.push_back({ "AMX_INT8", "1" });
        }
        if (ggml_cpu_has_neon()) {
            features.push_back({ "NEON", "1" });
        }
        if (ggml_cpu_has_arm_fma()) {
            features.push_back({ "ARM_FMA", "1" });
        }
        if (ggml_cpu_has_fp16_va()) {
            features.push_back({ "FP16_VA", "1" });
        }
        if (ggml_cpu_has_matmul_int8()) {
            features.push_back({ "MATMUL_INT8", "1" });
        }
        if (ggml_cpu_has_sve()) {
            features.push_back({ "SVE", "1" });
        }
        if (ggml_cpu_get_sve_cnt() > 0) {
            static std::string sve_cnt = std::to_string(ggml_cpu_get_sve_cnt());
            features.push_back({ "SVE_CNT", sve_cnt.c_str() });
        }
        if (ggml_cpu_has_riscv_v()) {
            features.push_back({ "RISCV_V", "1" });
        }
        if (ggml_cpu_has_vsx()) {
            features.push_back({ "VSX", "1" });
        }
        if (ggml_cpu_has_wasm_simd()) {
            features.push_back({ "WASM_SIMD", "1" });
        }
        if (ggml_cpu_has_llamafile()) {
            features.push_back({ "LLAMAFILE", "1" });
        }

        features.push_back({ nullptr, nullptr });

        return features;
    }();

    return features.data();

    GGML_UNUSED(reg);
}

static void * ggml_backend_cpu_get_proc_address(ggml_backend_reg_t reg, const char * name) {
    if (strcmp(name, "ggml_backend_set_n_threads") == 0) {
        return (void *)ggml_backend_cpu_set_n_threads;
    }
    if (strcmp(name, "ggml_backend_dev_get_extra_bufts") == 0) {
        return (void *)ggml_backend_cpu_get_extra_bufts;
    }

    return NULL;

    GGML_UNUSED(reg);
}

static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
    /* .get_proc_address = */ ggml_backend_cpu_get_proc_address,
};

ggml_backend_reg_t ggml_backend_cpu_reg(void) {
    static struct ggml_backend_reg ggml_backend_cpu_reg = {
        /* .iface   = */ ggml_backend_cpu_reg_i,
        /* .context = */ NULL,
    };

    return &ggml_backend_cpu_reg;
}