#include "ggml-blas.h"
#include "ggml-backend-impl.h"

#include <atomic>
#include <cassert>
#include <future>
#include <vector>

#if defined(GGML_USE_ACCELERATE)
#   include <Accelerate/Accelerate.h>
#elif defined(GGML_BLAS_USE_MKL)
#   include <mkl.h>
#else
#   include <cblas.h>
#   ifdef BLIS_ENABLE_CBLAS
#       include <blis.h>
#   endif
#endif

struct ggml_backend_blas_context {
    int n_threads = GGML_DEFAULT_N_THREADS;
    std::unique_ptr<char[]> work_data;
    size_t work_size = 0;
#ifndef GGML_USE_OPENMP
    std::vector<std::future<void>> tasks;
#endif
    std::atomic<int> current_chunk;
};

// helper function to determine if it is better to use BLAS or not
// for large matrices, BLAS is faster
static bool ggml_backend_blas_use_blas(const struct ggml_tensor * dst) {
    const struct ggml_tensor * src0 = dst->src[0];
    const struct ggml_tensor * src1 = dst->src[1];

    const int64_t ne10 = src1->ne[0];

    const int64_t ne0 = dst->ne[0];
    const int64_t ne1 = dst->ne[1];

    // TODO: find the optimal values for these
    if (ggml_is_contiguous(src0) &&
        ggml_is_contiguous(src1) &&
        src1->type == GGML_TYPE_F32 &&
        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {

        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
        return true;
    }

    return false;
}

static void ggml_compute_forward_mul_mat_one_chunk(
    ggml_backend_blas_context * ctx,
    struct ggml_tensor * dst,
    const int64_t num_rows_per_vec_dot,
    const int64_t ir0_start,
    const int64_t ir0_end,
    const int64_t ir1_start,
    const int64_t ir1_end) {

    const struct ggml_tensor * src0 = dst->src[0];
    const struct ggml_tensor * src1 = dst->src[1];

    GGML_TENSOR_BINARY_OP_LOCALS

    const enum ggml_type type = src0->type;

    const bool src1_cont = ggml_is_contiguous(src1);

    const ggml_type_traits_t * type_traits = ggml_internal_get_type_traits_ptr(type);

    ggml_vec_dot_t    const vec_dot = type_traits->vec_dot;
    enum ggml_type    const vec_dot_type = type_traits->vec_dot_type;

    // broadcast factors
    const int64_t r2 = ne12 / ne02;
    const int64_t r3 = ne13 / ne03;

    //printf("ir0_start = %6lld, ir0_end = %6lld, ir1_start = %6lld, ir1_end = %6lld\n", ir0_start, ir0_end, ir1_start, ir1_end);

    // threads with no work simply yield (not sure if it helps)
    if (ir0_start >= ir0_end || ir1_start >= ir1_end) {
        return;
    }

    const void * wdata = (src1->type == vec_dot_type) ? src1->data : ctx->work_data.get();
    const size_t row_size = ggml_row_size(vec_dot_type, ne10);

    assert(ne12 % ne02 == 0);
    assert(ne13 % ne03 == 0);

    // block-tiling attempt
    const int64_t blck_0 = 16;
    const int64_t blck_1 = 16;

    const size_t src1_col_stride = src1_cont || src1->type != vec_dot_type ? row_size : nb11;

    // attempt to reduce false-sharing (does not seem to make a difference)
    // 16 * 2, accounting for mmla kernels
    float tmp[32];

    for (int64_t iir1 = ir1_start; iir1 < ir1_end; iir1 += blck_1) {
        for (int64_t iir0 = ir0_start; iir0 < ir0_end; iir0 += blck_0) {
            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir1_end; ir1 += num_rows_per_vec_dot) {
                const int64_t i13 = (ir1 / (ne12 * ne1));
                const int64_t i12 = (ir1 - i13 * ne12 * ne1) / ne1;
                const int64_t i11 = (ir1 - i13 * ne12 * ne1 - i12 * ne1);

                // broadcast src0 into src1
                const int64_t i03 = i13 / r3;
                const int64_t i02 = i12 / r2;

                const int64_t i1 = i11;
                const int64_t i2 = i12;
                const int64_t i3 = i13;

                const char * src0_row = (const char*)src0->data + (0 + i02 * nb02 + i03 * nb03);

                // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
                //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
                //       the original src1 data pointer, so we should index using the indices directly
                // TODO: this is a bit of a hack, we should probably have a better way to handle this
                const char * src1_col = (const char*)wdata +
                    (src1_cont || src1->type != vec_dot_type
                        ? (i11 + i12 * ne11 + i13 * ne12 * ne11) * row_size
                        : (i11 * nb11 + i12 * nb12 + i13 * nb13));
                float * dst_col = (float*)((char*)dst->data + (i1 * nb1 + i2 * nb2 + i3 * nb3));

                //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ++ir0) {
                //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
                //}

                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ir0 += num_rows_per_vec_dot) {
                    vec_dot(ne00, &tmp[ir0 - iir0], (num_rows_per_vec_dot > 1 ? 16 : 0), src0_row + ir0 * nb01, (num_rows_per_vec_dot > 1 ? nb01 : 0), src1_col, (num_rows_per_vec_dot > 1 ? src1_col_stride : 0), num_rows_per_vec_dot);
                }

                for (int cn = 0; cn < num_rows_per_vec_dot; ++cn) {
                    memcpy(&dst_col[iir0 + cn * nb1 / nb0], tmp + (cn * 16), (std::min(iir0 + blck_0, ir0_end) - iir0) * sizeof(float));
                }
            }
        }
    }
}

static void ggml_compute_forward_mul_mat(
              ggml_backend_blas_context * ctx,
              struct ggml_tensor * dst) {

    const struct ggml_tensor * src0 = dst->src[0];
    const struct ggml_tensor * src1 = dst->src[1];


    GGML_TENSOR_BINARY_OP_LOCALS

    const enum ggml_type type = src0->type;

    const ggml_type_traits_t * type_traits = ggml_internal_get_type_traits_ptr(type);
    const ggml_type_traits_t * type_traits_vec_dot = ggml_internal_get_type_traits_ptr(type_traits->vec_dot_type);
    enum ggml_type    const vec_dot_type          = type_traits->vec_dot_type;
    ggml_from_float_t const from_float_to_vec_dot = type_traits_vec_dot->from_float;
    int64_t           const vec_dot_num_rows      = type_traits->nrows;

    GGML_ASSERT(ne0 == ne01);
    GGML_ASSERT(ne1 == ne11);
    GGML_ASSERT(ne2 == ne12);
    GGML_ASSERT(ne3 == ne13);

    // we don't support permuted src0 or src1
    GGML_ASSERT(nb00 == ggml_type_size(type));
    GGML_ASSERT(nb10 == ggml_type_size(src1->type));

    // dst cannot be transposed or permuted
    GGML_ASSERT(nb0 == sizeof(float));
    GGML_ASSERT(nb0 <= nb1);
    GGML_ASSERT(nb1 <= nb2);
    GGML_ASSERT(nb2 <= nb3);

    // broadcast factors
    const int64_t r2 = ne12 / ne02;
    const int64_t r3 = ne13 / ne03;
    GGML_UNUSED(r2);
    GGML_UNUSED(r3);

    // nb01 >= nb00 - src0 is not transposed
    //   compute by src0 rows

    if (src1->type != vec_dot_type) {
        const size_t row_size = ggml_row_size(vec_dot_type, ne10);
        if (ctx->work_size < ne13*ne12*ne11*row_size) {
            ctx->work_data.reset(new char[ne13*ne12*ne11*row_size]);
            ctx->work_size = ne13*ne12*ne11*row_size;
        }
        char * wdata = ctx->work_data.get();

        GGML_ASSERT(src1->type == GGML_TYPE_F32);
        int block_size = ggml_blck_size(vec_dot_type);
        int type_size = ggml_type_size(vec_dot_type);

        for (int64_t i13 = 0; i13 < ne13; ++i13) {
            for (int64_t i12 = 0; i12 < ne12; ++i12) {
                for (int64_t i11 = 0; i11 < ne11; ++i11) {
                    //from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
                    //#pragma omp parallel num_threads(ctx->n_threads)
                    {
                        int nth = omp_get_num_threads();
                        int ith = omp_get_thread_num();
                        int blocks_per_thread = (ne10 + block_size - 1) / block_size / nth;
                        int i10_start = ith * blocks_per_thread * block_size;
                        int i10_end = std::min(i10_start + blocks_per_thread * block_size, (int)ne10);
                        //printf("thread %d/%d: i10_start = %d, i10_end = %d\n", ith, nth, i10_start, i10_end);
                        from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10_start*nb10),
                                                (void *) ((char *) wdata + (type_size*i10_start/block_size)),
                                                i10_end - i10_start);

                    }

                    wdata += row_size;
                }
            }
        }
    }

    // This is the size of the first dimension of the result, so we can iterate that way. (see the ASSERT above, these are the same numbers)
    const int64_t nr0 = ne0;

    // This is the size of the rest of the dimensions of the result
    const int64_t nr1 = ne1 * ne2 * ne3;

    // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols
    int64_t num_rows_per_vec_dot = vec_dot_num_rows;
    // TODO: currently the mmla kernels support only even numbered rows/cols.
    // this check can be removed once they are extended to support odd numbered rows/cols too
    if ((nr0 % 2 != 0) || (ne11 % 2 != 0)) {
        num_rows_per_vec_dot = 1;
    }

    // Now select a reasonable chunk size.
    int chunk_size = 16;

    // We need to step up the size if it's small
    if (nr0 == 1 || nr1 == 1) {
        chunk_size = 64;
    }

    // distribute the work across the inner or outer loop based on which one is larger
    // The number of chunks in the 0/1 dim.
    // CEIL(nr0/chunk_size)
    int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
    int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;

    // If the chunking is poor for the number of threads on this setup, scrap the whole plan.  Re-chunk it by thread.
    //   Also, chunking by thread was measured to have perform better on NUMA systems.  See https://github.com/ggerganov/llama.cpp/pull/6915
    //   In theory, chunking should be just as useful on NUMA and non NUMA systems, but testing disagreed with that.

    //const int ith = 0; // params->ith;
    const int nth = ctx->n_threads;

    // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
    ctx->current_chunk.store(nth);

    if (nchunk0 * nchunk1 < nth * 4 || ggml_is_numa()) {
        // distribute the thread work across the inner or outer loop based on which one is larger
        nchunk0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
        nchunk1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
    }

    // The number of elements in each chunk
    const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
    const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;

    //if (ith == 0)
    //    printf("MUL_MAT = [%d, %d, %d, %d] x [%d, %d, %d, %d] = %d x %d = %d.  Fp Ops/Ch %d\n", ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, nchunk0, nchunk1, nchunk0 * nchunk1, ne00 * nr0 * nr1 / nchunk0 / nchunk1);

    // The first chunk comes from our thread_id, the rest will get auto-assigned.
    if (nth > 1) {
        #pragma omp parallel num_threads(nth)
        {
            int current_chunk = omp_get_thread_num();

            while (current_chunk < nchunk0 * nchunk1) {
                const int64_t ith0 = current_chunk % nchunk0;
                const int64_t ith1 = current_chunk / nchunk0;

                const int64_t ir0_start = dr0 * ith0;
                const int64_t ir0_end = std::min(ir0_start + dr0, nr0);

                const int64_t ir1_start = dr1 * ith1;
                const int64_t ir1_end = std::min(ir1_start + dr1, nr1);

                ggml_compute_forward_mul_mat_one_chunk(ctx, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);

                if (nth >= nchunk0 * nchunk1) {
                    break;
                }

                current_chunk = ctx->current_chunk.fetch_add(1);
            }
        }
    } else {
        ggml_compute_forward_mul_mat_one_chunk(ctx, dst, num_rows_per_vec_dot, 0, nr0, 0, nr1);
    }

#ifdef GGML_PERF
    // These numbers are useful when trying to measure how well the threading scheduling works.
    //int64_t workSize = (ne01 * ne11 * ne12 * ne13 * ne00) / nchunk0 / nchunk1;
    //float time = (ggml_perf_time_us() - t0);
    //printf("MUL_MAT = %f ms, [%d, %d, %d, %d] x [%d, %d, %d, %d] = %I64u, %f ops/usec in %d chunks.\n", time / 1000.0, ne00, ne01, ne02, ne03, ne10, ne11, ne12, ne13, workSize, (float)workSize/time, chunks_executed);
#endif
}

static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
    const struct ggml_tensor * src0 = dst->src[0];
    const struct ggml_tensor * src1 = dst->src[1];

    GGML_TENSOR_BINARY_OP_LOCALS

    const enum ggml_type type = src0->type;

    GGML_ASSERT(ne0 == ne01);
    GGML_ASSERT(ne1 == ne11);
    GGML_ASSERT(ne2 == ne12);
    GGML_ASSERT(ne3 == ne13);

    // we don't support permuted src0 or src1
    GGML_ASSERT(nb00 == ggml_type_size(type));
    GGML_ASSERT(nb10 == ggml_type_size(src1->type));

    // dst cannot be transposed or permuted
    GGML_ASSERT(nb0 == sizeof(float));
    GGML_ASSERT(nb0 <= nb1);
    GGML_ASSERT(nb1 <= nb2);
    GGML_ASSERT(nb2 <= nb3);

    // broadcast factors
    const int64_t r2 = ne12/ne02;
    const int64_t r3 = ne13/ne03;

    const int64_t ne_plane      = ne01*ne00;
    const size_t  desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);

    if (ctx->work_size < desired_wsize) {
        ctx->work_data.reset(new char[desired_wsize]);
        ctx->work_size = desired_wsize;
    }
    void * wdata = ctx->work_data.get();

    // convert src0 to float
    if (type != GGML_TYPE_F32) {
        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
        ggml_to_float_t const to_float = type_traits.to_float;

        for (int64_t i03 = 0; i03 < ne03; i03++) {
            for (int64_t i02 = 0; i02 < ne02; i02++) {
                const void  *       x      = (char *)  src0->data + i02*nb02          + i03*nb03;
                      float * const wplane = (float *) wdata      + i02*ne_plane      + i03*ne02*ne_plane;

                const int min_cols_per_thread = 4096;
                const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
                const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);

#ifdef GGML_USE_OPENMP
                #pragma omp parallel for num_threads(n_threads)
                for (int64_t i01 = 0; i01 < ne01; i01++) {
                    to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
                }
#else
                for (int i = 1; i < n_threads; i++) {
                    const int64_t start =       i*ne01/n_threads;
                    const int64_t end   = (i + 1)*ne01/n_threads;
                    if (start < end) {
                        ctx->tasks.push_back(std::async(std::launch::async, [=]() {
                            for (int64_t i01 = start; i01 < end; i01++) {
                                to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
                            }
                        }));
                    }
                }
                {
                    // reuse the current thread for the first task
                    const int64_t start = 0;
                    const int64_t end   = ne01/n_threads;
                    for (int64_t i01 = start; i01 < end; i01++) {
                        to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
                    }
                }
#endif
            }
        }

#ifndef GGML_USE_OPENMP
        // wait for all tasks to finish
        for (auto & task : ctx->tasks) {
            task.get();
        }
        ctx->tasks.clear();
#endif
    }

#if defined(OPENBLAS_VERSION)
    openblas_set_num_threads(ctx->n_threads);
#endif

#if defined(BLIS_ENABLE_CBLAS)
    bli_thread_set_num_threads(ctx->n_threads);
#endif

    for (int64_t i13 = 0; i13 < ne13; i13++) {
        for (int64_t i12 = 0; i12 < ne12; i12++) {
            const int64_t i03 = i13/r3;
            const int64_t i02 = i12/r2;

            const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
            const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
                  float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);

            if (type != GGML_TYPE_F32) {
                x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
            }

            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
                        ne1, ne01, ne10,
                        1.0f,   y, ne10,
                                x, ne00,
                        0.0f,   d, ne01);
        }
    }
}

static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
    const struct ggml_tensor * src0 = dst->src[0];
    const struct ggml_tensor * src1 = dst->src[1];

    GGML_TENSOR_BINARY_OP_LOCALS

    GGML_ASSERT(ne0  == ne00);
    GGML_ASSERT(ne1  == ne10);
    GGML_ASSERT(ne2  == ne02);
    GGML_ASSERT(ne02 == ne12);
    GGML_ASSERT(ne3  == ne13);
    GGML_ASSERT(ne03 == ne13);

    // we don't support permuted src0 or src1
    GGML_ASSERT(nb00 == sizeof(float));

    // dst cannot be transposed or permuted
    GGML_ASSERT(nb0 == sizeof(float));
    // GGML_ASSERT(nb0 <= nb1);
    // GGML_ASSERT(nb1 <= nb2);
    // GGML_ASSERT(nb2 <= nb3);

    // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
    // src0: (k,n)
    // src1: (k,m)
    // dst:  (m,n)
    //
    // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
    // Also expressed as (major,minor)
    // a: (m,k): so src1 transposed
    // b: (k,n): so src0
    // c: (m,n)
    //
    // However, if ggml_is_transposed(src1) is true, then
    // src1->data already contains a transposed version, so sgemm mustn't
    // transpose it further.

    int n = src0->ne[0];
    int k = src0->ne[1];
    int m = src1->ne[0];

    CBLAS_TRANSPOSE transposeA;
    int lda;

    if (!ggml_is_transposed(src1)) {
        transposeA = CblasTrans;
        lda = m;
    } else {
        transposeA = CblasNoTrans;
        lda = k;
    }

    float * a = (float *) ((char *) src1->data);
    float * b = (float *) ((char *) src0->data);
    float * c = (float *) ((char *) dst->data);

    cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);

    GGML_UNUSED(ctx);
}

// backend interface

GGML_CALL static const char * ggml_backend_blas_name(ggml_backend_t backend) {
    return "BLAS";

    GGML_UNUSED(backend);
}

GGML_CALL static void ggml_backend_blas_free(ggml_backend_t backend) {
    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
    delete ctx;
    delete backend;
}

GGML_CALL static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
    return ggml_backend_cpu_buffer_type();

    GGML_UNUSED(backend);
}

GGML_CALL static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;

    for (int i = 0; i < cgraph->n_nodes; i++) {
        struct ggml_tensor * node = cgraph->nodes[i];

        switch (node->op) {
            case GGML_OP_MUL_MAT:
                //ggml_backend_blas_mul_mat(ctx, node);
                ggml_compute_forward_mul_mat(ctx, node);
                break;

            case GGML_OP_OUT_PROD:
                ggml_backend_blas_out_prod(ctx, node);
                break;

            case GGML_OP_NONE:
            case GGML_OP_RESHAPE:
            case GGML_OP_VIEW:
            case GGML_OP_PERMUTE:
            case GGML_OP_TRANSPOSE:
                break;

            default:
                fprintf(stderr, "%s: unsupported op %s\n", __func__, ggml_op_desc(node));
                GGML_ASSERT(false);
        }
    }

    return GGML_STATUS_SUCCESS;

    GGML_UNUSED(backend);
}

GGML_CALL static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {

    return op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_OUT_PROD;

    /*
    const struct ggml_tensor * src0 = op->src[0];
    const struct ggml_tensor * src1 = op->src[1];

    return (op->op == GGML_OP_MUL_MAT  && ggml_backend_blas_use_blas(op)) ||
           (op->op == GGML_OP_OUT_PROD && op->src[0]->type == GGML_TYPE_F32 &&
                                          op->src[1]->type == GGML_TYPE_F32 &&
                                          ggml_is_matrix(src0) &&
                                          ggml_is_matrix(src1) &&
                                          ggml_is_contiguous(src0) &&
                                          (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
    */

    GGML_UNUSED(backend);
}

GGML_CALL static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
    return ggml_backend_buft_is_host(buft);

    GGML_UNUSED(backend);
}

static struct ggml_backend_i blas_backend_i = {
    /* .get_name                = */ ggml_backend_blas_name,
    /* .free                    = */ ggml_backend_blas_free,
    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
    /* .set_tensor_async        = */ NULL,
    /* .get_tensor_async        = */ NULL,
    /* .cpy_tensor_async        = */ NULL,
    /* .synchronize             = */ NULL,
    /* .graph_plan_create       = */ NULL,
    /* .graph_plan_free         = */ NULL,
    /* .graph_plan_update       = */ NULL,
    /* .graph_plan_compute      = */ NULL,
    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
    /* .supports_op             = */ ggml_backend_blas_supports_op,
    /* .supports_buft           = */ ggml_backend_blas_supports_buft,
    /* .offload_op              = */ NULL,
    /* .event_new               = */ NULL,
    /* .event_free              = */ NULL,
    /* .event_record            = */ NULL,
    /* .event_wait              = */ NULL,
    /* .event_synchronize       = */ NULL,
};

static ggml_guid_t ggml_backend_blas_guid(void) {
    static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
    return &guid;
}

ggml_backend_t ggml_backend_blas_init(void) {
    ggml_backend_blas_context * ctx = new ggml_backend_blas_context;

    ggml_backend_t backend = new ggml_backend {
        /* .guid      = */ ggml_backend_blas_guid(),
        /* .interface = */ blas_backend_i,
        /* .context   = */ ctx,
    };

#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
    if (openblas_get_parallel() != OPENBLAS_OPENMP) {
        fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
    }
#endif

#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
    fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
#endif

    return backend;
}

GGML_CALL bool ggml_backend_is_blas(ggml_backend_t backend) {
    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
}

void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
    GGML_ASSERT(ggml_backend_is_blas(backend_blas));

    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
    ctx->n_threads = n_threads;
}