@fugood/llama.node 0.3.16 → 0.3.17

package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp

@@ -313,7 +313,6 @@ struct ggml_backend_sycl_context {
     int device;
     std::string name;
     optimize_feature opt_feature;
-    bool optimized_graph=false;
 
     queue_ptr qptrs[GGML_SYCL_MAX_DEVICES][GGML_SYCL_MAX_STREAMS] = { { nullptr } };
 
@@ -494,298 +493,9 @@ static __dpct_inline__ Tp* get_pointer(sycl::local_accessor<Tp, dim> acc) {
 
 int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size);
 
-typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                       const ggml_tensor *src1,
-                                       ggml_tensor *dst, const float *src0_dd,
-                                       const float *src1_dd, float *dst_dd,
-                                       const queue_ptr &main_stream);
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
-        const sycl::nd_item<3> &item_ct1) {
-    const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                    item_ct1.get_local_id(2);
-    const int i1 = (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                    item_ct1.get_local_id(1));
-    const int i2 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
-                    item_ct1.get_local_id(0)) /
-                   ne3;
-    const int i3 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
-                    item_ct1.get_local_id(0)) %
-                   ne3;
-
-    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    for (int i0 = i0s; i0 < ne0;
-         i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) {
-        const int i10 = i0 % ne10;
-        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-    }
-}
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s00,*/ int s01, int s02, int s03,
-        /*int s10,*/ int s11, int s12, int s13,
-        const sycl::nd_item<3> &item_ct1) {
-
-    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
-                  item_ct1.get_local_id(2);
-
-    const int i3 = i/(ne2*ne1*ne0);
-    const int i2 = (i/(ne1*ne0)) % ne2;
-    const int i1 = (i/ne0) % ne1;
-    const int i0 = i % ne0;
-
-    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    const int i10 = i0 % ne10;
-    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-}
-
-
-template<float (*bin_op)(const float, const float)>
-struct bin_bcast_sycl {
-    template <typename src0_t, typename src1_t, typename dst_t>
-    void operator()(ggml_backend_sycl_context & ctx,
-                    const struct ggml_tensor *src0,
-                    const struct ggml_tensor *src1, struct ggml_tensor *dst,
-                    const src0_t *src0_dd, const src1_t *src1_dd, dst_t *dst_dd,
-                    queue_ptr stream) {
-
-        GGML_TENSOR_BINARY_OP_LOCALS
-
-        int nr0 = ne10/ne0;
-        int nr1 = ne11/ne1;
-        int nr2 = ne12/ne2;
-        int nr3 = ne13/ne3;
-
-        int nr[4] = { nr0, nr1, nr2, nr3 };
-
-        // collapse dimensions until first broadcast dimension
-        int64_t cne[] = {ne0, ne1, ne2, ne3};
-        int64_t cne0[] = {ne00, ne01, ne02, ne03};
-        int64_t cne1[] = {ne10, ne11, ne12, ne13};
-        size_t cnb[] = {nb0, nb1, nb2, nb3};
-        size_t cnb0[] = {nb00, nb01, nb02, nb03};
-        size_t cnb1[] = {nb10, nb11, nb12, nb13};
-        auto collapse = [](int64_t cne[]) {
-            cne[0] *= cne[1];
-            cne[1] = cne[2];
-            cne[2] = cne[3];
-            cne[3] = 1;
-        };
-
-        auto collapse_nb = [](size_t cnb[], int64_t cne[]) {
-            cnb[1] *= cne[1];
-            cnb[2] *= cne[2];
-            cnb[3] *= cne[3];
-        };
-
-        if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
-            for (int i = 0; i < 4; i++) {
-                if (nr[i] != 1) {
-                    break;
-                }
-                if (i > 0) {
-                    collapse_nb(cnb, cne);
-                    collapse_nb(cnb0, cne0);
-                    collapse_nb(cnb1, cne1);
-                    collapse(cne);
-                    collapse(cne0);
-                    collapse(cne1);
-                }
-            }
-        }
-        {
-            int64_t ne0 = cne[0];
-            int64_t ne1 = cne[1];
-            int64_t ne2 = cne[2];
-            int64_t ne3 = cne[3];
-
-            int64_t ne10 = cne1[0];
-            int64_t ne11 = cne1[1];
-            int64_t ne12 = cne1[2];
-            int64_t ne13 = cne1[3];
-
-            size_t nb0 = cnb[0];
-            size_t nb1 = cnb[1];
-            size_t nb2 = cnb[2];
-            size_t nb3 = cnb[3];
-
-            size_t nb00 = cnb0[0];
-            size_t nb01 = cnb0[1];
-            size_t nb02 = cnb0[2];
-            size_t nb03 = cnb0[3];
-
-            size_t nb10 = cnb1[0];
-            size_t nb11 = cnb1[1];
-            size_t nb12 = cnb1[2];
-            size_t nb13 = cnb1[3];
-
-            size_t s0 = nb0 / sizeof(dst_t);
-            size_t s1 = nb1 / sizeof(dst_t);
-            size_t s2 = nb2 / sizeof(dst_t);
-            size_t s3 = nb3 / sizeof(dst_t);
-
-            size_t s10 = nb10 / sizeof(src1_t);
-            size_t s11 = nb11 / sizeof(src1_t);
-            size_t s12 = nb12 / sizeof(src1_t);
-            size_t s13 = nb13 / sizeof(src1_t);
-
-            size_t s00 = nb00 / sizeof(src0_t);
-            size_t s01 = nb01 / sizeof(src0_t);
-            size_t s02 = nb02 / sizeof(src0_t);
-            size_t s03 = nb03 / sizeof(src0_t);
-
-            GGML_UNUSED(s00);
-
-            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
-            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
-
-            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
-            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
-
-            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
-            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
-
-            GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s10 == 1);
-
-            const int block_size = 128;
-
-            int64_t hne0 = std::max(ne0/2LL, 1LL);
-
-            sycl::range<3> block_dims(1, 1, 1);
-            block_dims[2] = std::min<unsigned int>(hne0, block_size);
-            block_dims[1] = std::min<unsigned int>(
-                ne1, block_size / (unsigned int)block_dims[2]);
-            block_dims[0] = std::min(
-                std::min<unsigned int>(
-                    ne2 * ne3, block_size / (unsigned int)block_dims[2] /
-                                   (unsigned int)block_dims[1]),
-                64U);
-
-            sycl::range<3> block_nums(
-                (ne2 * ne3 + block_dims[0] - 1) / block_dims[0],
-                (ne1 + block_dims[1] - 1) / block_dims[1],
-                (hne0 + block_dims[2] - 1) / block_dims[2]);
-
-            if (block_nums[0] > 65535) {
-                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
-                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
-                {
-                    dpct::has_capability_or_fail(stream->get_device(),
-                                                 {sycl::aspect::fp16});
-
-                    stream->parallel_for(
-                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
-                                              sycl::range<3>(1, 1, block_size),
-                                          sycl::range<3>(1, 1, block_size)),
-                        [=](sycl::nd_item<3> item_ct1) {
-                            k_bin_bcast_unravel<bin_op>(
-                                src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3,
-                                ne10, ne11, ne12, ne13, s1, s2, s3, s01, s02,
-                                s03, s11, s12, s13, item_ct1);
-                        });
-                }
-            } else {
-                /*
-                DPCT1049:16: The work-group size passed to the SYCL kernel may
-                exceed the limit. To get the device limit, query
-                info::device::max_work_group_size. Adjust the work-group size if
-                needed.
-                */
-                dpct::has_capability_or_fail(stream->get_device(),
-                                             {sycl::aspect::fp16});
-
-                stream->parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
-                        k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
-                                            ne2, ne3, ne10, ne11, ne12, ne13,
-                                            s1, s2, s3, s01, s02, s03, s11, s12, s13,
-                                            item_ct1);
-                    });
-            }
-        }
-        GGML_UNUSED(ctx);
-    }
-};
-
-template <class op>
-inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                   const ggml_tensor *src1, ggml_tensor *dst,
-                                   const float *src0_dd, const float *src1_dd,
-                                   float *dst_dd,
-                                   const queue_ptr &main_stream) {
-
-    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        op()(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-        op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd,
-             (sycl::half *)dst_dd, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-        op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, dst_dd,
-             main_stream);
-    } else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
-        op()(ctx, src0, src1, dst, (const int32_t *)src0_dd, (const int32_t *)src1_dd, (int32_t *)dst_dd,
-             main_stream);
-    } else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
-        op()(ctx, src0, src1, dst, (const int16_t *)src0_dd, (const int16_t *)src1_dd, (int16_t *)dst_dd,
-             main_stream);
-    } else {
-        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
-            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ABORT("fatal error");
-    }
+constexpr size_t ceil_div(const size_t m, const size_t n) {
+    return (m + n - 1) / n;
 }
 
 bool gpu_has_xmx(sycl::device &dev);
-
-void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
-                                 const ggml_tensor *src1, ggml_tensor *dst,
-                                 const ggml_sycl_op_flatten_t op);
 #endif // GGML_SYCL_COMMON_HPP

package/src/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp

@@ -16,9 +16,18 @@
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
 #include <syclcompat/math.hpp>
-#include <oneapi/mkl.hpp>
 #include <map>
 
+#ifdef GGML_SYCL_USE_INTEL_ONEMKL
+#include <oneapi/mkl.hpp>
+// Allow to use the same namespace for Intel oneMKL and oneMath
+namespace oneapi {
+    namespace math = mkl;
+}
+#else
+#include <oneapi/math.hpp>
+#endif
+
 #include "ggml.h"
 
 #if defined(__linux__)
@@ -83,13 +92,32 @@ inline std::string get_device_backend_and_type(const sycl::device &device) {
 }
 
 template <typename Ts> struct matrix_info_t {
-    oneapi::mkl::transpose transpose_info[2];
+    oneapi::math::transpose transpose_info[2];
     Ts                     value_info[2];
     std::int64_t           size_info[3];
     std::int64_t           ld_info[3];
     std::int64_t           groupsize_info;
 };
 
+inline auto get_onemath_backend(sycl::queue& queue)
+#if defined(GGML_SYCL_GENERIC) || defined(GGML_SYCL_USE_INTEL_ONEMKL)
+  -> sycl::queue&
+#endif
+{
+// If the backend is known at compile-time, use oneMath backend_selector to use
+// compile-time dispatching and avoid the need to dlopen libraries. Otherwise
+// fallback to runtime dispatching.
+#if defined(GGML_SYCL_NVIDIA)
+    return oneapi::math::backend_selector<oneapi::math::backend::cublas>{ queue };
+#elif defined(GGML_SYCL_AMD)
+    return oneapi::math::backend_selector<oneapi::math::backend::rocblas>{ queue };
+#elif defined(GGML_SYCL_GENERIC) || defined(GGML_SYCL_USE_INTEL_ONEMKL)
+    return queue;
+#else
+    static_assert(false, "Unsupported backend");
+#endif
+}
+
 namespace dpct
 {
     typedef sycl::queue *queue_ptr;
@@ -1686,26 +1714,18 @@ namespace dpct
 
     namespace detail
     {
-        template <class Ta, class Tb, class Tc, class Ts>
-        inline void gemm_impl(sycl::queue &q, oneapi::mkl::transpose a_trans,
-                              oneapi::mkl::transpose b_trans, int m, int n, int k,
-                              const void *alpha, const void *a, int lda, const void *b,
-                              int ldb, const void *beta, void *c, int ldc)
-        {
-            Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
-            Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
-            auto data_a = get_memory<const Ta>(a);
-            auto data_b = get_memory<const Tb>(b);
-            auto data_c = get_memory<Tc>(c);
-#ifdef GGML_SYCL_NVIDIA
-            oneapi::mkl::blas::column_major::gemm(oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q },
-                                                  a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb,
-                                                  beta_value, data_c, ldc);
-#else
-            oneapi::mkl::blas::column_major::gemm(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda, data_b, ldb,
-                                                  beta_value, data_c, ldc);
-#endif
-        }
+    template <class Ta, class Tb, class Tc, class Ts>
+    inline void gemm_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
+                          int n, int k, const void * alpha, const void * a, int lda, const void * b, int ldb,
+                          const void * beta, void * c, int ldc) {
+        Ts   alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
+        Ts   beta_value  = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
+        auto data_a      = get_memory<const Ta>(a);
+        auto data_b      = get_memory<const Tb>(b);
+        auto data_c      = get_memory<Tc>(c);
+        oneapi::math::blas::column_major::gemm(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value, data_a,
+                                               lda, data_b, ldb, beta_value, data_c, ldc);
+    }
 
         template <typename VecT, class BinaryOperation, class = void>
         class vectorized_binary
@@ -1735,7 +1755,7 @@ namespace dpct
         };
 
         template <class Ta, class Tb, class Tc, class Ts>
-        inline void gemm_batch_impl(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans,
+        inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans,
                                     int m, int n, int k, const void * alpha, const void ** a, int lda, const void ** b,
                                     int ldb, const void * beta, void ** c, int ldc, int batch_size,
                                     matrix_info_t<float> * matrix_info) {
@@ -1754,48 +1774,28 @@ namespace dpct
             matrix_info->ld_info[2] = ldc;
             matrix_info->groupsize_info = batch_size;
 
-#ifdef GGML_SYCL_NVIDIA
-            sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
-                oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, matrix_info->transpose_info,
-                matrix_info->transpose_info + 1, matrix_info->size_info, matrix_info->size_info + 1,
-                matrix_info->size_info + 2, reinterpret_cast<Ts *>(matrix_info->value_info),
-                reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
-                matrix_info->ld_info + 1, reinterpret_cast<Ts *>(matrix_info->value_info + 1),
-                reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
-#else
-            sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
-                q, matrix_info->transpose_info, matrix_info->transpose_info + 1, matrix_info->size_info,
-                matrix_info->size_info + 1, matrix_info->size_info + 2, reinterpret_cast<Ts *>(matrix_info->value_info),
-                reinterpret_cast<const Ta **>(a), matrix_info->ld_info, reinterpret_cast<const Tb **>(b),
-                matrix_info->ld_info + 1, reinterpret_cast<Ts *>(matrix_info->value_info + 1),
-                reinterpret_cast<Tc **>(c), matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
-#endif
+            sycl::event e = oneapi::math::blas::column_major::gemm_batch(
+                get_onemath_backend(q), matrix_info->transpose_info, matrix_info->transpose_info + 1,
+                matrix_info->size_info, matrix_info->size_info + 1, matrix_info->size_info + 2,
+                reinterpret_cast<Ts *>(matrix_info->value_info), reinterpret_cast<const Ta **>(a), matrix_info->ld_info,
+                reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
+                reinterpret_cast<Ts *>(matrix_info->value_info + 1), reinterpret_cast<Tc **>(c),
+                matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
         }
 
         template <class Ta, class Tb, class Tc, class Ts>
-        inline void
-        gemm_batch_impl(sycl::queue &q, oneapi::mkl::transpose a_trans,
-                        oneapi::mkl::transpose b_trans, int m, int n,
-                        int k, const void *alpha, const void *a, int lda,
-                        long long int stride_a, const void *b, int ldb,
-                        long long int stride_b, const void *beta, void *c,
-                        int ldc, long long int stride_c, int batch_size)
-        {
+        inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans,
+                                    int m, int n, int k, const void * alpha, const void * a, int lda,
+                                    long long int stride_a, const void * b, int ldb, long long int stride_b,
+                                    const void * beta, void * c, int ldc, long long int stride_c, int batch_size) {
             Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
             Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
             auto data_a = get_memory<const Ta>(a);
             auto data_b = get_memory<const Tb>(b);
             auto data_c = get_memory<Tc>(c);
-#ifdef GGML_SYCL_NVIDIA
-            oneapi::mkl::blas::column_major::gemm_batch(
-                oneapi::mkl::backend_selector<oneapi::mkl::backend::cublas>{ q }, a_trans, b_trans, m, n, k,
-                alpha_value, data_a, lda, stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc, stride_c,
-                batch_size);
-#else
-            oneapi::mkl::blas::column_major::gemm_batch(q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
-                                                        stride_a, data_b, ldb, stride_b, beta_value, data_c, ldc,
-                                                        stride_c, batch_size);
-#endif
+            oneapi::math::blas::column_major::gemm_batch(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value,
+                                                         data_a, lda, stride_a, data_b, ldb, stride_b, beta_value,
+                                                         data_c, ldc, stride_c, batch_size);
         }
 
     } // namespace detail
@@ -2259,13 +2259,10 @@ namespace dpct
                            sycl::range<3>(x, y, 1), direction);
     }
 
-    inline void gemm(sycl::queue &q, oneapi::mkl::transpose a_trans,
-                     oneapi::mkl::transpose b_trans, int m, int n, int k,
-                     const void *alpha, const void *a, library_data_t a_type,
-                     int lda, const void *b, library_data_t b_type, int ldb,
-                     const void *beta, void *c, library_data_t c_type, int ldc,
-                     library_data_t scaling_type)
-    {
+    inline void gemm(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, int n,
+                     int k, const void * alpha, const void * a, library_data_t a_type, int lda, const void * b,
+                     library_data_t b_type, int ldb, const void * beta, void * c, library_data_t c_type, int ldc,
+                     library_data_t scaling_type) {
         if (scaling_type == library_data_t::real_float &&
             c_type == library_data_t::complex_float)
         {
@@ -2329,9 +2326,8 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_float, library_data_t::real_float):
         {
-            detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float,
-                              float>(q, a_trans, b_trans, m, n, k, alpha, a, lda, b,
-                                     ldb, beta, c, ldc);
+            detail::gemm_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
             break;
         }
         case detail::get_type_combination_id(
@@ -2369,8 +2365,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_bfloat16, library_data_t::real_float):
         {
-            detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16,
-                              oneapi::mkl::bfloat16, float>(
+            detail::gemm_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
             break;
         }
@@ -2390,7 +2385,7 @@ namespace dpct
         default:
             throw std::runtime_error("the combination of data type is unsupported");
         }
-    } // gemm()
+    }  // gemm()
 
     /// Computes a batch of matrix-matrix product with general matrices.
     /// \param [in] q The queue where the routine should be executed.
@@ -2412,7 +2407,7 @@ namespace dpct
     /// \param [in] ldc Leading dimension of C.
     /// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
     /// \param [in] scaling_type Data type of the scaling factors.
-    inline void gemm_batch(sycl::queue & q, oneapi::mkl::transpose a_trans, oneapi::mkl::transpose b_trans, int m,
+    inline void gemm_batch(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
                            int n, int k, const void * alpha, const void * a[], library_data_t a_type, int lda,
                            const void * b[], library_data_t b_type, int ldb, const void * beta, void * c[],
                            library_data_t c_type, int ldc, int batch_size, library_data_t scaling_type,
@@ -2450,7 +2445,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_bfloat16, library_data_t::real_float):
         {
-            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float>(
+            detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
             break;
         }
@@ -2458,7 +2453,7 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_float, library_data_t::real_float):
         {
-            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float, float>(
+            detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
                 q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc, batch_size, matrix_info);
             break;
         }
@@ -2534,15 +2529,11 @@ namespace dpct
     /// \param [in] stride_c Stride between the different C matrices.
     /// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
     /// \param [in] scaling_type Data type of the scaling factors.
-    inline void gemm_batch(sycl::queue &q, oneapi::mkl::transpose a_trans,
-                           oneapi::mkl::transpose b_trans, int m, int n, int k,
-                           const void *alpha, const void *a, library_data_t a_type,
-                           int lda, long long int stride_a, const void *b,
-                           library_data_t b_type, int ldb, long long int stride_b,
-                           const void *beta, void *c, library_data_t c_type,
-                           int ldc, long long int stride_c, int batch_size,
-                           library_data_t scaling_type)
-    {
+    inline void gemm_batch(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
+                           int n, int k, const void * alpha, const void * a, library_data_t a_type, int lda,
+                           long long int stride_a, const void * b, library_data_t b_type, int ldb,
+                           long long int stride_b, const void * beta, void * c, library_data_t c_type, int ldc,
+                           long long int stride_c, int batch_size, library_data_t scaling_type) {
         if (scaling_type == library_data_t::real_float &&
             c_type == library_data_t::complex_float)
         {
@@ -2611,20 +2602,18 @@ namespace dpct
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_bfloat16, library_data_t::real_float):
         {
-            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16,
-                                    oneapi::mkl::bfloat16, float>(
-                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
-                beta, c, ldc, stride_c, batch_size);
+            detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, beta, c, ldc, stride_c,
+                batch_size);
             break;
         }
         case detail::get_type_combination_id(
             library_data_t::real_bfloat16, library_data_t::real_bfloat16,
             library_data_t::real_float, library_data_t::real_float):
         {
-            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float,
-                                    float>(q, a_trans, b_trans, m, n, k, alpha, a, lda,
-                                           stride_a, b, ldb, stride_b, beta, c, ldc,
-                                           stride_c, batch_size);
+            detail::gemm_batch_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b, beta, c, ldc, stride_c,
+                batch_size);
             break;
         }
 #endif