Remove support for Nvidia & AMD GPU, because the oneAPI plugin for Nvidia & AMD GPU is unavailable: download/installation channels are out of work. (#19246)

User can't build up the software for Nvidia & AMD GPU. rm the oneMath since it is only used in NV and AMD code path.
2026-02-02 21:06:21 +08:00
parent 4d5e972673
commit bf38346d13
7 changed files with 60 additions and 274 deletions
@@ -15,17 +15,9 @@

 #include <sycl/sycl.hpp>
 #include <sycl/half_type.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 <map>

 #include "ggml.h"

@@ -91,32 +83,13 @@ inline std::string get_device_backend_and_type(const sycl::device &device) {
 }

 template <typename Ts> struct matrix_info_t {
-    oneapi::math::transpose transpose_info[2];
+    oneapi::mkl::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;
@@ -1734,7 +1707,7 @@ namespace dpct
    namespace detail
    {
    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,
+    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);
@@ -1742,7 +1715,7 @@ namespace dpct
        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,
+        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);
    }

@@ -1774,7 +1747,7 @@ namespace dpct
        };

        template <class Ta, class Tb, class Tc, class Ts>
-        inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans,
+        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, const void ** b,
                                    int ldb, const void * beta, void ** c, int ldc, int batch_size,
                                    matrix_info_t<float> * matrix_info) {
@@ -1793,8 +1766,8 @@ namespace dpct
            matrix_info->ld_info[2] = ldc;
            matrix_info->groupsize_info = batch_size;

-            sycl::event e = oneapi::math::blas::column_major::gemm_batch(
-                get_onemath_backend(q), matrix_info->transpose_info, matrix_info->transpose_info + 1,
+            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,
@@ -1803,7 +1776,7 @@ namespace dpct
        }

        template <class Ta, class Tb, class Tc, class Ts>
-        inline void gemm_batch_impl(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans,
+        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) {
@@ -1812,7 +1785,7 @@ namespace dpct
            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_batch(get_onemath_backend(q), a_trans, b_trans, m, n, k, alpha_value,
+            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);
        }
@@ -2299,7 +2272,7 @@ namespace dpct
                           sycl::range<3>(x, y, 1), direction);
    }

-    inline void gemm(sycl::queue & q, oneapi::math::transpose a_trans, oneapi::math::transpose b_trans, int m, int n,
+    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) {
@@ -2366,7 +2339,7 @@ namespace dpct
            library_data_t::real_bfloat16, library_data_t::real_bfloat16,
            library_data_t::real_float, library_data_t::real_float):
        {
-            detail::gemm_impl<oneapi::math::bfloat16, oneapi::math::bfloat16, float, 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);
            break;
        }
@@ -2405,7 +2378,7 @@ 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::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, float>(
+            detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float>(
                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
            break;
        }
@@ -2447,7 +2420,7 @@ 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::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
+    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,
                           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,
@@ -2485,7 +2458,7 @@ 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::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, 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, b, ldb, beta, c, ldc, batch_size, matrix_info);
            break;
        }
@@ -2493,7 +2466,7 @@ 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::math::bfloat16, oneapi::math::bfloat16, float, float>(
+            detail::gemm_batch_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, batch_size, matrix_info);
            break;
        }
@@ -2569,7 +2542,7 @@ 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::math::transpose a_trans, oneapi::math::transpose b_trans, int m,
+    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,
@@ -2642,7 +2615,7 @@ 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::math::bfloat16, oneapi::math::bfloat16, oneapi::math::bfloat16, 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);
            break;
@@ -2651,7 +2624,7 @@ 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::math::bfloat16, oneapi::math::bfloat16, float, 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);
            break;