Implement a workaround to gemm issue in OneMKL

Author: antonwolfy

dpnp/backend/extensions/blas/blas_py.cpp

@@ -127,6 +127,13 @@ PYBIND11_MODULE(_blas_impl, m)
               py::arg("resultC"), py::arg("depends") = py::list());
     }
 
+    {
+        m.def("_is_lnl_arl_architecture", &blas_ns::_is_lnl_arl_architecture,
+              "Return ``True`` if SYCL device belongs to either Lunar Lake or "
+              "Arrow Lake architecture",
+              py::arg("device"));
+    }
+
     {
         m.def("_gemm_batch", &blas_ns::gemm_batch,
               "Call `gemm_batch` from OneMKL BLAS library to compute "

dpnp/backend/extensions/blas/gemm.cpp

@@ -323,6 +323,25 @@ std::tuple<sycl::event, sycl::event, bool>
     return std::make_tuple(args_ev, gemm_ev, is_row_major);
 }
 
+bool _is_lnl_arl_architecture(sycl::device &dev)
+{
+#if !defined(USE_ONEMKL_CUBLAS)
+    if (dev.ext_oneapi_architecture_is(
+            sycl::ext::oneapi::experimental::architecture::
+                intel_gpu_20_4_4)) /* Lunar Lake */
+    {
+        return true;
+    }
+    else if (dev.ext_oneapi_architecture_is(
+                 sycl::ext::oneapi::experimental::architecture::
+                     intel_gpu_12_74_4)) /* Arrow Lake */
+    {
+        return true;
+    }
+#endif // !defined(USE_ONEMKL_CUBLAS)
+    return false;
+}
+
 template <typename fnT, typename Tab, typename Tc>
 struct GemmContigFactory
 {

dpnp/backend/extensions/blas/gemm.hpp

@@ -39,6 +39,8 @@ extern std::tuple<sycl::event, sycl::event, bool>
          const dpctl::tensor::usm_ndarray &resultC,
          const std::vector<sycl::event> &depends);
 
+extern bool _is_lnl_arl_architecture(sycl::device &dev);
+
 extern std::tuple<sycl::event, sycl::event, bool>
     gemm_batch(sycl::queue &exec_q,
                const dpctl::tensor::usm_ndarray &matrixA,

dpnp/dpnp_utils/dpnp_utils_linearalgebra.py

@@ -894,6 +894,35 @@ def dpnp_matmul(
                 )
                 _manager.add_event_pair(ht_ev, gemv_ev)
             elif call_flag == "gemm":
+                # MKLD-17976: due to known issue in OneMKL on Lunar Lake and
+                # Arrow Lake architectures, it forces to implement a temporary
+                # workaround with extra copying of an input array in case when
+                # it has a small size and non-zero offset
+                # TODO: remove the workaraound once OneMKL issue is resolved
+                if (
+                    compute_dtype != dpnp.float32
+                    and bi._is_lnl_arl_architecture(exec_q.get_sycl_device())
+                ):
+
+                    def _need_to_copy(a):
+                        a_usm = dpnp.get_usm_ndarray(a)
+                        if a_usm._element_offset > 0 and a_usm.size < 16:
+                            return True
+                        return False
+
+                    x1 = _copy_array(
+                        x1,
+                        copy_flag=_need_to_copy(x1),
+                        dtype=compute_dtype,
+                        order=res_order,
+                    )
+                    x2 = _copy_array(
+                        x2,
+                        copy_flag=_need_to_copy(x2),
+                        dtype=compute_dtype,
+                        order=res_order,
+                    )
+
                 result = _gemm_matmul(
                     exec_q,
                     x1,