JuliaGPU · amontoison · Jul 21, 2025 · Jul 30, 2025 · Aug 8, 2025
diff --git a/.gitignore b/.gitignore
@@ -1,2 +1,8 @@
 LocalPreferences.toml
 Manifest.toml
+deps/onemkl_blas.cpp
+deps/onemkl_blas.h
+deps/onemkl_lapack.cpp
+deps/onemkl_lapack.h
+deps/onemkl_sparse.cpp
+deps/onemkl_sparse.h
diff --git a/deps/Project.toml b/deps/Project.toml
@@ -12,4 +12,4 @@ oneAPI_Level_Zero_Headers_jll = "f4bc562b-d309-54f8-9efb-476e56f0410d"
 oneAPI_Support_Headers_jll = "24f86df5-245d-5634-a4cc-32433d9800b3"
 
 [compat]
-oneAPI_Support_Headers_jll = "=2025.1.0"
+oneAPI_Support_Headers_jll = "=2025.2.0"
diff --git a/deps/build_local.jl b/deps/build_local.jl
@@ -45,7 +45,7 @@ if !isfile(joinpath(conda_dir, "condarc-julia.yml"))
     touch(joinpath(conda_dir, "conda-meta", "history"))
 end
 Conda.add_channel("https://software.repos.intel.com/python/conda/", conda_dir)
-Conda.add(["dpcpp_linux-64=2025.1.0", "mkl-devel-dpcpp=2025.1.0"], conda_dir)
+Conda.add(["dpcpp_linux-64=2025.2.0", "mkl-devel-dpcpp=2025.2.0"], conda_dir)
 
 Conda.list(conda_dir)
 

diff --git a/deps/generate_helpers.jl b/deps/generate_helpers.jl
@@ -1,6 +1,6 @@
 non_parametric_routines = ["init_matrix_handle", "release_matrix_handle", "set_matrix_property",
 "init_matmat_descr", "release_matmat_descr", "set_matmat_data", "get_matmat_data", "matmat",
-"omatcopy", "sort_matrix", "optimize_gemv", "optimize_trmv", "optimize_trsv", "optimize_trsm",
+"omatcopy", "sort_matrix", "optimize_gemv", "optimize_gemm", "optimize_trmv", "optimize_trsv", "optimize_trsm",
 "init_omatconvert_descr", "release_omatconvert_descr", "init_omatadd_descr", "release_omatadd_descr",
 "omatconvert_buffer_size", "omatconvert_analyze", "omatconvert_get_nnz", "omatconvert",
 "omatadd_buffer_size", "omatadd_analyze", "omatadd_get_nnz"]

diff --git a/deps/generate_interfaces.jl b/deps/generate_interfaces.jl
@@ -241,6 +241,7 @@ function generate_headers(library::String, filename::Vector{String}, output::Str
         # The routine "optimize_trsm" has two versions.
         suffix = ""
         (name_routine == "optimize_trsm") && occursin("columns", header) && (suffix = "_advanced")
+        (name_routine == "optimize_gemm") && occursin("columns", header) && (suffix = "_advanced")
         name_routine ∈ ("set_csr_data", "set_coo_data") && occursin("int64_t", header) && (suffix = "_64")
         occursin("batch", name_routine) && !occursin("**", header) && (suffix = "_strided")
 

diff --git a/deps/src/onemkl.cpp b/deps/src/onemkl.cpp
@@ -4038,6 +4038,16 @@ extern "C" int onemklXsparse_optimize_trsv(syclQueue_t device_queue, onemklUplo
    return 0;
 }
 
+extern "C" int onemklXsparse_optimize_gemm(syclQueue_t device_queue, onemklTranspose opA, matrix_handle_t A) {
+   auto status = oneapi::mkl::sparse::optimize_gemm(device_queue->val, convert(opA), (oneapi::mkl::sparse::matrix_handle_t) A, {});
+   return 0;
+}
+
+extern "C" int onemklXsparse_optimize_gemm_advanced(syclQueue_t device_queue, onemklLayout layout_val, onemklTranspose opA, onemklTranspose opB, matrix_handle_t A, int64_t columns) {
+   auto status = oneapi::mkl::sparse::optimize_gemm(device_queue->val, convert(layout_val), convert(opA), convert(opB), (oneapi::mkl::sparse::matrix_handle_t) A, columns, {});
+   return 0;
+}
+
 extern "C" int onemklXsparse_optimize_trsm(syclQueue_t device_queue, onemklUplo uplo_val, onemklTranspose opA, onemklDiag diag_val, matrix_handle_t A) {
    auto status = oneapi::mkl::sparse::optimize_trsm(device_queue->val, convert(uplo_val), convert(opA), convert(diag_val), (oneapi::mkl::sparse::matrix_handle_t) A, {});
    return 0;

diff --git a/deps/src/onemkl.h b/deps/src/onemkl.h
@@ -2842,6 +2842,13 @@ int onemklXsparse_optimize_trmv(syclQueue_t device_queue, onemklUplo uplo_val, o
 int onemklXsparse_optimize_trsv(syclQueue_t device_queue, onemklUplo uplo_val, onemklTranspose
                                 opA, onemklDiag diag_val, matrix_handle_t A);
 
+int onemklXsparse_optimize_gemm(syclQueue_t device_queue, onemklTranspose opA, matrix_handle_t
+                                A);
+
+int onemklXsparse_optimize_gemm_advanced(syclQueue_t device_queue, onemklLayout layout_val,
+                                         onemklTranspose opA, onemklTranspose opB,
+                                         matrix_handle_t A, int64_t columns);
+
 int onemklXsparse_optimize_trsm(syclQueue_t device_queue, onemklUplo uplo_val, onemklTranspose
                                 opA, onemklDiag diag_val, matrix_handle_t A);
 

diff --git a/lib/mkl/wrappers_sparse.jl b/lib/mkl/wrappers_sparse.jl
@@ -127,6 +127,18 @@ for (fname, elty) in ((:onemklSsparse_gemm, :Float32),
     end
 end
 
+function sparse_optimize_gemm!(trans::Char, A::oneSparseMatrixCSR)
+    queue = global_queue(context(A.nzVal), device(A.nzVal))
+    onemklXsparse_optimize_gemm(sycl_queue(queue), trans, A.handle)
+    return A
+end
+
+function sparse_optimize_gemm!(trans::Char, transB::Char, nrhs::Int, A::oneSparseMatrixCSR)
+    queue = global_queue(context(A.nzVal), device(A.nzVal))
+    onemklXsparse_optimize_gemm_advanced(sycl_queue(queue), 'C', trans, transB, A.handle, nrhs)
+    return A
+end
+
 for (fname, elty) in ((:onemklSsparse_symv, :Float32),
                       (:onemklDsparse_symv, :Float64),
                       (:onemklCsparse_symv, :ComplexF32),

diff --git a/lib/support/liboneapi_support.jl b/lib/support/liboneapi_support.jl
@@ -180,6 +180,14 @@ end
     ONEMKL_MATMAT_REQUEST_FINALIZE = 8
 end
 
+@cenum onemklOmatconvertAlg::UInt32 begin
+    ONEMKL_OMATCONVERT_DEFAULT_ALG = 0
+end
+
+@cenum onemklOmataddAlg::UInt32 begin
+    ONEMKL_OMATADD_DEFAULT_ALG = 0
+end
+
 mutable struct matrix_handle end
 
 const matrix_handle_t = Ptr{matrix_handle}
@@ -188,6 +196,14 @@ mutable struct matmat_descr end
 
 const matmat_descr_t = Ptr{matmat_descr}
 
+mutable struct omatconvert_descr end
+
+const omatconvert_descr_t = Ptr{omatconvert_descr}
+
+mutable struct omatadd_descr end
+
+const omatadd_descr_t = Ptr{omatadd_descr}
+
 function onemklHgemm_batch(device_queue, transa, transb, m, n, k, alpha, a, lda, b, ldb,
                            beta, c, ldc, group_count, group_size)
     @ccall liboneapi_support.onemklHgemm_batch(device_queue::syclQueue_t,
@@ -6596,6 +6612,26 @@ function onemklXsparse_release_matmat_descr(p_desc)
     @ccall liboneapi_support.onemklXsparse_release_matmat_descr(p_desc::Ptr{matmat_descr_t})::Cint
 end
 
+function onemklXsparse_init_omatconvert_descr(device_queue, p_descr)
+    @ccall liboneapi_support.onemklXsparse_init_omatconvert_descr(device_queue::syclQueue_t,
+                                                                  p_descr::Ptr{omatconvert_descr_t})::Cint
+end
+
+function onemklXsparse_release_omatconvert_descr(device_queue, descr)
+    @ccall liboneapi_support.onemklXsparse_release_omatconvert_descr(device_queue::syclQueue_t,
+                                                                     descr::omatconvert_descr_t)::Cint
+end
+
+function onemklXsparse_init_omatadd_descr(device_queue, p_omatadd_desc)
+    @ccall liboneapi_support.onemklXsparse_init_omatadd_descr(device_queue::syclQueue_t,
+                                                              p_omatadd_desc::Ptr{omatadd_descr_t})::Cint
+end
+
+function onemklXsparse_release_omatadd_descr(device_queue, omatadd_desc)
+    @ccall liboneapi_support.onemklXsparse_release_omatadd_descr(device_queue::syclQueue_t,
+                                                                 omatadd_desc::omatadd_descr_t)::Cint
+end
+
 function onemklXsparse_omatcopy(device_queue, transpose_val, spMat_in, spMat_out)
     @ccall liboneapi_support.onemklXsparse_omatcopy(device_queue::syclQueue_t,
                                                     transpose_val::onemklTranspose,
@@ -6658,6 +6694,22 @@ function onemklXsparse_optimize_trsv(device_queue, uplo_val, opA, diag_val, A)
                                                          A::matrix_handle_t)::Cint
 end
 
+function onemklXsparse_optimize_gemm(device_queue, opA, A)
+    @ccall liboneapi_support.onemklXsparse_optimize_gemm(device_queue::syclQueue_t,
+                                                         opA::onemklTranspose,
+                                                         A::matrix_handle_t)::Cint
+end
+
+function onemklXsparse_optimize_gemm_advanced(device_queue, layout_val, opA, opB, A,
+                                              columns)
+    @ccall liboneapi_support.onemklXsparse_optimize_gemm_advanced(device_queue::syclQueue_t,
+                                                                  layout_val::onemklLayout,
+                                                                  opA::onemklTranspose,
+                                                                  opB::onemklTranspose,
+                                                                  A::matrix_handle_t,
+                                                                  columns::Int64)::Cint
+end
+
 function onemklXsparse_optimize_trsm(device_queue, uplo_val, opA, diag_val, A)
     @ccall liboneapi_support.onemklXsparse_optimize_trsm(device_queue::syclQueue_t,
                                                          uplo_val::onemklUplo,

diff --git a/test/Project.toml b/test/Project.toml
@@ -18,4 +18,5 @@ StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
 Statistics = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 libigc_jll = "94295238-5935-5bd7-bb0f-b00942e9bdd5"
+oneAPI = "8f75cd03-7ff8-4ecb-9b8f-daf728133b1b"
 oneAPI_Support_jll = "b049733a-a71d-5ed3-8eba-7d323ac00b36"
diff --git a/test/onemkl.jl b/test/onemkl.jl
@@ -1132,6 +1132,7 @@ end
 
                     alpha = rand(T)
                     beta = rand(T)
+                    oneMKL.sparse_optimize_gemm!(transa, dA)
                     oneMKL.sparse_gemm!(transa, transb, alpha, dA, dB, beta, dC)
                     @test alpha * opa(A) * opb(B) + beta * C ≈ collect(dC)
                 end