Add Nvidia lapack support (cusolver backend) (#154)

* Add Nvidia lapack support (cusolver backend)

* add cusolver gebrd, geqrf, getrf, getrs, gesvd

* Add additional cusolver functions

* add cusolver potrf and potrs batch implementation

* Fix malloc_device

* update README

Author: AidanBeltonS

CMakeLists.txt

@@ -46,6 +46,9 @@ if(ENABLE_MKLCPU_BACKEND)
   option(ENABLE_MKLCPU_THREAD_TBB "" ON)
 endif()
 option(ENABLE_CUBLAS_BACKEND "" OFF)
+
+option(ENABLE_CUSOLVER_BACKEND "" OFF)
+
 option(ENABLE_CURAND_BACKEND "" OFF)
 option(ENABLE_NETLIB_BACKEND "" OFF)
 set(ONEMKL_SYCL_IMPLEMENTATION "dpc++" CACHE STRING "Name of the SYCL compiler")
@@ -65,7 +68,8 @@ if(ENABLE_MKLCPU_BACKEND
   list(APPEND DOMAINS_LIST "blas")
 endif()
 if(ENABLE_MKLCPU_BACKEND
-        OR ENABLE_MKLGPU_BACKEND)
+        OR ENABLE_MKLGPU_BACKEND
+	OR ENABLE_CUSOLVER_BACKEND)
   list(APPEND DOMAINS_LIST "lapack")
 endif()
 if(ENABLE_MKLCPU_BACKEND

README.md

@@ -31,6 +31,10 @@ oneMKL is part of [oneAPI](https://oneapi.io).
             <td align="center"><a href="https://developer.nvidia.com/cublas"> NVIDIA cuBLAS</a> for NVIDIA GPU </td>
             <td align="center">NVIDIA GPU</td>
         </tr>
+	<tr>
+            <td align="center"><a href="https://developer.nvidia.com/cusolver"> NVIDIA cuSOLVER</a> for NVIDIA GPU </td>
+            <td align="center">NVIDIA GPU</td>
+	</tr>
         <tr>
             <td align="center"><a href="https://developer.nvidia.com/curand"> NVIDIA cuRAND</a> for NVIDIA GPU </td>
             <td align="center">NVIDIA GPU</td>
@@ -159,6 +163,11 @@ Supported domains: BLAS, LAPACK, RNG
             <td align="center">Intel GPU</td>
             <td align="center">Dynamic, Static</td>
         </tr>
+        <tr>
+            <td align="center">NVIDIA GPU</td>
+            <td align="center">NVIDIA cuSOLVER</td>
+            <td align="center">Dynamic, Static</td>
+        </tr>
         <tr>
             <td rowspan=3 align="center">RNG</td>
             <td align="center">x86 CPU</td>
@@ -571,12 +580,18 @@ ctest
 cmake --install . --prefix <path_to_install_dir>
 ```
 
-To build with the cuRAND backend instead simply replace:
+To build with the cuSOLVER or cuRAND backend instead simply replace:
 ```bash
 -DENABLE_CUBLAS_BACKEND=True   \
 ```
 
 With:
+```bash
+-DENABLE_CUSOLVER_BACKEND=True   \
+```
+
+or
+
 ```bash
 -DENABLE_CURAND_BACKEND=True   \
 ```
@@ -592,6 +607,7 @@ build_shared_libs        | BUILD_SHARED_LIBS        | True, False         | True
 enable_mklcpu_backend    | ENABLE_MKLCPU_BACKEND    | True, False         | True
 enable_mklgpu_backend    | ENABLE_MKLGPU_BACKEND    | True, False         | True
 *Not Supported*          | ENABLE_CUBLAS_BACKEND    | True, False         | False
+*Not Supported*          | ENABLE_CUSOLVER_BACKEND  | True, False         | False
 *Not Supported*          | ENABLE_CURAND_BACKEND    | True, False         | False
 *Not Supported*          | ENABLE_NETLIB_BACKEND    | True, False         | False
 enable_mklcpu_thread_tbb | ENABLE_MKLCPU_THREAD_TBB | True, False         | True

cmake/FindCompiler.cmake

@@ -33,13 +33,13 @@ if(is_dpcpp)
   if(UNIX)
     set(UNIX_INTERFACE_COMPILE_OPTIONS -fsycl)
     set(UNIX_INTERFACE_LINK_OPTIONS -fsycl)
-    if(ENABLE_CURAND_BACKEND)
+    if(ENABLE_CURAND_BACKEND OR ENABLE_CUSOLVER_BACKEND)
       list(APPEND UNIX_INTERFACE_COMPILE_OPTIONS
-        -fsycl-targets=nvptx64-nvidia-cuda-sycldevice -fsycl-unnamed-lambda)
+        -fsycl-targets=nvptx64-nvidia-cuda -fsycl-unnamed-lambda)
       list(APPEND UNIX_INTERFACE_LINK_OPTIONS
-        -fsycl-targets=nvptx64-nvidia-cuda-sycldevice)
+        -fsycl-targets=nvptx64-nvidia-cuda)
     endif()
-    if(ENABLE_CURAND_BACKEND)
+    if(ENABLE_CURAND_BACKEND OR ENABLE_CUSOLVER_BACKEND)
       set_target_properties(ONEMKL::SYCL::SYCL PROPERTIES
         INTERFACE_COMPILE_OPTIONS "${UNIX_INTERFACE_COMPILE_OPTIONS}"
         INTERFACE_LINK_OPTIONS "${UNIX_INTERFACE_LINK_OPTIONS}"

cmake/FindcuSOLVER.cmake

@@ -0,0 +1,51 @@
+#==========================================================================
+#  Copyright (C) Codeplay Software Limited
+#  Licensed under the Apache License, Version 2.0 (the "License");
+#  you may not use this file except in compliance with the License.
+#  You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+#  For your convenience, a copy of the License has been included in this
+#  repository.
+#
+#  Unless required by applicable law or agreed to in writing, software
+#  distributed under the License is distributed on an "AS IS" BASIS,
+#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#  See the License for the specific language governing permissions and
+#  limitations under the License.
+#
+#=========================================================================
+
+find_package(CUDA 10.0 REQUIRED)
+get_filename_component(SYCL_BINARY_DIR ${CMAKE_CXX_COMPILER} DIRECTORY)
+# the OpenCL include file from cuda is opencl 1.1 and it is not compatible with DPC++
+# the OpenCL include headers 1.2 onward is required. This is used to bypass NVIDIA OpenCL headers
+find_path(OPENCL_INCLUDE_DIR CL/cl.h OpenCL/cl.h 
+HINTS 
+${OPENCL_INCLUDE_DIR}
+${SYCL_BINARY_DIR}/../include/sycl/
+)
+# this is work around to avoid duplication half creation in both cuda and SYCL
+add_compile_definitions(CUDA_NO_HALF)
+
+find_package(Threads REQUIRED)
+
+include(FindPackageHandleStandardArgs)
+find_package_handle_standard_args(cuSOLVER
+    REQUIRED_VARS
+	CUDA_TOOLKIT_INCLUDE
+	CUDA_cusolver_LIBRARY
+        CUDA_LIBRARIES
+        CUDA_CUDA_LIBRARY
+        OPENCL_INCLUDE_DIR
+)
+if(NOT TARGET ONEMKL::cuSOLVER::cuSOLVER)
+  add_library(ONEMKL::cuSOLVER::cuSOLVER SHARED IMPORTED)
+  set_target_properties(ONEMKL::cuSOLVER::cuSOLVER PROPERTIES
+      IMPORTED_LOCATION ${CUDA_cusolver_LIBRARY}
+      INTERFACE_INCLUDE_DIRECTORIES "${OPENCL_INCLUDE_DIR};${CUDA_TOOLKIT_INCLUDE}"
+      INTERFACE_LINK_LIBRARIES "Threads::Threads;${CUDA_CUDA_LIBRARY};${CUDA_LIBRARIES}"
+  )
+
+endif()

include/oneapi/mkl/detail/backend_selector_predicates.hpp

@@ -81,6 +81,19 @@ inline void backend_selector_precondition<backend::cublas>(cl::sycl::queue& queu
 #endif
 }
 
+template <>
+inline void backend_selector_precondition<backend::cusolver>(cl::sycl::queue& queue) {
+#ifndef ONEMKL_DISABLE_PREDICATES
+    unsigned int vendor_id =
+        static_cast<unsigned int>(queue.get_device().get_info<cl::sycl::info::device::vendor_id>());
+    if (!(queue.get_device().is_gpu() && vendor_id == NVIDIA_ID)) {
+        throw unsupported_device("",
+                                 "backend_selector<backend::" + backend_map[backend::cusolver] + ">",
+                                 queue.get_device());
+    }
+#endif
+}
+
 } // namespace mkl
 } // namespace oneapi
 

include/oneapi/mkl/detail/backends.hpp

@@ -26,14 +26,15 @@
 namespace oneapi {
 namespace mkl {
 
-enum class backend { mklcpu, mklgpu, cublas, curand, netlib, unsupported };
+enum class backend { mklcpu, mklgpu, cublas, cusolver, curand, netlib, unsupported };
 
 typedef std::map<backend, std::string> backendmap;
 
 static backendmap backend_map = {
     { backend::mklcpu, "mklcpu" }, { backend::mklgpu, "mklgpu" },
-    { backend::cublas, "cublas" }, { backend::curand, "curand" },
-    { backend::netlib, "netlib" }, { backend::unsupported, "unsupported" }
+    { backend::cublas, "cublas" }, { backend::cusolver, "cusolver" },
+    { backend::curand, "curand" }, { backend::netlib, "netlib" }, 
+    { backend::unsupported, "unsupported" }
 };
 
 } //namespace mkl

include/oneapi/mkl/detail/backends_table.hpp

@@ -74,6 +74,12 @@ static std::map<domain, std::map<device, std::vector<const char*>>> libraries =
           {
 #ifdef ENABLE_MKLGPU_BACKEND
               LIB_NAME("lapack_mklgpu")
+#endif
+	  } },
+	{ device::nvidiagpu,
+	  {
+#ifdef ENABLE_CUSOLVER_BACKEND
+              LIB_NAME("lapack_cusolver")
 #endif
           } } } },
 

include/oneapi/mkl/lapack.hpp

@@ -27,5 +27,8 @@
 #ifdef ENABLE_MKLGPU_BACKEND
 #include "oneapi/mkl/lapack/detail/mklgpu/lapack_ct.hpp"
 #endif
+#ifdef ENABLE_CUSOLVER_BACKEND
+#include "oneapi/mkl/lapack/detail/cusolver/lapack_ct.hpp"
+#endif
 
 #include "oneapi/mkl/lapack/detail/lapack_rt.hpp"

include/oneapi/mkl/lapack/detail/cusolver/lapack_ct.hpp

@@ -0,0 +1,44 @@
+/***************************************************************************
+*  Copyright (C) Codeplay Software Limited
+*  Licensed under the Apache License, Version 2.0 (the "License");
+*  you may not use this file except in compliance with the License.
+*  You may obtain a copy of the License at
+*
+*      http://www.apache.org/licenses/LICENSE-2.0
+*
+*  For your convenience, a copy of the License has been included in this
+*  repository.
+*
+*  Unless required by applicable law or agreed to in writing, software
+*  distributed under the License is distributed on an "AS IS" BASIS,
+*  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+*  See the License for the specific language governing permissions and
+*  limitations under the License.
+*
+**************************************************************************/
+
+#ifndef _DETAIL_CUSOLVER_LAPACK_CT_HPP_
+#define _DETAIL_CUSOLVER_LAPACK_CT_HPP_
+
+#include <CL/sycl.hpp>
+#include <complex>
+#include <cstdint>
+
+#include "oneapi/mkl/types.hpp"
+#include "oneapi/mkl/lapack/types.hpp"
+#include "oneapi/mkl/detail/backend_selector.hpp"
+#include "oneapi/mkl/lapack/detail/cusolver/onemkl_lapack_cusolver.hpp"
+
+namespace oneapi {
+namespace mkl {
+namespace lapack {
+
+#define LAPACK_BACKEND cusolver
+#include "lapack_ct.hxx"
+#undef LAPACK_BACKEND
+
+} // namespace lapack
+} // namespace mkl
+} // namespace oneapi
+
+#endif //_DETAIL_CUSOLVER_LAPACK_CT_HPP_