SYCL/oneMath backend

.github/workflows/ci.yml

@@ -64,14 +64,14 @@ jobs:
         cmake "--preset=ci-${{ matrix.os }}"
 
     - name: Build
-      run: cmake --build build --config Release -j 2
+      run: cmake --build build --config Release
 
     - name: Install
       run: cmake --install build --config Release --prefix prefix
 
     - name: Test
       working-directory: build
-      run: ctest --output-on-failure --no-tests=ignore -C Release -j 2
+      run: ctest --output-on-failure --no-tests=ignore -C Release
 
     - name: Upload
       uses: actions/upload-artifact@v4
@@ -90,7 +90,7 @@ jobs:
         cmake -G "Unix Makefiles" -S examples/gprat_cpp -B build_examples -DCMAKE_BUILD_TYPE=Release -DCMAKE_PREFIX_PATH=$PWD/prefix -DUSE_MKL=ON
 
     - name: Build example project
-      run: cmake --build build_examples --config Release -j 2
+      run: cmake --build build_examples --config Release
 
     # See: https://github.com/spack/setup-spack?tab=readme-ov-file#example-caching-your-own-binaries-for-public-repositories
     - name: Push packages and update index

.gitignore

@@ -195,7 +195,14 @@ examples/*/output.csv
 examples/gprat_*/include/
 examples/gpflow_reference/GPflow
 /.vscode
+benchmark_results_*
 
 # Editor related files
 ltex*
 compile_commands.json
+
+# Build files
+build*
+
+# Ignore folder
+ignore

CMakeLists.txt

@@ -1,3 +1,7 @@
+# Top-level CMakeLists.txt
+
+# CMake initialization
+# ##############################################################################
 cmake_minimum_required(VERSION 3.23)
 
 project(
@@ -9,21 +13,28 @@ project(
 
 include(CMakeDependentOption)
 
-# What to build?
+# Build options
+# ##############################################################################
 option(GPRAT_BUILD_CORE "Build the core library" ON)
+
 cmake_dependent_option(GPRAT_BUILD_BINDINGS "Build the Python bindings" ON
                        "GPRAT_BUILD_CORE" OFF)
+
 cmake_dependent_option(
   GPRAT_ENABLE_EXAMPLES "Build example applications as well?"
   ${PROJECT_IS_TOP_LEVEL} "GPRAT_BUILD_CORE" OFF)
+
 cmake_dependent_option(GPRAT_ENABLE_TESTS "Build unit and integration tests"
                        ${PROJECT_IS_TOP_LEVEL} "GPRAT_BUILD_CORE" OFF)
+
 cmake_dependent_option(GPRAT_ENABLE_MKL "Enable support for Intel oneMKL"
                        ${PROJECT_IS_TOP_LEVEL} "GPRAT_BUILD_CORE" OFF)
 
 option(GPRAT_ENABLE_FORMAT_TARGETS "Enable clang-format / cmake-format targets"
        ${PROJECT_IS_TOP_LEVEL})
 
+# Format settings
+# ##############################################################################
 if(GPRAT_ENABLE_FORMAT_TARGETS)
   set(CMAKE_FORMAT_EXCLUDE "^external_ports/")
 
@@ -39,44 +50,55 @@ if(GPRAT_ENABLE_FORMAT_TARGETS)
   endif()
 endif()
 
+# GNU install dirs
+# ##############################################################################
 if(NOT CMAKE_SKIP_INSTALL_RULES)
   # Our installs follow the standard GNU directory layout. This include needs to
   # come first since we need the CMAKE_INSTALL_* in the CMakeLists.txt of each
   # target.
   include(GNUInstallDirs)
 endif()
 
+# Building the GPRAT core
+# ##############################################################################
 if(GPRAT_BUILD_CORE)
+
+  # MKL backend
   if(GPRAT_ENABLE_MKL)
-    # Try to find Intel oneMKL
     set(MKL_INTERFACE_FULL "intel_lp64")
     set(MKL_THREADING "sequential")
     find_package(MKL CONFIG REQUIRED)
-
     if(MKL_FOUND)
       message(STATUS "Intel oneMKL Library found")
     else()
-      message(FATAL_ERROR "No BLAS Library found")
+      message(FATAL_ERROR "Intel oneMKL Library NOT found")
     endif()
+
+    # OpenBLAS backend
   else()
-    # Try to find OpenBLAS
     find_library(OpenBLAS_LIB NAMES openblas REQUIRED)
-
     if(OpenBLAS_LIB)
       message(STATUS "OpenBLAS Library found at ${OpenBLAS_LIB}")
     else()
       message(FATAL_ERROR "No BLAS Library found")
     endif()
   endif()
 
+  # HPX
   find_package(HPX REQUIRED)
 
+  # Add core subdirectiory
   add_subdirectory(core)
+
+  # Bindings subdirectory
   if(GPRAT_BUILD_BINDINGS)
     add_subdirectory(bindings)
   endif()
+
 endif()
 
+# Installation
+# ##############################################################################
 if(NOT CMAKE_SKIP_INSTALL_RULES AND GPRAT_BUILD_CORE)
   include(CMakePackageConfigHelpers)
 
@@ -116,11 +138,18 @@ if(NOT CMAKE_SKIP_INSTALL_RULES AND GPRAT_BUILD_CORE)
   endif()
 endif()
 
+# Examples
+# ##############################################################################
 if(GPRAT_ENABLE_EXAMPLES)
   add_subdirectory(examples/gprat_cpp)
 endif()
 
+# Tests
+# ##############################################################################
 if(GPRAT_ENABLE_TESTS)
   enable_testing()
   add_subdirectory(test)
 endif()
+
+# End of file
+# ##############################################################################

CMakePresets.json

@@ -192,8 +192,8 @@
       }
     },
     {
-      "name": "dev-linux-gpu",
-      "binaryDir": "${sourceDir}/build/dev-linux-gpu",
+      "name": "dev-linux-cuda",
+      "binaryDir": "${sourceDir}/build/dev-linux-cuda",
       "inherits": ["dev-linux"],
       "cacheVariables": {
         "GPRAT_WITH_CUDA": "ON",
@@ -202,15 +202,31 @@
       }
     },
     {
-      "name": "release-linux-gpu",
-      "binaryDir": "${sourceDir}/build/release-linux-gpu",
+      "name": "release-linux-cuda",
+      "binaryDir": "${sourceDir}/build/release-linux-cuda",
       "inherits": ["release-linux"],
       "cacheVariables": {
         "GPRAT_WITH_CUDA": "ON",
         "GPRAT_APEX_STEPS": "OFF",
         "GPRAT_APEX_CHOLESKY": "OFF"
       }
+    },
+    {
+      "name": "dev-linux-sycl",
+      "binaryDir": "${sourceDir}/build/dev-linux-sycl",
+      "inherits": ["dev-linux"],
+      "cacheVariables": {
+        "GPRAT_WITH_SYCL": "ON"
+      }
+    },
+    {
+    "name": "release-linux-sycl",
+    "binaryDir": "${sourceDir}/build/release-linux-sycl",
+    "inherits": ["release-linux"],
+    "cacheVariables": {
+      "GPRAT_WITH_SYCL": "ON"
     }
+  }
   ],
   "buildPresets": [
     {
@@ -224,13 +240,23 @@
       "configuration": "Release"
     },
     {
-      "name": "dev-linux-gpu",
-      "configurePreset": "dev-linux-gpu",
+      "name": "dev-linux-cuda",
+      "configurePreset": "dev-linux-cuda",
+      "configuration": "Debug"
+    },
+    {
+      "name": "release-linux-cuda",
+      "configurePreset": "release-linux-cuda",
+      "configuration": "Release"
+    },
+    {
+      "name": "dev-linux-sycl",
+      "configurePreset": "dev-linux-sycl",
       "configuration": "Debug"
     },
     {
-      "name": "release-linux-gpu",
-      "configurePreset": "release-linux-gpu",
+      "name": "release-linux-sycl",
+      "configurePreset": "release-linux-sycl",
       "configuration": "Release"
     }
   ],
@@ -258,8 +284,8 @@
       }
     },
     {
-      "name": "dev-linux-gpu",
-      "configurePreset": "dev-linux-gpu",
+      "name": "dev-linux-cuda",
+      "configurePreset": "dev-linux-cuda",
       "configuration": "Debug",
       "output": {
         "outputOnFailure": true
@@ -269,8 +295,8 @@
       }
     },
     {
-      "name": "release-linux-gpu",
-      "configurePreset": "release-linux-gpu",
+      "name": "release-linux-cuda",
+      "configurePreset": "release-linux-cuda",
       "configuration": "Release",
       "output": {
         "outputOnFailure": true

README.md

@@ -11,7 +11,8 @@ code.
 
 GPRat depends on [HPX](https://hpx-docs.stellar-group.org/latest/html/index.html) for asynchronous task-based parallelization.
 Furthermore, for CPU-only BLAS computation GPRat requires [OpenBLAS](http://www.openmathlib.org/OpenBLAS/) or [MKL](https://www.intel.com/content/www/us/en/developer/tools/oneapi/onemkl.html).
-A [CUDA](https://developer.nvidia.com/cuda-toolkit) installation is required for GPU-only BLAS computations.
+A [CUDA](https://developer.nvidia.com/cuda-toolkit) installation is required for GPU-only BLAS computations on NVIDIA hardware.
+For GPU computations on Intel and AMD hardware, GPRat supports [SYCL](https://www.khronos.org/sycl/) via [oneMath](https://github.com/uxlfoundation/oneMath).
 
 ### Install dependencies
 
@@ -34,8 +35,8 @@ ctest --preset=dev-linux
 
 As a developer, you may create a `CMakeUserPresets.json` file at the root of the project that contains additional
 presets local to your machine.
-In addition to the build configuration `dev-linux`, there are `release-linux`, `dev-linux-gpu`, and `release-linux-gpu`.
-The configurations suffixed with `-gpu` build the library with CUDA.
+In addition to the build configuration `dev-linux`, there are `release-linux`, `dev-linux-gpu`, `release-linux-gpu`, `dev-linux-sycl`, and `release-linux-sycl`.
+The configurations suffixed with `-gpu` build the library with CUDA for NVIDIA GPUs, and those suffixed with `-sycl` build it with SYCL support for Intel and AMD GPUs.
 
 GPRat can be build with or without Python bindings.
 The following options can be set to include / exclude parts of the project:
@@ -47,7 +48,8 @@ The following options can be set to include / exclude parts of the project:
 | GPRAT_ENABLE_FORMAT_TARGETS    | Enable/Disable code formatting helper targets                                        | ON if top-level |
 | GPRAT_ENABLE_EXAMPLES          | Enable/Disable example projects                                                      | ON if top-level |
 | GPRAT_USE_MKL                  | Enable/Disable usage of MKL library                                                  | OFF             |
-| GPRAT_WITH_CUDA                | Enable/disable compilation with CUDA support                                         | OFF             |
+| GPRAT_WITH_CUDA                | Enable/disable compilation with CUDA support (NVIDIA GPUs)                           | OFF             |
+| GPRAT_WITH_SYCL                | Enable/disable compilation with SYCL support (Intel and AMD GPUs via oneMath)        | OFF             |
 | GPRAT_APEX_STEPS               | Enable/disable compilation for steps duration measurement with APEX                  | OFF             |
 | GPRAT_APEX_CHOLESKY            | Enable/disable compilation for measuring cholesky assembly and computation with APEX | OFF             |
 
@@ -103,7 +105,10 @@ We specifically thank the follow contributors:
   [Optimization, Python bindings and reference implementations](https://doi.org/10.48550/arXiv.2505.00136).
 
 - [Henrik Möllmann](https://www.linkedin.com/in/moellh/):
-  [Accelerator Support](tbd.).
+  [CUDA backend via cuBLAS/cuSOLVER](tbd.).
+
+- Marcel Graf:
+  [SYCL backend via oneMath](tbd.).
 
 ## How To Cite
 

Testing/Temporary/CTestCostData.txt

@@ -0,0 +1 @@
+---

bindings/CMakeLists.txt

@@ -20,8 +20,11 @@ endif()
 
 # Option for GPU support with CUDA, cuSolver, cuBLAS
 option(GPRAT_WITH_CUDA "Enable GPU support with CUDA, cuSolver, cuBLAS" OFF)
+option(GPRAT_WITH_SYCL "Enable SYCL support with oneMath" OFF)
+
 # Pass variable to C++ code
-add_compile_definitions(GPRAT_WITH_CUDA=$<BOOL:${GPRAT_WITH_CUDA}>)
+add_compile_definitions(GPRAT_WITH_CUDA=$<BOOL:${GPRAT_WITH_CUDA}>
+                        GPRAT_WITH_SYCL=$<BOOL:${GPRAT_WITH_SYCL}>)
 
 # Option for steps duration measurement with APEX
 option(GPRAT_APEX_STEPS "Enable measuring duration of steps with APEX" OFF)

bindings/gprat_py.cpp

@@ -79,7 +79,7 @@ n_streams to a value enables computations on the GPU.
     trainable (list): List of booleans for trainable hyperparameters. Default is
         {true, true, true}.
     gpu_id (int): ID of the GPU to use. Default is 0.
-    n_streams (int): Number of streams for GPU computation. Default is 1.
+    n_units (int): Number of streams/queues for GPU computation. Default is 1.
              )pbdoc")
 
         // GPU constructor
@@ -100,11 +100,11 @@ n_streams to a value enables computations on the GPU.
              py::arg("kernel_params") = std::vector<double>{ 1.0, 1.0, 0.1 },
              py::arg("trainable") = std::vector<bool>{ true, true, true },
              py::arg("gpu_id") = 0,
-             py::arg("n_streams") = 1,
+             py::arg("n_units") = 1,
              R"pbdoc(
 Create Gaussian Process including its data, hyperparameters, and target. By
 default, the calculations are performed on the CPU. Setting at least gpu_id or
-n_streams to a value enables computations on the GPU.
+n_units to a value enables computations on the GPU.
 
 Parameters:
     input_data (list): Input data for the GP.
@@ -117,7 +117,7 @@ n_streams to a value enables computations on the GPU.
     trainable (list): List of booleans for trainable hyperparameters. Default is
         {true, true, true}.
     gpu_id (int): ID of the GPU to use. Default is 0.
-    n_streams (int): Number of streams for GPU computation. Default is 1.
+    n_units (int): Number of streams/queues for GPU computation. Default is 1.
 
              )pbdoc")
         .def_readwrite("n_reg", &gprat::GP::n_reg)

bindings/utils_py.cpp

@@ -103,6 +103,7 @@ void init_utils(py::module &m)
     m.def("stop_hpx", &utils::stop_hpx_runtime);
 
     m.def("compiled_with_cuda", &utils::compiled_with_cuda, "Check if the code was compiled with CUDA support");
+    m.def("compiled_with_sycl", &utils::compiled_with_sycl, "Check if the code was compiled with SYCL support");
 
     m.def("print_available_gpus", &gprat::print_available_gpus, "Print available GPUs with their properties");
     m.def("gpu_count", &gprat::gpu_count, "Return the number of available GPUs");

boilerplate-benchmark.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+#
+#SBATCH --job-name=gprat
+#SBATCH --output=output_gprat.txt
+#SBATCH --time=48:00:00
+
+source ~/Setup_Scripts/setup_breyerml.sh
+
+if [[ "$HOSTNAME" == "simcl1n1" || "$HOSTNAME" == "simcl1n2" ]]; then
+
+    ./execute-benchmark.sh yes yes yes yes gpu nvidia cuda
+    ./execute-benchmark.sh no no yes yes gpu nvidia sycl
+
+elif [[ "$HOSTNAME" == "simcl1n3" ]]; 
+then
+
+    ./execute-benchmark.sh yes yes yes yes gpu amd
+
+elif [[ "$HOSTNAME" == "simcl1n4" ]]; 
+then
+
+    ./execute-benchmark.sh yes yes yes yes cpu cpu
+
+fi