Make: USE_BLAS option

Author: ashvardanian

.vscode/settings.json

@@ -64,6 +64,7 @@
     "Klemens",
     "Kohlhoff",
     "Kulukundis",
+    "LAPACK",
     "Lelbach",
     "Lemire",
     "Lib",

CMakeLists.txt

@@ -72,11 +72,22 @@ else ()
     set(_SHOULD_USE_INTEL_TBB OFF)
 endif ()
 
+# Probe for BLAS support
+set(CMAKE_FIND_LIBRARY_PREFIXES ";lib")
+find_package(BLAS QUIET)
+if (BLAS_FOUND)
+    set(_SHOULD_USE_BLAS ON)
+else ()
+    set(_SHOULD_USE_BLAS OFF)
+endif ()
+
 option(USE_INTEL_TBB "Use Intel TBB for parallel STL algorithms" ${_SHOULD_USE_INTEL_TBB})
 option(USE_NVIDIA_CCCL "Use Nvidia CCCL for CUDA acceleration" ${_SHOULD_USE_NVIDIA_CCCL})
+option(USE_BLAS "Use BLAS for linear algebra" ${_SHOULD_USE_BLAS})
 
 message(STATUS "USE_INTEL_TBB: ${USE_INTEL_TBB}")
 message(STATUS "USE_NVIDIA_CCCL: ${USE_NVIDIA_CCCL}")
+message(STATUS "USE_BLAS: ${USE_BLAS}")
 
 # ------------------------------------------------------------------------------
 # Dependencies
@@ -102,13 +113,14 @@ set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -w")
 # ~~~
 #
 # Moreover, CMake sometimes fails to find it on Windows: https://stackoverflow.com/a/78335726/2766161
-set(CMAKE_FIND_LIBRARY_PREFIXES ";lib")
-find_package(BLAS REQUIRED)
+if (USE_BLAS)
+    find_package(BLAS REQUIRED)
 
-include(CheckFunctionExists)
-check_function_exists(openblas_set_num_threads LESS_SLOW_HAS_OPENBLAS_SET_NUM_THREADS)
-if (LESS_SLOW_HAS_OPENBLAS_SET_NUM_THREADS)
-    add_definitions(-DLESS_SLOW_HAS_OPENBLAS_SET_NUM_THREADS)
+    include(CheckFunctionExists)
+    check_function_exists(openblas_set_num_threads LESS_SLOW_HAS_OPENBLAS_SET_NUM_THREADS)
+    if (LESS_SLOW_HAS_OPENBLAS_SET_NUM_THREADS)
+        add_definitions(-D LESS_SLOW_HAS_OPENBLAS_SET_NUM_THREADS)
+    endif ()
 endif ()
 
 # GTest (required by Google Benchmark)
@@ -346,7 +358,7 @@ if (CMAKE_SYSTEM_PROCESSOR MATCHES "x86_64|amd64|AMD64|x64")
 elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "aarch64|ARM64|arm64")
     if (APPLE)
         set(CMAKE_ASM_FLAGS "${CMAKE_ASM_FLAGS} -march=armv8.6-a+bf16")
-    endif()
+    endif ()
     set_source_files_properties(less_slow_aarch64.S PROPERTIES LANGUAGE ASM)
     target_sources(less_slow PRIVATE less_slow_aarch64.S)
 endif ()
@@ -431,6 +443,7 @@ endif ()
 # ------------------------------------------------------------------------------
 target_compile_definitions(less_slow PRIVATE USE_NVIDIA_CCCL=$<BOOL:${USE_NVIDIA_CCCL}>)
 target_compile_definitions(less_slow PRIVATE USE_INTEL_TBB=$<BOOL:${USE_INTEL_TBB}>)
+target_compile_definitions(less_slow PRIVATE USE_BLAS=$<BOOL:${USE_BLAS}>)
 target_link_libraries(
     less_slow
     PRIVATE Threads::Threads
@@ -448,9 +461,12 @@ target_link_libraries(
             absl::flat_hash_map
             nlohmann_json::nlohmann_json
             Eigen3::Eigen
-            ${BLAS_LIBRARIES}
 )
 
+if (USE_BLAS)
+    target_link_libraries(less_slow PRIVATE ${BLAS_LIBRARIES})
+endif ()
+
 if (CMAKE_SYSTEM_NAME STREQUAL "Linux")
     # target_include_directories(less_slow PRIVATE ${LIBURING_INCLUDE_DIRS})
     target_link_libraries(less_slow PRIVATE ${LIBURING_LIBRARIES})

README.md

@@ -86,10 +86,10 @@ The build will pull and compile several third-party dependencies from the source
 - Nvidia's [CCCL](https://github.com/nvidia/cccl) for GPU-accelerated algorithms.
 - Nvidia's [CUTLASS](https://github.com/nvidia/cutlass) for GPU-accelerated Linear Algebra.
 
-To build without Parallel STL, Intel TBB, and CUDA:
+To build without Parallel STL, Intel TBB, BLAS, and CUDA:
 
 ```sh
-cmake -B build_release -D CMAKE_BUILD_TYPE=Release -D USE_INTEL_TBB=OFF -D USE_NVIDIA_CCCL=OFF
+cmake -B build_release -D CMAKE_BUILD_TYPE=Release -D USE_INTEL_TBB=OFF -D USE_NVIDIA_CCCL=OFF -D USE_BLAS=OFF
 cmake --build build_release --config Release
 ```
 

less_slow.cpp

@@ -2904,6 +2904,20 @@ std::size_t parse_size_string(std::string const &str) {
 #pragma endregion // Non Uniform Memory Access
 
 #pragma region Memory Bound Linear Algebra
+
+#if !defined(USE_BLAS)
+#if defined(__has_include)
+#if __has_include(<cblas.h>)
+#define USE_BLAS 1
+#else
+#define USE_BLAS 0
+#endif // __has_include(<cblas.h>)
+#else
+#define USE_BLAS 0
+#endif // defined(__has_include)
+#endif // !defined(USE_BLAS)
+
+#if USE_BLAS
 #include <cblas.h>
 /**
  *  ! OpenBLAS defines a `SIZE` macro for internal use, which conflicts with `fmt`
@@ -2947,6 +2961,8 @@ static void cblas_tops(bm::State &state) {
 BENCHMARK(cblas_tops<float>)->RangeMultiplier(2)->Range(8, 16384)->Complexity(benchmark::oNCubed);
 BENCHMARK(cblas_tops<double>)->RangeMultiplier(2)->Range(8, 16384)->Complexity(benchmark::oNCubed);
 
+#endif // USE_BLAS
+
 /**
  *  Eigen is a high-level C++ library for linear algebra that provides a
  *  convenient templated API for matrix operations.