cmake: add THREAD_TIMEOUT, prevent in-tree builds, fix ARCH handling

vtjnash · claude · vtjnash · commit ceb544c5444f · 2025-12-23T13:13:29.000-05:00
Changes: - Add THREAD_TIMEOUT option for thread spin-wait timeout configuration - Prevent in-tree builds with clear error message (protects Makefiles) - Fix ARCH variable to not redundantly re-set when user provides it - Remove experimental warning (CMake now has full Makefile.rule parity) All Makefile.rule options are now supported in CMake: | Makefile.rule Option | CMake Equivalent | |-------------------------------|-------------------------------------| | LIBNAMEPREFIX | LIBNAMEPREFIX | | LIBNAMESUFFIX | LIBNAMESUFFIX | | TARGET | TARGET | | DYNAMIC_ARCH | DYNAMIC_ARCH | | DYNAMIC_OLDER | DYNAMIC_OLDER | | BINARY | BINARY | | USE_THREAD | USE_THREAD | | USE_LOCKING | USE_LOCKING | | USE_OPENMP | USE_OPENMP | | OMP_SCHED | OMP_SCHED | | NUM_THREADS | NUM_THREADS | | NUM_PARALLEL | NUM_PARALLEL | | BUFFERSIZE | BUFFERSIZE | | NO_STATIC / NO_SHARED | BUILD_STATIC_LIBS / BUILD_SHARED_LIBS | | NO_CBLAS | BUILD_WITHOUT_CBLAS | | ONLY_CBLAS | ONLY_CBLAS | | NO_LAPACK | BUILD_WITHOUT_LAPACK | | NO_LAPACKE | BUILD_WITHOUT_LAPACKE | | BUILD_LAPACK_DEPRECATED | BUILD_LAPACK_DEPRECATED | | LAPACK_STRLEN | LAPACK_STRLEN | | BUILD_RELAPACK | BUILD_RELAPACK | | RELAPACK_REPLACE | RELAPACK_REPLACE | | USE_SIMPLE_THREADED_LEVEL3 | USE_SIMPLE_THREADED_LEVEL3 | | USE_TLS | USE_TLS | | INTERFACE64 | INTERFACE64 | | NO_WARMUP | NO_WARMUP | | NO_AFFINITY | NO_AFFINITY | | BIGNUMA | BIGNUMA | | EMBEDDED | EMBEDDED | | NO_AVX / NO_AVX2 / NO_AVX512 | NO_AVX / NO_AVX2 / NO_AVX512 | | FUNCTION_PROFILE | FUNCTION_PROFILE | | QUAD_PRECISION | QUAD_PRECISION | | THREAD_TIMEOUT | THREAD_TIMEOUT | | DEVICEDRIVER_ALLOCATION | DEVICEDRIVER_ALLOCATION | | HUGETLB_ALLOCATION | HUGETLB_ALLOCATION | | HUGETLBFILE_ALLOCATION | HUGETLBFILE_ALLOCATION | | CONSISTENT_FPCSR | CONSISTENT_FPCSR | | GEMM_MULTITHREAD_THRESHOLD | GEMM_MULTITHREAD_THRESHOLD | | SANITY_CHECK | SANITY_CHECK | | MAX_STACK_ALLOC | MAX_STACK_ALLOC | | SYMBOLPREFIX / SYMBOLSUFFIX | SYMBOLPREFIX / SYMBOLSUFFIX | | CPP_THREAD_SAFETY_TEST | CPP_THREAD_SAFETY_TEST | | CPP_THREAD_SAFETY_GEMV | CPP_THREAD_SAFETY_GEMV | | BUILD_BFLOAT16 / BUILD_HFLOAT16 | BUILD_BFLOAT16 / BUILD_HFLOAT16 | | BLAS3_MEM_ALLOC_THRESHOLD | BLAS3_MEM_ALLOC_THRESHOLD | | BUILD_SINGLE/DOUBLE/COMPLEX/COMPLEX16 | BUILD_SINGLE/DOUBLE/COMPLEX/COMPLEX16 | | ARM_SOFTFP_ABI | ARM_SOFTFP_ABI | | DEBUG | CMAKE_BUILD_TYPE=Debug | | PREFIX | CMAKE_INSTALL_PREFIX | | HOSTCC | CMake toolchain files | 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -7,6 +7,11 @@ cmake_minimum_required(VERSION 3.16.0)
 set (CMAKE_ASM_SOURCE_FILE_EXTENSIONS "S")
 project(OpenBLAS C ASM)
 
+# Prevent in-tree builds to avoid overwriting Makefiles
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_BINARY_DIR)
+  message(FATAL_ERROR "In-tree builds are not allowed. Use: cmake -B build -S .")
+endif ()
+
 set(OpenBLAS_MAJOR_VERSION 0)
 set(OpenBLAS_MINOR_VERSION 3)
 set(OpenBLAS_PATCH_VERSION 30.dev)
@@ -29,6 +34,10 @@ set(LAPACK_STRLEN "" CACHE STRING "When building LAPACK, use this type (e.g. \"i
 
 option(BUILD_TESTING "Build LAPACK testsuite when building LAPACK" ON)
 
+option(BUILD_TESTS "Build the BLAS test suite (test/ and ctest/ directories)" ON)
+
+option(BUILD_UTESTS "Build the unit tests (utest/ directory)" ON)
+
 option(BUILD_BENCHMARKS "Build the collection of BLAS/LAPACK benchmarks" OFF)
 
 option(C_LAPACK "Build LAPACK from C sources instead of the original Fortran" OFF)
@@ -39,8 +48,14 @@ option(DYNAMIC_ARCH "Include support for multiple CPU targets, with automatic se
 
 option(DYNAMIC_OLDER "Include specific support for older x86 cpu models (Penryn,Dunnington,Atom,Nano,Opteron) with DYNAMIC_ARCH" OFF)
 
+set(DYNAMIC_LIST "" CACHE STRING "Manually specify list of CPU targets for DYNAMIC_ARCH instead of default list (semicolon-separated, e.g., 'HASWELL;SKYLAKEX')")
+
+set(TARGET_CORE "" CACHE STRING "Override TARGET for DYNAMIC_ARCH kernel selection. Used to specify which CPU-specific kernels to build.")
+
 option(BUILD_RELAPACK "Build with ReLAPACK (recursive implementation of several LAPACK functions on top of standard LAPACK)" OFF)
 
+option(RELAPACK_REPLACE "Use ReLAPACK to replace standard LAPACK routines instead of adding RELAPACK_ prefixed equivalents (requires BUILD_RELAPACK)" OFF)
+
 option(USE_LOCKING "Use locks even in single-threaded builds to make them callable from multiple threads" OFF)
 
 option(USE_PERL "Use the older PERL scripts for build preparation instead of universal shell scripts" OFF)
@@ -51,6 +66,7 @@ option(FIXED_LIBNAME "Use a non-versioned name for the library and no symbolic l
 
 set(LIBNAMEPREFIX "" CACHE STRING  "Add a prefix to the openblas part of the library name" )
 set(LIBNAMESUFFIX "" CACHE STRING  "Add a suffix after the openblas part of the library name" )
+set(LIBSONAMEBASE "openblas" CACHE STRING "Base name for shared library soname (default: openblas)")
 
 if(${CMAKE_SYSTEM_NAME} MATCHES "Linux")
   option(NO_AFFINITY "Disable support for CPU affinity masks to avoid binding processes from e.g. R or numpy/scipy to a single core" ON)
@@ -81,6 +97,9 @@ set(SYMBOLSUFFIX "" CACHE STRING  "Add a suffix to all exported symbol names in
 # Target architecture - auto-detected if not specified
 set(TARGET "" CACHE STRING "Target CPU architecture (e.g. HASWELL, SANDYBRIDGE, NEHALEM, ARMV8, POWER9). Auto-detected if not specified.")
 
+# Force architecture (normally auto-detected from system)
+set(ARCH "" CACHE STRING "Force architecture (x86, x86_64, arm, arm64, power, mips, mips64, zarch, loongarch64, riscv64). Auto-detected if not specified.")
+
 # Binary type (32-bit or 64-bit)
 set(BINARY "" CACHE STRING "Build a 32-bit or 64-bit library (32 or 64). Auto-detected if not specified. Note: 32-bit disables AVX.")
 
@@ -89,6 +108,7 @@ set(USE_THREAD "" CACHE STRING "Enable multi-threading (0=disabled, 1=enabled).
 option(USE_OPENMP "Use OpenMP for threading instead of pthreads" OFF)
 set(NUM_THREADS "" CACHE STRING "Maximum number of threads. Auto-detected from CPU cores if not specified.")
 set(NUM_PARALLEL "1" CACHE STRING "Number of parallel OpenBLAS instances when using OpenMP (default: 1)")
+set(OMP_SCHED "static" CACHE STRING "OpenMP schedule type (static, dynamic, guided, auto, runtime). Default: static")
 
 # 64-bit integer interface
 option(INTERFACE64 "Use 64-bit integers for array indices (equivalent to -i8 in ifort)" OFF)
@@ -97,6 +117,8 @@ option(INTERFACE64 "Use 64-bit integers for array indices (equivalent to -i8 in
 option(NO_AVX "Disable AVX kernel support (use for compatibility with older systems)" OFF)
 option(NO_AVX2 "Disable AVX2 optimizations" OFF)
 option(NO_AVX512 "Disable AVX512 optimizations" OFF)
+option(NO_SVE "Disable ARM SVE (Scalable Vector Extension) optimizations" OFF)
+option(NO_SME "Disable ARM SME (Scalable Matrix Extension) optimizations" OFF)
 
 # Memory tuning options
 set(BUFFERSIZE "" CACHE STRING "Memory buffer size factor (32<<n bytes, default: architecture-dependent, typically 25)")
@@ -108,10 +130,12 @@ set(GEMM_MULTITHREAD_THRESHOLD "4" CACHE STRING "Threshold below which GEMM runs
 option(USE_SIMPLE_THREADED_LEVEL3 "Use legacy threaded Level 3 implementation" OFF)
 option(USE_TLS "Use thread-local storage instead of central memory buffer (requires glibc 2.21+)" OFF)
 option(CONSISTENT_FPCSR "Synchronize floating-point CSR between threads (x86/x86_64/aarch64 only)" OFF)
+set(THREAD_TIMEOUT "" CACHE STRING "Thread spin-wait timeout as power of 2 cycles, e.g. 26 means 2^26 cycles (~25ms at 3GHz). Range: 4-30. Empty = default (28).")
 
 # System configuration
 option(BIGNUMA "Support systems with more than 16 NUMA nodes or more than 256 CPUs (Linux only)" OFF)
 option(EMBEDDED "Build for embedded/bare-metal systems (requires custom malloc/free)" OFF)
+option(ARM_SOFTFP_ABI "Use soft floating-point ABI on ARM (for compatibility with soft-float systems)" OFF)
 
 # Precision type options (default: build all types if all OFF)
 option(BUILD_SINGLE "Build single precision (REAL) functions" OFF)
@@ -120,15 +144,19 @@ option(BUILD_COMPLEX "Build complex (COMPLEX) functions" OFF)
 option(BUILD_COMPLEX16 "Build double complex (COMPLEX*16) functions" OFF)
 option(BUILD_BFLOAT16 "Build experimental BFLOAT16 functions" OFF)
 option(BUILD_HFLOAT16 "Build experimental HFLOAT16 functions" OFF)
+option(QUAD_PRECISION "Build with IEEE quad precision support (experimental, x86_64 only)" OFF)
 
 # CBLAS-only mode
 option(ONLY_CBLAS "Build only CBLAS interface (no Fortran BLAS, implies NO_LAPACK)" OFF)
 
 # Profiling and debugging
 option(FUNCTION_PROFILE "Enable function-level performance profiling" OFF)
 option(SANITY_CHECK "Compare results against reference BLAS (slow, for testing only)" OFF)
+option(UTEST_CHECK "Enable unit test result checking (implies SANITY_CHECK)" OFF)
 
 # Memory allocation methods
+option(SHMEM_ALLOCATION "Use shared memory for buffer allocation" OFF)
+option(STATIC_ALLOCATION "Use static memory allocation (for single-threaded or embedded use)" OFF)
 option(HUGETLB_ALLOCATION "Use huge pages for thread buffers via shared memory" OFF)
 set(HUGETLBFILE_ALLOCATION "" CACHE STRING "Path to hugetlbfs mount for huge page allocation (e.g. /hugepages)")
 option(DEVICEDRIVER_ALLOCATION "Use device driver for physically contiguous memory allocation" OFF)
@@ -174,8 +202,6 @@ if(MSVC AND MSVC_STATIC_CRT)
   endforeach()
 endif()
 
-message(WARNING "CMake support is experimental. It does not yet support all build options and may not produce the same Makefiles that OpenBLAS ships with.")
-
 include("${PROJECT_SOURCE_DIR}/cmake/utils.cmake")
 include("${PROJECT_SOURCE_DIR}/cmake/system.cmake")
 
@@ -439,31 +465,41 @@ if (USE_THREAD)
   endif()
 endif()
 
-#if (MSVC OR NOT NOFORTRAN)
-if (NOT NO_CBLAS)
-  if (NOT ONLY_CBLAS)
-  # Broken without fortran on unix
-    add_subdirectory(utest)
+# Build unit tests
+if (BUILD_UTESTS)
+  if (NOT NO_CBLAS)
+    if (NOT ONLY_CBLAS)
+      # Broken without fortran on unix
+      add_subdirectory(utest)
+    endif()
+  endif()
 endif()
+
+# Build BLAS test suite
+if (BUILD_TESTS)
+  if (NOT NOFORTRAN)
+    if (NOT ONLY_CBLAS)
+      # Build test and ctest
+      add_subdirectory(test)
+    endif()
+  endif()
+  if(NOT NO_CBLAS)
+    if (NOT ONLY_CBLAS)
+      add_subdirectory(ctest)
+    endif()
+  endif()
 endif()
 
+# Build LAPACK test suite
 if (NOT NOFORTRAN)
- if (NOT ONLY_CBLAS)
-  # Build test and ctest
-  add_subdirectory(test)
- endif()
- if (BUILD_TESTING AND NOT BUILD_WITHOUT_LAPACK)
+  if (BUILD_TESTING AND NOT BUILD_WITHOUT_LAPACK)
     add_subdirectory(lapack-netlib/TESTING)
   endif()
 endif()
-  if(NOT NO_CBLAS)
-   if (NOT ONLY_CBLAS)
-    add_subdirectory(ctest)
-   endif()
-  endif()
-  if (CPP_THREAD_SAFETY_TEST OR CPP_THREAD_SAFETY_GEMV)
-    add_subdirectory(cpp_thread_test)
-  endif()
+
+if (CPP_THREAD_SAFETY_TEST OR CPP_THREAD_SAFETY_GEMV)
+  add_subdirectory(cpp_thread_test)
+endif()
 
 if (NOT FIXED_LIBNAME)
 set_target_properties(${OpenBLAS_LIBS} PROPERTIES
diff --git a/cmake/arch.cmake b/cmake/arch.cmake
@@ -49,7 +49,7 @@ if (DYNAMIC_ARCH)
         set(DYNAMIC_CORE ${DYNAMIC_CORE} ARMV9SME)
       endif()
     endif ()
-    if (DYNAMIC_LIST)
+    if (DEFINED DYNAMIC_LIST AND NOT "${DYNAMIC_LIST}" STREQUAL "")
 	  set(DYNAMIC_CORE ARMV8 ${DYNAMIC_LIST})
     endif ()
   endif ()
@@ -90,7 +90,7 @@ if (DYNAMIC_ARCH)
       set(DYNAMIC_CORE ${DYNAMIC_CORE} SKYLAKEX COOPERLAKE SAPPHIRERAPIDS)
       string(REGEX REPLACE "-march=native" "" CMAKE_C_FLAGS "${CMAKE_C_FLAGS}")
     endif ()
-    if (DYNAMIC_LIST)
+    if (DEFINED DYNAMIC_LIST AND NOT "${DYNAMIC_LIST}" STREQUAL "")
       set(DYNAMIC_CORE PRESCOTT ${DYNAMIC_LIST})
     endif ()
   endif ()
diff --git a/cmake/prebuild.cmake b/cmake/prebuild.cmake
@@ -128,7 +128,7 @@ endif ()
 # Cannot run getarch on target if we are cross-compiling
 if (DEFINED CORE AND CMAKE_CROSSCOMPILING AND NOT (${HOST_OS} STREQUAL "WINDOWSSTORE"))
   # Write to config as getarch would
-  if (DEFINED TARGET_CORE)
+  if (DEFINED TARGET_CORE AND NOT "${TARGET_CORE}" STREQUAL "")
   set(TCORE ${TARGET_CORE})
   else()
   set(TCORE ${CORE})
@@ -1589,7 +1589,7 @@ else(NOT CMAKE_CROSSCOMPILING)
     if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Darwin")
       list(APPEND GETARCH_SRC ${PROJECT_SOURCE_DIR}/cpuid.S)
     endif()
-    if (DEFINED TARGET_CORE)
+    if (DEFINED TARGET_CORE AND NOT "${TARGET_CORE}" STREQUAL "")
     set(GETARCH_FLAGS ${GETARCH_FLAGS} -DFORCE_${TARGET_CORE})
   endif ()
   endif ()
diff --git a/cmake/system.cmake b/cmake/system.cmake
@@ -84,7 +84,8 @@ if(NOT DEFINED CORE AND _TARGET_SET)
 endif()
 
 # TARGET_CORE will override TARGET which is used in DYNAMIC_ARCH=1.
-if (DEFINED TARGET_CORE)
+if (DEFINED TARGET_CORE AND NOT "${TARGET_CORE}" STREQUAL "")
+  message(STATUS "Using TARGET_CORE=${TARGET_CORE} for kernel selection")
   set(TARGET ${TARGET_CORE})
 endif ()
 
@@ -468,6 +469,10 @@ endif ()
 if (USE_OPENMP)
   find_package(OpenMP COMPONENTS C REQUIRED)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_OPENMP")
+  # Set OpenMP schedule type (default is static)
+  if (OMP_SCHED AND NOT "${OMP_SCHED}" STREQUAL "static")
+    set(CCOMMON_OPT "${CCOMMON_OPT} -DOMP_SCHED=${OMP_SCHED}")
+  endif ()
   if (NOT NOFORTRAN)
     find_package(OpenMP COMPONENTS Fortran REQUIRED)
     # Avoid mixed OpenMP linkage
@@ -496,6 +501,11 @@ if(EMBEDDED)
   set(CCOMMON_OPT "${CCOMMON_OPT} -mthumb -mcpu=cortex-m4 -mfloat-abi=hard -mfpu=fpv4-sp-d16")
 endif()
 
+if (ARM_SOFTFP_ABI)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -mfloat-abi=softfp")
+  set(FCOMMON_OPT "${FCOMMON_OPT} -mfloat-abi=softfp")
+endif ()
+
 if (NEED_PIC)
   if (${CMAKE_C_COMPILER} STREQUAL "IBM")
     set(CCOMMON_OPT "${CCOMMON_OPT} -qpic=large")
@@ -514,20 +524,21 @@ if (NEED_PIC)
   endif()
 endif ()
 
-if (X86_64 OR ${CORE} STREQUAL POWER10 OR ARM64 OR LOONGARCH64)
-  set(SMALL_MATRIX_OPT TRUE)
+# Auto-enable SMALL_MATRIX_OPT on supported architectures (internal, not user-configurable)
+if (X86_64 OR (DEFINED CORE AND ${CORE} STREQUAL POWER10) OR ARM64 OR LOONGARCH64)
+  set(SMALL_MATRIX_OPT ON)
 endif ()
+
+# Auto-enable GEMM_GEMV_FORWARD on supported architectures (internal, not user-configurable)
 if (ARM64)
-  set(GEMM_GEMV_FORWARD TRUE)
-  set(SBGEMM_GEMV_FORWARD TRUE)
-  set(BGEMM_GEMV_FORWARD TRUE)
-endif ()
-if (POWER)
-  set(GEMM_GEMV_FORWARD TRUE)
-  set(SBGEMM_GEMV_FORWARD TRUE)
-endif ()
-if (RISCV64)
-  set(GEMM_GEMV_FORWARD TRUE)
+  set(GEMM_GEMV_FORWARD ON)
+  set(SBGEMM_GEMV_FORWARD ON)
+  set(BGEMM_GEMV_FORWARD ON)
+elseif (POWER)
+  set(GEMM_GEMV_FORWARD ON)
+  set(SBGEMM_GEMV_FORWARD ON)
+elseif (RISCV64)
+  set(GEMM_GEMV_FORWARD ON)
 endif ()
 
 if (GEMM_GEMV_FORWARD)
@@ -555,7 +566,8 @@ if (DYNAMIC_ARCH)
   endif ()
 endif ()
 
-if (DYNAMIC_LIST)
+if (DEFINED DYNAMIC_LIST AND NOT "${DYNAMIC_LIST}" STREQUAL "")
+  message(STATUS "Using custom DYNAMIC_LIST: ${DYNAMIC_LIST}")
   set(CCOMMON_OPT "${CCOMMON_OPT} -DDYNAMIC_LIST")
   foreach(DCORE ${DYNAMIC_LIST})
     set(CCOMMON_OPT "${CCOMMON_OPT} -DDYN_${DCORE}")
@@ -588,6 +600,14 @@ if (NO_AVX512)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_AVX512")
 endif ()
 
+if (NO_SVE)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_SVE")
+endif ()
+
+if (NO_SME)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DNO_SME")
+endif ()
+
 if (USE_THREAD)
   # USE_SIMPLE_THREADED_LEVEL3 = 1
   # NO_AFFINITY = 1
@@ -616,6 +636,10 @@ if (USE_TLS)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DUSE_TLS")
 endif ()
 
+if (DEFINED THREAD_TIMEOUT AND NOT "${THREAD_TIMEOUT}" STREQUAL "")
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DTHREAD_TIMEOUT=${THREAD_TIMEOUT}")
+endif ()
+
 # Only for development
 # set(CCOMMON_OPT "${CCOMMON_OPT} -DPARAMTEST")
 # set(CCOMMON_OPT "${CCOMMON_OPT} -DPREFETCHTEST")
@@ -660,7 +684,12 @@ endif()
 endif()
 endif()
   
-set(LIBPREFIX "lib${LIBNAMEPREFIX}openblas")
+# Use LIBSONAMEBASE for library naming (default: openblas)
+if (NOT DEFINED LIBSONAMEBASE OR "${LIBSONAMEBASE}" STREQUAL "")
+  set(LIBSONAMEBASE "openblas")
+endif ()
+
+set(LIBPREFIX "lib${LIBNAMEPREFIX}${LIBSONAMEBASE}")
 
 if (DEFINED LIBNAMESUFFIX)
   set(LIBPREFIX "${LIBNAMEPREFIX}_${LIBNAMESUFFIX}")
@@ -711,6 +740,10 @@ if (DEFINED HUGETLBFILE_ALLOCATION AND NOT "${HUGETLBFILE_ALLOCATION}" STREQUAL
   set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_HUGETLBFILE -DHUGETLB_FILE_NAME=${HUGETLBFILE_ALLOCATION})")
 endif ()
 
+if (SHMEM_ALLOCATION)
+  set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_SHM")
+endif ()
+
 if (STATIC_ALLOCATION)
   set(CCOMMON_OPT "${CCOMMON_OPT} -DALLOC_STATIC")
 endif ()
diff --git a/cmake/system_check.cmake b/cmake/system_check.cmake
diff --git a/relapack/src/CMakeLists.txt b/relapack/src/CMakeLists.txt
