Unbreak packaging on BSD and non-x86_64 (colmap#348)

* VLfeat: use threads on more Unices (e.g., Solaris, BSDs) * PBA: GCC/Clang on aarch64 only defines __ARM_NEON * build: treat DragonFly as another BSD system * build: don't force SSE2/AVX, enable at runtime instead Make binary redistributable by default between machines of the same architecture by not enabling SIMD outside of routines gated by CPUID. No need to reinvent -march=native just for auto-vectorization. * VLFeat: actually use AVX when it should (cherry picked from commit vlfeat/vlfeat@df2a27cd5fb5) * VLFeat: enable SSE2 on Windows
dger1 · Apr 16, 2018 · 91ac87a · 91ac87a
1 parent 827bbb8
commit 91ac87a
Show file tree

Hide file tree

Showing 9 changed files with 48 additions and 197 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -18,8 +18,6 @@ include(CheckCXXCompilerFlag)
 # the CMake policies in this file
 include(${CMAKE_SOURCE_DIR}/cmake/CMakeHelper.cmake NO_POLICY_SCOPE)
 
-include(${CMAKE_SOURCE_DIR}/cmake/CheckSSEExtensions.cmake)
-
 if(EXISTS ${CMAKE_SOURCE_DIR}/LocalConfig.cmake)
     include(${CMAKE_SOURCE_DIR}/LocalConfig.cmake)
 endif()
@@ -29,6 +27,7 @@ endif()
 # Options
 ################################################################################
 
+option(SIMD_ENABLED "Whether to enable SIMD optimizations" ON)
 option(OPENMP_ENABLED "Whether to enable OpenMP" ON)
 option(IPO_ENABLED "Whether to enable interprocedural optimization" ON)
 option(CUDA_ENABLED "Whether to enable CUDA, if available" ON)

diff --git a/cmake/CMakeHelper.cmake b/cmake/CMakeHelper.cmake
@@ -16,9 +16,14 @@ if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
     set(IS_CLANG TRUE)
 endif()
 
+# Determine project architecture.
+if(CMAKE_SYSTEM_PROCESSOR MATCHES "[ix].?86|amd64|AMD64")
+    set(IS_X86 TRUE)
+endif()
+
 # Determine project operating system.
 string(REGEX MATCH "Linux" IS_LINUX ${CMAKE_SYSTEM_NAME})
-string(REGEX MATCH "BSD" IS_BSD ${CMAKE_SYSTEM_NAME})
+string(REGEX MATCH "DragonFly|BSD" IS_BSD ${CMAKE_SYSTEM_NAME})
 string(REGEX MATCH "SunOS" IS_SOLARIS ${CMAKE_SYSTEM_NAME})
 if(WIN32)
     SET(IS_WINDOWS TRUE BOOL INTERNAL)

diff --git a/cmake/CheckSSEExtensions.cmake b/cmake/CheckSSEExtensions.cmake
diff --git a/src/ext/PBA/CMakeLists.txt b/src/ext/PBA/CMakeLists.txt
@@ -1,5 +1,11 @@
 if(NOT IS_MSVC)
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fPIC -O3 -pthread -march=core2 -mfpmath=sse -Wno-c++11-narrowing")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fPIC -Wno-narrowing")
+endif()
+
+if(NOT SIMD_ENABLED)
+    add_definitions(-DDISABLE_CPU_NEON)
+    add_definitions(-DDISABLE_CPU_AVX)
+    add_definitions(-DDISABLE_CPU_SSE)
 endif()
 
 if(CUDA_ENABLED)

diff --git a/src/ext/PBA/SparseBundleCPU.cpp b/src/ext/PBA/SparseBundleCPU.cpp
@@ -52,15 +52,16 @@ using std::max;
 #if defined(_M_ARM) && _M_ARM >= 7 && !defined(DISABLE_CPU_NEON)
 #include <arm_neon.h>
 #define CPUPBA_USE_NEON
-#elif defined(__ARM_NEON__) && !defined(DISABLE_CPU_NEON)
+#elif defined(__ARM_NEON) && !defined(DISABLE_CPU_NEON)
 #include <arm_neon.h>
 #define CPUPBA_USE_NEON
 #endif
-#elif defined(CPUPBA_USE_AVX)  // Using AVX
+#elif defined(__AVX__) && !defined(DISABLE_CPU_AVX)
 #include <immintrin.h>
+#define CPUPBA_USE_AVX
 #undef CPUPBA_USE_SSE
 #undef POINT_DATA_ALIGN4
-#elif !defined(DISABLE_CPU_SSE)  // Using SSE
+#elif (defined(__SSE2__) || defined(_M_X64) || (defined(_M_IX86) && _M_IX86_FP >= 2)) && !defined(DISABLE_CPU_SSE)
 #define CPUPBA_USE_SSE
 #include <xmmintrin.h>
 #include <emmintrin.h>

diff --git a/src/ext/VLFeat/CMakeLists.txt b/src/ext/VLFeat/CMakeLists.txt
@@ -69,27 +69,35 @@ set(VLFEAT_SOURCE_FILES
     vlad.c
     vlad.h)
 
-if(IS_GNU OR IS_CLANG)
-    if(HAS_AVX_EXTENSION)
-        list(APPEND VLFEAT_SOURCE_FILES
-             mathop_avx.c
-             mathop_avx.h)
-    else()
+if(SIMD_ENABLED AND IS_X86)
+    if (MSVC)
+        # https://github.com/vlfeat/vlfeat/commit/4f0098fd47e9
         add_definitions(-DVL_DISABLE_AVX)
+    else()
+        set(AVX_SOURCES
+            mathop_avx.c
+            mathop_avx.h)
     endif()
 
-    if(HAS_SSE2_EXTENSION)
-        list(APPEND VLFEAT_SOURCE_FILES
-             mathop_sse2.c
-             mathop_sse2.h
-             imopv_sse2.c
-             imopv_sse2.h)
+    set(SSE2_SOURCES
+        imopv_sse2.c
+        imopv_sse2.h
+        mathop_sse2.c
+        mathop_sse2.h)
+
+    list(APPEND VLFEAT_SOURCE_FILES ${AVX_SOURCES} ${SSE2_SOURCES})
+
+    if (MSVC)
+        set_source_files_properties(${AVX_SOURCES}
+            PROPERTIES COMPILE_FLAGS "/arch:AVX")
+        set_source_files_properties(${SSE2_SOURCES}
+            PROPERTIES COMPILE_FLAGS "/arch:SSE2 /D__SSE2__")
     else()
-        add_definitions(-DVL_DISABLE_SSE2)
+        set_source_files_properties(${AVX_SOURCES}
+            PROPERTIES COMPILE_FLAGS "-mavx")
+        set_source_files_properties(${SSE2_SOURCES}
+            PROPERTIES COMPILE_FLAGS "-msse2")
     endif()
-
-    set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${SSE_FLAGS}")
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${SSE_FLAGS}")
 else()
     add_definitions(-DVL_DISABLE_AVX)
     add_definitions(-DVL_DISABLE_SSE2)

diff --git a/src/ext/VLFeat/generic.c b/src/ext/VLFeat/generic.c
@@ -777,16 +777,14 @@ calling @c omp_set_num_threads() in the application. Note that:
 
 #if defined(VL_OS_WIN)
 #include <Windows.h>
+#else
+#include <unistd.h>
 #endif
 
 #if ! defined(VL_DISABLE_THREADS) && defined(VL_THREADS_POSIX)
 #include <pthread.h>
 #endif
 
-#if defined(VL_OS_MACOSX) || defined(VL_OS_LINUX)
-#include <unistd.h>
-#endif
-
 #if defined(_OPENMP)
 #include <omp.h>
 #endif
@@ -1564,7 +1562,7 @@ vl_constructor (void)
     GetSystemInfo (&info) ;
     state->numCPUs = info.dwNumberOfProcessors ;
   }
-#elif defined(VL_OS_MACOSX) || defined(VL_OS_LINUX)
+#elif defined(_SC_NPROCESSORS_ONLN)
   state->numCPUs = sysconf(_SC_NPROCESSORS_ONLN) ;
 #else
   state->numCPUs = 1 ;

diff --git a/src/ext/VLFeat/gmm.c b/src/ext/VLFeat/gmm.c
@@ -1102,13 +1102,13 @@ VL_XCAT(_vl_gmm_maximization_, SFX)
 
         #ifndef VL_DISABLE_AVX
         if (vl_get_simd_enabled() && vl_cpu_has_avx()) {
-          VL_XCAT(_vl_weighted_mean_sse2_, SFX)
+          VL_XCAT(_vl_weighted_mean_avx_, SFX)
           (self->dimension,
            means_+ i_cl * self->dimension,
            data + i_d * self->dimension,
            p) ;
 
-          VL_XCAT(_vl_weighted_sigma_sse2_, SFX)
+          VL_XCAT(_vl_weighted_sigma_avx_, SFX)
           (self->dimension,
            covariances_ + i_cl * self->dimension,
            data + i_d * self->dimension,

diff --git a/src/ext/VLFeat/host.h b/src/ext/VLFeat/host.h
@@ -199,7 +199,7 @@ the terms of the BSD license (see the COPYING file).
 
 /** @name Identifying the host threading library
  ** @{ */
-#if defined(VL_OS_MACOSX) || defined(VL_OS_LINUX) || \
+#if !defined(VL_OS_WIN) && !defined(VL_OS_WIN64) || \
 defined(__DOXYGEN__)
 #define VL_THREADS_POSIX 1
 #endif