Merge pull request #2 from OpenGamma/MAT-287-switch-off-ffast-math

Mat 287 multilibalise lapack

BLAS/CMakeLists.txt

@@ -3,7 +3,3 @@ if(BUILD_TESTING)
 add_subdirectory(TESTING)
 endif(BUILD_TESTING)
 configure_file(${CMAKE_CURRENT_SOURCE_DIR}/blas.pc.in ${CMAKE_CURRENT_BINARY_DIR}/blas.pc)
-install(FILES
-  ${CMAKE_CURRENT_BINARY_DIR}/blas.pc
-  DESTINATION ${PKG_CONFIG_DIR}
-  )

BLAS/SRC/CMakeLists.txt

@@ -135,17 +135,27 @@ if(BLAS_COMPLEX16)
 	${ALLBLAS} ${ZBLAS2} ${ZBLAS3})
 endif()
 
-set(XERBLA xerbla.f)   
-
-add_library(ogxerbla SHARED ${XERBLA})
-set_target_properties(ogxerbla PROPERTIES SOVERSION ${LAPACK_VERSION_MAJOR} VERSION ${LAPACK_VERSION})
-lapack_install_library(ogxerbla)
-
-add_library(ogblas SHARED ${ALLOBJ})
-set_target_properties(ogblas PROPERTIES SOVERSION ${LAPACK_VERSION_MAJOR} VERSION ${LAPACK_VERSION})
-target_link_libraries(ogblas ogxerbla)
+set(XERBLA xerbla.f)
+
+add_multitarget_library(ogxerbla
+                        VERSION ${LAPACK_VERSION}
+                        SOVERSION ${LAPACK_VERSION_MAJOR}
+                        SOURCES ${XERBLA}
+                        TARGETS ${TARGET_TYPES}
+                        )
+
+set(LINK_MULTILIBRARIES ogblas ogxerbla)
 if(UNIX)
-  target_link_libraries(ogblas m)
+  set(LINK_LIBRARIES m)
+else()
+  set(LINK_LIBRARIES "")
 endif()
-target_link_libraries(ogblas)
-lapack_install_library(ogblas)
+
+add_multitarget_library(ogblas
+                        VERSION ${LAPACK_VERSION}
+                        SOVERSION ${LAPACK_VERSION_MAJOR}
+                        SOURCES ${ALLOBJ}
+                        TARGETS ${TARGET_TYPES}
+                        LINK_MULTILIBRARIES ${LINK_MULTILIBRARIES}
+                        LINK_LIBRARIES ${LINK_LIBRARIES}
+                        )

CMAKE/CheckCPUID.cmake

@@ -0,0 +1,50 @@
+#
+# Copyright (C) 2013 - present by OpenGamma Inc. and the OpenGamma group of companies
+#
+# Please see distribution for license.
+#
+
+# Inclusion of this module sets the targets supported by the machine on which
+# CMake is running. SUPPORT_${TARGET} is set to TRUE or FALSE for each target in
+# dbg, std, sse41, sse42, avx1 and avx2. std and sbg require no special CPU
+# support and are always set.
+
+# this module caches the fact it has run by setting CPUID_CACHED = TRUE
+
+macro( CheckCPUID )
+  if(NOT DEFINED CPUID_CACHED)
+    set(CPUID_CACHED TRUE CACHE INTERNAL "Whether CPUID has run")
+    try_run(CPUID_FLAG CPUID_COMPILE_RESULT
+            ${CMAKE_BINARY_DIR}
+            ${CMAKE_SOURCE_DIR}/CMAKE/cmcpuid.c)
+
+    set(SUPPORT_dbg TRUE CACHE INTERNAL "Support for debug build")
+    set(SUPPORT_std TRUE CACHE INTERNAL "Support for standard build")
+
+    # NOTE: These flags must match the numbers in cmcpuid.c
+    if(${CPUID_FLAG} GREATER 1)
+      set(SUPPORT_sse41 TRUE CACHE INTERNAL "Local CPU support for SSE4.1")
+      message(STATUS "CPU supports SSE4.1")
+    else()
+      message(STATUS "CPU does not support SSE4.1")
+    endif()
+    if(${CPUID_FLAG} GREATER 2)
+      set(SUPPORT_sse42 TRUE CACHE INTERNAL "Local CPU support for SSE4.2")
+      message(STATUS "CPU supports SSE4.2")
+    else()
+      message(STATUS "CPU does not support SSE4.2")
+    endif()
+    if(${CPUID_FLAG} GREATER 3)
+      set(SUPPORT_avx1 TRUE CACHE INTERNAL "Local CPU support for AVX1")
+      message(STATUS "CPU supports AVX1")
+    else()
+      message(STATUS "CPU does not support AVX1")
+    endif()
+    if(${CPUID_FLAG} GREATER 4)
+      set(SUPPORT_avx2 TRUE CACHE INTERNAL "Local CPU support for AVX2")
+      message(STATUS "CPU supports AVX2")
+    else()
+      message(STATUS "CPU does not support AVX2")
+    endif()
+  endif()
+endmacro()

CMAKE/MultiInstructionTarget.cmake

@@ -0,0 +1,75 @@
+#
+# Copyright (C) 2013 - present by OpenGamma Inc. and the OpenGamma group of companies
+#
+# Please see distribution for license.
+#
+
+include(CMakeParseArguments)
+
+include( CheckCPUID )
+CheckCPUID()
+
+# Create a library for multiple target instruction sets. Note that this is
+# specialised to its use in Nyqwk2 and does not fully provide all of the
+# functionality that can be provided by add_library and set_target_properties.
+# One instance of the library is created for each of the TARGETS.
+
+function(add_multitarget_library _TARGET_NAME)
+  cmake_parse_arguments(MTLIB
+                        ""
+                        "VERSION;SOVERSION"
+                        "SOURCES;TARGETS;LINK_MULTILIBRARIES;LINK_LIBRARIES;COMPILE_DEFINITIONS;MULTI_DEPENDS"
+                        ${ARGN})
+  foreach(TARGET ${MTLIB_TARGETS})
+    set(LIBNAME ${_TARGET_NAME}_${TARGET})
+    add_library(${LIBNAME} SHARED ${MTLIB_SOURCES})
+    set_target_properties(${LIBNAME} PROPERTIES
+                          VERSION ${MTLIB_VERSION}
+                          SOVERSION ${MTLIB_SOVERSION}
+                          COMPILE_FLAGS ${OPT_FLAGS_${TARGET}})
+    foreach(LINK_LIB ${MTLIB_LINK_MULTILIBRARIES})
+      target_link_libraries(${LIBNAME} ${LINK_LIB}_${TARGET})
+    endforeach()
+    foreach(LINK_LIB ${MTLIB_LINK_LIBRARIES})
+      target_link_libraries(${LIBNAME} ${LINK_LIB})
+    endforeach()
+    foreach(COMPILE_DEFINITION ${MTLIB_COMPILE_DEFINITIONS})
+      set_target_properties(${LIBNAME} PROPERTIES COMPILE_DEFINITIONS ${COMPILE_DEFINITION})
+    endforeach()
+    foreach(DEPENDENCY ${MTLIB_MULTI_DEPENDS})
+      add_dependencies(${LIBNAME} ${DEPENDENCY}_${TARGET})
+    endforeach()   
+  endforeach()
+endfunction()
+
+# Create a binary for multiple target instruction sets. Note that this is
+# specialised to its use in Nyqwk2 and does not fully provide all of the
+# functionality that can be provided by add_library and set_target_properties.
+# One instance of the library is created for each of the TARGETS.
+
+function(add_multitarget_binary _TARGET_NAME)
+  cmake_parse_arguments(MTBIN
+                        ""
+                        "VERSION;SOVERSION"
+                        "SOURCES;TARGETS;LINK_MULTILIBRARIES;LINK_LIBRARIES;COMPILE_DEFINITIONS;MULTI_DEPENDS"
+                        ${ARGN})
+  foreach(TARGET ${MTBIN_TARGETS})
+    set(BINNAME ${_TARGET_NAME}_${TARGET})
+    add_executable(${BINNAME} ${MTBIN_SOURCES})
+    set_target_properties(${BINNAME} PROPERTIES
+                          COMPILE_FLAGS ${OPT_FLAGS_${TARGET}})
+    foreach(LINK_LIB ${MTBIN_LINK_MULTILIBRARIES})
+      target_link_libraries(${BINNAME} ${LINK_LIB}_${TARGET})
+    endforeach()
+    foreach(LINK_LIB ${MTBIN_LINK_LIBRARIES})
+      target_link_libraries(${BINNAME} ${LINK_LIB})
+    endforeach()
+    foreach(COMPILE_DEFINITION ${MTBIN_COMPILE_DEFINITIONS})
+      set_target_properties(${BINNAME} PROPERTIES COMPILE_DEFINITIONS ${COMPILE_DEFINITION})
+    endforeach()
+    foreach(DEPENDENCY ${MTBIN_MULTI_DEPENDS})
+      add_dependencies(${BINNAME} ${DEPENDENCY}_${TARGET})
+    endforeach()
+  endforeach()
+endfunction()
+

CMAKE/cmcpuid.c

@@ -0,0 +1,113 @@
+/**
+ * Copyright (C) 2013 - present by OpenGamma Inc. and the OpenGamma group of companies
+ *
+ * Please see distribution for license.
+ */
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <stdint.h>
+#include <stdio.h>
+
+
+/**
+ * The supported instruction set on this machine for use in other functions.
+ *
+ * These must match up with the numbers used in MultiLib.cmake.
+ */
+typedef enum instructions_available_e
+{
+  supports_STANDARD = 1,
+  supports_SSE41    = 2,
+  supports_SSE42    = 3,
+  supports_AVX1     = 4,
+  supports_AVX2     = 5
+} instructions_available;
+
+
+instructions_available getSupportedInstructionSet() {
+
+  // probes of cpuid with %eax=0000_0001h
+  enum instructions_available_eax0000_0001h_e
+  {
+    probe01_SSE_3   = 1<<0,
+    probe01_SSE_4_1 = 1<<19,
+    probe01_SSE_4_2 = 1<<20,
+    probe01_AVX1     = 1<<28
+  };
+
+  // probes of cpuid with %eax=0000_0007h
+  enum instructions_available_eax0000_0007h_e
+  {
+    probe07_AVX2     = 1<<5
+  };
+
+  // the eax register
+  int32_t EAX;
+  // contends of the returned register
+  int32_t supported;
+
+  // Call cpuid with eax=0x00000001 and get ecx
+  EAX = 0x00000001;
+  __asm__("cpuid"
+        :"=c"(supported)         // %ecx contains large feature flag set
+        :"0"(EAX)                // call with 0x1
+        :"%eax","%ebx","%edx");  // clobbered
+
+  if(supported & probe01_AVX1) // we have at least AVX1
+  {
+    EAX = 0x00000007;
+    __asm__("cpuid"
+          :"=b"(supported)         // %ebx contains feature flag AVX2
+          :"0"(EAX)                // call with 0x7
+          :"%eax","%ecx","%edx");  // clobbered
+
+    if(supported & probe07_AVX2) // we have at least AVX2
+    {
+      printf("AVX2 SUPPORTED\n");
+      return supports_AVX2;
+    }
+    printf("AVX1 SUPPORTED\n");
+    return supports_AVX1;
+  }
+  else if(supported & probe01_SSE_4_1) // we have at least SSE4.1
+  {
+    printf("SSE4.2 SUPPORTED\n");
+    return supports_SSE42;
+  }
+  else // we have nothing specifically useful!
+  {
+    printf("STANDARD SUPPORTED\n");
+    return supports_STANDARD;
+  }
+}
+
+
+int main(void)
+{
+  instructions_available ia;
+  ia = getSupportedInstructionSet();
+
+  switch(ia)
+  {
+    case supports_AVX2:
+      printf("AVX2\n");
+      break;
+    case supports_AVX1:
+      printf("AVX1\n");
+      break;
+    case supports_SSE42:
+      printf("SSE42\n");
+      break;
+    case supports_SSE41:
+      printf("SSE41\n");
+      break;
+    case supports_STANDARD:
+      printf("STANDARD\n");
+      break;
+    default:
+      printf("Failed to find supported instruction set, this is an error!\n");
+      exit(-1);
+  }
+  return ia;
+}