Rewrite the SContruct.

- Works with new /strategy/ structure.
- Change architecture selection to use arch= instead of construction target.

SConstruct

@@ -1,116 +1,236 @@
 """ SConstruct file fore building Kvazaar with Scons.
 
-This file defines two targets, x86 and x64, and builds construction
+This is an additional cross platform way to build the program.
-environments for both.
+The main way is to use the Makefile and that is always kept up to date.
 
-TODO:
+This mostly exists so it can be used with a custom SConscript file that
-- add debug builds
+builds any old version of Kvazaar, ensuring that the compilation settings are
-- whole program optimization for gcc
+the same and any compilation from a specific version of Kvazaar will not
+affect the comparison. The SConscript included in git version only compiles
+the current version.
 
 """
 
 import os
 import platform
 
-Help("""
+
-Type: 'scons x86' to build a 32-bit release version,
+vars = Variables()
-     'scons x64' to build a 64-bit release version.
+
-     'scons x86 x64 -c' to clear all build directories.
+default_arch = 'amd64' if platform.machine().endswith('64') else 'x86'
-""")
+vars.Add(EnumVariable('arch', 'Set target arch.', default_arch,
+        allowed_values=('x86', 'x64', 'amd64', 'ppc'),
+        map={'x64': 'amd64'}))
+
+vars.Add(PathVariable('win32pthreads',
+        'Path to win32-pthreads dir.',
+        r'./../pthreads.2'))
+
+vars.Add(PathVariable('copyto',
+        'Copy exe and required DLLs to this dir.',
+        '',
+        PathVariable.PathAccept))
+
+vars.Add(BoolVariable('dump_env',
+        'Dump of construction environment to stdout.',
+        False))
+
+vars.Add(BoolVariable('use_yasm',
+        'Use yasm.',
+        True))
 
 
-# Create construction environments for 32 and 64 bit builds.
 # Visual studio needs the architecture to be set in this stage. It can not be
-# modified later.
+# modified later. Other tools ignore TARGET_ARCH.
-env_x86 = Environment(
+# The variable substitution is done in the environment construction so $arch
-        #tools=['mingw'],
+# will not work here. Get the value manually.
-        ASCOM='yasm $ASFLAGS -o $TARGET $SOURCES',
+arch = ARGUMENTS.get('arch', default_arch)
-        ENV={'PATH': os.environ['PATH']},  # to find yasm on Windows
+if arch == 'x64':
-        TARGET_ARCH='x86',  # for Visual Studio
+  arch = 'amd64'
-        )
+env = Environment(
-env_x64 = Environment(
+        variables=vars,
-        #tools=['mingw'],
+        tools=['msvc', 'mslink', 'nasm'],
-        ASCOM='yasm $ASFLAGS -o $TARGET $SOURCES',
+        ENV={'PATH': os.environ['PATH']},  # to find yasm
-        ENV={'PATH': os.environ['PATH']}, # to find yasm on Windows
+        TARGET_ARCH=arch,  # for Visual Studio
-        TARGET_ARCH='amd64',  # for Visual Studio
+        AS='yasm',
         )
 
+
+Help("""
+Example: 'scons arch=x64' to compile for amd64.
+         'scons --jobs=8' to compile in parallel.
+""" + vars.GenerateHelpText(env))
+
+if 'MSVS' in env:
+    compiler = 'msvs'
+elif 'MSYSTEM' in os.environ:
+	compiler = 'mingw'
+else:
+	compiler = 'gcc'
+
+env['use_yasm'] = env['use_yasm'] and env['arch'] not in ('ppc',)
+
+# pthreads_dll location for copyto
+pthreads_dll = None
+
+# Try and deal with all the remaining compiler specific differences that I
+# really wish scons would handle.
+if compiler in ('msvs',):
+    arch_dir = {'x86': r'x86', 'amd64': r'x64'}[env['arch']]
+    pthreads_dll = os.path.join(env['win32pthreads'], 'dll', arch_dir, 'pthreadVC2.dll')
+    
+    env.Append(
+            CCFLAGS=r'/MD /Ox /GL /wd"4028"',
+            LINKFLAGS=r'/LTCG',
+            LIBPATH=r'#$win32pthreads\lib\\' + arch_dir,
+            LIBS=r'pthreadVC2',
+            CPPPATH=r'#$win32pthreads\include')
+else:
+    env.MergeFlags('-O2 -pthread -lm -lrt -march=native')
+    if env['arch'] == 'x86':
+        env.MergeFlags('-m32')
+    elif env['arch'] == 'amd64':
+        env.MergeFlags('-m64')
+    
+    # platform.system() lies on msys2, so just try to detect msys/mingw.
+    if 'MSYSTEM' in os.environ:
+        # __USE_MINGW_ANSI_STDIO required on mingw for printf to function.
+        env.MergeFlags('-D' + '__USE_MINGW_ANSI_STDIO=1')
+
+
+# VS2010 linker and mingw64 need TMP.
+if compiler in ('msvs', 'mingw'):
+	if 'TMP' in os.environ:
+	    env['ENV']['TMP'] = os.environ['TMP']
+
+env.MergeFlags('-I. -Iextras -Istrategies')
+
+# Take comma separated list from 'D=' and pass them on to
+# preprocessor as defines.
+preprocessor_defines = ARGUMENTS.get('D', '')
+if preprocessor_defines:
+    for define in preprocessor_defines.split(','):
+        env.MergeFlags('-D' + define)
+
+# Declare build targets.
+variant_dir = 'scons_build'
+
+
+class EnvinronmentContainer(object):
+    """Class for initializing and holding optimization environments.
+    
+    As far as I know, making a separate environment is the only way to compile
+    different objects with different flags in scons. So this constructs all
+    the required environments.
+    
+    Yasm is also its own environment although it's not strictly necessary.
+    It does however allow KVZ_COMPILE_ASM to only those files that require
+    it, avoiding a complete recompile.
+    
+    """
+    
+    def __init__(self, env, compiler, arch):
+        # If optimization is not supported for arch, use default env.
+        self.env = env
+        self.sse2 = env
+        self.sse41 = env
+        self.avx = env
+        self.avx2 = env
+        self.altivec = env
+        self.asm = env
+        
+        if compiler == 'msvs':
+            self.avx = env.Clone()
+            self.avx.Append(CCFLAGS='/arch:AVX')
+            self.avx2 = env.Clone()
+            self.avx2.Append(CCFLAGS='/arch:AVX2')
+        elif compiler in ('gcc',) and arch in ('x86', 'x64'):
+            self.sse2 = env.Clone()
+            self.sse2.Append(CCFLAGS='-msse2')
+            self.sse41 = env.Clone().Append(CCFLAGS='-msse4.1')
+            self.avx = env.Clone().Append(CCFLAGS='-mavx')
+            self.avx2 = env.Clone().Append(CCFLAGS='-mavx2')
+        elif compiler in ('gcc',) and arch in ('ppc'):
+            self.altivec = env.Clone().Append(CCFLAGS='-maltivec')
+
+        if env['use_yasm']:
+            self._init_yasm()
+    
+    
+    def _init_yasm(self):
+        self.asm = env.Clone()
+        self.asm.MergeFlags('-D' + 'KVZ_COMPILE_ASM')
+        
         # YASM flags for different architectures. The object file format and name
         # mangling must be the same as used by the C compiler for that OS.
         # Indexed by platform.system().
         yasm_flags = {
             'Windows': {
-        'x86': '-f win32 -DPREFIX',
+                'x86':   '-f win32 -DPREFIX -DHAVE_ALIGNED_STACK=0 ',
-        'x64': '-f win64'},
+                'amd64': '-f win64 -DHAVE_ALIGNED_STACK=1',
+            },
             'Darwin': {
                 'x86':   '-f macho32 -DPREFIX',
-        'x64': '-f macho64 -DPREFIX'},
+                'amd64': '-f macho64 -DPREFIX',
+            },
             'Linux': {  # Flags for Unix-like.
                 'x86':   '-f elf32',
-        'x64': '-f elf64'},
+                'amd64': '-f elf64',
+            },
             'all': {  # Flags for all systems.
-        'x86': ' -DARCH_X86_64=0 -m x86',
+                'x86':   ' -I./src/extras -DARCH_X86_64=0 -m x86',
-        'x64': ' -DARCH_X86_64=1 -m amd64'},
+                'amd64': ' -I./src/extras -DARCH_X86_64=1 -m amd64',
+            },
         }
+
         # Set yasm flags for OS and architecture.
         target_yasm_flags = yasm_flags.get(platform.system(), yasm_flags['Linux'])
-env_x86.Replace(ASFLAGS=target_yasm_flags['x86'])
+        self.asm.Replace(ASFLAGS=target_yasm_flags[env['arch']])
-env_x64.Replace(ASFLAGS=target_yasm_flags['x64'])
+        self.asm.Append(ASFLAGS=yasm_flags['all'][env['arch']])
-env_x86.Append(ASFLAGS=yasm_flags['all']['x86'])
+        self.asm.Replace(AS='yasm')
-env_x64.Append(ASFLAGS=yasm_flags['all']['x64'])
 
 
-# Try and deal with all the remaining compiler specific differences that I
+envs = EnvinronmentContainer(env, compiler, env['arch'])
-# really with scons would handle.
-if 'MSVS' in env_x86:
-    # /MD = multithreaded DLL runtime
-    # /Ox = full optimization, sets /Ob2, /Og, /Oi, /Ot, /Oy
-    # /GL = enable whole program optimization
-    # /LTCG = link time code generation
-    # /arch:SSE2 = use SSE2 (x86 only)
-    # win32-pthreads is assumed to be in same level as kvazaar in dir pthreads
-    env_x86.Append(
-            CCFLAGS=r'/MD /Ox /GL /arch:SSE2 /I"..\..\pthreads\include"',
-            LINKFLAGS=r'/LTCG /LIBPATH:"..\..\pthreads\x86" "pthreadVC2.lib"')
-    env_x64.Append(
-            CCFLAGS=r'/MD /Ox /GL /I"..\..\pthreads\include"',
-            LINKFLAGS=r'/LTCG /LIBPATH:"..\pthreads\x64" "pthreadVC2.lib"')
-else:
-    # GCC flags
-    # -m for arch, -O2 for optimization, -lm for math lib
-    env_x86.MergeFlags('-m32 -O2 -lm -march=native -pthread')
-    env_x64.MergeFlags('-m64 -O2 -lm -march=native -pthread')
 
-# VS2010 linker and mingw64 need TMP.
-if 'TMP' in os.environ:
-    env_x86['ENV']['TMP'] = os.environ['TMP']
-    env_x64['ENV']['TMP'] = os.environ['TMP']
 
-env_x86.MergeFlags('-I. -Iextras -Istrategies')
+program = SConscript('src/SConscript',
-env_x64.MergeFlags('-I. -Iextras -Istrategies')
+        exports={'envs': envs},
-    
+        variant_dir=variant_dir,
-preprocessor_defines = ARGUMENTS.get('D', '')
-if preprocessor_defines:
-    for define in preprocessor_defines.split():
-        env_x86.MergeFlags('-D' + define)
-        env_x64.MergeFlags('-D' + define)
-
-# Declare build targets.
-x86 = SConscript('src/SConscript',
-        exports={'env': env_x86},
-        variant_dir='scons_build_x86',
         duplicate=False)
-Alias('x86', x86)
-
-x64 = SConscript('src/SConscript',
-        exports={'env': env_x64},
-        variant_dir='scons_build_x64',
-        duplicate=False)
-Alias('x64', x64)
 
 
-if platform.machine().endswith('64'):
+# Copy exe and dll's to the path from 'copyto=' argument.
-    Default(x64)
+copy_dir = env['copyto']
+if copy_dir:
+    env.Install(copy_dir, program)
+    if pthreads_dll:
+        env.Install(copy_dir, pthreads_dll)
+    env.Alias('copyto', copy_dir)
+    Default([program, 'copyto'])
 else:
-    Default(x86)
+    Default(program)
+
+
+# Dumpo environment to stdout.
+if env['dump_env']:
+    import difflib
+    
+    def get_diff(orig_env, new_env):
+        return "\n".join(difflib.unified_diff(orig_env, new_env.Dump().splitlines(), n=0))
+    
+    env_dump = env.Dump().splitlines()
+    print "== env"
+    print "\n".join(env_dump)
+    print "== diff env envs.sse2"
+    print get_diff(env_dump, envs.sse2)
+    print "== diff env envs.sse41"
+    print get_diff(env_dump, envs.sse41)
+    print "== diff env envs.avx"
+    print get_diff(env_dump, envs.avx)
+    print "== diff env envs.avx2"
+    print get_diff(env_dump, envs.avx2)
+    print "== diff env envs.altivec"
+    print get_diff(env_dump, envs.altivec)
+    print "== diff env envs.asm"
+    print get_diff(env_dump, envs.asm)
     

src/SConscript

@@ -3,14 +3,25 @@
 
 import os
 
-Import('env')
+Import('envs')
+env = envs.env
 
 sources = []
 sources += env.Glob('*.c')
 sources += env.Glob('extras/*.c')
+
 sources += env.Glob('strategies/*.c')
-sources += env.Glob('x86/*.asm')
+sources += env.Glob('strategies/generic/*.c')
-env.Depends('x86/cpu.asm', 'x86/x86inc.asm')
+sources += envs.sse2.Object(env.Glob('strategies/sse2/*.c'))
+sources += envs.sse41.Object(env.Glob('strategies/sse41/*.c'))
+sources += envs.avx.Object(env.Glob('strategies/avx/*.c'))
+sources += envs.avx2.Object(env.Glob('strategies/avx2/*.c'))
+sources += envs.altivec.Object(env.Glob('strategies/altivec/*.c'))
+
+sources += envs.asm.Object(env.Glob('strategies/x86_asm/*.c'))
+if env['use_yasm']:
+    sources += envs.asm.Object(env.Glob('strategies/x86_asm/*.asm'))
+    sources += envs.asm.Object(env.Glob('extras/*.asm'))
 
 prog = env.Program('kvazaar', sources)