Files
xash3d-fwgs/3rdparty/opus/wscript
Alibek Omarov 2557d5acaa 3rdparty: opus: limit assembly to SSE and SSE2
Otherwise whole engine builds with SSE4.1 and AVX support with LTO enabled. =/
2025-08-03 04:34:46 +05:00

141 lines
4.4 KiB
Python

#! /usr/bin/env python
# encoding: utf-8
import os
FRAGMENT_VLA='''int main (int argc, char **argv) {
char a[argc];
a[sizeof( a ) - 1] = 0;
int N;
return a[0];
}'''
FRAGMENT_ALLOCA_H='''#include <alloca.h>
int main (void) {
int foo=10;
int * array = alloca(foo);
}'''
FRAGMENT_STDLIB_H='''#include <malloc.h>
#include <stdlib.h>
int main (void) {
int foo=10;
int * array = alloca(foo);
}'''
FRAGMENT_LRINT='''#include <math.h>
#include <stdlib.h>
int main (int argc, char **argv) {
return lrint%s((%s)atof(argv[1]));
}'''
# assuming MSVC always enables NEON
FRAGMENT_NEON='''#if !defined __ARM_NEON__ && !defined _MSC_VER
#error
#endif'''
def options(opt):
pass
def configure(conf):
if not conf.path.find_dir('opus') or not conf.path.find_dir('opus/src'):
conf.fatal('Can\'t find opus submodule. Run `git submodule update --init --recursive`.')
return
if conf.check_cc(fragment=FRAGMENT_LRINT % ('', 'double'), msg = 'Checking for C99 lrint', use = 'M', mandatory = False):
conf.define('HAVE_LRINT', 1)
if conf.check_cc(fragment=FRAGMENT_LRINT % ('f', 'float'), msg = 'Checking for C99 lrintf', use = 'M', mandatory = False):
conf.define('HAVE_LRINTF', 1)
# Check for C99 variable-size arrays, or alloca() as fallback
if conf.check_cc(fragment=FRAGMENT_VLA, msg = 'Checking for C99 VLA support', mandatory = False):
conf.define('VAR_ARRAYS', 1)
elif conf.check_cc(fragment=FRAGMENT_ALLOCA_H, msg = 'Checking for alloca in alloca.h header', mandatory = False):
conf.define('USE_ALLOCA', 1)
conf.define('HAVE_ALLOCA_H', 1)
elif conf.check_cc(fragment=FRAGMENT_STDLIB_H, msg = 'Checking for alloca.h in stdlib.h', mandatory = False):
conf.define('USE_ALLOCA', 1)
if conf.env.DEST_CPU in ['thumb', 'arm']:
if conf.check(header_name='arm_neon.h', mandatory = False):
if conf.check(fragment=FRAGMENT_NEON, msg = 'Checking if compiler enabled NEON', mandatory = False):
conf.env.HAVE_NEON = True
conf.define('OPUS_ARM_MAY_HAVE_NEON_INTR', 1)
conf.define('OPUS_ARM_PRESUME_NEON', 1)
elif conf.env.DEST_CPU.startswith('x86'):
if conf.check(header_name='xmmintrin.h', mandatory = False):
if conf.env.COMPILER_CC != 'msvc':
conf.env.CFLAGS += ['-msse']
conf.define('OPUS_X86_MAY_HAVE_SSE', 1)
if conf.env.DEST_SIZEOF_VOID_P > 4:
conf.define('OPUS_X86_PRESUME_SSE', 1)
if conf.check(header_name='emmintrin.h', mandatory = False):
if conf.env.COMPILER_CC != 'msvc':
conf.env.CFLAGS += ['-msse2']
conf.define('OPUS_X86_MAY_HAVE_SSE2', 1)
if conf.env.DEST_SIZEOF_VOID_P > 4:
conf.define('OPUS_X86_PRESUME_SSE2', 1)
# on x86_64 we enable both SSE and SSE2, so RTCD can be disabled (it's actually doesn't even build with RTCD enabled)
if conf.env.DEST_SIZEOF_VOID_P == 4:
if conf.check(header_name='cpuid.h', mandatory=False):
conf.define('CPU_INFO_BY_C', 1)
else:
conf.define('CPU_INFO_BY_ASM', 1)
conf.define('OPUS_HAVE_RTCD', 1)
# if conf.check(header_name='smmintrin.h', mandatory = False):
# if conf.env.COMPILER_CC != 'msvc':
# conf.env.CFLAGS += ['-msse4.1']
# conf.define('OPUS_X86_MAY_HAVE_SSE4_1', 1)
# if conf.check(header_name='immintrin.h', mandatory = False):
# if conf.env.COMPILER_CC != 'msvc':
# conf.env.CFLAGS += ['-mavx']
# conf.define('OPUS_X86_MAY_HAVE_AVX', 1)
# TODO: ARM/x86 intrinsics detection
# TODO: maybe call autotools/cmake/meson instead?
def build(bld):
sources = bld.path.ant_glob([
'opus/src/*.c',
'opus/celt/*.c',
'opus/silk/*.c',
'opus/silk/float/*.c'
], excl = [
'opus/src/repacketizer_demo.c',
'opus/src/opus_demo.c',
'opus/src/opus_compare.c',
'opus/celt/opus_custom_demo.c'
])
includes = ['opus', 'opus/include/', 'opus/celt/', 'opus/silk/', 'opus/silk/float/']
if bld.env.DEST_CPU in ['thumb', 'arm']:
if bld.env.HAVE_NEON:
sources += bld.path.ant_glob(['opus/silk/arm/*.c', 'opus/celt/arm/*.c'])
includes += ['opus/silk/arm', 'opus/celt/arm']
elif bld.env.DEST_CPU.startswith('x86'):
sources += ['opus/silk/x86/x86_silk_map.c',
'opus/celt/x86/pitch_sse.c',
'opus/celt/x86/pitch_sse2.c',
'opus/celt/x86/vq_sse2.c',
'opus/celt/x86/x86_celt_map.c',
'opus/celt/x86/x86cpu.c']
includes += ['opus/silk/x86', 'opus/celt/x86']
defines = ['OPUS_BUILD', 'FLOAT_APPROX', 'PACKAGE_VERSION="1.4.0"', 'CUSTOM_MODES', 'ENABLE_HARDENING']
bld.stlib(
source = sources,
target = 'opus',
features = 'c',
includes = includes,
defines = defines,
export_includes = ['opus/include/']
)