Add an option to disable F16C

And add a compile time check for the presence of this
CPU feature.
This commit is contained in:
Matthias Clasen 2021-04-07 19:41:54 -04:00
parent 38c6fe2ce4
commit 9d81c129fc
5 changed files with 49 additions and 6 deletions

View File

@ -279,3 +279,5 @@
/* Define if tracker3 is available */
#mesondefine HAVE_TRACKER3
#mesondefine HAVE_F16C

View File

@ -221,7 +221,6 @@ libgsk = static_library('gsk',
'-DGTK_COMPILATION',
'-DG_LOG_DOMAIN="Gsk"',
'-DG_LOG_STRUCTURED=1',
'-mf16c'
] + common_cflags,
link_with: libgdk,
)

View File

@ -1,4 +1,4 @@
/* ninesliceprivate.h
/* fp16private.h
*
* Copyright 2021 Red Hat, Inc.
*
@ -21,10 +21,11 @@
#ifndef __FP16_PRIVATE_H__
#define __FP16_PRIVATE_H__
#include <config.h>
#include <glib.h>
#include <graphene.h>
#ifdef GRAPHENE_USE_SSE
#ifdef HAVE_F16C
#include <immintrin.h>
#endif
@ -34,7 +35,7 @@ G_BEGIN_DECLS
#define FP16_ONE ((guint16)15360)
#define FP16_MINUS_ONE ((guint16)48128)
#ifdef GRAPHENE_USE_SSE
#ifdef HAVE_F16C
static inline void
float_to_half4 (const float f[4],
@ -54,7 +55,7 @@ half_to_float4 (const guint16 h[4],
_mm_store_ps (f, s);
}
#else /* GRAPHENE_USE_SSE */
#else /* GTK_HAS_F16C */
static inline guint
as_uint (const float x)
@ -108,7 +109,7 @@ half_to_float4 (const guint16 h[4],
f[3] = half_to_float (h[3]);
}
#endif /* GRAPHENE_USE_SSE */
#endif /* HAVE_F16C */
G_END_DECLS

View File

@ -685,6 +685,42 @@ if graphene_has_sse2 or graphene_has_gcc
endif
endif
f16c_cflags = []
if get_option('f16c').enabled()
f16c_prog = '''
#if defined(__GNUC__)
# if !defined(__amd64__) && !defined(__x86_64__)
# error "F16C intrinsics are only available on x86_64"
# endif
#elif defined (_MSC_VER) && !defined (_M_X64) && !defined (_M_AMD64)
# error "F16C intrinsics not supported on x86 MSVC builds"
#endif
#if defined(__SSE__) || (_M_X64 > 0)
# include <immintrin.h>
#else
# error "No F16C intrinsics available"
#endif
int main () {
float f[4] = { 0, };
unsigned short h[4] = { 0, };
__m128 s = _mm_loadu_ps (f);
__m128i i = _mm_cvtps_ph (s, 0);
_mm_storel_epi64 ((__m128i*)h, i);
return 0;
}'''
if cc.get_id() != 'msvc'
test_f16c_cflags = [ '-mf16c' ]
else
test_f16c_cflags = []
endif
if cc.compiles(f16c_prog, args: test_f16c_cflags, name: 'F16C intrinsics')
cdata.set('HAVE_F16C', 1)
f16c_cflags = test_f16c_cflags
common_cflags += test_f16c_cflags
endif
endif
subdir('gtk/css')
subdir('gdk')
subdir('gsk')

View File

@ -86,6 +86,11 @@ option('sassc',
value: 'auto',
description: 'Rebuild themes using sassc')
option('f16c',
type: 'feature',
value: 'enabled',
description: 'Enable F16C fast paths (requires F16C)')
# Documentation and introspection
option('gtk_doc',