[gtk/half-float: 10/11] Add an option to disable F16C

[Matthias Clasen <matthiasc src gnome org>]

commit c966a08b676d80ec0ae3c46c29858f8f2542ba6f
Author: Matthias Clasen <mclasen redhat com>
Date:   Wed Apr 7 19:41:54 2021 -0400

    Add an option to disable F16C
    
    And add a compile time check for the presence of this
    CPU feature.

 config.h.meson        |  2 ++
 gsk/ngl/fp16private.h | 11 ++++++-----
 meson.build           | 36 ++++++++++++++++++++++++++++++++++++
 meson_options.txt     |  5 +++++
 4 files changed, 49 insertions(+), 5 deletions(-)
---
diff --git a/config.h.meson b/config.h.meson
index 549a677379..fc2dd0b3d2 100644
--- a/config.h.meson
+++ b/config.h.meson
@@ -279,3 +279,5 @@
 
 /* Define if tracker3 is available */
 #mesondefine HAVE_TRACKER3
+
+#mesondefine HAVE_F16C
diff --git a/gsk/ngl/fp16private.h b/gsk/ngl/fp16private.h
index f54f7307a6..2bc9f84eea 100644
--- a/gsk/ngl/fp16private.h
+++ b/gsk/ngl/fp16private.h
@@ -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
 
diff --git a/meson.build b/meson.build
index 38d23ed61c..c9586acc29 100644
--- a/meson.build
+++ b/meson.build
@@ -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')
diff --git a/meson_options.txt b/meson_options.txt
index 16a1cbd9fe..4193415151 100644
--- a/meson_options.txt
+++ b/meson_options.txt
@@ -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',