diff options
author | Matthias Clasen <mclasen@redhat.com> | 2021-04-12 20:35:10 +0000 |
---|---|---|
committer | Matthias Clasen <mclasen@redhat.com> | 2021-04-12 20:35:10 +0000 |
commit | 442f485591b65185538ab941b58743cadd1d2092 (patch) | |
tree | a3845cb1fd44aae0e22cf315b28ab9a361fd2098 | |
parent | 74238b542981ab9a63bcdad5d9bc12da5168e083 (diff) | |
parent | 65b35457fdfde69fcc8aaed516a6406855c0dbbc (diff) | |
download | gtk+-442f485591b65185538ab941b58743cadd1d2092.tar.gz |
Merge branch 'f16c.msvc' into 'master'
Fix F16C build/runtime detection on Visual Studio
See merge request GNOME/gtk!3429
-rw-r--r-- | gsk/ngl/fp16.c | 63 | ||||
-rw-r--r-- | meson.build | 6 |
2 files changed, 63 insertions, 6 deletions
diff --git a/gsk/ngl/fp16.c b/gsk/ngl/fp16.c index e9a8c8fbce..1e11faafd8 100644 --- a/gsk/ngl/fp16.c +++ b/gsk/ngl/fp16.c @@ -80,6 +80,12 @@ half_to_float4_c (const guint16 h[4], #ifdef HAVE_F16C +#if defined(_MSC_VER) && !defined(__clang__) +#define CAST_M128I_P(a) (__m128i const *) a +#else +#define CAST_M128I_P(a) (__m128i_u const *) a +#endif + static void float_to_half4_f16c (const float f[4], guint16 h[4]) @@ -93,11 +99,61 @@ static void half_to_float4_f16c (const guint16 h[4], float f[4]) { - __m128i i = _mm_loadl_epi64 ((__m128i_u const *)h); + __m128i i = _mm_loadl_epi64 (CAST_M128I_P (h)); __m128 s = _mm_cvtph_ps (i); + _mm_store_ps (f, s); } +#undef CAST_M128I_P + +#if defined(_MSC_VER) && !defined(__clang__) +/* based on info from https://walbourn.github.io/directxmath-f16c-and-fma/ */ +static gboolean +have_f16c_msvc (void) +{ + static gboolean result = FALSE; + static gsize inited = 0; + + if (g_once_init_enter (&inited)) + { + int cpuinfo[4] = { -1 }; + + __cpuid (cpuinfo, 0); + + if (cpuinfo[0] > 0) + { + __cpuid (cpuinfo, 1); + + if ((cpuinfo[2] & 0x8000000) != 0) + result = (cpuinfo[2] & 0x20000000) != 0; + } + + g_once_init_leave (&inited, 1); + } + + return result; +} + +void +float_to_half4 (const float f[4], guint16 h[4]) +{ + if (have_f16c_msvc ()) + float_to_half4_f16c (f, h); + else + float_to_half4_c (f, h); +} + +void +half_to_float4 (const guint16 h[4], float f[4]) +{ + if (have_f16c_msvc ()) + half_to_float4_f16c (h, f); + else + half_to_float4_c (h, f); +} + +#else void float_to_half4 (const float f[4], guint16 h[4]) __attribute__((ifunc ("resolve_float_to_half4"))); void half_to_float4 (const guint16 h[4], float f[4]) __attribute__((ifunc ("resolve_half_to_float4"))); @@ -120,11 +176,12 @@ resolve_half_to_float4 (void) else return half_to_float4_c; } +#endif #else -#ifdef __APPLE__ -// turns out aliases don't work on Darwin +#if defined(__APPLE__) || (defined(_MSC_VER) && !defined(__clang__)) +// turns out aliases don't work on Darwin nor Visual Studio void float_to_half4 (const float f[4], diff --git a/meson.build b/meson.build index 641208158c..26cda67aba 100644 --- a/meson.build +++ b/meson.build @@ -692,10 +692,8 @@ if get_option('f16c').enabled() # 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) +#if defined(__SSE__) || defined(_MSC_VER) # include <immintrin.h> #else # error "No F16C intrinsics available" @@ -707,8 +705,10 @@ int main () { __m128i i = _mm_cvtps_ph (s, 0); _mm_storel_epi64 ((__m128i*)h, i); +#if defined (__GNUC__) || defined (__clang__) __builtin_cpu_init (); __builtin_cpu_supports ("f16c"); +#endif return 0; }''' |