X-Git-Url: https://git.sesse.net/?p=movit;a=blobdiff_plain;f=fp16.h;h=c21153b3466fe71424283063cb7cf9b775a89153;hp=4e6652f689fb79affad6bb5d0c3cc0765afa5a3c;hb=0c821b2eb013de712041d2b337612b9a6297695f;hpb=cea222cc88f1e210cd73331d74d0b33a67de3d81

diff --git a/fp16.h b/fp16.h
index 4e6652f..c21153b 100644
--- a/fp16.h
+++ b/fp16.h
@@ -1,6 +1,10 @@
 #ifndef _MOVIT_FP16_H
 #define _MOVIT_FP16_H 1
 
+#ifdef __F16C__
+#include <immintrin.h>
+#endif
+
 // Code for converting to and from fp16 (from fp64), without any particular
 // machine support, with proper IEEE round-to-even behavior (and correct
 // handling of NaNs and infinities). This is needed because some OpenGL
@@ -10,12 +14,39 @@
 
 namespace movit {
 
-typedef unsigned int fp32_int_t;
-typedef unsigned short fp16_int_t;
+// structs instead of ints, so that they are not implicitly convertible.
+struct fp32_int_t {
+	unsigned int val;
+};
+struct fp16_int_t {
+	unsigned short val;
+};
+
+#ifdef __F16C__
+
+// Use the f16c instructions from Haswell if available (and we know that they
+// are at compile time).
+static inline double fp16_to_fp64(fp16_int_t x)
+{
+	return _cvtsh_ss(x.val);
+}
+
+static inline fp16_int_t fp64_to_fp16(double x)
+{
+	// NOTE: Strictly speaking, there are some select values where this isn't correct,
+	// since we first round to fp32 and then to fp16.
+	fp16_int_t ret;
+	ret.val = _cvtss_sh(x, 0);
+	return ret;
+}
+
+#else
 
 double fp16_to_fp64(fp16_int_t x);
 fp16_int_t fp64_to_fp16(double x);
 
+#endif
+
 // These are not very useful by themselves, but are implemented using the same
 // code as the fp16 ones (just with different constants), so they are useful
 // for verifying against the FPU in unit tests.