From 4fea7a091d7347b0c947011eded99da7bd0c2f23 Mon Sep 17 00:00:00 2001
From: Sam Hocevar <sam@hocevar.net>
Date: Mon, 29 Aug 2011 00:07:56 +0000
Subject: [PATCH] core: implement array versions of the float / half conversion
routines.
---
src/half.cpp | 32 +++++++++++++++++++++++++++++---
src/half.h | 5 +++++
2 files changed, 34 insertions(+), 3 deletions(-)
diff --git a/src/half.cpp b/src/half.cpp
index 1a1c062e..36065678 100644
--- a/src/half.cpp
+++ b/src/half.cpp
@@ -114,8 +114,8 @@ static uint32_t const shiftmagic = 0x07c4acddu;
/* Lookup table-based algorithm from “Fast Half Float Conversions”
* by Jeroen van der Zijp, November 2008. Tables are generated using
* the C++ preprocessor, thanks to a branchless implementation also
- * used in half_to_float_branch(). This code is actually almost always
- * slower than the branching one. */
+ * used in half_to_float_branch(). This code is very fast when performing
+ * conversions on arrays of values. */
static inline uint32_t half_to_float_nobranch(uint16_t x)
{
#define M3(i) ((i) | ((i) >> 1))
@@ -157,7 +157,9 @@ static inline uint32_t half_to_float_nobranch(uint16_t x)
}
/* This algorithm is similar to the OpenEXR implementation, except it
- * uses branchless code in the denormal path. */
+ * uses branchless code in the denormal path. This is slower than the
+ * table version, but will be more friendly to the cache for occasional
+ * uses. */
static inline uint32_t half_to_float_branch(uint16_t x)
{
uint32_t s = (x & 0x8000u) << 16;
@@ -213,5 +215,29 @@ half::operator float() const
return u.f;
}
+size_t half::copy(half *dst, float const *src, size_t nelem)
+{
+ for (size_t i = 0; i < nelem; i++)
+ {
+ union { float f; uint32_t x; } u;
+ u.f = *src++;
+ *dst++ = makebits(float_to_half_nobranch(u.x));
+ }
+
+ return nelem;
+}
+
+size_t half::copy(float *dst, half const *src, size_t nelem)
+{
+ for (size_t i = 0; i < nelem; i++)
+ {
+ union { float f; uint32_t x; } u;
+ u.x = half_to_float_nobranch((*src++).bits);
+ *dst++ = u.f;
+ }
+
+ return nelem;
+}
+
} /* namespace lol */
diff --git a/src/half.h b/src/half.h
index d3f68eae..e301d91b 100644
--- a/src/half.h
+++ b/src/half.h
@@ -16,6 +16,7 @@
#if !defined __LOL_HALF_H__
#define __LOL_HALF_H__
+#include <cstdio>
#include <stdint.h>
namespace lol
@@ -55,6 +56,10 @@ public:
operator float() const;
inline operator int() const { return (int)(float)*this; }
+ /* Array conversions */
+ static size_t copy(half *dst, float const *src, size_t nelem);
+ static size_t copy(float *dst, half const *src, size_t nelem);
+
/* Operations */
inline half operator -() { return makebits(bits ^ 0x8000u); }
inline half &operator +=(float f) { return (*this = (half)(*this + f)); }