Clamp incoming float values to half, instead of simply casting, on

encode.

Casting can introduce Infs, which are zero'ed later on, prior to the
forward DCT step. This can have the nasty side effect of forcing
bright values to zero, instead of clamping them to 65k.

OpenEXR/IlmImf/ImfDwaCompressor.cpp

@@ -149,6 +149,7 @@
 #include "ImathBox.h"
 #include "ImathVec.h"
 #include "half.h"
+#include "halfLimits.h"
 
 #include "dwaLookups.h"
 
@@ -1408,6 +1409,15 @@ DwaCompressor::LossyDctEncoderBase::execute ()
             {
 
                 Xdr::read<CharPtrIO> (srcXdr, src);
+
+                //
+                // Clamp to half ranges, instead of just casting. This
+                // avoids introducing Infs which end up getting zeroed later
+                //
+                src = std::max (
+                    std::min ((float) std::numeric_limits<half>::max(), src),
+                              (float)-std::numeric_limits<half>::max());
+
                 Xdr::write<CharPtrIO> (dstXdr, ((half)src).bits());
 
                 //