[PATCH] Move definition of INFINITY

Runtimes OpenCL
#1

We were missing a declaration of INFINITY in the clc headers and already
had one in gen_convert.py

NOTE: This may not necessarily be the correct value of INFINITY based on
the cl_platform.h headers

Signed-off-by: Aaron Watry <awatry@gmail.com>

#2

You probably want to put parentheses around that (just in case).

#define INFINITY (1.0f / 0.0f)

Otherwise, looks good to me. The same should probably be done for NAN, which I think also occurs somewhere already.

Jeroen

#3

You probably want to put parentheses around that (just in case).

#define INFINITY (1.0f / 0.0f)

Otherwise, looks good to me. The same should probably be done for NAN, which I think also occurs somewhere already.

Jeroen

This should be ___builtin_inf/ __builtin_inff

I believe these will have different warning behaviour

#4

You probably want to put parentheses around that (just in case).

#define INFINITY (1.0f / 0.0f)

Otherwise, looks good to me. The same should probably be done for NAN, which I think also occurs somewhere already.

Jeroen

This should be ___builtin_inf/ __builtin_inff

I believe these will have different warning behaviour

Yes, this changes at least -Wfloat-equal

kernel void foo(global int* out, float x)
{
     *out = (1.0f / 0.0f) == x; // warning: comparing floating point with == or != is unsafe [-Wfloat-equal]
}

kernel void bar(global int* out, float x)
{
     *out = __builtin_inff() == x; // No warning
}

#5

I'll change this to _builtin_inff() and also move the definition of
FLT_NAN/DBL_NAN from the nextafter implementation to definitions.h
(using _builtin_nan[f]), rename them to their proper names and then
send a v2.

--Aaron