[FFmpeg-devel] [PATCH] Fast half-float to float conversion

[Jimmy Christensen]

On 2009-07-01 08:38, Reimar D?ffinger wrote:
> On Wed, Jul 01, 2009 at 07:42:45AM +0200, Jimmy Christensen wrote:
>> On 2009-06-30 15:50, Reimar D?ffinger wrote:
>>> On Tue, Jun 30, 2009 at 03:09:33PM +0200, Jimmy Christensen wrote:
>>>>> Lastly it seems to me that this code would belong into
>>>>> libavutil/intfloat_readwrite.*
>>>>
>>>> You are probably right.
>>>
>>> Personally I'd suggest doing a stupid implementation for
>>> libavutil/intfloat_readwrite.*, something like this (certainly
>>> not a beauty, possibly buggy, does not support denormals, though since
>>> they are rare it probably wouldn't be a big issue to add handling for
>>> them in if (nosign<   0x0400)).
>>> float av_int2halflt(int16_t v){
>>>       uint16_t nosign = v+v;
>>>       if (nosign>= 0xfc00) {
>>>           if (nosign == 0xfc00) return v>>15 ? -1.0/0.0 : 1.0/0.0;
>>>           else return 0.0/0.0;
>>>       }
>>>       if (nosign<   0x0400) return 0; // denormal or 0
>>>       return ldexp((v&0x3ff) + (1<<11)) * (v>>15|1), (v>>10&0x1f)-26);
>>> }
>>>
>>>
>>> And consider an optimized version only after it is actually used and
>>> thus it is clear what kind of optimizations are necessary/most useful.
>>> E.g. the table based approach probably is hard to SIMD and thus might be
>>> a bad idea if you want to convert a lot of data.
>>
>> I tried testing with the code that you posted but it doesn't seem to
>> work? The image looks quite clamped.
>
> Well, I already said it is possibly buggy. E.g. the (1<<11) should be
> (1<<10).

Looks more correct now, but still a very dark image. I multiplied with 2 
and it seems to match the table lookup method. Is this correct?

float av_int2halflt(int16_t v){
      uint16_t nosign = v+v;
      if (nosign>= 0xfc00) {
          if (nosign == 0xfc00) return v>>15 ? -1.0/0.0 : 1.0/0.0;
          else return 0.0/0.0;
      }
      if (nosign<  0x0400) return 0; // denormal or 0
      return ldexp((v&0x3ff) + (1<<10) * (v>>15|1), (v>>10&0x1f)-26)*2;
}

I really appreciate this code sample.

>
>> Also according to the original author the table conversion does
>> implement denormals, zero, infinite and NaN.
>
> Yes, I know now. Implementing denormals for this code isn't hard either.
> But optimizations when the code that is to be optimized is not even
> available yet makes very little sense to me.
> If you can get a away with ignoring the INF/NaN and denormal case, a
> SIMD version can probably convert at least 4 floats in as many cycles as
> the table code needs for 1, not even counting possible memory latencies.

I would love to do such optimizations if I was able to. Right now I'm 
just looking for some reference code which can be used to do the job 
which is simple and not too bloated coding wise. Implementing the table 
based is not a good solution for smaller systems, but never had these 
systems in mind when proposing the code for half-float > float.