Re: Event Invitation: PNG marketing meeting

> (Oh, and also if the floats compress roughly as well.)

I think lossless half-float .PNG can actually beat .EXR in Deflate mode for
several reasons, but I need to do some testing with well-configured
software for both formats and see what happens. PNG's predictors are
valuable here.

Note lossy half-float PNG using a RDO (rate distortion optimization)
encoder with a HDR colorspace error metric is also solidly on our radar for
the future. We use such technology today on various GPU texture formats.





On Mon, Nov 13, 2023 at 5:27 PM Richard Geldreich <rich@binomial.info>
wrote:

> > The reason I said "traditionally" above is Intel's new chips have
> hardware zlib decompression.
>
> AMD also has supported hardware Deflate for many years, which is exposed
> on at least one major game console. I don't know when (if ever) it'll be
> exposed on common OS's, but hopefully one day.
>
> On Mon, Nov 13, 2023 at 4:46 PM Chris Blume (ProgramMax) <
> programmax@gmail.com> wrote:
>
>>   RE: hardware acceleration & PNG
>>
>> You are right that traditionally, a pre-computed texture in a game might
>> use a block compression that the GPU can handle directly. Streaming
>> compressions (like zlib, which PNG uses) aren't traditionally used in the
>> hardware itself. These block compressions have major limitations, though.
>> They are lossy in ways that are sometimes unacceptable. Which is where PNG
>> (and also TGA) fill the void.
>>
>> Rich and Stephanie at Binomial created a transcoder that goes from a
>> streaming compression to a block compression that the GPU uses. So the line
>> between hardware acceleration and PNG is already becoming blurred.
>>
>>
>> The reason I said "traditionally" above is Intel's new chips have
>> hardware zlib decompression.
>> https://www.intel.com/content/www/us/en/architecture-and-technology/intel-quick-assist-technology-overview.html
>> So the line is becoming even more blurred than it already was.
>>
>>
>>
>> RE: Windows compositing HDR using half floats
>>
>> When PNG was first created, floats were slow and rarely ever useful for
>> images. Ints were used throughout the pipeline. During that time, it was
>> natural for PNG to use ints for storage. As (half) floats become more
>> common, the conversion from int to float might become a needless step. This
>> really hinges on if an int texture rendering to a (half) float target
>> introduces inefficiencies. I don't know the answer to that. (Oh, and also
>> if the floats compress roughly as well.)
>>
>> On Mon, Nov 13, 2023 at 4:34 PM Richard Geldreich <rich@binomial.info>
>> wrote:
>>
>>> >But is PNG really the right choice of format for real-time stuff?
>>>
>>> BTW - from my perspective, which is browser/video games/real-time
>>> rendering oriented, I could also ask "is PNG really the right choice for
>>> broadcast stuff"?
>>>
>>>
>>> On Mon, Nov 13, 2023 at 2:45 PM Richard Geldreich <rich@binomial.info>
>>> wrote:
>>>
>>>> .PNG is already incredibly popular in video games, real-time rendering,
>>>> visualization, etc. They already use half-floats at various stages in their
>>>> pipelines.
>>>>
>>>> Note the Windows compositer internally now uses IEEE half-floats:
>>>>
>>>>
>>>> https://learn.microsoft.com/en-us/windows/win32/direct3darticles/high-dynamic-range
>>>>
>>>> FP16: "When Advanced Color is enabled, the DWM performs its
>>>> composition using IEEE half-precision floating point (FP16), eliminating
>>>> any bottlenecks, and allowing the full precision of the display to be used."
>>>>
>>>> [image: image.png]
>>>>
>>>> On Mon, Nov 13, 2023 at 2:12 PM Leo Barnes <lbarnes@apple.com> wrote:
>>>>
>>>>>
>>>>>
>>>>> On Nov 13, 2023, at 18:56, Richard Geldreich <rich@binomial.info>
>>>>> wrote:
>>>>>
>>>>> A note about marketing "HDR" PNG: What PNG is calling "HDR" does not
>>>>> have nearly enough dynamic range to be called "HDR" in some segments of the
>>>>> industry, such as in real-time rendering. Even 12-bits with a transfer
>>>>> function is nowhere near enough for our needs. Importantly, we need at
>>>>> least 5-bit exponents.
>>>>>
>>>>>
>>>>> But is PNG really the right choice of format for real-time stuff?
>>>>> Real-time stuff tends to require HW acceleration and PNG is not exactly
>>>>> easy to accelerate with HW.
>>>>>
>>>>> I don't disagree with the use-case, but I don't necessarily think all
>>>>> formats have to support all use-cases.
>>>>>
>>>>> Cheers,
>>>>> //Leo
>>>>>
>>>>>
>>>>>
>>>>>
>>>>> On Mon, Nov 13, 2023 at 12:23 PM Chris Blume (W3C Calendar) <
>>>>> noreply+calendar@w3.org> wrote:
>>>>>
>>>>>> View this event in your browser
>>>>>> <https://www.w3.org/events/meetings/77dd00ce-0351-4282-bab9-bdef33c6d076/>
>>>>>> PNG marketing meeting Upcoming Confirmed
>>>>>>
>>>>>> 27 November 2023, 10:00 -11:00 EST5EDT
>>>>>>
>>>>>> Event is recurring every other week on Monday, starting from
>>>>>> 2023-11-27, until 2024-11-13
>>>>>> Portable Network Graphics (PNG) Working Group
>>>>>> <https://www.w3.org/groups/wg/png/calendar/>
>>>>>>
>>>>>> The primary meeting link ishttps://
>>>>>> w3c.zoom.us/j/85136933003?pwd=SW9ZaFV0UEpydkU2RXNMYnEraE85QT09
>>>>>>   Joining Instructions
>>>>>>
>>>>>> Instructions are restricted to meeting participants. You need to log
>>>>>> in
>>>>>> <https://auth.w3.org/?url=https%3A%2F%2Fwww.w3.org%2Fevents%2Fmeetings%2F77dd00ce-0351-4282-bab9-bdef33c6d076%2F>
>>>>>> to see them.
>>>>>>   Participants Organizers
>>>>>>
>>>>>>    - Chris Blume
>>>>>>    - Stephanie Hurlburt
>>>>>>
>>>>>> Groups
>>>>>>
>>>>>>    - Portable Network Graphics (PNG) Working Group
>>>>>>    <https://www.w3.org/groups/wg/png/> (View Calendar
>>>>>>    <https://www.w3.org/groups/wg/png/calendar/>)
>>>>>>
>>>>>> Report feedback and issues on GitHub
>>>>>> <https://github.com/w3c/calendar>.
>>>>>>
>>>>>
>>>>>