Incorrect decode of 32 bit DWA encoded EXR

[slipperyseal]

I created 16 and 32 bit DWA encoded EXR images from the same source image (from Blender).

16 bit DWA files decode fine in ffmpeg while there seems to be an issue with 32 bit DWA.

Looking at the code I can see when decoding DWA EXR files, ffmpeg runs 16 bit or 32 bit decoders but always targets a 32bit space. There is common code which then transforms that intermediate output to linear space.

libavcodec/exr.c

On line 1122 is a scaling factor for the linear space conversion.

const float scale = s->pixel_type == EXR_FLOAT ? 2.f : 1.f;

When pixels are EXR_FLOAT (32 bit) the scaling factor is set to 2.f which produces underexposed images.
If I set the scaling factor to always be 1.0, both file conversions appear correct and "identical".

const float scale = 1.f;

With this change applied, I logged pixel values (prior to apply_trc) and saw the results almost identical (probably because of float size precision).

Perhaps this scaling factor was originally required at one time but now is compensated for in the 32 bit decoder?

I havn't been able to locate any DWA files other than those I generate with Blender, but I can confirm that when importing both types into DaVinci Resolve, they appear identical. Therefor I don't think the 32 bit file is at fault.

Note that a RGB colour space conversion also needs to be applied when converting EXRs.

# Convert dwaa_16_0001.exr to png

ffmpeg -y -start_number 1 -f image2 -apply_trc iec61966_2_1 -i dwaa_16_%04d.exr dwaa_16_%04d.png

# Convert dwaa_32_0001.exr to png

ffmpeg -y -start_number 1 -f image2 -apply_trc iec61966_2_1 -i dwaa_32_%04d.exr dwaa_32_%04d.png

[slipperyseal]

16 bit DWA EXR file for testing

[slipperyseal]

32 bit DWA EXR file for testing

[Elon Musk]

Fixed in 024c30aa3b3cdb7d32fa5d79b0973ab8f3065ead