texSubImage2D with sRGB texture on Metal

[amshv]

Hi,

I am using texStorage2D to create textures with internalformat SRGB8_ALPHA8 and then calling texSubImage2D with format RGBA and type UNSIGNED_BYTE. On the Metal debugger I noticed that there are several replaceRegion calls apparently for a row by row format conversion.

I put some print statements where this is happening, here https://github.com/google/angle/blob/91d2d125ec0054f970594b29660f1cf3d172a8f1/src/libANGLE/renderer/metal/TextureMtl.mm#L2324 it's trying to convert to R8G8B8A8_UNORM_SRGB from R8G8B8A8_UNORM which it gets from the format and type. But R8G8B8A8_UNORM_SRGB also corresponds to the same format and type. I can work around this by changing the internal format to RGBA8 and decoding in the fragment shader but I'm wondering if the conversion could be skipped in this case.

Thanks.

[Shahbaz Youssefi]

Hi,

This could be a bug, but I'm looking at the `SRGBTextureTestES3.SRGBFormatCombinationValidation` test, and it looks like creating a texture with `GL_SRGB_ALPHA_EXT` and uploading to it with GL_RGBA+GL_UNSIGNED_BYTE is tested and apparently passing on Metal (test uploads sRGB, expects sampling to return linear). I don't see the sized GL_SRGB8_ALPHA8 format tested with GL_RGBA though; the test is using GL_SRGB_ALPHA_EXT+GL_UNSIGNED_BYTE instead. Perhaps GL_SRGB_ALPHA_EXT is what you need to use in GL_RGBA in glTexSubImage2D?

[amshv]

Thanks for looking into it. For me the values are correct both ways, ie I also get linear values in the shader as expected. However the row conversion loop is taking 5-10 ms for a 512x512 texture so I assume it's copying the data unchanged. I don't know if it's a bug necessarily rather than a performance thing, or if it's something with the way this data has to be stored.

[Shahbaz Youssefi]

Right, I checked with an expert and using GL_RGBA is fine, and this is just a matter of the metal backend not implementing the conversion efficiently.

[amshv]

I see, thank you. I opened https://issues.angleproject.org/issues/542677296 in case this is worth tracking.