Optimal Texture Transfer Formats and GL_BGRA

OpenGL OpenGL: Advanced Coding
1

On Windows, on an NVidia GPU, I’m seeing:


glGetInternalformativ( GL_TEXTURE_2D, GL_RGBA8, GL_TEXTURE_IMAGE_FORMAT     , 1, &format[0] );  // GL_RGBA
glGetInternalformativ( GL_TEXTURE_2D, GL_RGBA8, GL_TEXTURE_IMAGE_TYPE       , 1, &type  [0] );  // GL_UNSIGNED_INT_8_8_8_8_REV
glGetInternalformativ( GL_TEXTURE_2D, GL_RGBA8, GL_READ_PIXELS_FORMAT       , 1, &format[1] );  // GL_RGBA
glGetInternalformativ( GL_TEXTURE_2D, GL_RGBA8, GL_READ_PIXELS_TYPE         , 1, &type  [1] );  // GL_UNSIGNED_INT_8_8_8_8_REV

That is, for GL_RGBA8, the driver recommends uploads and downloads using GL_RGBA / GL_UNSIGNED_INT_8_8_8_8_REV.

Does that makes sense to anyone?

I recall repeatedly folks benchmarking and finding that GL_BGRA is the external format to use on Windows (either with GL_UNSIGNED_BYTE or possibly GL_UNSIGNED_INT_8_8_8_8_REV (for instance, in mhagain’s post here). Not to mention that NVidia recommends this in their developer literature too.

Correct me if I’m wrong, but the NVidia driver’s recommendation (via glGetInternalformativ()) doesn’t match that:


ORDER     FORMAT    TYPE
-------- ---------- -----------
RGBA      RGBA      UNSIGNED_BYTE
RGBA      RGBA      UINT_8_8_8_8_REV   <-- NV driver recommends
BGRA      BGRA      UNSIGNED_BYTE      <-- Often recommended
BGRA      BGRA      UINT_8_8_8_8_REV   <-- Sometimes recommended
ABGR      RGBA      UINT_8_8_8_8
ARGB      BGRA      UINT_8_8_8_8

Note here that “ORDER” is the memory order on a little-endian machine like x86 (with GL_{PACK,UNPACK}_SWAP_BYTES = FALSE).

1 Like
2

I believe, and this is without having done any formal testing to back it up, that ever since Microsoft shoved RGBA down the vendors throats in the D3D10 timeframe this is not as important as it used to be.

It’s certainly the case however that in my old tests BGRA was primarily optimal on Intel, significantly less so on NV whereas on AMD it didn’t matter at all.

3

Thanks for the reply. That’s a better theory than I had. In this test case, I’m not seeing any better perf with BGRA order than with RGBA order (in fact, RGBA is better).