GLhalf [ 8 ] -> glVertexAttribPointer?

[Dark_Photon]

I know this isn’t explicitly supported, but do we have the tools to pass a GUhalf[8] in a single 4-component vertex attribute despite the API?

glVertexAttribPointer size <= 4, so the nieve approach doesn’t work.

But what about packing 2 halfs in a single float (or uint)?
Question is, once you get them in the shader, can you shift/mask off 16-bits of a float/uint into a half nowadays?

(ARB_shader_bit_encoding only provides for 32-bit float<->int AFAICT)

[imported_Groovounet]

Do you want to use half float like any other type?

In this case I would say no without a hand written conversion from this raw half data (passed as int without normalization) to float with the bits operations and functions. With GLSL 4.0 and the pretty extract*** functions it should be easy. It should be possible with GLSL 3.3 with & | << to actually write these extract functions.

Maybe an explicit support of half in GLSL would be nice. We have double, so why not?

[Dark_Photon]

@Groovounet

Do you want to use half float like any other type?

Yeah, the goal being just to save vertex attributes.

I could grab the right bits using bitfieldExtract, but then how to cast that to a half?

GLSL used to be pretty limited in casting capability. Is this supported now? Would you literally just do:

half( bitfieldExtract( packed_attrib, 0, 16 ) )

I suspect not, as it’ll interpret the half operand as a genuine integer, not an encoded half. In C++ you might have done something like:

*(half *)( &half_in_unsigned_short )

…but this is GLSL.

[RGHP]

I use the half type that comes with OpenEXR, works perfectly. Yo don’t need to embed all the OpenEXR library, just two files.

Bye!

[imported_Groovounet]

What I meant is a by “hand cast”.

Extracting the bits (sign, mentissa, exponent) from the raw data, manipulating the value with shiftings (and maybe scaling?) and assembling the 3 bits fields in a int to finally use the extension you mention to get a float. It might not that sample actually as for example the half to float cast in OpenEXR is quite big relatively to a shader program at least.

Beside saving attribute slot for maybe some heavy attribute based instancing, I not sure it could really be useful. Maybe… I love to hear the use cases!

[Dark_Photon]

@Groovounet

What I meant is a by “hand cast”.

Extracting the bits (sign, mentissa, exponent) from the raw data, manipulating the value with shiftings (and maybe scaling?) and assembling the 3 bits fields in a int to finally use the extension you mention to get a float.

Ah, I see what you’re seeing. Could work. Was just hoping there was a simpler way.

Thanks.