Ok I think this should be standarized in OGL:
*a comprehensive spec exposing standard way to query support and expose ability to access host mem (buffers textures) from GPU and/or viceversa unifying INTEL_map_texture AMD_pinned_memory and abusing stream flags NV tricks for ensuring pinned mem..
*portable binaries like OCL is going to expose with SPIR like D3D bytecode..
*a state object extension like NV seems is doing should ARBfied exposing state through objects like D3D does..
*a way to access from one GPU mem from other GPU seems supported on Fermi and up (exposed on CUDA p2p ext) and also on GCN and up at least (AMD_bus_addresable_mem)
(altough will be useful on consumer GPUs only on Linux or Windows and ATI as on Windows and NV you can't control individual GPUs on consumer cards)
*by last limitation standarize AMD_gpu_association and WGL_NV_affinity and as byproduct exposing to consumer cards?
*in sampler_objects ext we get:

Should sampler objects be made visible to the shading language. This is left for a future extension

It's time to expose this!
for OGL 5:
standarize and expose pointers in GLSL like NV has done for some years now that AMD GCN support seem to support it in HW also Intel for Gen8 next year should have it?
NV seems to be working on a bindless_multidraw_indirect using bindless features
Thanks,