If we're making slots part of the device API, then I think it makes sense to not only do it for bindless CUDA textures, but for CPU textures too, which are now decoding the slot from the name string.

Doesn't have to be done as part of this patch or even soon, but would make the code nicer I think.

Not sure whats wrong here?

This is giving me a compile error.

error: invalid conversion from ‘int’ to ‘CUresourcetype {aka CUresourcetype_enum}’ [-fpermissive]

We don't need two arrays, we can have just a single device_vector<uint> bindless_mapping;.

And then to add slots:

if (flat_slot >= bindless_mapping.size())
    bindless_mapping.resize(max(bindless_mapping.size() * 2, 1));

bindless_mapping[flat_slot] = tex;

Push/pops don't seem to be matched anymore, or at least the logic is hard the follow. Better to do it more often than have confusing logic.

This architecture check is unnecessary as we should never have sync_bindless_mapping = true with other architectures.

And the copying code can be simplified if we have a single array.

Also, this code should be moved to a separate function and called also from shader(), since baking can access textures too.

This should only be added for if needed, and will also not require the number of image textures to be reduced on other cards.

The removal of the 3 slots here is fine, the /* Kepler and above */ comment was outdated. Atm we have 5 float4 and 88 byte4 textures on Fermi, so 93 slots in total, starting with index 0 we are then at 092. :)

This is effectively way to a memory leak, you modified tex_free() signature here, which is now different from base class, meaning this code will never run.

Eventually we can merge this two maps into one like this:

struct MemInfo {
   bool is_array_data;
   CUtexObject bindless_texture;
};
map<device_ptr, MemInfo> tex_meminfo_map;

This can be done later. For now i would only strongly suggest using CUtexObject instead of uint for non-kernel code.

Use need_ prefix to distinguish flag from method?

Will it work correct if the texture was never allocated?

Use parenthesis btw.

This we should add a check that tex fits into uint and print a warning or abort rendering otherwise. CUtexObject is unsigned long long after all which is not necessarily gonna to fit into a smaller type.

Screw it then for now.

Shall we do a call from thread_run() instead? path_trace() is run for each sample, and trying to load bindless mapping for each of them is a bit redundant.

Yes, because in that case there is no valid mem.device_pointer and the function will do nothing.