Compiling gridencoder on windows with cuda 12.5

[Duodecimus]

this was a chore, but I got it working. compiled, at least.

first, setup some environment variables.
specifically, you need CUDA_ARCH to exist. you'll need to look up exactly what the value should be for your video card. my Nvidia 3050 is 'sm_86'
Some others that you might need are TCNN_CUDA_ARCHITECTURES which depends on your card again, mine is '86'
and also CUDA_HOME which should be the path to your cuda install

Start with downloading the repo, then navigate to torch-ngp\gridencoder and run python setup.py install

This will almost certainly fail. you'll probably get a bunch of errors about atomicAdd.
we needed to do this to generate a build.ninja file. it would be somewhere like \torch-ngp\gridencoder\build\temp.win-amd64-cpython-312\Release

edit the file.
you need to double check all the paths you see in here. mine failed to have correct paths for cuda and visual studio. there is also a reference to std=c++. make sure that you change this to a version you have installed. I used c++17
save the file

now open gridencoder.cu
near the top, around like 20 you'll see a definition for atomicadd

now, depending on the errors you get you'll need to be creative here, but basically we need to write a function to convert what is getting put into atomic add into what cuda 12.5 is expecting.
My errors looked like this:
function "atomicAdd(int *, int)" does not match because argument #1 does not match parameter
static __inline __declspec(__device__) int atomicAdd(int *address, int val)

so my functions ended up looking like this

__device__ inline at::Half atomicAdd(at::Half *address, at::Half val) {
return atomicAdd(address, val);
}

template <typename T>
__device__ inline T atomicAdd(T *address, T val) {
return atomicAdd(address, val);
}

now open a cmd window next to the build.ninja file and run ninja -f build.ninja

and then it compiled. yay

[Duodecimus]

you may also need to edit setup.py to have the correct c++ versions and visual studio paths