Support CUDA 11 + linux fixes

[paulmelis]

This adds support for CUDA 11, but still checks for CUDA 12 in the main cmake config. The few places in the CUDA code where CUDA12-specific features are used are behind #include guards.

The patches in the patches/ dir need to be applied to the recursively checked out repository, as we need to patches to the glad, tiny-cuda-nn and pybind11 source repos in include/.

~~The first is needed as a shared library (and so the Python module) on Linux requires the code in the tiny-cuda-nn library to be built as position-independent code (-fPIC). I did not see the required options to enable this in the upstream code, hence the patches like this. ~~

The patch to pybind11 is needed when building against CUDA12 to work around pybind/pybind11#4606

Finally, I had an error related to the use of #pragma once, which I worked around using include guards.

[JamesPerlman]

instead of doing this, i'm curious if removing this line from main.cu has the same effect?

#include "core/occupancy-grid.cuh"

[paulmelis]

Just tried, it doesn't get rid of the error:

[ 36%] Building CUDA object CMakeFiles/PyTurboNeRF.dir/src/main.cu.o
/home/paulm/c/TurboNeRF/src/utils/bit-utils.cu:1:9: error: #pragma once in main file [-Werror]
    1 | #pragma once
      |         ^~~~
/home/paulm/c/TurboNeRF/src/utils/bit-utils.cu:1:9: warning: #pragma once in main file
    1 | #pragma once
      |         ^~~~

To be sure, here's what I changed for the test:

snellius paulm@gcn65 20:58 ~/c/TurboNeRF/build$ git diff
diff --git a/src/main.cu b/src/main.cu
index 6caa331..ab6dcdb 100644
--- a/src/main.cu
+++ b/src/main.cu
@@ -11,7 +11,7 @@
 #include "main.h"
 #include "controllers/nerf-training-controller.h"
 #include "controllers/nerf-rendering-controller.h"
-#include "core/occupancy-grid.cuh"
+//#include "core/occupancy-grid.cuh"
 #include "models/camera.cuh"
 #include "models/dataset.h"
 #include "models/render-request.cuh"
@@ -215,4 +215,4 @@ bool check_is_json_file(const std::string &path_str) {
                std::cout << fmt::format("Fileformat not JSON : {}", path_str);
                return false;
        }
-}
\ No newline at end of file
+}
diff --git a/src/utils/bit-utils.cu b/src/utils/bit-utils.cu
index 4108a79..29f41e1 100644
--- a/src/utils/bit-utils.cu
+++ b/src/utils/bit-utils.cu
@@ -1,7 +1,4 @@
-//#pragma once
-// error: #pragma once in main file
-#ifndef TURBONERF_BIT_UTILS_CU
-#define TURBONERF_BIT_UTILS_CU
+#pragma once
 
 #include "bit-utils.cuh"
 
@@ -25,4 +22,3 @@ __global__ void get_1s_per_uint32(
 
 TURBO_NAMESPACE_END
 
-#endif

[JamesPerlman]

What about also removing the include for occupancy-grid from nerf-training-controller.cuh?

(you'll need to add it to nerf-training-controller.cu the implementation file)

I just don't like that only one file uses header guards while every other file uses #pragma once

[JamesPerlman]

LGTM. @paulmelis are you ready for me to merge this?

[paulmelis]

Yep, all done!

[paulmelis]

Found a small error in the pybind11 patch while trying to apply it on a different workstation. Fixed now.