Merge pull request #272 from RobJY/cuda_build_fixes

Cuda build fixes

.gitignore

@@ -4,6 +4,7 @@
 .Ruserdata
 *.Rproj
 .*.swp
+tests/gpu/misc/hosts
 
 tests/gpu/misc/hosts
 

tests/gpu/misc/build-fix-notes.md

@@ -0,0 +1,42 @@
+# Build combination notes
+
+## Build combination summary
+
+The following combinations of toolkit and tensorflow versions were tested for Nvidia driver versions 460 and 470 and had the same results.
+
+| tensorflow version / toolkit version | 10 | 11 |
+| ----------------------------------- | -- | -- |
+| 1.15 | succeeds | succeeds |
+| 2.2 | succeeds | fails |
+| 2.5 | fails | succeeds |
+| 2.6 | fails | succeeds |
+
+## Failed build combination notes
+
+* Driver 470, toolkit 11, tensorflow 2.2:
+  * Tensorflow 2.2 seems to require toolkit 10.
+  * I was able to load the required libraries in version 10 with apt. They are:
+    * libcudart10
+    * libcusolver10
+    * libcublas10
+    * libcusparse10
+  * Unfortunately, it also requires cudnn7 which isn't available from apt.
+    * maybe a repo can be added that has cudnn7
+    * other option is to build from download from NVIDIA, but I don't know how this would work within the build
+  * unsupported; tensorflow 2.2 seems to require toolkit 10
+* Driver 470, toolkit 10, tensorflow 2.5:
+  * fails on missing libraries:
+    * libcudart11
+    * libcublas11
+    * libcublasLt11
+    * libcusolver11
+    * libcusparse11
+  * unsupported; tensorflow 2.5 seems to require toolkit 11
+* Driver 470, toolkit 10, tensorflow 2.6:
+  * fails on missing libraries:
+    * libcudart11
+    * libcublas11
+    * libcublasLt11
+    * libcusolver11
+    * libcusparse11
+  * unsupported; tensorflow 2.6 seems to require toolkit 11