Support for older GPUs

[kagalenko-m-b]

At present, there is no way to install CUDA for older GPU (compute capacity 3.0) on either current or LTS julia versions. The latest supported CUDA version for cc 3.0 is 4.0, but it requires GPUCompiler that is compatible only with julia version no greater than 1.6

[maleadt]

As explained in JuliaGPU/GPUCompiler.jl#675 (comment), this is not going to happen. I'd be happy to take PRs for additional releases of older CUDA.jl versions, but since NVIDIA doesn't support these GPUs anymore, we are not going to re-instate support for a 12 year old device.

[kagalenko-m-b]

I'd be happy to take PRs for additional releases of older CUDA.jl versions

How extensive are the necessary changes?

[maleadt]

I don't remember by heart. Have a look at the changes in GPUCompiler, but if anything, the compathelper PRs here looked very straightforward: #2462 and #2521.

[kagalenko-m-b]

I don't remember by heart. Have a look at the changes in GPUCompiler, but if anything, the compathelper PRs here looked very straightforward: #2462 and #2521.

I have tried to change the branch release-4.0 :

diff --git i/Project.toml w/Project.toml
index bc534da3b..f79a68f3c 100644
--- i/Project.toml
+++ w/Project.toml
@@ -39,8 +39,8 @@ CUDA_Runtime_Discovery = "0.1"
 CUDA_Runtime_jll = "0.2.3"
 ExprTools = "0.1"
 GPUArrays = "8.6"
-GPUCompiler = "0.17"
-LLVM = "4.15"
+GPUCompiler = "0.17, 0.24, 0.25, 0.26, 0.27"
+LLVM = "4,5,6,7,8,9"
 Preferences = "1"
 Random123 = "1.2"
 RandomNumbers = "1.5.3"

but for some reason precompiling CUDA.jl errors within GPUCompiler.__init__() . Specifically, Base.pkgversion(GPUCompiler) returns nothing

[maleadt]

You have to actually use the latest GPUCompiler, not simply bump the compat (to include JuliaGPU/GPUCompiler.jl#594).

[kagalenko-m-b]

You have to actually use the latest GPUCompiler, not simply bump the compat (to include JuliaGPU/GPUCompiler.jl#594).

This patch allows release-4.0 to precompile on Julia 1.11.4 , but tests fail to start. I tried to
fix them, but this seems to be getting into the internals.