Fix 5090 support (#1488)

* Fix 5090 support

* Apply suggestion

* fix

---------

Co-authored-by: Mosè Giordano <[email protected]>

Author: wsmoses

ext/ReactantCUDAExt.jl

@@ -1455,8 +1455,8 @@ end
 
 function __init__()
     if CUDA.functional() && !Reactant.precompiling()
-        target = CUDA._compiler_config(CUDA.device()).target
-        Reactant.Compiler.cubinChip[] = "sm_$(target.cap.major)$(target.cap.minor)"
+	cap =  CUDA.capability(CUDA.device())
+        Reactant.Compiler.cubinChip[] = "sm_$(cap.major)$(cap.minor)"
     end
     return nothing
 end

src/Compiler.jl

@@ -1375,8 +1375,9 @@ function cubinFeatures()
     major, ver = divrem(ver, 1000)
     minor, patch = divrem(ver, 10)
     version = VersionNumber(major, minor, patch)
-    # From https://github.com/llvm/llvm-project/blob/106c483a102e1328f11e2b1d9398f4ad2826b59f/clang/lib/Driver/ToolChains/Cuda.cpp#L685
+    # From https://github.com/llvm/llvm-project/blob/b60aed6fbabc291a7afbcb460453f9dcdce76f34/clang/lib/Driver/ToolChains/Cuda.cpp#L686
     cuver_map = Dict([
+        (128, 87),
         (126, 85),
         (125, 85),
         (124, 84),
@@ -1401,7 +1402,7 @@ function cubinFeatures()
         (90, 60),
     ])
     mver = major * 10 + minor
-    if mver > 126
+    if !in(mver, keys(cuver_map))
         return 86
     end
     ptx = cuver_map[mver]

test/ops.jl

@@ -23,7 +23,7 @@ end
 @testset "add" begin
     a = Reactant.to_rarray([false, false, true, true])
     b = Reactant.to_rarray([false, true, false, true])
-    @test [false, true, true, false] ≈ @jit Ops.add(a, b)
+    @test [false, true, true, true] ≈ @jit Ops.add(a, b)
 
     a = Reactant.to_rarray([1, 2, 3, 4])
     b = Reactant.to_rarray([5, 6, -7, -8])