Initialize CUDA target lazily (#541)

terralang · Apr 22, 2022 · 868748e · 868748e
1 parent 43c883c
commit 868748e
Show file tree

Hide file tree

Showing 2 changed files with 7 additions and 3 deletions.
diff --git a/src/cudalib.lua b/src/cudalib.lua
@@ -29,7 +29,7 @@ terralib.CUDAParams.entries = { { "gridDimX", uint },
 function cudalib.toptx(module,dumpmodule,version,profile)
     dumpmodule,version = not not dumpmodule,assert(tonumber(version))
     profile = profile or {fastmath=false}
-    local cu = terralib.newcompilationunit(terra.cudatarget, false, profile) -- TODO: add nvptx target options here
+    local cu = terralib.newcompilationunit(terra.getcudatarget(), false, profile) -- TODO: add nvptx target options here
     local annotations = terra.newlist{} -- list of annotations { functionname, annotationname, annotationvalue } to be tagged
     local function addkernel(k,fn)
         fn:setinlined(true)

diff --git a/src/terralib.lua b/src/terralib.lua
@@ -4239,8 +4239,12 @@ if terra.cudalibpaths and terra.cudahome then
 	end
 end                       
 
-if terra.cudahome then
-  terra.cudatarget = terra.newtarget {Triple = 'nvptx64-nvidia-cuda', FloatABIHard = true}
+local cudatarget
+function terra.getcudatarget()
+    if cudatarget == nil then
+        cudatarget = terra.newtarget {Triple = 'nvptx64-nvidia-cuda', FloatABIHard = true}
+    end
+    return cudatarget
 end