autodock-gpu: build with the specified cuda_arch (#42244)

The CUDA target should be specified at build time, otherwise
by default `autodock-gpu` will be built for compute capabilities
52, 60, 61, 70, which may cause errors on unsopported cards.

var/spack/repos/builtin/packages/autodock-gpu/package.py

@@ -6,7 +6,7 @@
 from spack.package import *
 
 
-class AutodockGpu(MakefilePackage):
+class AutodockGpu(MakefilePackage, CudaPackage):
     """AutoDock-GPU: AutoDock for GPUs and other accelerators.
     OpenCL and Cuda accelerated version of AutoDock 4.2.6. It
     leverages its embarrasingly parallelizable LGA by processing
@@ -30,14 +30,19 @@ class AutodockGpu(MakefilePackage):
         multi=False,
     )
     variant("overlap", default=False, description="Overlap CPU and GPU operations")
+    variant("cuda", default=True, description="Build with CUDA")
 
     depends_on("cuda")
 
+    conflicts("~cuda")  # the cuda variant is mandatory
+    conflicts("+cuda", when="cuda_arch=none")
+
     @property
     def build_targets(self):
         spec = self.spec
         return [
             "DEVICE={0}".format(spec.variants["device"].value.upper()),
+            "TARGETS={0}".format(" ".join(spec.variants["cuda_arch"].value)),
             "GPU_INCLUDE_PATH={0}".format(spec["cuda"].prefix.include),
             "GPU_LIBRARY_PATH={0}".format(spec["cuda"].libs.directories[0]),
             "OVERLAP={0}".format("ON" if "+overlap" in spec else "OFF"),