[CODEGEN NVPTX] Initialize llvm before lookup for the target

[denis0x0D]

LLVMTargetInfo, TargetMachine and so on should be initialized before lookup for the specific target.

 n = tvm.var ("n")
 A = tvm.placeholder ((n ,n), name='A', dtype="float32")
 B = tvm.placeholder ((n, n), name='B', dtype="float32")
 C = tvm.compute (A.shape, lambda *i: A(*i) + B(*i), name='C')
 s = tvm.create_schedule (C.op)

 bx, tx = s[C].split (C.op.axis[0], factor=64)
 s[C].bind (bx, tvm.thread_axis("blockIdx.x"))
 s[C].bind (tx, tvm.thread_axis("threadIdx.x"))
 module = tvm.build(s, [A, B, C], "nvptx", target_host="llvm")

This code fails with error:
Unable to find target for this triple (no targets are registered) target_triple=nvptx64_nvidia_cuda

[masahi]

can you also update AMDGPU codegen here? I remember being puzzled by this error when target_host="llvm" is explicitly passed.

[denis0x0D]

Ok, I see the same problem for AMD GPU codegen.

[masahi]

thanks @denis0x0D this is merged.