Use int for int8x4 due to performance overhead of char4 (#1569)

* Use int for int8x4 due to performance overhead of char4

* Add a comment about using int

* Remove invalid test

src/codegen/codegen_cuda.cc

@@ -90,7 +90,11 @@ void CodeGenCUDA::PrintType(Type t, std::ostream& os) {  // NOLINT(*)
         if (t.lanes() == 4) {
           // directly 4 8 bit int in integer.
           enable_int8_ = true;
-          os << "char4"; return;
+
+          // We use int for int8x4 instead of char4 because using char4 is
+          // likely to produce extra instructions to pack four int8 elements
+          // into 32-bit data.
+          os << "int"; return;
         } else if (t.lanes() == 8) {
           enable_int8_ = true;
           os << "int2"; return;

tests/python/unittest/test_codegen_cuda.py

@@ -31,7 +31,6 @@ def test_cuda_vectorize_add():
         
     check_cuda("float32", 64, 2)
     check_cuda("float16", 64, 2)
-    check_cuda("int8", 64, 4)
 
 
 def test_cuda_multiply_add():