Update tests to accept variants of generated PTX (#585)

Author: gmarkall

numba_cuda/numba/cuda/tests/cudapy/test_casting.py

@@ -255,10 +255,10 @@ def test_float_to_complex(self):
     @skip_on_cudasim("Compilation unsupported in the simulator")
     def test_native_cast(self):
         float32_ptx, _ = cuda.compile_ptx(native_cast, (float32,), device=True)
-        self.assertIn("st.f32", float32_ptx)
+        self.assertRegex(float32_ptx, r"st\.[bf]32")
 
         float16_ptx, _ = cuda.compile_ptx(native_cast, (float16,), device=True)
-        self.assertIn("st.u16", float16_ptx)
+        self.assertRegex(float16_ptx, r"st\.[bu]16", float16_ptx)
 
 
 if __name__ == "__main__":

numba_cuda/numba/cuda/tests/cudapy/test_compiler.py

@@ -661,7 +661,9 @@ def f():
         sig = "void()"
         ptx, resty = cuda.compile_ptx(f, sig, launch_bounds=launch_bounds)
         self.assertIsInstance(resty, types.NoneType)
-        self.assertRegex(ptx, r".maxntid\s+128,\s+1,\s+1")
+        # Match either `.maxntid, 128, 1, 1` or `.maxntid 128` on a line by
+        # itself:
+        self.assertRegex(ptx, r".maxntid\s+128(?:,\s+1,\s+1)?\s*\n")
         return ptx
 
     def test_launch_bounds_scalar(self):

numba_cuda/numba/cuda/tests/cudapy/test_constmem.py

@@ -98,10 +98,10 @@ def test_const_array(self):
         self.assertTrue(np.all(A == CONST1D + 1))
 
         if not ENABLE_CUDASIM:
-            self.assertIn(
-                "ld.const.f64",
+            self.assertRegex(
                 jcuconst.inspect_asm(sig),
-                "as we're adding to it, load as a double",
+                r"ld\.const\.[bf]64",
+                "load as a 64-bit value",
             )
 
     def test_const_empty(self):
@@ -124,10 +124,10 @@ def test_const_array_2d(self):
         self.assertTrue(np.all(A == CONST2D))
 
         if not ENABLE_CUDASIM:
-            self.assertIn(
-                "ld.const.u32",
+            self.assertRegex(
                 jcuconst2d.inspect_asm(sig),
-                "load the ints as ints",
+                r"ld\.const\.[bu]32",
+                "load as a 32-bit value",
             )
 
     def test_const_array_3d(self):
@@ -139,9 +139,9 @@ def test_const_array_3d(self):
 
         if not ENABLE_CUDASIM:
             asm = jcuconst3d.inspect_asm(sig)
-            complex_load = "ld.const.v2.f32"
-            description = "Load the complex as a vector of 2x f32"
-            self.assertIn(complex_load, asm, description)
+            complex_load = r"ld\.const\.v2\.[bf]32"
+            description = "Load the complex as a vector of 2x 32-bits"
+            self.assertRegex(asm, complex_load, description)
 
     def test_const_record_empty(self):
         jcuconstRecEmpty = cuda.jit("void(int64[:])")(cuconstRecEmpty)

numba_cuda/numba/cuda/tests/cudapy/test_dispatcher.py

@@ -773,7 +773,9 @@ def f():
 
         sig = f.signatures[0]
         ptx = f.inspect_asm(sig)
-        self.assertRegex(ptx, r".maxntid\s+128,\s+1,\s+1")
+        # Match either `.maxntid, 128, 1, 1` or `.maxntid 128` on a line by
+        # itself:
+        self.assertRegex(ptx, r".maxntid\s+128(?:,\s+1,\s+1)?\s*\n")
 
         return ptx
 

numba_cuda/numba/cuda/tests/nrt/test_nrt.py

@@ -128,11 +128,11 @@ def g():
         match = re.search(p1, ptx)
         assert match is not None
 
-        p2 = r"call\.uni.*\n.*NRT_incref"
+        p2 = r"call\.uni.*\n?.*NRT_incref"
         match = re.search(p2, ptx)
         assert match is not None
 
-        p3 = r"call\.uni.*\n.*NRT_decref"
+        p3 = r"call\.uni.*\n?.*NRT_decref"
         match = re.search(p3, ptx)
         assert match is not None