Adding trace functionality

mcdc/adapt.py

@@ -106,11 +106,9 @@ def local_array(shape, dtype):
 
 @numba.extending.type_callable(local_array)
 def type_local_array(context):
-
     from numba.core.typing.npydecl import parse_dtype, parse_shape
 
     if isinstance(context, numba.core.typing.context.Context):
-
         # Function repurposed from Numba's ol_np_empty.
         def typer(shape, dtype):
             numba.np.arrayobj._check_const_str_dtype("empty", dtype)
@@ -148,10 +146,8 @@ def typer(shape, dtype):
         return typer
 
     elif isinstance(context, numba.cuda.target.CUDATypingContext):
-
         # Function repurposed from Numba's Cuda_array_decl.
         def typer(shape, dtype):
-
             # Only integer literals and tuples of integer literals are valid
             # shapes
             if isinstance(shape, types.Integer):
@@ -200,14 +196,12 @@ def typer(shape, dtype):
 
 @numba.extending.lower_builtin(local_array, types.IntegerLiteral, types.Any)
 def builtin_local_array(context, builder, sig, args):
-
     shape, dtype = sig.args
 
     from numba.core.typing.npydecl import parse_dtype, parse_shape
     import numba.np.arrayobj as arrayobj
 
     if isinstance(context, numba.core.cpu.CPUContext):
-
         # No default arguments.
         nb_dtype = parse_dtype(dtype)
         nb_shape = parse_shape(shape)
@@ -420,7 +414,6 @@ def nopython_mode(is_on):
 
 
 def gpu_forward_declare(args):
-
     if args.gpu_rocm_path != None:
         harm.config.set_rocm_path(args.gpu_rocm_path)
 
@@ -651,7 +644,6 @@ def finalize(prog: numba.uintp):
         final_fn(prog)
 
     def step(prog: numba.uintp, arg: arg_type):
-
         step_async()
 
     (step_async,) = harm.RuntimeSpec.async_dispatch(step)

mcdc/config.py

@@ -66,6 +66,13 @@
 parser.add_argument("--clear_cache", action="store_true")
 parser.add_argument("--caching", action="store_true")
 parser.add_argument("--no_caching", dest="caching", action="store_false")
+parser.add_argument("--trace", action="store_true")
+parser.add_argument(
+    "--trace_slot_limit",
+    type=int,
+    help="Maximum number of functions that may be traced.",
+    default=4096,
+)
 parser.add_argument("--runtime_output", default=False, action="store_true")
 parser.set_defaults(caching=False)
 args, unargs = parser.parse_known_args()
@@ -75,6 +82,8 @@
 target = args.target
 caching = args.caching
 clear_cache = args.clear_cache
+trace = args.trace
+trace_slot_limit = args.trace_slot_limit
 
 from mpi4py import MPI
 import shutil